4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
25 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
26 * Copyright (c) 2012 by Delphix. All rights reserved
27 * Use is subject to license terms.
30 #pragma ident "%Z%%M% %I% %E% SMI"
33 * DTrace - Dynamic Tracing for Solaris
35 * This is the implementation of the Solaris Dynamic Tracing framework
36 * (DTrace). The user-visible interface to DTrace is described at length in
37 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
38 * library, the in-kernel DTrace framework, and the DTrace providers are
39 * described in the block comments in the <sys/dtrace.h> header file. The
40 * internal architecture of DTrace is described in the block comments in the
41 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
42 * implementation very much assume mastery of all of these sources; if one has
43 * an unanswered question about the implementation, one should consult them
46 * The functions here are ordered roughly as follows:
48 * - Probe context functions
49 * - Probe hashing functions
50 * - Non-probe context utility functions
51 * - Matching functions
52 * - Provider-to-Framework API functions
53 * - Probe management functions
54 * - DIF object functions
56 * - Predicate functions
59 * - Enabling functions
61 * - Anonymous enabling functions
62 * - Consumer state functions
65 * - Driver cookbook functions
67 * Each group of functions begins with a block comment labelled the "DTrace
68 * [Group] Functions", allowing one to find each block by searching forward
69 * on capital-f functions.
71 #include <sys/errno.h>
76 #include <sys/modctl.h>
78 #include <sys/systm.h>
81 #include <sys/sunddi.h>
83 #include <sys/cpuvar.h>
86 #include <sys/strsubr.h>
88 #include <sys/sysmacros.h>
89 #include <sys/dtrace_impl.h>
90 #include <sys/atomic.h>
91 #include <sys/cmn_err.h>
93 #include <sys/mutex_impl.h>
94 #include <sys/rwlock_impl.h>
96 #include <sys/ctf_api.h>
98 #include <sys/panic.h>
99 #include <sys/priv_impl.h>
101 #include <sys/policy.h>
103 #include <sys/cred_impl.h>
104 #include <sys/procfs_isa.h>
106 #include <sys/taskq.h>
108 #include <sys/mkdev.h>
111 #include <sys/zone.h>
112 #include <sys/socket.h>
113 #include <netinet/in.h>
115 /* FreeBSD includes: */
117 #include <sys/callout.h>
118 #include <sys/ctype.h>
119 #include <sys/eventhandler.h>
120 #include <sys/limits.h>
122 #include <sys/kernel.h>
123 #include <sys/malloc.h>
124 #include <sys/sysctl.h>
125 #include <sys/lock.h>
126 #include <sys/mutex.h>
127 #include <sys/rwlock.h>
129 #include <sys/dtrace_bsd.h>
130 #include <netinet/in.h>
131 #include "dtrace_cddl.h"
132 #include "dtrace_debug.c"
136 * DTrace Tunable Variables
138 * The following variables may be tuned by adding a line to /etc/system that
139 * includes both the name of the DTrace module ("dtrace") and the name of the
140 * variable. For example:
142 * set dtrace:dtrace_destructive_disallow = 1
144 * In general, the only variables that one should be tuning this way are those
145 * that affect system-wide DTrace behavior, and for which the default behavior
146 * is undesirable. Most of these variables are tunable on a per-consumer
147 * basis using DTrace options, and need not be tuned on a system-wide basis.
148 * When tuning these variables, avoid pathological values; while some attempt
149 * is made to verify the integrity of these variables, they are not considered
150 * part of the supported interface to DTrace, and they are therefore not
151 * checked comprehensively. Further, these variables should not be tuned
152 * dynamically via "mdb -kw" or other means; they should only be tuned via
155 int dtrace_destructive_disallow = 0;
156 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
157 size_t dtrace_difo_maxsize = (256 * 1024);
158 dtrace_optval_t dtrace_dof_maxsize = (8 * 1024 * 1024);
159 size_t dtrace_global_maxsize = (16 * 1024);
160 size_t dtrace_actions_max = (16 * 1024);
161 size_t dtrace_retain_max = 1024;
162 dtrace_optval_t dtrace_helper_actions_max = 128;
163 dtrace_optval_t dtrace_helper_providers_max = 32;
164 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
165 size_t dtrace_strsize_default = 256;
166 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
167 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
168 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
169 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
170 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
171 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
172 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
173 dtrace_optval_t dtrace_nspec_default = 1;
174 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
175 dtrace_optval_t dtrace_stackframes_default = 20;
176 dtrace_optval_t dtrace_ustackframes_default = 20;
177 dtrace_optval_t dtrace_jstackframes_default = 50;
178 dtrace_optval_t dtrace_jstackstrsize_default = 512;
179 int dtrace_msgdsize_max = 128;
180 hrtime_t dtrace_chill_max = 500 * (NANOSEC / MILLISEC); /* 500 ms */
181 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
182 int dtrace_devdepth_max = 32;
183 int dtrace_err_verbose;
184 hrtime_t dtrace_deadman_interval = NANOSEC;
185 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
186 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
187 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
189 int dtrace_memstr_max = 4096;
193 * DTrace External Variables
195 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
196 * available to DTrace consumers via the backtick (`) syntax. One of these,
197 * dtrace_zero, is made deliberately so: it is provided as a source of
198 * well-known, zero-filled memory. While this variable is not documented,
199 * it is used by some translators as an implementation detail.
201 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
204 * DTrace Internal Variables
207 static dev_info_t *dtrace_devi; /* device info */
210 static vmem_t *dtrace_arena; /* probe ID arena */
211 static vmem_t *dtrace_minor; /* minor number arena */
213 static taskq_t *dtrace_taskq; /* task queue */
214 static struct unrhdr *dtrace_arena; /* Probe ID number. */
216 static dtrace_probe_t **dtrace_probes; /* array of all probes */
217 static int dtrace_nprobes; /* number of probes */
218 static dtrace_provider_t *dtrace_provider; /* provider list */
219 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
220 static int dtrace_opens; /* number of opens */
221 static int dtrace_helpers; /* number of helpers */
223 static void *dtrace_softstate; /* softstate pointer */
225 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
226 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
227 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
228 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
229 static int dtrace_toxranges; /* number of toxic ranges */
230 static int dtrace_toxranges_max; /* size of toxic range array */
231 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
232 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
233 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
234 static kthread_t *dtrace_panicked; /* panicking thread */
235 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
236 static dtrace_genid_t dtrace_probegen; /* current probe generation */
237 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
238 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
239 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
241 static struct mtx dtrace_unr_mtx;
242 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
243 int dtrace_in_probe; /* non-zero if executing a probe */
244 #if defined(__i386__) || defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
245 uintptr_t dtrace_in_probe_addr; /* Address of invop when already in probe */
247 static eventhandler_tag dtrace_kld_load_tag;
248 static eventhandler_tag dtrace_kld_unload_try_tag;
253 * DTrace is protected by three (relatively coarse-grained) locks:
255 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
256 * including enabling state, probes, ECBs, consumer state, helper state,
257 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
258 * probe context is lock-free -- synchronization is handled via the
259 * dtrace_sync() cross call mechanism.
261 * (2) dtrace_provider_lock is required when manipulating provider state, or
262 * when provider state must be held constant.
264 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
265 * when meta provider state must be held constant.
267 * The lock ordering between these three locks is dtrace_meta_lock before
268 * dtrace_provider_lock before dtrace_lock. (In particular, there are
269 * several places where dtrace_provider_lock is held by the framework as it
270 * calls into the providers -- which then call back into the framework,
271 * grabbing dtrace_lock.)
273 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
274 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
275 * role as a coarse-grained lock; it is acquired before both of these locks.
276 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
277 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
278 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
279 * acquired _between_ dtrace_provider_lock and dtrace_lock.
281 static kmutex_t dtrace_lock; /* probe state lock */
282 static kmutex_t dtrace_provider_lock; /* provider state lock */
283 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
286 /* XXX FreeBSD hacks. */
287 #define cr_suid cr_svuid
288 #define cr_sgid cr_svgid
289 #define ipaddr_t in_addr_t
290 #define mod_modname pathname
291 #define vuprintf vprintf
292 #define ttoproc(_a) ((_a)->td_proc)
293 #define crgetzoneid(_a) 0
296 #define CPU_ON_INTR(_a) 0
298 #define PRIV_EFFECTIVE (1 << 0)
299 #define PRIV_DTRACE_KERNEL (1 << 1)
300 #define PRIV_DTRACE_PROC (1 << 2)
301 #define PRIV_DTRACE_USER (1 << 3)
302 #define PRIV_PROC_OWNER (1 << 4)
303 #define PRIV_PROC_ZONE (1 << 5)
306 SYSCTL_NODE(_debug, OID_AUTO, dtrace, CTLFLAG_RD, 0, "DTrace Information");
310 #define curcpu CPU->cpu_id
315 * DTrace Provider Variables
317 * These are the variables relating to DTrace as a provider (that is, the
318 * provider of the BEGIN, END, and ERROR probes).
320 static dtrace_pattr_t dtrace_provider_attr = {
321 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
322 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
323 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
324 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
325 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
332 static dtrace_pops_t dtrace_provider_ops = {
333 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
334 (void (*)(void *, modctl_t *))dtrace_nullop,
335 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
336 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
337 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
338 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
342 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
345 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
346 static dtrace_id_t dtrace_probeid_end; /* special END probe */
347 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
350 * DTrace Helper Tracing Variables
352 uint32_t dtrace_helptrace_next = 0;
353 uint32_t dtrace_helptrace_nlocals;
354 char *dtrace_helptrace_buffer;
355 int dtrace_helptrace_bufsize = 512 * 1024;
358 int dtrace_helptrace_enabled = 1;
360 int dtrace_helptrace_enabled = 0;
364 * DTrace Error Hashing
366 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
367 * table. This is very useful for checking coverage of tests that are
368 * expected to induce DIF or DOF processing errors, and may be useful for
369 * debugging problems in the DIF code generator or in DOF generation . The
370 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
373 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
374 static const char *dtrace_errlast;
375 static kthread_t *dtrace_errthread;
376 static kmutex_t dtrace_errlock;
380 * DTrace Macros and Constants
382 * These are various macros that are useful in various spots in the
383 * implementation, along with a few random constants that have no meaning
384 * outside of the implementation. There is no real structure to this cpp
385 * mishmash -- but is there ever?
387 #define DTRACE_HASHSTR(hash, probe) \
388 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
390 #define DTRACE_HASHNEXT(hash, probe) \
391 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
393 #define DTRACE_HASHPREV(hash, probe) \
394 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
396 #define DTRACE_HASHEQ(hash, lhs, rhs) \
397 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
398 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
400 #define DTRACE_AGGHASHSIZE_SLEW 17
402 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
405 * The key for a thread-local variable consists of the lower 61 bits of the
406 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
407 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
408 * equal to a variable identifier. This is necessary (but not sufficient) to
409 * assure that global associative arrays never collide with thread-local
410 * variables. To guarantee that they cannot collide, we must also define the
411 * order for keying dynamic variables. That order is:
413 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
415 * Because the variable-key and the tls-key are in orthogonal spaces, there is
416 * no way for a global variable key signature to match a thread-local key
420 #define DTRACE_TLS_THRKEY(where) { \
422 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
423 for (; actv; actv >>= 1) \
425 ASSERT(intr < (1 << 3)); \
426 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
427 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
430 #define DTRACE_TLS_THRKEY(where) { \
431 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
433 uint_t actv = _c->cpu_intr_actv; \
434 for (; actv; actv >>= 1) \
436 ASSERT(intr < (1 << 3)); \
437 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
438 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
442 #define DT_BSWAP_8(x) ((x) & 0xff)
443 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
444 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
445 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
447 #define DT_MASK_LO 0x00000000FFFFFFFFULL
449 #define DTRACE_STORE(type, tomax, offset, what) \
450 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
453 #define DTRACE_ALIGNCHECK(addr, size, flags) \
454 if (addr & (size - 1)) { \
455 *flags |= CPU_DTRACE_BADALIGN; \
456 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
460 #define DTRACE_ALIGNCHECK(addr, size, flags)
464 * Test whether a range of memory starting at testaddr of size testsz falls
465 * within the range of memory described by addr, sz. We take care to avoid
466 * problems with overflow and underflow of the unsigned quantities, and
467 * disallow all negative sizes. Ranges of size 0 are allowed.
469 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
470 ((testaddr) - (baseaddr) < (basesz) && \
471 (testaddr) + (testsz) - (baseaddr) <= (basesz) && \
472 (testaddr) + (testsz) >= (testaddr))
475 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
476 * alloc_sz on the righthand side of the comparison in order to avoid overflow
477 * or underflow in the comparison with it. This is simpler than the INRANGE
478 * check above, because we know that the dtms_scratch_ptr is valid in the
479 * range. Allocations of size zero are allowed.
481 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
482 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
483 (mstate)->dtms_scratch_ptr >= (alloc_sz))
485 #define DTRACE_LOADFUNC(bits) \
488 dtrace_load##bits(uintptr_t addr) \
490 size_t size = bits / NBBY; \
492 uint##bits##_t rval; \
494 volatile uint16_t *flags = (volatile uint16_t *) \
495 &cpu_core[curcpu].cpuc_dtrace_flags; \
497 DTRACE_ALIGNCHECK(addr, size, flags); \
499 for (i = 0; i < dtrace_toxranges; i++) { \
500 if (addr >= dtrace_toxrange[i].dtt_limit) \
503 if (addr + size <= dtrace_toxrange[i].dtt_base) \
507 * This address falls within a toxic region; return 0. \
509 *flags |= CPU_DTRACE_BADADDR; \
510 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
514 *flags |= CPU_DTRACE_NOFAULT; \
516 rval = *((volatile uint##bits##_t *)addr); \
517 *flags &= ~CPU_DTRACE_NOFAULT; \
519 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
523 #define dtrace_loadptr dtrace_load64
525 #define dtrace_loadptr dtrace_load32
528 #define DTRACE_DYNHASH_FREE 0
529 #define DTRACE_DYNHASH_SINK 1
530 #define DTRACE_DYNHASH_VALID 2
532 #define DTRACE_MATCH_NEXT 0
533 #define DTRACE_MATCH_DONE 1
534 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
535 #define DTRACE_STATE_ALIGN 64
537 #define DTRACE_FLAGS2FLT(flags) \
538 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
539 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
540 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
541 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
542 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
543 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
544 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
545 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
546 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
549 #define DTRACEACT_ISSTRING(act) \
550 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
551 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
553 /* Function prototype definitions: */
554 static size_t dtrace_strlen(const char *, size_t);
555 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
556 static void dtrace_enabling_provide(dtrace_provider_t *);
557 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
558 static void dtrace_enabling_matchall(void);
559 static void dtrace_enabling_reap(void);
560 static dtrace_state_t *dtrace_anon_grab(void);
561 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
562 dtrace_state_t *, uint64_t, uint64_t);
563 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
564 static void dtrace_buffer_drop(dtrace_buffer_t *);
565 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
566 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
567 dtrace_state_t *, dtrace_mstate_t *);
568 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
570 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
571 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
572 uint16_t dtrace_load16(uintptr_t);
573 uint32_t dtrace_load32(uintptr_t);
574 uint64_t dtrace_load64(uintptr_t);
575 uint8_t dtrace_load8(uintptr_t);
576 void dtrace_dynvar_clean(dtrace_dstate_t *);
577 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
578 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
579 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
582 * DTrace Probe Context Functions
584 * These functions are called from probe context. Because probe context is
585 * any context in which C may be called, arbitrarily locks may be held,
586 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
587 * As a result, functions called from probe context may only call other DTrace
588 * support functions -- they may not interact at all with the system at large.
589 * (Note that the ASSERT macro is made probe-context safe by redefining it in
590 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
591 * loads are to be performed from probe context, they _must_ be in terms of
592 * the safe dtrace_load*() variants.
594 * Some functions in this block are not actually called from probe context;
595 * for these functions, there will be a comment above the function reading
596 * "Note: not called from probe context."
599 dtrace_panic(const char *format, ...)
603 va_start(alist, format);
604 dtrace_vpanic(format, alist);
609 dtrace_assfail(const char *a, const char *f, int l)
611 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
614 * We just need something here that even the most clever compiler
615 * cannot optimize away.
617 return (a[(uintptr_t)f]);
621 * Atomically increment a specified error counter from probe context.
624 dtrace_error(uint32_t *counter)
627 * Most counters stored to in probe context are per-CPU counters.
628 * However, there are some error conditions that are sufficiently
629 * arcane that they don't merit per-CPU storage. If these counters
630 * are incremented concurrently on different CPUs, scalability will be
631 * adversely affected -- but we don't expect them to be white-hot in a
632 * correctly constructed enabling...
639 if ((nval = oval + 1) == 0) {
641 * If the counter would wrap, set it to 1 -- assuring
642 * that the counter is never zero when we have seen
643 * errors. (The counter must be 32-bits because we
644 * aren't guaranteed a 64-bit compare&swap operation.)
645 * To save this code both the infamy of being fingered
646 * by a priggish news story and the indignity of being
647 * the target of a neo-puritan witch trial, we're
648 * carefully avoiding any colorful description of the
649 * likelihood of this condition -- but suffice it to
650 * say that it is only slightly more likely than the
651 * overflow of predicate cache IDs, as discussed in
652 * dtrace_predicate_create().
656 } while (dtrace_cas32(counter, oval, nval) != oval);
660 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
661 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
669 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
671 if (dest < mstate->dtms_scratch_base)
674 if (dest + size < dest)
677 if (dest + size > mstate->dtms_scratch_ptr)
684 dtrace_canstore_statvar(uint64_t addr, size_t sz,
685 dtrace_statvar_t **svars, int nsvars)
689 for (i = 0; i < nsvars; i++) {
690 dtrace_statvar_t *svar = svars[i];
692 if (svar == NULL || svar->dtsv_size == 0)
695 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
703 * Check to see if the address is within a memory region to which a store may
704 * be issued. This includes the DTrace scratch areas, and any DTrace variable
705 * region. The caller of dtrace_canstore() is responsible for performing any
706 * alignment checks that are needed before stores are actually executed.
709 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
710 dtrace_vstate_t *vstate)
713 * First, check to see if the address is in scratch space...
715 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
716 mstate->dtms_scratch_size))
720 * Now check to see if it's a dynamic variable. This check will pick
721 * up both thread-local variables and any global dynamically-allocated
724 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
725 vstate->dtvs_dynvars.dtds_size)) {
726 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
727 uintptr_t base = (uintptr_t)dstate->dtds_base +
728 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
732 * Before we assume that we can store here, we need to make
733 * sure that it isn't in our metadata -- storing to our
734 * dynamic variable metadata would corrupt our state. For
735 * the range to not include any dynamic variable metadata,
738 * (1) Start above the hash table that is at the base of
739 * the dynamic variable space
741 * (2) Have a starting chunk offset that is beyond the
742 * dtrace_dynvar_t that is at the base of every chunk
744 * (3) Not span a chunk boundary
750 chunkoffs = (addr - base) % dstate->dtds_chunksize;
752 if (chunkoffs < sizeof (dtrace_dynvar_t))
755 if (chunkoffs + sz > dstate->dtds_chunksize)
762 * Finally, check the static local and global variables. These checks
763 * take the longest, so we perform them last.
765 if (dtrace_canstore_statvar(addr, sz,
766 vstate->dtvs_locals, vstate->dtvs_nlocals))
769 if (dtrace_canstore_statvar(addr, sz,
770 vstate->dtvs_globals, vstate->dtvs_nglobals))
778 * Convenience routine to check to see if the address is within a memory
779 * region in which a load may be issued given the user's privilege level;
780 * if not, it sets the appropriate error flags and loads 'addr' into the
781 * illegal value slot.
783 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
784 * appropriate memory access protection.
787 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
788 dtrace_vstate_t *vstate)
790 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
793 * If we hold the privilege to read from kernel memory, then
794 * everything is readable.
796 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
800 * You can obviously read that which you can store.
802 if (dtrace_canstore(addr, sz, mstate, vstate))
806 * We're allowed to read from our own string table.
808 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
809 mstate->dtms_difo->dtdo_strlen))
812 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
818 * Convenience routine to check to see if a given string is within a memory
819 * region in which a load may be issued given the user's privilege level;
820 * this exists so that we don't need to issue unnecessary dtrace_strlen()
821 * calls in the event that the user has all privileges.
824 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
825 dtrace_vstate_t *vstate)
830 * If we hold the privilege to read from kernel memory, then
831 * everything is readable.
833 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
836 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
837 if (dtrace_canload(addr, strsz, mstate, vstate))
844 * Convenience routine to check to see if a given variable is within a memory
845 * region in which a load may be issued given the user's privilege level.
848 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
849 dtrace_vstate_t *vstate)
852 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
855 * If we hold the privilege to read from kernel memory, then
856 * everything is readable.
858 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
861 if (type->dtdt_kind == DIF_TYPE_STRING)
862 sz = dtrace_strlen(src,
863 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
865 sz = type->dtdt_size;
867 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
871 * Compare two strings using safe loads.
874 dtrace_strncmp(char *s1, char *s2, size_t limit)
877 volatile uint16_t *flags;
879 if (s1 == s2 || limit == 0)
882 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
888 c1 = dtrace_load8((uintptr_t)s1++);
894 c2 = dtrace_load8((uintptr_t)s2++);
899 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
905 * Compute strlen(s) for a string using safe memory accesses. The additional
906 * len parameter is used to specify a maximum length to ensure completion.
909 dtrace_strlen(const char *s, size_t lim)
913 for (len = 0; len != lim; len++) {
914 if (dtrace_load8((uintptr_t)s++) == '\0')
922 * Check if an address falls within a toxic region.
925 dtrace_istoxic(uintptr_t kaddr, size_t size)
927 uintptr_t taddr, tsize;
930 for (i = 0; i < dtrace_toxranges; i++) {
931 taddr = dtrace_toxrange[i].dtt_base;
932 tsize = dtrace_toxrange[i].dtt_limit - taddr;
934 if (kaddr - taddr < tsize) {
935 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
936 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
940 if (taddr - kaddr < size) {
941 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
942 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
951 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
952 * memory specified by the DIF program. The dst is assumed to be safe memory
953 * that we can store to directly because it is managed by DTrace. As with
954 * standard bcopy, overlapping copies are handled properly.
957 dtrace_bcopy(const void *src, void *dst, size_t len)
961 const uint8_t *s2 = src;
965 *s1++ = dtrace_load8((uintptr_t)s2++);
966 } while (--len != 0);
972 *--s1 = dtrace_load8((uintptr_t)--s2);
973 } while (--len != 0);
979 * Copy src to dst using safe memory accesses, up to either the specified
980 * length, or the point that a nul byte is encountered. The src is assumed to
981 * be unsafe memory specified by the DIF program. The dst is assumed to be
982 * safe memory that we can store to directly because it is managed by DTrace.
983 * Unlike dtrace_bcopy(), overlapping regions are not handled.
986 dtrace_strcpy(const void *src, void *dst, size_t len)
989 uint8_t *s1 = dst, c;
990 const uint8_t *s2 = src;
993 *s1++ = c = dtrace_load8((uintptr_t)s2++);
994 } while (--len != 0 && c != '\0');
999 * Copy src to dst, deriving the size and type from the specified (BYREF)
1000 * variable type. The src is assumed to be unsafe memory specified by the DIF
1001 * program. The dst is assumed to be DTrace variable memory that is of the
1002 * specified type; we assume that we can store to directly.
1005 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1007 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1009 if (type->dtdt_kind == DIF_TYPE_STRING) {
1010 dtrace_strcpy(src, dst, type->dtdt_size);
1012 dtrace_bcopy(src, dst, type->dtdt_size);
1017 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1018 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1019 * safe memory that we can access directly because it is managed by DTrace.
1022 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1024 volatile uint16_t *flags;
1026 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1031 if (s1 == NULL || s2 == NULL)
1034 if (s1 != s2 && len != 0) {
1035 const uint8_t *ps1 = s1;
1036 const uint8_t *ps2 = s2;
1039 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1041 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1047 * Zero the specified region using a simple byte-by-byte loop. Note that this
1048 * is for safe DTrace-managed memory only.
1051 dtrace_bzero(void *dst, size_t len)
1055 for (cp = dst; len != 0; len--)
1060 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1064 result[0] = addend1[0] + addend2[0];
1065 result[1] = addend1[1] + addend2[1] +
1066 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1073 * Shift the 128-bit value in a by b. If b is positive, shift left.
1074 * If b is negative, shift right.
1077 dtrace_shift_128(uint64_t *a, int b)
1087 a[0] = a[1] >> (b - 64);
1091 mask = 1LL << (64 - b);
1093 a[0] |= ((a[1] & mask) << (64 - b));
1098 a[1] = a[0] << (b - 64);
1102 mask = a[0] >> (64 - b);
1110 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1111 * use native multiplication on those, and then re-combine into the
1112 * resulting 128-bit value.
1114 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1121 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1123 uint64_t hi1, hi2, lo1, lo2;
1126 hi1 = factor1 >> 32;
1127 hi2 = factor2 >> 32;
1129 lo1 = factor1 & DT_MASK_LO;
1130 lo2 = factor2 & DT_MASK_LO;
1132 product[0] = lo1 * lo2;
1133 product[1] = hi1 * hi2;
1137 dtrace_shift_128(tmp, 32);
1138 dtrace_add_128(product, tmp, product);
1142 dtrace_shift_128(tmp, 32);
1143 dtrace_add_128(product, tmp, product);
1147 * This privilege check should be used by actions and subroutines to
1148 * verify that the user credentials of the process that enabled the
1149 * invoking ECB match the target credentials
1152 dtrace_priv_proc_common_user(dtrace_state_t *state)
1154 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1157 * We should always have a non-NULL state cred here, since if cred
1158 * is null (anonymous tracing), we fast-path bypass this routine.
1160 ASSERT(s_cr != NULL);
1162 if ((cr = CRED()) != NULL &&
1163 s_cr->cr_uid == cr->cr_uid &&
1164 s_cr->cr_uid == cr->cr_ruid &&
1165 s_cr->cr_uid == cr->cr_suid &&
1166 s_cr->cr_gid == cr->cr_gid &&
1167 s_cr->cr_gid == cr->cr_rgid &&
1168 s_cr->cr_gid == cr->cr_sgid)
1175 * This privilege check should be used by actions and subroutines to
1176 * verify that the zone of the process that enabled the invoking ECB
1177 * matches the target credentials
1180 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1183 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1186 * We should always have a non-NULL state cred here, since if cred
1187 * is null (anonymous tracing), we fast-path bypass this routine.
1189 ASSERT(s_cr != NULL);
1191 if ((cr = CRED()) != NULL &&
1192 s_cr->cr_zone == cr->cr_zone)
1202 * This privilege check should be used by actions and subroutines to
1203 * verify that the process has not setuid or changed credentials.
1206 dtrace_priv_proc_common_nocd(void)
1210 if ((proc = ttoproc(curthread)) != NULL &&
1211 !(proc->p_flag & SNOCD))
1218 dtrace_priv_proc_destructive(dtrace_state_t *state)
1220 int action = state->dts_cred.dcr_action;
1222 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1223 dtrace_priv_proc_common_zone(state) == 0)
1226 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1227 dtrace_priv_proc_common_user(state) == 0)
1230 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1231 dtrace_priv_proc_common_nocd() == 0)
1237 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1243 dtrace_priv_proc_control(dtrace_state_t *state)
1245 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1248 if (dtrace_priv_proc_common_zone(state) &&
1249 dtrace_priv_proc_common_user(state) &&
1250 dtrace_priv_proc_common_nocd())
1253 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1259 dtrace_priv_proc(dtrace_state_t *state)
1261 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1264 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1270 dtrace_priv_kernel(dtrace_state_t *state)
1272 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1275 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1281 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1283 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1286 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1292 * Note: not called from probe context. This function is called
1293 * asynchronously (and at a regular interval) from outside of probe context to
1294 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1295 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1298 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1300 dtrace_dynvar_t *dirty;
1301 dtrace_dstate_percpu_t *dcpu;
1304 for (i = 0; i < NCPU; i++) {
1305 dcpu = &dstate->dtds_percpu[i];
1307 ASSERT(dcpu->dtdsc_rinsing == NULL);
1310 * If the dirty list is NULL, there is no dirty work to do.
1312 if (dcpu->dtdsc_dirty == NULL)
1316 * If the clean list is non-NULL, then we're not going to do
1317 * any work for this CPU -- it means that there has not been
1318 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1319 * since the last time we cleaned house.
1321 if (dcpu->dtdsc_clean != NULL)
1327 * Atomically move the dirty list aside.
1330 dirty = dcpu->dtdsc_dirty;
1333 * Before we zap the dirty list, set the rinsing list.
1334 * (This allows for a potential assertion in
1335 * dtrace_dynvar(): if a free dynamic variable appears
1336 * on a hash chain, either the dirty list or the
1337 * rinsing list for some CPU must be non-NULL.)
1339 dcpu->dtdsc_rinsing = dirty;
1340 dtrace_membar_producer();
1341 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1342 dirty, NULL) != dirty);
1347 * We have no work to do; we can simply return.
1354 for (i = 0; i < NCPU; i++) {
1355 dcpu = &dstate->dtds_percpu[i];
1357 if (dcpu->dtdsc_rinsing == NULL)
1361 * We are now guaranteed that no hash chain contains a pointer
1362 * into this dirty list; we can make it clean.
1364 ASSERT(dcpu->dtdsc_clean == NULL);
1365 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1366 dcpu->dtdsc_rinsing = NULL;
1370 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1371 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1372 * This prevents a race whereby a CPU incorrectly decides that
1373 * the state should be something other than DTRACE_DSTATE_CLEAN
1374 * after dtrace_dynvar_clean() has completed.
1378 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1382 * Depending on the value of the op parameter, this function looks-up,
1383 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1384 * allocation is requested, this function will return a pointer to a
1385 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1386 * variable can be allocated. If NULL is returned, the appropriate counter
1387 * will be incremented.
1390 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1391 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1392 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1394 uint64_t hashval = DTRACE_DYNHASH_VALID;
1395 dtrace_dynhash_t *hash = dstate->dtds_hash;
1396 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1397 processorid_t me = curcpu, cpu = me;
1398 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1399 size_t bucket, ksize;
1400 size_t chunksize = dstate->dtds_chunksize;
1401 uintptr_t kdata, lock, nstate;
1407 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1408 * algorithm. For the by-value portions, we perform the algorithm in
1409 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1410 * bit, and seems to have only a minute effect on distribution. For
1411 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1412 * over each referenced byte. It's painful to do this, but it's much
1413 * better than pathological hash distribution. The efficacy of the
1414 * hashing algorithm (and a comparison with other algorithms) may be
1415 * found by running the ::dtrace_dynstat MDB dcmd.
1417 for (i = 0; i < nkeys; i++) {
1418 if (key[i].dttk_size == 0) {
1419 uint64_t val = key[i].dttk_value;
1421 hashval += (val >> 48) & 0xffff;
1422 hashval += (hashval << 10);
1423 hashval ^= (hashval >> 6);
1425 hashval += (val >> 32) & 0xffff;
1426 hashval += (hashval << 10);
1427 hashval ^= (hashval >> 6);
1429 hashval += (val >> 16) & 0xffff;
1430 hashval += (hashval << 10);
1431 hashval ^= (hashval >> 6);
1433 hashval += val & 0xffff;
1434 hashval += (hashval << 10);
1435 hashval ^= (hashval >> 6);
1438 * This is incredibly painful, but it beats the hell
1439 * out of the alternative.
1441 uint64_t j, size = key[i].dttk_size;
1442 uintptr_t base = (uintptr_t)key[i].dttk_value;
1444 if (!dtrace_canload(base, size, mstate, vstate))
1447 for (j = 0; j < size; j++) {
1448 hashval += dtrace_load8(base + j);
1449 hashval += (hashval << 10);
1450 hashval ^= (hashval >> 6);
1455 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1458 hashval += (hashval << 3);
1459 hashval ^= (hashval >> 11);
1460 hashval += (hashval << 15);
1463 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1464 * comes out to be one of our two sentinel hash values. If this
1465 * actually happens, we set the hashval to be a value known to be a
1466 * non-sentinel value.
1468 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1469 hashval = DTRACE_DYNHASH_VALID;
1472 * Yes, it's painful to do a divide here. If the cycle count becomes
1473 * important here, tricks can be pulled to reduce it. (However, it's
1474 * critical that hash collisions be kept to an absolute minimum;
1475 * they're much more painful than a divide.) It's better to have a
1476 * solution that generates few collisions and still keeps things
1477 * relatively simple.
1479 bucket = hashval % dstate->dtds_hashsize;
1481 if (op == DTRACE_DYNVAR_DEALLOC) {
1482 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1485 while ((lock = *lockp) & 1)
1488 if (dtrace_casptr((volatile void *)lockp,
1489 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1493 dtrace_membar_producer();
1498 lock = hash[bucket].dtdh_lock;
1500 dtrace_membar_consumer();
1502 start = hash[bucket].dtdh_chain;
1503 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1504 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1505 op != DTRACE_DYNVAR_DEALLOC));
1507 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1508 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1509 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1511 if (dvar->dtdv_hashval != hashval) {
1512 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1514 * We've reached the sink, and therefore the
1515 * end of the hash chain; we can kick out of
1516 * the loop knowing that we have seen a valid
1517 * snapshot of state.
1519 ASSERT(dvar->dtdv_next == NULL);
1520 ASSERT(dvar == &dtrace_dynhash_sink);
1524 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1526 * We've gone off the rails: somewhere along
1527 * the line, one of the members of this hash
1528 * chain was deleted. Note that we could also
1529 * detect this by simply letting this loop run
1530 * to completion, as we would eventually hit
1531 * the end of the dirty list. However, we
1532 * want to avoid running the length of the
1533 * dirty list unnecessarily (it might be quite
1534 * long), so we catch this as early as
1535 * possible by detecting the hash marker. In
1536 * this case, we simply set dvar to NULL and
1537 * break; the conditional after the loop will
1538 * send us back to top.
1547 if (dtuple->dtt_nkeys != nkeys)
1550 for (i = 0; i < nkeys; i++, dkey++) {
1551 if (dkey->dttk_size != key[i].dttk_size)
1552 goto next; /* size or type mismatch */
1554 if (dkey->dttk_size != 0) {
1556 (void *)(uintptr_t)key[i].dttk_value,
1557 (void *)(uintptr_t)dkey->dttk_value,
1561 if (dkey->dttk_value != key[i].dttk_value)
1566 if (op != DTRACE_DYNVAR_DEALLOC)
1569 ASSERT(dvar->dtdv_next == NULL ||
1570 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1573 ASSERT(hash[bucket].dtdh_chain != dvar);
1574 ASSERT(start != dvar);
1575 ASSERT(prev->dtdv_next == dvar);
1576 prev->dtdv_next = dvar->dtdv_next;
1578 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1579 start, dvar->dtdv_next) != start) {
1581 * We have failed to atomically swing the
1582 * hash table head pointer, presumably because
1583 * of a conflicting allocation on another CPU.
1584 * We need to reread the hash chain and try
1591 dtrace_membar_producer();
1594 * Now set the hash value to indicate that it's free.
1596 ASSERT(hash[bucket].dtdh_chain != dvar);
1597 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1599 dtrace_membar_producer();
1602 * Set the next pointer to point at the dirty list, and
1603 * atomically swing the dirty pointer to the newly freed dvar.
1606 next = dcpu->dtdsc_dirty;
1607 dvar->dtdv_next = next;
1608 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1611 * Finally, unlock this hash bucket.
1613 ASSERT(hash[bucket].dtdh_lock == lock);
1615 hash[bucket].dtdh_lock++;
1625 * If dvar is NULL, it is because we went off the rails:
1626 * one of the elements that we traversed in the hash chain
1627 * was deleted while we were traversing it. In this case,
1628 * we assert that we aren't doing a dealloc (deallocs lock
1629 * the hash bucket to prevent themselves from racing with
1630 * one another), and retry the hash chain traversal.
1632 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1636 if (op != DTRACE_DYNVAR_ALLOC) {
1638 * If we are not to allocate a new variable, we want to
1639 * return NULL now. Before we return, check that the value
1640 * of the lock word hasn't changed. If it has, we may have
1641 * seen an inconsistent snapshot.
1643 if (op == DTRACE_DYNVAR_NOALLOC) {
1644 if (hash[bucket].dtdh_lock != lock)
1647 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1648 ASSERT(hash[bucket].dtdh_lock == lock);
1650 hash[bucket].dtdh_lock++;
1657 * We need to allocate a new dynamic variable. The size we need is the
1658 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1659 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1660 * the size of any referred-to data (dsize). We then round the final
1661 * size up to the chunksize for allocation.
1663 for (ksize = 0, i = 0; i < nkeys; i++)
1664 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1667 * This should be pretty much impossible, but could happen if, say,
1668 * strange DIF specified the tuple. Ideally, this should be an
1669 * assertion and not an error condition -- but that requires that the
1670 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1671 * bullet-proof. (That is, it must not be able to be fooled by
1672 * malicious DIF.) Given the lack of backwards branches in DIF,
1673 * solving this would presumably not amount to solving the Halting
1674 * Problem -- but it still seems awfully hard.
1676 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1677 ksize + dsize > chunksize) {
1678 dcpu->dtdsc_drops++;
1682 nstate = DTRACE_DSTATE_EMPTY;
1686 free = dcpu->dtdsc_free;
1689 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1692 if (clean == NULL) {
1694 * We're out of dynamic variable space on
1695 * this CPU. Unless we have tried all CPUs,
1696 * we'll try to allocate from a different
1699 switch (dstate->dtds_state) {
1700 case DTRACE_DSTATE_CLEAN: {
1701 void *sp = &dstate->dtds_state;
1706 if (dcpu->dtdsc_dirty != NULL &&
1707 nstate == DTRACE_DSTATE_EMPTY)
1708 nstate = DTRACE_DSTATE_DIRTY;
1710 if (dcpu->dtdsc_rinsing != NULL)
1711 nstate = DTRACE_DSTATE_RINSING;
1713 dcpu = &dstate->dtds_percpu[cpu];
1718 (void) dtrace_cas32(sp,
1719 DTRACE_DSTATE_CLEAN, nstate);
1722 * To increment the correct bean
1723 * counter, take another lap.
1728 case DTRACE_DSTATE_DIRTY:
1729 dcpu->dtdsc_dirty_drops++;
1732 case DTRACE_DSTATE_RINSING:
1733 dcpu->dtdsc_rinsing_drops++;
1736 case DTRACE_DSTATE_EMPTY:
1737 dcpu->dtdsc_drops++;
1741 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1746 * The clean list appears to be non-empty. We want to
1747 * move the clean list to the free list; we start by
1748 * moving the clean pointer aside.
1750 if (dtrace_casptr(&dcpu->dtdsc_clean,
1751 clean, NULL) != clean) {
1753 * We are in one of two situations:
1755 * (a) The clean list was switched to the
1756 * free list by another CPU.
1758 * (b) The clean list was added to by the
1761 * In either of these situations, we can
1762 * just reattempt the free list allocation.
1767 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1770 * Now we'll move the clean list to the free list.
1771 * It's impossible for this to fail: the only way
1772 * the free list can be updated is through this
1773 * code path, and only one CPU can own the clean list.
1774 * Thus, it would only be possible for this to fail if
1775 * this code were racing with dtrace_dynvar_clean().
1776 * (That is, if dtrace_dynvar_clean() updated the clean
1777 * list, and we ended up racing to update the free
1778 * list.) This race is prevented by the dtrace_sync()
1779 * in dtrace_dynvar_clean() -- which flushes the
1780 * owners of the clean lists out before resetting
1783 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1784 ASSERT(rval == NULL);
1789 new_free = dvar->dtdv_next;
1790 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1793 * We have now allocated a new chunk. We copy the tuple keys into the
1794 * tuple array and copy any referenced key data into the data space
1795 * following the tuple array. As we do this, we relocate dttk_value
1796 * in the final tuple to point to the key data address in the chunk.
1798 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1799 dvar->dtdv_data = (void *)(kdata + ksize);
1800 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1802 for (i = 0; i < nkeys; i++) {
1803 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1804 size_t kesize = key[i].dttk_size;
1808 (const void *)(uintptr_t)key[i].dttk_value,
1809 (void *)kdata, kesize);
1810 dkey->dttk_value = kdata;
1811 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1813 dkey->dttk_value = key[i].dttk_value;
1816 dkey->dttk_size = kesize;
1819 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1820 dvar->dtdv_hashval = hashval;
1821 dvar->dtdv_next = start;
1823 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1827 * The cas has failed. Either another CPU is adding an element to
1828 * this hash chain, or another CPU is deleting an element from this
1829 * hash chain. The simplest way to deal with both of these cases
1830 * (though not necessarily the most efficient) is to free our
1831 * allocated block and tail-call ourselves. Note that the free is
1832 * to the dirty list and _not_ to the free list. This is to prevent
1833 * races with allocators, above.
1835 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1837 dtrace_membar_producer();
1840 free = dcpu->dtdsc_dirty;
1841 dvar->dtdv_next = free;
1842 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1844 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1849 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1851 if ((int64_t)nval < (int64_t)*oval)
1857 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1859 if ((int64_t)nval > (int64_t)*oval)
1864 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1866 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1867 int64_t val = (int64_t)nval;
1870 for (i = 0; i < zero; i++) {
1871 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1877 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1878 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1879 quanta[i - 1] += incr;
1884 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1892 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1894 uint64_t arg = *lquanta++;
1895 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1896 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1897 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1898 int32_t val = (int32_t)nval, level;
1901 ASSERT(levels != 0);
1905 * This is an underflow.
1911 level = (val - base) / step;
1913 if (level < levels) {
1914 lquanta[level + 1] += incr;
1919 * This is an overflow.
1921 lquanta[levels + 1] += incr;
1925 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1926 uint16_t high, uint16_t nsteps, int64_t value)
1928 int64_t this = 1, last, next;
1929 int base = 1, order;
1931 ASSERT(factor <= nsteps);
1932 ASSERT(nsteps % factor == 0);
1934 for (order = 0; order < low; order++)
1938 * If our value is less than our factor taken to the power of the
1939 * low order of magnitude, it goes into the zeroth bucket.
1941 if (value < (last = this))
1944 for (this *= factor; order <= high; order++) {
1945 int nbuckets = this > nsteps ? nsteps : this;
1947 if ((next = this * factor) < this) {
1949 * We should not generally get log/linear quantizations
1950 * with a high magnitude that allows 64-bits to
1951 * overflow, but we nonetheless protect against this
1952 * by explicitly checking for overflow, and clamping
1953 * our value accordingly.
1960 * If our value lies within this order of magnitude,
1961 * determine its position by taking the offset within
1962 * the order of magnitude, dividing by the bucket
1963 * width, and adding to our (accumulated) base.
1965 return (base + (value - last) / (this / nbuckets));
1968 base += nbuckets - (nbuckets / factor);
1974 * Our value is greater than or equal to our factor taken to the
1975 * power of one plus the high magnitude -- return the top bucket.
1981 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
1983 uint64_t arg = *llquanta++;
1984 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
1985 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
1986 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
1987 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
1989 llquanta[dtrace_aggregate_llquantize_bucket(factor,
1990 low, high, nsteps, nval)] += incr;
1995 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2003 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2005 int64_t snval = (int64_t)nval;
2012 * What we want to say here is:
2014 * data[2] += nval * nval;
2016 * But given that nval is 64-bit, we could easily overflow, so
2017 * we do this as 128-bit arithmetic.
2022 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2023 dtrace_add_128(data + 2, tmp, data + 2);
2028 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2035 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2041 * Aggregate given the tuple in the principal data buffer, and the aggregating
2042 * action denoted by the specified dtrace_aggregation_t. The aggregation
2043 * buffer is specified as the buf parameter. This routine does not return
2044 * failure; if there is no space in the aggregation buffer, the data will be
2045 * dropped, and a corresponding counter incremented.
2048 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2049 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2051 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2052 uint32_t i, ndx, size, fsize;
2053 uint32_t align = sizeof (uint64_t) - 1;
2054 dtrace_aggbuffer_t *agb;
2055 dtrace_aggkey_t *key;
2056 uint32_t hashval = 0, limit, isstr;
2057 caddr_t tomax, data, kdata;
2058 dtrace_actkind_t action;
2059 dtrace_action_t *act;
2065 if (!agg->dtag_hasarg) {
2067 * Currently, only quantize() and lquantize() take additional
2068 * arguments, and they have the same semantics: an increment
2069 * value that defaults to 1 when not present. If additional
2070 * aggregating actions take arguments, the setting of the
2071 * default argument value will presumably have to become more
2077 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2078 size = rec->dtrd_offset - agg->dtag_base;
2079 fsize = size + rec->dtrd_size;
2081 ASSERT(dbuf->dtb_tomax != NULL);
2082 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2084 if ((tomax = buf->dtb_tomax) == NULL) {
2085 dtrace_buffer_drop(buf);
2090 * The metastructure is always at the bottom of the buffer.
2092 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2093 sizeof (dtrace_aggbuffer_t));
2095 if (buf->dtb_offset == 0) {
2097 * We just kludge up approximately 1/8th of the size to be
2098 * buckets. If this guess ends up being routinely
2099 * off-the-mark, we may need to dynamically readjust this
2100 * based on past performance.
2102 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2104 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2105 (uintptr_t)tomax || hashsize == 0) {
2107 * We've been given a ludicrously small buffer;
2108 * increment our drop count and leave.
2110 dtrace_buffer_drop(buf);
2115 * And now, a pathetic attempt to try to get a an odd (or
2116 * perchance, a prime) hash size for better hash distribution.
2118 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2119 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2121 agb->dtagb_hashsize = hashsize;
2122 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2123 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2124 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2126 for (i = 0; i < agb->dtagb_hashsize; i++)
2127 agb->dtagb_hash[i] = NULL;
2130 ASSERT(agg->dtag_first != NULL);
2131 ASSERT(agg->dtag_first->dta_intuple);
2134 * Calculate the hash value based on the key. Note that we _don't_
2135 * include the aggid in the hashing (but we will store it as part of
2136 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2137 * algorithm: a simple, quick algorithm that has no known funnels, and
2138 * gets good distribution in practice. The efficacy of the hashing
2139 * algorithm (and a comparison with other algorithms) may be found by
2140 * running the ::dtrace_aggstat MDB dcmd.
2142 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2143 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2144 limit = i + act->dta_rec.dtrd_size;
2145 ASSERT(limit <= size);
2146 isstr = DTRACEACT_ISSTRING(act);
2148 for (; i < limit; i++) {
2150 hashval += (hashval << 10);
2151 hashval ^= (hashval >> 6);
2153 if (isstr && data[i] == '\0')
2158 hashval += (hashval << 3);
2159 hashval ^= (hashval >> 11);
2160 hashval += (hashval << 15);
2163 * Yes, the divide here is expensive -- but it's generally the least
2164 * of the performance issues given the amount of data that we iterate
2165 * over to compute hash values, compare data, etc.
2167 ndx = hashval % agb->dtagb_hashsize;
2169 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2170 ASSERT((caddr_t)key >= tomax);
2171 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2173 if (hashval != key->dtak_hashval || key->dtak_size != size)
2176 kdata = key->dtak_data;
2177 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2179 for (act = agg->dtag_first; act->dta_intuple;
2180 act = act->dta_next) {
2181 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2182 limit = i + act->dta_rec.dtrd_size;
2183 ASSERT(limit <= size);
2184 isstr = DTRACEACT_ISSTRING(act);
2186 for (; i < limit; i++) {
2187 if (kdata[i] != data[i])
2190 if (isstr && data[i] == '\0')
2195 if (action != key->dtak_action) {
2197 * We are aggregating on the same value in the same
2198 * aggregation with two different aggregating actions.
2199 * (This should have been picked up in the compiler,
2200 * so we may be dealing with errant or devious DIF.)
2201 * This is an error condition; we indicate as much,
2204 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2209 * This is a hit: we need to apply the aggregator to
2210 * the value at this key.
2212 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2219 * We didn't find it. We need to allocate some zero-filled space,
2220 * link it into the hash table appropriately, and apply the aggregator
2221 * to the (zero-filled) value.
2223 offs = buf->dtb_offset;
2224 while (offs & (align - 1))
2225 offs += sizeof (uint32_t);
2228 * If we don't have enough room to both allocate a new key _and_
2229 * its associated data, increment the drop count and return.
2231 if ((uintptr_t)tomax + offs + fsize >
2232 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2233 dtrace_buffer_drop(buf);
2238 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2239 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2240 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2242 key->dtak_data = kdata = tomax + offs;
2243 buf->dtb_offset = offs + fsize;
2246 * Now copy the data across.
2248 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2250 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2254 * Because strings are not zeroed out by default, we need to iterate
2255 * looking for actions that store strings, and we need to explicitly
2256 * pad these strings out with zeroes.
2258 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2261 if (!DTRACEACT_ISSTRING(act))
2264 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2265 limit = i + act->dta_rec.dtrd_size;
2266 ASSERT(limit <= size);
2268 for (nul = 0; i < limit; i++) {
2274 if (data[i] != '\0')
2281 for (i = size; i < fsize; i++)
2284 key->dtak_hashval = hashval;
2285 key->dtak_size = size;
2286 key->dtak_action = action;
2287 key->dtak_next = agb->dtagb_hash[ndx];
2288 agb->dtagb_hash[ndx] = key;
2291 * Finally, apply the aggregator.
2293 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2294 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2298 * Given consumer state, this routine finds a speculation in the INACTIVE
2299 * state and transitions it into the ACTIVE state. If there is no speculation
2300 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2301 * incremented -- it is up to the caller to take appropriate action.
2304 dtrace_speculation(dtrace_state_t *state)
2307 dtrace_speculation_state_t current;
2308 uint32_t *stat = &state->dts_speculations_unavail, count;
2310 while (i < state->dts_nspeculations) {
2311 dtrace_speculation_t *spec = &state->dts_speculations[i];
2313 current = spec->dtsp_state;
2315 if (current != DTRACESPEC_INACTIVE) {
2316 if (current == DTRACESPEC_COMMITTINGMANY ||
2317 current == DTRACESPEC_COMMITTING ||
2318 current == DTRACESPEC_DISCARDING)
2319 stat = &state->dts_speculations_busy;
2324 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2325 current, DTRACESPEC_ACTIVE) == current)
2330 * We couldn't find a speculation. If we found as much as a single
2331 * busy speculation buffer, we'll attribute this failure as "busy"
2332 * instead of "unavail".
2336 } while (dtrace_cas32(stat, count, count + 1) != count);
2342 * This routine commits an active speculation. If the specified speculation
2343 * is not in a valid state to perform a commit(), this routine will silently do
2344 * nothing. The state of the specified speculation is transitioned according
2345 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2348 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2349 dtrace_specid_t which)
2351 dtrace_speculation_t *spec;
2352 dtrace_buffer_t *src, *dest;
2353 uintptr_t daddr, saddr, dlimit, slimit;
2354 dtrace_speculation_state_t current, new = 0;
2361 if (which > state->dts_nspeculations) {
2362 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2366 spec = &state->dts_speculations[which - 1];
2367 src = &spec->dtsp_buffer[cpu];
2368 dest = &state->dts_buffer[cpu];
2371 current = spec->dtsp_state;
2373 if (current == DTRACESPEC_COMMITTINGMANY)
2377 case DTRACESPEC_INACTIVE:
2378 case DTRACESPEC_DISCARDING:
2381 case DTRACESPEC_COMMITTING:
2383 * This is only possible if we are (a) commit()'ing
2384 * without having done a prior speculate() on this CPU
2385 * and (b) racing with another commit() on a different
2386 * CPU. There's nothing to do -- we just assert that
2389 ASSERT(src->dtb_offset == 0);
2392 case DTRACESPEC_ACTIVE:
2393 new = DTRACESPEC_COMMITTING;
2396 case DTRACESPEC_ACTIVEONE:
2398 * This speculation is active on one CPU. If our
2399 * buffer offset is non-zero, we know that the one CPU
2400 * must be us. Otherwise, we are committing on a
2401 * different CPU from the speculate(), and we must
2402 * rely on being asynchronously cleaned.
2404 if (src->dtb_offset != 0) {
2405 new = DTRACESPEC_COMMITTING;
2410 case DTRACESPEC_ACTIVEMANY:
2411 new = DTRACESPEC_COMMITTINGMANY;
2417 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2418 current, new) != current);
2421 * We have set the state to indicate that we are committing this
2422 * speculation. Now reserve the necessary space in the destination
2425 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2426 sizeof (uint64_t), state, NULL)) < 0) {
2427 dtrace_buffer_drop(dest);
2432 * We have sufficient space to copy the speculative buffer into the
2433 * primary buffer. First, modify the speculative buffer, filling
2434 * in the timestamp of all entries with the current time. The data
2435 * must have the commit() time rather than the time it was traced,
2436 * so that all entries in the primary buffer are in timestamp order.
2438 timestamp = dtrace_gethrtime();
2439 saddr = (uintptr_t)src->dtb_tomax;
2440 slimit = saddr + src->dtb_offset;
2441 while (saddr < slimit) {
2443 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2445 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2446 saddr += sizeof (dtrace_epid_t);
2449 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2450 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2452 ASSERT3U(saddr + size, <=, slimit);
2453 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2454 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2456 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2462 * Copy the buffer across. (Note that this is a
2463 * highly subobtimal bcopy(); in the unlikely event that this becomes
2464 * a serious performance issue, a high-performance DTrace-specific
2465 * bcopy() should obviously be invented.)
2467 daddr = (uintptr_t)dest->dtb_tomax + offs;
2468 dlimit = daddr + src->dtb_offset;
2469 saddr = (uintptr_t)src->dtb_tomax;
2472 * First, the aligned portion.
2474 while (dlimit - daddr >= sizeof (uint64_t)) {
2475 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2477 daddr += sizeof (uint64_t);
2478 saddr += sizeof (uint64_t);
2482 * Now any left-over bit...
2484 while (dlimit - daddr)
2485 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2488 * Finally, commit the reserved space in the destination buffer.
2490 dest->dtb_offset = offs + src->dtb_offset;
2494 * If we're lucky enough to be the only active CPU on this speculation
2495 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2497 if (current == DTRACESPEC_ACTIVE ||
2498 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2499 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2500 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2502 ASSERT(rval == DTRACESPEC_COMMITTING);
2505 src->dtb_offset = 0;
2506 src->dtb_xamot_drops += src->dtb_drops;
2511 * This routine discards an active speculation. If the specified speculation
2512 * is not in a valid state to perform a discard(), this routine will silently
2513 * do nothing. The state of the specified speculation is transitioned
2514 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2517 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2518 dtrace_specid_t which)
2520 dtrace_speculation_t *spec;
2521 dtrace_speculation_state_t current, new = 0;
2522 dtrace_buffer_t *buf;
2527 if (which > state->dts_nspeculations) {
2528 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2532 spec = &state->dts_speculations[which - 1];
2533 buf = &spec->dtsp_buffer[cpu];
2536 current = spec->dtsp_state;
2539 case DTRACESPEC_INACTIVE:
2540 case DTRACESPEC_COMMITTINGMANY:
2541 case DTRACESPEC_COMMITTING:
2542 case DTRACESPEC_DISCARDING:
2545 case DTRACESPEC_ACTIVE:
2546 case DTRACESPEC_ACTIVEMANY:
2547 new = DTRACESPEC_DISCARDING;
2550 case DTRACESPEC_ACTIVEONE:
2551 if (buf->dtb_offset != 0) {
2552 new = DTRACESPEC_INACTIVE;
2554 new = DTRACESPEC_DISCARDING;
2561 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2562 current, new) != current);
2564 buf->dtb_offset = 0;
2569 * Note: not called from probe context. This function is called
2570 * asynchronously from cross call context to clean any speculations that are
2571 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2572 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2576 dtrace_speculation_clean_here(dtrace_state_t *state)
2578 dtrace_icookie_t cookie;
2579 processorid_t cpu = curcpu;
2580 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2583 cookie = dtrace_interrupt_disable();
2585 if (dest->dtb_tomax == NULL) {
2586 dtrace_interrupt_enable(cookie);
2590 for (i = 0; i < state->dts_nspeculations; i++) {
2591 dtrace_speculation_t *spec = &state->dts_speculations[i];
2592 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2594 if (src->dtb_tomax == NULL)
2597 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2598 src->dtb_offset = 0;
2602 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2605 if (src->dtb_offset == 0)
2608 dtrace_speculation_commit(state, cpu, i + 1);
2611 dtrace_interrupt_enable(cookie);
2615 * Note: not called from probe context. This function is called
2616 * asynchronously (and at a regular interval) to clean any speculations that
2617 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2618 * is work to be done, it cross calls all CPUs to perform that work;
2619 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2620 * INACTIVE state until they have been cleaned by all CPUs.
2623 dtrace_speculation_clean(dtrace_state_t *state)
2628 for (i = 0; i < state->dts_nspeculations; i++) {
2629 dtrace_speculation_t *spec = &state->dts_speculations[i];
2631 ASSERT(!spec->dtsp_cleaning);
2633 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2634 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2638 spec->dtsp_cleaning = 1;
2644 dtrace_xcall(DTRACE_CPUALL,
2645 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2648 * We now know that all CPUs have committed or discarded their
2649 * speculation buffers, as appropriate. We can now set the state
2652 for (i = 0; i < state->dts_nspeculations; i++) {
2653 dtrace_speculation_t *spec = &state->dts_speculations[i];
2654 dtrace_speculation_state_t current, new;
2656 if (!spec->dtsp_cleaning)
2659 current = spec->dtsp_state;
2660 ASSERT(current == DTRACESPEC_DISCARDING ||
2661 current == DTRACESPEC_COMMITTINGMANY);
2663 new = DTRACESPEC_INACTIVE;
2665 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2666 ASSERT(rv == current);
2667 spec->dtsp_cleaning = 0;
2672 * Called as part of a speculate() to get the speculative buffer associated
2673 * with a given speculation. Returns NULL if the specified speculation is not
2674 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2675 * the active CPU is not the specified CPU -- the speculation will be
2676 * atomically transitioned into the ACTIVEMANY state.
2678 static dtrace_buffer_t *
2679 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2680 dtrace_specid_t which)
2682 dtrace_speculation_t *spec;
2683 dtrace_speculation_state_t current, new = 0;
2684 dtrace_buffer_t *buf;
2689 if (which > state->dts_nspeculations) {
2690 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2694 spec = &state->dts_speculations[which - 1];
2695 buf = &spec->dtsp_buffer[cpuid];
2698 current = spec->dtsp_state;
2701 case DTRACESPEC_INACTIVE:
2702 case DTRACESPEC_COMMITTINGMANY:
2703 case DTRACESPEC_DISCARDING:
2706 case DTRACESPEC_COMMITTING:
2707 ASSERT(buf->dtb_offset == 0);
2710 case DTRACESPEC_ACTIVEONE:
2712 * This speculation is currently active on one CPU.
2713 * Check the offset in the buffer; if it's non-zero,
2714 * that CPU must be us (and we leave the state alone).
2715 * If it's zero, assume that we're starting on a new
2716 * CPU -- and change the state to indicate that the
2717 * speculation is active on more than one CPU.
2719 if (buf->dtb_offset != 0)
2722 new = DTRACESPEC_ACTIVEMANY;
2725 case DTRACESPEC_ACTIVEMANY:
2728 case DTRACESPEC_ACTIVE:
2729 new = DTRACESPEC_ACTIVEONE;
2735 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2736 current, new) != current);
2738 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2743 * Return a string. In the event that the user lacks the privilege to access
2744 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2745 * don't fail access checking.
2747 * dtrace_dif_variable() uses this routine as a helper for various
2748 * builtin values such as 'execname' and 'probefunc.'
2751 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2752 dtrace_mstate_t *mstate)
2754 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2759 * The easy case: this probe is allowed to read all of memory, so
2760 * we can just return this as a vanilla pointer.
2762 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2766 * This is the tougher case: we copy the string in question from
2767 * kernel memory into scratch memory and return it that way: this
2768 * ensures that we won't trip up when access checking tests the
2769 * BYREF return value.
2771 strsz = dtrace_strlen((char *)addr, size) + 1;
2773 if (mstate->dtms_scratch_ptr + strsz >
2774 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2775 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2779 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2781 ret = mstate->dtms_scratch_ptr;
2782 mstate->dtms_scratch_ptr += strsz;
2787 * Return a string from a memoy address which is known to have one or
2788 * more concatenated, individually zero terminated, sub-strings.
2789 * In the event that the user lacks the privilege to access
2790 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2791 * don't fail access checking.
2793 * dtrace_dif_variable() uses this routine as a helper for various
2794 * builtin values such as 'execargs'.
2797 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2798 dtrace_mstate_t *mstate)
2804 if (mstate->dtms_scratch_ptr + strsz >
2805 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2806 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2810 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2813 /* Replace sub-string termination characters with a space. */
2814 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2819 ret = mstate->dtms_scratch_ptr;
2820 mstate->dtms_scratch_ptr += strsz;
2825 * This function implements the DIF emulator's variable lookups. The emulator
2826 * passes a reserved variable identifier and optional built-in array index.
2829 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2833 * If we're accessing one of the uncached arguments, we'll turn this
2834 * into a reference in the args array.
2836 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2837 ndx = v - DIF_VAR_ARG0;
2843 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2844 if (ndx >= sizeof (mstate->dtms_arg) /
2845 sizeof (mstate->dtms_arg[0])) {
2846 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2847 dtrace_provider_t *pv;
2850 pv = mstate->dtms_probe->dtpr_provider;
2851 if (pv->dtpv_pops.dtps_getargval != NULL)
2852 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2853 mstate->dtms_probe->dtpr_id,
2854 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2856 val = dtrace_getarg(ndx, aframes);
2859 * This is regrettably required to keep the compiler
2860 * from tail-optimizing the call to dtrace_getarg().
2861 * The condition always evaluates to true, but the
2862 * compiler has no way of figuring that out a priori.
2863 * (None of this would be necessary if the compiler
2864 * could be relied upon to _always_ tail-optimize
2865 * the call to dtrace_getarg() -- but it can't.)
2867 if (mstate->dtms_probe != NULL)
2873 return (mstate->dtms_arg[ndx]);
2876 case DIF_VAR_UREGS: {
2879 if (!dtrace_priv_proc(state))
2882 if ((lwp = curthread->t_lwp) == NULL) {
2883 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2884 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2888 return (dtrace_getreg(lwp->lwp_regs, ndx));
2892 case DIF_VAR_UREGS: {
2893 struct trapframe *tframe;
2895 if (!dtrace_priv_proc(state))
2898 if ((tframe = curthread->td_frame) == NULL) {
2899 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2900 cpu_core[curcpu].cpuc_dtrace_illval = 0;
2904 return (dtrace_getreg(tframe, ndx));
2908 case DIF_VAR_CURTHREAD:
2909 if (!dtrace_priv_kernel(state))
2911 return ((uint64_t)(uintptr_t)curthread);
2913 case DIF_VAR_TIMESTAMP:
2914 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2915 mstate->dtms_timestamp = dtrace_gethrtime();
2916 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2918 return (mstate->dtms_timestamp);
2920 case DIF_VAR_VTIMESTAMP:
2921 ASSERT(dtrace_vtime_references != 0);
2922 return (curthread->t_dtrace_vtime);
2924 case DIF_VAR_WALLTIMESTAMP:
2925 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2926 mstate->dtms_walltimestamp = dtrace_gethrestime();
2927 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2929 return (mstate->dtms_walltimestamp);
2933 if (!dtrace_priv_kernel(state))
2935 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2936 mstate->dtms_ipl = dtrace_getipl();
2937 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2939 return (mstate->dtms_ipl);
2943 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2944 return (mstate->dtms_epid);
2947 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2948 return (mstate->dtms_probe->dtpr_id);
2950 case DIF_VAR_STACKDEPTH:
2951 if (!dtrace_priv_kernel(state))
2953 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2954 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2956 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2957 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2959 return (mstate->dtms_stackdepth);
2961 case DIF_VAR_USTACKDEPTH:
2962 if (!dtrace_priv_proc(state))
2964 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2966 * See comment in DIF_VAR_PID.
2968 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2970 mstate->dtms_ustackdepth = 0;
2972 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2973 mstate->dtms_ustackdepth =
2974 dtrace_getustackdepth();
2975 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2977 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2979 return (mstate->dtms_ustackdepth);
2981 case DIF_VAR_CALLER:
2982 if (!dtrace_priv_kernel(state))
2984 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2985 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2987 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2989 * If this is an unanchored probe, we are
2990 * required to go through the slow path:
2991 * dtrace_caller() only guarantees correct
2992 * results for anchored probes.
2994 pc_t caller[2] = {0, 0};
2996 dtrace_getpcstack(caller, 2, aframes,
2997 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2998 mstate->dtms_caller = caller[1];
2999 } else if ((mstate->dtms_caller =
3000 dtrace_caller(aframes)) == -1) {
3002 * We have failed to do this the quick way;
3003 * we must resort to the slower approach of
3004 * calling dtrace_getpcstack().
3008 dtrace_getpcstack(&caller, 1, aframes, NULL);
3009 mstate->dtms_caller = caller;
3012 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3014 return (mstate->dtms_caller);
3016 case DIF_VAR_UCALLER:
3017 if (!dtrace_priv_proc(state))
3020 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3024 * dtrace_getupcstack() fills in the first uint64_t
3025 * with the current PID. The second uint64_t will
3026 * be the program counter at user-level. The third
3027 * uint64_t will contain the caller, which is what
3031 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3032 dtrace_getupcstack(ustack, 3);
3033 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3034 mstate->dtms_ucaller = ustack[2];
3035 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3038 return (mstate->dtms_ucaller);
3040 case DIF_VAR_PROBEPROV:
3041 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3042 return (dtrace_dif_varstr(
3043 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3046 case DIF_VAR_PROBEMOD:
3047 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3048 return (dtrace_dif_varstr(
3049 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3052 case DIF_VAR_PROBEFUNC:
3053 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3054 return (dtrace_dif_varstr(
3055 (uintptr_t)mstate->dtms_probe->dtpr_func,
3058 case DIF_VAR_PROBENAME:
3059 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3060 return (dtrace_dif_varstr(
3061 (uintptr_t)mstate->dtms_probe->dtpr_name,
3065 if (!dtrace_priv_proc(state))
3070 * Note that we are assuming that an unanchored probe is
3071 * always due to a high-level interrupt. (And we're assuming
3072 * that there is only a single high level interrupt.)
3074 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3075 return (pid0.pid_id);
3078 * It is always safe to dereference one's own t_procp pointer:
3079 * it always points to a valid, allocated proc structure.
3080 * Further, it is always safe to dereference the p_pidp member
3081 * of one's own proc structure. (These are truisms becuase
3082 * threads and processes don't clean up their own state --
3083 * they leave that task to whomever reaps them.)
3085 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3087 return ((uint64_t)curproc->p_pid);
3091 if (!dtrace_priv_proc(state))
3096 * See comment in DIF_VAR_PID.
3098 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3099 return (pid0.pid_id);
3102 * It is always safe to dereference one's own t_procp pointer:
3103 * it always points to a valid, allocated proc structure.
3104 * (This is true because threads don't clean up their own
3105 * state -- they leave that task to whomever reaps them.)
3107 return ((uint64_t)curthread->t_procp->p_ppid);
3109 return ((uint64_t)curproc->p_pptr->p_pid);
3115 * See comment in DIF_VAR_PID.
3117 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3121 return ((uint64_t)curthread->t_tid);
3123 case DIF_VAR_EXECARGS: {
3124 struct pargs *p_args = curthread->td_proc->p_args;
3129 return (dtrace_dif_varstrz(
3130 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3133 case DIF_VAR_EXECNAME:
3135 if (!dtrace_priv_proc(state))
3139 * See comment in DIF_VAR_PID.
3141 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3142 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3145 * It is always safe to dereference one's own t_procp pointer:
3146 * it always points to a valid, allocated proc structure.
3147 * (This is true because threads don't clean up their own
3148 * state -- they leave that task to whomever reaps them.)
3150 return (dtrace_dif_varstr(
3151 (uintptr_t)curthread->t_procp->p_user.u_comm,
3154 return (dtrace_dif_varstr(
3155 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3158 case DIF_VAR_ZONENAME:
3160 if (!dtrace_priv_proc(state))
3164 * See comment in DIF_VAR_PID.
3166 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3167 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3170 * It is always safe to dereference one's own t_procp pointer:
3171 * it always points to a valid, allocated proc structure.
3172 * (This is true because threads don't clean up their own
3173 * state -- they leave that task to whomever reaps them.)
3175 return (dtrace_dif_varstr(
3176 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3183 if (!dtrace_priv_proc(state))
3188 * See comment in DIF_VAR_PID.
3190 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3191 return ((uint64_t)p0.p_cred->cr_uid);
3195 * It is always safe to dereference one's own t_procp pointer:
3196 * it always points to a valid, allocated proc structure.
3197 * (This is true because threads don't clean up their own
3198 * state -- they leave that task to whomever reaps them.)
3200 * Additionally, it is safe to dereference one's own process
3201 * credential, since this is never NULL after process birth.
3203 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3206 if (!dtrace_priv_proc(state))
3211 * See comment in DIF_VAR_PID.
3213 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3214 return ((uint64_t)p0.p_cred->cr_gid);
3218 * It is always safe to dereference one's own t_procp pointer:
3219 * it always points to a valid, allocated proc structure.
3220 * (This is true because threads don't clean up their own
3221 * state -- they leave that task to whomever reaps them.)
3223 * Additionally, it is safe to dereference one's own process
3224 * credential, since this is never NULL after process birth.
3226 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3228 case DIF_VAR_ERRNO: {
3231 if (!dtrace_priv_proc(state))
3235 * See comment in DIF_VAR_PID.
3237 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3241 * It is always safe to dereference one's own t_lwp pointer in
3242 * the event that this pointer is non-NULL. (This is true
3243 * because threads and lwps don't clean up their own state --
3244 * they leave that task to whomever reaps them.)
3246 if ((lwp = curthread->t_lwp) == NULL)
3249 return ((uint64_t)lwp->lwp_errno);
3251 return (curthread->td_errno);
3260 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3266 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3267 * Notice that we don't bother validating the proper number of arguments or
3268 * their types in the tuple stack. This isn't needed because all argument
3269 * interpretation is safe because of our load safety -- the worst that can
3270 * happen is that a bogus program can obtain bogus results.
3273 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3274 dtrace_key_t *tupregs, int nargs,
3275 dtrace_mstate_t *mstate, dtrace_state_t *state)
3277 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3278 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3279 dtrace_vstate_t *vstate = &state->dts_vstate;
3292 struct thread *lowner;
3294 struct lock_object *li;
3301 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3305 case DIF_SUBR_MUTEX_OWNED:
3306 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3312 m.mx = dtrace_load64(tupregs[0].dttk_value);
3313 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3314 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3316 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3319 case DIF_SUBR_MUTEX_OWNER:
3320 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3326 m.mx = dtrace_load64(tupregs[0].dttk_value);
3327 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3328 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3329 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3334 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3335 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3341 m.mx = dtrace_load64(tupregs[0].dttk_value);
3342 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3345 case DIF_SUBR_MUTEX_TYPE_SPIN:
3346 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3352 m.mx = dtrace_load64(tupregs[0].dttk_value);
3353 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3356 case DIF_SUBR_RW_READ_HELD: {
3359 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3365 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3366 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3370 case DIF_SUBR_RW_WRITE_HELD:
3371 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3377 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3378 regs[rd] = _RW_WRITE_HELD(&r.ri);
3381 case DIF_SUBR_RW_ISWRITER:
3382 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3388 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3389 regs[rd] = _RW_ISWRITER(&r.ri);
3393 case DIF_SUBR_MUTEX_OWNED:
3394 if (!dtrace_canload(tupregs[0].dttk_value,
3395 sizeof (struct lock_object), mstate, vstate)) {
3399 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3400 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3403 case DIF_SUBR_MUTEX_OWNER:
3404 if (!dtrace_canload(tupregs[0].dttk_value,
3405 sizeof (struct lock_object), mstate, vstate)) {
3409 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3410 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3411 regs[rd] = (uintptr_t)lowner;
3414 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3415 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3420 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3421 /* XXX - should be only LC_SLEEPABLE? */
3422 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3423 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3426 case DIF_SUBR_MUTEX_TYPE_SPIN:
3427 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3432 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3433 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3436 case DIF_SUBR_RW_READ_HELD:
3437 case DIF_SUBR_SX_SHARED_HELD:
3438 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3443 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3444 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3448 case DIF_SUBR_RW_WRITE_HELD:
3449 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3450 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3455 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3456 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3457 regs[rd] = (lowner == curthread);
3460 case DIF_SUBR_RW_ISWRITER:
3461 case DIF_SUBR_SX_ISEXCLUSIVE:
3462 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3467 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3468 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3471 #endif /* ! defined(sun) */
3473 case DIF_SUBR_BCOPY: {
3475 * We need to be sure that the destination is in the scratch
3476 * region -- no other region is allowed.
3478 uintptr_t src = tupregs[0].dttk_value;
3479 uintptr_t dest = tupregs[1].dttk_value;
3480 size_t size = tupregs[2].dttk_value;
3482 if (!dtrace_inscratch(dest, size, mstate)) {
3483 *flags |= CPU_DTRACE_BADADDR;
3488 if (!dtrace_canload(src, size, mstate, vstate)) {
3493 dtrace_bcopy((void *)src, (void *)dest, size);
3497 case DIF_SUBR_ALLOCA:
3498 case DIF_SUBR_COPYIN: {
3499 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3501 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3502 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3505 * This action doesn't require any credential checks since
3506 * probes will not activate in user contexts to which the
3507 * enabling user does not have permissions.
3511 * Rounding up the user allocation size could have overflowed
3512 * a large, bogus allocation (like -1ULL) to 0.
3514 if (scratch_size < size ||
3515 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3516 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3521 if (subr == DIF_SUBR_COPYIN) {
3522 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3523 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3524 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3527 mstate->dtms_scratch_ptr += scratch_size;
3532 case DIF_SUBR_COPYINTO: {
3533 uint64_t size = tupregs[1].dttk_value;
3534 uintptr_t dest = tupregs[2].dttk_value;
3537 * This action doesn't require any credential checks since
3538 * probes will not activate in user contexts to which the
3539 * enabling user does not have permissions.
3541 if (!dtrace_inscratch(dest, size, mstate)) {
3542 *flags |= CPU_DTRACE_BADADDR;
3547 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3548 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3549 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3553 case DIF_SUBR_COPYINSTR: {
3554 uintptr_t dest = mstate->dtms_scratch_ptr;
3555 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3557 if (nargs > 1 && tupregs[1].dttk_value < size)
3558 size = tupregs[1].dttk_value + 1;
3561 * This action doesn't require any credential checks since
3562 * probes will not activate in user contexts to which the
3563 * enabling user does not have permissions.
3565 if (!DTRACE_INSCRATCH(mstate, size)) {
3566 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3571 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3572 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3573 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3575 ((char *)dest)[size - 1] = '\0';
3576 mstate->dtms_scratch_ptr += size;
3582 case DIF_SUBR_MSGSIZE:
3583 case DIF_SUBR_MSGDSIZE: {
3584 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3585 uintptr_t wptr, rptr;
3589 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3591 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3597 wptr = dtrace_loadptr(baddr +
3598 offsetof(mblk_t, b_wptr));
3600 rptr = dtrace_loadptr(baddr +
3601 offsetof(mblk_t, b_rptr));
3604 *flags |= CPU_DTRACE_BADADDR;
3605 *illval = tupregs[0].dttk_value;
3609 daddr = dtrace_loadptr(baddr +
3610 offsetof(mblk_t, b_datap));
3612 baddr = dtrace_loadptr(baddr +
3613 offsetof(mblk_t, b_cont));
3616 * We want to prevent against denial-of-service here,
3617 * so we're only going to search the list for
3618 * dtrace_msgdsize_max mblks.
3620 if (cont++ > dtrace_msgdsize_max) {
3621 *flags |= CPU_DTRACE_ILLOP;
3625 if (subr == DIF_SUBR_MSGDSIZE) {
3626 if (dtrace_load8(daddr +
3627 offsetof(dblk_t, db_type)) != M_DATA)
3631 count += wptr - rptr;
3634 if (!(*flags & CPU_DTRACE_FAULT))
3641 case DIF_SUBR_PROGENYOF: {
3642 pid_t pid = tupregs[0].dttk_value;
3646 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3648 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3650 if (p->p_pidp->pid_id == pid) {
3652 if (p->p_pid == pid) {
3659 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3665 case DIF_SUBR_SPECULATION:
3666 regs[rd] = dtrace_speculation(state);
3669 case DIF_SUBR_COPYOUT: {
3670 uintptr_t kaddr = tupregs[0].dttk_value;
3671 uintptr_t uaddr = tupregs[1].dttk_value;
3672 uint64_t size = tupregs[2].dttk_value;
3674 if (!dtrace_destructive_disallow &&
3675 dtrace_priv_proc_control(state) &&
3676 !dtrace_istoxic(kaddr, size)) {
3677 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3678 dtrace_copyout(kaddr, uaddr, size, flags);
3679 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3684 case DIF_SUBR_COPYOUTSTR: {
3685 uintptr_t kaddr = tupregs[0].dttk_value;
3686 uintptr_t uaddr = tupregs[1].dttk_value;
3687 uint64_t size = tupregs[2].dttk_value;
3689 if (!dtrace_destructive_disallow &&
3690 dtrace_priv_proc_control(state) &&
3691 !dtrace_istoxic(kaddr, size)) {
3692 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3693 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3694 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3699 case DIF_SUBR_STRLEN: {
3701 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3702 sz = dtrace_strlen((char *)addr,
3703 state->dts_options[DTRACEOPT_STRSIZE]);
3705 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3715 case DIF_SUBR_STRCHR:
3716 case DIF_SUBR_STRRCHR: {
3718 * We're going to iterate over the string looking for the
3719 * specified character. We will iterate until we have reached
3720 * the string length or we have found the character. If this
3721 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3722 * of the specified character instead of the first.
3724 uintptr_t saddr = tupregs[0].dttk_value;
3725 uintptr_t addr = tupregs[0].dttk_value;
3726 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3727 char c, target = (char)tupregs[1].dttk_value;
3729 for (regs[rd] = 0; addr < limit; addr++) {
3730 if ((c = dtrace_load8(addr)) == target) {
3733 if (subr == DIF_SUBR_STRCHR)
3741 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3749 case DIF_SUBR_STRSTR:
3750 case DIF_SUBR_INDEX:
3751 case DIF_SUBR_RINDEX: {
3753 * We're going to iterate over the string looking for the
3754 * specified string. We will iterate until we have reached
3755 * the string length or we have found the string. (Yes, this
3756 * is done in the most naive way possible -- but considering
3757 * that the string we're searching for is likely to be
3758 * relatively short, the complexity of Rabin-Karp or similar
3759 * hardly seems merited.)
3761 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3762 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3763 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3764 size_t len = dtrace_strlen(addr, size);
3765 size_t sublen = dtrace_strlen(substr, size);
3766 char *limit = addr + len, *orig = addr;
3767 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3770 regs[rd] = notfound;
3772 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3777 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3784 * strstr() and index()/rindex() have similar semantics if
3785 * both strings are the empty string: strstr() returns a
3786 * pointer to the (empty) string, and index() and rindex()
3787 * both return index 0 (regardless of any position argument).
3789 if (sublen == 0 && len == 0) {
3790 if (subr == DIF_SUBR_STRSTR)
3791 regs[rd] = (uintptr_t)addr;
3797 if (subr != DIF_SUBR_STRSTR) {
3798 if (subr == DIF_SUBR_RINDEX) {
3805 * Both index() and rindex() take an optional position
3806 * argument that denotes the starting position.
3809 int64_t pos = (int64_t)tupregs[2].dttk_value;
3812 * If the position argument to index() is
3813 * negative, Perl implicitly clamps it at
3814 * zero. This semantic is a little surprising
3815 * given the special meaning of negative
3816 * positions to similar Perl functions like
3817 * substr(), but it appears to reflect a
3818 * notion that index() can start from a
3819 * negative index and increment its way up to
3820 * the string. Given this notion, Perl's
3821 * rindex() is at least self-consistent in
3822 * that it implicitly clamps positions greater
3823 * than the string length to be the string
3824 * length. Where Perl completely loses
3825 * coherence, however, is when the specified
3826 * substring is the empty string (""). In
3827 * this case, even if the position is
3828 * negative, rindex() returns 0 -- and even if
3829 * the position is greater than the length,
3830 * index() returns the string length. These
3831 * semantics violate the notion that index()
3832 * should never return a value less than the
3833 * specified position and that rindex() should
3834 * never return a value greater than the
3835 * specified position. (One assumes that
3836 * these semantics are artifacts of Perl's
3837 * implementation and not the results of
3838 * deliberate design -- it beggars belief that
3839 * even Larry Wall could desire such oddness.)
3840 * While in the abstract one would wish for
3841 * consistent position semantics across
3842 * substr(), index() and rindex() -- or at the
3843 * very least self-consistent position
3844 * semantics for index() and rindex() -- we
3845 * instead opt to keep with the extant Perl
3846 * semantics, in all their broken glory. (Do
3847 * we have more desire to maintain Perl's
3848 * semantics than Perl does? Probably.)
3850 if (subr == DIF_SUBR_RINDEX) {
3874 for (regs[rd] = notfound; addr != limit; addr += inc) {
3875 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3876 if (subr != DIF_SUBR_STRSTR) {
3878 * As D index() and rindex() are
3879 * modeled on Perl (and not on awk),
3880 * we return a zero-based (and not a
3881 * one-based) index. (For you Perl
3882 * weenies: no, we're not going to add
3883 * $[ -- and shouldn't you be at a con
3886 regs[rd] = (uintptr_t)(addr - orig);
3890 ASSERT(subr == DIF_SUBR_STRSTR);
3891 regs[rd] = (uintptr_t)addr;
3899 case DIF_SUBR_STRTOK: {
3900 uintptr_t addr = tupregs[0].dttk_value;
3901 uintptr_t tokaddr = tupregs[1].dttk_value;
3902 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3903 uintptr_t limit, toklimit = tokaddr + size;
3904 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3905 char *dest = (char *)mstate->dtms_scratch_ptr;
3909 * Check both the token buffer and (later) the input buffer,
3910 * since both could be non-scratch addresses.
3912 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3917 if (!DTRACE_INSCRATCH(mstate, size)) {
3918 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3925 * If the address specified is NULL, we use our saved
3926 * strtok pointer from the mstate. Note that this
3927 * means that the saved strtok pointer is _only_
3928 * valid within multiple enablings of the same probe --
3929 * it behaves like an implicit clause-local variable.
3931 addr = mstate->dtms_strtok;
3934 * If the user-specified address is non-NULL we must
3935 * access check it. This is the only time we have
3936 * a chance to do so, since this address may reside
3937 * in the string table of this clause-- future calls
3938 * (when we fetch addr from mstate->dtms_strtok)
3939 * would fail this access check.
3941 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3948 * First, zero the token map, and then process the token
3949 * string -- setting a bit in the map for every character
3950 * found in the token string.
3952 for (i = 0; i < sizeof (tokmap); i++)
3955 for (; tokaddr < toklimit; tokaddr++) {
3956 if ((c = dtrace_load8(tokaddr)) == '\0')
3959 ASSERT((c >> 3) < sizeof (tokmap));
3960 tokmap[c >> 3] |= (1 << (c & 0x7));
3963 for (limit = addr + size; addr < limit; addr++) {
3965 * We're looking for a character that is _not_ contained
3966 * in the token string.
3968 if ((c = dtrace_load8(addr)) == '\0')
3971 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3977 * We reached the end of the string without finding
3978 * any character that was not in the token string.
3979 * We return NULL in this case, and we set the saved
3980 * address to NULL as well.
3983 mstate->dtms_strtok = 0;
3988 * From here on, we're copying into the destination string.
3990 for (i = 0; addr < limit && i < size - 1; addr++) {
3991 if ((c = dtrace_load8(addr)) == '\0')
3994 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4003 regs[rd] = (uintptr_t)dest;
4004 mstate->dtms_scratch_ptr += size;
4005 mstate->dtms_strtok = addr;
4009 case DIF_SUBR_SUBSTR: {
4010 uintptr_t s = tupregs[0].dttk_value;
4011 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4012 char *d = (char *)mstate->dtms_scratch_ptr;
4013 int64_t index = (int64_t)tupregs[1].dttk_value;
4014 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4015 size_t len = dtrace_strlen((char *)s, size);
4018 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4023 if (!DTRACE_INSCRATCH(mstate, size)) {
4024 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4030 remaining = (int64_t)size;
4035 if (index < 0 && index + remaining > 0) {
4041 if (index >= len || index < 0) {
4043 } else if (remaining < 0) {
4044 remaining += len - index;
4045 } else if (index + remaining > size) {
4046 remaining = size - index;
4049 for (i = 0; i < remaining; i++) {
4050 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4056 mstate->dtms_scratch_ptr += size;
4057 regs[rd] = (uintptr_t)d;
4061 case DIF_SUBR_TOUPPER:
4062 case DIF_SUBR_TOLOWER: {
4063 uintptr_t s = tupregs[0].dttk_value;
4064 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4065 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4066 size_t len = dtrace_strlen((char *)s, size);
4067 char lower, upper, convert;
4070 if (subr == DIF_SUBR_TOUPPER) {
4080 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4085 if (!DTRACE_INSCRATCH(mstate, size)) {
4086 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4091 for (i = 0; i < size - 1; i++) {
4092 if ((c = dtrace_load8(s + i)) == '\0')
4095 if (c >= lower && c <= upper)
4096 c = convert + (c - lower);
4103 regs[rd] = (uintptr_t)dest;
4104 mstate->dtms_scratch_ptr += size;
4109 case DIF_SUBR_GETMAJOR:
4111 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4113 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4117 case DIF_SUBR_GETMINOR:
4119 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4121 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4125 case DIF_SUBR_DDI_PATHNAME: {
4127 * This one is a galactic mess. We are going to roughly
4128 * emulate ddi_pathname(), but it's made more complicated
4129 * by the fact that we (a) want to include the minor name and
4130 * (b) must proceed iteratively instead of recursively.
4132 uintptr_t dest = mstate->dtms_scratch_ptr;
4133 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4134 char *start = (char *)dest, *end = start + size - 1;
4135 uintptr_t daddr = tupregs[0].dttk_value;
4136 int64_t minor = (int64_t)tupregs[1].dttk_value;
4138 int i, len, depth = 0;
4141 * Due to all the pointer jumping we do and context we must
4142 * rely upon, we just mandate that the user must have kernel
4143 * read privileges to use this routine.
4145 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4146 *flags |= CPU_DTRACE_KPRIV;
4151 if (!DTRACE_INSCRATCH(mstate, size)) {
4152 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4160 * We want to have a name for the minor. In order to do this,
4161 * we need to walk the minor list from the devinfo. We want
4162 * to be sure that we don't infinitely walk a circular list,
4163 * so we check for circularity by sending a scout pointer
4164 * ahead two elements for every element that we iterate over;
4165 * if the list is circular, these will ultimately point to the
4166 * same element. You may recognize this little trick as the
4167 * answer to a stupid interview question -- one that always
4168 * seems to be asked by those who had to have it laboriously
4169 * explained to them, and who can't even concisely describe
4170 * the conditions under which one would be forced to resort to
4171 * this technique. Needless to say, those conditions are
4172 * found here -- and probably only here. Is this the only use
4173 * of this infamous trick in shipping, production code? If it
4174 * isn't, it probably should be...
4177 uintptr_t maddr = dtrace_loadptr(daddr +
4178 offsetof(struct dev_info, devi_minor));
4180 uintptr_t next = offsetof(struct ddi_minor_data, next);
4181 uintptr_t name = offsetof(struct ddi_minor_data,
4182 d_minor) + offsetof(struct ddi_minor, name);
4183 uintptr_t dev = offsetof(struct ddi_minor_data,
4184 d_minor) + offsetof(struct ddi_minor, dev);
4188 scout = dtrace_loadptr(maddr + next);
4190 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4193 m = dtrace_load64(maddr + dev) & MAXMIN64;
4195 m = dtrace_load32(maddr + dev) & MAXMIN;
4198 maddr = dtrace_loadptr(maddr + next);
4203 scout = dtrace_loadptr(scout + next);
4208 scout = dtrace_loadptr(scout + next);
4213 if (scout == maddr) {
4214 *flags |= CPU_DTRACE_ILLOP;
4222 * We have the minor data. Now we need to
4223 * copy the minor's name into the end of the
4226 s = (char *)dtrace_loadptr(maddr + name);
4227 len = dtrace_strlen(s, size);
4229 if (*flags & CPU_DTRACE_FAULT)
4233 if ((end -= (len + 1)) < start)
4239 for (i = 1; i <= len; i++)
4240 end[i] = dtrace_load8((uintptr_t)s++);
4245 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4246 ddi_node_state_t devi_state;
4248 devi_state = dtrace_load32(daddr +
4249 offsetof(struct dev_info, devi_node_state));
4251 if (*flags & CPU_DTRACE_FAULT)
4254 if (devi_state >= DS_INITIALIZED) {
4255 s = (char *)dtrace_loadptr(daddr +
4256 offsetof(struct dev_info, devi_addr));
4257 len = dtrace_strlen(s, size);
4259 if (*flags & CPU_DTRACE_FAULT)
4263 if ((end -= (len + 1)) < start)
4269 for (i = 1; i <= len; i++)
4270 end[i] = dtrace_load8((uintptr_t)s++);
4274 * Now for the node name...
4276 s = (char *)dtrace_loadptr(daddr +
4277 offsetof(struct dev_info, devi_node_name));
4279 daddr = dtrace_loadptr(daddr +
4280 offsetof(struct dev_info, devi_parent));
4283 * If our parent is NULL (that is, if we're the root
4284 * node), we're going to use the special path
4290 len = dtrace_strlen(s, size);
4291 if (*flags & CPU_DTRACE_FAULT)
4294 if ((end -= (len + 1)) < start)
4297 for (i = 1; i <= len; i++)
4298 end[i] = dtrace_load8((uintptr_t)s++);
4301 if (depth++ > dtrace_devdepth_max) {
4302 *flags |= CPU_DTRACE_ILLOP;
4308 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4311 regs[rd] = (uintptr_t)end;
4312 mstate->dtms_scratch_ptr += size;
4319 case DIF_SUBR_STRJOIN: {
4320 char *d = (char *)mstate->dtms_scratch_ptr;
4321 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4322 uintptr_t s1 = tupregs[0].dttk_value;
4323 uintptr_t s2 = tupregs[1].dttk_value;
4326 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4327 !dtrace_strcanload(s2, size, mstate, vstate)) {
4332 if (!DTRACE_INSCRATCH(mstate, size)) {
4333 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4340 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4345 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4353 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4358 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4363 mstate->dtms_scratch_ptr += i;
4364 regs[rd] = (uintptr_t)d;
4370 case DIF_SUBR_LLTOSTR: {
4371 int64_t i = (int64_t)tupregs[0].dttk_value;
4372 uint64_t val, digit;
4373 uint64_t size = 65; /* enough room for 2^64 in binary */
4374 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4378 if ((base = tupregs[1].dttk_value) <= 1 ||
4379 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4380 *flags |= CPU_DTRACE_ILLOP;
4385 val = (base == 10 && i < 0) ? i * -1 : i;
4387 if (!DTRACE_INSCRATCH(mstate, size)) {
4388 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4393 for (*end-- = '\0'; val; val /= base) {
4394 if ((digit = val % base) <= '9' - '0') {
4395 *end-- = '0' + digit;
4397 *end-- = 'a' + (digit - ('9' - '0') - 1);
4401 if (i == 0 && base == 16)
4407 if (i == 0 || base == 8 || base == 16)
4410 if (i < 0 && base == 10)
4413 regs[rd] = (uintptr_t)end + 1;
4414 mstate->dtms_scratch_ptr += size;
4418 case DIF_SUBR_HTONS:
4419 case DIF_SUBR_NTOHS:
4420 #if BYTE_ORDER == BIG_ENDIAN
4421 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4423 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4428 case DIF_SUBR_HTONL:
4429 case DIF_SUBR_NTOHL:
4430 #if BYTE_ORDER == BIG_ENDIAN
4431 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4433 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4438 case DIF_SUBR_HTONLL:
4439 case DIF_SUBR_NTOHLL:
4440 #if BYTE_ORDER == BIG_ENDIAN
4441 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4443 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4448 case DIF_SUBR_DIRNAME:
4449 case DIF_SUBR_BASENAME: {
4450 char *dest = (char *)mstate->dtms_scratch_ptr;
4451 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4452 uintptr_t src = tupregs[0].dttk_value;
4453 int i, j, len = dtrace_strlen((char *)src, size);
4454 int lastbase = -1, firstbase = -1, lastdir = -1;
4457 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4462 if (!DTRACE_INSCRATCH(mstate, size)) {
4463 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4469 * The basename and dirname for a zero-length string is
4474 src = (uintptr_t)".";
4478 * Start from the back of the string, moving back toward the
4479 * front until we see a character that isn't a slash. That
4480 * character is the last character in the basename.
4482 for (i = len - 1; i >= 0; i--) {
4483 if (dtrace_load8(src + i) != '/')
4491 * Starting from the last character in the basename, move
4492 * towards the front until we find a slash. The character
4493 * that we processed immediately before that is the first
4494 * character in the basename.
4496 for (; i >= 0; i--) {
4497 if (dtrace_load8(src + i) == '/')
4505 * Now keep going until we find a non-slash character. That
4506 * character is the last character in the dirname.
4508 for (; i >= 0; i--) {
4509 if (dtrace_load8(src + i) != '/')
4516 ASSERT(!(lastbase == -1 && firstbase != -1));
4517 ASSERT(!(firstbase == -1 && lastdir != -1));
4519 if (lastbase == -1) {
4521 * We didn't find a non-slash character. We know that
4522 * the length is non-zero, so the whole string must be
4523 * slashes. In either the dirname or the basename
4524 * case, we return '/'.
4526 ASSERT(firstbase == -1);
4527 firstbase = lastbase = lastdir = 0;
4530 if (firstbase == -1) {
4532 * The entire string consists only of a basename
4533 * component. If we're looking for dirname, we need
4534 * to change our string to be just "."; if we're
4535 * looking for a basename, we'll just set the first
4536 * character of the basename to be 0.
4538 if (subr == DIF_SUBR_DIRNAME) {
4539 ASSERT(lastdir == -1);
4540 src = (uintptr_t)".";
4547 if (subr == DIF_SUBR_DIRNAME) {
4548 if (lastdir == -1) {
4550 * We know that we have a slash in the name --
4551 * or lastdir would be set to 0, above. And
4552 * because lastdir is -1, we know that this
4553 * slash must be the first character. (That
4554 * is, the full string must be of the form
4555 * "/basename".) In this case, the last
4556 * character of the directory name is 0.
4564 ASSERT(subr == DIF_SUBR_BASENAME);
4565 ASSERT(firstbase != -1 && lastbase != -1);
4570 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4571 dest[j] = dtrace_load8(src + i);
4574 regs[rd] = (uintptr_t)dest;
4575 mstate->dtms_scratch_ptr += size;
4579 case DIF_SUBR_CLEANPATH: {
4580 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4581 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4582 uintptr_t src = tupregs[0].dttk_value;
4585 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4590 if (!DTRACE_INSCRATCH(mstate, size)) {
4591 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4597 * Move forward, loading each character.
4600 c = dtrace_load8(src + i++);
4602 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4610 c = dtrace_load8(src + i++);
4614 * We have two slashes -- we can just advance
4615 * to the next character.
4622 * This is not "." and it's not ".." -- we can
4623 * just store the "/" and this character and
4631 c = dtrace_load8(src + i++);
4635 * This is a "/./" component. We're not going
4636 * to store anything in the destination buffer;
4637 * we're just going to go to the next component.
4644 * This is not ".." -- we can just store the
4645 * "/." and this character and continue
4654 c = dtrace_load8(src + i++);
4656 if (c != '/' && c != '\0') {
4658 * This is not ".." -- it's "..[mumble]".
4659 * We'll store the "/.." and this character
4660 * and continue processing.
4670 * This is "/../" or "/..\0". We need to back up
4671 * our destination pointer until we find a "/".
4674 while (j != 0 && dest[--j] != '/')
4679 } while (c != '\0');
4682 regs[rd] = (uintptr_t)dest;
4683 mstate->dtms_scratch_ptr += size;
4687 case DIF_SUBR_INET_NTOA:
4688 case DIF_SUBR_INET_NTOA6:
4689 case DIF_SUBR_INET_NTOP: {
4694 if (subr == DIF_SUBR_INET_NTOP) {
4695 af = (int)tupregs[0].dttk_value;
4698 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4702 if (af == AF_INET) {
4707 * Safely load the IPv4 address.
4709 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4712 * Check an IPv4 string will fit in scratch.
4714 size = INET_ADDRSTRLEN;
4715 if (!DTRACE_INSCRATCH(mstate, size)) {
4716 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4720 base = (char *)mstate->dtms_scratch_ptr;
4721 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4724 * Stringify as a dotted decimal quad.
4727 ptr8 = (uint8_t *)&ip4;
4728 for (i = 3; i >= 0; i--) {
4734 for (; val; val /= 10) {
4735 *end-- = '0' + (val % 10);
4742 ASSERT(end + 1 >= base);
4744 } else if (af == AF_INET6) {
4745 struct in6_addr ip6;
4746 int firstzero, tryzero, numzero, v6end;
4748 const char digits[] = "0123456789abcdef";
4751 * Stringify using RFC 1884 convention 2 - 16 bit
4752 * hexadecimal values with a zero-run compression.
4753 * Lower case hexadecimal digits are used.
4754 * eg, fe80::214:4fff:fe0b:76c8.
4755 * The IPv4 embedded form is returned for inet_ntop,
4756 * just the IPv4 string is returned for inet_ntoa6.
4760 * Safely load the IPv6 address.
4763 (void *)(uintptr_t)tupregs[argi].dttk_value,
4764 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4767 * Check an IPv6 string will fit in scratch.
4769 size = INET6_ADDRSTRLEN;
4770 if (!DTRACE_INSCRATCH(mstate, size)) {
4771 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4775 base = (char *)mstate->dtms_scratch_ptr;
4776 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4780 * Find the longest run of 16 bit zero values
4781 * for the single allowed zero compression - "::".
4786 for (i = 0; i < sizeof (struct in6_addr); i++) {
4788 if (ip6._S6_un._S6_u8[i] == 0 &&
4790 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4792 tryzero == -1 && i % 2 == 0) {
4797 if (tryzero != -1 &&
4799 (ip6._S6_un._S6_u8[i] != 0 ||
4801 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4803 i == sizeof (struct in6_addr) - 1)) {
4805 if (i - tryzero <= numzero) {
4810 firstzero = tryzero;
4811 numzero = i - i % 2 - tryzero;
4815 if (ip6._S6_un._S6_u8[i] == 0 &&
4817 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4819 i == sizeof (struct in6_addr) - 1)
4823 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4826 * Check for an IPv4 embedded address.
4828 v6end = sizeof (struct in6_addr) - 2;
4829 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4830 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4831 for (i = sizeof (struct in6_addr) - 1;
4832 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4833 ASSERT(end >= base);
4836 val = ip6._S6_un._S6_u8[i];
4838 val = ip6.__u6_addr.__u6_addr8[i];
4844 for (; val; val /= 10) {
4845 *end-- = '0' + val % 10;
4849 if (i > DTRACE_V4MAPPED_OFFSET)
4853 if (subr == DIF_SUBR_INET_NTOA6)
4857 * Set v6end to skip the IPv4 address that
4858 * we have already stringified.
4864 * Build the IPv6 string by working through the
4865 * address in reverse.
4867 for (i = v6end; i >= 0; i -= 2) {
4868 ASSERT(end >= base);
4870 if (i == firstzero + numzero - 2) {
4877 if (i < 14 && i != firstzero - 2)
4881 val = (ip6._S6_un._S6_u8[i] << 8) +
4882 ip6._S6_un._S6_u8[i + 1];
4884 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4885 ip6.__u6_addr.__u6_addr8[i + 1];
4891 for (; val; val /= 16) {
4892 *end-- = digits[val % 16];
4896 ASSERT(end + 1 >= base);
4900 * The user didn't use AH_INET or AH_INET6.
4902 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4907 inetout: regs[rd] = (uintptr_t)end + 1;
4908 mstate->dtms_scratch_ptr += size;
4912 case DIF_SUBR_MEMREF: {
4913 uintptr_t size = 2 * sizeof(uintptr_t);
4914 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4915 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4917 /* address and length */
4918 memref[0] = tupregs[0].dttk_value;
4919 memref[1] = tupregs[1].dttk_value;
4921 regs[rd] = (uintptr_t) memref;
4922 mstate->dtms_scratch_ptr += scratch_size;
4927 case DIF_SUBR_MEMSTR: {
4928 char *str = (char *)mstate->dtms_scratch_ptr;
4929 uintptr_t mem = tupregs[0].dttk_value;
4930 char c = tupregs[1].dttk_value;
4931 size_t size = tupregs[2].dttk_value;
4940 if (!dtrace_canload(mem, size - 1, mstate, vstate))
4943 if (!DTRACE_INSCRATCH(mstate, size)) {
4944 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4948 if (dtrace_memstr_max != 0 && size > dtrace_memstr_max) {
4949 *flags |= CPU_DTRACE_ILLOP;
4953 for (i = 0; i < size - 1; i++) {
4954 n = dtrace_load8(mem++);
4955 str[i] = (n == 0) ? c : n;
4959 regs[rd] = (uintptr_t)str;
4960 mstate->dtms_scratch_ptr += size;
4965 case DIF_SUBR_TYPEREF: {
4966 uintptr_t size = 4 * sizeof(uintptr_t);
4967 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4968 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4970 /* address, num_elements, type_str, type_len */
4971 typeref[0] = tupregs[0].dttk_value;
4972 typeref[1] = tupregs[1].dttk_value;
4973 typeref[2] = tupregs[2].dttk_value;
4974 typeref[3] = tupregs[3].dttk_value;
4976 regs[rd] = (uintptr_t) typeref;
4977 mstate->dtms_scratch_ptr += scratch_size;
4984 * Emulate the execution of DTrace IR instructions specified by the given
4985 * DIF object. This function is deliberately void of assertions as all of
4986 * the necessary checks are handled by a call to dtrace_difo_validate().
4989 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4990 dtrace_vstate_t *vstate, dtrace_state_t *state)
4992 const dif_instr_t *text = difo->dtdo_buf;
4993 const uint_t textlen = difo->dtdo_len;
4994 const char *strtab = difo->dtdo_strtab;
4995 const uint64_t *inttab = difo->dtdo_inttab;
4998 dtrace_statvar_t *svar;
4999 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5001 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5002 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
5004 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5005 uint64_t regs[DIF_DIR_NREGS];
5008 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5010 uint_t pc = 0, id, opc = 0;
5016 * We stash the current DIF object into the machine state: we need it
5017 * for subsequent access checking.
5019 mstate->dtms_difo = difo;
5021 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
5023 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5027 r1 = DIF_INSTR_R1(instr);
5028 r2 = DIF_INSTR_R2(instr);
5029 rd = DIF_INSTR_RD(instr);
5031 switch (DIF_INSTR_OP(instr)) {
5033 regs[rd] = regs[r1] | regs[r2];
5036 regs[rd] = regs[r1] ^ regs[r2];
5039 regs[rd] = regs[r1] & regs[r2];
5042 regs[rd] = regs[r1] << regs[r2];
5045 regs[rd] = regs[r1] >> regs[r2];
5048 regs[rd] = regs[r1] - regs[r2];
5051 regs[rd] = regs[r1] + regs[r2];
5054 regs[rd] = regs[r1] * regs[r2];
5057 if (regs[r2] == 0) {
5059 *flags |= CPU_DTRACE_DIVZERO;
5061 regs[rd] = (int64_t)regs[r1] /
5067 if (regs[r2] == 0) {
5069 *flags |= CPU_DTRACE_DIVZERO;
5071 regs[rd] = regs[r1] / regs[r2];
5076 if (regs[r2] == 0) {
5078 *flags |= CPU_DTRACE_DIVZERO;
5080 regs[rd] = (int64_t)regs[r1] %
5086 if (regs[r2] == 0) {
5088 *flags |= CPU_DTRACE_DIVZERO;
5090 regs[rd] = regs[r1] % regs[r2];
5095 regs[rd] = ~regs[r1];
5098 regs[rd] = regs[r1];
5101 cc_r = regs[r1] - regs[r2];
5105 cc_c = regs[r1] < regs[r2];
5108 cc_n = cc_v = cc_c = 0;
5109 cc_z = regs[r1] == 0;
5112 pc = DIF_INSTR_LABEL(instr);
5116 pc = DIF_INSTR_LABEL(instr);
5120 pc = DIF_INSTR_LABEL(instr);
5123 if ((cc_z | (cc_n ^ cc_v)) == 0)
5124 pc = DIF_INSTR_LABEL(instr);
5127 if ((cc_c | cc_z) == 0)
5128 pc = DIF_INSTR_LABEL(instr);
5131 if ((cc_n ^ cc_v) == 0)
5132 pc = DIF_INSTR_LABEL(instr);
5136 pc = DIF_INSTR_LABEL(instr);
5140 pc = DIF_INSTR_LABEL(instr);
5144 pc = DIF_INSTR_LABEL(instr);
5147 if (cc_z | (cc_n ^ cc_v))
5148 pc = DIF_INSTR_LABEL(instr);
5152 pc = DIF_INSTR_LABEL(instr);
5155 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5156 *flags |= CPU_DTRACE_KPRIV;
5162 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5165 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5166 *flags |= CPU_DTRACE_KPRIV;
5172 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5175 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5176 *flags |= CPU_DTRACE_KPRIV;
5182 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5185 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5186 *flags |= CPU_DTRACE_KPRIV;
5192 regs[rd] = dtrace_load8(regs[r1]);
5195 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5196 *flags |= CPU_DTRACE_KPRIV;
5202 regs[rd] = dtrace_load16(regs[r1]);
5205 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5206 *flags |= CPU_DTRACE_KPRIV;
5212 regs[rd] = dtrace_load32(regs[r1]);
5215 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5216 *flags |= CPU_DTRACE_KPRIV;
5222 regs[rd] = dtrace_load64(regs[r1]);
5226 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5229 regs[rd] = (int16_t)
5230 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5233 regs[rd] = (int32_t)
5234 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5238 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5242 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5246 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5250 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5259 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5262 regs[rd] = (uint64_t)(uintptr_t)
5263 (strtab + DIF_INSTR_STRING(instr));
5266 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5267 uintptr_t s1 = regs[r1];
5268 uintptr_t s2 = regs[r2];
5271 !dtrace_strcanload(s1, sz, mstate, vstate))
5274 !dtrace_strcanload(s2, sz, mstate, vstate))
5277 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5285 regs[rd] = dtrace_dif_variable(mstate, state,
5289 id = DIF_INSTR_VAR(instr);
5291 if (id >= DIF_VAR_OTHER_UBASE) {
5294 id -= DIF_VAR_OTHER_UBASE;
5295 svar = vstate->dtvs_globals[id];
5296 ASSERT(svar != NULL);
5297 v = &svar->dtsv_var;
5299 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5300 regs[rd] = svar->dtsv_data;
5304 a = (uintptr_t)svar->dtsv_data;
5306 if (*(uint8_t *)a == UINT8_MAX) {
5308 * If the 0th byte is set to UINT8_MAX
5309 * then this is to be treated as a
5310 * reference to a NULL variable.
5314 regs[rd] = a + sizeof (uint64_t);
5320 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5324 id = DIF_INSTR_VAR(instr);
5326 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5327 id -= DIF_VAR_OTHER_UBASE;
5329 svar = vstate->dtvs_globals[id];
5330 ASSERT(svar != NULL);
5331 v = &svar->dtsv_var;
5333 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5334 uintptr_t a = (uintptr_t)svar->dtsv_data;
5337 ASSERT(svar->dtsv_size != 0);
5339 if (regs[rd] == 0) {
5340 *(uint8_t *)a = UINT8_MAX;
5344 a += sizeof (uint64_t);
5346 if (!dtrace_vcanload(
5347 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5351 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5352 (void *)a, &v->dtdv_type);
5356 svar->dtsv_data = regs[rd];
5361 * There are no DTrace built-in thread-local arrays at
5362 * present. This opcode is saved for future work.
5364 *flags |= CPU_DTRACE_ILLOP;
5369 id = DIF_INSTR_VAR(instr);
5371 if (id < DIF_VAR_OTHER_UBASE) {
5373 * For now, this has no meaning.
5379 id -= DIF_VAR_OTHER_UBASE;
5381 ASSERT(id < vstate->dtvs_nlocals);
5382 ASSERT(vstate->dtvs_locals != NULL);
5384 svar = vstate->dtvs_locals[id];
5385 ASSERT(svar != NULL);
5386 v = &svar->dtsv_var;
5388 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5389 uintptr_t a = (uintptr_t)svar->dtsv_data;
5390 size_t sz = v->dtdv_type.dtdt_size;
5392 sz += sizeof (uint64_t);
5393 ASSERT(svar->dtsv_size == NCPU * sz);
5396 if (*(uint8_t *)a == UINT8_MAX) {
5398 * If the 0th byte is set to UINT8_MAX
5399 * then this is to be treated as a
5400 * reference to a NULL variable.
5404 regs[rd] = a + sizeof (uint64_t);
5410 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5411 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5412 regs[rd] = tmp[curcpu];
5416 id = DIF_INSTR_VAR(instr);
5418 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5419 id -= DIF_VAR_OTHER_UBASE;
5420 ASSERT(id < vstate->dtvs_nlocals);
5422 ASSERT(vstate->dtvs_locals != NULL);
5423 svar = vstate->dtvs_locals[id];
5424 ASSERT(svar != NULL);
5425 v = &svar->dtsv_var;
5427 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5428 uintptr_t a = (uintptr_t)svar->dtsv_data;
5429 size_t sz = v->dtdv_type.dtdt_size;
5431 sz += sizeof (uint64_t);
5432 ASSERT(svar->dtsv_size == NCPU * sz);
5435 if (regs[rd] == 0) {
5436 *(uint8_t *)a = UINT8_MAX;
5440 a += sizeof (uint64_t);
5443 if (!dtrace_vcanload(
5444 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5448 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5449 (void *)a, &v->dtdv_type);
5453 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5454 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5455 tmp[curcpu] = regs[rd];
5459 dtrace_dynvar_t *dvar;
5462 id = DIF_INSTR_VAR(instr);
5463 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5464 id -= DIF_VAR_OTHER_UBASE;
5465 v = &vstate->dtvs_tlocals[id];
5467 key = &tupregs[DIF_DTR_NREGS];
5468 key[0].dttk_value = (uint64_t)id;
5469 key[0].dttk_size = 0;
5470 DTRACE_TLS_THRKEY(key[1].dttk_value);
5471 key[1].dttk_size = 0;
5473 dvar = dtrace_dynvar(dstate, 2, key,
5474 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5482 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5483 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5485 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5492 dtrace_dynvar_t *dvar;
5495 id = DIF_INSTR_VAR(instr);
5496 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5497 id -= DIF_VAR_OTHER_UBASE;
5499 key = &tupregs[DIF_DTR_NREGS];
5500 key[0].dttk_value = (uint64_t)id;
5501 key[0].dttk_size = 0;
5502 DTRACE_TLS_THRKEY(key[1].dttk_value);
5503 key[1].dttk_size = 0;
5504 v = &vstate->dtvs_tlocals[id];
5506 dvar = dtrace_dynvar(dstate, 2, key,
5507 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5508 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5509 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5510 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5513 * Given that we're storing to thread-local data,
5514 * we need to flush our predicate cache.
5516 curthread->t_predcache = 0;
5521 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5522 if (!dtrace_vcanload(
5523 (void *)(uintptr_t)regs[rd],
5524 &v->dtdv_type, mstate, vstate))
5527 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5528 dvar->dtdv_data, &v->dtdv_type);
5530 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5537 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5541 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5542 regs, tupregs, ttop, mstate, state);
5546 if (ttop == DIF_DTR_NREGS) {
5547 *flags |= CPU_DTRACE_TUPOFLOW;
5551 if (r1 == DIF_TYPE_STRING) {
5553 * If this is a string type and the size is 0,
5554 * we'll use the system-wide default string
5555 * size. Note that we are _not_ looking at
5556 * the value of the DTRACEOPT_STRSIZE option;
5557 * had this been set, we would expect to have
5558 * a non-zero size value in the "pushtr".
5560 tupregs[ttop].dttk_size =
5561 dtrace_strlen((char *)(uintptr_t)regs[rd],
5562 regs[r2] ? regs[r2] :
5563 dtrace_strsize_default) + 1;
5565 tupregs[ttop].dttk_size = regs[r2];
5568 tupregs[ttop++].dttk_value = regs[rd];
5572 if (ttop == DIF_DTR_NREGS) {
5573 *flags |= CPU_DTRACE_TUPOFLOW;
5577 tupregs[ttop].dttk_value = regs[rd];
5578 tupregs[ttop++].dttk_size = 0;
5586 case DIF_OP_FLUSHTS:
5591 case DIF_OP_LDTAA: {
5592 dtrace_dynvar_t *dvar;
5593 dtrace_key_t *key = tupregs;
5594 uint_t nkeys = ttop;
5596 id = DIF_INSTR_VAR(instr);
5597 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5598 id -= DIF_VAR_OTHER_UBASE;
5600 key[nkeys].dttk_value = (uint64_t)id;
5601 key[nkeys++].dttk_size = 0;
5603 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5604 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5605 key[nkeys++].dttk_size = 0;
5606 v = &vstate->dtvs_tlocals[id];
5608 v = &vstate->dtvs_globals[id]->dtsv_var;
5611 dvar = dtrace_dynvar(dstate, nkeys, key,
5612 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5613 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5614 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5621 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5622 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5624 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5631 case DIF_OP_STTAA: {
5632 dtrace_dynvar_t *dvar;
5633 dtrace_key_t *key = tupregs;
5634 uint_t nkeys = ttop;
5636 id = DIF_INSTR_VAR(instr);
5637 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5638 id -= DIF_VAR_OTHER_UBASE;
5640 key[nkeys].dttk_value = (uint64_t)id;
5641 key[nkeys++].dttk_size = 0;
5643 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5644 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5645 key[nkeys++].dttk_size = 0;
5646 v = &vstate->dtvs_tlocals[id];
5648 v = &vstate->dtvs_globals[id]->dtsv_var;
5651 dvar = dtrace_dynvar(dstate, nkeys, key,
5652 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5653 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5654 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5655 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5660 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5661 if (!dtrace_vcanload(
5662 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5666 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5667 dvar->dtdv_data, &v->dtdv_type);
5669 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5675 case DIF_OP_ALLOCS: {
5676 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5677 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5680 * Rounding up the user allocation size could have
5681 * overflowed large, bogus allocations (like -1ULL) to
5684 if (size < regs[r1] ||
5685 !DTRACE_INSCRATCH(mstate, size)) {
5686 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5691 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5692 mstate->dtms_scratch_ptr += size;
5698 if (!dtrace_canstore(regs[rd], regs[r2],
5700 *flags |= CPU_DTRACE_BADADDR;
5705 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5708 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5709 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5713 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5714 *flags |= CPU_DTRACE_BADADDR;
5718 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5722 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5723 *flags |= CPU_DTRACE_BADADDR;
5728 *flags |= CPU_DTRACE_BADALIGN;
5732 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5736 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5737 *flags |= CPU_DTRACE_BADADDR;
5742 *flags |= CPU_DTRACE_BADALIGN;
5746 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5750 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5751 *flags |= CPU_DTRACE_BADADDR;
5756 *flags |= CPU_DTRACE_BADALIGN;
5760 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5765 if (!(*flags & CPU_DTRACE_FAULT))
5768 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5769 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5775 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5777 dtrace_probe_t *probe = ecb->dte_probe;
5778 dtrace_provider_t *prov = probe->dtpr_provider;
5779 char c[DTRACE_FULLNAMELEN + 80], *str;
5780 char *msg = "dtrace: breakpoint action at probe ";
5781 char *ecbmsg = " (ecb ";
5782 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5783 uintptr_t val = (uintptr_t)ecb;
5784 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5786 if (dtrace_destructive_disallow)
5790 * It's impossible to be taking action on the NULL probe.
5792 ASSERT(probe != NULL);
5795 * This is a poor man's (destitute man's?) sprintf(): we want to
5796 * print the provider name, module name, function name and name of
5797 * the probe, along with the hex address of the ECB with the breakpoint
5798 * action -- all of which we must place in the character buffer by
5801 while (*msg != '\0')
5804 for (str = prov->dtpv_name; *str != '\0'; str++)
5808 for (str = probe->dtpr_mod; *str != '\0'; str++)
5812 for (str = probe->dtpr_func; *str != '\0'; str++)
5816 for (str = probe->dtpr_name; *str != '\0'; str++)
5819 while (*ecbmsg != '\0')
5822 while (shift >= 0) {
5823 mask = (uintptr_t)0xf << shift;
5825 if (val >= ((uintptr_t)1 << shift))
5826 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5836 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5841 dtrace_action_panic(dtrace_ecb_t *ecb)
5843 dtrace_probe_t *probe = ecb->dte_probe;
5846 * It's impossible to be taking action on the NULL probe.
5848 ASSERT(probe != NULL);
5850 if (dtrace_destructive_disallow)
5853 if (dtrace_panicked != NULL)
5856 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5860 * We won the right to panic. (We want to be sure that only one
5861 * thread calls panic() from dtrace_probe(), and that panic() is
5862 * called exactly once.)
5864 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5865 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5866 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5870 dtrace_action_raise(uint64_t sig)
5872 if (dtrace_destructive_disallow)
5876 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5882 * raise() has a queue depth of 1 -- we ignore all subsequent
5883 * invocations of the raise() action.
5885 if (curthread->t_dtrace_sig == 0)
5886 curthread->t_dtrace_sig = (uint8_t)sig;
5888 curthread->t_sig_check = 1;
5891 struct proc *p = curproc;
5893 kern_psignal(p, sig);
5899 dtrace_action_stop(void)
5901 if (dtrace_destructive_disallow)
5905 if (!curthread->t_dtrace_stop) {
5906 curthread->t_dtrace_stop = 1;
5907 curthread->t_sig_check = 1;
5911 struct proc *p = curproc;
5913 kern_psignal(p, SIGSTOP);
5919 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5922 volatile uint16_t *flags;
5926 cpu_t *cpu = &solaris_cpu[curcpu];
5929 if (dtrace_destructive_disallow)
5932 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5934 now = dtrace_gethrtime();
5936 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5938 * We need to advance the mark to the current time.
5940 cpu->cpu_dtrace_chillmark = now;
5941 cpu->cpu_dtrace_chilled = 0;
5945 * Now check to see if the requested chill time would take us over
5946 * the maximum amount of time allowed in the chill interval. (Or
5947 * worse, if the calculation itself induces overflow.)
5949 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5950 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5951 *flags |= CPU_DTRACE_ILLOP;
5955 while (dtrace_gethrtime() - now < val)
5959 * Normally, we assure that the value of the variable "timestamp" does
5960 * not change within an ECB. The presence of chill() represents an
5961 * exception to this rule, however.
5963 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5964 cpu->cpu_dtrace_chilled += val;
5968 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5969 uint64_t *buf, uint64_t arg)
5971 int nframes = DTRACE_USTACK_NFRAMES(arg);
5972 int strsize = DTRACE_USTACK_STRSIZE(arg);
5973 uint64_t *pcs = &buf[1], *fps;
5974 char *str = (char *)&pcs[nframes];
5975 int size, offs = 0, i, j;
5976 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5977 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5981 * Should be taking a faster path if string space has not been
5984 ASSERT(strsize != 0);
5987 * We will first allocate some temporary space for the frame pointers.
5989 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5990 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5991 (nframes * sizeof (uint64_t));
5993 if (!DTRACE_INSCRATCH(mstate, size)) {
5995 * Not enough room for our frame pointers -- need to indicate
5996 * that we ran out of scratch space.
5998 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6002 mstate->dtms_scratch_ptr += size;
6003 saved = mstate->dtms_scratch_ptr;
6006 * Now get a stack with both program counters and frame pointers.
6008 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6009 dtrace_getufpstack(buf, fps, nframes + 1);
6010 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6013 * If that faulted, we're cooked.
6015 if (*flags & CPU_DTRACE_FAULT)
6019 * Now we want to walk up the stack, calling the USTACK helper. For
6020 * each iteration, we restore the scratch pointer.
6022 for (i = 0; i < nframes; i++) {
6023 mstate->dtms_scratch_ptr = saved;
6025 if (offs >= strsize)
6028 sym = (char *)(uintptr_t)dtrace_helper(
6029 DTRACE_HELPER_ACTION_USTACK,
6030 mstate, state, pcs[i], fps[i]);
6033 * If we faulted while running the helper, we're going to
6034 * clear the fault and null out the corresponding string.
6036 if (*flags & CPU_DTRACE_FAULT) {
6037 *flags &= ~CPU_DTRACE_FAULT;
6047 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6050 * Now copy in the string that the helper returned to us.
6052 for (j = 0; offs + j < strsize; j++) {
6053 if ((str[offs + j] = sym[j]) == '\0')
6057 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6062 if (offs >= strsize) {
6064 * If we didn't have room for all of the strings, we don't
6065 * abort processing -- this needn't be a fatal error -- but we
6066 * still want to increment a counter (dts_stkstroverflows) to
6067 * allow this condition to be warned about. (If this is from
6068 * a jstack() action, it is easily tuned via jstackstrsize.)
6070 dtrace_error(&state->dts_stkstroverflows);
6073 while (offs < strsize)
6077 mstate->dtms_scratch_ptr = old;
6081 * If you're looking for the epicenter of DTrace, you just found it. This
6082 * is the function called by the provider to fire a probe -- from which all
6083 * subsequent probe-context DTrace activity emanates.
6086 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6087 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6089 processorid_t cpuid;
6090 dtrace_icookie_t cookie;
6091 dtrace_probe_t *probe;
6092 dtrace_mstate_t mstate;
6094 dtrace_action_t *act;
6098 volatile uint16_t *flags;
6101 if (panicstr != NULL)
6106 * Kick out immediately if this CPU is still being born (in which case
6107 * curthread will be set to -1) or the current thread can't allow
6108 * probes in its current context.
6110 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6114 cookie = dtrace_interrupt_disable();
6115 probe = dtrace_probes[id - 1];
6117 onintr = CPU_ON_INTR(CPU);
6119 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6120 probe->dtpr_predcache == curthread->t_predcache) {
6122 * We have hit in the predicate cache; we know that
6123 * this predicate would evaluate to be false.
6125 dtrace_interrupt_enable(cookie);
6130 if (panic_quiesce) {
6132 if (panicstr != NULL) {
6135 * We don't trace anything if we're panicking.
6137 dtrace_interrupt_enable(cookie);
6141 now = dtrace_gethrtime();
6142 vtime = dtrace_vtime_references != 0;
6144 if (vtime && curthread->t_dtrace_start)
6145 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6147 mstate.dtms_difo = NULL;
6148 mstate.dtms_probe = probe;
6149 mstate.dtms_strtok = 0;
6150 mstate.dtms_arg[0] = arg0;
6151 mstate.dtms_arg[1] = arg1;
6152 mstate.dtms_arg[2] = arg2;
6153 mstate.dtms_arg[3] = arg3;
6154 mstate.dtms_arg[4] = arg4;
6156 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6158 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6159 dtrace_predicate_t *pred = ecb->dte_predicate;
6160 dtrace_state_t *state = ecb->dte_state;
6161 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6162 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6163 dtrace_vstate_t *vstate = &state->dts_vstate;
6164 dtrace_provider_t *prov = probe->dtpr_provider;
6165 uint64_t tracememsize = 0;
6170 * A little subtlety with the following (seemingly innocuous)
6171 * declaration of the automatic 'val': by looking at the
6172 * code, you might think that it could be declared in the
6173 * action processing loop, below. (That is, it's only used in
6174 * the action processing loop.) However, it must be declared
6175 * out of that scope because in the case of DIF expression
6176 * arguments to aggregating actions, one iteration of the
6177 * action loop will use the last iteration's value.
6181 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6182 *flags &= ~CPU_DTRACE_ERROR;
6184 if (prov == dtrace_provider) {
6186 * If dtrace itself is the provider of this probe,
6187 * we're only going to continue processing the ECB if
6188 * arg0 (the dtrace_state_t) is equal to the ECB's
6189 * creating state. (This prevents disjoint consumers
6190 * from seeing one another's metaprobes.)
6192 if (arg0 != (uint64_t)(uintptr_t)state)
6196 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6198 * We're not currently active. If our provider isn't
6199 * the dtrace pseudo provider, we're not interested.
6201 if (prov != dtrace_provider)
6205 * Now we must further check if we are in the BEGIN
6206 * probe. If we are, we will only continue processing
6207 * if we're still in WARMUP -- if one BEGIN enabling
6208 * has invoked the exit() action, we don't want to
6209 * evaluate subsequent BEGIN enablings.
6211 if (probe->dtpr_id == dtrace_probeid_begin &&
6212 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6213 ASSERT(state->dts_activity ==
6214 DTRACE_ACTIVITY_DRAINING);
6219 if (ecb->dte_cond) {
6221 * If the dte_cond bits indicate that this
6222 * consumer is only allowed to see user-mode firings
6223 * of this probe, call the provider's dtps_usermode()
6224 * entry point to check that the probe was fired
6225 * while in a user context. Skip this ECB if that's
6228 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6229 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6230 probe->dtpr_id, probe->dtpr_arg) == 0)
6235 * This is more subtle than it looks. We have to be
6236 * absolutely certain that CRED() isn't going to
6237 * change out from under us so it's only legit to
6238 * examine that structure if we're in constrained
6239 * situations. Currently, the only times we'll this
6240 * check is if a non-super-user has enabled the
6241 * profile or syscall providers -- providers that
6242 * allow visibility of all processes. For the
6243 * profile case, the check above will ensure that
6244 * we're examining a user context.
6246 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6249 ecb->dte_state->dts_cred.dcr_cred;
6252 ASSERT(s_cr != NULL);
6254 if ((cr = CRED()) == NULL ||
6255 s_cr->cr_uid != cr->cr_uid ||
6256 s_cr->cr_uid != cr->cr_ruid ||
6257 s_cr->cr_uid != cr->cr_suid ||
6258 s_cr->cr_gid != cr->cr_gid ||
6259 s_cr->cr_gid != cr->cr_rgid ||
6260 s_cr->cr_gid != cr->cr_sgid ||
6261 (proc = ttoproc(curthread)) == NULL ||
6262 (proc->p_flag & SNOCD))
6266 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6269 ecb->dte_state->dts_cred.dcr_cred;
6271 ASSERT(s_cr != NULL);
6273 if ((cr = CRED()) == NULL ||
6274 s_cr->cr_zone->zone_id !=
6275 cr->cr_zone->zone_id)
6281 if (now - state->dts_alive > dtrace_deadman_timeout) {
6283 * We seem to be dead. Unless we (a) have kernel
6284 * destructive permissions (b) have explicitly enabled
6285 * destructive actions and (c) destructive actions have
6286 * not been disabled, we're going to transition into
6287 * the KILLED state, from which no further processing
6288 * on this state will be performed.
6290 if (!dtrace_priv_kernel_destructive(state) ||
6291 !state->dts_cred.dcr_destructive ||
6292 dtrace_destructive_disallow) {
6293 void *activity = &state->dts_activity;
6294 dtrace_activity_t current;
6297 current = state->dts_activity;
6298 } while (dtrace_cas32(activity, current,
6299 DTRACE_ACTIVITY_KILLED) != current);
6305 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6306 ecb->dte_alignment, state, &mstate)) < 0)
6309 tomax = buf->dtb_tomax;
6310 ASSERT(tomax != NULL);
6312 if (ecb->dte_size != 0) {
6313 dtrace_rechdr_t dtrh;
6314 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6315 mstate.dtms_timestamp = dtrace_gethrtime();
6316 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6318 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6319 dtrh.dtrh_epid = ecb->dte_epid;
6320 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6321 mstate.dtms_timestamp);
6322 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6325 mstate.dtms_epid = ecb->dte_epid;
6326 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6328 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6329 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6331 mstate.dtms_access = 0;
6334 dtrace_difo_t *dp = pred->dtp_difo;
6337 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6339 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6340 dtrace_cacheid_t cid = probe->dtpr_predcache;
6342 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6344 * Update the predicate cache...
6346 ASSERT(cid == pred->dtp_cacheid);
6347 curthread->t_predcache = cid;
6354 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6355 act != NULL; act = act->dta_next) {
6358 dtrace_recdesc_t *rec = &act->dta_rec;
6360 size = rec->dtrd_size;
6361 valoffs = offs + rec->dtrd_offset;
6363 if (DTRACEACT_ISAGG(act->dta_kind)) {
6365 dtrace_aggregation_t *agg;
6367 agg = (dtrace_aggregation_t *)act;
6369 if ((dp = act->dta_difo) != NULL)
6370 v = dtrace_dif_emulate(dp,
6371 &mstate, vstate, state);
6373 if (*flags & CPU_DTRACE_ERROR)
6377 * Note that we always pass the expression
6378 * value from the previous iteration of the
6379 * action loop. This value will only be used
6380 * if there is an expression argument to the
6381 * aggregating action, denoted by the
6382 * dtag_hasarg field.
6384 dtrace_aggregate(agg, buf,
6385 offs, aggbuf, v, val);
6389 switch (act->dta_kind) {
6390 case DTRACEACT_STOP:
6391 if (dtrace_priv_proc_destructive(state))
6392 dtrace_action_stop();
6395 case DTRACEACT_BREAKPOINT:
6396 if (dtrace_priv_kernel_destructive(state))
6397 dtrace_action_breakpoint(ecb);
6400 case DTRACEACT_PANIC:
6401 if (dtrace_priv_kernel_destructive(state))
6402 dtrace_action_panic(ecb);
6405 case DTRACEACT_STACK:
6406 if (!dtrace_priv_kernel(state))
6409 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6410 size / sizeof (pc_t), probe->dtpr_aframes,
6411 DTRACE_ANCHORED(probe) ? NULL :
6415 case DTRACEACT_JSTACK:
6416 case DTRACEACT_USTACK:
6417 if (!dtrace_priv_proc(state))
6421 * See comment in DIF_VAR_PID.
6423 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6425 int depth = DTRACE_USTACK_NFRAMES(
6428 dtrace_bzero((void *)(tomax + valoffs),
6429 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6430 + depth * sizeof (uint64_t));
6435 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6436 curproc->p_dtrace_helpers != NULL) {
6438 * This is the slow path -- we have
6439 * allocated string space, and we're
6440 * getting the stack of a process that
6441 * has helpers. Call into a separate
6442 * routine to perform this processing.
6444 dtrace_action_ustack(&mstate, state,
6445 (uint64_t *)(tomax + valoffs),
6450 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6451 dtrace_getupcstack((uint64_t *)
6453 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6454 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6464 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6466 if (*flags & CPU_DTRACE_ERROR)
6469 switch (act->dta_kind) {
6470 case DTRACEACT_SPECULATE: {
6471 dtrace_rechdr_t *dtrh;
6473 ASSERT(buf == &state->dts_buffer[cpuid]);
6474 buf = dtrace_speculation_buffer(state,
6478 *flags |= CPU_DTRACE_DROP;
6482 offs = dtrace_buffer_reserve(buf,
6483 ecb->dte_needed, ecb->dte_alignment,
6487 *flags |= CPU_DTRACE_DROP;
6491 tomax = buf->dtb_tomax;
6492 ASSERT(tomax != NULL);
6494 if (ecb->dte_size == 0)
6497 ASSERT3U(ecb->dte_size, >=,
6498 sizeof (dtrace_rechdr_t));
6499 dtrh = ((void *)(tomax + offs));
6500 dtrh->dtrh_epid = ecb->dte_epid;
6502 * When the speculation is committed, all of
6503 * the records in the speculative buffer will
6504 * have their timestamps set to the commit
6505 * time. Until then, it is set to a sentinel
6506 * value, for debugability.
6508 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6512 case DTRACEACT_PRINTM: {
6513 /* The DIF returns a 'memref'. */
6514 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6516 /* Get the size from the memref. */
6520 * Check if the size exceeds the allocated
6523 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6525 *flags |= CPU_DTRACE_DROP;
6529 /* Store the size in the buffer first. */
6530 DTRACE_STORE(uintptr_t, tomax,
6534 * Offset the buffer address to the start
6537 valoffs += sizeof(uintptr_t);
6540 * Reset to the memory address rather than
6541 * the memref array, then let the BYREF
6542 * code below do the work to store the
6543 * memory data in the buffer.
6549 case DTRACEACT_PRINTT: {
6550 /* The DIF returns a 'typeref'. */
6551 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6556 * Get the type string length and round it
6557 * up so that the data that follows is
6558 * aligned for easy access.
6560 size_t typs = strlen((char *) typeref[2]) + 1;
6561 typs = roundup(typs, sizeof(uintptr_t));
6564 *Get the size from the typeref using the
6565 * number of elements and the type size.
6567 size = typeref[1] * typeref[3];
6570 * Check if the size exceeds the allocated
6573 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6575 *flags |= CPU_DTRACE_DROP;
6579 /* Store the size in the buffer first. */
6580 DTRACE_STORE(uintptr_t, tomax,
6582 valoffs += sizeof(uintptr_t);
6584 /* Store the type size in the buffer. */
6585 DTRACE_STORE(uintptr_t, tomax,
6586 valoffs, typeref[3]);
6587 valoffs += sizeof(uintptr_t);
6591 for (s = 0; s < typs; s++) {
6593 c = dtrace_load8(val++);
6595 DTRACE_STORE(uint8_t, tomax,
6600 * Reset to the memory address rather than
6601 * the typeref array, then let the BYREF
6602 * code below do the work to store the
6603 * memory data in the buffer.
6609 case DTRACEACT_CHILL:
6610 if (dtrace_priv_kernel_destructive(state))
6611 dtrace_action_chill(&mstate, val);
6614 case DTRACEACT_RAISE:
6615 if (dtrace_priv_proc_destructive(state))
6616 dtrace_action_raise(val);
6619 case DTRACEACT_COMMIT:
6623 * We need to commit our buffer state.
6626 buf->dtb_offset = offs + ecb->dte_size;
6627 buf = &state->dts_buffer[cpuid];
6628 dtrace_speculation_commit(state, cpuid, val);
6632 case DTRACEACT_DISCARD:
6633 dtrace_speculation_discard(state, cpuid, val);
6636 case DTRACEACT_DIFEXPR:
6637 case DTRACEACT_LIBACT:
6638 case DTRACEACT_PRINTF:
6639 case DTRACEACT_PRINTA:
6640 case DTRACEACT_SYSTEM:
6641 case DTRACEACT_FREOPEN:
6642 case DTRACEACT_TRACEMEM:
6645 case DTRACEACT_TRACEMEM_DYNSIZE:
6651 if (!dtrace_priv_kernel(state))
6655 case DTRACEACT_USYM:
6656 case DTRACEACT_UMOD:
6657 case DTRACEACT_UADDR: {
6659 struct pid *pid = curthread->t_procp->p_pidp;
6662 if (!dtrace_priv_proc(state))
6665 DTRACE_STORE(uint64_t, tomax,
6667 valoffs, (uint64_t)pid->pid_id);
6669 valoffs, (uint64_t) curproc->p_pid);
6671 DTRACE_STORE(uint64_t, tomax,
6672 valoffs + sizeof (uint64_t), val);
6677 case DTRACEACT_EXIT: {
6679 * For the exit action, we are going to attempt
6680 * to atomically set our activity to be
6681 * draining. If this fails (either because
6682 * another CPU has beat us to the exit action,
6683 * or because our current activity is something
6684 * other than ACTIVE or WARMUP), we will
6685 * continue. This assures that the exit action
6686 * can be successfully recorded at most once
6687 * when we're in the ACTIVE state. If we're
6688 * encountering the exit() action while in
6689 * COOLDOWN, however, we want to honor the new
6690 * status code. (We know that we're the only
6691 * thread in COOLDOWN, so there is no race.)
6693 void *activity = &state->dts_activity;
6694 dtrace_activity_t current = state->dts_activity;
6696 if (current == DTRACE_ACTIVITY_COOLDOWN)
6699 if (current != DTRACE_ACTIVITY_WARMUP)
6700 current = DTRACE_ACTIVITY_ACTIVE;
6702 if (dtrace_cas32(activity, current,
6703 DTRACE_ACTIVITY_DRAINING) != current) {
6704 *flags |= CPU_DTRACE_DROP;
6715 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6716 uintptr_t end = valoffs + size;
6718 if (tracememsize != 0 &&
6719 valoffs + tracememsize < end) {
6720 end = valoffs + tracememsize;
6724 if (!dtrace_vcanload((void *)(uintptr_t)val,
6725 &dp->dtdo_rtype, &mstate, vstate))
6729 * If this is a string, we're going to only
6730 * load until we find the zero byte -- after
6731 * which we'll store zero bytes.
6733 if (dp->dtdo_rtype.dtdt_kind ==
6736 int intuple = act->dta_intuple;
6739 for (s = 0; s < size; s++) {
6741 c = dtrace_load8(val++);
6743 DTRACE_STORE(uint8_t, tomax,
6746 if (c == '\0' && intuple)
6753 while (valoffs < end) {
6754 DTRACE_STORE(uint8_t, tomax, valoffs++,
6755 dtrace_load8(val++));
6765 case sizeof (uint8_t):
6766 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6768 case sizeof (uint16_t):
6769 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6771 case sizeof (uint32_t):
6772 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6774 case sizeof (uint64_t):
6775 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6779 * Any other size should have been returned by
6780 * reference, not by value.
6787 if (*flags & CPU_DTRACE_DROP)
6790 if (*flags & CPU_DTRACE_FAULT) {
6792 dtrace_action_t *err;
6796 if (probe->dtpr_id == dtrace_probeid_error) {
6798 * There's nothing we can do -- we had an
6799 * error on the error probe. We bump an
6800 * error counter to at least indicate that
6801 * this condition happened.
6803 dtrace_error(&state->dts_dblerrors);
6809 * Before recursing on dtrace_probe(), we
6810 * need to explicitly clear out our start
6811 * time to prevent it from being accumulated
6812 * into t_dtrace_vtime.
6814 curthread->t_dtrace_start = 0;
6818 * Iterate over the actions to figure out which action
6819 * we were processing when we experienced the error.
6820 * Note that act points _past_ the faulting action; if
6821 * act is ecb->dte_action, the fault was in the
6822 * predicate, if it's ecb->dte_action->dta_next it's
6823 * in action #1, and so on.
6825 for (err = ecb->dte_action, ndx = 0;
6826 err != act; err = err->dta_next, ndx++)
6829 dtrace_probe_error(state, ecb->dte_epid, ndx,
6830 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6831 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6832 cpu_core[cpuid].cpuc_dtrace_illval);
6838 buf->dtb_offset = offs + ecb->dte_size;
6842 curthread->t_dtrace_start = dtrace_gethrtime();
6844 dtrace_interrupt_enable(cookie);
6848 * DTrace Probe Hashing Functions
6850 * The functions in this section (and indeed, the functions in remaining
6851 * sections) are not _called_ from probe context. (Any exceptions to this are
6852 * marked with a "Note:".) Rather, they are called from elsewhere in the
6853 * DTrace framework to look-up probes in, add probes to and remove probes from
6854 * the DTrace probe hashes. (Each probe is hashed by each element of the
6855 * probe tuple -- allowing for fast lookups, regardless of what was
6859 dtrace_hash_str(const char *p)
6865 hval = (hval << 4) + *p++;
6866 if ((g = (hval & 0xf0000000)) != 0)
6873 static dtrace_hash_t *
6874 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6876 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6878 hash->dth_stroffs = stroffs;
6879 hash->dth_nextoffs = nextoffs;
6880 hash->dth_prevoffs = prevoffs;
6883 hash->dth_mask = hash->dth_size - 1;
6885 hash->dth_tab = kmem_zalloc(hash->dth_size *
6886 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6892 dtrace_hash_destroy(dtrace_hash_t *hash)
6897 for (i = 0; i < hash->dth_size; i++)
6898 ASSERT(hash->dth_tab[i] == NULL);
6901 kmem_free(hash->dth_tab,
6902 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6903 kmem_free(hash, sizeof (dtrace_hash_t));
6907 dtrace_hash_resize(dtrace_hash_t *hash)
6909 int size = hash->dth_size, i, ndx;
6910 int new_size = hash->dth_size << 1;
6911 int new_mask = new_size - 1;
6912 dtrace_hashbucket_t **new_tab, *bucket, *next;
6914 ASSERT((new_size & new_mask) == 0);
6916 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6918 for (i = 0; i < size; i++) {
6919 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6920 dtrace_probe_t *probe = bucket->dthb_chain;
6922 ASSERT(probe != NULL);
6923 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6925 next = bucket->dthb_next;
6926 bucket->dthb_next = new_tab[ndx];
6927 new_tab[ndx] = bucket;
6931 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6932 hash->dth_tab = new_tab;
6933 hash->dth_size = new_size;
6934 hash->dth_mask = new_mask;
6938 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6940 int hashval = DTRACE_HASHSTR(hash, new);
6941 int ndx = hashval & hash->dth_mask;
6942 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6943 dtrace_probe_t **nextp, **prevp;
6945 for (; bucket != NULL; bucket = bucket->dthb_next) {
6946 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6950 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6951 dtrace_hash_resize(hash);
6952 dtrace_hash_add(hash, new);
6956 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6957 bucket->dthb_next = hash->dth_tab[ndx];
6958 hash->dth_tab[ndx] = bucket;
6959 hash->dth_nbuckets++;
6962 nextp = DTRACE_HASHNEXT(hash, new);
6963 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6964 *nextp = bucket->dthb_chain;
6966 if (bucket->dthb_chain != NULL) {
6967 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6968 ASSERT(*prevp == NULL);
6972 bucket->dthb_chain = new;
6976 static dtrace_probe_t *
6977 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6979 int hashval = DTRACE_HASHSTR(hash, template);
6980 int ndx = hashval & hash->dth_mask;
6981 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6983 for (; bucket != NULL; bucket = bucket->dthb_next) {
6984 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6985 return (bucket->dthb_chain);
6992 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6994 int hashval = DTRACE_HASHSTR(hash, template);
6995 int ndx = hashval & hash->dth_mask;
6996 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6998 for (; bucket != NULL; bucket = bucket->dthb_next) {
6999 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7000 return (bucket->dthb_len);
7007 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7009 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7010 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7012 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7013 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7016 * Find the bucket that we're removing this probe from.
7018 for (; bucket != NULL; bucket = bucket->dthb_next) {
7019 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7023 ASSERT(bucket != NULL);
7025 if (*prevp == NULL) {
7026 if (*nextp == NULL) {
7028 * The removed probe was the only probe on this
7029 * bucket; we need to remove the bucket.
7031 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7033 ASSERT(bucket->dthb_chain == probe);
7037 hash->dth_tab[ndx] = bucket->dthb_next;
7039 while (b->dthb_next != bucket)
7041 b->dthb_next = bucket->dthb_next;
7044 ASSERT(hash->dth_nbuckets > 0);
7045 hash->dth_nbuckets--;
7046 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7050 bucket->dthb_chain = *nextp;
7052 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7056 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7060 * DTrace Utility Functions
7062 * These are random utility functions that are _not_ called from probe context.
7065 dtrace_badattr(const dtrace_attribute_t *a)
7067 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7068 a->dtat_data > DTRACE_STABILITY_MAX ||
7069 a->dtat_class > DTRACE_CLASS_MAX);
7073 * Return a duplicate copy of a string. If the specified string is NULL,
7074 * this function returns a zero-length string.
7077 dtrace_strdup(const char *str)
7079 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7082 (void) strcpy(new, str);
7087 #define DTRACE_ISALPHA(c) \
7088 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7091 dtrace_badname(const char *s)
7095 if (s == NULL || (c = *s++) == '\0')
7098 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7101 while ((c = *s++) != '\0') {
7102 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7103 c != '-' && c != '_' && c != '.' && c != '`')
7111 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7116 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7118 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7120 priv = DTRACE_PRIV_ALL;
7122 *uidp = crgetuid(cr);
7123 *zoneidp = crgetzoneid(cr);
7126 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7127 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7128 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7129 priv |= DTRACE_PRIV_USER;
7130 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7131 priv |= DTRACE_PRIV_PROC;
7132 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7133 priv |= DTRACE_PRIV_OWNER;
7134 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7135 priv |= DTRACE_PRIV_ZONEOWNER;
7138 priv = DTRACE_PRIV_ALL;
7144 #ifdef DTRACE_ERRDEBUG
7146 dtrace_errdebug(const char *str)
7148 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
7151 mutex_enter(&dtrace_errlock);
7152 dtrace_errlast = str;
7153 dtrace_errthread = curthread;
7155 while (occupied++ < DTRACE_ERRHASHSZ) {
7156 if (dtrace_errhash[hval].dter_msg == str) {
7157 dtrace_errhash[hval].dter_count++;
7161 if (dtrace_errhash[hval].dter_msg != NULL) {
7162 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7166 dtrace_errhash[hval].dter_msg = str;
7167 dtrace_errhash[hval].dter_count = 1;
7171 panic("dtrace: undersized error hash");
7173 mutex_exit(&dtrace_errlock);
7178 * DTrace Matching Functions
7180 * These functions are used to match groups of probes, given some elements of
7181 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7184 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7187 if (priv != DTRACE_PRIV_ALL) {
7188 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7189 uint32_t match = priv & ppriv;
7192 * No PRIV_DTRACE_* privileges...
7194 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7195 DTRACE_PRIV_KERNEL)) == 0)
7199 * No matching bits, but there were bits to match...
7201 if (match == 0 && ppriv != 0)
7205 * Need to have permissions to the process, but don't...
7207 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7208 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7213 * Need to be in the same zone unless we possess the
7214 * privilege to examine all zones.
7216 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7217 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7226 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7227 * consists of input pattern strings and an ops-vector to evaluate them.
7228 * This function returns >0 for match, 0 for no match, and <0 for error.
7231 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7232 uint32_t priv, uid_t uid, zoneid_t zoneid)
7234 dtrace_provider_t *pvp = prp->dtpr_provider;
7237 if (pvp->dtpv_defunct)
7240 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7243 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7246 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7249 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7252 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7259 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7260 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7261 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7262 * In addition, all of the recursion cases except for '*' matching have been
7263 * unwound. For '*', we still implement recursive evaluation, but a depth
7264 * counter is maintained and matching is aborted if we recurse too deep.
7265 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7268 dtrace_match_glob(const char *s, const char *p, int depth)
7274 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7278 s = ""; /* treat NULL as empty string */
7287 if ((c = *p++) == '\0')
7288 return (s1 == '\0');
7292 int ok = 0, notflag = 0;
7303 if ((c = *p++) == '\0')
7307 if (c == '-' && lc != '\0' && *p != ']') {
7308 if ((c = *p++) == '\0')
7310 if (c == '\\' && (c = *p++) == '\0')
7314 if (s1 < lc || s1 > c)
7318 } else if (lc <= s1 && s1 <= c)
7321 } else if (c == '\\' && (c = *p++) == '\0')
7324 lc = c; /* save left-hand 'c' for next iteration */
7334 if ((c = *p++) == '\0')
7346 if ((c = *p++) == '\0')
7362 p++; /* consecutive *'s are identical to a single one */
7367 for (s = olds; *s != '\0'; s++) {
7368 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7378 dtrace_match_string(const char *s, const char *p, int depth)
7380 return (s != NULL && strcmp(s, p) == 0);
7385 dtrace_match_nul(const char *s, const char *p, int depth)
7387 return (1); /* always match the empty pattern */
7392 dtrace_match_nonzero(const char *s, const char *p, int depth)
7394 return (s != NULL && s[0] != '\0');
7398 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7399 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7401 dtrace_probe_t template, *probe;
7402 dtrace_hash_t *hash = NULL;
7403 int len, best = INT_MAX, nmatched = 0;
7406 ASSERT(MUTEX_HELD(&dtrace_lock));
7409 * If the probe ID is specified in the key, just lookup by ID and
7410 * invoke the match callback once if a matching probe is found.
7412 if (pkp->dtpk_id != DTRACE_IDNONE) {
7413 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7414 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7415 (void) (*matched)(probe, arg);
7421 template.dtpr_mod = (char *)pkp->dtpk_mod;
7422 template.dtpr_func = (char *)pkp->dtpk_func;
7423 template.dtpr_name = (char *)pkp->dtpk_name;
7426 * We want to find the most distinct of the module name, function
7427 * name, and name. So for each one that is not a glob pattern or
7428 * empty string, we perform a lookup in the corresponding hash and
7429 * use the hash table with the fewest collisions to do our search.
7431 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7432 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7434 hash = dtrace_bymod;
7437 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7438 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7440 hash = dtrace_byfunc;
7443 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7444 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7446 hash = dtrace_byname;
7450 * If we did not select a hash table, iterate over every probe and
7451 * invoke our callback for each one that matches our input probe key.
7454 for (i = 0; i < dtrace_nprobes; i++) {
7455 if ((probe = dtrace_probes[i]) == NULL ||
7456 dtrace_match_probe(probe, pkp, priv, uid,
7462 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7470 * If we selected a hash table, iterate over each probe of the same key
7471 * name and invoke the callback for every probe that matches the other
7472 * attributes of our input probe key.
7474 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7475 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7477 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7482 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7490 * Return the function pointer dtrace_probecmp() should use to compare the
7491 * specified pattern with a string. For NULL or empty patterns, we select
7492 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7493 * For non-empty non-glob strings, we use dtrace_match_string().
7495 static dtrace_probekey_f *
7496 dtrace_probekey_func(const char *p)
7500 if (p == NULL || *p == '\0')
7501 return (&dtrace_match_nul);
7503 while ((c = *p++) != '\0') {
7504 if (c == '[' || c == '?' || c == '*' || c == '\\')
7505 return (&dtrace_match_glob);
7508 return (&dtrace_match_string);
7512 * Build a probe comparison key for use with dtrace_match_probe() from the
7513 * given probe description. By convention, a null key only matches anchored
7514 * probes: if each field is the empty string, reset dtpk_fmatch to
7515 * dtrace_match_nonzero().
7518 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7520 pkp->dtpk_prov = pdp->dtpd_provider;
7521 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7523 pkp->dtpk_mod = pdp->dtpd_mod;
7524 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7526 pkp->dtpk_func = pdp->dtpd_func;
7527 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7529 pkp->dtpk_name = pdp->dtpd_name;
7530 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7532 pkp->dtpk_id = pdp->dtpd_id;
7534 if (pkp->dtpk_id == DTRACE_IDNONE &&
7535 pkp->dtpk_pmatch == &dtrace_match_nul &&
7536 pkp->dtpk_mmatch == &dtrace_match_nul &&
7537 pkp->dtpk_fmatch == &dtrace_match_nul &&
7538 pkp->dtpk_nmatch == &dtrace_match_nul)
7539 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7543 * DTrace Provider-to-Framework API Functions
7545 * These functions implement much of the Provider-to-Framework API, as
7546 * described in <sys/dtrace.h>. The parts of the API not in this section are
7547 * the functions in the API for probe management (found below), and
7548 * dtrace_probe() itself (found above).
7552 * Register the calling provider with the DTrace framework. This should
7553 * generally be called by DTrace providers in their attach(9E) entry point.
7556 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7557 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7559 dtrace_provider_t *provider;
7561 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7562 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7563 "arguments", name ? name : "<NULL>");
7567 if (name[0] == '\0' || dtrace_badname(name)) {
7568 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7569 "provider name", name);
7573 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7574 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7575 pops->dtps_destroy == NULL ||
7576 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7577 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7578 "provider ops", name);
7582 if (dtrace_badattr(&pap->dtpa_provider) ||
7583 dtrace_badattr(&pap->dtpa_mod) ||
7584 dtrace_badattr(&pap->dtpa_func) ||
7585 dtrace_badattr(&pap->dtpa_name) ||
7586 dtrace_badattr(&pap->dtpa_args)) {
7587 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7588 "provider attributes", name);
7592 if (priv & ~DTRACE_PRIV_ALL) {
7593 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7594 "privilege attributes", name);
7598 if ((priv & DTRACE_PRIV_KERNEL) &&
7599 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7600 pops->dtps_usermode == NULL) {
7601 cmn_err(CE_WARN, "failed to register provider '%s': need "
7602 "dtps_usermode() op for given privilege attributes", name);
7606 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7607 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7608 (void) strcpy(provider->dtpv_name, name);
7610 provider->dtpv_attr = *pap;
7611 provider->dtpv_priv.dtpp_flags = priv;
7613 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7614 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7616 provider->dtpv_pops = *pops;
7618 if (pops->dtps_provide == NULL) {
7619 ASSERT(pops->dtps_provide_module != NULL);
7620 provider->dtpv_pops.dtps_provide =
7621 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7624 if (pops->dtps_provide_module == NULL) {
7625 ASSERT(pops->dtps_provide != NULL);
7626 provider->dtpv_pops.dtps_provide_module =
7627 (void (*)(void *, modctl_t *))dtrace_nullop;
7630 if (pops->dtps_suspend == NULL) {
7631 ASSERT(pops->dtps_resume == NULL);
7632 provider->dtpv_pops.dtps_suspend =
7633 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7634 provider->dtpv_pops.dtps_resume =
7635 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7638 provider->dtpv_arg = arg;
7639 *idp = (dtrace_provider_id_t)provider;
7641 if (pops == &dtrace_provider_ops) {
7642 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7643 ASSERT(MUTEX_HELD(&dtrace_lock));
7644 ASSERT(dtrace_anon.dta_enabling == NULL);
7647 * We make sure that the DTrace provider is at the head of
7648 * the provider chain.
7650 provider->dtpv_next = dtrace_provider;
7651 dtrace_provider = provider;
7655 mutex_enter(&dtrace_provider_lock);
7656 mutex_enter(&dtrace_lock);
7659 * If there is at least one provider registered, we'll add this
7660 * provider after the first provider.
7662 if (dtrace_provider != NULL) {
7663 provider->dtpv_next = dtrace_provider->dtpv_next;
7664 dtrace_provider->dtpv_next = provider;
7666 dtrace_provider = provider;
7669 if (dtrace_retained != NULL) {
7670 dtrace_enabling_provide(provider);
7673 * Now we need to call dtrace_enabling_matchall() -- which
7674 * will acquire cpu_lock and dtrace_lock. We therefore need
7675 * to drop all of our locks before calling into it...
7677 mutex_exit(&dtrace_lock);
7678 mutex_exit(&dtrace_provider_lock);
7679 dtrace_enabling_matchall();
7684 mutex_exit(&dtrace_lock);
7685 mutex_exit(&dtrace_provider_lock);
7691 * Unregister the specified provider from the DTrace framework. This should
7692 * generally be called by DTrace providers in their detach(9E) entry point.
7695 dtrace_unregister(dtrace_provider_id_t id)
7697 dtrace_provider_t *old = (dtrace_provider_t *)id;
7698 dtrace_provider_t *prev = NULL;
7699 int i, self = 0, noreap = 0;
7700 dtrace_probe_t *probe, *first = NULL;
7702 if (old->dtpv_pops.dtps_enable ==
7703 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7705 * If DTrace itself is the provider, we're called with locks
7708 ASSERT(old == dtrace_provider);
7710 ASSERT(dtrace_devi != NULL);
7712 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7713 ASSERT(MUTEX_HELD(&dtrace_lock));
7716 if (dtrace_provider->dtpv_next != NULL) {
7718 * There's another provider here; return failure.
7723 mutex_enter(&dtrace_provider_lock);
7725 mutex_enter(&mod_lock);
7727 mutex_enter(&dtrace_lock);
7731 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7732 * probes, we refuse to let providers slither away, unless this
7733 * provider has already been explicitly invalidated.
7735 if (!old->dtpv_defunct &&
7736 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7737 dtrace_anon.dta_state->dts_necbs > 0))) {
7739 mutex_exit(&dtrace_lock);
7741 mutex_exit(&mod_lock);
7743 mutex_exit(&dtrace_provider_lock);
7749 * Attempt to destroy the probes associated with this provider.
7751 for (i = 0; i < dtrace_nprobes; i++) {
7752 if ((probe = dtrace_probes[i]) == NULL)
7755 if (probe->dtpr_provider != old)
7758 if (probe->dtpr_ecb == NULL)
7762 * If we are trying to unregister a defunct provider, and the
7763 * provider was made defunct within the interval dictated by
7764 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7765 * attempt to reap our enablings. To denote that the provider
7766 * should reattempt to unregister itself at some point in the
7767 * future, we will return a differentiable error code (EAGAIN
7768 * instead of EBUSY) in this case.
7770 if (dtrace_gethrtime() - old->dtpv_defunct >
7771 dtrace_unregister_defunct_reap)
7775 mutex_exit(&dtrace_lock);
7777 mutex_exit(&mod_lock);
7779 mutex_exit(&dtrace_provider_lock);
7785 (void) taskq_dispatch(dtrace_taskq,
7786 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7792 * All of the probes for this provider are disabled; we can safely
7793 * remove all of them from their hash chains and from the probe array.
7795 for (i = 0; i < dtrace_nprobes; i++) {
7796 if ((probe = dtrace_probes[i]) == NULL)
7799 if (probe->dtpr_provider != old)
7802 dtrace_probes[i] = NULL;
7804 dtrace_hash_remove(dtrace_bymod, probe);
7805 dtrace_hash_remove(dtrace_byfunc, probe);
7806 dtrace_hash_remove(dtrace_byname, probe);
7808 if (first == NULL) {
7810 probe->dtpr_nextmod = NULL;
7812 probe->dtpr_nextmod = first;
7818 * The provider's probes have been removed from the hash chains and
7819 * from the probe array. Now issue a dtrace_sync() to be sure that
7820 * everyone has cleared out from any probe array processing.
7824 for (probe = first; probe != NULL; probe = first) {
7825 first = probe->dtpr_nextmod;
7827 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7829 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7830 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7831 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7833 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7835 free_unr(dtrace_arena, probe->dtpr_id);
7837 kmem_free(probe, sizeof (dtrace_probe_t));
7840 if ((prev = dtrace_provider) == old) {
7842 ASSERT(self || dtrace_devi == NULL);
7843 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7845 dtrace_provider = old->dtpv_next;
7847 while (prev != NULL && prev->dtpv_next != old)
7848 prev = prev->dtpv_next;
7851 panic("attempt to unregister non-existent "
7852 "dtrace provider %p\n", (void *)id);
7855 prev->dtpv_next = old->dtpv_next;
7859 mutex_exit(&dtrace_lock);
7861 mutex_exit(&mod_lock);
7863 mutex_exit(&dtrace_provider_lock);
7866 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7867 kmem_free(old, sizeof (dtrace_provider_t));
7873 * Invalidate the specified provider. All subsequent probe lookups for the
7874 * specified provider will fail, but its probes will not be removed.
7877 dtrace_invalidate(dtrace_provider_id_t id)
7879 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7881 ASSERT(pvp->dtpv_pops.dtps_enable !=
7882 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7884 mutex_enter(&dtrace_provider_lock);
7885 mutex_enter(&dtrace_lock);
7887 pvp->dtpv_defunct = dtrace_gethrtime();
7889 mutex_exit(&dtrace_lock);
7890 mutex_exit(&dtrace_provider_lock);
7894 * Indicate whether or not DTrace has attached.
7897 dtrace_attached(void)
7900 * dtrace_provider will be non-NULL iff the DTrace driver has
7901 * attached. (It's non-NULL because DTrace is always itself a
7904 return (dtrace_provider != NULL);
7908 * Remove all the unenabled probes for the given provider. This function is
7909 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7910 * -- just as many of its associated probes as it can.
7913 dtrace_condense(dtrace_provider_id_t id)
7915 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7917 dtrace_probe_t *probe;
7920 * Make sure this isn't the dtrace provider itself.
7922 ASSERT(prov->dtpv_pops.dtps_enable !=
7923 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7925 mutex_enter(&dtrace_provider_lock);
7926 mutex_enter(&dtrace_lock);
7929 * Attempt to destroy the probes associated with this provider.
7931 for (i = 0; i < dtrace_nprobes; i++) {
7932 if ((probe = dtrace_probes[i]) == NULL)
7935 if (probe->dtpr_provider != prov)
7938 if (probe->dtpr_ecb != NULL)
7941 dtrace_probes[i] = NULL;
7943 dtrace_hash_remove(dtrace_bymod, probe);
7944 dtrace_hash_remove(dtrace_byfunc, probe);
7945 dtrace_hash_remove(dtrace_byname, probe);
7947 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7949 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7950 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7951 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7952 kmem_free(probe, sizeof (dtrace_probe_t));
7954 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7956 free_unr(dtrace_arena, i + 1);
7960 mutex_exit(&dtrace_lock);
7961 mutex_exit(&dtrace_provider_lock);
7967 * DTrace Probe Management Functions
7969 * The functions in this section perform the DTrace probe management,
7970 * including functions to create probes, look-up probes, and call into the
7971 * providers to request that probes be provided. Some of these functions are
7972 * in the Provider-to-Framework API; these functions can be identified by the
7973 * fact that they are not declared "static".
7977 * Create a probe with the specified module name, function name, and name.
7980 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7981 const char *func, const char *name, int aframes, void *arg)
7983 dtrace_probe_t *probe, **probes;
7984 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7987 if (provider == dtrace_provider) {
7988 ASSERT(MUTEX_HELD(&dtrace_lock));
7990 mutex_enter(&dtrace_lock);
7994 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7995 VM_BESTFIT | VM_SLEEP);
7997 id = alloc_unr(dtrace_arena);
7999 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8001 probe->dtpr_id = id;
8002 probe->dtpr_gen = dtrace_probegen++;
8003 probe->dtpr_mod = dtrace_strdup(mod);
8004 probe->dtpr_func = dtrace_strdup(func);
8005 probe->dtpr_name = dtrace_strdup(name);
8006 probe->dtpr_arg = arg;
8007 probe->dtpr_aframes = aframes;
8008 probe->dtpr_provider = provider;
8010 dtrace_hash_add(dtrace_bymod, probe);
8011 dtrace_hash_add(dtrace_byfunc, probe);
8012 dtrace_hash_add(dtrace_byname, probe);
8014 if (id - 1 >= dtrace_nprobes) {
8015 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8016 size_t nsize = osize << 1;
8020 ASSERT(dtrace_probes == NULL);
8021 nsize = sizeof (dtrace_probe_t *);
8024 probes = kmem_zalloc(nsize, KM_SLEEP);
8026 if (dtrace_probes == NULL) {
8028 dtrace_probes = probes;
8031 dtrace_probe_t **oprobes = dtrace_probes;
8033 bcopy(oprobes, probes, osize);
8034 dtrace_membar_producer();
8035 dtrace_probes = probes;
8040 * All CPUs are now seeing the new probes array; we can
8041 * safely free the old array.
8043 kmem_free(oprobes, osize);
8044 dtrace_nprobes <<= 1;
8047 ASSERT(id - 1 < dtrace_nprobes);
8050 ASSERT(dtrace_probes[id - 1] == NULL);
8051 dtrace_probes[id - 1] = probe;
8053 if (provider != dtrace_provider)
8054 mutex_exit(&dtrace_lock);
8059 static dtrace_probe_t *
8060 dtrace_probe_lookup_id(dtrace_id_t id)
8062 ASSERT(MUTEX_HELD(&dtrace_lock));
8064 if (id == 0 || id > dtrace_nprobes)
8067 return (dtrace_probes[id - 1]);
8071 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8073 *((dtrace_id_t *)arg) = probe->dtpr_id;
8075 return (DTRACE_MATCH_DONE);
8079 * Look up a probe based on provider and one or more of module name, function
8080 * name and probe name.
8083 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
8084 char *func, char *name)
8086 dtrace_probekey_t pkey;
8090 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8091 pkey.dtpk_pmatch = &dtrace_match_string;
8092 pkey.dtpk_mod = mod;
8093 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8094 pkey.dtpk_func = func;
8095 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8096 pkey.dtpk_name = name;
8097 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8098 pkey.dtpk_id = DTRACE_IDNONE;
8100 mutex_enter(&dtrace_lock);
8101 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8102 dtrace_probe_lookup_match, &id);
8103 mutex_exit(&dtrace_lock);
8105 ASSERT(match == 1 || match == 0);
8106 return (match ? id : 0);
8110 * Returns the probe argument associated with the specified probe.
8113 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8115 dtrace_probe_t *probe;
8118 mutex_enter(&dtrace_lock);
8120 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8121 probe->dtpr_provider == (dtrace_provider_t *)id)
8122 rval = probe->dtpr_arg;
8124 mutex_exit(&dtrace_lock);
8130 * Copy a probe into a probe description.
8133 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8135 bzero(pdp, sizeof (dtrace_probedesc_t));
8136 pdp->dtpd_id = prp->dtpr_id;
8138 (void) strncpy(pdp->dtpd_provider,
8139 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8141 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8142 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8143 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8147 * Called to indicate that a probe -- or probes -- should be provided by a
8148 * specfied provider. If the specified description is NULL, the provider will
8149 * be told to provide all of its probes. (This is done whenever a new
8150 * consumer comes along, or whenever a retained enabling is to be matched.) If
8151 * the specified description is non-NULL, the provider is given the
8152 * opportunity to dynamically provide the specified probe, allowing providers
8153 * to support the creation of probes on-the-fly. (So-called _autocreated_
8154 * probes.) If the provider is NULL, the operations will be applied to all
8155 * providers; if the provider is non-NULL the operations will only be applied
8156 * to the specified provider. The dtrace_provider_lock must be held, and the
8157 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8158 * will need to grab the dtrace_lock when it reenters the framework through
8159 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8162 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8169 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8173 prv = dtrace_provider;
8178 * First, call the blanket provide operation.
8180 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8184 * Now call the per-module provide operation. We will grab
8185 * mod_lock to prevent the list from being modified. Note
8186 * that this also prevents the mod_busy bits from changing.
8187 * (mod_busy can only be changed with mod_lock held.)
8189 mutex_enter(&mod_lock);
8193 if (ctl->mod_busy || ctl->mod_mp == NULL)
8196 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8198 } while ((ctl = ctl->mod_next) != &modules);
8200 mutex_exit(&mod_lock);
8202 } while (all && (prv = prv->dtpv_next) != NULL);
8207 * Iterate over each probe, and call the Framework-to-Provider API function
8211 dtrace_probe_foreach(uintptr_t offs)
8213 dtrace_provider_t *prov;
8214 void (*func)(void *, dtrace_id_t, void *);
8215 dtrace_probe_t *probe;
8216 dtrace_icookie_t cookie;
8220 * We disable interrupts to walk through the probe array. This is
8221 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8222 * won't see stale data.
8224 cookie = dtrace_interrupt_disable();
8226 for (i = 0; i < dtrace_nprobes; i++) {
8227 if ((probe = dtrace_probes[i]) == NULL)
8230 if (probe->dtpr_ecb == NULL) {
8232 * This probe isn't enabled -- don't call the function.
8237 prov = probe->dtpr_provider;
8238 func = *((void(**)(void *, dtrace_id_t, void *))
8239 ((uintptr_t)&prov->dtpv_pops + offs));
8241 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8244 dtrace_interrupt_enable(cookie);
8249 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8251 dtrace_probekey_t pkey;
8256 ASSERT(MUTEX_HELD(&dtrace_lock));
8257 dtrace_ecb_create_cache = NULL;
8261 * If we're passed a NULL description, we're being asked to
8262 * create an ECB with a NULL probe.
8264 (void) dtrace_ecb_create_enable(NULL, enab);
8268 dtrace_probekey(desc, &pkey);
8269 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8270 &priv, &uid, &zoneid);
8272 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8277 * DTrace Helper Provider Functions
8280 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8282 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8283 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8284 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8288 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8289 const dof_provider_t *dofprov, char *strtab)
8291 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8292 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8293 dofprov->dofpv_provattr);
8294 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8295 dofprov->dofpv_modattr);
8296 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8297 dofprov->dofpv_funcattr);
8298 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8299 dofprov->dofpv_nameattr);
8300 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8301 dofprov->dofpv_argsattr);
8305 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8307 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8308 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8309 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8310 dof_provider_t *provider;
8312 uint32_t *off, *enoff;
8316 dtrace_helper_provdesc_t dhpv;
8317 dtrace_helper_probedesc_t dhpb;
8318 dtrace_meta_t *meta = dtrace_meta_pid;
8319 dtrace_mops_t *mops = &meta->dtm_mops;
8322 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8323 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8324 provider->dofpv_strtab * dof->dofh_secsize);
8325 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8326 provider->dofpv_probes * dof->dofh_secsize);
8327 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8328 provider->dofpv_prargs * dof->dofh_secsize);
8329 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8330 provider->dofpv_proffs * dof->dofh_secsize);
8332 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8333 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8334 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8338 * See dtrace_helper_provider_validate().
8340 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8341 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8342 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8343 provider->dofpv_prenoffs * dof->dofh_secsize);
8344 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8347 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8350 * Create the provider.
8352 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8354 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8360 * Create the probes.
8362 for (i = 0; i < nprobes; i++) {
8363 probe = (dof_probe_t *)(uintptr_t)(daddr +
8364 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8366 dhpb.dthpb_mod = dhp->dofhp_mod;
8367 dhpb.dthpb_func = strtab + probe->dofpr_func;
8368 dhpb.dthpb_name = strtab + probe->dofpr_name;
8369 dhpb.dthpb_base = probe->dofpr_addr;
8370 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8371 dhpb.dthpb_noffs = probe->dofpr_noffs;
8372 if (enoff != NULL) {
8373 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8374 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8376 dhpb.dthpb_enoffs = NULL;
8377 dhpb.dthpb_nenoffs = 0;
8379 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8380 dhpb.dthpb_nargc = probe->dofpr_nargc;
8381 dhpb.dthpb_xargc = probe->dofpr_xargc;
8382 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8383 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8385 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8390 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8392 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8393 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8396 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8398 for (i = 0; i < dof->dofh_secnum; i++) {
8399 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8400 dof->dofh_secoff + i * dof->dofh_secsize);
8402 if (sec->dofs_type != DOF_SECT_PROVIDER)
8405 dtrace_helper_provide_one(dhp, sec, pid);
8409 * We may have just created probes, so we must now rematch against
8410 * any retained enablings. Note that this call will acquire both
8411 * cpu_lock and dtrace_lock; the fact that we are holding
8412 * dtrace_meta_lock now is what defines the ordering with respect to
8413 * these three locks.
8415 dtrace_enabling_matchall();
8419 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8421 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8422 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8424 dof_provider_t *provider;
8426 dtrace_helper_provdesc_t dhpv;
8427 dtrace_meta_t *meta = dtrace_meta_pid;
8428 dtrace_mops_t *mops = &meta->dtm_mops;
8430 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8431 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8432 provider->dofpv_strtab * dof->dofh_secsize);
8434 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8437 * Create the provider.
8439 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8441 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8447 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8449 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8450 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8453 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8455 for (i = 0; i < dof->dofh_secnum; i++) {
8456 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8457 dof->dofh_secoff + i * dof->dofh_secsize);
8459 if (sec->dofs_type != DOF_SECT_PROVIDER)
8462 dtrace_helper_provider_remove_one(dhp, sec, pid);
8467 * DTrace Meta Provider-to-Framework API Functions
8469 * These functions implement the Meta Provider-to-Framework API, as described
8470 * in <sys/dtrace.h>.
8473 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8474 dtrace_meta_provider_id_t *idp)
8476 dtrace_meta_t *meta;
8477 dtrace_helpers_t *help, *next;
8480 *idp = DTRACE_METAPROVNONE;
8483 * We strictly don't need the name, but we hold onto it for
8484 * debuggability. All hail error queues!
8487 cmn_err(CE_WARN, "failed to register meta-provider: "
8493 mops->dtms_create_probe == NULL ||
8494 mops->dtms_provide_pid == NULL ||
8495 mops->dtms_remove_pid == NULL) {
8496 cmn_err(CE_WARN, "failed to register meta-register %s: "
8497 "invalid ops", name);
8501 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8502 meta->dtm_mops = *mops;
8503 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8504 (void) strcpy(meta->dtm_name, name);
8505 meta->dtm_arg = arg;
8507 mutex_enter(&dtrace_meta_lock);
8508 mutex_enter(&dtrace_lock);
8510 if (dtrace_meta_pid != NULL) {
8511 mutex_exit(&dtrace_lock);
8512 mutex_exit(&dtrace_meta_lock);
8513 cmn_err(CE_WARN, "failed to register meta-register %s: "
8514 "user-land meta-provider exists", name);
8515 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8516 kmem_free(meta, sizeof (dtrace_meta_t));
8520 dtrace_meta_pid = meta;
8521 *idp = (dtrace_meta_provider_id_t)meta;
8524 * If there are providers and probes ready to go, pass them
8525 * off to the new meta provider now.
8528 help = dtrace_deferred_pid;
8529 dtrace_deferred_pid = NULL;
8531 mutex_exit(&dtrace_lock);
8533 while (help != NULL) {
8534 for (i = 0; i < help->dthps_nprovs; i++) {
8535 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8539 next = help->dthps_next;
8540 help->dthps_next = NULL;
8541 help->dthps_prev = NULL;
8542 help->dthps_deferred = 0;
8546 mutex_exit(&dtrace_meta_lock);
8552 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8554 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8556 mutex_enter(&dtrace_meta_lock);
8557 mutex_enter(&dtrace_lock);
8559 if (old == dtrace_meta_pid) {
8560 pp = &dtrace_meta_pid;
8562 panic("attempt to unregister non-existent "
8563 "dtrace meta-provider %p\n", (void *)old);
8566 if (old->dtm_count != 0) {
8567 mutex_exit(&dtrace_lock);
8568 mutex_exit(&dtrace_meta_lock);
8574 mutex_exit(&dtrace_lock);
8575 mutex_exit(&dtrace_meta_lock);
8577 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8578 kmem_free(old, sizeof (dtrace_meta_t));
8585 * DTrace DIF Object Functions
8588 dtrace_difo_err(uint_t pc, const char *format, ...)
8590 if (dtrace_err_verbose) {
8593 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8594 va_start(alist, format);
8595 (void) vuprintf(format, alist);
8599 #ifdef DTRACE_ERRDEBUG
8600 dtrace_errdebug(format);
8606 * Validate a DTrace DIF object by checking the IR instructions. The following
8607 * rules are currently enforced by dtrace_difo_validate():
8609 * 1. Each instruction must have a valid opcode
8610 * 2. Each register, string, variable, or subroutine reference must be valid
8611 * 3. No instruction can modify register %r0 (must be zero)
8612 * 4. All instruction reserved bits must be set to zero
8613 * 5. The last instruction must be a "ret" instruction
8614 * 6. All branch targets must reference a valid instruction _after_ the branch
8617 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8621 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8625 kcheckload = cr == NULL ||
8626 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8628 dp->dtdo_destructive = 0;
8630 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8631 dif_instr_t instr = dp->dtdo_buf[pc];
8633 uint_t r1 = DIF_INSTR_R1(instr);
8634 uint_t r2 = DIF_INSTR_R2(instr);
8635 uint_t rd = DIF_INSTR_RD(instr);
8636 uint_t rs = DIF_INSTR_RS(instr);
8637 uint_t label = DIF_INSTR_LABEL(instr);
8638 uint_t v = DIF_INSTR_VAR(instr);
8639 uint_t subr = DIF_INSTR_SUBR(instr);
8640 uint_t type = DIF_INSTR_TYPE(instr);
8641 uint_t op = DIF_INSTR_OP(instr);
8659 err += efunc(pc, "invalid register %u\n", r1);
8661 err += efunc(pc, "invalid register %u\n", r2);
8663 err += efunc(pc, "invalid register %u\n", rd);
8665 err += efunc(pc, "cannot write to %r0\n");
8671 err += efunc(pc, "invalid register %u\n", r1);
8673 err += efunc(pc, "non-zero reserved bits\n");
8675 err += efunc(pc, "invalid register %u\n", rd);
8677 err += efunc(pc, "cannot write to %r0\n");
8687 err += efunc(pc, "invalid register %u\n", r1);
8689 err += efunc(pc, "non-zero reserved bits\n");
8691 err += efunc(pc, "invalid register %u\n", rd);
8693 err += efunc(pc, "cannot write to %r0\n");
8695 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8696 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8706 err += efunc(pc, "invalid register %u\n", r1);
8708 err += efunc(pc, "non-zero reserved bits\n");
8710 err += efunc(pc, "invalid register %u\n", rd);
8712 err += efunc(pc, "cannot write to %r0\n");
8722 err += efunc(pc, "invalid register %u\n", r1);
8724 err += efunc(pc, "non-zero reserved bits\n");
8726 err += efunc(pc, "invalid register %u\n", rd);
8728 err += efunc(pc, "cannot write to %r0\n");
8735 err += efunc(pc, "invalid register %u\n", r1);
8737 err += efunc(pc, "non-zero reserved bits\n");
8739 err += efunc(pc, "invalid register %u\n", rd);
8741 err += efunc(pc, "cannot write to 0 address\n");
8746 err += efunc(pc, "invalid register %u\n", r1);
8748 err += efunc(pc, "invalid register %u\n", r2);
8750 err += efunc(pc, "non-zero reserved bits\n");
8754 err += efunc(pc, "invalid register %u\n", r1);
8755 if (r2 != 0 || rd != 0)
8756 err += efunc(pc, "non-zero reserved bits\n");
8769 if (label >= dp->dtdo_len) {
8770 err += efunc(pc, "invalid branch target %u\n",
8774 err += efunc(pc, "backward branch to %u\n",
8779 if (r1 != 0 || r2 != 0)
8780 err += efunc(pc, "non-zero reserved bits\n");
8782 err += efunc(pc, "invalid register %u\n", rd);
8786 case DIF_OP_FLUSHTS:
8787 if (r1 != 0 || r2 != 0 || rd != 0)
8788 err += efunc(pc, "non-zero reserved bits\n");
8791 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8792 err += efunc(pc, "invalid integer ref %u\n",
8793 DIF_INSTR_INTEGER(instr));
8796 err += efunc(pc, "invalid register %u\n", rd);
8798 err += efunc(pc, "cannot write to %r0\n");
8801 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8802 err += efunc(pc, "invalid string ref %u\n",
8803 DIF_INSTR_STRING(instr));
8806 err += efunc(pc, "invalid register %u\n", rd);
8808 err += efunc(pc, "cannot write to %r0\n");
8812 if (r1 > DIF_VAR_ARRAY_MAX)
8813 err += efunc(pc, "invalid array %u\n", r1);
8815 err += efunc(pc, "invalid register %u\n", r2);
8817 err += efunc(pc, "invalid register %u\n", rd);
8819 err += efunc(pc, "cannot write to %r0\n");
8826 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8827 err += efunc(pc, "invalid variable %u\n", v);
8829 err += efunc(pc, "invalid register %u\n", rd);
8831 err += efunc(pc, "cannot write to %r0\n");
8838 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8839 err += efunc(pc, "invalid variable %u\n", v);
8841 err += efunc(pc, "invalid register %u\n", rd);
8844 if (subr > DIF_SUBR_MAX)
8845 err += efunc(pc, "invalid subr %u\n", subr);
8847 err += efunc(pc, "invalid register %u\n", rd);
8849 err += efunc(pc, "cannot write to %r0\n");
8851 if (subr == DIF_SUBR_COPYOUT ||
8852 subr == DIF_SUBR_COPYOUTSTR) {
8853 dp->dtdo_destructive = 1;
8857 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8858 err += efunc(pc, "invalid ref type %u\n", type);
8860 err += efunc(pc, "invalid register %u\n", r2);
8862 err += efunc(pc, "invalid register %u\n", rs);
8865 if (type != DIF_TYPE_CTF)
8866 err += efunc(pc, "invalid val type %u\n", type);
8868 err += efunc(pc, "invalid register %u\n", r2);
8870 err += efunc(pc, "invalid register %u\n", rs);
8873 err += efunc(pc, "invalid opcode %u\n",
8874 DIF_INSTR_OP(instr));
8878 if (dp->dtdo_len != 0 &&
8879 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8880 err += efunc(dp->dtdo_len - 1,
8881 "expected 'ret' as last DIF instruction\n");
8884 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8886 * If we're not returning by reference, the size must be either
8887 * 0 or the size of one of the base types.
8889 switch (dp->dtdo_rtype.dtdt_size) {
8891 case sizeof (uint8_t):
8892 case sizeof (uint16_t):
8893 case sizeof (uint32_t):
8894 case sizeof (uint64_t):
8898 err += efunc(dp->dtdo_len - 1, "bad return size");
8902 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8903 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8904 dtrace_diftype_t *vt, *et;
8907 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8908 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8909 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8910 err += efunc(i, "unrecognized variable scope %d\n",
8915 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8916 v->dtdv_kind != DIFV_KIND_SCALAR) {
8917 err += efunc(i, "unrecognized variable type %d\n",
8922 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8923 err += efunc(i, "%d exceeds variable id limit\n", id);
8927 if (id < DIF_VAR_OTHER_UBASE)
8931 * For user-defined variables, we need to check that this
8932 * definition is identical to any previous definition that we
8935 ndx = id - DIF_VAR_OTHER_UBASE;
8937 switch (v->dtdv_scope) {
8938 case DIFV_SCOPE_GLOBAL:
8939 if (ndx < vstate->dtvs_nglobals) {
8940 dtrace_statvar_t *svar;
8942 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8943 existing = &svar->dtsv_var;
8948 case DIFV_SCOPE_THREAD:
8949 if (ndx < vstate->dtvs_ntlocals)
8950 existing = &vstate->dtvs_tlocals[ndx];
8953 case DIFV_SCOPE_LOCAL:
8954 if (ndx < vstate->dtvs_nlocals) {
8955 dtrace_statvar_t *svar;
8957 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8958 existing = &svar->dtsv_var;
8966 if (vt->dtdt_flags & DIF_TF_BYREF) {
8967 if (vt->dtdt_size == 0) {
8968 err += efunc(i, "zero-sized variable\n");
8972 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8973 vt->dtdt_size > dtrace_global_maxsize) {
8974 err += efunc(i, "oversized by-ref global\n");
8979 if (existing == NULL || existing->dtdv_id == 0)
8982 ASSERT(existing->dtdv_id == v->dtdv_id);
8983 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8985 if (existing->dtdv_kind != v->dtdv_kind)
8986 err += efunc(i, "%d changed variable kind\n", id);
8988 et = &existing->dtdv_type;
8990 if (vt->dtdt_flags != et->dtdt_flags) {
8991 err += efunc(i, "%d changed variable type flags\n", id);
8995 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8996 err += efunc(i, "%d changed variable type size\n", id);
9005 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
9006 * are much more constrained than normal DIFOs. Specifically, they may
9009 * 1. Make calls to subroutines other than copyin(), copyinstr() or
9010 * miscellaneous string routines
9011 * 2. Access DTrace variables other than the args[] array, and the
9012 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9013 * 3. Have thread-local variables.
9014 * 4. Have dynamic variables.
9017 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9019 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9023 for (pc = 0; pc < dp->dtdo_len; pc++) {
9024 dif_instr_t instr = dp->dtdo_buf[pc];
9026 uint_t v = DIF_INSTR_VAR(instr);
9027 uint_t subr = DIF_INSTR_SUBR(instr);
9028 uint_t op = DIF_INSTR_OP(instr);
9083 case DIF_OP_FLUSHTS:
9095 if (v >= DIF_VAR_OTHER_UBASE)
9098 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9101 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9102 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9103 v == DIF_VAR_EXECARGS ||
9104 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9105 v == DIF_VAR_UID || v == DIF_VAR_GID)
9108 err += efunc(pc, "illegal variable %u\n", v);
9115 err += efunc(pc, "illegal dynamic variable load\n");
9121 err += efunc(pc, "illegal dynamic variable store\n");
9125 if (subr == DIF_SUBR_ALLOCA ||
9126 subr == DIF_SUBR_BCOPY ||
9127 subr == DIF_SUBR_COPYIN ||
9128 subr == DIF_SUBR_COPYINTO ||
9129 subr == DIF_SUBR_COPYINSTR ||
9130 subr == DIF_SUBR_INDEX ||
9131 subr == DIF_SUBR_INET_NTOA ||
9132 subr == DIF_SUBR_INET_NTOA6 ||
9133 subr == DIF_SUBR_INET_NTOP ||
9134 subr == DIF_SUBR_LLTOSTR ||
9135 subr == DIF_SUBR_RINDEX ||
9136 subr == DIF_SUBR_STRCHR ||
9137 subr == DIF_SUBR_STRJOIN ||
9138 subr == DIF_SUBR_STRRCHR ||
9139 subr == DIF_SUBR_STRSTR ||
9140 subr == DIF_SUBR_HTONS ||
9141 subr == DIF_SUBR_HTONL ||
9142 subr == DIF_SUBR_HTONLL ||
9143 subr == DIF_SUBR_NTOHS ||
9144 subr == DIF_SUBR_NTOHL ||
9145 subr == DIF_SUBR_NTOHLL ||
9146 subr == DIF_SUBR_MEMREF ||
9148 subr == DIF_SUBR_MEMSTR ||
9150 subr == DIF_SUBR_TYPEREF)
9153 err += efunc(pc, "invalid subr %u\n", subr);
9157 err += efunc(pc, "invalid opcode %u\n",
9158 DIF_INSTR_OP(instr));
9166 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9170 dtrace_difo_cacheable(dtrace_difo_t *dp)
9177 for (i = 0; i < dp->dtdo_varlen; i++) {
9178 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9180 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9183 switch (v->dtdv_id) {
9184 case DIF_VAR_CURTHREAD:
9187 case DIF_VAR_EXECARGS:
9188 case DIF_VAR_EXECNAME:
9189 case DIF_VAR_ZONENAME:
9198 * This DIF object may be cacheable. Now we need to look for any
9199 * array loading instructions, any memory loading instructions, or
9200 * any stores to thread-local variables.
9202 for (i = 0; i < dp->dtdo_len; i++) {
9203 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9205 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9206 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9207 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9208 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9216 dtrace_difo_hold(dtrace_difo_t *dp)
9220 ASSERT(MUTEX_HELD(&dtrace_lock));
9223 ASSERT(dp->dtdo_refcnt != 0);
9226 * We need to check this DIF object for references to the variable
9227 * DIF_VAR_VTIMESTAMP.
9229 for (i = 0; i < dp->dtdo_varlen; i++) {
9230 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9232 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9235 if (dtrace_vtime_references++ == 0)
9236 dtrace_vtime_enable();
9241 * This routine calculates the dynamic variable chunksize for a given DIF
9242 * object. The calculation is not fool-proof, and can probably be tricked by
9243 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9244 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9245 * if a dynamic variable size exceeds the chunksize.
9248 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9251 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9252 const dif_instr_t *text = dp->dtdo_buf;
9258 for (pc = 0; pc < dp->dtdo_len; pc++) {
9259 dif_instr_t instr = text[pc];
9260 uint_t op = DIF_INSTR_OP(instr);
9261 uint_t rd = DIF_INSTR_RD(instr);
9262 uint_t r1 = DIF_INSTR_R1(instr);
9266 dtrace_key_t *key = tupregs;
9270 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9275 key = &tupregs[DIF_DTR_NREGS];
9276 key[0].dttk_size = 0;
9277 key[1].dttk_size = 0;
9279 scope = DIFV_SCOPE_THREAD;
9286 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9287 key[nkeys++].dttk_size = 0;
9289 key[nkeys++].dttk_size = 0;
9291 if (op == DIF_OP_STTAA) {
9292 scope = DIFV_SCOPE_THREAD;
9294 scope = DIFV_SCOPE_GLOBAL;
9300 if (ttop == DIF_DTR_NREGS)
9303 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9305 * If the register for the size of the "pushtr"
9306 * is %r0 (or the value is 0) and the type is
9307 * a string, we'll use the system-wide default
9310 tupregs[ttop++].dttk_size =
9311 dtrace_strsize_default;
9316 tupregs[ttop++].dttk_size = sval;
9322 if (ttop == DIF_DTR_NREGS)
9325 tupregs[ttop++].dttk_size = 0;
9328 case DIF_OP_FLUSHTS:
9345 * We have a dynamic variable allocation; calculate its size.
9347 for (ksize = 0, i = 0; i < nkeys; i++)
9348 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9350 size = sizeof (dtrace_dynvar_t);
9351 size += sizeof (dtrace_key_t) * (nkeys - 1);
9355 * Now we need to determine the size of the stored data.
9357 id = DIF_INSTR_VAR(instr);
9359 for (i = 0; i < dp->dtdo_varlen; i++) {
9360 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9362 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9363 size += v->dtdv_type.dtdt_size;
9368 if (i == dp->dtdo_varlen)
9372 * We have the size. If this is larger than the chunk size
9373 * for our dynamic variable state, reset the chunk size.
9375 size = P2ROUNDUP(size, sizeof (uint64_t));
9377 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9378 vstate->dtvs_dynvars.dtds_chunksize = size;
9383 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9385 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9388 ASSERT(MUTEX_HELD(&dtrace_lock));
9389 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9391 for (i = 0; i < dp->dtdo_varlen; i++) {
9392 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9393 dtrace_statvar_t *svar, ***svarp = NULL;
9395 uint8_t scope = v->dtdv_scope;
9398 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9401 id -= DIF_VAR_OTHER_UBASE;
9404 case DIFV_SCOPE_THREAD:
9405 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9406 dtrace_difv_t *tlocals;
9408 if ((ntlocals = (otlocals << 1)) == 0)
9411 osz = otlocals * sizeof (dtrace_difv_t);
9412 nsz = ntlocals * sizeof (dtrace_difv_t);
9414 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9417 bcopy(vstate->dtvs_tlocals,
9419 kmem_free(vstate->dtvs_tlocals, osz);
9422 vstate->dtvs_tlocals = tlocals;
9423 vstate->dtvs_ntlocals = ntlocals;
9426 vstate->dtvs_tlocals[id] = *v;
9429 case DIFV_SCOPE_LOCAL:
9430 np = &vstate->dtvs_nlocals;
9431 svarp = &vstate->dtvs_locals;
9433 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9434 dsize = NCPU * (v->dtdv_type.dtdt_size +
9437 dsize = NCPU * sizeof (uint64_t);
9441 case DIFV_SCOPE_GLOBAL:
9442 np = &vstate->dtvs_nglobals;
9443 svarp = &vstate->dtvs_globals;
9445 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9446 dsize = v->dtdv_type.dtdt_size +
9455 while (id >= (oldsvars = *np)) {
9456 dtrace_statvar_t **statics;
9457 int newsvars, oldsize, newsize;
9459 if ((newsvars = (oldsvars << 1)) == 0)
9462 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9463 newsize = newsvars * sizeof (dtrace_statvar_t *);
9465 statics = kmem_zalloc(newsize, KM_SLEEP);
9468 bcopy(*svarp, statics, oldsize);
9469 kmem_free(*svarp, oldsize);
9476 if ((svar = (*svarp)[id]) == NULL) {
9477 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9478 svar->dtsv_var = *v;
9480 if ((svar->dtsv_size = dsize) != 0) {
9481 svar->dtsv_data = (uint64_t)(uintptr_t)
9482 kmem_zalloc(dsize, KM_SLEEP);
9485 (*svarp)[id] = svar;
9488 svar->dtsv_refcnt++;
9491 dtrace_difo_chunksize(dp, vstate);
9492 dtrace_difo_hold(dp);
9495 static dtrace_difo_t *
9496 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9501 ASSERT(dp->dtdo_buf != NULL);
9502 ASSERT(dp->dtdo_refcnt != 0);
9504 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9506 ASSERT(dp->dtdo_buf != NULL);
9507 sz = dp->dtdo_len * sizeof (dif_instr_t);
9508 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9509 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9510 new->dtdo_len = dp->dtdo_len;
9512 if (dp->dtdo_strtab != NULL) {
9513 ASSERT(dp->dtdo_strlen != 0);
9514 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9515 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9516 new->dtdo_strlen = dp->dtdo_strlen;
9519 if (dp->dtdo_inttab != NULL) {
9520 ASSERT(dp->dtdo_intlen != 0);
9521 sz = dp->dtdo_intlen * sizeof (uint64_t);
9522 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9523 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9524 new->dtdo_intlen = dp->dtdo_intlen;
9527 if (dp->dtdo_vartab != NULL) {
9528 ASSERT(dp->dtdo_varlen != 0);
9529 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9530 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9531 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9532 new->dtdo_varlen = dp->dtdo_varlen;
9535 dtrace_difo_init(new, vstate);
9540 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9544 ASSERT(dp->dtdo_refcnt == 0);
9546 for (i = 0; i < dp->dtdo_varlen; i++) {
9547 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9548 dtrace_statvar_t *svar, **svarp = NULL;
9550 uint8_t scope = v->dtdv_scope;
9554 case DIFV_SCOPE_THREAD:
9557 case DIFV_SCOPE_LOCAL:
9558 np = &vstate->dtvs_nlocals;
9559 svarp = vstate->dtvs_locals;
9562 case DIFV_SCOPE_GLOBAL:
9563 np = &vstate->dtvs_nglobals;
9564 svarp = vstate->dtvs_globals;
9571 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9574 id -= DIF_VAR_OTHER_UBASE;
9578 ASSERT(svar != NULL);
9579 ASSERT(svar->dtsv_refcnt > 0);
9581 if (--svar->dtsv_refcnt > 0)
9584 if (svar->dtsv_size != 0) {
9585 ASSERT(svar->dtsv_data != 0);
9586 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9590 kmem_free(svar, sizeof (dtrace_statvar_t));
9594 if (dp->dtdo_buf != NULL)
9595 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9596 if (dp->dtdo_inttab != NULL)
9597 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9598 if (dp->dtdo_strtab != NULL)
9599 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9600 if (dp->dtdo_vartab != NULL)
9601 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9603 kmem_free(dp, sizeof (dtrace_difo_t));
9607 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9611 ASSERT(MUTEX_HELD(&dtrace_lock));
9612 ASSERT(dp->dtdo_refcnt != 0);
9614 for (i = 0; i < dp->dtdo_varlen; i++) {
9615 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9617 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9620 ASSERT(dtrace_vtime_references > 0);
9621 if (--dtrace_vtime_references == 0)
9622 dtrace_vtime_disable();
9625 if (--dp->dtdo_refcnt == 0)
9626 dtrace_difo_destroy(dp, vstate);
9630 * DTrace Format Functions
9633 dtrace_format_add(dtrace_state_t *state, char *str)
9636 uint16_t ndx, len = strlen(str) + 1;
9638 fmt = kmem_zalloc(len, KM_SLEEP);
9639 bcopy(str, fmt, len);
9641 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9642 if (state->dts_formats[ndx] == NULL) {
9643 state->dts_formats[ndx] = fmt;
9648 if (state->dts_nformats == USHRT_MAX) {
9650 * This is only likely if a denial-of-service attack is being
9651 * attempted. As such, it's okay to fail silently here.
9653 kmem_free(fmt, len);
9658 * For simplicity, we always resize the formats array to be exactly the
9659 * number of formats.
9661 ndx = state->dts_nformats++;
9662 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9664 if (state->dts_formats != NULL) {
9666 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9667 kmem_free(state->dts_formats, ndx * sizeof (char *));
9670 state->dts_formats = new;
9671 state->dts_formats[ndx] = fmt;
9677 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9681 ASSERT(state->dts_formats != NULL);
9682 ASSERT(format <= state->dts_nformats);
9683 ASSERT(state->dts_formats[format - 1] != NULL);
9685 fmt = state->dts_formats[format - 1];
9686 kmem_free(fmt, strlen(fmt) + 1);
9687 state->dts_formats[format - 1] = NULL;
9691 dtrace_format_destroy(dtrace_state_t *state)
9695 if (state->dts_nformats == 0) {
9696 ASSERT(state->dts_formats == NULL);
9700 ASSERT(state->dts_formats != NULL);
9702 for (i = 0; i < state->dts_nformats; i++) {
9703 char *fmt = state->dts_formats[i];
9708 kmem_free(fmt, strlen(fmt) + 1);
9711 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9712 state->dts_nformats = 0;
9713 state->dts_formats = NULL;
9717 * DTrace Predicate Functions
9719 static dtrace_predicate_t *
9720 dtrace_predicate_create(dtrace_difo_t *dp)
9722 dtrace_predicate_t *pred;
9724 ASSERT(MUTEX_HELD(&dtrace_lock));
9725 ASSERT(dp->dtdo_refcnt != 0);
9727 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9728 pred->dtp_difo = dp;
9729 pred->dtp_refcnt = 1;
9731 if (!dtrace_difo_cacheable(dp))
9734 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9736 * This is only theoretically possible -- we have had 2^32
9737 * cacheable predicates on this machine. We cannot allow any
9738 * more predicates to become cacheable: as unlikely as it is,
9739 * there may be a thread caching a (now stale) predicate cache
9740 * ID. (N.B.: the temptation is being successfully resisted to
9741 * have this cmn_err() "Holy shit -- we executed this code!")
9746 pred->dtp_cacheid = dtrace_predcache_id++;
9752 dtrace_predicate_hold(dtrace_predicate_t *pred)
9754 ASSERT(MUTEX_HELD(&dtrace_lock));
9755 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9756 ASSERT(pred->dtp_refcnt > 0);
9762 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9764 dtrace_difo_t *dp = pred->dtp_difo;
9766 ASSERT(MUTEX_HELD(&dtrace_lock));
9767 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9768 ASSERT(pred->dtp_refcnt > 0);
9770 if (--pred->dtp_refcnt == 0) {
9771 dtrace_difo_release(pred->dtp_difo, vstate);
9772 kmem_free(pred, sizeof (dtrace_predicate_t));
9777 * DTrace Action Description Functions
9779 static dtrace_actdesc_t *
9780 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9781 uint64_t uarg, uint64_t arg)
9783 dtrace_actdesc_t *act;
9786 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9787 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9790 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9791 act->dtad_kind = kind;
9792 act->dtad_ntuple = ntuple;
9793 act->dtad_uarg = uarg;
9794 act->dtad_arg = arg;
9795 act->dtad_refcnt = 1;
9801 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9803 ASSERT(act->dtad_refcnt >= 1);
9808 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9810 dtrace_actkind_t kind = act->dtad_kind;
9813 ASSERT(act->dtad_refcnt >= 1);
9815 if (--act->dtad_refcnt != 0)
9818 if ((dp = act->dtad_difo) != NULL)
9819 dtrace_difo_release(dp, vstate);
9821 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9822 char *str = (char *)(uintptr_t)act->dtad_arg;
9825 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9826 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9830 kmem_free(str, strlen(str) + 1);
9833 kmem_free(act, sizeof (dtrace_actdesc_t));
9837 * DTrace ECB Functions
9839 static dtrace_ecb_t *
9840 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9845 ASSERT(MUTEX_HELD(&dtrace_lock));
9847 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9848 ecb->dte_predicate = NULL;
9849 ecb->dte_probe = probe;
9852 * The default size is the size of the default action: recording
9855 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
9856 ecb->dte_alignment = sizeof (dtrace_epid_t);
9858 epid = state->dts_epid++;
9860 if (epid - 1 >= state->dts_necbs) {
9861 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9862 int necbs = state->dts_necbs << 1;
9864 ASSERT(epid == state->dts_necbs + 1);
9867 ASSERT(oecbs == NULL);
9871 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9874 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9876 dtrace_membar_producer();
9877 state->dts_ecbs = ecbs;
9879 if (oecbs != NULL) {
9881 * If this state is active, we must dtrace_sync()
9882 * before we can free the old dts_ecbs array: we're
9883 * coming in hot, and there may be active ring
9884 * buffer processing (which indexes into the dts_ecbs
9885 * array) on another CPU.
9887 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9890 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9893 dtrace_membar_producer();
9894 state->dts_necbs = necbs;
9897 ecb->dte_state = state;
9899 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9900 dtrace_membar_producer();
9901 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9907 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9909 dtrace_probe_t *probe = ecb->dte_probe;
9911 ASSERT(MUTEX_HELD(&cpu_lock));
9912 ASSERT(MUTEX_HELD(&dtrace_lock));
9913 ASSERT(ecb->dte_next == NULL);
9915 if (probe == NULL) {
9917 * This is the NULL probe -- there's nothing to do.
9922 if (probe->dtpr_ecb == NULL) {
9923 dtrace_provider_t *prov = probe->dtpr_provider;
9926 * We're the first ECB on this probe.
9928 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9930 if (ecb->dte_predicate != NULL)
9931 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9933 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9934 probe->dtpr_id, probe->dtpr_arg);
9937 * This probe is already active. Swing the last pointer to
9938 * point to the new ECB, and issue a dtrace_sync() to assure
9939 * that all CPUs have seen the change.
9941 ASSERT(probe->dtpr_ecb_last != NULL);
9942 probe->dtpr_ecb_last->dte_next = ecb;
9943 probe->dtpr_ecb_last = ecb;
9944 probe->dtpr_predcache = 0;
9951 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9953 dtrace_action_t *act;
9954 uint32_t curneeded = UINT32_MAX;
9955 uint32_t aggbase = UINT32_MAX;
9958 * If we record anything, we always record the dtrace_rechdr_t. (And
9959 * we always record it first.)
9961 ecb->dte_size = sizeof (dtrace_rechdr_t);
9962 ecb->dte_alignment = sizeof (dtrace_epid_t);
9964 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9965 dtrace_recdesc_t *rec = &act->dta_rec;
9966 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
9968 ecb->dte_alignment = MAX(ecb->dte_alignment,
9969 rec->dtrd_alignment);
9971 if (DTRACEACT_ISAGG(act->dta_kind)) {
9972 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9974 ASSERT(rec->dtrd_size != 0);
9975 ASSERT(agg->dtag_first != NULL);
9976 ASSERT(act->dta_prev->dta_intuple);
9977 ASSERT(aggbase != UINT32_MAX);
9978 ASSERT(curneeded != UINT32_MAX);
9980 agg->dtag_base = aggbase;
9982 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9983 rec->dtrd_offset = curneeded;
9984 curneeded += rec->dtrd_size;
9985 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
9987 aggbase = UINT32_MAX;
9988 curneeded = UINT32_MAX;
9989 } else if (act->dta_intuple) {
9990 if (curneeded == UINT32_MAX) {
9992 * This is the first record in a tuple. Align
9993 * curneeded to be at offset 4 in an 8-byte
9996 ASSERT(act->dta_prev == NULL ||
9997 !act->dta_prev->dta_intuple);
9998 ASSERT3U(aggbase, ==, UINT32_MAX);
9999 curneeded = P2PHASEUP(ecb->dte_size,
10000 sizeof (uint64_t), sizeof (dtrace_aggid_t));
10002 aggbase = curneeded - sizeof (dtrace_aggid_t);
10003 ASSERT(IS_P2ALIGNED(aggbase,
10004 sizeof (uint64_t)));
10006 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10007 rec->dtrd_offset = curneeded;
10008 curneeded += rec->dtrd_size;
10010 /* tuples must be followed by an aggregation */
10011 ASSERT(act->dta_prev == NULL ||
10012 !act->dta_prev->dta_intuple);
10014 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10015 rec->dtrd_alignment);
10016 rec->dtrd_offset = ecb->dte_size;
10017 ecb->dte_size += rec->dtrd_size;
10018 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10022 if ((act = ecb->dte_action) != NULL &&
10023 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10024 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10026 * If the size is still sizeof (dtrace_rechdr_t), then all
10027 * actions store no data; set the size to 0.
10032 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10033 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10034 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10038 static dtrace_action_t *
10039 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10041 dtrace_aggregation_t *agg;
10042 size_t size = sizeof (uint64_t);
10043 int ntuple = desc->dtad_ntuple;
10044 dtrace_action_t *act;
10045 dtrace_recdesc_t *frec;
10046 dtrace_aggid_t aggid;
10047 dtrace_state_t *state = ecb->dte_state;
10049 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10050 agg->dtag_ecb = ecb;
10052 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10054 switch (desc->dtad_kind) {
10055 case DTRACEAGG_MIN:
10056 agg->dtag_initial = INT64_MAX;
10057 agg->dtag_aggregate = dtrace_aggregate_min;
10060 case DTRACEAGG_MAX:
10061 agg->dtag_initial = INT64_MIN;
10062 agg->dtag_aggregate = dtrace_aggregate_max;
10065 case DTRACEAGG_COUNT:
10066 agg->dtag_aggregate = dtrace_aggregate_count;
10069 case DTRACEAGG_QUANTIZE:
10070 agg->dtag_aggregate = dtrace_aggregate_quantize;
10071 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10075 case DTRACEAGG_LQUANTIZE: {
10076 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10077 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10079 agg->dtag_initial = desc->dtad_arg;
10080 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10082 if (step == 0 || levels == 0)
10085 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10089 case DTRACEAGG_LLQUANTIZE: {
10090 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10091 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10092 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10093 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10096 agg->dtag_initial = desc->dtad_arg;
10097 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10099 if (factor < 2 || low >= high || nsteps < factor)
10103 * Now check that the number of steps evenly divides a power
10104 * of the factor. (This assures both integer bucket size and
10105 * linearity within each magnitude.)
10107 for (v = factor; v < nsteps; v *= factor)
10110 if ((v % nsteps) || (nsteps % factor))
10113 size = (dtrace_aggregate_llquantize_bucket(factor,
10114 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10118 case DTRACEAGG_AVG:
10119 agg->dtag_aggregate = dtrace_aggregate_avg;
10120 size = sizeof (uint64_t) * 2;
10123 case DTRACEAGG_STDDEV:
10124 agg->dtag_aggregate = dtrace_aggregate_stddev;
10125 size = sizeof (uint64_t) * 4;
10128 case DTRACEAGG_SUM:
10129 agg->dtag_aggregate = dtrace_aggregate_sum;
10136 agg->dtag_action.dta_rec.dtrd_size = size;
10142 * We must make sure that we have enough actions for the n-tuple.
10144 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10145 if (DTRACEACT_ISAGG(act->dta_kind))
10148 if (--ntuple == 0) {
10150 * This is the action with which our n-tuple begins.
10152 agg->dtag_first = act;
10158 * This n-tuple is short by ntuple elements. Return failure.
10160 ASSERT(ntuple != 0);
10162 kmem_free(agg, sizeof (dtrace_aggregation_t));
10167 * If the last action in the tuple has a size of zero, it's actually
10168 * an expression argument for the aggregating action.
10170 ASSERT(ecb->dte_action_last != NULL);
10171 act = ecb->dte_action_last;
10173 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10174 ASSERT(act->dta_difo != NULL);
10176 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10177 agg->dtag_hasarg = 1;
10181 * We need to allocate an id for this aggregation.
10184 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10185 VM_BESTFIT | VM_SLEEP);
10187 aggid = alloc_unr(state->dts_aggid_arena);
10190 if (aggid - 1 >= state->dts_naggregations) {
10191 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10192 dtrace_aggregation_t **aggs;
10193 int naggs = state->dts_naggregations << 1;
10194 int onaggs = state->dts_naggregations;
10196 ASSERT(aggid == state->dts_naggregations + 1);
10199 ASSERT(oaggs == NULL);
10203 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10205 if (oaggs != NULL) {
10206 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10207 kmem_free(oaggs, onaggs * sizeof (*aggs));
10210 state->dts_aggregations = aggs;
10211 state->dts_naggregations = naggs;
10214 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10215 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10217 frec = &agg->dtag_first->dta_rec;
10218 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10219 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10221 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10222 ASSERT(!act->dta_intuple);
10223 act->dta_intuple = 1;
10226 return (&agg->dtag_action);
10230 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10232 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10233 dtrace_state_t *state = ecb->dte_state;
10234 dtrace_aggid_t aggid = agg->dtag_id;
10236 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10238 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10240 free_unr(state->dts_aggid_arena, aggid);
10243 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10244 state->dts_aggregations[aggid - 1] = NULL;
10246 kmem_free(agg, sizeof (dtrace_aggregation_t));
10250 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10252 dtrace_action_t *action, *last;
10253 dtrace_difo_t *dp = desc->dtad_difo;
10254 uint32_t size = 0, align = sizeof (uint8_t), mask;
10255 uint16_t format = 0;
10256 dtrace_recdesc_t *rec;
10257 dtrace_state_t *state = ecb->dte_state;
10258 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
10259 uint64_t arg = desc->dtad_arg;
10261 ASSERT(MUTEX_HELD(&dtrace_lock));
10262 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10264 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10266 * If this is an aggregating action, there must be neither
10267 * a speculate nor a commit on the action chain.
10269 dtrace_action_t *act;
10271 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10272 if (act->dta_kind == DTRACEACT_COMMIT)
10275 if (act->dta_kind == DTRACEACT_SPECULATE)
10279 action = dtrace_ecb_aggregation_create(ecb, desc);
10281 if (action == NULL)
10284 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10285 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10286 dp != NULL && dp->dtdo_destructive)) {
10287 state->dts_destructive = 1;
10290 switch (desc->dtad_kind) {
10291 case DTRACEACT_PRINTF:
10292 case DTRACEACT_PRINTA:
10293 case DTRACEACT_SYSTEM:
10294 case DTRACEACT_FREOPEN:
10295 case DTRACEACT_DIFEXPR:
10297 * We know that our arg is a string -- turn it into a
10301 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10302 desc->dtad_kind == DTRACEACT_DIFEXPR);
10307 ASSERT(arg > KERNELBASE);
10309 format = dtrace_format_add(state,
10310 (char *)(uintptr_t)arg);
10314 case DTRACEACT_LIBACT:
10315 case DTRACEACT_TRACEMEM:
10316 case DTRACEACT_TRACEMEM_DYNSIZE:
10320 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10323 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10324 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10327 size = opt[DTRACEOPT_STRSIZE];
10332 case DTRACEACT_STACK:
10333 if ((nframes = arg) == 0) {
10334 nframes = opt[DTRACEOPT_STACKFRAMES];
10335 ASSERT(nframes > 0);
10339 size = nframes * sizeof (pc_t);
10342 case DTRACEACT_JSTACK:
10343 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10344 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10346 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10347 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10349 arg = DTRACE_USTACK_ARG(nframes, strsize);
10352 case DTRACEACT_USTACK:
10353 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10354 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10355 strsize = DTRACE_USTACK_STRSIZE(arg);
10356 nframes = opt[DTRACEOPT_USTACKFRAMES];
10357 ASSERT(nframes > 0);
10358 arg = DTRACE_USTACK_ARG(nframes, strsize);
10362 * Save a slot for the pid.
10364 size = (nframes + 1) * sizeof (uint64_t);
10365 size += DTRACE_USTACK_STRSIZE(arg);
10366 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10370 case DTRACEACT_SYM:
10371 case DTRACEACT_MOD:
10372 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10373 sizeof (uint64_t)) ||
10374 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10378 case DTRACEACT_USYM:
10379 case DTRACEACT_UMOD:
10380 case DTRACEACT_UADDR:
10382 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10383 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10387 * We have a slot for the pid, plus a slot for the
10388 * argument. To keep things simple (aligned with
10389 * bitness-neutral sizing), we store each as a 64-bit
10392 size = 2 * sizeof (uint64_t);
10395 case DTRACEACT_STOP:
10396 case DTRACEACT_BREAKPOINT:
10397 case DTRACEACT_PANIC:
10400 case DTRACEACT_CHILL:
10401 case DTRACEACT_DISCARD:
10402 case DTRACEACT_RAISE:
10407 case DTRACEACT_EXIT:
10409 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10410 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10414 case DTRACEACT_SPECULATE:
10415 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10421 state->dts_speculates = 1;
10424 case DTRACEACT_PRINTM:
10425 size = dp->dtdo_rtype.dtdt_size;
10428 case DTRACEACT_PRINTT:
10429 size = dp->dtdo_rtype.dtdt_size;
10432 case DTRACEACT_COMMIT: {
10433 dtrace_action_t *act = ecb->dte_action;
10435 for (; act != NULL; act = act->dta_next) {
10436 if (act->dta_kind == DTRACEACT_COMMIT)
10449 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10451 * If this is a data-storing action or a speculate,
10452 * we must be sure that there isn't a commit on the
10455 dtrace_action_t *act = ecb->dte_action;
10457 for (; act != NULL; act = act->dta_next) {
10458 if (act->dta_kind == DTRACEACT_COMMIT)
10463 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10464 action->dta_rec.dtrd_size = size;
10467 action->dta_refcnt = 1;
10468 rec = &action->dta_rec;
10469 size = rec->dtrd_size;
10471 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10472 if (!(size & mask)) {
10478 action->dta_kind = desc->dtad_kind;
10480 if ((action->dta_difo = dp) != NULL)
10481 dtrace_difo_hold(dp);
10483 rec->dtrd_action = action->dta_kind;
10484 rec->dtrd_arg = arg;
10485 rec->dtrd_uarg = desc->dtad_uarg;
10486 rec->dtrd_alignment = (uint16_t)align;
10487 rec->dtrd_format = format;
10489 if ((last = ecb->dte_action_last) != NULL) {
10490 ASSERT(ecb->dte_action != NULL);
10491 action->dta_prev = last;
10492 last->dta_next = action;
10494 ASSERT(ecb->dte_action == NULL);
10495 ecb->dte_action = action;
10498 ecb->dte_action_last = action;
10504 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10506 dtrace_action_t *act = ecb->dte_action, *next;
10507 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10511 if (act != NULL && act->dta_refcnt > 1) {
10512 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10515 for (; act != NULL; act = next) {
10516 next = act->dta_next;
10517 ASSERT(next != NULL || act == ecb->dte_action_last);
10518 ASSERT(act->dta_refcnt == 1);
10520 if ((format = act->dta_rec.dtrd_format) != 0)
10521 dtrace_format_remove(ecb->dte_state, format);
10523 if ((dp = act->dta_difo) != NULL)
10524 dtrace_difo_release(dp, vstate);
10526 if (DTRACEACT_ISAGG(act->dta_kind)) {
10527 dtrace_ecb_aggregation_destroy(ecb, act);
10529 kmem_free(act, sizeof (dtrace_action_t));
10534 ecb->dte_action = NULL;
10535 ecb->dte_action_last = NULL;
10540 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10543 * We disable the ECB by removing it from its probe.
10545 dtrace_ecb_t *pecb, *prev = NULL;
10546 dtrace_probe_t *probe = ecb->dte_probe;
10548 ASSERT(MUTEX_HELD(&dtrace_lock));
10550 if (probe == NULL) {
10552 * This is the NULL probe; there is nothing to disable.
10557 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10563 ASSERT(pecb != NULL);
10565 if (prev == NULL) {
10566 probe->dtpr_ecb = ecb->dte_next;
10568 prev->dte_next = ecb->dte_next;
10571 if (ecb == probe->dtpr_ecb_last) {
10572 ASSERT(ecb->dte_next == NULL);
10573 probe->dtpr_ecb_last = prev;
10577 * The ECB has been disconnected from the probe; now sync to assure
10578 * that all CPUs have seen the change before returning.
10582 if (probe->dtpr_ecb == NULL) {
10584 * That was the last ECB on the probe; clear the predicate
10585 * cache ID for the probe, disable it and sync one more time
10586 * to assure that we'll never hit it again.
10588 dtrace_provider_t *prov = probe->dtpr_provider;
10590 ASSERT(ecb->dte_next == NULL);
10591 ASSERT(probe->dtpr_ecb_last == NULL);
10592 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10593 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10594 probe->dtpr_id, probe->dtpr_arg);
10598 * There is at least one ECB remaining on the probe. If there
10599 * is _exactly_ one, set the probe's predicate cache ID to be
10600 * the predicate cache ID of the remaining ECB.
10602 ASSERT(probe->dtpr_ecb_last != NULL);
10603 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10605 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10606 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10608 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10611 probe->dtpr_predcache = p->dtp_cacheid;
10614 ecb->dte_next = NULL;
10619 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10621 dtrace_state_t *state = ecb->dte_state;
10622 dtrace_vstate_t *vstate = &state->dts_vstate;
10623 dtrace_predicate_t *pred;
10624 dtrace_epid_t epid = ecb->dte_epid;
10626 ASSERT(MUTEX_HELD(&dtrace_lock));
10627 ASSERT(ecb->dte_next == NULL);
10628 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10630 if ((pred = ecb->dte_predicate) != NULL)
10631 dtrace_predicate_release(pred, vstate);
10633 dtrace_ecb_action_remove(ecb);
10635 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10636 state->dts_ecbs[epid - 1] = NULL;
10638 kmem_free(ecb, sizeof (dtrace_ecb_t));
10641 static dtrace_ecb_t *
10642 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10643 dtrace_enabling_t *enab)
10646 dtrace_predicate_t *pred;
10647 dtrace_actdesc_t *act;
10648 dtrace_provider_t *prov;
10649 dtrace_ecbdesc_t *desc = enab->dten_current;
10651 ASSERT(MUTEX_HELD(&dtrace_lock));
10652 ASSERT(state != NULL);
10654 ecb = dtrace_ecb_add(state, probe);
10655 ecb->dte_uarg = desc->dted_uarg;
10657 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10658 dtrace_predicate_hold(pred);
10659 ecb->dte_predicate = pred;
10662 if (probe != NULL) {
10664 * If the provider shows more leg than the consumer is old
10665 * enough to see, we need to enable the appropriate implicit
10666 * predicate bits to prevent the ecb from activating at
10669 * Providers specifying DTRACE_PRIV_USER at register time
10670 * are stating that they need the /proc-style privilege
10671 * model to be enforced, and this is what DTRACE_COND_OWNER
10672 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10674 prov = probe->dtpr_provider;
10675 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10676 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10677 ecb->dte_cond |= DTRACE_COND_OWNER;
10679 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10680 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10681 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10684 * If the provider shows us kernel innards and the user
10685 * is lacking sufficient privilege, enable the
10686 * DTRACE_COND_USERMODE implicit predicate.
10688 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10689 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10690 ecb->dte_cond |= DTRACE_COND_USERMODE;
10693 if (dtrace_ecb_create_cache != NULL) {
10695 * If we have a cached ecb, we'll use its action list instead
10696 * of creating our own (saving both time and space).
10698 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10699 dtrace_action_t *act = cached->dte_action;
10702 ASSERT(act->dta_refcnt > 0);
10704 ecb->dte_action = act;
10705 ecb->dte_action_last = cached->dte_action_last;
10706 ecb->dte_needed = cached->dte_needed;
10707 ecb->dte_size = cached->dte_size;
10708 ecb->dte_alignment = cached->dte_alignment;
10714 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10715 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10716 dtrace_ecb_destroy(ecb);
10721 dtrace_ecb_resize(ecb);
10723 return (dtrace_ecb_create_cache = ecb);
10727 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10730 dtrace_enabling_t *enab = arg;
10731 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10733 ASSERT(state != NULL);
10735 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10737 * This probe was created in a generation for which this
10738 * enabling has previously created ECBs; we don't want to
10739 * enable it again, so just kick out.
10741 return (DTRACE_MATCH_NEXT);
10744 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10745 return (DTRACE_MATCH_DONE);
10747 dtrace_ecb_enable(ecb);
10748 return (DTRACE_MATCH_NEXT);
10751 static dtrace_ecb_t *
10752 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10756 ASSERT(MUTEX_HELD(&dtrace_lock));
10758 if (id == 0 || id > state->dts_necbs)
10761 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10762 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10764 return (state->dts_ecbs[id - 1]);
10767 static dtrace_aggregation_t *
10768 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10770 dtrace_aggregation_t *agg;
10772 ASSERT(MUTEX_HELD(&dtrace_lock));
10774 if (id == 0 || id > state->dts_naggregations)
10777 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10778 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10779 agg->dtag_id == id);
10781 return (state->dts_aggregations[id - 1]);
10785 * DTrace Buffer Functions
10787 * The following functions manipulate DTrace buffers. Most of these functions
10788 * are called in the context of establishing or processing consumer state;
10789 * exceptions are explicitly noted.
10793 * Note: called from cross call context. This function switches the two
10794 * buffers on a given CPU. The atomicity of this operation is assured by
10795 * disabling interrupts while the actual switch takes place; the disabling of
10796 * interrupts serializes the execution with any execution of dtrace_probe() on
10800 dtrace_buffer_switch(dtrace_buffer_t *buf)
10802 caddr_t tomax = buf->dtb_tomax;
10803 caddr_t xamot = buf->dtb_xamot;
10804 dtrace_icookie_t cookie;
10807 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10808 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10810 cookie = dtrace_interrupt_disable();
10811 now = dtrace_gethrtime();
10812 buf->dtb_tomax = xamot;
10813 buf->dtb_xamot = tomax;
10814 buf->dtb_xamot_drops = buf->dtb_drops;
10815 buf->dtb_xamot_offset = buf->dtb_offset;
10816 buf->dtb_xamot_errors = buf->dtb_errors;
10817 buf->dtb_xamot_flags = buf->dtb_flags;
10818 buf->dtb_offset = 0;
10819 buf->dtb_drops = 0;
10820 buf->dtb_errors = 0;
10821 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10822 buf->dtb_interval = now - buf->dtb_switched;
10823 buf->dtb_switched = now;
10824 dtrace_interrupt_enable(cookie);
10828 * Note: called from cross call context. This function activates a buffer
10829 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10830 * is guaranteed by the disabling of interrupts.
10833 dtrace_buffer_activate(dtrace_state_t *state)
10835 dtrace_buffer_t *buf;
10836 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10838 buf = &state->dts_buffer[curcpu];
10840 if (buf->dtb_tomax != NULL) {
10842 * We might like to assert that the buffer is marked inactive,
10843 * but this isn't necessarily true: the buffer for the CPU
10844 * that processes the BEGIN probe has its buffer activated
10845 * manually. In this case, we take the (harmless) action
10846 * re-clearing the bit INACTIVE bit.
10848 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10851 dtrace_interrupt_enable(cookie);
10855 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10856 processorid_t cpu, int *factor)
10861 dtrace_buffer_t *buf;
10862 int allocated = 0, desired = 0;
10865 ASSERT(MUTEX_HELD(&cpu_lock));
10866 ASSERT(MUTEX_HELD(&dtrace_lock));
10870 if (size > dtrace_nonroot_maxsize &&
10871 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10877 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10880 buf = &bufs[cp->cpu_id];
10883 * If there is already a buffer allocated for this CPU, it
10884 * is only possible that this is a DR event. In this case,
10886 if (buf->dtb_tomax != NULL) {
10887 ASSERT(buf->dtb_size == size);
10891 ASSERT(buf->dtb_xamot == NULL);
10893 if ((buf->dtb_tomax = kmem_zalloc(size,
10894 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10897 buf->dtb_size = size;
10898 buf->dtb_flags = flags;
10899 buf->dtb_offset = 0;
10900 buf->dtb_drops = 0;
10902 if (flags & DTRACEBUF_NOSWITCH)
10905 if ((buf->dtb_xamot = kmem_zalloc(size,
10906 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10908 } while ((cp = cp->cpu_next) != cpu_list);
10916 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10919 buf = &bufs[cp->cpu_id];
10922 if (buf->dtb_xamot != NULL) {
10923 ASSERT(buf->dtb_tomax != NULL);
10924 ASSERT(buf->dtb_size == size);
10925 kmem_free(buf->dtb_xamot, size);
10929 if (buf->dtb_tomax != NULL) {
10930 ASSERT(buf->dtb_size == size);
10931 kmem_free(buf->dtb_tomax, size);
10935 buf->dtb_tomax = NULL;
10936 buf->dtb_xamot = NULL;
10938 } while ((cp = cp->cpu_next) != cpu_list);
10943 #if defined(__amd64__) || defined(__mips__) || defined(__powerpc__)
10945 * FreeBSD isn't good at limiting the amount of memory we
10946 * ask to malloc, so let's place a limit here before trying
10947 * to do something that might well end in tears at bedtime.
10949 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10953 ASSERT(MUTEX_HELD(&dtrace_lock));
10955 if (cpu != DTRACE_CPUALL && cpu != i)
10961 * If there is already a buffer allocated for this CPU, it
10962 * is only possible that this is a DR event. In this case,
10963 * the buffer size must match our specified size.
10965 if (buf->dtb_tomax != NULL) {
10966 ASSERT(buf->dtb_size == size);
10970 ASSERT(buf->dtb_xamot == NULL);
10972 if ((buf->dtb_tomax = kmem_zalloc(size,
10973 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10976 buf->dtb_size = size;
10977 buf->dtb_flags = flags;
10978 buf->dtb_offset = 0;
10979 buf->dtb_drops = 0;
10981 if (flags & DTRACEBUF_NOSWITCH)
10984 if ((buf->dtb_xamot = kmem_zalloc(size,
10985 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10993 * Error allocating memory, so free the buffers that were
10994 * allocated before the failed allocation.
10997 if (cpu != DTRACE_CPUALL && cpu != i)
11003 if (buf->dtb_xamot != NULL) {
11004 ASSERT(buf->dtb_tomax != NULL);
11005 ASSERT(buf->dtb_size == size);
11006 kmem_free(buf->dtb_xamot, size);
11010 if (buf->dtb_tomax != NULL) {
11011 ASSERT(buf->dtb_size == size);
11012 kmem_free(buf->dtb_tomax, size);
11016 buf->dtb_tomax = NULL;
11017 buf->dtb_xamot = NULL;
11022 *factor = desired / (allocated > 0 ? allocated : 1);
11028 * Note: called from probe context. This function just increments the drop
11029 * count on a buffer. It has been made a function to allow for the
11030 * possibility of understanding the source of mysterious drop counts. (A
11031 * problem for which one may be particularly disappointed that DTrace cannot
11032 * be used to understand DTrace.)
11035 dtrace_buffer_drop(dtrace_buffer_t *buf)
11041 * Note: called from probe context. This function is called to reserve space
11042 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
11043 * mstate. Returns the new offset in the buffer, or a negative value if an
11044 * error has occurred.
11047 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11048 dtrace_state_t *state, dtrace_mstate_t *mstate)
11050 intptr_t offs = buf->dtb_offset, soffs;
11055 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11058 if ((tomax = buf->dtb_tomax) == NULL) {
11059 dtrace_buffer_drop(buf);
11063 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11064 while (offs & (align - 1)) {
11066 * Assert that our alignment is off by a number which
11067 * is itself sizeof (uint32_t) aligned.
11069 ASSERT(!((align - (offs & (align - 1))) &
11070 (sizeof (uint32_t) - 1)));
11071 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11072 offs += sizeof (uint32_t);
11075 if ((soffs = offs + needed) > buf->dtb_size) {
11076 dtrace_buffer_drop(buf);
11080 if (mstate == NULL)
11083 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11084 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11085 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11090 if (buf->dtb_flags & DTRACEBUF_FILL) {
11091 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11092 (buf->dtb_flags & DTRACEBUF_FULL))
11097 total = needed + (offs & (align - 1));
11100 * For a ring buffer, life is quite a bit more complicated. Before
11101 * we can store any padding, we need to adjust our wrapping offset.
11102 * (If we've never before wrapped or we're not about to, no adjustment
11105 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11106 offs + total > buf->dtb_size) {
11107 woffs = buf->dtb_xamot_offset;
11109 if (offs + total > buf->dtb_size) {
11111 * We can't fit in the end of the buffer. First, a
11112 * sanity check that we can fit in the buffer at all.
11114 if (total > buf->dtb_size) {
11115 dtrace_buffer_drop(buf);
11120 * We're going to be storing at the top of the buffer,
11121 * so now we need to deal with the wrapped offset. We
11122 * only reset our wrapped offset to 0 if it is
11123 * currently greater than the current offset. If it
11124 * is less than the current offset, it is because a
11125 * previous allocation induced a wrap -- but the
11126 * allocation didn't subsequently take the space due
11127 * to an error or false predicate evaluation. In this
11128 * case, we'll just leave the wrapped offset alone: if
11129 * the wrapped offset hasn't been advanced far enough
11130 * for this allocation, it will be adjusted in the
11133 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11141 * Now we know that we're going to be storing to the
11142 * top of the buffer and that there is room for us
11143 * there. We need to clear the buffer from the current
11144 * offset to the end (there may be old gunk there).
11146 while (offs < buf->dtb_size)
11150 * We need to set our offset to zero. And because we
11151 * are wrapping, we need to set the bit indicating as
11152 * much. We can also adjust our needed space back
11153 * down to the space required by the ECB -- we know
11154 * that the top of the buffer is aligned.
11158 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11161 * There is room for us in the buffer, so we simply
11162 * need to check the wrapped offset.
11164 if (woffs < offs) {
11166 * The wrapped offset is less than the offset.
11167 * This can happen if we allocated buffer space
11168 * that induced a wrap, but then we didn't
11169 * subsequently take the space due to an error
11170 * or false predicate evaluation. This is
11171 * okay; we know that _this_ allocation isn't
11172 * going to induce a wrap. We still can't
11173 * reset the wrapped offset to be zero,
11174 * however: the space may have been trashed in
11175 * the previous failed probe attempt. But at
11176 * least the wrapped offset doesn't need to
11177 * be adjusted at all...
11183 while (offs + total > woffs) {
11184 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11187 if (epid == DTRACE_EPIDNONE) {
11188 size = sizeof (uint32_t);
11190 ASSERT3U(epid, <=, state->dts_necbs);
11191 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11193 size = state->dts_ecbs[epid - 1]->dte_size;
11196 ASSERT(woffs + size <= buf->dtb_size);
11199 if (woffs + size == buf->dtb_size) {
11201 * We've reached the end of the buffer; we want
11202 * to set the wrapped offset to 0 and break
11203 * out. However, if the offs is 0, then we're
11204 * in a strange edge-condition: the amount of
11205 * space that we want to reserve plus the size
11206 * of the record that we're overwriting is
11207 * greater than the size of the buffer. This
11208 * is problematic because if we reserve the
11209 * space but subsequently don't consume it (due
11210 * to a failed predicate or error) the wrapped
11211 * offset will be 0 -- yet the EPID at offset 0
11212 * will not be committed. This situation is
11213 * relatively easy to deal with: if we're in
11214 * this case, the buffer is indistinguishable
11215 * from one that hasn't wrapped; we need only
11216 * finish the job by clearing the wrapped bit,
11217 * explicitly setting the offset to be 0, and
11218 * zero'ing out the old data in the buffer.
11221 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11222 buf->dtb_offset = 0;
11225 while (woffs < buf->dtb_size)
11226 tomax[woffs++] = 0;
11237 * We have a wrapped offset. It may be that the wrapped offset
11238 * has become zero -- that's okay.
11240 buf->dtb_xamot_offset = woffs;
11245 * Now we can plow the buffer with any necessary padding.
11247 while (offs & (align - 1)) {
11249 * Assert that our alignment is off by a number which
11250 * is itself sizeof (uint32_t) aligned.
11252 ASSERT(!((align - (offs & (align - 1))) &
11253 (sizeof (uint32_t) - 1)));
11254 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11255 offs += sizeof (uint32_t);
11258 if (buf->dtb_flags & DTRACEBUF_FILL) {
11259 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11260 buf->dtb_flags |= DTRACEBUF_FULL;
11265 if (mstate == NULL)
11269 * For ring buffers and fill buffers, the scratch space is always
11270 * the inactive buffer.
11272 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11273 mstate->dtms_scratch_size = buf->dtb_size;
11274 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11280 dtrace_buffer_polish(dtrace_buffer_t *buf)
11282 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11283 ASSERT(MUTEX_HELD(&dtrace_lock));
11285 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11289 * We need to polish the ring buffer. There are three cases:
11291 * - The first (and presumably most common) is that there is no gap
11292 * between the buffer offset and the wrapped offset. In this case,
11293 * there is nothing in the buffer that isn't valid data; we can
11294 * mark the buffer as polished and return.
11296 * - The second (less common than the first but still more common
11297 * than the third) is that there is a gap between the buffer offset
11298 * and the wrapped offset, and the wrapped offset is larger than the
11299 * buffer offset. This can happen because of an alignment issue, or
11300 * can happen because of a call to dtrace_buffer_reserve() that
11301 * didn't subsequently consume the buffer space. In this case,
11302 * we need to zero the data from the buffer offset to the wrapped
11305 * - The third (and least common) is that there is a gap between the
11306 * buffer offset and the wrapped offset, but the wrapped offset is
11307 * _less_ than the buffer offset. This can only happen because a
11308 * call to dtrace_buffer_reserve() induced a wrap, but the space
11309 * was not subsequently consumed. In this case, we need to zero the
11310 * space from the offset to the end of the buffer _and_ from the
11311 * top of the buffer to the wrapped offset.
11313 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11314 bzero(buf->dtb_tomax + buf->dtb_offset,
11315 buf->dtb_xamot_offset - buf->dtb_offset);
11318 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11319 bzero(buf->dtb_tomax + buf->dtb_offset,
11320 buf->dtb_size - buf->dtb_offset);
11321 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11326 * This routine determines if data generated at the specified time has likely
11327 * been entirely consumed at user-level. This routine is called to determine
11328 * if an ECB on a defunct probe (but for an active enabling) can be safely
11329 * disabled and destroyed.
11332 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11336 for (i = 0; i < NCPU; i++) {
11337 dtrace_buffer_t *buf = &bufs[i];
11339 if (buf->dtb_size == 0)
11342 if (buf->dtb_flags & DTRACEBUF_RING)
11345 if (!buf->dtb_switched && buf->dtb_offset != 0)
11348 if (buf->dtb_switched - buf->dtb_interval < when)
11356 dtrace_buffer_free(dtrace_buffer_t *bufs)
11360 for (i = 0; i < NCPU; i++) {
11361 dtrace_buffer_t *buf = &bufs[i];
11363 if (buf->dtb_tomax == NULL) {
11364 ASSERT(buf->dtb_xamot == NULL);
11365 ASSERT(buf->dtb_size == 0);
11369 if (buf->dtb_xamot != NULL) {
11370 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11371 kmem_free(buf->dtb_xamot, buf->dtb_size);
11374 kmem_free(buf->dtb_tomax, buf->dtb_size);
11376 buf->dtb_tomax = NULL;
11377 buf->dtb_xamot = NULL;
11382 * DTrace Enabling Functions
11384 static dtrace_enabling_t *
11385 dtrace_enabling_create(dtrace_vstate_t *vstate)
11387 dtrace_enabling_t *enab;
11389 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11390 enab->dten_vstate = vstate;
11396 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11398 dtrace_ecbdesc_t **ndesc;
11399 size_t osize, nsize;
11402 * We can't add to enablings after we've enabled them, or after we've
11405 ASSERT(enab->dten_probegen == 0);
11406 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11408 if (enab->dten_ndesc < enab->dten_maxdesc) {
11409 enab->dten_desc[enab->dten_ndesc++] = ecb;
11413 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11415 if (enab->dten_maxdesc == 0) {
11416 enab->dten_maxdesc = 1;
11418 enab->dten_maxdesc <<= 1;
11421 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11423 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11424 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11425 bcopy(enab->dten_desc, ndesc, osize);
11426 if (enab->dten_desc != NULL)
11427 kmem_free(enab->dten_desc, osize);
11429 enab->dten_desc = ndesc;
11430 enab->dten_desc[enab->dten_ndesc++] = ecb;
11434 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11435 dtrace_probedesc_t *pd)
11437 dtrace_ecbdesc_t *new;
11438 dtrace_predicate_t *pred;
11439 dtrace_actdesc_t *act;
11442 * We're going to create a new ECB description that matches the
11443 * specified ECB in every way, but has the specified probe description.
11445 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11447 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11448 dtrace_predicate_hold(pred);
11450 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11451 dtrace_actdesc_hold(act);
11453 new->dted_action = ecb->dted_action;
11454 new->dted_pred = ecb->dted_pred;
11455 new->dted_probe = *pd;
11456 new->dted_uarg = ecb->dted_uarg;
11458 dtrace_enabling_add(enab, new);
11462 dtrace_enabling_dump(dtrace_enabling_t *enab)
11466 for (i = 0; i < enab->dten_ndesc; i++) {
11467 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11469 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11470 desc->dtpd_provider, desc->dtpd_mod,
11471 desc->dtpd_func, desc->dtpd_name);
11476 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11479 dtrace_ecbdesc_t *ep;
11480 dtrace_vstate_t *vstate = enab->dten_vstate;
11482 ASSERT(MUTEX_HELD(&dtrace_lock));
11484 for (i = 0; i < enab->dten_ndesc; i++) {
11485 dtrace_actdesc_t *act, *next;
11486 dtrace_predicate_t *pred;
11488 ep = enab->dten_desc[i];
11490 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11491 dtrace_predicate_release(pred, vstate);
11493 for (act = ep->dted_action; act != NULL; act = next) {
11494 next = act->dtad_next;
11495 dtrace_actdesc_release(act, vstate);
11498 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11501 if (enab->dten_desc != NULL)
11502 kmem_free(enab->dten_desc,
11503 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11506 * If this was a retained enabling, decrement the dts_nretained count
11507 * and take it off of the dtrace_retained list.
11509 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11510 dtrace_retained == enab) {
11511 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11512 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11513 enab->dten_vstate->dtvs_state->dts_nretained--;
11516 if (enab->dten_prev == NULL) {
11517 if (dtrace_retained == enab) {
11518 dtrace_retained = enab->dten_next;
11520 if (dtrace_retained != NULL)
11521 dtrace_retained->dten_prev = NULL;
11524 ASSERT(enab != dtrace_retained);
11525 ASSERT(dtrace_retained != NULL);
11526 enab->dten_prev->dten_next = enab->dten_next;
11529 if (enab->dten_next != NULL) {
11530 ASSERT(dtrace_retained != NULL);
11531 enab->dten_next->dten_prev = enab->dten_prev;
11534 kmem_free(enab, sizeof (dtrace_enabling_t));
11538 dtrace_enabling_retain(dtrace_enabling_t *enab)
11540 dtrace_state_t *state;
11542 ASSERT(MUTEX_HELD(&dtrace_lock));
11543 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11544 ASSERT(enab->dten_vstate != NULL);
11546 state = enab->dten_vstate->dtvs_state;
11547 ASSERT(state != NULL);
11550 * We only allow each state to retain dtrace_retain_max enablings.
11552 if (state->dts_nretained >= dtrace_retain_max)
11555 state->dts_nretained++;
11557 if (dtrace_retained == NULL) {
11558 dtrace_retained = enab;
11562 enab->dten_next = dtrace_retained;
11563 dtrace_retained->dten_prev = enab;
11564 dtrace_retained = enab;
11570 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11571 dtrace_probedesc_t *create)
11573 dtrace_enabling_t *new, *enab;
11574 int found = 0, err = ENOENT;
11576 ASSERT(MUTEX_HELD(&dtrace_lock));
11577 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11578 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11579 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11580 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11582 new = dtrace_enabling_create(&state->dts_vstate);
11585 * Iterate over all retained enablings, looking for enablings that
11586 * match the specified state.
11588 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11592 * dtvs_state can only be NULL for helper enablings -- and
11593 * helper enablings can't be retained.
11595 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11597 if (enab->dten_vstate->dtvs_state != state)
11601 * Now iterate over each probe description; we're looking for
11602 * an exact match to the specified probe description.
11604 for (i = 0; i < enab->dten_ndesc; i++) {
11605 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11606 dtrace_probedesc_t *pd = &ep->dted_probe;
11608 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11611 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11614 if (strcmp(pd->dtpd_func, match->dtpd_func))
11617 if (strcmp(pd->dtpd_name, match->dtpd_name))
11621 * We have a winning probe! Add it to our growing
11625 dtrace_enabling_addlike(new, ep, create);
11629 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11630 dtrace_enabling_destroy(new);
11638 dtrace_enabling_retract(dtrace_state_t *state)
11640 dtrace_enabling_t *enab, *next;
11642 ASSERT(MUTEX_HELD(&dtrace_lock));
11645 * Iterate over all retained enablings, destroy the enablings retained
11646 * for the specified state.
11648 for (enab = dtrace_retained; enab != NULL; enab = next) {
11649 next = enab->dten_next;
11652 * dtvs_state can only be NULL for helper enablings -- and
11653 * helper enablings can't be retained.
11655 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11657 if (enab->dten_vstate->dtvs_state == state) {
11658 ASSERT(state->dts_nretained > 0);
11659 dtrace_enabling_destroy(enab);
11663 ASSERT(state->dts_nretained == 0);
11667 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11672 ASSERT(MUTEX_HELD(&cpu_lock));
11673 ASSERT(MUTEX_HELD(&dtrace_lock));
11675 for (i = 0; i < enab->dten_ndesc; i++) {
11676 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11678 enab->dten_current = ep;
11679 enab->dten_error = 0;
11681 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11683 if (enab->dten_error != 0) {
11685 * If we get an error half-way through enabling the
11686 * probes, we kick out -- perhaps with some number of
11687 * them enabled. Leaving enabled probes enabled may
11688 * be slightly confusing for user-level, but we expect
11689 * that no one will attempt to actually drive on in
11690 * the face of such errors. If this is an anonymous
11691 * enabling (indicated with a NULL nmatched pointer),
11692 * we cmn_err() a message. We aren't expecting to
11693 * get such an error -- such as it can exist at all,
11694 * it would be a result of corrupted DOF in the driver
11697 if (nmatched == NULL) {
11698 cmn_err(CE_WARN, "dtrace_enabling_match() "
11699 "error on %p: %d", (void *)ep,
11703 return (enab->dten_error);
11707 enab->dten_probegen = dtrace_probegen;
11708 if (nmatched != NULL)
11709 *nmatched = matched;
11715 dtrace_enabling_matchall(void)
11717 dtrace_enabling_t *enab;
11719 mutex_enter(&cpu_lock);
11720 mutex_enter(&dtrace_lock);
11723 * Iterate over all retained enablings to see if any probes match
11724 * against them. We only perform this operation on enablings for which
11725 * we have sufficient permissions by virtue of being in the global zone
11726 * or in the same zone as the DTrace client. Because we can be called
11727 * after dtrace_detach() has been called, we cannot assert that there
11728 * are retained enablings. We can safely load from dtrace_retained,
11729 * however: the taskq_destroy() at the end of dtrace_detach() will
11730 * block pending our completion.
11732 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11734 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11736 if (INGLOBALZONE(curproc) || getzoneid() == crgetzoneid(cr))
11738 (void) dtrace_enabling_match(enab, NULL);
11741 mutex_exit(&dtrace_lock);
11742 mutex_exit(&cpu_lock);
11746 * If an enabling is to be enabled without having matched probes (that is, if
11747 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11748 * enabling must be _primed_ by creating an ECB for every ECB description.
11749 * This must be done to assure that we know the number of speculations, the
11750 * number of aggregations, the minimum buffer size needed, etc. before we
11751 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11752 * enabling any probes, we create ECBs for every ECB decription, but with a
11753 * NULL probe -- which is exactly what this function does.
11756 dtrace_enabling_prime(dtrace_state_t *state)
11758 dtrace_enabling_t *enab;
11761 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11762 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11764 if (enab->dten_vstate->dtvs_state != state)
11768 * We don't want to prime an enabling more than once, lest
11769 * we allow a malicious user to induce resource exhaustion.
11770 * (The ECBs that result from priming an enabling aren't
11771 * leaked -- but they also aren't deallocated until the
11772 * consumer state is destroyed.)
11774 if (enab->dten_primed)
11777 for (i = 0; i < enab->dten_ndesc; i++) {
11778 enab->dten_current = enab->dten_desc[i];
11779 (void) dtrace_probe_enable(NULL, enab);
11782 enab->dten_primed = 1;
11787 * Called to indicate that probes should be provided due to retained
11788 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11789 * must take an initial lap through the enabling calling the dtps_provide()
11790 * entry point explicitly to allow for autocreated probes.
11793 dtrace_enabling_provide(dtrace_provider_t *prv)
11796 dtrace_probedesc_t desc;
11798 ASSERT(MUTEX_HELD(&dtrace_lock));
11799 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11803 prv = dtrace_provider;
11807 dtrace_enabling_t *enab = dtrace_retained;
11808 void *parg = prv->dtpv_arg;
11810 for (; enab != NULL; enab = enab->dten_next) {
11811 for (i = 0; i < enab->dten_ndesc; i++) {
11812 desc = enab->dten_desc[i]->dted_probe;
11813 mutex_exit(&dtrace_lock);
11814 prv->dtpv_pops.dtps_provide(parg, &desc);
11815 mutex_enter(&dtrace_lock);
11818 } while (all && (prv = prv->dtpv_next) != NULL);
11820 mutex_exit(&dtrace_lock);
11821 dtrace_probe_provide(NULL, all ? NULL : prv);
11822 mutex_enter(&dtrace_lock);
11826 * Called to reap ECBs that are attached to probes from defunct providers.
11829 dtrace_enabling_reap(void)
11831 dtrace_provider_t *prov;
11832 dtrace_probe_t *probe;
11837 mutex_enter(&cpu_lock);
11838 mutex_enter(&dtrace_lock);
11840 for (i = 0; i < dtrace_nprobes; i++) {
11841 if ((probe = dtrace_probes[i]) == NULL)
11844 if (probe->dtpr_ecb == NULL)
11847 prov = probe->dtpr_provider;
11849 if ((when = prov->dtpv_defunct) == 0)
11853 * We have ECBs on a defunct provider: we want to reap these
11854 * ECBs to allow the provider to unregister. The destruction
11855 * of these ECBs must be done carefully: if we destroy the ECB
11856 * and the consumer later wishes to consume an EPID that
11857 * corresponds to the destroyed ECB (and if the EPID metadata
11858 * has not been previously consumed), the consumer will abort
11859 * processing on the unknown EPID. To reduce (but not, sadly,
11860 * eliminate) the possibility of this, we will only destroy an
11861 * ECB for a defunct provider if, for the state that
11862 * corresponds to the ECB:
11864 * (a) There is no speculative tracing (which can effectively
11865 * cache an EPID for an arbitrary amount of time).
11867 * (b) The principal buffers have been switched twice since the
11868 * provider became defunct.
11870 * (c) The aggregation buffers are of zero size or have been
11871 * switched twice since the provider became defunct.
11873 * We use dts_speculates to determine (a) and call a function
11874 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
11875 * that as soon as we've been unable to destroy one of the ECBs
11876 * associated with the probe, we quit trying -- reaping is only
11877 * fruitful in as much as we can destroy all ECBs associated
11878 * with the defunct provider's probes.
11880 while ((ecb = probe->dtpr_ecb) != NULL) {
11881 dtrace_state_t *state = ecb->dte_state;
11882 dtrace_buffer_t *buf = state->dts_buffer;
11883 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11885 if (state->dts_speculates)
11888 if (!dtrace_buffer_consumed(buf, when))
11891 if (!dtrace_buffer_consumed(aggbuf, when))
11894 dtrace_ecb_disable(ecb);
11895 ASSERT(probe->dtpr_ecb != ecb);
11896 dtrace_ecb_destroy(ecb);
11900 mutex_exit(&dtrace_lock);
11901 mutex_exit(&cpu_lock);
11905 * DTrace DOF Functions
11909 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11911 if (dtrace_err_verbose)
11912 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11914 #ifdef DTRACE_ERRDEBUG
11915 dtrace_errdebug(str);
11920 * Create DOF out of a currently enabled state. Right now, we only create
11921 * DOF containing the run-time options -- but this could be expanded to create
11922 * complete DOF representing the enabled state.
11925 dtrace_dof_create(dtrace_state_t *state)
11929 dof_optdesc_t *opt;
11930 int i, len = sizeof (dof_hdr_t) +
11931 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11932 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11934 ASSERT(MUTEX_HELD(&dtrace_lock));
11936 dof = kmem_zalloc(len, KM_SLEEP);
11937 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11938 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11939 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11940 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11942 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11943 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11944 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11945 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11946 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11947 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11949 dof->dofh_flags = 0;
11950 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11951 dof->dofh_secsize = sizeof (dof_sec_t);
11952 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11953 dof->dofh_secoff = sizeof (dof_hdr_t);
11954 dof->dofh_loadsz = len;
11955 dof->dofh_filesz = len;
11959 * Fill in the option section header...
11961 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11962 sec->dofs_type = DOF_SECT_OPTDESC;
11963 sec->dofs_align = sizeof (uint64_t);
11964 sec->dofs_flags = DOF_SECF_LOAD;
11965 sec->dofs_entsize = sizeof (dof_optdesc_t);
11967 opt = (dof_optdesc_t *)((uintptr_t)sec +
11968 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11970 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11971 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11973 for (i = 0; i < DTRACEOPT_MAX; i++) {
11974 opt[i].dofo_option = i;
11975 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11976 opt[i].dofo_value = state->dts_options[i];
11983 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11985 dof_hdr_t hdr, *dof;
11987 ASSERT(!MUTEX_HELD(&dtrace_lock));
11990 * First, we're going to copyin() the sizeof (dof_hdr_t).
11992 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11993 dtrace_dof_error(NULL, "failed to copyin DOF header");
11999 * Now we'll allocate the entire DOF and copy it in -- provided
12000 * that the length isn't outrageous.
12002 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12003 dtrace_dof_error(&hdr, "load size exceeds maximum");
12008 if (hdr.dofh_loadsz < sizeof (hdr)) {
12009 dtrace_dof_error(&hdr, "invalid load size");
12014 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12016 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
12017 kmem_free(dof, hdr.dofh_loadsz);
12026 static __inline uchar_t
12027 dtrace_dof_char(char c) {
12046 return (c - 'A' + 10);
12053 return (c - 'a' + 10);
12055 /* Should not reach here. */
12061 dtrace_dof_property(const char *name)
12065 unsigned int len, i;
12070 * Unfortunately, array of values in .conf files are always (and
12071 * only) interpreted to be integer arrays. We must read our DOF
12072 * as an integer array, and then squeeze it into a byte array.
12074 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12075 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12078 for (i = 0; i < len; i++)
12079 buf[i] = (uchar_t)(((int *)buf)[i]);
12081 if (len < sizeof (dof_hdr_t)) {
12082 ddi_prop_free(buf);
12083 dtrace_dof_error(NULL, "truncated header");
12087 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12088 ddi_prop_free(buf);
12089 dtrace_dof_error(NULL, "truncated DOF");
12093 if (loadsz >= dtrace_dof_maxsize) {
12094 ddi_prop_free(buf);
12095 dtrace_dof_error(NULL, "oversized DOF");
12099 dof = kmem_alloc(loadsz, KM_SLEEP);
12100 bcopy(buf, dof, loadsz);
12101 ddi_prop_free(buf);
12106 if ((p_env = getenv(name)) == NULL)
12109 len = strlen(p_env) / 2;
12111 buf = kmem_alloc(len, KM_SLEEP);
12113 dof = (dof_hdr_t *) buf;
12117 for (i = 0; i < len; i++) {
12118 buf[i] = (dtrace_dof_char(p[0]) << 4) |
12119 dtrace_dof_char(p[1]);
12125 if (len < sizeof (dof_hdr_t)) {
12127 dtrace_dof_error(NULL, "truncated header");
12131 if (len < (loadsz = dof->dofh_loadsz)) {
12133 dtrace_dof_error(NULL, "truncated DOF");
12137 if (loadsz >= dtrace_dof_maxsize) {
12139 dtrace_dof_error(NULL, "oversized DOF");
12148 dtrace_dof_destroy(dof_hdr_t *dof)
12150 kmem_free(dof, dof->dofh_loadsz);
12154 * Return the dof_sec_t pointer corresponding to a given section index. If the
12155 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12156 * a type other than DOF_SECT_NONE is specified, the header is checked against
12157 * this type and NULL is returned if the types do not match.
12160 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12162 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12163 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12165 if (i >= dof->dofh_secnum) {
12166 dtrace_dof_error(dof, "referenced section index is invalid");
12170 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12171 dtrace_dof_error(dof, "referenced section is not loadable");
12175 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12176 dtrace_dof_error(dof, "referenced section is the wrong type");
12183 static dtrace_probedesc_t *
12184 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12186 dof_probedesc_t *probe;
12188 uintptr_t daddr = (uintptr_t)dof;
12192 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12193 dtrace_dof_error(dof, "invalid probe section");
12197 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12198 dtrace_dof_error(dof, "bad alignment in probe description");
12202 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12203 dtrace_dof_error(dof, "truncated probe description");
12207 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12208 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12210 if (strtab == NULL)
12213 str = daddr + strtab->dofs_offset;
12214 size = strtab->dofs_size;
12216 if (probe->dofp_provider >= strtab->dofs_size) {
12217 dtrace_dof_error(dof, "corrupt probe provider");
12221 (void) strncpy(desc->dtpd_provider,
12222 (char *)(str + probe->dofp_provider),
12223 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12225 if (probe->dofp_mod >= strtab->dofs_size) {
12226 dtrace_dof_error(dof, "corrupt probe module");
12230 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12231 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12233 if (probe->dofp_func >= strtab->dofs_size) {
12234 dtrace_dof_error(dof, "corrupt probe function");
12238 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12239 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12241 if (probe->dofp_name >= strtab->dofs_size) {
12242 dtrace_dof_error(dof, "corrupt probe name");
12246 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12247 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12252 static dtrace_difo_t *
12253 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12258 dof_difohdr_t *dofd;
12259 uintptr_t daddr = (uintptr_t)dof;
12260 size_t max = dtrace_difo_maxsize;
12263 static const struct {
12271 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12272 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12273 sizeof (dif_instr_t), "multiple DIF sections" },
12275 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12276 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12277 sizeof (uint64_t), "multiple integer tables" },
12279 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12280 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12281 sizeof (char), "multiple string tables" },
12283 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12284 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12285 sizeof (uint_t), "multiple variable tables" },
12287 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
12290 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12291 dtrace_dof_error(dof, "invalid DIFO header section");
12295 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12296 dtrace_dof_error(dof, "bad alignment in DIFO header");
12300 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12301 sec->dofs_size % sizeof (dof_secidx_t)) {
12302 dtrace_dof_error(dof, "bad size in DIFO header");
12306 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12307 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12309 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12310 dp->dtdo_rtype = dofd->dofd_rtype;
12312 for (l = 0; l < n; l++) {
12317 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12318 dofd->dofd_links[l])) == NULL)
12319 goto err; /* invalid section link */
12321 if (ttl + subsec->dofs_size > max) {
12322 dtrace_dof_error(dof, "exceeds maximum size");
12326 ttl += subsec->dofs_size;
12328 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12329 if (subsec->dofs_type != difo[i].section)
12332 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12333 dtrace_dof_error(dof, "section not loaded");
12337 if (subsec->dofs_align != difo[i].align) {
12338 dtrace_dof_error(dof, "bad alignment");
12342 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12343 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12345 if (*bufp != NULL) {
12346 dtrace_dof_error(dof, difo[i].msg);
12350 if (difo[i].entsize != subsec->dofs_entsize) {
12351 dtrace_dof_error(dof, "entry size mismatch");
12355 if (subsec->dofs_entsize != 0 &&
12356 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12357 dtrace_dof_error(dof, "corrupt entry size");
12361 *lenp = subsec->dofs_size;
12362 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12363 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12364 *bufp, subsec->dofs_size);
12366 if (subsec->dofs_entsize != 0)
12367 *lenp /= subsec->dofs_entsize;
12373 * If we encounter a loadable DIFO sub-section that is not
12374 * known to us, assume this is a broken program and fail.
12376 if (difo[i].section == DOF_SECT_NONE &&
12377 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12378 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12383 if (dp->dtdo_buf == NULL) {
12385 * We can't have a DIF object without DIF text.
12387 dtrace_dof_error(dof, "missing DIF text");
12392 * Before we validate the DIF object, run through the variable table
12393 * looking for the strings -- if any of their size are under, we'll set
12394 * their size to be the system-wide default string size. Note that
12395 * this should _not_ happen if the "strsize" option has been set --
12396 * in this case, the compiler should have set the size to reflect the
12397 * setting of the option.
12399 for (i = 0; i < dp->dtdo_varlen; i++) {
12400 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12401 dtrace_diftype_t *t = &v->dtdv_type;
12403 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12406 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12407 t->dtdt_size = dtrace_strsize_default;
12410 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12413 dtrace_difo_init(dp, vstate);
12417 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12418 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12419 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12420 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12422 kmem_free(dp, sizeof (dtrace_difo_t));
12426 static dtrace_predicate_t *
12427 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12432 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12435 return (dtrace_predicate_create(dp));
12438 static dtrace_actdesc_t *
12439 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12442 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12443 dof_actdesc_t *desc;
12444 dof_sec_t *difosec;
12446 uintptr_t daddr = (uintptr_t)dof;
12448 dtrace_actkind_t kind;
12450 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12451 dtrace_dof_error(dof, "invalid action section");
12455 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12456 dtrace_dof_error(dof, "truncated action description");
12460 if (sec->dofs_align != sizeof (uint64_t)) {
12461 dtrace_dof_error(dof, "bad alignment in action description");
12465 if (sec->dofs_size < sec->dofs_entsize) {
12466 dtrace_dof_error(dof, "section entry size exceeds total size");
12470 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12471 dtrace_dof_error(dof, "bad entry size in action description");
12475 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12476 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12480 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12481 desc = (dof_actdesc_t *)(daddr +
12482 (uintptr_t)sec->dofs_offset + offs);
12483 kind = (dtrace_actkind_t)desc->dofa_kind;
12485 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12486 (kind != DTRACEACT_PRINTA ||
12487 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12488 (kind == DTRACEACT_DIFEXPR &&
12489 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12495 * The argument to these actions is an index into the
12496 * DOF string table. For printf()-like actions, this
12497 * is the format string. For print(), this is the
12498 * CTF type of the expression result.
12500 if ((strtab = dtrace_dof_sect(dof,
12501 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12504 str = (char *)((uintptr_t)dof +
12505 (uintptr_t)strtab->dofs_offset);
12507 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12508 if (str[i] == '\0')
12512 if (i >= strtab->dofs_size) {
12513 dtrace_dof_error(dof, "bogus format string");
12517 if (i == desc->dofa_arg) {
12518 dtrace_dof_error(dof, "empty format string");
12522 i -= desc->dofa_arg;
12523 fmt = kmem_alloc(i + 1, KM_SLEEP);
12524 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12525 arg = (uint64_t)(uintptr_t)fmt;
12527 if (kind == DTRACEACT_PRINTA) {
12528 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12531 arg = desc->dofa_arg;
12535 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12536 desc->dofa_uarg, arg);
12538 if (last != NULL) {
12539 last->dtad_next = act;
12546 if (desc->dofa_difo == DOF_SECIDX_NONE)
12549 if ((difosec = dtrace_dof_sect(dof,
12550 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12553 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12555 if (act->dtad_difo == NULL)
12559 ASSERT(first != NULL);
12563 for (act = first; act != NULL; act = next) {
12564 next = act->dtad_next;
12565 dtrace_actdesc_release(act, vstate);
12571 static dtrace_ecbdesc_t *
12572 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12575 dtrace_ecbdesc_t *ep;
12576 dof_ecbdesc_t *ecb;
12577 dtrace_probedesc_t *desc;
12578 dtrace_predicate_t *pred = NULL;
12580 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12581 dtrace_dof_error(dof, "truncated ECB description");
12585 if (sec->dofs_align != sizeof (uint64_t)) {
12586 dtrace_dof_error(dof, "bad alignment in ECB description");
12590 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12591 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12596 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12597 ep->dted_uarg = ecb->dofe_uarg;
12598 desc = &ep->dted_probe;
12600 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12603 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12604 if ((sec = dtrace_dof_sect(dof,
12605 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12608 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12611 ep->dted_pred.dtpdd_predicate = pred;
12614 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12615 if ((sec = dtrace_dof_sect(dof,
12616 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12619 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12621 if (ep->dted_action == NULL)
12629 dtrace_predicate_release(pred, vstate);
12630 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12635 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12636 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12637 * site of any user SETX relocations to account for load object base address.
12638 * In the future, if we need other relocations, this function can be extended.
12641 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12643 uintptr_t daddr = (uintptr_t)dof;
12644 dof_relohdr_t *dofr =
12645 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12646 dof_sec_t *ss, *rs, *ts;
12650 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12651 sec->dofs_align != sizeof (dof_secidx_t)) {
12652 dtrace_dof_error(dof, "invalid relocation header");
12656 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12657 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12658 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12660 if (ss == NULL || rs == NULL || ts == NULL)
12661 return (-1); /* dtrace_dof_error() has been called already */
12663 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12664 rs->dofs_align != sizeof (uint64_t)) {
12665 dtrace_dof_error(dof, "invalid relocation section");
12669 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12670 n = rs->dofs_size / rs->dofs_entsize;
12672 for (i = 0; i < n; i++) {
12673 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12675 switch (r->dofr_type) {
12676 case DOF_RELO_NONE:
12678 case DOF_RELO_SETX:
12679 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12680 sizeof (uint64_t) > ts->dofs_size) {
12681 dtrace_dof_error(dof, "bad relocation offset");
12685 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12686 dtrace_dof_error(dof, "misaligned setx relo");
12690 *(uint64_t *)taddr += ubase;
12693 dtrace_dof_error(dof, "invalid relocation type");
12697 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12704 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12705 * header: it should be at the front of a memory region that is at least
12706 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12707 * size. It need not be validated in any other way.
12710 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12711 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12713 uint64_t len = dof->dofh_loadsz, seclen;
12714 uintptr_t daddr = (uintptr_t)dof;
12715 dtrace_ecbdesc_t *ep;
12716 dtrace_enabling_t *enab;
12719 ASSERT(MUTEX_HELD(&dtrace_lock));
12720 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12723 * Check the DOF header identification bytes. In addition to checking
12724 * valid settings, we also verify that unused bits/bytes are zeroed so
12725 * we can use them later without fear of regressing existing binaries.
12727 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12728 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12729 dtrace_dof_error(dof, "DOF magic string mismatch");
12733 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12734 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12735 dtrace_dof_error(dof, "DOF has invalid data model");
12739 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12740 dtrace_dof_error(dof, "DOF encoding mismatch");
12744 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12745 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12746 dtrace_dof_error(dof, "DOF version mismatch");
12750 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12751 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12755 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12756 dtrace_dof_error(dof, "DOF uses too many integer registers");
12760 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12761 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12765 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12766 if (dof->dofh_ident[i] != 0) {
12767 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12772 if (dof->dofh_flags & ~DOF_FL_VALID) {
12773 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12777 if (dof->dofh_secsize == 0) {
12778 dtrace_dof_error(dof, "zero section header size");
12783 * Check that the section headers don't exceed the amount of DOF
12784 * data. Note that we cast the section size and number of sections
12785 * to uint64_t's to prevent possible overflow in the multiplication.
12787 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12789 if (dof->dofh_secoff > len || seclen > len ||
12790 dof->dofh_secoff + seclen > len) {
12791 dtrace_dof_error(dof, "truncated section headers");
12795 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12796 dtrace_dof_error(dof, "misaligned section headers");
12800 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12801 dtrace_dof_error(dof, "misaligned section size");
12806 * Take an initial pass through the section headers to be sure that
12807 * the headers don't have stray offsets. If the 'noprobes' flag is
12808 * set, do not permit sections relating to providers, probes, or args.
12810 for (i = 0; i < dof->dofh_secnum; i++) {
12811 dof_sec_t *sec = (dof_sec_t *)(daddr +
12812 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12815 switch (sec->dofs_type) {
12816 case DOF_SECT_PROVIDER:
12817 case DOF_SECT_PROBES:
12818 case DOF_SECT_PRARGS:
12819 case DOF_SECT_PROFFS:
12820 dtrace_dof_error(dof, "illegal sections "
12826 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12827 continue; /* just ignore non-loadable sections */
12829 if (sec->dofs_align & (sec->dofs_align - 1)) {
12830 dtrace_dof_error(dof, "bad section alignment");
12834 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12835 dtrace_dof_error(dof, "misaligned section");
12839 if (sec->dofs_offset > len || sec->dofs_size > len ||
12840 sec->dofs_offset + sec->dofs_size > len) {
12841 dtrace_dof_error(dof, "corrupt section header");
12845 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12846 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12847 dtrace_dof_error(dof, "non-terminating string table");
12853 * Take a second pass through the sections and locate and perform any
12854 * relocations that are present. We do this after the first pass to
12855 * be sure that all sections have had their headers validated.
12857 for (i = 0; i < dof->dofh_secnum; i++) {
12858 dof_sec_t *sec = (dof_sec_t *)(daddr +
12859 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12861 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12862 continue; /* skip sections that are not loadable */
12864 switch (sec->dofs_type) {
12865 case DOF_SECT_URELHDR:
12866 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12872 if ((enab = *enabp) == NULL)
12873 enab = *enabp = dtrace_enabling_create(vstate);
12875 for (i = 0; i < dof->dofh_secnum; i++) {
12876 dof_sec_t *sec = (dof_sec_t *)(daddr +
12877 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12879 if (sec->dofs_type != DOF_SECT_ECBDESC)
12882 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12883 dtrace_enabling_destroy(enab);
12888 dtrace_enabling_add(enab, ep);
12895 * Process DOF for any options. This routine assumes that the DOF has been
12896 * at least processed by dtrace_dof_slurp().
12899 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12904 dof_optdesc_t *desc;
12906 for (i = 0; i < dof->dofh_secnum; i++) {
12907 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12908 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12910 if (sec->dofs_type != DOF_SECT_OPTDESC)
12913 if (sec->dofs_align != sizeof (uint64_t)) {
12914 dtrace_dof_error(dof, "bad alignment in "
12915 "option description");
12919 if ((entsize = sec->dofs_entsize) == 0) {
12920 dtrace_dof_error(dof, "zeroed option entry size");
12924 if (entsize < sizeof (dof_optdesc_t)) {
12925 dtrace_dof_error(dof, "bad option entry size");
12929 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12930 desc = (dof_optdesc_t *)((uintptr_t)dof +
12931 (uintptr_t)sec->dofs_offset + offs);
12933 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12934 dtrace_dof_error(dof, "non-zero option string");
12938 if (desc->dofo_value == DTRACEOPT_UNSET) {
12939 dtrace_dof_error(dof, "unset option");
12943 if ((rval = dtrace_state_option(state,
12944 desc->dofo_option, desc->dofo_value)) != 0) {
12945 dtrace_dof_error(dof, "rejected option");
12955 * DTrace Consumer State Functions
12958 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12960 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12963 dtrace_dynvar_t *dvar, *next, *start;
12966 ASSERT(MUTEX_HELD(&dtrace_lock));
12967 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12969 bzero(dstate, sizeof (dtrace_dstate_t));
12971 if ((dstate->dtds_chunksize = chunksize) == 0)
12972 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12974 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12977 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12980 dstate->dtds_size = size;
12981 dstate->dtds_base = base;
12982 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12983 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12985 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12987 if (hashsize != 1 && (hashsize & 1))
12990 dstate->dtds_hashsize = hashsize;
12991 dstate->dtds_hash = dstate->dtds_base;
12994 * Set all of our hash buckets to point to the single sink, and (if
12995 * it hasn't already been set), set the sink's hash value to be the
12996 * sink sentinel value. The sink is needed for dynamic variable
12997 * lookups to know that they have iterated over an entire, valid hash
13000 for (i = 0; i < hashsize; i++)
13001 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13003 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13004 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13007 * Determine number of active CPUs. Divide free list evenly among
13010 start = (dtrace_dynvar_t *)
13011 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13012 limit = (uintptr_t)base + size;
13014 maxper = (limit - (uintptr_t)start) / NCPU;
13015 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13020 for (i = 0; i < NCPU; i++) {
13022 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13025 * If we don't even have enough chunks to make it once through
13026 * NCPUs, we're just going to allocate everything to the first
13027 * CPU. And if we're on the last CPU, we're going to allocate
13028 * whatever is left over. In either case, we set the limit to
13029 * be the limit of the dynamic variable space.
13031 if (maxper == 0 || i == NCPU - 1) {
13032 limit = (uintptr_t)base + size;
13035 limit = (uintptr_t)start + maxper;
13036 start = (dtrace_dynvar_t *)limit;
13039 ASSERT(limit <= (uintptr_t)base + size);
13042 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13043 dstate->dtds_chunksize);
13045 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13048 dvar->dtdv_next = next;
13060 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13062 ASSERT(MUTEX_HELD(&cpu_lock));
13064 if (dstate->dtds_base == NULL)
13067 kmem_free(dstate->dtds_base, dstate->dtds_size);
13068 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13072 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13075 * Logical XOR, where are you?
13077 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13079 if (vstate->dtvs_nglobals > 0) {
13080 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13081 sizeof (dtrace_statvar_t *));
13084 if (vstate->dtvs_ntlocals > 0) {
13085 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13086 sizeof (dtrace_difv_t));
13089 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13091 if (vstate->dtvs_nlocals > 0) {
13092 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13093 sizeof (dtrace_statvar_t *));
13099 dtrace_state_clean(dtrace_state_t *state)
13101 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13104 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13105 dtrace_speculation_clean(state);
13109 dtrace_state_deadman(dtrace_state_t *state)
13115 now = dtrace_gethrtime();
13117 if (state != dtrace_anon.dta_state &&
13118 now - state->dts_laststatus >= dtrace_deadman_user)
13122 * We must be sure that dts_alive never appears to be less than the
13123 * value upon entry to dtrace_state_deadman(), and because we lack a
13124 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13125 * store INT64_MAX to it, followed by a memory barrier, followed by
13126 * the new value. This assures that dts_alive never appears to be
13127 * less than its true value, regardless of the order in which the
13128 * stores to the underlying storage are issued.
13130 state->dts_alive = INT64_MAX;
13131 dtrace_membar_producer();
13132 state->dts_alive = now;
13136 dtrace_state_clean(void *arg)
13138 dtrace_state_t *state = arg;
13139 dtrace_optval_t *opt = state->dts_options;
13141 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13144 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13145 dtrace_speculation_clean(state);
13147 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13148 dtrace_state_clean, state);
13152 dtrace_state_deadman(void *arg)
13154 dtrace_state_t *state = arg;
13159 dtrace_debug_output();
13161 now = dtrace_gethrtime();
13163 if (state != dtrace_anon.dta_state &&
13164 now - state->dts_laststatus >= dtrace_deadman_user)
13168 * We must be sure that dts_alive never appears to be less than the
13169 * value upon entry to dtrace_state_deadman(), and because we lack a
13170 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13171 * store INT64_MAX to it, followed by a memory barrier, followed by
13172 * the new value. This assures that dts_alive never appears to be
13173 * less than its true value, regardless of the order in which the
13174 * stores to the underlying storage are issued.
13176 state->dts_alive = INT64_MAX;
13177 dtrace_membar_producer();
13178 state->dts_alive = now;
13180 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13181 dtrace_state_deadman, state);
13185 static dtrace_state_t *
13187 dtrace_state_create(dev_t *devp, cred_t *cr)
13189 dtrace_state_create(struct cdev *dev)
13200 dtrace_state_t *state;
13201 dtrace_optval_t *opt;
13202 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13204 ASSERT(MUTEX_HELD(&dtrace_lock));
13205 ASSERT(MUTEX_HELD(&cpu_lock));
13208 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13209 VM_BESTFIT | VM_SLEEP);
13211 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13212 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13216 state = ddi_get_soft_state(dtrace_softstate, minor);
13223 /* Allocate memory for the state. */
13224 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
13227 state->dts_epid = DTRACE_EPIDNONE + 1;
13229 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
13231 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13232 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13234 if (devp != NULL) {
13235 major = getemajor(*devp);
13237 major = ddi_driver_major(dtrace_devi);
13240 state->dts_dev = makedevice(major, minor);
13243 *devp = state->dts_dev;
13245 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
13246 state->dts_dev = dev;
13250 * We allocate NCPU buffers. On the one hand, this can be quite
13251 * a bit of memory per instance (nearly 36K on a Starcat). On the
13252 * other hand, it saves an additional memory reference in the probe
13255 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13256 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13259 state->dts_cleaner = CYCLIC_NONE;
13260 state->dts_deadman = CYCLIC_NONE;
13262 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
13263 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
13265 state->dts_vstate.dtvs_state = state;
13267 for (i = 0; i < DTRACEOPT_MAX; i++)
13268 state->dts_options[i] = DTRACEOPT_UNSET;
13271 * Set the default options.
13273 opt = state->dts_options;
13274 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13275 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13276 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13277 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13278 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13279 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13280 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13281 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13282 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13283 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13284 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13285 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13286 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13287 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13289 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13292 * Depending on the user credentials, we set flag bits which alter probe
13293 * visibility or the amount of destructiveness allowed. In the case of
13294 * actual anonymous tracing, or the possession of all privileges, all of
13295 * the normal checks are bypassed.
13297 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13298 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13299 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13302 * Set up the credentials for this instantiation. We take a
13303 * hold on the credential to prevent it from disappearing on
13304 * us; this in turn prevents the zone_t referenced by this
13305 * credential from disappearing. This means that we can
13306 * examine the credential and the zone from probe context.
13309 state->dts_cred.dcr_cred = cr;
13312 * CRA_PROC means "we have *some* privilege for dtrace" and
13313 * unlocks the use of variables like pid, zonename, etc.
13315 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13316 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13317 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13321 * dtrace_user allows use of syscall and profile providers.
13322 * If the user also has proc_owner and/or proc_zone, we
13323 * extend the scope to include additional visibility and
13324 * destructive power.
13326 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13327 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13328 state->dts_cred.dcr_visible |=
13329 DTRACE_CRV_ALLPROC;
13331 state->dts_cred.dcr_action |=
13332 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13335 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13336 state->dts_cred.dcr_visible |=
13337 DTRACE_CRV_ALLZONE;
13339 state->dts_cred.dcr_action |=
13340 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13344 * If we have all privs in whatever zone this is,
13345 * we can do destructive things to processes which
13346 * have altered credentials.
13349 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13350 cr->cr_zone->zone_privset)) {
13351 state->dts_cred.dcr_action |=
13352 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13358 * Holding the dtrace_kernel privilege also implies that
13359 * the user has the dtrace_user privilege from a visibility
13360 * perspective. But without further privileges, some
13361 * destructive actions are not available.
13363 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13365 * Make all probes in all zones visible. However,
13366 * this doesn't mean that all actions become available
13369 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13370 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13372 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13375 * Holding proc_owner means that destructive actions
13376 * for *this* zone are allowed.
13378 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13379 state->dts_cred.dcr_action |=
13380 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13383 * Holding proc_zone means that destructive actions
13384 * for this user/group ID in all zones is allowed.
13386 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13387 state->dts_cred.dcr_action |=
13388 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13392 * If we have all privs in whatever zone this is,
13393 * we can do destructive things to processes which
13394 * have altered credentials.
13396 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13397 cr->cr_zone->zone_privset)) {
13398 state->dts_cred.dcr_action |=
13399 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13405 * Holding the dtrace_proc privilege gives control over fasttrap
13406 * and pid providers. We need to grant wider destructive
13407 * privileges in the event that the user has proc_owner and/or
13410 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13411 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13412 state->dts_cred.dcr_action |=
13413 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13415 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13416 state->dts_cred.dcr_action |=
13417 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13425 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13427 dtrace_optval_t *opt = state->dts_options, size;
13428 processorid_t cpu = 0;;
13429 int flags = 0, rval, factor, divisor = 1;
13431 ASSERT(MUTEX_HELD(&dtrace_lock));
13432 ASSERT(MUTEX_HELD(&cpu_lock));
13433 ASSERT(which < DTRACEOPT_MAX);
13434 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13435 (state == dtrace_anon.dta_state &&
13436 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13438 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13441 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13442 cpu = opt[DTRACEOPT_CPU];
13444 if (which == DTRACEOPT_SPECSIZE)
13445 flags |= DTRACEBUF_NOSWITCH;
13447 if (which == DTRACEOPT_BUFSIZE) {
13448 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13449 flags |= DTRACEBUF_RING;
13451 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13452 flags |= DTRACEBUF_FILL;
13454 if (state != dtrace_anon.dta_state ||
13455 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13456 flags |= DTRACEBUF_INACTIVE;
13459 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13461 * The size must be 8-byte aligned. If the size is not 8-byte
13462 * aligned, drop it down by the difference.
13464 if (size & (sizeof (uint64_t) - 1))
13465 size -= size & (sizeof (uint64_t) - 1);
13467 if (size < state->dts_reserve) {
13469 * Buffers always must be large enough to accommodate
13470 * their prereserved space. We return E2BIG instead
13471 * of ENOMEM in this case to allow for user-level
13472 * software to differentiate the cases.
13477 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13479 if (rval != ENOMEM) {
13484 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13487 for (divisor = 2; divisor < factor; divisor <<= 1)
13495 dtrace_state_buffers(dtrace_state_t *state)
13497 dtrace_speculation_t *spec = state->dts_speculations;
13500 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13501 DTRACEOPT_BUFSIZE)) != 0)
13504 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13505 DTRACEOPT_AGGSIZE)) != 0)
13508 for (i = 0; i < state->dts_nspeculations; i++) {
13509 if ((rval = dtrace_state_buffer(state,
13510 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13518 dtrace_state_prereserve(dtrace_state_t *state)
13521 dtrace_probe_t *probe;
13523 state->dts_reserve = 0;
13525 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13529 * If our buffer policy is a "fill" buffer policy, we need to set the
13530 * prereserved space to be the space required by the END probes.
13532 probe = dtrace_probes[dtrace_probeid_end - 1];
13533 ASSERT(probe != NULL);
13535 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13536 if (ecb->dte_state != state)
13539 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13544 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13546 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13547 dtrace_speculation_t *spec;
13548 dtrace_buffer_t *buf;
13550 cyc_handler_t hdlr;
13553 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13554 dtrace_icookie_t cookie;
13556 mutex_enter(&cpu_lock);
13557 mutex_enter(&dtrace_lock);
13559 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13565 * Before we can perform any checks, we must prime all of the
13566 * retained enablings that correspond to this state.
13568 dtrace_enabling_prime(state);
13570 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13575 dtrace_state_prereserve(state);
13578 * Now we want to do is try to allocate our speculations.
13579 * We do not automatically resize the number of speculations; if
13580 * this fails, we will fail the operation.
13582 nspec = opt[DTRACEOPT_NSPEC];
13583 ASSERT(nspec != DTRACEOPT_UNSET);
13585 if (nspec > INT_MAX) {
13590 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13591 KM_NOSLEEP | KM_NORMALPRI);
13593 if (spec == NULL) {
13598 state->dts_speculations = spec;
13599 state->dts_nspeculations = (int)nspec;
13601 for (i = 0; i < nspec; i++) {
13602 if ((buf = kmem_zalloc(bufsize,
13603 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
13608 spec[i].dtsp_buffer = buf;
13611 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13612 if (dtrace_anon.dta_state == NULL) {
13617 if (state->dts_necbs != 0) {
13622 state->dts_anon = dtrace_anon_grab();
13623 ASSERT(state->dts_anon != NULL);
13624 state = state->dts_anon;
13627 * We want "grabanon" to be set in the grabbed state, so we'll
13628 * copy that option value from the grabbing state into the
13631 state->dts_options[DTRACEOPT_GRABANON] =
13632 opt[DTRACEOPT_GRABANON];
13634 *cpu = dtrace_anon.dta_beganon;
13637 * If the anonymous state is active (as it almost certainly
13638 * is if the anonymous enabling ultimately matched anything),
13639 * we don't allow any further option processing -- but we
13640 * don't return failure.
13642 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13646 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13647 opt[DTRACEOPT_AGGSIZE] != 0) {
13648 if (state->dts_aggregations == NULL) {
13650 * We're not going to create an aggregation buffer
13651 * because we don't have any ECBs that contain
13652 * aggregations -- set this option to 0.
13654 opt[DTRACEOPT_AGGSIZE] = 0;
13657 * If we have an aggregation buffer, we must also have
13658 * a buffer to use as scratch.
13660 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13661 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13662 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13667 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13668 opt[DTRACEOPT_SPECSIZE] != 0) {
13669 if (!state->dts_speculates) {
13671 * We're not going to create speculation buffers
13672 * because we don't have any ECBs that actually
13673 * speculate -- set the speculation size to 0.
13675 opt[DTRACEOPT_SPECSIZE] = 0;
13680 * The bare minimum size for any buffer that we're actually going to
13681 * do anything to is sizeof (uint64_t).
13683 sz = sizeof (uint64_t);
13685 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13686 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13687 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13689 * A buffer size has been explicitly set to 0 (or to a size
13690 * that will be adjusted to 0) and we need the space -- we
13691 * need to return failure. We return ENOSPC to differentiate
13692 * it from failing to allocate a buffer due to failure to meet
13693 * the reserve (for which we return E2BIG).
13699 if ((rval = dtrace_state_buffers(state)) != 0)
13702 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13703 sz = dtrace_dstate_defsize;
13706 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13711 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13713 } while (sz >>= 1);
13715 opt[DTRACEOPT_DYNVARSIZE] = sz;
13720 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13721 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13723 if (opt[DTRACEOPT_CLEANRATE] == 0)
13724 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13726 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13727 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13729 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13730 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13732 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13734 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13735 hdlr.cyh_arg = state;
13736 hdlr.cyh_level = CY_LOW_LEVEL;
13739 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13741 state->dts_cleaner = cyclic_add(&hdlr, &when);
13743 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13744 hdlr.cyh_arg = state;
13745 hdlr.cyh_level = CY_LOW_LEVEL;
13748 when.cyt_interval = dtrace_deadman_interval;
13750 state->dts_deadman = cyclic_add(&hdlr, &when);
13752 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13753 dtrace_state_clean, state);
13754 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13755 dtrace_state_deadman, state);
13758 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13761 * Now it's time to actually fire the BEGIN probe. We need to disable
13762 * interrupts here both to record the CPU on which we fired the BEGIN
13763 * probe (the data from this CPU will be processed first at user
13764 * level) and to manually activate the buffer for this CPU.
13766 cookie = dtrace_interrupt_disable();
13768 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13769 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13771 dtrace_probe(dtrace_probeid_begin,
13772 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13773 dtrace_interrupt_enable(cookie);
13775 * We may have had an exit action from a BEGIN probe; only change our
13776 * state to ACTIVE if we're still in WARMUP.
13778 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13779 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13781 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13782 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13785 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13786 * want each CPU to transition its principal buffer out of the
13787 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13788 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13789 * atomically transition from processing none of a state's ECBs to
13790 * processing all of them.
13792 dtrace_xcall(DTRACE_CPUALL,
13793 (dtrace_xcall_t)dtrace_buffer_activate, state);
13797 dtrace_buffer_free(state->dts_buffer);
13798 dtrace_buffer_free(state->dts_aggbuffer);
13800 if ((nspec = state->dts_nspeculations) == 0) {
13801 ASSERT(state->dts_speculations == NULL);
13805 spec = state->dts_speculations;
13806 ASSERT(spec != NULL);
13808 for (i = 0; i < state->dts_nspeculations; i++) {
13809 if ((buf = spec[i].dtsp_buffer) == NULL)
13812 dtrace_buffer_free(buf);
13813 kmem_free(buf, bufsize);
13816 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13817 state->dts_nspeculations = 0;
13818 state->dts_speculations = NULL;
13821 mutex_exit(&dtrace_lock);
13822 mutex_exit(&cpu_lock);
13828 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13830 dtrace_icookie_t cookie;
13832 ASSERT(MUTEX_HELD(&dtrace_lock));
13834 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13835 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13839 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13840 * to be sure that every CPU has seen it. See below for the details
13841 * on why this is done.
13843 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13847 * By this point, it is impossible for any CPU to be still processing
13848 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13849 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13850 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13851 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13852 * iff we're in the END probe.
13854 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13856 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13859 * Finally, we can release the reserve and call the END probe. We
13860 * disable interrupts across calling the END probe to allow us to
13861 * return the CPU on which we actually called the END probe. This
13862 * allows user-land to be sure that this CPU's principal buffer is
13865 state->dts_reserve = 0;
13867 cookie = dtrace_interrupt_disable();
13869 dtrace_probe(dtrace_probeid_end,
13870 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13871 dtrace_interrupt_enable(cookie);
13873 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13880 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13881 dtrace_optval_t val)
13883 ASSERT(MUTEX_HELD(&dtrace_lock));
13885 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13888 if (option >= DTRACEOPT_MAX)
13891 if (option != DTRACEOPT_CPU && val < 0)
13895 case DTRACEOPT_DESTRUCTIVE:
13896 if (dtrace_destructive_disallow)
13899 state->dts_cred.dcr_destructive = 1;
13902 case DTRACEOPT_BUFSIZE:
13903 case DTRACEOPT_DYNVARSIZE:
13904 case DTRACEOPT_AGGSIZE:
13905 case DTRACEOPT_SPECSIZE:
13906 case DTRACEOPT_STRSIZE:
13910 if (val >= LONG_MAX) {
13912 * If this is an otherwise negative value, set it to
13913 * the highest multiple of 128m less than LONG_MAX.
13914 * Technically, we're adjusting the size without
13915 * regard to the buffer resizing policy, but in fact,
13916 * this has no effect -- if we set the buffer size to
13917 * ~LONG_MAX and the buffer policy is ultimately set to
13918 * be "manual", the buffer allocation is guaranteed to
13919 * fail, if only because the allocation requires two
13920 * buffers. (We set the the size to the highest
13921 * multiple of 128m because it ensures that the size
13922 * will remain a multiple of a megabyte when
13923 * repeatedly halved -- all the way down to 15m.)
13925 val = LONG_MAX - (1 << 27) + 1;
13929 state->dts_options[option] = val;
13935 dtrace_state_destroy(dtrace_state_t *state)
13938 dtrace_vstate_t *vstate = &state->dts_vstate;
13940 minor_t minor = getminor(state->dts_dev);
13942 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13943 dtrace_speculation_t *spec = state->dts_speculations;
13944 int nspec = state->dts_nspeculations;
13947 ASSERT(MUTEX_HELD(&dtrace_lock));
13948 ASSERT(MUTEX_HELD(&cpu_lock));
13951 * First, retract any retained enablings for this state.
13953 dtrace_enabling_retract(state);
13954 ASSERT(state->dts_nretained == 0);
13956 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13957 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13959 * We have managed to come into dtrace_state_destroy() on a
13960 * hot enabling -- almost certainly because of a disorderly
13961 * shutdown of a consumer. (That is, a consumer that is
13962 * exiting without having called dtrace_stop().) In this case,
13963 * we're going to set our activity to be KILLED, and then
13964 * issue a sync to be sure that everyone is out of probe
13965 * context before we start blowing away ECBs.
13967 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13972 * Release the credential hold we took in dtrace_state_create().
13974 if (state->dts_cred.dcr_cred != NULL)
13975 crfree(state->dts_cred.dcr_cred);
13978 * Now we can safely disable and destroy any enabled probes. Because
13979 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13980 * (especially if they're all enabled), we take two passes through the
13981 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13982 * in the second we disable whatever is left over.
13984 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13985 for (i = 0; i < state->dts_necbs; i++) {
13986 if ((ecb = state->dts_ecbs[i]) == NULL)
13989 if (match && ecb->dte_probe != NULL) {
13990 dtrace_probe_t *probe = ecb->dte_probe;
13991 dtrace_provider_t *prov = probe->dtpr_provider;
13993 if (!(prov->dtpv_priv.dtpp_flags & match))
13997 dtrace_ecb_disable(ecb);
13998 dtrace_ecb_destroy(ecb);
14006 * Before we free the buffers, perform one more sync to assure that
14007 * every CPU is out of probe context.
14011 dtrace_buffer_free(state->dts_buffer);
14012 dtrace_buffer_free(state->dts_aggbuffer);
14014 for (i = 0; i < nspec; i++)
14015 dtrace_buffer_free(spec[i].dtsp_buffer);
14018 if (state->dts_cleaner != CYCLIC_NONE)
14019 cyclic_remove(state->dts_cleaner);
14021 if (state->dts_deadman != CYCLIC_NONE)
14022 cyclic_remove(state->dts_deadman);
14024 callout_stop(&state->dts_cleaner);
14025 callout_drain(&state->dts_cleaner);
14026 callout_stop(&state->dts_deadman);
14027 callout_drain(&state->dts_deadman);
14030 dtrace_dstate_fini(&vstate->dtvs_dynvars);
14031 dtrace_vstate_fini(vstate);
14032 if (state->dts_ecbs != NULL)
14033 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14035 if (state->dts_aggregations != NULL) {
14037 for (i = 0; i < state->dts_naggregations; i++)
14038 ASSERT(state->dts_aggregations[i] == NULL);
14040 ASSERT(state->dts_naggregations > 0);
14041 kmem_free(state->dts_aggregations,
14042 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14045 kmem_free(state->dts_buffer, bufsize);
14046 kmem_free(state->dts_aggbuffer, bufsize);
14048 for (i = 0; i < nspec; i++)
14049 kmem_free(spec[i].dtsp_buffer, bufsize);
14052 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14054 dtrace_format_destroy(state);
14056 if (state->dts_aggid_arena != NULL) {
14058 vmem_destroy(state->dts_aggid_arena);
14060 delete_unrhdr(state->dts_aggid_arena);
14062 state->dts_aggid_arena = NULL;
14065 ddi_soft_state_free(dtrace_softstate, minor);
14066 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14071 * DTrace Anonymous Enabling Functions
14073 static dtrace_state_t *
14074 dtrace_anon_grab(void)
14076 dtrace_state_t *state;
14078 ASSERT(MUTEX_HELD(&dtrace_lock));
14080 if ((state = dtrace_anon.dta_state) == NULL) {
14081 ASSERT(dtrace_anon.dta_enabling == NULL);
14085 ASSERT(dtrace_anon.dta_enabling != NULL);
14086 ASSERT(dtrace_retained != NULL);
14088 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14089 dtrace_anon.dta_enabling = NULL;
14090 dtrace_anon.dta_state = NULL;
14096 dtrace_anon_property(void)
14099 dtrace_state_t *state;
14101 char c[32]; /* enough for "dof-data-" + digits */
14103 ASSERT(MUTEX_HELD(&dtrace_lock));
14104 ASSERT(MUTEX_HELD(&cpu_lock));
14106 for (i = 0; ; i++) {
14107 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14109 dtrace_err_verbose = 1;
14111 if ((dof = dtrace_dof_property(c)) == NULL) {
14112 dtrace_err_verbose = 0;
14118 * We want to create anonymous state, so we need to transition
14119 * the kernel debugger to indicate that DTrace is active. If
14120 * this fails (e.g. because the debugger has modified text in
14121 * some way), we won't continue with the processing.
14123 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14124 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14125 "enabling ignored.");
14126 dtrace_dof_destroy(dof);
14132 * If we haven't allocated an anonymous state, we'll do so now.
14134 if ((state = dtrace_anon.dta_state) == NULL) {
14136 state = dtrace_state_create(NULL, NULL);
14138 state = dtrace_state_create(NULL);
14140 dtrace_anon.dta_state = state;
14142 if (state == NULL) {
14144 * This basically shouldn't happen: the only
14145 * failure mode from dtrace_state_create() is a
14146 * failure of ddi_soft_state_zalloc() that
14147 * itself should never happen. Still, the
14148 * interface allows for a failure mode, and
14149 * we want to fail as gracefully as possible:
14150 * we'll emit an error message and cease
14151 * processing anonymous state in this case.
14153 cmn_err(CE_WARN, "failed to create "
14154 "anonymous state");
14155 dtrace_dof_destroy(dof);
14160 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14161 &dtrace_anon.dta_enabling, 0, B_TRUE);
14164 rv = dtrace_dof_options(dof, state);
14166 dtrace_err_verbose = 0;
14167 dtrace_dof_destroy(dof);
14171 * This is malformed DOF; chuck any anonymous state
14174 ASSERT(dtrace_anon.dta_enabling == NULL);
14175 dtrace_state_destroy(state);
14176 dtrace_anon.dta_state = NULL;
14180 ASSERT(dtrace_anon.dta_enabling != NULL);
14183 if (dtrace_anon.dta_enabling != NULL) {
14187 * dtrace_enabling_retain() can only fail because we are
14188 * trying to retain more enablings than are allowed -- but
14189 * we only have one anonymous enabling, and we are guaranteed
14190 * to be allowed at least one retained enabling; we assert
14191 * that dtrace_enabling_retain() returns success.
14193 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14196 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14201 * DTrace Helper Functions
14204 dtrace_helper_trace(dtrace_helper_action_t *helper,
14205 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14207 uint32_t size, next, nnext, i;
14208 dtrace_helptrace_t *ent;
14209 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
14211 if (!dtrace_helptrace_enabled)
14214 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14217 * What would a tracing framework be without its own tracing
14218 * framework? (Well, a hell of a lot simpler, for starters...)
14220 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14221 sizeof (uint64_t) - sizeof (uint64_t);
14224 * Iterate until we can allocate a slot in the trace buffer.
14227 next = dtrace_helptrace_next;
14229 if (next + size < dtrace_helptrace_bufsize) {
14230 nnext = next + size;
14234 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14237 * We have our slot; fill it in.
14242 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14243 ent->dtht_helper = helper;
14244 ent->dtht_where = where;
14245 ent->dtht_nlocals = vstate->dtvs_nlocals;
14247 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14248 mstate->dtms_fltoffs : -1;
14249 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14250 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
14252 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14253 dtrace_statvar_t *svar;
14255 if ((svar = vstate->dtvs_locals[i]) == NULL)
14258 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14259 ent->dtht_locals[i] =
14260 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
14265 dtrace_helper(int which, dtrace_mstate_t *mstate,
14266 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14268 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
14269 uint64_t sarg0 = mstate->dtms_arg[0];
14270 uint64_t sarg1 = mstate->dtms_arg[1];
14272 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14273 dtrace_helper_action_t *helper;
14274 dtrace_vstate_t *vstate;
14275 dtrace_difo_t *pred;
14276 int i, trace = dtrace_helptrace_enabled;
14278 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14280 if (helpers == NULL)
14283 if ((helper = helpers->dthps_actions[which]) == NULL)
14286 vstate = &helpers->dthps_vstate;
14287 mstate->dtms_arg[0] = arg0;
14288 mstate->dtms_arg[1] = arg1;
14291 * Now iterate over each helper. If its predicate evaluates to 'true',
14292 * we'll call the corresponding actions. Note that the below calls
14293 * to dtrace_dif_emulate() may set faults in machine state. This is
14294 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14295 * the stored DIF offset with its own (which is the desired behavior).
14296 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14297 * from machine state; this is okay, too.
14299 for (; helper != NULL; helper = helper->dtha_next) {
14300 if ((pred = helper->dtha_predicate) != NULL) {
14302 dtrace_helper_trace(helper, mstate, vstate, 0);
14304 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14307 if (*flags & CPU_DTRACE_FAULT)
14311 for (i = 0; i < helper->dtha_nactions; i++) {
14313 dtrace_helper_trace(helper,
14314 mstate, vstate, i + 1);
14316 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14317 mstate, vstate, state);
14319 if (*flags & CPU_DTRACE_FAULT)
14325 dtrace_helper_trace(helper, mstate, vstate,
14326 DTRACE_HELPTRACE_NEXT);
14330 dtrace_helper_trace(helper, mstate, vstate,
14331 DTRACE_HELPTRACE_DONE);
14334 * Restore the arg0 that we saved upon entry.
14336 mstate->dtms_arg[0] = sarg0;
14337 mstate->dtms_arg[1] = sarg1;
14343 dtrace_helper_trace(helper, mstate, vstate,
14344 DTRACE_HELPTRACE_ERR);
14347 * Restore the arg0 that we saved upon entry.
14349 mstate->dtms_arg[0] = sarg0;
14350 mstate->dtms_arg[1] = sarg1;
14356 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14357 dtrace_vstate_t *vstate)
14361 if (helper->dtha_predicate != NULL)
14362 dtrace_difo_release(helper->dtha_predicate, vstate);
14364 for (i = 0; i < helper->dtha_nactions; i++) {
14365 ASSERT(helper->dtha_actions[i] != NULL);
14366 dtrace_difo_release(helper->dtha_actions[i], vstate);
14369 kmem_free(helper->dtha_actions,
14370 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14371 kmem_free(helper, sizeof (dtrace_helper_action_t));
14375 dtrace_helper_destroygen(int gen)
14377 proc_t *p = curproc;
14378 dtrace_helpers_t *help = p->p_dtrace_helpers;
14379 dtrace_vstate_t *vstate;
14382 ASSERT(MUTEX_HELD(&dtrace_lock));
14384 if (help == NULL || gen > help->dthps_generation)
14387 vstate = &help->dthps_vstate;
14389 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14390 dtrace_helper_action_t *last = NULL, *h, *next;
14392 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14393 next = h->dtha_next;
14395 if (h->dtha_generation == gen) {
14396 if (last != NULL) {
14397 last->dtha_next = next;
14399 help->dthps_actions[i] = next;
14402 dtrace_helper_action_destroy(h, vstate);
14410 * Interate until we've cleared out all helper providers with the
14411 * given generation number.
14414 dtrace_helper_provider_t *prov;
14417 * Look for a helper provider with the right generation. We
14418 * have to start back at the beginning of the list each time
14419 * because we drop dtrace_lock. It's unlikely that we'll make
14420 * more than two passes.
14422 for (i = 0; i < help->dthps_nprovs; i++) {
14423 prov = help->dthps_provs[i];
14425 if (prov->dthp_generation == gen)
14430 * If there were no matches, we're done.
14432 if (i == help->dthps_nprovs)
14436 * Move the last helper provider into this slot.
14438 help->dthps_nprovs--;
14439 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14440 help->dthps_provs[help->dthps_nprovs] = NULL;
14442 mutex_exit(&dtrace_lock);
14445 * If we have a meta provider, remove this helper provider.
14447 mutex_enter(&dtrace_meta_lock);
14448 if (dtrace_meta_pid != NULL) {
14449 ASSERT(dtrace_deferred_pid == NULL);
14450 dtrace_helper_provider_remove(&prov->dthp_prov,
14453 mutex_exit(&dtrace_meta_lock);
14455 dtrace_helper_provider_destroy(prov);
14457 mutex_enter(&dtrace_lock);
14464 dtrace_helper_validate(dtrace_helper_action_t *helper)
14469 if ((dp = helper->dtha_predicate) != NULL)
14470 err += dtrace_difo_validate_helper(dp);
14472 for (i = 0; i < helper->dtha_nactions; i++)
14473 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14479 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14481 dtrace_helpers_t *help;
14482 dtrace_helper_action_t *helper, *last;
14483 dtrace_actdesc_t *act;
14484 dtrace_vstate_t *vstate;
14485 dtrace_predicate_t *pred;
14486 int count = 0, nactions = 0, i;
14488 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14491 help = curproc->p_dtrace_helpers;
14492 last = help->dthps_actions[which];
14493 vstate = &help->dthps_vstate;
14495 for (count = 0; last != NULL; last = last->dtha_next) {
14497 if (last->dtha_next == NULL)
14502 * If we already have dtrace_helper_actions_max helper actions for this
14503 * helper action type, we'll refuse to add a new one.
14505 if (count >= dtrace_helper_actions_max)
14508 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14509 helper->dtha_generation = help->dthps_generation;
14511 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14512 ASSERT(pred->dtp_difo != NULL);
14513 dtrace_difo_hold(pred->dtp_difo);
14514 helper->dtha_predicate = pred->dtp_difo;
14517 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14518 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14521 if (act->dtad_difo == NULL)
14527 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14528 (helper->dtha_nactions = nactions), KM_SLEEP);
14530 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14531 dtrace_difo_hold(act->dtad_difo);
14532 helper->dtha_actions[i++] = act->dtad_difo;
14535 if (!dtrace_helper_validate(helper))
14538 if (last == NULL) {
14539 help->dthps_actions[which] = helper;
14541 last->dtha_next = helper;
14544 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14545 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14546 dtrace_helptrace_next = 0;
14551 dtrace_helper_action_destroy(helper, vstate);
14556 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14557 dof_helper_t *dofhp)
14559 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14561 mutex_enter(&dtrace_meta_lock);
14562 mutex_enter(&dtrace_lock);
14564 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14566 * If the dtrace module is loaded but not attached, or if
14567 * there aren't isn't a meta provider registered to deal with
14568 * these provider descriptions, we need to postpone creating
14569 * the actual providers until later.
14572 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14573 dtrace_deferred_pid != help) {
14574 help->dthps_deferred = 1;
14575 help->dthps_pid = p->p_pid;
14576 help->dthps_next = dtrace_deferred_pid;
14577 help->dthps_prev = NULL;
14578 if (dtrace_deferred_pid != NULL)
14579 dtrace_deferred_pid->dthps_prev = help;
14580 dtrace_deferred_pid = help;
14583 mutex_exit(&dtrace_lock);
14585 } else if (dofhp != NULL) {
14587 * If the dtrace module is loaded and we have a particular
14588 * helper provider description, pass that off to the
14592 mutex_exit(&dtrace_lock);
14594 dtrace_helper_provide(dofhp, p->p_pid);
14598 * Otherwise, just pass all the helper provider descriptions
14599 * off to the meta provider.
14603 mutex_exit(&dtrace_lock);
14605 for (i = 0; i < help->dthps_nprovs; i++) {
14606 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14611 mutex_exit(&dtrace_meta_lock);
14615 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14617 dtrace_helpers_t *help;
14618 dtrace_helper_provider_t *hprov, **tmp_provs;
14619 uint_t tmp_maxprovs, i;
14621 ASSERT(MUTEX_HELD(&dtrace_lock));
14623 help = curproc->p_dtrace_helpers;
14624 ASSERT(help != NULL);
14627 * If we already have dtrace_helper_providers_max helper providers,
14628 * we're refuse to add a new one.
14630 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14634 * Check to make sure this isn't a duplicate.
14636 for (i = 0; i < help->dthps_nprovs; i++) {
14637 if (dofhp->dofhp_addr ==
14638 help->dthps_provs[i]->dthp_prov.dofhp_addr)
14642 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14643 hprov->dthp_prov = *dofhp;
14644 hprov->dthp_ref = 1;
14645 hprov->dthp_generation = gen;
14648 * Allocate a bigger table for helper providers if it's already full.
14650 if (help->dthps_maxprovs == help->dthps_nprovs) {
14651 tmp_maxprovs = help->dthps_maxprovs;
14652 tmp_provs = help->dthps_provs;
14654 if (help->dthps_maxprovs == 0)
14655 help->dthps_maxprovs = 2;
14657 help->dthps_maxprovs *= 2;
14658 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14659 help->dthps_maxprovs = dtrace_helper_providers_max;
14661 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14663 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14664 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14666 if (tmp_provs != NULL) {
14667 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14668 sizeof (dtrace_helper_provider_t *));
14669 kmem_free(tmp_provs, tmp_maxprovs *
14670 sizeof (dtrace_helper_provider_t *));
14674 help->dthps_provs[help->dthps_nprovs] = hprov;
14675 help->dthps_nprovs++;
14681 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14683 mutex_enter(&dtrace_lock);
14685 if (--hprov->dthp_ref == 0) {
14687 mutex_exit(&dtrace_lock);
14688 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14689 dtrace_dof_destroy(dof);
14690 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14692 mutex_exit(&dtrace_lock);
14697 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14699 uintptr_t daddr = (uintptr_t)dof;
14700 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14701 dof_provider_t *provider;
14702 dof_probe_t *probe;
14704 char *strtab, *typestr;
14705 dof_stridx_t typeidx;
14707 uint_t nprobes, j, k;
14709 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14711 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14712 dtrace_dof_error(dof, "misaligned section offset");
14717 * The section needs to be large enough to contain the DOF provider
14718 * structure appropriate for the given version.
14720 if (sec->dofs_size <
14721 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14722 offsetof(dof_provider_t, dofpv_prenoffs) :
14723 sizeof (dof_provider_t))) {
14724 dtrace_dof_error(dof, "provider section too small");
14728 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14729 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14730 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14731 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14732 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14734 if (str_sec == NULL || prb_sec == NULL ||
14735 arg_sec == NULL || off_sec == NULL)
14740 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14741 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14742 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14743 provider->dofpv_prenoffs)) == NULL)
14746 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14748 if (provider->dofpv_name >= str_sec->dofs_size ||
14749 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14750 dtrace_dof_error(dof, "invalid provider name");
14754 if (prb_sec->dofs_entsize == 0 ||
14755 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14756 dtrace_dof_error(dof, "invalid entry size");
14760 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14761 dtrace_dof_error(dof, "misaligned entry size");
14765 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14766 dtrace_dof_error(dof, "invalid entry size");
14770 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14771 dtrace_dof_error(dof, "misaligned section offset");
14775 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14776 dtrace_dof_error(dof, "invalid entry size");
14780 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14782 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14785 * Take a pass through the probes to check for errors.
14787 for (j = 0; j < nprobes; j++) {
14788 probe = (dof_probe_t *)(uintptr_t)(daddr +
14789 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14791 if (probe->dofpr_func >= str_sec->dofs_size) {
14792 dtrace_dof_error(dof, "invalid function name");
14796 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14797 dtrace_dof_error(dof, "function name too long");
14801 if (probe->dofpr_name >= str_sec->dofs_size ||
14802 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14803 dtrace_dof_error(dof, "invalid probe name");
14808 * The offset count must not wrap the index, and the offsets
14809 * must also not overflow the section's data.
14811 if (probe->dofpr_offidx + probe->dofpr_noffs <
14812 probe->dofpr_offidx ||
14813 (probe->dofpr_offidx + probe->dofpr_noffs) *
14814 off_sec->dofs_entsize > off_sec->dofs_size) {
14815 dtrace_dof_error(dof, "invalid probe offset");
14819 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14821 * If there's no is-enabled offset section, make sure
14822 * there aren't any is-enabled offsets. Otherwise
14823 * perform the same checks as for probe offsets
14824 * (immediately above).
14826 if (enoff_sec == NULL) {
14827 if (probe->dofpr_enoffidx != 0 ||
14828 probe->dofpr_nenoffs != 0) {
14829 dtrace_dof_error(dof, "is-enabled "
14830 "offsets with null section");
14833 } else if (probe->dofpr_enoffidx +
14834 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14835 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14836 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14837 dtrace_dof_error(dof, "invalid is-enabled "
14842 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14843 dtrace_dof_error(dof, "zero probe and "
14844 "is-enabled offsets");
14847 } else if (probe->dofpr_noffs == 0) {
14848 dtrace_dof_error(dof, "zero probe offsets");
14852 if (probe->dofpr_argidx + probe->dofpr_xargc <
14853 probe->dofpr_argidx ||
14854 (probe->dofpr_argidx + probe->dofpr_xargc) *
14855 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14856 dtrace_dof_error(dof, "invalid args");
14860 typeidx = probe->dofpr_nargv;
14861 typestr = strtab + probe->dofpr_nargv;
14862 for (k = 0; k < probe->dofpr_nargc; k++) {
14863 if (typeidx >= str_sec->dofs_size) {
14864 dtrace_dof_error(dof, "bad "
14865 "native argument type");
14869 typesz = strlen(typestr) + 1;
14870 if (typesz > DTRACE_ARGTYPELEN) {
14871 dtrace_dof_error(dof, "native "
14872 "argument type too long");
14879 typeidx = probe->dofpr_xargv;
14880 typestr = strtab + probe->dofpr_xargv;
14881 for (k = 0; k < probe->dofpr_xargc; k++) {
14882 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14883 dtrace_dof_error(dof, "bad "
14884 "native argument index");
14888 if (typeidx >= str_sec->dofs_size) {
14889 dtrace_dof_error(dof, "bad "
14890 "translated argument type");
14894 typesz = strlen(typestr) + 1;
14895 if (typesz > DTRACE_ARGTYPELEN) {
14896 dtrace_dof_error(dof, "translated argument "
14910 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14912 dtrace_helpers_t *help;
14913 dtrace_vstate_t *vstate;
14914 dtrace_enabling_t *enab = NULL;
14915 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14916 uintptr_t daddr = (uintptr_t)dof;
14918 ASSERT(MUTEX_HELD(&dtrace_lock));
14920 if ((help = curproc->p_dtrace_helpers) == NULL)
14921 help = dtrace_helpers_create(curproc);
14923 vstate = &help->dthps_vstate;
14925 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14926 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14927 dtrace_dof_destroy(dof);
14932 * Look for helper providers and validate their descriptions.
14935 for (i = 0; i < dof->dofh_secnum; i++) {
14936 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14937 dof->dofh_secoff + i * dof->dofh_secsize);
14939 if (sec->dofs_type != DOF_SECT_PROVIDER)
14942 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14943 dtrace_enabling_destroy(enab);
14944 dtrace_dof_destroy(dof);
14953 * Now we need to walk through the ECB descriptions in the enabling.
14955 for (i = 0; i < enab->dten_ndesc; i++) {
14956 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14957 dtrace_probedesc_t *desc = &ep->dted_probe;
14959 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14962 if (strcmp(desc->dtpd_mod, "helper") != 0)
14965 if (strcmp(desc->dtpd_func, "ustack") != 0)
14968 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14971 * Adding this helper action failed -- we are now going
14972 * to rip out the entire generation and return failure.
14974 (void) dtrace_helper_destroygen(help->dthps_generation);
14975 dtrace_enabling_destroy(enab);
14976 dtrace_dof_destroy(dof);
14983 if (nhelpers < enab->dten_ndesc)
14984 dtrace_dof_error(dof, "unmatched helpers");
14986 gen = help->dthps_generation++;
14987 dtrace_enabling_destroy(enab);
14989 if (dhp != NULL && nprovs > 0) {
14990 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14991 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14992 mutex_exit(&dtrace_lock);
14993 dtrace_helper_provider_register(curproc, help, dhp);
14994 mutex_enter(&dtrace_lock);
15001 dtrace_dof_destroy(dof);
15006 static dtrace_helpers_t *
15007 dtrace_helpers_create(proc_t *p)
15009 dtrace_helpers_t *help;
15011 ASSERT(MUTEX_HELD(&dtrace_lock));
15012 ASSERT(p->p_dtrace_helpers == NULL);
15014 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
15015 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
15016 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
15018 p->p_dtrace_helpers = help;
15028 dtrace_helpers_destroy(proc_t *p)
15030 dtrace_helpers_t *help;
15031 dtrace_vstate_t *vstate;
15033 proc_t *p = curproc;
15037 mutex_enter(&dtrace_lock);
15039 ASSERT(p->p_dtrace_helpers != NULL);
15040 ASSERT(dtrace_helpers > 0);
15042 help = p->p_dtrace_helpers;
15043 vstate = &help->dthps_vstate;
15046 * We're now going to lose the help from this process.
15048 p->p_dtrace_helpers = NULL;
15052 * Destory the helper actions.
15054 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15055 dtrace_helper_action_t *h, *next;
15057 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15058 next = h->dtha_next;
15059 dtrace_helper_action_destroy(h, vstate);
15064 mutex_exit(&dtrace_lock);
15067 * Destroy the helper providers.
15069 if (help->dthps_maxprovs > 0) {
15070 mutex_enter(&dtrace_meta_lock);
15071 if (dtrace_meta_pid != NULL) {
15072 ASSERT(dtrace_deferred_pid == NULL);
15074 for (i = 0; i < help->dthps_nprovs; i++) {
15075 dtrace_helper_provider_remove(
15076 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15079 mutex_enter(&dtrace_lock);
15080 ASSERT(help->dthps_deferred == 0 ||
15081 help->dthps_next != NULL ||
15082 help->dthps_prev != NULL ||
15083 help == dtrace_deferred_pid);
15086 * Remove the helper from the deferred list.
15088 if (help->dthps_next != NULL)
15089 help->dthps_next->dthps_prev = help->dthps_prev;
15090 if (help->dthps_prev != NULL)
15091 help->dthps_prev->dthps_next = help->dthps_next;
15092 if (dtrace_deferred_pid == help) {
15093 dtrace_deferred_pid = help->dthps_next;
15094 ASSERT(help->dthps_prev == NULL);
15097 mutex_exit(&dtrace_lock);
15100 mutex_exit(&dtrace_meta_lock);
15102 for (i = 0; i < help->dthps_nprovs; i++) {
15103 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15106 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15107 sizeof (dtrace_helper_provider_t *));
15110 mutex_enter(&dtrace_lock);
15112 dtrace_vstate_fini(&help->dthps_vstate);
15113 kmem_free(help->dthps_actions,
15114 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15115 kmem_free(help, sizeof (dtrace_helpers_t));
15118 mutex_exit(&dtrace_lock);
15125 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15127 dtrace_helpers_t *help, *newhelp;
15128 dtrace_helper_action_t *helper, *new, *last;
15130 dtrace_vstate_t *vstate;
15131 int i, j, sz, hasprovs = 0;
15133 mutex_enter(&dtrace_lock);
15134 ASSERT(from->p_dtrace_helpers != NULL);
15135 ASSERT(dtrace_helpers > 0);
15137 help = from->p_dtrace_helpers;
15138 newhelp = dtrace_helpers_create(to);
15139 ASSERT(to->p_dtrace_helpers != NULL);
15141 newhelp->dthps_generation = help->dthps_generation;
15142 vstate = &newhelp->dthps_vstate;
15145 * Duplicate the helper actions.
15147 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15148 if ((helper = help->dthps_actions[i]) == NULL)
15151 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15152 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15154 new->dtha_generation = helper->dtha_generation;
15156 if ((dp = helper->dtha_predicate) != NULL) {
15157 dp = dtrace_difo_duplicate(dp, vstate);
15158 new->dtha_predicate = dp;
15161 new->dtha_nactions = helper->dtha_nactions;
15162 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15163 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15165 for (j = 0; j < new->dtha_nactions; j++) {
15166 dtrace_difo_t *dp = helper->dtha_actions[j];
15168 ASSERT(dp != NULL);
15169 dp = dtrace_difo_duplicate(dp, vstate);
15170 new->dtha_actions[j] = dp;
15173 if (last != NULL) {
15174 last->dtha_next = new;
15176 newhelp->dthps_actions[i] = new;
15184 * Duplicate the helper providers and register them with the
15185 * DTrace framework.
15187 if (help->dthps_nprovs > 0) {
15188 newhelp->dthps_nprovs = help->dthps_nprovs;
15189 newhelp->dthps_maxprovs = help->dthps_nprovs;
15190 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15191 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15192 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15193 newhelp->dthps_provs[i] = help->dthps_provs[i];
15194 newhelp->dthps_provs[i]->dthp_ref++;
15200 mutex_exit(&dtrace_lock);
15203 dtrace_helper_provider_register(to, newhelp, NULL);
15207 * DTrace Hook Functions
15210 dtrace_module_loaded(modctl_t *ctl)
15212 dtrace_provider_t *prv;
15214 mutex_enter(&dtrace_provider_lock);
15216 mutex_enter(&mod_lock);
15220 ASSERT(ctl->mod_busy);
15224 * We're going to call each providers per-module provide operation
15225 * specifying only this module.
15227 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15228 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15231 mutex_exit(&mod_lock);
15233 mutex_exit(&dtrace_provider_lock);
15236 * If we have any retained enablings, we need to match against them.
15237 * Enabling probes requires that cpu_lock be held, and we cannot hold
15238 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15239 * module. (In particular, this happens when loading scheduling
15240 * classes.) So if we have any retained enablings, we need to dispatch
15241 * our task queue to do the match for us.
15243 mutex_enter(&dtrace_lock);
15245 if (dtrace_retained == NULL) {
15246 mutex_exit(&dtrace_lock);
15250 (void) taskq_dispatch(dtrace_taskq,
15251 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15253 mutex_exit(&dtrace_lock);
15256 * And now, for a little heuristic sleaze: in general, we want to
15257 * match modules as soon as they load. However, we cannot guarantee
15258 * this, because it would lead us to the lock ordering violation
15259 * outlined above. The common case, of course, is that cpu_lock is
15260 * _not_ held -- so we delay here for a clock tick, hoping that that's
15261 * long enough for the task queue to do its work. If it's not, it's
15262 * not a serious problem -- it just means that the module that we
15263 * just loaded may not be immediately instrumentable.
15270 dtrace_module_unloaded(modctl_t *ctl)
15272 dtrace_module_unloaded(modctl_t *ctl, int *error)
15275 dtrace_probe_t template, *probe, *first, *next;
15276 dtrace_provider_t *prov;
15278 char modname[DTRACE_MODNAMELEN];
15283 template.dtpr_mod = ctl->mod_modname;
15285 /* Handle the fact that ctl->filename may end in ".ko". */
15286 strlcpy(modname, ctl->filename, sizeof(modname));
15287 len = strlen(ctl->filename);
15288 if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
15289 modname[len - 3] = '\0';
15290 template.dtpr_mod = modname;
15293 mutex_enter(&dtrace_provider_lock);
15295 mutex_enter(&mod_lock);
15297 mutex_enter(&dtrace_lock);
15300 if (ctl->nenabled > 0) {
15301 /* Don't allow unloads if a probe is enabled. */
15302 mutex_exit(&dtrace_provider_lock);
15303 mutex_exit(&dtrace_lock);
15306 "kldunload: attempt to unload module that has DTrace probes enabled\n");
15311 if (dtrace_bymod == NULL) {
15313 * The DTrace module is loaded (obviously) but not attached;
15314 * we don't have any work to do.
15316 mutex_exit(&dtrace_provider_lock);
15318 mutex_exit(&mod_lock);
15320 mutex_exit(&dtrace_lock);
15324 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15325 probe != NULL; probe = probe->dtpr_nextmod) {
15326 if (probe->dtpr_ecb != NULL) {
15327 mutex_exit(&dtrace_provider_lock);
15329 mutex_exit(&mod_lock);
15331 mutex_exit(&dtrace_lock);
15334 * This shouldn't _actually_ be possible -- we're
15335 * unloading a module that has an enabled probe in it.
15336 * (It's normally up to the provider to make sure that
15337 * this can't happen.) However, because dtps_enable()
15338 * doesn't have a failure mode, there can be an
15339 * enable/unload race. Upshot: we don't want to
15340 * assert, but we're not going to disable the
15343 if (dtrace_err_verbose) {
15345 cmn_err(CE_WARN, "unloaded module '%s' had "
15346 "enabled probes", ctl->mod_modname);
15348 cmn_err(CE_WARN, "unloaded module '%s' had "
15349 "enabled probes", modname);
15359 for (first = NULL; probe != NULL; probe = next) {
15360 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15362 dtrace_probes[probe->dtpr_id - 1] = NULL;
15364 next = probe->dtpr_nextmod;
15365 dtrace_hash_remove(dtrace_bymod, probe);
15366 dtrace_hash_remove(dtrace_byfunc, probe);
15367 dtrace_hash_remove(dtrace_byname, probe);
15369 if (first == NULL) {
15371 probe->dtpr_nextmod = NULL;
15373 probe->dtpr_nextmod = first;
15379 * We've removed all of the module's probes from the hash chains and
15380 * from the probe array. Now issue a dtrace_sync() to be sure that
15381 * everyone has cleared out from any probe array processing.
15385 for (probe = first; probe != NULL; probe = first) {
15386 first = probe->dtpr_nextmod;
15387 prov = probe->dtpr_provider;
15388 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15390 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15391 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15392 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15394 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15396 free_unr(dtrace_arena, probe->dtpr_id);
15398 kmem_free(probe, sizeof (dtrace_probe_t));
15401 mutex_exit(&dtrace_lock);
15403 mutex_exit(&mod_lock);
15405 mutex_exit(&dtrace_provider_lock);
15410 dtrace_kld_load(void *arg __unused, linker_file_t lf)
15413 dtrace_module_loaded(lf);
15417 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
15421 /* We already have an error, so don't do anything. */
15423 dtrace_module_unloaded(lf, error);
15429 dtrace_suspend(void)
15431 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15435 dtrace_resume(void)
15437 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15442 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15444 ASSERT(MUTEX_HELD(&cpu_lock));
15445 mutex_enter(&dtrace_lock);
15449 dtrace_state_t *state;
15450 dtrace_optval_t *opt, rs, c;
15453 * For now, we only allocate a new buffer for anonymous state.
15455 if ((state = dtrace_anon.dta_state) == NULL)
15458 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15461 opt = state->dts_options;
15462 c = opt[DTRACEOPT_CPU];
15464 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15468 * Regardless of what the actual policy is, we're going to
15469 * temporarily set our resize policy to be manual. We're
15470 * also going to temporarily set our CPU option to denote
15471 * the newly configured CPU.
15473 rs = opt[DTRACEOPT_BUFRESIZE];
15474 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15475 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15477 (void) dtrace_state_buffers(state);
15479 opt[DTRACEOPT_BUFRESIZE] = rs;
15480 opt[DTRACEOPT_CPU] = c;
15487 * We don't free the buffer in the CPU_UNCONFIG case. (The
15488 * buffer will be freed when the consumer exits.)
15496 mutex_exit(&dtrace_lock);
15502 dtrace_cpu_setup_initial(processorid_t cpu)
15504 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15509 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15511 if (dtrace_toxranges >= dtrace_toxranges_max) {
15513 dtrace_toxrange_t *range;
15515 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15518 ASSERT(dtrace_toxrange == NULL);
15519 ASSERT(dtrace_toxranges_max == 0);
15520 dtrace_toxranges_max = 1;
15522 dtrace_toxranges_max <<= 1;
15525 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15526 range = kmem_zalloc(nsize, KM_SLEEP);
15528 if (dtrace_toxrange != NULL) {
15529 ASSERT(osize != 0);
15530 bcopy(dtrace_toxrange, range, osize);
15531 kmem_free(dtrace_toxrange, osize);
15534 dtrace_toxrange = range;
15537 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15538 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15540 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15541 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15542 dtrace_toxranges++;
15546 * DTrace Driver Cookbook Functions
15551 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15553 dtrace_provider_id_t id;
15554 dtrace_state_t *state = NULL;
15555 dtrace_enabling_t *enab;
15557 mutex_enter(&cpu_lock);
15558 mutex_enter(&dtrace_provider_lock);
15559 mutex_enter(&dtrace_lock);
15561 if (ddi_soft_state_init(&dtrace_softstate,
15562 sizeof (dtrace_state_t), 0) != 0) {
15563 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15564 mutex_exit(&cpu_lock);
15565 mutex_exit(&dtrace_provider_lock);
15566 mutex_exit(&dtrace_lock);
15567 return (DDI_FAILURE);
15570 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15571 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15572 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15573 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15574 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15575 ddi_remove_minor_node(devi, NULL);
15576 ddi_soft_state_fini(&dtrace_softstate);
15577 mutex_exit(&cpu_lock);
15578 mutex_exit(&dtrace_provider_lock);
15579 mutex_exit(&dtrace_lock);
15580 return (DDI_FAILURE);
15583 ddi_report_dev(devi);
15584 dtrace_devi = devi;
15586 dtrace_modload = dtrace_module_loaded;
15587 dtrace_modunload = dtrace_module_unloaded;
15588 dtrace_cpu_init = dtrace_cpu_setup_initial;
15589 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15590 dtrace_helpers_fork = dtrace_helpers_duplicate;
15591 dtrace_cpustart_init = dtrace_suspend;
15592 dtrace_cpustart_fini = dtrace_resume;
15593 dtrace_debugger_init = dtrace_suspend;
15594 dtrace_debugger_fini = dtrace_resume;
15596 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15598 ASSERT(MUTEX_HELD(&cpu_lock));
15600 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15601 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15602 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15603 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15604 VM_SLEEP | VMC_IDENTIFIER);
15605 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15608 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15609 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15610 NULL, NULL, NULL, NULL, NULL, 0);
15612 ASSERT(MUTEX_HELD(&cpu_lock));
15613 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15614 offsetof(dtrace_probe_t, dtpr_nextmod),
15615 offsetof(dtrace_probe_t, dtpr_prevmod));
15617 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15618 offsetof(dtrace_probe_t, dtpr_nextfunc),
15619 offsetof(dtrace_probe_t, dtpr_prevfunc));
15621 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15622 offsetof(dtrace_probe_t, dtpr_nextname),
15623 offsetof(dtrace_probe_t, dtpr_prevname));
15625 if (dtrace_retain_max < 1) {
15626 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15627 "setting to 1", dtrace_retain_max);
15628 dtrace_retain_max = 1;
15632 * Now discover our toxic ranges.
15634 dtrace_toxic_ranges(dtrace_toxrange_add);
15637 * Before we register ourselves as a provider to our own framework,
15638 * we would like to assert that dtrace_provider is NULL -- but that's
15639 * not true if we were loaded as a dependency of a DTrace provider.
15640 * Once we've registered, we can assert that dtrace_provider is our
15643 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15644 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15646 ASSERT(dtrace_provider != NULL);
15647 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15649 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15650 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15651 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15652 dtrace_provider, NULL, NULL, "END", 0, NULL);
15653 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15654 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15656 dtrace_anon_property();
15657 mutex_exit(&cpu_lock);
15660 * If DTrace helper tracing is enabled, we need to allocate the
15661 * trace buffer and initialize the values.
15663 if (dtrace_helptrace_enabled) {
15664 ASSERT(dtrace_helptrace_buffer == NULL);
15665 dtrace_helptrace_buffer =
15666 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15667 dtrace_helptrace_next = 0;
15671 * If there are already providers, we must ask them to provide their
15672 * probes, and then match any anonymous enabling against them. Note
15673 * that there should be no other retained enablings at this time:
15674 * the only retained enablings at this time should be the anonymous
15677 if (dtrace_anon.dta_enabling != NULL) {
15678 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15680 dtrace_enabling_provide(NULL);
15681 state = dtrace_anon.dta_state;
15684 * We couldn't hold cpu_lock across the above call to
15685 * dtrace_enabling_provide(), but we must hold it to actually
15686 * enable the probes. We have to drop all of our locks, pick
15687 * up cpu_lock, and regain our locks before matching the
15688 * retained anonymous enabling.
15690 mutex_exit(&dtrace_lock);
15691 mutex_exit(&dtrace_provider_lock);
15693 mutex_enter(&cpu_lock);
15694 mutex_enter(&dtrace_provider_lock);
15695 mutex_enter(&dtrace_lock);
15697 if ((enab = dtrace_anon.dta_enabling) != NULL)
15698 (void) dtrace_enabling_match(enab, NULL);
15700 mutex_exit(&cpu_lock);
15703 mutex_exit(&dtrace_lock);
15704 mutex_exit(&dtrace_provider_lock);
15706 if (state != NULL) {
15708 * If we created any anonymous state, set it going now.
15710 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15713 return (DDI_SUCCESS);
15718 #if __FreeBSD_version >= 800039
15719 static void dtrace_dtr(void *);
15726 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15728 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15731 dtrace_state_t *state;
15737 if (getminor(*devp) == DTRACEMNRN_HELPER)
15741 * If this wasn't an open with the "helper" minor, then it must be
15742 * the "dtrace" minor.
15744 ASSERT(getminor(*devp) == DTRACEMNRN_DTRACE);
15746 cred_t *cred_p = NULL;
15748 #if __FreeBSD_version < 800039
15750 * The first minor device is the one that is cloned so there is
15751 * nothing more to do here.
15753 if (dev2unit(dev) == 0)
15757 * Devices are cloned, so if the DTrace state has already
15758 * been allocated, that means this device belongs to a
15759 * different client. Each client should open '/dev/dtrace'
15760 * to get a cloned device.
15762 if (dev->si_drv1 != NULL)
15766 cred_p = dev->si_cred;
15770 * If no DTRACE_PRIV_* bits are set in the credential, then the
15771 * caller lacks sufficient permission to do anything with DTrace.
15773 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15774 if (priv == DTRACE_PRIV_NONE) {
15776 #if __FreeBSD_version < 800039
15777 /* Destroy the cloned device. */
15786 * Ask all providers to provide all their probes.
15788 mutex_enter(&dtrace_provider_lock);
15789 dtrace_probe_provide(NULL, NULL);
15790 mutex_exit(&dtrace_provider_lock);
15792 mutex_enter(&cpu_lock);
15793 mutex_enter(&dtrace_lock);
15795 dtrace_membar_producer();
15799 * If the kernel debugger is active (that is, if the kernel debugger
15800 * modified text in some way), we won't allow the open.
15802 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15804 mutex_exit(&cpu_lock);
15805 mutex_exit(&dtrace_lock);
15809 state = dtrace_state_create(devp, cred_p);
15811 state = dtrace_state_create(dev);
15812 #if __FreeBSD_version < 800039
15813 dev->si_drv1 = state;
15815 devfs_set_cdevpriv(state, dtrace_dtr);
15819 mutex_exit(&cpu_lock);
15821 if (state == NULL) {
15823 if (--dtrace_opens == 0)
15824 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15828 mutex_exit(&dtrace_lock);
15830 #if __FreeBSD_version < 800039
15831 /* Destroy the cloned device. */
15838 mutex_exit(&dtrace_lock);
15846 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15847 #elif __FreeBSD_version < 800039
15849 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15852 dtrace_dtr(void *data)
15856 minor_t minor = getminor(dev);
15857 dtrace_state_t *state;
15859 if (minor == DTRACEMNRN_HELPER)
15862 state = ddi_get_soft_state(dtrace_softstate, minor);
15864 #if __FreeBSD_version < 800039
15865 dtrace_state_t *state = dev->si_drv1;
15867 /* Check if this is not a cloned device. */
15868 if (dev2unit(dev) == 0)
15871 dtrace_state_t *state = data;
15876 mutex_enter(&cpu_lock);
15877 mutex_enter(&dtrace_lock);
15879 if (state != NULL) {
15880 if (state->dts_anon) {
15882 * There is anonymous state. Destroy that first.
15884 ASSERT(dtrace_anon.dta_state == NULL);
15885 dtrace_state_destroy(state->dts_anon);
15888 dtrace_state_destroy(state);
15891 kmem_free(state, 0);
15892 #if __FreeBSD_version < 800039
15893 dev->si_drv1 = NULL;
15898 ASSERT(dtrace_opens > 0);
15900 if (--dtrace_opens == 0)
15901 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15906 mutex_exit(&dtrace_lock);
15907 mutex_exit(&cpu_lock);
15909 #if __FreeBSD_version < 800039
15910 /* Schedule this cloned device to be destroyed. */
15911 destroy_dev_sched(dev);
15914 #if defined(sun) || __FreeBSD_version < 800039
15922 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15925 dof_helper_t help, *dhp = NULL;
15928 case DTRACEHIOC_ADDDOF:
15929 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15930 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15935 arg = (intptr_t)help.dofhp_dof;
15938 case DTRACEHIOC_ADD: {
15939 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15944 mutex_enter(&dtrace_lock);
15947 * dtrace_helper_slurp() takes responsibility for the dof --
15948 * it may free it now or it may save it and free it later.
15950 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15957 mutex_exit(&dtrace_lock);
15961 case DTRACEHIOC_REMOVE: {
15962 mutex_enter(&dtrace_lock);
15963 rval = dtrace_helper_destroygen(arg);
15964 mutex_exit(&dtrace_lock);
15978 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15980 minor_t minor = getminor(dev);
15981 dtrace_state_t *state;
15984 if (minor == DTRACEMNRN_HELPER)
15985 return (dtrace_ioctl_helper(cmd, arg, rv));
15987 state = ddi_get_soft_state(dtrace_softstate, minor);
15989 if (state->dts_anon) {
15990 ASSERT(dtrace_anon.dta_state == NULL);
15991 state = state->dts_anon;
15995 case DTRACEIOC_PROVIDER: {
15996 dtrace_providerdesc_t pvd;
15997 dtrace_provider_t *pvp;
15999 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
16002 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
16003 mutex_enter(&dtrace_provider_lock);
16005 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
16006 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
16010 mutex_exit(&dtrace_provider_lock);
16015 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
16016 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
16018 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
16024 case DTRACEIOC_EPROBE: {
16025 dtrace_eprobedesc_t epdesc;
16027 dtrace_action_t *act;
16033 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
16036 mutex_enter(&dtrace_lock);
16038 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
16039 mutex_exit(&dtrace_lock);
16043 if (ecb->dte_probe == NULL) {
16044 mutex_exit(&dtrace_lock);
16048 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
16049 epdesc.dtepd_uarg = ecb->dte_uarg;
16050 epdesc.dtepd_size = ecb->dte_size;
16052 nrecs = epdesc.dtepd_nrecs;
16053 epdesc.dtepd_nrecs = 0;
16054 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16055 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16058 epdesc.dtepd_nrecs++;
16062 * Now that we have the size, we need to allocate a temporary
16063 * buffer in which to store the complete description. We need
16064 * the temporary buffer to be able to drop dtrace_lock()
16065 * across the copyout(), below.
16067 size = sizeof (dtrace_eprobedesc_t) +
16068 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16070 buf = kmem_alloc(size, KM_SLEEP);
16071 dest = (uintptr_t)buf;
16073 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16074 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16076 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16077 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16083 bcopy(&act->dta_rec, (void *)dest,
16084 sizeof (dtrace_recdesc_t));
16085 dest += sizeof (dtrace_recdesc_t);
16088 mutex_exit(&dtrace_lock);
16090 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16091 kmem_free(buf, size);
16095 kmem_free(buf, size);
16099 case DTRACEIOC_AGGDESC: {
16100 dtrace_aggdesc_t aggdesc;
16101 dtrace_action_t *act;
16102 dtrace_aggregation_t *agg;
16105 dtrace_recdesc_t *lrec;
16110 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16113 mutex_enter(&dtrace_lock);
16115 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16116 mutex_exit(&dtrace_lock);
16120 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16122 nrecs = aggdesc.dtagd_nrecs;
16123 aggdesc.dtagd_nrecs = 0;
16125 offs = agg->dtag_base;
16126 lrec = &agg->dtag_action.dta_rec;
16127 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16129 for (act = agg->dtag_first; ; act = act->dta_next) {
16130 ASSERT(act->dta_intuple ||
16131 DTRACEACT_ISAGG(act->dta_kind));
16134 * If this action has a record size of zero, it
16135 * denotes an argument to the aggregating action.
16136 * Because the presence of this record doesn't (or
16137 * shouldn't) affect the way the data is interpreted,
16138 * we don't copy it out to save user-level the
16139 * confusion of dealing with a zero-length record.
16141 if (act->dta_rec.dtrd_size == 0) {
16142 ASSERT(agg->dtag_hasarg);
16146 aggdesc.dtagd_nrecs++;
16148 if (act == &agg->dtag_action)
16153 * Now that we have the size, we need to allocate a temporary
16154 * buffer in which to store the complete description. We need
16155 * the temporary buffer to be able to drop dtrace_lock()
16156 * across the copyout(), below.
16158 size = sizeof (dtrace_aggdesc_t) +
16159 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16161 buf = kmem_alloc(size, KM_SLEEP);
16162 dest = (uintptr_t)buf;
16164 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16165 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16167 for (act = agg->dtag_first; ; act = act->dta_next) {
16168 dtrace_recdesc_t rec = act->dta_rec;
16171 * See the comment in the above loop for why we pass
16172 * over zero-length records.
16174 if (rec.dtrd_size == 0) {
16175 ASSERT(agg->dtag_hasarg);
16182 rec.dtrd_offset -= offs;
16183 bcopy(&rec, (void *)dest, sizeof (rec));
16184 dest += sizeof (dtrace_recdesc_t);
16186 if (act == &agg->dtag_action)
16190 mutex_exit(&dtrace_lock);
16192 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16193 kmem_free(buf, size);
16197 kmem_free(buf, size);
16201 case DTRACEIOC_ENABLE: {
16203 dtrace_enabling_t *enab = NULL;
16204 dtrace_vstate_t *vstate;
16210 * If a NULL argument has been passed, we take this as our
16211 * cue to reevaluate our enablings.
16214 dtrace_enabling_matchall();
16219 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16222 mutex_enter(&cpu_lock);
16223 mutex_enter(&dtrace_lock);
16224 vstate = &state->dts_vstate;
16226 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16227 mutex_exit(&dtrace_lock);
16228 mutex_exit(&cpu_lock);
16229 dtrace_dof_destroy(dof);
16233 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16234 mutex_exit(&dtrace_lock);
16235 mutex_exit(&cpu_lock);
16236 dtrace_dof_destroy(dof);
16240 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16241 dtrace_enabling_destroy(enab);
16242 mutex_exit(&dtrace_lock);
16243 mutex_exit(&cpu_lock);
16244 dtrace_dof_destroy(dof);
16248 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16249 err = dtrace_enabling_retain(enab);
16251 dtrace_enabling_destroy(enab);
16254 mutex_exit(&cpu_lock);
16255 mutex_exit(&dtrace_lock);
16256 dtrace_dof_destroy(dof);
16261 case DTRACEIOC_REPLICATE: {
16262 dtrace_repldesc_t desc;
16263 dtrace_probedesc_t *match = &desc.dtrpd_match;
16264 dtrace_probedesc_t *create = &desc.dtrpd_create;
16267 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16270 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16271 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16272 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16273 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16275 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16276 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16277 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16278 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16280 mutex_enter(&dtrace_lock);
16281 err = dtrace_enabling_replicate(state, match, create);
16282 mutex_exit(&dtrace_lock);
16287 case DTRACEIOC_PROBEMATCH:
16288 case DTRACEIOC_PROBES: {
16289 dtrace_probe_t *probe = NULL;
16290 dtrace_probedesc_t desc;
16291 dtrace_probekey_t pkey;
16298 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16301 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16302 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16303 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16304 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16307 * Before we attempt to match this probe, we want to give
16308 * all providers the opportunity to provide it.
16310 if (desc.dtpd_id == DTRACE_IDNONE) {
16311 mutex_enter(&dtrace_provider_lock);
16312 dtrace_probe_provide(&desc, NULL);
16313 mutex_exit(&dtrace_provider_lock);
16317 if (cmd == DTRACEIOC_PROBEMATCH) {
16318 dtrace_probekey(&desc, &pkey);
16319 pkey.dtpk_id = DTRACE_IDNONE;
16322 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16324 mutex_enter(&dtrace_lock);
16326 if (cmd == DTRACEIOC_PROBEMATCH) {
16327 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16328 if ((probe = dtrace_probes[i - 1]) != NULL &&
16329 (m = dtrace_match_probe(probe, &pkey,
16330 priv, uid, zoneid)) != 0)
16335 mutex_exit(&dtrace_lock);
16340 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16341 if ((probe = dtrace_probes[i - 1]) != NULL &&
16342 dtrace_match_priv(probe, priv, uid, zoneid))
16347 if (probe == NULL) {
16348 mutex_exit(&dtrace_lock);
16352 dtrace_probe_description(probe, &desc);
16353 mutex_exit(&dtrace_lock);
16355 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16361 case DTRACEIOC_PROBEARG: {
16362 dtrace_argdesc_t desc;
16363 dtrace_probe_t *probe;
16364 dtrace_provider_t *prov;
16366 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16369 if (desc.dtargd_id == DTRACE_IDNONE)
16372 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16375 mutex_enter(&dtrace_provider_lock);
16376 mutex_enter(&mod_lock);
16377 mutex_enter(&dtrace_lock);
16379 if (desc.dtargd_id > dtrace_nprobes) {
16380 mutex_exit(&dtrace_lock);
16381 mutex_exit(&mod_lock);
16382 mutex_exit(&dtrace_provider_lock);
16386 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16387 mutex_exit(&dtrace_lock);
16388 mutex_exit(&mod_lock);
16389 mutex_exit(&dtrace_provider_lock);
16393 mutex_exit(&dtrace_lock);
16395 prov = probe->dtpr_provider;
16397 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16399 * There isn't any typed information for this probe.
16400 * Set the argument number to DTRACE_ARGNONE.
16402 desc.dtargd_ndx = DTRACE_ARGNONE;
16404 desc.dtargd_native[0] = '\0';
16405 desc.dtargd_xlate[0] = '\0';
16406 desc.dtargd_mapping = desc.dtargd_ndx;
16408 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16409 probe->dtpr_id, probe->dtpr_arg, &desc);
16412 mutex_exit(&mod_lock);
16413 mutex_exit(&dtrace_provider_lock);
16415 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16421 case DTRACEIOC_GO: {
16422 processorid_t cpuid;
16423 rval = dtrace_state_go(state, &cpuid);
16428 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16434 case DTRACEIOC_STOP: {
16435 processorid_t cpuid;
16437 mutex_enter(&dtrace_lock);
16438 rval = dtrace_state_stop(state, &cpuid);
16439 mutex_exit(&dtrace_lock);
16444 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16450 case DTRACEIOC_DOFGET: {
16451 dof_hdr_t hdr, *dof;
16454 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16457 mutex_enter(&dtrace_lock);
16458 dof = dtrace_dof_create(state);
16459 mutex_exit(&dtrace_lock);
16461 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16462 rval = copyout(dof, (void *)arg, len);
16463 dtrace_dof_destroy(dof);
16465 return (rval == 0 ? 0 : EFAULT);
16468 case DTRACEIOC_AGGSNAP:
16469 case DTRACEIOC_BUFSNAP: {
16470 dtrace_bufdesc_t desc;
16472 dtrace_buffer_t *buf;
16474 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16477 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16480 mutex_enter(&dtrace_lock);
16482 if (cmd == DTRACEIOC_BUFSNAP) {
16483 buf = &state->dts_buffer[desc.dtbd_cpu];
16485 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16488 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16489 size_t sz = buf->dtb_offset;
16491 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16492 mutex_exit(&dtrace_lock);
16497 * If this buffer has already been consumed, we're
16498 * going to indicate that there's nothing left here
16501 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16502 mutex_exit(&dtrace_lock);
16504 desc.dtbd_size = 0;
16505 desc.dtbd_drops = 0;
16506 desc.dtbd_errors = 0;
16507 desc.dtbd_oldest = 0;
16508 sz = sizeof (desc);
16510 if (copyout(&desc, (void *)arg, sz) != 0)
16517 * If this is a ring buffer that has wrapped, we want
16518 * to copy the whole thing out.
16520 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16521 dtrace_buffer_polish(buf);
16522 sz = buf->dtb_size;
16525 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16526 mutex_exit(&dtrace_lock);
16530 desc.dtbd_size = sz;
16531 desc.dtbd_drops = buf->dtb_drops;
16532 desc.dtbd_errors = buf->dtb_errors;
16533 desc.dtbd_oldest = buf->dtb_xamot_offset;
16534 desc.dtbd_timestamp = dtrace_gethrtime();
16536 mutex_exit(&dtrace_lock);
16538 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16541 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16546 if (buf->dtb_tomax == NULL) {
16547 ASSERT(buf->dtb_xamot == NULL);
16548 mutex_exit(&dtrace_lock);
16552 cached = buf->dtb_tomax;
16553 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16555 dtrace_xcall(desc.dtbd_cpu,
16556 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16558 state->dts_errors += buf->dtb_xamot_errors;
16561 * If the buffers did not actually switch, then the cross call
16562 * did not take place -- presumably because the given CPU is
16563 * not in the ready set. If this is the case, we'll return
16566 if (buf->dtb_tomax == cached) {
16567 ASSERT(buf->dtb_xamot != cached);
16568 mutex_exit(&dtrace_lock);
16572 ASSERT(cached == buf->dtb_xamot);
16575 * We have our snapshot; now copy it out.
16577 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16578 buf->dtb_xamot_offset) != 0) {
16579 mutex_exit(&dtrace_lock);
16583 desc.dtbd_size = buf->dtb_xamot_offset;
16584 desc.dtbd_drops = buf->dtb_xamot_drops;
16585 desc.dtbd_errors = buf->dtb_xamot_errors;
16586 desc.dtbd_oldest = 0;
16587 desc.dtbd_timestamp = buf->dtb_switched;
16589 mutex_exit(&dtrace_lock);
16592 * Finally, copy out the buffer description.
16594 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16600 case DTRACEIOC_CONF: {
16601 dtrace_conf_t conf;
16603 bzero(&conf, sizeof (conf));
16604 conf.dtc_difversion = DIF_VERSION;
16605 conf.dtc_difintregs = DIF_DIR_NREGS;
16606 conf.dtc_diftupregs = DIF_DTR_NREGS;
16607 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16609 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16615 case DTRACEIOC_STATUS: {
16616 dtrace_status_t stat;
16617 dtrace_dstate_t *dstate;
16622 * See the comment in dtrace_state_deadman() for the reason
16623 * for setting dts_laststatus to INT64_MAX before setting
16624 * it to the correct value.
16626 state->dts_laststatus = INT64_MAX;
16627 dtrace_membar_producer();
16628 state->dts_laststatus = dtrace_gethrtime();
16630 bzero(&stat, sizeof (stat));
16632 mutex_enter(&dtrace_lock);
16634 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16635 mutex_exit(&dtrace_lock);
16639 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16640 stat.dtst_exiting = 1;
16642 nerrs = state->dts_errors;
16643 dstate = &state->dts_vstate.dtvs_dynvars;
16645 for (i = 0; i < NCPU; i++) {
16646 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16648 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16649 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16650 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16652 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16653 stat.dtst_filled++;
16655 nerrs += state->dts_buffer[i].dtb_errors;
16657 for (j = 0; j < state->dts_nspeculations; j++) {
16658 dtrace_speculation_t *spec;
16659 dtrace_buffer_t *buf;
16661 spec = &state->dts_speculations[j];
16662 buf = &spec->dtsp_buffer[i];
16663 stat.dtst_specdrops += buf->dtb_xamot_drops;
16667 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16668 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16669 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16670 stat.dtst_dblerrors = state->dts_dblerrors;
16672 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16673 stat.dtst_errors = nerrs;
16675 mutex_exit(&dtrace_lock);
16677 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16683 case DTRACEIOC_FORMAT: {
16684 dtrace_fmtdesc_t fmt;
16688 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16691 mutex_enter(&dtrace_lock);
16693 if (fmt.dtfd_format == 0 ||
16694 fmt.dtfd_format > state->dts_nformats) {
16695 mutex_exit(&dtrace_lock);
16700 * Format strings are allocated contiguously and they are
16701 * never freed; if a format index is less than the number
16702 * of formats, we can assert that the format map is non-NULL
16703 * and that the format for the specified index is non-NULL.
16705 ASSERT(state->dts_formats != NULL);
16706 str = state->dts_formats[fmt.dtfd_format - 1];
16707 ASSERT(str != NULL);
16709 len = strlen(str) + 1;
16711 if (len > fmt.dtfd_length) {
16712 fmt.dtfd_length = len;
16714 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16715 mutex_exit(&dtrace_lock);
16719 if (copyout(str, fmt.dtfd_string, len) != 0) {
16720 mutex_exit(&dtrace_lock);
16725 mutex_exit(&dtrace_lock);
16738 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16740 dtrace_state_t *state;
16747 return (DDI_SUCCESS);
16750 return (DDI_FAILURE);
16753 mutex_enter(&cpu_lock);
16754 mutex_enter(&dtrace_provider_lock);
16755 mutex_enter(&dtrace_lock);
16757 ASSERT(dtrace_opens == 0);
16759 if (dtrace_helpers > 0) {
16760 mutex_exit(&dtrace_provider_lock);
16761 mutex_exit(&dtrace_lock);
16762 mutex_exit(&cpu_lock);
16763 return (DDI_FAILURE);
16766 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16767 mutex_exit(&dtrace_provider_lock);
16768 mutex_exit(&dtrace_lock);
16769 mutex_exit(&cpu_lock);
16770 return (DDI_FAILURE);
16773 dtrace_provider = NULL;
16775 if ((state = dtrace_anon_grab()) != NULL) {
16777 * If there were ECBs on this state, the provider should
16778 * have not been allowed to detach; assert that there is
16781 ASSERT(state->dts_necbs == 0);
16782 dtrace_state_destroy(state);
16785 * If we're being detached with anonymous state, we need to
16786 * indicate to the kernel debugger that DTrace is now inactive.
16788 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16791 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16792 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16793 dtrace_cpu_init = NULL;
16794 dtrace_helpers_cleanup = NULL;
16795 dtrace_helpers_fork = NULL;
16796 dtrace_cpustart_init = NULL;
16797 dtrace_cpustart_fini = NULL;
16798 dtrace_debugger_init = NULL;
16799 dtrace_debugger_fini = NULL;
16800 dtrace_modload = NULL;
16801 dtrace_modunload = NULL;
16803 mutex_exit(&cpu_lock);
16805 if (dtrace_helptrace_enabled) {
16806 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16807 dtrace_helptrace_buffer = NULL;
16810 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16811 dtrace_probes = NULL;
16812 dtrace_nprobes = 0;
16814 dtrace_hash_destroy(dtrace_bymod);
16815 dtrace_hash_destroy(dtrace_byfunc);
16816 dtrace_hash_destroy(dtrace_byname);
16817 dtrace_bymod = NULL;
16818 dtrace_byfunc = NULL;
16819 dtrace_byname = NULL;
16821 kmem_cache_destroy(dtrace_state_cache);
16822 vmem_destroy(dtrace_minor);
16823 vmem_destroy(dtrace_arena);
16825 if (dtrace_toxrange != NULL) {
16826 kmem_free(dtrace_toxrange,
16827 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16828 dtrace_toxrange = NULL;
16829 dtrace_toxranges = 0;
16830 dtrace_toxranges_max = 0;
16833 ddi_remove_minor_node(dtrace_devi, NULL);
16834 dtrace_devi = NULL;
16836 ddi_soft_state_fini(&dtrace_softstate);
16838 ASSERT(dtrace_vtime_references == 0);
16839 ASSERT(dtrace_opens == 0);
16840 ASSERT(dtrace_retained == NULL);
16842 mutex_exit(&dtrace_lock);
16843 mutex_exit(&dtrace_provider_lock);
16846 * We don't destroy the task queue until after we have dropped our
16847 * locks (taskq_destroy() may block on running tasks). To prevent
16848 * attempting to do work after we have effectively detached but before
16849 * the task queue has been destroyed, all tasks dispatched via the
16850 * task queue must check that DTrace is still attached before
16851 * performing any operation.
16853 taskq_destroy(dtrace_taskq);
16854 dtrace_taskq = NULL;
16856 return (DDI_SUCCESS);
16863 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16868 case DDI_INFO_DEVT2DEVINFO:
16869 *result = (void *)dtrace_devi;
16870 error = DDI_SUCCESS;
16872 case DDI_INFO_DEVT2INSTANCE:
16873 *result = (void *)0;
16874 error = DDI_SUCCESS;
16877 error = DDI_FAILURE;
16884 static struct cb_ops dtrace_cb_ops = {
16885 dtrace_open, /* open */
16886 dtrace_close, /* close */
16887 nulldev, /* strategy */
16888 nulldev, /* print */
16892 dtrace_ioctl, /* ioctl */
16893 nodev, /* devmap */
16895 nodev, /* segmap */
16896 nochpoll, /* poll */
16897 ddi_prop_op, /* cb_prop_op */
16899 D_NEW | D_MP /* Driver compatibility flag */
16902 static struct dev_ops dtrace_ops = {
16903 DEVO_REV, /* devo_rev */
16905 dtrace_info, /* get_dev_info */
16906 nulldev, /* identify */
16907 nulldev, /* probe */
16908 dtrace_attach, /* attach */
16909 dtrace_detach, /* detach */
16911 &dtrace_cb_ops, /* driver operations */
16912 NULL, /* bus operations */
16913 nodev /* dev power */
16916 static struct modldrv modldrv = {
16917 &mod_driverops, /* module type (this is a pseudo driver) */
16918 "Dynamic Tracing", /* name of module */
16919 &dtrace_ops, /* driver ops */
16922 static struct modlinkage modlinkage = {
16931 return (mod_install(&modlinkage));
16935 _info(struct modinfo *modinfop)
16937 return (mod_info(&modlinkage, modinfop));
16943 return (mod_remove(&modlinkage));
16947 static d_ioctl_t dtrace_ioctl;
16948 static d_ioctl_t dtrace_ioctl_helper;
16949 static void dtrace_load(void *);
16950 static int dtrace_unload(void);
16951 #if __FreeBSD_version < 800039
16952 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16953 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16954 static eventhandler_tag eh_tag; /* Event handler tag. */
16956 static struct cdev *dtrace_dev;
16957 static struct cdev *helper_dev;
16960 void dtrace_invop_init(void);
16961 void dtrace_invop_uninit(void);
16963 static struct cdevsw dtrace_cdevsw = {
16964 .d_version = D_VERSION,
16965 #if __FreeBSD_version < 800039
16966 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16967 .d_close = dtrace_close,
16969 .d_ioctl = dtrace_ioctl,
16970 .d_open = dtrace_open,
16971 .d_name = "dtrace",
16974 static struct cdevsw helper_cdevsw = {
16975 .d_version = D_VERSION,
16976 .d_ioctl = dtrace_ioctl_helper,
16977 .d_name = "helper",
16980 #include <dtrace_anon.c>
16981 #if __FreeBSD_version < 800039
16982 #include <dtrace_clone.c>
16984 #include <dtrace_ioctl.c>
16985 #include <dtrace_load.c>
16986 #include <dtrace_modevent.c>
16987 #include <dtrace_sysctl.c>
16988 #include <dtrace_unload.c>
16989 #include <dtrace_vtime.c>
16990 #include <dtrace_hacks.c>
16991 #include <dtrace_isa.c>
16993 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
16994 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
16995 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
16997 DEV_MODULE(dtrace, dtrace_modevent, NULL);
16998 MODULE_VERSION(dtrace, 1);
16999 MODULE_DEPEND(dtrace, cyclic, 1, 1, 1);
17000 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);