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) 2013, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2012 by Delphix. All rights reserved.
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__)
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_DECL(_debug_dtrace);
307 SYSCTL_DECL(_kern_dtrace);
311 #define curcpu CPU->cpu_id
316 * DTrace Provider Variables
318 * These are the variables relating to DTrace as a provider (that is, the
319 * provider of the BEGIN, END, and ERROR probes).
321 static dtrace_pattr_t dtrace_provider_attr = {
322 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
323 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
324 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
325 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
326 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
333 static dtrace_pops_t dtrace_provider_ops = {
334 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
335 (void (*)(void *, modctl_t *))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,
339 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
343 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop
346 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
347 static dtrace_id_t dtrace_probeid_end; /* special END probe */
348 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
351 * DTrace Helper Tracing Variables
353 uint32_t dtrace_helptrace_next = 0;
354 uint32_t dtrace_helptrace_nlocals;
355 char *dtrace_helptrace_buffer;
356 int dtrace_helptrace_bufsize = 512 * 1024;
359 int dtrace_helptrace_enabled = 1;
361 int dtrace_helptrace_enabled = 0;
365 * DTrace Error Hashing
367 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
368 * table. This is very useful for checking coverage of tests that are
369 * expected to induce DIF or DOF processing errors, and may be useful for
370 * debugging problems in the DIF code generator or in DOF generation . The
371 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
374 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
375 static const char *dtrace_errlast;
376 static kthread_t *dtrace_errthread;
377 static kmutex_t dtrace_errlock;
381 * DTrace Macros and Constants
383 * These are various macros that are useful in various spots in the
384 * implementation, along with a few random constants that have no meaning
385 * outside of the implementation. There is no real structure to this cpp
386 * mishmash -- but is there ever?
388 #define DTRACE_HASHSTR(hash, probe) \
389 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
391 #define DTRACE_HASHNEXT(hash, probe) \
392 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
394 #define DTRACE_HASHPREV(hash, probe) \
395 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
397 #define DTRACE_HASHEQ(hash, lhs, rhs) \
398 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
399 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
401 #define DTRACE_AGGHASHSIZE_SLEW 17
403 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
406 * The key for a thread-local variable consists of the lower 61 bits of the
407 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
408 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
409 * equal to a variable identifier. This is necessary (but not sufficient) to
410 * assure that global associative arrays never collide with thread-local
411 * variables. To guarantee that they cannot collide, we must also define the
412 * order for keying dynamic variables. That order is:
414 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
416 * Because the variable-key and the tls-key are in orthogonal spaces, there is
417 * no way for a global variable key signature to match a thread-local key
421 #define DTRACE_TLS_THRKEY(where) { \
423 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
424 for (; actv; actv >>= 1) \
426 ASSERT(intr < (1 << 3)); \
427 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
428 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
431 #define DTRACE_TLS_THRKEY(where) { \
432 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
434 uint_t actv = _c->cpu_intr_actv; \
435 for (; actv; actv >>= 1) \
437 ASSERT(intr < (1 << 3)); \
438 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
439 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
443 #define DT_BSWAP_8(x) ((x) & 0xff)
444 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
445 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
446 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
448 #define DT_MASK_LO 0x00000000FFFFFFFFULL
450 #define DTRACE_STORE(type, tomax, offset, what) \
451 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
454 #define DTRACE_ALIGNCHECK(addr, size, flags) \
455 if (addr & (size - 1)) { \
456 *flags |= CPU_DTRACE_BADALIGN; \
457 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
461 #define DTRACE_ALIGNCHECK(addr, size, flags)
465 * Test whether a range of memory starting at testaddr of size testsz falls
466 * within the range of memory described by addr, sz. We take care to avoid
467 * problems with overflow and underflow of the unsigned quantities, and
468 * disallow all negative sizes. Ranges of size 0 are allowed.
470 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
471 ((testaddr) - (baseaddr) < (basesz) && \
472 (testaddr) + (testsz) - (baseaddr) <= (basesz) && \
473 (testaddr) + (testsz) >= (testaddr))
476 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
477 * alloc_sz on the righthand side of the comparison in order to avoid overflow
478 * or underflow in the comparison with it. This is simpler than the INRANGE
479 * check above, because we know that the dtms_scratch_ptr is valid in the
480 * range. Allocations of size zero are allowed.
482 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
483 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
484 (mstate)->dtms_scratch_ptr >= (alloc_sz))
486 #define DTRACE_LOADFUNC(bits) \
489 dtrace_load##bits(uintptr_t addr) \
491 size_t size = bits / NBBY; \
493 uint##bits##_t rval; \
495 volatile uint16_t *flags = (volatile uint16_t *) \
496 &cpu_core[curcpu].cpuc_dtrace_flags; \
498 DTRACE_ALIGNCHECK(addr, size, flags); \
500 for (i = 0; i < dtrace_toxranges; i++) { \
501 if (addr >= dtrace_toxrange[i].dtt_limit) \
504 if (addr + size <= dtrace_toxrange[i].dtt_base) \
508 * This address falls within a toxic region; return 0. \
510 *flags |= CPU_DTRACE_BADADDR; \
511 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
515 *flags |= CPU_DTRACE_NOFAULT; \
517 rval = *((volatile uint##bits##_t *)addr); \
518 *flags &= ~CPU_DTRACE_NOFAULT; \
520 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
524 #define dtrace_loadptr dtrace_load64
526 #define dtrace_loadptr dtrace_load32
529 #define DTRACE_DYNHASH_FREE 0
530 #define DTRACE_DYNHASH_SINK 1
531 #define DTRACE_DYNHASH_VALID 2
533 #define DTRACE_MATCH_NEXT 0
534 #define DTRACE_MATCH_DONE 1
535 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
536 #define DTRACE_STATE_ALIGN 64
538 #define DTRACE_FLAGS2FLT(flags) \
539 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
540 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
541 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
542 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
543 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
544 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
545 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
546 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
547 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
550 #define DTRACEACT_ISSTRING(act) \
551 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
552 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
554 /* Function prototype definitions: */
555 static size_t dtrace_strlen(const char *, size_t);
556 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
557 static void dtrace_enabling_provide(dtrace_provider_t *);
558 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
559 static void dtrace_enabling_matchall(void);
560 static void dtrace_enabling_reap(void);
561 static dtrace_state_t *dtrace_anon_grab(void);
562 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
563 dtrace_state_t *, uint64_t, uint64_t);
564 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
565 static void dtrace_buffer_drop(dtrace_buffer_t *);
566 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
567 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
568 dtrace_state_t *, dtrace_mstate_t *);
569 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
571 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
572 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
573 uint16_t dtrace_load16(uintptr_t);
574 uint32_t dtrace_load32(uintptr_t);
575 uint64_t dtrace_load64(uintptr_t);
576 uint8_t dtrace_load8(uintptr_t);
577 void dtrace_dynvar_clean(dtrace_dstate_t *);
578 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
579 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
580 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
583 * DTrace Probe Context Functions
585 * These functions are called from probe context. Because probe context is
586 * any context in which C may be called, arbitrarily locks may be held,
587 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
588 * As a result, functions called from probe context may only call other DTrace
589 * support functions -- they may not interact at all with the system at large.
590 * (Note that the ASSERT macro is made probe-context safe by redefining it in
591 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
592 * loads are to be performed from probe context, they _must_ be in terms of
593 * the safe dtrace_load*() variants.
595 * Some functions in this block are not actually called from probe context;
596 * for these functions, there will be a comment above the function reading
597 * "Note: not called from probe context."
600 dtrace_panic(const char *format, ...)
604 va_start(alist, format);
605 dtrace_vpanic(format, alist);
610 dtrace_assfail(const char *a, const char *f, int l)
612 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
615 * We just need something here that even the most clever compiler
616 * cannot optimize away.
618 return (a[(uintptr_t)f]);
622 * Atomically increment a specified error counter from probe context.
625 dtrace_error(uint32_t *counter)
628 * Most counters stored to in probe context are per-CPU counters.
629 * However, there are some error conditions that are sufficiently
630 * arcane that they don't merit per-CPU storage. If these counters
631 * are incremented concurrently on different CPUs, scalability will be
632 * adversely affected -- but we don't expect them to be white-hot in a
633 * correctly constructed enabling...
640 if ((nval = oval + 1) == 0) {
642 * If the counter would wrap, set it to 1 -- assuring
643 * that the counter is never zero when we have seen
644 * errors. (The counter must be 32-bits because we
645 * aren't guaranteed a 64-bit compare&swap operation.)
646 * To save this code both the infamy of being fingered
647 * by a priggish news story and the indignity of being
648 * the target of a neo-puritan witch trial, we're
649 * carefully avoiding any colorful description of the
650 * likelihood of this condition -- but suffice it to
651 * say that it is only slightly more likely than the
652 * overflow of predicate cache IDs, as discussed in
653 * dtrace_predicate_create().
657 } while (dtrace_cas32(counter, oval, nval) != oval);
661 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
662 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
670 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
672 if (dest < mstate->dtms_scratch_base)
675 if (dest + size < dest)
678 if (dest + size > mstate->dtms_scratch_ptr)
685 dtrace_canstore_statvar(uint64_t addr, size_t sz,
686 dtrace_statvar_t **svars, int nsvars)
690 for (i = 0; i < nsvars; i++) {
691 dtrace_statvar_t *svar = svars[i];
693 if (svar == NULL || svar->dtsv_size == 0)
696 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
704 * Check to see if the address is within a memory region to which a store may
705 * be issued. This includes the DTrace scratch areas, and any DTrace variable
706 * region. The caller of dtrace_canstore() is responsible for performing any
707 * alignment checks that are needed before stores are actually executed.
710 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
711 dtrace_vstate_t *vstate)
714 * First, check to see if the address is in scratch space...
716 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
717 mstate->dtms_scratch_size))
721 * Now check to see if it's a dynamic variable. This check will pick
722 * up both thread-local variables and any global dynamically-allocated
725 if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
726 vstate->dtvs_dynvars.dtds_size)) {
727 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
728 uintptr_t base = (uintptr_t)dstate->dtds_base +
729 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
733 * Before we assume that we can store here, we need to make
734 * sure that it isn't in our metadata -- storing to our
735 * dynamic variable metadata would corrupt our state. For
736 * the range to not include any dynamic variable metadata,
739 * (1) Start above the hash table that is at the base of
740 * the dynamic variable space
742 * (2) Have a starting chunk offset that is beyond the
743 * dtrace_dynvar_t that is at the base of every chunk
745 * (3) Not span a chunk boundary
751 chunkoffs = (addr - base) % dstate->dtds_chunksize;
753 if (chunkoffs < sizeof (dtrace_dynvar_t))
756 if (chunkoffs + sz > dstate->dtds_chunksize)
763 * Finally, check the static local and global variables. These checks
764 * take the longest, so we perform them last.
766 if (dtrace_canstore_statvar(addr, sz,
767 vstate->dtvs_locals, vstate->dtvs_nlocals))
770 if (dtrace_canstore_statvar(addr, sz,
771 vstate->dtvs_globals, vstate->dtvs_nglobals))
779 * Convenience routine to check to see if the address is within a memory
780 * region in which a load may be issued given the user's privilege level;
781 * if not, it sets the appropriate error flags and loads 'addr' into the
782 * illegal value slot.
784 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
785 * appropriate memory access protection.
788 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
789 dtrace_vstate_t *vstate)
791 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
794 * If we hold the privilege to read from kernel memory, then
795 * everything is readable.
797 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
801 * You can obviously read that which you can store.
803 if (dtrace_canstore(addr, sz, mstate, vstate))
807 * We're allowed to read from our own string table.
809 if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
810 mstate->dtms_difo->dtdo_strlen))
813 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
819 * Convenience routine to check to see if a given string is within a memory
820 * region in which a load may be issued given the user's privilege level;
821 * this exists so that we don't need to issue unnecessary dtrace_strlen()
822 * calls in the event that the user has all privileges.
825 dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
826 dtrace_vstate_t *vstate)
831 * If we hold the privilege to read from kernel memory, then
832 * everything is readable.
834 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
837 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
838 if (dtrace_canload(addr, strsz, mstate, vstate))
845 * Convenience routine to check to see if a given variable is within a memory
846 * region in which a load may be issued given the user's privilege level.
849 dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
850 dtrace_vstate_t *vstate)
853 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
856 * If we hold the privilege to read from kernel memory, then
857 * everything is readable.
859 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
862 if (type->dtdt_kind == DIF_TYPE_STRING)
863 sz = dtrace_strlen(src,
864 vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
866 sz = type->dtdt_size;
868 return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
872 * Compare two strings using safe loads.
875 dtrace_strncmp(char *s1, char *s2, size_t limit)
878 volatile uint16_t *flags;
880 if (s1 == s2 || limit == 0)
883 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
889 c1 = dtrace_load8((uintptr_t)s1++);
895 c2 = dtrace_load8((uintptr_t)s2++);
900 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
906 * Compute strlen(s) for a string using safe memory accesses. The additional
907 * len parameter is used to specify a maximum length to ensure completion.
910 dtrace_strlen(const char *s, size_t lim)
914 for (len = 0; len != lim; len++) {
915 if (dtrace_load8((uintptr_t)s++) == '\0')
923 * Check if an address falls within a toxic region.
926 dtrace_istoxic(uintptr_t kaddr, size_t size)
928 uintptr_t taddr, tsize;
931 for (i = 0; i < dtrace_toxranges; i++) {
932 taddr = dtrace_toxrange[i].dtt_base;
933 tsize = dtrace_toxrange[i].dtt_limit - taddr;
935 if (kaddr - taddr < tsize) {
936 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
937 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
941 if (taddr - kaddr < size) {
942 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
943 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
952 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
953 * memory specified by the DIF program. The dst is assumed to be safe memory
954 * that we can store to directly because it is managed by DTrace. As with
955 * standard bcopy, overlapping copies are handled properly.
958 dtrace_bcopy(const void *src, void *dst, size_t len)
962 const uint8_t *s2 = src;
966 *s1++ = dtrace_load8((uintptr_t)s2++);
967 } while (--len != 0);
973 *--s1 = dtrace_load8((uintptr_t)--s2);
974 } while (--len != 0);
980 * Copy src to dst using safe memory accesses, up to either the specified
981 * length, or the point that a nul byte is encountered. The src is assumed to
982 * be unsafe memory specified by the DIF program. The dst is assumed to be
983 * safe memory that we can store to directly because it is managed by DTrace.
984 * Unlike dtrace_bcopy(), overlapping regions are not handled.
987 dtrace_strcpy(const void *src, void *dst, size_t len)
990 uint8_t *s1 = dst, c;
991 const uint8_t *s2 = src;
994 *s1++ = c = dtrace_load8((uintptr_t)s2++);
995 } while (--len != 0 && c != '\0');
1000 * Copy src to dst, deriving the size and type from the specified (BYREF)
1001 * variable type. The src is assumed to be unsafe memory specified by the DIF
1002 * program. The dst is assumed to be DTrace variable memory that is of the
1003 * specified type; we assume that we can store to directly.
1006 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
1008 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1010 if (type->dtdt_kind == DIF_TYPE_STRING) {
1011 dtrace_strcpy(src, dst, type->dtdt_size);
1013 dtrace_bcopy(src, dst, type->dtdt_size);
1018 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1019 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1020 * safe memory that we can access directly because it is managed by DTrace.
1023 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1025 volatile uint16_t *flags;
1027 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1032 if (s1 == NULL || s2 == NULL)
1035 if (s1 != s2 && len != 0) {
1036 const uint8_t *ps1 = s1;
1037 const uint8_t *ps2 = s2;
1040 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1042 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1048 * Zero the specified region using a simple byte-by-byte loop. Note that this
1049 * is for safe DTrace-managed memory only.
1052 dtrace_bzero(void *dst, size_t len)
1056 for (cp = dst; len != 0; len--)
1061 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1065 result[0] = addend1[0] + addend2[0];
1066 result[1] = addend1[1] + addend2[1] +
1067 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1074 * Shift the 128-bit value in a by b. If b is positive, shift left.
1075 * If b is negative, shift right.
1078 dtrace_shift_128(uint64_t *a, int b)
1088 a[0] = a[1] >> (b - 64);
1092 mask = 1LL << (64 - b);
1094 a[0] |= ((a[1] & mask) << (64 - b));
1099 a[1] = a[0] << (b - 64);
1103 mask = a[0] >> (64 - b);
1111 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1112 * use native multiplication on those, and then re-combine into the
1113 * resulting 128-bit value.
1115 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1122 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1124 uint64_t hi1, hi2, lo1, lo2;
1127 hi1 = factor1 >> 32;
1128 hi2 = factor2 >> 32;
1130 lo1 = factor1 & DT_MASK_LO;
1131 lo2 = factor2 & DT_MASK_LO;
1133 product[0] = lo1 * lo2;
1134 product[1] = hi1 * hi2;
1138 dtrace_shift_128(tmp, 32);
1139 dtrace_add_128(product, tmp, product);
1143 dtrace_shift_128(tmp, 32);
1144 dtrace_add_128(product, tmp, product);
1148 * This privilege check should be used by actions and subroutines to
1149 * verify that the user credentials of the process that enabled the
1150 * invoking ECB match the target credentials
1153 dtrace_priv_proc_common_user(dtrace_state_t *state)
1155 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1158 * We should always have a non-NULL state cred here, since if cred
1159 * is null (anonymous tracing), we fast-path bypass this routine.
1161 ASSERT(s_cr != NULL);
1163 if ((cr = CRED()) != NULL &&
1164 s_cr->cr_uid == cr->cr_uid &&
1165 s_cr->cr_uid == cr->cr_ruid &&
1166 s_cr->cr_uid == cr->cr_suid &&
1167 s_cr->cr_gid == cr->cr_gid &&
1168 s_cr->cr_gid == cr->cr_rgid &&
1169 s_cr->cr_gid == cr->cr_sgid)
1176 * This privilege check should be used by actions and subroutines to
1177 * verify that the zone of the process that enabled the invoking ECB
1178 * matches the target credentials
1181 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1184 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1187 * We should always have a non-NULL state cred here, since if cred
1188 * is null (anonymous tracing), we fast-path bypass this routine.
1190 ASSERT(s_cr != NULL);
1192 if ((cr = CRED()) != NULL &&
1193 s_cr->cr_zone == cr->cr_zone)
1203 * This privilege check should be used by actions and subroutines to
1204 * verify that the process has not setuid or changed credentials.
1207 dtrace_priv_proc_common_nocd(void)
1211 if ((proc = ttoproc(curthread)) != NULL &&
1212 !(proc->p_flag & SNOCD))
1219 dtrace_priv_proc_destructive(dtrace_state_t *state)
1221 int action = state->dts_cred.dcr_action;
1223 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1224 dtrace_priv_proc_common_zone(state) == 0)
1227 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1228 dtrace_priv_proc_common_user(state) == 0)
1231 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1232 dtrace_priv_proc_common_nocd() == 0)
1238 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1244 dtrace_priv_proc_control(dtrace_state_t *state)
1246 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1249 if (dtrace_priv_proc_common_zone(state) &&
1250 dtrace_priv_proc_common_user(state) &&
1251 dtrace_priv_proc_common_nocd())
1254 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1260 dtrace_priv_proc(dtrace_state_t *state)
1262 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1265 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1271 dtrace_priv_kernel(dtrace_state_t *state)
1273 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1276 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1282 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1284 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1287 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1293 * Note: not called from probe context. This function is called
1294 * asynchronously (and at a regular interval) from outside of probe context to
1295 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1296 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1299 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1301 dtrace_dynvar_t *dirty;
1302 dtrace_dstate_percpu_t *dcpu;
1305 for (i = 0; i < NCPU; i++) {
1306 dcpu = &dstate->dtds_percpu[i];
1308 ASSERT(dcpu->dtdsc_rinsing == NULL);
1311 * If the dirty list is NULL, there is no dirty work to do.
1313 if (dcpu->dtdsc_dirty == NULL)
1317 * If the clean list is non-NULL, then we're not going to do
1318 * any work for this CPU -- it means that there has not been
1319 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
1320 * since the last time we cleaned house.
1322 if (dcpu->dtdsc_clean != NULL)
1328 * Atomically move the dirty list aside.
1331 dirty = dcpu->dtdsc_dirty;
1334 * Before we zap the dirty list, set the rinsing list.
1335 * (This allows for a potential assertion in
1336 * dtrace_dynvar(): if a free dynamic variable appears
1337 * on a hash chain, either the dirty list or the
1338 * rinsing list for some CPU must be non-NULL.)
1340 dcpu->dtdsc_rinsing = dirty;
1341 dtrace_membar_producer();
1342 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1343 dirty, NULL) != dirty);
1348 * We have no work to do; we can simply return.
1355 for (i = 0; i < NCPU; i++) {
1356 dcpu = &dstate->dtds_percpu[i];
1358 if (dcpu->dtdsc_rinsing == NULL)
1362 * We are now guaranteed that no hash chain contains a pointer
1363 * into this dirty list; we can make it clean.
1365 ASSERT(dcpu->dtdsc_clean == NULL);
1366 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1367 dcpu->dtdsc_rinsing = NULL;
1371 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1372 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1373 * This prevents a race whereby a CPU incorrectly decides that
1374 * the state should be something other than DTRACE_DSTATE_CLEAN
1375 * after dtrace_dynvar_clean() has completed.
1379 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1383 * Depending on the value of the op parameter, this function looks-up,
1384 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1385 * allocation is requested, this function will return a pointer to a
1386 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1387 * variable can be allocated. If NULL is returned, the appropriate counter
1388 * will be incremented.
1391 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1392 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1393 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1395 uint64_t hashval = DTRACE_DYNHASH_VALID;
1396 dtrace_dynhash_t *hash = dstate->dtds_hash;
1397 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1398 processorid_t me = curcpu, cpu = me;
1399 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1400 size_t bucket, ksize;
1401 size_t chunksize = dstate->dtds_chunksize;
1402 uintptr_t kdata, lock, nstate;
1408 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1409 * algorithm. For the by-value portions, we perform the algorithm in
1410 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1411 * bit, and seems to have only a minute effect on distribution. For
1412 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1413 * over each referenced byte. It's painful to do this, but it's much
1414 * better than pathological hash distribution. The efficacy of the
1415 * hashing algorithm (and a comparison with other algorithms) may be
1416 * found by running the ::dtrace_dynstat MDB dcmd.
1418 for (i = 0; i < nkeys; i++) {
1419 if (key[i].dttk_size == 0) {
1420 uint64_t val = key[i].dttk_value;
1422 hashval += (val >> 48) & 0xffff;
1423 hashval += (hashval << 10);
1424 hashval ^= (hashval >> 6);
1426 hashval += (val >> 32) & 0xffff;
1427 hashval += (hashval << 10);
1428 hashval ^= (hashval >> 6);
1430 hashval += (val >> 16) & 0xffff;
1431 hashval += (hashval << 10);
1432 hashval ^= (hashval >> 6);
1434 hashval += val & 0xffff;
1435 hashval += (hashval << 10);
1436 hashval ^= (hashval >> 6);
1439 * This is incredibly painful, but it beats the hell
1440 * out of the alternative.
1442 uint64_t j, size = key[i].dttk_size;
1443 uintptr_t base = (uintptr_t)key[i].dttk_value;
1445 if (!dtrace_canload(base, size, mstate, vstate))
1448 for (j = 0; j < size; j++) {
1449 hashval += dtrace_load8(base + j);
1450 hashval += (hashval << 10);
1451 hashval ^= (hashval >> 6);
1456 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1459 hashval += (hashval << 3);
1460 hashval ^= (hashval >> 11);
1461 hashval += (hashval << 15);
1464 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1465 * comes out to be one of our two sentinel hash values. If this
1466 * actually happens, we set the hashval to be a value known to be a
1467 * non-sentinel value.
1469 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1470 hashval = DTRACE_DYNHASH_VALID;
1473 * Yes, it's painful to do a divide here. If the cycle count becomes
1474 * important here, tricks can be pulled to reduce it. (However, it's
1475 * critical that hash collisions be kept to an absolute minimum;
1476 * they're much more painful than a divide.) It's better to have a
1477 * solution that generates few collisions and still keeps things
1478 * relatively simple.
1480 bucket = hashval % dstate->dtds_hashsize;
1482 if (op == DTRACE_DYNVAR_DEALLOC) {
1483 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1486 while ((lock = *lockp) & 1)
1489 if (dtrace_casptr((volatile void *)lockp,
1490 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1494 dtrace_membar_producer();
1499 lock = hash[bucket].dtdh_lock;
1501 dtrace_membar_consumer();
1503 start = hash[bucket].dtdh_chain;
1504 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1505 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1506 op != DTRACE_DYNVAR_DEALLOC));
1508 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1509 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1510 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1512 if (dvar->dtdv_hashval != hashval) {
1513 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
1515 * We've reached the sink, and therefore the
1516 * end of the hash chain; we can kick out of
1517 * the loop knowing that we have seen a valid
1518 * snapshot of state.
1520 ASSERT(dvar->dtdv_next == NULL);
1521 ASSERT(dvar == &dtrace_dynhash_sink);
1525 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
1527 * We've gone off the rails: somewhere along
1528 * the line, one of the members of this hash
1529 * chain was deleted. Note that we could also
1530 * detect this by simply letting this loop run
1531 * to completion, as we would eventually hit
1532 * the end of the dirty list. However, we
1533 * want to avoid running the length of the
1534 * dirty list unnecessarily (it might be quite
1535 * long), so we catch this as early as
1536 * possible by detecting the hash marker. In
1537 * this case, we simply set dvar to NULL and
1538 * break; the conditional after the loop will
1539 * send us back to top.
1548 if (dtuple->dtt_nkeys != nkeys)
1551 for (i = 0; i < nkeys; i++, dkey++) {
1552 if (dkey->dttk_size != key[i].dttk_size)
1553 goto next; /* size or type mismatch */
1555 if (dkey->dttk_size != 0) {
1557 (void *)(uintptr_t)key[i].dttk_value,
1558 (void *)(uintptr_t)dkey->dttk_value,
1562 if (dkey->dttk_value != key[i].dttk_value)
1567 if (op != DTRACE_DYNVAR_DEALLOC)
1570 ASSERT(dvar->dtdv_next == NULL ||
1571 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
1574 ASSERT(hash[bucket].dtdh_chain != dvar);
1575 ASSERT(start != dvar);
1576 ASSERT(prev->dtdv_next == dvar);
1577 prev->dtdv_next = dvar->dtdv_next;
1579 if (dtrace_casptr(&hash[bucket].dtdh_chain,
1580 start, dvar->dtdv_next) != start) {
1582 * We have failed to atomically swing the
1583 * hash table head pointer, presumably because
1584 * of a conflicting allocation on another CPU.
1585 * We need to reread the hash chain and try
1592 dtrace_membar_producer();
1595 * Now set the hash value to indicate that it's free.
1597 ASSERT(hash[bucket].dtdh_chain != dvar);
1598 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1600 dtrace_membar_producer();
1603 * Set the next pointer to point at the dirty list, and
1604 * atomically swing the dirty pointer to the newly freed dvar.
1607 next = dcpu->dtdsc_dirty;
1608 dvar->dtdv_next = next;
1609 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
1612 * Finally, unlock this hash bucket.
1614 ASSERT(hash[bucket].dtdh_lock == lock);
1616 hash[bucket].dtdh_lock++;
1626 * If dvar is NULL, it is because we went off the rails:
1627 * one of the elements that we traversed in the hash chain
1628 * was deleted while we were traversing it. In this case,
1629 * we assert that we aren't doing a dealloc (deallocs lock
1630 * the hash bucket to prevent themselves from racing with
1631 * one another), and retry the hash chain traversal.
1633 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
1637 if (op != DTRACE_DYNVAR_ALLOC) {
1639 * If we are not to allocate a new variable, we want to
1640 * return NULL now. Before we return, check that the value
1641 * of the lock word hasn't changed. If it has, we may have
1642 * seen an inconsistent snapshot.
1644 if (op == DTRACE_DYNVAR_NOALLOC) {
1645 if (hash[bucket].dtdh_lock != lock)
1648 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
1649 ASSERT(hash[bucket].dtdh_lock == lock);
1651 hash[bucket].dtdh_lock++;
1658 * We need to allocate a new dynamic variable. The size we need is the
1659 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
1660 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
1661 * the size of any referred-to data (dsize). We then round the final
1662 * size up to the chunksize for allocation.
1664 for (ksize = 0, i = 0; i < nkeys; i++)
1665 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
1668 * This should be pretty much impossible, but could happen if, say,
1669 * strange DIF specified the tuple. Ideally, this should be an
1670 * assertion and not an error condition -- but that requires that the
1671 * chunksize calculation in dtrace_difo_chunksize() be absolutely
1672 * bullet-proof. (That is, it must not be able to be fooled by
1673 * malicious DIF.) Given the lack of backwards branches in DIF,
1674 * solving this would presumably not amount to solving the Halting
1675 * Problem -- but it still seems awfully hard.
1677 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
1678 ksize + dsize > chunksize) {
1679 dcpu->dtdsc_drops++;
1683 nstate = DTRACE_DSTATE_EMPTY;
1687 free = dcpu->dtdsc_free;
1690 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
1693 if (clean == NULL) {
1695 * We're out of dynamic variable space on
1696 * this CPU. Unless we have tried all CPUs,
1697 * we'll try to allocate from a different
1700 switch (dstate->dtds_state) {
1701 case DTRACE_DSTATE_CLEAN: {
1702 void *sp = &dstate->dtds_state;
1707 if (dcpu->dtdsc_dirty != NULL &&
1708 nstate == DTRACE_DSTATE_EMPTY)
1709 nstate = DTRACE_DSTATE_DIRTY;
1711 if (dcpu->dtdsc_rinsing != NULL)
1712 nstate = DTRACE_DSTATE_RINSING;
1714 dcpu = &dstate->dtds_percpu[cpu];
1719 (void) dtrace_cas32(sp,
1720 DTRACE_DSTATE_CLEAN, nstate);
1723 * To increment the correct bean
1724 * counter, take another lap.
1729 case DTRACE_DSTATE_DIRTY:
1730 dcpu->dtdsc_dirty_drops++;
1733 case DTRACE_DSTATE_RINSING:
1734 dcpu->dtdsc_rinsing_drops++;
1737 case DTRACE_DSTATE_EMPTY:
1738 dcpu->dtdsc_drops++;
1742 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
1747 * The clean list appears to be non-empty. We want to
1748 * move the clean list to the free list; we start by
1749 * moving the clean pointer aside.
1751 if (dtrace_casptr(&dcpu->dtdsc_clean,
1752 clean, NULL) != clean) {
1754 * We are in one of two situations:
1756 * (a) The clean list was switched to the
1757 * free list by another CPU.
1759 * (b) The clean list was added to by the
1762 * In either of these situations, we can
1763 * just reattempt the free list allocation.
1768 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
1771 * Now we'll move the clean list to the free list.
1772 * It's impossible for this to fail: the only way
1773 * the free list can be updated is through this
1774 * code path, and only one CPU can own the clean list.
1775 * Thus, it would only be possible for this to fail if
1776 * this code were racing with dtrace_dynvar_clean().
1777 * (That is, if dtrace_dynvar_clean() updated the clean
1778 * list, and we ended up racing to update the free
1779 * list.) This race is prevented by the dtrace_sync()
1780 * in dtrace_dynvar_clean() -- which flushes the
1781 * owners of the clean lists out before resetting
1784 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
1785 ASSERT(rval == NULL);
1790 new_free = dvar->dtdv_next;
1791 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
1794 * We have now allocated a new chunk. We copy the tuple keys into the
1795 * tuple array and copy any referenced key data into the data space
1796 * following the tuple array. As we do this, we relocate dttk_value
1797 * in the final tuple to point to the key data address in the chunk.
1799 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
1800 dvar->dtdv_data = (void *)(kdata + ksize);
1801 dvar->dtdv_tuple.dtt_nkeys = nkeys;
1803 for (i = 0; i < nkeys; i++) {
1804 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
1805 size_t kesize = key[i].dttk_size;
1809 (const void *)(uintptr_t)key[i].dttk_value,
1810 (void *)kdata, kesize);
1811 dkey->dttk_value = kdata;
1812 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
1814 dkey->dttk_value = key[i].dttk_value;
1817 dkey->dttk_size = kesize;
1820 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
1821 dvar->dtdv_hashval = hashval;
1822 dvar->dtdv_next = start;
1824 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
1828 * The cas has failed. Either another CPU is adding an element to
1829 * this hash chain, or another CPU is deleting an element from this
1830 * hash chain. The simplest way to deal with both of these cases
1831 * (though not necessarily the most efficient) is to free our
1832 * allocated block and tail-call ourselves. Note that the free is
1833 * to the dirty list and _not_ to the free list. This is to prevent
1834 * races with allocators, above.
1836 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
1838 dtrace_membar_producer();
1841 free = dcpu->dtdsc_dirty;
1842 dvar->dtdv_next = free;
1843 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
1845 return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
1850 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
1852 if ((int64_t)nval < (int64_t)*oval)
1858 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
1860 if ((int64_t)nval > (int64_t)*oval)
1865 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
1867 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
1868 int64_t val = (int64_t)nval;
1871 for (i = 0; i < zero; i++) {
1872 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
1878 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
1879 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
1880 quanta[i - 1] += incr;
1885 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
1893 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
1895 uint64_t arg = *lquanta++;
1896 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
1897 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
1898 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
1899 int32_t val = (int32_t)nval, level;
1902 ASSERT(levels != 0);
1906 * This is an underflow.
1912 level = (val - base) / step;
1914 if (level < levels) {
1915 lquanta[level + 1] += incr;
1920 * This is an overflow.
1922 lquanta[levels + 1] += incr;
1926 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
1927 uint16_t high, uint16_t nsteps, int64_t value)
1929 int64_t this = 1, last, next;
1930 int base = 1, order;
1932 ASSERT(factor <= nsteps);
1933 ASSERT(nsteps % factor == 0);
1935 for (order = 0; order < low; order++)
1939 * If our value is less than our factor taken to the power of the
1940 * low order of magnitude, it goes into the zeroth bucket.
1942 if (value < (last = this))
1945 for (this *= factor; order <= high; order++) {
1946 int nbuckets = this > nsteps ? nsteps : this;
1948 if ((next = this * factor) < this) {
1950 * We should not generally get log/linear quantizations
1951 * with a high magnitude that allows 64-bits to
1952 * overflow, but we nonetheless protect against this
1953 * by explicitly checking for overflow, and clamping
1954 * our value accordingly.
1961 * If our value lies within this order of magnitude,
1962 * determine its position by taking the offset within
1963 * the order of magnitude, dividing by the bucket
1964 * width, and adding to our (accumulated) base.
1966 return (base + (value - last) / (this / nbuckets));
1969 base += nbuckets - (nbuckets / factor);
1975 * Our value is greater than or equal to our factor taken to the
1976 * power of one plus the high magnitude -- return the top bucket.
1982 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
1984 uint64_t arg = *llquanta++;
1985 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
1986 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
1987 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
1988 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
1990 llquanta[dtrace_aggregate_llquantize_bucket(factor,
1991 low, high, nsteps, nval)] += incr;
1996 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2004 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2006 int64_t snval = (int64_t)nval;
2013 * What we want to say here is:
2015 * data[2] += nval * nval;
2017 * But given that nval is 64-bit, we could easily overflow, so
2018 * we do this as 128-bit arithmetic.
2023 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2024 dtrace_add_128(data + 2, tmp, data + 2);
2029 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2036 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2042 * Aggregate given the tuple in the principal data buffer, and the aggregating
2043 * action denoted by the specified dtrace_aggregation_t. The aggregation
2044 * buffer is specified as the buf parameter. This routine does not return
2045 * failure; if there is no space in the aggregation buffer, the data will be
2046 * dropped, and a corresponding counter incremented.
2049 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2050 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2052 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2053 uint32_t i, ndx, size, fsize;
2054 uint32_t align = sizeof (uint64_t) - 1;
2055 dtrace_aggbuffer_t *agb;
2056 dtrace_aggkey_t *key;
2057 uint32_t hashval = 0, limit, isstr;
2058 caddr_t tomax, data, kdata;
2059 dtrace_actkind_t action;
2060 dtrace_action_t *act;
2066 if (!agg->dtag_hasarg) {
2068 * Currently, only quantize() and lquantize() take additional
2069 * arguments, and they have the same semantics: an increment
2070 * value that defaults to 1 when not present. If additional
2071 * aggregating actions take arguments, the setting of the
2072 * default argument value will presumably have to become more
2078 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2079 size = rec->dtrd_offset - agg->dtag_base;
2080 fsize = size + rec->dtrd_size;
2082 ASSERT(dbuf->dtb_tomax != NULL);
2083 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2085 if ((tomax = buf->dtb_tomax) == NULL) {
2086 dtrace_buffer_drop(buf);
2091 * The metastructure is always at the bottom of the buffer.
2093 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2094 sizeof (dtrace_aggbuffer_t));
2096 if (buf->dtb_offset == 0) {
2098 * We just kludge up approximately 1/8th of the size to be
2099 * buckets. If this guess ends up being routinely
2100 * off-the-mark, we may need to dynamically readjust this
2101 * based on past performance.
2103 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2105 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2106 (uintptr_t)tomax || hashsize == 0) {
2108 * We've been given a ludicrously small buffer;
2109 * increment our drop count and leave.
2111 dtrace_buffer_drop(buf);
2116 * And now, a pathetic attempt to try to get a an odd (or
2117 * perchance, a prime) hash size for better hash distribution.
2119 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2120 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2122 agb->dtagb_hashsize = hashsize;
2123 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2124 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2125 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2127 for (i = 0; i < agb->dtagb_hashsize; i++)
2128 agb->dtagb_hash[i] = NULL;
2131 ASSERT(agg->dtag_first != NULL);
2132 ASSERT(agg->dtag_first->dta_intuple);
2135 * Calculate the hash value based on the key. Note that we _don't_
2136 * include the aggid in the hashing (but we will store it as part of
2137 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2138 * algorithm: a simple, quick algorithm that has no known funnels, and
2139 * gets good distribution in practice. The efficacy of the hashing
2140 * algorithm (and a comparison with other algorithms) may be found by
2141 * running the ::dtrace_aggstat MDB dcmd.
2143 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2144 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2145 limit = i + act->dta_rec.dtrd_size;
2146 ASSERT(limit <= size);
2147 isstr = DTRACEACT_ISSTRING(act);
2149 for (; i < limit; i++) {
2151 hashval += (hashval << 10);
2152 hashval ^= (hashval >> 6);
2154 if (isstr && data[i] == '\0')
2159 hashval += (hashval << 3);
2160 hashval ^= (hashval >> 11);
2161 hashval += (hashval << 15);
2164 * Yes, the divide here is expensive -- but it's generally the least
2165 * of the performance issues given the amount of data that we iterate
2166 * over to compute hash values, compare data, etc.
2168 ndx = hashval % agb->dtagb_hashsize;
2170 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2171 ASSERT((caddr_t)key >= tomax);
2172 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2174 if (hashval != key->dtak_hashval || key->dtak_size != size)
2177 kdata = key->dtak_data;
2178 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2180 for (act = agg->dtag_first; act->dta_intuple;
2181 act = act->dta_next) {
2182 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2183 limit = i + act->dta_rec.dtrd_size;
2184 ASSERT(limit <= size);
2185 isstr = DTRACEACT_ISSTRING(act);
2187 for (; i < limit; i++) {
2188 if (kdata[i] != data[i])
2191 if (isstr && data[i] == '\0')
2196 if (action != key->dtak_action) {
2198 * We are aggregating on the same value in the same
2199 * aggregation with two different aggregating actions.
2200 * (This should have been picked up in the compiler,
2201 * so we may be dealing with errant or devious DIF.)
2202 * This is an error condition; we indicate as much,
2205 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2210 * This is a hit: we need to apply the aggregator to
2211 * the value at this key.
2213 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2220 * We didn't find it. We need to allocate some zero-filled space,
2221 * link it into the hash table appropriately, and apply the aggregator
2222 * to the (zero-filled) value.
2224 offs = buf->dtb_offset;
2225 while (offs & (align - 1))
2226 offs += sizeof (uint32_t);
2229 * If we don't have enough room to both allocate a new key _and_
2230 * its associated data, increment the drop count and return.
2232 if ((uintptr_t)tomax + offs + fsize >
2233 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2234 dtrace_buffer_drop(buf);
2239 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2240 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2241 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2243 key->dtak_data = kdata = tomax + offs;
2244 buf->dtb_offset = offs + fsize;
2247 * Now copy the data across.
2249 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2251 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2255 * Because strings are not zeroed out by default, we need to iterate
2256 * looking for actions that store strings, and we need to explicitly
2257 * pad these strings out with zeroes.
2259 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2262 if (!DTRACEACT_ISSTRING(act))
2265 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2266 limit = i + act->dta_rec.dtrd_size;
2267 ASSERT(limit <= size);
2269 for (nul = 0; i < limit; i++) {
2275 if (data[i] != '\0')
2282 for (i = size; i < fsize; i++)
2285 key->dtak_hashval = hashval;
2286 key->dtak_size = size;
2287 key->dtak_action = action;
2288 key->dtak_next = agb->dtagb_hash[ndx];
2289 agb->dtagb_hash[ndx] = key;
2292 * Finally, apply the aggregator.
2294 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2295 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2299 * Given consumer state, this routine finds a speculation in the INACTIVE
2300 * state and transitions it into the ACTIVE state. If there is no speculation
2301 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2302 * incremented -- it is up to the caller to take appropriate action.
2305 dtrace_speculation(dtrace_state_t *state)
2308 dtrace_speculation_state_t current;
2309 uint32_t *stat = &state->dts_speculations_unavail, count;
2311 while (i < state->dts_nspeculations) {
2312 dtrace_speculation_t *spec = &state->dts_speculations[i];
2314 current = spec->dtsp_state;
2316 if (current != DTRACESPEC_INACTIVE) {
2317 if (current == DTRACESPEC_COMMITTINGMANY ||
2318 current == DTRACESPEC_COMMITTING ||
2319 current == DTRACESPEC_DISCARDING)
2320 stat = &state->dts_speculations_busy;
2325 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2326 current, DTRACESPEC_ACTIVE) == current)
2331 * We couldn't find a speculation. If we found as much as a single
2332 * busy speculation buffer, we'll attribute this failure as "busy"
2333 * instead of "unavail".
2337 } while (dtrace_cas32(stat, count, count + 1) != count);
2343 * This routine commits an active speculation. If the specified speculation
2344 * is not in a valid state to perform a commit(), this routine will silently do
2345 * nothing. The state of the specified speculation is transitioned according
2346 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2349 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2350 dtrace_specid_t which)
2352 dtrace_speculation_t *spec;
2353 dtrace_buffer_t *src, *dest;
2354 uintptr_t daddr, saddr, dlimit, slimit;
2355 dtrace_speculation_state_t current, new = 0;
2362 if (which > state->dts_nspeculations) {
2363 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2367 spec = &state->dts_speculations[which - 1];
2368 src = &spec->dtsp_buffer[cpu];
2369 dest = &state->dts_buffer[cpu];
2372 current = spec->dtsp_state;
2374 if (current == DTRACESPEC_COMMITTINGMANY)
2378 case DTRACESPEC_INACTIVE:
2379 case DTRACESPEC_DISCARDING:
2382 case DTRACESPEC_COMMITTING:
2384 * This is only possible if we are (a) commit()'ing
2385 * without having done a prior speculate() on this CPU
2386 * and (b) racing with another commit() on a different
2387 * CPU. There's nothing to do -- we just assert that
2390 ASSERT(src->dtb_offset == 0);
2393 case DTRACESPEC_ACTIVE:
2394 new = DTRACESPEC_COMMITTING;
2397 case DTRACESPEC_ACTIVEONE:
2399 * This speculation is active on one CPU. If our
2400 * buffer offset is non-zero, we know that the one CPU
2401 * must be us. Otherwise, we are committing on a
2402 * different CPU from the speculate(), and we must
2403 * rely on being asynchronously cleaned.
2405 if (src->dtb_offset != 0) {
2406 new = DTRACESPEC_COMMITTING;
2411 case DTRACESPEC_ACTIVEMANY:
2412 new = DTRACESPEC_COMMITTINGMANY;
2418 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2419 current, new) != current);
2422 * We have set the state to indicate that we are committing this
2423 * speculation. Now reserve the necessary space in the destination
2426 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2427 sizeof (uint64_t), state, NULL)) < 0) {
2428 dtrace_buffer_drop(dest);
2433 * We have sufficient space to copy the speculative buffer into the
2434 * primary buffer. First, modify the speculative buffer, filling
2435 * in the timestamp of all entries with the current time. The data
2436 * must have the commit() time rather than the time it was traced,
2437 * so that all entries in the primary buffer are in timestamp order.
2439 timestamp = dtrace_gethrtime();
2440 saddr = (uintptr_t)src->dtb_tomax;
2441 slimit = saddr + src->dtb_offset;
2442 while (saddr < slimit) {
2444 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2446 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2447 saddr += sizeof (dtrace_epid_t);
2450 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2451 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2453 ASSERT3U(saddr + size, <=, slimit);
2454 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2455 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2457 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2463 * Copy the buffer across. (Note that this is a
2464 * highly subobtimal bcopy(); in the unlikely event that this becomes
2465 * a serious performance issue, a high-performance DTrace-specific
2466 * bcopy() should obviously be invented.)
2468 daddr = (uintptr_t)dest->dtb_tomax + offs;
2469 dlimit = daddr + src->dtb_offset;
2470 saddr = (uintptr_t)src->dtb_tomax;
2473 * First, the aligned portion.
2475 while (dlimit - daddr >= sizeof (uint64_t)) {
2476 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2478 daddr += sizeof (uint64_t);
2479 saddr += sizeof (uint64_t);
2483 * Now any left-over bit...
2485 while (dlimit - daddr)
2486 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2489 * Finally, commit the reserved space in the destination buffer.
2491 dest->dtb_offset = offs + src->dtb_offset;
2495 * If we're lucky enough to be the only active CPU on this speculation
2496 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2498 if (current == DTRACESPEC_ACTIVE ||
2499 (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2500 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2501 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2503 ASSERT(rval == DTRACESPEC_COMMITTING);
2506 src->dtb_offset = 0;
2507 src->dtb_xamot_drops += src->dtb_drops;
2512 * This routine discards an active speculation. If the specified speculation
2513 * is not in a valid state to perform a discard(), this routine will silently
2514 * do nothing. The state of the specified speculation is transitioned
2515 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
2518 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
2519 dtrace_specid_t which)
2521 dtrace_speculation_t *spec;
2522 dtrace_speculation_state_t current, new = 0;
2523 dtrace_buffer_t *buf;
2528 if (which > state->dts_nspeculations) {
2529 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2533 spec = &state->dts_speculations[which - 1];
2534 buf = &spec->dtsp_buffer[cpu];
2537 current = spec->dtsp_state;
2540 case DTRACESPEC_INACTIVE:
2541 case DTRACESPEC_COMMITTINGMANY:
2542 case DTRACESPEC_COMMITTING:
2543 case DTRACESPEC_DISCARDING:
2546 case DTRACESPEC_ACTIVE:
2547 case DTRACESPEC_ACTIVEMANY:
2548 new = DTRACESPEC_DISCARDING;
2551 case DTRACESPEC_ACTIVEONE:
2552 if (buf->dtb_offset != 0) {
2553 new = DTRACESPEC_INACTIVE;
2555 new = DTRACESPEC_DISCARDING;
2562 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2563 current, new) != current);
2565 buf->dtb_offset = 0;
2570 * Note: not called from probe context. This function is called
2571 * asynchronously from cross call context to clean any speculations that are
2572 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
2573 * transitioned back to the INACTIVE state until all CPUs have cleaned the
2577 dtrace_speculation_clean_here(dtrace_state_t *state)
2579 dtrace_icookie_t cookie;
2580 processorid_t cpu = curcpu;
2581 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
2584 cookie = dtrace_interrupt_disable();
2586 if (dest->dtb_tomax == NULL) {
2587 dtrace_interrupt_enable(cookie);
2591 for (i = 0; i < state->dts_nspeculations; i++) {
2592 dtrace_speculation_t *spec = &state->dts_speculations[i];
2593 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
2595 if (src->dtb_tomax == NULL)
2598 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
2599 src->dtb_offset = 0;
2603 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2606 if (src->dtb_offset == 0)
2609 dtrace_speculation_commit(state, cpu, i + 1);
2612 dtrace_interrupt_enable(cookie);
2616 * Note: not called from probe context. This function is called
2617 * asynchronously (and at a regular interval) to clean any speculations that
2618 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
2619 * is work to be done, it cross calls all CPUs to perform that work;
2620 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
2621 * INACTIVE state until they have been cleaned by all CPUs.
2624 dtrace_speculation_clean(dtrace_state_t *state)
2629 for (i = 0; i < state->dts_nspeculations; i++) {
2630 dtrace_speculation_t *spec = &state->dts_speculations[i];
2632 ASSERT(!spec->dtsp_cleaning);
2634 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
2635 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
2639 spec->dtsp_cleaning = 1;
2645 dtrace_xcall(DTRACE_CPUALL,
2646 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
2649 * We now know that all CPUs have committed or discarded their
2650 * speculation buffers, as appropriate. We can now set the state
2653 for (i = 0; i < state->dts_nspeculations; i++) {
2654 dtrace_speculation_t *spec = &state->dts_speculations[i];
2655 dtrace_speculation_state_t current, new;
2657 if (!spec->dtsp_cleaning)
2660 current = spec->dtsp_state;
2661 ASSERT(current == DTRACESPEC_DISCARDING ||
2662 current == DTRACESPEC_COMMITTINGMANY);
2664 new = DTRACESPEC_INACTIVE;
2666 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
2667 ASSERT(rv == current);
2668 spec->dtsp_cleaning = 0;
2673 * Called as part of a speculate() to get the speculative buffer associated
2674 * with a given speculation. Returns NULL if the specified speculation is not
2675 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
2676 * the active CPU is not the specified CPU -- the speculation will be
2677 * atomically transitioned into the ACTIVEMANY state.
2679 static dtrace_buffer_t *
2680 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
2681 dtrace_specid_t which)
2683 dtrace_speculation_t *spec;
2684 dtrace_speculation_state_t current, new = 0;
2685 dtrace_buffer_t *buf;
2690 if (which > state->dts_nspeculations) {
2691 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2695 spec = &state->dts_speculations[which - 1];
2696 buf = &spec->dtsp_buffer[cpuid];
2699 current = spec->dtsp_state;
2702 case DTRACESPEC_INACTIVE:
2703 case DTRACESPEC_COMMITTINGMANY:
2704 case DTRACESPEC_DISCARDING:
2707 case DTRACESPEC_COMMITTING:
2708 ASSERT(buf->dtb_offset == 0);
2711 case DTRACESPEC_ACTIVEONE:
2713 * This speculation is currently active on one CPU.
2714 * Check the offset in the buffer; if it's non-zero,
2715 * that CPU must be us (and we leave the state alone).
2716 * If it's zero, assume that we're starting on a new
2717 * CPU -- and change the state to indicate that the
2718 * speculation is active on more than one CPU.
2720 if (buf->dtb_offset != 0)
2723 new = DTRACESPEC_ACTIVEMANY;
2726 case DTRACESPEC_ACTIVEMANY:
2729 case DTRACESPEC_ACTIVE:
2730 new = DTRACESPEC_ACTIVEONE;
2736 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2737 current, new) != current);
2739 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
2744 * Return a string. In the event that the user lacks the privilege to access
2745 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2746 * don't fail access checking.
2748 * dtrace_dif_variable() uses this routine as a helper for various
2749 * builtin values such as 'execname' and 'probefunc.'
2752 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
2753 dtrace_mstate_t *mstate)
2755 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
2760 * The easy case: this probe is allowed to read all of memory, so
2761 * we can just return this as a vanilla pointer.
2763 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
2767 * This is the tougher case: we copy the string in question from
2768 * kernel memory into scratch memory and return it that way: this
2769 * ensures that we won't trip up when access checking tests the
2770 * BYREF return value.
2772 strsz = dtrace_strlen((char *)addr, size) + 1;
2774 if (mstate->dtms_scratch_ptr + strsz >
2775 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2776 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2780 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2782 ret = mstate->dtms_scratch_ptr;
2783 mstate->dtms_scratch_ptr += strsz;
2788 * Return a string from a memoy address which is known to have one or
2789 * more concatenated, individually zero terminated, sub-strings.
2790 * In the event that the user lacks the privilege to access
2791 * arbitrary kernel memory, we copy the string out to scratch memory so that we
2792 * don't fail access checking.
2794 * dtrace_dif_variable() uses this routine as a helper for various
2795 * builtin values such as 'execargs'.
2798 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
2799 dtrace_mstate_t *mstate)
2805 if (mstate->dtms_scratch_ptr + strsz >
2806 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
2807 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
2811 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
2814 /* Replace sub-string termination characters with a space. */
2815 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
2820 ret = mstate->dtms_scratch_ptr;
2821 mstate->dtms_scratch_ptr += strsz;
2826 * This function implements the DIF emulator's variable lookups. The emulator
2827 * passes a reserved variable identifier and optional built-in array index.
2830 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
2834 * If we're accessing one of the uncached arguments, we'll turn this
2835 * into a reference in the args array.
2837 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
2838 ndx = v - DIF_VAR_ARG0;
2844 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
2845 if (ndx >= sizeof (mstate->dtms_arg) /
2846 sizeof (mstate->dtms_arg[0])) {
2847 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2848 dtrace_provider_t *pv;
2851 pv = mstate->dtms_probe->dtpr_provider;
2852 if (pv->dtpv_pops.dtps_getargval != NULL)
2853 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
2854 mstate->dtms_probe->dtpr_id,
2855 mstate->dtms_probe->dtpr_arg, ndx, aframes);
2857 val = dtrace_getarg(ndx, aframes);
2860 * This is regrettably required to keep the compiler
2861 * from tail-optimizing the call to dtrace_getarg().
2862 * The condition always evaluates to true, but the
2863 * compiler has no way of figuring that out a priori.
2864 * (None of this would be necessary if the compiler
2865 * could be relied upon to _always_ tail-optimize
2866 * the call to dtrace_getarg() -- but it can't.)
2868 if (mstate->dtms_probe != NULL)
2874 return (mstate->dtms_arg[ndx]);
2877 case DIF_VAR_UREGS: {
2880 if (!dtrace_priv_proc(state))
2883 if ((lwp = curthread->t_lwp) == NULL) {
2884 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2885 cpu_core[curcpu].cpuc_dtrace_illval = NULL;
2889 return (dtrace_getreg(lwp->lwp_regs, ndx));
2893 case DIF_VAR_UREGS: {
2894 struct trapframe *tframe;
2896 if (!dtrace_priv_proc(state))
2899 if ((tframe = curthread->td_frame) == NULL) {
2900 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
2901 cpu_core[curcpu].cpuc_dtrace_illval = 0;
2905 return (dtrace_getreg(tframe, ndx));
2909 case DIF_VAR_CURTHREAD:
2910 if (!dtrace_priv_kernel(state))
2912 return ((uint64_t)(uintptr_t)curthread);
2914 case DIF_VAR_TIMESTAMP:
2915 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
2916 mstate->dtms_timestamp = dtrace_gethrtime();
2917 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
2919 return (mstate->dtms_timestamp);
2921 case DIF_VAR_VTIMESTAMP:
2922 ASSERT(dtrace_vtime_references != 0);
2923 return (curthread->t_dtrace_vtime);
2925 case DIF_VAR_WALLTIMESTAMP:
2926 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
2927 mstate->dtms_walltimestamp = dtrace_gethrestime();
2928 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
2930 return (mstate->dtms_walltimestamp);
2934 if (!dtrace_priv_kernel(state))
2936 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
2937 mstate->dtms_ipl = dtrace_getipl();
2938 mstate->dtms_present |= DTRACE_MSTATE_IPL;
2940 return (mstate->dtms_ipl);
2944 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
2945 return (mstate->dtms_epid);
2948 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
2949 return (mstate->dtms_probe->dtpr_id);
2951 case DIF_VAR_STACKDEPTH:
2952 if (!dtrace_priv_kernel(state))
2954 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
2955 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2957 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
2958 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
2960 return (mstate->dtms_stackdepth);
2962 case DIF_VAR_USTACKDEPTH:
2963 if (!dtrace_priv_proc(state))
2965 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
2967 * See comment in DIF_VAR_PID.
2969 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
2971 mstate->dtms_ustackdepth = 0;
2973 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
2974 mstate->dtms_ustackdepth =
2975 dtrace_getustackdepth();
2976 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
2978 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
2980 return (mstate->dtms_ustackdepth);
2982 case DIF_VAR_CALLER:
2983 if (!dtrace_priv_kernel(state))
2985 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
2986 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
2988 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
2990 * If this is an unanchored probe, we are
2991 * required to go through the slow path:
2992 * dtrace_caller() only guarantees correct
2993 * results for anchored probes.
2995 pc_t caller[2] = {0, 0};
2997 dtrace_getpcstack(caller, 2, aframes,
2998 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
2999 mstate->dtms_caller = caller[1];
3000 } else if ((mstate->dtms_caller =
3001 dtrace_caller(aframes)) == -1) {
3003 * We have failed to do this the quick way;
3004 * we must resort to the slower approach of
3005 * calling dtrace_getpcstack().
3009 dtrace_getpcstack(&caller, 1, aframes, NULL);
3010 mstate->dtms_caller = caller;
3013 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3015 return (mstate->dtms_caller);
3017 case DIF_VAR_UCALLER:
3018 if (!dtrace_priv_proc(state))
3021 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3025 * dtrace_getupcstack() fills in the first uint64_t
3026 * with the current PID. The second uint64_t will
3027 * be the program counter at user-level. The third
3028 * uint64_t will contain the caller, which is what
3032 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3033 dtrace_getupcstack(ustack, 3);
3034 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3035 mstate->dtms_ucaller = ustack[2];
3036 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3039 return (mstate->dtms_ucaller);
3041 case DIF_VAR_PROBEPROV:
3042 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3043 return (dtrace_dif_varstr(
3044 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3047 case DIF_VAR_PROBEMOD:
3048 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3049 return (dtrace_dif_varstr(
3050 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3053 case DIF_VAR_PROBEFUNC:
3054 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3055 return (dtrace_dif_varstr(
3056 (uintptr_t)mstate->dtms_probe->dtpr_func,
3059 case DIF_VAR_PROBENAME:
3060 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3061 return (dtrace_dif_varstr(
3062 (uintptr_t)mstate->dtms_probe->dtpr_name,
3066 if (!dtrace_priv_proc(state))
3071 * Note that we are assuming that an unanchored probe is
3072 * always due to a high-level interrupt. (And we're assuming
3073 * that there is only a single high level interrupt.)
3075 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3076 return (pid0.pid_id);
3079 * It is always safe to dereference one's own t_procp pointer:
3080 * it always points to a valid, allocated proc structure.
3081 * Further, it is always safe to dereference the p_pidp member
3082 * of one's own proc structure. (These are truisms becuase
3083 * threads and processes don't clean up their own state --
3084 * they leave that task to whomever reaps them.)
3086 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3088 return ((uint64_t)curproc->p_pid);
3092 if (!dtrace_priv_proc(state))
3097 * See comment in DIF_VAR_PID.
3099 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3100 return (pid0.pid_id);
3103 * It is always safe to dereference one's own t_procp pointer:
3104 * it always points to a valid, allocated proc structure.
3105 * (This is true because threads don't clean up their own
3106 * state -- they leave that task to whomever reaps them.)
3108 return ((uint64_t)curthread->t_procp->p_ppid);
3110 return ((uint64_t)curproc->p_pptr->p_pid);
3116 * See comment in DIF_VAR_PID.
3118 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3122 return ((uint64_t)curthread->t_tid);
3124 case DIF_VAR_EXECARGS: {
3125 struct pargs *p_args = curthread->td_proc->p_args;
3130 return (dtrace_dif_varstrz(
3131 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3134 case DIF_VAR_EXECNAME:
3136 if (!dtrace_priv_proc(state))
3140 * See comment in DIF_VAR_PID.
3142 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3143 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3146 * It is always safe to dereference one's own t_procp pointer:
3147 * it always points to a valid, allocated proc structure.
3148 * (This is true because threads don't clean up their own
3149 * state -- they leave that task to whomever reaps them.)
3151 return (dtrace_dif_varstr(
3152 (uintptr_t)curthread->t_procp->p_user.u_comm,
3155 return (dtrace_dif_varstr(
3156 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3159 case DIF_VAR_ZONENAME:
3161 if (!dtrace_priv_proc(state))
3165 * See comment in DIF_VAR_PID.
3167 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3168 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3171 * It is always safe to dereference one's own t_procp pointer:
3172 * it always points to a valid, allocated proc structure.
3173 * (This is true because threads don't clean up their own
3174 * state -- they leave that task to whomever reaps them.)
3176 return (dtrace_dif_varstr(
3177 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3184 if (!dtrace_priv_proc(state))
3189 * See comment in DIF_VAR_PID.
3191 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3192 return ((uint64_t)p0.p_cred->cr_uid);
3196 * It is always safe to dereference one's own t_procp pointer:
3197 * it always points to a valid, allocated proc structure.
3198 * (This is true because threads don't clean up their own
3199 * state -- they leave that task to whomever reaps them.)
3201 * Additionally, it is safe to dereference one's own process
3202 * credential, since this is never NULL after process birth.
3204 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3207 if (!dtrace_priv_proc(state))
3212 * See comment in DIF_VAR_PID.
3214 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3215 return ((uint64_t)p0.p_cred->cr_gid);
3219 * It is always safe to dereference one's own t_procp pointer:
3220 * it always points to a valid, allocated proc structure.
3221 * (This is true because threads don't clean up their own
3222 * state -- they leave that task to whomever reaps them.)
3224 * Additionally, it is safe to dereference one's own process
3225 * credential, since this is never NULL after process birth.
3227 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3229 case DIF_VAR_ERRNO: {
3232 if (!dtrace_priv_proc(state))
3236 * See comment in DIF_VAR_PID.
3238 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3242 * It is always safe to dereference one's own t_lwp pointer in
3243 * the event that this pointer is non-NULL. (This is true
3244 * because threads and lwps don't clean up their own state --
3245 * they leave that task to whomever reaps them.)
3247 if ((lwp = curthread->t_lwp) == NULL)
3250 return ((uint64_t)lwp->lwp_errno);
3252 return (curthread->td_errno);
3261 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3267 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
3268 * Notice that we don't bother validating the proper number of arguments or
3269 * their types in the tuple stack. This isn't needed because all argument
3270 * interpretation is safe because of our load safety -- the worst that can
3271 * happen is that a bogus program can obtain bogus results.
3274 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
3275 dtrace_key_t *tupregs, int nargs,
3276 dtrace_mstate_t *mstate, dtrace_state_t *state)
3278 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
3279 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
3280 dtrace_vstate_t *vstate = &state->dts_vstate;
3293 struct thread *lowner;
3295 struct lock_object *li;
3302 regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
3306 case DIF_SUBR_MUTEX_OWNED:
3307 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3313 m.mx = dtrace_load64(tupregs[0].dttk_value);
3314 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
3315 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
3317 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
3320 case DIF_SUBR_MUTEX_OWNER:
3321 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3327 m.mx = dtrace_load64(tupregs[0].dttk_value);
3328 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
3329 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
3330 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
3335 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3336 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3342 m.mx = dtrace_load64(tupregs[0].dttk_value);
3343 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
3346 case DIF_SUBR_MUTEX_TYPE_SPIN:
3347 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
3353 m.mx = dtrace_load64(tupregs[0].dttk_value);
3354 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
3357 case DIF_SUBR_RW_READ_HELD: {
3360 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3366 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3367 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
3371 case DIF_SUBR_RW_WRITE_HELD:
3372 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3378 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3379 regs[rd] = _RW_WRITE_HELD(&r.ri);
3382 case DIF_SUBR_RW_ISWRITER:
3383 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
3389 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
3390 regs[rd] = _RW_ISWRITER(&r.ri);
3394 case DIF_SUBR_MUTEX_OWNED:
3395 if (!dtrace_canload(tupregs[0].dttk_value,
3396 sizeof (struct lock_object), mstate, vstate)) {
3400 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3401 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3404 case DIF_SUBR_MUTEX_OWNER:
3405 if (!dtrace_canload(tupregs[0].dttk_value,
3406 sizeof (struct lock_object), mstate, vstate)) {
3410 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3411 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3412 regs[rd] = (uintptr_t)lowner;
3415 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
3416 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3421 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3422 /* XXX - should be only LC_SLEEPABLE? */
3423 regs[rd] = (LOCK_CLASS(l.li)->lc_flags &
3424 (LC_SLEEPLOCK | LC_SLEEPABLE)) != 0;
3427 case DIF_SUBR_MUTEX_TYPE_SPIN:
3428 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
3433 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3434 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
3437 case DIF_SUBR_RW_READ_HELD:
3438 case DIF_SUBR_SX_SHARED_HELD:
3439 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3444 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
3445 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3449 case DIF_SUBR_RW_WRITE_HELD:
3450 case DIF_SUBR_SX_EXCLUSIVE_HELD:
3451 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3456 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3457 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
3458 regs[rd] = (lowner == curthread);
3461 case DIF_SUBR_RW_ISWRITER:
3462 case DIF_SUBR_SX_ISEXCLUSIVE:
3463 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
3468 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
3469 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
3472 #endif /* ! defined(sun) */
3474 case DIF_SUBR_BCOPY: {
3476 * We need to be sure that the destination is in the scratch
3477 * region -- no other region is allowed.
3479 uintptr_t src = tupregs[0].dttk_value;
3480 uintptr_t dest = tupregs[1].dttk_value;
3481 size_t size = tupregs[2].dttk_value;
3483 if (!dtrace_inscratch(dest, size, mstate)) {
3484 *flags |= CPU_DTRACE_BADADDR;
3489 if (!dtrace_canload(src, size, mstate, vstate)) {
3494 dtrace_bcopy((void *)src, (void *)dest, size);
3498 case DIF_SUBR_ALLOCA:
3499 case DIF_SUBR_COPYIN: {
3500 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
3502 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
3503 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
3506 * This action doesn't require any credential checks since
3507 * probes will not activate in user contexts to which the
3508 * enabling user does not have permissions.
3512 * Rounding up the user allocation size could have overflowed
3513 * a large, bogus allocation (like -1ULL) to 0.
3515 if (scratch_size < size ||
3516 !DTRACE_INSCRATCH(mstate, scratch_size)) {
3517 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3522 if (subr == DIF_SUBR_COPYIN) {
3523 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3524 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3525 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3528 mstate->dtms_scratch_ptr += scratch_size;
3533 case DIF_SUBR_COPYINTO: {
3534 uint64_t size = tupregs[1].dttk_value;
3535 uintptr_t dest = tupregs[2].dttk_value;
3538 * This action doesn't require any credential checks since
3539 * probes will not activate in user contexts to which the
3540 * enabling user does not have permissions.
3542 if (!dtrace_inscratch(dest, size, mstate)) {
3543 *flags |= CPU_DTRACE_BADADDR;
3548 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3549 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
3550 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3554 case DIF_SUBR_COPYINSTR: {
3555 uintptr_t dest = mstate->dtms_scratch_ptr;
3556 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3558 if (nargs > 1 && tupregs[1].dttk_value < size)
3559 size = tupregs[1].dttk_value + 1;
3562 * This action doesn't require any credential checks since
3563 * probes will not activate in user contexts to which the
3564 * enabling user does not have permissions.
3566 if (!DTRACE_INSCRATCH(mstate, size)) {
3567 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3572 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3573 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
3574 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3576 ((char *)dest)[size - 1] = '\0';
3577 mstate->dtms_scratch_ptr += size;
3583 case DIF_SUBR_MSGSIZE:
3584 case DIF_SUBR_MSGDSIZE: {
3585 uintptr_t baddr = tupregs[0].dttk_value, daddr;
3586 uintptr_t wptr, rptr;
3590 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
3592 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
3598 wptr = dtrace_loadptr(baddr +
3599 offsetof(mblk_t, b_wptr));
3601 rptr = dtrace_loadptr(baddr +
3602 offsetof(mblk_t, b_rptr));
3605 *flags |= CPU_DTRACE_BADADDR;
3606 *illval = tupregs[0].dttk_value;
3610 daddr = dtrace_loadptr(baddr +
3611 offsetof(mblk_t, b_datap));
3613 baddr = dtrace_loadptr(baddr +
3614 offsetof(mblk_t, b_cont));
3617 * We want to prevent against denial-of-service here,
3618 * so we're only going to search the list for
3619 * dtrace_msgdsize_max mblks.
3621 if (cont++ > dtrace_msgdsize_max) {
3622 *flags |= CPU_DTRACE_ILLOP;
3626 if (subr == DIF_SUBR_MSGDSIZE) {
3627 if (dtrace_load8(daddr +
3628 offsetof(dblk_t, db_type)) != M_DATA)
3632 count += wptr - rptr;
3635 if (!(*flags & CPU_DTRACE_FAULT))
3642 case DIF_SUBR_PROGENYOF: {
3643 pid_t pid = tupregs[0].dttk_value;
3647 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3649 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
3651 if (p->p_pidp->pid_id == pid) {
3653 if (p->p_pid == pid) {
3660 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3666 case DIF_SUBR_SPECULATION:
3667 regs[rd] = dtrace_speculation(state);
3670 case DIF_SUBR_COPYOUT: {
3671 uintptr_t kaddr = tupregs[0].dttk_value;
3672 uintptr_t uaddr = tupregs[1].dttk_value;
3673 uint64_t size = tupregs[2].dttk_value;
3675 if (!dtrace_destructive_disallow &&
3676 dtrace_priv_proc_control(state) &&
3677 !dtrace_istoxic(kaddr, size)) {
3678 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3679 dtrace_copyout(kaddr, uaddr, size, flags);
3680 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3685 case DIF_SUBR_COPYOUTSTR: {
3686 uintptr_t kaddr = tupregs[0].dttk_value;
3687 uintptr_t uaddr = tupregs[1].dttk_value;
3688 uint64_t size = tupregs[2].dttk_value;
3690 if (!dtrace_destructive_disallow &&
3691 dtrace_priv_proc_control(state) &&
3692 !dtrace_istoxic(kaddr, size)) {
3693 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3694 dtrace_copyoutstr(kaddr, uaddr, size, flags);
3695 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3700 case DIF_SUBR_STRLEN: {
3702 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
3703 sz = dtrace_strlen((char *)addr,
3704 state->dts_options[DTRACEOPT_STRSIZE]);
3706 if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
3716 case DIF_SUBR_STRCHR:
3717 case DIF_SUBR_STRRCHR: {
3719 * We're going to iterate over the string looking for the
3720 * specified character. We will iterate until we have reached
3721 * the string length or we have found the character. If this
3722 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
3723 * of the specified character instead of the first.
3725 uintptr_t saddr = tupregs[0].dttk_value;
3726 uintptr_t addr = tupregs[0].dttk_value;
3727 uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
3728 char c, target = (char)tupregs[1].dttk_value;
3730 for (regs[rd] = 0; addr < limit; addr++) {
3731 if ((c = dtrace_load8(addr)) == target) {
3734 if (subr == DIF_SUBR_STRCHR)
3742 if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
3750 case DIF_SUBR_STRSTR:
3751 case DIF_SUBR_INDEX:
3752 case DIF_SUBR_RINDEX: {
3754 * We're going to iterate over the string looking for the
3755 * specified string. We will iterate until we have reached
3756 * the string length or we have found the string. (Yes, this
3757 * is done in the most naive way possible -- but considering
3758 * that the string we're searching for is likely to be
3759 * relatively short, the complexity of Rabin-Karp or similar
3760 * hardly seems merited.)
3762 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
3763 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
3764 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3765 size_t len = dtrace_strlen(addr, size);
3766 size_t sublen = dtrace_strlen(substr, size);
3767 char *limit = addr + len, *orig = addr;
3768 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
3771 regs[rd] = notfound;
3773 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
3778 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
3785 * strstr() and index()/rindex() have similar semantics if
3786 * both strings are the empty string: strstr() returns a
3787 * pointer to the (empty) string, and index() and rindex()
3788 * both return index 0 (regardless of any position argument).
3790 if (sublen == 0 && len == 0) {
3791 if (subr == DIF_SUBR_STRSTR)
3792 regs[rd] = (uintptr_t)addr;
3798 if (subr != DIF_SUBR_STRSTR) {
3799 if (subr == DIF_SUBR_RINDEX) {
3806 * Both index() and rindex() take an optional position
3807 * argument that denotes the starting position.
3810 int64_t pos = (int64_t)tupregs[2].dttk_value;
3813 * If the position argument to index() is
3814 * negative, Perl implicitly clamps it at
3815 * zero. This semantic is a little surprising
3816 * given the special meaning of negative
3817 * positions to similar Perl functions like
3818 * substr(), but it appears to reflect a
3819 * notion that index() can start from a
3820 * negative index and increment its way up to
3821 * the string. Given this notion, Perl's
3822 * rindex() is at least self-consistent in
3823 * that it implicitly clamps positions greater
3824 * than the string length to be the string
3825 * length. Where Perl completely loses
3826 * coherence, however, is when the specified
3827 * substring is the empty string (""). In
3828 * this case, even if the position is
3829 * negative, rindex() returns 0 -- and even if
3830 * the position is greater than the length,
3831 * index() returns the string length. These
3832 * semantics violate the notion that index()
3833 * should never return a value less than the
3834 * specified position and that rindex() should
3835 * never return a value greater than the
3836 * specified position. (One assumes that
3837 * these semantics are artifacts of Perl's
3838 * implementation and not the results of
3839 * deliberate design -- it beggars belief that
3840 * even Larry Wall could desire such oddness.)
3841 * While in the abstract one would wish for
3842 * consistent position semantics across
3843 * substr(), index() and rindex() -- or at the
3844 * very least self-consistent position
3845 * semantics for index() and rindex() -- we
3846 * instead opt to keep with the extant Perl
3847 * semantics, in all their broken glory. (Do
3848 * we have more desire to maintain Perl's
3849 * semantics than Perl does? Probably.)
3851 if (subr == DIF_SUBR_RINDEX) {
3875 for (regs[rd] = notfound; addr != limit; addr += inc) {
3876 if (dtrace_strncmp(addr, substr, sublen) == 0) {
3877 if (subr != DIF_SUBR_STRSTR) {
3879 * As D index() and rindex() are
3880 * modeled on Perl (and not on awk),
3881 * we return a zero-based (and not a
3882 * one-based) index. (For you Perl
3883 * weenies: no, we're not going to add
3884 * $[ -- and shouldn't you be at a con
3887 regs[rd] = (uintptr_t)(addr - orig);
3891 ASSERT(subr == DIF_SUBR_STRSTR);
3892 regs[rd] = (uintptr_t)addr;
3900 case DIF_SUBR_STRTOK: {
3901 uintptr_t addr = tupregs[0].dttk_value;
3902 uintptr_t tokaddr = tupregs[1].dttk_value;
3903 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3904 uintptr_t limit, toklimit = tokaddr + size;
3905 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
3906 char *dest = (char *)mstate->dtms_scratch_ptr;
3910 * Check both the token buffer and (later) the input buffer,
3911 * since both could be non-scratch addresses.
3913 if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
3918 if (!DTRACE_INSCRATCH(mstate, size)) {
3919 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3926 * If the address specified is NULL, we use our saved
3927 * strtok pointer from the mstate. Note that this
3928 * means that the saved strtok pointer is _only_
3929 * valid within multiple enablings of the same probe --
3930 * it behaves like an implicit clause-local variable.
3932 addr = mstate->dtms_strtok;
3935 * If the user-specified address is non-NULL we must
3936 * access check it. This is the only time we have
3937 * a chance to do so, since this address may reside
3938 * in the string table of this clause-- future calls
3939 * (when we fetch addr from mstate->dtms_strtok)
3940 * would fail this access check.
3942 if (!dtrace_strcanload(addr, size, mstate, vstate)) {
3949 * First, zero the token map, and then process the token
3950 * string -- setting a bit in the map for every character
3951 * found in the token string.
3953 for (i = 0; i < sizeof (tokmap); i++)
3956 for (; tokaddr < toklimit; tokaddr++) {
3957 if ((c = dtrace_load8(tokaddr)) == '\0')
3960 ASSERT((c >> 3) < sizeof (tokmap));
3961 tokmap[c >> 3] |= (1 << (c & 0x7));
3964 for (limit = addr + size; addr < limit; addr++) {
3966 * We're looking for a character that is _not_ contained
3967 * in the token string.
3969 if ((c = dtrace_load8(addr)) == '\0')
3972 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
3978 * We reached the end of the string without finding
3979 * any character that was not in the token string.
3980 * We return NULL in this case, and we set the saved
3981 * address to NULL as well.
3984 mstate->dtms_strtok = 0;
3989 * From here on, we're copying into the destination string.
3991 for (i = 0; addr < limit && i < size - 1; addr++) {
3992 if ((c = dtrace_load8(addr)) == '\0')
3995 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4004 regs[rd] = (uintptr_t)dest;
4005 mstate->dtms_scratch_ptr += size;
4006 mstate->dtms_strtok = addr;
4010 case DIF_SUBR_SUBSTR: {
4011 uintptr_t s = tupregs[0].dttk_value;
4012 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4013 char *d = (char *)mstate->dtms_scratch_ptr;
4014 int64_t index = (int64_t)tupregs[1].dttk_value;
4015 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4016 size_t len = dtrace_strlen((char *)s, size);
4019 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4024 if (!DTRACE_INSCRATCH(mstate, size)) {
4025 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4031 remaining = (int64_t)size;
4036 if (index < 0 && index + remaining > 0) {
4042 if (index >= len || index < 0) {
4044 } else if (remaining < 0) {
4045 remaining += len - index;
4046 } else if (index + remaining > size) {
4047 remaining = size - index;
4050 for (i = 0; i < remaining; i++) {
4051 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
4057 mstate->dtms_scratch_ptr += size;
4058 regs[rd] = (uintptr_t)d;
4062 case DIF_SUBR_TOUPPER:
4063 case DIF_SUBR_TOLOWER: {
4064 uintptr_t s = tupregs[0].dttk_value;
4065 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4066 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4067 size_t len = dtrace_strlen((char *)s, size);
4068 char lower, upper, convert;
4071 if (subr == DIF_SUBR_TOUPPER) {
4081 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4086 if (!DTRACE_INSCRATCH(mstate, size)) {
4087 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4092 for (i = 0; i < size - 1; i++) {
4093 if ((c = dtrace_load8(s + i)) == '\0')
4096 if (c >= lower && c <= upper)
4097 c = convert + (c - lower);
4104 regs[rd] = (uintptr_t)dest;
4105 mstate->dtms_scratch_ptr += size;
4110 case DIF_SUBR_GETMAJOR:
4112 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
4114 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
4118 case DIF_SUBR_GETMINOR:
4120 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
4122 regs[rd] = tupregs[0].dttk_value & MAXMIN;
4126 case DIF_SUBR_DDI_PATHNAME: {
4128 * This one is a galactic mess. We are going to roughly
4129 * emulate ddi_pathname(), but it's made more complicated
4130 * by the fact that we (a) want to include the minor name and
4131 * (b) must proceed iteratively instead of recursively.
4133 uintptr_t dest = mstate->dtms_scratch_ptr;
4134 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4135 char *start = (char *)dest, *end = start + size - 1;
4136 uintptr_t daddr = tupregs[0].dttk_value;
4137 int64_t minor = (int64_t)tupregs[1].dttk_value;
4139 int i, len, depth = 0;
4142 * Due to all the pointer jumping we do and context we must
4143 * rely upon, we just mandate that the user must have kernel
4144 * read privileges to use this routine.
4146 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
4147 *flags |= CPU_DTRACE_KPRIV;
4152 if (!DTRACE_INSCRATCH(mstate, size)) {
4153 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4161 * We want to have a name for the minor. In order to do this,
4162 * we need to walk the minor list from the devinfo. We want
4163 * to be sure that we don't infinitely walk a circular list,
4164 * so we check for circularity by sending a scout pointer
4165 * ahead two elements for every element that we iterate over;
4166 * if the list is circular, these will ultimately point to the
4167 * same element. You may recognize this little trick as the
4168 * answer to a stupid interview question -- one that always
4169 * seems to be asked by those who had to have it laboriously
4170 * explained to them, and who can't even concisely describe
4171 * the conditions under which one would be forced to resort to
4172 * this technique. Needless to say, those conditions are
4173 * found here -- and probably only here. Is this the only use
4174 * of this infamous trick in shipping, production code? If it
4175 * isn't, it probably should be...
4178 uintptr_t maddr = dtrace_loadptr(daddr +
4179 offsetof(struct dev_info, devi_minor));
4181 uintptr_t next = offsetof(struct ddi_minor_data, next);
4182 uintptr_t name = offsetof(struct ddi_minor_data,
4183 d_minor) + offsetof(struct ddi_minor, name);
4184 uintptr_t dev = offsetof(struct ddi_minor_data,
4185 d_minor) + offsetof(struct ddi_minor, dev);
4189 scout = dtrace_loadptr(maddr + next);
4191 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4194 m = dtrace_load64(maddr + dev) & MAXMIN64;
4196 m = dtrace_load32(maddr + dev) & MAXMIN;
4199 maddr = dtrace_loadptr(maddr + next);
4204 scout = dtrace_loadptr(scout + next);
4209 scout = dtrace_loadptr(scout + next);
4214 if (scout == maddr) {
4215 *flags |= CPU_DTRACE_ILLOP;
4223 * We have the minor data. Now we need to
4224 * copy the minor's name into the end of the
4227 s = (char *)dtrace_loadptr(maddr + name);
4228 len = dtrace_strlen(s, size);
4230 if (*flags & CPU_DTRACE_FAULT)
4234 if ((end -= (len + 1)) < start)
4240 for (i = 1; i <= len; i++)
4241 end[i] = dtrace_load8((uintptr_t)s++);
4246 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
4247 ddi_node_state_t devi_state;
4249 devi_state = dtrace_load32(daddr +
4250 offsetof(struct dev_info, devi_node_state));
4252 if (*flags & CPU_DTRACE_FAULT)
4255 if (devi_state >= DS_INITIALIZED) {
4256 s = (char *)dtrace_loadptr(daddr +
4257 offsetof(struct dev_info, devi_addr));
4258 len = dtrace_strlen(s, size);
4260 if (*flags & CPU_DTRACE_FAULT)
4264 if ((end -= (len + 1)) < start)
4270 for (i = 1; i <= len; i++)
4271 end[i] = dtrace_load8((uintptr_t)s++);
4275 * Now for the node name...
4277 s = (char *)dtrace_loadptr(daddr +
4278 offsetof(struct dev_info, devi_node_name));
4280 daddr = dtrace_loadptr(daddr +
4281 offsetof(struct dev_info, devi_parent));
4284 * If our parent is NULL (that is, if we're the root
4285 * node), we're going to use the special path
4291 len = dtrace_strlen(s, size);
4292 if (*flags & CPU_DTRACE_FAULT)
4295 if ((end -= (len + 1)) < start)
4298 for (i = 1; i <= len; i++)
4299 end[i] = dtrace_load8((uintptr_t)s++);
4302 if (depth++ > dtrace_devdepth_max) {
4303 *flags |= CPU_DTRACE_ILLOP;
4309 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4312 regs[rd] = (uintptr_t)end;
4313 mstate->dtms_scratch_ptr += size;
4320 case DIF_SUBR_STRJOIN: {
4321 char *d = (char *)mstate->dtms_scratch_ptr;
4322 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4323 uintptr_t s1 = tupregs[0].dttk_value;
4324 uintptr_t s2 = tupregs[1].dttk_value;
4327 if (!dtrace_strcanload(s1, size, mstate, vstate) ||
4328 !dtrace_strcanload(s2, size, mstate, vstate)) {
4333 if (!DTRACE_INSCRATCH(mstate, size)) {
4334 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4341 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4346 if ((d[i++] = dtrace_load8(s1++)) == '\0') {
4354 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4359 if ((d[i++] = dtrace_load8(s2++)) == '\0')
4364 mstate->dtms_scratch_ptr += i;
4365 regs[rd] = (uintptr_t)d;
4371 case DIF_SUBR_LLTOSTR: {
4372 int64_t i = (int64_t)tupregs[0].dttk_value;
4373 uint64_t val, digit;
4374 uint64_t size = 65; /* enough room for 2^64 in binary */
4375 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
4379 if ((base = tupregs[1].dttk_value) <= 1 ||
4380 base > ('z' - 'a' + 1) + ('9' - '0' + 1)) {
4381 *flags |= CPU_DTRACE_ILLOP;
4386 val = (base == 10 && i < 0) ? i * -1 : i;
4388 if (!DTRACE_INSCRATCH(mstate, size)) {
4389 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4394 for (*end-- = '\0'; val; val /= base) {
4395 if ((digit = val % base) <= '9' - '0') {
4396 *end-- = '0' + digit;
4398 *end-- = 'a' + (digit - ('9' - '0') - 1);
4402 if (i == 0 && base == 16)
4408 if (i == 0 || base == 8 || base == 16)
4411 if (i < 0 && base == 10)
4414 regs[rd] = (uintptr_t)end + 1;
4415 mstate->dtms_scratch_ptr += size;
4419 case DIF_SUBR_HTONS:
4420 case DIF_SUBR_NTOHS:
4421 #if BYTE_ORDER == BIG_ENDIAN
4422 regs[rd] = (uint16_t)tupregs[0].dttk_value;
4424 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
4429 case DIF_SUBR_HTONL:
4430 case DIF_SUBR_NTOHL:
4431 #if BYTE_ORDER == BIG_ENDIAN
4432 regs[rd] = (uint32_t)tupregs[0].dttk_value;
4434 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
4439 case DIF_SUBR_HTONLL:
4440 case DIF_SUBR_NTOHLL:
4441 #if BYTE_ORDER == BIG_ENDIAN
4442 regs[rd] = (uint64_t)tupregs[0].dttk_value;
4444 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
4449 case DIF_SUBR_DIRNAME:
4450 case DIF_SUBR_BASENAME: {
4451 char *dest = (char *)mstate->dtms_scratch_ptr;
4452 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4453 uintptr_t src = tupregs[0].dttk_value;
4454 int i, j, len = dtrace_strlen((char *)src, size);
4455 int lastbase = -1, firstbase = -1, lastdir = -1;
4458 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
4463 if (!DTRACE_INSCRATCH(mstate, size)) {
4464 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4470 * The basename and dirname for a zero-length string is
4475 src = (uintptr_t)".";
4479 * Start from the back of the string, moving back toward the
4480 * front until we see a character that isn't a slash. That
4481 * character is the last character in the basename.
4483 for (i = len - 1; i >= 0; i--) {
4484 if (dtrace_load8(src + i) != '/')
4492 * Starting from the last character in the basename, move
4493 * towards the front until we find a slash. The character
4494 * that we processed immediately before that is the first
4495 * character in the basename.
4497 for (; i >= 0; i--) {
4498 if (dtrace_load8(src + i) == '/')
4506 * Now keep going until we find a non-slash character. That
4507 * character is the last character in the dirname.
4509 for (; i >= 0; i--) {
4510 if (dtrace_load8(src + i) != '/')
4517 ASSERT(!(lastbase == -1 && firstbase != -1));
4518 ASSERT(!(firstbase == -1 && lastdir != -1));
4520 if (lastbase == -1) {
4522 * We didn't find a non-slash character. We know that
4523 * the length is non-zero, so the whole string must be
4524 * slashes. In either the dirname or the basename
4525 * case, we return '/'.
4527 ASSERT(firstbase == -1);
4528 firstbase = lastbase = lastdir = 0;
4531 if (firstbase == -1) {
4533 * The entire string consists only of a basename
4534 * component. If we're looking for dirname, we need
4535 * to change our string to be just "."; if we're
4536 * looking for a basename, we'll just set the first
4537 * character of the basename to be 0.
4539 if (subr == DIF_SUBR_DIRNAME) {
4540 ASSERT(lastdir == -1);
4541 src = (uintptr_t)".";
4548 if (subr == DIF_SUBR_DIRNAME) {
4549 if (lastdir == -1) {
4551 * We know that we have a slash in the name --
4552 * or lastdir would be set to 0, above. And
4553 * because lastdir is -1, we know that this
4554 * slash must be the first character. (That
4555 * is, the full string must be of the form
4556 * "/basename".) In this case, the last
4557 * character of the directory name is 0.
4565 ASSERT(subr == DIF_SUBR_BASENAME);
4566 ASSERT(firstbase != -1 && lastbase != -1);
4571 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
4572 dest[j] = dtrace_load8(src + i);
4575 regs[rd] = (uintptr_t)dest;
4576 mstate->dtms_scratch_ptr += size;
4580 case DIF_SUBR_CLEANPATH: {
4581 char *dest = (char *)mstate->dtms_scratch_ptr, c;
4582 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4583 uintptr_t src = tupregs[0].dttk_value;
4586 if (!dtrace_strcanload(src, size, mstate, vstate)) {
4591 if (!DTRACE_INSCRATCH(mstate, size)) {
4592 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4598 * Move forward, loading each character.
4601 c = dtrace_load8(src + i++);
4603 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
4611 c = dtrace_load8(src + i++);
4615 * We have two slashes -- we can just advance
4616 * to the next character.
4623 * This is not "." and it's not ".." -- we can
4624 * just store the "/" and this character and
4632 c = dtrace_load8(src + i++);
4636 * This is a "/./" component. We're not going
4637 * to store anything in the destination buffer;
4638 * we're just going to go to the next component.
4645 * This is not ".." -- we can just store the
4646 * "/." and this character and continue
4655 c = dtrace_load8(src + i++);
4657 if (c != '/' && c != '\0') {
4659 * This is not ".." -- it's "..[mumble]".
4660 * We'll store the "/.." and this character
4661 * and continue processing.
4671 * This is "/../" or "/..\0". We need to back up
4672 * our destination pointer until we find a "/".
4675 while (j != 0 && dest[--j] != '/')
4680 } while (c != '\0');
4683 regs[rd] = (uintptr_t)dest;
4684 mstate->dtms_scratch_ptr += size;
4688 case DIF_SUBR_INET_NTOA:
4689 case DIF_SUBR_INET_NTOA6:
4690 case DIF_SUBR_INET_NTOP: {
4695 if (subr == DIF_SUBR_INET_NTOP) {
4696 af = (int)tupregs[0].dttk_value;
4699 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
4703 if (af == AF_INET) {
4708 * Safely load the IPv4 address.
4710 ip4 = dtrace_load32(tupregs[argi].dttk_value);
4713 * Check an IPv4 string will fit in scratch.
4715 size = INET_ADDRSTRLEN;
4716 if (!DTRACE_INSCRATCH(mstate, size)) {
4717 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4721 base = (char *)mstate->dtms_scratch_ptr;
4722 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4725 * Stringify as a dotted decimal quad.
4728 ptr8 = (uint8_t *)&ip4;
4729 for (i = 3; i >= 0; i--) {
4735 for (; val; val /= 10) {
4736 *end-- = '0' + (val % 10);
4743 ASSERT(end + 1 >= base);
4745 } else if (af == AF_INET6) {
4746 struct in6_addr ip6;
4747 int firstzero, tryzero, numzero, v6end;
4749 const char digits[] = "0123456789abcdef";
4752 * Stringify using RFC 1884 convention 2 - 16 bit
4753 * hexadecimal values with a zero-run compression.
4754 * Lower case hexadecimal digits are used.
4755 * eg, fe80::214:4fff:fe0b:76c8.
4756 * The IPv4 embedded form is returned for inet_ntop,
4757 * just the IPv4 string is returned for inet_ntoa6.
4761 * Safely load the IPv6 address.
4764 (void *)(uintptr_t)tupregs[argi].dttk_value,
4765 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
4768 * Check an IPv6 string will fit in scratch.
4770 size = INET6_ADDRSTRLEN;
4771 if (!DTRACE_INSCRATCH(mstate, size)) {
4772 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4776 base = (char *)mstate->dtms_scratch_ptr;
4777 end = (char *)mstate->dtms_scratch_ptr + size - 1;
4781 * Find the longest run of 16 bit zero values
4782 * for the single allowed zero compression - "::".
4787 for (i = 0; i < sizeof (struct in6_addr); i++) {
4789 if (ip6._S6_un._S6_u8[i] == 0 &&
4791 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4793 tryzero == -1 && i % 2 == 0) {
4798 if (tryzero != -1 &&
4800 (ip6._S6_un._S6_u8[i] != 0 ||
4802 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
4804 i == sizeof (struct in6_addr) - 1)) {
4806 if (i - tryzero <= numzero) {
4811 firstzero = tryzero;
4812 numzero = i - i % 2 - tryzero;
4816 if (ip6._S6_un._S6_u8[i] == 0 &&
4818 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
4820 i == sizeof (struct in6_addr) - 1)
4824 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
4827 * Check for an IPv4 embedded address.
4829 v6end = sizeof (struct in6_addr) - 2;
4830 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
4831 IN6_IS_ADDR_V4COMPAT(&ip6)) {
4832 for (i = sizeof (struct in6_addr) - 1;
4833 i >= DTRACE_V4MAPPED_OFFSET; i--) {
4834 ASSERT(end >= base);
4837 val = ip6._S6_un._S6_u8[i];
4839 val = ip6.__u6_addr.__u6_addr8[i];
4845 for (; val; val /= 10) {
4846 *end-- = '0' + val % 10;
4850 if (i > DTRACE_V4MAPPED_OFFSET)
4854 if (subr == DIF_SUBR_INET_NTOA6)
4858 * Set v6end to skip the IPv4 address that
4859 * we have already stringified.
4865 * Build the IPv6 string by working through the
4866 * address in reverse.
4868 for (i = v6end; i >= 0; i -= 2) {
4869 ASSERT(end >= base);
4871 if (i == firstzero + numzero - 2) {
4878 if (i < 14 && i != firstzero - 2)
4882 val = (ip6._S6_un._S6_u8[i] << 8) +
4883 ip6._S6_un._S6_u8[i + 1];
4885 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
4886 ip6.__u6_addr.__u6_addr8[i + 1];
4892 for (; val; val /= 16) {
4893 *end-- = digits[val % 16];
4897 ASSERT(end + 1 >= base);
4901 * The user didn't use AH_INET or AH_INET6.
4903 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
4908 inetout: regs[rd] = (uintptr_t)end + 1;
4909 mstate->dtms_scratch_ptr += size;
4913 case DIF_SUBR_MEMREF: {
4914 uintptr_t size = 2 * sizeof(uintptr_t);
4915 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4916 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
4918 /* address and length */
4919 memref[0] = tupregs[0].dttk_value;
4920 memref[1] = tupregs[1].dttk_value;
4922 regs[rd] = (uintptr_t) memref;
4923 mstate->dtms_scratch_ptr += scratch_size;
4928 case DIF_SUBR_MEMSTR: {
4929 char *str = (char *)mstate->dtms_scratch_ptr;
4930 uintptr_t mem = tupregs[0].dttk_value;
4931 char c = tupregs[1].dttk_value;
4932 size_t size = tupregs[2].dttk_value;
4941 if (!dtrace_canload(mem, size - 1, mstate, vstate))
4944 if (!DTRACE_INSCRATCH(mstate, size)) {
4945 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4949 if (dtrace_memstr_max != 0 && size > dtrace_memstr_max) {
4950 *flags |= CPU_DTRACE_ILLOP;
4954 for (i = 0; i < size - 1; i++) {
4955 n = dtrace_load8(mem++);
4956 str[i] = (n == 0) ? c : n;
4960 regs[rd] = (uintptr_t)str;
4961 mstate->dtms_scratch_ptr += size;
4966 case DIF_SUBR_TYPEREF: {
4967 uintptr_t size = 4 * sizeof(uintptr_t);
4968 uintptr_t *typeref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
4969 size_t scratch_size = ((uintptr_t) typeref - mstate->dtms_scratch_ptr) + size;
4971 /* address, num_elements, type_str, type_len */
4972 typeref[0] = tupregs[0].dttk_value;
4973 typeref[1] = tupregs[1].dttk_value;
4974 typeref[2] = tupregs[2].dttk_value;
4975 typeref[3] = tupregs[3].dttk_value;
4977 regs[rd] = (uintptr_t) typeref;
4978 mstate->dtms_scratch_ptr += scratch_size;
4985 * Emulate the execution of DTrace IR instructions specified by the given
4986 * DIF object. This function is deliberately void of assertions as all of
4987 * the necessary checks are handled by a call to dtrace_difo_validate().
4990 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
4991 dtrace_vstate_t *vstate, dtrace_state_t *state)
4993 const dif_instr_t *text = difo->dtdo_buf;
4994 const uint_t textlen = difo->dtdo_len;
4995 const char *strtab = difo->dtdo_strtab;
4996 const uint64_t *inttab = difo->dtdo_inttab;
4999 dtrace_statvar_t *svar;
5000 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
5002 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5003 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
5005 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
5006 uint64_t regs[DIF_DIR_NREGS];
5009 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
5011 uint_t pc = 0, id, opc = 0;
5017 * We stash the current DIF object into the machine state: we need it
5018 * for subsequent access checking.
5020 mstate->dtms_difo = difo;
5022 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
5024 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
5028 r1 = DIF_INSTR_R1(instr);
5029 r2 = DIF_INSTR_R2(instr);
5030 rd = DIF_INSTR_RD(instr);
5032 switch (DIF_INSTR_OP(instr)) {
5034 regs[rd] = regs[r1] | regs[r2];
5037 regs[rd] = regs[r1] ^ regs[r2];
5040 regs[rd] = regs[r1] & regs[r2];
5043 regs[rd] = regs[r1] << regs[r2];
5046 regs[rd] = regs[r1] >> regs[r2];
5049 regs[rd] = regs[r1] - regs[r2];
5052 regs[rd] = regs[r1] + regs[r2];
5055 regs[rd] = regs[r1] * regs[r2];
5058 if (regs[r2] == 0) {
5060 *flags |= CPU_DTRACE_DIVZERO;
5062 regs[rd] = (int64_t)regs[r1] /
5068 if (regs[r2] == 0) {
5070 *flags |= CPU_DTRACE_DIVZERO;
5072 regs[rd] = regs[r1] / regs[r2];
5077 if (regs[r2] == 0) {
5079 *flags |= CPU_DTRACE_DIVZERO;
5081 regs[rd] = (int64_t)regs[r1] %
5087 if (regs[r2] == 0) {
5089 *flags |= CPU_DTRACE_DIVZERO;
5091 regs[rd] = regs[r1] % regs[r2];
5096 regs[rd] = ~regs[r1];
5099 regs[rd] = regs[r1];
5102 cc_r = regs[r1] - regs[r2];
5106 cc_c = regs[r1] < regs[r2];
5109 cc_n = cc_v = cc_c = 0;
5110 cc_z = regs[r1] == 0;
5113 pc = DIF_INSTR_LABEL(instr);
5117 pc = DIF_INSTR_LABEL(instr);
5121 pc = DIF_INSTR_LABEL(instr);
5124 if ((cc_z | (cc_n ^ cc_v)) == 0)
5125 pc = DIF_INSTR_LABEL(instr);
5128 if ((cc_c | cc_z) == 0)
5129 pc = DIF_INSTR_LABEL(instr);
5132 if ((cc_n ^ cc_v) == 0)
5133 pc = DIF_INSTR_LABEL(instr);
5137 pc = DIF_INSTR_LABEL(instr);
5141 pc = DIF_INSTR_LABEL(instr);
5145 pc = DIF_INSTR_LABEL(instr);
5148 if (cc_z | (cc_n ^ cc_v))
5149 pc = DIF_INSTR_LABEL(instr);
5153 pc = DIF_INSTR_LABEL(instr);
5156 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5157 *flags |= CPU_DTRACE_KPRIV;
5163 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
5166 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5167 *flags |= CPU_DTRACE_KPRIV;
5173 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
5176 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5177 *flags |= CPU_DTRACE_KPRIV;
5183 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
5186 if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
5187 *flags |= CPU_DTRACE_KPRIV;
5193 regs[rd] = dtrace_load8(regs[r1]);
5196 if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
5197 *flags |= CPU_DTRACE_KPRIV;
5203 regs[rd] = dtrace_load16(regs[r1]);
5206 if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
5207 *flags |= CPU_DTRACE_KPRIV;
5213 regs[rd] = dtrace_load32(regs[r1]);
5216 if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
5217 *flags |= CPU_DTRACE_KPRIV;
5223 regs[rd] = dtrace_load64(regs[r1]);
5227 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5230 regs[rd] = (int16_t)
5231 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5234 regs[rd] = (int32_t)
5235 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5239 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
5243 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
5247 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
5251 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
5260 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
5263 regs[rd] = (uint64_t)(uintptr_t)
5264 (strtab + DIF_INSTR_STRING(instr));
5267 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
5268 uintptr_t s1 = regs[r1];
5269 uintptr_t s2 = regs[r2];
5272 !dtrace_strcanload(s1, sz, mstate, vstate))
5275 !dtrace_strcanload(s2, sz, mstate, vstate))
5278 cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);
5286 regs[rd] = dtrace_dif_variable(mstate, state,
5290 id = DIF_INSTR_VAR(instr);
5292 if (id >= DIF_VAR_OTHER_UBASE) {
5295 id -= DIF_VAR_OTHER_UBASE;
5296 svar = vstate->dtvs_globals[id];
5297 ASSERT(svar != NULL);
5298 v = &svar->dtsv_var;
5300 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
5301 regs[rd] = svar->dtsv_data;
5305 a = (uintptr_t)svar->dtsv_data;
5307 if (*(uint8_t *)a == UINT8_MAX) {
5309 * If the 0th byte is set to UINT8_MAX
5310 * then this is to be treated as a
5311 * reference to a NULL variable.
5315 regs[rd] = a + sizeof (uint64_t);
5321 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
5325 id = DIF_INSTR_VAR(instr);
5327 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5328 id -= DIF_VAR_OTHER_UBASE;
5330 svar = vstate->dtvs_globals[id];
5331 ASSERT(svar != NULL);
5332 v = &svar->dtsv_var;
5334 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5335 uintptr_t a = (uintptr_t)svar->dtsv_data;
5338 ASSERT(svar->dtsv_size != 0);
5340 if (regs[rd] == 0) {
5341 *(uint8_t *)a = UINT8_MAX;
5345 a += sizeof (uint64_t);
5347 if (!dtrace_vcanload(
5348 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5352 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5353 (void *)a, &v->dtdv_type);
5357 svar->dtsv_data = regs[rd];
5362 * There are no DTrace built-in thread-local arrays at
5363 * present. This opcode is saved for future work.
5365 *flags |= CPU_DTRACE_ILLOP;
5370 id = DIF_INSTR_VAR(instr);
5372 if (id < DIF_VAR_OTHER_UBASE) {
5374 * For now, this has no meaning.
5380 id -= DIF_VAR_OTHER_UBASE;
5382 ASSERT(id < vstate->dtvs_nlocals);
5383 ASSERT(vstate->dtvs_locals != NULL);
5385 svar = vstate->dtvs_locals[id];
5386 ASSERT(svar != NULL);
5387 v = &svar->dtsv_var;
5389 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5390 uintptr_t a = (uintptr_t)svar->dtsv_data;
5391 size_t sz = v->dtdv_type.dtdt_size;
5393 sz += sizeof (uint64_t);
5394 ASSERT(svar->dtsv_size == NCPU * sz);
5397 if (*(uint8_t *)a == UINT8_MAX) {
5399 * If the 0th byte is set to UINT8_MAX
5400 * then this is to be treated as a
5401 * reference to a NULL variable.
5405 regs[rd] = a + sizeof (uint64_t);
5411 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5412 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5413 regs[rd] = tmp[curcpu];
5417 id = DIF_INSTR_VAR(instr);
5419 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5420 id -= DIF_VAR_OTHER_UBASE;
5421 ASSERT(id < vstate->dtvs_nlocals);
5423 ASSERT(vstate->dtvs_locals != NULL);
5424 svar = vstate->dtvs_locals[id];
5425 ASSERT(svar != NULL);
5426 v = &svar->dtsv_var;
5428 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5429 uintptr_t a = (uintptr_t)svar->dtsv_data;
5430 size_t sz = v->dtdv_type.dtdt_size;
5432 sz += sizeof (uint64_t);
5433 ASSERT(svar->dtsv_size == NCPU * sz);
5436 if (regs[rd] == 0) {
5437 *(uint8_t *)a = UINT8_MAX;
5441 a += sizeof (uint64_t);
5444 if (!dtrace_vcanload(
5445 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5449 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5450 (void *)a, &v->dtdv_type);
5454 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
5455 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
5456 tmp[curcpu] = regs[rd];
5460 dtrace_dynvar_t *dvar;
5463 id = DIF_INSTR_VAR(instr);
5464 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5465 id -= DIF_VAR_OTHER_UBASE;
5466 v = &vstate->dtvs_tlocals[id];
5468 key = &tupregs[DIF_DTR_NREGS];
5469 key[0].dttk_value = (uint64_t)id;
5470 key[0].dttk_size = 0;
5471 DTRACE_TLS_THRKEY(key[1].dttk_value);
5472 key[1].dttk_size = 0;
5474 dvar = dtrace_dynvar(dstate, 2, key,
5475 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
5483 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5484 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5486 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5493 dtrace_dynvar_t *dvar;
5496 id = DIF_INSTR_VAR(instr);
5497 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5498 id -= DIF_VAR_OTHER_UBASE;
5500 key = &tupregs[DIF_DTR_NREGS];
5501 key[0].dttk_value = (uint64_t)id;
5502 key[0].dttk_size = 0;
5503 DTRACE_TLS_THRKEY(key[1].dttk_value);
5504 key[1].dttk_size = 0;
5505 v = &vstate->dtvs_tlocals[id];
5507 dvar = dtrace_dynvar(dstate, 2, key,
5508 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5509 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5510 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5511 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5514 * Given that we're storing to thread-local data,
5515 * we need to flush our predicate cache.
5517 curthread->t_predcache = 0;
5522 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5523 if (!dtrace_vcanload(
5524 (void *)(uintptr_t)regs[rd],
5525 &v->dtdv_type, mstate, vstate))
5528 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5529 dvar->dtdv_data, &v->dtdv_type);
5531 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5538 regs[rd] = (int64_t)regs[r1] >> regs[r2];
5542 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
5543 regs, tupregs, ttop, mstate, state);
5547 if (ttop == DIF_DTR_NREGS) {
5548 *flags |= CPU_DTRACE_TUPOFLOW;
5552 if (r1 == DIF_TYPE_STRING) {
5554 * If this is a string type and the size is 0,
5555 * we'll use the system-wide default string
5556 * size. Note that we are _not_ looking at
5557 * the value of the DTRACEOPT_STRSIZE option;
5558 * had this been set, we would expect to have
5559 * a non-zero size value in the "pushtr".
5561 tupregs[ttop].dttk_size =
5562 dtrace_strlen((char *)(uintptr_t)regs[rd],
5563 regs[r2] ? regs[r2] :
5564 dtrace_strsize_default) + 1;
5566 tupregs[ttop].dttk_size = regs[r2];
5569 tupregs[ttop++].dttk_value = regs[rd];
5573 if (ttop == DIF_DTR_NREGS) {
5574 *flags |= CPU_DTRACE_TUPOFLOW;
5578 tupregs[ttop].dttk_value = regs[rd];
5579 tupregs[ttop++].dttk_size = 0;
5587 case DIF_OP_FLUSHTS:
5592 case DIF_OP_LDTAA: {
5593 dtrace_dynvar_t *dvar;
5594 dtrace_key_t *key = tupregs;
5595 uint_t nkeys = ttop;
5597 id = DIF_INSTR_VAR(instr);
5598 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5599 id -= DIF_VAR_OTHER_UBASE;
5601 key[nkeys].dttk_value = (uint64_t)id;
5602 key[nkeys++].dttk_size = 0;
5604 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
5605 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5606 key[nkeys++].dttk_size = 0;
5607 v = &vstate->dtvs_tlocals[id];
5609 v = &vstate->dtvs_globals[id]->dtsv_var;
5612 dvar = dtrace_dynvar(dstate, nkeys, key,
5613 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5614 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5615 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
5622 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5623 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
5625 regs[rd] = *((uint64_t *)dvar->dtdv_data);
5632 case DIF_OP_STTAA: {
5633 dtrace_dynvar_t *dvar;
5634 dtrace_key_t *key = tupregs;
5635 uint_t nkeys = ttop;
5637 id = DIF_INSTR_VAR(instr);
5638 ASSERT(id >= DIF_VAR_OTHER_UBASE);
5639 id -= DIF_VAR_OTHER_UBASE;
5641 key[nkeys].dttk_value = (uint64_t)id;
5642 key[nkeys++].dttk_size = 0;
5644 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
5645 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
5646 key[nkeys++].dttk_size = 0;
5647 v = &vstate->dtvs_tlocals[id];
5649 v = &vstate->dtvs_globals[id]->dtsv_var;
5652 dvar = dtrace_dynvar(dstate, nkeys, key,
5653 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
5654 v->dtdv_type.dtdt_size : sizeof (uint64_t),
5655 regs[rd] ? DTRACE_DYNVAR_ALLOC :
5656 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
5661 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
5662 if (!dtrace_vcanload(
5663 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
5667 dtrace_vcopy((void *)(uintptr_t)regs[rd],
5668 dvar->dtdv_data, &v->dtdv_type);
5670 *((uint64_t *)dvar->dtdv_data) = regs[rd];
5676 case DIF_OP_ALLOCS: {
5677 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5678 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
5681 * Rounding up the user allocation size could have
5682 * overflowed large, bogus allocations (like -1ULL) to
5685 if (size < regs[r1] ||
5686 !DTRACE_INSCRATCH(mstate, size)) {
5687 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5692 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
5693 mstate->dtms_scratch_ptr += size;
5699 if (!dtrace_canstore(regs[rd], regs[r2],
5701 *flags |= CPU_DTRACE_BADADDR;
5706 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
5709 dtrace_bcopy((void *)(uintptr_t)regs[r1],
5710 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
5714 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
5715 *flags |= CPU_DTRACE_BADADDR;
5719 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
5723 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
5724 *flags |= CPU_DTRACE_BADADDR;
5729 *flags |= CPU_DTRACE_BADALIGN;
5733 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
5737 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
5738 *flags |= CPU_DTRACE_BADADDR;
5743 *flags |= CPU_DTRACE_BADALIGN;
5747 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
5751 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
5752 *flags |= CPU_DTRACE_BADADDR;
5757 *flags |= CPU_DTRACE_BADALIGN;
5761 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
5766 if (!(*flags & CPU_DTRACE_FAULT))
5769 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
5770 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
5776 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
5778 dtrace_probe_t *probe = ecb->dte_probe;
5779 dtrace_provider_t *prov = probe->dtpr_provider;
5780 char c[DTRACE_FULLNAMELEN + 80], *str;
5781 char *msg = "dtrace: breakpoint action at probe ";
5782 char *ecbmsg = " (ecb ";
5783 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
5784 uintptr_t val = (uintptr_t)ecb;
5785 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
5787 if (dtrace_destructive_disallow)
5791 * It's impossible to be taking action on the NULL probe.
5793 ASSERT(probe != NULL);
5796 * This is a poor man's (destitute man's?) sprintf(): we want to
5797 * print the provider name, module name, function name and name of
5798 * the probe, along with the hex address of the ECB with the breakpoint
5799 * action -- all of which we must place in the character buffer by
5802 while (*msg != '\0')
5805 for (str = prov->dtpv_name; *str != '\0'; str++)
5809 for (str = probe->dtpr_mod; *str != '\0'; str++)
5813 for (str = probe->dtpr_func; *str != '\0'; str++)
5817 for (str = probe->dtpr_name; *str != '\0'; str++)
5820 while (*ecbmsg != '\0')
5823 while (shift >= 0) {
5824 mask = (uintptr_t)0xf << shift;
5826 if (val >= ((uintptr_t)1 << shift))
5827 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
5837 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
5842 dtrace_action_panic(dtrace_ecb_t *ecb)
5844 dtrace_probe_t *probe = ecb->dte_probe;
5847 * It's impossible to be taking action on the NULL probe.
5849 ASSERT(probe != NULL);
5851 if (dtrace_destructive_disallow)
5854 if (dtrace_panicked != NULL)
5857 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
5861 * We won the right to panic. (We want to be sure that only one
5862 * thread calls panic() from dtrace_probe(), and that panic() is
5863 * called exactly once.)
5865 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
5866 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
5867 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
5871 dtrace_action_raise(uint64_t sig)
5873 if (dtrace_destructive_disallow)
5877 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
5883 * raise() has a queue depth of 1 -- we ignore all subsequent
5884 * invocations of the raise() action.
5886 if (curthread->t_dtrace_sig == 0)
5887 curthread->t_dtrace_sig = (uint8_t)sig;
5889 curthread->t_sig_check = 1;
5892 struct proc *p = curproc;
5894 kern_psignal(p, sig);
5900 dtrace_action_stop(void)
5902 if (dtrace_destructive_disallow)
5906 if (!curthread->t_dtrace_stop) {
5907 curthread->t_dtrace_stop = 1;
5908 curthread->t_sig_check = 1;
5912 struct proc *p = curproc;
5914 kern_psignal(p, SIGSTOP);
5920 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
5923 volatile uint16_t *flags;
5927 cpu_t *cpu = &solaris_cpu[curcpu];
5930 if (dtrace_destructive_disallow)
5933 flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;
5935 now = dtrace_gethrtime();
5937 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
5939 * We need to advance the mark to the current time.
5941 cpu->cpu_dtrace_chillmark = now;
5942 cpu->cpu_dtrace_chilled = 0;
5946 * Now check to see if the requested chill time would take us over
5947 * the maximum amount of time allowed in the chill interval. (Or
5948 * worse, if the calculation itself induces overflow.)
5950 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
5951 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
5952 *flags |= CPU_DTRACE_ILLOP;
5956 while (dtrace_gethrtime() - now < val)
5960 * Normally, we assure that the value of the variable "timestamp" does
5961 * not change within an ECB. The presence of chill() represents an
5962 * exception to this rule, however.
5964 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
5965 cpu->cpu_dtrace_chilled += val;
5969 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
5970 uint64_t *buf, uint64_t arg)
5972 int nframes = DTRACE_USTACK_NFRAMES(arg);
5973 int strsize = DTRACE_USTACK_STRSIZE(arg);
5974 uint64_t *pcs = &buf[1], *fps;
5975 char *str = (char *)&pcs[nframes];
5976 int size, offs = 0, i, j;
5977 uintptr_t old = mstate->dtms_scratch_ptr, saved;
5978 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
5982 * Should be taking a faster path if string space has not been
5985 ASSERT(strsize != 0);
5988 * We will first allocate some temporary space for the frame pointers.
5990 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
5991 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
5992 (nframes * sizeof (uint64_t));
5994 if (!DTRACE_INSCRATCH(mstate, size)) {
5996 * Not enough room for our frame pointers -- need to indicate
5997 * that we ran out of scratch space.
5999 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6003 mstate->dtms_scratch_ptr += size;
6004 saved = mstate->dtms_scratch_ptr;
6007 * Now get a stack with both program counters and frame pointers.
6009 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6010 dtrace_getufpstack(buf, fps, nframes + 1);
6011 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6014 * If that faulted, we're cooked.
6016 if (*flags & CPU_DTRACE_FAULT)
6020 * Now we want to walk up the stack, calling the USTACK helper. For
6021 * each iteration, we restore the scratch pointer.
6023 for (i = 0; i < nframes; i++) {
6024 mstate->dtms_scratch_ptr = saved;
6026 if (offs >= strsize)
6029 sym = (char *)(uintptr_t)dtrace_helper(
6030 DTRACE_HELPER_ACTION_USTACK,
6031 mstate, state, pcs[i], fps[i]);
6034 * If we faulted while running the helper, we're going to
6035 * clear the fault and null out the corresponding string.
6037 if (*flags & CPU_DTRACE_FAULT) {
6038 *flags &= ~CPU_DTRACE_FAULT;
6048 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6051 * Now copy in the string that the helper returned to us.
6053 for (j = 0; offs + j < strsize; j++) {
6054 if ((str[offs + j] = sym[j]) == '\0')
6058 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6063 if (offs >= strsize) {
6065 * If we didn't have room for all of the strings, we don't
6066 * abort processing -- this needn't be a fatal error -- but we
6067 * still want to increment a counter (dts_stkstroverflows) to
6068 * allow this condition to be warned about. (If this is from
6069 * a jstack() action, it is easily tuned via jstackstrsize.)
6071 dtrace_error(&state->dts_stkstroverflows);
6074 while (offs < strsize)
6078 mstate->dtms_scratch_ptr = old;
6082 * If you're looking for the epicenter of DTrace, you just found it. This
6083 * is the function called by the provider to fire a probe -- from which all
6084 * subsequent probe-context DTrace activity emanates.
6087 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
6088 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
6090 processorid_t cpuid;
6091 dtrace_icookie_t cookie;
6092 dtrace_probe_t *probe;
6093 dtrace_mstate_t mstate;
6095 dtrace_action_t *act;
6099 volatile uint16_t *flags;
6102 if (panicstr != NULL)
6107 * Kick out immediately if this CPU is still being born (in which case
6108 * curthread will be set to -1) or the current thread can't allow
6109 * probes in its current context.
6111 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
6115 cookie = dtrace_interrupt_disable();
6116 probe = dtrace_probes[id - 1];
6118 onintr = CPU_ON_INTR(CPU);
6120 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
6121 probe->dtpr_predcache == curthread->t_predcache) {
6123 * We have hit in the predicate cache; we know that
6124 * this predicate would evaluate to be false.
6126 dtrace_interrupt_enable(cookie);
6131 if (panic_quiesce) {
6133 if (panicstr != NULL) {
6136 * We don't trace anything if we're panicking.
6138 dtrace_interrupt_enable(cookie);
6142 now = dtrace_gethrtime();
6143 vtime = dtrace_vtime_references != 0;
6145 if (vtime && curthread->t_dtrace_start)
6146 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
6148 mstate.dtms_difo = NULL;
6149 mstate.dtms_probe = probe;
6150 mstate.dtms_strtok = 0;
6151 mstate.dtms_arg[0] = arg0;
6152 mstate.dtms_arg[1] = arg1;
6153 mstate.dtms_arg[2] = arg2;
6154 mstate.dtms_arg[3] = arg3;
6155 mstate.dtms_arg[4] = arg4;
6157 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
6159 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
6160 dtrace_predicate_t *pred = ecb->dte_predicate;
6161 dtrace_state_t *state = ecb->dte_state;
6162 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
6163 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
6164 dtrace_vstate_t *vstate = &state->dts_vstate;
6165 dtrace_provider_t *prov = probe->dtpr_provider;
6166 uint64_t tracememsize = 0;
6171 * A little subtlety with the following (seemingly innocuous)
6172 * declaration of the automatic 'val': by looking at the
6173 * code, you might think that it could be declared in the
6174 * action processing loop, below. (That is, it's only used in
6175 * the action processing loop.) However, it must be declared
6176 * out of that scope because in the case of DIF expression
6177 * arguments to aggregating actions, one iteration of the
6178 * action loop will use the last iteration's value.
6182 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
6183 *flags &= ~CPU_DTRACE_ERROR;
6185 if (prov == dtrace_provider) {
6187 * If dtrace itself is the provider of this probe,
6188 * we're only going to continue processing the ECB if
6189 * arg0 (the dtrace_state_t) is equal to the ECB's
6190 * creating state. (This prevents disjoint consumers
6191 * from seeing one another's metaprobes.)
6193 if (arg0 != (uint64_t)(uintptr_t)state)
6197 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
6199 * We're not currently active. If our provider isn't
6200 * the dtrace pseudo provider, we're not interested.
6202 if (prov != dtrace_provider)
6206 * Now we must further check if we are in the BEGIN
6207 * probe. If we are, we will only continue processing
6208 * if we're still in WARMUP -- if one BEGIN enabling
6209 * has invoked the exit() action, we don't want to
6210 * evaluate subsequent BEGIN enablings.
6212 if (probe->dtpr_id == dtrace_probeid_begin &&
6213 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
6214 ASSERT(state->dts_activity ==
6215 DTRACE_ACTIVITY_DRAINING);
6220 if (ecb->dte_cond) {
6222 * If the dte_cond bits indicate that this
6223 * consumer is only allowed to see user-mode firings
6224 * of this probe, call the provider's dtps_usermode()
6225 * entry point to check that the probe was fired
6226 * while in a user context. Skip this ECB if that's
6229 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
6230 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
6231 probe->dtpr_id, probe->dtpr_arg) == 0)
6236 * This is more subtle than it looks. We have to be
6237 * absolutely certain that CRED() isn't going to
6238 * change out from under us so it's only legit to
6239 * examine that structure if we're in constrained
6240 * situations. Currently, the only times we'll this
6241 * check is if a non-super-user has enabled the
6242 * profile or syscall providers -- providers that
6243 * allow visibility of all processes. For the
6244 * profile case, the check above will ensure that
6245 * we're examining a user context.
6247 if (ecb->dte_cond & DTRACE_COND_OWNER) {
6250 ecb->dte_state->dts_cred.dcr_cred;
6253 ASSERT(s_cr != NULL);
6255 if ((cr = CRED()) == NULL ||
6256 s_cr->cr_uid != cr->cr_uid ||
6257 s_cr->cr_uid != cr->cr_ruid ||
6258 s_cr->cr_uid != cr->cr_suid ||
6259 s_cr->cr_gid != cr->cr_gid ||
6260 s_cr->cr_gid != cr->cr_rgid ||
6261 s_cr->cr_gid != cr->cr_sgid ||
6262 (proc = ttoproc(curthread)) == NULL ||
6263 (proc->p_flag & SNOCD))
6267 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
6270 ecb->dte_state->dts_cred.dcr_cred;
6272 ASSERT(s_cr != NULL);
6274 if ((cr = CRED()) == NULL ||
6275 s_cr->cr_zone->zone_id !=
6276 cr->cr_zone->zone_id)
6282 if (now - state->dts_alive > dtrace_deadman_timeout) {
6284 * We seem to be dead. Unless we (a) have kernel
6285 * destructive permissions (b) have explicitly enabled
6286 * destructive actions and (c) destructive actions have
6287 * not been disabled, we're going to transition into
6288 * the KILLED state, from which no further processing
6289 * on this state will be performed.
6291 if (!dtrace_priv_kernel_destructive(state) ||
6292 !state->dts_cred.dcr_destructive ||
6293 dtrace_destructive_disallow) {
6294 void *activity = &state->dts_activity;
6295 dtrace_activity_t current;
6298 current = state->dts_activity;
6299 } while (dtrace_cas32(activity, current,
6300 DTRACE_ACTIVITY_KILLED) != current);
6306 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
6307 ecb->dte_alignment, state, &mstate)) < 0)
6310 tomax = buf->dtb_tomax;
6311 ASSERT(tomax != NULL);
6313 if (ecb->dte_size != 0) {
6314 dtrace_rechdr_t dtrh;
6315 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
6316 mstate.dtms_timestamp = dtrace_gethrtime();
6317 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
6319 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
6320 dtrh.dtrh_epid = ecb->dte_epid;
6321 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
6322 mstate.dtms_timestamp);
6323 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
6326 mstate.dtms_epid = ecb->dte_epid;
6327 mstate.dtms_present |= DTRACE_MSTATE_EPID;
6329 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
6330 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
6332 mstate.dtms_access = 0;
6335 dtrace_difo_t *dp = pred->dtp_difo;
6338 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
6340 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
6341 dtrace_cacheid_t cid = probe->dtpr_predcache;
6343 if (cid != DTRACE_CACHEIDNONE && !onintr) {
6345 * Update the predicate cache...
6347 ASSERT(cid == pred->dtp_cacheid);
6348 curthread->t_predcache = cid;
6355 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
6356 act != NULL; act = act->dta_next) {
6359 dtrace_recdesc_t *rec = &act->dta_rec;
6361 size = rec->dtrd_size;
6362 valoffs = offs + rec->dtrd_offset;
6364 if (DTRACEACT_ISAGG(act->dta_kind)) {
6366 dtrace_aggregation_t *agg;
6368 agg = (dtrace_aggregation_t *)act;
6370 if ((dp = act->dta_difo) != NULL)
6371 v = dtrace_dif_emulate(dp,
6372 &mstate, vstate, state);
6374 if (*flags & CPU_DTRACE_ERROR)
6378 * Note that we always pass the expression
6379 * value from the previous iteration of the
6380 * action loop. This value will only be used
6381 * if there is an expression argument to the
6382 * aggregating action, denoted by the
6383 * dtag_hasarg field.
6385 dtrace_aggregate(agg, buf,
6386 offs, aggbuf, v, val);
6390 switch (act->dta_kind) {
6391 case DTRACEACT_STOP:
6392 if (dtrace_priv_proc_destructive(state))
6393 dtrace_action_stop();
6396 case DTRACEACT_BREAKPOINT:
6397 if (dtrace_priv_kernel_destructive(state))
6398 dtrace_action_breakpoint(ecb);
6401 case DTRACEACT_PANIC:
6402 if (dtrace_priv_kernel_destructive(state))
6403 dtrace_action_panic(ecb);
6406 case DTRACEACT_STACK:
6407 if (!dtrace_priv_kernel(state))
6410 dtrace_getpcstack((pc_t *)(tomax + valoffs),
6411 size / sizeof (pc_t), probe->dtpr_aframes,
6412 DTRACE_ANCHORED(probe) ? NULL :
6416 case DTRACEACT_JSTACK:
6417 case DTRACEACT_USTACK:
6418 if (!dtrace_priv_proc(state))
6422 * See comment in DIF_VAR_PID.
6424 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
6426 int depth = DTRACE_USTACK_NFRAMES(
6429 dtrace_bzero((void *)(tomax + valoffs),
6430 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
6431 + depth * sizeof (uint64_t));
6436 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
6437 curproc->p_dtrace_helpers != NULL) {
6439 * This is the slow path -- we have
6440 * allocated string space, and we're
6441 * getting the stack of a process that
6442 * has helpers. Call into a separate
6443 * routine to perform this processing.
6445 dtrace_action_ustack(&mstate, state,
6446 (uint64_t *)(tomax + valoffs),
6451 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6452 dtrace_getupcstack((uint64_t *)
6454 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
6455 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6465 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
6467 if (*flags & CPU_DTRACE_ERROR)
6470 switch (act->dta_kind) {
6471 case DTRACEACT_SPECULATE: {
6472 dtrace_rechdr_t *dtrh;
6474 ASSERT(buf == &state->dts_buffer[cpuid]);
6475 buf = dtrace_speculation_buffer(state,
6479 *flags |= CPU_DTRACE_DROP;
6483 offs = dtrace_buffer_reserve(buf,
6484 ecb->dte_needed, ecb->dte_alignment,
6488 *flags |= CPU_DTRACE_DROP;
6492 tomax = buf->dtb_tomax;
6493 ASSERT(tomax != NULL);
6495 if (ecb->dte_size == 0)
6498 ASSERT3U(ecb->dte_size, >=,
6499 sizeof (dtrace_rechdr_t));
6500 dtrh = ((void *)(tomax + offs));
6501 dtrh->dtrh_epid = ecb->dte_epid;
6503 * When the speculation is committed, all of
6504 * the records in the speculative buffer will
6505 * have their timestamps set to the commit
6506 * time. Until then, it is set to a sentinel
6507 * value, for debugability.
6509 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
6513 case DTRACEACT_PRINTM: {
6514 /* The DIF returns a 'memref'. */
6515 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
6517 /* Get the size from the memref. */
6521 * Check if the size exceeds the allocated
6524 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6526 *flags |= CPU_DTRACE_DROP;
6530 /* Store the size in the buffer first. */
6531 DTRACE_STORE(uintptr_t, tomax,
6535 * Offset the buffer address to the start
6538 valoffs += sizeof(uintptr_t);
6541 * Reset to the memory address rather than
6542 * the memref array, then let the BYREF
6543 * code below do the work to store the
6544 * memory data in the buffer.
6550 case DTRACEACT_PRINTT: {
6551 /* The DIF returns a 'typeref'. */
6552 uintptr_t *typeref = (uintptr_t *)(uintptr_t) val;
6557 * Get the type string length and round it
6558 * up so that the data that follows is
6559 * aligned for easy access.
6561 size_t typs = strlen((char *) typeref[2]) + 1;
6562 typs = roundup(typs, sizeof(uintptr_t));
6565 *Get the size from the typeref using the
6566 * number of elements and the type size.
6568 size = typeref[1] * typeref[3];
6571 * Check if the size exceeds the allocated
6574 if (size + typs + 2 * sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
6576 *flags |= CPU_DTRACE_DROP;
6580 /* Store the size in the buffer first. */
6581 DTRACE_STORE(uintptr_t, tomax,
6583 valoffs += sizeof(uintptr_t);
6585 /* Store the type size in the buffer. */
6586 DTRACE_STORE(uintptr_t, tomax,
6587 valoffs, typeref[3]);
6588 valoffs += sizeof(uintptr_t);
6592 for (s = 0; s < typs; s++) {
6594 c = dtrace_load8(val++);
6596 DTRACE_STORE(uint8_t, tomax,
6601 * Reset to the memory address rather than
6602 * the typeref array, then let the BYREF
6603 * code below do the work to store the
6604 * memory data in the buffer.
6610 case DTRACEACT_CHILL:
6611 if (dtrace_priv_kernel_destructive(state))
6612 dtrace_action_chill(&mstate, val);
6615 case DTRACEACT_RAISE:
6616 if (dtrace_priv_proc_destructive(state))
6617 dtrace_action_raise(val);
6620 case DTRACEACT_COMMIT:
6624 * We need to commit our buffer state.
6627 buf->dtb_offset = offs + ecb->dte_size;
6628 buf = &state->dts_buffer[cpuid];
6629 dtrace_speculation_commit(state, cpuid, val);
6633 case DTRACEACT_DISCARD:
6634 dtrace_speculation_discard(state, cpuid, val);
6637 case DTRACEACT_DIFEXPR:
6638 case DTRACEACT_LIBACT:
6639 case DTRACEACT_PRINTF:
6640 case DTRACEACT_PRINTA:
6641 case DTRACEACT_SYSTEM:
6642 case DTRACEACT_FREOPEN:
6643 case DTRACEACT_TRACEMEM:
6646 case DTRACEACT_TRACEMEM_DYNSIZE:
6652 if (!dtrace_priv_kernel(state))
6656 case DTRACEACT_USYM:
6657 case DTRACEACT_UMOD:
6658 case DTRACEACT_UADDR: {
6660 struct pid *pid = curthread->t_procp->p_pidp;
6663 if (!dtrace_priv_proc(state))
6666 DTRACE_STORE(uint64_t, tomax,
6668 valoffs, (uint64_t)pid->pid_id);
6670 valoffs, (uint64_t) curproc->p_pid);
6672 DTRACE_STORE(uint64_t, tomax,
6673 valoffs + sizeof (uint64_t), val);
6678 case DTRACEACT_EXIT: {
6680 * For the exit action, we are going to attempt
6681 * to atomically set our activity to be
6682 * draining. If this fails (either because
6683 * another CPU has beat us to the exit action,
6684 * or because our current activity is something
6685 * other than ACTIVE or WARMUP), we will
6686 * continue. This assures that the exit action
6687 * can be successfully recorded at most once
6688 * when we're in the ACTIVE state. If we're
6689 * encountering the exit() action while in
6690 * COOLDOWN, however, we want to honor the new
6691 * status code. (We know that we're the only
6692 * thread in COOLDOWN, so there is no race.)
6694 void *activity = &state->dts_activity;
6695 dtrace_activity_t current = state->dts_activity;
6697 if (current == DTRACE_ACTIVITY_COOLDOWN)
6700 if (current != DTRACE_ACTIVITY_WARMUP)
6701 current = DTRACE_ACTIVITY_ACTIVE;
6703 if (dtrace_cas32(activity, current,
6704 DTRACE_ACTIVITY_DRAINING) != current) {
6705 *flags |= CPU_DTRACE_DROP;
6716 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
6717 uintptr_t end = valoffs + size;
6719 if (tracememsize != 0 &&
6720 valoffs + tracememsize < end) {
6721 end = valoffs + tracememsize;
6725 if (!dtrace_vcanload((void *)(uintptr_t)val,
6726 &dp->dtdo_rtype, &mstate, vstate))
6730 * If this is a string, we're going to only
6731 * load until we find the zero byte -- after
6732 * which we'll store zero bytes.
6734 if (dp->dtdo_rtype.dtdt_kind ==
6737 int intuple = act->dta_intuple;
6740 for (s = 0; s < size; s++) {
6742 c = dtrace_load8(val++);
6744 DTRACE_STORE(uint8_t, tomax,
6747 if (c == '\0' && intuple)
6754 while (valoffs < end) {
6755 DTRACE_STORE(uint8_t, tomax, valoffs++,
6756 dtrace_load8(val++));
6766 case sizeof (uint8_t):
6767 DTRACE_STORE(uint8_t, tomax, valoffs, val);
6769 case sizeof (uint16_t):
6770 DTRACE_STORE(uint16_t, tomax, valoffs, val);
6772 case sizeof (uint32_t):
6773 DTRACE_STORE(uint32_t, tomax, valoffs, val);
6775 case sizeof (uint64_t):
6776 DTRACE_STORE(uint64_t, tomax, valoffs, val);
6780 * Any other size should have been returned by
6781 * reference, not by value.
6788 if (*flags & CPU_DTRACE_DROP)
6791 if (*flags & CPU_DTRACE_FAULT) {
6793 dtrace_action_t *err;
6797 if (probe->dtpr_id == dtrace_probeid_error) {
6799 * There's nothing we can do -- we had an
6800 * error on the error probe. We bump an
6801 * error counter to at least indicate that
6802 * this condition happened.
6804 dtrace_error(&state->dts_dblerrors);
6810 * Before recursing on dtrace_probe(), we
6811 * need to explicitly clear out our start
6812 * time to prevent it from being accumulated
6813 * into t_dtrace_vtime.
6815 curthread->t_dtrace_start = 0;
6819 * Iterate over the actions to figure out which action
6820 * we were processing when we experienced the error.
6821 * Note that act points _past_ the faulting action; if
6822 * act is ecb->dte_action, the fault was in the
6823 * predicate, if it's ecb->dte_action->dta_next it's
6824 * in action #1, and so on.
6826 for (err = ecb->dte_action, ndx = 0;
6827 err != act; err = err->dta_next, ndx++)
6830 dtrace_probe_error(state, ecb->dte_epid, ndx,
6831 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
6832 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
6833 cpu_core[cpuid].cpuc_dtrace_illval);
6839 buf->dtb_offset = offs + ecb->dte_size;
6843 curthread->t_dtrace_start = dtrace_gethrtime();
6845 dtrace_interrupt_enable(cookie);
6849 * DTrace Probe Hashing Functions
6851 * The functions in this section (and indeed, the functions in remaining
6852 * sections) are not _called_ from probe context. (Any exceptions to this are
6853 * marked with a "Note:".) Rather, they are called from elsewhere in the
6854 * DTrace framework to look-up probes in, add probes to and remove probes from
6855 * the DTrace probe hashes. (Each probe is hashed by each element of the
6856 * probe tuple -- allowing for fast lookups, regardless of what was
6860 dtrace_hash_str(const char *p)
6866 hval = (hval << 4) + *p++;
6867 if ((g = (hval & 0xf0000000)) != 0)
6874 static dtrace_hash_t *
6875 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
6877 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
6879 hash->dth_stroffs = stroffs;
6880 hash->dth_nextoffs = nextoffs;
6881 hash->dth_prevoffs = prevoffs;
6884 hash->dth_mask = hash->dth_size - 1;
6886 hash->dth_tab = kmem_zalloc(hash->dth_size *
6887 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
6893 dtrace_hash_destroy(dtrace_hash_t *hash)
6898 for (i = 0; i < hash->dth_size; i++)
6899 ASSERT(hash->dth_tab[i] == NULL);
6902 kmem_free(hash->dth_tab,
6903 hash->dth_size * sizeof (dtrace_hashbucket_t *));
6904 kmem_free(hash, sizeof (dtrace_hash_t));
6908 dtrace_hash_resize(dtrace_hash_t *hash)
6910 int size = hash->dth_size, i, ndx;
6911 int new_size = hash->dth_size << 1;
6912 int new_mask = new_size - 1;
6913 dtrace_hashbucket_t **new_tab, *bucket, *next;
6915 ASSERT((new_size & new_mask) == 0);
6917 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
6919 for (i = 0; i < size; i++) {
6920 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
6921 dtrace_probe_t *probe = bucket->dthb_chain;
6923 ASSERT(probe != NULL);
6924 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
6926 next = bucket->dthb_next;
6927 bucket->dthb_next = new_tab[ndx];
6928 new_tab[ndx] = bucket;
6932 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
6933 hash->dth_tab = new_tab;
6934 hash->dth_size = new_size;
6935 hash->dth_mask = new_mask;
6939 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
6941 int hashval = DTRACE_HASHSTR(hash, new);
6942 int ndx = hashval & hash->dth_mask;
6943 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6944 dtrace_probe_t **nextp, **prevp;
6946 for (; bucket != NULL; bucket = bucket->dthb_next) {
6947 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
6951 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
6952 dtrace_hash_resize(hash);
6953 dtrace_hash_add(hash, new);
6957 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
6958 bucket->dthb_next = hash->dth_tab[ndx];
6959 hash->dth_tab[ndx] = bucket;
6960 hash->dth_nbuckets++;
6963 nextp = DTRACE_HASHNEXT(hash, new);
6964 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
6965 *nextp = bucket->dthb_chain;
6967 if (bucket->dthb_chain != NULL) {
6968 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
6969 ASSERT(*prevp == NULL);
6973 bucket->dthb_chain = new;
6977 static dtrace_probe_t *
6978 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
6980 int hashval = DTRACE_HASHSTR(hash, template);
6981 int ndx = hashval & hash->dth_mask;
6982 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6984 for (; bucket != NULL; bucket = bucket->dthb_next) {
6985 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
6986 return (bucket->dthb_chain);
6993 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
6995 int hashval = DTRACE_HASHSTR(hash, template);
6996 int ndx = hashval & hash->dth_mask;
6997 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
6999 for (; bucket != NULL; bucket = bucket->dthb_next) {
7000 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
7001 return (bucket->dthb_len);
7008 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
7010 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
7011 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
7013 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
7014 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
7017 * Find the bucket that we're removing this probe from.
7019 for (; bucket != NULL; bucket = bucket->dthb_next) {
7020 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
7024 ASSERT(bucket != NULL);
7026 if (*prevp == NULL) {
7027 if (*nextp == NULL) {
7029 * The removed probe was the only probe on this
7030 * bucket; we need to remove the bucket.
7032 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
7034 ASSERT(bucket->dthb_chain == probe);
7038 hash->dth_tab[ndx] = bucket->dthb_next;
7040 while (b->dthb_next != bucket)
7042 b->dthb_next = bucket->dthb_next;
7045 ASSERT(hash->dth_nbuckets > 0);
7046 hash->dth_nbuckets--;
7047 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
7051 bucket->dthb_chain = *nextp;
7053 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
7057 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
7061 * DTrace Utility Functions
7063 * These are random utility functions that are _not_ called from probe context.
7066 dtrace_badattr(const dtrace_attribute_t *a)
7068 return (a->dtat_name > DTRACE_STABILITY_MAX ||
7069 a->dtat_data > DTRACE_STABILITY_MAX ||
7070 a->dtat_class > DTRACE_CLASS_MAX);
7074 * Return a duplicate copy of a string. If the specified string is NULL,
7075 * this function returns a zero-length string.
7078 dtrace_strdup(const char *str)
7080 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
7083 (void) strcpy(new, str);
7088 #define DTRACE_ISALPHA(c) \
7089 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
7092 dtrace_badname(const char *s)
7096 if (s == NULL || (c = *s++) == '\0')
7099 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
7102 while ((c = *s++) != '\0') {
7103 if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
7104 c != '-' && c != '_' && c != '.' && c != '`')
7112 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
7117 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
7119 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
7121 priv = DTRACE_PRIV_ALL;
7123 *uidp = crgetuid(cr);
7124 *zoneidp = crgetzoneid(cr);
7127 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
7128 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
7129 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
7130 priv |= DTRACE_PRIV_USER;
7131 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
7132 priv |= DTRACE_PRIV_PROC;
7133 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
7134 priv |= DTRACE_PRIV_OWNER;
7135 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
7136 priv |= DTRACE_PRIV_ZONEOWNER;
7139 priv = DTRACE_PRIV_ALL;
7145 #ifdef DTRACE_ERRDEBUG
7147 dtrace_errdebug(const char *str)
7149 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
7152 mutex_enter(&dtrace_errlock);
7153 dtrace_errlast = str;
7154 dtrace_errthread = curthread;
7156 while (occupied++ < DTRACE_ERRHASHSZ) {
7157 if (dtrace_errhash[hval].dter_msg == str) {
7158 dtrace_errhash[hval].dter_count++;
7162 if (dtrace_errhash[hval].dter_msg != NULL) {
7163 hval = (hval + 1) % DTRACE_ERRHASHSZ;
7167 dtrace_errhash[hval].dter_msg = str;
7168 dtrace_errhash[hval].dter_count = 1;
7172 panic("dtrace: undersized error hash");
7174 mutex_exit(&dtrace_errlock);
7179 * DTrace Matching Functions
7181 * These functions are used to match groups of probes, given some elements of
7182 * a probe tuple, or some globbed expressions for elements of a probe tuple.
7185 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
7188 if (priv != DTRACE_PRIV_ALL) {
7189 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
7190 uint32_t match = priv & ppriv;
7193 * No PRIV_DTRACE_* privileges...
7195 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
7196 DTRACE_PRIV_KERNEL)) == 0)
7200 * No matching bits, but there were bits to match...
7202 if (match == 0 && ppriv != 0)
7206 * Need to have permissions to the process, but don't...
7208 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
7209 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
7214 * Need to be in the same zone unless we possess the
7215 * privilege to examine all zones.
7217 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
7218 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
7227 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
7228 * consists of input pattern strings and an ops-vector to evaluate them.
7229 * This function returns >0 for match, 0 for no match, and <0 for error.
7232 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
7233 uint32_t priv, uid_t uid, zoneid_t zoneid)
7235 dtrace_provider_t *pvp = prp->dtpr_provider;
7238 if (pvp->dtpv_defunct)
7241 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
7244 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
7247 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
7250 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
7253 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
7260 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
7261 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
7262 * libc's version, the kernel version only applies to 8-bit ASCII strings.
7263 * In addition, all of the recursion cases except for '*' matching have been
7264 * unwound. For '*', we still implement recursive evaluation, but a depth
7265 * counter is maintained and matching is aborted if we recurse too deep.
7266 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
7269 dtrace_match_glob(const char *s, const char *p, int depth)
7275 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
7279 s = ""; /* treat NULL as empty string */
7288 if ((c = *p++) == '\0')
7289 return (s1 == '\0');
7293 int ok = 0, notflag = 0;
7304 if ((c = *p++) == '\0')
7308 if (c == '-' && lc != '\0' && *p != ']') {
7309 if ((c = *p++) == '\0')
7311 if (c == '\\' && (c = *p++) == '\0')
7315 if (s1 < lc || s1 > c)
7319 } else if (lc <= s1 && s1 <= c)
7322 } else if (c == '\\' && (c = *p++) == '\0')
7325 lc = c; /* save left-hand 'c' for next iteration */
7335 if ((c = *p++) == '\0')
7347 if ((c = *p++) == '\0')
7363 p++; /* consecutive *'s are identical to a single one */
7368 for (s = olds; *s != '\0'; s++) {
7369 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
7379 dtrace_match_string(const char *s, const char *p, int depth)
7381 return (s != NULL && strcmp(s, p) == 0);
7386 dtrace_match_nul(const char *s, const char *p, int depth)
7388 return (1); /* always match the empty pattern */
7393 dtrace_match_nonzero(const char *s, const char *p, int depth)
7395 return (s != NULL && s[0] != '\0');
7399 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
7400 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
7402 dtrace_probe_t template, *probe;
7403 dtrace_hash_t *hash = NULL;
7404 int len, best = INT_MAX, nmatched = 0;
7407 ASSERT(MUTEX_HELD(&dtrace_lock));
7410 * If the probe ID is specified in the key, just lookup by ID and
7411 * invoke the match callback once if a matching probe is found.
7413 if (pkp->dtpk_id != DTRACE_IDNONE) {
7414 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
7415 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
7416 (void) (*matched)(probe, arg);
7422 template.dtpr_mod = (char *)pkp->dtpk_mod;
7423 template.dtpr_func = (char *)pkp->dtpk_func;
7424 template.dtpr_name = (char *)pkp->dtpk_name;
7427 * We want to find the most distinct of the module name, function
7428 * name, and name. So for each one that is not a glob pattern or
7429 * empty string, we perform a lookup in the corresponding hash and
7430 * use the hash table with the fewest collisions to do our search.
7432 if (pkp->dtpk_mmatch == &dtrace_match_string &&
7433 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
7435 hash = dtrace_bymod;
7438 if (pkp->dtpk_fmatch == &dtrace_match_string &&
7439 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
7441 hash = dtrace_byfunc;
7444 if (pkp->dtpk_nmatch == &dtrace_match_string &&
7445 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
7447 hash = dtrace_byname;
7451 * If we did not select a hash table, iterate over every probe and
7452 * invoke our callback for each one that matches our input probe key.
7455 for (i = 0; i < dtrace_nprobes; i++) {
7456 if ((probe = dtrace_probes[i]) == NULL ||
7457 dtrace_match_probe(probe, pkp, priv, uid,
7463 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7471 * If we selected a hash table, iterate over each probe of the same key
7472 * name and invoke the callback for every probe that matches the other
7473 * attributes of our input probe key.
7475 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
7476 probe = *(DTRACE_HASHNEXT(hash, probe))) {
7478 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
7483 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
7491 * Return the function pointer dtrace_probecmp() should use to compare the
7492 * specified pattern with a string. For NULL or empty patterns, we select
7493 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
7494 * For non-empty non-glob strings, we use dtrace_match_string().
7496 static dtrace_probekey_f *
7497 dtrace_probekey_func(const char *p)
7501 if (p == NULL || *p == '\0')
7502 return (&dtrace_match_nul);
7504 while ((c = *p++) != '\0') {
7505 if (c == '[' || c == '?' || c == '*' || c == '\\')
7506 return (&dtrace_match_glob);
7509 return (&dtrace_match_string);
7513 * Build a probe comparison key for use with dtrace_match_probe() from the
7514 * given probe description. By convention, a null key only matches anchored
7515 * probes: if each field is the empty string, reset dtpk_fmatch to
7516 * dtrace_match_nonzero().
7519 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
7521 pkp->dtpk_prov = pdp->dtpd_provider;
7522 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
7524 pkp->dtpk_mod = pdp->dtpd_mod;
7525 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
7527 pkp->dtpk_func = pdp->dtpd_func;
7528 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
7530 pkp->dtpk_name = pdp->dtpd_name;
7531 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
7533 pkp->dtpk_id = pdp->dtpd_id;
7535 if (pkp->dtpk_id == DTRACE_IDNONE &&
7536 pkp->dtpk_pmatch == &dtrace_match_nul &&
7537 pkp->dtpk_mmatch == &dtrace_match_nul &&
7538 pkp->dtpk_fmatch == &dtrace_match_nul &&
7539 pkp->dtpk_nmatch == &dtrace_match_nul)
7540 pkp->dtpk_fmatch = &dtrace_match_nonzero;
7544 * DTrace Provider-to-Framework API Functions
7546 * These functions implement much of the Provider-to-Framework API, as
7547 * described in <sys/dtrace.h>. The parts of the API not in this section are
7548 * the functions in the API for probe management (found below), and
7549 * dtrace_probe() itself (found above).
7553 * Register the calling provider with the DTrace framework. This should
7554 * generally be called by DTrace providers in their attach(9E) entry point.
7557 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
7558 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
7560 dtrace_provider_t *provider;
7562 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
7563 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7564 "arguments", name ? name : "<NULL>");
7568 if (name[0] == '\0' || dtrace_badname(name)) {
7569 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7570 "provider name", name);
7574 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
7575 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
7576 pops->dtps_destroy == NULL ||
7577 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
7578 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7579 "provider ops", name);
7583 if (dtrace_badattr(&pap->dtpa_provider) ||
7584 dtrace_badattr(&pap->dtpa_mod) ||
7585 dtrace_badattr(&pap->dtpa_func) ||
7586 dtrace_badattr(&pap->dtpa_name) ||
7587 dtrace_badattr(&pap->dtpa_args)) {
7588 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7589 "provider attributes", name);
7593 if (priv & ~DTRACE_PRIV_ALL) {
7594 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
7595 "privilege attributes", name);
7599 if ((priv & DTRACE_PRIV_KERNEL) &&
7600 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
7601 pops->dtps_usermode == NULL) {
7602 cmn_err(CE_WARN, "failed to register provider '%s': need "
7603 "dtps_usermode() op for given privilege attributes", name);
7607 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
7608 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
7609 (void) strcpy(provider->dtpv_name, name);
7611 provider->dtpv_attr = *pap;
7612 provider->dtpv_priv.dtpp_flags = priv;
7614 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
7615 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
7617 provider->dtpv_pops = *pops;
7619 if (pops->dtps_provide == NULL) {
7620 ASSERT(pops->dtps_provide_module != NULL);
7621 provider->dtpv_pops.dtps_provide =
7622 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
7625 if (pops->dtps_provide_module == NULL) {
7626 ASSERT(pops->dtps_provide != NULL);
7627 provider->dtpv_pops.dtps_provide_module =
7628 (void (*)(void *, modctl_t *))dtrace_nullop;
7631 if (pops->dtps_suspend == NULL) {
7632 ASSERT(pops->dtps_resume == NULL);
7633 provider->dtpv_pops.dtps_suspend =
7634 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7635 provider->dtpv_pops.dtps_resume =
7636 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
7639 provider->dtpv_arg = arg;
7640 *idp = (dtrace_provider_id_t)provider;
7642 if (pops == &dtrace_provider_ops) {
7643 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7644 ASSERT(MUTEX_HELD(&dtrace_lock));
7645 ASSERT(dtrace_anon.dta_enabling == NULL);
7648 * We make sure that the DTrace provider is at the head of
7649 * the provider chain.
7651 provider->dtpv_next = dtrace_provider;
7652 dtrace_provider = provider;
7656 mutex_enter(&dtrace_provider_lock);
7657 mutex_enter(&dtrace_lock);
7660 * If there is at least one provider registered, we'll add this
7661 * provider after the first provider.
7663 if (dtrace_provider != NULL) {
7664 provider->dtpv_next = dtrace_provider->dtpv_next;
7665 dtrace_provider->dtpv_next = provider;
7667 dtrace_provider = provider;
7670 if (dtrace_retained != NULL) {
7671 dtrace_enabling_provide(provider);
7674 * Now we need to call dtrace_enabling_matchall() -- which
7675 * will acquire cpu_lock and dtrace_lock. We therefore need
7676 * to drop all of our locks before calling into it...
7678 mutex_exit(&dtrace_lock);
7679 mutex_exit(&dtrace_provider_lock);
7680 dtrace_enabling_matchall();
7685 mutex_exit(&dtrace_lock);
7686 mutex_exit(&dtrace_provider_lock);
7692 * Unregister the specified provider from the DTrace framework. This should
7693 * generally be called by DTrace providers in their detach(9E) entry point.
7696 dtrace_unregister(dtrace_provider_id_t id)
7698 dtrace_provider_t *old = (dtrace_provider_t *)id;
7699 dtrace_provider_t *prev = NULL;
7700 int i, self = 0, noreap = 0;
7701 dtrace_probe_t *probe, *first = NULL;
7703 if (old->dtpv_pops.dtps_enable ==
7704 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
7706 * If DTrace itself is the provider, we're called with locks
7709 ASSERT(old == dtrace_provider);
7711 ASSERT(dtrace_devi != NULL);
7713 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
7714 ASSERT(MUTEX_HELD(&dtrace_lock));
7717 if (dtrace_provider->dtpv_next != NULL) {
7719 * There's another provider here; return failure.
7724 mutex_enter(&dtrace_provider_lock);
7726 mutex_enter(&mod_lock);
7728 mutex_enter(&dtrace_lock);
7732 * If anyone has /dev/dtrace open, or if there are anonymous enabled
7733 * probes, we refuse to let providers slither away, unless this
7734 * provider has already been explicitly invalidated.
7736 if (!old->dtpv_defunct &&
7737 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
7738 dtrace_anon.dta_state->dts_necbs > 0))) {
7740 mutex_exit(&dtrace_lock);
7742 mutex_exit(&mod_lock);
7744 mutex_exit(&dtrace_provider_lock);
7750 * Attempt to destroy the probes associated with this provider.
7752 for (i = 0; i < dtrace_nprobes; i++) {
7753 if ((probe = dtrace_probes[i]) == NULL)
7756 if (probe->dtpr_provider != old)
7759 if (probe->dtpr_ecb == NULL)
7763 * If we are trying to unregister a defunct provider, and the
7764 * provider was made defunct within the interval dictated by
7765 * dtrace_unregister_defunct_reap, we'll (asynchronously)
7766 * attempt to reap our enablings. To denote that the provider
7767 * should reattempt to unregister itself at some point in the
7768 * future, we will return a differentiable error code (EAGAIN
7769 * instead of EBUSY) in this case.
7771 if (dtrace_gethrtime() - old->dtpv_defunct >
7772 dtrace_unregister_defunct_reap)
7776 mutex_exit(&dtrace_lock);
7778 mutex_exit(&mod_lock);
7780 mutex_exit(&dtrace_provider_lock);
7786 (void) taskq_dispatch(dtrace_taskq,
7787 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
7793 * All of the probes for this provider are disabled; we can safely
7794 * remove all of them from their hash chains and from the probe array.
7796 for (i = 0; i < dtrace_nprobes; i++) {
7797 if ((probe = dtrace_probes[i]) == NULL)
7800 if (probe->dtpr_provider != old)
7803 dtrace_probes[i] = NULL;
7805 dtrace_hash_remove(dtrace_bymod, probe);
7806 dtrace_hash_remove(dtrace_byfunc, probe);
7807 dtrace_hash_remove(dtrace_byname, probe);
7809 if (first == NULL) {
7811 probe->dtpr_nextmod = NULL;
7813 probe->dtpr_nextmod = first;
7819 * The provider's probes have been removed from the hash chains and
7820 * from the probe array. Now issue a dtrace_sync() to be sure that
7821 * everyone has cleared out from any probe array processing.
7825 for (probe = first; probe != NULL; probe = first) {
7826 first = probe->dtpr_nextmod;
7828 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
7830 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7831 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7832 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7834 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
7836 free_unr(dtrace_arena, probe->dtpr_id);
7838 kmem_free(probe, sizeof (dtrace_probe_t));
7841 if ((prev = dtrace_provider) == old) {
7843 ASSERT(self || dtrace_devi == NULL);
7844 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
7846 dtrace_provider = old->dtpv_next;
7848 while (prev != NULL && prev->dtpv_next != old)
7849 prev = prev->dtpv_next;
7852 panic("attempt to unregister non-existent "
7853 "dtrace provider %p\n", (void *)id);
7856 prev->dtpv_next = old->dtpv_next;
7860 mutex_exit(&dtrace_lock);
7862 mutex_exit(&mod_lock);
7864 mutex_exit(&dtrace_provider_lock);
7867 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
7868 kmem_free(old, sizeof (dtrace_provider_t));
7874 * Invalidate the specified provider. All subsequent probe lookups for the
7875 * specified provider will fail, but its probes will not be removed.
7878 dtrace_invalidate(dtrace_provider_id_t id)
7880 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
7882 ASSERT(pvp->dtpv_pops.dtps_enable !=
7883 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7885 mutex_enter(&dtrace_provider_lock);
7886 mutex_enter(&dtrace_lock);
7888 pvp->dtpv_defunct = dtrace_gethrtime();
7890 mutex_exit(&dtrace_lock);
7891 mutex_exit(&dtrace_provider_lock);
7895 * Indicate whether or not DTrace has attached.
7898 dtrace_attached(void)
7901 * dtrace_provider will be non-NULL iff the DTrace driver has
7902 * attached. (It's non-NULL because DTrace is always itself a
7905 return (dtrace_provider != NULL);
7909 * Remove all the unenabled probes for the given provider. This function is
7910 * not unlike dtrace_unregister(), except that it doesn't remove the provider
7911 * -- just as many of its associated probes as it can.
7914 dtrace_condense(dtrace_provider_id_t id)
7916 dtrace_provider_t *prov = (dtrace_provider_t *)id;
7918 dtrace_probe_t *probe;
7921 * Make sure this isn't the dtrace provider itself.
7923 ASSERT(prov->dtpv_pops.dtps_enable !=
7924 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
7926 mutex_enter(&dtrace_provider_lock);
7927 mutex_enter(&dtrace_lock);
7930 * Attempt to destroy the probes associated with this provider.
7932 for (i = 0; i < dtrace_nprobes; i++) {
7933 if ((probe = dtrace_probes[i]) == NULL)
7936 if (probe->dtpr_provider != prov)
7939 if (probe->dtpr_ecb != NULL)
7942 dtrace_probes[i] = NULL;
7944 dtrace_hash_remove(dtrace_bymod, probe);
7945 dtrace_hash_remove(dtrace_byfunc, probe);
7946 dtrace_hash_remove(dtrace_byname, probe);
7948 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
7950 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
7951 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
7952 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
7953 kmem_free(probe, sizeof (dtrace_probe_t));
7955 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
7957 free_unr(dtrace_arena, i + 1);
7961 mutex_exit(&dtrace_lock);
7962 mutex_exit(&dtrace_provider_lock);
7968 * DTrace Probe Management Functions
7970 * The functions in this section perform the DTrace probe management,
7971 * including functions to create probes, look-up probes, and call into the
7972 * providers to request that probes be provided. Some of these functions are
7973 * in the Provider-to-Framework API; these functions can be identified by the
7974 * fact that they are not declared "static".
7978 * Create a probe with the specified module name, function name, and name.
7981 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
7982 const char *func, const char *name, int aframes, void *arg)
7984 dtrace_probe_t *probe, **probes;
7985 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
7988 if (provider == dtrace_provider) {
7989 ASSERT(MUTEX_HELD(&dtrace_lock));
7991 mutex_enter(&dtrace_lock);
7995 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
7996 VM_BESTFIT | VM_SLEEP);
7998 id = alloc_unr(dtrace_arena);
8000 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
8002 probe->dtpr_id = id;
8003 probe->dtpr_gen = dtrace_probegen++;
8004 probe->dtpr_mod = dtrace_strdup(mod);
8005 probe->dtpr_func = dtrace_strdup(func);
8006 probe->dtpr_name = dtrace_strdup(name);
8007 probe->dtpr_arg = arg;
8008 probe->dtpr_aframes = aframes;
8009 probe->dtpr_provider = provider;
8011 dtrace_hash_add(dtrace_bymod, probe);
8012 dtrace_hash_add(dtrace_byfunc, probe);
8013 dtrace_hash_add(dtrace_byname, probe);
8015 if (id - 1 >= dtrace_nprobes) {
8016 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
8017 size_t nsize = osize << 1;
8021 ASSERT(dtrace_probes == NULL);
8022 nsize = sizeof (dtrace_probe_t *);
8025 probes = kmem_zalloc(nsize, KM_SLEEP);
8027 if (dtrace_probes == NULL) {
8029 dtrace_probes = probes;
8032 dtrace_probe_t **oprobes = dtrace_probes;
8034 bcopy(oprobes, probes, osize);
8035 dtrace_membar_producer();
8036 dtrace_probes = probes;
8041 * All CPUs are now seeing the new probes array; we can
8042 * safely free the old array.
8044 kmem_free(oprobes, osize);
8045 dtrace_nprobes <<= 1;
8048 ASSERT(id - 1 < dtrace_nprobes);
8051 ASSERT(dtrace_probes[id - 1] == NULL);
8052 dtrace_probes[id - 1] = probe;
8054 if (provider != dtrace_provider)
8055 mutex_exit(&dtrace_lock);
8060 static dtrace_probe_t *
8061 dtrace_probe_lookup_id(dtrace_id_t id)
8063 ASSERT(MUTEX_HELD(&dtrace_lock));
8065 if (id == 0 || id > dtrace_nprobes)
8068 return (dtrace_probes[id - 1]);
8072 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
8074 *((dtrace_id_t *)arg) = probe->dtpr_id;
8076 return (DTRACE_MATCH_DONE);
8080 * Look up a probe based on provider and one or more of module name, function
8081 * name and probe name.
8084 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
8085 char *func, char *name)
8087 dtrace_probekey_t pkey;
8091 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
8092 pkey.dtpk_pmatch = &dtrace_match_string;
8093 pkey.dtpk_mod = mod;
8094 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
8095 pkey.dtpk_func = func;
8096 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
8097 pkey.dtpk_name = name;
8098 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
8099 pkey.dtpk_id = DTRACE_IDNONE;
8101 mutex_enter(&dtrace_lock);
8102 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
8103 dtrace_probe_lookup_match, &id);
8104 mutex_exit(&dtrace_lock);
8106 ASSERT(match == 1 || match == 0);
8107 return (match ? id : 0);
8111 * Returns the probe argument associated with the specified probe.
8114 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
8116 dtrace_probe_t *probe;
8119 mutex_enter(&dtrace_lock);
8121 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
8122 probe->dtpr_provider == (dtrace_provider_t *)id)
8123 rval = probe->dtpr_arg;
8125 mutex_exit(&dtrace_lock);
8131 * Copy a probe into a probe description.
8134 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
8136 bzero(pdp, sizeof (dtrace_probedesc_t));
8137 pdp->dtpd_id = prp->dtpr_id;
8139 (void) strncpy(pdp->dtpd_provider,
8140 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
8142 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
8143 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
8144 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
8148 * Called to indicate that a probe -- or probes -- should be provided by a
8149 * specfied provider. If the specified description is NULL, the provider will
8150 * be told to provide all of its probes. (This is done whenever a new
8151 * consumer comes along, or whenever a retained enabling is to be matched.) If
8152 * the specified description is non-NULL, the provider is given the
8153 * opportunity to dynamically provide the specified probe, allowing providers
8154 * to support the creation of probes on-the-fly. (So-called _autocreated_
8155 * probes.) If the provider is NULL, the operations will be applied to all
8156 * providers; if the provider is non-NULL the operations will only be applied
8157 * to the specified provider. The dtrace_provider_lock must be held, and the
8158 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
8159 * will need to grab the dtrace_lock when it reenters the framework through
8160 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
8163 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
8170 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8174 prv = dtrace_provider;
8179 * First, call the blanket provide operation.
8181 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
8185 * Now call the per-module provide operation. We will grab
8186 * mod_lock to prevent the list from being modified. Note
8187 * that this also prevents the mod_busy bits from changing.
8188 * (mod_busy can only be changed with mod_lock held.)
8190 mutex_enter(&mod_lock);
8194 if (ctl->mod_busy || ctl->mod_mp == NULL)
8197 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
8199 } while ((ctl = ctl->mod_next) != &modules);
8201 mutex_exit(&mod_lock);
8203 } while (all && (prv = prv->dtpv_next) != NULL);
8208 * Iterate over each probe, and call the Framework-to-Provider API function
8212 dtrace_probe_foreach(uintptr_t offs)
8214 dtrace_provider_t *prov;
8215 void (*func)(void *, dtrace_id_t, void *);
8216 dtrace_probe_t *probe;
8217 dtrace_icookie_t cookie;
8221 * We disable interrupts to walk through the probe array. This is
8222 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
8223 * won't see stale data.
8225 cookie = dtrace_interrupt_disable();
8227 for (i = 0; i < dtrace_nprobes; i++) {
8228 if ((probe = dtrace_probes[i]) == NULL)
8231 if (probe->dtpr_ecb == NULL) {
8233 * This probe isn't enabled -- don't call the function.
8238 prov = probe->dtpr_provider;
8239 func = *((void(**)(void *, dtrace_id_t, void *))
8240 ((uintptr_t)&prov->dtpv_pops + offs));
8242 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
8245 dtrace_interrupt_enable(cookie);
8250 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
8252 dtrace_probekey_t pkey;
8257 ASSERT(MUTEX_HELD(&dtrace_lock));
8258 dtrace_ecb_create_cache = NULL;
8262 * If we're passed a NULL description, we're being asked to
8263 * create an ECB with a NULL probe.
8265 (void) dtrace_ecb_create_enable(NULL, enab);
8269 dtrace_probekey(desc, &pkey);
8270 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
8271 &priv, &uid, &zoneid);
8273 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
8278 * DTrace Helper Provider Functions
8281 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
8283 attr->dtat_name = DOF_ATTR_NAME(dofattr);
8284 attr->dtat_data = DOF_ATTR_DATA(dofattr);
8285 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
8289 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
8290 const dof_provider_t *dofprov, char *strtab)
8292 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
8293 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
8294 dofprov->dofpv_provattr);
8295 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
8296 dofprov->dofpv_modattr);
8297 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
8298 dofprov->dofpv_funcattr);
8299 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
8300 dofprov->dofpv_nameattr);
8301 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
8302 dofprov->dofpv_argsattr);
8306 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8308 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8309 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8310 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
8311 dof_provider_t *provider;
8313 uint32_t *off, *enoff;
8317 dtrace_helper_provdesc_t dhpv;
8318 dtrace_helper_probedesc_t dhpb;
8319 dtrace_meta_t *meta = dtrace_meta_pid;
8320 dtrace_mops_t *mops = &meta->dtm_mops;
8323 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8324 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8325 provider->dofpv_strtab * dof->dofh_secsize);
8326 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8327 provider->dofpv_probes * dof->dofh_secsize);
8328 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8329 provider->dofpv_prargs * dof->dofh_secsize);
8330 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8331 provider->dofpv_proffs * dof->dofh_secsize);
8333 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8334 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
8335 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
8339 * See dtrace_helper_provider_validate().
8341 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
8342 provider->dofpv_prenoffs != DOF_SECT_NONE) {
8343 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8344 provider->dofpv_prenoffs * dof->dofh_secsize);
8345 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
8348 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
8351 * Create the provider.
8353 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8355 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
8361 * Create the probes.
8363 for (i = 0; i < nprobes; i++) {
8364 probe = (dof_probe_t *)(uintptr_t)(daddr +
8365 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
8367 dhpb.dthpb_mod = dhp->dofhp_mod;
8368 dhpb.dthpb_func = strtab + probe->dofpr_func;
8369 dhpb.dthpb_name = strtab + probe->dofpr_name;
8370 dhpb.dthpb_base = probe->dofpr_addr;
8371 dhpb.dthpb_offs = off + probe->dofpr_offidx;
8372 dhpb.dthpb_noffs = probe->dofpr_noffs;
8373 if (enoff != NULL) {
8374 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
8375 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
8377 dhpb.dthpb_enoffs = NULL;
8378 dhpb.dthpb_nenoffs = 0;
8380 dhpb.dthpb_args = arg + probe->dofpr_argidx;
8381 dhpb.dthpb_nargc = probe->dofpr_nargc;
8382 dhpb.dthpb_xargc = probe->dofpr_xargc;
8383 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
8384 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
8386 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
8391 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
8393 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8394 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8397 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8399 for (i = 0; i < dof->dofh_secnum; i++) {
8400 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8401 dof->dofh_secoff + i * dof->dofh_secsize);
8403 if (sec->dofs_type != DOF_SECT_PROVIDER)
8406 dtrace_helper_provide_one(dhp, sec, pid);
8410 * We may have just created probes, so we must now rematch against
8411 * any retained enablings. Note that this call will acquire both
8412 * cpu_lock and dtrace_lock; the fact that we are holding
8413 * dtrace_meta_lock now is what defines the ordering with respect to
8414 * these three locks.
8416 dtrace_enabling_matchall();
8420 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
8422 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8423 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8425 dof_provider_t *provider;
8427 dtrace_helper_provdesc_t dhpv;
8428 dtrace_meta_t *meta = dtrace_meta_pid;
8429 dtrace_mops_t *mops = &meta->dtm_mops;
8431 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
8432 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
8433 provider->dofpv_strtab * dof->dofh_secsize);
8435 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
8438 * Create the provider.
8440 dtrace_dofprov2hprov(&dhpv, provider, strtab);
8442 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
8448 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
8450 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
8451 dof_hdr_t *dof = (dof_hdr_t *)daddr;
8454 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
8456 for (i = 0; i < dof->dofh_secnum; i++) {
8457 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
8458 dof->dofh_secoff + i * dof->dofh_secsize);
8460 if (sec->dofs_type != DOF_SECT_PROVIDER)
8463 dtrace_helper_provider_remove_one(dhp, sec, pid);
8468 * DTrace Meta Provider-to-Framework API Functions
8470 * These functions implement the Meta Provider-to-Framework API, as described
8471 * in <sys/dtrace.h>.
8474 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
8475 dtrace_meta_provider_id_t *idp)
8477 dtrace_meta_t *meta;
8478 dtrace_helpers_t *help, *next;
8481 *idp = DTRACE_METAPROVNONE;
8484 * We strictly don't need the name, but we hold onto it for
8485 * debuggability. All hail error queues!
8488 cmn_err(CE_WARN, "failed to register meta-provider: "
8494 mops->dtms_create_probe == NULL ||
8495 mops->dtms_provide_pid == NULL ||
8496 mops->dtms_remove_pid == NULL) {
8497 cmn_err(CE_WARN, "failed to register meta-register %s: "
8498 "invalid ops", name);
8502 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
8503 meta->dtm_mops = *mops;
8504 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8505 (void) strcpy(meta->dtm_name, name);
8506 meta->dtm_arg = arg;
8508 mutex_enter(&dtrace_meta_lock);
8509 mutex_enter(&dtrace_lock);
8511 if (dtrace_meta_pid != NULL) {
8512 mutex_exit(&dtrace_lock);
8513 mutex_exit(&dtrace_meta_lock);
8514 cmn_err(CE_WARN, "failed to register meta-register %s: "
8515 "user-land meta-provider exists", name);
8516 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
8517 kmem_free(meta, sizeof (dtrace_meta_t));
8521 dtrace_meta_pid = meta;
8522 *idp = (dtrace_meta_provider_id_t)meta;
8525 * If there are providers and probes ready to go, pass them
8526 * off to the new meta provider now.
8529 help = dtrace_deferred_pid;
8530 dtrace_deferred_pid = NULL;
8532 mutex_exit(&dtrace_lock);
8534 while (help != NULL) {
8535 for (i = 0; i < help->dthps_nprovs; i++) {
8536 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
8540 next = help->dthps_next;
8541 help->dthps_next = NULL;
8542 help->dthps_prev = NULL;
8543 help->dthps_deferred = 0;
8547 mutex_exit(&dtrace_meta_lock);
8553 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
8555 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
8557 mutex_enter(&dtrace_meta_lock);
8558 mutex_enter(&dtrace_lock);
8560 if (old == dtrace_meta_pid) {
8561 pp = &dtrace_meta_pid;
8563 panic("attempt to unregister non-existent "
8564 "dtrace meta-provider %p\n", (void *)old);
8567 if (old->dtm_count != 0) {
8568 mutex_exit(&dtrace_lock);
8569 mutex_exit(&dtrace_meta_lock);
8575 mutex_exit(&dtrace_lock);
8576 mutex_exit(&dtrace_meta_lock);
8578 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
8579 kmem_free(old, sizeof (dtrace_meta_t));
8586 * DTrace DIF Object Functions
8589 dtrace_difo_err(uint_t pc, const char *format, ...)
8591 if (dtrace_err_verbose) {
8594 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
8595 va_start(alist, format);
8596 (void) vuprintf(format, alist);
8600 #ifdef DTRACE_ERRDEBUG
8601 dtrace_errdebug(format);
8607 * Validate a DTrace DIF object by checking the IR instructions. The following
8608 * rules are currently enforced by dtrace_difo_validate():
8610 * 1. Each instruction must have a valid opcode
8611 * 2. Each register, string, variable, or subroutine reference must be valid
8612 * 3. No instruction can modify register %r0 (must be zero)
8613 * 4. All instruction reserved bits must be set to zero
8614 * 5. The last instruction must be a "ret" instruction
8615 * 6. All branch targets must reference a valid instruction _after_ the branch
8618 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
8622 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
8626 kcheckload = cr == NULL ||
8627 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
8629 dp->dtdo_destructive = 0;
8631 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
8632 dif_instr_t instr = dp->dtdo_buf[pc];
8634 uint_t r1 = DIF_INSTR_R1(instr);
8635 uint_t r2 = DIF_INSTR_R2(instr);
8636 uint_t rd = DIF_INSTR_RD(instr);
8637 uint_t rs = DIF_INSTR_RS(instr);
8638 uint_t label = DIF_INSTR_LABEL(instr);
8639 uint_t v = DIF_INSTR_VAR(instr);
8640 uint_t subr = DIF_INSTR_SUBR(instr);
8641 uint_t type = DIF_INSTR_TYPE(instr);
8642 uint_t op = DIF_INSTR_OP(instr);
8660 err += efunc(pc, "invalid register %u\n", r1);
8662 err += efunc(pc, "invalid register %u\n", r2);
8664 err += efunc(pc, "invalid register %u\n", rd);
8666 err += efunc(pc, "cannot write to %r0\n");
8672 err += efunc(pc, "invalid register %u\n", r1);
8674 err += efunc(pc, "non-zero reserved bits\n");
8676 err += efunc(pc, "invalid register %u\n", rd);
8678 err += efunc(pc, "cannot write to %r0\n");
8688 err += efunc(pc, "invalid register %u\n", r1);
8690 err += efunc(pc, "non-zero reserved bits\n");
8692 err += efunc(pc, "invalid register %u\n", rd);
8694 err += efunc(pc, "cannot write to %r0\n");
8696 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
8697 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
8707 err += efunc(pc, "invalid register %u\n", r1);
8709 err += efunc(pc, "non-zero reserved bits\n");
8711 err += efunc(pc, "invalid register %u\n", rd);
8713 err += efunc(pc, "cannot write to %r0\n");
8723 err += efunc(pc, "invalid register %u\n", r1);
8725 err += efunc(pc, "non-zero reserved bits\n");
8727 err += efunc(pc, "invalid register %u\n", rd);
8729 err += efunc(pc, "cannot write to %r0\n");
8736 err += efunc(pc, "invalid register %u\n", r1);
8738 err += efunc(pc, "non-zero reserved bits\n");
8740 err += efunc(pc, "invalid register %u\n", rd);
8742 err += efunc(pc, "cannot write to 0 address\n");
8747 err += efunc(pc, "invalid register %u\n", r1);
8749 err += efunc(pc, "invalid register %u\n", r2);
8751 err += efunc(pc, "non-zero reserved bits\n");
8755 err += efunc(pc, "invalid register %u\n", r1);
8756 if (r2 != 0 || rd != 0)
8757 err += efunc(pc, "non-zero reserved bits\n");
8770 if (label >= dp->dtdo_len) {
8771 err += efunc(pc, "invalid branch target %u\n",
8775 err += efunc(pc, "backward branch to %u\n",
8780 if (r1 != 0 || r2 != 0)
8781 err += efunc(pc, "non-zero reserved bits\n");
8783 err += efunc(pc, "invalid register %u\n", rd);
8787 case DIF_OP_FLUSHTS:
8788 if (r1 != 0 || r2 != 0 || rd != 0)
8789 err += efunc(pc, "non-zero reserved bits\n");
8792 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
8793 err += efunc(pc, "invalid integer ref %u\n",
8794 DIF_INSTR_INTEGER(instr));
8797 err += efunc(pc, "invalid register %u\n", rd);
8799 err += efunc(pc, "cannot write to %r0\n");
8802 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
8803 err += efunc(pc, "invalid string ref %u\n",
8804 DIF_INSTR_STRING(instr));
8807 err += efunc(pc, "invalid register %u\n", rd);
8809 err += efunc(pc, "cannot write to %r0\n");
8813 if (r1 > DIF_VAR_ARRAY_MAX)
8814 err += efunc(pc, "invalid array %u\n", r1);
8816 err += efunc(pc, "invalid register %u\n", r2);
8818 err += efunc(pc, "invalid register %u\n", rd);
8820 err += efunc(pc, "cannot write to %r0\n");
8827 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
8828 err += efunc(pc, "invalid variable %u\n", v);
8830 err += efunc(pc, "invalid register %u\n", rd);
8832 err += efunc(pc, "cannot write to %r0\n");
8839 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
8840 err += efunc(pc, "invalid variable %u\n", v);
8842 err += efunc(pc, "invalid register %u\n", rd);
8845 if (subr > DIF_SUBR_MAX)
8846 err += efunc(pc, "invalid subr %u\n", subr);
8848 err += efunc(pc, "invalid register %u\n", rd);
8850 err += efunc(pc, "cannot write to %r0\n");
8852 if (subr == DIF_SUBR_COPYOUT ||
8853 subr == DIF_SUBR_COPYOUTSTR) {
8854 dp->dtdo_destructive = 1;
8858 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
8859 err += efunc(pc, "invalid ref type %u\n", type);
8861 err += efunc(pc, "invalid register %u\n", r2);
8863 err += efunc(pc, "invalid register %u\n", rs);
8866 if (type != DIF_TYPE_CTF)
8867 err += efunc(pc, "invalid val type %u\n", type);
8869 err += efunc(pc, "invalid register %u\n", r2);
8871 err += efunc(pc, "invalid register %u\n", rs);
8874 err += efunc(pc, "invalid opcode %u\n",
8875 DIF_INSTR_OP(instr));
8879 if (dp->dtdo_len != 0 &&
8880 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
8881 err += efunc(dp->dtdo_len - 1,
8882 "expected 'ret' as last DIF instruction\n");
8885 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
8887 * If we're not returning by reference, the size must be either
8888 * 0 or the size of one of the base types.
8890 switch (dp->dtdo_rtype.dtdt_size) {
8892 case sizeof (uint8_t):
8893 case sizeof (uint16_t):
8894 case sizeof (uint32_t):
8895 case sizeof (uint64_t):
8899 err += efunc(dp->dtdo_len - 1, "bad return size");
8903 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
8904 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
8905 dtrace_diftype_t *vt, *et;
8908 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
8909 v->dtdv_scope != DIFV_SCOPE_THREAD &&
8910 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
8911 err += efunc(i, "unrecognized variable scope %d\n",
8916 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
8917 v->dtdv_kind != DIFV_KIND_SCALAR) {
8918 err += efunc(i, "unrecognized variable type %d\n",
8923 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
8924 err += efunc(i, "%d exceeds variable id limit\n", id);
8928 if (id < DIF_VAR_OTHER_UBASE)
8932 * For user-defined variables, we need to check that this
8933 * definition is identical to any previous definition that we
8936 ndx = id - DIF_VAR_OTHER_UBASE;
8938 switch (v->dtdv_scope) {
8939 case DIFV_SCOPE_GLOBAL:
8940 if (ndx < vstate->dtvs_nglobals) {
8941 dtrace_statvar_t *svar;
8943 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
8944 existing = &svar->dtsv_var;
8949 case DIFV_SCOPE_THREAD:
8950 if (ndx < vstate->dtvs_ntlocals)
8951 existing = &vstate->dtvs_tlocals[ndx];
8954 case DIFV_SCOPE_LOCAL:
8955 if (ndx < vstate->dtvs_nlocals) {
8956 dtrace_statvar_t *svar;
8958 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
8959 existing = &svar->dtsv_var;
8967 if (vt->dtdt_flags & DIF_TF_BYREF) {
8968 if (vt->dtdt_size == 0) {
8969 err += efunc(i, "zero-sized variable\n");
8973 if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
8974 vt->dtdt_size > dtrace_global_maxsize) {
8975 err += efunc(i, "oversized by-ref global\n");
8980 if (existing == NULL || existing->dtdv_id == 0)
8983 ASSERT(existing->dtdv_id == v->dtdv_id);
8984 ASSERT(existing->dtdv_scope == v->dtdv_scope);
8986 if (existing->dtdv_kind != v->dtdv_kind)
8987 err += efunc(i, "%d changed variable kind\n", id);
8989 et = &existing->dtdv_type;
8991 if (vt->dtdt_flags != et->dtdt_flags) {
8992 err += efunc(i, "%d changed variable type flags\n", id);
8996 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
8997 err += efunc(i, "%d changed variable type size\n", id);
9006 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
9007 * are much more constrained than normal DIFOs. Specifically, they may
9010 * 1. Make calls to subroutines other than copyin(), copyinstr() or
9011 * miscellaneous string routines
9012 * 2. Access DTrace variables other than the args[] array, and the
9013 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
9014 * 3. Have thread-local variables.
9015 * 4. Have dynamic variables.
9018 dtrace_difo_validate_helper(dtrace_difo_t *dp)
9020 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9024 for (pc = 0; pc < dp->dtdo_len; pc++) {
9025 dif_instr_t instr = dp->dtdo_buf[pc];
9027 uint_t v = DIF_INSTR_VAR(instr);
9028 uint_t subr = DIF_INSTR_SUBR(instr);
9029 uint_t op = DIF_INSTR_OP(instr);
9084 case DIF_OP_FLUSHTS:
9096 if (v >= DIF_VAR_OTHER_UBASE)
9099 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
9102 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
9103 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
9104 v == DIF_VAR_EXECARGS ||
9105 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
9106 v == DIF_VAR_UID || v == DIF_VAR_GID)
9109 err += efunc(pc, "illegal variable %u\n", v);
9116 err += efunc(pc, "illegal dynamic variable load\n");
9122 err += efunc(pc, "illegal dynamic variable store\n");
9126 if (subr == DIF_SUBR_ALLOCA ||
9127 subr == DIF_SUBR_BCOPY ||
9128 subr == DIF_SUBR_COPYIN ||
9129 subr == DIF_SUBR_COPYINTO ||
9130 subr == DIF_SUBR_COPYINSTR ||
9131 subr == DIF_SUBR_INDEX ||
9132 subr == DIF_SUBR_INET_NTOA ||
9133 subr == DIF_SUBR_INET_NTOA6 ||
9134 subr == DIF_SUBR_INET_NTOP ||
9135 subr == DIF_SUBR_LLTOSTR ||
9136 subr == DIF_SUBR_RINDEX ||
9137 subr == DIF_SUBR_STRCHR ||
9138 subr == DIF_SUBR_STRJOIN ||
9139 subr == DIF_SUBR_STRRCHR ||
9140 subr == DIF_SUBR_STRSTR ||
9141 subr == DIF_SUBR_HTONS ||
9142 subr == DIF_SUBR_HTONL ||
9143 subr == DIF_SUBR_HTONLL ||
9144 subr == DIF_SUBR_NTOHS ||
9145 subr == DIF_SUBR_NTOHL ||
9146 subr == DIF_SUBR_NTOHLL ||
9147 subr == DIF_SUBR_MEMREF ||
9149 subr == DIF_SUBR_MEMSTR ||
9151 subr == DIF_SUBR_TYPEREF)
9154 err += efunc(pc, "invalid subr %u\n", subr);
9158 err += efunc(pc, "invalid opcode %u\n",
9159 DIF_INSTR_OP(instr));
9167 * Returns 1 if the expression in the DIF object can be cached on a per-thread
9171 dtrace_difo_cacheable(dtrace_difo_t *dp)
9178 for (i = 0; i < dp->dtdo_varlen; i++) {
9179 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9181 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
9184 switch (v->dtdv_id) {
9185 case DIF_VAR_CURTHREAD:
9188 case DIF_VAR_EXECARGS:
9189 case DIF_VAR_EXECNAME:
9190 case DIF_VAR_ZONENAME:
9199 * This DIF object may be cacheable. Now we need to look for any
9200 * array loading instructions, any memory loading instructions, or
9201 * any stores to thread-local variables.
9203 for (i = 0; i < dp->dtdo_len; i++) {
9204 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
9206 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
9207 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
9208 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
9209 op == DIF_OP_LDGA || op == DIF_OP_STTS)
9217 dtrace_difo_hold(dtrace_difo_t *dp)
9221 ASSERT(MUTEX_HELD(&dtrace_lock));
9224 ASSERT(dp->dtdo_refcnt != 0);
9227 * We need to check this DIF object for references to the variable
9228 * DIF_VAR_VTIMESTAMP.
9230 for (i = 0; i < dp->dtdo_varlen; i++) {
9231 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9233 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9236 if (dtrace_vtime_references++ == 0)
9237 dtrace_vtime_enable();
9242 * This routine calculates the dynamic variable chunksize for a given DIF
9243 * object. The calculation is not fool-proof, and can probably be tricked by
9244 * malicious DIF -- but it works for all compiler-generated DIF. Because this
9245 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
9246 * if a dynamic variable size exceeds the chunksize.
9249 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9252 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
9253 const dif_instr_t *text = dp->dtdo_buf;
9259 for (pc = 0; pc < dp->dtdo_len; pc++) {
9260 dif_instr_t instr = text[pc];
9261 uint_t op = DIF_INSTR_OP(instr);
9262 uint_t rd = DIF_INSTR_RD(instr);
9263 uint_t r1 = DIF_INSTR_R1(instr);
9267 dtrace_key_t *key = tupregs;
9271 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
9276 key = &tupregs[DIF_DTR_NREGS];
9277 key[0].dttk_size = 0;
9278 key[1].dttk_size = 0;
9280 scope = DIFV_SCOPE_THREAD;
9287 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
9288 key[nkeys++].dttk_size = 0;
9290 key[nkeys++].dttk_size = 0;
9292 if (op == DIF_OP_STTAA) {
9293 scope = DIFV_SCOPE_THREAD;
9295 scope = DIFV_SCOPE_GLOBAL;
9301 if (ttop == DIF_DTR_NREGS)
9304 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
9306 * If the register for the size of the "pushtr"
9307 * is %r0 (or the value is 0) and the type is
9308 * a string, we'll use the system-wide default
9311 tupregs[ttop++].dttk_size =
9312 dtrace_strsize_default;
9317 tupregs[ttop++].dttk_size = sval;
9323 if (ttop == DIF_DTR_NREGS)
9326 tupregs[ttop++].dttk_size = 0;
9329 case DIF_OP_FLUSHTS:
9346 * We have a dynamic variable allocation; calculate its size.
9348 for (ksize = 0, i = 0; i < nkeys; i++)
9349 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
9351 size = sizeof (dtrace_dynvar_t);
9352 size += sizeof (dtrace_key_t) * (nkeys - 1);
9356 * Now we need to determine the size of the stored data.
9358 id = DIF_INSTR_VAR(instr);
9360 for (i = 0; i < dp->dtdo_varlen; i++) {
9361 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9363 if (v->dtdv_id == id && v->dtdv_scope == scope) {
9364 size += v->dtdv_type.dtdt_size;
9369 if (i == dp->dtdo_varlen)
9373 * We have the size. If this is larger than the chunk size
9374 * for our dynamic variable state, reset the chunk size.
9376 size = P2ROUNDUP(size, sizeof (uint64_t));
9378 if (size > vstate->dtvs_dynvars.dtds_chunksize)
9379 vstate->dtvs_dynvars.dtds_chunksize = size;
9384 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9386 int i, oldsvars, osz, nsz, otlocals, ntlocals;
9389 ASSERT(MUTEX_HELD(&dtrace_lock));
9390 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
9392 for (i = 0; i < dp->dtdo_varlen; i++) {
9393 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9394 dtrace_statvar_t *svar, ***svarp = NULL;
9396 uint8_t scope = v->dtdv_scope;
9399 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9402 id -= DIF_VAR_OTHER_UBASE;
9405 case DIFV_SCOPE_THREAD:
9406 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
9407 dtrace_difv_t *tlocals;
9409 if ((ntlocals = (otlocals << 1)) == 0)
9412 osz = otlocals * sizeof (dtrace_difv_t);
9413 nsz = ntlocals * sizeof (dtrace_difv_t);
9415 tlocals = kmem_zalloc(nsz, KM_SLEEP);
9418 bcopy(vstate->dtvs_tlocals,
9420 kmem_free(vstate->dtvs_tlocals, osz);
9423 vstate->dtvs_tlocals = tlocals;
9424 vstate->dtvs_ntlocals = ntlocals;
9427 vstate->dtvs_tlocals[id] = *v;
9430 case DIFV_SCOPE_LOCAL:
9431 np = &vstate->dtvs_nlocals;
9432 svarp = &vstate->dtvs_locals;
9434 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9435 dsize = NCPU * (v->dtdv_type.dtdt_size +
9438 dsize = NCPU * sizeof (uint64_t);
9442 case DIFV_SCOPE_GLOBAL:
9443 np = &vstate->dtvs_nglobals;
9444 svarp = &vstate->dtvs_globals;
9446 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
9447 dsize = v->dtdv_type.dtdt_size +
9456 while (id >= (oldsvars = *np)) {
9457 dtrace_statvar_t **statics;
9458 int newsvars, oldsize, newsize;
9460 if ((newsvars = (oldsvars << 1)) == 0)
9463 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
9464 newsize = newsvars * sizeof (dtrace_statvar_t *);
9466 statics = kmem_zalloc(newsize, KM_SLEEP);
9469 bcopy(*svarp, statics, oldsize);
9470 kmem_free(*svarp, oldsize);
9477 if ((svar = (*svarp)[id]) == NULL) {
9478 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
9479 svar->dtsv_var = *v;
9481 if ((svar->dtsv_size = dsize) != 0) {
9482 svar->dtsv_data = (uint64_t)(uintptr_t)
9483 kmem_zalloc(dsize, KM_SLEEP);
9486 (*svarp)[id] = svar;
9489 svar->dtsv_refcnt++;
9492 dtrace_difo_chunksize(dp, vstate);
9493 dtrace_difo_hold(dp);
9496 static dtrace_difo_t *
9497 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9502 ASSERT(dp->dtdo_buf != NULL);
9503 ASSERT(dp->dtdo_refcnt != 0);
9505 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
9507 ASSERT(dp->dtdo_buf != NULL);
9508 sz = dp->dtdo_len * sizeof (dif_instr_t);
9509 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
9510 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
9511 new->dtdo_len = dp->dtdo_len;
9513 if (dp->dtdo_strtab != NULL) {
9514 ASSERT(dp->dtdo_strlen != 0);
9515 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
9516 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
9517 new->dtdo_strlen = dp->dtdo_strlen;
9520 if (dp->dtdo_inttab != NULL) {
9521 ASSERT(dp->dtdo_intlen != 0);
9522 sz = dp->dtdo_intlen * sizeof (uint64_t);
9523 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
9524 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
9525 new->dtdo_intlen = dp->dtdo_intlen;
9528 if (dp->dtdo_vartab != NULL) {
9529 ASSERT(dp->dtdo_varlen != 0);
9530 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
9531 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
9532 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
9533 new->dtdo_varlen = dp->dtdo_varlen;
9536 dtrace_difo_init(new, vstate);
9541 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9545 ASSERT(dp->dtdo_refcnt == 0);
9547 for (i = 0; i < dp->dtdo_varlen; i++) {
9548 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9549 dtrace_statvar_t *svar, **svarp = NULL;
9551 uint8_t scope = v->dtdv_scope;
9555 case DIFV_SCOPE_THREAD:
9558 case DIFV_SCOPE_LOCAL:
9559 np = &vstate->dtvs_nlocals;
9560 svarp = vstate->dtvs_locals;
9563 case DIFV_SCOPE_GLOBAL:
9564 np = &vstate->dtvs_nglobals;
9565 svarp = vstate->dtvs_globals;
9572 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
9575 id -= DIF_VAR_OTHER_UBASE;
9579 ASSERT(svar != NULL);
9580 ASSERT(svar->dtsv_refcnt > 0);
9582 if (--svar->dtsv_refcnt > 0)
9585 if (svar->dtsv_size != 0) {
9586 ASSERT(svar->dtsv_data != 0);
9587 kmem_free((void *)(uintptr_t)svar->dtsv_data,
9591 kmem_free(svar, sizeof (dtrace_statvar_t));
9595 if (dp->dtdo_buf != NULL)
9596 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
9597 if (dp->dtdo_inttab != NULL)
9598 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
9599 if (dp->dtdo_strtab != NULL)
9600 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
9601 if (dp->dtdo_vartab != NULL)
9602 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
9604 kmem_free(dp, sizeof (dtrace_difo_t));
9608 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
9612 ASSERT(MUTEX_HELD(&dtrace_lock));
9613 ASSERT(dp->dtdo_refcnt != 0);
9615 for (i = 0; i < dp->dtdo_varlen; i++) {
9616 dtrace_difv_t *v = &dp->dtdo_vartab[i];
9618 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
9621 ASSERT(dtrace_vtime_references > 0);
9622 if (--dtrace_vtime_references == 0)
9623 dtrace_vtime_disable();
9626 if (--dp->dtdo_refcnt == 0)
9627 dtrace_difo_destroy(dp, vstate);
9631 * DTrace Format Functions
9634 dtrace_format_add(dtrace_state_t *state, char *str)
9637 uint16_t ndx, len = strlen(str) + 1;
9639 fmt = kmem_zalloc(len, KM_SLEEP);
9640 bcopy(str, fmt, len);
9642 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
9643 if (state->dts_formats[ndx] == NULL) {
9644 state->dts_formats[ndx] = fmt;
9649 if (state->dts_nformats == USHRT_MAX) {
9651 * This is only likely if a denial-of-service attack is being
9652 * attempted. As such, it's okay to fail silently here.
9654 kmem_free(fmt, len);
9659 * For simplicity, we always resize the formats array to be exactly the
9660 * number of formats.
9662 ndx = state->dts_nformats++;
9663 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
9665 if (state->dts_formats != NULL) {
9667 bcopy(state->dts_formats, new, ndx * sizeof (char *));
9668 kmem_free(state->dts_formats, ndx * sizeof (char *));
9671 state->dts_formats = new;
9672 state->dts_formats[ndx] = fmt;
9678 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
9682 ASSERT(state->dts_formats != NULL);
9683 ASSERT(format <= state->dts_nformats);
9684 ASSERT(state->dts_formats[format - 1] != NULL);
9686 fmt = state->dts_formats[format - 1];
9687 kmem_free(fmt, strlen(fmt) + 1);
9688 state->dts_formats[format - 1] = NULL;
9692 dtrace_format_destroy(dtrace_state_t *state)
9696 if (state->dts_nformats == 0) {
9697 ASSERT(state->dts_formats == NULL);
9701 ASSERT(state->dts_formats != NULL);
9703 for (i = 0; i < state->dts_nformats; i++) {
9704 char *fmt = state->dts_formats[i];
9709 kmem_free(fmt, strlen(fmt) + 1);
9712 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
9713 state->dts_nformats = 0;
9714 state->dts_formats = NULL;
9718 * DTrace Predicate Functions
9720 static dtrace_predicate_t *
9721 dtrace_predicate_create(dtrace_difo_t *dp)
9723 dtrace_predicate_t *pred;
9725 ASSERT(MUTEX_HELD(&dtrace_lock));
9726 ASSERT(dp->dtdo_refcnt != 0);
9728 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
9729 pred->dtp_difo = dp;
9730 pred->dtp_refcnt = 1;
9732 if (!dtrace_difo_cacheable(dp))
9735 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
9737 * This is only theoretically possible -- we have had 2^32
9738 * cacheable predicates on this machine. We cannot allow any
9739 * more predicates to become cacheable: as unlikely as it is,
9740 * there may be a thread caching a (now stale) predicate cache
9741 * ID. (N.B.: the temptation is being successfully resisted to
9742 * have this cmn_err() "Holy shit -- we executed this code!")
9747 pred->dtp_cacheid = dtrace_predcache_id++;
9753 dtrace_predicate_hold(dtrace_predicate_t *pred)
9755 ASSERT(MUTEX_HELD(&dtrace_lock));
9756 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
9757 ASSERT(pred->dtp_refcnt > 0);
9763 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
9765 dtrace_difo_t *dp = pred->dtp_difo;
9767 ASSERT(MUTEX_HELD(&dtrace_lock));
9768 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
9769 ASSERT(pred->dtp_refcnt > 0);
9771 if (--pred->dtp_refcnt == 0) {
9772 dtrace_difo_release(pred->dtp_difo, vstate);
9773 kmem_free(pred, sizeof (dtrace_predicate_t));
9778 * DTrace Action Description Functions
9780 static dtrace_actdesc_t *
9781 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
9782 uint64_t uarg, uint64_t arg)
9784 dtrace_actdesc_t *act;
9787 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
9788 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
9791 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
9792 act->dtad_kind = kind;
9793 act->dtad_ntuple = ntuple;
9794 act->dtad_uarg = uarg;
9795 act->dtad_arg = arg;
9796 act->dtad_refcnt = 1;
9802 dtrace_actdesc_hold(dtrace_actdesc_t *act)
9804 ASSERT(act->dtad_refcnt >= 1);
9809 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
9811 dtrace_actkind_t kind = act->dtad_kind;
9814 ASSERT(act->dtad_refcnt >= 1);
9816 if (--act->dtad_refcnt != 0)
9819 if ((dp = act->dtad_difo) != NULL)
9820 dtrace_difo_release(dp, vstate);
9822 if (DTRACEACT_ISPRINTFLIKE(kind)) {
9823 char *str = (char *)(uintptr_t)act->dtad_arg;
9826 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
9827 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
9831 kmem_free(str, strlen(str) + 1);
9834 kmem_free(act, sizeof (dtrace_actdesc_t));
9838 * DTrace ECB Functions
9840 static dtrace_ecb_t *
9841 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
9846 ASSERT(MUTEX_HELD(&dtrace_lock));
9848 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
9849 ecb->dte_predicate = NULL;
9850 ecb->dte_probe = probe;
9853 * The default size is the size of the default action: recording
9856 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
9857 ecb->dte_alignment = sizeof (dtrace_epid_t);
9859 epid = state->dts_epid++;
9861 if (epid - 1 >= state->dts_necbs) {
9862 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
9863 int necbs = state->dts_necbs << 1;
9865 ASSERT(epid == state->dts_necbs + 1);
9868 ASSERT(oecbs == NULL);
9872 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
9875 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
9877 dtrace_membar_producer();
9878 state->dts_ecbs = ecbs;
9880 if (oecbs != NULL) {
9882 * If this state is active, we must dtrace_sync()
9883 * before we can free the old dts_ecbs array: we're
9884 * coming in hot, and there may be active ring
9885 * buffer processing (which indexes into the dts_ecbs
9886 * array) on another CPU.
9888 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
9891 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
9894 dtrace_membar_producer();
9895 state->dts_necbs = necbs;
9898 ecb->dte_state = state;
9900 ASSERT(state->dts_ecbs[epid - 1] == NULL);
9901 dtrace_membar_producer();
9902 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
9908 dtrace_ecb_enable(dtrace_ecb_t *ecb)
9910 dtrace_probe_t *probe = ecb->dte_probe;
9912 ASSERT(MUTEX_HELD(&cpu_lock));
9913 ASSERT(MUTEX_HELD(&dtrace_lock));
9914 ASSERT(ecb->dte_next == NULL);
9916 if (probe == NULL) {
9918 * This is the NULL probe -- there's nothing to do.
9923 if (probe->dtpr_ecb == NULL) {
9924 dtrace_provider_t *prov = probe->dtpr_provider;
9927 * We're the first ECB on this probe.
9929 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
9931 if (ecb->dte_predicate != NULL)
9932 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
9934 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
9935 probe->dtpr_id, probe->dtpr_arg);
9938 * This probe is already active. Swing the last pointer to
9939 * point to the new ECB, and issue a dtrace_sync() to assure
9940 * that all CPUs have seen the change.
9942 ASSERT(probe->dtpr_ecb_last != NULL);
9943 probe->dtpr_ecb_last->dte_next = ecb;
9944 probe->dtpr_ecb_last = ecb;
9945 probe->dtpr_predcache = 0;
9952 dtrace_ecb_resize(dtrace_ecb_t *ecb)
9954 dtrace_action_t *act;
9955 uint32_t curneeded = UINT32_MAX;
9956 uint32_t aggbase = UINT32_MAX;
9959 * If we record anything, we always record the dtrace_rechdr_t. (And
9960 * we always record it first.)
9962 ecb->dte_size = sizeof (dtrace_rechdr_t);
9963 ecb->dte_alignment = sizeof (dtrace_epid_t);
9965 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
9966 dtrace_recdesc_t *rec = &act->dta_rec;
9967 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
9969 ecb->dte_alignment = MAX(ecb->dte_alignment,
9970 rec->dtrd_alignment);
9972 if (DTRACEACT_ISAGG(act->dta_kind)) {
9973 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
9975 ASSERT(rec->dtrd_size != 0);
9976 ASSERT(agg->dtag_first != NULL);
9977 ASSERT(act->dta_prev->dta_intuple);
9978 ASSERT(aggbase != UINT32_MAX);
9979 ASSERT(curneeded != UINT32_MAX);
9981 agg->dtag_base = aggbase;
9983 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
9984 rec->dtrd_offset = curneeded;
9985 curneeded += rec->dtrd_size;
9986 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
9988 aggbase = UINT32_MAX;
9989 curneeded = UINT32_MAX;
9990 } else if (act->dta_intuple) {
9991 if (curneeded == UINT32_MAX) {
9993 * This is the first record in a tuple. Align
9994 * curneeded to be at offset 4 in an 8-byte
9997 ASSERT(act->dta_prev == NULL ||
9998 !act->dta_prev->dta_intuple);
9999 ASSERT3U(aggbase, ==, UINT32_MAX);
10000 curneeded = P2PHASEUP(ecb->dte_size,
10001 sizeof (uint64_t), sizeof (dtrace_aggid_t));
10003 aggbase = curneeded - sizeof (dtrace_aggid_t);
10004 ASSERT(IS_P2ALIGNED(aggbase,
10005 sizeof (uint64_t)));
10007 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
10008 rec->dtrd_offset = curneeded;
10009 curneeded += rec->dtrd_size;
10011 /* tuples must be followed by an aggregation */
10012 ASSERT(act->dta_prev == NULL ||
10013 !act->dta_prev->dta_intuple);
10015 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
10016 rec->dtrd_alignment);
10017 rec->dtrd_offset = ecb->dte_size;
10018 ecb->dte_size += rec->dtrd_size;
10019 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
10023 if ((act = ecb->dte_action) != NULL &&
10024 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
10025 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
10027 * If the size is still sizeof (dtrace_rechdr_t), then all
10028 * actions store no data; set the size to 0.
10033 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
10034 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
10035 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
10039 static dtrace_action_t *
10040 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10042 dtrace_aggregation_t *agg;
10043 size_t size = sizeof (uint64_t);
10044 int ntuple = desc->dtad_ntuple;
10045 dtrace_action_t *act;
10046 dtrace_recdesc_t *frec;
10047 dtrace_aggid_t aggid;
10048 dtrace_state_t *state = ecb->dte_state;
10050 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
10051 agg->dtag_ecb = ecb;
10053 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
10055 switch (desc->dtad_kind) {
10056 case DTRACEAGG_MIN:
10057 agg->dtag_initial = INT64_MAX;
10058 agg->dtag_aggregate = dtrace_aggregate_min;
10061 case DTRACEAGG_MAX:
10062 agg->dtag_initial = INT64_MIN;
10063 agg->dtag_aggregate = dtrace_aggregate_max;
10066 case DTRACEAGG_COUNT:
10067 agg->dtag_aggregate = dtrace_aggregate_count;
10070 case DTRACEAGG_QUANTIZE:
10071 agg->dtag_aggregate = dtrace_aggregate_quantize;
10072 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
10076 case DTRACEAGG_LQUANTIZE: {
10077 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
10078 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
10080 agg->dtag_initial = desc->dtad_arg;
10081 agg->dtag_aggregate = dtrace_aggregate_lquantize;
10083 if (step == 0 || levels == 0)
10086 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
10090 case DTRACEAGG_LLQUANTIZE: {
10091 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
10092 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
10093 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
10094 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
10097 agg->dtag_initial = desc->dtad_arg;
10098 agg->dtag_aggregate = dtrace_aggregate_llquantize;
10100 if (factor < 2 || low >= high || nsteps < factor)
10104 * Now check that the number of steps evenly divides a power
10105 * of the factor. (This assures both integer bucket size and
10106 * linearity within each magnitude.)
10108 for (v = factor; v < nsteps; v *= factor)
10111 if ((v % nsteps) || (nsteps % factor))
10114 size = (dtrace_aggregate_llquantize_bucket(factor,
10115 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
10119 case DTRACEAGG_AVG:
10120 agg->dtag_aggregate = dtrace_aggregate_avg;
10121 size = sizeof (uint64_t) * 2;
10124 case DTRACEAGG_STDDEV:
10125 agg->dtag_aggregate = dtrace_aggregate_stddev;
10126 size = sizeof (uint64_t) * 4;
10129 case DTRACEAGG_SUM:
10130 agg->dtag_aggregate = dtrace_aggregate_sum;
10137 agg->dtag_action.dta_rec.dtrd_size = size;
10143 * We must make sure that we have enough actions for the n-tuple.
10145 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
10146 if (DTRACEACT_ISAGG(act->dta_kind))
10149 if (--ntuple == 0) {
10151 * This is the action with which our n-tuple begins.
10153 agg->dtag_first = act;
10159 * This n-tuple is short by ntuple elements. Return failure.
10161 ASSERT(ntuple != 0);
10163 kmem_free(agg, sizeof (dtrace_aggregation_t));
10168 * If the last action in the tuple has a size of zero, it's actually
10169 * an expression argument for the aggregating action.
10171 ASSERT(ecb->dte_action_last != NULL);
10172 act = ecb->dte_action_last;
10174 if (act->dta_kind == DTRACEACT_DIFEXPR) {
10175 ASSERT(act->dta_difo != NULL);
10177 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
10178 agg->dtag_hasarg = 1;
10182 * We need to allocate an id for this aggregation.
10185 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
10186 VM_BESTFIT | VM_SLEEP);
10188 aggid = alloc_unr(state->dts_aggid_arena);
10191 if (aggid - 1 >= state->dts_naggregations) {
10192 dtrace_aggregation_t **oaggs = state->dts_aggregations;
10193 dtrace_aggregation_t **aggs;
10194 int naggs = state->dts_naggregations << 1;
10195 int onaggs = state->dts_naggregations;
10197 ASSERT(aggid == state->dts_naggregations + 1);
10200 ASSERT(oaggs == NULL);
10204 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
10206 if (oaggs != NULL) {
10207 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
10208 kmem_free(oaggs, onaggs * sizeof (*aggs));
10211 state->dts_aggregations = aggs;
10212 state->dts_naggregations = naggs;
10215 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
10216 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
10218 frec = &agg->dtag_first->dta_rec;
10219 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
10220 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
10222 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
10223 ASSERT(!act->dta_intuple);
10224 act->dta_intuple = 1;
10227 return (&agg->dtag_action);
10231 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
10233 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
10234 dtrace_state_t *state = ecb->dte_state;
10235 dtrace_aggid_t aggid = agg->dtag_id;
10237 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
10239 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
10241 free_unr(state->dts_aggid_arena, aggid);
10244 ASSERT(state->dts_aggregations[aggid - 1] == agg);
10245 state->dts_aggregations[aggid - 1] = NULL;
10247 kmem_free(agg, sizeof (dtrace_aggregation_t));
10251 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
10253 dtrace_action_t *action, *last;
10254 dtrace_difo_t *dp = desc->dtad_difo;
10255 uint32_t size = 0, align = sizeof (uint8_t), mask;
10256 uint16_t format = 0;
10257 dtrace_recdesc_t *rec;
10258 dtrace_state_t *state = ecb->dte_state;
10259 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
10260 uint64_t arg = desc->dtad_arg;
10262 ASSERT(MUTEX_HELD(&dtrace_lock));
10263 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
10265 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
10267 * If this is an aggregating action, there must be neither
10268 * a speculate nor a commit on the action chain.
10270 dtrace_action_t *act;
10272 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
10273 if (act->dta_kind == DTRACEACT_COMMIT)
10276 if (act->dta_kind == DTRACEACT_SPECULATE)
10280 action = dtrace_ecb_aggregation_create(ecb, desc);
10282 if (action == NULL)
10285 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
10286 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
10287 dp != NULL && dp->dtdo_destructive)) {
10288 state->dts_destructive = 1;
10291 switch (desc->dtad_kind) {
10292 case DTRACEACT_PRINTF:
10293 case DTRACEACT_PRINTA:
10294 case DTRACEACT_SYSTEM:
10295 case DTRACEACT_FREOPEN:
10296 case DTRACEACT_DIFEXPR:
10298 * We know that our arg is a string -- turn it into a
10302 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
10303 desc->dtad_kind == DTRACEACT_DIFEXPR);
10308 ASSERT(arg > KERNELBASE);
10310 format = dtrace_format_add(state,
10311 (char *)(uintptr_t)arg);
10315 case DTRACEACT_LIBACT:
10316 case DTRACEACT_TRACEMEM:
10317 case DTRACEACT_TRACEMEM_DYNSIZE:
10321 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
10324 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
10325 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10328 size = opt[DTRACEOPT_STRSIZE];
10333 case DTRACEACT_STACK:
10334 if ((nframes = arg) == 0) {
10335 nframes = opt[DTRACEOPT_STACKFRAMES];
10336 ASSERT(nframes > 0);
10340 size = nframes * sizeof (pc_t);
10343 case DTRACEACT_JSTACK:
10344 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
10345 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
10347 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
10348 nframes = opt[DTRACEOPT_JSTACKFRAMES];
10350 arg = DTRACE_USTACK_ARG(nframes, strsize);
10353 case DTRACEACT_USTACK:
10354 if (desc->dtad_kind != DTRACEACT_JSTACK &&
10355 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
10356 strsize = DTRACE_USTACK_STRSIZE(arg);
10357 nframes = opt[DTRACEOPT_USTACKFRAMES];
10358 ASSERT(nframes > 0);
10359 arg = DTRACE_USTACK_ARG(nframes, strsize);
10363 * Save a slot for the pid.
10365 size = (nframes + 1) * sizeof (uint64_t);
10366 size += DTRACE_USTACK_STRSIZE(arg);
10367 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
10371 case DTRACEACT_SYM:
10372 case DTRACEACT_MOD:
10373 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
10374 sizeof (uint64_t)) ||
10375 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10379 case DTRACEACT_USYM:
10380 case DTRACEACT_UMOD:
10381 case DTRACEACT_UADDR:
10383 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
10384 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10388 * We have a slot for the pid, plus a slot for the
10389 * argument. To keep things simple (aligned with
10390 * bitness-neutral sizing), we store each as a 64-bit
10393 size = 2 * sizeof (uint64_t);
10396 case DTRACEACT_STOP:
10397 case DTRACEACT_BREAKPOINT:
10398 case DTRACEACT_PANIC:
10401 case DTRACEACT_CHILL:
10402 case DTRACEACT_DISCARD:
10403 case DTRACEACT_RAISE:
10408 case DTRACEACT_EXIT:
10410 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
10411 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
10415 case DTRACEACT_SPECULATE:
10416 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
10422 state->dts_speculates = 1;
10425 case DTRACEACT_PRINTM:
10426 size = dp->dtdo_rtype.dtdt_size;
10429 case DTRACEACT_PRINTT:
10430 size = dp->dtdo_rtype.dtdt_size;
10433 case DTRACEACT_COMMIT: {
10434 dtrace_action_t *act = ecb->dte_action;
10436 for (; act != NULL; act = act->dta_next) {
10437 if (act->dta_kind == DTRACEACT_COMMIT)
10450 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
10452 * If this is a data-storing action or a speculate,
10453 * we must be sure that there isn't a commit on the
10456 dtrace_action_t *act = ecb->dte_action;
10458 for (; act != NULL; act = act->dta_next) {
10459 if (act->dta_kind == DTRACEACT_COMMIT)
10464 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
10465 action->dta_rec.dtrd_size = size;
10468 action->dta_refcnt = 1;
10469 rec = &action->dta_rec;
10470 size = rec->dtrd_size;
10472 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
10473 if (!(size & mask)) {
10479 action->dta_kind = desc->dtad_kind;
10481 if ((action->dta_difo = dp) != NULL)
10482 dtrace_difo_hold(dp);
10484 rec->dtrd_action = action->dta_kind;
10485 rec->dtrd_arg = arg;
10486 rec->dtrd_uarg = desc->dtad_uarg;
10487 rec->dtrd_alignment = (uint16_t)align;
10488 rec->dtrd_format = format;
10490 if ((last = ecb->dte_action_last) != NULL) {
10491 ASSERT(ecb->dte_action != NULL);
10492 action->dta_prev = last;
10493 last->dta_next = action;
10495 ASSERT(ecb->dte_action == NULL);
10496 ecb->dte_action = action;
10499 ecb->dte_action_last = action;
10505 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
10507 dtrace_action_t *act = ecb->dte_action, *next;
10508 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
10512 if (act != NULL && act->dta_refcnt > 1) {
10513 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
10516 for (; act != NULL; act = next) {
10517 next = act->dta_next;
10518 ASSERT(next != NULL || act == ecb->dte_action_last);
10519 ASSERT(act->dta_refcnt == 1);
10521 if ((format = act->dta_rec.dtrd_format) != 0)
10522 dtrace_format_remove(ecb->dte_state, format);
10524 if ((dp = act->dta_difo) != NULL)
10525 dtrace_difo_release(dp, vstate);
10527 if (DTRACEACT_ISAGG(act->dta_kind)) {
10528 dtrace_ecb_aggregation_destroy(ecb, act);
10530 kmem_free(act, sizeof (dtrace_action_t));
10535 ecb->dte_action = NULL;
10536 ecb->dte_action_last = NULL;
10541 dtrace_ecb_disable(dtrace_ecb_t *ecb)
10544 * We disable the ECB by removing it from its probe.
10546 dtrace_ecb_t *pecb, *prev = NULL;
10547 dtrace_probe_t *probe = ecb->dte_probe;
10549 ASSERT(MUTEX_HELD(&dtrace_lock));
10551 if (probe == NULL) {
10553 * This is the NULL probe; there is nothing to disable.
10558 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
10564 ASSERT(pecb != NULL);
10566 if (prev == NULL) {
10567 probe->dtpr_ecb = ecb->dte_next;
10569 prev->dte_next = ecb->dte_next;
10572 if (ecb == probe->dtpr_ecb_last) {
10573 ASSERT(ecb->dte_next == NULL);
10574 probe->dtpr_ecb_last = prev;
10578 * The ECB has been disconnected from the probe; now sync to assure
10579 * that all CPUs have seen the change before returning.
10583 if (probe->dtpr_ecb == NULL) {
10585 * That was the last ECB on the probe; clear the predicate
10586 * cache ID for the probe, disable it and sync one more time
10587 * to assure that we'll never hit it again.
10589 dtrace_provider_t *prov = probe->dtpr_provider;
10591 ASSERT(ecb->dte_next == NULL);
10592 ASSERT(probe->dtpr_ecb_last == NULL);
10593 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
10594 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
10595 probe->dtpr_id, probe->dtpr_arg);
10599 * There is at least one ECB remaining on the probe. If there
10600 * is _exactly_ one, set the probe's predicate cache ID to be
10601 * the predicate cache ID of the remaining ECB.
10603 ASSERT(probe->dtpr_ecb_last != NULL);
10604 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
10606 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
10607 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
10609 ASSERT(probe->dtpr_ecb->dte_next == NULL);
10612 probe->dtpr_predcache = p->dtp_cacheid;
10615 ecb->dte_next = NULL;
10620 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
10622 dtrace_state_t *state = ecb->dte_state;
10623 dtrace_vstate_t *vstate = &state->dts_vstate;
10624 dtrace_predicate_t *pred;
10625 dtrace_epid_t epid = ecb->dte_epid;
10627 ASSERT(MUTEX_HELD(&dtrace_lock));
10628 ASSERT(ecb->dte_next == NULL);
10629 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
10631 if ((pred = ecb->dte_predicate) != NULL)
10632 dtrace_predicate_release(pred, vstate);
10634 dtrace_ecb_action_remove(ecb);
10636 ASSERT(state->dts_ecbs[epid - 1] == ecb);
10637 state->dts_ecbs[epid - 1] = NULL;
10639 kmem_free(ecb, sizeof (dtrace_ecb_t));
10642 static dtrace_ecb_t *
10643 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
10644 dtrace_enabling_t *enab)
10647 dtrace_predicate_t *pred;
10648 dtrace_actdesc_t *act;
10649 dtrace_provider_t *prov;
10650 dtrace_ecbdesc_t *desc = enab->dten_current;
10652 ASSERT(MUTEX_HELD(&dtrace_lock));
10653 ASSERT(state != NULL);
10655 ecb = dtrace_ecb_add(state, probe);
10656 ecb->dte_uarg = desc->dted_uarg;
10658 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
10659 dtrace_predicate_hold(pred);
10660 ecb->dte_predicate = pred;
10663 if (probe != NULL) {
10665 * If the provider shows more leg than the consumer is old
10666 * enough to see, we need to enable the appropriate implicit
10667 * predicate bits to prevent the ecb from activating at
10670 * Providers specifying DTRACE_PRIV_USER at register time
10671 * are stating that they need the /proc-style privilege
10672 * model to be enforced, and this is what DTRACE_COND_OWNER
10673 * and DTRACE_COND_ZONEOWNER will then do at probe time.
10675 prov = probe->dtpr_provider;
10676 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
10677 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10678 ecb->dte_cond |= DTRACE_COND_OWNER;
10680 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
10681 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
10682 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
10685 * If the provider shows us kernel innards and the user
10686 * is lacking sufficient privilege, enable the
10687 * DTRACE_COND_USERMODE implicit predicate.
10689 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
10690 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
10691 ecb->dte_cond |= DTRACE_COND_USERMODE;
10694 if (dtrace_ecb_create_cache != NULL) {
10696 * If we have a cached ecb, we'll use its action list instead
10697 * of creating our own (saving both time and space).
10699 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
10700 dtrace_action_t *act = cached->dte_action;
10703 ASSERT(act->dta_refcnt > 0);
10705 ecb->dte_action = act;
10706 ecb->dte_action_last = cached->dte_action_last;
10707 ecb->dte_needed = cached->dte_needed;
10708 ecb->dte_size = cached->dte_size;
10709 ecb->dte_alignment = cached->dte_alignment;
10715 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
10716 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
10717 dtrace_ecb_destroy(ecb);
10722 dtrace_ecb_resize(ecb);
10724 return (dtrace_ecb_create_cache = ecb);
10728 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
10731 dtrace_enabling_t *enab = arg;
10732 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
10734 ASSERT(state != NULL);
10736 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
10738 * This probe was created in a generation for which this
10739 * enabling has previously created ECBs; we don't want to
10740 * enable it again, so just kick out.
10742 return (DTRACE_MATCH_NEXT);
10745 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
10746 return (DTRACE_MATCH_DONE);
10748 dtrace_ecb_enable(ecb);
10749 return (DTRACE_MATCH_NEXT);
10752 static dtrace_ecb_t *
10753 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
10757 ASSERT(MUTEX_HELD(&dtrace_lock));
10759 if (id == 0 || id > state->dts_necbs)
10762 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
10763 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
10765 return (state->dts_ecbs[id - 1]);
10768 static dtrace_aggregation_t *
10769 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
10771 dtrace_aggregation_t *agg;
10773 ASSERT(MUTEX_HELD(&dtrace_lock));
10775 if (id == 0 || id > state->dts_naggregations)
10778 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
10779 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
10780 agg->dtag_id == id);
10782 return (state->dts_aggregations[id - 1]);
10786 * DTrace Buffer Functions
10788 * The following functions manipulate DTrace buffers. Most of these functions
10789 * are called in the context of establishing or processing consumer state;
10790 * exceptions are explicitly noted.
10794 * Note: called from cross call context. This function switches the two
10795 * buffers on a given CPU. The atomicity of this operation is assured by
10796 * disabling interrupts while the actual switch takes place; the disabling of
10797 * interrupts serializes the execution with any execution of dtrace_probe() on
10801 dtrace_buffer_switch(dtrace_buffer_t *buf)
10803 caddr_t tomax = buf->dtb_tomax;
10804 caddr_t xamot = buf->dtb_xamot;
10805 dtrace_icookie_t cookie;
10808 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
10809 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
10811 cookie = dtrace_interrupt_disable();
10812 now = dtrace_gethrtime();
10813 buf->dtb_tomax = xamot;
10814 buf->dtb_xamot = tomax;
10815 buf->dtb_xamot_drops = buf->dtb_drops;
10816 buf->dtb_xamot_offset = buf->dtb_offset;
10817 buf->dtb_xamot_errors = buf->dtb_errors;
10818 buf->dtb_xamot_flags = buf->dtb_flags;
10819 buf->dtb_offset = 0;
10820 buf->dtb_drops = 0;
10821 buf->dtb_errors = 0;
10822 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
10823 buf->dtb_interval = now - buf->dtb_switched;
10824 buf->dtb_switched = now;
10825 dtrace_interrupt_enable(cookie);
10829 * Note: called from cross call context. This function activates a buffer
10830 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
10831 * is guaranteed by the disabling of interrupts.
10834 dtrace_buffer_activate(dtrace_state_t *state)
10836 dtrace_buffer_t *buf;
10837 dtrace_icookie_t cookie = dtrace_interrupt_disable();
10839 buf = &state->dts_buffer[curcpu];
10841 if (buf->dtb_tomax != NULL) {
10843 * We might like to assert that the buffer is marked inactive,
10844 * but this isn't necessarily true: the buffer for the CPU
10845 * that processes the BEGIN probe has its buffer activated
10846 * manually. In this case, we take the (harmless) action
10847 * re-clearing the bit INACTIVE bit.
10849 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
10852 dtrace_interrupt_enable(cookie);
10856 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
10857 processorid_t cpu, int *factor)
10862 dtrace_buffer_t *buf;
10863 int allocated = 0, desired = 0;
10866 ASSERT(MUTEX_HELD(&cpu_lock));
10867 ASSERT(MUTEX_HELD(&dtrace_lock));
10871 if (size > dtrace_nonroot_maxsize &&
10872 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
10878 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10881 buf = &bufs[cp->cpu_id];
10884 * If there is already a buffer allocated for this CPU, it
10885 * is only possible that this is a DR event. In this case,
10887 if (buf->dtb_tomax != NULL) {
10888 ASSERT(buf->dtb_size == size);
10892 ASSERT(buf->dtb_xamot == NULL);
10894 if ((buf->dtb_tomax = kmem_zalloc(size,
10895 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10898 buf->dtb_size = size;
10899 buf->dtb_flags = flags;
10900 buf->dtb_offset = 0;
10901 buf->dtb_drops = 0;
10903 if (flags & DTRACEBUF_NOSWITCH)
10906 if ((buf->dtb_xamot = kmem_zalloc(size,
10907 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10909 } while ((cp = cp->cpu_next) != cpu_list);
10917 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
10920 buf = &bufs[cp->cpu_id];
10923 if (buf->dtb_xamot != NULL) {
10924 ASSERT(buf->dtb_tomax != NULL);
10925 ASSERT(buf->dtb_size == size);
10926 kmem_free(buf->dtb_xamot, size);
10930 if (buf->dtb_tomax != NULL) {
10931 ASSERT(buf->dtb_size == size);
10932 kmem_free(buf->dtb_tomax, size);
10936 buf->dtb_tomax = NULL;
10937 buf->dtb_xamot = NULL;
10939 } while ((cp = cp->cpu_next) != cpu_list);
10944 #if defined(__amd64__)
10946 * FreeBSD isn't good at limiting the amount of memory we
10947 * ask to malloc, so let's place a limit here before trying
10948 * to do something that might well end in tears at bedtime.
10950 if (size > physmem * PAGE_SIZE / (128 * (mp_maxid + 1)))
10954 ASSERT(MUTEX_HELD(&dtrace_lock));
10956 if (cpu != DTRACE_CPUALL && cpu != i)
10962 * If there is already a buffer allocated for this CPU, it
10963 * is only possible that this is a DR event. In this case,
10964 * the buffer size must match our specified size.
10966 if (buf->dtb_tomax != NULL) {
10967 ASSERT(buf->dtb_size == size);
10971 ASSERT(buf->dtb_xamot == NULL);
10973 if ((buf->dtb_tomax = kmem_zalloc(size,
10974 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10977 buf->dtb_size = size;
10978 buf->dtb_flags = flags;
10979 buf->dtb_offset = 0;
10980 buf->dtb_drops = 0;
10982 if (flags & DTRACEBUF_NOSWITCH)
10985 if ((buf->dtb_xamot = kmem_zalloc(size,
10986 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
10994 * Error allocating memory, so free the buffers that were
10995 * allocated before the failed allocation.
10998 if (cpu != DTRACE_CPUALL && cpu != i)
11004 if (buf->dtb_xamot != NULL) {
11005 ASSERT(buf->dtb_tomax != NULL);
11006 ASSERT(buf->dtb_size == size);
11007 kmem_free(buf->dtb_xamot, size);
11011 if (buf->dtb_tomax != NULL) {
11012 ASSERT(buf->dtb_size == size);
11013 kmem_free(buf->dtb_tomax, size);
11017 buf->dtb_tomax = NULL;
11018 buf->dtb_xamot = NULL;
11023 *factor = desired / (allocated > 0 ? allocated : 1);
11029 * Note: called from probe context. This function just increments the drop
11030 * count on a buffer. It has been made a function to allow for the
11031 * possibility of understanding the source of mysterious drop counts. (A
11032 * problem for which one may be particularly disappointed that DTrace cannot
11033 * be used to understand DTrace.)
11036 dtrace_buffer_drop(dtrace_buffer_t *buf)
11042 * Note: called from probe context. This function is called to reserve space
11043 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
11044 * mstate. Returns the new offset in the buffer, or a negative value if an
11045 * error has occurred.
11048 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
11049 dtrace_state_t *state, dtrace_mstate_t *mstate)
11051 intptr_t offs = buf->dtb_offset, soffs;
11056 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
11059 if ((tomax = buf->dtb_tomax) == NULL) {
11060 dtrace_buffer_drop(buf);
11064 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
11065 while (offs & (align - 1)) {
11067 * Assert that our alignment is off by a number which
11068 * is itself sizeof (uint32_t) aligned.
11070 ASSERT(!((align - (offs & (align - 1))) &
11071 (sizeof (uint32_t) - 1)));
11072 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11073 offs += sizeof (uint32_t);
11076 if ((soffs = offs + needed) > buf->dtb_size) {
11077 dtrace_buffer_drop(buf);
11081 if (mstate == NULL)
11084 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
11085 mstate->dtms_scratch_size = buf->dtb_size - soffs;
11086 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11091 if (buf->dtb_flags & DTRACEBUF_FILL) {
11092 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
11093 (buf->dtb_flags & DTRACEBUF_FULL))
11098 total = needed + (offs & (align - 1));
11101 * For a ring buffer, life is quite a bit more complicated. Before
11102 * we can store any padding, we need to adjust our wrapping offset.
11103 * (If we've never before wrapped or we're not about to, no adjustment
11106 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
11107 offs + total > buf->dtb_size) {
11108 woffs = buf->dtb_xamot_offset;
11110 if (offs + total > buf->dtb_size) {
11112 * We can't fit in the end of the buffer. First, a
11113 * sanity check that we can fit in the buffer at all.
11115 if (total > buf->dtb_size) {
11116 dtrace_buffer_drop(buf);
11121 * We're going to be storing at the top of the buffer,
11122 * so now we need to deal with the wrapped offset. We
11123 * only reset our wrapped offset to 0 if it is
11124 * currently greater than the current offset. If it
11125 * is less than the current offset, it is because a
11126 * previous allocation induced a wrap -- but the
11127 * allocation didn't subsequently take the space due
11128 * to an error or false predicate evaluation. In this
11129 * case, we'll just leave the wrapped offset alone: if
11130 * the wrapped offset hasn't been advanced far enough
11131 * for this allocation, it will be adjusted in the
11134 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
11142 * Now we know that we're going to be storing to the
11143 * top of the buffer and that there is room for us
11144 * there. We need to clear the buffer from the current
11145 * offset to the end (there may be old gunk there).
11147 while (offs < buf->dtb_size)
11151 * We need to set our offset to zero. And because we
11152 * are wrapping, we need to set the bit indicating as
11153 * much. We can also adjust our needed space back
11154 * down to the space required by the ECB -- we know
11155 * that the top of the buffer is aligned.
11159 buf->dtb_flags |= DTRACEBUF_WRAPPED;
11162 * There is room for us in the buffer, so we simply
11163 * need to check the wrapped offset.
11165 if (woffs < offs) {
11167 * The wrapped offset is less than the offset.
11168 * This can happen if we allocated buffer space
11169 * that induced a wrap, but then we didn't
11170 * subsequently take the space due to an error
11171 * or false predicate evaluation. This is
11172 * okay; we know that _this_ allocation isn't
11173 * going to induce a wrap. We still can't
11174 * reset the wrapped offset to be zero,
11175 * however: the space may have been trashed in
11176 * the previous failed probe attempt. But at
11177 * least the wrapped offset doesn't need to
11178 * be adjusted at all...
11184 while (offs + total > woffs) {
11185 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
11188 if (epid == DTRACE_EPIDNONE) {
11189 size = sizeof (uint32_t);
11191 ASSERT3U(epid, <=, state->dts_necbs);
11192 ASSERT(state->dts_ecbs[epid - 1] != NULL);
11194 size = state->dts_ecbs[epid - 1]->dte_size;
11197 ASSERT(woffs + size <= buf->dtb_size);
11200 if (woffs + size == buf->dtb_size) {
11202 * We've reached the end of the buffer; we want
11203 * to set the wrapped offset to 0 and break
11204 * out. However, if the offs is 0, then we're
11205 * in a strange edge-condition: the amount of
11206 * space that we want to reserve plus the size
11207 * of the record that we're overwriting is
11208 * greater than the size of the buffer. This
11209 * is problematic because if we reserve the
11210 * space but subsequently don't consume it (due
11211 * to a failed predicate or error) the wrapped
11212 * offset will be 0 -- yet the EPID at offset 0
11213 * will not be committed. This situation is
11214 * relatively easy to deal with: if we're in
11215 * this case, the buffer is indistinguishable
11216 * from one that hasn't wrapped; we need only
11217 * finish the job by clearing the wrapped bit,
11218 * explicitly setting the offset to be 0, and
11219 * zero'ing out the old data in the buffer.
11222 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
11223 buf->dtb_offset = 0;
11226 while (woffs < buf->dtb_size)
11227 tomax[woffs++] = 0;
11238 * We have a wrapped offset. It may be that the wrapped offset
11239 * has become zero -- that's okay.
11241 buf->dtb_xamot_offset = woffs;
11246 * Now we can plow the buffer with any necessary padding.
11248 while (offs & (align - 1)) {
11250 * Assert that our alignment is off by a number which
11251 * is itself sizeof (uint32_t) aligned.
11253 ASSERT(!((align - (offs & (align - 1))) &
11254 (sizeof (uint32_t) - 1)));
11255 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
11256 offs += sizeof (uint32_t);
11259 if (buf->dtb_flags & DTRACEBUF_FILL) {
11260 if (offs + needed > buf->dtb_size - state->dts_reserve) {
11261 buf->dtb_flags |= DTRACEBUF_FULL;
11266 if (mstate == NULL)
11270 * For ring buffers and fill buffers, the scratch space is always
11271 * the inactive buffer.
11273 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
11274 mstate->dtms_scratch_size = buf->dtb_size;
11275 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
11281 dtrace_buffer_polish(dtrace_buffer_t *buf)
11283 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
11284 ASSERT(MUTEX_HELD(&dtrace_lock));
11286 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
11290 * We need to polish the ring buffer. There are three cases:
11292 * - The first (and presumably most common) is that there is no gap
11293 * between the buffer offset and the wrapped offset. In this case,
11294 * there is nothing in the buffer that isn't valid data; we can
11295 * mark the buffer as polished and return.
11297 * - The second (less common than the first but still more common
11298 * than the third) is that there is a gap between the buffer offset
11299 * and the wrapped offset, and the wrapped offset is larger than the
11300 * buffer offset. This can happen because of an alignment issue, or
11301 * can happen because of a call to dtrace_buffer_reserve() that
11302 * didn't subsequently consume the buffer space. In this case,
11303 * we need to zero the data from the buffer offset to the wrapped
11306 * - The third (and least common) is that there is a gap between the
11307 * buffer offset and the wrapped offset, but the wrapped offset is
11308 * _less_ than the buffer offset. This can only happen because a
11309 * call to dtrace_buffer_reserve() induced a wrap, but the space
11310 * was not subsequently consumed. In this case, we need to zero the
11311 * space from the offset to the end of the buffer _and_ from the
11312 * top of the buffer to the wrapped offset.
11314 if (buf->dtb_offset < buf->dtb_xamot_offset) {
11315 bzero(buf->dtb_tomax + buf->dtb_offset,
11316 buf->dtb_xamot_offset - buf->dtb_offset);
11319 if (buf->dtb_offset > buf->dtb_xamot_offset) {
11320 bzero(buf->dtb_tomax + buf->dtb_offset,
11321 buf->dtb_size - buf->dtb_offset);
11322 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
11327 * This routine determines if data generated at the specified time has likely
11328 * been entirely consumed at user-level. This routine is called to determine
11329 * if an ECB on a defunct probe (but for an active enabling) can be safely
11330 * disabled and destroyed.
11333 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
11337 for (i = 0; i < NCPU; i++) {
11338 dtrace_buffer_t *buf = &bufs[i];
11340 if (buf->dtb_size == 0)
11343 if (buf->dtb_flags & DTRACEBUF_RING)
11346 if (!buf->dtb_switched && buf->dtb_offset != 0)
11349 if (buf->dtb_switched - buf->dtb_interval < when)
11357 dtrace_buffer_free(dtrace_buffer_t *bufs)
11361 for (i = 0; i < NCPU; i++) {
11362 dtrace_buffer_t *buf = &bufs[i];
11364 if (buf->dtb_tomax == NULL) {
11365 ASSERT(buf->dtb_xamot == NULL);
11366 ASSERT(buf->dtb_size == 0);
11370 if (buf->dtb_xamot != NULL) {
11371 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
11372 kmem_free(buf->dtb_xamot, buf->dtb_size);
11375 kmem_free(buf->dtb_tomax, buf->dtb_size);
11377 buf->dtb_tomax = NULL;
11378 buf->dtb_xamot = NULL;
11383 * DTrace Enabling Functions
11385 static dtrace_enabling_t *
11386 dtrace_enabling_create(dtrace_vstate_t *vstate)
11388 dtrace_enabling_t *enab;
11390 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
11391 enab->dten_vstate = vstate;
11397 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
11399 dtrace_ecbdesc_t **ndesc;
11400 size_t osize, nsize;
11403 * We can't add to enablings after we've enabled them, or after we've
11406 ASSERT(enab->dten_probegen == 0);
11407 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11409 if (enab->dten_ndesc < enab->dten_maxdesc) {
11410 enab->dten_desc[enab->dten_ndesc++] = ecb;
11414 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11416 if (enab->dten_maxdesc == 0) {
11417 enab->dten_maxdesc = 1;
11419 enab->dten_maxdesc <<= 1;
11422 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
11424 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
11425 ndesc = kmem_zalloc(nsize, KM_SLEEP);
11426 bcopy(enab->dten_desc, ndesc, osize);
11427 if (enab->dten_desc != NULL)
11428 kmem_free(enab->dten_desc, osize);
11430 enab->dten_desc = ndesc;
11431 enab->dten_desc[enab->dten_ndesc++] = ecb;
11435 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
11436 dtrace_probedesc_t *pd)
11438 dtrace_ecbdesc_t *new;
11439 dtrace_predicate_t *pred;
11440 dtrace_actdesc_t *act;
11443 * We're going to create a new ECB description that matches the
11444 * specified ECB in every way, but has the specified probe description.
11446 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
11448 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
11449 dtrace_predicate_hold(pred);
11451 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
11452 dtrace_actdesc_hold(act);
11454 new->dted_action = ecb->dted_action;
11455 new->dted_pred = ecb->dted_pred;
11456 new->dted_probe = *pd;
11457 new->dted_uarg = ecb->dted_uarg;
11459 dtrace_enabling_add(enab, new);
11463 dtrace_enabling_dump(dtrace_enabling_t *enab)
11467 for (i = 0; i < enab->dten_ndesc; i++) {
11468 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
11470 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
11471 desc->dtpd_provider, desc->dtpd_mod,
11472 desc->dtpd_func, desc->dtpd_name);
11477 dtrace_enabling_destroy(dtrace_enabling_t *enab)
11480 dtrace_ecbdesc_t *ep;
11481 dtrace_vstate_t *vstate = enab->dten_vstate;
11483 ASSERT(MUTEX_HELD(&dtrace_lock));
11485 for (i = 0; i < enab->dten_ndesc; i++) {
11486 dtrace_actdesc_t *act, *next;
11487 dtrace_predicate_t *pred;
11489 ep = enab->dten_desc[i];
11491 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
11492 dtrace_predicate_release(pred, vstate);
11494 for (act = ep->dted_action; act != NULL; act = next) {
11495 next = act->dtad_next;
11496 dtrace_actdesc_release(act, vstate);
11499 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
11502 if (enab->dten_desc != NULL)
11503 kmem_free(enab->dten_desc,
11504 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
11507 * If this was a retained enabling, decrement the dts_nretained count
11508 * and take it off of the dtrace_retained list.
11510 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
11511 dtrace_retained == enab) {
11512 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11513 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
11514 enab->dten_vstate->dtvs_state->dts_nretained--;
11517 if (enab->dten_prev == NULL) {
11518 if (dtrace_retained == enab) {
11519 dtrace_retained = enab->dten_next;
11521 if (dtrace_retained != NULL)
11522 dtrace_retained->dten_prev = NULL;
11525 ASSERT(enab != dtrace_retained);
11526 ASSERT(dtrace_retained != NULL);
11527 enab->dten_prev->dten_next = enab->dten_next;
11530 if (enab->dten_next != NULL) {
11531 ASSERT(dtrace_retained != NULL);
11532 enab->dten_next->dten_prev = enab->dten_prev;
11535 kmem_free(enab, sizeof (dtrace_enabling_t));
11539 dtrace_enabling_retain(dtrace_enabling_t *enab)
11541 dtrace_state_t *state;
11543 ASSERT(MUTEX_HELD(&dtrace_lock));
11544 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
11545 ASSERT(enab->dten_vstate != NULL);
11547 state = enab->dten_vstate->dtvs_state;
11548 ASSERT(state != NULL);
11551 * We only allow each state to retain dtrace_retain_max enablings.
11553 if (state->dts_nretained >= dtrace_retain_max)
11556 state->dts_nretained++;
11558 if (dtrace_retained == NULL) {
11559 dtrace_retained = enab;
11563 enab->dten_next = dtrace_retained;
11564 dtrace_retained->dten_prev = enab;
11565 dtrace_retained = enab;
11571 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
11572 dtrace_probedesc_t *create)
11574 dtrace_enabling_t *new, *enab;
11575 int found = 0, err = ENOENT;
11577 ASSERT(MUTEX_HELD(&dtrace_lock));
11578 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
11579 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
11580 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
11581 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
11583 new = dtrace_enabling_create(&state->dts_vstate);
11586 * Iterate over all retained enablings, looking for enablings that
11587 * match the specified state.
11589 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11593 * dtvs_state can only be NULL for helper enablings -- and
11594 * helper enablings can't be retained.
11596 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11598 if (enab->dten_vstate->dtvs_state != state)
11602 * Now iterate over each probe description; we're looking for
11603 * an exact match to the specified probe description.
11605 for (i = 0; i < enab->dten_ndesc; i++) {
11606 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11607 dtrace_probedesc_t *pd = &ep->dted_probe;
11609 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
11612 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
11615 if (strcmp(pd->dtpd_func, match->dtpd_func))
11618 if (strcmp(pd->dtpd_name, match->dtpd_name))
11622 * We have a winning probe! Add it to our growing
11626 dtrace_enabling_addlike(new, ep, create);
11630 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
11631 dtrace_enabling_destroy(new);
11639 dtrace_enabling_retract(dtrace_state_t *state)
11641 dtrace_enabling_t *enab, *next;
11643 ASSERT(MUTEX_HELD(&dtrace_lock));
11646 * Iterate over all retained enablings, destroy the enablings retained
11647 * for the specified state.
11649 for (enab = dtrace_retained; enab != NULL; enab = next) {
11650 next = enab->dten_next;
11653 * dtvs_state can only be NULL for helper enablings -- and
11654 * helper enablings can't be retained.
11656 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11658 if (enab->dten_vstate->dtvs_state == state) {
11659 ASSERT(state->dts_nretained > 0);
11660 dtrace_enabling_destroy(enab);
11664 ASSERT(state->dts_nretained == 0);
11668 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
11673 ASSERT(MUTEX_HELD(&cpu_lock));
11674 ASSERT(MUTEX_HELD(&dtrace_lock));
11676 for (i = 0; i < enab->dten_ndesc; i++) {
11677 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
11679 enab->dten_current = ep;
11680 enab->dten_error = 0;
11682 matched += dtrace_probe_enable(&ep->dted_probe, enab);
11684 if (enab->dten_error != 0) {
11686 * If we get an error half-way through enabling the
11687 * probes, we kick out -- perhaps with some number of
11688 * them enabled. Leaving enabled probes enabled may
11689 * be slightly confusing for user-level, but we expect
11690 * that no one will attempt to actually drive on in
11691 * the face of such errors. If this is an anonymous
11692 * enabling (indicated with a NULL nmatched pointer),
11693 * we cmn_err() a message. We aren't expecting to
11694 * get such an error -- such as it can exist at all,
11695 * it would be a result of corrupted DOF in the driver
11698 if (nmatched == NULL) {
11699 cmn_err(CE_WARN, "dtrace_enabling_match() "
11700 "error on %p: %d", (void *)ep,
11704 return (enab->dten_error);
11708 enab->dten_probegen = dtrace_probegen;
11709 if (nmatched != NULL)
11710 *nmatched = matched;
11716 dtrace_enabling_matchall(void)
11718 dtrace_enabling_t *enab;
11720 mutex_enter(&cpu_lock);
11721 mutex_enter(&dtrace_lock);
11724 * Iterate over all retained enablings to see if any probes match
11725 * against them. We only perform this operation on enablings for which
11726 * we have sufficient permissions by virtue of being in the global zone
11727 * or in the same zone as the DTrace client. Because we can be called
11728 * after dtrace_detach() has been called, we cannot assert that there
11729 * are retained enablings. We can safely load from dtrace_retained,
11730 * however: the taskq_destroy() at the end of dtrace_detach() will
11731 * block pending our completion.
11733 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11735 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
11737 if (INGLOBALZONE(curproc) ||
11738 cr != NULL && getzoneid() == crgetzoneid(cr))
11740 (void) dtrace_enabling_match(enab, NULL);
11743 mutex_exit(&dtrace_lock);
11744 mutex_exit(&cpu_lock);
11748 * If an enabling is to be enabled without having matched probes (that is, if
11749 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
11750 * enabling must be _primed_ by creating an ECB for every ECB description.
11751 * This must be done to assure that we know the number of speculations, the
11752 * number of aggregations, the minimum buffer size needed, etc. before we
11753 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
11754 * enabling any probes, we create ECBs for every ECB decription, but with a
11755 * NULL probe -- which is exactly what this function does.
11758 dtrace_enabling_prime(dtrace_state_t *state)
11760 dtrace_enabling_t *enab;
11763 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
11764 ASSERT(enab->dten_vstate->dtvs_state != NULL);
11766 if (enab->dten_vstate->dtvs_state != state)
11770 * We don't want to prime an enabling more than once, lest
11771 * we allow a malicious user to induce resource exhaustion.
11772 * (The ECBs that result from priming an enabling aren't
11773 * leaked -- but they also aren't deallocated until the
11774 * consumer state is destroyed.)
11776 if (enab->dten_primed)
11779 for (i = 0; i < enab->dten_ndesc; i++) {
11780 enab->dten_current = enab->dten_desc[i];
11781 (void) dtrace_probe_enable(NULL, enab);
11784 enab->dten_primed = 1;
11789 * Called to indicate that probes should be provided due to retained
11790 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
11791 * must take an initial lap through the enabling calling the dtps_provide()
11792 * entry point explicitly to allow for autocreated probes.
11795 dtrace_enabling_provide(dtrace_provider_t *prv)
11798 dtrace_probedesc_t desc;
11800 ASSERT(MUTEX_HELD(&dtrace_lock));
11801 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
11805 prv = dtrace_provider;
11809 dtrace_enabling_t *enab = dtrace_retained;
11810 void *parg = prv->dtpv_arg;
11812 for (; enab != NULL; enab = enab->dten_next) {
11813 for (i = 0; i < enab->dten_ndesc; i++) {
11814 desc = enab->dten_desc[i]->dted_probe;
11815 mutex_exit(&dtrace_lock);
11816 prv->dtpv_pops.dtps_provide(parg, &desc);
11817 mutex_enter(&dtrace_lock);
11820 } while (all && (prv = prv->dtpv_next) != NULL);
11822 mutex_exit(&dtrace_lock);
11823 dtrace_probe_provide(NULL, all ? NULL : prv);
11824 mutex_enter(&dtrace_lock);
11828 * Called to reap ECBs that are attached to probes from defunct providers.
11831 dtrace_enabling_reap(void)
11833 dtrace_provider_t *prov;
11834 dtrace_probe_t *probe;
11839 mutex_enter(&cpu_lock);
11840 mutex_enter(&dtrace_lock);
11842 for (i = 0; i < dtrace_nprobes; i++) {
11843 if ((probe = dtrace_probes[i]) == NULL)
11846 if (probe->dtpr_ecb == NULL)
11849 prov = probe->dtpr_provider;
11851 if ((when = prov->dtpv_defunct) == 0)
11855 * We have ECBs on a defunct provider: we want to reap these
11856 * ECBs to allow the provider to unregister. The destruction
11857 * of these ECBs must be done carefully: if we destroy the ECB
11858 * and the consumer later wishes to consume an EPID that
11859 * corresponds to the destroyed ECB (and if the EPID metadata
11860 * has not been previously consumed), the consumer will abort
11861 * processing on the unknown EPID. To reduce (but not, sadly,
11862 * eliminate) the possibility of this, we will only destroy an
11863 * ECB for a defunct provider if, for the state that
11864 * corresponds to the ECB:
11866 * (a) There is no speculative tracing (which can effectively
11867 * cache an EPID for an arbitrary amount of time).
11869 * (b) The principal buffers have been switched twice since the
11870 * provider became defunct.
11872 * (c) The aggregation buffers are of zero size or have been
11873 * switched twice since the provider became defunct.
11875 * We use dts_speculates to determine (a) and call a function
11876 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
11877 * that as soon as we've been unable to destroy one of the ECBs
11878 * associated with the probe, we quit trying -- reaping is only
11879 * fruitful in as much as we can destroy all ECBs associated
11880 * with the defunct provider's probes.
11882 while ((ecb = probe->dtpr_ecb) != NULL) {
11883 dtrace_state_t *state = ecb->dte_state;
11884 dtrace_buffer_t *buf = state->dts_buffer;
11885 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
11887 if (state->dts_speculates)
11890 if (!dtrace_buffer_consumed(buf, when))
11893 if (!dtrace_buffer_consumed(aggbuf, when))
11896 dtrace_ecb_disable(ecb);
11897 ASSERT(probe->dtpr_ecb != ecb);
11898 dtrace_ecb_destroy(ecb);
11902 mutex_exit(&dtrace_lock);
11903 mutex_exit(&cpu_lock);
11907 * DTrace DOF Functions
11911 dtrace_dof_error(dof_hdr_t *dof, const char *str)
11913 if (dtrace_err_verbose)
11914 cmn_err(CE_WARN, "failed to process DOF: %s", str);
11916 #ifdef DTRACE_ERRDEBUG
11917 dtrace_errdebug(str);
11922 * Create DOF out of a currently enabled state. Right now, we only create
11923 * DOF containing the run-time options -- but this could be expanded to create
11924 * complete DOF representing the enabled state.
11927 dtrace_dof_create(dtrace_state_t *state)
11931 dof_optdesc_t *opt;
11932 int i, len = sizeof (dof_hdr_t) +
11933 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
11934 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11936 ASSERT(MUTEX_HELD(&dtrace_lock));
11938 dof = kmem_zalloc(len, KM_SLEEP);
11939 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
11940 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
11941 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
11942 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
11944 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
11945 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
11946 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
11947 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
11948 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
11949 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
11951 dof->dofh_flags = 0;
11952 dof->dofh_hdrsize = sizeof (dof_hdr_t);
11953 dof->dofh_secsize = sizeof (dof_sec_t);
11954 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
11955 dof->dofh_secoff = sizeof (dof_hdr_t);
11956 dof->dofh_loadsz = len;
11957 dof->dofh_filesz = len;
11961 * Fill in the option section header...
11963 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
11964 sec->dofs_type = DOF_SECT_OPTDESC;
11965 sec->dofs_align = sizeof (uint64_t);
11966 sec->dofs_flags = DOF_SECF_LOAD;
11967 sec->dofs_entsize = sizeof (dof_optdesc_t);
11969 opt = (dof_optdesc_t *)((uintptr_t)sec +
11970 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
11972 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
11973 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
11975 for (i = 0; i < DTRACEOPT_MAX; i++) {
11976 opt[i].dofo_option = i;
11977 opt[i].dofo_strtab = DOF_SECIDX_NONE;
11978 opt[i].dofo_value = state->dts_options[i];
11985 dtrace_dof_copyin(uintptr_t uarg, int *errp)
11987 dof_hdr_t hdr, *dof;
11989 ASSERT(!MUTEX_HELD(&dtrace_lock));
11992 * First, we're going to copyin() the sizeof (dof_hdr_t).
11994 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
11995 dtrace_dof_error(NULL, "failed to copyin DOF header");
12001 * Now we'll allocate the entire DOF and copy it in -- provided
12002 * that the length isn't outrageous.
12004 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
12005 dtrace_dof_error(&hdr, "load size exceeds maximum");
12010 if (hdr.dofh_loadsz < sizeof (hdr)) {
12011 dtrace_dof_error(&hdr, "invalid load size");
12016 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
12018 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0) {
12019 kmem_free(dof, hdr.dofh_loadsz);
12028 static __inline uchar_t
12029 dtrace_dof_char(char c) {
12048 return (c - 'A' + 10);
12055 return (c - 'a' + 10);
12057 /* Should not reach here. */
12063 dtrace_dof_property(const char *name)
12067 unsigned int len, i;
12072 * Unfortunately, array of values in .conf files are always (and
12073 * only) interpreted to be integer arrays. We must read our DOF
12074 * as an integer array, and then squeeze it into a byte array.
12076 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
12077 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
12080 for (i = 0; i < len; i++)
12081 buf[i] = (uchar_t)(((int *)buf)[i]);
12083 if (len < sizeof (dof_hdr_t)) {
12084 ddi_prop_free(buf);
12085 dtrace_dof_error(NULL, "truncated header");
12089 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
12090 ddi_prop_free(buf);
12091 dtrace_dof_error(NULL, "truncated DOF");
12095 if (loadsz >= dtrace_dof_maxsize) {
12096 ddi_prop_free(buf);
12097 dtrace_dof_error(NULL, "oversized DOF");
12101 dof = kmem_alloc(loadsz, KM_SLEEP);
12102 bcopy(buf, dof, loadsz);
12103 ddi_prop_free(buf);
12108 if ((p_env = getenv(name)) == NULL)
12111 len = strlen(p_env) / 2;
12113 buf = kmem_alloc(len, KM_SLEEP);
12115 dof = (dof_hdr_t *) buf;
12119 for (i = 0; i < len; i++) {
12120 buf[i] = (dtrace_dof_char(p[0]) << 4) |
12121 dtrace_dof_char(p[1]);
12127 if (len < sizeof (dof_hdr_t)) {
12129 dtrace_dof_error(NULL, "truncated header");
12133 if (len < (loadsz = dof->dofh_loadsz)) {
12135 dtrace_dof_error(NULL, "truncated DOF");
12139 if (loadsz >= dtrace_dof_maxsize) {
12141 dtrace_dof_error(NULL, "oversized DOF");
12150 dtrace_dof_destroy(dof_hdr_t *dof)
12152 kmem_free(dof, dof->dofh_loadsz);
12156 * Return the dof_sec_t pointer corresponding to a given section index. If the
12157 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
12158 * a type other than DOF_SECT_NONE is specified, the header is checked against
12159 * this type and NULL is returned if the types do not match.
12162 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
12164 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
12165 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
12167 if (i >= dof->dofh_secnum) {
12168 dtrace_dof_error(dof, "referenced section index is invalid");
12172 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
12173 dtrace_dof_error(dof, "referenced section is not loadable");
12177 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
12178 dtrace_dof_error(dof, "referenced section is the wrong type");
12185 static dtrace_probedesc_t *
12186 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
12188 dof_probedesc_t *probe;
12190 uintptr_t daddr = (uintptr_t)dof;
12194 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
12195 dtrace_dof_error(dof, "invalid probe section");
12199 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12200 dtrace_dof_error(dof, "bad alignment in probe description");
12204 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
12205 dtrace_dof_error(dof, "truncated probe description");
12209 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
12210 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
12212 if (strtab == NULL)
12215 str = daddr + strtab->dofs_offset;
12216 size = strtab->dofs_size;
12218 if (probe->dofp_provider >= strtab->dofs_size) {
12219 dtrace_dof_error(dof, "corrupt probe provider");
12223 (void) strncpy(desc->dtpd_provider,
12224 (char *)(str + probe->dofp_provider),
12225 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
12227 if (probe->dofp_mod >= strtab->dofs_size) {
12228 dtrace_dof_error(dof, "corrupt probe module");
12232 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
12233 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
12235 if (probe->dofp_func >= strtab->dofs_size) {
12236 dtrace_dof_error(dof, "corrupt probe function");
12240 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
12241 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
12243 if (probe->dofp_name >= strtab->dofs_size) {
12244 dtrace_dof_error(dof, "corrupt probe name");
12248 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
12249 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
12254 static dtrace_difo_t *
12255 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12260 dof_difohdr_t *dofd;
12261 uintptr_t daddr = (uintptr_t)dof;
12262 size_t max = dtrace_difo_maxsize;
12265 static const struct {
12273 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
12274 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
12275 sizeof (dif_instr_t), "multiple DIF sections" },
12277 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
12278 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
12279 sizeof (uint64_t), "multiple integer tables" },
12281 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
12282 offsetof(dtrace_difo_t, dtdo_strlen), 0,
12283 sizeof (char), "multiple string tables" },
12285 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
12286 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
12287 sizeof (uint_t), "multiple variable tables" },
12289 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
12292 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
12293 dtrace_dof_error(dof, "invalid DIFO header section");
12297 if (sec->dofs_align != sizeof (dof_secidx_t)) {
12298 dtrace_dof_error(dof, "bad alignment in DIFO header");
12302 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
12303 sec->dofs_size % sizeof (dof_secidx_t)) {
12304 dtrace_dof_error(dof, "bad size in DIFO header");
12308 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12309 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
12311 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
12312 dp->dtdo_rtype = dofd->dofd_rtype;
12314 for (l = 0; l < n; l++) {
12319 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
12320 dofd->dofd_links[l])) == NULL)
12321 goto err; /* invalid section link */
12323 if (ttl + subsec->dofs_size > max) {
12324 dtrace_dof_error(dof, "exceeds maximum size");
12328 ttl += subsec->dofs_size;
12330 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
12331 if (subsec->dofs_type != difo[i].section)
12334 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
12335 dtrace_dof_error(dof, "section not loaded");
12339 if (subsec->dofs_align != difo[i].align) {
12340 dtrace_dof_error(dof, "bad alignment");
12344 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
12345 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
12347 if (*bufp != NULL) {
12348 dtrace_dof_error(dof, difo[i].msg);
12352 if (difo[i].entsize != subsec->dofs_entsize) {
12353 dtrace_dof_error(dof, "entry size mismatch");
12357 if (subsec->dofs_entsize != 0 &&
12358 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
12359 dtrace_dof_error(dof, "corrupt entry size");
12363 *lenp = subsec->dofs_size;
12364 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
12365 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
12366 *bufp, subsec->dofs_size);
12368 if (subsec->dofs_entsize != 0)
12369 *lenp /= subsec->dofs_entsize;
12375 * If we encounter a loadable DIFO sub-section that is not
12376 * known to us, assume this is a broken program and fail.
12378 if (difo[i].section == DOF_SECT_NONE &&
12379 (subsec->dofs_flags & DOF_SECF_LOAD)) {
12380 dtrace_dof_error(dof, "unrecognized DIFO subsection");
12385 if (dp->dtdo_buf == NULL) {
12387 * We can't have a DIF object without DIF text.
12389 dtrace_dof_error(dof, "missing DIF text");
12394 * Before we validate the DIF object, run through the variable table
12395 * looking for the strings -- if any of their size are under, we'll set
12396 * their size to be the system-wide default string size. Note that
12397 * this should _not_ happen if the "strsize" option has been set --
12398 * in this case, the compiler should have set the size to reflect the
12399 * setting of the option.
12401 for (i = 0; i < dp->dtdo_varlen; i++) {
12402 dtrace_difv_t *v = &dp->dtdo_vartab[i];
12403 dtrace_diftype_t *t = &v->dtdv_type;
12405 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
12408 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
12409 t->dtdt_size = dtrace_strsize_default;
12412 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
12415 dtrace_difo_init(dp, vstate);
12419 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
12420 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
12421 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
12422 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
12424 kmem_free(dp, sizeof (dtrace_difo_t));
12428 static dtrace_predicate_t *
12429 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12434 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
12437 return (dtrace_predicate_create(dp));
12440 static dtrace_actdesc_t *
12441 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12444 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
12445 dof_actdesc_t *desc;
12446 dof_sec_t *difosec;
12448 uintptr_t daddr = (uintptr_t)dof;
12450 dtrace_actkind_t kind;
12452 if (sec->dofs_type != DOF_SECT_ACTDESC) {
12453 dtrace_dof_error(dof, "invalid action section");
12457 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
12458 dtrace_dof_error(dof, "truncated action description");
12462 if (sec->dofs_align != sizeof (uint64_t)) {
12463 dtrace_dof_error(dof, "bad alignment in action description");
12467 if (sec->dofs_size < sec->dofs_entsize) {
12468 dtrace_dof_error(dof, "section entry size exceeds total size");
12472 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
12473 dtrace_dof_error(dof, "bad entry size in action description");
12477 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
12478 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
12482 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
12483 desc = (dof_actdesc_t *)(daddr +
12484 (uintptr_t)sec->dofs_offset + offs);
12485 kind = (dtrace_actkind_t)desc->dofa_kind;
12487 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
12488 (kind != DTRACEACT_PRINTA ||
12489 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
12490 (kind == DTRACEACT_DIFEXPR &&
12491 desc->dofa_strtab != DOF_SECIDX_NONE)) {
12497 * The argument to these actions is an index into the
12498 * DOF string table. For printf()-like actions, this
12499 * is the format string. For print(), this is the
12500 * CTF type of the expression result.
12502 if ((strtab = dtrace_dof_sect(dof,
12503 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
12506 str = (char *)((uintptr_t)dof +
12507 (uintptr_t)strtab->dofs_offset);
12509 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
12510 if (str[i] == '\0')
12514 if (i >= strtab->dofs_size) {
12515 dtrace_dof_error(dof, "bogus format string");
12519 if (i == desc->dofa_arg) {
12520 dtrace_dof_error(dof, "empty format string");
12524 i -= desc->dofa_arg;
12525 fmt = kmem_alloc(i + 1, KM_SLEEP);
12526 bcopy(&str[desc->dofa_arg], fmt, i + 1);
12527 arg = (uint64_t)(uintptr_t)fmt;
12529 if (kind == DTRACEACT_PRINTA) {
12530 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
12533 arg = desc->dofa_arg;
12537 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
12538 desc->dofa_uarg, arg);
12540 if (last != NULL) {
12541 last->dtad_next = act;
12548 if (desc->dofa_difo == DOF_SECIDX_NONE)
12551 if ((difosec = dtrace_dof_sect(dof,
12552 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
12555 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
12557 if (act->dtad_difo == NULL)
12561 ASSERT(first != NULL);
12565 for (act = first; act != NULL; act = next) {
12566 next = act->dtad_next;
12567 dtrace_actdesc_release(act, vstate);
12573 static dtrace_ecbdesc_t *
12574 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
12577 dtrace_ecbdesc_t *ep;
12578 dof_ecbdesc_t *ecb;
12579 dtrace_probedesc_t *desc;
12580 dtrace_predicate_t *pred = NULL;
12582 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
12583 dtrace_dof_error(dof, "truncated ECB description");
12587 if (sec->dofs_align != sizeof (uint64_t)) {
12588 dtrace_dof_error(dof, "bad alignment in ECB description");
12592 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
12593 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
12598 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12599 ep->dted_uarg = ecb->dofe_uarg;
12600 desc = &ep->dted_probe;
12602 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
12605 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
12606 if ((sec = dtrace_dof_sect(dof,
12607 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
12610 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
12613 ep->dted_pred.dtpdd_predicate = pred;
12616 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
12617 if ((sec = dtrace_dof_sect(dof,
12618 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
12621 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
12623 if (ep->dted_action == NULL)
12631 dtrace_predicate_release(pred, vstate);
12632 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12637 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
12638 * specified DOF. At present, this amounts to simply adding 'ubase' to the
12639 * site of any user SETX relocations to account for load object base address.
12640 * In the future, if we need other relocations, this function can be extended.
12643 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase)
12645 uintptr_t daddr = (uintptr_t)dof;
12646 dof_relohdr_t *dofr =
12647 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
12648 dof_sec_t *ss, *rs, *ts;
12652 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
12653 sec->dofs_align != sizeof (dof_secidx_t)) {
12654 dtrace_dof_error(dof, "invalid relocation header");
12658 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
12659 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
12660 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
12662 if (ss == NULL || rs == NULL || ts == NULL)
12663 return (-1); /* dtrace_dof_error() has been called already */
12665 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
12666 rs->dofs_align != sizeof (uint64_t)) {
12667 dtrace_dof_error(dof, "invalid relocation section");
12671 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
12672 n = rs->dofs_size / rs->dofs_entsize;
12674 for (i = 0; i < n; i++) {
12675 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
12677 switch (r->dofr_type) {
12678 case DOF_RELO_NONE:
12680 case DOF_RELO_SETX:
12681 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
12682 sizeof (uint64_t) > ts->dofs_size) {
12683 dtrace_dof_error(dof, "bad relocation offset");
12687 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
12688 dtrace_dof_error(dof, "misaligned setx relo");
12692 *(uint64_t *)taddr += ubase;
12695 dtrace_dof_error(dof, "invalid relocation type");
12699 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
12706 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
12707 * header: it should be at the front of a memory region that is at least
12708 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
12709 * size. It need not be validated in any other way.
12712 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
12713 dtrace_enabling_t **enabp, uint64_t ubase, int noprobes)
12715 uint64_t len = dof->dofh_loadsz, seclen;
12716 uintptr_t daddr = (uintptr_t)dof;
12717 dtrace_ecbdesc_t *ep;
12718 dtrace_enabling_t *enab;
12721 ASSERT(MUTEX_HELD(&dtrace_lock));
12722 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
12725 * Check the DOF header identification bytes. In addition to checking
12726 * valid settings, we also verify that unused bits/bytes are zeroed so
12727 * we can use them later without fear of regressing existing binaries.
12729 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
12730 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
12731 dtrace_dof_error(dof, "DOF magic string mismatch");
12735 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
12736 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
12737 dtrace_dof_error(dof, "DOF has invalid data model");
12741 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
12742 dtrace_dof_error(dof, "DOF encoding mismatch");
12746 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
12747 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
12748 dtrace_dof_error(dof, "DOF version mismatch");
12752 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
12753 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
12757 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
12758 dtrace_dof_error(dof, "DOF uses too many integer registers");
12762 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
12763 dtrace_dof_error(dof, "DOF uses too many tuple registers");
12767 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
12768 if (dof->dofh_ident[i] != 0) {
12769 dtrace_dof_error(dof, "DOF has invalid ident byte set");
12774 if (dof->dofh_flags & ~DOF_FL_VALID) {
12775 dtrace_dof_error(dof, "DOF has invalid flag bits set");
12779 if (dof->dofh_secsize == 0) {
12780 dtrace_dof_error(dof, "zero section header size");
12785 * Check that the section headers don't exceed the amount of DOF
12786 * data. Note that we cast the section size and number of sections
12787 * to uint64_t's to prevent possible overflow in the multiplication.
12789 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
12791 if (dof->dofh_secoff > len || seclen > len ||
12792 dof->dofh_secoff + seclen > len) {
12793 dtrace_dof_error(dof, "truncated section headers");
12797 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
12798 dtrace_dof_error(dof, "misaligned section headers");
12802 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
12803 dtrace_dof_error(dof, "misaligned section size");
12808 * Take an initial pass through the section headers to be sure that
12809 * the headers don't have stray offsets. If the 'noprobes' flag is
12810 * set, do not permit sections relating to providers, probes, or args.
12812 for (i = 0; i < dof->dofh_secnum; i++) {
12813 dof_sec_t *sec = (dof_sec_t *)(daddr +
12814 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12817 switch (sec->dofs_type) {
12818 case DOF_SECT_PROVIDER:
12819 case DOF_SECT_PROBES:
12820 case DOF_SECT_PRARGS:
12821 case DOF_SECT_PROFFS:
12822 dtrace_dof_error(dof, "illegal sections "
12828 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12829 continue; /* just ignore non-loadable sections */
12831 if (sec->dofs_align & (sec->dofs_align - 1)) {
12832 dtrace_dof_error(dof, "bad section alignment");
12836 if (sec->dofs_offset & (sec->dofs_align - 1)) {
12837 dtrace_dof_error(dof, "misaligned section");
12841 if (sec->dofs_offset > len || sec->dofs_size > len ||
12842 sec->dofs_offset + sec->dofs_size > len) {
12843 dtrace_dof_error(dof, "corrupt section header");
12847 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
12848 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
12849 dtrace_dof_error(dof, "non-terminating string table");
12855 * Take a second pass through the sections and locate and perform any
12856 * relocations that are present. We do this after the first pass to
12857 * be sure that all sections have had their headers validated.
12859 for (i = 0; i < dof->dofh_secnum; i++) {
12860 dof_sec_t *sec = (dof_sec_t *)(daddr +
12861 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12863 if (!(sec->dofs_flags & DOF_SECF_LOAD))
12864 continue; /* skip sections that are not loadable */
12866 switch (sec->dofs_type) {
12867 case DOF_SECT_URELHDR:
12868 if (dtrace_dof_relocate(dof, sec, ubase) != 0)
12874 if ((enab = *enabp) == NULL)
12875 enab = *enabp = dtrace_enabling_create(vstate);
12877 for (i = 0; i < dof->dofh_secnum; i++) {
12878 dof_sec_t *sec = (dof_sec_t *)(daddr +
12879 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12881 if (sec->dofs_type != DOF_SECT_ECBDESC)
12884 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
12885 dtrace_enabling_destroy(enab);
12890 dtrace_enabling_add(enab, ep);
12897 * Process DOF for any options. This routine assumes that the DOF has been
12898 * at least processed by dtrace_dof_slurp().
12901 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
12906 dof_optdesc_t *desc;
12908 for (i = 0; i < dof->dofh_secnum; i++) {
12909 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
12910 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
12912 if (sec->dofs_type != DOF_SECT_OPTDESC)
12915 if (sec->dofs_align != sizeof (uint64_t)) {
12916 dtrace_dof_error(dof, "bad alignment in "
12917 "option description");
12921 if ((entsize = sec->dofs_entsize) == 0) {
12922 dtrace_dof_error(dof, "zeroed option entry size");
12926 if (entsize < sizeof (dof_optdesc_t)) {
12927 dtrace_dof_error(dof, "bad option entry size");
12931 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
12932 desc = (dof_optdesc_t *)((uintptr_t)dof +
12933 (uintptr_t)sec->dofs_offset + offs);
12935 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
12936 dtrace_dof_error(dof, "non-zero option string");
12940 if (desc->dofo_value == DTRACEOPT_UNSET) {
12941 dtrace_dof_error(dof, "unset option");
12945 if ((rval = dtrace_state_option(state,
12946 desc->dofo_option, desc->dofo_value)) != 0) {
12947 dtrace_dof_error(dof, "rejected option");
12957 * DTrace Consumer State Functions
12960 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
12962 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
12965 dtrace_dynvar_t *dvar, *next, *start;
12968 ASSERT(MUTEX_HELD(&dtrace_lock));
12969 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
12971 bzero(dstate, sizeof (dtrace_dstate_t));
12973 if ((dstate->dtds_chunksize = chunksize) == 0)
12974 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
12976 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
12979 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12982 dstate->dtds_size = size;
12983 dstate->dtds_base = base;
12984 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
12985 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
12987 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
12989 if (hashsize != 1 && (hashsize & 1))
12992 dstate->dtds_hashsize = hashsize;
12993 dstate->dtds_hash = dstate->dtds_base;
12996 * Set all of our hash buckets to point to the single sink, and (if
12997 * it hasn't already been set), set the sink's hash value to be the
12998 * sink sentinel value. The sink is needed for dynamic variable
12999 * lookups to know that they have iterated over an entire, valid hash
13002 for (i = 0; i < hashsize; i++)
13003 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
13005 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
13006 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
13009 * Determine number of active CPUs. Divide free list evenly among
13012 start = (dtrace_dynvar_t *)
13013 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
13014 limit = (uintptr_t)base + size;
13016 maxper = (limit - (uintptr_t)start) / NCPU;
13017 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
13022 for (i = 0; i < NCPU; i++) {
13024 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
13027 * If we don't even have enough chunks to make it once through
13028 * NCPUs, we're just going to allocate everything to the first
13029 * CPU. And if we're on the last CPU, we're going to allocate
13030 * whatever is left over. In either case, we set the limit to
13031 * be the limit of the dynamic variable space.
13033 if (maxper == 0 || i == NCPU - 1) {
13034 limit = (uintptr_t)base + size;
13037 limit = (uintptr_t)start + maxper;
13038 start = (dtrace_dynvar_t *)limit;
13041 ASSERT(limit <= (uintptr_t)base + size);
13044 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
13045 dstate->dtds_chunksize);
13047 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
13050 dvar->dtdv_next = next;
13062 dtrace_dstate_fini(dtrace_dstate_t *dstate)
13064 ASSERT(MUTEX_HELD(&cpu_lock));
13066 if (dstate->dtds_base == NULL)
13069 kmem_free(dstate->dtds_base, dstate->dtds_size);
13070 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
13074 dtrace_vstate_fini(dtrace_vstate_t *vstate)
13077 * Logical XOR, where are you?
13079 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
13081 if (vstate->dtvs_nglobals > 0) {
13082 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
13083 sizeof (dtrace_statvar_t *));
13086 if (vstate->dtvs_ntlocals > 0) {
13087 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
13088 sizeof (dtrace_difv_t));
13091 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
13093 if (vstate->dtvs_nlocals > 0) {
13094 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
13095 sizeof (dtrace_statvar_t *));
13101 dtrace_state_clean(dtrace_state_t *state)
13103 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13106 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13107 dtrace_speculation_clean(state);
13111 dtrace_state_deadman(dtrace_state_t *state)
13117 now = dtrace_gethrtime();
13119 if (state != dtrace_anon.dta_state &&
13120 now - state->dts_laststatus >= dtrace_deadman_user)
13124 * We must be sure that dts_alive never appears to be less than the
13125 * value upon entry to dtrace_state_deadman(), and because we lack a
13126 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13127 * store INT64_MAX to it, followed by a memory barrier, followed by
13128 * the new value. This assures that dts_alive never appears to be
13129 * less than its true value, regardless of the order in which the
13130 * stores to the underlying storage are issued.
13132 state->dts_alive = INT64_MAX;
13133 dtrace_membar_producer();
13134 state->dts_alive = now;
13138 dtrace_state_clean(void *arg)
13140 dtrace_state_t *state = arg;
13141 dtrace_optval_t *opt = state->dts_options;
13143 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
13146 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
13147 dtrace_speculation_clean(state);
13149 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13150 dtrace_state_clean, state);
13154 dtrace_state_deadman(void *arg)
13156 dtrace_state_t *state = arg;
13161 dtrace_debug_output();
13163 now = dtrace_gethrtime();
13165 if (state != dtrace_anon.dta_state &&
13166 now - state->dts_laststatus >= dtrace_deadman_user)
13170 * We must be sure that dts_alive never appears to be less than the
13171 * value upon entry to dtrace_state_deadman(), and because we lack a
13172 * dtrace_cas64(), we cannot store to it atomically. We thus instead
13173 * store INT64_MAX to it, followed by a memory barrier, followed by
13174 * the new value. This assures that dts_alive never appears to be
13175 * less than its true value, regardless of the order in which the
13176 * stores to the underlying storage are issued.
13178 state->dts_alive = INT64_MAX;
13179 dtrace_membar_producer();
13180 state->dts_alive = now;
13182 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13183 dtrace_state_deadman, state);
13187 static dtrace_state_t *
13189 dtrace_state_create(dev_t *devp, cred_t *cr)
13191 dtrace_state_create(struct cdev *dev)
13202 dtrace_state_t *state;
13203 dtrace_optval_t *opt;
13204 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
13206 ASSERT(MUTEX_HELD(&dtrace_lock));
13207 ASSERT(MUTEX_HELD(&cpu_lock));
13210 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
13211 VM_BESTFIT | VM_SLEEP);
13213 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
13214 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
13218 state = ddi_get_soft_state(dtrace_softstate, minor);
13225 /* Allocate memory for the state. */
13226 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
13229 state->dts_epid = DTRACE_EPIDNONE + 1;
13231 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
13233 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
13234 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
13236 if (devp != NULL) {
13237 major = getemajor(*devp);
13239 major = ddi_driver_major(dtrace_devi);
13242 state->dts_dev = makedevice(major, minor);
13245 *devp = state->dts_dev;
13247 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
13248 state->dts_dev = dev;
13252 * We allocate NCPU buffers. On the one hand, this can be quite
13253 * a bit of memory per instance (nearly 36K on a Starcat). On the
13254 * other hand, it saves an additional memory reference in the probe
13257 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
13258 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
13261 state->dts_cleaner = CYCLIC_NONE;
13262 state->dts_deadman = CYCLIC_NONE;
13264 callout_init(&state->dts_cleaner, CALLOUT_MPSAFE);
13265 callout_init(&state->dts_deadman, CALLOUT_MPSAFE);
13267 state->dts_vstate.dtvs_state = state;
13269 for (i = 0; i < DTRACEOPT_MAX; i++)
13270 state->dts_options[i] = DTRACEOPT_UNSET;
13273 * Set the default options.
13275 opt = state->dts_options;
13276 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
13277 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
13278 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
13279 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
13280 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
13281 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
13282 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
13283 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
13284 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
13285 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
13286 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
13287 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
13288 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
13289 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
13291 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
13294 * Depending on the user credentials, we set flag bits which alter probe
13295 * visibility or the amount of destructiveness allowed. In the case of
13296 * actual anonymous tracing, or the possession of all privileges, all of
13297 * the normal checks are bypassed.
13299 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
13300 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
13301 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
13304 * Set up the credentials for this instantiation. We take a
13305 * hold on the credential to prevent it from disappearing on
13306 * us; this in turn prevents the zone_t referenced by this
13307 * credential from disappearing. This means that we can
13308 * examine the credential and the zone from probe context.
13311 state->dts_cred.dcr_cred = cr;
13314 * CRA_PROC means "we have *some* privilege for dtrace" and
13315 * unlocks the use of variables like pid, zonename, etc.
13317 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
13318 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13319 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
13323 * dtrace_user allows use of syscall and profile providers.
13324 * If the user also has proc_owner and/or proc_zone, we
13325 * extend the scope to include additional visibility and
13326 * destructive power.
13328 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
13329 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
13330 state->dts_cred.dcr_visible |=
13331 DTRACE_CRV_ALLPROC;
13333 state->dts_cred.dcr_action |=
13334 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13337 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
13338 state->dts_cred.dcr_visible |=
13339 DTRACE_CRV_ALLZONE;
13341 state->dts_cred.dcr_action |=
13342 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13346 * If we have all privs in whatever zone this is,
13347 * we can do destructive things to processes which
13348 * have altered credentials.
13351 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13352 cr->cr_zone->zone_privset)) {
13353 state->dts_cred.dcr_action |=
13354 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13360 * Holding the dtrace_kernel privilege also implies that
13361 * the user has the dtrace_user privilege from a visibility
13362 * perspective. But without further privileges, some
13363 * destructive actions are not available.
13365 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
13367 * Make all probes in all zones visible. However,
13368 * this doesn't mean that all actions become available
13371 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
13372 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
13374 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
13377 * Holding proc_owner means that destructive actions
13378 * for *this* zone are allowed.
13380 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13381 state->dts_cred.dcr_action |=
13382 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13385 * Holding proc_zone means that destructive actions
13386 * for this user/group ID in all zones is allowed.
13388 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13389 state->dts_cred.dcr_action |=
13390 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13394 * If we have all privs in whatever zone this is,
13395 * we can do destructive things to processes which
13396 * have altered credentials.
13398 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
13399 cr->cr_zone->zone_privset)) {
13400 state->dts_cred.dcr_action |=
13401 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
13407 * Holding the dtrace_proc privilege gives control over fasttrap
13408 * and pid providers. We need to grant wider destructive
13409 * privileges in the event that the user has proc_owner and/or
13412 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
13413 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
13414 state->dts_cred.dcr_action |=
13415 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
13417 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
13418 state->dts_cred.dcr_action |=
13419 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
13427 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
13429 dtrace_optval_t *opt = state->dts_options, size;
13430 processorid_t cpu = 0;;
13431 int flags = 0, rval, factor, divisor = 1;
13433 ASSERT(MUTEX_HELD(&dtrace_lock));
13434 ASSERT(MUTEX_HELD(&cpu_lock));
13435 ASSERT(which < DTRACEOPT_MAX);
13436 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
13437 (state == dtrace_anon.dta_state &&
13438 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
13440 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
13443 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
13444 cpu = opt[DTRACEOPT_CPU];
13446 if (which == DTRACEOPT_SPECSIZE)
13447 flags |= DTRACEBUF_NOSWITCH;
13449 if (which == DTRACEOPT_BUFSIZE) {
13450 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
13451 flags |= DTRACEBUF_RING;
13453 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
13454 flags |= DTRACEBUF_FILL;
13456 if (state != dtrace_anon.dta_state ||
13457 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
13458 flags |= DTRACEBUF_INACTIVE;
13461 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
13463 * The size must be 8-byte aligned. If the size is not 8-byte
13464 * aligned, drop it down by the difference.
13466 if (size & (sizeof (uint64_t) - 1))
13467 size -= size & (sizeof (uint64_t) - 1);
13469 if (size < state->dts_reserve) {
13471 * Buffers always must be large enough to accommodate
13472 * their prereserved space. We return E2BIG instead
13473 * of ENOMEM in this case to allow for user-level
13474 * software to differentiate the cases.
13479 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
13481 if (rval != ENOMEM) {
13486 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13489 for (divisor = 2; divisor < factor; divisor <<= 1)
13497 dtrace_state_buffers(dtrace_state_t *state)
13499 dtrace_speculation_t *spec = state->dts_speculations;
13502 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
13503 DTRACEOPT_BUFSIZE)) != 0)
13506 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
13507 DTRACEOPT_AGGSIZE)) != 0)
13510 for (i = 0; i < state->dts_nspeculations; i++) {
13511 if ((rval = dtrace_state_buffer(state,
13512 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
13520 dtrace_state_prereserve(dtrace_state_t *state)
13523 dtrace_probe_t *probe;
13525 state->dts_reserve = 0;
13527 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
13531 * If our buffer policy is a "fill" buffer policy, we need to set the
13532 * prereserved space to be the space required by the END probes.
13534 probe = dtrace_probes[dtrace_probeid_end - 1];
13535 ASSERT(probe != NULL);
13537 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
13538 if (ecb->dte_state != state)
13541 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
13546 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
13548 dtrace_optval_t *opt = state->dts_options, sz, nspec;
13549 dtrace_speculation_t *spec;
13550 dtrace_buffer_t *buf;
13552 cyc_handler_t hdlr;
13555 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13556 dtrace_icookie_t cookie;
13558 mutex_enter(&cpu_lock);
13559 mutex_enter(&dtrace_lock);
13561 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
13567 * Before we can perform any checks, we must prime all of the
13568 * retained enablings that correspond to this state.
13570 dtrace_enabling_prime(state);
13572 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
13577 dtrace_state_prereserve(state);
13580 * Now we want to do is try to allocate our speculations.
13581 * We do not automatically resize the number of speculations; if
13582 * this fails, we will fail the operation.
13584 nspec = opt[DTRACEOPT_NSPEC];
13585 ASSERT(nspec != DTRACEOPT_UNSET);
13587 if (nspec > INT_MAX) {
13592 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
13593 KM_NOSLEEP | KM_NORMALPRI);
13595 if (spec == NULL) {
13600 state->dts_speculations = spec;
13601 state->dts_nspeculations = (int)nspec;
13603 for (i = 0; i < nspec; i++) {
13604 if ((buf = kmem_zalloc(bufsize,
13605 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
13610 spec[i].dtsp_buffer = buf;
13613 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
13614 if (dtrace_anon.dta_state == NULL) {
13619 if (state->dts_necbs != 0) {
13624 state->dts_anon = dtrace_anon_grab();
13625 ASSERT(state->dts_anon != NULL);
13626 state = state->dts_anon;
13629 * We want "grabanon" to be set in the grabbed state, so we'll
13630 * copy that option value from the grabbing state into the
13633 state->dts_options[DTRACEOPT_GRABANON] =
13634 opt[DTRACEOPT_GRABANON];
13636 *cpu = dtrace_anon.dta_beganon;
13639 * If the anonymous state is active (as it almost certainly
13640 * is if the anonymous enabling ultimately matched anything),
13641 * we don't allow any further option processing -- but we
13642 * don't return failure.
13644 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13648 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
13649 opt[DTRACEOPT_AGGSIZE] != 0) {
13650 if (state->dts_aggregations == NULL) {
13652 * We're not going to create an aggregation buffer
13653 * because we don't have any ECBs that contain
13654 * aggregations -- set this option to 0.
13656 opt[DTRACEOPT_AGGSIZE] = 0;
13659 * If we have an aggregation buffer, we must also have
13660 * a buffer to use as scratch.
13662 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
13663 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
13664 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
13669 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
13670 opt[DTRACEOPT_SPECSIZE] != 0) {
13671 if (!state->dts_speculates) {
13673 * We're not going to create speculation buffers
13674 * because we don't have any ECBs that actually
13675 * speculate -- set the speculation size to 0.
13677 opt[DTRACEOPT_SPECSIZE] = 0;
13682 * The bare minimum size for any buffer that we're actually going to
13683 * do anything to is sizeof (uint64_t).
13685 sz = sizeof (uint64_t);
13687 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
13688 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
13689 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
13691 * A buffer size has been explicitly set to 0 (or to a size
13692 * that will be adjusted to 0) and we need the space -- we
13693 * need to return failure. We return ENOSPC to differentiate
13694 * it from failing to allocate a buffer due to failure to meet
13695 * the reserve (for which we return E2BIG).
13701 if ((rval = dtrace_state_buffers(state)) != 0)
13704 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
13705 sz = dtrace_dstate_defsize;
13708 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
13713 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
13715 } while (sz >>= 1);
13717 opt[DTRACEOPT_DYNVARSIZE] = sz;
13722 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
13723 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
13725 if (opt[DTRACEOPT_CLEANRATE] == 0)
13726 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13728 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
13729 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
13731 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
13732 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
13734 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
13736 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
13737 hdlr.cyh_arg = state;
13738 hdlr.cyh_level = CY_LOW_LEVEL;
13741 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
13743 state->dts_cleaner = cyclic_add(&hdlr, &when);
13745 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
13746 hdlr.cyh_arg = state;
13747 hdlr.cyh_level = CY_LOW_LEVEL;
13750 when.cyt_interval = dtrace_deadman_interval;
13752 state->dts_deadman = cyclic_add(&hdlr, &when);
13754 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
13755 dtrace_state_clean, state);
13756 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
13757 dtrace_state_deadman, state);
13760 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
13763 * Now it's time to actually fire the BEGIN probe. We need to disable
13764 * interrupts here both to record the CPU on which we fired the BEGIN
13765 * probe (the data from this CPU will be processed first at user
13766 * level) and to manually activate the buffer for this CPU.
13768 cookie = dtrace_interrupt_disable();
13770 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
13771 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
13773 dtrace_probe(dtrace_probeid_begin,
13774 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13775 dtrace_interrupt_enable(cookie);
13777 * We may have had an exit action from a BEGIN probe; only change our
13778 * state to ACTIVE if we're still in WARMUP.
13780 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
13781 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
13783 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
13784 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
13787 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
13788 * want each CPU to transition its principal buffer out of the
13789 * INACTIVE state. Doing this assures that no CPU will suddenly begin
13790 * processing an ECB halfway down a probe's ECB chain; all CPUs will
13791 * atomically transition from processing none of a state's ECBs to
13792 * processing all of them.
13794 dtrace_xcall(DTRACE_CPUALL,
13795 (dtrace_xcall_t)dtrace_buffer_activate, state);
13799 dtrace_buffer_free(state->dts_buffer);
13800 dtrace_buffer_free(state->dts_aggbuffer);
13802 if ((nspec = state->dts_nspeculations) == 0) {
13803 ASSERT(state->dts_speculations == NULL);
13807 spec = state->dts_speculations;
13808 ASSERT(spec != NULL);
13810 for (i = 0; i < state->dts_nspeculations; i++) {
13811 if ((buf = spec[i].dtsp_buffer) == NULL)
13814 dtrace_buffer_free(buf);
13815 kmem_free(buf, bufsize);
13818 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
13819 state->dts_nspeculations = 0;
13820 state->dts_speculations = NULL;
13823 mutex_exit(&dtrace_lock);
13824 mutex_exit(&cpu_lock);
13830 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
13832 dtrace_icookie_t cookie;
13834 ASSERT(MUTEX_HELD(&dtrace_lock));
13836 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
13837 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
13841 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
13842 * to be sure that every CPU has seen it. See below for the details
13843 * on why this is done.
13845 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
13849 * By this point, it is impossible for any CPU to be still processing
13850 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
13851 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
13852 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
13853 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
13854 * iff we're in the END probe.
13856 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
13858 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
13861 * Finally, we can release the reserve and call the END probe. We
13862 * disable interrupts across calling the END probe to allow us to
13863 * return the CPU on which we actually called the END probe. This
13864 * allows user-land to be sure that this CPU's principal buffer is
13867 state->dts_reserve = 0;
13869 cookie = dtrace_interrupt_disable();
13871 dtrace_probe(dtrace_probeid_end,
13872 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
13873 dtrace_interrupt_enable(cookie);
13875 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
13882 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
13883 dtrace_optval_t val)
13885 ASSERT(MUTEX_HELD(&dtrace_lock));
13887 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
13890 if (option >= DTRACEOPT_MAX)
13893 if (option != DTRACEOPT_CPU && val < 0)
13897 case DTRACEOPT_DESTRUCTIVE:
13898 if (dtrace_destructive_disallow)
13901 state->dts_cred.dcr_destructive = 1;
13904 case DTRACEOPT_BUFSIZE:
13905 case DTRACEOPT_DYNVARSIZE:
13906 case DTRACEOPT_AGGSIZE:
13907 case DTRACEOPT_SPECSIZE:
13908 case DTRACEOPT_STRSIZE:
13912 if (val >= LONG_MAX) {
13914 * If this is an otherwise negative value, set it to
13915 * the highest multiple of 128m less than LONG_MAX.
13916 * Technically, we're adjusting the size without
13917 * regard to the buffer resizing policy, but in fact,
13918 * this has no effect -- if we set the buffer size to
13919 * ~LONG_MAX and the buffer policy is ultimately set to
13920 * be "manual", the buffer allocation is guaranteed to
13921 * fail, if only because the allocation requires two
13922 * buffers. (We set the the size to the highest
13923 * multiple of 128m because it ensures that the size
13924 * will remain a multiple of a megabyte when
13925 * repeatedly halved -- all the way down to 15m.)
13927 val = LONG_MAX - (1 << 27) + 1;
13931 state->dts_options[option] = val;
13937 dtrace_state_destroy(dtrace_state_t *state)
13940 dtrace_vstate_t *vstate = &state->dts_vstate;
13942 minor_t minor = getminor(state->dts_dev);
13944 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
13945 dtrace_speculation_t *spec = state->dts_speculations;
13946 int nspec = state->dts_nspeculations;
13949 ASSERT(MUTEX_HELD(&dtrace_lock));
13950 ASSERT(MUTEX_HELD(&cpu_lock));
13953 * First, retract any retained enablings for this state.
13955 dtrace_enabling_retract(state);
13956 ASSERT(state->dts_nretained == 0);
13958 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
13959 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
13961 * We have managed to come into dtrace_state_destroy() on a
13962 * hot enabling -- almost certainly because of a disorderly
13963 * shutdown of a consumer. (That is, a consumer that is
13964 * exiting without having called dtrace_stop().) In this case,
13965 * we're going to set our activity to be KILLED, and then
13966 * issue a sync to be sure that everyone is out of probe
13967 * context before we start blowing away ECBs.
13969 state->dts_activity = DTRACE_ACTIVITY_KILLED;
13974 * Release the credential hold we took in dtrace_state_create().
13976 if (state->dts_cred.dcr_cred != NULL)
13977 crfree(state->dts_cred.dcr_cred);
13980 * Now we can safely disable and destroy any enabled probes. Because
13981 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
13982 * (especially if they're all enabled), we take two passes through the
13983 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
13984 * in the second we disable whatever is left over.
13986 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
13987 for (i = 0; i < state->dts_necbs; i++) {
13988 if ((ecb = state->dts_ecbs[i]) == NULL)
13991 if (match && ecb->dte_probe != NULL) {
13992 dtrace_probe_t *probe = ecb->dte_probe;
13993 dtrace_provider_t *prov = probe->dtpr_provider;
13995 if (!(prov->dtpv_priv.dtpp_flags & match))
13999 dtrace_ecb_disable(ecb);
14000 dtrace_ecb_destroy(ecb);
14008 * Before we free the buffers, perform one more sync to assure that
14009 * every CPU is out of probe context.
14013 dtrace_buffer_free(state->dts_buffer);
14014 dtrace_buffer_free(state->dts_aggbuffer);
14016 for (i = 0; i < nspec; i++)
14017 dtrace_buffer_free(spec[i].dtsp_buffer);
14020 if (state->dts_cleaner != CYCLIC_NONE)
14021 cyclic_remove(state->dts_cleaner);
14023 if (state->dts_deadman != CYCLIC_NONE)
14024 cyclic_remove(state->dts_deadman);
14026 callout_stop(&state->dts_cleaner);
14027 callout_drain(&state->dts_cleaner);
14028 callout_stop(&state->dts_deadman);
14029 callout_drain(&state->dts_deadman);
14032 dtrace_dstate_fini(&vstate->dtvs_dynvars);
14033 dtrace_vstate_fini(vstate);
14034 if (state->dts_ecbs != NULL)
14035 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
14037 if (state->dts_aggregations != NULL) {
14039 for (i = 0; i < state->dts_naggregations; i++)
14040 ASSERT(state->dts_aggregations[i] == NULL);
14042 ASSERT(state->dts_naggregations > 0);
14043 kmem_free(state->dts_aggregations,
14044 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
14047 kmem_free(state->dts_buffer, bufsize);
14048 kmem_free(state->dts_aggbuffer, bufsize);
14050 for (i = 0; i < nspec; i++)
14051 kmem_free(spec[i].dtsp_buffer, bufsize);
14054 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
14056 dtrace_format_destroy(state);
14058 if (state->dts_aggid_arena != NULL) {
14060 vmem_destroy(state->dts_aggid_arena);
14062 delete_unrhdr(state->dts_aggid_arena);
14064 state->dts_aggid_arena = NULL;
14067 ddi_soft_state_free(dtrace_softstate, minor);
14068 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14073 * DTrace Anonymous Enabling Functions
14075 static dtrace_state_t *
14076 dtrace_anon_grab(void)
14078 dtrace_state_t *state;
14080 ASSERT(MUTEX_HELD(&dtrace_lock));
14082 if ((state = dtrace_anon.dta_state) == NULL) {
14083 ASSERT(dtrace_anon.dta_enabling == NULL);
14087 ASSERT(dtrace_anon.dta_enabling != NULL);
14088 ASSERT(dtrace_retained != NULL);
14090 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
14091 dtrace_anon.dta_enabling = NULL;
14092 dtrace_anon.dta_state = NULL;
14098 dtrace_anon_property(void)
14101 dtrace_state_t *state;
14103 char c[32]; /* enough for "dof-data-" + digits */
14105 ASSERT(MUTEX_HELD(&dtrace_lock));
14106 ASSERT(MUTEX_HELD(&cpu_lock));
14108 for (i = 0; ; i++) {
14109 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
14111 dtrace_err_verbose = 1;
14113 if ((dof = dtrace_dof_property(c)) == NULL) {
14114 dtrace_err_verbose = 0;
14120 * We want to create anonymous state, so we need to transition
14121 * the kernel debugger to indicate that DTrace is active. If
14122 * this fails (e.g. because the debugger has modified text in
14123 * some way), we won't continue with the processing.
14125 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
14126 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
14127 "enabling ignored.");
14128 dtrace_dof_destroy(dof);
14134 * If we haven't allocated an anonymous state, we'll do so now.
14136 if ((state = dtrace_anon.dta_state) == NULL) {
14138 state = dtrace_state_create(NULL, NULL);
14140 state = dtrace_state_create(NULL);
14142 dtrace_anon.dta_state = state;
14144 if (state == NULL) {
14146 * This basically shouldn't happen: the only
14147 * failure mode from dtrace_state_create() is a
14148 * failure of ddi_soft_state_zalloc() that
14149 * itself should never happen. Still, the
14150 * interface allows for a failure mode, and
14151 * we want to fail as gracefully as possible:
14152 * we'll emit an error message and cease
14153 * processing anonymous state in this case.
14155 cmn_err(CE_WARN, "failed to create "
14156 "anonymous state");
14157 dtrace_dof_destroy(dof);
14162 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
14163 &dtrace_anon.dta_enabling, 0, B_TRUE);
14166 rv = dtrace_dof_options(dof, state);
14168 dtrace_err_verbose = 0;
14169 dtrace_dof_destroy(dof);
14173 * This is malformed DOF; chuck any anonymous state
14176 ASSERT(dtrace_anon.dta_enabling == NULL);
14177 dtrace_state_destroy(state);
14178 dtrace_anon.dta_state = NULL;
14182 ASSERT(dtrace_anon.dta_enabling != NULL);
14185 if (dtrace_anon.dta_enabling != NULL) {
14189 * dtrace_enabling_retain() can only fail because we are
14190 * trying to retain more enablings than are allowed -- but
14191 * we only have one anonymous enabling, and we are guaranteed
14192 * to be allowed at least one retained enabling; we assert
14193 * that dtrace_enabling_retain() returns success.
14195 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
14198 dtrace_enabling_dump(dtrace_anon.dta_enabling);
14203 * DTrace Helper Functions
14206 dtrace_helper_trace(dtrace_helper_action_t *helper,
14207 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
14209 uint32_t size, next, nnext, i;
14210 dtrace_helptrace_t *ent;
14211 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
14213 if (!dtrace_helptrace_enabled)
14216 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
14219 * What would a tracing framework be without its own tracing
14220 * framework? (Well, a hell of a lot simpler, for starters...)
14222 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
14223 sizeof (uint64_t) - sizeof (uint64_t);
14226 * Iterate until we can allocate a slot in the trace buffer.
14229 next = dtrace_helptrace_next;
14231 if (next + size < dtrace_helptrace_bufsize) {
14232 nnext = next + size;
14236 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
14239 * We have our slot; fill it in.
14244 ent = (dtrace_helptrace_t *)&dtrace_helptrace_buffer[next];
14245 ent->dtht_helper = helper;
14246 ent->dtht_where = where;
14247 ent->dtht_nlocals = vstate->dtvs_nlocals;
14249 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
14250 mstate->dtms_fltoffs : -1;
14251 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
14252 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
14254 for (i = 0; i < vstate->dtvs_nlocals; i++) {
14255 dtrace_statvar_t *svar;
14257 if ((svar = vstate->dtvs_locals[i]) == NULL)
14260 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
14261 ent->dtht_locals[i] =
14262 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
14267 dtrace_helper(int which, dtrace_mstate_t *mstate,
14268 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
14270 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
14271 uint64_t sarg0 = mstate->dtms_arg[0];
14272 uint64_t sarg1 = mstate->dtms_arg[1];
14274 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
14275 dtrace_helper_action_t *helper;
14276 dtrace_vstate_t *vstate;
14277 dtrace_difo_t *pred;
14278 int i, trace = dtrace_helptrace_enabled;
14280 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
14282 if (helpers == NULL)
14285 if ((helper = helpers->dthps_actions[which]) == NULL)
14288 vstate = &helpers->dthps_vstate;
14289 mstate->dtms_arg[0] = arg0;
14290 mstate->dtms_arg[1] = arg1;
14293 * Now iterate over each helper. If its predicate evaluates to 'true',
14294 * we'll call the corresponding actions. Note that the below calls
14295 * to dtrace_dif_emulate() may set faults in machine state. This is
14296 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
14297 * the stored DIF offset with its own (which is the desired behavior).
14298 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
14299 * from machine state; this is okay, too.
14301 for (; helper != NULL; helper = helper->dtha_next) {
14302 if ((pred = helper->dtha_predicate) != NULL) {
14304 dtrace_helper_trace(helper, mstate, vstate, 0);
14306 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
14309 if (*flags & CPU_DTRACE_FAULT)
14313 for (i = 0; i < helper->dtha_nactions; i++) {
14315 dtrace_helper_trace(helper,
14316 mstate, vstate, i + 1);
14318 rval = dtrace_dif_emulate(helper->dtha_actions[i],
14319 mstate, vstate, state);
14321 if (*flags & CPU_DTRACE_FAULT)
14327 dtrace_helper_trace(helper, mstate, vstate,
14328 DTRACE_HELPTRACE_NEXT);
14332 dtrace_helper_trace(helper, mstate, vstate,
14333 DTRACE_HELPTRACE_DONE);
14336 * Restore the arg0 that we saved upon entry.
14338 mstate->dtms_arg[0] = sarg0;
14339 mstate->dtms_arg[1] = sarg1;
14345 dtrace_helper_trace(helper, mstate, vstate,
14346 DTRACE_HELPTRACE_ERR);
14349 * Restore the arg0 that we saved upon entry.
14351 mstate->dtms_arg[0] = sarg0;
14352 mstate->dtms_arg[1] = sarg1;
14358 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
14359 dtrace_vstate_t *vstate)
14363 if (helper->dtha_predicate != NULL)
14364 dtrace_difo_release(helper->dtha_predicate, vstate);
14366 for (i = 0; i < helper->dtha_nactions; i++) {
14367 ASSERT(helper->dtha_actions[i] != NULL);
14368 dtrace_difo_release(helper->dtha_actions[i], vstate);
14371 kmem_free(helper->dtha_actions,
14372 helper->dtha_nactions * sizeof (dtrace_difo_t *));
14373 kmem_free(helper, sizeof (dtrace_helper_action_t));
14377 dtrace_helper_destroygen(int gen)
14379 proc_t *p = curproc;
14380 dtrace_helpers_t *help = p->p_dtrace_helpers;
14381 dtrace_vstate_t *vstate;
14384 ASSERT(MUTEX_HELD(&dtrace_lock));
14386 if (help == NULL || gen > help->dthps_generation)
14389 vstate = &help->dthps_vstate;
14391 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
14392 dtrace_helper_action_t *last = NULL, *h, *next;
14394 for (h = help->dthps_actions[i]; h != NULL; h = next) {
14395 next = h->dtha_next;
14397 if (h->dtha_generation == gen) {
14398 if (last != NULL) {
14399 last->dtha_next = next;
14401 help->dthps_actions[i] = next;
14404 dtrace_helper_action_destroy(h, vstate);
14412 * Interate until we've cleared out all helper providers with the
14413 * given generation number.
14416 dtrace_helper_provider_t *prov;
14419 * Look for a helper provider with the right generation. We
14420 * have to start back at the beginning of the list each time
14421 * because we drop dtrace_lock. It's unlikely that we'll make
14422 * more than two passes.
14424 for (i = 0; i < help->dthps_nprovs; i++) {
14425 prov = help->dthps_provs[i];
14427 if (prov->dthp_generation == gen)
14432 * If there were no matches, we're done.
14434 if (i == help->dthps_nprovs)
14438 * Move the last helper provider into this slot.
14440 help->dthps_nprovs--;
14441 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
14442 help->dthps_provs[help->dthps_nprovs] = NULL;
14444 mutex_exit(&dtrace_lock);
14447 * If we have a meta provider, remove this helper provider.
14449 mutex_enter(&dtrace_meta_lock);
14450 if (dtrace_meta_pid != NULL) {
14451 ASSERT(dtrace_deferred_pid == NULL);
14452 dtrace_helper_provider_remove(&prov->dthp_prov,
14455 mutex_exit(&dtrace_meta_lock);
14457 dtrace_helper_provider_destroy(prov);
14459 mutex_enter(&dtrace_lock);
14466 dtrace_helper_validate(dtrace_helper_action_t *helper)
14471 if ((dp = helper->dtha_predicate) != NULL)
14472 err += dtrace_difo_validate_helper(dp);
14474 for (i = 0; i < helper->dtha_nactions; i++)
14475 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
14481 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep)
14483 dtrace_helpers_t *help;
14484 dtrace_helper_action_t *helper, *last;
14485 dtrace_actdesc_t *act;
14486 dtrace_vstate_t *vstate;
14487 dtrace_predicate_t *pred;
14488 int count = 0, nactions = 0, i;
14490 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
14493 help = curproc->p_dtrace_helpers;
14494 last = help->dthps_actions[which];
14495 vstate = &help->dthps_vstate;
14497 for (count = 0; last != NULL; last = last->dtha_next) {
14499 if (last->dtha_next == NULL)
14504 * If we already have dtrace_helper_actions_max helper actions for this
14505 * helper action type, we'll refuse to add a new one.
14507 if (count >= dtrace_helper_actions_max)
14510 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
14511 helper->dtha_generation = help->dthps_generation;
14513 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
14514 ASSERT(pred->dtp_difo != NULL);
14515 dtrace_difo_hold(pred->dtp_difo);
14516 helper->dtha_predicate = pred->dtp_difo;
14519 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
14520 if (act->dtad_kind != DTRACEACT_DIFEXPR)
14523 if (act->dtad_difo == NULL)
14529 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
14530 (helper->dtha_nactions = nactions), KM_SLEEP);
14532 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
14533 dtrace_difo_hold(act->dtad_difo);
14534 helper->dtha_actions[i++] = act->dtad_difo;
14537 if (!dtrace_helper_validate(helper))
14540 if (last == NULL) {
14541 help->dthps_actions[which] = helper;
14543 last->dtha_next = helper;
14546 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
14547 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
14548 dtrace_helptrace_next = 0;
14553 dtrace_helper_action_destroy(helper, vstate);
14558 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
14559 dof_helper_t *dofhp)
14561 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
14563 mutex_enter(&dtrace_meta_lock);
14564 mutex_enter(&dtrace_lock);
14566 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
14568 * If the dtrace module is loaded but not attached, or if
14569 * there aren't isn't a meta provider registered to deal with
14570 * these provider descriptions, we need to postpone creating
14571 * the actual providers until later.
14574 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
14575 dtrace_deferred_pid != help) {
14576 help->dthps_deferred = 1;
14577 help->dthps_pid = p->p_pid;
14578 help->dthps_next = dtrace_deferred_pid;
14579 help->dthps_prev = NULL;
14580 if (dtrace_deferred_pid != NULL)
14581 dtrace_deferred_pid->dthps_prev = help;
14582 dtrace_deferred_pid = help;
14585 mutex_exit(&dtrace_lock);
14587 } else if (dofhp != NULL) {
14589 * If the dtrace module is loaded and we have a particular
14590 * helper provider description, pass that off to the
14594 mutex_exit(&dtrace_lock);
14596 dtrace_helper_provide(dofhp, p->p_pid);
14600 * Otherwise, just pass all the helper provider descriptions
14601 * off to the meta provider.
14605 mutex_exit(&dtrace_lock);
14607 for (i = 0; i < help->dthps_nprovs; i++) {
14608 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
14613 mutex_exit(&dtrace_meta_lock);
14617 dtrace_helper_provider_add(dof_helper_t *dofhp, int gen)
14619 dtrace_helpers_t *help;
14620 dtrace_helper_provider_t *hprov, **tmp_provs;
14621 uint_t tmp_maxprovs, i;
14623 ASSERT(MUTEX_HELD(&dtrace_lock));
14625 help = curproc->p_dtrace_helpers;
14626 ASSERT(help != NULL);
14629 * If we already have dtrace_helper_providers_max helper providers,
14630 * we're refuse to add a new one.
14632 if (help->dthps_nprovs >= dtrace_helper_providers_max)
14636 * Check to make sure this isn't a duplicate.
14638 for (i = 0; i < help->dthps_nprovs; i++) {
14639 if (dofhp->dofhp_dof ==
14640 help->dthps_provs[i]->dthp_prov.dofhp_dof)
14644 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
14645 hprov->dthp_prov = *dofhp;
14646 hprov->dthp_ref = 1;
14647 hprov->dthp_generation = gen;
14650 * Allocate a bigger table for helper providers if it's already full.
14652 if (help->dthps_maxprovs == help->dthps_nprovs) {
14653 tmp_maxprovs = help->dthps_maxprovs;
14654 tmp_provs = help->dthps_provs;
14656 if (help->dthps_maxprovs == 0)
14657 help->dthps_maxprovs = 2;
14659 help->dthps_maxprovs *= 2;
14660 if (help->dthps_maxprovs > dtrace_helper_providers_max)
14661 help->dthps_maxprovs = dtrace_helper_providers_max;
14663 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
14665 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
14666 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
14668 if (tmp_provs != NULL) {
14669 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
14670 sizeof (dtrace_helper_provider_t *));
14671 kmem_free(tmp_provs, tmp_maxprovs *
14672 sizeof (dtrace_helper_provider_t *));
14676 help->dthps_provs[help->dthps_nprovs] = hprov;
14677 help->dthps_nprovs++;
14683 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
14685 mutex_enter(&dtrace_lock);
14687 if (--hprov->dthp_ref == 0) {
14689 mutex_exit(&dtrace_lock);
14690 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
14691 dtrace_dof_destroy(dof);
14692 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
14694 mutex_exit(&dtrace_lock);
14699 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
14701 uintptr_t daddr = (uintptr_t)dof;
14702 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
14703 dof_provider_t *provider;
14704 dof_probe_t *probe;
14706 char *strtab, *typestr;
14707 dof_stridx_t typeidx;
14709 uint_t nprobes, j, k;
14711 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
14713 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
14714 dtrace_dof_error(dof, "misaligned section offset");
14719 * The section needs to be large enough to contain the DOF provider
14720 * structure appropriate for the given version.
14722 if (sec->dofs_size <
14723 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
14724 offsetof(dof_provider_t, dofpv_prenoffs) :
14725 sizeof (dof_provider_t))) {
14726 dtrace_dof_error(dof, "provider section too small");
14730 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
14731 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
14732 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
14733 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
14734 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
14736 if (str_sec == NULL || prb_sec == NULL ||
14737 arg_sec == NULL || off_sec == NULL)
14742 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14743 provider->dofpv_prenoffs != DOF_SECT_NONE &&
14744 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
14745 provider->dofpv_prenoffs)) == NULL)
14748 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
14750 if (provider->dofpv_name >= str_sec->dofs_size ||
14751 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
14752 dtrace_dof_error(dof, "invalid provider name");
14756 if (prb_sec->dofs_entsize == 0 ||
14757 prb_sec->dofs_entsize > prb_sec->dofs_size) {
14758 dtrace_dof_error(dof, "invalid entry size");
14762 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
14763 dtrace_dof_error(dof, "misaligned entry size");
14767 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
14768 dtrace_dof_error(dof, "invalid entry size");
14772 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
14773 dtrace_dof_error(dof, "misaligned section offset");
14777 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
14778 dtrace_dof_error(dof, "invalid entry size");
14782 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
14784 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
14787 * Take a pass through the probes to check for errors.
14789 for (j = 0; j < nprobes; j++) {
14790 probe = (dof_probe_t *)(uintptr_t)(daddr +
14791 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
14793 if (probe->dofpr_func >= str_sec->dofs_size) {
14794 dtrace_dof_error(dof, "invalid function name");
14798 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
14799 dtrace_dof_error(dof, "function name too long");
14803 if (probe->dofpr_name >= str_sec->dofs_size ||
14804 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
14805 dtrace_dof_error(dof, "invalid probe name");
14810 * The offset count must not wrap the index, and the offsets
14811 * must also not overflow the section's data.
14813 if (probe->dofpr_offidx + probe->dofpr_noffs <
14814 probe->dofpr_offidx ||
14815 (probe->dofpr_offidx + probe->dofpr_noffs) *
14816 off_sec->dofs_entsize > off_sec->dofs_size) {
14817 dtrace_dof_error(dof, "invalid probe offset");
14821 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
14823 * If there's no is-enabled offset section, make sure
14824 * there aren't any is-enabled offsets. Otherwise
14825 * perform the same checks as for probe offsets
14826 * (immediately above).
14828 if (enoff_sec == NULL) {
14829 if (probe->dofpr_enoffidx != 0 ||
14830 probe->dofpr_nenoffs != 0) {
14831 dtrace_dof_error(dof, "is-enabled "
14832 "offsets with null section");
14835 } else if (probe->dofpr_enoffidx +
14836 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
14837 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
14838 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
14839 dtrace_dof_error(dof, "invalid is-enabled "
14844 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
14845 dtrace_dof_error(dof, "zero probe and "
14846 "is-enabled offsets");
14849 } else if (probe->dofpr_noffs == 0) {
14850 dtrace_dof_error(dof, "zero probe offsets");
14854 if (probe->dofpr_argidx + probe->dofpr_xargc <
14855 probe->dofpr_argidx ||
14856 (probe->dofpr_argidx + probe->dofpr_xargc) *
14857 arg_sec->dofs_entsize > arg_sec->dofs_size) {
14858 dtrace_dof_error(dof, "invalid args");
14862 typeidx = probe->dofpr_nargv;
14863 typestr = strtab + probe->dofpr_nargv;
14864 for (k = 0; k < probe->dofpr_nargc; k++) {
14865 if (typeidx >= str_sec->dofs_size) {
14866 dtrace_dof_error(dof, "bad "
14867 "native argument type");
14871 typesz = strlen(typestr) + 1;
14872 if (typesz > DTRACE_ARGTYPELEN) {
14873 dtrace_dof_error(dof, "native "
14874 "argument type too long");
14881 typeidx = probe->dofpr_xargv;
14882 typestr = strtab + probe->dofpr_xargv;
14883 for (k = 0; k < probe->dofpr_xargc; k++) {
14884 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
14885 dtrace_dof_error(dof, "bad "
14886 "native argument index");
14890 if (typeidx >= str_sec->dofs_size) {
14891 dtrace_dof_error(dof, "bad "
14892 "translated argument type");
14896 typesz = strlen(typestr) + 1;
14897 if (typesz > DTRACE_ARGTYPELEN) {
14898 dtrace_dof_error(dof, "translated argument "
14912 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp)
14914 dtrace_helpers_t *help;
14915 dtrace_vstate_t *vstate;
14916 dtrace_enabling_t *enab = NULL;
14917 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
14918 uintptr_t daddr = (uintptr_t)dof;
14920 ASSERT(MUTEX_HELD(&dtrace_lock));
14922 if ((help = curproc->p_dtrace_helpers) == NULL)
14923 help = dtrace_helpers_create(curproc);
14925 vstate = &help->dthps_vstate;
14927 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab,
14928 dhp != NULL ? dhp->dofhp_addr : 0, B_FALSE)) != 0) {
14929 dtrace_dof_destroy(dof);
14934 * Look for helper providers and validate their descriptions.
14937 for (i = 0; i < dof->dofh_secnum; i++) {
14938 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
14939 dof->dofh_secoff + i * dof->dofh_secsize);
14941 if (sec->dofs_type != DOF_SECT_PROVIDER)
14944 if (dtrace_helper_provider_validate(dof, sec) != 0) {
14945 dtrace_enabling_destroy(enab);
14946 dtrace_dof_destroy(dof);
14955 * Now we need to walk through the ECB descriptions in the enabling.
14957 for (i = 0; i < enab->dten_ndesc; i++) {
14958 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
14959 dtrace_probedesc_t *desc = &ep->dted_probe;
14961 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
14964 if (strcmp(desc->dtpd_mod, "helper") != 0)
14967 if (strcmp(desc->dtpd_func, "ustack") != 0)
14970 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
14973 * Adding this helper action failed -- we are now going
14974 * to rip out the entire generation and return failure.
14976 (void) dtrace_helper_destroygen(help->dthps_generation);
14977 dtrace_enabling_destroy(enab);
14978 dtrace_dof_destroy(dof);
14985 if (nhelpers < enab->dten_ndesc)
14986 dtrace_dof_error(dof, "unmatched helpers");
14988 gen = help->dthps_generation++;
14989 dtrace_enabling_destroy(enab);
14991 if (dhp != NULL && nprovs > 0) {
14992 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
14993 if (dtrace_helper_provider_add(dhp, gen) == 0) {
14994 mutex_exit(&dtrace_lock);
14995 dtrace_helper_provider_register(curproc, help, dhp);
14996 mutex_enter(&dtrace_lock);
15003 dtrace_dof_destroy(dof);
15008 static dtrace_helpers_t *
15009 dtrace_helpers_create(proc_t *p)
15011 dtrace_helpers_t *help;
15013 ASSERT(MUTEX_HELD(&dtrace_lock));
15014 ASSERT(p->p_dtrace_helpers == NULL);
15016 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
15017 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
15018 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
15020 p->p_dtrace_helpers = help;
15030 dtrace_helpers_destroy(proc_t *p)
15032 dtrace_helpers_t *help;
15033 dtrace_vstate_t *vstate;
15035 proc_t *p = curproc;
15039 mutex_enter(&dtrace_lock);
15041 ASSERT(p->p_dtrace_helpers != NULL);
15042 ASSERT(dtrace_helpers > 0);
15044 help = p->p_dtrace_helpers;
15045 vstate = &help->dthps_vstate;
15048 * We're now going to lose the help from this process.
15050 p->p_dtrace_helpers = NULL;
15054 * Destory the helper actions.
15056 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15057 dtrace_helper_action_t *h, *next;
15059 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15060 next = h->dtha_next;
15061 dtrace_helper_action_destroy(h, vstate);
15066 mutex_exit(&dtrace_lock);
15069 * Destroy the helper providers.
15071 if (help->dthps_maxprovs > 0) {
15072 mutex_enter(&dtrace_meta_lock);
15073 if (dtrace_meta_pid != NULL) {
15074 ASSERT(dtrace_deferred_pid == NULL);
15076 for (i = 0; i < help->dthps_nprovs; i++) {
15077 dtrace_helper_provider_remove(
15078 &help->dthps_provs[i]->dthp_prov, p->p_pid);
15081 mutex_enter(&dtrace_lock);
15082 ASSERT(help->dthps_deferred == 0 ||
15083 help->dthps_next != NULL ||
15084 help->dthps_prev != NULL ||
15085 help == dtrace_deferred_pid);
15088 * Remove the helper from the deferred list.
15090 if (help->dthps_next != NULL)
15091 help->dthps_next->dthps_prev = help->dthps_prev;
15092 if (help->dthps_prev != NULL)
15093 help->dthps_prev->dthps_next = help->dthps_next;
15094 if (dtrace_deferred_pid == help) {
15095 dtrace_deferred_pid = help->dthps_next;
15096 ASSERT(help->dthps_prev == NULL);
15099 mutex_exit(&dtrace_lock);
15102 mutex_exit(&dtrace_meta_lock);
15104 for (i = 0; i < help->dthps_nprovs; i++) {
15105 dtrace_helper_provider_destroy(help->dthps_provs[i]);
15108 kmem_free(help->dthps_provs, help->dthps_maxprovs *
15109 sizeof (dtrace_helper_provider_t *));
15112 mutex_enter(&dtrace_lock);
15114 dtrace_vstate_fini(&help->dthps_vstate);
15115 kmem_free(help->dthps_actions,
15116 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
15117 kmem_free(help, sizeof (dtrace_helpers_t));
15120 mutex_exit(&dtrace_lock);
15127 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
15129 dtrace_helpers_t *help, *newhelp;
15130 dtrace_helper_action_t *helper, *new, *last;
15132 dtrace_vstate_t *vstate;
15133 int i, j, sz, hasprovs = 0;
15135 mutex_enter(&dtrace_lock);
15136 ASSERT(from->p_dtrace_helpers != NULL);
15137 ASSERT(dtrace_helpers > 0);
15139 help = from->p_dtrace_helpers;
15140 newhelp = dtrace_helpers_create(to);
15141 ASSERT(to->p_dtrace_helpers != NULL);
15143 newhelp->dthps_generation = help->dthps_generation;
15144 vstate = &newhelp->dthps_vstate;
15147 * Duplicate the helper actions.
15149 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15150 if ((helper = help->dthps_actions[i]) == NULL)
15153 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
15154 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
15156 new->dtha_generation = helper->dtha_generation;
15158 if ((dp = helper->dtha_predicate) != NULL) {
15159 dp = dtrace_difo_duplicate(dp, vstate);
15160 new->dtha_predicate = dp;
15163 new->dtha_nactions = helper->dtha_nactions;
15164 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
15165 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
15167 for (j = 0; j < new->dtha_nactions; j++) {
15168 dtrace_difo_t *dp = helper->dtha_actions[j];
15170 ASSERT(dp != NULL);
15171 dp = dtrace_difo_duplicate(dp, vstate);
15172 new->dtha_actions[j] = dp;
15175 if (last != NULL) {
15176 last->dtha_next = new;
15178 newhelp->dthps_actions[i] = new;
15186 * Duplicate the helper providers and register them with the
15187 * DTrace framework.
15189 if (help->dthps_nprovs > 0) {
15190 newhelp->dthps_nprovs = help->dthps_nprovs;
15191 newhelp->dthps_maxprovs = help->dthps_nprovs;
15192 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
15193 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
15194 for (i = 0; i < newhelp->dthps_nprovs; i++) {
15195 newhelp->dthps_provs[i] = help->dthps_provs[i];
15196 newhelp->dthps_provs[i]->dthp_ref++;
15202 mutex_exit(&dtrace_lock);
15205 dtrace_helper_provider_register(to, newhelp, NULL);
15209 * DTrace Hook Functions
15212 dtrace_module_loaded(modctl_t *ctl)
15214 dtrace_provider_t *prv;
15216 mutex_enter(&dtrace_provider_lock);
15218 mutex_enter(&mod_lock);
15222 ASSERT(ctl->mod_busy);
15226 * We're going to call each providers per-module provide operation
15227 * specifying only this module.
15229 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
15230 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
15233 mutex_exit(&mod_lock);
15235 mutex_exit(&dtrace_provider_lock);
15238 * If we have any retained enablings, we need to match against them.
15239 * Enabling probes requires that cpu_lock be held, and we cannot hold
15240 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
15241 * module. (In particular, this happens when loading scheduling
15242 * classes.) So if we have any retained enablings, we need to dispatch
15243 * our task queue to do the match for us.
15245 mutex_enter(&dtrace_lock);
15247 if (dtrace_retained == NULL) {
15248 mutex_exit(&dtrace_lock);
15252 (void) taskq_dispatch(dtrace_taskq,
15253 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
15255 mutex_exit(&dtrace_lock);
15258 * And now, for a little heuristic sleaze: in general, we want to
15259 * match modules as soon as they load. However, we cannot guarantee
15260 * this, because it would lead us to the lock ordering violation
15261 * outlined above. The common case, of course, is that cpu_lock is
15262 * _not_ held -- so we delay here for a clock tick, hoping that that's
15263 * long enough for the task queue to do its work. If it's not, it's
15264 * not a serious problem -- it just means that the module that we
15265 * just loaded may not be immediately instrumentable.
15272 dtrace_module_unloaded(modctl_t *ctl)
15274 dtrace_module_unloaded(modctl_t *ctl, int *error)
15277 dtrace_probe_t template, *probe, *first, *next;
15278 dtrace_provider_t *prov;
15280 char modname[DTRACE_MODNAMELEN];
15285 template.dtpr_mod = ctl->mod_modname;
15287 /* Handle the fact that ctl->filename may end in ".ko". */
15288 strlcpy(modname, ctl->filename, sizeof(modname));
15289 len = strlen(ctl->filename);
15290 if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
15291 modname[len - 3] = '\0';
15292 template.dtpr_mod = modname;
15295 mutex_enter(&dtrace_provider_lock);
15297 mutex_enter(&mod_lock);
15299 mutex_enter(&dtrace_lock);
15302 if (ctl->nenabled > 0) {
15303 /* Don't allow unloads if a probe is enabled. */
15304 mutex_exit(&dtrace_provider_lock);
15305 mutex_exit(&dtrace_lock);
15308 "kldunload: attempt to unload module that has DTrace probes enabled\n");
15313 if (dtrace_bymod == NULL) {
15315 * The DTrace module is loaded (obviously) but not attached;
15316 * we don't have any work to do.
15318 mutex_exit(&dtrace_provider_lock);
15320 mutex_exit(&mod_lock);
15322 mutex_exit(&dtrace_lock);
15326 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
15327 probe != NULL; probe = probe->dtpr_nextmod) {
15328 if (probe->dtpr_ecb != NULL) {
15329 mutex_exit(&dtrace_provider_lock);
15331 mutex_exit(&mod_lock);
15333 mutex_exit(&dtrace_lock);
15336 * This shouldn't _actually_ be possible -- we're
15337 * unloading a module that has an enabled probe in it.
15338 * (It's normally up to the provider to make sure that
15339 * this can't happen.) However, because dtps_enable()
15340 * doesn't have a failure mode, there can be an
15341 * enable/unload race. Upshot: we don't want to
15342 * assert, but we're not going to disable the
15345 if (dtrace_err_verbose) {
15347 cmn_err(CE_WARN, "unloaded module '%s' had "
15348 "enabled probes", ctl->mod_modname);
15350 cmn_err(CE_WARN, "unloaded module '%s' had "
15351 "enabled probes", modname);
15361 for (first = NULL; probe != NULL; probe = next) {
15362 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
15364 dtrace_probes[probe->dtpr_id - 1] = NULL;
15366 next = probe->dtpr_nextmod;
15367 dtrace_hash_remove(dtrace_bymod, probe);
15368 dtrace_hash_remove(dtrace_byfunc, probe);
15369 dtrace_hash_remove(dtrace_byname, probe);
15371 if (first == NULL) {
15373 probe->dtpr_nextmod = NULL;
15375 probe->dtpr_nextmod = first;
15381 * We've removed all of the module's probes from the hash chains and
15382 * from the probe array. Now issue a dtrace_sync() to be sure that
15383 * everyone has cleared out from any probe array processing.
15387 for (probe = first; probe != NULL; probe = first) {
15388 first = probe->dtpr_nextmod;
15389 prov = probe->dtpr_provider;
15390 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
15392 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
15393 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
15394 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
15396 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
15398 free_unr(dtrace_arena, probe->dtpr_id);
15400 kmem_free(probe, sizeof (dtrace_probe_t));
15403 mutex_exit(&dtrace_lock);
15405 mutex_exit(&mod_lock);
15407 mutex_exit(&dtrace_provider_lock);
15412 dtrace_kld_load(void *arg __unused, linker_file_t lf)
15415 dtrace_module_loaded(lf);
15419 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
15423 /* We already have an error, so don't do anything. */
15425 dtrace_module_unloaded(lf, error);
15431 dtrace_suspend(void)
15433 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
15437 dtrace_resume(void)
15439 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
15444 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
15446 ASSERT(MUTEX_HELD(&cpu_lock));
15447 mutex_enter(&dtrace_lock);
15451 dtrace_state_t *state;
15452 dtrace_optval_t *opt, rs, c;
15455 * For now, we only allocate a new buffer for anonymous state.
15457 if ((state = dtrace_anon.dta_state) == NULL)
15460 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
15463 opt = state->dts_options;
15464 c = opt[DTRACEOPT_CPU];
15466 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
15470 * Regardless of what the actual policy is, we're going to
15471 * temporarily set our resize policy to be manual. We're
15472 * also going to temporarily set our CPU option to denote
15473 * the newly configured CPU.
15475 rs = opt[DTRACEOPT_BUFRESIZE];
15476 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
15477 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
15479 (void) dtrace_state_buffers(state);
15481 opt[DTRACEOPT_BUFRESIZE] = rs;
15482 opt[DTRACEOPT_CPU] = c;
15489 * We don't free the buffer in the CPU_UNCONFIG case. (The
15490 * buffer will be freed when the consumer exits.)
15498 mutex_exit(&dtrace_lock);
15504 dtrace_cpu_setup_initial(processorid_t cpu)
15506 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
15511 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
15513 if (dtrace_toxranges >= dtrace_toxranges_max) {
15515 dtrace_toxrange_t *range;
15517 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15520 ASSERT(dtrace_toxrange == NULL);
15521 ASSERT(dtrace_toxranges_max == 0);
15522 dtrace_toxranges_max = 1;
15524 dtrace_toxranges_max <<= 1;
15527 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
15528 range = kmem_zalloc(nsize, KM_SLEEP);
15530 if (dtrace_toxrange != NULL) {
15531 ASSERT(osize != 0);
15532 bcopy(dtrace_toxrange, range, osize);
15533 kmem_free(dtrace_toxrange, osize);
15536 dtrace_toxrange = range;
15539 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
15540 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
15542 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
15543 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
15544 dtrace_toxranges++;
15548 * DTrace Driver Cookbook Functions
15553 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
15555 dtrace_provider_id_t id;
15556 dtrace_state_t *state = NULL;
15557 dtrace_enabling_t *enab;
15559 mutex_enter(&cpu_lock);
15560 mutex_enter(&dtrace_provider_lock);
15561 mutex_enter(&dtrace_lock);
15563 if (ddi_soft_state_init(&dtrace_softstate,
15564 sizeof (dtrace_state_t), 0) != 0) {
15565 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
15566 mutex_exit(&cpu_lock);
15567 mutex_exit(&dtrace_provider_lock);
15568 mutex_exit(&dtrace_lock);
15569 return (DDI_FAILURE);
15572 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
15573 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
15574 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
15575 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
15576 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
15577 ddi_remove_minor_node(devi, NULL);
15578 ddi_soft_state_fini(&dtrace_softstate);
15579 mutex_exit(&cpu_lock);
15580 mutex_exit(&dtrace_provider_lock);
15581 mutex_exit(&dtrace_lock);
15582 return (DDI_FAILURE);
15585 ddi_report_dev(devi);
15586 dtrace_devi = devi;
15588 dtrace_modload = dtrace_module_loaded;
15589 dtrace_modunload = dtrace_module_unloaded;
15590 dtrace_cpu_init = dtrace_cpu_setup_initial;
15591 dtrace_helpers_cleanup = dtrace_helpers_destroy;
15592 dtrace_helpers_fork = dtrace_helpers_duplicate;
15593 dtrace_cpustart_init = dtrace_suspend;
15594 dtrace_cpustart_fini = dtrace_resume;
15595 dtrace_debugger_init = dtrace_suspend;
15596 dtrace_debugger_fini = dtrace_resume;
15598 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
15600 ASSERT(MUTEX_HELD(&cpu_lock));
15602 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
15603 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
15604 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
15605 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
15606 VM_SLEEP | VMC_IDENTIFIER);
15607 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
15610 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
15611 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
15612 NULL, NULL, NULL, NULL, NULL, 0);
15614 ASSERT(MUTEX_HELD(&cpu_lock));
15615 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
15616 offsetof(dtrace_probe_t, dtpr_nextmod),
15617 offsetof(dtrace_probe_t, dtpr_prevmod));
15619 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
15620 offsetof(dtrace_probe_t, dtpr_nextfunc),
15621 offsetof(dtrace_probe_t, dtpr_prevfunc));
15623 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
15624 offsetof(dtrace_probe_t, dtpr_nextname),
15625 offsetof(dtrace_probe_t, dtpr_prevname));
15627 if (dtrace_retain_max < 1) {
15628 cmn_err(CE_WARN, "illegal value (%lu) for dtrace_retain_max; "
15629 "setting to 1", dtrace_retain_max);
15630 dtrace_retain_max = 1;
15634 * Now discover our toxic ranges.
15636 dtrace_toxic_ranges(dtrace_toxrange_add);
15639 * Before we register ourselves as a provider to our own framework,
15640 * we would like to assert that dtrace_provider is NULL -- but that's
15641 * not true if we were loaded as a dependency of a DTrace provider.
15642 * Once we've registered, we can assert that dtrace_provider is our
15645 (void) dtrace_register("dtrace", &dtrace_provider_attr,
15646 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
15648 ASSERT(dtrace_provider != NULL);
15649 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
15651 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
15652 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
15653 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
15654 dtrace_provider, NULL, NULL, "END", 0, NULL);
15655 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
15656 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
15658 dtrace_anon_property();
15659 mutex_exit(&cpu_lock);
15662 * If DTrace helper tracing is enabled, we need to allocate the
15663 * trace buffer and initialize the values.
15665 if (dtrace_helptrace_enabled) {
15666 ASSERT(dtrace_helptrace_buffer == NULL);
15667 dtrace_helptrace_buffer =
15668 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
15669 dtrace_helptrace_next = 0;
15673 * If there are already providers, we must ask them to provide their
15674 * probes, and then match any anonymous enabling against them. Note
15675 * that there should be no other retained enablings at this time:
15676 * the only retained enablings at this time should be the anonymous
15679 if (dtrace_anon.dta_enabling != NULL) {
15680 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
15682 dtrace_enabling_provide(NULL);
15683 state = dtrace_anon.dta_state;
15686 * We couldn't hold cpu_lock across the above call to
15687 * dtrace_enabling_provide(), but we must hold it to actually
15688 * enable the probes. We have to drop all of our locks, pick
15689 * up cpu_lock, and regain our locks before matching the
15690 * retained anonymous enabling.
15692 mutex_exit(&dtrace_lock);
15693 mutex_exit(&dtrace_provider_lock);
15695 mutex_enter(&cpu_lock);
15696 mutex_enter(&dtrace_provider_lock);
15697 mutex_enter(&dtrace_lock);
15699 if ((enab = dtrace_anon.dta_enabling) != NULL)
15700 (void) dtrace_enabling_match(enab, NULL);
15702 mutex_exit(&cpu_lock);
15705 mutex_exit(&dtrace_lock);
15706 mutex_exit(&dtrace_provider_lock);
15708 if (state != NULL) {
15710 * If we created any anonymous state, set it going now.
15712 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
15715 return (DDI_SUCCESS);
15720 #if __FreeBSD_version >= 800039
15722 dtrace_dtr(void *data __unused)
15731 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
15733 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
15736 dtrace_state_t *state;
15742 if (getminor(*devp) == DTRACEMNRN_HELPER)
15746 * If this wasn't an open with the "helper" minor, then it must be
15747 * the "dtrace" minor.
15749 if (getminor(*devp) == DTRACEMNRN_DTRACE)
15752 cred_t *cred_p = NULL;
15754 #if __FreeBSD_version < 800039
15756 * The first minor device is the one that is cloned so there is
15757 * nothing more to do here.
15759 if (dev2unit(dev) == 0)
15763 * Devices are cloned, so if the DTrace state has already
15764 * been allocated, that means this device belongs to a
15765 * different client. Each client should open '/dev/dtrace'
15766 * to get a cloned device.
15768 if (dev->si_drv1 != NULL)
15772 cred_p = dev->si_cred;
15776 * If no DTRACE_PRIV_* bits are set in the credential, then the
15777 * caller lacks sufficient permission to do anything with DTrace.
15779 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
15780 if (priv == DTRACE_PRIV_NONE) {
15782 #if __FreeBSD_version < 800039
15783 /* Destroy the cloned device. */
15792 * Ask all providers to provide all their probes.
15794 mutex_enter(&dtrace_provider_lock);
15795 dtrace_probe_provide(NULL, NULL);
15796 mutex_exit(&dtrace_provider_lock);
15798 mutex_enter(&cpu_lock);
15799 mutex_enter(&dtrace_lock);
15801 dtrace_membar_producer();
15805 * If the kernel debugger is active (that is, if the kernel debugger
15806 * modified text in some way), we won't allow the open.
15808 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15810 mutex_exit(&cpu_lock);
15811 mutex_exit(&dtrace_lock);
15815 state = dtrace_state_create(devp, cred_p);
15817 state = dtrace_state_create(dev);
15818 #if __FreeBSD_version < 800039
15819 dev->si_drv1 = state;
15821 devfs_set_cdevpriv(state, dtrace_dtr);
15825 mutex_exit(&cpu_lock);
15827 if (state == NULL) {
15829 if (--dtrace_opens == 0)
15830 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15834 mutex_exit(&dtrace_lock);
15836 #if __FreeBSD_version < 800039
15837 /* Destroy the cloned device. */
15844 mutex_exit(&dtrace_lock);
15852 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
15854 dtrace_close(struct cdev *dev, int flags, int fmt __unused, struct thread *td)
15858 minor_t minor = getminor(dev);
15859 dtrace_state_t *state;
15861 if (minor == DTRACEMNRN_HELPER)
15864 state = ddi_get_soft_state(dtrace_softstate, minor);
15866 #if __FreeBSD_version < 800039
15867 dtrace_state_t *state = dev->si_drv1;
15869 /* Check if this is not a cloned device. */
15870 if (dev2unit(dev) == 0)
15873 dtrace_state_t *state;
15874 devfs_get_cdevpriv((void **) &state);
15879 mutex_enter(&cpu_lock);
15880 mutex_enter(&dtrace_lock);
15882 if (state != NULL) {
15883 if (state->dts_anon) {
15885 * There is anonymous state. Destroy that first.
15887 ASSERT(dtrace_anon.dta_state == NULL);
15888 dtrace_state_destroy(state->dts_anon);
15891 dtrace_state_destroy(state);
15894 kmem_free(state, 0);
15895 #if __FreeBSD_version < 800039
15896 dev->si_drv1 = NULL;
15901 ASSERT(dtrace_opens > 0);
15903 if (--dtrace_opens == 0)
15904 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
15909 mutex_exit(&dtrace_lock);
15910 mutex_exit(&cpu_lock);
15912 #if __FreeBSD_version < 800039
15913 /* Schedule this cloned device to be destroyed. */
15914 destroy_dev_sched(dev);
15923 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
15926 dof_helper_t help, *dhp = NULL;
15929 case DTRACEHIOC_ADDDOF:
15930 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
15931 dtrace_dof_error(NULL, "failed to copyin DOF helper");
15936 arg = (intptr_t)help.dofhp_dof;
15939 case DTRACEHIOC_ADD: {
15940 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
15945 mutex_enter(&dtrace_lock);
15948 * dtrace_helper_slurp() takes responsibility for the dof --
15949 * it may free it now or it may save it and free it later.
15951 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
15958 mutex_exit(&dtrace_lock);
15962 case DTRACEHIOC_REMOVE: {
15963 mutex_enter(&dtrace_lock);
15964 rval = dtrace_helper_destroygen(arg);
15965 mutex_exit(&dtrace_lock);
15979 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
15981 minor_t minor = getminor(dev);
15982 dtrace_state_t *state;
15985 if (minor == DTRACEMNRN_HELPER)
15986 return (dtrace_ioctl_helper(cmd, arg, rv));
15988 state = ddi_get_soft_state(dtrace_softstate, minor);
15990 if (state->dts_anon) {
15991 ASSERT(dtrace_anon.dta_state == NULL);
15992 state = state->dts_anon;
15996 case DTRACEIOC_PROVIDER: {
15997 dtrace_providerdesc_t pvd;
15998 dtrace_provider_t *pvp;
16000 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
16003 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
16004 mutex_enter(&dtrace_provider_lock);
16006 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
16007 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
16011 mutex_exit(&dtrace_provider_lock);
16016 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
16017 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
16019 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
16025 case DTRACEIOC_EPROBE: {
16026 dtrace_eprobedesc_t epdesc;
16028 dtrace_action_t *act;
16034 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
16037 mutex_enter(&dtrace_lock);
16039 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
16040 mutex_exit(&dtrace_lock);
16044 if (ecb->dte_probe == NULL) {
16045 mutex_exit(&dtrace_lock);
16049 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
16050 epdesc.dtepd_uarg = ecb->dte_uarg;
16051 epdesc.dtepd_size = ecb->dte_size;
16053 nrecs = epdesc.dtepd_nrecs;
16054 epdesc.dtepd_nrecs = 0;
16055 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16056 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16059 epdesc.dtepd_nrecs++;
16063 * Now that we have the size, we need to allocate a temporary
16064 * buffer in which to store the complete description. We need
16065 * the temporary buffer to be able to drop dtrace_lock()
16066 * across the copyout(), below.
16068 size = sizeof (dtrace_eprobedesc_t) +
16069 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
16071 buf = kmem_alloc(size, KM_SLEEP);
16072 dest = (uintptr_t)buf;
16074 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
16075 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
16077 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
16078 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
16084 bcopy(&act->dta_rec, (void *)dest,
16085 sizeof (dtrace_recdesc_t));
16086 dest += sizeof (dtrace_recdesc_t);
16089 mutex_exit(&dtrace_lock);
16091 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16092 kmem_free(buf, size);
16096 kmem_free(buf, size);
16100 case DTRACEIOC_AGGDESC: {
16101 dtrace_aggdesc_t aggdesc;
16102 dtrace_action_t *act;
16103 dtrace_aggregation_t *agg;
16106 dtrace_recdesc_t *lrec;
16111 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
16114 mutex_enter(&dtrace_lock);
16116 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
16117 mutex_exit(&dtrace_lock);
16121 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
16123 nrecs = aggdesc.dtagd_nrecs;
16124 aggdesc.dtagd_nrecs = 0;
16126 offs = agg->dtag_base;
16127 lrec = &agg->dtag_action.dta_rec;
16128 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
16130 for (act = agg->dtag_first; ; act = act->dta_next) {
16131 ASSERT(act->dta_intuple ||
16132 DTRACEACT_ISAGG(act->dta_kind));
16135 * If this action has a record size of zero, it
16136 * denotes an argument to the aggregating action.
16137 * Because the presence of this record doesn't (or
16138 * shouldn't) affect the way the data is interpreted,
16139 * we don't copy it out to save user-level the
16140 * confusion of dealing with a zero-length record.
16142 if (act->dta_rec.dtrd_size == 0) {
16143 ASSERT(agg->dtag_hasarg);
16147 aggdesc.dtagd_nrecs++;
16149 if (act == &agg->dtag_action)
16154 * Now that we have the size, we need to allocate a temporary
16155 * buffer in which to store the complete description. We need
16156 * the temporary buffer to be able to drop dtrace_lock()
16157 * across the copyout(), below.
16159 size = sizeof (dtrace_aggdesc_t) +
16160 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
16162 buf = kmem_alloc(size, KM_SLEEP);
16163 dest = (uintptr_t)buf;
16165 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
16166 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
16168 for (act = agg->dtag_first; ; act = act->dta_next) {
16169 dtrace_recdesc_t rec = act->dta_rec;
16172 * See the comment in the above loop for why we pass
16173 * over zero-length records.
16175 if (rec.dtrd_size == 0) {
16176 ASSERT(agg->dtag_hasarg);
16183 rec.dtrd_offset -= offs;
16184 bcopy(&rec, (void *)dest, sizeof (rec));
16185 dest += sizeof (dtrace_recdesc_t);
16187 if (act == &agg->dtag_action)
16191 mutex_exit(&dtrace_lock);
16193 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
16194 kmem_free(buf, size);
16198 kmem_free(buf, size);
16202 case DTRACEIOC_ENABLE: {
16204 dtrace_enabling_t *enab = NULL;
16205 dtrace_vstate_t *vstate;
16211 * If a NULL argument has been passed, we take this as our
16212 * cue to reevaluate our enablings.
16215 dtrace_enabling_matchall();
16220 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
16223 mutex_enter(&cpu_lock);
16224 mutex_enter(&dtrace_lock);
16225 vstate = &state->dts_vstate;
16227 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
16228 mutex_exit(&dtrace_lock);
16229 mutex_exit(&cpu_lock);
16230 dtrace_dof_destroy(dof);
16234 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
16235 mutex_exit(&dtrace_lock);
16236 mutex_exit(&cpu_lock);
16237 dtrace_dof_destroy(dof);
16241 if ((rval = dtrace_dof_options(dof, state)) != 0) {
16242 dtrace_enabling_destroy(enab);
16243 mutex_exit(&dtrace_lock);
16244 mutex_exit(&cpu_lock);
16245 dtrace_dof_destroy(dof);
16249 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
16250 err = dtrace_enabling_retain(enab);
16252 dtrace_enabling_destroy(enab);
16255 mutex_exit(&cpu_lock);
16256 mutex_exit(&dtrace_lock);
16257 dtrace_dof_destroy(dof);
16262 case DTRACEIOC_REPLICATE: {
16263 dtrace_repldesc_t desc;
16264 dtrace_probedesc_t *match = &desc.dtrpd_match;
16265 dtrace_probedesc_t *create = &desc.dtrpd_create;
16268 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16271 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16272 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16273 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16274 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16276 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16277 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16278 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16279 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16281 mutex_enter(&dtrace_lock);
16282 err = dtrace_enabling_replicate(state, match, create);
16283 mutex_exit(&dtrace_lock);
16288 case DTRACEIOC_PROBEMATCH:
16289 case DTRACEIOC_PROBES: {
16290 dtrace_probe_t *probe = NULL;
16291 dtrace_probedesc_t desc;
16292 dtrace_probekey_t pkey;
16299 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16302 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
16303 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
16304 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
16305 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
16308 * Before we attempt to match this probe, we want to give
16309 * all providers the opportunity to provide it.
16311 if (desc.dtpd_id == DTRACE_IDNONE) {
16312 mutex_enter(&dtrace_provider_lock);
16313 dtrace_probe_provide(&desc, NULL);
16314 mutex_exit(&dtrace_provider_lock);
16318 if (cmd == DTRACEIOC_PROBEMATCH) {
16319 dtrace_probekey(&desc, &pkey);
16320 pkey.dtpk_id = DTRACE_IDNONE;
16323 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
16325 mutex_enter(&dtrace_lock);
16327 if (cmd == DTRACEIOC_PROBEMATCH) {
16328 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16329 if ((probe = dtrace_probes[i - 1]) != NULL &&
16330 (m = dtrace_match_probe(probe, &pkey,
16331 priv, uid, zoneid)) != 0)
16336 mutex_exit(&dtrace_lock);
16341 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
16342 if ((probe = dtrace_probes[i - 1]) != NULL &&
16343 dtrace_match_priv(probe, priv, uid, zoneid))
16348 if (probe == NULL) {
16349 mutex_exit(&dtrace_lock);
16353 dtrace_probe_description(probe, &desc);
16354 mutex_exit(&dtrace_lock);
16356 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16362 case DTRACEIOC_PROBEARG: {
16363 dtrace_argdesc_t desc;
16364 dtrace_probe_t *probe;
16365 dtrace_provider_t *prov;
16367 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16370 if (desc.dtargd_id == DTRACE_IDNONE)
16373 if (desc.dtargd_ndx == DTRACE_ARGNONE)
16376 mutex_enter(&dtrace_provider_lock);
16377 mutex_enter(&mod_lock);
16378 mutex_enter(&dtrace_lock);
16380 if (desc.dtargd_id > dtrace_nprobes) {
16381 mutex_exit(&dtrace_lock);
16382 mutex_exit(&mod_lock);
16383 mutex_exit(&dtrace_provider_lock);
16387 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
16388 mutex_exit(&dtrace_lock);
16389 mutex_exit(&mod_lock);
16390 mutex_exit(&dtrace_provider_lock);
16394 mutex_exit(&dtrace_lock);
16396 prov = probe->dtpr_provider;
16398 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
16400 * There isn't any typed information for this probe.
16401 * Set the argument number to DTRACE_ARGNONE.
16403 desc.dtargd_ndx = DTRACE_ARGNONE;
16405 desc.dtargd_native[0] = '\0';
16406 desc.dtargd_xlate[0] = '\0';
16407 desc.dtargd_mapping = desc.dtargd_ndx;
16409 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
16410 probe->dtpr_id, probe->dtpr_arg, &desc);
16413 mutex_exit(&mod_lock);
16414 mutex_exit(&dtrace_provider_lock);
16416 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16422 case DTRACEIOC_GO: {
16423 processorid_t cpuid;
16424 rval = dtrace_state_go(state, &cpuid);
16429 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16435 case DTRACEIOC_STOP: {
16436 processorid_t cpuid;
16438 mutex_enter(&dtrace_lock);
16439 rval = dtrace_state_stop(state, &cpuid);
16440 mutex_exit(&dtrace_lock);
16445 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
16451 case DTRACEIOC_DOFGET: {
16452 dof_hdr_t hdr, *dof;
16455 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
16458 mutex_enter(&dtrace_lock);
16459 dof = dtrace_dof_create(state);
16460 mutex_exit(&dtrace_lock);
16462 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
16463 rval = copyout(dof, (void *)arg, len);
16464 dtrace_dof_destroy(dof);
16466 return (rval == 0 ? 0 : EFAULT);
16469 case DTRACEIOC_AGGSNAP:
16470 case DTRACEIOC_BUFSNAP: {
16471 dtrace_bufdesc_t desc;
16473 dtrace_buffer_t *buf;
16475 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
16478 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
16481 mutex_enter(&dtrace_lock);
16483 if (cmd == DTRACEIOC_BUFSNAP) {
16484 buf = &state->dts_buffer[desc.dtbd_cpu];
16486 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
16489 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
16490 size_t sz = buf->dtb_offset;
16492 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
16493 mutex_exit(&dtrace_lock);
16498 * If this buffer has already been consumed, we're
16499 * going to indicate that there's nothing left here
16502 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
16503 mutex_exit(&dtrace_lock);
16505 desc.dtbd_size = 0;
16506 desc.dtbd_drops = 0;
16507 desc.dtbd_errors = 0;
16508 desc.dtbd_oldest = 0;
16509 sz = sizeof (desc);
16511 if (copyout(&desc, (void *)arg, sz) != 0)
16518 * If this is a ring buffer that has wrapped, we want
16519 * to copy the whole thing out.
16521 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
16522 dtrace_buffer_polish(buf);
16523 sz = buf->dtb_size;
16526 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
16527 mutex_exit(&dtrace_lock);
16531 desc.dtbd_size = sz;
16532 desc.dtbd_drops = buf->dtb_drops;
16533 desc.dtbd_errors = buf->dtb_errors;
16534 desc.dtbd_oldest = buf->dtb_xamot_offset;
16535 desc.dtbd_timestamp = dtrace_gethrtime();
16537 mutex_exit(&dtrace_lock);
16539 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16542 buf->dtb_flags |= DTRACEBUF_CONSUMED;
16547 if (buf->dtb_tomax == NULL) {
16548 ASSERT(buf->dtb_xamot == NULL);
16549 mutex_exit(&dtrace_lock);
16553 cached = buf->dtb_tomax;
16554 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
16556 dtrace_xcall(desc.dtbd_cpu,
16557 (dtrace_xcall_t)dtrace_buffer_switch, buf);
16559 state->dts_errors += buf->dtb_xamot_errors;
16562 * If the buffers did not actually switch, then the cross call
16563 * did not take place -- presumably because the given CPU is
16564 * not in the ready set. If this is the case, we'll return
16567 if (buf->dtb_tomax == cached) {
16568 ASSERT(buf->dtb_xamot != cached);
16569 mutex_exit(&dtrace_lock);
16573 ASSERT(cached == buf->dtb_xamot);
16576 * We have our snapshot; now copy it out.
16578 if (copyout(buf->dtb_xamot, desc.dtbd_data,
16579 buf->dtb_xamot_offset) != 0) {
16580 mutex_exit(&dtrace_lock);
16584 desc.dtbd_size = buf->dtb_xamot_offset;
16585 desc.dtbd_drops = buf->dtb_xamot_drops;
16586 desc.dtbd_errors = buf->dtb_xamot_errors;
16587 desc.dtbd_oldest = 0;
16588 desc.dtbd_timestamp = buf->dtb_switched;
16590 mutex_exit(&dtrace_lock);
16593 * Finally, copy out the buffer description.
16595 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
16601 case DTRACEIOC_CONF: {
16602 dtrace_conf_t conf;
16604 bzero(&conf, sizeof (conf));
16605 conf.dtc_difversion = DIF_VERSION;
16606 conf.dtc_difintregs = DIF_DIR_NREGS;
16607 conf.dtc_diftupregs = DIF_DTR_NREGS;
16608 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
16610 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
16616 case DTRACEIOC_STATUS: {
16617 dtrace_status_t stat;
16618 dtrace_dstate_t *dstate;
16623 * See the comment in dtrace_state_deadman() for the reason
16624 * for setting dts_laststatus to INT64_MAX before setting
16625 * it to the correct value.
16627 state->dts_laststatus = INT64_MAX;
16628 dtrace_membar_producer();
16629 state->dts_laststatus = dtrace_gethrtime();
16631 bzero(&stat, sizeof (stat));
16633 mutex_enter(&dtrace_lock);
16635 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
16636 mutex_exit(&dtrace_lock);
16640 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
16641 stat.dtst_exiting = 1;
16643 nerrs = state->dts_errors;
16644 dstate = &state->dts_vstate.dtvs_dynvars;
16646 for (i = 0; i < NCPU; i++) {
16647 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
16649 stat.dtst_dyndrops += dcpu->dtdsc_drops;
16650 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
16651 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
16653 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
16654 stat.dtst_filled++;
16656 nerrs += state->dts_buffer[i].dtb_errors;
16658 for (j = 0; j < state->dts_nspeculations; j++) {
16659 dtrace_speculation_t *spec;
16660 dtrace_buffer_t *buf;
16662 spec = &state->dts_speculations[j];
16663 buf = &spec->dtsp_buffer[i];
16664 stat.dtst_specdrops += buf->dtb_xamot_drops;
16668 stat.dtst_specdrops_busy = state->dts_speculations_busy;
16669 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
16670 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
16671 stat.dtst_dblerrors = state->dts_dblerrors;
16673 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
16674 stat.dtst_errors = nerrs;
16676 mutex_exit(&dtrace_lock);
16678 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
16684 case DTRACEIOC_FORMAT: {
16685 dtrace_fmtdesc_t fmt;
16689 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
16692 mutex_enter(&dtrace_lock);
16694 if (fmt.dtfd_format == 0 ||
16695 fmt.dtfd_format > state->dts_nformats) {
16696 mutex_exit(&dtrace_lock);
16701 * Format strings are allocated contiguously and they are
16702 * never freed; if a format index is less than the number
16703 * of formats, we can assert that the format map is non-NULL
16704 * and that the format for the specified index is non-NULL.
16706 ASSERT(state->dts_formats != NULL);
16707 str = state->dts_formats[fmt.dtfd_format - 1];
16708 ASSERT(str != NULL);
16710 len = strlen(str) + 1;
16712 if (len > fmt.dtfd_length) {
16713 fmt.dtfd_length = len;
16715 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
16716 mutex_exit(&dtrace_lock);
16720 if (copyout(str, fmt.dtfd_string, len) != 0) {
16721 mutex_exit(&dtrace_lock);
16726 mutex_exit(&dtrace_lock);
16739 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
16741 dtrace_state_t *state;
16748 return (DDI_SUCCESS);
16751 return (DDI_FAILURE);
16754 mutex_enter(&cpu_lock);
16755 mutex_enter(&dtrace_provider_lock);
16756 mutex_enter(&dtrace_lock);
16758 ASSERT(dtrace_opens == 0);
16760 if (dtrace_helpers > 0) {
16761 mutex_exit(&dtrace_provider_lock);
16762 mutex_exit(&dtrace_lock);
16763 mutex_exit(&cpu_lock);
16764 return (DDI_FAILURE);
16767 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
16768 mutex_exit(&dtrace_provider_lock);
16769 mutex_exit(&dtrace_lock);
16770 mutex_exit(&cpu_lock);
16771 return (DDI_FAILURE);
16774 dtrace_provider = NULL;
16776 if ((state = dtrace_anon_grab()) != NULL) {
16778 * If there were ECBs on this state, the provider should
16779 * have not been allowed to detach; assert that there is
16782 ASSERT(state->dts_necbs == 0);
16783 dtrace_state_destroy(state);
16786 * If we're being detached with anonymous state, we need to
16787 * indicate to the kernel debugger that DTrace is now inactive.
16789 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
16792 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
16793 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
16794 dtrace_cpu_init = NULL;
16795 dtrace_helpers_cleanup = NULL;
16796 dtrace_helpers_fork = NULL;
16797 dtrace_cpustart_init = NULL;
16798 dtrace_cpustart_fini = NULL;
16799 dtrace_debugger_init = NULL;
16800 dtrace_debugger_fini = NULL;
16801 dtrace_modload = NULL;
16802 dtrace_modunload = NULL;
16804 mutex_exit(&cpu_lock);
16806 if (dtrace_helptrace_enabled) {
16807 kmem_free(dtrace_helptrace_buffer, dtrace_helptrace_bufsize);
16808 dtrace_helptrace_buffer = NULL;
16811 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
16812 dtrace_probes = NULL;
16813 dtrace_nprobes = 0;
16815 dtrace_hash_destroy(dtrace_bymod);
16816 dtrace_hash_destroy(dtrace_byfunc);
16817 dtrace_hash_destroy(dtrace_byname);
16818 dtrace_bymod = NULL;
16819 dtrace_byfunc = NULL;
16820 dtrace_byname = NULL;
16822 kmem_cache_destroy(dtrace_state_cache);
16823 vmem_destroy(dtrace_minor);
16824 vmem_destroy(dtrace_arena);
16826 if (dtrace_toxrange != NULL) {
16827 kmem_free(dtrace_toxrange,
16828 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
16829 dtrace_toxrange = NULL;
16830 dtrace_toxranges = 0;
16831 dtrace_toxranges_max = 0;
16834 ddi_remove_minor_node(dtrace_devi, NULL);
16835 dtrace_devi = NULL;
16837 ddi_soft_state_fini(&dtrace_softstate);
16839 ASSERT(dtrace_vtime_references == 0);
16840 ASSERT(dtrace_opens == 0);
16841 ASSERT(dtrace_retained == NULL);
16843 mutex_exit(&dtrace_lock);
16844 mutex_exit(&dtrace_provider_lock);
16847 * We don't destroy the task queue until after we have dropped our
16848 * locks (taskq_destroy() may block on running tasks). To prevent
16849 * attempting to do work after we have effectively detached but before
16850 * the task queue has been destroyed, all tasks dispatched via the
16851 * task queue must check that DTrace is still attached before
16852 * performing any operation.
16854 taskq_destroy(dtrace_taskq);
16855 dtrace_taskq = NULL;
16857 return (DDI_SUCCESS);
16864 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
16869 case DDI_INFO_DEVT2DEVINFO:
16870 *result = (void *)dtrace_devi;
16871 error = DDI_SUCCESS;
16873 case DDI_INFO_DEVT2INSTANCE:
16874 *result = (void *)0;
16875 error = DDI_SUCCESS;
16878 error = DDI_FAILURE;
16885 static struct cb_ops dtrace_cb_ops = {
16886 dtrace_open, /* open */
16887 dtrace_close, /* close */
16888 nulldev, /* strategy */
16889 nulldev, /* print */
16893 dtrace_ioctl, /* ioctl */
16894 nodev, /* devmap */
16896 nodev, /* segmap */
16897 nochpoll, /* poll */
16898 ddi_prop_op, /* cb_prop_op */
16900 D_NEW | D_MP /* Driver compatibility flag */
16903 static struct dev_ops dtrace_ops = {
16904 DEVO_REV, /* devo_rev */
16906 dtrace_info, /* get_dev_info */
16907 nulldev, /* identify */
16908 nulldev, /* probe */
16909 dtrace_attach, /* attach */
16910 dtrace_detach, /* detach */
16912 &dtrace_cb_ops, /* driver operations */
16913 NULL, /* bus operations */
16914 nodev /* dev power */
16917 static struct modldrv modldrv = {
16918 &mod_driverops, /* module type (this is a pseudo driver) */
16919 "Dynamic Tracing", /* name of module */
16920 &dtrace_ops, /* driver ops */
16923 static struct modlinkage modlinkage = {
16932 return (mod_install(&modlinkage));
16936 _info(struct modinfo *modinfop)
16938 return (mod_info(&modlinkage, modinfop));
16944 return (mod_remove(&modlinkage));
16948 static d_ioctl_t dtrace_ioctl;
16949 static d_ioctl_t dtrace_ioctl_helper;
16950 static void dtrace_load(void *);
16951 static int dtrace_unload(void);
16952 #if __FreeBSD_version < 800039
16953 static void dtrace_clone(void *, struct ucred *, char *, int , struct cdev **);
16954 static struct clonedevs *dtrace_clones; /* Ptr to the array of cloned devices. */
16955 static eventhandler_tag eh_tag; /* Event handler tag. */
16957 static struct cdev *dtrace_dev;
16958 static struct cdev *helper_dev;
16961 void dtrace_invop_init(void);
16962 void dtrace_invop_uninit(void);
16964 static struct cdevsw dtrace_cdevsw = {
16965 .d_version = D_VERSION,
16966 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
16967 .d_close = dtrace_close,
16968 .d_ioctl = dtrace_ioctl,
16969 .d_open = dtrace_open,
16970 .d_name = "dtrace",
16973 static struct cdevsw helper_cdevsw = {
16974 .d_version = D_VERSION,
16975 .d_flags = D_TRACKCLOSE | D_NEEDMINOR,
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);