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/endian.h>
38 #include <sys/fcntl.h>
39 #include <sys/filio.h>
41 #include <sys/kernel.h>
43 #include <sys/kthread.h>
44 #include <sys/limits.h>
45 #include <sys/linker.h>
47 #include <sys/malloc.h>
48 #include <sys/module.h>
49 #include <sys/mutex.h>
52 #include <sys/selinfo.h>
54 #include <sys/sysctl.h>
56 #include <sys/unistd.h>
57 #include <machine/cpu.h>
58 #include <machine/stdarg.h>
60 #include <sys/dtrace.h>
61 #include <sys/dtrace_bsd.h>
63 #define PROF_NAMELEN 15
65 #define PROF_PROFILE 0
67 #define PROF_PREFIX_PROFILE "profile-"
68 #define PROF_PREFIX_TICK "tick-"
71 * Regardless of platform, there are five artificial frames in the case of the
80 * On amd64, there are two frames associated with locore: one in locore, and
81 * another in common interrupt dispatch code. (i386 has not been modified to
82 * use this common layer.) Further, on i386, the interrupted instruction
83 * appears as its own stack frame. All of this means that we need to add one
84 * frame for amd64, and then take one away for both amd64 and i386.
86 * On SPARC, the picture is further complicated because the compiler
87 * optimizes away tail-calls -- so the following frames are optimized away:
92 * This gives three frames. However, on DEBUG kernels, the cyclic_expire
93 * frame cannot be tail-call eliminated, yielding four frames in this case.
95 * All of the above constraints lead to the mess below. Yes, the profile
96 * provider should ideally figure this out on-the-fly by hiting one of its own
97 * probes and then walking its own stack trace. This is complicated, however,
98 * and the static definition doesn't seem to be overly brittle. Still, we
99 * allow for a manual override in case we get it completely wrong.
102 #define PROF_ARTIFICIAL_FRAMES 10
105 #define PROF_ARTIFICIAL_FRAMES 6
109 #define PROF_ARTIFICIAL_FRAMES 4
111 #define PROF_ARTIFICIAL_FRAMES 3
119 * This value is bogus just to make module compilable on mips
121 #define PROF_ARTIFICIAL_FRAMES 3
126 * This value is bogus just to make module compilable on powerpc
128 #define PROF_ARTIFICIAL_FRAMES 3
131 struct profile_probe_percpu;
135 #define PROF_ARTIFICIAL_FRAMES 3
139 #define PROF_ARTIFICIAL_FRAMES 3
144 #define PROF_ARTIFICIAL_FRAMES 10
149 #define PROF_ARTIFICIAL_FRAMES 10
152 typedef struct profile_probe {
153 char prof_name[PROF_NAMELEN];
157 hrtime_t prof_interval;
158 cyclic_id_t prof_cyclic;
160 sbintime_t prof_interval;
161 struct callout prof_cyclic;
162 sbintime_t prof_expected;
163 struct profile_probe_percpu **prof_pcpus;
167 typedef struct profile_probe_percpu {
168 hrtime_t profc_expected;
169 hrtime_t profc_interval;
170 profile_probe_t *profc_probe;
172 struct callout profc_cyclic;
174 } profile_probe_percpu_t;
176 static d_open_t profile_open;
177 static int profile_unload(void);
178 static void profile_create(hrtime_t, char *, int);
179 static void profile_destroy(void *, dtrace_id_t, void *);
180 static void profile_enable(void *, dtrace_id_t, void *);
181 static void profile_disable(void *, dtrace_id_t, void *);
182 static void profile_load(void *);
183 static void profile_provide(void *, dtrace_probedesc_t *);
185 static int profile_rates[] = {
186 97, 199, 499, 997, 1999,
192 static int profile_ticks[] = {
193 1, 10, 100, 500, 1000,
199 * profile_max defines the upper bound on the number of profile probes that
200 * can exist (this is to prevent malicious or clumsy users from exhausing
201 * system resources by creating a slew of profile probes). At mod load time,
202 * this gets its value from PROFILE_MAX_DEFAULT or profile-max-probes if it's
203 * present in the profile.conf file.
205 #define PROFILE_MAX_DEFAULT 1000 /* default max. number of probes */
206 static uint32_t profile_max = PROFILE_MAX_DEFAULT;
207 /* maximum number of profile probes */
208 static uint32_t profile_total; /* current number of profile probes */
210 static struct cdevsw profile_cdevsw = {
211 .d_version = D_VERSION,
212 .d_open = profile_open,
216 static dtrace_pattr_t profile_attr = {
217 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
218 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
219 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
220 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
221 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
224 static dtrace_pops_t profile_pops = {
225 .dtps_provide = profile_provide,
226 .dtps_provide_module = NULL,
227 .dtps_enable = profile_enable,
228 .dtps_disable = profile_disable,
229 .dtps_suspend = NULL,
231 .dtps_getargdesc = NULL,
232 .dtps_getargval = NULL,
233 .dtps_usermode = NULL,
234 .dtps_destroy = profile_destroy
237 static struct cdev *profile_cdev;
238 static dtrace_provider_id_t profile_id;
239 static hrtime_t profile_interval_min = NANOSEC / 5000; /* 5000 hz */
240 static int profile_aframes = PROF_ARTIFICIAL_FRAMES;
242 SYSCTL_DECL(_kern_dtrace);
243 SYSCTL_NODE(_kern_dtrace, OID_AUTO, profile, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
244 "DTrace profile parameters");
245 SYSCTL_INT(_kern_dtrace_profile, OID_AUTO, aframes, CTLFLAG_RW, &profile_aframes,
246 0, "Skipped frames for profile provider");
249 nsec_to_sbt(hrtime_t nsec)
254 * We need to calculate nsec * 2^32 / 10^9
255 * Seconds and nanoseconds are split to avoid overflow.
257 sec = nsec / NANOSEC;
258 nsec = nsec % NANOSEC;
259 return (((sbintime_t)sec << 32) | ((sbintime_t)nsec << 32) / NANOSEC);
263 sbt_to_nsec(sbintime_t sbt)
266 return ((sbt >> 32) * NANOSEC +
267 (((uint32_t)sbt * (hrtime_t)NANOSEC) >> 32));
271 profile_probe(profile_probe_t *prof, hrtime_t late)
274 struct trapframe *frame;
281 * td_intr_frame can be unset if this is a catch-up event upon waking up
282 * from idle sleep. This can only happen on a CPU idle thread. Use a
283 * representative arg0 value in this case so that one of the probe
284 * arguments is non-zero.
286 frame = td->td_intr_frame;
288 if (TRAPF_USERMODE(frame))
289 upc = TRAPF_PC(frame);
291 pc = TRAPF_PC(frame);
292 } else if (TD_IS_IDLETHREAD(td))
293 pc = (uintfptr_t)&cpu_idle;
295 dtrace_probe(prof->prof_id, pc, upc, late, 0, 0);
299 profile_fire(void *arg)
301 profile_probe_percpu_t *pcpu = arg;
302 profile_probe_t *prof = pcpu->profc_probe;
305 late = sbt_to_nsec(sbinuptime() - pcpu->profc_expected);
307 profile_probe(prof, late);
308 pcpu->profc_expected += pcpu->profc_interval;
309 callout_schedule_sbt_curcpu(&pcpu->profc_cyclic,
310 pcpu->profc_expected, 0, C_DIRECT_EXEC | C_ABSOLUTE);
314 profile_tick(void *arg)
316 profile_probe_t *prof = arg;
318 profile_probe(prof, 0);
319 prof->prof_expected += prof->prof_interval;
320 callout_schedule_sbt(&prof->prof_cyclic,
321 prof->prof_expected, 0, C_DIRECT_EXEC | C_ABSOLUTE);
325 profile_create(hrtime_t interval, char *name, int kind)
327 profile_probe_t *prof;
329 if (interval < profile_interval_min)
332 if (dtrace_probe_lookup(profile_id, NULL, NULL, name) != 0)
335 atomic_add_32(&profile_total, 1);
336 if (profile_total > profile_max) {
337 atomic_add_32(&profile_total, -1);
341 prof = kmem_zalloc(sizeof (profile_probe_t), KM_SLEEP);
342 (void) strcpy(prof->prof_name, name);
344 prof->prof_interval = interval;
345 prof->prof_cyclic = CYCLIC_NONE;
347 prof->prof_interval = nsec_to_sbt(interval);
348 callout_init(&prof->prof_cyclic, 1);
350 prof->prof_kind = kind;
351 prof->prof_id = dtrace_probe_create(profile_id,
353 profile_aframes, prof);
358 profile_provide(void *arg, dtrace_probedesc_t *desc)
360 int i, j, rate, kind;
361 hrtime_t val = 0, mult = 1, len = 0;
362 char *name, *suffix = NULL;
368 { PROF_PREFIX_PROFILE, PROF_PROFILE },
369 { PROF_PREFIX_TICK, PROF_TICK },
377 { "ns", NANOSEC / NANOSEC },
378 { "nsec", NANOSEC / NANOSEC },
379 { "us", NANOSEC / MICROSEC },
380 { "usec", NANOSEC / MICROSEC },
381 { "ms", NANOSEC / MILLISEC },
382 { "msec", NANOSEC / MILLISEC },
383 { "s", NANOSEC / SEC },
384 { "sec", NANOSEC / SEC },
385 { "m", NANOSEC * (hrtime_t)60 },
386 { "min", NANOSEC * (hrtime_t)60 },
387 { "h", NANOSEC * (hrtime_t)(60 * 60) },
388 { "hour", NANOSEC * (hrtime_t)(60 * 60) },
389 { "d", NANOSEC * (hrtime_t)(24 * 60 * 60) },
390 { "day", NANOSEC * (hrtime_t)(24 * 60 * 60) },
396 char n[PROF_NAMELEN];
399 * If no description was provided, provide all of our probes.
401 for (i = 0; i < sizeof (profile_rates) / sizeof (int); i++) {
402 if ((rate = profile_rates[i]) == 0)
405 (void) snprintf(n, PROF_NAMELEN, "%s%d",
406 PROF_PREFIX_PROFILE, rate);
407 profile_create(NANOSEC / rate, n, PROF_PROFILE);
410 for (i = 0; i < sizeof (profile_ticks) / sizeof (int); i++) {
411 if ((rate = profile_ticks[i]) == 0)
414 (void) snprintf(n, PROF_NAMELEN, "%s%d",
415 PROF_PREFIX_TICK, rate);
416 profile_create(NANOSEC / rate, n, PROF_TICK);
422 name = desc->dtpd_name;
424 for (i = 0; types[i].prefix != NULL; i++) {
425 len = strlen(types[i].prefix);
427 if (strncmp(name, types[i].prefix, len) != 0)
432 if (types[i].prefix == NULL)
435 kind = types[i].kind;
436 j = strlen(name) - len;
439 * We need to start before any time suffix.
441 for (j = strlen(name); j >= len; j--) {
442 if (name[j] >= '0' && name[j] <= '9')
447 ASSERT(suffix != NULL);
450 * Now determine the numerical value present in the probe name.
452 for (; j >= len; j--) {
453 if (name[j] < '0' || name[j] > '9')
456 val += (name[j] - '0') * mult;
457 mult *= (hrtime_t)10;
464 * Look-up the suffix to determine the multiplier.
466 for (i = 0, mult = 0; suffixes[i].name != NULL; i++) {
467 if (strcasecmp(suffixes[i].name, suffix) == 0) {
468 mult = suffixes[i].mult;
473 if (suffixes[i].name == NULL && *suffix != '\0')
478 * The default is frequency-per-second.
485 profile_create(val, name, kind);
490 profile_destroy(void *arg, dtrace_id_t id, void *parg)
492 profile_probe_t *prof = parg;
495 ASSERT(prof->prof_cyclic == CYCLIC_NONE);
497 ASSERT(!callout_active(&prof->prof_cyclic) && prof->prof_pcpus == NULL);
499 kmem_free(prof, sizeof (profile_probe_t));
501 ASSERT(profile_total >= 1);
502 atomic_add_32(&profile_total, -1);
508 profile_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
510 profile_probe_t *prof = arg;
511 profile_probe_percpu_t *pcpu;
513 pcpu = kmem_zalloc(sizeof (profile_probe_percpu_t), KM_SLEEP);
514 pcpu->profc_probe = prof;
516 hdlr->cyh_func = profile_fire;
517 hdlr->cyh_arg = pcpu;
519 when->cyt_interval = prof->prof_interval;
520 when->cyt_when = gethrtime() + when->cyt_interval;
522 pcpu->profc_expected = when->cyt_when;
523 pcpu->profc_interval = when->cyt_interval;
528 profile_offline(void *arg, cpu_t *cpu, void *oarg)
530 profile_probe_percpu_t *pcpu = oarg;
532 ASSERT(pcpu->profc_probe == arg);
533 kmem_free(pcpu, sizeof (profile_probe_percpu_t));
538 profile_enable(void *arg, dtrace_id_t id, void *parg)
540 profile_probe_t *prof = parg;
541 cyc_omni_handler_t omni;
545 ASSERT(prof->prof_interval != 0);
546 ASSERT(MUTEX_HELD(&cpu_lock));
548 if (prof->prof_kind == PROF_TICK) {
549 hdlr.cyh_func = profile_tick;
552 when.cyt_interval = prof->prof_interval;
553 when.cyt_when = gethrtime() + when.cyt_interval;
555 ASSERT(prof->prof_kind == PROF_PROFILE);
556 omni.cyo_online = profile_online;
557 omni.cyo_offline = profile_offline;
561 if (prof->prof_kind == PROF_TICK) {
562 prof->prof_cyclic = cyclic_add(&hdlr, &when);
564 prof->prof_cyclic = cyclic_add_omni(&omni);
570 profile_disable(void *arg, dtrace_id_t id, void *parg)
572 profile_probe_t *prof = parg;
574 ASSERT(prof->prof_cyclic != CYCLIC_NONE);
575 ASSERT(MUTEX_HELD(&cpu_lock));
577 cyclic_remove(prof->prof_cyclic);
578 prof->prof_cyclic = CYCLIC_NONE;
584 profile_enable_omni(profile_probe_t *prof)
586 profile_probe_percpu_t *pcpu;
589 prof->prof_pcpus = kmem_zalloc((mp_maxid + 1) * sizeof(pcpu), KM_SLEEP);
591 pcpu = kmem_zalloc(sizeof(profile_probe_percpu_t), KM_SLEEP);
592 prof->prof_pcpus[cpu] = pcpu;
593 pcpu->profc_probe = prof;
594 pcpu->profc_expected = sbinuptime() + prof->prof_interval;
595 pcpu->profc_interval = prof->prof_interval;
596 callout_init(&pcpu->profc_cyclic, 1);
597 callout_reset_sbt_on(&pcpu->profc_cyclic,
598 pcpu->profc_expected, 0, profile_fire, pcpu,
599 cpu, C_DIRECT_EXEC | C_ABSOLUTE);
604 profile_disable_omni(profile_probe_t *prof)
606 profile_probe_percpu_t *pcpu;
609 ASSERT(prof->prof_pcpus != NULL);
611 pcpu = prof->prof_pcpus[cpu];
612 ASSERT(pcpu->profc_probe == prof);
613 ASSERT(callout_active(&pcpu->profc_cyclic));
614 callout_stop(&pcpu->profc_cyclic);
615 callout_drain(&pcpu->profc_cyclic);
616 kmem_free(pcpu, sizeof(profile_probe_percpu_t));
618 kmem_free(prof->prof_pcpus, (mp_maxid + 1) * sizeof(pcpu));
619 prof->prof_pcpus = NULL;
624 profile_enable(void *arg, dtrace_id_t id, void *parg)
626 profile_probe_t *prof = parg;
628 if (prof->prof_kind == PROF_TICK) {
629 prof->prof_expected = sbinuptime() + prof->prof_interval;
630 callout_reset_sbt(&prof->prof_cyclic,
631 prof->prof_expected, 0, profile_tick, prof,
632 C_DIRECT_EXEC | C_ABSOLUTE);
634 ASSERT(prof->prof_kind == PROF_PROFILE);
635 profile_enable_omni(prof);
641 profile_disable(void *arg, dtrace_id_t id, void *parg)
643 profile_probe_t *prof = parg;
645 if (prof->prof_kind == PROF_TICK) {
646 ASSERT(callout_active(&prof->prof_cyclic));
647 callout_stop(&prof->prof_cyclic);
648 callout_drain(&prof->prof_cyclic);
650 ASSERT(prof->prof_kind == PROF_PROFILE);
651 profile_disable_omni(prof);
657 profile_load(void *dummy)
659 /* Create the /dev/dtrace/profile entry. */
660 profile_cdev = make_dev(&profile_cdevsw, 0, UID_ROOT, GID_WHEEL, 0600,
663 if (dtrace_register("profile", &profile_attr, DTRACE_PRIV_USER,
664 NULL, &profile_pops, NULL, &profile_id) != 0)
674 if ((error = dtrace_unregister(profile_id)) != 0)
677 destroy_dev(profile_cdev);
684 profile_modevent(module_t mod __unused, int type, void *data __unused)
708 profile_open(struct cdev *dev __unused, int oflags __unused, int devtype __unused, struct thread *td __unused)
713 SYSINIT(profile_load, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_load, NULL);
714 SYSUNINIT(profile_unload, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_unload, NULL);
716 DEV_MODULE(profile, profile_modevent, NULL);
717 MODULE_VERSION(profile, 1);
718 MODULE_DEPEND(profile, dtrace, 1, 1, 1);
719 MODULE_DEPEND(profile, opensolaris, 1, 1, 1);