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/eventhandler.h>
119 #include <sys/limits.h>
121 #include <sys/kernel.h>
122 #include <sys/malloc.h>
123 #include <sys/sysctl.h>
124 #include <sys/lock.h>
125 #include <sys/mutex.h>
126 #include <sys/rwlock.h>
128 #include <sys/dtrace_bsd.h>
129 #include <netinet/in.h>
130 #include "dtrace_cddl.h"
131 #include "dtrace_debug.c"
135 * DTrace Tunable Variables
137 * The following variables may be tuned by adding a line to /etc/system that
138 * includes both the name of the DTrace module ("dtrace") and the name of the
139 * variable. For example:
141 * set dtrace:dtrace_destructive_disallow = 1
143 * In general, the only variables that one should be tuning this way are those
144 * that affect system-wide DTrace behavior, and for which the default behavior
145 * is undesirable. Most of these variables are tunable on a per-consumer
146 * basis using DTrace options, and need not be tuned on a system-wide basis.
147 * When tuning these variables, avoid pathological values; while some attempt
148 * is made to verify the integrity of these variables, they are not considered
149 * part of the supported interface to DTrace, and they are therefore not
150 * checked comprehensively. Further, these variables should not be tuned
151 * dynamically via "mdb -kw" or other means; they should only be tuned via
154 int dtrace_destructive_disallow = 0;
155 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
156 size_t dtrace_difo_maxsize = (256 * 1024);
157 dtrace_optval_t dtrace_dof_maxsize = (256 * 1024);
158 size_t dtrace_global_maxsize = (16 * 1024);
159 size_t dtrace_actions_max = (16 * 1024);
160 size_t dtrace_retain_max = 1024;
161 dtrace_optval_t dtrace_helper_actions_max = 128;
162 dtrace_optval_t dtrace_helper_providers_max = 32;
163 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
164 size_t dtrace_strsize_default = 256;
165 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
166 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
167 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
168 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
169 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
170 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
171 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
172 dtrace_optval_t dtrace_nspec_default = 1;
173 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
174 dtrace_optval_t dtrace_stackframes_default = 20;
175 dtrace_optval_t dtrace_ustackframes_default = 20;
176 dtrace_optval_t dtrace_jstackframes_default = 50;
177 dtrace_optval_t dtrace_jstackstrsize_default = 512;
178 int dtrace_msgdsize_max = 128;
179 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
180 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
181 int dtrace_devdepth_max = 32;
182 int dtrace_err_verbose;
183 hrtime_t dtrace_deadman_interval = NANOSEC;
184 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
185 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
186 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
189 * DTrace External Variables
191 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
192 * available to DTrace consumers via the backtick (`) syntax. One of these,
193 * dtrace_zero, is made deliberately so: it is provided as a source of
194 * well-known, zero-filled memory. While this variable is not documented,
195 * it is used by some translators as an implementation detail.
197 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
200 * DTrace Internal Variables
203 static dev_info_t *dtrace_devi; /* device info */
206 static vmem_t *dtrace_arena; /* probe ID arena */
207 static vmem_t *dtrace_minor; /* minor number arena */
209 static taskq_t *dtrace_taskq; /* task queue */
210 static struct unrhdr *dtrace_arena; /* Probe ID number. */
212 static dtrace_probe_t **dtrace_probes; /* array of all probes */
213 static int dtrace_nprobes; /* number of probes */
214 static dtrace_provider_t *dtrace_provider; /* provider list */
215 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
216 static int dtrace_opens; /* number of opens */
217 static int dtrace_helpers; /* number of helpers */
219 static void *dtrace_softstate; /* softstate pointer */
221 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
222 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
223 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
224 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
225 static int dtrace_toxranges; /* number of toxic ranges */
226 static int dtrace_toxranges_max; /* size of toxic range array */
227 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
228 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
229 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
230 static kthread_t *dtrace_panicked; /* panicking thread */
231 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
232 static dtrace_genid_t dtrace_probegen; /* current probe generation */
233 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
234 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
235 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
237 static struct mtx dtrace_unr_mtx;
238 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
239 int dtrace_in_probe; /* non-zero if executing a probe */
240 #if defined(__i386__) || defined(__amd64__)
241 uintptr_t dtrace_in_probe_addr; /* Address of invop when already in probe */
243 static eventhandler_tag dtrace_kld_load_tag;
244 static eventhandler_tag dtrace_kld_unload_tag;
249 * DTrace is protected by three (relatively coarse-grained) locks:
251 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
252 * including enabling state, probes, ECBs, consumer state, helper state,
253 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
254 * probe context is lock-free -- synchronization is handled via the
255 * dtrace_sync() cross call mechanism.
257 * (2) dtrace_provider_lock is required when manipulating provider state, or
258 * when provider state must be held constant.
260 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
261 * when meta provider state must be held constant.
263 * The lock ordering between these three locks is dtrace_meta_lock before
264 * dtrace_provider_lock before dtrace_lock. (In particular, there are
265 * several places where dtrace_provider_lock is held by the framework as it
266 * calls into the providers -- which then call back into the framework,
267 * grabbing dtrace_lock.)
269 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
270 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
271 * role as a coarse-grained lock; it is acquired before both of these locks.
272 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
273 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
274 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
275 * acquired _between_ dtrace_provider_lock and dtrace_lock.
277 static kmutex_t dtrace_lock; /* probe state lock */
278 static kmutex_t dtrace_provider_lock; /* provider state lock */
279 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
282 /* XXX FreeBSD hacks. */
283 static kmutex_t mod_lock;
285 #define cr_suid cr_svuid
286 #define cr_sgid cr_svgid
287 #define ipaddr_t in_addr_t
288 #define mod_modname pathname
289 #define vuprintf vprintf
290 #define ttoproc(_a) ((_a)->td_proc)
291 #define crgetzoneid(_a) 0
294 #define CPU_ON_INTR(_a) 0
296 #define PRIV_EFFECTIVE (1 << 0)
297 #define PRIV_DTRACE_KERNEL (1 << 1)
298 #define PRIV_DTRACE_PROC (1 << 2)
299 #define PRIV_DTRACE_USER (1 << 3)
300 #define PRIV_PROC_OWNER (1 << 4)
301 #define PRIV_PROC_ZONE (1 << 5)
304 SYSCTL_NODE(_debug, OID_AUTO, dtrace, CTLFLAG_RD, 0, "DTrace Information");
308 #define curcpu CPU->cpu_id
313 * DTrace Provider Variables
315 * These are the variables relating to DTrace as a provider (that is, the
316 * provider of the BEGIN, END, and ERROR probes).
318 static dtrace_pattr_t dtrace_provider_attr = {
319 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
320 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
321 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
322 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
323 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
330 static dtrace_pops_t dtrace_provider_ops = {
331 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
332 (void (*)(void *, modctl_t *))dtrace_nullop,
333 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
334 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
335 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
336 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
340 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
343 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
344 static dtrace_id_t dtrace_probeid_end; /* special END probe */
345 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
348 * DTrace Helper Tracing Variables
350 uint32_t dtrace_helptrace_next = 0;
351 uint32_t dtrace_helptrace_nlocals;
352 char *dtrace_helptrace_buffer;
353 int dtrace_helptrace_bufsize = 512 * 1024;
356 int dtrace_helptrace_enabled = 1;
358 int dtrace_helptrace_enabled = 0;
362 * DTrace Error Hashing
364 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
365 * table. This is very useful for checking coverage of tests that are
366 * expected to induce DIF or DOF processing errors, and may be useful for
367 * debugging problems in the DIF code generator or in DOF generation . The
368 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
371 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
372 static const char *dtrace_errlast;
373 static kthread_t *dtrace_errthread;
374 static kmutex_t dtrace_errlock;
378 * DTrace Macros and Constants
380 * These are various macros that are useful in various spots in the
381 * implementation, along with a few random constants that have no meaning
382 * outside of the implementation. There is no real structure to this cpp
383 * mishmash -- but is there ever?
385 #define DTRACE_HASHSTR(hash, probe) \
386 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
388 #define DTRACE_HASHNEXT(hash, probe) \
389 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
391 #define DTRACE_HASHPREV(hash, probe) \
392 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
394 #define DTRACE_HASHEQ(hash, lhs, rhs) \
395 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
396 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
398 #define DTRACE_AGGHASHSIZE_SLEW 17
400 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
403 * The key for a thread-local variable consists of the lower 61 bits of the
404 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
405 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
406 * equal to a variable identifier. This is necessary (but not sufficient) to
407 * assure that global associative arrays never collide with thread-local
408 * variables. To guarantee that they cannot collide, we must also define the
409 * order for keying dynamic variables. That order is:
411 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
413 * Because the variable-key and the tls-key are in orthogonal spaces, there is
414 * no way for a global variable key signature to match a thread-local key
418 #define DTRACE_TLS_THRKEY(where) { \
420 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
421 for (; actv; actv >>= 1) \
423 ASSERT(intr < (1 << 3)); \
424 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
425 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
428 #define DTRACE_TLS_THRKEY(where) { \
429 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
431 uint_t actv = _c->cpu_intr_actv; \
432 for (; actv; actv >>= 1) \
434 ASSERT(intr < (1 << 3)); \
435 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
436 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
440 #define DT_BSWAP_8(x) ((x) & 0xff)
441 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
442 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
443 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
445 #define DT_MASK_LO 0x00000000FFFFFFFFULL
447 #define DTRACE_STORE(type, tomax, offset, what) \
448 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
451 #define DTRACE_ALIGNCHECK(addr, size, flags) \
452 if (addr & (size - 1)) { \
453 *flags |= CPU_DTRACE_BADALIGN; \
454 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
458 #define DTRACE_ALIGNCHECK(addr, size, flags)
462 * Test whether a range of memory starting at testaddr of size testsz falls
463 * within the range of memory described by addr, sz. We take care to avoid
464 * problems with overflow and underflow of the unsigned quantities, and
465 * disallow all negative sizes. Ranges of size 0 are allowed.
467 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
468 ((testaddr) - (baseaddr) < (basesz) && \
469 (testaddr) + (testsz) - (baseaddr) <= (basesz) && \
470 (testaddr) + (testsz) >= (testaddr))
473 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
474 * alloc_sz on the righthand side of the comparison in order to avoid overflow
475 * or underflow in the comparison with it. This is simpler than the INRANGE
476 * check above, because we know that the dtms_scratch_ptr is valid in the
477 * range. Allocations of size zero are allowed.
479 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
480 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
481 (mstate)->dtms_scratch_ptr >= (alloc_sz))
483 #define DTRACE_LOADFUNC(bits) \
486 dtrace_load##bits(uintptr_t addr) \
488 size_t size = bits / NBBY; \
490 uint##bits##_t rval; \
492 volatile uint16_t *flags = (volatile uint16_t *) \
493 &cpu_core[curcpu].cpuc_dtrace_flags; \
495 DTRACE_ALIGNCHECK(addr, size, flags); \
497 for (i = 0; i < dtrace_toxranges; i++) { \
498 if (addr >= dtrace_toxrange[i].dtt_limit) \
501 if (addr + size <= dtrace_toxrange[i].dtt_base) \
505 * This address falls within a toxic region; return 0. \
507 *flags |= CPU_DTRACE_BADADDR; \
508 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
512 *flags |= CPU_DTRACE_NOFAULT; \
514 rval = *((volatile uint##bits##_t *)addr); \
515 *flags &= ~CPU_DTRACE_NOFAULT; \
517 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
521 #define dtrace_loadptr dtrace_load64
523 #define dtrace_loadptr dtrace_load32
526 #define DTRACE_DYNHASH_FREE 0
527 #define DTRACE_DYNHASH_SINK 1
528 #define DTRACE_DYNHASH_VALID 2
530 #define DTRACE_MATCH_NEXT 0
531 #define DTRACE_MATCH_DONE 1
532 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
533 #define DTRACE_STATE_ALIGN 64
535 #define DTRACE_FLAGS2FLT(flags) \
536 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
537 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
538 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
539 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
540 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
541 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
542 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
543 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
544 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
547 #define DTRACEACT_ISSTRING(act) \
548 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
549 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
551 /* Function prototype definitions: */
552 static size_t dtrace_strlen(const char *, size_t);
553 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
554 static void dtrace_enabling_provide(dtrace_provider_t *);
555 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
556 static void dtrace_enabling_matchall(void);
557 static void dtrace_enabling_reap(void);
558 static dtrace_state_t *dtrace_anon_grab(void);
559 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
560 dtrace_state_t *, uint64_t, uint64_t);
561 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
562 static void dtrace_buffer_drop(dtrace_buffer_t *);
563 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
564 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
565 dtrace_state_t *, dtrace_mstate_t *);
566 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
568 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
569 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
570 uint16_t dtrace_load16(uintptr_t);
571 uint32_t dtrace_load32(uintptr_t);
572 uint64_t dtrace_load64(uintptr_t);
573 uint8_t dtrace_load8(uintptr_t);
574 void dtrace_dynvar_clean(dtrace_dstate_t *);
575 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
576 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
577 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
580 * DTrace Probe Context Functions
582 * These functions are called from probe context. Because probe context is
583 * any context in which C may be called, arbitrarily locks may be held,
584 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
585 * As a result, functions called from probe context may only call other DTrace
586 * support functions -- they may not interact at all with the system at large.
587 * (Note that the ASSERT macro is made probe-context safe by redefining it in
588 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
589 * loads are to be performed from probe context, they _must_ be in terms of
590 * the safe dtrace_load*() variants.
592 * Some functions in this block are not actually called from probe context;
593 * for these functions, there will be a comment above the function reading
594 * "Note: not called from probe context."
597 dtrace_panic(const char *format, ...)
601 va_start(alist, format);
602 dtrace_vpanic(format, alist);
607 dtrace_assfail(const char *a, const char *f, int l)
609 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
612 * We just need something here that even the most clever compiler
613 * cannot optimize away.
615 return (a[(uintptr_t)f]);
619 * Atomically increment a specified error counter from probe context.
622 dtrace_error(uint32_t *counter)
625 * Most counters stored to in probe context are per-CPU counters.
626 * However, there are some error conditions that are sufficiently
627 * arcane that they don't merit per-CPU storage. If these counters
628 * are incremented concurrently on different CPUs, scalability will be
629 * adversely affected -- but we don't expect them to be white-hot in a
630 * correctly constructed enabling...
637 if ((nval = oval + 1) == 0) {
639 * If the counter would wrap, set it to 1 -- assuring
640 * that the counter is never zero when we have seen
641 * errors. (The counter must be 32-bits because we
642 * aren't guaranteed a 64-bit compare&swap operation.)
643 * To save this code both the infamy of being fingered
644 * by a priggish news story and the indignity of being
645 * the target of a neo-puritan witch trial, we're
646 * carefully avoiding any colorful description of the
647 * likelihood of this condition -- but suffice it to
648 * say that it is only slightly more likely than the
649 * overflow of predicate cache IDs, as discussed in
650 * dtrace_predicate_create().
654 } while (dtrace_cas32(counter, oval, nval) != oval);
658 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
659 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
667 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
669 if (dest < mstate->dtms_scratch_base)
672 if (dest + size < dest)
675 if (dest + size > mstate->dtms_scratch_ptr)
682 dtrace_canstore_statvar(uint64_t addr, size_t sz,
683 dtrace_statvar_t **svars, int nsvars)
687 for (i = 0; i < nsvars; i++) {
688 dtrace_statvar_t *svar = svars[i];
690 if (svar == NULL || svar->dtsv_size == 0)
693 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
701 * Check to see if the address is within a memory region to which a store may
702 * be issued. This includes the DTrace scratch areas, and any DTrace variable
703 * region. The caller of dtrace_canstore() is responsible for performing any
704 * alignment checks that are needed before stores are actually executed.
707 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
708 dtrace_vstate_t *vstate)
711 * First, check to see if the address is in scratch space...
713 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
714 mstate->dtms_scratch_size))
718 * Now check to see if it's a dynamic variable. This check will pick
719 * up both thread-local variables and any global dynamically-allocated
722 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
723 vstate->dtvs_dynvars.dtds_size)) {
724 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
725 uintptr_t base = (uintptr_t)dstate->dtds_base +
726 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
730 * Before we assume that we can store here, we need to make
731 * sure that it isn't in our metadata -- storing to our
732 * dynamic variable metadata would corrupt our state. For
733 * the range to not include any dynamic variable metadata,
736 * (1) Start above the hash table that is at the base of
737 * the dynamic variable space
739 * (2) Have a starting chunk offset that is beyond the
740 * dtrace_dynvar_t that is at the base of every chunk
742 * (3) Not span a chunk boundary
748 chunkoffs = (addr - base) % dstate->dtds_chunksize;
750 if (chunkoffs < sizeof (dtrace_dynvar_t))
753 if (chunkoffs + sz > dstate->dtds_chunksize)
760 * Finally, check the static local and global variables. These checks
761 * take the longest, so we perform them last.
763 if (dtrace_canstore_statvar(addr, sz,
764 vstate->dtvs_locals, vstate->dtvs_nlocals))
767 if (dtrace_canstore_statvar(addr, sz,
768 vstate->dtvs_globals, vstate->dtvs_nglobals))
776 * Convenience routine to check to see if the address is within a memory
777 * region in which a load may be issued given the user's privilege level;
778 * if not, it sets the appropriate error flags and loads 'addr' into the
779 * illegal value slot.
781 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
782 * appropriate memory access protection.
785 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
786 dtrace_vstate_t *vstate)
788 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
791 * If we hold the privilege to read from kernel memory, then
792 * everything is readable.
794 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
798 * You can obviously read that which you can store.
800 if (dtrace_canstore(addr, sz, mstate, vstate))
804 * We're allowed to read from our own string table.
806 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
807 mstate->dtms_difo->dtdo_strlen))
810 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
816 * Convenience routine to check to see if a given string is within a memory
817 * region in which a load may be issued given the user's privilege level;
818 * this exists so that we don't need to issue unnecessary dtrace_strlen()
819 * calls in the event that the user has all privileges.
822 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
823 dtrace_vstate_t *vstate)
828 * If we hold the privilege to read from kernel memory, then
829 * everything is readable.
831 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
834 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
835 if (dtrace_canload(addr, strsz, mstate, vstate))
842 * Convenience routine to check to see if a given variable is within a memory
843 * region in which a load may be issued given the user's privilege level.
846 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
847 dtrace_vstate_t *vstate)
850 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
853 * If we hold the privilege to read from kernel memory, then
854 * everything is readable.
856 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
859 if (type->dtdt_kind == DIF_TYPE_STRING)
860 sz = dtrace_strlen(src,
861 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
863 sz = type->dtdt_size;
865 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
869 * Compare two strings using safe loads.
872 dtrace_strncmp(char *s1, char *s2, size_t limit)
875 volatile uint16_t *flags;
877 if (s1 == s2 || limit == 0)
880 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
886 c1 = dtrace_load8((uintptr_t)s1++);
892 c2 = dtrace_load8((uintptr_t)s2++);
897 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
903 * Compute strlen(s) for a string using safe memory accesses. The additional
904 * len parameter is used to specify a maximum length to ensure completion.
907 dtrace_strlen(const char *s, size_t lim)
911 for (len = 0; len != lim; len++) {
912 if (dtrace_load8((uintptr_t)s++) == '\0')
920 * Check if an address falls within a toxic region.
923 dtrace_istoxic(uintptr_t kaddr, size_t size)
925 uintptr_t taddr, tsize;
928 for (i = 0; i < dtrace_toxranges; i++) {
929 taddr = dtrace_toxrange[i].dtt_base;
930 tsize = dtrace_toxrange[i].dtt_limit - taddr;
932 if (kaddr - taddr < tsize) {
933 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
934 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
938 if (taddr - kaddr < size) {
939 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
940 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
949 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
950 * memory specified by the DIF program. The dst is assumed to be safe memory
951 * that we can store to directly because it is managed by DTrace. As with
952 * standard bcopy, overlapping copies are handled properly.
955 dtrace_bcopy(const void *src, void *dst, size_t len)
959 const uint8_t *s2 = src;
963 *s1++ = dtrace_load8((uintptr_t)s2++);
964 } while (--len != 0);
970 *--s1 = dtrace_load8((uintptr_t)--s2);
971 } while (--len != 0);
977 * Copy src to dst using safe memory accesses, up to either the specified
978 * length, or the point that a nul byte is encountered. The src is assumed to
979 * be unsafe memory specified by the DIF program. The dst is assumed to be
980 * safe memory that we can store to directly because it is managed by DTrace.
981 * Unlike dtrace_bcopy(), overlapping regions are not handled.
984 dtrace_strcpy(const void *src, void *dst, size_t len)
987 uint8_t *s1 = dst, c;
988 const uint8_t *s2 = src;
991 *s1++ = c = dtrace_load8((uintptr_t)s2++);
992 } while (--len != 0 && c != '\0');
997 * Copy src to dst, deriving the size and type from the specified (BYREF)
998 * variable type. The src is assumed to be unsafe memory specified by the DIF
999 * program. The dst is assumed to be DTrace variable memory that is of the
1000 * specified type; we assume that we can store to directly.
1003 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1005 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1007 if (type->dtdt_kind == DIF_TYPE_STRING) {
1008 dtrace_strcpy(src, dst, type->dtdt_size);
1010 dtrace_bcopy(src, dst, type->dtdt_size);
1015 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1016 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1017 * safe memory that we can access directly because it is managed by DTrace.
1020 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1022 volatile uint16_t *flags;
1024 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1029 if (s1 == NULL || s2 == NULL)
1032 if (s1 != s2 && len != 0) {
1033 const uint8_t *ps1 = s1;
1034 const uint8_t *ps2 = s2;
1037 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1039 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1045 * Zero the specified region using a simple byte-by-byte loop. Note that this
1046 * is for safe DTrace-managed memory only.
1049 dtrace_bzero(void *dst, size_t len)
1053 for (cp = dst; len != 0; len--)
1058 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1062 result[0] = addend1[0] + addend2[0];
1063 result[1] = addend1[1] + addend2[1] +
1064 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1071 * Shift the 128-bit value in a by b. If b is positive, shift left.
1072 * If b is negative, shift right.
1075 dtrace_shift_128(uint64_t *a, int b)
1085 a[0] = a[1] >> (b - 64);
1089 mask = 1LL << (64 - b);
1091 a[0] |= ((a[1] & mask) << (64 - b));
1096 a[1] = a[0] << (b - 64);
1100 mask = a[0] >> (64 - b);
1108 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1109 * use native multiplication on those, and then re-combine into the
1110 * resulting 128-bit value.
1112 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1119 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1121 uint64_t hi1, hi2, lo1, lo2;
1124 hi1 = factor1 >> 32;
1125 hi2 = factor2 >> 32;
1127 lo1 = factor1 & DT_MASK_LO;
1128 lo2 = factor2 & DT_MASK_LO;
1130 product[0] = lo1 * lo2;
1131 product[1] = hi1 * hi2;
1135 dtrace_shift_128(tmp, 32);
1136 dtrace_add_128(product, tmp, product);
1140 dtrace_shift_128(tmp, 32);
1141 dtrace_add_128(product, tmp, product);
1145 * This privilege check should be used by actions and subroutines to
1146 * verify that the user credentials of the process that enabled the
1147 * invoking ECB match the target credentials
1150 dtrace_priv_proc_common_user(dtrace_state_t *state)
1152 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1155 * We should always have a non-NULL state cred here, since if cred
1156 * is null (anonymous tracing), we fast-path bypass this routine.
1158 ASSERT(s_cr != NULL);
1160 if ((cr = CRED()) != NULL &&
1161 s_cr->cr_uid == cr->cr_uid &&
1162 s_cr->cr_uid == cr->cr_ruid &&
1163 s_cr->cr_uid == cr->cr_suid &&
1164 s_cr->cr_gid == cr->cr_gid &&
1165 s_cr->cr_gid == cr->cr_rgid &&
1166 s_cr->cr_gid == cr->cr_sgid)
1173 * This privilege check should be used by actions and subroutines to
1174 * verify that the zone of the process that enabled the invoking ECB
1175 * matches the target credentials
1178 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1181 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1184 * We should always have a non-NULL state cred here, since if cred
1185 * is null (anonymous tracing), we fast-path bypass this routine.
1187 ASSERT(s_cr != NULL);
1189 if ((cr = CRED()) != NULL &&
1190 s_cr->cr_zone == cr->cr_zone)
1200 * This privilege check should be used by actions and subroutines to
1201 * verify that the process has not setuid or changed credentials.
1204 dtrace_priv_proc_common_nocd(void)
1208 if ((proc = ttoproc(curthread)) != NULL &&
1209 !(proc->p_flag & SNOCD))
1216 dtrace_priv_proc_destructive(dtrace_state_t *state)
1218 int action = state->dts_cred.dcr_action;
1220 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1221 dtrace_priv_proc_common_zone(state) == 0)
1224 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1225 dtrace_priv_proc_common_user(state) == 0)
1228 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1229 dtrace_priv_proc_common_nocd() == 0)
1235 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1241 dtrace_priv_proc_control(dtrace_state_t *state)
1243 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1246 if (dtrace_priv_proc_common_zone(state) &&
1247 dtrace_priv_proc_common_user(state) &&
1248 dtrace_priv_proc_common_nocd())
1251 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1257 dtrace_priv_proc(dtrace_state_t *state)
1259 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1262 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1268 dtrace_priv_kernel(dtrace_state_t *state)
1270 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1273 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1279 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1281 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1284 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1290 * Note: not called from probe context. This function is called
1291 * asynchronously (and at a regular interval) from outside of probe context to
1292 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1293 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1296 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1298 dtrace_dynvar_t *dirty;
1299 dtrace_dstate_percpu_t *dcpu;
1302 for (i = 0; i < NCPU; i++) {
1303 dcpu = &dstate->dtds_percpu[i];
1305 ASSERT(dcpu->dtdsc_rinsing == NULL);
1308 * If the dirty list is NULL, there is no dirty work to do.
1310 if (dcpu->dtdsc_dirty == NULL)
1314 * If the clean list is non-NULL, then we're not going to do
1315 * any work for this CPU -- it means that there has not been
1316 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1317 * since the last time we cleaned house.
1319 if (dcpu->dtdsc_clean != NULL)
1325 * Atomically move the dirty list aside.
1328 dirty = dcpu->dtdsc_dirty;
1331 * Before we zap the dirty list, set the rinsing list.
1332 * (This allows for a potential assertion in
1333 * dtrace_dynvar(): if a free dynamic variable appears
1334 * on a hash chain, either the dirty list or the
1335 * rinsing list for some CPU must be non-NULL.)
1337 dcpu->dtdsc_rinsing = dirty;
1338 dtrace_membar_producer();
1339 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1340 dirty, NULL) != dirty);
1345 * We have no work to do; we can simply return.
1352 for (i = 0; i < NCPU; i++) {
1353 dcpu = &dstate->dtds_percpu[i];
1355 if (dcpu->dtdsc_rinsing == NULL)
1359 * We are now guaranteed that no hash chain contains a pointer
1360 * into this dirty list; we can make it clean.
1362 ASSERT(dcpu->dtdsc_clean == NULL);
1363 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1364 dcpu->dtdsc_rinsing = NULL;
1368 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1369 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1370 * This prevents a race whereby a CPU incorrectly decides that
1371 * the state should be something other than DTRACE_DSTATE_CLEAN
1372 * after dtrace_dynvar_clean() has completed.
1376 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1380 * Depending on the value of the op parameter, this function looks-up,
1381 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1382 * allocation is requested, this function will return a pointer to a
1383 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1384 * variable can be allocated. If NULL is returned, the appropriate counter
1385 * will be incremented.
1388 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1389 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1390 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1392 uint64_t hashval = DTRACE_DYNHASH_VALID;
1393 dtrace_dynhash_t *hash = dstate->dtds_hash;
1394 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1395 processorid_t me = curcpu, cpu = me;
1396 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1397 size_t bucket, ksize;
1398 size_t chunksize = dstate->dtds_chunksize;
1399 uintptr_t kdata, lock, nstate;
1405 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1406 * algorithm. For the by-value portions, we perform the algorithm in
1407 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1408 * bit, and seems to have only a minute effect on distribution. For
1409 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1410 * over each referenced byte. It's painful to do this, but it's much
1411 * better than pathological hash distribution. The efficacy of the
1412 * hashing algorithm (and a comparison with other algorithms) may be
1413 * found by running the ::dtrace_dynstat MDB dcmd.
1415 for (i = 0; i < nkeys; i++) {
1416 if (key[i].dttk_size == 0) {
1417 uint64_t val = key[i].dttk_value;
1419 hashval += (val >> 48) & 0xffff;
1420 hashval += (hashval << 10);
1421 hashval ^= (hashval >> 6);
1423 hashval += (val >> 32) & 0xffff;
1424 hashval += (hashval << 10);
1425 hashval ^= (hashval >> 6);
1427 hashval += (val >> 16) & 0xffff;
1428 hashval += (hashval << 10);
1429 hashval ^= (hashval >> 6);
1431 hashval += val & 0xffff;
1432 hashval += (hashval << 10);
1433 hashval ^= (hashval >> 6);
1436 * This is incredibly painful, but it beats the hell
1437 * out of the alternative.
1439 uint64_t j, size = key[i].dttk_size;
1440 uintptr_t base = (uintptr_t)key[i].dttk_value;
1442 if (!dtrace_canload(base, size, mstate, vstate))
1445 for (j = 0; j < size; j++) {
1446 hashval += dtrace_load8(base + j);
1447 hashval += (hashval << 10);
1448 hashval ^= (hashval >> 6);
1453 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1456 hashval += (hashval << 3);
1457 hashval ^= (hashval >> 11);
1458 hashval += (hashval << 15);
1461 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1462 * comes out to be one of our two sentinel hash values. If this
1463 * actually happens, we set the hashval to be a value known to be a
1464 * non-sentinel value.
1466 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1467 hashval = DTRACE_DYNHASH_VALID;
1470 * Yes, it's painful to do a divide here. If the cycle count becomes
1471 * important here, tricks can be pulled to reduce it. (However, it's
1472 * critical that hash collisions be kept to an absolute minimum;
1473 * they're much more painful than a divide.) It's better to have a
1474 * solution that generates few collisions and still keeps things
1475 * relatively simple.
1477 bucket = hashval % dstate->dtds_hashsize;
1479 if (op == DTRACE_DYNVAR_DEALLOC) {
1480 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1483 while ((lock = *lockp) & 1)
1486 if (dtrace_casptr((volatile void *)lockp,
1487 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1491 dtrace_membar_producer();
1496 lock = hash[bucket].dtdh_lock;
1498 dtrace_membar_consumer();
1500 start = hash[bucket].dtdh_chain;
1501 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1502 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1503 op != DTRACE_DYNVAR_DEALLOC));
1505 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1506 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1507 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1509 if (dvar->dtdv_hashval != hashval) {
1510 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1512 * We've reached the sink, and therefore the
1513 * end of the hash chain; we can kick out of
1514 * the loop knowing that we have seen a valid
1515 * snapshot of state.
1517 ASSERT(dvar->dtdv_next == NULL);
1518 ASSERT(dvar == &dtrace_dynhash_sink);
1522 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1524 * We've gone off the rails: somewhere along
1525 * the line, one of the members of this hash
1526 * chain was deleted. Note that we could also
1527 * detect this by simply letting this loop run
1528 * to completion, as we would eventually hit
1529 * the end of the dirty list. However, we
1530 * want to avoid running the length of the
1531 * dirty list unnecessarily (it might be quite
1532 * long), so we catch this as early as
1533 * possible by detecting the hash marker. In
1534 * this case, we simply set dvar to NULL and
1535 * break; the conditional after the loop will
1536 * send us back to top.
1545 if (dtuple->dtt_nkeys != nkeys)
1548 for (i = 0; i < nkeys; i++, dkey++) {
1549 if (dkey->dttk_size != key[i].dttk_size)
1550 goto next; /* size or type mismatch */
1552 if (dkey->dttk_size != 0) {
1554 (void *)(uintptr_t)key[i].dttk_value,
1555 (void *)(uintptr_t)dkey->dttk_value,
1559 if (dkey->dttk_value != key[i].dttk_value)
1564 if (op != DTRACE_DYNVAR_DEALLOC)
1567 ASSERT(dvar->dtdv_next == NULL ||
1568 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1571 ASSERT(hash[bucket].dtdh_chain != dvar);
1572 ASSERT(start != dvar);
1573 ASSERT(prev->dtdv_next == dvar);
1574 prev->dtdv_next = dvar->dtdv_next;
1576 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1577 start, dvar->dtdv_next) != start) {
1579 * We have failed to atomically swing the
1580 * hash table head pointer, presumably because
1581 * of a conflicting allocation on another CPU.
1582 * We need to reread the hash chain and try
1589 dtrace_membar_producer();
1592 * Now set the hash value to indicate that it's free.
1594 ASSERT(hash[bucket].dtdh_chain != dvar);
1595 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1597 dtrace_membar_producer();
1600 * Set the next pointer to point at the dirty list, and
1601 * atomically swing the dirty pointer to the newly freed dvar.
1604 next = dcpu->dtdsc_dirty;
1605 dvar->dtdv_next = next;
1606 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1609 * Finally, unlock this hash bucket.
1611 ASSERT(hash[bucket].dtdh_lock == lock);
1613 hash[bucket].dtdh_lock++;
1623 * If dvar is NULL, it is because we went off the rails:
1624 * one of the elements that we traversed in the hash chain
1625 * was deleted while we were traversing it. In this case,
1626 * we assert that we aren't doing a dealloc (deallocs lock
1627 * the hash bucket to prevent themselves from racing with
1628 * one another), and retry the hash chain traversal.
1630 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1634 if (op != DTRACE_DYNVAR_ALLOC) {
1636 * If we are not to allocate a new variable, we want to
1637 * return NULL now. Before we return, check that the value
1638 * of the lock word hasn't changed. If it has, we may have
1639 * seen an inconsistent snapshot.
1641 if (op == DTRACE_DYNVAR_NOALLOC) {
1642 if (hash[bucket].dtdh_lock != lock)
1645 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1646 ASSERT(hash[bucket].dtdh_lock == lock);
1648 hash[bucket].dtdh_lock++;
1655 * We need to allocate a new dynamic variable. The size we need is the
1656 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1657 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1658 * the size of any referred-to data (dsize). We then round the final
1659 * size up to the chunksize for allocation.
1661 for (ksize = 0, i = 0; i < nkeys; i++)
1662 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1665 * This should be pretty much impossible, but could happen if, say,
1666 * strange DIF specified the tuple. Ideally, this should be an
1667 * assertion and not an error condition -- but that requires that the
1668 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1669 * bullet-proof. (That is, it must not be able to be fooled by
1670 * malicious DIF.) Given the lack of backwards branches in DIF,
1671 * solving this would presumably not amount to solving the Halting
1672 * Problem -- but it still seems awfully hard.
1674 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1675 ksize + dsize > chunksize) {
1676 dcpu->dtdsc_drops++;
1680 nstate = DTRACE_DSTATE_EMPTY;
1684 free = dcpu->dtdsc_free;
1687 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1690 if (clean == NULL) {
1692 * We're out of dynamic variable space on
1693 * this CPU. Unless we have tried all CPUs,
1694 * we'll try to allocate from a different
1697 switch (dstate->dtds_state) {
1698 case DTRACE_DSTATE_CLEAN: {
1699 void *sp = &dstate->dtds_state;
1704 if (dcpu->dtdsc_dirty != NULL &&
1705 nstate == DTRACE_DSTATE_EMPTY)
1706 nstate = DTRACE_DSTATE_DIRTY;
1708 if (dcpu->dtdsc_rinsing != NULL)
1709 nstate = DTRACE_DSTATE_RINSING;
1711 dcpu = &dstate->dtds_percpu[cpu];
1716 (void) dtrace_cas32(sp,
1717 DTRACE_DSTATE_CLEAN, nstate);
1720 * To increment the correct bean
1721 * counter, take another lap.
1726 case DTRACE_DSTATE_DIRTY:
1727 dcpu->dtdsc_dirty_drops++;
1730 case DTRACE_DSTATE_RINSING:
1731 dcpu->dtdsc_rinsing_drops++;
1734 case DTRACE_DSTATE_EMPTY:
1735 dcpu->dtdsc_drops++;
1739 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1744 * The clean list appears to be non-empty. We want to
1745 * move the clean list to the free list; we start by
1746 * moving the clean pointer aside.
1748 if (dtrace_casptr(&dcpu->dtdsc_clean,
1749 clean, NULL) != clean) {
1751 * We are in one of two situations:
1753 * (a) The clean list was switched to the
1754 * free list by another CPU.
1756 * (b) The clean list was added to by the
1759 * In either of these situations, we can
1760 * just reattempt the free list allocation.
1765 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1768 * Now we'll move the clean list to the free list.
1769 * It's impossible for this to fail: the only way
1770 * the free list can be updated is through this
1771 * code path, and only one CPU can own the clean list.
1772 * Thus, it would only be possible for this to fail if
1773 * this code were racing with dtrace_dynvar_clean().
1774 * (That is, if dtrace_dynvar_clean() updated the clean
1775 * list, and we ended up racing to update the free
1776 * list.) This race is prevented by the dtrace_sync()
1777 * in dtrace_dynvar_clean() -- which flushes the
1778 * owners of the clean lists out before resetting
1781 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1782 ASSERT(rval == NULL);
1787 new_free = dvar->dtdv_next;
1788 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1791 * We have now allocated a new chunk. We copy the tuple keys into the
1792 * tuple array and copy any referenced key data into the data space
1793 * following the tuple array. As we do this, we relocate dttk_value
1794 * in the final tuple to point to the key data address in the chunk.
1796 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1797 dvar->dtdv_data = (void *)(kdata + ksize);
1798 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1800 for (i = 0; i < nkeys; i++) {
1801 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1802 size_t kesize = key[i].dttk_size;
1806 (const void *)(uintptr_t)key[i].dttk_value,
1807 (void *)kdata, kesize);
1808 dkey->dttk_value = kdata;
1809 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1811 dkey->dttk_value = key[i].dttk_value;
1814 dkey->dttk_size = kesize;
1817 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1818 dvar->dtdv_hashval = hashval;
1819 dvar->dtdv_next = start;
1821 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1825 * The cas has failed. Either another CPU is adding an element to
1826 * this hash chain, or another CPU is deleting an element from this
1827 * hash chain. The simplest way to deal with both of these cases
1828 * (though not necessarily the most efficient) is to free our
1829 * allocated block and tail-call ourselves. Note that the free is
1830 * to the dirty list and _not_ to the free list. This is to prevent
1831 * races with allocators, above.
1833 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1835 dtrace_membar_producer();
1838 free = dcpu->dtdsc_dirty;
1839 dvar->dtdv_next = free;
1840 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1842 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1847 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1849 if ((int64_t)nval < (int64_t)*oval)
1855 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1857 if ((int64_t)nval > (int64_t)*oval)
1862 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1864 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1865 int64_t val = (int64_t)nval;
1868 for (i = 0; i < zero; i++) {
1869 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1875 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1876 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1877 quanta[i - 1] += incr;
1882 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1890 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1892 uint64_t arg = *lquanta++;
1893 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1894 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1895 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1896 int32_t val = (int32_t)nval, level;
1899 ASSERT(levels != 0);
1903 * This is an underflow.
1909 level = (val - base) / step;
1911 if (level < levels) {
1912 lquanta[level + 1] += incr;
1917 * This is an overflow.
1919 lquanta[levels + 1] += incr;
1923 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1924 uint16_t high, uint16_t nsteps, int64_t value)
1926 int64_t this = 1, last, next;
1927 int base = 1, order;
1929 ASSERT(factor <= nsteps);
1930 ASSERT(nsteps % factor == 0);
1932 for (order = 0; order < low; order++)
1936 * If our value is less than our factor taken to the power of the
1937 * low order of magnitude, it goes into the zeroth bucket.
1939 if (value < (last = this))
1942 for (this *= factor; order <= high; order++) {
1943 int nbuckets = this > nsteps ? nsteps : this;
1945 if ((next = this * factor) < this) {
1947 * We should not generally get log/linear quantizations
1948 * with a high magnitude that allows 64-bits to
1949 * overflow, but we nonetheless protect against this
1950 * by explicitly checking for overflow, and clamping
1951 * our value accordingly.
1958 * If our value lies within this order of magnitude,
1959 * determine its position by taking the offset within
1960 * the order of magnitude, dividing by the bucket
1961 * width, and adding to our (accumulated) base.
1963 return (base + (value - last) / (this / nbuckets));
1966 base += nbuckets - (nbuckets / factor);
1972 * Our value is greater than or equal to our factor taken to the
1973 * power of one plus the high magnitude -- return the top bucket.
1979 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
1981 uint64_t arg = *llquanta++;
1982 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
1983 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
1984 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
1985 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
1987 llquanta[dtrace_aggregate_llquantize_bucket(factor,
1988 low, high, nsteps, nval)] += incr;
1993 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2001 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2003 int64_t snval = (int64_t)nval;
2010 * What we want to say here is:
2012 * data[2] += nval * nval;
2014 * But given that nval is 64-bit, we could easily overflow, so
2015 * we do this as 128-bit arithmetic.
2020 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2021 dtrace_add_128(data + 2, tmp, data + 2);
2026 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2033 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2039 * Aggregate given the tuple in the principal data buffer, and the aggregating
2040 * action denoted by the specified dtrace_aggregation_t. The aggregation
2041 * buffer is specified as the buf parameter. This routine does not return
2042 * failure; if there is no space in the aggregation buffer, the data will be
2043 * dropped, and a corresponding counter incremented.
2046 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2047 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2049 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2050 uint32_t i, ndx, size, fsize;
2051 uint32_t align = sizeof (uint64_t) - 1;
2052 dtrace_aggbuffer_t *agb;
2053 dtrace_aggkey_t *key;
2054 uint32_t hashval = 0, limit, isstr;
2055 caddr_t tomax, data, kdata;
2056 dtrace_actkind_t action;
2057 dtrace_action_t *act;
2063 if (!agg->dtag_hasarg) {
2065 * Currently, only quantize() and lquantize() take additional
2066 * arguments, and they have the same semantics: an increment
2067 * value that defaults to 1 when not present. If additional
2068 * aggregating actions take arguments, the setting of the
2069 * default argument value will presumably have to become more
2075 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2076 size = rec->dtrd_offset - agg->dtag_base;
2077 fsize = size + rec->dtrd_size;
2079 ASSERT(dbuf->dtb_tomax != NULL);
2080 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2082 if ((tomax = buf->dtb_tomax) == NULL) {
2083 dtrace_buffer_drop(buf);
2088 * The metastructure is always at the bottom of the buffer.
2090 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2091 sizeof (dtrace_aggbuffer_t));
2093 if (buf->dtb_offset == 0) {
2095 * We just kludge up approximately 1/8th of the size to be
2096 * buckets. If this guess ends up being routinely
2097 * off-the-mark, we may need to dynamically readjust this
2098 * based on past performance.
2100 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2102 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2103 (uintptr_t)tomax || hashsize == 0) {
2105 * We've been given a ludicrously small buffer;
2106 * increment our drop count and leave.
2108 dtrace_buffer_drop(buf);
2113 * And now, a pathetic attempt to try to get a an odd (or
2114 * perchance, a prime) hash size for better hash distribution.
2116 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2117 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2119 agb->dtagb_hashsize = hashsize;
2120 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2121 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2122 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2124 for (i = 0; i < agb->dtagb_hashsize; i++)
2125 agb->dtagb_hash[i] = NULL;
2128 ASSERT(agg->dtag_first != NULL);
2129 ASSERT(agg->dtag_first->dta_intuple);
2132 * Calculate the hash value based on the key. Note that we _don't_
2133 * include the aggid in the hashing (but we will store it as part of
2134 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2135 * algorithm: a simple, quick algorithm that has no known funnels, and
2136 * gets good distribution in practice. The efficacy of the hashing
2137 * algorithm (and a comparison with other algorithms) may be found by
2138 * running the ::dtrace_aggstat MDB dcmd.
2140 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2141 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2142 limit = i + act->dta_rec.dtrd_size;
2143 ASSERT(limit <= size);
2144 isstr = DTRACEACT_ISSTRING(act);
2146 for (; i < limit; i++) {
2148 hashval += (hashval << 10);
2149 hashval ^= (hashval >> 6);
2151 if (isstr && data[i] == '\0')
2156 hashval += (hashval << 3);
2157 hashval ^= (hashval >> 11);
2158 hashval += (hashval << 15);
2161 * Yes, the divide here is expensive -- but it's generally the least
2162 * of the performance issues given the amount of data that we iterate
2163 * over to compute hash values, compare data, etc.
2165 ndx = hashval % agb->dtagb_hashsize;
2167 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2168 ASSERT((caddr_t)key >= tomax);
2169 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2171 if (hashval != key->dtak_hashval || key->dtak_size != size)
2174 kdata = key->dtak_data;
2175 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2177 for (act = agg->dtag_first; act->dta_intuple;
2178 act = act->dta_next) {
2179 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2180 limit = i + act->dta_rec.dtrd_size;
2181 ASSERT(limit <= size);
2182 isstr = DTRACEACT_ISSTRING(act);
2184 for (; i < limit; i++) {
2185 if (kdata[i] != data[i])
2188 if (isstr && data[i] == '\0')
2193 if (action != key->dtak_action) {
2195 * We are aggregating on the same value in the same
2196 * aggregation with two different aggregating actions.
2197 * (This should have been picked up in the compiler,
2198 * so we may be dealing with errant or devious DIF.)
2199 * This is an error condition; we indicate as much,
2202 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2207 * This is a hit: we need to apply the aggregator to
2208 * the value at this key.
2210 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2217 * We didn't find it. We need to allocate some zero-filled space,
2218 * link it into the hash table appropriately, and apply the aggregator
2219 * to the (zero-filled) value.
2221 offs = buf->dtb_offset;
2222 while (offs & (align - 1))
2223 offs += sizeof (uint32_t);
2226 * If we don't have enough room to both allocate a new key _and_
2227 * its associated data, increment the drop count and return.
2229 if ((uintptr_t)tomax + offs + fsize >
2230 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2231 dtrace_buffer_drop(buf);
2236 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2237 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2238 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2240 key->dtak_data = kdata = tomax + offs;
2241 buf->dtb_offset = offs + fsize;
2244 * Now copy the data across.
2246 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2248 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2252 * Because strings are not zeroed out by default, we need to iterate
2253 * looking for actions that store strings, and we need to explicitly
2254 * pad these strings out with zeroes.
2256 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2259 if (!DTRACEACT_ISSTRING(act))
2262 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2263 limit = i + act->dta_rec.dtrd_size;
2264 ASSERT(limit <= size);
2266 for (nul = 0; i < limit; i++) {
2272 if (data[i] != '\0')
2279 for (i = size; i < fsize; i++)
2282 key->dtak_hashval = hashval;
2283 key->dtak_size = size;
2284 key->dtak_action = action;
2285 key->dtak_next = agb->dtagb_hash[ndx];
2286 agb->dtagb_hash[ndx] = key;
2289 * Finally, apply the aggregator.
2291 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2292 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2296 * Given consumer state, this routine finds a speculation in the INACTIVE
2297 * state and transitions it into the ACTIVE state. If there is no speculation
2298 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2299 * incremented -- it is up to the caller to take appropriate action.
2302 dtrace_speculation(dtrace_state_t *state)
2305 dtrace_speculation_state_t current;
2306 uint32_t *stat = &state->dts_speculations_unavail, count;
2308 while (i < state->dts_nspeculations) {
2309 dtrace_speculation_t *spec = &state->dts_speculations[i];
2311 current = spec->dtsp_state;
2313 if (current != DTRACESPEC_INACTIVE) {
2314 if (current == DTRACESPEC_COMMITTINGMANY ||
2315 current == DTRACESPEC_COMMITTING ||
2316 current == DTRACESPEC_DISCARDING)
2317 stat = &state->dts_speculations_busy;
2322 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2323 current, DTRACESPEC_ACTIVE) == current)
2328 * We couldn't find a speculation. If we found as much as a single
2329 * busy speculation buffer, we'll attribute this failure as "busy"
2330 * instead of "unavail".
2334 } while (dtrace_cas32(stat, count, count + 1) != count);
2340 * This routine commits an active speculation. If the specified speculation
2341 * is not in a valid state to perform a commit(), this routine will silently do
2342 * nothing. The state of the specified speculation is transitioned according
2343 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2346 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2347 dtrace_specid_t which)
2349 dtrace_speculation_t *spec;
2350 dtrace_buffer_t *src, *dest;
2351 uintptr_t daddr, saddr, dlimit;
2352 dtrace_speculation_state_t current, new = 0;
2358 if (which > state->dts_nspeculations) {
2359 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2363 spec = &state->dts_speculations[which - 1];
2364 src = &spec->dtsp_buffer[cpu];
2365 dest = &state->dts_buffer[cpu];
2368 current = spec->dtsp_state;
2370 if (current == DTRACESPEC_COMMITTINGMANY)
2374 case DTRACESPEC_INACTIVE:
2375 case DTRACESPEC_DISCARDING:
2378 case DTRACESPEC_COMMITTING:
2380 * This is only possible if we are (a) commit()'ing
2381 * without having done a prior speculate() on this CPU
2382 * and (b) racing with another commit() on a different
2383 * CPU. There's nothing to do -- we just assert that
2386 ASSERT(src->dtb_offset == 0);
2389 case DTRACESPEC_ACTIVE:
2390 new = DTRACESPEC_COMMITTING;
2393 case DTRACESPEC_ACTIVEONE:
2395 * This speculation is active on one CPU. If our
2396 * buffer offset is non-zero, we know that the one CPU
2397 * must be us. Otherwise, we are committing on a
2398 * different CPU from the speculate(), and we must
2399 * rely on being asynchronously cleaned.
2401 if (src->dtb_offset != 0) {
2402 new = DTRACESPEC_COMMITTING;
2407 case DTRACESPEC_ACTIVEMANY:
2408 new = DTRACESPEC_COMMITTINGMANY;
2414 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2415 current, new) != current);
2418 * We have set the state to indicate that we are committing this
2419 * speculation. Now reserve the necessary space in the destination
2422 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2423 sizeof (uint64_t), state, NULL)) < 0) {
2424 dtrace_buffer_drop(dest);
2429 * We have the space; copy the buffer across. (Note that this is a
2430 * highly subobtimal bcopy(); in the unlikely event that this becomes
2431 * a serious performance issue, a high-performance DTrace-specific
2432 * bcopy() should obviously be invented.)
2434 daddr = (uintptr_t)dest->dtb_tomax + offs;
2435 dlimit = daddr + src->dtb_offset;
2436 saddr = (uintptr_t)src->dtb_tomax;
2439 * First, the aligned portion.
2441 while (dlimit - daddr >= sizeof (uint64_t)) {
2442 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2444 daddr += sizeof (uint64_t);
2445 saddr += sizeof (uint64_t);
2449 * Now any left-over bit...
2451 while (dlimit - daddr)
2452 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2455 * Finally, commit the reserved space in the destination buffer.
2457 dest->dtb_offset = offs + src->dtb_offset;
2461 * If we're lucky enough to be the only active CPU on this speculation
2462 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2464 if (current == DTRACESPEC_ACTIVE ||
2465 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2466 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2467 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2469 ASSERT(rval == DTRACESPEC_COMMITTING);
2472 src->dtb_offset = 0;
2473 src->dtb_xamot_drops += src->dtb_drops;
2478 * This routine discards an active speculation. If the specified speculation
2479 * is not in a valid state to perform a discard(), this routine will silently
2480 * do nothing. The state of the specified speculation is transitioned
2481 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2484 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2485 dtrace_specid_t which)
2487 dtrace_speculation_t *spec;
2488 dtrace_speculation_state_t current, new = 0;
2489 dtrace_buffer_t *buf;
2494 if (which > state->dts_nspeculations) {
2495 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2499 spec = &state->dts_speculations[which - 1];
2500 buf = &spec->dtsp_buffer[cpu];
2503 current = spec->dtsp_state;
2506 case DTRACESPEC_INACTIVE:
2507 case DTRACESPEC_COMMITTINGMANY:
2508 case DTRACESPEC_COMMITTING:
2509 case DTRACESPEC_DISCARDING:
2512 case DTRACESPEC_ACTIVE:
2513 case DTRACESPEC_ACTIVEMANY:
2514 new = DTRACESPEC_DISCARDING;
2517 case DTRACESPEC_ACTIVEONE:
2518 if (buf->dtb_offset != 0) {
2519 new = DTRACESPEC_INACTIVE;
2521 new = DTRACESPEC_DISCARDING;
2528 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2529 current, new) != current);
2531 buf->dtb_offset = 0;
2536 * Note: not called from probe context. This function is called
2537 * asynchronously from cross call context to clean any speculations that are
2538 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2539 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2543 dtrace_speculation_clean_here(dtrace_state_t *state)
2545 dtrace_icookie_t cookie;
2546 processorid_t cpu = curcpu;
2547 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2550 cookie = dtrace_interrupt_disable();
2552 if (dest->dtb_tomax == NULL) {
2553 dtrace_interrupt_enable(cookie);
2557 for (i = 0; i < state->dts_nspeculations; i++) {
2558 dtrace_speculation_t *spec = &state->dts_speculations[i];
2559 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2561 if (src->dtb_tomax == NULL)
2564 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2565 src->dtb_offset = 0;
2569 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2572 if (src->dtb_offset == 0)
2575 dtrace_speculation_commit(state, cpu, i + 1);
2578 dtrace_interrupt_enable(cookie);
2582 * Note: not called from probe context. This function is called
2583 * asynchronously (and at a regular interval) to clean any speculations that
2584 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2585 * is work to be done, it cross calls all CPUs to perform that work;
2586 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2587 * INACTIVE state until they have been cleaned by all CPUs.
2590 dtrace_speculation_clean(dtrace_state_t *state)
2595 for (i = 0; i < state->dts_nspeculations; i++) {
2596 dtrace_speculation_t *spec = &state->dts_speculations[i];
2598 ASSERT(!spec->dtsp_cleaning);
2600 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2601 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2605 spec->dtsp_cleaning = 1;
2611 dtrace_xcall(DTRACE_CPUALL,
2612 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2615 * We now know that all CPUs have committed or discarded their
2616 * speculation buffers, as appropriate. We can now set the state
2619 for (i = 0; i < state->dts_nspeculations; i++) {
2620 dtrace_speculation_t *spec = &state->dts_speculations[i];
2621 dtrace_speculation_state_t current, new;
2623 if (!spec->dtsp_cleaning)
2626 current = spec->dtsp_state;
2627 ASSERT(current == DTRACESPEC_DISCARDING ||
2628 current == DTRACESPEC_COMMITTINGMANY);
2630 new = DTRACESPEC_INACTIVE;
2632 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2633 ASSERT(rv == current);
2634 spec->dtsp_cleaning = 0;
2639 * Called as part of a speculate() to get the speculative buffer associated
2640 * with a given speculation. Returns NULL if the specified speculation is not
2641 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2642 * the active CPU is not the specified CPU -- the speculation will be
2643 * atomically transitioned into the ACTIVEMANY state.
2645 static dtrace_buffer_t *
2646 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2647 dtrace_specid_t which)
2649 dtrace_speculation_t *spec;
2650 dtrace_speculation_state_t current, new = 0;
2651 dtrace_buffer_t *buf;
2656 if (which > state->dts_nspeculations) {
2657 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2661 spec = &state->dts_speculations[which - 1];
2662 buf = &spec->dtsp_buffer[cpuid];
2665 current = spec->dtsp_state;
2668 case DTRACESPEC_INACTIVE:
2669 case DTRACESPEC_COMMITTINGMANY:
2670 case DTRACESPEC_DISCARDING:
2673 case DTRACESPEC_COMMITTING:
2674 ASSERT(buf->dtb_offset == 0);
2677 case DTRACESPEC_ACTIVEONE:
2679 * This speculation is currently active on one CPU.
2680 * Check the offset in the buffer; if it's non-zero,
2681 * that CPU must be us (and we leave the state alone).
2682 * If it's zero, assume that we're starting on a new
2683 * CPU -- and change the state to indicate that the
2684 * speculation is active on more than one CPU.
2686 if (buf->dtb_offset != 0)
2689 new = DTRACESPEC_ACTIVEMANY;
2692 case DTRACESPEC_ACTIVEMANY:
2695 case DTRACESPEC_ACTIVE:
2696 new = DTRACESPEC_ACTIVEONE;
2702 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2703 current, new) != current);
2705 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2710 * Return a string. In the event that the user lacks the privilege to access
2711 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2712 * don't fail access checking.
2714 * dtrace_dif_variable() uses this routine as a helper for various
2715 * builtin values such as 'execname' and 'probefunc.'
2718 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2719 dtrace_mstate_t *mstate)
2721 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2726 * The easy case: this probe is allowed to read all of memory, so
2727 * we can just return this as a vanilla pointer.
2729 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2733 * This is the tougher case: we copy the string in question from
2734 * kernel memory into scratch memory and return it that way: this
2735 * ensures that we won't trip up when access checking tests the
2736 * BYREF return value.
2738 strsz = dtrace_strlen((char *)addr, size) + 1;
2740 if (mstate->dtms_scratch_ptr + strsz >
2741 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2742 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2746 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2748 ret = mstate->dtms_scratch_ptr;
2749 mstate->dtms_scratch_ptr += strsz;
2754 * Return a string from a memoy address which is known to have one or
2755 * more concatenated, individually zero terminated, sub-strings.
2756 * In the event that the user lacks the privilege to access
2757 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2758 * don't fail access checking.
2760 * dtrace_dif_variable() uses this routine as a helper for various
2761 * builtin values such as 'execargs'.
2764 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2765 dtrace_mstate_t *mstate)
2771 if (mstate->dtms_scratch_ptr + strsz >
2772 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2773 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2777 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2780 /* Replace sub-string termination characters with a space. */
2781 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2786 ret = mstate->dtms_scratch_ptr;
2787 mstate->dtms_scratch_ptr += strsz;
2792 * This function implements the DIF emulator's variable lookups. The emulator
2793 * passes a reserved variable identifier and optional built-in array index.
2796 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2800 * If we're accessing one of the uncached arguments, we'll turn this
2801 * into a reference in the args array.
2803 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2804 ndx = v - DIF_VAR_ARG0;
2810 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2811 if (ndx >= sizeof (mstate->dtms_arg) /
2812 sizeof (mstate->dtms_arg[0])) {
2813 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2814 dtrace_provider_t *pv;
2817 pv = mstate->dtms_probe->dtpr_provider;
2818 if (pv->dtpv_pops.dtps_getargval != NULL)
2819 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2820 mstate->dtms_probe->dtpr_id,
2821 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2823 val = dtrace_getarg(ndx, aframes);
2826 * This is regrettably required to keep the compiler
2827 * from tail-optimizing the call to dtrace_getarg().
2828 * The condition always evaluates to true, but the
2829 * compiler has no way of figuring that out a priori.
2830 * (None of this would be necessary if the compiler
2831 * could be relied upon to _always_ tail-optimize
2832 * the call to dtrace_getarg() -- but it can't.)
2834 if (mstate->dtms_probe != NULL)
2840 return (mstate->dtms_arg[ndx]);
2843 case DIF_VAR_UREGS: {
2846 if (!dtrace_priv_proc(state))
2849 if ((lwp = curthread->t_lwp) == NULL) {
2850 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2851 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2855 return (dtrace_getreg(lwp->lwp_regs, ndx));
2859 case DIF_VAR_UREGS: {
2860 struct trapframe *tframe;
2862 if (!dtrace_priv_proc(state))
2865 if ((tframe = curthread->td_frame) == NULL) {
2866 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2867 cpu_core[curcpu].cpuc_dtrace_illval = 0;
2871 return (dtrace_getreg(tframe, ndx));
2875 case DIF_VAR_CURTHREAD:
2876 if (!dtrace_priv_kernel(state))
2878 return ((uint64_t)(uintptr_t)curthread);
2880 case DIF_VAR_TIMESTAMP:
2881 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2882 mstate->dtms_timestamp = dtrace_gethrtime();
2883 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2885 return (mstate->dtms_timestamp);
2887 case DIF_VAR_VTIMESTAMP:
2888 ASSERT(dtrace_vtime_references != 0);
2889 return (curthread->t_dtrace_vtime);
2891 case DIF_VAR_WALLTIMESTAMP:
2892 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2893 mstate->dtms_walltimestamp = dtrace_gethrestime();
2894 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2896 return (mstate->dtms_walltimestamp);
2900 if (!dtrace_priv_kernel(state))
2902 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2903 mstate->dtms_ipl = dtrace_getipl();
2904 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2906 return (mstate->dtms_ipl);
2910 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2911 return (mstate->dtms_epid);
2914 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2915 return (mstate->dtms_probe->dtpr_id);
2917 case DIF_VAR_STACKDEPTH:
2918 if (!dtrace_priv_kernel(state))
2920 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2921 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2923 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2924 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2926 return (mstate->dtms_stackdepth);
2928 case DIF_VAR_USTACKDEPTH:
2929 if (!dtrace_priv_proc(state))
2931 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2933 * See comment in DIF_VAR_PID.
2935 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2937 mstate->dtms_ustackdepth = 0;
2939 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2940 mstate->dtms_ustackdepth =
2941 dtrace_getustackdepth();
2942 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2944 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2946 return (mstate->dtms_ustackdepth);
2948 case DIF_VAR_CALLER:
2949 if (!dtrace_priv_kernel(state))
2951 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2952 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2954 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2956 * If this is an unanchored probe, we are
2957 * required to go through the slow path:
2958 * dtrace_caller() only guarantees correct
2959 * results for anchored probes.
2961 pc_t caller[2] = {0, 0};
2963 dtrace_getpcstack(caller, 2, aframes,
2964 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2965 mstate->dtms_caller = caller[1];
2966 } else if ((mstate->dtms_caller =
2967 dtrace_caller(aframes)) == -1) {
2969 * We have failed to do this the quick way;
2970 * we must resort to the slower approach of
2971 * calling dtrace_getpcstack().
2975 dtrace_getpcstack(&caller, 1, aframes, NULL);
2976 mstate->dtms_caller = caller;
2979 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
2981 return (mstate->dtms_caller);
2983 case DIF_VAR_UCALLER:
2984 if (!dtrace_priv_proc(state))
2987 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
2991 * dtrace_getupcstack() fills in the first uint64_t
2992 * with the current PID. The second uint64_t will
2993 * be the program counter at user-level. The third
2994 * uint64_t will contain the caller, which is what
2998 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2999 dtrace_getupcstack(ustack, 3);
3000 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3001 mstate->dtms_ucaller = ustack[2];
3002 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3005 return (mstate->dtms_ucaller);
3007 case DIF_VAR_PROBEPROV:
3008 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3009 return (dtrace_dif_varstr(
3010 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3013 case DIF_VAR_PROBEMOD:
3014 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3015 return (dtrace_dif_varstr(
3016 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3019 case DIF_VAR_PROBEFUNC:
3020 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3021 return (dtrace_dif_varstr(
3022 (uintptr_t)mstate->dtms_probe->dtpr_func,
3025 case DIF_VAR_PROBENAME:
3026 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3027 return (dtrace_dif_varstr(
3028 (uintptr_t)mstate->dtms_probe->dtpr_name,
3032 if (!dtrace_priv_proc(state))
3037 * Note that we are assuming that an unanchored probe is
3038 * always due to a high-level interrupt. (And we're assuming
3039 * that there is only a single high level interrupt.)
3041 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3042 return (pid0.pid_id);
3045 * It is always safe to dereference one's own t_procp pointer:
3046 * it always points to a valid, allocated proc structure.
3047 * Further, it is always safe to dereference the p_pidp member
3048 * of one's own proc structure. (These are truisms becuase
3049 * threads and processes don't clean up their own state --
3050 * they leave that task to whomever reaps them.)
3052 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3054 return ((uint64_t)curproc->p_pid);
3058 if (!dtrace_priv_proc(state))
3063 * See comment in DIF_VAR_PID.
3065 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3066 return (pid0.pid_id);
3069 * It is always safe to dereference one's own t_procp pointer:
3070 * it always points to a valid, allocated proc structure.
3071 * (This is true because threads don't clean up their own
3072 * state -- they leave that task to whomever reaps them.)
3074 return ((uint64_t)curthread->t_procp->p_ppid);
3076 return ((uint64_t)curproc->p_pptr->p_pid);
3082 * See comment in DIF_VAR_PID.
3084 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3088 return ((uint64_t)curthread->t_tid);
3090 case DIF_VAR_EXECARGS: {
3091 struct pargs *p_args = curthread->td_proc->p_args;
3096 return (dtrace_dif_varstrz(
3097 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3100 case DIF_VAR_EXECNAME:
3102 if (!dtrace_priv_proc(state))
3106 * See comment in DIF_VAR_PID.
3108 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3109 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3112 * It is always safe to dereference one's own t_procp pointer:
3113 * it always points to a valid, allocated proc structure.
3114 * (This is true because threads don't clean up their own
3115 * state -- they leave that task to whomever reaps them.)
3117 return (dtrace_dif_varstr(
3118 (uintptr_t)curthread->t_procp->p_user.u_comm,
3121 return (dtrace_dif_varstr(
3122 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3125 case DIF_VAR_ZONENAME:
3127 if (!dtrace_priv_proc(state))
3131 * See comment in DIF_VAR_PID.
3133 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3134 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3137 * It is always safe to dereference one's own t_procp pointer:
3138 * it always points to a valid, allocated proc structure.
3139 * (This is true because threads don't clean up their own
3140 * state -- they leave that task to whomever reaps them.)
3142 return (dtrace_dif_varstr(
3143 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3150 if (!dtrace_priv_proc(state))
3155 * See comment in DIF_VAR_PID.
3157 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3158 return ((uint64_t)p0.p_cred->cr_uid);
3162 * It is always safe to dereference one's own t_procp pointer:
3163 * it always points to a valid, allocated proc structure.
3164 * (This is true because threads don't clean up their own
3165 * state -- they leave that task to whomever reaps them.)
3167 * Additionally, it is safe to dereference one's own process
3168 * credential, since this is never NULL after process birth.
3170 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3173 if (!dtrace_priv_proc(state))
3178 * See comment in DIF_VAR_PID.
3180 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3181 return ((uint64_t)p0.p_cred->cr_gid);
3185 * It is always safe to dereference one's own t_procp pointer:
3186 * it always points to a valid, allocated proc structure.
3187 * (This is true because threads don't clean up their own
3188 * state -- they leave that task to whomever reaps them.)
3190 * Additionally, it is safe to dereference one's own process
3191 * credential, since this is never NULL after process birth.
3193 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3195 case DIF_VAR_ERRNO: {
3198 if (!dtrace_priv_proc(state))
3202 * See comment in DIF_VAR_PID.
3204 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3208 * It is always safe to dereference one's own t_lwp pointer in
3209 * the event that this pointer is non-NULL. (This is true
3210 * because threads and lwps don't clean up their own state --
3211 * they leave that task to whomever reaps them.)
3213 if ((lwp = curthread->t_lwp) == NULL)
3216 return ((uint64_t)lwp->lwp_errno);
3218 return (curthread->td_errno);
3227 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3233 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3234 * Notice that we don't bother validating the proper number of arguments or
3235 * their types in the tuple stack. This isn't needed because all argument
3236 * interpretation is safe because of our load safety -- the worst that can
3237 * happen is that a bogus program can obtain bogus results.
3240 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3241 dtrace_key_t *tupregs, int nargs,
3242 dtrace_mstate_t *mstate, dtrace_state_t *state)
3244 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3245 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3246 dtrace_vstate_t *vstate = &state->dts_vstate;
3259 struct thread *lowner;
3261 struct lock_object *li;
3268 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3272 case DIF_SUBR_MUTEX_OWNED:
3273 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3279 m.mx = dtrace_load64(tupregs[0].dttk_value);
3280 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3281 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3283 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3286 case DIF_SUBR_MUTEX_OWNER:
3287 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3293 m.mx = dtrace_load64(tupregs[0].dttk_value);
3294 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3295 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3296 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3301 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3302 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3308 m.mx = dtrace_load64(tupregs[0].dttk_value);
3309 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3312 case DIF_SUBR_MUTEX_TYPE_SPIN:
3313 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3319 m.mx = dtrace_load64(tupregs[0].dttk_value);
3320 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3323 case DIF_SUBR_RW_READ_HELD: {
3326 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3332 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3333 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3337 case DIF_SUBR_RW_WRITE_HELD:
3338 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3344 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3345 regs[rd] = _RW_WRITE_HELD(&r.ri);
3348 case DIF_SUBR_RW_ISWRITER:
3349 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3355 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3356 regs[rd] = _RW_ISWRITER(&r.ri);
3360 case DIF_SUBR_MUTEX_OWNED:
3361 if (!dtrace_canload(tupregs[0].dttk_value,
3362 sizeof (struct lock_object), mstate, vstate)) {
3366 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3367 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3370 case DIF_SUBR_MUTEX_OWNER:
3371 if (!dtrace_canload(tupregs[0].dttk_value,
3372 sizeof (struct lock_object), mstate, vstate)) {
3376 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3377 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3378 regs[rd] = (uintptr_t)lowner;
3381 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3382 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3387 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3388 /* XXX - should be only LC_SLEEPABLE? */
3389 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3390 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3393 case DIF_SUBR_MUTEX_TYPE_SPIN:
3394 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3399 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3400 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3403 case DIF_SUBR_RW_READ_HELD:
3404 case DIF_SUBR_SX_SHARED_HELD:
3405 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3410 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3411 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3415 case DIF_SUBR_RW_WRITE_HELD:
3416 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3417 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3422 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3423 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3424 regs[rd] = (lowner == curthread);
3427 case DIF_SUBR_RW_ISWRITER:
3428 case DIF_SUBR_SX_ISEXCLUSIVE:
3429 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3434 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3435 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3438 #endif /* ! defined(sun) */
3440 case DIF_SUBR_BCOPY: {
3442 * We need to be sure that the destination is in the scratch
3443 * region -- no other region is allowed.
3445 uintptr_t src = tupregs[0].dttk_value;
3446 uintptr_t dest = tupregs[1].dttk_value;
3447 size_t size = tupregs[2].dttk_value;
3449 if (!dtrace_inscratch(dest, size, mstate)) {
3450 *flags |= CPU_DTRACE_BADADDR;
3455 if (!dtrace_canload(src, size, mstate, vstate)) {
3460 dtrace_bcopy((void *)src, (void *)dest, size);
3464 case DIF_SUBR_ALLOCA:
3465 case DIF_SUBR_COPYIN: {
3466 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3468 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3469 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3472 * This action doesn't require any credential checks since
3473 * probes will not activate in user contexts to which the
3474 * enabling user does not have permissions.
3478 * Rounding up the user allocation size could have overflowed
3479 * a large, bogus allocation (like -1ULL) to 0.
3481 if (scratch_size < size ||
3482 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3483 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3488 if (subr == DIF_SUBR_COPYIN) {
3489 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3490 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3491 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3494 mstate->dtms_scratch_ptr += scratch_size;
3499 case DIF_SUBR_COPYINTO: {
3500 uint64_t size = tupregs[1].dttk_value;
3501 uintptr_t dest = tupregs[2].dttk_value;
3504 * This action doesn't require any credential checks since
3505 * probes will not activate in user contexts to which the
3506 * enabling user does not have permissions.
3508 if (!dtrace_inscratch(dest, size, mstate)) {
3509 *flags |= CPU_DTRACE_BADADDR;
3514 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3515 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3516 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3520 case DIF_SUBR_COPYINSTR: {
3521 uintptr_t dest = mstate->dtms_scratch_ptr;
3522 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3524 if (nargs > 1 && tupregs[1].dttk_value < size)
3525 size = tupregs[1].dttk_value + 1;
3528 * This action doesn't require any credential checks since
3529 * probes will not activate in user contexts to which the
3530 * enabling user does not have permissions.
3532 if (!DTRACE_INSCRATCH(mstate, size)) {
3533 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3538 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3539 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3540 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3542 ((char *)dest)[size - 1] = '\0';
3543 mstate->dtms_scratch_ptr += size;
3549 case DIF_SUBR_MSGSIZE:
3550 case DIF_SUBR_MSGDSIZE: {
3551 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3552 uintptr_t wptr, rptr;
3556 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3558 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3564 wptr = dtrace_loadptr(baddr +
3565 offsetof(mblk_t, b_wptr));
3567 rptr = dtrace_loadptr(baddr +
3568 offsetof(mblk_t, b_rptr));
3571 *flags |= CPU_DTRACE_BADADDR;
3572 *illval = tupregs[0].dttk_value;
3576 daddr = dtrace_loadptr(baddr +
3577 offsetof(mblk_t, b_datap));
3579 baddr = dtrace_loadptr(baddr +
3580 offsetof(mblk_t, b_cont));
3583 * We want to prevent against denial-of-service here,
3584 * so we're only going to search the list for
3585 * dtrace_msgdsize_max mblks.
3587 if (cont++ > dtrace_msgdsize_max) {
3588 *flags |= CPU_DTRACE_ILLOP;
3592 if (subr == DIF_SUBR_MSGDSIZE) {
3593 if (dtrace_load8(daddr +
3594 offsetof(dblk_t, db_type)) != M_DATA)
3598 count += wptr - rptr;
3601 if (!(*flags & CPU_DTRACE_FAULT))
3608 case DIF_SUBR_PROGENYOF: {
3609 pid_t pid = tupregs[0].dttk_value;
3613 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3615 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3617 if (p->p_pidp->pid_id == pid) {
3619 if (p->p_pid == pid) {
3626 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3632 case DIF_SUBR_SPECULATION:
3633 regs[rd] = dtrace_speculation(state);
3636 case DIF_SUBR_COPYOUT: {
3637 uintptr_t kaddr = tupregs[0].dttk_value;
3638 uintptr_t uaddr = tupregs[1].dttk_value;
3639 uint64_t size = tupregs[2].dttk_value;
3641 if (!dtrace_destructive_disallow &&
3642 dtrace_priv_proc_control(state) &&
3643 !dtrace_istoxic(kaddr, size)) {
3644 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3645 dtrace_copyout(kaddr, uaddr, size, flags);
3646 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3651 case DIF_SUBR_COPYOUTSTR: {
3652 uintptr_t kaddr = tupregs[0].dttk_value;
3653 uintptr_t uaddr = tupregs[1].dttk_value;
3654 uint64_t size = tupregs[2].dttk_value;
3656 if (!dtrace_destructive_disallow &&
3657 dtrace_priv_proc_control(state) &&
3658 !dtrace_istoxic(kaddr, size)) {
3659 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3660 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3661 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3666 case DIF_SUBR_STRLEN: {
3668 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3669 sz = dtrace_strlen((char *)addr,
3670 state->dts_options[DTRACEOPT_STRSIZE]);
3672 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3682 case DIF_SUBR_STRCHR:
3683 case DIF_SUBR_STRRCHR: {
3685 * We're going to iterate over the string looking for the
3686 * specified character. We will iterate until we have reached
3687 * the string length or we have found the character. If this
3688 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3689 * of the specified character instead of the first.
3691 uintptr_t saddr = tupregs[0].dttk_value;
3692 uintptr_t addr = tupregs[0].dttk_value;
3693 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3694 char c, target = (char)tupregs[1].dttk_value;
3696 for (regs[rd] = 0; addr < limit; addr++) {
3697 if ((c = dtrace_load8(addr)) == target) {
3700 if (subr == DIF_SUBR_STRCHR)
3708 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3716 case DIF_SUBR_STRSTR:
3717 case DIF_SUBR_INDEX:
3718 case DIF_SUBR_RINDEX: {
3720 * We're going to iterate over the string looking for the
3721 * specified string. We will iterate until we have reached
3722 * the string length or we have found the string. (Yes, this
3723 * is done in the most naive way possible -- but considering
3724 * that the string we're searching for is likely to be
3725 * relatively short, the complexity of Rabin-Karp or similar
3726 * hardly seems merited.)
3728 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3729 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3730 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3731 size_t len = dtrace_strlen(addr, size);
3732 size_t sublen = dtrace_strlen(substr, size);
3733 char *limit = addr + len, *orig = addr;
3734 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3737 regs[rd] = notfound;
3739 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3744 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3751 * strstr() and index()/rindex() have similar semantics if
3752 * both strings are the empty string: strstr() returns a
3753 * pointer to the (empty) string, and index() and rindex()
3754 * both return index 0 (regardless of any position argument).
3756 if (sublen == 0 && len == 0) {
3757 if (subr == DIF_SUBR_STRSTR)
3758 regs[rd] = (uintptr_t)addr;
3764 if (subr != DIF_SUBR_STRSTR) {
3765 if (subr == DIF_SUBR_RINDEX) {
3772 * Both index() and rindex() take an optional position
3773 * argument that denotes the starting position.
3776 int64_t pos = (int64_t)tupregs[2].dttk_value;
3779 * If the position argument to index() is
3780 * negative, Perl implicitly clamps it at
3781 * zero. This semantic is a little surprising
3782 * given the special meaning of negative
3783 * positions to similar Perl functions like
3784 * substr(), but it appears to reflect a
3785 * notion that index() can start from a
3786 * negative index and increment its way up to
3787 * the string. Given this notion, Perl's
3788 * rindex() is at least self-consistent in
3789 * that it implicitly clamps positions greater
3790 * than the string length to be the string
3791 * length. Where Perl completely loses
3792 * coherence, however, is when the specified
3793 * substring is the empty string (""). In
3794 * this case, even if the position is
3795 * negative, rindex() returns 0 -- and even if
3796 * the position is greater than the length,
3797 * index() returns the string length. These
3798 * semantics violate the notion that index()
3799 * should never return a value less than the
3800 * specified position and that rindex() should
3801 * never return a value greater than the
3802 * specified position. (One assumes that
3803 * these semantics are artifacts of Perl's
3804 * implementation and not the results of
3805 * deliberate design -- it beggars belief that
3806 * even Larry Wall could desire such oddness.)
3807 * While in the abstract one would wish for
3808 * consistent position semantics across
3809 * substr(), index() and rindex() -- or at the
3810 * very least self-consistent position
3811 * semantics for index() and rindex() -- we
3812 * instead opt to keep with the extant Perl
3813 * semantics, in all their broken glory. (Do
3814 * we have more desire to maintain Perl's
3815 * semantics than Perl does? Probably.)
3817 if (subr == DIF_SUBR_RINDEX) {
3841 for (regs[rd] = notfound; addr != limit; addr += inc) {
3842 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3843 if (subr != DIF_SUBR_STRSTR) {
3845 * As D index() and rindex() are
3846 * modeled on Perl (and not on awk),
3847 * we return a zero-based (and not a
3848 * one-based) index. (For you Perl
3849 * weenies: no, we're not going to add
3850 * $[ -- and shouldn't you be at a con
3853 regs[rd] = (uintptr_t)(addr - orig);
3857 ASSERT(subr == DIF_SUBR_STRSTR);
3858 regs[rd] = (uintptr_t)addr;
3866 case DIF_SUBR_STRTOK: {
3867 uintptr_t addr = tupregs[0].dttk_value;
3868 uintptr_t tokaddr = tupregs[1].dttk_value;
3869 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3870 uintptr_t limit, toklimit = tokaddr + size;
3871 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3872 char *dest = (char *)mstate->dtms_scratch_ptr;
3876 * Check both the token buffer and (later) the input buffer,
3877 * since both could be non-scratch addresses.
3879 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3884 if (!DTRACE_INSCRATCH(mstate, size)) {
3885 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3892 * If the address specified is NULL, we use our saved
3893 * strtok pointer from the mstate. Note that this
3894 * means that the saved strtok pointer is _only_
3895 * valid within multiple enablings of the same probe --
3896 * it behaves like an implicit clause-local variable.
3898 addr = mstate->dtms_strtok;
3901 * If the user-specified address is non-NULL we must
3902 * access check it. This is the only time we have
3903 * a chance to do so, since this address may reside
3904 * in the string table of this clause-- future calls
3905 * (when we fetch addr from mstate->dtms_strtok)
3906 * would fail this access check.
3908 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3915 * First, zero the token map, and then process the token
3916 * string -- setting a bit in the map for every character
3917 * found in the token string.
3919 for (i = 0; i < sizeof (tokmap); i++)
3922 for (; tokaddr < toklimit; tokaddr++) {
3923 if ((c = dtrace_load8(tokaddr)) == '\0')
3926 ASSERT((c >> 3) < sizeof (tokmap));
3927 tokmap[c >> 3] |= (1 << (c & 0x7));
3930 for (limit = addr + size; addr < limit; addr++) {
3932 * We're looking for a character that is _not_ contained
3933 * in the token string.
3935 if ((c = dtrace_load8(addr)) == '\0')
3938 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3944 * We reached the end of the string without finding
3945 * any character that was not in the token string.
3946 * We return NULL in this case, and we set the saved
3947 * address to NULL as well.
3950 mstate->dtms_strtok = 0;
3955 * From here on, we're copying into the destination string.
3957 for (i = 0; addr < limit && i < size - 1; addr++) {
3958 if ((c = dtrace_load8(addr)) == '\0')
3961 if (tokmap[c >> 3] & (1 << (c & 0x7)))
3970 regs[rd] = (uintptr_t)dest;
3971 mstate->dtms_scratch_ptr += size;
3972 mstate->dtms_strtok = addr;
3976 case DIF_SUBR_SUBSTR: {
3977 uintptr_t s = tupregs[0].dttk_value;
3978 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3979 char *d = (char *)mstate->dtms_scratch_ptr;
3980 int64_t index = (int64_t)tupregs[1].dttk_value;
3981 int64_t remaining = (int64_t)tupregs[2].dttk_value;
3982 size_t len = dtrace_strlen((char *)s, size);
3985 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
3990 if (!DTRACE_INSCRATCH(mstate, size)) {
3991 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3997 remaining = (int64_t)size;
4002 if (index < 0 && index + remaining > 0) {
4008 if (index >= len || index < 0) {
4010 } else if (remaining < 0) {
4011 remaining += len - index;
4012 } else if (index + remaining > size) {
4013 remaining = size - index;
4016 for (i = 0; i < remaining; i++) {
4017 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4023 mstate->dtms_scratch_ptr += size;
4024 regs[rd] = (uintptr_t)d;
4028 case DIF_SUBR_TOUPPER:
4029 case DIF_SUBR_TOLOWER: {
4030 uintptr_t s = tupregs[0].dttk_value;
4031 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4032 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4033 size_t len = dtrace_strlen((char *)s, size);
4034 char lower, upper, convert;
4037 if (subr == DIF_SUBR_TOUPPER) {
4047 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4052 if (!DTRACE_INSCRATCH(mstate, size)) {
4053 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4058 for (i = 0; i < size - 1; i++) {
4059 if ((c = dtrace_load8(s + i)) == '\0')
4062 if (c >= lower && c <= upper)
4063 c = convert + (c - lower);
4070 regs[rd] = (uintptr_t)dest;
4071 mstate->dtms_scratch_ptr += size;
4076 case DIF_SUBR_GETMAJOR:
4078 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4080 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4084 case DIF_SUBR_GETMINOR:
4086 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4088 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4092 case DIF_SUBR_DDI_PATHNAME: {
4094 * This one is a galactic mess. We are going to roughly
4095 * emulate ddi_pathname(), but it's made more complicated
4096 * by the fact that we (a) want to include the minor name and
4097 * (b) must proceed iteratively instead of recursively.
4099 uintptr_t dest = mstate->dtms_scratch_ptr;
4100 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4101 char *start = (char *)dest, *end = start + size - 1;
4102 uintptr_t daddr = tupregs[0].dttk_value;
4103 int64_t minor = (int64_t)tupregs[1].dttk_value;
4105 int i, len, depth = 0;
4108 * Due to all the pointer jumping we do and context we must
4109 * rely upon, we just mandate that the user must have kernel
4110 * read privileges to use this routine.
4112 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4113 *flags |= CPU_DTRACE_KPRIV;
4118 if (!DTRACE_INSCRATCH(mstate, size)) {
4119 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4127 * We want to have a name for the minor. In order to do this,
4128 * we need to walk the minor list from the devinfo. We want
4129 * to be sure that we don't infinitely walk a circular list,
4130 * so we check for circularity by sending a scout pointer
4131 * ahead two elements for every element that we iterate over;
4132 * if the list is circular, these will ultimately point to the
4133 * same element. You may recognize this little trick as the
4134 * answer to a stupid interview question -- one that always
4135 * seems to be asked by those who had to have it laboriously
4136 * explained to them, and who can't even concisely describe
4137 * the conditions under which one would be forced to resort to
4138 * this technique. Needless to say, those conditions are
4139 * found here -- and probably only here. Is this the only use
4140 * of this infamous trick in shipping, production code? If it
4141 * isn't, it probably should be...
4144 uintptr_t maddr = dtrace_loadptr(daddr +
4145 offsetof(struct dev_info, devi_minor));
4147 uintptr_t next = offsetof(struct ddi_minor_data, next);
4148 uintptr_t name = offsetof(struct ddi_minor_data,
4149 d_minor) + offsetof(struct ddi_minor, name);
4150 uintptr_t dev = offsetof(struct ddi_minor_data,
4151 d_minor) + offsetof(struct ddi_minor, dev);
4155 scout = dtrace_loadptr(maddr + next);
4157 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4160 m = dtrace_load64(maddr + dev) & MAXMIN64;
4162 m = dtrace_load32(maddr + dev) & MAXMIN;
4165 maddr = dtrace_loadptr(maddr + next);
4170 scout = dtrace_loadptr(scout + next);
4175 scout = dtrace_loadptr(scout + next);
4180 if (scout == maddr) {
4181 *flags |= CPU_DTRACE_ILLOP;
4189 * We have the minor data. Now we need to
4190 * copy the minor's name into the end of the
4193 s = (char *)dtrace_loadptr(maddr + name);
4194 len = dtrace_strlen(s, size);
4196 if (*flags & CPU_DTRACE_FAULT)
4200 if ((end -= (len + 1)) < start)
4206 for (i = 1; i <= len; i++)
4207 end[i] = dtrace_load8((uintptr_t)s++);
4212 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4213 ddi_node_state_t devi_state;
4215 devi_state = dtrace_load32(daddr +
4216 offsetof(struct dev_info, devi_node_state));
4218 if (*flags & CPU_DTRACE_FAULT)
4221 if (devi_state >= DS_INITIALIZED) {
4222 s = (char *)dtrace_loadptr(daddr +
4223 offsetof(struct dev_info, devi_addr));
4224 len = dtrace_strlen(s, size);
4226 if (*flags & CPU_DTRACE_FAULT)
4230 if ((end -= (len + 1)) < start)
4236 for (i = 1; i <= len; i++)
4237 end[i] = dtrace_load8((uintptr_t)s++);
4241 * Now for the node name...
4243 s = (char *)dtrace_loadptr(daddr +
4244 offsetof(struct dev_info, devi_node_name));
4246 daddr = dtrace_loadptr(daddr +
4247 offsetof(struct dev_info, devi_parent));
4250 * If our parent is NULL (that is, if we're the root
4251 * node), we're going to use the special path
4257 len = dtrace_strlen(s, size);
4258 if (*flags & CPU_DTRACE_FAULT)
4261 if ((end -= (len + 1)) < start)
4264 for (i = 1; i <= len; i++)
4265 end[i] = dtrace_load8((uintptr_t)s++);
4268 if (depth++ > dtrace_devdepth_max) {
4269 *flags |= CPU_DTRACE_ILLOP;
4275 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4278 regs[rd] = (uintptr_t)end;
4279 mstate->dtms_scratch_ptr += size;
4286 case DIF_SUBR_STRJOIN: {
4287 char *d = (char *)mstate->dtms_scratch_ptr;
4288 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4289 uintptr_t s1 = tupregs[0].dttk_value;
4290 uintptr_t s2 = tupregs[1].dttk_value;
4293 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4294 !dtrace_strcanload(s2, size, mstate, vstate)) {
4299 if (!DTRACE_INSCRATCH(mstate, size)) {
4300 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4307 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4312 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4320 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4325 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4330 mstate->dtms_scratch_ptr += i;
4331 regs[rd] = (uintptr_t)d;
4337 case DIF_SUBR_LLTOSTR: {
4338 int64_t i = (int64_t)tupregs[0].dttk_value;
4339 uint64_t val, digit;
4340 uint64_t size = 65; /* enough room for 2^64 in binary */
4341 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4345 if ((base = tupregs[1].dttk_value) <= 1 ||
4346 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4347 *flags |= CPU_DTRACE_ILLOP;
4352 val = (base == 10 && i < 0) ? i * -1 : i;
4354 if (!DTRACE_INSCRATCH(mstate, size)) {
4355 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4360 for (*end-- = '\0'; val; val /= base) {
4361 if ((digit = val % base) <= '9' - '0') {
4362 *end-- = '0' + digit;
4364 *end-- = 'a' + (digit - ('9' - '0') - 1);
4368 if (i == 0 && base == 16)
4374 if (i == 0 || base == 8 || base == 16)
4377 if (i < 0 && base == 10)
4380 regs[rd] = (uintptr_t)end + 1;
4381 mstate->dtms_scratch_ptr += size;
4385 case DIF_SUBR_HTONS:
4386 case DIF_SUBR_NTOHS:
4387 #if BYTE_ORDER == BIG_ENDIAN
4388 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4390 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4395 case DIF_SUBR_HTONL:
4396 case DIF_SUBR_NTOHL:
4397 #if BYTE_ORDER == BIG_ENDIAN
4398 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4400 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4405 case DIF_SUBR_HTONLL:
4406 case DIF_SUBR_NTOHLL:
4407 #if BYTE_ORDER == BIG_ENDIAN
4408 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4410 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4415 case DIF_SUBR_DIRNAME:
4416 case DIF_SUBR_BASENAME: {
4417 char *dest = (char *)mstate->dtms_scratch_ptr;
4418 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4419 uintptr_t src = tupregs[0].dttk_value;
4420 int i, j, len = dtrace_strlen((char *)src, size);
4421 int lastbase = -1, firstbase = -1, lastdir = -1;
4424 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4429 if (!DTRACE_INSCRATCH(mstate, size)) {
4430 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4436 * The basename and dirname for a zero-length string is
4441 src = (uintptr_t)".";
4445 * Start from the back of the string, moving back toward the
4446 * front until we see a character that isn't a slash. That
4447 * character is the last character in the basename.
4449 for (i = len - 1; i >= 0; i--) {
4450 if (dtrace_load8(src + i) != '/')
4458 * Starting from the last character in the basename, move
4459 * towards the front until we find a slash. The character
4460 * that we processed immediately before that is the first
4461 * character in the basename.
4463 for (; i >= 0; i--) {
4464 if (dtrace_load8(src + i) == '/')
4472 * Now keep going until we find a non-slash character. That
4473 * character is the last character in the dirname.
4475 for (; i >= 0; i--) {
4476 if (dtrace_load8(src + i) != '/')
4483 ASSERT(!(lastbase == -1 && firstbase != -1));
4484 ASSERT(!(firstbase == -1 && lastdir != -1));
4486 if (lastbase == -1) {
4488 * We didn't find a non-slash character. We know that
4489 * the length is non-zero, so the whole string must be
4490 * slashes. In either the dirname or the basename
4491 * case, we return '/'.
4493 ASSERT(firstbase == -1);
4494 firstbase = lastbase = lastdir = 0;
4497 if (firstbase == -1) {
4499 * The entire string consists only of a basename
4500 * component. If we're looking for dirname, we need
4501 * to change our string to be just "."; if we're
4502 * looking for a basename, we'll just set the first
4503 * character of the basename to be 0.
4505 if (subr == DIF_SUBR_DIRNAME) {
4506 ASSERT(lastdir == -1);
4507 src = (uintptr_t)".";
4514 if (subr == DIF_SUBR_DIRNAME) {
4515 if (lastdir == -1) {
4517 * We know that we have a slash in the name --
4518 * or lastdir would be set to 0, above. And
4519 * because lastdir is -1, we know that this
4520 * slash must be the first character. (That
4521 * is, the full string must be of the form
4522 * "/basename".) In this case, the last
4523 * character of the directory name is 0.
4531 ASSERT(subr == DIF_SUBR_BASENAME);
4532 ASSERT(firstbase != -1 && lastbase != -1);
4537 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4538 dest[j] = dtrace_load8(src + i);
4541 regs[rd] = (uintptr_t)dest;
4542 mstate->dtms_scratch_ptr += size;
4546 case DIF_SUBR_CLEANPATH: {
4547 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4548 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4549 uintptr_t src = tupregs[0].dttk_value;
4552 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4557 if (!DTRACE_INSCRATCH(mstate, size)) {
4558 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4564 * Move forward, loading each character.
4567 c = dtrace_load8(src + i++);
4569 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4577 c = dtrace_load8(src + i++);
4581 * We have two slashes -- we can just advance
4582 * to the next character.
4589 * This is not "." and it's not ".." -- we can
4590 * just store the "/" and this character and
4598 c = dtrace_load8(src + i++);
4602 * This is a "/./" component. We're not going
4603 * to store anything in the destination buffer;
4604 * we're just going to go to the next component.
4611 * This is not ".." -- we can just store the
4612 * "/." and this character and continue
4621 c = dtrace_load8(src + i++);
4623 if (c != '/' && c != '\0') {
4625 * This is not ".." -- it's "..[mumble]".
4626 * We'll store the "/.." and this character
4627 * and continue processing.
4637 * This is "/../" or "/..\0". We need to back up
4638 * our destination pointer until we find a "/".
4641 while (j != 0 && dest[--j] != '/')
4646 } while (c != '\0');
4649 regs[rd] = (uintptr_t)dest;
4650 mstate->dtms_scratch_ptr += size;
4654 case DIF_SUBR_INET_NTOA:
4655 case DIF_SUBR_INET_NTOA6:
4656 case DIF_SUBR_INET_NTOP: {
4661 if (subr == DIF_SUBR_INET_NTOP) {
4662 af = (int)tupregs[0].dttk_value;
4665 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4669 if (af == AF_INET) {
4674 * Safely load the IPv4 address.
4676 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4679 * Check an IPv4 string will fit in scratch.
4681 size = INET_ADDRSTRLEN;
4682 if (!DTRACE_INSCRATCH(mstate, size)) {
4683 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4687 base = (char *)mstate->dtms_scratch_ptr;
4688 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4691 * Stringify as a dotted decimal quad.
4694 ptr8 = (uint8_t *)&ip4;
4695 for (i = 3; i >= 0; i--) {
4701 for (; val; val /= 10) {
4702 *end-- = '0' + (val % 10);
4709 ASSERT(end + 1 >= base);
4711 } else if (af == AF_INET6) {
4712 struct in6_addr ip6;
4713 int firstzero, tryzero, numzero, v6end;
4715 const char digits[] = "0123456789abcdef";
4718 * Stringify using RFC 1884 convention 2 - 16 bit
4719 * hexadecimal values with a zero-run compression.
4720 * Lower case hexadecimal digits are used.
4721 * eg, fe80::214:4fff:fe0b:76c8.
4722 * The IPv4 embedded form is returned for inet_ntop,
4723 * just the IPv4 string is returned for inet_ntoa6.
4727 * Safely load the IPv6 address.
4730 (void *)(uintptr_t)tupregs[argi].dttk_value,
4731 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4734 * Check an IPv6 string will fit in scratch.
4736 size = INET6_ADDRSTRLEN;
4737 if (!DTRACE_INSCRATCH(mstate, size)) {
4738 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4742 base = (char *)mstate->dtms_scratch_ptr;
4743 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4747 * Find the longest run of 16 bit zero values
4748 * for the single allowed zero compression - "::".
4753 for (i = 0; i < sizeof (struct in6_addr); i++) {
4755 if (ip6._S6_un._S6_u8[i] == 0 &&
4757 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4759 tryzero == -1 && i % 2 == 0) {
4764 if (tryzero != -1 &&
4766 (ip6._S6_un._S6_u8[i] != 0 ||
4768 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4770 i == sizeof (struct in6_addr) - 1)) {
4772 if (i - tryzero <= numzero) {
4777 firstzero = tryzero;
4778 numzero = i - i % 2 - tryzero;
4782 if (ip6._S6_un._S6_u8[i] == 0 &&
4784 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4786 i == sizeof (struct in6_addr) - 1)
4790 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4793 * Check for an IPv4 embedded address.
4795 v6end = sizeof (struct in6_addr) - 2;
4796 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4797 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4798 for (i = sizeof (struct in6_addr) - 1;
4799 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4800 ASSERT(end >= base);
4803 val = ip6._S6_un._S6_u8[i];
4805 val = ip6.__u6_addr.__u6_addr8[i];
4811 for (; val; val /= 10) {
4812 *end-- = '0' + val % 10;
4816 if (i > DTRACE_V4MAPPED_OFFSET)
4820 if (subr == DIF_SUBR_INET_NTOA6)
4824 * Set v6end to skip the IPv4 address that
4825 * we have already stringified.
4831 * Build the IPv6 string by working through the
4832 * address in reverse.
4834 for (i = v6end; i >= 0; i -= 2) {
4835 ASSERT(end >= base);
4837 if (i == firstzero + numzero - 2) {
4844 if (i < 14 && i != firstzero - 2)
4848 val = (ip6._S6_un._S6_u8[i] << 8) +
4849 ip6._S6_un._S6_u8[i + 1];
4851 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4852 ip6.__u6_addr.__u6_addr8[i + 1];
4858 for (; val; val /= 16) {
4859 *end-- = digits[val % 16];
4863 ASSERT(end + 1 >= base);
4867 * The user didn't use AH_INET or AH_INET6.
4869 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4874 inetout: regs[rd] = (uintptr_t)end + 1;
4875 mstate->dtms_scratch_ptr += size;
4879 case DIF_SUBR_MEMREF: {
4880 uintptr_t size = 2 * sizeof(uintptr_t);
4881 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4882 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4884 /* address and length */
4885 memref[0] = tupregs[0].dttk_value;
4886 memref[1] = tupregs[1].dttk_value;
4888 regs[rd] = (uintptr_t) memref;
4889 mstate->dtms_scratch_ptr += scratch_size;
4893 case DIF_SUBR_TYPEREF: {
4894 uintptr_t size = 4 * sizeof(uintptr_t);
4895 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4896 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4898 /* address, num_elements, type_str, type_len */
4899 typeref[0] = tupregs[0].dttk_value;
4900 typeref[1] = tupregs[1].dttk_value;
4901 typeref[2] = tupregs[2].dttk_value;
4902 typeref[3] = tupregs[3].dttk_value;
4904 regs[rd] = (uintptr_t) typeref;
4905 mstate->dtms_scratch_ptr += scratch_size;
4912 * Emulate the execution of DTrace IR instructions specified by the given
4913 * DIF object. This function is deliberately void of assertions as all of
4914 * the necessary checks are handled by a call to dtrace_difo_validate().
4917 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4918 dtrace_vstate_t *vstate, dtrace_state_t *state)
4920 const dif_instr_t *text = difo->dtdo_buf;
4921 const uint_t textlen = difo->dtdo_len;
4922 const char *strtab = difo->dtdo_strtab;
4923 const uint64_t *inttab = difo->dtdo_inttab;
4926 dtrace_statvar_t *svar;
4927 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
4929 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4930 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4932 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
4933 uint64_t regs[DIF_DIR_NREGS];
4936 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
4938 uint_t pc = 0, id, opc = 0;
4944 * We stash the current DIF object into the machine state: we need it
4945 * for subsequent access checking.
4947 mstate->dtms_difo = difo;
4949 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
4951 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
4955 r1 = DIF_INSTR_R1(instr);
4956 r2 = DIF_INSTR_R2(instr);
4957 rd = DIF_INSTR_RD(instr);
4959 switch (DIF_INSTR_OP(instr)) {
4961 regs[rd] = regs[r1] | regs[r2];
4964 regs[rd] = regs[r1] ^ regs[r2];
4967 regs[rd] = regs[r1] & regs[r2];
4970 regs[rd] = regs[r1] << regs[r2];
4973 regs[rd] = regs[r1] >> regs[r2];
4976 regs[rd] = regs[r1] - regs[r2];
4979 regs[rd] = regs[r1] + regs[r2];
4982 regs[rd] = regs[r1] * regs[r2];
4985 if (regs[r2] == 0) {
4987 *flags |= CPU_DTRACE_DIVZERO;
4989 regs[rd] = (int64_t)regs[r1] /
4995 if (regs[r2] == 0) {
4997 *flags |= CPU_DTRACE_DIVZERO;
4999 regs[rd] = regs[r1] / regs[r2];
5004 if (regs[r2] == 0) {
5006 *flags |= CPU_DTRACE_DIVZERO;
5008 regs[rd] = (int64_t)regs[r1] %
5014 if (regs[r2] == 0) {
5016 *flags |= CPU_DTRACE_DIVZERO;
5018 regs[rd] = regs[r1] % regs[r2];
5023 regs[rd] = ~regs[r1];
5026 regs[rd] = regs[r1];
5029 cc_r = regs[r1] - regs[r2];
5033 cc_c = regs[r1] < regs[r2];
5036 cc_n = cc_v = cc_c = 0;
5037 cc_z = regs[r1] == 0;
5040 pc = DIF_INSTR_LABEL(instr);
5044 pc = DIF_INSTR_LABEL(instr);
5048 pc = DIF_INSTR_LABEL(instr);
5051 if ((cc_z | (cc_n ^ cc_v)) == 0)
5052 pc = DIF_INSTR_LABEL(instr);
5055 if ((cc_c | cc_z) == 0)
5056 pc = DIF_INSTR_LABEL(instr);
5059 if ((cc_n ^ cc_v) == 0)
5060 pc = DIF_INSTR_LABEL(instr);
5064 pc = DIF_INSTR_LABEL(instr);
5068 pc = DIF_INSTR_LABEL(instr);
5072 pc = DIF_INSTR_LABEL(instr);
5075 if (cc_z | (cc_n ^ cc_v))
5076 pc = DIF_INSTR_LABEL(instr);
5080 pc = DIF_INSTR_LABEL(instr);
5083 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5084 *flags |= CPU_DTRACE_KPRIV;
5090 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5093 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5094 *flags |= CPU_DTRACE_KPRIV;
5100 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5103 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5104 *flags |= CPU_DTRACE_KPRIV;
5110 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5113 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5114 *flags |= CPU_DTRACE_KPRIV;
5120 regs[rd] = dtrace_load8(regs[r1]);
5123 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5124 *flags |= CPU_DTRACE_KPRIV;
5130 regs[rd] = dtrace_load16(regs[r1]);
5133 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5134 *flags |= CPU_DTRACE_KPRIV;
5140 regs[rd] = dtrace_load32(regs[r1]);
5143 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5144 *flags |= CPU_DTRACE_KPRIV;
5150 regs[rd] = dtrace_load64(regs[r1]);
5154 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5157 regs[rd] = (int16_t)
5158 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5161 regs[rd] = (int32_t)
5162 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5166 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5170 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5174 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5178 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5187 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5190 regs[rd] = (uint64_t)(uintptr_t)
5191 (strtab + DIF_INSTR_STRING(instr));
5194 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5195 uintptr_t s1 = regs[r1];
5196 uintptr_t s2 = regs[r2];
5199 !dtrace_strcanload(s1, sz, mstate, vstate))
5202 !dtrace_strcanload(s2, sz, mstate, vstate))
5205 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5213 regs[rd] = dtrace_dif_variable(mstate, state,
5217 id = DIF_INSTR_VAR(instr);
5219 if (id >= DIF_VAR_OTHER_UBASE) {
5222 id -= DIF_VAR_OTHER_UBASE;
5223 svar = vstate->dtvs_globals[id];
5224 ASSERT(svar != NULL);
5225 v = &svar->dtsv_var;
5227 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5228 regs[rd] = svar->dtsv_data;
5232 a = (uintptr_t)svar->dtsv_data;
5234 if (*(uint8_t *)a == UINT8_MAX) {
5236 * If the 0th byte is set to UINT8_MAX
5237 * then this is to be treated as a
5238 * reference to a NULL variable.
5242 regs[rd] = a + sizeof (uint64_t);
5248 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5252 id = DIF_INSTR_VAR(instr);
5254 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5255 id -= DIF_VAR_OTHER_UBASE;
5257 svar = vstate->dtvs_globals[id];
5258 ASSERT(svar != NULL);
5259 v = &svar->dtsv_var;
5261 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5262 uintptr_t a = (uintptr_t)svar->dtsv_data;
5265 ASSERT(svar->dtsv_size != 0);
5267 if (regs[rd] == 0) {
5268 *(uint8_t *)a = UINT8_MAX;
5272 a += sizeof (uint64_t);
5274 if (!dtrace_vcanload(
5275 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5279 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5280 (void *)a, &v->dtdv_type);
5284 svar->dtsv_data = regs[rd];
5289 * There are no DTrace built-in thread-local arrays at
5290 * present. This opcode is saved for future work.
5292 *flags |= CPU_DTRACE_ILLOP;
5297 id = DIF_INSTR_VAR(instr);
5299 if (id < DIF_VAR_OTHER_UBASE) {
5301 * For now, this has no meaning.
5307 id -= DIF_VAR_OTHER_UBASE;
5309 ASSERT(id < vstate->dtvs_nlocals);
5310 ASSERT(vstate->dtvs_locals != NULL);
5312 svar = vstate->dtvs_locals[id];
5313 ASSERT(svar != NULL);
5314 v = &svar->dtsv_var;
5316 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5317 uintptr_t a = (uintptr_t)svar->dtsv_data;
5318 size_t sz = v->dtdv_type.dtdt_size;
5320 sz += sizeof (uint64_t);
5321 ASSERT(svar->dtsv_size == NCPU * sz);
5324 if (*(uint8_t *)a == UINT8_MAX) {
5326 * If the 0th byte is set to UINT8_MAX
5327 * then this is to be treated as a
5328 * reference to a NULL variable.
5332 regs[rd] = a + sizeof (uint64_t);
5338 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5339 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5340 regs[rd] = tmp[curcpu];
5344 id = DIF_INSTR_VAR(instr);
5346 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5347 id -= DIF_VAR_OTHER_UBASE;
5348 ASSERT(id < vstate->dtvs_nlocals);
5350 ASSERT(vstate->dtvs_locals != NULL);
5351 svar = vstate->dtvs_locals[id];
5352 ASSERT(svar != NULL);
5353 v = &svar->dtsv_var;
5355 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5356 uintptr_t a = (uintptr_t)svar->dtsv_data;
5357 size_t sz = v->dtdv_type.dtdt_size;
5359 sz += sizeof (uint64_t);
5360 ASSERT(svar->dtsv_size == NCPU * sz);
5363 if (regs[rd] == 0) {
5364 *(uint8_t *)a = UINT8_MAX;
5368 a += sizeof (uint64_t);
5371 if (!dtrace_vcanload(
5372 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5376 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5377 (void *)a, &v->dtdv_type);
5381 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5382 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5383 tmp[curcpu] = regs[rd];
5387 dtrace_dynvar_t *dvar;
5390 id = DIF_INSTR_VAR(instr);
5391 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5392 id -= DIF_VAR_OTHER_UBASE;
5393 v = &vstate->dtvs_tlocals[id];
5395 key = &tupregs[DIF_DTR_NREGS];
5396 key[0].dttk_value = (uint64_t)id;
5397 key[0].dttk_size = 0;
5398 DTRACE_TLS_THRKEY(key[1].dttk_value);
5399 key[1].dttk_size = 0;
5401 dvar = dtrace_dynvar(dstate, 2, key,
5402 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5410 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5411 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5413 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5420 dtrace_dynvar_t *dvar;
5423 id = DIF_INSTR_VAR(instr);
5424 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5425 id -= DIF_VAR_OTHER_UBASE;
5427 key = &tupregs[DIF_DTR_NREGS];
5428 key[0].dttk_value = (uint64_t)id;
5429 key[0].dttk_size = 0;
5430 DTRACE_TLS_THRKEY(key[1].dttk_value);
5431 key[1].dttk_size = 0;
5432 v = &vstate->dtvs_tlocals[id];
5434 dvar = dtrace_dynvar(dstate, 2, key,
5435 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5436 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5437 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5438 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5441 * Given that we're storing to thread-local data,
5442 * we need to flush our predicate cache.
5444 curthread->t_predcache = 0;
5449 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5450 if (!dtrace_vcanload(
5451 (void *)(uintptr_t)regs[rd],
5452 &v->dtdv_type, mstate, vstate))
5455 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5456 dvar->dtdv_data, &v->dtdv_type);
5458 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5465 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5469 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5470 regs, tupregs, ttop, mstate, state);
5474 if (ttop == DIF_DTR_NREGS) {
5475 *flags |= CPU_DTRACE_TUPOFLOW;
5479 if (r1 == DIF_TYPE_STRING) {
5481 * If this is a string type and the size is 0,
5482 * we'll use the system-wide default string
5483 * size. Note that we are _not_ looking at
5484 * the value of the DTRACEOPT_STRSIZE option;
5485 * had this been set, we would expect to have
5486 * a non-zero size value in the "pushtr".
5488 tupregs[ttop].dttk_size =
5489 dtrace_strlen((char *)(uintptr_t)regs[rd],
5490 regs[r2] ? regs[r2] :
5491 dtrace_strsize_default) + 1;
5493 tupregs[ttop].dttk_size = regs[r2];
5496 tupregs[ttop++].dttk_value = regs[rd];
5500 if (ttop == DIF_DTR_NREGS) {
5501 *flags |= CPU_DTRACE_TUPOFLOW;
5505 tupregs[ttop].dttk_value = regs[rd];
5506 tupregs[ttop++].dttk_size = 0;
5514 case DIF_OP_FLUSHTS:
5519 case DIF_OP_LDTAA: {
5520 dtrace_dynvar_t *dvar;
5521 dtrace_key_t *key = tupregs;
5522 uint_t nkeys = ttop;
5524 id = DIF_INSTR_VAR(instr);
5525 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5526 id -= DIF_VAR_OTHER_UBASE;
5528 key[nkeys].dttk_value = (uint64_t)id;
5529 key[nkeys++].dttk_size = 0;
5531 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5532 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5533 key[nkeys++].dttk_size = 0;
5534 v = &vstate->dtvs_tlocals[id];
5536 v = &vstate->dtvs_globals[id]->dtsv_var;
5539 dvar = dtrace_dynvar(dstate, nkeys, key,
5540 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5541 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5542 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5549 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5550 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5552 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5559 case DIF_OP_STTAA: {
5560 dtrace_dynvar_t *dvar;
5561 dtrace_key_t *key = tupregs;
5562 uint_t nkeys = ttop;
5564 id = DIF_INSTR_VAR(instr);
5565 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5566 id -= DIF_VAR_OTHER_UBASE;
5568 key[nkeys].dttk_value = (uint64_t)id;
5569 key[nkeys++].dttk_size = 0;
5571 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5572 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5573 key[nkeys++].dttk_size = 0;
5574 v = &vstate->dtvs_tlocals[id];
5576 v = &vstate->dtvs_globals[id]->dtsv_var;
5579 dvar = dtrace_dynvar(dstate, nkeys, key,
5580 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5581 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5582 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5583 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5588 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5589 if (!dtrace_vcanload(
5590 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5594 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5595 dvar->dtdv_data, &v->dtdv_type);
5597 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5603 case DIF_OP_ALLOCS: {
5604 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5605 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5608 * Rounding up the user allocation size could have
5609 * overflowed large, bogus allocations (like -1ULL) to
5612 if (size < regs[r1] ||
5613 !DTRACE_INSCRATCH(mstate, size)) {
5614 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5619 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5620 mstate->dtms_scratch_ptr += size;
5626 if (!dtrace_canstore(regs[rd], regs[r2],
5628 *flags |= CPU_DTRACE_BADADDR;
5633 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5636 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5637 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5641 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5642 *flags |= CPU_DTRACE_BADADDR;
5646 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5650 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5651 *flags |= CPU_DTRACE_BADADDR;
5656 *flags |= CPU_DTRACE_BADALIGN;
5660 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5664 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5665 *flags |= CPU_DTRACE_BADADDR;
5670 *flags |= CPU_DTRACE_BADALIGN;
5674 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5678 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5679 *flags |= CPU_DTRACE_BADADDR;
5684 *flags |= CPU_DTRACE_BADALIGN;
5688 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5693 if (!(*flags & CPU_DTRACE_FAULT))
5696 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5697 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5703 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5705 dtrace_probe_t *probe = ecb->dte_probe;
5706 dtrace_provider_t *prov = probe->dtpr_provider;
5707 char c[DTRACE_FULLNAMELEN + 80], *str;
5708 char *msg = "dtrace: breakpoint action at probe ";
5709 char *ecbmsg = " (ecb ";
5710 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5711 uintptr_t val = (uintptr_t)ecb;
5712 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5714 if (dtrace_destructive_disallow)
5718 * It's impossible to be taking action on the NULL probe.
5720 ASSERT(probe != NULL);
5723 * This is a poor man's (destitute man's?) sprintf(): we want to
5724 * print the provider name, module name, function name and name of
5725 * the probe, along with the hex address of the ECB with the breakpoint
5726 * action -- all of which we must place in the character buffer by
5729 while (*msg != '\0')
5732 for (str = prov->dtpv_name; *str != '\0'; str++)
5736 for (str = probe->dtpr_mod; *str != '\0'; str++)
5740 for (str = probe->dtpr_func; *str != '\0'; str++)
5744 for (str = probe->dtpr_name; *str != '\0'; str++)
5747 while (*ecbmsg != '\0')
5750 while (shift >= 0) {
5751 mask = (uintptr_t)0xf << shift;
5753 if (val >= ((uintptr_t)1 << shift))
5754 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5764 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5769 dtrace_action_panic(dtrace_ecb_t *ecb)
5771 dtrace_probe_t *probe = ecb->dte_probe;
5774 * It's impossible to be taking action on the NULL probe.
5776 ASSERT(probe != NULL);
5778 if (dtrace_destructive_disallow)
5781 if (dtrace_panicked != NULL)
5784 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5788 * We won the right to panic. (We want to be sure that only one
5789 * thread calls panic() from dtrace_probe(), and that panic() is
5790 * called exactly once.)
5792 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5793 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5794 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5798 dtrace_action_raise(uint64_t sig)
5800 if (dtrace_destructive_disallow)
5804 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5810 * raise() has a queue depth of 1 -- we ignore all subsequent
5811 * invocations of the raise() action.
5813 if (curthread->t_dtrace_sig == 0)
5814 curthread->t_dtrace_sig = (uint8_t)sig;
5816 curthread->t_sig_check = 1;
5819 struct proc *p = curproc;
5821 kern_psignal(p, sig);
5827 dtrace_action_stop(void)
5829 if (dtrace_destructive_disallow)
5833 if (!curthread->t_dtrace_stop) {
5834 curthread->t_dtrace_stop = 1;
5835 curthread->t_sig_check = 1;
5839 struct proc *p = curproc;
5841 kern_psignal(p, SIGSTOP);
5847 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5850 volatile uint16_t *flags;
5854 cpu_t *cpu = &solaris_cpu[curcpu];
5857 if (dtrace_destructive_disallow)
5860 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5862 now = dtrace_gethrtime();
5864 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5866 * We need to advance the mark to the current time.
5868 cpu->cpu_dtrace_chillmark = now;
5869 cpu->cpu_dtrace_chilled = 0;
5873 * Now check to see if the requested chill time would take us over
5874 * the maximum amount of time allowed in the chill interval. (Or
5875 * worse, if the calculation itself induces overflow.)
5877 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5878 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5879 *flags |= CPU_DTRACE_ILLOP;
5883 while (dtrace_gethrtime() - now < val)
5887 * Normally, we assure that the value of the variable "timestamp" does
5888 * not change within an ECB. The presence of chill() represents an
5889 * exception to this rule, however.
5891 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5892 cpu->cpu_dtrace_chilled += val;
5896 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5897 uint64_t *buf, uint64_t arg)
5899 int nframes = DTRACE_USTACK_NFRAMES(arg);
5900 int strsize = DTRACE_USTACK_STRSIZE(arg);
5901 uint64_t *pcs = &buf[1], *fps;
5902 char *str = (char *)&pcs[nframes];
5903 int size, offs = 0, i, j;
5904 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5905 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5909 * Should be taking a faster path if string space has not been
5912 ASSERT(strsize != 0);
5915 * We will first allocate some temporary space for the frame pointers.
5917 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5918 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5919 (nframes * sizeof (uint64_t));
5921 if (!DTRACE_INSCRATCH(mstate, size)) {
5923 * Not enough room for our frame pointers -- need to indicate
5924 * that we ran out of scratch space.
5926 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5930 mstate->dtms_scratch_ptr += size;
5931 saved = mstate->dtms_scratch_ptr;
5934 * Now get a stack with both program counters and frame pointers.
5936 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5937 dtrace_getufpstack(buf, fps, nframes + 1);
5938 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5941 * If that faulted, we're cooked.
5943 if (*flags & CPU_DTRACE_FAULT)
5947 * Now we want to walk up the stack, calling the USTACK helper. For
5948 * each iteration, we restore the scratch pointer.
5950 for (i = 0; i < nframes; i++) {
5951 mstate->dtms_scratch_ptr = saved;
5953 if (offs >= strsize)
5956 sym = (char *)(uintptr_t)dtrace_helper(
5957 DTRACE_HELPER_ACTION_USTACK,
5958 mstate, state, pcs[i], fps[i]);
5961 * If we faulted while running the helper, we're going to
5962 * clear the fault and null out the corresponding string.
5964 if (*flags & CPU_DTRACE_FAULT) {
5965 *flags &= ~CPU_DTRACE_FAULT;
5975 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
5978 * Now copy in the string that the helper returned to us.
5980 for (j = 0; offs + j < strsize; j++) {
5981 if ((str[offs + j] = sym[j]) == '\0')
5985 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
5990 if (offs >= strsize) {
5992 * If we didn't have room for all of the strings, we don't
5993 * abort processing -- this needn't be a fatal error -- but we
5994 * still want to increment a counter (dts_stkstroverflows) to
5995 * allow this condition to be warned about. (If this is from
5996 * a jstack() action, it is easily tuned via jstackstrsize.)
5998 dtrace_error(&state->dts_stkstroverflows);
6001 while (offs < strsize)
6005 mstate->dtms_scratch_ptr = old;
6009 * If you're looking for the epicenter of DTrace, you just found it. This
6010 * is the function called by the provider to fire a probe -- from which all
6011 * subsequent probe-context DTrace activity emanates.
6014 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6015 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6017 processorid_t cpuid;
6018 dtrace_icookie_t cookie;
6019 dtrace_probe_t *probe;
6020 dtrace_mstate_t mstate;
6022 dtrace_action_t *act;
6026 volatile uint16_t *flags;
6029 if (panicstr != NULL)
6034 * Kick out immediately if this CPU is still being born (in which case
6035 * curthread will be set to -1) or the current thread can't allow
6036 * probes in its current context.
6038 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6042 cookie = dtrace_interrupt_disable();
6043 probe = dtrace_probes[id - 1];
6045 onintr = CPU_ON_INTR(CPU);
6047 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6048 probe->dtpr_predcache == curthread->t_predcache) {
6050 * We have hit in the predicate cache; we know that
6051 * this predicate would evaluate to be false.
6053 dtrace_interrupt_enable(cookie);
6058 if (panic_quiesce) {
6060 if (panicstr != NULL) {
6063 * We don't trace anything if we're panicking.
6065 dtrace_interrupt_enable(cookie);
6069 now = dtrace_gethrtime();
6070 vtime = dtrace_vtime_references != 0;
6072 if (vtime && curthread->t_dtrace_start)
6073 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6075 mstate.dtms_difo = NULL;
6076 mstate.dtms_probe = probe;
6077 mstate.dtms_strtok = 0;
6078 mstate.dtms_arg[0] = arg0;
6079 mstate.dtms_arg[1] = arg1;
6080 mstate.dtms_arg[2] = arg2;
6081 mstate.dtms_arg[3] = arg3;
6082 mstate.dtms_arg[4] = arg4;
6084 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6086 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6087 dtrace_predicate_t *pred = ecb->dte_predicate;
6088 dtrace_state_t *state = ecb->dte_state;
6089 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6090 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6091 dtrace_vstate_t *vstate = &state->dts_vstate;
6092 dtrace_provider_t *prov = probe->dtpr_provider;
6093 uint64_t tracememsize = 0;
6098 * A little subtlety with the following (seemingly innocuous)
6099 * declaration of the automatic 'val': by looking at the
6100 * code, you might think that it could be declared in the
6101 * action processing loop, below. (That is, it's only used in
6102 * the action processing loop.) However, it must be declared
6103 * out of that scope because in the case of DIF expression
6104 * arguments to aggregating actions, one iteration of the
6105 * action loop will use the last iteration's value.
6109 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6110 *flags &= ~CPU_DTRACE_ERROR;
6112 if (prov == dtrace_provider) {
6114 * If dtrace itself is the provider of this probe,
6115 * we're only going to continue processing the ECB if
6116 * arg0 (the dtrace_state_t) is equal to the ECB's
6117 * creating state. (This prevents disjoint consumers
6118 * from seeing one another's metaprobes.)
6120 if (arg0 != (uint64_t)(uintptr_t)state)
6124 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6126 * We're not currently active. If our provider isn't
6127 * the dtrace pseudo provider, we're not interested.
6129 if (prov != dtrace_provider)
6133 * Now we must further check if we are in the BEGIN
6134 * probe. If we are, we will only continue processing
6135 * if we're still in WARMUP -- if one BEGIN enabling
6136 * has invoked the exit() action, we don't want to
6137 * evaluate subsequent BEGIN enablings.
6139 if (probe->dtpr_id == dtrace_probeid_begin &&
6140 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6141 ASSERT(state->dts_activity ==
6142 DTRACE_ACTIVITY_DRAINING);
6147 if (ecb->dte_cond) {
6149 * If the dte_cond bits indicate that this
6150 * consumer is only allowed to see user-mode firings
6151 * of this probe, call the provider's dtps_usermode()
6152 * entry point to check that the probe was fired
6153 * while in a user context. Skip this ECB if that's
6156 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6157 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6158 probe->dtpr_id, probe->dtpr_arg) == 0)
6163 * This is more subtle than it looks. We have to be
6164 * absolutely certain that CRED() isn't going to
6165 * change out from under us so it's only legit to
6166 * examine that structure if we're in constrained
6167 * situations. Currently, the only times we'll this
6168 * check is if a non-super-user has enabled the
6169 * profile or syscall providers -- providers that
6170 * allow visibility of all processes. For the
6171 * profile case, the check above will ensure that
6172 * we're examining a user context.
6174 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6177 ecb->dte_state->dts_cred.dcr_cred;
6180 ASSERT(s_cr != NULL);
6182 if ((cr = CRED()) == NULL ||
6183 s_cr->cr_uid != cr->cr_uid ||
6184 s_cr->cr_uid != cr->cr_ruid ||
6185 s_cr->cr_uid != cr->cr_suid ||
6186 s_cr->cr_gid != cr->cr_gid ||
6187 s_cr->cr_gid != cr->cr_rgid ||
6188 s_cr->cr_gid != cr->cr_sgid ||
6189 (proc = ttoproc(curthread)) == NULL ||
6190 (proc->p_flag & SNOCD))
6194 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6197 ecb->dte_state->dts_cred.dcr_cred;
6199 ASSERT(s_cr != NULL);
6201 if ((cr = CRED()) == NULL ||
6202 s_cr->cr_zone->zone_id !=
6203 cr->cr_zone->zone_id)
6209 if (now - state->dts_alive > dtrace_deadman_timeout) {
6211 * We seem to be dead. Unless we (a) have kernel
6212 * destructive permissions (b) have expicitly enabled
6213 * destructive actions and (c) destructive actions have
6214 * not been disabled, we're going to transition into
6215 * the KILLED state, from which no further processing
6216 * on this state will be performed.
6218 if (!dtrace_priv_kernel_destructive(state) ||
6219 !state->dts_cred.dcr_destructive ||
6220 dtrace_destructive_disallow) {
6221 void *activity = &state->dts_activity;
6222 dtrace_activity_t current;
6225 current = state->dts_activity;
6226 } while (dtrace_cas32(activity, current,
6227 DTRACE_ACTIVITY_KILLED) != current);
6233 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6234 ecb->dte_alignment, state, &mstate)) < 0)
6237 tomax = buf->dtb_tomax;
6238 ASSERT(tomax != NULL);
6240 if (ecb->dte_size != 0)
6241 DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);
6243 mstate.dtms_epid = ecb->dte_epid;
6244 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6246 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6247 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6249 mstate.dtms_access = 0;
6252 dtrace_difo_t *dp = pred->dtp_difo;
6255 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6257 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6258 dtrace_cacheid_t cid = probe->dtpr_predcache;
6260 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6262 * Update the predicate cache...
6264 ASSERT(cid == pred->dtp_cacheid);
6265 curthread->t_predcache = cid;
6272 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6273 act != NULL; act = act->dta_next) {
6276 dtrace_recdesc_t *rec = &act->dta_rec;
6278 size = rec->dtrd_size;
6279 valoffs = offs + rec->dtrd_offset;
6281 if (DTRACEACT_ISAGG(act->dta_kind)) {
6283 dtrace_aggregation_t *agg;
6285 agg = (dtrace_aggregation_t *)act;
6287 if ((dp = act->dta_difo) != NULL)
6288 v = dtrace_dif_emulate(dp,
6289 &mstate, vstate, state);
6291 if (*flags & CPU_DTRACE_ERROR)
6295 * Note that we always pass the expression
6296 * value from the previous iteration of the
6297 * action loop. This value will only be used
6298 * if there is an expression argument to the
6299 * aggregating action, denoted by the
6300 * dtag_hasarg field.
6302 dtrace_aggregate(agg, buf,
6303 offs, aggbuf, v, val);
6307 switch (act->dta_kind) {
6308 case DTRACEACT_STOP:
6309 if (dtrace_priv_proc_destructive(state))
6310 dtrace_action_stop();
6313 case DTRACEACT_BREAKPOINT:
6314 if (dtrace_priv_kernel_destructive(state))
6315 dtrace_action_breakpoint(ecb);
6318 case DTRACEACT_PANIC:
6319 if (dtrace_priv_kernel_destructive(state))
6320 dtrace_action_panic(ecb);
6323 case DTRACEACT_STACK:
6324 if (!dtrace_priv_kernel(state))
6327 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6328 size / sizeof (pc_t), probe->dtpr_aframes,
6329 DTRACE_ANCHORED(probe) ? NULL :
6333 case DTRACEACT_JSTACK:
6334 case DTRACEACT_USTACK:
6335 if (!dtrace_priv_proc(state))
6339 * See comment in DIF_VAR_PID.
6341 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6343 int depth = DTRACE_USTACK_NFRAMES(
6346 dtrace_bzero((void *)(tomax + valoffs),
6347 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6348 + depth * sizeof (uint64_t));
6353 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6354 curproc->p_dtrace_helpers != NULL) {
6356 * This is the slow path -- we have
6357 * allocated string space, and we're
6358 * getting the stack of a process that
6359 * has helpers. Call into a separate
6360 * routine to perform this processing.
6362 dtrace_action_ustack(&mstate, state,
6363 (uint64_t *)(tomax + valoffs),
6368 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6369 dtrace_getupcstack((uint64_t *)
6371 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6372 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6382 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6384 if (*flags & CPU_DTRACE_ERROR)
6387 switch (act->dta_kind) {
6388 case DTRACEACT_SPECULATE:
6389 ASSERT(buf == &state->dts_buffer[cpuid]);
6390 buf = dtrace_speculation_buffer(state,
6394 *flags |= CPU_DTRACE_DROP;
6398 offs = dtrace_buffer_reserve(buf,
6399 ecb->dte_needed, ecb->dte_alignment,
6403 *flags |= CPU_DTRACE_DROP;
6407 tomax = buf->dtb_tomax;
6408 ASSERT(tomax != NULL);
6410 if (ecb->dte_size != 0)
6411 DTRACE_STORE(uint32_t, tomax, offs,
6415 case DTRACEACT_PRINTM: {
6416 /* The DIF returns a 'memref'. */
6417 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6419 /* Get the size from the memref. */
6423 * Check if the size exceeds the allocated
6426 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6428 *flags |= CPU_DTRACE_DROP;
6432 /* Store the size in the buffer first. */
6433 DTRACE_STORE(uintptr_t, tomax,
6437 * Offset the buffer address to the start
6440 valoffs += sizeof(uintptr_t);
6443 * Reset to the memory address rather than
6444 * the memref array, then let the BYREF
6445 * code below do the work to store the
6446 * memory data in the buffer.
6452 case DTRACEACT_PRINTT: {
6453 /* The DIF returns a 'typeref'. */
6454 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6459 * Get the type string length and round it
6460 * up so that the data that follows is
6461 * aligned for easy access.
6463 size_t typs = strlen((char *) typeref[2]) + 1;
6464 typs = roundup(typs, sizeof(uintptr_t));
6467 *Get the size from the typeref using the
6468 * number of elements and the type size.
6470 size = typeref[1] * typeref[3];
6473 * Check if the size exceeds the allocated
6476 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6478 *flags |= CPU_DTRACE_DROP;
6482 /* Store the size in the buffer first. */
6483 DTRACE_STORE(uintptr_t, tomax,
6485 valoffs += sizeof(uintptr_t);
6487 /* Store the type size in the buffer. */
6488 DTRACE_STORE(uintptr_t, tomax,
6489 valoffs, typeref[3]);
6490 valoffs += sizeof(uintptr_t);
6494 for (s = 0; s < typs; s++) {
6496 c = dtrace_load8(val++);
6498 DTRACE_STORE(uint8_t, tomax,
6503 * Reset to the memory address rather than
6504 * the typeref array, then let the BYREF
6505 * code below do the work to store the
6506 * memory data in the buffer.
6512 case DTRACEACT_CHILL:
6513 if (dtrace_priv_kernel_destructive(state))
6514 dtrace_action_chill(&mstate, val);
6517 case DTRACEACT_RAISE:
6518 if (dtrace_priv_proc_destructive(state))
6519 dtrace_action_raise(val);
6522 case DTRACEACT_COMMIT:
6526 * We need to commit our buffer state.
6529 buf->dtb_offset = offs + ecb->dte_size;
6530 buf = &state->dts_buffer[cpuid];
6531 dtrace_speculation_commit(state, cpuid, val);
6535 case DTRACEACT_DISCARD:
6536 dtrace_speculation_discard(state, cpuid, val);
6539 case DTRACEACT_DIFEXPR:
6540 case DTRACEACT_LIBACT:
6541 case DTRACEACT_PRINTF:
6542 case DTRACEACT_PRINTA:
6543 case DTRACEACT_SYSTEM:
6544 case DTRACEACT_FREOPEN:
6545 case DTRACEACT_TRACEMEM:
6548 case DTRACEACT_TRACEMEM_DYNSIZE:
6554 if (!dtrace_priv_kernel(state))
6558 case DTRACEACT_USYM:
6559 case DTRACEACT_UMOD:
6560 case DTRACEACT_UADDR: {
6562 struct pid *pid = curthread->t_procp->p_pidp;
6565 if (!dtrace_priv_proc(state))
6568 DTRACE_STORE(uint64_t, tomax,
6570 valoffs, (uint64_t)pid->pid_id);
6572 valoffs, (uint64_t) curproc->p_pid);
6574 DTRACE_STORE(uint64_t, tomax,
6575 valoffs + sizeof (uint64_t), val);
6580 case DTRACEACT_EXIT: {
6582 * For the exit action, we are going to attempt
6583 * to atomically set our activity to be
6584 * draining. If this fails (either because
6585 * another CPU has beat us to the exit action,
6586 * or because our current activity is something
6587 * other than ACTIVE or WARMUP), we will
6588 * continue. This assures that the exit action
6589 * can be successfully recorded at most once
6590 * when we're in the ACTIVE state. If we're
6591 * encountering the exit() action while in
6592 * COOLDOWN, however, we want to honor the new
6593 * status code. (We know that we're the only
6594 * thread in COOLDOWN, so there is no race.)
6596 void *activity = &state->dts_activity;
6597 dtrace_activity_t current = state->dts_activity;
6599 if (current == DTRACE_ACTIVITY_COOLDOWN)
6602 if (current != DTRACE_ACTIVITY_WARMUP)
6603 current = DTRACE_ACTIVITY_ACTIVE;
6605 if (dtrace_cas32(activity, current,
6606 DTRACE_ACTIVITY_DRAINING) != current) {
6607 *flags |= CPU_DTRACE_DROP;
6618 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6619 uintptr_t end = valoffs + size;
6621 if (tracememsize != 0 &&
6622 valoffs + tracememsize < end) {
6623 end = valoffs + tracememsize;
6627 if (!dtrace_vcanload((void *)(uintptr_t)val,
6628 &dp->dtdo_rtype, &mstate, vstate))
6632 * If this is a string, we're going to only
6633 * load until we find the zero byte -- after
6634 * which we'll store zero bytes.
6636 if (dp->dtdo_rtype.dtdt_kind ==
6639 int intuple = act->dta_intuple;
6642 for (s = 0; s < size; s++) {
6644 c = dtrace_load8(val++);
6646 DTRACE_STORE(uint8_t, tomax,
6649 if (c == '\0' && intuple)
6656 while (valoffs < end) {
6657 DTRACE_STORE(uint8_t, tomax, valoffs++,
6658 dtrace_load8(val++));
6668 case sizeof (uint8_t):
6669 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6671 case sizeof (uint16_t):
6672 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6674 case sizeof (uint32_t):
6675 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6677 case sizeof (uint64_t):
6678 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6682 * Any other size should have been returned by
6683 * reference, not by value.
6690 if (*flags & CPU_DTRACE_DROP)
6693 if (*flags & CPU_DTRACE_FAULT) {
6695 dtrace_action_t *err;
6699 if (probe->dtpr_id == dtrace_probeid_error) {
6701 * There's nothing we can do -- we had an
6702 * error on the error probe. We bump an
6703 * error counter to at least indicate that
6704 * this condition happened.
6706 dtrace_error(&state->dts_dblerrors);
6712 * Before recursing on dtrace_probe(), we
6713 * need to explicitly clear out our start
6714 * time to prevent it from being accumulated
6715 * into t_dtrace_vtime.
6717 curthread->t_dtrace_start = 0;
6721 * Iterate over the actions to figure out which action
6722 * we were processing when we experienced the error.
6723 * Note that act points _past_ the faulting action; if
6724 * act is ecb->dte_action, the fault was in the
6725 * predicate, if it's ecb->dte_action->dta_next it's
6726 * in action #1, and so on.
6728 for (err = ecb->dte_action, ndx = 0;
6729 err != act; err = err->dta_next, ndx++)
6732 dtrace_probe_error(state, ecb->dte_epid, ndx,
6733 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6734 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6735 cpu_core[cpuid].cpuc_dtrace_illval);
6741 buf->dtb_offset = offs + ecb->dte_size;
6745 curthread->t_dtrace_start = dtrace_gethrtime();
6747 dtrace_interrupt_enable(cookie);
6751 * DTrace Probe Hashing Functions
6753 * The functions in this section (and indeed, the functions in remaining
6754 * sections) are not _called_ from probe context. (Any exceptions to this are
6755 * marked with a "Note:".) Rather, they are called from elsewhere in the
6756 * DTrace framework to look-up probes in, add probes to and remove probes from
6757 * the DTrace probe hashes. (Each probe is hashed by each element of the
6758 * probe tuple -- allowing for fast lookups, regardless of what was
6762 dtrace_hash_str(const char *p)
6768 hval = (hval << 4) + *p++;
6769 if ((g = (hval & 0xf0000000)) != 0)
6776 static dtrace_hash_t *
6777 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6779 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6781 hash->dth_stroffs = stroffs;
6782 hash->dth_nextoffs = nextoffs;
6783 hash->dth_prevoffs = prevoffs;
6786 hash->dth_mask = hash->dth_size - 1;
6788 hash->dth_tab = kmem_zalloc(hash->dth_size *
6789 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6795 dtrace_hash_destroy(dtrace_hash_t *hash)
6800 for (i = 0; i < hash->dth_size; i++)
6801 ASSERT(hash->dth_tab[i] == NULL);
6804 kmem_free(hash->dth_tab,
6805 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6806 kmem_free(hash, sizeof (dtrace_hash_t));
6810 dtrace_hash_resize(dtrace_hash_t *hash)
6812 int size = hash->dth_size, i, ndx;
6813 int new_size = hash->dth_size << 1;
6814 int new_mask = new_size - 1;
6815 dtrace_hashbucket_t **new_tab, *bucket, *next;
6817 ASSERT((new_size & new_mask) == 0);
6819 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6821 for (i = 0; i < size; i++) {
6822 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6823 dtrace_probe_t *probe = bucket->dthb_chain;
6825 ASSERT(probe != NULL);
6826 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6828 next = bucket->dthb_next;
6829 bucket->dthb_next = new_tab[ndx];
6830 new_tab[ndx] = bucket;
6834 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6835 hash->dth_tab = new_tab;
6836 hash->dth_size = new_size;
6837 hash->dth_mask = new_mask;
6841 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6843 int hashval = DTRACE_HASHSTR(hash, new);
6844 int ndx = hashval & hash->dth_mask;
6845 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6846 dtrace_probe_t **nextp, **prevp;
6848 for (; bucket != NULL; bucket = bucket->dthb_next) {
6849 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6853 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6854 dtrace_hash_resize(hash);
6855 dtrace_hash_add(hash, new);
6859 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6860 bucket->dthb_next = hash->dth_tab[ndx];
6861 hash->dth_tab[ndx] = bucket;
6862 hash->dth_nbuckets++;
6865 nextp = DTRACE_HASHNEXT(hash, new);
6866 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6867 *nextp = bucket->dthb_chain;
6869 if (bucket->dthb_chain != NULL) {
6870 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6871 ASSERT(*prevp == NULL);
6875 bucket->dthb_chain = new;
6879 static dtrace_probe_t *
6880 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6882 int hashval = DTRACE_HASHSTR(hash, template);
6883 int ndx = hashval & hash->dth_mask;
6884 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6886 for (; bucket != NULL; bucket = bucket->dthb_next) {
6887 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6888 return (bucket->dthb_chain);
6895 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6897 int hashval = DTRACE_HASHSTR(hash, template);
6898 int ndx = hashval & hash->dth_mask;
6899 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6901 for (; bucket != NULL; bucket = bucket->dthb_next) {
6902 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6903 return (bucket->dthb_len);
6910 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
6912 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
6913 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6915 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
6916 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
6919 * Find the bucket that we're removing this probe from.
6921 for (; bucket != NULL; bucket = bucket->dthb_next) {
6922 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
6926 ASSERT(bucket != NULL);
6928 if (*prevp == NULL) {
6929 if (*nextp == NULL) {
6931 * The removed probe was the only probe on this
6932 * bucket; we need to remove the bucket.
6934 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
6936 ASSERT(bucket->dthb_chain == probe);
6940 hash->dth_tab[ndx] = bucket->dthb_next;
6942 while (b->dthb_next != bucket)
6944 b->dthb_next = bucket->dthb_next;
6947 ASSERT(hash->dth_nbuckets > 0);
6948 hash->dth_nbuckets--;
6949 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
6953 bucket->dthb_chain = *nextp;
6955 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
6959 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
6963 * DTrace Utility Functions
6965 * These are random utility functions that are _not_ called from probe context.
6968 dtrace_badattr(const dtrace_attribute_t *a)
6970 return (a->dtat_name > DTRACE_STABILITY_MAX ||
6971 a->dtat_data > DTRACE_STABILITY_MAX ||
6972 a->dtat_class > DTRACE_CLASS_MAX);
6976 * Return a duplicate copy of a string. If the specified string is NULL,
6977 * this function returns a zero-length string.
6980 dtrace_strdup(const char *str)
6982 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
6985 (void) strcpy(new, str);
6990 #define DTRACE_ISALPHA(c) \
6991 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
6994 dtrace_badname(const char *s)
6998 if (s == NULL || (c = *s++) == '\0')
7001 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7004 while ((c = *s++) != '\0') {
7005 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7006 c != '-' && c != '_' && c != '.' && c != '`')
7014 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7019 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7021 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7023 priv = DTRACE_PRIV_ALL;
7025 *uidp = crgetuid(cr);
7026 *zoneidp = crgetzoneid(cr);
7029 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7030 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7031 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7032 priv |= DTRACE_PRIV_USER;
7033 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7034 priv |= DTRACE_PRIV_PROC;
7035 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7036 priv |= DTRACE_PRIV_OWNER;
7037 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7038 priv |= DTRACE_PRIV_ZONEOWNER;
7041 priv = DTRACE_PRIV_ALL;
7047 #ifdef DTRACE_ERRDEBUG
7049 dtrace_errdebug(const char *str)
7051 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
7054 mutex_enter(&dtrace_errlock);
7055 dtrace_errlast = str;
7056 dtrace_errthread = curthread;
7058 while (occupied++ < DTRACE_ERRHASHSZ) {
7059 if (dtrace_errhash[hval].dter_msg == str) {
7060 dtrace_errhash[hval].dter_count++;
7064 if (dtrace_errhash[hval].dter_msg != NULL) {
7065 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7069 dtrace_errhash[hval].dter_msg = str;
7070 dtrace_errhash[hval].dter_count = 1;
7074 panic("dtrace: undersized error hash");
7076 mutex_exit(&dtrace_errlock);
7081 * DTrace Matching Functions
7083 * These functions are used to match groups of probes, given some elements of
7084 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7087 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7090 if (priv != DTRACE_PRIV_ALL) {
7091 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7092 uint32_t match = priv & ppriv;
7095 * No PRIV_DTRACE_* privileges...
7097 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7098 DTRACE_PRIV_KERNEL)) == 0)
7102 * No matching bits, but there were bits to match...
7104 if (match == 0 && ppriv != 0)
7108 * Need to have permissions to the process, but don't...
7110 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7111 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7116 * Need to be in the same zone unless we possess the
7117 * privilege to examine all zones.
7119 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7120 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7129 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7130 * consists of input pattern strings and an ops-vector to evaluate them.
7131 * This function returns >0 for match, 0 for no match, and <0 for error.
7134 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7135 uint32_t priv, uid_t uid, zoneid_t zoneid)
7137 dtrace_provider_t *pvp = prp->dtpr_provider;
7140 if (pvp->dtpv_defunct)
7143 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7146 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7149 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7152 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7155 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7162 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7163 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7164 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7165 * In addition, all of the recursion cases except for '*' matching have been
7166 * unwound. For '*', we still implement recursive evaluation, but a depth
7167 * counter is maintained and matching is aborted if we recurse too deep.
7168 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7171 dtrace_match_glob(const char *s, const char *p, int depth)
7177 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7181 s = ""; /* treat NULL as empty string */
7190 if ((c = *p++) == '\0')
7191 return (s1 == '\0');
7195 int ok = 0, notflag = 0;
7206 if ((c = *p++) == '\0')
7210 if (c == '-' && lc != '\0' && *p != ']') {
7211 if ((c = *p++) == '\0')
7213 if (c == '\\' && (c = *p++) == '\0')
7217 if (s1 < lc || s1 > c)
7221 } else if (lc <= s1 && s1 <= c)
7224 } else if (c == '\\' && (c = *p++) == '\0')
7227 lc = c; /* save left-hand 'c' for next iteration */
7237 if ((c = *p++) == '\0')
7249 if ((c = *p++) == '\0')
7265 p++; /* consecutive *'s are identical to a single one */
7270 for (s = olds; *s != '\0'; s++) {
7271 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7281 dtrace_match_string(const char *s, const char *p, int depth)
7283 return (s != NULL && strcmp(s, p) == 0);
7288 dtrace_match_nul(const char *s, const char *p, int depth)
7290 return (1); /* always match the empty pattern */
7295 dtrace_match_nonzero(const char *s, const char *p, int depth)
7297 return (s != NULL && s[0] != '\0');
7301 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7302 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7304 dtrace_probe_t template, *probe;
7305 dtrace_hash_t *hash = NULL;
7306 int len, best = INT_MAX, nmatched = 0;
7309 ASSERT(MUTEX_HELD(&dtrace_lock));
7312 * If the probe ID is specified in the key, just lookup by ID and
7313 * invoke the match callback once if a matching probe is found.
7315 if (pkp->dtpk_id != DTRACE_IDNONE) {
7316 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7317 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7318 (void) (*matched)(probe, arg);
7324 template.dtpr_mod = (char *)pkp->dtpk_mod;
7325 template.dtpr_func = (char *)pkp->dtpk_func;
7326 template.dtpr_name = (char *)pkp->dtpk_name;
7329 * We want to find the most distinct of the module name, function
7330 * name, and name. So for each one that is not a glob pattern or
7331 * empty string, we perform a lookup in the corresponding hash and
7332 * use the hash table with the fewest collisions to do our search.
7334 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7335 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7337 hash = dtrace_bymod;
7340 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7341 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7343 hash = dtrace_byfunc;
7346 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7347 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7349 hash = dtrace_byname;
7353 * If we did not select a hash table, iterate over every probe and
7354 * invoke our callback for each one that matches our input probe key.
7357 for (i = 0; i < dtrace_nprobes; i++) {
7358 if ((probe = dtrace_probes[i]) == NULL ||
7359 dtrace_match_probe(probe, pkp, priv, uid,
7365 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7373 * If we selected a hash table, iterate over each probe of the same key
7374 * name and invoke the callback for every probe that matches the other
7375 * attributes of our input probe key.
7377 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7378 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7380 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7385 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7393 * Return the function pointer dtrace_probecmp() should use to compare the
7394 * specified pattern with a string. For NULL or empty patterns, we select
7395 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7396 * For non-empty non-glob strings, we use dtrace_match_string().
7398 static dtrace_probekey_f *
7399 dtrace_probekey_func(const char *p)
7403 if (p == NULL || *p == '\0')
7404 return (&dtrace_match_nul);
7406 while ((c = *p++) != '\0') {
7407 if (c == '[' || c == '?' || c == '*' || c == '\\')
7408 return (&dtrace_match_glob);
7411 return (&dtrace_match_string);
7415 * Build a probe comparison key for use with dtrace_match_probe() from the
7416 * given probe description. By convention, a null key only matches anchored
7417 * probes: if each field is the empty string, reset dtpk_fmatch to
7418 * dtrace_match_nonzero().
7421 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7423 pkp->dtpk_prov = pdp->dtpd_provider;
7424 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7426 pkp->dtpk_mod = pdp->dtpd_mod;
7427 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7429 pkp->dtpk_func = pdp->dtpd_func;
7430 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7432 pkp->dtpk_name = pdp->dtpd_name;
7433 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7435 pkp->dtpk_id = pdp->dtpd_id;
7437 if (pkp->dtpk_id == DTRACE_IDNONE &&
7438 pkp->dtpk_pmatch == &dtrace_match_nul &&
7439 pkp->dtpk_mmatch == &dtrace_match_nul &&
7440 pkp->dtpk_fmatch == &dtrace_match_nul &&
7441 pkp->dtpk_nmatch == &dtrace_match_nul)
7442 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7446 * DTrace Provider-to-Framework API Functions
7448 * These functions implement much of the Provider-to-Framework API, as
7449 * described in <sys/dtrace.h>. The parts of the API not in this section are
7450 * the functions in the API for probe management (found below), and
7451 * dtrace_probe() itself (found above).
7455 * Register the calling provider with the DTrace framework. This should
7456 * generally be called by DTrace providers in their attach(9E) entry point.
7459 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7460 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7462 dtrace_provider_t *provider;
7464 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7465 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7466 "arguments", name ? name : "<NULL>");
7470 if (name[0] == '\0' || dtrace_badname(name)) {
7471 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7472 "provider name", name);
7476 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7477 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7478 pops->dtps_destroy == NULL ||
7479 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7480 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7481 "provider ops", name);
7485 if (dtrace_badattr(&pap->dtpa_provider) ||
7486 dtrace_badattr(&pap->dtpa_mod) ||
7487 dtrace_badattr(&pap->dtpa_func) ||
7488 dtrace_badattr(&pap->dtpa_name) ||
7489 dtrace_badattr(&pap->dtpa_args)) {
7490 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7491 "provider attributes", name);
7495 if (priv & ~DTRACE_PRIV_ALL) {
7496 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7497 "privilege attributes", name);
7501 if ((priv & DTRACE_PRIV_KERNEL) &&
7502 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7503 pops->dtps_usermode == NULL) {
7504 cmn_err(CE_WARN, "failed to register provider '%s': need "
7505 "dtps_usermode() op for given privilege attributes", name);
7509 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7510 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7511 (void) strcpy(provider->dtpv_name, name);
7513 provider->dtpv_attr = *pap;
7514 provider->dtpv_priv.dtpp_flags = priv;
7516 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7517 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7519 provider->dtpv_pops = *pops;
7521 if (pops->dtps_provide == NULL) {
7522 ASSERT(pops->dtps_provide_module != NULL);
7523 provider->dtpv_pops.dtps_provide =
7524 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7527 if (pops->dtps_provide_module == NULL) {
7528 ASSERT(pops->dtps_provide != NULL);
7529 provider->dtpv_pops.dtps_provide_module =
7530 (void (*)(void *, modctl_t *))dtrace_nullop;
7533 if (pops->dtps_suspend == NULL) {
7534 ASSERT(pops->dtps_resume == NULL);
7535 provider->dtpv_pops.dtps_suspend =
7536 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7537 provider->dtpv_pops.dtps_resume =
7538 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7541 provider->dtpv_arg = arg;
7542 *idp = (dtrace_provider_id_t)provider;
7544 if (pops == &dtrace_provider_ops) {
7545 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7546 ASSERT(MUTEX_HELD(&dtrace_lock));
7547 ASSERT(dtrace_anon.dta_enabling == NULL);
7550 * We make sure that the DTrace provider is at the head of
7551 * the provider chain.
7553 provider->dtpv_next = dtrace_provider;
7554 dtrace_provider = provider;
7558 mutex_enter(&dtrace_provider_lock);
7559 mutex_enter(&dtrace_lock);
7562 * If there is at least one provider registered, we'll add this
7563 * provider after the first provider.
7565 if (dtrace_provider != NULL) {
7566 provider->dtpv_next = dtrace_provider->dtpv_next;
7567 dtrace_provider->dtpv_next = provider;
7569 dtrace_provider = provider;
7572 if (dtrace_retained != NULL) {
7573 dtrace_enabling_provide(provider);
7576 * Now we need to call dtrace_enabling_matchall() -- which
7577 * will acquire cpu_lock and dtrace_lock. We therefore need
7578 * to drop all of our locks before calling into it...
7580 mutex_exit(&dtrace_lock);
7581 mutex_exit(&dtrace_provider_lock);
7582 dtrace_enabling_matchall();
7587 mutex_exit(&dtrace_lock);
7588 mutex_exit(&dtrace_provider_lock);
7594 * Unregister the specified provider from the DTrace framework. This should
7595 * generally be called by DTrace providers in their detach(9E) entry point.
7598 dtrace_unregister(dtrace_provider_id_t id)
7600 dtrace_provider_t *old = (dtrace_provider_t *)id;
7601 dtrace_provider_t *prev = NULL;
7602 int i, self = 0, noreap = 0;
7603 dtrace_probe_t *probe, *first = NULL;
7605 if (old->dtpv_pops.dtps_enable ==
7606 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7608 * If DTrace itself is the provider, we're called with locks
7611 ASSERT(old == dtrace_provider);
7613 ASSERT(dtrace_devi != NULL);
7615 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7616 ASSERT(MUTEX_HELD(&dtrace_lock));
7619 if (dtrace_provider->dtpv_next != NULL) {
7621 * There's another provider here; return failure.
7626 mutex_enter(&dtrace_provider_lock);
7627 mutex_enter(&mod_lock);
7628 mutex_enter(&dtrace_lock);
7632 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7633 * probes, we refuse to let providers slither away, unless this
7634 * provider has already been explicitly invalidated.
7636 if (!old->dtpv_defunct &&
7637 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7638 dtrace_anon.dta_state->dts_necbs > 0))) {
7640 mutex_exit(&dtrace_lock);
7641 mutex_exit(&mod_lock);
7642 mutex_exit(&dtrace_provider_lock);
7648 * Attempt to destroy the probes associated with this provider.
7650 for (i = 0; i < dtrace_nprobes; i++) {
7651 if ((probe = dtrace_probes[i]) == NULL)
7654 if (probe->dtpr_provider != old)
7657 if (probe->dtpr_ecb == NULL)
7661 * If we are trying to unregister a defunct provider, and the
7662 * provider was made defunct within the interval dictated by
7663 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7664 * attempt to reap our enablings. To denote that the provider
7665 * should reattempt to unregister itself at some point in the
7666 * future, we will return a differentiable error code (EAGAIN
7667 * instead of EBUSY) in this case.
7669 if (dtrace_gethrtime() - old->dtpv_defunct >
7670 dtrace_unregister_defunct_reap)
7674 mutex_exit(&dtrace_lock);
7675 mutex_exit(&mod_lock);
7676 mutex_exit(&dtrace_provider_lock);
7682 (void) taskq_dispatch(dtrace_taskq,
7683 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7689 * All of the probes for this provider are disabled; we can safely
7690 * remove all of them from their hash chains and from the probe array.
7692 for (i = 0; i < dtrace_nprobes; i++) {
7693 if ((probe = dtrace_probes[i]) == NULL)
7696 if (probe->dtpr_provider != old)
7699 dtrace_probes[i] = NULL;
7701 dtrace_hash_remove(dtrace_bymod, probe);
7702 dtrace_hash_remove(dtrace_byfunc, probe);
7703 dtrace_hash_remove(dtrace_byname, probe);
7705 if (first == NULL) {
7707 probe->dtpr_nextmod = NULL;
7709 probe->dtpr_nextmod = first;
7715 * The provider's probes have been removed from the hash chains and
7716 * from the probe array. Now issue a dtrace_sync() to be sure that
7717 * everyone has cleared out from any probe array processing.
7721 for (probe = first; probe != NULL; probe = first) {
7722 first = probe->dtpr_nextmod;
7724 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7726 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7727 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7728 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7730 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7732 free_unr(dtrace_arena, probe->dtpr_id);
7734 kmem_free(probe, sizeof (dtrace_probe_t));
7737 if ((prev = dtrace_provider) == old) {
7739 ASSERT(self || dtrace_devi == NULL);
7740 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7742 dtrace_provider = old->dtpv_next;
7744 while (prev != NULL && prev->dtpv_next != old)
7745 prev = prev->dtpv_next;
7748 panic("attempt to unregister non-existent "
7749 "dtrace provider %p\n", (void *)id);
7752 prev->dtpv_next = old->dtpv_next;
7756 mutex_exit(&dtrace_lock);
7757 mutex_exit(&mod_lock);
7758 mutex_exit(&dtrace_provider_lock);
7761 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7762 kmem_free(old, sizeof (dtrace_provider_t));
7768 * Invalidate the specified provider. All subsequent probe lookups for the
7769 * specified provider will fail, but its probes will not be removed.
7772 dtrace_invalidate(dtrace_provider_id_t id)
7774 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7776 ASSERT(pvp->dtpv_pops.dtps_enable !=
7777 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7779 mutex_enter(&dtrace_provider_lock);
7780 mutex_enter(&dtrace_lock);
7782 pvp->dtpv_defunct = dtrace_gethrtime();
7784 mutex_exit(&dtrace_lock);
7785 mutex_exit(&dtrace_provider_lock);
7789 * Indicate whether or not DTrace has attached.
7792 dtrace_attached(void)
7795 * dtrace_provider will be non-NULL iff the DTrace driver has
7796 * attached. (It's non-NULL because DTrace is always itself a
7799 return (dtrace_provider != NULL);
7803 * Remove all the unenabled probes for the given provider. This function is
7804 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7805 * -- just as many of its associated probes as it can.
7808 dtrace_condense(dtrace_provider_id_t id)
7810 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7812 dtrace_probe_t *probe;
7815 * Make sure this isn't the dtrace provider itself.
7817 ASSERT(prov->dtpv_pops.dtps_enable !=
7818 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7820 mutex_enter(&dtrace_provider_lock);
7821 mutex_enter(&dtrace_lock);
7824 * Attempt to destroy the probes associated with this provider.
7826 for (i = 0; i < dtrace_nprobes; i++) {
7827 if ((probe = dtrace_probes[i]) == NULL)
7830 if (probe->dtpr_provider != prov)
7833 if (probe->dtpr_ecb != NULL)
7836 dtrace_probes[i] = NULL;
7838 dtrace_hash_remove(dtrace_bymod, probe);
7839 dtrace_hash_remove(dtrace_byfunc, probe);
7840 dtrace_hash_remove(dtrace_byname, probe);
7842 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7844 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7845 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7846 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7847 kmem_free(probe, sizeof (dtrace_probe_t));
7849 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7851 free_unr(dtrace_arena, i + 1);
7855 mutex_exit(&dtrace_lock);
7856 mutex_exit(&dtrace_provider_lock);
7862 * DTrace Probe Management Functions
7864 * The functions in this section perform the DTrace probe management,
7865 * including functions to create probes, look-up probes, and call into the
7866 * providers to request that probes be provided. Some of these functions are
7867 * in the Provider-to-Framework API; these functions can be identified by the
7868 * fact that they are not declared "static".
7872 * Create a probe with the specified module name, function name, and name.
7875 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7876 const char *func, const char *name, int aframes, void *arg)
7878 dtrace_probe_t *probe, **probes;
7879 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7882 if (provider == dtrace_provider) {
7883 ASSERT(MUTEX_HELD(&dtrace_lock));
7885 mutex_enter(&dtrace_lock);
7889 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7890 VM_BESTFIT | VM_SLEEP);
7892 id = alloc_unr(dtrace_arena);
7894 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
7896 probe->dtpr_id = id;
7897 probe->dtpr_gen = dtrace_probegen++;
7898 probe->dtpr_mod = dtrace_strdup(mod);
7899 probe->dtpr_func = dtrace_strdup(func);
7900 probe->dtpr_name = dtrace_strdup(name);
7901 probe->dtpr_arg = arg;
7902 probe->dtpr_aframes = aframes;
7903 probe->dtpr_provider = provider;
7905 dtrace_hash_add(dtrace_bymod, probe);
7906 dtrace_hash_add(dtrace_byfunc, probe);
7907 dtrace_hash_add(dtrace_byname, probe);
7909 if (id - 1 >= dtrace_nprobes) {
7910 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
7911 size_t nsize = osize << 1;
7915 ASSERT(dtrace_probes == NULL);
7916 nsize = sizeof (dtrace_probe_t *);
7919 probes = kmem_zalloc(nsize, KM_SLEEP);
7921 if (dtrace_probes == NULL) {
7923 dtrace_probes = probes;
7926 dtrace_probe_t **oprobes = dtrace_probes;
7928 bcopy(oprobes, probes, osize);
7929 dtrace_membar_producer();
7930 dtrace_probes = probes;
7935 * All CPUs are now seeing the new probes array; we can
7936 * safely free the old array.
7938 kmem_free(oprobes, osize);
7939 dtrace_nprobes <<= 1;
7942 ASSERT(id - 1 < dtrace_nprobes);
7945 ASSERT(dtrace_probes[id - 1] == NULL);
7946 dtrace_probes[id - 1] = probe;
7948 if (provider != dtrace_provider)
7949 mutex_exit(&dtrace_lock);
7954 static dtrace_probe_t *
7955 dtrace_probe_lookup_id(dtrace_id_t id)
7957 ASSERT(MUTEX_HELD(&dtrace_lock));
7959 if (id == 0 || id > dtrace_nprobes)
7962 return (dtrace_probes[id - 1]);
7966 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
7968 *((dtrace_id_t *)arg) = probe->dtpr_id;
7970 return (DTRACE_MATCH_DONE);
7974 * Look up a probe based on provider and one or more of module name, function
7975 * name and probe name.
7978 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
7979 char *func, char *name)
7981 dtrace_probekey_t pkey;
7985 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
7986 pkey.dtpk_pmatch = &dtrace_match_string;
7987 pkey.dtpk_mod = mod;
7988 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
7989 pkey.dtpk_func = func;
7990 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
7991 pkey.dtpk_name = name;
7992 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
7993 pkey.dtpk_id = DTRACE_IDNONE;
7995 mutex_enter(&dtrace_lock);
7996 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
7997 dtrace_probe_lookup_match, &id);
7998 mutex_exit(&dtrace_lock);
8000 ASSERT(match == 1 || match == 0);
8001 return (match ? id : 0);
8005 * Returns the probe argument associated with the specified probe.
8008 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8010 dtrace_probe_t *probe;
8013 mutex_enter(&dtrace_lock);
8015 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8016 probe->dtpr_provider == (dtrace_provider_t *)id)
8017 rval = probe->dtpr_arg;
8019 mutex_exit(&dtrace_lock);
8025 * Copy a probe into a probe description.
8028 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8030 bzero(pdp, sizeof (dtrace_probedesc_t));
8031 pdp->dtpd_id = prp->dtpr_id;
8033 (void) strncpy(pdp->dtpd_provider,
8034 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8036 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8037 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8038 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8042 * Called to indicate that a probe -- or probes -- should be provided by a
8043 * specfied provider. If the specified description is NULL, the provider will
8044 * be told to provide all of its probes. (This is done whenever a new
8045 * consumer comes along, or whenever a retained enabling is to be matched.) If
8046 * the specified description is non-NULL, the provider is given the
8047 * opportunity to dynamically provide the specified probe, allowing providers
8048 * to support the creation of probes on-the-fly. (So-called _autocreated_
8049 * probes.) If the provider is NULL, the operations will be applied to all
8050 * providers; if the provider is non-NULL the operations will only be applied
8051 * to the specified provider. The dtrace_provider_lock must be held, and the
8052 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8053 * will need to grab the dtrace_lock when it reenters the framework through
8054 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8057 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8064 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8068 prv = dtrace_provider;
8073 * First, call the blanket provide operation.
8075 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8078 * Now call the per-module provide operation. We will grab
8079 * mod_lock to prevent the list from being modified. Note
8080 * that this also prevents the mod_busy bits from changing.
8081 * (mod_busy can only be changed with mod_lock held.)
8083 mutex_enter(&mod_lock);
8088 if (ctl->mod_busy || ctl->mod_mp == NULL)
8091 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8093 } while ((ctl = ctl->mod_next) != &modules);
8096 mutex_exit(&mod_lock);
8097 } while (all && (prv = prv->dtpv_next) != NULL);
8102 * Iterate over each probe, and call the Framework-to-Provider API function
8106 dtrace_probe_foreach(uintptr_t offs)
8108 dtrace_provider_t *prov;
8109 void (*func)(void *, dtrace_id_t, void *);
8110 dtrace_probe_t *probe;
8111 dtrace_icookie_t cookie;
8115 * We disable interrupts to walk through the probe array. This is
8116 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8117 * won't see stale data.
8119 cookie = dtrace_interrupt_disable();
8121 for (i = 0; i < dtrace_nprobes; i++) {
8122 if ((probe = dtrace_probes[i]) == NULL)
8125 if (probe->dtpr_ecb == NULL) {
8127 * This probe isn't enabled -- don't call the function.
8132 prov = probe->dtpr_provider;
8133 func = *((void(**)(void *, dtrace_id_t, void *))
8134 ((uintptr_t)&prov->dtpv_pops + offs));
8136 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8139 dtrace_interrupt_enable(cookie);
8144 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8146 dtrace_probekey_t pkey;
8151 ASSERT(MUTEX_HELD(&dtrace_lock));
8152 dtrace_ecb_create_cache = NULL;
8156 * If we're passed a NULL description, we're being asked to
8157 * create an ECB with a NULL probe.
8159 (void) dtrace_ecb_create_enable(NULL, enab);
8163 dtrace_probekey(desc, &pkey);
8164 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8165 &priv, &uid, &zoneid);
8167 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8172 * DTrace Helper Provider Functions
8175 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8177 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8178 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8179 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8183 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8184 const dof_provider_t *dofprov, char *strtab)
8186 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8187 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8188 dofprov->dofpv_provattr);
8189 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8190 dofprov->dofpv_modattr);
8191 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8192 dofprov->dofpv_funcattr);
8193 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8194 dofprov->dofpv_nameattr);
8195 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8196 dofprov->dofpv_argsattr);
8200 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8202 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8203 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8204 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8205 dof_provider_t *provider;
8207 uint32_t *off, *enoff;
8211 dtrace_helper_provdesc_t dhpv;
8212 dtrace_helper_probedesc_t dhpb;
8213 dtrace_meta_t *meta = dtrace_meta_pid;
8214 dtrace_mops_t *mops = &meta->dtm_mops;
8217 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8218 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8219 provider->dofpv_strtab * dof->dofh_secsize);
8220 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8221 provider->dofpv_probes * dof->dofh_secsize);
8222 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8223 provider->dofpv_prargs * dof->dofh_secsize);
8224 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8225 provider->dofpv_proffs * dof->dofh_secsize);
8227 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8228 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8229 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8233 * See dtrace_helper_provider_validate().
8235 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8236 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8237 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8238 provider->dofpv_prenoffs * dof->dofh_secsize);
8239 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8242 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8245 * Create the provider.
8247 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8249 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8255 * Create the probes.
8257 for (i = 0; i < nprobes; i++) {
8258 probe = (dof_probe_t *)(uintptr_t)(daddr +
8259 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8261 dhpb.dthpb_mod = dhp->dofhp_mod;
8262 dhpb.dthpb_func = strtab + probe->dofpr_func;
8263 dhpb.dthpb_name = strtab + probe->dofpr_name;
8264 dhpb.dthpb_base = probe->dofpr_addr;
8265 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8266 dhpb.dthpb_noffs = probe->dofpr_noffs;
8267 if (enoff != NULL) {
8268 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8269 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8271 dhpb.dthpb_enoffs = NULL;
8272 dhpb.dthpb_nenoffs = 0;
8274 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8275 dhpb.dthpb_nargc = probe->dofpr_nargc;
8276 dhpb.dthpb_xargc = probe->dofpr_xargc;
8277 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8278 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8280 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8285 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8287 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8288 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8291 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8293 for (i = 0; i < dof->dofh_secnum; i++) {
8294 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8295 dof->dofh_secoff + i * dof->dofh_secsize);
8297 if (sec->dofs_type != DOF_SECT_PROVIDER)
8300 dtrace_helper_provide_one(dhp, sec, pid);
8304 * We may have just created probes, so we must now rematch against
8305 * any retained enablings. Note that this call will acquire both
8306 * cpu_lock and dtrace_lock; the fact that we are holding
8307 * dtrace_meta_lock now is what defines the ordering with respect to
8308 * these three locks.
8310 dtrace_enabling_matchall();
8314 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8316 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8317 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8319 dof_provider_t *provider;
8321 dtrace_helper_provdesc_t dhpv;
8322 dtrace_meta_t *meta = dtrace_meta_pid;
8323 dtrace_mops_t *mops = &meta->dtm_mops;
8325 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8326 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8327 provider->dofpv_strtab * dof->dofh_secsize);
8329 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8332 * Create the provider.
8334 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8336 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8342 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8344 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8345 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8348 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8350 for (i = 0; i < dof->dofh_secnum; i++) {
8351 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8352 dof->dofh_secoff + i * dof->dofh_secsize);
8354 if (sec->dofs_type != DOF_SECT_PROVIDER)
8357 dtrace_helper_provider_remove_one(dhp, sec, pid);
8362 * DTrace Meta Provider-to-Framework API Functions
8364 * These functions implement the Meta Provider-to-Framework API, as described
8365 * in <sys/dtrace.h>.
8368 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8369 dtrace_meta_provider_id_t *idp)
8371 dtrace_meta_t *meta;
8372 dtrace_helpers_t *help, *next;
8375 *idp = DTRACE_METAPROVNONE;
8378 * We strictly don't need the name, but we hold onto it for
8379 * debuggability. All hail error queues!
8382 cmn_err(CE_WARN, "failed to register meta-provider: "
8388 mops->dtms_create_probe == NULL ||
8389 mops->dtms_provide_pid == NULL ||
8390 mops->dtms_remove_pid == NULL) {
8391 cmn_err(CE_WARN, "failed to register meta-register %s: "
8392 "invalid ops", name);
8396 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8397 meta->dtm_mops = *mops;
8398 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8399 (void) strcpy(meta->dtm_name, name);
8400 meta->dtm_arg = arg;
8402 mutex_enter(&dtrace_meta_lock);
8403 mutex_enter(&dtrace_lock);
8405 if (dtrace_meta_pid != NULL) {
8406 mutex_exit(&dtrace_lock);
8407 mutex_exit(&dtrace_meta_lock);
8408 cmn_err(CE_WARN, "failed to register meta-register %s: "
8409 "user-land meta-provider exists", name);
8410 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8411 kmem_free(meta, sizeof (dtrace_meta_t));
8415 dtrace_meta_pid = meta;
8416 *idp = (dtrace_meta_provider_id_t)meta;
8419 * If there are providers and probes ready to go, pass them
8420 * off to the new meta provider now.
8423 help = dtrace_deferred_pid;
8424 dtrace_deferred_pid = NULL;
8426 mutex_exit(&dtrace_lock);
8428 while (help != NULL) {
8429 for (i = 0; i < help->dthps_nprovs; i++) {
8430 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8434 next = help->dthps_next;
8435 help->dthps_next = NULL;
8436 help->dthps_prev = NULL;
8437 help->dthps_deferred = 0;
8441 mutex_exit(&dtrace_meta_lock);
8447 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8449 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8451 mutex_enter(&dtrace_meta_lock);
8452 mutex_enter(&dtrace_lock);
8454 if (old == dtrace_meta_pid) {
8455 pp = &dtrace_meta_pid;
8457 panic("attempt to unregister non-existent "
8458 "dtrace meta-provider %p\n", (void *)old);
8461 if (old->dtm_count != 0) {
8462 mutex_exit(&dtrace_lock);
8463 mutex_exit(&dtrace_meta_lock);
8469 mutex_exit(&dtrace_lock);
8470 mutex_exit(&dtrace_meta_lock);
8472 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8473 kmem_free(old, sizeof (dtrace_meta_t));
8480 * DTrace DIF Object Functions
8483 dtrace_difo_err(uint_t pc, const char *format, ...)
8485 if (dtrace_err_verbose) {
8488 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8489 va_start(alist, format);
8490 (void) vuprintf(format, alist);
8494 #ifdef DTRACE_ERRDEBUG
8495 dtrace_errdebug(format);
8501 * Validate a DTrace DIF object by checking the IR instructions. The following
8502 * rules are currently enforced by dtrace_difo_validate():
8504 * 1. Each instruction must have a valid opcode
8505 * 2. Each register, string, variable, or subroutine reference must be valid
8506 * 3. No instruction can modify register %r0 (must be zero)
8507 * 4. All instruction reserved bits must be set to zero
8508 * 5. The last instruction must be a "ret" instruction
8509 * 6. All branch targets must reference a valid instruction _after_ the branch
8512 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8516 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8520 kcheckload = cr == NULL ||
8521 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8523 dp->dtdo_destructive = 0;
8525 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8526 dif_instr_t instr = dp->dtdo_buf[pc];
8528 uint_t r1 = DIF_INSTR_R1(instr);
8529 uint_t r2 = DIF_INSTR_R2(instr);
8530 uint_t rd = DIF_INSTR_RD(instr);
8531 uint_t rs = DIF_INSTR_RS(instr);
8532 uint_t label = DIF_INSTR_LABEL(instr);
8533 uint_t v = DIF_INSTR_VAR(instr);
8534 uint_t subr = DIF_INSTR_SUBR(instr);
8535 uint_t type = DIF_INSTR_TYPE(instr);
8536 uint_t op = DIF_INSTR_OP(instr);
8554 err += efunc(pc, "invalid register %u\n", r1);
8556 err += efunc(pc, "invalid register %u\n", r2);
8558 err += efunc(pc, "invalid register %u\n", rd);
8560 err += efunc(pc, "cannot write to %r0\n");
8566 err += efunc(pc, "invalid register %u\n", r1);
8568 err += efunc(pc, "non-zero reserved bits\n");
8570 err += efunc(pc, "invalid register %u\n", rd);
8572 err += efunc(pc, "cannot write to %r0\n");
8582 err += efunc(pc, "invalid register %u\n", r1);
8584 err += efunc(pc, "non-zero reserved bits\n");
8586 err += efunc(pc, "invalid register %u\n", rd);
8588 err += efunc(pc, "cannot write to %r0\n");
8590 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8591 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8601 err += efunc(pc, "invalid register %u\n", r1);
8603 err += efunc(pc, "non-zero reserved bits\n");
8605 err += efunc(pc, "invalid register %u\n", rd);
8607 err += efunc(pc, "cannot write to %r0\n");
8617 err += efunc(pc, "invalid register %u\n", r1);
8619 err += efunc(pc, "non-zero reserved bits\n");
8621 err += efunc(pc, "invalid register %u\n", rd);
8623 err += efunc(pc, "cannot write to %r0\n");
8630 err += efunc(pc, "invalid register %u\n", r1);
8632 err += efunc(pc, "non-zero reserved bits\n");
8634 err += efunc(pc, "invalid register %u\n", rd);
8636 err += efunc(pc, "cannot write to 0 address\n");
8641 err += efunc(pc, "invalid register %u\n", r1);
8643 err += efunc(pc, "invalid register %u\n", r2);
8645 err += efunc(pc, "non-zero reserved bits\n");
8649 err += efunc(pc, "invalid register %u\n", r1);
8650 if (r2 != 0 || rd != 0)
8651 err += efunc(pc, "non-zero reserved bits\n");
8664 if (label >= dp->dtdo_len) {
8665 err += efunc(pc, "invalid branch target %u\n",
8669 err += efunc(pc, "backward branch to %u\n",
8674 if (r1 != 0 || r2 != 0)
8675 err += efunc(pc, "non-zero reserved bits\n");
8677 err += efunc(pc, "invalid register %u\n", rd);
8681 case DIF_OP_FLUSHTS:
8682 if (r1 != 0 || r2 != 0 || rd != 0)
8683 err += efunc(pc, "non-zero reserved bits\n");
8686 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8687 err += efunc(pc, "invalid integer ref %u\n",
8688 DIF_INSTR_INTEGER(instr));
8691 err += efunc(pc, "invalid register %u\n", rd);
8693 err += efunc(pc, "cannot write to %r0\n");
8696 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8697 err += efunc(pc, "invalid string ref %u\n",
8698 DIF_INSTR_STRING(instr));
8701 err += efunc(pc, "invalid register %u\n", rd);
8703 err += efunc(pc, "cannot write to %r0\n");
8707 if (r1 > DIF_VAR_ARRAY_MAX)
8708 err += efunc(pc, "invalid array %u\n", r1);
8710 err += efunc(pc, "invalid register %u\n", r2);
8712 err += efunc(pc, "invalid register %u\n", rd);
8714 err += efunc(pc, "cannot write to %r0\n");
8721 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8722 err += efunc(pc, "invalid variable %u\n", v);
8724 err += efunc(pc, "invalid register %u\n", rd);
8726 err += efunc(pc, "cannot write to %r0\n");
8733 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8734 err += efunc(pc, "invalid variable %u\n", v);
8736 err += efunc(pc, "invalid register %u\n", rd);
8739 if (subr > DIF_SUBR_MAX)
8740 err += efunc(pc, "invalid subr %u\n", subr);
8742 err += efunc(pc, "invalid register %u\n", rd);
8744 err += efunc(pc, "cannot write to %r0\n");
8746 if (subr == DIF_SUBR_COPYOUT ||
8747 subr == DIF_SUBR_COPYOUTSTR) {
8748 dp->dtdo_destructive = 1;
8752 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8753 err += efunc(pc, "invalid ref type %u\n", type);
8755 err += efunc(pc, "invalid register %u\n", r2);
8757 err += efunc(pc, "invalid register %u\n", rs);
8760 if (type != DIF_TYPE_CTF)
8761 err += efunc(pc, "invalid val type %u\n", type);
8763 err += efunc(pc, "invalid register %u\n", r2);
8765 err += efunc(pc, "invalid register %u\n", rs);
8768 err += efunc(pc, "invalid opcode %u\n",
8769 DIF_INSTR_OP(instr));
8773 if (dp->dtdo_len != 0 &&
8774 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8775 err += efunc(dp->dtdo_len - 1,
8776 "expected 'ret' as last DIF instruction\n");
8779 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8781 * If we're not returning by reference, the size must be either
8782 * 0 or the size of one of the base types.
8784 switch (dp->dtdo_rtype.dtdt_size) {
8786 case sizeof (uint8_t):
8787 case sizeof (uint16_t):
8788 case sizeof (uint32_t):
8789 case sizeof (uint64_t):
8793 err += efunc(dp->dtdo_len - 1, "bad return size");
8797 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8798 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8799 dtrace_diftype_t *vt, *et;
8802 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8803 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8804 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8805 err += efunc(i, "unrecognized variable scope %d\n",
8810 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8811 v->dtdv_kind != DIFV_KIND_SCALAR) {
8812 err += efunc(i, "unrecognized variable type %d\n",
8817 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8818 err += efunc(i, "%d exceeds variable id limit\n", id);
8822 if (id < DIF_VAR_OTHER_UBASE)
8826 * For user-defined variables, we need to check that this
8827 * definition is identical to any previous definition that we
8830 ndx = id - DIF_VAR_OTHER_UBASE;
8832 switch (v->dtdv_scope) {
8833 case DIFV_SCOPE_GLOBAL:
8834 if (ndx < vstate->dtvs_nglobals) {
8835 dtrace_statvar_t *svar;
8837 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8838 existing = &svar->dtsv_var;
8843 case DIFV_SCOPE_THREAD:
8844 if (ndx < vstate->dtvs_ntlocals)
8845 existing = &vstate->dtvs_tlocals[ndx];
8848 case DIFV_SCOPE_LOCAL:
8849 if (ndx < vstate->dtvs_nlocals) {
8850 dtrace_statvar_t *svar;
8852 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8853 existing = &svar->dtsv_var;
8861 if (vt->dtdt_flags & DIF_TF_BYREF) {
8862 if (vt->dtdt_size == 0) {
8863 err += efunc(i, "zero-sized variable\n");
8867 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8868 vt->dtdt_size > dtrace_global_maxsize) {
8869 err += efunc(i, "oversized by-ref global\n");
8874 if (existing == NULL || existing->dtdv_id == 0)
8877 ASSERT(existing->dtdv_id == v->dtdv_id);
8878 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8880 if (existing->dtdv_kind != v->dtdv_kind)
8881 err += efunc(i, "%d changed variable kind\n", id);
8883 et = &existing->dtdv_type;
8885 if (vt->dtdt_flags != et->dtdt_flags) {
8886 err += efunc(i, "%d changed variable type flags\n", id);
8890 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8891 err += efunc(i, "%d changed variable type size\n", id);
8900 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
8901 * are much more constrained than normal DIFOs. Specifically, they may
8904 * 1. Make calls to subroutines other than copyin(), copyinstr() or
8905 * miscellaneous string routines
8906 * 2. Access DTrace variables other than the args[] array, and the
8907 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
8908 * 3. Have thread-local variables.
8909 * 4. Have dynamic variables.
8912 dtrace_difo_validate_helper(dtrace_difo_t *dp)
8914 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8918 for (pc = 0; pc < dp->dtdo_len; pc++) {
8919 dif_instr_t instr = dp->dtdo_buf[pc];
8921 uint_t v = DIF_INSTR_VAR(instr);
8922 uint_t subr = DIF_INSTR_SUBR(instr);
8923 uint_t op = DIF_INSTR_OP(instr);
8978 case DIF_OP_FLUSHTS:
8990 if (v >= DIF_VAR_OTHER_UBASE)
8993 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
8996 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
8997 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
8998 v == DIF_VAR_EXECARGS ||
8999 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9000 v == DIF_VAR_UID || v == DIF_VAR_GID)
9003 err += efunc(pc, "illegal variable %u\n", v);
9010 err += efunc(pc, "illegal dynamic variable load\n");
9016 err += efunc(pc, "illegal dynamic variable store\n");
9020 if (subr == DIF_SUBR_ALLOCA ||
9021 subr == DIF_SUBR_BCOPY ||
9022 subr == DIF_SUBR_COPYIN ||
9023 subr == DIF_SUBR_COPYINTO ||
9024 subr == DIF_SUBR_COPYINSTR ||
9025 subr == DIF_SUBR_INDEX ||
9026 subr == DIF_SUBR_INET_NTOA ||
9027 subr == DIF_SUBR_INET_NTOA6 ||
9028 subr == DIF_SUBR_INET_NTOP ||
9029 subr == DIF_SUBR_LLTOSTR ||
9030 subr == DIF_SUBR_RINDEX ||
9031 subr == DIF_SUBR_STRCHR ||
9032 subr == DIF_SUBR_STRJOIN ||
9033 subr == DIF_SUBR_STRRCHR ||
9034 subr == DIF_SUBR_STRSTR ||
9035 subr == DIF_SUBR_HTONS ||
9036 subr == DIF_SUBR_HTONL ||
9037 subr == DIF_SUBR_HTONLL ||
9038 subr == DIF_SUBR_NTOHS ||
9039 subr == DIF_SUBR_NTOHL ||
9040 subr == DIF_SUBR_NTOHLL ||
9041 subr == DIF_SUBR_MEMREF ||
9042 subr == DIF_SUBR_TYPEREF)
9045 err += efunc(pc, "invalid subr %u\n", subr);
9049 err += efunc(pc, "invalid opcode %u\n",
9050 DIF_INSTR_OP(instr));
9058 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9062 dtrace_difo_cacheable(dtrace_difo_t *dp)
9069 for (i = 0; i < dp->dtdo_varlen; i++) {
9070 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9072 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9075 switch (v->dtdv_id) {
9076 case DIF_VAR_CURTHREAD:
9079 case DIF_VAR_EXECARGS:
9080 case DIF_VAR_EXECNAME:
9081 case DIF_VAR_ZONENAME:
9090 * This DIF object may be cacheable. Now we need to look for any
9091 * array loading instructions, any memory loading instructions, or
9092 * any stores to thread-local variables.
9094 for (i = 0; i < dp->dtdo_len; i++) {
9095 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9097 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9098 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9099 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9100 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9108 dtrace_difo_hold(dtrace_difo_t *dp)
9112 ASSERT(MUTEX_HELD(&dtrace_lock));
9115 ASSERT(dp->dtdo_refcnt != 0);
9118 * We need to check this DIF object for references to the variable
9119 * DIF_VAR_VTIMESTAMP.
9121 for (i = 0; i < dp->dtdo_varlen; i++) {
9122 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9124 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9127 if (dtrace_vtime_references++ == 0)
9128 dtrace_vtime_enable();
9133 * This routine calculates the dynamic variable chunksize for a given DIF
9134 * object. The calculation is not fool-proof, and can probably be tricked by
9135 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9136 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9137 * if a dynamic variable size exceeds the chunksize.
9140 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9143 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9144 const dif_instr_t *text = dp->dtdo_buf;
9150 for (pc = 0; pc < dp->dtdo_len; pc++) {
9151 dif_instr_t instr = text[pc];
9152 uint_t op = DIF_INSTR_OP(instr);
9153 uint_t rd = DIF_INSTR_RD(instr);
9154 uint_t r1 = DIF_INSTR_R1(instr);
9158 dtrace_key_t *key = tupregs;
9162 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9167 key = &tupregs[DIF_DTR_NREGS];
9168 key[0].dttk_size = 0;
9169 key[1].dttk_size = 0;
9171 scope = DIFV_SCOPE_THREAD;
9178 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9179 key[nkeys++].dttk_size = 0;
9181 key[nkeys++].dttk_size = 0;
9183 if (op == DIF_OP_STTAA) {
9184 scope = DIFV_SCOPE_THREAD;
9186 scope = DIFV_SCOPE_GLOBAL;
9192 if (ttop == DIF_DTR_NREGS)
9195 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9197 * If the register for the size of the "pushtr"
9198 * is %r0 (or the value is 0) and the type is
9199 * a string, we'll use the system-wide default
9202 tupregs[ttop++].dttk_size =
9203 dtrace_strsize_default;
9208 tupregs[ttop++].dttk_size = sval;
9214 if (ttop == DIF_DTR_NREGS)
9217 tupregs[ttop++].dttk_size = 0;
9220 case DIF_OP_FLUSHTS:
9237 * We have a dynamic variable allocation; calculate its size.
9239 for (ksize = 0, i = 0; i < nkeys; i++)
9240 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9242 size = sizeof (dtrace_dynvar_t);
9243 size += sizeof (dtrace_key_t) * (nkeys - 1);
9247 * Now we need to determine the size of the stored data.
9249 id = DIF_INSTR_VAR(instr);
9251 for (i = 0; i < dp->dtdo_varlen; i++) {
9252 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9254 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9255 size += v->dtdv_type.dtdt_size;
9260 if (i == dp->dtdo_varlen)
9264 * We have the size. If this is larger than the chunk size
9265 * for our dynamic variable state, reset the chunk size.
9267 size = P2ROUNDUP(size, sizeof (uint64_t));
9269 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9270 vstate->dtvs_dynvars.dtds_chunksize = size;
9275 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9277 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9280 ASSERT(MUTEX_HELD(&dtrace_lock));
9281 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9283 for (i = 0; i < dp->dtdo_varlen; i++) {
9284 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9285 dtrace_statvar_t *svar, ***svarp = NULL;
9287 uint8_t scope = v->dtdv_scope;
9290 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9293 id -= DIF_VAR_OTHER_UBASE;
9296 case DIFV_SCOPE_THREAD:
9297 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9298 dtrace_difv_t *tlocals;
9300 if ((ntlocals = (otlocals << 1)) == 0)
9303 osz = otlocals * sizeof (dtrace_difv_t);
9304 nsz = ntlocals * sizeof (dtrace_difv_t);
9306 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9309 bcopy(vstate->dtvs_tlocals,
9311 kmem_free(vstate->dtvs_tlocals, osz);
9314 vstate->dtvs_tlocals = tlocals;
9315 vstate->dtvs_ntlocals = ntlocals;
9318 vstate->dtvs_tlocals[id] = *v;
9321 case DIFV_SCOPE_LOCAL:
9322 np = &vstate->dtvs_nlocals;
9323 svarp = &vstate->dtvs_locals;
9325 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9326 dsize = NCPU * (v->dtdv_type.dtdt_size +
9329 dsize = NCPU * sizeof (uint64_t);
9333 case DIFV_SCOPE_GLOBAL:
9334 np = &vstate->dtvs_nglobals;
9335 svarp = &vstate->dtvs_globals;
9337 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9338 dsize = v->dtdv_type.dtdt_size +
9347 while (id >= (oldsvars = *np)) {
9348 dtrace_statvar_t **statics;
9349 int newsvars, oldsize, newsize;
9351 if ((newsvars = (oldsvars << 1)) == 0)
9354 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9355 newsize = newsvars * sizeof (dtrace_statvar_t *);
9357 statics = kmem_zalloc(newsize, KM_SLEEP);
9360 bcopy(*svarp, statics, oldsize);
9361 kmem_free(*svarp, oldsize);
9368 if ((svar = (*svarp)[id]) == NULL) {
9369 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9370 svar->dtsv_var = *v;
9372 if ((svar->dtsv_size = dsize) != 0) {
9373 svar->dtsv_data = (uint64_t)(uintptr_t)
9374 kmem_zalloc(dsize, KM_SLEEP);
9377 (*svarp)[id] = svar;
9380 svar->dtsv_refcnt++;
9383 dtrace_difo_chunksize(dp, vstate);
9384 dtrace_difo_hold(dp);
9387 static dtrace_difo_t *
9388 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9393 ASSERT(dp->dtdo_buf != NULL);
9394 ASSERT(dp->dtdo_refcnt != 0);
9396 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9398 ASSERT(dp->dtdo_buf != NULL);
9399 sz = dp->dtdo_len * sizeof (dif_instr_t);
9400 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9401 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9402 new->dtdo_len = dp->dtdo_len;
9404 if (dp->dtdo_strtab != NULL) {
9405 ASSERT(dp->dtdo_strlen != 0);
9406 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9407 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9408 new->dtdo_strlen = dp->dtdo_strlen;
9411 if (dp->dtdo_inttab != NULL) {
9412 ASSERT(dp->dtdo_intlen != 0);
9413 sz = dp->dtdo_intlen * sizeof (uint64_t);
9414 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9415 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9416 new->dtdo_intlen = dp->dtdo_intlen;
9419 if (dp->dtdo_vartab != NULL) {
9420 ASSERT(dp->dtdo_varlen != 0);
9421 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9422 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9423 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9424 new->dtdo_varlen = dp->dtdo_varlen;
9427 dtrace_difo_init(new, vstate);
9432 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9436 ASSERT(dp->dtdo_refcnt == 0);
9438 for (i = 0; i < dp->dtdo_varlen; i++) {
9439 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9440 dtrace_statvar_t *svar, **svarp = NULL;
9442 uint8_t scope = v->dtdv_scope;
9446 case DIFV_SCOPE_THREAD:
9449 case DIFV_SCOPE_LOCAL:
9450 np = &vstate->dtvs_nlocals;
9451 svarp = vstate->dtvs_locals;
9454 case DIFV_SCOPE_GLOBAL:
9455 np = &vstate->dtvs_nglobals;
9456 svarp = vstate->dtvs_globals;
9463 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9466 id -= DIF_VAR_OTHER_UBASE;
9470 ASSERT(svar != NULL);
9471 ASSERT(svar->dtsv_refcnt > 0);
9473 if (--svar->dtsv_refcnt > 0)
9476 if (svar->dtsv_size != 0) {
9477 ASSERT(svar->dtsv_data != 0);
9478 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9482 kmem_free(svar, sizeof (dtrace_statvar_t));
9486 if (dp->dtdo_buf != NULL)
9487 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9488 if (dp->dtdo_inttab != NULL)
9489 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9490 if (dp->dtdo_strtab != NULL)
9491 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9492 if (dp->dtdo_vartab != NULL)
9493 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9495 kmem_free(dp, sizeof (dtrace_difo_t));
9499 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9503 ASSERT(MUTEX_HELD(&dtrace_lock));
9504 ASSERT(dp->dtdo_refcnt != 0);
9506 for (i = 0; i < dp->dtdo_varlen; i++) {
9507 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9509 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9512 ASSERT(dtrace_vtime_references > 0);
9513 if (--dtrace_vtime_references == 0)
9514 dtrace_vtime_disable();
9517 if (--dp->dtdo_refcnt == 0)
9518 dtrace_difo_destroy(dp, vstate);
9522 * DTrace Format Functions
9525 dtrace_format_add(dtrace_state_t *state, char *str)
9528 uint16_t ndx, len = strlen(str) + 1;
9530 fmt = kmem_zalloc(len, KM_SLEEP);
9531 bcopy(str, fmt, len);
9533 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9534 if (state->dts_formats[ndx] == NULL) {
9535 state->dts_formats[ndx] = fmt;
9540 if (state->dts_nformats == USHRT_MAX) {
9542 * This is only likely if a denial-of-service attack is being
9543 * attempted. As such, it's okay to fail silently here.
9545 kmem_free(fmt, len);
9550 * For simplicity, we always resize the formats array to be exactly the
9551 * number of formats.
9553 ndx = state->dts_nformats++;
9554 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9556 if (state->dts_formats != NULL) {
9558 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9559 kmem_free(state->dts_formats, ndx * sizeof (char *));
9562 state->dts_formats = new;
9563 state->dts_formats[ndx] = fmt;
9569 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9573 ASSERT(state->dts_formats != NULL);
9574 ASSERT(format <= state->dts_nformats);
9575 ASSERT(state->dts_formats[format - 1] != NULL);
9577 fmt = state->dts_formats[format - 1];
9578 kmem_free(fmt, strlen(fmt) + 1);
9579 state->dts_formats[format - 1] = NULL;
9583 dtrace_format_destroy(dtrace_state_t *state)
9587 if (state->dts_nformats == 0) {
9588 ASSERT(state->dts_formats == NULL);
9592 ASSERT(state->dts_formats != NULL);
9594 for (i = 0; i < state->dts_nformats; i++) {
9595 char *fmt = state->dts_formats[i];
9600 kmem_free(fmt, strlen(fmt) + 1);
9603 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9604 state->dts_nformats = 0;
9605 state->dts_formats = NULL;
9609 * DTrace Predicate Functions
9611 static dtrace_predicate_t *
9612 dtrace_predicate_create(dtrace_difo_t *dp)
9614 dtrace_predicate_t *pred;
9616 ASSERT(MUTEX_HELD(&dtrace_lock));
9617 ASSERT(dp->dtdo_refcnt != 0);
9619 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9620 pred->dtp_difo = dp;
9621 pred->dtp_refcnt = 1;
9623 if (!dtrace_difo_cacheable(dp))
9626 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9628 * This is only theoretically possible -- we have had 2^32
9629 * cacheable predicates on this machine. We cannot allow any
9630 * more predicates to become cacheable: as unlikely as it is,
9631 * there may be a thread caching a (now stale) predicate cache
9632 * ID. (N.B.: the temptation is being successfully resisted to
9633 * have this cmn_err() "Holy shit -- we executed this code!")
9638 pred->dtp_cacheid = dtrace_predcache_id++;
9644 dtrace_predicate_hold(dtrace_predicate_t *pred)
9646 ASSERT(MUTEX_HELD(&dtrace_lock));
9647 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9648 ASSERT(pred->dtp_refcnt > 0);
9654 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9656 dtrace_difo_t *dp = pred->dtp_difo;
9658 ASSERT(MUTEX_HELD(&dtrace_lock));
9659 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9660 ASSERT(pred->dtp_refcnt > 0);
9662 if (--pred->dtp_refcnt == 0) {
9663 dtrace_difo_release(pred->dtp_difo, vstate);
9664 kmem_free(pred, sizeof (dtrace_predicate_t));
9669 * DTrace Action Description Functions
9671 static dtrace_actdesc_t *
9672 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9673 uint64_t uarg, uint64_t arg)
9675 dtrace_actdesc_t *act;
9678 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9679 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9682 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9683 act->dtad_kind = kind;
9684 act->dtad_ntuple = ntuple;
9685 act->dtad_uarg = uarg;
9686 act->dtad_arg = arg;
9687 act->dtad_refcnt = 1;
9693 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9695 ASSERT(act->dtad_refcnt >= 1);
9700 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9702 dtrace_actkind_t kind = act->dtad_kind;
9705 ASSERT(act->dtad_refcnt >= 1);
9707 if (--act->dtad_refcnt != 0)
9710 if ((dp = act->dtad_difo) != NULL)
9711 dtrace_difo_release(dp, vstate);
9713 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9714 char *str = (char *)(uintptr_t)act->dtad_arg;
9717 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9718 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9722 kmem_free(str, strlen(str) + 1);
9725 kmem_free(act, sizeof (dtrace_actdesc_t));
9729 * DTrace ECB Functions
9731 static dtrace_ecb_t *
9732 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9737 ASSERT(MUTEX_HELD(&dtrace_lock));
9739 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9740 ecb->dte_predicate = NULL;
9741 ecb->dte_probe = probe;
9744 * The default size is the size of the default action: recording
9747 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9748 ecb->dte_alignment = sizeof (dtrace_epid_t);
9750 epid = state->dts_epid++;
9752 if (epid - 1 >= state->dts_necbs) {
9753 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9754 int necbs = state->dts_necbs << 1;
9756 ASSERT(epid == state->dts_necbs + 1);
9759 ASSERT(oecbs == NULL);
9763 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9766 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9768 dtrace_membar_producer();
9769 state->dts_ecbs = ecbs;
9771 if (oecbs != NULL) {
9773 * If this state is active, we must dtrace_sync()
9774 * before we can free the old dts_ecbs array: we're
9775 * coming in hot, and there may be active ring
9776 * buffer processing (which indexes into the dts_ecbs
9777 * array) on another CPU.
9779 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9782 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9785 dtrace_membar_producer();
9786 state->dts_necbs = necbs;
9789 ecb->dte_state = state;
9791 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9792 dtrace_membar_producer();
9793 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9799 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9801 dtrace_probe_t *probe = ecb->dte_probe;
9803 ASSERT(MUTEX_HELD(&cpu_lock));
9804 ASSERT(MUTEX_HELD(&dtrace_lock));
9805 ASSERT(ecb->dte_next == NULL);
9807 if (probe == NULL) {
9809 * This is the NULL probe -- there's nothing to do.
9814 if (probe->dtpr_ecb == NULL) {
9815 dtrace_provider_t *prov = probe->dtpr_provider;
9818 * We're the first ECB on this probe.
9820 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9822 if (ecb->dte_predicate != NULL)
9823 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9825 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9826 probe->dtpr_id, probe->dtpr_arg);
9829 * This probe is already active. Swing the last pointer to
9830 * point to the new ECB, and issue a dtrace_sync() to assure
9831 * that all CPUs have seen the change.
9833 ASSERT(probe->dtpr_ecb_last != NULL);
9834 probe->dtpr_ecb_last->dte_next = ecb;
9835 probe->dtpr_ecb_last = ecb;
9836 probe->dtpr_predcache = 0;
9843 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9845 uint32_t maxalign = sizeof (dtrace_epid_t);
9846 uint32_t align = sizeof (uint8_t), offs, diff;
9847 dtrace_action_t *act;
9849 uint32_t aggbase = UINT32_MAX;
9850 dtrace_state_t *state = ecb->dte_state;
9853 * If we record anything, we always record the epid. (And we always
9856 offs = sizeof (dtrace_epid_t);
9857 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_epid_t);
9859 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9860 dtrace_recdesc_t *rec = &act->dta_rec;
9862 if ((align = rec->dtrd_alignment) > maxalign)
9865 if (!wastuple && act->dta_intuple) {
9867 * This is the first record in a tuple. Align the
9868 * offset to be at offset 4 in an 8-byte aligned
9871 diff = offs + sizeof (dtrace_aggid_t);
9873 if ((diff = (diff & (sizeof (uint64_t) - 1))))
9874 offs += sizeof (uint64_t) - diff;
9876 aggbase = offs - sizeof (dtrace_aggid_t);
9877 ASSERT(!(aggbase & (sizeof (uint64_t) - 1)));
9881 if (rec->dtrd_size != 0 && (diff = (offs & (align - 1)))) {
9883 * The current offset is not properly aligned; align it.
9885 offs += align - diff;
9888 rec->dtrd_offset = offs;
9890 if (offs + rec->dtrd_size > ecb->dte_needed) {
9891 ecb->dte_needed = offs + rec->dtrd_size;
9893 if (ecb->dte_needed > state->dts_needed)
9894 state->dts_needed = ecb->dte_needed;
9897 if (DTRACEACT_ISAGG(act->dta_kind)) {
9898 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9899 dtrace_action_t *first = agg->dtag_first, *prev;
9901 ASSERT(rec->dtrd_size != 0 && first != NULL);
9903 ASSERT(aggbase != UINT32_MAX);
9905 agg->dtag_base = aggbase;
9907 while ((prev = first->dta_prev) != NULL &&
9908 DTRACEACT_ISAGG(prev->dta_kind)) {
9909 agg = (dtrace_aggregation_t *)prev;
9910 first = agg->dtag_first;
9914 offs = prev->dta_rec.dtrd_offset +
9915 prev->dta_rec.dtrd_size;
9917 offs = sizeof (dtrace_epid_t);
9921 if (!act->dta_intuple)
9922 ecb->dte_size = offs + rec->dtrd_size;
9924 offs += rec->dtrd_size;
9927 wastuple = act->dta_intuple;
9930 if ((act = ecb->dte_action) != NULL &&
9931 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
9932 ecb->dte_size == sizeof (dtrace_epid_t)) {
9934 * If the size is still sizeof (dtrace_epid_t), then all
9935 * actions store no data; set the size to 0.
9937 ecb->dte_alignment = maxalign;
9941 * If the needed space is still sizeof (dtrace_epid_t), then
9942 * all actions need no additional space; set the needed
9945 if (ecb->dte_needed == sizeof (dtrace_epid_t))
9946 ecb->dte_needed = 0;
9952 * Set our alignment, and make sure that the dte_size and dte_needed
9953 * are aligned to the size of an EPID.
9955 ecb->dte_alignment = maxalign;
9956 ecb->dte_size = (ecb->dte_size + (sizeof (dtrace_epid_t) - 1)) &
9957 ~(sizeof (dtrace_epid_t) - 1);
9958 ecb->dte_needed = (ecb->dte_needed + (sizeof (dtrace_epid_t) - 1)) &
9959 ~(sizeof (dtrace_epid_t) - 1);
9960 ASSERT(ecb->dte_size <= ecb->dte_needed);
9963 static dtrace_action_t *
9964 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
9966 dtrace_aggregation_t *agg;
9967 size_t size = sizeof (uint64_t);
9968 int ntuple = desc->dtad_ntuple;
9969 dtrace_action_t *act;
9970 dtrace_recdesc_t *frec;
9971 dtrace_aggid_t aggid;
9972 dtrace_state_t *state = ecb->dte_state;
9974 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
9975 agg->dtag_ecb = ecb;
9977 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
9979 switch (desc->dtad_kind) {
9981 agg->dtag_initial = INT64_MAX;
9982 agg->dtag_aggregate = dtrace_aggregate_min;
9986 agg->dtag_initial = INT64_MIN;
9987 agg->dtag_aggregate = dtrace_aggregate_max;
9990 case DTRACEAGG_COUNT:
9991 agg->dtag_aggregate = dtrace_aggregate_count;
9994 case DTRACEAGG_QUANTIZE:
9995 agg->dtag_aggregate = dtrace_aggregate_quantize;
9996 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10000 case DTRACEAGG_LQUANTIZE: {
10001 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10002 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10004 agg->dtag_initial = desc->dtad_arg;
10005 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10007 if (step == 0 || levels == 0)
10010 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10014 case DTRACEAGG_LLQUANTIZE: {
10015 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10016 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10017 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10018 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10021 agg->dtag_initial = desc->dtad_arg;
10022 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10024 if (factor < 2 || low >= high || nsteps < factor)
10028 * Now check that the number of steps evenly divides a power
10029 * of the factor. (This assures both integer bucket size and
10030 * linearity within each magnitude.)
10032 for (v = factor; v < nsteps; v *= factor)
10035 if ((v % nsteps) || (nsteps % factor))
10038 size = (dtrace_aggregate_llquantize_bucket(factor,
10039 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10043 case DTRACEAGG_AVG:
10044 agg->dtag_aggregate = dtrace_aggregate_avg;
10045 size = sizeof (uint64_t) * 2;
10048 case DTRACEAGG_STDDEV:
10049 agg->dtag_aggregate = dtrace_aggregate_stddev;
10050 size = sizeof (uint64_t) * 4;
10053 case DTRACEAGG_SUM:
10054 agg->dtag_aggregate = dtrace_aggregate_sum;
10061 agg->dtag_action.dta_rec.dtrd_size = size;
10067 * We must make sure that we have enough actions for the n-tuple.
10069 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10070 if (DTRACEACT_ISAGG(act->dta_kind))
10073 if (--ntuple == 0) {
10075 * This is the action with which our n-tuple begins.
10077 agg->dtag_first = act;
10083 * This n-tuple is short by ntuple elements. Return failure.
10085 ASSERT(ntuple != 0);
10087 kmem_free(agg, sizeof (dtrace_aggregation_t));
10092 * If the last action in the tuple has a size of zero, it's actually
10093 * an expression argument for the aggregating action.
10095 ASSERT(ecb->dte_action_last != NULL);
10096 act = ecb->dte_action_last;
10098 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10099 ASSERT(act->dta_difo != NULL);
10101 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10102 agg->dtag_hasarg = 1;
10106 * We need to allocate an id for this aggregation.
10109 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10110 VM_BESTFIT | VM_SLEEP);
10112 aggid = alloc_unr(state->dts_aggid_arena);
10115 if (aggid - 1 >= state->dts_naggregations) {
10116 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10117 dtrace_aggregation_t **aggs;
10118 int naggs = state->dts_naggregations << 1;
10119 int onaggs = state->dts_naggregations;
10121 ASSERT(aggid == state->dts_naggregations + 1);
10124 ASSERT(oaggs == NULL);
10128 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10130 if (oaggs != NULL) {
10131 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10132 kmem_free(oaggs, onaggs * sizeof (*aggs));
10135 state->dts_aggregations = aggs;
10136 state->dts_naggregations = naggs;
10139 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10140 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10142 frec = &agg->dtag_first->dta_rec;
10143 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10144 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10146 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10147 ASSERT(!act->dta_intuple);
10148 act->dta_intuple = 1;
10151 return (&agg->dtag_action);
10155 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10157 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10158 dtrace_state_t *state = ecb->dte_state;
10159 dtrace_aggid_t aggid = agg->dtag_id;
10161 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10163 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10165 free_unr(state->dts_aggid_arena, aggid);
10168 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10169 state->dts_aggregations[aggid - 1] = NULL;
10171 kmem_free(agg, sizeof (dtrace_aggregation_t));
10175 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10177 dtrace_action_t *action, *last;
10178 dtrace_difo_t *dp = desc->dtad_difo;
10179 uint32_t size = 0, align = sizeof (uint8_t), mask;
10180 uint16_t format = 0;
10181 dtrace_recdesc_t *rec;
10182 dtrace_state_t *state = ecb->dte_state;
10183 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
10184 uint64_t arg = desc->dtad_arg;
10186 ASSERT(MUTEX_HELD(&dtrace_lock));
10187 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10189 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10191 * If this is an aggregating action, there must be neither
10192 * a speculate nor a commit on the action chain.
10194 dtrace_action_t *act;
10196 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10197 if (act->dta_kind == DTRACEACT_COMMIT)
10200 if (act->dta_kind == DTRACEACT_SPECULATE)
10204 action = dtrace_ecb_aggregation_create(ecb, desc);
10206 if (action == NULL)
10209 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10210 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10211 dp != NULL && dp->dtdo_destructive)) {
10212 state->dts_destructive = 1;
10215 switch (desc->dtad_kind) {
10216 case DTRACEACT_PRINTF:
10217 case DTRACEACT_PRINTA:
10218 case DTRACEACT_SYSTEM:
10219 case DTRACEACT_FREOPEN:
10220 case DTRACEACT_DIFEXPR:
10222 * We know that our arg is a string -- turn it into a
10226 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10227 desc->dtad_kind == DTRACEACT_DIFEXPR);
10232 ASSERT(arg > KERNELBASE);
10234 format = dtrace_format_add(state,
10235 (char *)(uintptr_t)arg);
10239 case DTRACEACT_LIBACT:
10240 case DTRACEACT_TRACEMEM:
10241 case DTRACEACT_TRACEMEM_DYNSIZE:
10245 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10248 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10249 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10252 size = opt[DTRACEOPT_STRSIZE];
10257 case DTRACEACT_STACK:
10258 if ((nframes = arg) == 0) {
10259 nframes = opt[DTRACEOPT_STACKFRAMES];
10260 ASSERT(nframes > 0);
10264 size = nframes * sizeof (pc_t);
10267 case DTRACEACT_JSTACK:
10268 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10269 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10271 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10272 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10274 arg = DTRACE_USTACK_ARG(nframes, strsize);
10277 case DTRACEACT_USTACK:
10278 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10279 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10280 strsize = DTRACE_USTACK_STRSIZE(arg);
10281 nframes = opt[DTRACEOPT_USTACKFRAMES];
10282 ASSERT(nframes > 0);
10283 arg = DTRACE_USTACK_ARG(nframes, strsize);
10287 * Save a slot for the pid.
10289 size = (nframes + 1) * sizeof (uint64_t);
10290 size += DTRACE_USTACK_STRSIZE(arg);
10291 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10295 case DTRACEACT_SYM:
10296 case DTRACEACT_MOD:
10297 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10298 sizeof (uint64_t)) ||
10299 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10303 case DTRACEACT_USYM:
10304 case DTRACEACT_UMOD:
10305 case DTRACEACT_UADDR:
10307 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10308 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10312 * We have a slot for the pid, plus a slot for the
10313 * argument. To keep things simple (aligned with
10314 * bitness-neutral sizing), we store each as a 64-bit
10317 size = 2 * sizeof (uint64_t);
10320 case DTRACEACT_STOP:
10321 case DTRACEACT_BREAKPOINT:
10322 case DTRACEACT_PANIC:
10325 case DTRACEACT_CHILL:
10326 case DTRACEACT_DISCARD:
10327 case DTRACEACT_RAISE:
10332 case DTRACEACT_EXIT:
10334 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10335 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10339 case DTRACEACT_SPECULATE:
10340 if (ecb->dte_size > sizeof (dtrace_epid_t))
10346 state->dts_speculates = 1;
10349 case DTRACEACT_PRINTM:
10350 size = dp->dtdo_rtype.dtdt_size;
10353 case DTRACEACT_PRINTT:
10354 size = dp->dtdo_rtype.dtdt_size;
10357 case DTRACEACT_COMMIT: {
10358 dtrace_action_t *act = ecb->dte_action;
10360 for (; act != NULL; act = act->dta_next) {
10361 if (act->dta_kind == DTRACEACT_COMMIT)
10374 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10376 * If this is a data-storing action or a speculate,
10377 * we must be sure that there isn't a commit on the
10380 dtrace_action_t *act = ecb->dte_action;
10382 for (; act != NULL; act = act->dta_next) {
10383 if (act->dta_kind == DTRACEACT_COMMIT)
10388 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10389 action->dta_rec.dtrd_size = size;
10392 action->dta_refcnt = 1;
10393 rec = &action->dta_rec;
10394 size = rec->dtrd_size;
10396 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10397 if (!(size & mask)) {
10403 action->dta_kind = desc->dtad_kind;
10405 if ((action->dta_difo = dp) != NULL)
10406 dtrace_difo_hold(dp);
10408 rec->dtrd_action = action->dta_kind;
10409 rec->dtrd_arg = arg;
10410 rec->dtrd_uarg = desc->dtad_uarg;
10411 rec->dtrd_alignment = (uint16_t)align;
10412 rec->dtrd_format = format;
10414 if ((last = ecb->dte_action_last) != NULL) {
10415 ASSERT(ecb->dte_action != NULL);
10416 action->dta_prev = last;
10417 last->dta_next = action;
10419 ASSERT(ecb->dte_action == NULL);
10420 ecb->dte_action = action;
10423 ecb->dte_action_last = action;
10429 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10431 dtrace_action_t *act = ecb->dte_action, *next;
10432 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10436 if (act != NULL && act->dta_refcnt > 1) {
10437 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10440 for (; act != NULL; act = next) {
10441 next = act->dta_next;
10442 ASSERT(next != NULL || act == ecb->dte_action_last);
10443 ASSERT(act->dta_refcnt == 1);
10445 if ((format = act->dta_rec.dtrd_format) != 0)
10446 dtrace_format_remove(ecb->dte_state, format);
10448 if ((dp = act->dta_difo) != NULL)
10449 dtrace_difo_release(dp, vstate);
10451 if (DTRACEACT_ISAGG(act->dta_kind)) {
10452 dtrace_ecb_aggregation_destroy(ecb, act);
10454 kmem_free(act, sizeof (dtrace_action_t));
10459 ecb->dte_action = NULL;
10460 ecb->dte_action_last = NULL;
10461 ecb->dte_size = sizeof (dtrace_epid_t);
10465 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10468 * We disable the ECB by removing it from its probe.
10470 dtrace_ecb_t *pecb, *prev = NULL;
10471 dtrace_probe_t *probe = ecb->dte_probe;
10473 ASSERT(MUTEX_HELD(&dtrace_lock));
10475 if (probe == NULL) {
10477 * This is the NULL probe; there is nothing to disable.
10482 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10488 ASSERT(pecb != NULL);
10490 if (prev == NULL) {
10491 probe->dtpr_ecb = ecb->dte_next;
10493 prev->dte_next = ecb->dte_next;
10496 if (ecb == probe->dtpr_ecb_last) {
10497 ASSERT(ecb->dte_next == NULL);
10498 probe->dtpr_ecb_last = prev;
10502 * The ECB has been disconnected from the probe; now sync to assure
10503 * that all CPUs have seen the change before returning.
10507 if (probe->dtpr_ecb == NULL) {
10509 * That was the last ECB on the probe; clear the predicate
10510 * cache ID for the probe, disable it and sync one more time
10511 * to assure that we'll never hit it again.
10513 dtrace_provider_t *prov = probe->dtpr_provider;
10515 ASSERT(ecb->dte_next == NULL);
10516 ASSERT(probe->dtpr_ecb_last == NULL);
10517 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10518 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10519 probe->dtpr_id, probe->dtpr_arg);
10523 * There is at least one ECB remaining on the probe. If there
10524 * is _exactly_ one, set the probe's predicate cache ID to be
10525 * the predicate cache ID of the remaining ECB.
10527 ASSERT(probe->dtpr_ecb_last != NULL);
10528 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10530 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10531 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10533 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10536 probe->dtpr_predcache = p->dtp_cacheid;
10539 ecb->dte_next = NULL;
10544 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10546 dtrace_state_t *state = ecb->dte_state;
10547 dtrace_vstate_t *vstate = &state->dts_vstate;
10548 dtrace_predicate_t *pred;
10549 dtrace_epid_t epid = ecb->dte_epid;
10551 ASSERT(MUTEX_HELD(&dtrace_lock));
10552 ASSERT(ecb->dte_next == NULL);
10553 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10555 if ((pred = ecb->dte_predicate) != NULL)
10556 dtrace_predicate_release(pred, vstate);
10558 dtrace_ecb_action_remove(ecb);
10560 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10561 state->dts_ecbs[epid - 1] = NULL;
10563 kmem_free(ecb, sizeof (dtrace_ecb_t));
10566 static dtrace_ecb_t *
10567 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10568 dtrace_enabling_t *enab)
10571 dtrace_predicate_t *pred;
10572 dtrace_actdesc_t *act;
10573 dtrace_provider_t *prov;
10574 dtrace_ecbdesc_t *desc = enab->dten_current;
10576 ASSERT(MUTEX_HELD(&dtrace_lock));
10577 ASSERT(state != NULL);
10579 ecb = dtrace_ecb_add(state, probe);
10580 ecb->dte_uarg = desc->dted_uarg;
10582 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10583 dtrace_predicate_hold(pred);
10584 ecb->dte_predicate = pred;
10587 if (probe != NULL) {
10589 * If the provider shows more leg than the consumer is old
10590 * enough to see, we need to enable the appropriate implicit
10591 * predicate bits to prevent the ecb from activating at
10594 * Providers specifying DTRACE_PRIV_USER at register time
10595 * are stating that they need the /proc-style privilege
10596 * model to be enforced, and this is what DTRACE_COND_OWNER
10597 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10599 prov = probe->dtpr_provider;
10600 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10601 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10602 ecb->dte_cond |= DTRACE_COND_OWNER;
10604 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10605 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10606 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10609 * If the provider shows us kernel innards and the user
10610 * is lacking sufficient privilege, enable the
10611 * DTRACE_COND_USERMODE implicit predicate.
10613 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10614 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10615 ecb->dte_cond |= DTRACE_COND_USERMODE;
10618 if (dtrace_ecb_create_cache != NULL) {
10620 * If we have a cached ecb, we'll use its action list instead
10621 * of creating our own (saving both time and space).
10623 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10624 dtrace_action_t *act = cached->dte_action;
10627 ASSERT(act->dta_refcnt > 0);
10629 ecb->dte_action = act;
10630 ecb->dte_action_last = cached->dte_action_last;
10631 ecb->dte_needed = cached->dte_needed;
10632 ecb->dte_size = cached->dte_size;
10633 ecb->dte_alignment = cached->dte_alignment;
10639 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10640 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10641 dtrace_ecb_destroy(ecb);
10646 dtrace_ecb_resize(ecb);
10648 return (dtrace_ecb_create_cache = ecb);
10652 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10655 dtrace_enabling_t *enab = arg;
10656 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10658 ASSERT(state != NULL);
10660 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10662 * This probe was created in a generation for which this
10663 * enabling has previously created ECBs; we don't want to
10664 * enable it again, so just kick out.
10666 return (DTRACE_MATCH_NEXT);
10669 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10670 return (DTRACE_MATCH_DONE);
10672 dtrace_ecb_enable(ecb);
10673 return (DTRACE_MATCH_NEXT);
10676 static dtrace_ecb_t *
10677 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10681 ASSERT(MUTEX_HELD(&dtrace_lock));
10683 if (id == 0 || id > state->dts_necbs)
10686 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10687 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10689 return (state->dts_ecbs[id - 1]);
10692 static dtrace_aggregation_t *
10693 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10695 dtrace_aggregation_t *agg;
10697 ASSERT(MUTEX_HELD(&dtrace_lock));
10699 if (id == 0 || id > state->dts_naggregations)
10702 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10703 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10704 agg->dtag_id == id);
10706 return (state->dts_aggregations[id - 1]);
10710 * DTrace Buffer Functions
10712 * The following functions manipulate DTrace buffers. Most of these functions
10713 * are called in the context of establishing or processing consumer state;
10714 * exceptions are explicitly noted.
10718 * Note: called from cross call context. This function switches the two
10719 * buffers on a given CPU. The atomicity of this operation is assured by
10720 * disabling interrupts while the actual switch takes place; the disabling of
10721 * interrupts serializes the execution with any execution of dtrace_probe() on
10725 dtrace_buffer_switch(dtrace_buffer_t *buf)
10727 caddr_t tomax = buf->dtb_tomax;
10728 caddr_t xamot = buf->dtb_xamot;
10729 dtrace_icookie_t cookie;
10730 hrtime_t now = dtrace_gethrtime();
10732 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10733 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10735 cookie = dtrace_interrupt_disable();
10736 buf->dtb_tomax = xamot;
10737 buf->dtb_xamot = tomax;
10738 buf->dtb_xamot_drops = buf->dtb_drops;
10739 buf->dtb_xamot_offset = buf->dtb_offset;
10740 buf->dtb_xamot_errors = buf->dtb_errors;
10741 buf->dtb_xamot_flags = buf->dtb_flags;
10742 buf->dtb_offset = 0;
10743 buf->dtb_drops = 0;
10744 buf->dtb_errors = 0;
10745 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10746 buf->dtb_interval = now - buf->dtb_switched;
10747 buf->dtb_switched = now;
10748 dtrace_interrupt_enable(cookie);
10752 * Note: called from cross call context. This function activates a buffer
10753 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10754 * is guaranteed by the disabling of interrupts.
10757 dtrace_buffer_activate(dtrace_state_t *state)
10759 dtrace_buffer_t *buf;
10760 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10762 buf = &state->dts_buffer[curcpu];
10764 if (buf->dtb_tomax != NULL) {
10766 * We might like to assert that the buffer is marked inactive,
10767 * but this isn't necessarily true: the buffer for the CPU
10768 * that processes the BEGIN probe has its buffer activated
10769 * manually. In this case, we take the (harmless) action
10770 * re-clearing the bit INACTIVE bit.
10772 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10775 dtrace_interrupt_enable(cookie);
10779 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10785 dtrace_buffer_t *buf;
10788 ASSERT(MUTEX_HELD(&cpu_lock));
10789 ASSERT(MUTEX_HELD(&dtrace_lock));
10791 if (size > dtrace_nonroot_maxsize &&
10792 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10798 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10801 buf = &bufs[cp->cpu_id];
10804 * If there is already a buffer allocated for this CPU, it
10805 * is only possible that this is a DR event. In this case,
10807 if (buf->dtb_tomax != NULL) {
10808 ASSERT(buf->dtb_size == size);
10812 ASSERT(buf->dtb_xamot == NULL);
10814 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10817 buf->dtb_size = size;
10818 buf->dtb_flags = flags;
10819 buf->dtb_offset = 0;
10820 buf->dtb_drops = 0;
10822 if (flags & DTRACEBUF_NOSWITCH)
10825 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10827 } while ((cp = cp->cpu_next) != cpu_list);
10835 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10838 buf = &bufs[cp->cpu_id];
10840 if (buf->dtb_xamot != NULL) {
10841 ASSERT(buf->dtb_tomax != NULL);
10842 ASSERT(buf->dtb_size == size);
10843 kmem_free(buf->dtb_xamot, size);
10846 if (buf->dtb_tomax != NULL) {
10847 ASSERT(buf->dtb_size == size);
10848 kmem_free(buf->dtb_tomax, size);
10851 buf->dtb_tomax = NULL;
10852 buf->dtb_xamot = NULL;
10854 } while ((cp = cp->cpu_next) != cpu_list);
10860 #if defined(__amd64__)
10862 * FreeBSD isn't good at limiting the amount of memory we
10863 * ask to malloc, so let's place a limit here before trying
10864 * to do something that might well end in tears at bedtime.
10866 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10870 ASSERT(MUTEX_HELD(&dtrace_lock));
10872 if (cpu != DTRACE_CPUALL && cpu != i)
10878 * If there is already a buffer allocated for this CPU, it
10879 * is only possible that this is a DR event. In this case,
10880 * the buffer size must match our specified size.
10882 if (buf->dtb_tomax != NULL) {
10883 ASSERT(buf->dtb_size == size);
10887 ASSERT(buf->dtb_xamot == NULL);
10889 if ((buf->dtb_tomax = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10892 buf->dtb_size = size;
10893 buf->dtb_flags = flags;
10894 buf->dtb_offset = 0;
10895 buf->dtb_drops = 0;
10897 if (flags & DTRACEBUF_NOSWITCH)
10900 if ((buf->dtb_xamot = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
10908 * Error allocating memory, so free the buffers that were
10909 * allocated before the failed allocation.
10912 if (cpu != DTRACE_CPUALL && cpu != i)
10917 if (buf->dtb_xamot != NULL) {
10918 ASSERT(buf->dtb_tomax != NULL);
10919 ASSERT(buf->dtb_size == size);
10920 kmem_free(buf->dtb_xamot, size);
10923 if (buf->dtb_tomax != NULL) {
10924 ASSERT(buf->dtb_size == size);
10925 kmem_free(buf->dtb_tomax, size);
10928 buf->dtb_tomax = NULL;
10929 buf->dtb_xamot = NULL;
10939 * Note: called from probe context. This function just increments the drop
10940 * count on a buffer. It has been made a function to allow for the
10941 * possibility of understanding the source of mysterious drop counts. (A
10942 * problem for which one may be particularly disappointed that DTrace cannot
10943 * be used to understand DTrace.)
10946 dtrace_buffer_drop(dtrace_buffer_t *buf)
10952 * Note: called from probe context. This function is called to reserve space
10953 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
10954 * mstate. Returns the new offset in the buffer, or a negative value if an
10955 * error has occurred.
10958 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
10959 dtrace_state_t *state, dtrace_mstate_t *mstate)
10961 intptr_t offs = buf->dtb_offset, soffs;
10966 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
10969 if ((tomax = buf->dtb_tomax) == NULL) {
10970 dtrace_buffer_drop(buf);
10974 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
10975 while (offs & (align - 1)) {
10977 * Assert that our alignment is off by a number which
10978 * is itself sizeof (uint32_t) aligned.
10980 ASSERT(!((align - (offs & (align - 1))) &
10981 (sizeof (uint32_t) - 1)));
10982 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
10983 offs += sizeof (uint32_t);
10986 if ((soffs = offs + needed) > buf->dtb_size) {
10987 dtrace_buffer_drop(buf);
10991 if (mstate == NULL)
10994 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
10995 mstate->dtms_scratch_size = buf->dtb_size - soffs;
10996 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11001 if (buf->dtb_flags & DTRACEBUF_FILL) {
11002 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11003 (buf->dtb_flags & DTRACEBUF_FULL))
11008 total = needed + (offs & (align - 1));
11011 * For a ring buffer, life is quite a bit more complicated. Before
11012 * we can store any padding, we need to adjust our wrapping offset.
11013 * (If we've never before wrapped or we're not about to, no adjustment
11016 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11017 offs + total > buf->dtb_size) {
11018 woffs = buf->dtb_xamot_offset;
11020 if (offs + total > buf->dtb_size) {
11022 * We can't fit in the end of the buffer. First, a
11023 * sanity check that we can fit in the buffer at all.
11025 if (total > buf->dtb_size) {
11026 dtrace_buffer_drop(buf);
11031 * We're going to be storing at the top of the buffer,
11032 * so now we need to deal with the wrapped offset. We
11033 * only reset our wrapped offset to 0 if it is
11034 * currently greater than the current offset. If it
11035 * is less than the current offset, it is because a
11036 * previous allocation induced a wrap -- but the
11037 * allocation didn't subsequently take the space due
11038 * to an error or false predicate evaluation. In this
11039 * case, we'll just leave the wrapped offset alone: if
11040 * the wrapped offset hasn't been advanced far enough
11041 * for this allocation, it will be adjusted in the
11044 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11052 * Now we know that we're going to be storing to the
11053 * top of the buffer and that there is room for us
11054 * there. We need to clear the buffer from the current
11055 * offset to the end (there may be old gunk there).
11057 while (offs < buf->dtb_size)
11061 * We need to set our offset to zero. And because we
11062 * are wrapping, we need to set the bit indicating as
11063 * much. We can also adjust our needed space back
11064 * down to the space required by the ECB -- we know
11065 * that the top of the buffer is aligned.
11069 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11072 * There is room for us in the buffer, so we simply
11073 * need to check the wrapped offset.
11075 if (woffs < offs) {
11077 * The wrapped offset is less than the offset.
11078 * This can happen if we allocated buffer space
11079 * that induced a wrap, but then we didn't
11080 * subsequently take the space due to an error
11081 * or false predicate evaluation. This is
11082 * okay; we know that _this_ allocation isn't
11083 * going to induce a wrap. We still can't
11084 * reset the wrapped offset to be zero,
11085 * however: the space may have been trashed in
11086 * the previous failed probe attempt. But at
11087 * least the wrapped offset doesn't need to
11088 * be adjusted at all...
11094 while (offs + total > woffs) {
11095 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11098 if (epid == DTRACE_EPIDNONE) {
11099 size = sizeof (uint32_t);
11101 ASSERT(epid <= state->dts_necbs);
11102 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11104 size = state->dts_ecbs[epid - 1]->dte_size;
11107 ASSERT(woffs + size <= buf->dtb_size);
11110 if (woffs + size == buf->dtb_size) {
11112 * We've reached the end of the buffer; we want
11113 * to set the wrapped offset to 0 and break
11114 * out. However, if the offs is 0, then we're
11115 * in a strange edge-condition: the amount of
11116 * space that we want to reserve plus the size
11117 * of the record that we're overwriting is
11118 * greater than the size of the buffer. This
11119 * is problematic because if we reserve the
11120 * space but subsequently don't consume it (due
11121 * to a failed predicate or error) the wrapped
11122 * offset will be 0 -- yet the EPID at offset 0
11123 * will not be committed. This situation is
11124 * relatively easy to deal with: if we're in
11125 * this case, the buffer is indistinguishable
11126 * from one that hasn't wrapped; we need only
11127 * finish the job by clearing the wrapped bit,
11128 * explicitly setting the offset to be 0, and
11129 * zero'ing out the old data in the buffer.
11132 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11133 buf->dtb_offset = 0;
11136 while (woffs < buf->dtb_size)
11137 tomax[woffs++] = 0;
11148 * We have a wrapped offset. It may be that the wrapped offset
11149 * has become zero -- that's okay.
11151 buf->dtb_xamot_offset = woffs;
11156 * Now we can plow the buffer with any necessary padding.
11158 while (offs & (align - 1)) {
11160 * Assert that our alignment is off by a number which
11161 * is itself sizeof (uint32_t) aligned.
11163 ASSERT(!((align - (offs & (align - 1))) &
11164 (sizeof (uint32_t) - 1)));
11165 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11166 offs += sizeof (uint32_t);
11169 if (buf->dtb_flags & DTRACEBUF_FILL) {
11170 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11171 buf->dtb_flags |= DTRACEBUF_FULL;
11176 if (mstate == NULL)
11180 * For ring buffers and fill buffers, the scratch space is always
11181 * the inactive buffer.
11183 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11184 mstate->dtms_scratch_size = buf->dtb_size;
11185 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11191 dtrace_buffer_polish(dtrace_buffer_t *buf)
11193 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11194 ASSERT(MUTEX_HELD(&dtrace_lock));
11196 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11200 * We need to polish the ring buffer. There are three cases:
11202 * - The first (and presumably most common) is that there is no gap
11203 * between the buffer offset and the wrapped offset. In this case,
11204 * there is nothing in the buffer that isn't valid data; we can
11205 * mark the buffer as polished and return.
11207 * - The second (less common than the first but still more common
11208 * than the third) is that there is a gap between the buffer offset
11209 * and the wrapped offset, and the wrapped offset is larger than the
11210 * buffer offset. This can happen because of an alignment issue, or
11211 * can happen because of a call to dtrace_buffer_reserve() that
11212 * didn't subsequently consume the buffer space. In this case,
11213 * we need to zero the data from the buffer offset to the wrapped
11216 * - The third (and least common) is that there is a gap between the
11217 * buffer offset and the wrapped offset, but the wrapped offset is
11218 * _less_ than the buffer offset. This can only happen because a
11219 * call to dtrace_buffer_reserve() induced a wrap, but the space
11220 * was not subsequently consumed. In this case, we need to zero the
11221 * space from the offset to the end of the buffer _and_ from the
11222 * top of the buffer to the wrapped offset.
11224 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11225 bzero(buf->dtb_tomax + buf->dtb_offset,
11226 buf->dtb_xamot_offset - buf->dtb_offset);
11229 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11230 bzero(buf->dtb_tomax + buf->dtb_offset,
11231 buf->dtb_size - buf->dtb_offset);
11232 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11237 * This routine determines if data generated at the specified time has likely
11238 * been entirely consumed at user-level. This routine is called to determine
11239 * if an ECB on a defunct probe (but for an active enabling) can be safely
11240 * disabled and destroyed.
11243 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11247 for (i = 0; i < NCPU; i++) {
11248 dtrace_buffer_t *buf = &bufs[i];
11250 if (buf->dtb_size == 0)
11253 if (buf->dtb_flags & DTRACEBUF_RING)
11256 if (!buf->dtb_switched && buf->dtb_offset != 0)
11259 if (buf->dtb_switched - buf->dtb_interval < when)
11267 dtrace_buffer_free(dtrace_buffer_t *bufs)
11271 for (i = 0; i < NCPU; i++) {
11272 dtrace_buffer_t *buf = &bufs[i];
11274 if (buf->dtb_tomax == NULL) {
11275 ASSERT(buf->dtb_xamot == NULL);
11276 ASSERT(buf->dtb_size == 0);
11280 if (buf->dtb_xamot != NULL) {
11281 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11282 kmem_free(buf->dtb_xamot, buf->dtb_size);
11285 kmem_free(buf->dtb_tomax, buf->dtb_size);
11287 buf->dtb_tomax = NULL;
11288 buf->dtb_xamot = NULL;
11293 * DTrace Enabling Functions
11295 static dtrace_enabling_t *
11296 dtrace_enabling_create(dtrace_vstate_t *vstate)
11298 dtrace_enabling_t *enab;
11300 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11301 enab->dten_vstate = vstate;
11307 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11309 dtrace_ecbdesc_t **ndesc;
11310 size_t osize, nsize;
11313 * We can't add to enablings after we've enabled them, or after we've
11316 ASSERT(enab->dten_probegen == 0);
11317 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11319 if (enab->dten_ndesc < enab->dten_maxdesc) {
11320 enab->dten_desc[enab->dten_ndesc++] = ecb;
11324 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11326 if (enab->dten_maxdesc == 0) {
11327 enab->dten_maxdesc = 1;
11329 enab->dten_maxdesc <<= 1;
11332 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11334 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11335 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11336 bcopy(enab->dten_desc, ndesc, osize);
11337 if (enab->dten_desc != NULL)
11338 kmem_free(enab->dten_desc, osize);
11340 enab->dten_desc = ndesc;
11341 enab->dten_desc[enab->dten_ndesc++] = ecb;
11345 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11346 dtrace_probedesc_t *pd)
11348 dtrace_ecbdesc_t *new;
11349 dtrace_predicate_t *pred;
11350 dtrace_actdesc_t *act;
11353 * We're going to create a new ECB description that matches the
11354 * specified ECB in every way, but has the specified probe description.
11356 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11358 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11359 dtrace_predicate_hold(pred);
11361 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11362 dtrace_actdesc_hold(act);
11364 new->dted_action = ecb->dted_action;
11365 new->dted_pred = ecb->dted_pred;
11366 new->dted_probe = *pd;
11367 new->dted_uarg = ecb->dted_uarg;
11369 dtrace_enabling_add(enab, new);
11373 dtrace_enabling_dump(dtrace_enabling_t *enab)
11377 for (i = 0; i < enab->dten_ndesc; i++) {
11378 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11380 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11381 desc->dtpd_provider, desc->dtpd_mod,
11382 desc->dtpd_func, desc->dtpd_name);
11387 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11390 dtrace_ecbdesc_t *ep;
11391 dtrace_vstate_t *vstate = enab->dten_vstate;
11393 ASSERT(MUTEX_HELD(&dtrace_lock));
11395 for (i = 0; i < enab->dten_ndesc; i++) {
11396 dtrace_actdesc_t *act, *next;
11397 dtrace_predicate_t *pred;
11399 ep = enab->dten_desc[i];
11401 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11402 dtrace_predicate_release(pred, vstate);
11404 for (act = ep->dted_action; act != NULL; act = next) {
11405 next = act->dtad_next;
11406 dtrace_actdesc_release(act, vstate);
11409 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11412 if (enab->dten_desc != NULL)
11413 kmem_free(enab->dten_desc,
11414 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11417 * If this was a retained enabling, decrement the dts_nretained count
11418 * and take it off of the dtrace_retained list.
11420 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11421 dtrace_retained == enab) {
11422 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11423 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11424 enab->dten_vstate->dtvs_state->dts_nretained--;
11427 if (enab->dten_prev == NULL) {
11428 if (dtrace_retained == enab) {
11429 dtrace_retained = enab->dten_next;
11431 if (dtrace_retained != NULL)
11432 dtrace_retained->dten_prev = NULL;
11435 ASSERT(enab != dtrace_retained);
11436 ASSERT(dtrace_retained != NULL);
11437 enab->dten_prev->dten_next = enab->dten_next;
11440 if (enab->dten_next != NULL) {
11441 ASSERT(dtrace_retained != NULL);
11442 enab->dten_next->dten_prev = enab->dten_prev;
11445 kmem_free(enab, sizeof (dtrace_enabling_t));
11449 dtrace_enabling_retain(dtrace_enabling_t *enab)
11451 dtrace_state_t *state;
11453 ASSERT(MUTEX_HELD(&dtrace_lock));
11454 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11455 ASSERT(enab->dten_vstate != NULL);
11457 state = enab->dten_vstate->dtvs_state;
11458 ASSERT(state != NULL);
11461 * We only allow each state to retain dtrace_retain_max enablings.
11463 if (state->dts_nretained >= dtrace_retain_max)
11466 state->dts_nretained++;
11468 if (dtrace_retained == NULL) {
11469 dtrace_retained = enab;
11473 enab->dten_next = dtrace_retained;
11474 dtrace_retained->dten_prev = enab;
11475 dtrace_retained = enab;
11481 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11482 dtrace_probedesc_t *create)
11484 dtrace_enabling_t *new, *enab;
11485 int found = 0, err = ENOENT;
11487 ASSERT(MUTEX_HELD(&dtrace_lock));
11488 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11489 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11490 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11491 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11493 new = dtrace_enabling_create(&state->dts_vstate);
11496 * Iterate over all retained enablings, looking for enablings that
11497 * match the specified state.
11499 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11503 * dtvs_state can only be NULL for helper enablings -- and
11504 * helper enablings can't be retained.
11506 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11508 if (enab->dten_vstate->dtvs_state != state)
11512 * Now iterate over each probe description; we're looking for
11513 * an exact match to the specified probe description.
11515 for (i = 0; i < enab->dten_ndesc; i++) {
11516 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11517 dtrace_probedesc_t *pd = &ep->dted_probe;
11519 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11522 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11525 if (strcmp(pd->dtpd_func, match->dtpd_func))
11528 if (strcmp(pd->dtpd_name, match->dtpd_name))
11532 * We have a winning probe! Add it to our growing
11536 dtrace_enabling_addlike(new, ep, create);
11540 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11541 dtrace_enabling_destroy(new);
11549 dtrace_enabling_retract(dtrace_state_t *state)
11551 dtrace_enabling_t *enab, *next;
11553 ASSERT(MUTEX_HELD(&dtrace_lock));
11556 * Iterate over all retained enablings, destroy the enablings retained
11557 * for the specified state.
11559 for (enab = dtrace_retained; enab != NULL; enab = next) {
11560 next = enab->dten_next;
11563 * dtvs_state can only be NULL for helper enablings -- and
11564 * helper enablings can't be retained.
11566 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11568 if (enab->dten_vstate->dtvs_state == state) {
11569 ASSERT(state->dts_nretained > 0);
11570 dtrace_enabling_destroy(enab);
11574 ASSERT(state->dts_nretained == 0);
11578 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11583 ASSERT(MUTEX_HELD(&cpu_lock));
11584 ASSERT(MUTEX_HELD(&dtrace_lock));
11586 for (i = 0; i < enab->dten_ndesc; i++) {
11587 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11589 enab->dten_current = ep;
11590 enab->dten_error = 0;
11592 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11594 if (enab->dten_error != 0) {
11596 * If we get an error half-way through enabling the
11597 * probes, we kick out -- perhaps with some number of
11598 * them enabled. Leaving enabled probes enabled may
11599 * be slightly confusing for user-level, but we expect
11600 * that no one will attempt to actually drive on in
11601 * the face of such errors. If this is an anonymous
11602 * enabling (indicated with a NULL nmatched pointer),
11603 * we cmn_err() a message. We aren't expecting to
11604 * get such an error -- such as it can exist at all,
11605 * it would be a result of corrupted DOF in the driver
11608 if (nmatched == NULL) {
11609 cmn_err(CE_WARN, "dtrace_enabling_match() "
11610 "error on %p: %d", (void *)ep,
11614 return (enab->dten_error);
11618 enab->dten_probegen = dtrace_probegen;
11619 if (nmatched != NULL)
11620 *nmatched = matched;
11626 dtrace_enabling_matchall(void)
11628 dtrace_enabling_t *enab;
11630 mutex_enter(&cpu_lock);
11631 mutex_enter(&dtrace_lock);
11634 * Iterate over all retained enablings to see if any probes match
11635 * against them. We only perform this operation on enablings for which
11636 * we have sufficient permissions by virtue of being in the global zone
11637 * or in the same zone as the DTrace client. Because we can be called
11638 * after dtrace_detach() has been called, we cannot assert that there
11639 * are retained enablings. We can safely load from dtrace_retained,
11640 * however: the taskq_destroy() at the end of dtrace_detach() will
11641 * block pending our completion.
11643 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11645 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11647 if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11649 (void) dtrace_enabling_match(enab, NULL);
11652 mutex_exit(&dtrace_lock);
11653 mutex_exit(&cpu_lock);
11657 * If an enabling is to be enabled without having matched probes (that is, if
11658 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11659 * enabling must be _primed_ by creating an ECB for every ECB description.
11660 * This must be done to assure that we know the number of speculations, the
11661 * number of aggregations, the minimum buffer size needed, etc. before we
11662 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11663 * enabling any probes, we create ECBs for every ECB decription, but with a
11664 * NULL probe -- which is exactly what this function does.
11667 dtrace_enabling_prime(dtrace_state_t *state)
11669 dtrace_enabling_t *enab;
11672 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11673 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11675 if (enab->dten_vstate->dtvs_state != state)
11679 * We don't want to prime an enabling more than once, lest
11680 * we allow a malicious user to induce resource exhaustion.
11681 * (The ECBs that result from priming an enabling aren't
11682 * leaked -- but they also aren't deallocated until the
11683 * consumer state is destroyed.)
11685 if (enab->dten_primed)
11688 for (i = 0; i < enab->dten_ndesc; i++) {
11689 enab->dten_current = enab->dten_desc[i];
11690 (void) dtrace_probe_enable(NULL, enab);
11693 enab->dten_primed = 1;
11698 * Called to indicate that probes should be provided due to retained
11699 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11700 * must take an initial lap through the enabling calling the dtps_provide()
11701 * entry point explicitly to allow for autocreated probes.
11704 dtrace_enabling_provide(dtrace_provider_t *prv)
11707 dtrace_probedesc_t desc;
11709 ASSERT(MUTEX_HELD(&dtrace_lock));
11710 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11714 prv = dtrace_provider;
11718 dtrace_enabling_t *enab = dtrace_retained;
11719 void *parg = prv->dtpv_arg;
11721 for (; enab != NULL; enab = enab->dten_next) {
11722 for (i = 0; i < enab->dten_ndesc; i++) {
11723 desc = enab->dten_desc[i]->dted_probe;
11724 mutex_exit(&dtrace_lock);
11725 prv->dtpv_pops.dtps_provide(parg, &desc);
11726 mutex_enter(&dtrace_lock);
11729 } while (all && (prv = prv->dtpv_next) != NULL);
11731 mutex_exit(&dtrace_lock);
11732 dtrace_probe_provide(NULL, all ? NULL : prv);
11733 mutex_enter(&dtrace_lock);
11737 * Called to reap ECBs that are attached to probes from defunct providers.
11740 dtrace_enabling_reap(void)
11742 dtrace_provider_t *prov;
11743 dtrace_probe_t *probe;
11748 mutex_enter(&cpu_lock);
11749 mutex_enter(&dtrace_lock);
11751 for (i = 0; i < dtrace_nprobes; i++) {
11752 if ((probe = dtrace_probes[i]) == NULL)
11755 if (probe->dtpr_ecb == NULL)
11758 prov = probe->dtpr_provider;
11760 if ((when = prov->dtpv_defunct) == 0)
11764 * We have ECBs on a defunct provider: we want to reap these
11765 * ECBs to allow the provider to unregister. The destruction
11766 * of these ECBs must be done carefully: if we destroy the ECB
11767 * and the consumer later wishes to consume an EPID that
11768 * corresponds to the destroyed ECB (and if the EPID metadata
11769 * has not been previously consumed), the consumer will abort
11770 * processing on the unknown EPID. To reduce (but not, sadly,
11771 * eliminate) the possibility of this, we will only destroy an
11772 * ECB for a defunct provider if, for the state that
11773 * corresponds to the ECB:
11775 * (a) There is no speculative tracing (which can effectively
11776 * cache an EPID for an arbitrary amount of time).
11778 * (b) The principal buffers have been switched twice since the
11779 * provider became defunct.
11781 * (c) The aggregation buffers are of zero size or have been
11782 * switched twice since the provider became defunct.
11784 * We use dts_speculates to determine (a) and call a function
11785 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
11786 * that as soon as we've been unable to destroy one of the ECBs
11787 * associated with the probe, we quit trying -- reaping is only
11788 * fruitful in as much as we can destroy all ECBs associated
11789 * with the defunct provider's probes.
11791 while ((ecb = probe->dtpr_ecb) != NULL) {
11792 dtrace_state_t *state = ecb->dte_state;
11793 dtrace_buffer_t *buf = state->dts_buffer;
11794 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11796 if (state->dts_speculates)
11799 if (!dtrace_buffer_consumed(buf, when))
11802 if (!dtrace_buffer_consumed(aggbuf, when))
11805 dtrace_ecb_disable(ecb);
11806 ASSERT(probe->dtpr_ecb != ecb);
11807 dtrace_ecb_destroy(ecb);
11811 mutex_exit(&dtrace_lock);
11812 mutex_exit(&cpu_lock);
11816 * DTrace DOF Functions
11820 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11822 if (dtrace_err_verbose)
11823 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11825 #ifdef DTRACE_ERRDEBUG
11826 dtrace_errdebug(str);
11831 * Create DOF out of a currently enabled state. Right now, we only create
11832 * DOF containing the run-time options -- but this could be expanded to create
11833 * complete DOF representing the enabled state.
11836 dtrace_dof_create(dtrace_state_t *state)
11840 dof_optdesc_t *opt;
11841 int i, len = sizeof (dof_hdr_t) +
11842 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11843 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11845 ASSERT(MUTEX_HELD(&dtrace_lock));
11847 dof = kmem_zalloc(len, KM_SLEEP);
11848 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11849 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11850 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11851 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11853 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11854 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11855 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11856 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11857 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11858 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11860 dof->dofh_flags = 0;
11861 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11862 dof->dofh_secsize = sizeof (dof_sec_t);
11863 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11864 dof->dofh_secoff = sizeof (dof_hdr_t);
11865 dof->dofh_loadsz = len;
11866 dof->dofh_filesz = len;
11870 * Fill in the option section header...
11872 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11873 sec->dofs_type = DOF_SECT_OPTDESC;
11874 sec->dofs_align = sizeof (uint64_t);
11875 sec->dofs_flags = DOF_SECF_LOAD;
11876 sec->dofs_entsize = sizeof (dof_optdesc_t);
11878 opt = (dof_optdesc_t *)((uintptr_t)sec +
11879 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11881 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11882 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11884 for (i = 0; i < DTRACEOPT_MAX; i++) {
11885 opt[i].dofo_option = i;
11886 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11887 opt[i].dofo_value = state->dts_options[i];
11894 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11896 dof_hdr_t hdr, *dof;
11898 ASSERT(!MUTEX_HELD(&dtrace_lock));
11901 * First, we're going to copyin() the sizeof (dof_hdr_t).
11903 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11904 dtrace_dof_error(NULL, "failed to copyin DOF header");
11910 * Now we'll allocate the entire DOF and copy it in -- provided
11911 * that the length isn't outrageous.
11913 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
11914 dtrace_dof_error(&hdr, "load size exceeds maximum");
11919 if (hdr.dofh_loadsz < sizeof (hdr)) {
11920 dtrace_dof_error(&hdr, "invalid load size");
11925 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
11927 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
11928 kmem_free(dof, hdr.dofh_loadsz);
11937 static __inline uchar_t
11938 dtrace_dof_char(char c) {
11957 return (c - 'A' + 10);
11964 return (c - 'a' + 10);
11966 /* Should not reach here. */
11972 dtrace_dof_property(const char *name)
11976 unsigned int len, i;
11981 * Unfortunately, array of values in .conf files are always (and
11982 * only) interpreted to be integer arrays. We must read our DOF
11983 * as an integer array, and then squeeze it into a byte array.
11985 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
11986 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
11989 for (i = 0; i < len; i++)
11990 buf[i] = (uchar_t)(((int *)buf)[i]);
11992 if (len < sizeof (dof_hdr_t)) {
11993 ddi_prop_free(buf);
11994 dtrace_dof_error(NULL, "truncated header");
11998 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
11999 ddi_prop_free(buf);
12000 dtrace_dof_error(NULL, "truncated DOF");
12004 if (loadsz >= dtrace_dof_maxsize) {
12005 ddi_prop_free(buf);
12006 dtrace_dof_error(NULL, "oversized DOF");
12010 dof = kmem_alloc(loadsz, KM_SLEEP);
12011 bcopy(buf, dof, loadsz);
12012 ddi_prop_free(buf);
12017 if ((p_env = getenv(name)) == NULL)
12020 len = strlen(p_env) / 2;
12022 buf = kmem_alloc(len, KM_SLEEP);
12024 dof = (dof_hdr_t *) buf;
12028 for (i = 0; i < len; i++) {
12029 buf[i] = (dtrace_dof_char(p[0]) << 4) |
12030 dtrace_dof_char(p[1]);
12036 if (len < sizeof (dof_hdr_t)) {
12038 dtrace_dof_error(NULL, "truncated header");
12042 if (len < (loadsz = dof->dofh_loadsz)) {
12044 dtrace_dof_error(NULL, "truncated DOF");
12048 if (loadsz >= dtrace_dof_maxsize) {
12050 dtrace_dof_error(NULL, "oversized DOF");
12059 dtrace_dof_destroy(dof_hdr_t *dof)
12061 kmem_free(dof, dof->dofh_loadsz);
12065 * Return the dof_sec_t pointer corresponding to a given section index. If the
12066 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12067 * a type other than DOF_SECT_NONE is specified, the header is checked against
12068 * this type and NULL is returned if the types do not match.
12071 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12073 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12074 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12076 if (i >= dof->dofh_secnum) {
12077 dtrace_dof_error(dof, "referenced section index is invalid");
12081 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12082 dtrace_dof_error(dof, "referenced section is not loadable");
12086 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12087 dtrace_dof_error(dof, "referenced section is the wrong type");
12094 static dtrace_probedesc_t *
12095 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12097 dof_probedesc_t *probe;
12099 uintptr_t daddr = (uintptr_t)dof;
12103 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12104 dtrace_dof_error(dof, "invalid probe section");
12108 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12109 dtrace_dof_error(dof, "bad alignment in probe description");
12113 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12114 dtrace_dof_error(dof, "truncated probe description");
12118 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12119 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12121 if (strtab == NULL)
12124 str = daddr + strtab->dofs_offset;
12125 size = strtab->dofs_size;
12127 if (probe->dofp_provider >= strtab->dofs_size) {
12128 dtrace_dof_error(dof, "corrupt probe provider");
12132 (void) strncpy(desc->dtpd_provider,
12133 (char *)(str + probe->dofp_provider),
12134 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12136 if (probe->dofp_mod >= strtab->dofs_size) {
12137 dtrace_dof_error(dof, "corrupt probe module");
12141 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12142 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12144 if (probe->dofp_func >= strtab->dofs_size) {
12145 dtrace_dof_error(dof, "corrupt probe function");
12149 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12150 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12152 if (probe->dofp_name >= strtab->dofs_size) {
12153 dtrace_dof_error(dof, "corrupt probe name");
12157 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12158 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12163 static dtrace_difo_t *
12164 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12169 dof_difohdr_t *dofd;
12170 uintptr_t daddr = (uintptr_t)dof;
12171 size_t max = dtrace_difo_maxsize;
12174 static const struct {
12182 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12183 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12184 sizeof (dif_instr_t), "multiple DIF sections" },
12186 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12187 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12188 sizeof (uint64_t), "multiple integer tables" },
12190 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12191 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12192 sizeof (char), "multiple string tables" },
12194 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12195 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12196 sizeof (uint_t), "multiple variable tables" },
12198 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
12201 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12202 dtrace_dof_error(dof, "invalid DIFO header section");
12206 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12207 dtrace_dof_error(dof, "bad alignment in DIFO header");
12211 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12212 sec->dofs_size % sizeof (dof_secidx_t)) {
12213 dtrace_dof_error(dof, "bad size in DIFO header");
12217 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12218 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12220 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12221 dp->dtdo_rtype = dofd->dofd_rtype;
12223 for (l = 0; l < n; l++) {
12228 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12229 dofd->dofd_links[l])) == NULL)
12230 goto err; /* invalid section link */
12232 if (ttl + subsec->dofs_size > max) {
12233 dtrace_dof_error(dof, "exceeds maximum size");
12237 ttl += subsec->dofs_size;
12239 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12240 if (subsec->dofs_type != difo[i].section)
12243 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12244 dtrace_dof_error(dof, "section not loaded");
12248 if (subsec->dofs_align != difo[i].align) {
12249 dtrace_dof_error(dof, "bad alignment");
12253 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12254 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12256 if (*bufp != NULL) {
12257 dtrace_dof_error(dof, difo[i].msg);
12261 if (difo[i].entsize != subsec->dofs_entsize) {
12262 dtrace_dof_error(dof, "entry size mismatch");
12266 if (subsec->dofs_entsize != 0 &&
12267 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12268 dtrace_dof_error(dof, "corrupt entry size");
12272 *lenp = subsec->dofs_size;
12273 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12274 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12275 *bufp, subsec->dofs_size);
12277 if (subsec->dofs_entsize != 0)
12278 *lenp /= subsec->dofs_entsize;
12284 * If we encounter a loadable DIFO sub-section that is not
12285 * known to us, assume this is a broken program and fail.
12287 if (difo[i].section == DOF_SECT_NONE &&
12288 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12289 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12294 if (dp->dtdo_buf == NULL) {
12296 * We can't have a DIF object without DIF text.
12298 dtrace_dof_error(dof, "missing DIF text");
12303 * Before we validate the DIF object, run through the variable table
12304 * looking for the strings -- if any of their size are under, we'll set
12305 * their size to be the system-wide default string size. Note that
12306 * this should _not_ happen if the "strsize" option has been set --
12307 * in this case, the compiler should have set the size to reflect the
12308 * setting of the option.
12310 for (i = 0; i < dp->dtdo_varlen; i++) {
12311 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12312 dtrace_diftype_t *t = &v->dtdv_type;
12314 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12317 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12318 t->dtdt_size = dtrace_strsize_default;
12321 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12324 dtrace_difo_init(dp, vstate);
12328 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12329 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12330 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12331 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12333 kmem_free(dp, sizeof (dtrace_difo_t));
12337 static dtrace_predicate_t *
12338 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12343 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12346 return (dtrace_predicate_create(dp));
12349 static dtrace_actdesc_t *
12350 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12353 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12354 dof_actdesc_t *desc;
12355 dof_sec_t *difosec;
12357 uintptr_t daddr = (uintptr_t)dof;
12359 dtrace_actkind_t kind;
12361 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12362 dtrace_dof_error(dof, "invalid action section");
12366 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12367 dtrace_dof_error(dof, "truncated action description");
12371 if (sec->dofs_align != sizeof (uint64_t)) {
12372 dtrace_dof_error(dof, "bad alignment in action description");
12376 if (sec->dofs_size < sec->dofs_entsize) {
12377 dtrace_dof_error(dof, "section entry size exceeds total size");
12381 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12382 dtrace_dof_error(dof, "bad entry size in action description");
12386 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12387 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12391 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12392 desc = (dof_actdesc_t *)(daddr +
12393 (uintptr_t)sec->dofs_offset + offs);
12394 kind = (dtrace_actkind_t)desc->dofa_kind;
12396 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12397 (kind != DTRACEACT_PRINTA ||
12398 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12399 (kind == DTRACEACT_DIFEXPR &&
12400 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12406 * The argument to these actions is an index into the
12407 * DOF string table. For printf()-like actions, this
12408 * is the format string. For print(), this is the
12409 * CTF type of the expression result.
12411 if ((strtab = dtrace_dof_sect(dof,
12412 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12415 str = (char *)((uintptr_t)dof +
12416 (uintptr_t)strtab->dofs_offset);
12418 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12419 if (str[i] == '\0')
12423 if (i >= strtab->dofs_size) {
12424 dtrace_dof_error(dof, "bogus format string");
12428 if (i == desc->dofa_arg) {
12429 dtrace_dof_error(dof, "empty format string");
12433 i -= desc->dofa_arg;
12434 fmt = kmem_alloc(i + 1, KM_SLEEP);
12435 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12436 arg = (uint64_t)(uintptr_t)fmt;
12438 if (kind == DTRACEACT_PRINTA) {
12439 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12442 arg = desc->dofa_arg;
12446 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12447 desc->dofa_uarg, arg);
12449 if (last != NULL) {
12450 last->dtad_next = act;
12457 if (desc->dofa_difo == DOF_SECIDX_NONE)
12460 if ((difosec = dtrace_dof_sect(dof,
12461 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12464 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12466 if (act->dtad_difo == NULL)
12470 ASSERT(first != NULL);
12474 for (act = first; act != NULL; act = next) {
12475 next = act->dtad_next;
12476 dtrace_actdesc_release(act, vstate);
12482 static dtrace_ecbdesc_t *
12483 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12486 dtrace_ecbdesc_t *ep;
12487 dof_ecbdesc_t *ecb;
12488 dtrace_probedesc_t *desc;
12489 dtrace_predicate_t *pred = NULL;
12491 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12492 dtrace_dof_error(dof, "truncated ECB description");
12496 if (sec->dofs_align != sizeof (uint64_t)) {
12497 dtrace_dof_error(dof, "bad alignment in ECB description");
12501 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12502 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12507 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12508 ep->dted_uarg = ecb->dofe_uarg;
12509 desc = &ep->dted_probe;
12511 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12514 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12515 if ((sec = dtrace_dof_sect(dof,
12516 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12519 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12522 ep->dted_pred.dtpdd_predicate = pred;
12525 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12526 if ((sec = dtrace_dof_sect(dof,
12527 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12530 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12532 if (ep->dted_action == NULL)
12540 dtrace_predicate_release(pred, vstate);
12541 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12546 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12547 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12548 * site of any user SETX relocations to account for load object base address.
12549 * In the future, if we need other relocations, this function can be extended.
12552 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12554 uintptr_t daddr = (uintptr_t)dof;
12555 dof_relohdr_t *dofr =
12556 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12557 dof_sec_t *ss, *rs, *ts;
12561 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12562 sec->dofs_align != sizeof (dof_secidx_t)) {
12563 dtrace_dof_error(dof, "invalid relocation header");
12567 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12568 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12569 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12571 if (ss == NULL || rs == NULL || ts == NULL)
12572 return (-1); /* dtrace_dof_error() has been called already */
12574 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12575 rs->dofs_align != sizeof (uint64_t)) {
12576 dtrace_dof_error(dof, "invalid relocation section");
12580 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12581 n = rs->dofs_size / rs->dofs_entsize;
12583 for (i = 0; i < n; i++) {
12584 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12586 switch (r->dofr_type) {
12587 case DOF_RELO_NONE:
12589 case DOF_RELO_SETX:
12590 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12591 sizeof (uint64_t) > ts->dofs_size) {
12592 dtrace_dof_error(dof, "bad relocation offset");
12596 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12597 dtrace_dof_error(dof, "misaligned setx relo");
12601 *(uint64_t *)taddr += ubase;
12604 dtrace_dof_error(dof, "invalid relocation type");
12608 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12615 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12616 * header: it should be at the front of a memory region that is at least
12617 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12618 * size. It need not be validated in any other way.
12621 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12622 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12624 uint64_t len = dof->dofh_loadsz, seclen;
12625 uintptr_t daddr = (uintptr_t)dof;
12626 dtrace_ecbdesc_t *ep;
12627 dtrace_enabling_t *enab;
12630 ASSERT(MUTEX_HELD(&dtrace_lock));
12631 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12634 * Check the DOF header identification bytes. In addition to checking
12635 * valid settings, we also verify that unused bits/bytes are zeroed so
12636 * we can use them later without fear of regressing existing binaries.
12638 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12639 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12640 dtrace_dof_error(dof, "DOF magic string mismatch");
12644 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12645 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12646 dtrace_dof_error(dof, "DOF has invalid data model");
12650 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12651 dtrace_dof_error(dof, "DOF encoding mismatch");
12655 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12656 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12657 dtrace_dof_error(dof, "DOF version mismatch");
12661 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12662 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12666 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12667 dtrace_dof_error(dof, "DOF uses too many integer registers");
12671 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12672 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12676 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12677 if (dof->dofh_ident[i] != 0) {
12678 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12683 if (dof->dofh_flags & ~DOF_FL_VALID) {
12684 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12688 if (dof->dofh_secsize == 0) {
12689 dtrace_dof_error(dof, "zero section header size");
12694 * Check that the section headers don't exceed the amount of DOF
12695 * data. Note that we cast the section size and number of sections
12696 * to uint64_t's to prevent possible overflow in the multiplication.
12698 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12700 if (dof->dofh_secoff > len || seclen > len ||
12701 dof->dofh_secoff + seclen > len) {
12702 dtrace_dof_error(dof, "truncated section headers");
12706 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12707 dtrace_dof_error(dof, "misaligned section headers");
12711 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12712 dtrace_dof_error(dof, "misaligned section size");
12717 * Take an initial pass through the section headers to be sure that
12718 * the headers don't have stray offsets. If the 'noprobes' flag is
12719 * set, do not permit sections relating to providers, probes, or args.
12721 for (i = 0; i < dof->dofh_secnum; i++) {
12722 dof_sec_t *sec = (dof_sec_t *)(daddr +
12723 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12726 switch (sec->dofs_type) {
12727 case DOF_SECT_PROVIDER:
12728 case DOF_SECT_PROBES:
12729 case DOF_SECT_PRARGS:
12730 case DOF_SECT_PROFFS:
12731 dtrace_dof_error(dof, "illegal sections "
12737 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12738 continue; /* just ignore non-loadable sections */
12740 if (sec->dofs_align & (sec->dofs_align - 1)) {
12741 dtrace_dof_error(dof, "bad section alignment");
12745 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12746 dtrace_dof_error(dof, "misaligned section");
12750 if (sec->dofs_offset > len || sec->dofs_size > len ||
12751 sec->dofs_offset + sec->dofs_size > len) {
12752 dtrace_dof_error(dof, "corrupt section header");
12756 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12757 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12758 dtrace_dof_error(dof, "non-terminating string table");
12764 * Take a second pass through the sections and locate and perform any
12765 * relocations that are present. We do this after the first pass to
12766 * be sure that all sections have had their headers validated.
12768 for (i = 0; i < dof->dofh_secnum; i++) {
12769 dof_sec_t *sec = (dof_sec_t *)(daddr +
12770 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12772 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12773 continue; /* skip sections that are not loadable */
12775 switch (sec->dofs_type) {
12776 case DOF_SECT_URELHDR:
12777 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12783 if ((enab = *enabp) == NULL)
12784 enab = *enabp = dtrace_enabling_create(vstate);
12786 for (i = 0; i < dof->dofh_secnum; i++) {
12787 dof_sec_t *sec = (dof_sec_t *)(daddr +
12788 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12790 if (sec->dofs_type != DOF_SECT_ECBDESC)
12793 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12794 dtrace_enabling_destroy(enab);
12799 dtrace_enabling_add(enab, ep);
12806 * Process DOF for any options. This routine assumes that the DOF has been
12807 * at least processed by dtrace_dof_slurp().
12810 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12815 dof_optdesc_t *desc;
12817 for (i = 0; i < dof->dofh_secnum; i++) {
12818 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12819 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12821 if (sec->dofs_type != DOF_SECT_OPTDESC)
12824 if (sec->dofs_align != sizeof (uint64_t)) {
12825 dtrace_dof_error(dof, "bad alignment in "
12826 "option description");
12830 if ((entsize = sec->dofs_entsize) == 0) {
12831 dtrace_dof_error(dof, "zeroed option entry size");
12835 if (entsize < sizeof (dof_optdesc_t)) {
12836 dtrace_dof_error(dof, "bad option entry size");
12840 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12841 desc = (dof_optdesc_t *)((uintptr_t)dof +
12842 (uintptr_t)sec->dofs_offset + offs);
12844 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12845 dtrace_dof_error(dof, "non-zero option string");
12849 if (desc->dofo_value == DTRACEOPT_UNSET) {
12850 dtrace_dof_error(dof, "unset option");
12854 if ((rval = dtrace_state_option(state,
12855 desc->dofo_option, desc->dofo_value)) != 0) {
12856 dtrace_dof_error(dof, "rejected option");
12866 * DTrace Consumer State Functions
12869 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12871 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12874 dtrace_dynvar_t *dvar, *next, *start;
12877 ASSERT(MUTEX_HELD(&dtrace_lock));
12878 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12880 bzero(dstate, sizeof (dtrace_dstate_t));
12882 if ((dstate->dtds_chunksize = chunksize) == 0)
12883 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12885 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12888 if ((base = kmem_zalloc(size, KM_NOSLEEP)) == NULL)
12891 dstate->dtds_size = size;
12892 dstate->dtds_base = base;
12893 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12894 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12896 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12898 if (hashsize != 1 && (hashsize & 1))
12901 dstate->dtds_hashsize = hashsize;
12902 dstate->dtds_hash = dstate->dtds_base;
12905 * Set all of our hash buckets to point to the single sink, and (if
12906 * it hasn't already been set), set the sink's hash value to be the
12907 * sink sentinel value. The sink is needed for dynamic variable
12908 * lookups to know that they have iterated over an entire, valid hash
12911 for (i = 0; i < hashsize; i++)
12912 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
12914 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
12915 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
12918 * Determine number of active CPUs. Divide free list evenly among
12921 start = (dtrace_dynvar_t *)
12922 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
12923 limit = (uintptr_t)base + size;
12925 maxper = (limit - (uintptr_t)start) / NCPU;
12926 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
12931 for (i = 0; i < NCPU; i++) {
12933 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
12936 * If we don't even have enough chunks to make it once through
12937 * NCPUs, we're just going to allocate everything to the first
12938 * CPU. And if we're on the last CPU, we're going to allocate
12939 * whatever is left over. In either case, we set the limit to
12940 * be the limit of the dynamic variable space.
12942 if (maxper == 0 || i == NCPU - 1) {
12943 limit = (uintptr_t)base + size;
12946 limit = (uintptr_t)start + maxper;
12947 start = (dtrace_dynvar_t *)limit;
12950 ASSERT(limit <= (uintptr_t)base + size);
12953 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
12954 dstate->dtds_chunksize);
12956 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
12959 dvar->dtdv_next = next;
12971 dtrace_dstate_fini(dtrace_dstate_t *dstate)
12973 ASSERT(MUTEX_HELD(&cpu_lock));
12975 if (dstate->dtds_base == NULL)
12978 kmem_free(dstate->dtds_base, dstate->dtds_size);
12979 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
12983 dtrace_vstate_fini(dtrace_vstate_t *vstate)
12986 * Logical XOR, where are you?
12988 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
12990 if (vstate->dtvs_nglobals > 0) {
12991 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
12992 sizeof (dtrace_statvar_t *));
12995 if (vstate->dtvs_ntlocals > 0) {
12996 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
12997 sizeof (dtrace_difv_t));
13000 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13002 if (vstate->dtvs_nlocals > 0) {
13003 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13004 sizeof (dtrace_statvar_t *));
13010 dtrace_state_clean(dtrace_state_t *state)
13012 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13015 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13016 dtrace_speculation_clean(state);
13020 dtrace_state_deadman(dtrace_state_t *state)
13026 now = dtrace_gethrtime();
13028 if (state != dtrace_anon.dta_state &&
13029 now - state->dts_laststatus >= dtrace_deadman_user)
13033 * We must be sure that dts_alive never appears to be less than the
13034 * value upon entry to dtrace_state_deadman(), and because we lack a
13035 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13036 * store INT64_MAX to it, followed by a memory barrier, followed by
13037 * the new value. This assures that dts_alive never appears to be
13038 * less than its true value, regardless of the order in which the
13039 * stores to the underlying storage are issued.
13041 state->dts_alive = INT64_MAX;
13042 dtrace_membar_producer();
13043 state->dts_alive = now;
13047 dtrace_state_clean(void *arg)
13049 dtrace_state_t *state = arg;
13050 dtrace_optval_t *opt = state->dts_options;
13052 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13055 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13056 dtrace_speculation_clean(state);
13058 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13059 dtrace_state_clean, state);
13063 dtrace_state_deadman(void *arg)
13065 dtrace_state_t *state = arg;
13070 dtrace_debug_output();
13072 now = dtrace_gethrtime();
13074 if (state != dtrace_anon.dta_state &&
13075 now - state->dts_laststatus >= dtrace_deadman_user)
13079 * We must be sure that dts_alive never appears to be less than the
13080 * value upon entry to dtrace_state_deadman(), and because we lack a
13081 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13082 * store INT64_MAX to it, followed by a memory barrier, followed by
13083 * the new value. This assures that dts_alive never appears to be
13084 * less than its true value, regardless of the order in which the
13085 * stores to the underlying storage are issued.
13087 state->dts_alive = INT64_MAX;
13088 dtrace_membar_producer();
13089 state->dts_alive = now;
13091 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13092 dtrace_state_deadman, state);
13096 static dtrace_state_t *
13098 dtrace_state_create(dev_t *devp, cred_t *cr)
13100 dtrace_state_create(struct cdev *dev)
13111 dtrace_state_t *state;
13112 dtrace_optval_t *opt;
13113 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13115 ASSERT(MUTEX_HELD(&dtrace_lock));
13116 ASSERT(MUTEX_HELD(&cpu_lock));
13119 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13120 VM_BESTFIT | VM_SLEEP);
13122 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13123 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13127 state = ddi_get_soft_state(dtrace_softstate, minor);
13134 /* Allocate memory for the state. */
13135 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
13138 state->dts_epid = DTRACE_EPIDNONE + 1;
13140 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
13142 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13143 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13145 if (devp != NULL) {
13146 major = getemajor(*devp);
13148 major = ddi_driver_major(dtrace_devi);
13151 state->dts_dev = makedevice(major, minor);
13154 *devp = state->dts_dev;
13156 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
13157 state->dts_dev = dev;
13161 * We allocate NCPU buffers. On the one hand, this can be quite
13162 * a bit of memory per instance (nearly 36K on a Starcat). On the
13163 * other hand, it saves an additional memory reference in the probe
13166 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13167 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13170 state->dts_cleaner = CYCLIC_NONE;
13171 state->dts_deadman = CYCLIC_NONE;
13173 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
13174 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
13176 state->dts_vstate.dtvs_state = state;
13178 for (i = 0; i < DTRACEOPT_MAX; i++)
13179 state->dts_options[i] = DTRACEOPT_UNSET;
13182 * Set the default options.
13184 opt = state->dts_options;
13185 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13186 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13187 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13188 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13189 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13190 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13191 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13192 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13193 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13194 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13195 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13196 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13197 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13198 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13200 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13203 * Depending on the user credentials, we set flag bits which alter probe
13204 * visibility or the amount of destructiveness allowed. In the case of
13205 * actual anonymous tracing, or the possession of all privileges, all of
13206 * the normal checks are bypassed.
13208 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13209 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13210 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13213 * Set up the credentials for this instantiation. We take a
13214 * hold on the credential to prevent it from disappearing on
13215 * us; this in turn prevents the zone_t referenced by this
13216 * credential from disappearing. This means that we can
13217 * examine the credential and the zone from probe context.
13220 state->dts_cred.dcr_cred = cr;
13223 * CRA_PROC means "we have *some* privilege for dtrace" and
13224 * unlocks the use of variables like pid, zonename, etc.
13226 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13227 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13228 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13232 * dtrace_user allows use of syscall and profile providers.
13233 * If the user also has proc_owner and/or proc_zone, we
13234 * extend the scope to include additional visibility and
13235 * destructive power.
13237 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13238 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13239 state->dts_cred.dcr_visible |=
13240 DTRACE_CRV_ALLPROC;
13242 state->dts_cred.dcr_action |=
13243 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13246 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13247 state->dts_cred.dcr_visible |=
13248 DTRACE_CRV_ALLZONE;
13250 state->dts_cred.dcr_action |=
13251 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13255 * If we have all privs in whatever zone this is,
13256 * we can do destructive things to processes which
13257 * have altered credentials.
13260 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13261 cr->cr_zone->zone_privset)) {
13262 state->dts_cred.dcr_action |=
13263 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13269 * Holding the dtrace_kernel privilege also implies that
13270 * the user has the dtrace_user privilege from a visibility
13271 * perspective. But without further privileges, some
13272 * destructive actions are not available.
13274 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13276 * Make all probes in all zones visible. However,
13277 * this doesn't mean that all actions become available
13280 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13281 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13283 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13286 * Holding proc_owner means that destructive actions
13287 * for *this* zone are allowed.
13289 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13290 state->dts_cred.dcr_action |=
13291 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13294 * Holding proc_zone means that destructive actions
13295 * for this user/group ID in all zones is allowed.
13297 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13298 state->dts_cred.dcr_action |=
13299 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13303 * If we have all privs in whatever zone this is,
13304 * we can do destructive things to processes which
13305 * have altered credentials.
13307 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13308 cr->cr_zone->zone_privset)) {
13309 state->dts_cred.dcr_action |=
13310 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13316 * Holding the dtrace_proc privilege gives control over fasttrap
13317 * and pid providers. We need to grant wider destructive
13318 * privileges in the event that the user has proc_owner and/or
13321 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13322 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13323 state->dts_cred.dcr_action |=
13324 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13326 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13327 state->dts_cred.dcr_action |=
13328 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13336 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13338 dtrace_optval_t *opt = state->dts_options, size;
13339 processorid_t cpu = 0;;
13340 int flags = 0, rval;
13342 ASSERT(MUTEX_HELD(&dtrace_lock));
13343 ASSERT(MUTEX_HELD(&cpu_lock));
13344 ASSERT(which < DTRACEOPT_MAX);
13345 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13346 (state == dtrace_anon.dta_state &&
13347 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13349 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13352 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13353 cpu = opt[DTRACEOPT_CPU];
13355 if (which == DTRACEOPT_SPECSIZE)
13356 flags |= DTRACEBUF_NOSWITCH;
13358 if (which == DTRACEOPT_BUFSIZE) {
13359 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13360 flags |= DTRACEBUF_RING;
13362 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13363 flags |= DTRACEBUF_FILL;
13365 if (state != dtrace_anon.dta_state ||
13366 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13367 flags |= DTRACEBUF_INACTIVE;
13370 for (size = opt[which]; size >= sizeof (uint64_t); size >>= 1) {
13372 * The size must be 8-byte aligned. If the size is not 8-byte
13373 * aligned, drop it down by the difference.
13375 if (size & (sizeof (uint64_t) - 1))
13376 size -= size & (sizeof (uint64_t) - 1);
13378 if (size < state->dts_reserve) {
13380 * Buffers always must be large enough to accommodate
13381 * their prereserved space. We return E2BIG instead
13382 * of ENOMEM in this case to allow for user-level
13383 * software to differentiate the cases.
13388 rval = dtrace_buffer_alloc(buf, size, flags, cpu);
13390 if (rval != ENOMEM) {
13395 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13403 dtrace_state_buffers(dtrace_state_t *state)
13405 dtrace_speculation_t *spec = state->dts_speculations;
13408 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13409 DTRACEOPT_BUFSIZE)) != 0)
13412 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13413 DTRACEOPT_AGGSIZE)) != 0)
13416 for (i = 0; i < state->dts_nspeculations; i++) {
13417 if ((rval = dtrace_state_buffer(state,
13418 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13426 dtrace_state_prereserve(dtrace_state_t *state)
13429 dtrace_probe_t *probe;
13431 state->dts_reserve = 0;
13433 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13437 * If our buffer policy is a "fill" buffer policy, we need to set the
13438 * prereserved space to be the space required by the END probes.
13440 probe = dtrace_probes[dtrace_probeid_end - 1];
13441 ASSERT(probe != NULL);
13443 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13444 if (ecb->dte_state != state)
13447 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13452 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13454 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13455 dtrace_speculation_t *spec;
13456 dtrace_buffer_t *buf;
13458 cyc_handler_t hdlr;
13461 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13462 dtrace_icookie_t cookie;
13464 mutex_enter(&cpu_lock);
13465 mutex_enter(&dtrace_lock);
13467 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13473 * Before we can perform any checks, we must prime all of the
13474 * retained enablings that correspond to this state.
13476 dtrace_enabling_prime(state);
13478 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13483 dtrace_state_prereserve(state);
13486 * Now we want to do is try to allocate our speculations.
13487 * We do not automatically resize the number of speculations; if
13488 * this fails, we will fail the operation.
13490 nspec = opt[DTRACEOPT_NSPEC];
13491 ASSERT(nspec != DTRACEOPT_UNSET);
13493 if (nspec > INT_MAX) {
13498 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t), KM_NOSLEEP);
13500 if (spec == NULL) {
13505 state->dts_speculations = spec;
13506 state->dts_nspeculations = (int)nspec;
13508 for (i = 0; i < nspec; i++) {
13509 if ((buf = kmem_zalloc(bufsize, KM_NOSLEEP)) == NULL) {
13514 spec[i].dtsp_buffer = buf;
13517 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13518 if (dtrace_anon.dta_state == NULL) {
13523 if (state->dts_necbs != 0) {
13528 state->dts_anon = dtrace_anon_grab();
13529 ASSERT(state->dts_anon != NULL);
13530 state = state->dts_anon;
13533 * We want "grabanon" to be set in the grabbed state, so we'll
13534 * copy that option value from the grabbing state into the
13537 state->dts_options[DTRACEOPT_GRABANON] =
13538 opt[DTRACEOPT_GRABANON];
13540 *cpu = dtrace_anon.dta_beganon;
13543 * If the anonymous state is active (as it almost certainly
13544 * is if the anonymous enabling ultimately matched anything),
13545 * we don't allow any further option processing -- but we
13546 * don't return failure.
13548 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13552 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13553 opt[DTRACEOPT_AGGSIZE] != 0) {
13554 if (state->dts_aggregations == NULL) {
13556 * We're not going to create an aggregation buffer
13557 * because we don't have any ECBs that contain
13558 * aggregations -- set this option to 0.
13560 opt[DTRACEOPT_AGGSIZE] = 0;
13563 * If we have an aggregation buffer, we must also have
13564 * a buffer to use as scratch.
13566 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13567 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13568 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13573 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13574 opt[DTRACEOPT_SPECSIZE] != 0) {
13575 if (!state->dts_speculates) {
13577 * We're not going to create speculation buffers
13578 * because we don't have any ECBs that actually
13579 * speculate -- set the speculation size to 0.
13581 opt[DTRACEOPT_SPECSIZE] = 0;
13586 * The bare minimum size for any buffer that we're actually going to
13587 * do anything to is sizeof (uint64_t).
13589 sz = sizeof (uint64_t);
13591 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13592 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13593 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13595 * A buffer size has been explicitly set to 0 (or to a size
13596 * that will be adjusted to 0) and we need the space -- we
13597 * need to return failure. We return ENOSPC to differentiate
13598 * it from failing to allocate a buffer due to failure to meet
13599 * the reserve (for which we return E2BIG).
13605 if ((rval = dtrace_state_buffers(state)) != 0)
13608 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13609 sz = dtrace_dstate_defsize;
13612 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13617 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13619 } while (sz >>= 1);
13621 opt[DTRACEOPT_DYNVARSIZE] = sz;
13626 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13627 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13629 if (opt[DTRACEOPT_CLEANRATE] == 0)
13630 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13632 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13633 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13635 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13636 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13638 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13640 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13641 hdlr.cyh_arg = state;
13642 hdlr.cyh_level = CY_LOW_LEVEL;
13645 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13647 state->dts_cleaner = cyclic_add(&hdlr, &when);
13649 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13650 hdlr.cyh_arg = state;
13651 hdlr.cyh_level = CY_LOW_LEVEL;
13654 when.cyt_interval = dtrace_deadman_interval;
13656 state->dts_deadman = cyclic_add(&hdlr, &when);
13658 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13659 dtrace_state_clean, state);
13660 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13661 dtrace_state_deadman, state);
13664 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13667 * Now it's time to actually fire the BEGIN probe. We need to disable
13668 * interrupts here both to record the CPU on which we fired the BEGIN
13669 * probe (the data from this CPU will be processed first at user
13670 * level) and to manually activate the buffer for this CPU.
13672 cookie = dtrace_interrupt_disable();
13674 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13675 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13677 dtrace_probe(dtrace_probeid_begin,
13678 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13679 dtrace_interrupt_enable(cookie);
13681 * We may have had an exit action from a BEGIN probe; only change our
13682 * state to ACTIVE if we're still in WARMUP.
13684 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13685 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13687 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13688 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13691 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13692 * want each CPU to transition its principal buffer out of the
13693 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13694 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13695 * atomically transition from processing none of a state's ECBs to
13696 * processing all of them.
13698 dtrace_xcall(DTRACE_CPUALL,
13699 (dtrace_xcall_t)dtrace_buffer_activate, state);
13703 dtrace_buffer_free(state->dts_buffer);
13704 dtrace_buffer_free(state->dts_aggbuffer);
13706 if ((nspec = state->dts_nspeculations) == 0) {
13707 ASSERT(state->dts_speculations == NULL);
13711 spec = state->dts_speculations;
13712 ASSERT(spec != NULL);
13714 for (i = 0; i < state->dts_nspeculations; i++) {
13715 if ((buf = spec[i].dtsp_buffer) == NULL)
13718 dtrace_buffer_free(buf);
13719 kmem_free(buf, bufsize);
13722 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13723 state->dts_nspeculations = 0;
13724 state->dts_speculations = NULL;
13727 mutex_exit(&dtrace_lock);
13728 mutex_exit(&cpu_lock);
13734 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13736 dtrace_icookie_t cookie;
13738 ASSERT(MUTEX_HELD(&dtrace_lock));
13740 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13741 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13745 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13746 * to be sure that every CPU has seen it. See below for the details
13747 * on why this is done.
13749 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13753 * By this point, it is impossible for any CPU to be still processing
13754 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13755 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13756 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13757 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13758 * iff we're in the END probe.
13760 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13762 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13765 * Finally, we can release the reserve and call the END probe. We
13766 * disable interrupts across calling the END probe to allow us to
13767 * return the CPU on which we actually called the END probe. This
13768 * allows user-land to be sure that this CPU's principal buffer is
13771 state->dts_reserve = 0;
13773 cookie = dtrace_interrupt_disable();
13775 dtrace_probe(dtrace_probeid_end,
13776 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13777 dtrace_interrupt_enable(cookie);
13779 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13786 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13787 dtrace_optval_t val)
13789 ASSERT(MUTEX_HELD(&dtrace_lock));
13791 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13794 if (option >= DTRACEOPT_MAX)
13797 if (option != DTRACEOPT_CPU && val < 0)
13801 case DTRACEOPT_DESTRUCTIVE:
13802 if (dtrace_destructive_disallow)
13805 state->dts_cred.dcr_destructive = 1;
13808 case DTRACEOPT_BUFSIZE:
13809 case DTRACEOPT_DYNVARSIZE:
13810 case DTRACEOPT_AGGSIZE:
13811 case DTRACEOPT_SPECSIZE:
13812 case DTRACEOPT_STRSIZE:
13816 if (val >= LONG_MAX) {
13818 * If this is an otherwise negative value, set it to
13819 * the highest multiple of 128m less than LONG_MAX.
13820 * Technically, we're adjusting the size without
13821 * regard to the buffer resizing policy, but in fact,
13822 * this has no effect -- if we set the buffer size to
13823 * ~LONG_MAX and the buffer policy is ultimately set to
13824 * be "manual", the buffer allocation is guaranteed to
13825 * fail, if only because the allocation requires two
13826 * buffers. (We set the the size to the highest
13827 * multiple of 128m because it ensures that the size
13828 * will remain a multiple of a megabyte when
13829 * repeatedly halved -- all the way down to 15m.)
13831 val = LONG_MAX - (1 << 27) + 1;
13835 state->dts_options[option] = val;
13841 dtrace_state_destroy(dtrace_state_t *state)
13844 dtrace_vstate_t *vstate = &state->dts_vstate;
13846 minor_t minor = getminor(state->dts_dev);
13848 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13849 dtrace_speculation_t *spec = state->dts_speculations;
13850 int nspec = state->dts_nspeculations;
13853 ASSERT(MUTEX_HELD(&dtrace_lock));
13854 ASSERT(MUTEX_HELD(&cpu_lock));
13857 * First, retract any retained enablings for this state.
13859 dtrace_enabling_retract(state);
13860 ASSERT(state->dts_nretained == 0);
13862 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13863 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13865 * We have managed to come into dtrace_state_destroy() on a
13866 * hot enabling -- almost certainly because of a disorderly
13867 * shutdown of a consumer. (That is, a consumer that is
13868 * exiting without having called dtrace_stop().) In this case,
13869 * we're going to set our activity to be KILLED, and then
13870 * issue a sync to be sure that everyone is out of probe
13871 * context before we start blowing away ECBs.
13873 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13878 * Release the credential hold we took in dtrace_state_create().
13880 if (state->dts_cred.dcr_cred != NULL)
13881 crfree(state->dts_cred.dcr_cred);
13884 * Now we can safely disable and destroy any enabled probes. Because
13885 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13886 * (especially if they're all enabled), we take two passes through the
13887 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13888 * in the second we disable whatever is left over.
13890 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13891 for (i = 0; i < state->dts_necbs; i++) {
13892 if ((ecb = state->dts_ecbs[i]) == NULL)
13895 if (match && ecb->dte_probe != NULL) {
13896 dtrace_probe_t *probe = ecb->dte_probe;
13897 dtrace_provider_t *prov = probe->dtpr_provider;
13899 if (!(prov->dtpv_priv.dtpp_flags & match))
13903 dtrace_ecb_disable(ecb);
13904 dtrace_ecb_destroy(ecb);
13912 * Before we free the buffers, perform one more sync to assure that
13913 * every CPU is out of probe context.
13917 dtrace_buffer_free(state->dts_buffer);
13918 dtrace_buffer_free(state->dts_aggbuffer);
13920 for (i = 0; i < nspec; i++)
13921 dtrace_buffer_free(spec[i].dtsp_buffer);
13924 if (state->dts_cleaner != CYCLIC_NONE)
13925 cyclic_remove(state->dts_cleaner);
13927 if (state->dts_deadman != CYCLIC_NONE)
13928 cyclic_remove(state->dts_deadman);
13930 callout_stop(&state->dts_cleaner);
13931 callout_drain(&state->dts_cleaner);
13932 callout_stop(&state->dts_deadman);
13933 callout_drain(&state->dts_deadman);
13936 dtrace_dstate_fini(&vstate->dtvs_dynvars);
13937 dtrace_vstate_fini(vstate);
13938 if (state->dts_ecbs != NULL)
13939 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
13941 if (state->dts_aggregations != NULL) {
13943 for (i = 0; i < state->dts_naggregations; i++)
13944 ASSERT(state->dts_aggregations[i] == NULL);
13946 ASSERT(state->dts_naggregations > 0);
13947 kmem_free(state->dts_aggregations,
13948 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
13951 kmem_free(state->dts_buffer, bufsize);
13952 kmem_free(state->dts_aggbuffer, bufsize);
13954 for (i = 0; i < nspec; i++)
13955 kmem_free(spec[i].dtsp_buffer, bufsize);
13958 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13960 dtrace_format_destroy(state);
13962 if (state->dts_aggid_arena != NULL) {
13964 vmem_destroy(state->dts_aggid_arena);
13966 delete_unrhdr(state->dts_aggid_arena);
13968 state->dts_aggid_arena = NULL;
13971 ddi_soft_state_free(dtrace_softstate, minor);
13972 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13977 * DTrace Anonymous Enabling Functions
13979 static dtrace_state_t *
13980 dtrace_anon_grab(void)
13982 dtrace_state_t *state;
13984 ASSERT(MUTEX_HELD(&dtrace_lock));
13986 if ((state = dtrace_anon.dta_state) == NULL) {
13987 ASSERT(dtrace_anon.dta_enabling == NULL);
13991 ASSERT(dtrace_anon.dta_enabling != NULL);
13992 ASSERT(dtrace_retained != NULL);
13994 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
13995 dtrace_anon.dta_enabling = NULL;
13996 dtrace_anon.dta_state = NULL;
14002 dtrace_anon_property(void)
14005 dtrace_state_t *state;
14007 char c[32]; /* enough for "dof-data-" + digits */
14009 ASSERT(MUTEX_HELD(&dtrace_lock));
14010 ASSERT(MUTEX_HELD(&cpu_lock));
14012 for (i = 0; ; i++) {
14013 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14015 dtrace_err_verbose = 1;
14017 if ((dof = dtrace_dof_property(c)) == NULL) {
14018 dtrace_err_verbose = 0;
14024 * We want to create anonymous state, so we need to transition
14025 * the kernel debugger to indicate that DTrace is active. If
14026 * this fails (e.g. because the debugger has modified text in
14027 * some way), we won't continue with the processing.
14029 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14030 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14031 "enabling ignored.");
14032 dtrace_dof_destroy(dof);
14038 * If we haven't allocated an anonymous state, we'll do so now.
14040 if ((state = dtrace_anon.dta_state) == NULL) {
14042 state = dtrace_state_create(NULL, NULL);
14044 state = dtrace_state_create(NULL);
14046 dtrace_anon.dta_state = state;
14048 if (state == NULL) {
14050 * This basically shouldn't happen: the only
14051 * failure mode from dtrace_state_create() is a
14052 * failure of ddi_soft_state_zalloc() that
14053 * itself should never happen. Still, the
14054 * interface allows for a failure mode, and
14055 * we want to fail as gracefully as possible:
14056 * we'll emit an error message and cease
14057 * processing anonymous state in this case.
14059 cmn_err(CE_WARN, "failed to create "
14060 "anonymous state");
14061 dtrace_dof_destroy(dof);
14066 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14067 &dtrace_anon.dta_enabling, 0, B_TRUE);
14070 rv = dtrace_dof_options(dof, state);
14072 dtrace_err_verbose = 0;
14073 dtrace_dof_destroy(dof);
14077 * This is malformed DOF; chuck any anonymous state
14080 ASSERT(dtrace_anon.dta_enabling == NULL);
14081 dtrace_state_destroy(state);
14082 dtrace_anon.dta_state = NULL;
14086 ASSERT(dtrace_anon.dta_enabling != NULL);
14089 if (dtrace_anon.dta_enabling != NULL) {
14093 * dtrace_enabling_retain() can only fail because we are
14094 * trying to retain more enablings than are allowed -- but
14095 * we only have one anonymous enabling, and we are guaranteed
14096 * to be allowed at least one retained enabling; we assert
14097 * that dtrace_enabling_retain() returns success.
14099 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14102 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14107 * DTrace Helper Functions
14110 dtrace_helper_trace(dtrace_helper_action_t *helper,
14111 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14113 uint32_t size, next, nnext, i;
14114 dtrace_helptrace_t *ent;
14115 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
14117 if (!dtrace_helptrace_enabled)
14120 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14123 * What would a tracing framework be without its own tracing
14124 * framework? (Well, a hell of a lot simpler, for starters...)
14126 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14127 sizeof (uint64_t) - sizeof (uint64_t);
14130 * Iterate until we can allocate a slot in the trace buffer.
14133 next = dtrace_helptrace_next;
14135 if (next + size < dtrace_helptrace_bufsize) {
14136 nnext = next + size;
14140 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14143 * We have our slot; fill it in.
14148 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14149 ent->dtht_helper = helper;
14150 ent->dtht_where = where;
14151 ent->dtht_nlocals = vstate->dtvs_nlocals;
14153 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14154 mstate->dtms_fltoffs : -1;
14155 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14156 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
14158 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14159 dtrace_statvar_t *svar;
14161 if ((svar = vstate->dtvs_locals[i]) == NULL)
14164 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14165 ent->dtht_locals[i] =
14166 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
14171 dtrace_helper(int which, dtrace_mstate_t *mstate,
14172 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14174 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
14175 uint64_t sarg0 = mstate->dtms_arg[0];
14176 uint64_t sarg1 = mstate->dtms_arg[1];
14178 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14179 dtrace_helper_action_t *helper;
14180 dtrace_vstate_t *vstate;
14181 dtrace_difo_t *pred;
14182 int i, trace = dtrace_helptrace_enabled;
14184 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14186 if (helpers == NULL)
14189 if ((helper = helpers->dthps_actions[which]) == NULL)
14192 vstate = &helpers->dthps_vstate;
14193 mstate->dtms_arg[0] = arg0;
14194 mstate->dtms_arg[1] = arg1;
14197 * Now iterate over each helper. If its predicate evaluates to 'true',
14198 * we'll call the corresponding actions. Note that the below calls
14199 * to dtrace_dif_emulate() may set faults in machine state. This is
14200 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14201 * the stored DIF offset with its own (which is the desired behavior).
14202 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14203 * from machine state; this is okay, too.
14205 for (; helper != NULL; helper = helper->dtha_next) {
14206 if ((pred = helper->dtha_predicate) != NULL) {
14208 dtrace_helper_trace(helper, mstate, vstate, 0);
14210 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14213 if (*flags & CPU_DTRACE_FAULT)
14217 for (i = 0; i < helper->dtha_nactions; i++) {
14219 dtrace_helper_trace(helper,
14220 mstate, vstate, i + 1);
14222 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14223 mstate, vstate, state);
14225 if (*flags & CPU_DTRACE_FAULT)
14231 dtrace_helper_trace(helper, mstate, vstate,
14232 DTRACE_HELPTRACE_NEXT);
14236 dtrace_helper_trace(helper, mstate, vstate,
14237 DTRACE_HELPTRACE_DONE);
14240 * Restore the arg0 that we saved upon entry.
14242 mstate->dtms_arg[0] = sarg0;
14243 mstate->dtms_arg[1] = sarg1;
14249 dtrace_helper_trace(helper, mstate, vstate,
14250 DTRACE_HELPTRACE_ERR);
14253 * Restore the arg0 that we saved upon entry.
14255 mstate->dtms_arg[0] = sarg0;
14256 mstate->dtms_arg[1] = sarg1;
14262 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14263 dtrace_vstate_t *vstate)
14267 if (helper->dtha_predicate != NULL)
14268 dtrace_difo_release(helper->dtha_predicate, vstate);
14270 for (i = 0; i < helper->dtha_nactions; i++) {
14271 ASSERT(helper->dtha_actions[i] != NULL);
14272 dtrace_difo_release(helper->dtha_actions[i], vstate);
14275 kmem_free(helper->dtha_actions,
14276 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14277 kmem_free(helper, sizeof (dtrace_helper_action_t));
14281 dtrace_helper_destroygen(int gen)
14283 proc_t *p = curproc;
14284 dtrace_helpers_t *help = p->p_dtrace_helpers;
14285 dtrace_vstate_t *vstate;
14288 ASSERT(MUTEX_HELD(&dtrace_lock));
14290 if (help == NULL || gen > help->dthps_generation)
14293 vstate = &help->dthps_vstate;
14295 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14296 dtrace_helper_action_t *last = NULL, *h, *next;
14298 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14299 next = h->dtha_next;
14301 if (h->dtha_generation == gen) {
14302 if (last != NULL) {
14303 last->dtha_next = next;
14305 help->dthps_actions[i] = next;
14308 dtrace_helper_action_destroy(h, vstate);
14316 * Interate until we've cleared out all helper providers with the
14317 * given generation number.
14320 dtrace_helper_provider_t *prov;
14323 * Look for a helper provider with the right generation. We
14324 * have to start back at the beginning of the list each time
14325 * because we drop dtrace_lock. It's unlikely that we'll make
14326 * more than two passes.
14328 for (i = 0; i < help->dthps_nprovs; i++) {
14329 prov = help->dthps_provs[i];
14331 if (prov->dthp_generation == gen)
14336 * If there were no matches, we're done.
14338 if (i == help->dthps_nprovs)
14342 * Move the last helper provider into this slot.
14344 help->dthps_nprovs--;
14345 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14346 help->dthps_provs[help->dthps_nprovs] = NULL;
14348 mutex_exit(&dtrace_lock);
14351 * If we have a meta provider, remove this helper provider.
14353 mutex_enter(&dtrace_meta_lock);
14354 if (dtrace_meta_pid != NULL) {
14355 ASSERT(dtrace_deferred_pid == NULL);
14356 dtrace_helper_provider_remove(&prov->dthp_prov,
14359 mutex_exit(&dtrace_meta_lock);
14361 dtrace_helper_provider_destroy(prov);
14363 mutex_enter(&dtrace_lock);
14370 dtrace_helper_validate(dtrace_helper_action_t *helper)
14375 if ((dp = helper->dtha_predicate) != NULL)
14376 err += dtrace_difo_validate_helper(dp);
14378 for (i = 0; i < helper->dtha_nactions; i++)
14379 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14385 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14387 dtrace_helpers_t *help;
14388 dtrace_helper_action_t *helper, *last;
14389 dtrace_actdesc_t *act;
14390 dtrace_vstate_t *vstate;
14391 dtrace_predicate_t *pred;
14392 int count = 0, nactions = 0, i;
14394 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14397 help = curproc->p_dtrace_helpers;
14398 last = help->dthps_actions[which];
14399 vstate = &help->dthps_vstate;
14401 for (count = 0; last != NULL; last = last->dtha_next) {
14403 if (last->dtha_next == NULL)
14408 * If we already have dtrace_helper_actions_max helper actions for this
14409 * helper action type, we'll refuse to add a new one.
14411 if (count >= dtrace_helper_actions_max)
14414 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14415 helper->dtha_generation = help->dthps_generation;
14417 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14418 ASSERT(pred->dtp_difo != NULL);
14419 dtrace_difo_hold(pred->dtp_difo);
14420 helper->dtha_predicate = pred->dtp_difo;
14423 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14424 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14427 if (act->dtad_difo == NULL)
14433 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14434 (helper->dtha_nactions = nactions), KM_SLEEP);
14436 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14437 dtrace_difo_hold(act->dtad_difo);
14438 helper->dtha_actions[i++] = act->dtad_difo;
14441 if (!dtrace_helper_validate(helper))
14444 if (last == NULL) {
14445 help->dthps_actions[which] = helper;
14447 last->dtha_next = helper;
14450 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14451 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14452 dtrace_helptrace_next = 0;
14457 dtrace_helper_action_destroy(helper, vstate);
14462 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14463 dof_helper_t *dofhp)
14465 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14467 mutex_enter(&dtrace_meta_lock);
14468 mutex_enter(&dtrace_lock);
14470 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14472 * If the dtrace module is loaded but not attached, or if
14473 * there aren't isn't a meta provider registered to deal with
14474 * these provider descriptions, we need to postpone creating
14475 * the actual providers until later.
14478 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14479 dtrace_deferred_pid != help) {
14480 help->dthps_deferred = 1;
14481 help->dthps_pid = p->p_pid;
14482 help->dthps_next = dtrace_deferred_pid;
14483 help->dthps_prev = NULL;
14484 if (dtrace_deferred_pid != NULL)
14485 dtrace_deferred_pid->dthps_prev = help;
14486 dtrace_deferred_pid = help;
14489 mutex_exit(&dtrace_lock);
14491 } else if (dofhp != NULL) {
14493 * If the dtrace module is loaded and we have a particular
14494 * helper provider description, pass that off to the
14498 mutex_exit(&dtrace_lock);
14500 dtrace_helper_provide(dofhp, p->p_pid);
14504 * Otherwise, just pass all the helper provider descriptions
14505 * off to the meta provider.
14509 mutex_exit(&dtrace_lock);
14511 for (i = 0; i < help->dthps_nprovs; i++) {
14512 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14517 mutex_exit(&dtrace_meta_lock);
14521 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14523 dtrace_helpers_t *help;
14524 dtrace_helper_provider_t *hprov, **tmp_provs;
14525 uint_t tmp_maxprovs, i;
14527 ASSERT(MUTEX_HELD(&dtrace_lock));
14529 help = curproc->p_dtrace_helpers;
14530 ASSERT(help != NULL);
14533 * If we already have dtrace_helper_providers_max helper providers,
14534 * we're refuse to add a new one.
14536 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14540 * Check to make sure this isn't a duplicate.
14542 for (i = 0; i < help->dthps_nprovs; i++) {
14543 if (dofhp->dofhp_addr ==
14544 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14548 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14549 hprov->dthp_prov = *dofhp;
14550 hprov->dthp_ref = 1;
14551 hprov->dthp_generation = gen;
14554 * Allocate a bigger table for helper providers if it's already full.
14556 if (help->dthps_maxprovs == help->dthps_nprovs) {
14557 tmp_maxprovs = help->dthps_maxprovs;
14558 tmp_provs = help->dthps_provs;
14560 if (help->dthps_maxprovs == 0)
14561 help->dthps_maxprovs = 2;
14563 help->dthps_maxprovs *= 2;
14564 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14565 help->dthps_maxprovs = dtrace_helper_providers_max;
14567 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14569 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14570 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14572 if (tmp_provs != NULL) {
14573 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14574 sizeof (dtrace_helper_provider_t *));
14575 kmem_free(tmp_provs, tmp_maxprovs *
14576 sizeof (dtrace_helper_provider_t *));
14580 help->dthps_provs[help->dthps_nprovs] = hprov;
14581 help->dthps_nprovs++;
14587 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14589 mutex_enter(&dtrace_lock);
14591 if (--hprov->dthp_ref == 0) {
14593 mutex_exit(&dtrace_lock);
14594 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14595 dtrace_dof_destroy(dof);
14596 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14598 mutex_exit(&dtrace_lock);
14603 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14605 uintptr_t daddr = (uintptr_t)dof;
14606 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14607 dof_provider_t *provider;
14608 dof_probe_t *probe;
14610 char *strtab, *typestr;
14611 dof_stridx_t typeidx;
14613 uint_t nprobes, j, k;
14615 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14617 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14618 dtrace_dof_error(dof, "misaligned section offset");
14623 * The section needs to be large enough to contain the DOF provider
14624 * structure appropriate for the given version.
14626 if (sec->dofs_size <
14627 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14628 offsetof(dof_provider_t, dofpv_prenoffs) :
14629 sizeof (dof_provider_t))) {
14630 dtrace_dof_error(dof, "provider section too small");
14634 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14635 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14636 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14637 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14638 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14640 if (str_sec == NULL || prb_sec == NULL ||
14641 arg_sec == NULL || off_sec == NULL)
14646 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14647 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14648 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14649 provider->dofpv_prenoffs)) == NULL)
14652 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14654 if (provider->dofpv_name >= str_sec->dofs_size ||
14655 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14656 dtrace_dof_error(dof, "invalid provider name");
14660 if (prb_sec->dofs_entsize == 0 ||
14661 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14662 dtrace_dof_error(dof, "invalid entry size");
14666 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14667 dtrace_dof_error(dof, "misaligned entry size");
14671 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14672 dtrace_dof_error(dof, "invalid entry size");
14676 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14677 dtrace_dof_error(dof, "misaligned section offset");
14681 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14682 dtrace_dof_error(dof, "invalid entry size");
14686 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14688 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14691 * Take a pass through the probes to check for errors.
14693 for (j = 0; j < nprobes; j++) {
14694 probe = (dof_probe_t *)(uintptr_t)(daddr +
14695 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14697 if (probe->dofpr_func >= str_sec->dofs_size) {
14698 dtrace_dof_error(dof, "invalid function name");
14702 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14703 dtrace_dof_error(dof, "function name too long");
14707 if (probe->dofpr_name >= str_sec->dofs_size ||
14708 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14709 dtrace_dof_error(dof, "invalid probe name");
14714 * The offset count must not wrap the index, and the offsets
14715 * must also not overflow the section's data.
14717 if (probe->dofpr_offidx + probe->dofpr_noffs <
14718 probe->dofpr_offidx ||
14719 (probe->dofpr_offidx + probe->dofpr_noffs) *
14720 off_sec->dofs_entsize > off_sec->dofs_size) {
14721 dtrace_dof_error(dof, "invalid probe offset");
14725 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14727 * If there's no is-enabled offset section, make sure
14728 * there aren't any is-enabled offsets. Otherwise
14729 * perform the same checks as for probe offsets
14730 * (immediately above).
14732 if (enoff_sec == NULL) {
14733 if (probe->dofpr_enoffidx != 0 ||
14734 probe->dofpr_nenoffs != 0) {
14735 dtrace_dof_error(dof, "is-enabled "
14736 "offsets with null section");
14739 } else if (probe->dofpr_enoffidx +
14740 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14741 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14742 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14743 dtrace_dof_error(dof, "invalid is-enabled "
14748 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14749 dtrace_dof_error(dof, "zero probe and "
14750 "is-enabled offsets");
14753 } else if (probe->dofpr_noffs == 0) {
14754 dtrace_dof_error(dof, "zero probe offsets");
14758 if (probe->dofpr_argidx + probe->dofpr_xargc <
14759 probe->dofpr_argidx ||
14760 (probe->dofpr_argidx + probe->dofpr_xargc) *
14761 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14762 dtrace_dof_error(dof, "invalid args");
14766 typeidx = probe->dofpr_nargv;
14767 typestr = strtab + probe->dofpr_nargv;
14768 for (k = 0; k < probe->dofpr_nargc; k++) {
14769 if (typeidx >= str_sec->dofs_size) {
14770 dtrace_dof_error(dof, "bad "
14771 "native argument type");
14775 typesz = strlen(typestr) + 1;
14776 if (typesz > DTRACE_ARGTYPELEN) {
14777 dtrace_dof_error(dof, "native "
14778 "argument type too long");
14785 typeidx = probe->dofpr_xargv;
14786 typestr = strtab + probe->dofpr_xargv;
14787 for (k = 0; k < probe->dofpr_xargc; k++) {
14788 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14789 dtrace_dof_error(dof, "bad "
14790 "native argument index");
14794 if (typeidx >= str_sec->dofs_size) {
14795 dtrace_dof_error(dof, "bad "
14796 "translated argument type");
14800 typesz = strlen(typestr) + 1;
14801 if (typesz > DTRACE_ARGTYPELEN) {
14802 dtrace_dof_error(dof, "translated argument "
14816 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14818 dtrace_helpers_t *help;
14819 dtrace_vstate_t *vstate;
14820 dtrace_enabling_t *enab = NULL;
14821 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14822 uintptr_t daddr = (uintptr_t)dof;
14824 ASSERT(MUTEX_HELD(&dtrace_lock));
14826 if ((help = curproc->p_dtrace_helpers) == NULL)
14827 help = dtrace_helpers_create(curproc);
14829 vstate = &help->dthps_vstate;
14831 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14832 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14833 dtrace_dof_destroy(dof);
14838 * Look for helper providers and validate their descriptions.
14841 for (i = 0; i < dof->dofh_secnum; i++) {
14842 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14843 dof->dofh_secoff + i * dof->dofh_secsize);
14845 if (sec->dofs_type != DOF_SECT_PROVIDER)
14848 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14849 dtrace_enabling_destroy(enab);
14850 dtrace_dof_destroy(dof);
14859 * Now we need to walk through the ECB descriptions in the enabling.
14861 for (i = 0; i < enab->dten_ndesc; i++) {
14862 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14863 dtrace_probedesc_t *desc = &ep->dted_probe;
14865 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14868 if (strcmp(desc->dtpd_mod, "helper") != 0)
14871 if (strcmp(desc->dtpd_func, "ustack") != 0)
14874 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14877 * Adding this helper action failed -- we are now going
14878 * to rip out the entire generation and return failure.
14880 (void) dtrace_helper_destroygen(help->dthps_generation);
14881 dtrace_enabling_destroy(enab);
14882 dtrace_dof_destroy(dof);
14889 if (nhelpers < enab->dten_ndesc)
14890 dtrace_dof_error(dof, "unmatched helpers");
14892 gen = help->dthps_generation++;
14893 dtrace_enabling_destroy(enab);
14895 if (dhp != NULL && nprovs > 0) {
14896 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14897 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14898 mutex_exit(&dtrace_lock);
14899 dtrace_helper_provider_register(curproc, help, dhp);
14900 mutex_enter(&dtrace_lock);
14907 dtrace_dof_destroy(dof);
14912 static dtrace_helpers_t *
14913 dtrace_helpers_create(proc_t *p)
14915 dtrace_helpers_t *help;
14917 ASSERT(MUTEX_HELD(&dtrace_lock));
14918 ASSERT(p->p_dtrace_helpers == NULL);
14920 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
14921 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
14922 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
14924 p->p_dtrace_helpers = help;
14934 dtrace_helpers_destroy(proc_t *p)
14936 dtrace_helpers_t *help;
14937 dtrace_vstate_t *vstate;
14939 proc_t *p = curproc;
14943 mutex_enter(&dtrace_lock);
14945 ASSERT(p->p_dtrace_helpers != NULL);
14946 ASSERT(dtrace_helpers > 0);
14948 help = p->p_dtrace_helpers;
14949 vstate = &help->dthps_vstate;
14952 * We're now going to lose the help from this process.
14954 p->p_dtrace_helpers = NULL;
14958 * Destory the helper actions.
14960 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14961 dtrace_helper_action_t *h, *next;
14963 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14964 next = h->dtha_next;
14965 dtrace_helper_action_destroy(h, vstate);
14970 mutex_exit(&dtrace_lock);
14973 * Destroy the helper providers.
14975 if (help->dthps_maxprovs > 0) {
14976 mutex_enter(&dtrace_meta_lock);
14977 if (dtrace_meta_pid != NULL) {
14978 ASSERT(dtrace_deferred_pid == NULL);
14980 for (i = 0; i < help->dthps_nprovs; i++) {
14981 dtrace_helper_provider_remove(
14982 &help->dthps_provs[i]->dthp_prov, p->p_pid);
14985 mutex_enter(&dtrace_lock);
14986 ASSERT(help->dthps_deferred == 0 ||
14987 help->dthps_next != NULL ||
14988 help->dthps_prev != NULL ||
14989 help == dtrace_deferred_pid);
14992 * Remove the helper from the deferred list.
14994 if (help->dthps_next != NULL)
14995 help->dthps_next->dthps_prev = help->dthps_prev;
14996 if (help->dthps_prev != NULL)
14997 help->dthps_prev->dthps_next = help->dthps_next;
14998 if (dtrace_deferred_pid == help) {
14999 dtrace_deferred_pid = help->dthps_next;
15000 ASSERT(help->dthps_prev == NULL);
15003 mutex_exit(&dtrace_lock);
15006 mutex_exit(&dtrace_meta_lock);
15008 for (i = 0; i < help->dthps_nprovs; i++) {
15009 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15012 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15013 sizeof (dtrace_helper_provider_t *));
15016 mutex_enter(&dtrace_lock);
15018 dtrace_vstate_fini(&help->dthps_vstate);
15019 kmem_free(help->dthps_actions,
15020 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15021 kmem_free(help, sizeof (dtrace_helpers_t));
15024 mutex_exit(&dtrace_lock);
15031 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15033 dtrace_helpers_t *help, *newhelp;
15034 dtrace_helper_action_t *helper, *new, *last;
15036 dtrace_vstate_t *vstate;
15037 int i, j, sz, hasprovs = 0;
15039 mutex_enter(&dtrace_lock);
15040 ASSERT(from->p_dtrace_helpers != NULL);
15041 ASSERT(dtrace_helpers > 0);
15043 help = from->p_dtrace_helpers;
15044 newhelp = dtrace_helpers_create(to);
15045 ASSERT(to->p_dtrace_helpers != NULL);
15047 newhelp->dthps_generation = help->dthps_generation;
15048 vstate = &newhelp->dthps_vstate;
15051 * Duplicate the helper actions.
15053 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15054 if ((helper = help->dthps_actions[i]) == NULL)
15057 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15058 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15060 new->dtha_generation = helper->dtha_generation;
15062 if ((dp = helper->dtha_predicate) != NULL) {
15063 dp = dtrace_difo_duplicate(dp, vstate);
15064 new->dtha_predicate = dp;
15067 new->dtha_nactions = helper->dtha_nactions;
15068 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15069 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15071 for (j = 0; j < new->dtha_nactions; j++) {
15072 dtrace_difo_t *dp = helper->dtha_actions[j];
15074 ASSERT(dp != NULL);
15075 dp = dtrace_difo_duplicate(dp, vstate);
15076 new->dtha_actions[j] = dp;
15079 if (last != NULL) {
15080 last->dtha_next = new;
15082 newhelp->dthps_actions[i] = new;
15090 * Duplicate the helper providers and register them with the
15091 * DTrace framework.
15093 if (help->dthps_nprovs > 0) {
15094 newhelp->dthps_nprovs = help->dthps_nprovs;
15095 newhelp->dthps_maxprovs = help->dthps_nprovs;
15096 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15097 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15098 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15099 newhelp->dthps_provs[i] = help->dthps_provs[i];
15100 newhelp->dthps_provs[i]->dthp_ref++;
15106 mutex_exit(&dtrace_lock);
15109 dtrace_helper_provider_register(to, newhelp, NULL);
15113 * DTrace Hook Functions
15116 dtrace_module_loaded(modctl_t *ctl)
15118 dtrace_provider_t *prv;
15120 mutex_enter(&dtrace_provider_lock);
15121 mutex_enter(&mod_lock);
15124 ASSERT(ctl->mod_busy);
15128 * We're going to call each providers per-module provide operation
15129 * specifying only this module.
15131 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15132 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15134 mutex_exit(&mod_lock);
15135 mutex_exit(&dtrace_provider_lock);
15138 * If we have any retained enablings, we need to match against them.
15139 * Enabling probes requires that cpu_lock be held, and we cannot hold
15140 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15141 * module. (In particular, this happens when loading scheduling
15142 * classes.) So if we have any retained enablings, we need to dispatch
15143 * our task queue to do the match for us.
15145 mutex_enter(&dtrace_lock);
15147 if (dtrace_retained == NULL) {
15148 mutex_exit(&dtrace_lock);
15152 (void) taskq_dispatch(dtrace_taskq,
15153 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15155 mutex_exit(&dtrace_lock);
15158 * And now, for a little heuristic sleaze: in general, we want to
15159 * match modules as soon as they load. However, we cannot guarantee
15160 * this, because it would lead us to the lock ordering violation
15161 * outlined above. The common case, of course, is that cpu_lock is
15162 * _not_ held -- so we delay here for a clock tick, hoping that that's
15163 * long enough for the task queue to do its work. If it's not, it's
15164 * not a serious problem -- it just means that the module that we
15165 * just loaded may not be immediately instrumentable.
15172 dtrace_module_unloaded(modctl_t *ctl)
15174 dtrace_module_unloaded(modctl_t *ctl, int *error)
15177 dtrace_probe_t template, *probe, *first, *next;
15178 dtrace_provider_t *prov;
15180 char modname[DTRACE_MODNAMELEN];
15185 template.dtpr_mod = ctl->mod_modname;
15187 /* Handle the fact that ctl->filename may end in ".ko". */
15188 strlcpy(modname, ctl->filename, sizeof(modname));
15189 len = strlen(ctl->filename);
15190 if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
15191 modname[len - 3] = '\0';
15192 template.dtpr_mod = modname;
15195 mutex_enter(&dtrace_provider_lock);
15196 mutex_enter(&mod_lock);
15197 mutex_enter(&dtrace_lock);
15200 if (ctl->nenabled > 0) {
15201 /* Don't allow unloads if a probe is enabled. */
15202 mutex_exit(&dtrace_provider_lock);
15203 mutex_exit(&dtrace_lock);
15206 "kldunload: attempt to unload module that has DTrace probes enabled\n");
15211 if (dtrace_bymod == NULL) {
15213 * The DTrace module is loaded (obviously) but not attached;
15214 * we don't have any work to do.
15216 mutex_exit(&dtrace_provider_lock);
15217 mutex_exit(&mod_lock);
15218 mutex_exit(&dtrace_lock);
15222 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15223 probe != NULL; probe = probe->dtpr_nextmod) {
15224 if (probe->dtpr_ecb != NULL) {
15225 mutex_exit(&dtrace_provider_lock);
15226 mutex_exit(&mod_lock);
15227 mutex_exit(&dtrace_lock);
15230 * This shouldn't _actually_ be possible -- we're
15231 * unloading a module that has an enabled probe in it.
15232 * (It's normally up to the provider to make sure that
15233 * this can't happen.) However, because dtps_enable()
15234 * doesn't have a failure mode, there can be an
15235 * enable/unload race. Upshot: we don't want to
15236 * assert, but we're not going to disable the
15239 if (dtrace_err_verbose) {
15241 cmn_err(CE_WARN, "unloaded module '%s' had "
15242 "enabled probes", ctl->mod_modname);
15244 cmn_err(CE_WARN, "unloaded module '%s' had "
15245 "enabled probes", modname);
15255 for (first = NULL; probe != NULL; probe = next) {
15256 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15258 dtrace_probes[probe->dtpr_id - 1] = NULL;
15260 next = probe->dtpr_nextmod;
15261 dtrace_hash_remove(dtrace_bymod, probe);
15262 dtrace_hash_remove(dtrace_byfunc, probe);
15263 dtrace_hash_remove(dtrace_byname, probe);
15265 if (first == NULL) {
15267 probe->dtpr_nextmod = NULL;
15269 probe->dtpr_nextmod = first;
15275 * We've removed all of the module's probes from the hash chains and
15276 * from the probe array. Now issue a dtrace_sync() to be sure that
15277 * everyone has cleared out from any probe array processing.
15281 for (probe = first; probe != NULL; probe = first) {
15282 first = probe->dtpr_nextmod;
15283 prov = probe->dtpr_provider;
15284 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15286 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15287 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15288 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15290 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15292 free_unr(dtrace_arena, probe->dtpr_id);
15294 kmem_free(probe, sizeof (dtrace_probe_t));
15297 mutex_exit(&dtrace_lock);
15298 mutex_exit(&mod_lock);
15299 mutex_exit(&dtrace_provider_lock);
15304 dtrace_kld_load(void *arg __unused, linker_file_t lf)
15307 dtrace_module_loaded(lf);
15311 dtrace_kld_unload(void *arg __unused, linker_file_t lf, int *error)
15315 /* We already have an error, so don't do anything. */
15317 dtrace_module_unloaded(lf, error);
15323 dtrace_suspend(void)
15325 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15329 dtrace_resume(void)
15331 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15336 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15338 ASSERT(MUTEX_HELD(&cpu_lock));
15339 mutex_enter(&dtrace_lock);
15343 dtrace_state_t *state;
15344 dtrace_optval_t *opt, rs, c;
15347 * For now, we only allocate a new buffer for anonymous state.
15349 if ((state = dtrace_anon.dta_state) == NULL)
15352 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15355 opt = state->dts_options;
15356 c = opt[DTRACEOPT_CPU];
15358 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15362 * Regardless of what the actual policy is, we're going to
15363 * temporarily set our resize policy to be manual. We're
15364 * also going to temporarily set our CPU option to denote
15365 * the newly configured CPU.
15367 rs = opt[DTRACEOPT_BUFRESIZE];
15368 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15369 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15371 (void) dtrace_state_buffers(state);
15373 opt[DTRACEOPT_BUFRESIZE] = rs;
15374 opt[DTRACEOPT_CPU] = c;
15381 * We don't free the buffer in the CPU_UNCONFIG case. (The
15382 * buffer will be freed when the consumer exits.)
15390 mutex_exit(&dtrace_lock);
15396 dtrace_cpu_setup_initial(processorid_t cpu)
15398 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15403 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15405 if (dtrace_toxranges >= dtrace_toxranges_max) {
15407 dtrace_toxrange_t *range;
15409 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15412 ASSERT(dtrace_toxrange == NULL);
15413 ASSERT(dtrace_toxranges_max == 0);
15414 dtrace_toxranges_max = 1;
15416 dtrace_toxranges_max <<= 1;
15419 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15420 range = kmem_zalloc(nsize, KM_SLEEP);
15422 if (dtrace_toxrange != NULL) {
15423 ASSERT(osize != 0);
15424 bcopy(dtrace_toxrange, range, osize);
15425 kmem_free(dtrace_toxrange, osize);
15428 dtrace_toxrange = range;
15431 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15432 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15434 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15435 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15436 dtrace_toxranges++;
15440 * DTrace Driver Cookbook Functions
15445 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15447 dtrace_provider_id_t id;
15448 dtrace_state_t *state = NULL;
15449 dtrace_enabling_t *enab;
15451 mutex_enter(&cpu_lock);
15452 mutex_enter(&dtrace_provider_lock);
15453 mutex_enter(&dtrace_lock);
15455 if (ddi_soft_state_init(&dtrace_softstate,
15456 sizeof (dtrace_state_t), 0) != 0) {
15457 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15458 mutex_exit(&cpu_lock);
15459 mutex_exit(&dtrace_provider_lock);
15460 mutex_exit(&dtrace_lock);
15461 return (DDI_FAILURE);
15464 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15465 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15466 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15467 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15468 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15469 ddi_remove_minor_node(devi, NULL);
15470 ddi_soft_state_fini(&dtrace_softstate);
15471 mutex_exit(&cpu_lock);
15472 mutex_exit(&dtrace_provider_lock);
15473 mutex_exit(&dtrace_lock);
15474 return (DDI_FAILURE);
15477 ddi_report_dev(devi);
15478 dtrace_devi = devi;
15480 dtrace_modload = dtrace_module_loaded;
15481 dtrace_modunload = dtrace_module_unloaded;
15482 dtrace_cpu_init = dtrace_cpu_setup_initial;
15483 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15484 dtrace_helpers_fork = dtrace_helpers_duplicate;
15485 dtrace_cpustart_init = dtrace_suspend;
15486 dtrace_cpustart_fini = dtrace_resume;
15487 dtrace_debugger_init = dtrace_suspend;
15488 dtrace_debugger_fini = dtrace_resume;
15490 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15492 ASSERT(MUTEX_HELD(&cpu_lock));
15494 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15495 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15496 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15497 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15498 VM_SLEEP | VMC_IDENTIFIER);
15499 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15502 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15503 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15504 NULL, NULL, NULL, NULL, NULL, 0);
15506 ASSERT(MUTEX_HELD(&cpu_lock));
15507 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15508 offsetof(dtrace_probe_t, dtpr_nextmod),
15509 offsetof(dtrace_probe_t, dtpr_prevmod));
15511 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15512 offsetof(dtrace_probe_t, dtpr_nextfunc),
15513 offsetof(dtrace_probe_t, dtpr_prevfunc));
15515 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15516 offsetof(dtrace_probe_t, dtpr_nextname),
15517 offsetof(dtrace_probe_t, dtpr_prevname));
15519 if (dtrace_retain_max < 1) {
15520 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15521 "setting to 1", dtrace_retain_max);
15522 dtrace_retain_max = 1;
15526 * Now discover our toxic ranges.
15528 dtrace_toxic_ranges(dtrace_toxrange_add);
15531 * Before we register ourselves as a provider to our own framework,
15532 * we would like to assert that dtrace_provider is NULL -- but that's
15533 * not true if we were loaded as a dependency of a DTrace provider.
15534 * Once we've registered, we can assert that dtrace_provider is our
15537 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15538 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15540 ASSERT(dtrace_provider != NULL);
15541 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15543 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15544 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15545 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15546 dtrace_provider, NULL, NULL, "END", 0, NULL);
15547 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15548 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15550 dtrace_anon_property();
15551 mutex_exit(&cpu_lock);
15554 * If DTrace helper tracing is enabled, we need to allocate the
15555 * trace buffer and initialize the values.
15557 if (dtrace_helptrace_enabled) {
15558 ASSERT(dtrace_helptrace_buffer == NULL);
15559 dtrace_helptrace_buffer =
15560 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15561 dtrace_helptrace_next = 0;
15565 * If there are already providers, we must ask them to provide their
15566 * probes, and then match any anonymous enabling against them. Note
15567 * that there should be no other retained enablings at this time:
15568 * the only retained enablings at this time should be the anonymous
15571 if (dtrace_anon.dta_enabling != NULL) {
15572 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15574 dtrace_enabling_provide(NULL);
15575 state = dtrace_anon.dta_state;
15578 * We couldn't hold cpu_lock across the above call to
15579 * dtrace_enabling_provide(), but we must hold it to actually
15580 * enable the probes. We have to drop all of our locks, pick
15581 * up cpu_lock, and regain our locks before matching the
15582 * retained anonymous enabling.
15584 mutex_exit(&dtrace_lock);
15585 mutex_exit(&dtrace_provider_lock);
15587 mutex_enter(&cpu_lock);
15588 mutex_enter(&dtrace_provider_lock);
15589 mutex_enter(&dtrace_lock);
15591 if ((enab = dtrace_anon.dta_enabling) != NULL)
15592 (void) dtrace_enabling_match(enab, NULL);
15594 mutex_exit(&cpu_lock);
15597 mutex_exit(&dtrace_lock);
15598 mutex_exit(&dtrace_provider_lock);
15600 if (state != NULL) {
15602 * If we created any anonymous state, set it going now.
15604 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15607 return (DDI_SUCCESS);
15612 #if __FreeBSD_version >= 800039
15614 dtrace_dtr(void *data __unused)
15623 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15625 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15628 dtrace_state_t *state;
15634 if (getminor(*devp) == DTRACEMNRN_HELPER)
15638 * If this wasn't an open with the "helper" minor, then it must be
15639 * the "dtrace" minor.
15641 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15643 cred_t *cred_p = NULL;
15645 #if __FreeBSD_version < 800039
15647 * The first minor device is the one that is cloned so there is
15648 * nothing more to do here.
15650 if (dev2unit(dev) == 0)
15654 * Devices are cloned, so if the DTrace state has already
15655 * been allocated, that means this device belongs to a
15656 * different client. Each client should open '/dev/dtrace'
15657 * to get a cloned device.
15659 if (dev->si_drv1 != NULL)
15663 cred_p = dev->si_cred;
15667 * If no DTRACE_PRIV_* bits are set in the credential, then the
15668 * caller lacks sufficient permission to do anything with DTrace.
15670 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15671 if (priv == DTRACE_PRIV_NONE) {
15673 #if __FreeBSD_version < 800039
15674 /* Destroy the cloned device. */
15683 * Ask all providers to provide all their probes.
15685 mutex_enter(&dtrace_provider_lock);
15686 dtrace_probe_provide(NULL, NULL);
15687 mutex_exit(&dtrace_provider_lock);
15689 mutex_enter(&cpu_lock);
15690 mutex_enter(&dtrace_lock);
15692 dtrace_membar_producer();
15696 * If the kernel debugger is active (that is, if the kernel debugger
15697 * modified text in some way), we won't allow the open.
15699 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15701 mutex_exit(&cpu_lock);
15702 mutex_exit(&dtrace_lock);
15706 state = dtrace_state_create(devp, cred_p);
15708 state = dtrace_state_create(dev);
15709 #if __FreeBSD_version < 800039
15710 dev->si_drv1 = state;
15712 devfs_set_cdevpriv(state, dtrace_dtr);
15714 /* This code actually belongs in dtrace_attach() */
15715 if (dtrace_opens == 1)
15716 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15720 mutex_exit(&cpu_lock);
15722 if (state == NULL) {
15724 if (--dtrace_opens == 0)
15725 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15729 mutex_exit(&dtrace_lock);
15731 #if __FreeBSD_version < 800039
15732 /* Destroy the cloned device. */
15739 mutex_exit(&dtrace_lock);
15747 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15749 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15753 minor_t minor = getminor(dev);
15754 dtrace_state_t *state;
15756 if (minor == DTRACEMNRN_HELPER)
15759 state = ddi_get_soft_state(dtrace_softstate, minor);
15761 #if __FreeBSD_version < 800039
15762 dtrace_state_t *state = dev->si_drv1;
15764 /* Check if this is not a cloned device. */
15765 if (dev2unit(dev) == 0)
15768 dtrace_state_t *state;
15769 devfs_get_cdevpriv((void **) &state);
15774 mutex_enter(&cpu_lock);
15775 mutex_enter(&dtrace_lock);
15777 if (state != NULL) {
15778 if (state->dts_anon) {
15780 * There is anonymous state. Destroy that first.
15782 ASSERT(dtrace_anon.dta_state == NULL);
15783 dtrace_state_destroy(state->dts_anon);
15786 dtrace_state_destroy(state);
15789 kmem_free(state, 0);
15790 #if __FreeBSD_version < 800039
15791 dev->si_drv1 = NULL;
15796 ASSERT(dtrace_opens > 0);
15798 if (--dtrace_opens == 0)
15799 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15802 /* This code actually belongs in dtrace_detach() */
15803 if ((dtrace_opens == 0) && (dtrace_taskq != NULL)) {
15804 taskq_destroy(dtrace_taskq);
15805 dtrace_taskq = NULL;
15809 mutex_exit(&dtrace_lock);
15810 mutex_exit(&cpu_lock);
15812 #if __FreeBSD_version < 800039
15813 /* Schedule this cloned device to be destroyed. */
15814 destroy_dev_sched(dev);
15823 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15826 dof_helper_t help, *dhp = NULL;
15829 case DTRACEHIOC_ADDDOF:
15830 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15831 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15836 arg = (intptr_t)help.dofhp_dof;
15839 case DTRACEHIOC_ADD: {
15840 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15845 mutex_enter(&dtrace_lock);
15848 * dtrace_helper_slurp() takes responsibility for the dof --
15849 * it may free it now or it may save it and free it later.
15851 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15858 mutex_exit(&dtrace_lock);
15862 case DTRACEHIOC_REMOVE: {
15863 mutex_enter(&dtrace_lock);
15864 rval = dtrace_helper_destroygen(arg);
15865 mutex_exit(&dtrace_lock);
15879 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15881 minor_t minor = getminor(dev);
15882 dtrace_state_t *state;
15885 if (minor == DTRACEMNRN_HELPER)
15886 return (dtrace_ioctl_helper(cmd, arg, rv));
15888 state = ddi_get_soft_state(dtrace_softstate, minor);
15890 if (state->dts_anon) {
15891 ASSERT(dtrace_anon.dta_state == NULL);
15892 state = state->dts_anon;
15896 case DTRACEIOC_PROVIDER: {
15897 dtrace_providerdesc_t pvd;
15898 dtrace_provider_t *pvp;
15900 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
15903 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
15904 mutex_enter(&dtrace_provider_lock);
15906 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
15907 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
15911 mutex_exit(&dtrace_provider_lock);
15916 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
15917 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
15919 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
15925 case DTRACEIOC_EPROBE: {
15926 dtrace_eprobedesc_t epdesc;
15928 dtrace_action_t *act;
15934 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
15937 mutex_enter(&dtrace_lock);
15939 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
15940 mutex_exit(&dtrace_lock);
15944 if (ecb->dte_probe == NULL) {
15945 mutex_exit(&dtrace_lock);
15949 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
15950 epdesc.dtepd_uarg = ecb->dte_uarg;
15951 epdesc.dtepd_size = ecb->dte_size;
15953 nrecs = epdesc.dtepd_nrecs;
15954 epdesc.dtepd_nrecs = 0;
15955 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15956 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15959 epdesc.dtepd_nrecs++;
15963 * Now that we have the size, we need to allocate a temporary
15964 * buffer in which to store the complete description. We need
15965 * the temporary buffer to be able to drop dtrace_lock()
15966 * across the copyout(), below.
15968 size = sizeof (dtrace_eprobedesc_t) +
15969 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
15971 buf = kmem_alloc(size, KM_SLEEP);
15972 dest = (uintptr_t)buf;
15974 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
15975 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
15977 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
15978 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
15984 bcopy(&act->dta_rec, (void *)dest,
15985 sizeof (dtrace_recdesc_t));
15986 dest += sizeof (dtrace_recdesc_t);
15989 mutex_exit(&dtrace_lock);
15991 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
15992 kmem_free(buf, size);
15996 kmem_free(buf, size);
16000 case DTRACEIOC_AGGDESC: {
16001 dtrace_aggdesc_t aggdesc;
16002 dtrace_action_t *act;
16003 dtrace_aggregation_t *agg;
16006 dtrace_recdesc_t *lrec;
16011 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16014 mutex_enter(&dtrace_lock);
16016 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16017 mutex_exit(&dtrace_lock);
16021 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16023 nrecs = aggdesc.dtagd_nrecs;
16024 aggdesc.dtagd_nrecs = 0;
16026 offs = agg->dtag_base;
16027 lrec = &agg->dtag_action.dta_rec;
16028 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16030 for (act = agg->dtag_first; ; act = act->dta_next) {
16031 ASSERT(act->dta_intuple ||
16032 DTRACEACT_ISAGG(act->dta_kind));
16035 * If this action has a record size of zero, it
16036 * denotes an argument to the aggregating action.
16037 * Because the presence of this record doesn't (or
16038 * shouldn't) affect the way the data is interpreted,
16039 * we don't copy it out to save user-level the
16040 * confusion of dealing with a zero-length record.
16042 if (act->dta_rec.dtrd_size == 0) {
16043 ASSERT(agg->dtag_hasarg);
16047 aggdesc.dtagd_nrecs++;
16049 if (act == &agg->dtag_action)
16054 * Now that we have the size, we need to allocate a temporary
16055 * buffer in which to store the complete description. We need
16056 * the temporary buffer to be able to drop dtrace_lock()
16057 * across the copyout(), below.
16059 size = sizeof (dtrace_aggdesc_t) +
16060 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16062 buf = kmem_alloc(size, KM_SLEEP);
16063 dest = (uintptr_t)buf;
16065 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16066 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16068 for (act = agg->dtag_first; ; act = act->dta_next) {
16069 dtrace_recdesc_t rec = act->dta_rec;
16072 * See the comment in the above loop for why we pass
16073 * over zero-length records.
16075 if (rec.dtrd_size == 0) {
16076 ASSERT(agg->dtag_hasarg);
16083 rec.dtrd_offset -= offs;
16084 bcopy(&rec, (void *)dest, sizeof (rec));
16085 dest += sizeof (dtrace_recdesc_t);
16087 if (act == &agg->dtag_action)
16091 mutex_exit(&dtrace_lock);
16093 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16094 kmem_free(buf, size);
16098 kmem_free(buf, size);
16102 case DTRACEIOC_ENABLE: {
16104 dtrace_enabling_t *enab = NULL;
16105 dtrace_vstate_t *vstate;
16111 * If a NULL argument has been passed, we take this as our
16112 * cue to reevaluate our enablings.
16115 dtrace_enabling_matchall();
16120 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16123 mutex_enter(&cpu_lock);
16124 mutex_enter(&dtrace_lock);
16125 vstate = &state->dts_vstate;
16127 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16128 mutex_exit(&dtrace_lock);
16129 mutex_exit(&cpu_lock);
16130 dtrace_dof_destroy(dof);
16134 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16135 mutex_exit(&dtrace_lock);
16136 mutex_exit(&cpu_lock);
16137 dtrace_dof_destroy(dof);
16141 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16142 dtrace_enabling_destroy(enab);
16143 mutex_exit(&dtrace_lock);
16144 mutex_exit(&cpu_lock);
16145 dtrace_dof_destroy(dof);
16149 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16150 err = dtrace_enabling_retain(enab);
16152 dtrace_enabling_destroy(enab);
16155 mutex_exit(&cpu_lock);
16156 mutex_exit(&dtrace_lock);
16157 dtrace_dof_destroy(dof);
16162 case DTRACEIOC_REPLICATE: {
16163 dtrace_repldesc_t desc;
16164 dtrace_probedesc_t *match = &desc.dtrpd_match;
16165 dtrace_probedesc_t *create = &desc.dtrpd_create;
16168 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16171 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16172 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16173 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16174 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16176 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16177 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16178 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16179 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16181 mutex_enter(&dtrace_lock);
16182 err = dtrace_enabling_replicate(state, match, create);
16183 mutex_exit(&dtrace_lock);
16188 case DTRACEIOC_PROBEMATCH:
16189 case DTRACEIOC_PROBES: {
16190 dtrace_probe_t *probe = NULL;
16191 dtrace_probedesc_t desc;
16192 dtrace_probekey_t pkey;
16199 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16202 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16203 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16204 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16205 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16208 * Before we attempt to match this probe, we want to give
16209 * all providers the opportunity to provide it.
16211 if (desc.dtpd_id == DTRACE_IDNONE) {
16212 mutex_enter(&dtrace_provider_lock);
16213 dtrace_probe_provide(&desc, NULL);
16214 mutex_exit(&dtrace_provider_lock);
16218 if (cmd == DTRACEIOC_PROBEMATCH) {
16219 dtrace_probekey(&desc, &pkey);
16220 pkey.dtpk_id = DTRACE_IDNONE;
16223 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16225 mutex_enter(&dtrace_lock);
16227 if (cmd == DTRACEIOC_PROBEMATCH) {
16228 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16229 if ((probe = dtrace_probes[i - 1]) != NULL &&
16230 (m = dtrace_match_probe(probe, &pkey,
16231 priv, uid, zoneid)) != 0)
16236 mutex_exit(&dtrace_lock);
16241 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16242 if ((probe = dtrace_probes[i - 1]) != NULL &&
16243 dtrace_match_priv(probe, priv, uid, zoneid))
16248 if (probe == NULL) {
16249 mutex_exit(&dtrace_lock);
16253 dtrace_probe_description(probe, &desc);
16254 mutex_exit(&dtrace_lock);
16256 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16262 case DTRACEIOC_PROBEARG: {
16263 dtrace_argdesc_t desc;
16264 dtrace_probe_t *probe;
16265 dtrace_provider_t *prov;
16267 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16270 if (desc.dtargd_id == DTRACE_IDNONE)
16273 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16276 mutex_enter(&dtrace_provider_lock);
16277 mutex_enter(&mod_lock);
16278 mutex_enter(&dtrace_lock);
16280 if (desc.dtargd_id > dtrace_nprobes) {
16281 mutex_exit(&dtrace_lock);
16282 mutex_exit(&mod_lock);
16283 mutex_exit(&dtrace_provider_lock);
16287 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16288 mutex_exit(&dtrace_lock);
16289 mutex_exit(&mod_lock);
16290 mutex_exit(&dtrace_provider_lock);
16294 mutex_exit(&dtrace_lock);
16296 prov = probe->dtpr_provider;
16298 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16300 * There isn't any typed information for this probe.
16301 * Set the argument number to DTRACE_ARGNONE.
16303 desc.dtargd_ndx = DTRACE_ARGNONE;
16305 desc.dtargd_native[0] = '\0';
16306 desc.dtargd_xlate[0] = '\0';
16307 desc.dtargd_mapping = desc.dtargd_ndx;
16309 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16310 probe->dtpr_id, probe->dtpr_arg, &desc);
16313 mutex_exit(&mod_lock);
16314 mutex_exit(&dtrace_provider_lock);
16316 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16322 case DTRACEIOC_GO: {
16323 processorid_t cpuid;
16324 rval = dtrace_state_go(state, &cpuid);
16329 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16335 case DTRACEIOC_STOP: {
16336 processorid_t cpuid;
16338 mutex_enter(&dtrace_lock);
16339 rval = dtrace_state_stop(state, &cpuid);
16340 mutex_exit(&dtrace_lock);
16345 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16351 case DTRACEIOC_DOFGET: {
16352 dof_hdr_t hdr, *dof;
16355 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16358 mutex_enter(&dtrace_lock);
16359 dof = dtrace_dof_create(state);
16360 mutex_exit(&dtrace_lock);
16362 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16363 rval = copyout(dof, (void *)arg, len);
16364 dtrace_dof_destroy(dof);
16366 return (rval == 0 ? 0 : EFAULT);
16369 case DTRACEIOC_AGGSNAP:
16370 case DTRACEIOC_BUFSNAP: {
16371 dtrace_bufdesc_t desc;
16373 dtrace_buffer_t *buf;
16375 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16378 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16381 mutex_enter(&dtrace_lock);
16383 if (cmd == DTRACEIOC_BUFSNAP) {
16384 buf = &state->dts_buffer[desc.dtbd_cpu];
16386 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16389 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16390 size_t sz = buf->dtb_offset;
16392 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16393 mutex_exit(&dtrace_lock);
16398 * If this buffer has already been consumed, we're
16399 * going to indicate that there's nothing left here
16402 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16403 mutex_exit(&dtrace_lock);
16405 desc.dtbd_size = 0;
16406 desc.dtbd_drops = 0;
16407 desc.dtbd_errors = 0;
16408 desc.dtbd_oldest = 0;
16409 sz = sizeof (desc);
16411 if (copyout(&desc, (void *)arg, sz) != 0)
16418 * If this is a ring buffer that has wrapped, we want
16419 * to copy the whole thing out.
16421 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16422 dtrace_buffer_polish(buf);
16423 sz = buf->dtb_size;
16426 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16427 mutex_exit(&dtrace_lock);
16431 desc.dtbd_size = sz;
16432 desc.dtbd_drops = buf->dtb_drops;
16433 desc.dtbd_errors = buf->dtb_errors;
16434 desc.dtbd_oldest = buf->dtb_xamot_offset;
16436 mutex_exit(&dtrace_lock);
16438 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16441 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16446 if (buf->dtb_tomax == NULL) {
16447 ASSERT(buf->dtb_xamot == NULL);
16448 mutex_exit(&dtrace_lock);
16452 cached = buf->dtb_tomax;
16453 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16455 dtrace_xcall(desc.dtbd_cpu,
16456 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16458 state->dts_errors += buf->dtb_xamot_errors;
16461 * If the buffers did not actually switch, then the cross call
16462 * did not take place -- presumably because the given CPU is
16463 * not in the ready set. If this is the case, we'll return
16466 if (buf->dtb_tomax == cached) {
16467 ASSERT(buf->dtb_xamot != cached);
16468 mutex_exit(&dtrace_lock);
16472 ASSERT(cached == buf->dtb_xamot);
16475 * We have our snapshot; now copy it out.
16477 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16478 buf->dtb_xamot_offset) != 0) {
16479 mutex_exit(&dtrace_lock);
16483 desc.dtbd_size = buf->dtb_xamot_offset;
16484 desc.dtbd_drops = buf->dtb_xamot_drops;
16485 desc.dtbd_errors = buf->dtb_xamot_errors;
16486 desc.dtbd_oldest = 0;
16488 mutex_exit(&dtrace_lock);
16491 * Finally, copy out the buffer description.
16493 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16499 case DTRACEIOC_CONF: {
16500 dtrace_conf_t conf;
16502 bzero(&conf, sizeof (conf));
16503 conf.dtc_difversion = DIF_VERSION;
16504 conf.dtc_difintregs = DIF_DIR_NREGS;
16505 conf.dtc_diftupregs = DIF_DTR_NREGS;
16506 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16508 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16514 case DTRACEIOC_STATUS: {
16515 dtrace_status_t stat;
16516 dtrace_dstate_t *dstate;
16521 * See the comment in dtrace_state_deadman() for the reason
16522 * for setting dts_laststatus to INT64_MAX before setting
16523 * it to the correct value.
16525 state->dts_laststatus = INT64_MAX;
16526 dtrace_membar_producer();
16527 state->dts_laststatus = dtrace_gethrtime();
16529 bzero(&stat, sizeof (stat));
16531 mutex_enter(&dtrace_lock);
16533 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16534 mutex_exit(&dtrace_lock);
16538 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16539 stat.dtst_exiting = 1;
16541 nerrs = state->dts_errors;
16542 dstate = &state->dts_vstate.dtvs_dynvars;
16544 for (i = 0; i < NCPU; i++) {
16545 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16547 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16548 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16549 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16551 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16552 stat.dtst_filled++;
16554 nerrs += state->dts_buffer[i].dtb_errors;
16556 for (j = 0; j < state->dts_nspeculations; j++) {
16557 dtrace_speculation_t *spec;
16558 dtrace_buffer_t *buf;
16560 spec = &state->dts_speculations[j];
16561 buf = &spec->dtsp_buffer[i];
16562 stat.dtst_specdrops += buf->dtb_xamot_drops;
16566 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16567 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16568 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16569 stat.dtst_dblerrors = state->dts_dblerrors;
16571 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16572 stat.dtst_errors = nerrs;
16574 mutex_exit(&dtrace_lock);
16576 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16582 case DTRACEIOC_FORMAT: {
16583 dtrace_fmtdesc_t fmt;
16587 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16590 mutex_enter(&dtrace_lock);
16592 if (fmt.dtfd_format == 0 ||
16593 fmt.dtfd_format > state->dts_nformats) {
16594 mutex_exit(&dtrace_lock);
16599 * Format strings are allocated contiguously and they are
16600 * never freed; if a format index is less than the number
16601 * of formats, we can assert that the format map is non-NULL
16602 * and that the format for the specified index is non-NULL.
16604 ASSERT(state->dts_formats != NULL);
16605 str = state->dts_formats[fmt.dtfd_format - 1];
16606 ASSERT(str != NULL);
16608 len = strlen(str) + 1;
16610 if (len > fmt.dtfd_length) {
16611 fmt.dtfd_length = len;
16613 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16614 mutex_exit(&dtrace_lock);
16618 if (copyout(str, fmt.dtfd_string, len) != 0) {
16619 mutex_exit(&dtrace_lock);
16624 mutex_exit(&dtrace_lock);
16637 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16639 dtrace_state_t *state;
16646 return (DDI_SUCCESS);
16649 return (DDI_FAILURE);
16652 mutex_enter(&cpu_lock);
16653 mutex_enter(&dtrace_provider_lock);
16654 mutex_enter(&dtrace_lock);
16656 ASSERT(dtrace_opens == 0);
16658 if (dtrace_helpers > 0) {
16659 mutex_exit(&dtrace_provider_lock);
16660 mutex_exit(&dtrace_lock);
16661 mutex_exit(&cpu_lock);
16662 return (DDI_FAILURE);
16665 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16666 mutex_exit(&dtrace_provider_lock);
16667 mutex_exit(&dtrace_lock);
16668 mutex_exit(&cpu_lock);
16669 return (DDI_FAILURE);
16672 dtrace_provider = NULL;
16674 if ((state = dtrace_anon_grab()) != NULL) {
16676 * If there were ECBs on this state, the provider should
16677 * have not been allowed to detach; assert that there is
16680 ASSERT(state->dts_necbs == 0);
16681 dtrace_state_destroy(state);
16684 * If we're being detached with anonymous state, we need to
16685 * indicate to the kernel debugger that DTrace is now inactive.
16687 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16690 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16691 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16692 dtrace_cpu_init = NULL;
16693 dtrace_helpers_cleanup = NULL;
16694 dtrace_helpers_fork = NULL;
16695 dtrace_cpustart_init = NULL;
16696 dtrace_cpustart_fini = NULL;
16697 dtrace_debugger_init = NULL;
16698 dtrace_debugger_fini = NULL;
16699 dtrace_modload = NULL;
16700 dtrace_modunload = NULL;
16702 mutex_exit(&cpu_lock);
16704 if (dtrace_helptrace_enabled) {
16705 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16706 dtrace_helptrace_buffer = NULL;
16709 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16710 dtrace_probes = NULL;
16711 dtrace_nprobes = 0;
16713 dtrace_hash_destroy(dtrace_bymod);
16714 dtrace_hash_destroy(dtrace_byfunc);
16715 dtrace_hash_destroy(dtrace_byname);
16716 dtrace_bymod = NULL;
16717 dtrace_byfunc = NULL;
16718 dtrace_byname = NULL;
16720 kmem_cache_destroy(dtrace_state_cache);
16721 vmem_destroy(dtrace_minor);
16722 vmem_destroy(dtrace_arena);
16724 if (dtrace_toxrange != NULL) {
16725 kmem_free(dtrace_toxrange,
16726 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16727 dtrace_toxrange = NULL;
16728 dtrace_toxranges = 0;
16729 dtrace_toxranges_max = 0;
16732 ddi_remove_minor_node(dtrace_devi, NULL);
16733 dtrace_devi = NULL;
16735 ddi_soft_state_fini(&dtrace_softstate);
16737 ASSERT(dtrace_vtime_references == 0);
16738 ASSERT(dtrace_opens == 0);
16739 ASSERT(dtrace_retained == NULL);
16741 mutex_exit(&dtrace_lock);
16742 mutex_exit(&dtrace_provider_lock);
16745 * We don't destroy the task queue until after we have dropped our
16746 * locks (taskq_destroy() may block on running tasks). To prevent
16747 * attempting to do work after we have effectively detached but before
16748 * the task queue has been destroyed, all tasks dispatched via the
16749 * task queue must check that DTrace is still attached before
16750 * performing any operation.
16752 taskq_destroy(dtrace_taskq);
16753 dtrace_taskq = NULL;
16755 return (DDI_SUCCESS);
16762 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16767 case DDI_INFO_DEVT2DEVINFO:
16768 *result = (void *)dtrace_devi;
16769 error = DDI_SUCCESS;
16771 case DDI_INFO_DEVT2INSTANCE:
16772 *result = (void *)0;
16773 error = DDI_SUCCESS;
16776 error = DDI_FAILURE;
16783 static struct cb_ops dtrace_cb_ops = {
16784 dtrace_open, /* open */
16785 dtrace_close, /* close */
16786 nulldev, /* strategy */
16787 nulldev, /* print */
16791 dtrace_ioctl, /* ioctl */
16792 nodev, /* devmap */
16794 nodev, /* segmap */
16795 nochpoll, /* poll */
16796 ddi_prop_op, /* cb_prop_op */
16798 D_NEW | D_MP /* Driver compatibility flag */
16801 static struct dev_ops dtrace_ops = {
16802 DEVO_REV, /* devo_rev */
16804 dtrace_info, /* get_dev_info */
16805 nulldev, /* identify */
16806 nulldev, /* probe */
16807 dtrace_attach, /* attach */
16808 dtrace_detach, /* detach */
16810 &dtrace_cb_ops, /* driver operations */
16811 NULL, /* bus operations */
16812 nodev /* dev power */
16815 static struct modldrv modldrv = {
16816 &mod_driverops, /* module type (this is a pseudo driver) */
16817 "Dynamic Tracing", /* name of module */
16818 &dtrace_ops, /* driver ops */
16821 static struct modlinkage modlinkage = {
16830 return (mod_install(&modlinkage));
16834 _info(struct modinfo *modinfop)
16836 return (mod_info(&modlinkage, modinfop));
16842 return (mod_remove(&modlinkage));
16846 static d_ioctl_t dtrace_ioctl;
16847 static d_ioctl_t dtrace_ioctl_helper;
16848 static void dtrace_load(void *);
16849 static int dtrace_unload(void);
16850 #if __FreeBSD_version < 800039
16851 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16852 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16853 static eventhandler_tag eh_tag; /* Event handler tag. */
16855 static struct cdev *dtrace_dev;
16856 static struct cdev *helper_dev;
16859 void dtrace_invop_init(void);
16860 void dtrace_invop_uninit(void);
16862 static struct cdevsw dtrace_cdevsw = {
16863 .d_version = D_VERSION,
16864 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16865 .d_close = dtrace_close,
16866 .d_ioctl = dtrace_ioctl,
16867 .d_open = dtrace_open,
16868 .d_name = "dtrace",
16871 static struct cdevsw helper_cdevsw = {
16872 .d_version = D_VERSION,
16873 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16874 .d_ioctl = dtrace_ioctl_helper,
16875 .d_name = "helper",
16878 #include <dtrace_anon.c>
16879 #if __FreeBSD_version < 800039
16880 #include <dtrace_clone.c>
16882 #include <dtrace_ioctl.c>
16883 #include <dtrace_load.c>
16884 #include <dtrace_modevent.c>
16885 #include <dtrace_sysctl.c>
16886 #include <dtrace_unload.c>
16887 #include <dtrace_vtime.c>
16888 #include <dtrace_hacks.c>
16889 #include <dtrace_isa.c>
16891 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16892 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16893 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16895 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16896 MODULE_VERSION(dtrace, 1);
16897 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
16898 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);