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 * All of the above constraints lead to the mess below. Yes, the profile
87 * provider should ideally figure this out on-the-fly by hiting one of its own
88 * probes and then walking its own stack trace. This is complicated, however,
89 * and the static definition doesn't seem to be overly brittle. Still, we
90 * allow for a manual override in case we get it completely wrong.
93 #define PROF_ARTIFICIAL_FRAMES 10
96 #define PROF_ARTIFICIAL_FRAMES 6
102 * This value is bogus just to make module compilable on mips
104 #define PROF_ARTIFICIAL_FRAMES 3
109 * This value is bogus just to make module compilable on powerpc
111 #define PROF_ARTIFICIAL_FRAMES 3
114 struct profile_probe_percpu;
118 #define PROF_ARTIFICIAL_FRAMES 3
122 #define PROF_ARTIFICIAL_FRAMES 3
126 #define PROF_ARTIFICIAL_FRAMES 12
131 #define PROF_ARTIFICIAL_FRAMES 10
134 typedef struct profile_probe {
135 char prof_name[PROF_NAMELEN];
139 hrtime_t prof_interval;
140 cyclic_id_t prof_cyclic;
142 sbintime_t prof_interval;
143 struct callout prof_cyclic;
144 sbintime_t prof_expected;
145 struct profile_probe_percpu **prof_pcpus;
149 typedef struct profile_probe_percpu {
150 hrtime_t profc_expected;
151 hrtime_t profc_interval;
152 profile_probe_t *profc_probe;
154 struct callout profc_cyclic;
156 } profile_probe_percpu_t;
158 static d_open_t profile_open;
159 static int profile_unload(void);
160 static void profile_create(hrtime_t, char *, int);
161 static void profile_destroy(void *, dtrace_id_t, void *);
162 static void profile_enable(void *, dtrace_id_t, void *);
163 static void profile_disable(void *, dtrace_id_t, void *);
164 static void profile_load(void *);
165 static void profile_provide(void *, dtrace_probedesc_t *);
167 static int profile_rates[] = {
168 97, 199, 499, 997, 1999,
174 static int profile_ticks[] = {
175 1, 10, 100, 500, 1000,
181 * profile_max defines the upper bound on the number of profile probes that
182 * can exist (this is to prevent malicious or clumsy users from exhausing
183 * system resources by creating a slew of profile probes). At mod load time,
184 * this gets its value from PROFILE_MAX_DEFAULT or profile-max-probes if it's
185 * present in the profile.conf file.
187 #define PROFILE_MAX_DEFAULT 1000 /* default max. number of probes */
188 static uint32_t profile_max = PROFILE_MAX_DEFAULT;
189 /* maximum number of profile probes */
190 static uint32_t profile_total; /* current number of profile probes */
192 static struct cdevsw profile_cdevsw = {
193 .d_version = D_VERSION,
194 .d_open = profile_open,
198 static dtrace_pattr_t profile_attr = {
199 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
200 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
201 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
202 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
203 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
206 static dtrace_pops_t profile_pops = {
207 .dtps_provide = profile_provide,
208 .dtps_provide_module = NULL,
209 .dtps_enable = profile_enable,
210 .dtps_disable = profile_disable,
211 .dtps_suspend = NULL,
213 .dtps_getargdesc = NULL,
214 .dtps_getargval = NULL,
215 .dtps_usermode = NULL,
216 .dtps_destroy = profile_destroy
219 static struct cdev *profile_cdev;
220 static dtrace_provider_id_t profile_id;
221 static hrtime_t profile_interval_min = NANOSEC / 5000; /* 5000 hz */
222 static int profile_aframes = PROF_ARTIFICIAL_FRAMES;
224 SYSCTL_DECL(_kern_dtrace);
225 SYSCTL_NODE(_kern_dtrace, OID_AUTO, profile, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
226 "DTrace profile parameters");
227 SYSCTL_INT(_kern_dtrace_profile, OID_AUTO, aframes, CTLFLAG_RW, &profile_aframes,
228 0, "Skipped frames for profile provider");
231 nsec_to_sbt(hrtime_t nsec)
236 * We need to calculate nsec * 2^32 / 10^9
237 * Seconds and nanoseconds are split to avoid overflow.
239 sec = nsec / NANOSEC;
240 nsec = nsec % NANOSEC;
241 return (((sbintime_t)sec << 32) | ((sbintime_t)nsec << 32) / NANOSEC);
245 sbt_to_nsec(sbintime_t sbt)
248 return ((sbt >> 32) * NANOSEC +
249 (((uint32_t)sbt * (hrtime_t)NANOSEC) >> 32));
253 profile_probe(profile_probe_t *prof, hrtime_t late)
256 struct trapframe *frame;
263 * td_intr_frame can be unset if this is a catch-up event upon waking up
264 * from idle sleep. This can only happen on a CPU idle thread. Use a
265 * representative arg0 value in this case so that one of the probe
266 * arguments is non-zero.
268 frame = td->td_intr_frame;
270 if (TRAPF_USERMODE(frame))
271 upc = TRAPF_PC(frame);
273 pc = TRAPF_PC(frame);
274 } else if (TD_IS_IDLETHREAD(td))
275 pc = (uintfptr_t)&cpu_idle;
277 dtrace_probe(prof->prof_id, pc, upc, late, 0, 0);
281 profile_fire(void *arg)
283 profile_probe_percpu_t *pcpu = arg;
284 profile_probe_t *prof = pcpu->profc_probe;
287 late = sbt_to_nsec(sbinuptime() - pcpu->profc_expected);
289 profile_probe(prof, late);
290 pcpu->profc_expected += pcpu->profc_interval;
291 callout_schedule_sbt_curcpu(&pcpu->profc_cyclic,
292 pcpu->profc_expected, 0, C_DIRECT_EXEC | C_ABSOLUTE);
296 profile_tick(void *arg)
298 profile_probe_t *prof = arg;
300 profile_probe(prof, 0);
301 prof->prof_expected += prof->prof_interval;
302 callout_schedule_sbt(&prof->prof_cyclic,
303 prof->prof_expected, 0, C_DIRECT_EXEC | C_ABSOLUTE);
307 profile_create(hrtime_t interval, char *name, int kind)
309 profile_probe_t *prof;
311 if (interval < profile_interval_min)
314 if (dtrace_probe_lookup(profile_id, NULL, NULL, name) != 0)
317 atomic_add_32(&profile_total, 1);
318 if (profile_total > profile_max) {
319 atomic_add_32(&profile_total, -1);
323 prof = kmem_zalloc(sizeof (profile_probe_t), KM_SLEEP);
324 (void) strcpy(prof->prof_name, name);
326 prof->prof_interval = interval;
327 prof->prof_cyclic = CYCLIC_NONE;
329 prof->prof_interval = nsec_to_sbt(interval);
330 callout_init(&prof->prof_cyclic, 1);
332 prof->prof_kind = kind;
333 prof->prof_id = dtrace_probe_create(profile_id,
335 profile_aframes, prof);
340 profile_provide(void *arg, dtrace_probedesc_t *desc)
342 int i, j, rate, kind;
343 hrtime_t val = 0, mult = 1, len = 0;
344 char *name, *suffix = NULL;
350 { PROF_PREFIX_PROFILE, PROF_PROFILE },
351 { PROF_PREFIX_TICK, PROF_TICK },
359 { "ns", NANOSEC / NANOSEC },
360 { "nsec", NANOSEC / NANOSEC },
361 { "us", NANOSEC / MICROSEC },
362 { "usec", NANOSEC / MICROSEC },
363 { "ms", NANOSEC / MILLISEC },
364 { "msec", NANOSEC / MILLISEC },
365 { "s", NANOSEC / SEC },
366 { "sec", NANOSEC / SEC },
367 { "m", NANOSEC * (hrtime_t)60 },
368 { "min", NANOSEC * (hrtime_t)60 },
369 { "h", NANOSEC * (hrtime_t)(60 * 60) },
370 { "hour", NANOSEC * (hrtime_t)(60 * 60) },
371 { "d", NANOSEC * (hrtime_t)(24 * 60 * 60) },
372 { "day", NANOSEC * (hrtime_t)(24 * 60 * 60) },
378 char n[PROF_NAMELEN];
381 * If no description was provided, provide all of our probes.
383 for (i = 0; i < sizeof (profile_rates) / sizeof (int); i++) {
384 if ((rate = profile_rates[i]) == 0)
387 (void) snprintf(n, PROF_NAMELEN, "%s%d",
388 PROF_PREFIX_PROFILE, rate);
389 profile_create(NANOSEC / rate, n, PROF_PROFILE);
392 for (i = 0; i < sizeof (profile_ticks) / sizeof (int); i++) {
393 if ((rate = profile_ticks[i]) == 0)
396 (void) snprintf(n, PROF_NAMELEN, "%s%d",
397 PROF_PREFIX_TICK, rate);
398 profile_create(NANOSEC / rate, n, PROF_TICK);
404 name = desc->dtpd_name;
406 for (i = 0; types[i].prefix != NULL; i++) {
407 len = strlen(types[i].prefix);
409 if (strncmp(name, types[i].prefix, len) != 0)
414 if (types[i].prefix == NULL)
417 kind = types[i].kind;
418 j = strlen(name) - len;
421 * We need to start before any time suffix.
423 for (j = strlen(name); j >= len; j--) {
424 if (name[j] >= '0' && name[j] <= '9')
429 ASSERT(suffix != NULL);
432 * Now determine the numerical value present in the probe name.
434 for (; j >= len; j--) {
435 if (name[j] < '0' || name[j] > '9')
438 val += (name[j] - '0') * mult;
439 mult *= (hrtime_t)10;
446 * Look-up the suffix to determine the multiplier.
448 for (i = 0, mult = 0; suffixes[i].name != NULL; i++) {
449 if (strcasecmp(suffixes[i].name, suffix) == 0) {
450 mult = suffixes[i].mult;
455 if (suffixes[i].name == NULL && *suffix != '\0')
460 * The default is frequency-per-second.
467 profile_create(val, name, kind);
472 profile_destroy(void *arg, dtrace_id_t id, void *parg)
474 profile_probe_t *prof = parg;
477 ASSERT(prof->prof_cyclic == CYCLIC_NONE);
479 ASSERT(!callout_active(&prof->prof_cyclic) && prof->prof_pcpus == NULL);
481 kmem_free(prof, sizeof (profile_probe_t));
483 ASSERT(profile_total >= 1);
484 atomic_add_32(&profile_total, -1);
490 profile_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
492 profile_probe_t *prof = arg;
493 profile_probe_percpu_t *pcpu;
495 pcpu = kmem_zalloc(sizeof (profile_probe_percpu_t), KM_SLEEP);
496 pcpu->profc_probe = prof;
498 hdlr->cyh_func = profile_fire;
499 hdlr->cyh_arg = pcpu;
501 when->cyt_interval = prof->prof_interval;
502 when->cyt_when = gethrtime() + when->cyt_interval;
504 pcpu->profc_expected = when->cyt_when;
505 pcpu->profc_interval = when->cyt_interval;
510 profile_offline(void *arg, cpu_t *cpu, void *oarg)
512 profile_probe_percpu_t *pcpu = oarg;
514 ASSERT(pcpu->profc_probe == arg);
515 kmem_free(pcpu, sizeof (profile_probe_percpu_t));
520 profile_enable(void *arg, dtrace_id_t id, void *parg)
522 profile_probe_t *prof = parg;
523 cyc_omni_handler_t omni;
527 ASSERT(prof->prof_interval != 0);
528 ASSERT(MUTEX_HELD(&cpu_lock));
530 if (prof->prof_kind == PROF_TICK) {
531 hdlr.cyh_func = profile_tick;
534 when.cyt_interval = prof->prof_interval;
535 when.cyt_when = gethrtime() + when.cyt_interval;
537 ASSERT(prof->prof_kind == PROF_PROFILE);
538 omni.cyo_online = profile_online;
539 omni.cyo_offline = profile_offline;
543 if (prof->prof_kind == PROF_TICK) {
544 prof->prof_cyclic = cyclic_add(&hdlr, &when);
546 prof->prof_cyclic = cyclic_add_omni(&omni);
552 profile_disable(void *arg, dtrace_id_t id, void *parg)
554 profile_probe_t *prof = parg;
556 ASSERT(prof->prof_cyclic != CYCLIC_NONE);
557 ASSERT(MUTEX_HELD(&cpu_lock));
559 cyclic_remove(prof->prof_cyclic);
560 prof->prof_cyclic = CYCLIC_NONE;
566 profile_enable_omni(profile_probe_t *prof)
568 profile_probe_percpu_t *pcpu;
571 prof->prof_pcpus = kmem_zalloc((mp_maxid + 1) * sizeof(pcpu), KM_SLEEP);
573 pcpu = kmem_zalloc(sizeof(profile_probe_percpu_t), KM_SLEEP);
574 prof->prof_pcpus[cpu] = pcpu;
575 pcpu->profc_probe = prof;
576 pcpu->profc_expected = sbinuptime() + prof->prof_interval;
577 pcpu->profc_interval = prof->prof_interval;
578 callout_init(&pcpu->profc_cyclic, 1);
579 callout_reset_sbt_on(&pcpu->profc_cyclic,
580 pcpu->profc_expected, 0, profile_fire, pcpu,
581 cpu, C_DIRECT_EXEC | C_ABSOLUTE);
586 profile_disable_omni(profile_probe_t *prof)
588 profile_probe_percpu_t *pcpu;
591 ASSERT(prof->prof_pcpus != NULL);
593 pcpu = prof->prof_pcpus[cpu];
594 ASSERT(pcpu->profc_probe == prof);
595 ASSERT(callout_active(&pcpu->profc_cyclic));
596 callout_stop(&pcpu->profc_cyclic);
597 callout_drain(&pcpu->profc_cyclic);
598 kmem_free(pcpu, sizeof(profile_probe_percpu_t));
600 kmem_free(prof->prof_pcpus, (mp_maxid + 1) * sizeof(pcpu));
601 prof->prof_pcpus = NULL;
606 profile_enable(void *arg, dtrace_id_t id, void *parg)
608 profile_probe_t *prof = parg;
610 if (prof->prof_kind == PROF_TICK) {
611 prof->prof_expected = sbinuptime() + prof->prof_interval;
612 callout_reset_sbt(&prof->prof_cyclic,
613 prof->prof_expected, 0, profile_tick, prof,
614 C_DIRECT_EXEC | C_ABSOLUTE);
616 ASSERT(prof->prof_kind == PROF_PROFILE);
617 profile_enable_omni(prof);
623 profile_disable(void *arg, dtrace_id_t id, void *parg)
625 profile_probe_t *prof = parg;
627 if (prof->prof_kind == PROF_TICK) {
628 ASSERT(callout_active(&prof->prof_cyclic));
629 callout_stop(&prof->prof_cyclic);
630 callout_drain(&prof->prof_cyclic);
632 ASSERT(prof->prof_kind == PROF_PROFILE);
633 profile_disable_omni(prof);
639 profile_load(void *dummy)
641 /* Create the /dev/dtrace/profile entry. */
642 profile_cdev = make_dev(&profile_cdevsw, 0, UID_ROOT, GID_WHEEL, 0600,
645 if (dtrace_register("profile", &profile_attr, DTRACE_PRIV_USER,
646 NULL, &profile_pops, NULL, &profile_id) != 0)
656 if ((error = dtrace_unregister(profile_id)) != 0)
659 destroy_dev(profile_cdev);
666 profile_modevent(module_t mod __unused, int type, void *data __unused)
690 profile_open(struct cdev *dev __unused, int oflags __unused, int devtype __unused, struct thread *td __unused)
695 SYSINIT(profile_load, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_load, NULL);
696 SYSUNINIT(profile_unload, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_unload, NULL);
698 DEV_MODULE(profile, profile_modevent, NULL);
699 MODULE_VERSION(profile, 1);
700 MODULE_DEPEND(profile, dtrace, 1, 1, 1);
701 MODULE_DEPEND(profile, opensolaris, 1, 1, 1);