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]
21 * Portions Copyright 2006-2008 John Birrell jb@freebsd.org
28 * Copyright 2006 Sun Microsystems, Inc. All rights reserved.
29 * Use is subject to license terms.
32 #include <sys/cdefs.h>
33 #include <sys/param.h>
34 #include <sys/systm.h>
36 #include <sys/cpuvar.h>
37 #include <sys/fcntl.h>
38 #include <sys/filio.h>
40 #include <sys/kernel.h>
42 #include <sys/kthread.h>
43 #include <sys/limits.h>
44 #include <sys/linker.h>
46 #include <sys/malloc.h>
47 #include <sys/module.h>
48 #include <sys/mutex.h>
51 #include <sys/selinfo.h>
53 #include <sys/sysctl.h>
55 #include <sys/unistd.h>
56 #include <machine/cpu.h>
57 #include <machine/stdarg.h>
59 #include <sys/dtrace.h>
60 #include <sys/dtrace_bsd.h>
62 #define PROF_NAMELEN 15
64 #define PROF_PROFILE 0
66 #define PROF_PREFIX_PROFILE "profile-"
67 #define PROF_PREFIX_TICK "tick-"
70 * Regardless of platform, there are five artificial frames in the case of the
79 * On amd64, there are two frames associated with locore: one in locore, and
80 * another in common interrupt dispatch code. (i386 has not been modified to
81 * use this common layer.) Further, on i386, the interrupted instruction
82 * appears as its own stack frame. All of this means that we need to add one
83 * frame for amd64, and then take one away for both amd64 and i386.
85 * On SPARC, the picture is further complicated because the compiler
86 * optimizes away tail-calls -- so the following frames are optimized away:
91 * This gives three frames. However, on DEBUG kernels, the cyclic_expire
92 * frame cannot be tail-call eliminated, yielding four frames in this case.
94 * All of the above constraints lead to the mess below. Yes, the profile
95 * provider should ideally figure this out on-the-fly by hiting one of its own
96 * probes and then walking its own stack trace. This is complicated, however,
97 * and the static definition doesn't seem to be overly brittle. Still, we
98 * allow for a manual override in case we get it completely wrong.
101 #define PROF_ARTIFICIAL_FRAMES 10
104 #define PROF_ARTIFICIAL_FRAMES 6
108 #define PROF_ARTIFICIAL_FRAMES 4
110 #define PROF_ARTIFICIAL_FRAMES 3
118 * This value is bogus just to make module compilable on mips
120 #define PROF_ARTIFICIAL_FRAMES 3
125 * This value is bogus just to make module compilable on powerpc
127 #define PROF_ARTIFICIAL_FRAMES 3
130 struct profile_probe_percpu;
134 #define PROF_ARTIFICIAL_FRAMES 3
138 #define PROF_ARTIFICIAL_FRAMES 3
143 #define PROF_ARTIFICIAL_FRAMES 10
148 #define PROF_ARTIFICIAL_FRAMES 10
151 typedef struct profile_probe {
152 char prof_name[PROF_NAMELEN];
156 hrtime_t prof_interval;
157 cyclic_id_t prof_cyclic;
159 sbintime_t prof_interval;
160 struct callout prof_cyclic;
161 sbintime_t prof_expected;
162 struct profile_probe_percpu **prof_pcpus;
166 typedef struct profile_probe_percpu {
167 hrtime_t profc_expected;
168 hrtime_t profc_interval;
169 profile_probe_t *profc_probe;
171 struct callout profc_cyclic;
173 } profile_probe_percpu_t;
175 static d_open_t profile_open;
176 static int profile_unload(void);
177 static void profile_create(hrtime_t, char *, int);
178 static void profile_destroy(void *, dtrace_id_t, void *);
179 static void profile_enable(void *, dtrace_id_t, void *);
180 static void profile_disable(void *, dtrace_id_t, void *);
181 static void profile_load(void *);
182 static void profile_provide(void *, dtrace_probedesc_t *);
184 static int profile_rates[] = {
185 97, 199, 499, 997, 1999,
191 static int profile_ticks[] = {
192 1, 10, 100, 500, 1000,
198 * profile_max defines the upper bound on the number of profile probes that
199 * can exist (this is to prevent malicious or clumsy users from exhausing
200 * system resources by creating a slew of profile probes). At mod load time,
201 * this gets its value from PROFILE_MAX_DEFAULT or profile-max-probes if it's
202 * present in the profile.conf file.
204 #define PROFILE_MAX_DEFAULT 1000 /* default max. number of probes */
205 static uint32_t profile_max = PROFILE_MAX_DEFAULT;
206 /* maximum number of profile probes */
207 static uint32_t profile_total; /* current number of profile probes */
209 static struct cdevsw profile_cdevsw = {
210 .d_version = D_VERSION,
211 .d_open = profile_open,
215 static dtrace_pattr_t profile_attr = {
216 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
217 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
218 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
219 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
220 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
223 static dtrace_pops_t profile_pops = {
224 .dtps_provide = profile_provide,
225 .dtps_provide_module = NULL,
226 .dtps_enable = profile_enable,
227 .dtps_disable = profile_disable,
228 .dtps_suspend = NULL,
230 .dtps_getargdesc = NULL,
231 .dtps_getargval = NULL,
232 .dtps_usermode = NULL,
233 .dtps_destroy = profile_destroy
236 static struct cdev *profile_cdev;
237 static dtrace_provider_id_t profile_id;
238 static hrtime_t profile_interval_min = NANOSEC / 5000; /* 5000 hz */
239 static int profile_aframes = PROF_ARTIFICIAL_FRAMES;
241 SYSCTL_DECL(_kern_dtrace);
242 SYSCTL_NODE(_kern_dtrace, OID_AUTO, profile, CTLFLAG_RD, 0, "DTrace profile parameters");
243 SYSCTL_INT(_kern_dtrace_profile, OID_AUTO, aframes, CTLFLAG_RW, &profile_aframes,
244 0, "Skipped frames for profile provider");
247 nsec_to_sbt(hrtime_t nsec)
252 * We need to calculate nsec * 2^32 / 10^9
253 * Seconds and nanoseconds are split to avoid overflow.
255 sec = nsec / NANOSEC;
256 nsec = nsec % NANOSEC;
257 return (((sbintime_t)sec << 32) | ((sbintime_t)nsec << 32) / NANOSEC);
261 sbt_to_nsec(sbintime_t sbt)
264 return ((sbt >> 32) * NANOSEC +
265 (((uint32_t)sbt * (hrtime_t)NANOSEC) >> 32));
269 profile_probe(profile_probe_t *prof, hrtime_t late)
272 struct trapframe *frame;
279 * td_intr_frame can be unset if this is a catch-up event upon waking up
280 * from idle sleep. This can only happen on a CPU idle thread. Use a
281 * representative arg0 value in this case so that one of the probe
282 * arguments is non-zero.
284 frame = td->td_intr_frame;
286 if (TRAPF_USERMODE(frame))
287 upc = TRAPF_PC(frame);
289 pc = TRAPF_PC(frame);
290 } else if (TD_IS_IDLETHREAD(td))
291 pc = (uintfptr_t)&cpu_idle;
293 dtrace_probe(prof->prof_id, pc, upc, late, 0, 0);
297 profile_fire(void *arg)
299 profile_probe_percpu_t *pcpu = arg;
300 profile_probe_t *prof = pcpu->profc_probe;
303 late = sbt_to_nsec(sbinuptime() - pcpu->profc_expected);
305 profile_probe(prof, late);
306 pcpu->profc_expected += pcpu->profc_interval;
307 callout_schedule_sbt_curcpu(&pcpu->profc_cyclic,
308 pcpu->profc_expected, 0, C_DIRECT_EXEC | C_ABSOLUTE);
312 profile_tick(void *arg)
314 profile_probe_t *prof = arg;
316 profile_probe(prof, 0);
317 prof->prof_expected += prof->prof_interval;
318 callout_schedule_sbt(&prof->prof_cyclic,
319 prof->prof_expected, 0, C_DIRECT_EXEC | C_ABSOLUTE);
323 profile_create(hrtime_t interval, char *name, int kind)
325 profile_probe_t *prof;
327 if (interval < profile_interval_min)
330 if (dtrace_probe_lookup(profile_id, NULL, NULL, name) != 0)
333 atomic_add_32(&profile_total, 1);
334 if (profile_total > profile_max) {
335 atomic_add_32(&profile_total, -1);
339 prof = kmem_zalloc(sizeof (profile_probe_t), KM_SLEEP);
340 (void) strcpy(prof->prof_name, name);
342 prof->prof_interval = interval;
343 prof->prof_cyclic = CYCLIC_NONE;
345 prof->prof_interval = nsec_to_sbt(interval);
346 callout_init(&prof->prof_cyclic, 1);
348 prof->prof_kind = kind;
349 prof->prof_id = dtrace_probe_create(profile_id,
351 profile_aframes, prof);
356 profile_provide(void *arg, dtrace_probedesc_t *desc)
358 int i, j, rate, kind;
359 hrtime_t val = 0, mult = 1, len = 0;
360 char *name, *suffix = NULL;
366 { PROF_PREFIX_PROFILE, PROF_PROFILE },
367 { PROF_PREFIX_TICK, PROF_TICK },
375 { "ns", NANOSEC / NANOSEC },
376 { "nsec", NANOSEC / NANOSEC },
377 { "us", NANOSEC / MICROSEC },
378 { "usec", NANOSEC / MICROSEC },
379 { "ms", NANOSEC / MILLISEC },
380 { "msec", NANOSEC / MILLISEC },
381 { "s", NANOSEC / SEC },
382 { "sec", NANOSEC / SEC },
383 { "m", NANOSEC * (hrtime_t)60 },
384 { "min", NANOSEC * (hrtime_t)60 },
385 { "h", NANOSEC * (hrtime_t)(60 * 60) },
386 { "hour", NANOSEC * (hrtime_t)(60 * 60) },
387 { "d", NANOSEC * (hrtime_t)(24 * 60 * 60) },
388 { "day", NANOSEC * (hrtime_t)(24 * 60 * 60) },
394 char n[PROF_NAMELEN];
397 * If no description was provided, provide all of our probes.
399 for (i = 0; i < sizeof (profile_rates) / sizeof (int); i++) {
400 if ((rate = profile_rates[i]) == 0)
403 (void) snprintf(n, PROF_NAMELEN, "%s%d",
404 PROF_PREFIX_PROFILE, rate);
405 profile_create(NANOSEC / rate, n, PROF_PROFILE);
408 for (i = 0; i < sizeof (profile_ticks) / sizeof (int); i++) {
409 if ((rate = profile_ticks[i]) == 0)
412 (void) snprintf(n, PROF_NAMELEN, "%s%d",
413 PROF_PREFIX_TICK, rate);
414 profile_create(NANOSEC / rate, n, PROF_TICK);
420 name = desc->dtpd_name;
422 for (i = 0; types[i].prefix != NULL; i++) {
423 len = strlen(types[i].prefix);
425 if (strncmp(name, types[i].prefix, len) != 0)
430 if (types[i].prefix == NULL)
433 kind = types[i].kind;
434 j = strlen(name) - len;
437 * We need to start before any time suffix.
439 for (j = strlen(name); j >= len; j--) {
440 if (name[j] >= '0' && name[j] <= '9')
445 ASSERT(suffix != NULL);
448 * Now determine the numerical value present in the probe name.
450 for (; j >= len; j--) {
451 if (name[j] < '0' || name[j] > '9')
454 val += (name[j] - '0') * mult;
455 mult *= (hrtime_t)10;
462 * Look-up the suffix to determine the multiplier.
464 for (i = 0, mult = 0; suffixes[i].name != NULL; i++) {
465 if (strcasecmp(suffixes[i].name, suffix) == 0) {
466 mult = suffixes[i].mult;
471 if (suffixes[i].name == NULL && *suffix != '\0')
476 * The default is frequency-per-second.
483 profile_create(val, name, kind);
488 profile_destroy(void *arg, dtrace_id_t id, void *parg)
490 profile_probe_t *prof = parg;
493 ASSERT(prof->prof_cyclic == CYCLIC_NONE);
495 ASSERT(!callout_active(&prof->prof_cyclic) && prof->prof_pcpus == NULL);
497 kmem_free(prof, sizeof (profile_probe_t));
499 ASSERT(profile_total >= 1);
500 atomic_add_32(&profile_total, -1);
506 profile_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
508 profile_probe_t *prof = arg;
509 profile_probe_percpu_t *pcpu;
511 pcpu = kmem_zalloc(sizeof (profile_probe_percpu_t), KM_SLEEP);
512 pcpu->profc_probe = prof;
514 hdlr->cyh_func = profile_fire;
515 hdlr->cyh_arg = pcpu;
517 when->cyt_interval = prof->prof_interval;
518 when->cyt_when = gethrtime() + when->cyt_interval;
520 pcpu->profc_expected = when->cyt_when;
521 pcpu->profc_interval = when->cyt_interval;
526 profile_offline(void *arg, cpu_t *cpu, void *oarg)
528 profile_probe_percpu_t *pcpu = oarg;
530 ASSERT(pcpu->profc_probe == arg);
531 kmem_free(pcpu, sizeof (profile_probe_percpu_t));
536 profile_enable(void *arg, dtrace_id_t id, void *parg)
538 profile_probe_t *prof = parg;
539 cyc_omni_handler_t omni;
543 ASSERT(prof->prof_interval != 0);
544 ASSERT(MUTEX_HELD(&cpu_lock));
546 if (prof->prof_kind == PROF_TICK) {
547 hdlr.cyh_func = profile_tick;
550 when.cyt_interval = prof->prof_interval;
551 when.cyt_when = gethrtime() + when.cyt_interval;
553 ASSERT(prof->prof_kind == PROF_PROFILE);
554 omni.cyo_online = profile_online;
555 omni.cyo_offline = profile_offline;
559 if (prof->prof_kind == PROF_TICK) {
560 prof->prof_cyclic = cyclic_add(&hdlr, &when);
562 prof->prof_cyclic = cyclic_add_omni(&omni);
568 profile_disable(void *arg, dtrace_id_t id, void *parg)
570 profile_probe_t *prof = parg;
572 ASSERT(prof->prof_cyclic != CYCLIC_NONE);
573 ASSERT(MUTEX_HELD(&cpu_lock));
575 cyclic_remove(prof->prof_cyclic);
576 prof->prof_cyclic = CYCLIC_NONE;
582 profile_enable_omni(profile_probe_t *prof)
584 profile_probe_percpu_t *pcpu;
587 prof->prof_pcpus = kmem_zalloc((mp_maxid + 1) * sizeof(pcpu), KM_SLEEP);
589 pcpu = kmem_zalloc(sizeof(profile_probe_percpu_t), KM_SLEEP);
590 prof->prof_pcpus[cpu] = pcpu;
591 pcpu->profc_probe = prof;
592 pcpu->profc_expected = sbinuptime() + prof->prof_interval;
593 pcpu->profc_interval = prof->prof_interval;
594 callout_init(&pcpu->profc_cyclic, 1);
595 callout_reset_sbt_on(&pcpu->profc_cyclic,
596 pcpu->profc_expected, 0, profile_fire, pcpu,
597 cpu, C_DIRECT_EXEC | C_ABSOLUTE);
602 profile_disable_omni(profile_probe_t *prof)
604 profile_probe_percpu_t *pcpu;
607 ASSERT(prof->prof_pcpus != NULL);
609 pcpu = prof->prof_pcpus[cpu];
610 ASSERT(pcpu->profc_probe == prof);
611 ASSERT(callout_active(&pcpu->profc_cyclic));
612 callout_stop(&pcpu->profc_cyclic);
613 callout_drain(&pcpu->profc_cyclic);
614 kmem_free(pcpu, sizeof(profile_probe_percpu_t));
616 kmem_free(prof->prof_pcpus, (mp_maxid + 1) * sizeof(pcpu));
617 prof->prof_pcpus = NULL;
622 profile_enable(void *arg, dtrace_id_t id, void *parg)
624 profile_probe_t *prof = parg;
626 if (prof->prof_kind == PROF_TICK) {
627 prof->prof_expected = sbinuptime() + prof->prof_interval;
628 callout_reset_sbt(&prof->prof_cyclic,
629 prof->prof_expected, 0, profile_tick, prof,
630 C_DIRECT_EXEC | C_ABSOLUTE);
632 ASSERT(prof->prof_kind == PROF_PROFILE);
633 profile_enable_omni(prof);
639 profile_disable(void *arg, dtrace_id_t id, void *parg)
641 profile_probe_t *prof = parg;
643 if (prof->prof_kind == PROF_TICK) {
644 ASSERT(callout_active(&prof->prof_cyclic));
645 callout_stop(&prof->prof_cyclic);
646 callout_drain(&prof->prof_cyclic);
648 ASSERT(prof->prof_kind == PROF_PROFILE);
649 profile_disable_omni(prof);
655 profile_load(void *dummy)
657 /* Create the /dev/dtrace/profile entry. */
658 profile_cdev = make_dev(&profile_cdevsw, 0, UID_ROOT, GID_WHEEL, 0600,
661 if (dtrace_register("profile", &profile_attr, DTRACE_PRIV_USER,
662 NULL, &profile_pops, NULL, &profile_id) != 0)
672 if ((error = dtrace_unregister(profile_id)) != 0)
675 destroy_dev(profile_cdev);
682 profile_modevent(module_t mod __unused, int type, void *data __unused)
706 profile_open(struct cdev *dev __unused, int oflags __unused, int devtype __unused, struct thread *td __unused)
711 SYSINIT(profile_load, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_load, NULL);
712 SYSUNINIT(profile_unload, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_unload, NULL);
714 DEV_MODULE(profile, profile_modevent, NULL);
715 MODULE_VERSION(profile, 1);
716 MODULE_DEPEND(profile, dtrace, 1, 1, 1);
717 MODULE_DEPEND(profile, opensolaris, 1, 1, 1);