2 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * This module holds the global variables and functions used to maintain
29 * lock_object structures.
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
36 #include "opt_mprof.h"
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
43 #include <sys/lock_profile.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
49 #include <sys/sched.h>
51 #include <sys/sysctl.h>
57 #include <machine/cpufunc.h>
59 SDT_PROVIDER_DEFINE(lock);
60 SDT_PROBE_DEFINE1(lock, , , starvation, "u_int");
62 CTASSERT(LOCK_CLASS_MAX == 15);
64 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
66 &lock_class_mtx_sleep,
69 &lock_class_rm_sleepable,
75 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
76 const char *type, int flags)
80 /* Check for double-init and zero object. */
81 KASSERT(flags & LO_NEW || !lock_initialized(lock),
82 ("lock \"%s\" %p already initialized", name, lock));
84 /* Look up lock class to find its index. */
85 for (i = 0; i < LOCK_CLASS_MAX; i++)
86 if (lock_classes[i] == class) {
87 lock->lo_flags = i << LO_CLASSSHIFT;
90 KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
92 /* Initialize the lock object. */
94 lock->lo_flags |= flags | LO_INITIALIZED;
95 LOCK_LOG_INIT(lock, 0);
96 WITNESS_INIT(lock, (type != NULL) ? type : name);
100 lock_destroy(struct lock_object *lock)
103 KASSERT(lock_initialized(lock), ("lock %p is not initialized", lock));
104 WITNESS_DESTROY(lock);
105 LOCK_LOG_DESTROY(lock, 0);
106 lock->lo_flags &= ~LO_INITIALIZED;
109 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD, NULL, "lock debugging");
110 static SYSCTL_NODE(_debug_lock, OID_AUTO, delay, CTLFLAG_RD, NULL,
113 static u_int __read_mostly starvation_limit = 131072;
114 SYSCTL_INT(_debug_lock_delay, OID_AUTO, starvation_limit, CTLFLAG_RW,
115 &starvation_limit, 0, "");
117 static u_int __read_mostly restrict_starvation = 0;
118 SYSCTL_INT(_debug_lock_delay, OID_AUTO, restrict_starvation, CTLFLAG_RW,
119 &restrict_starvation, 0, "");
122 lock_delay(struct lock_delay_arg *la)
124 struct lock_delay_config *lc = la->config;
128 if (__predict_false(la->delay > lc->max))
131 for (i = la->delay; i > 0; i--)
134 la->spin_cnt += la->delay;
135 if (__predict_false(la->spin_cnt > starvation_limit)) {
136 SDT_PROBE1(lock, , , starvation, la->delay);
137 if (restrict_starvation)
138 la->delay = lc->base;
143 lock_roundup_2(u_int val)
147 for (res = 1; res <= val; res <<= 1)
154 lock_delay_default_init(struct lock_delay_config *lc)
157 lc->base = lock_roundup_2(mp_ncpus) / 4;
158 lc->max = lc->base * 1024;
162 DB_SHOW_COMMAND(lock, db_show_lock)
164 struct lock_object *lock;
165 struct lock_class *class;
169 lock = (struct lock_object *)addr;
170 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
171 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
174 class = LOCK_CLASS(lock);
175 db_printf(" class: %s\n", class->lc_name);
176 db_printf(" name: %s\n", lock->lo_name);
177 class->lc_ddb_show(lock);
181 #ifdef LOCK_PROFILING
184 * One object per-thread for each lock the thread owns. Tracks individual
187 struct lock_profile_object {
188 LIST_ENTRY(lock_profile_object) lpo_link;
189 struct lock_object *lpo_obj;
190 const char *lpo_file;
194 uint64_t lpo_acqtime;
195 uint64_t lpo_waittime;
196 u_int lpo_contest_locking;
200 * One lock_prof for each (file, line, lock object) triple.
203 SLIST_ENTRY(lock_prof) link;
204 struct lock_class *class;
209 uintmax_t cnt_wait_max;
214 uintmax_t cnt_contest_locking;
217 SLIST_HEAD(lphead, lock_prof);
219 #define LPROF_HASH_SIZE 4096
220 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
221 #define LPROF_CACHE_SIZE 4096
224 * Array of objects and profs for each type of object for each cpu. Spinlocks
225 * are handled separately because a thread may be preempted and acquire a
226 * spinlock while in the lock profiling code of a non-spinlock. In this way
227 * we only need a critical section to protect the per-cpu lists.
229 struct lock_prof_type {
230 struct lphead lpt_lpalloc;
231 struct lpohead lpt_lpoalloc;
232 struct lphead lpt_hash[LPROF_HASH_SIZE];
233 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
234 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
237 struct lock_prof_cpu {
238 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
241 struct lock_prof_cpu *lp_cpu[MAXCPU];
243 volatile int __read_mostly lock_prof_enable;
244 static volatile int lock_prof_resetting;
246 #define LPROF_SBUF_SIZE 256
248 static int lock_prof_rejected;
249 static int lock_prof_skipspin;
250 static int lock_prof_skipcount;
252 #ifndef USE_CPU_NANOSECONDS
260 /* From bintime2timespec */
261 ns = bt.sec * (uint64_t)1000000000;
262 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
268 lock_prof_init_type(struct lock_prof_type *type)
272 SLIST_INIT(&type->lpt_lpalloc);
273 LIST_INIT(&type->lpt_lpoalloc);
274 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
275 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
277 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
283 lock_prof_init(void *arg)
287 for (cpu = 0; cpu <= mp_maxid; cpu++) {
288 lp_cpu[cpu] = malloc(sizeof(*lp_cpu[cpu]), M_DEVBUF,
290 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[0]);
291 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[1]);
294 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
297 lock_prof_reset_wait(void)
301 * Spin relinquishing our cpu so that quiesce_all_cpus may
304 while (lock_prof_resetting)
305 sched_relinquish(curthread);
309 lock_prof_reset(void)
311 struct lock_prof_cpu *lpc;
315 * We not only race with acquiring and releasing locks but also
316 * thread exit. To be certain that threads exit without valid head
317 * pointers they must see resetting set before enabled is cleared.
318 * Otherwise a lock may not be removed from a per-thread list due
319 * to disabled being set but not wait for reset() to remove it below.
321 atomic_store_rel_int(&lock_prof_resetting, 1);
322 enabled = lock_prof_enable;
323 lock_prof_enable = 0;
324 quiesce_all_cpus("profreset", 0);
326 * Some objects may have migrated between CPUs. Clear all links
327 * before we zero the structures. Some items may still be linked
328 * into per-thread lists as well.
330 for (cpu = 0; cpu <= mp_maxid; cpu++) {
332 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
333 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
334 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
337 for (cpu = 0; cpu <= mp_maxid; cpu++) {
339 bzero(lpc, sizeof(*lpc));
340 lock_prof_init_type(&lpc->lpc_types[0]);
341 lock_prof_init_type(&lpc->lpc_types[1]);
343 atomic_store_rel_int(&lock_prof_resetting, 0);
344 lock_prof_enable = enabled;
348 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
352 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
354 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
355 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
356 lp->cnt_wait / 1000, lp->cnt_cur,
357 lp->cnt_cur == 0 ? (uintmax_t)0 :
358 lp->cnt_tot / (lp->cnt_cur * 1000),
359 lp->cnt_cur == 0 ? (uintmax_t)0 :
360 lp->cnt_wait / (lp->cnt_cur * 1000),
361 (uintmax_t)0, lp->cnt_contest_locking,
362 p, lp->line, lp->class->lc_name, lp->name);
366 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
369 struct lock_prof_type *type;
373 dst->file = match->file;
374 dst->line = match->line;
375 dst->class = match->class;
376 dst->name = match->name;
378 for (cpu = 0; cpu <= mp_maxid; cpu++) {
379 if (lp_cpu[cpu] == NULL)
381 type = &lp_cpu[cpu]->lpc_types[spin];
382 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
385 if (l->file != match->file || l->line != match->line ||
386 l->name != match->name)
389 if (l->cnt_max > dst->cnt_max)
390 dst->cnt_max = l->cnt_max;
391 if (l->cnt_wait_max > dst->cnt_wait_max)
392 dst->cnt_wait_max = l->cnt_wait_max;
393 dst->cnt_tot += l->cnt_tot;
394 dst->cnt_wait += l->cnt_wait;
395 dst->cnt_cur += l->cnt_cur;
396 dst->cnt_contest_locking += l->cnt_contest_locking;
403 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
409 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
410 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
411 struct lock_prof lp = {};
415 lock_prof_sum(l, &lp, i, spin, t);
416 lock_prof_output(&lp, sb);
422 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
428 error = sysctl_wire_old_buffer(req, 0);
431 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
432 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
433 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
434 enabled = lock_prof_enable;
435 lock_prof_enable = 0;
436 quiesce_all_cpus("profstat", 0);
438 for (cpu = 0; cpu <= mp_maxid; cpu++) {
439 if (lp_cpu[cpu] == NULL)
441 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[0], sb, 0, t);
442 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[1], sb, 1, t);
444 lock_prof_enable = enabled;
446 error = sbuf_finish(sb);
447 /* Output a trailing NUL. */
449 error = SYSCTL_OUT(req, "", 1);
455 enable_lock_prof(SYSCTL_HANDLER_ARGS)
459 v = lock_prof_enable;
460 error = sysctl_handle_int(oidp, &v, v, req);
463 if (req->newptr == NULL)
465 if (v == lock_prof_enable)
469 lock_prof_enable = !!v;
475 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
480 error = sysctl_handle_int(oidp, &v, 0, req);
483 if (req->newptr == NULL)
492 static struct lock_prof *
493 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
496 const char *unknown = "(unknown)";
497 struct lock_prof_type *type;
498 struct lock_prof *lp;
504 if (p == NULL || *p == '\0')
506 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
507 hash &= LPROF_HASH_MASK;
508 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
509 head = &type->lpt_hash[hash];
510 SLIST_FOREACH(lp, head, link) {
511 if (lp->line == line && lp->file == p &&
512 lp->name == lo->lo_name)
516 lp = SLIST_FIRST(&type->lpt_lpalloc);
518 lock_prof_rejected++;
521 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
524 lp->class = LOCK_CLASS(lo);
525 lp->name = lo->lo_name;
526 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
530 static struct lock_profile_object *
531 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
534 struct lock_profile_object *l;
535 struct lock_prof_type *type;
536 struct lpohead *head;
538 head = &curthread->td_lprof[spin];
539 LIST_FOREACH(l, head, lpo_link)
540 if (l->lpo_obj == lo && l->lpo_file == file &&
543 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
544 l = LIST_FIRST(&type->lpt_lpoalloc);
546 lock_prof_rejected++;
549 LIST_REMOVE(l, lpo_link);
554 LIST_INSERT_HEAD(head, l, lpo_link);
560 lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
561 uint64_t waittime, const char *file, int line)
563 static int lock_prof_count;
564 struct lock_profile_object *l;
567 if (SCHEDULER_STOPPED())
570 /* don't reset the timer when/if recursing */
571 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
573 if (lock_prof_skipcount &&
574 (++lock_prof_count % lock_prof_skipcount) != 0)
576 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
577 if (spin && lock_prof_skipspin == 1)
580 /* Recheck enabled now that we're in a critical section. */
581 if (lock_prof_enable == 0)
583 l = lock_profile_object_lookup(lo, spin, file, line);
587 if (++l->lpo_ref > 1)
589 l->lpo_contest_locking = contested;
590 l->lpo_acqtime = nanoseconds();
591 if (waittime && (l->lpo_acqtime > waittime))
592 l->lpo_waittime = l->lpo_acqtime - waittime;
600 lock_profile_thread_exit(struct thread *td)
603 struct lock_profile_object *l;
605 MPASS(curthread->td_critnest == 0);
608 * If lock profiling was disabled we have to wait for reset to
609 * clear our pointers before we can exit safely.
611 lock_prof_reset_wait();
613 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
614 printf("thread still holds lock acquired at %s:%d\n",
615 l->lpo_file, l->lpo_line);
616 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
617 printf("thread still holds lock acquired at %s:%d\n",
618 l->lpo_file, l->lpo_line);
620 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
621 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
625 lock_profile_release_lock(struct lock_object *lo)
627 struct lock_profile_object *l;
628 struct lock_prof_type *type;
629 struct lock_prof *lp;
630 uint64_t curtime, holdtime;
631 struct lpohead *head;
634 if (SCHEDULER_STOPPED())
636 if (lo->lo_flags & LO_NOPROFILE)
638 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
639 head = &curthread->td_lprof[spin];
640 if (LIST_FIRST(head) == NULL)
643 /* Recheck enabled now that we're in a critical section. */
644 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
647 * If lock profiling is not enabled we still want to remove the
648 * lpo from our queue.
650 LIST_FOREACH(l, head, lpo_link)
651 if (l->lpo_obj == lo)
655 if (--l->lpo_ref > 0)
657 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
660 curtime = nanoseconds();
661 if (curtime < l->lpo_acqtime)
663 holdtime = curtime - l->lpo_acqtime;
666 * Record if the lock has been held longer now than ever
669 if (holdtime > lp->cnt_max)
670 lp->cnt_max = holdtime;
671 if (l->lpo_waittime > lp->cnt_wait_max)
672 lp->cnt_wait_max = l->lpo_waittime;
673 lp->cnt_tot += holdtime;
674 lp->cnt_wait += l->lpo_waittime;
675 lp->cnt_contest_locking += l->lpo_contest_locking;
676 lp->cnt_cur += l->lpo_cnt;
678 LIST_REMOVE(l, lpo_link);
679 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
680 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
685 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL,
687 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
688 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
689 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
690 &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
691 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
692 &lock_prof_rejected, 0, "Number of rejected profiling records");
693 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
694 NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
695 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
696 NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
697 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
698 NULL, 0, enable_lock_prof, "I", "Enable lock profiling");