2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * This module holds the global variables and functions used to maintain
30 * lock_object structures.
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
37 #include "opt_mprof.h"
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
43 #include <sys/limits.h>
45 #include <sys/lock_profile.h>
46 #include <sys/malloc.h>
47 #include <sys/mutex.h>
51 #include <sys/sched.h>
53 #include <sys/sysctl.h>
59 #include <machine/cpufunc.h>
62 * Uncomment to validate that spin argument to acquire/release routines matches
63 * the flag in the lock
65 //#define LOCK_PROFILING_DEBUG_SPIN
67 SDT_PROVIDER_DEFINE(lock);
68 SDT_PROBE_DEFINE1(lock, , , starvation, "u_int");
70 CTASSERT(LOCK_CLASS_MAX == 15);
72 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
74 &lock_class_mtx_sleep,
77 &lock_class_rm_sleepable,
83 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
84 const char *type, int flags)
88 /* Check for double-init and zero object. */
89 KASSERT(flags & LO_NEW || !lock_initialized(lock),
90 ("lock \"%s\" %p already initialized", name, lock));
92 /* Look up lock class to find its index. */
93 for (i = 0; i < LOCK_CLASS_MAX; i++)
94 if (lock_classes[i] == class) {
95 lock->lo_flags = i << LO_CLASSSHIFT;
98 KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
100 /* Initialize the lock object. */
101 lock->lo_name = name;
102 lock->lo_flags |= flags | LO_INITIALIZED;
103 LOCK_LOG_INIT(lock, 0);
104 WITNESS_INIT(lock, (type != NULL) ? type : name);
108 lock_destroy(struct lock_object *lock)
111 KASSERT(lock_initialized(lock), ("lock %p is not initialized", lock));
112 WITNESS_DESTROY(lock);
113 LOCK_LOG_DESTROY(lock, 0);
114 lock->lo_flags &= ~LO_INITIALIZED;
117 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
119 static SYSCTL_NODE(_debug_lock, OID_AUTO, delay,
120 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
123 static u_int __read_mostly starvation_limit = 131072;
124 SYSCTL_INT(_debug_lock_delay, OID_AUTO, starvation_limit, CTLFLAG_RW,
125 &starvation_limit, 0, "");
127 static u_int __read_mostly restrict_starvation = 0;
128 SYSCTL_INT(_debug_lock_delay, OID_AUTO, restrict_starvation, CTLFLAG_RW,
129 &restrict_starvation, 0, "");
132 lock_delay(struct lock_delay_arg *la)
134 struct lock_delay_config *lc = la->config;
137 for (i = la->delay; i > 0; i--)
139 la->spin_cnt += la->delay;
142 if (__predict_false(la->delay > lc->max))
145 if (__predict_false(la->spin_cnt > starvation_limit)) {
146 SDT_PROBE1(lock, , , starvation, la->delay);
147 if (restrict_starvation)
148 la->delay = lc->base;
153 lock_roundup_2(u_int val)
157 for (res = 1; res <= val; res <<= 1)
164 lock_delay_default_init(struct lock_delay_config *lc)
168 lc->max = lock_roundup_2(mp_ncpus) * 256;
173 struct lock_delay_config __read_frequently locks_delay;
174 u_short __read_frequently locks_delay_retries;
175 u_short __read_frequently locks_delay_loops;
177 SYSCTL_U16(_debug_lock, OID_AUTO, delay_base, CTLFLAG_RW, &locks_delay.base,
179 SYSCTL_U16(_debug_lock, OID_AUTO, delay_max, CTLFLAG_RW, &locks_delay.max,
181 SYSCTL_U16(_debug_lock, OID_AUTO, delay_retries, CTLFLAG_RW, &locks_delay_retries,
183 SYSCTL_U16(_debug_lock, OID_AUTO, delay_loops, CTLFLAG_RW, &locks_delay_loops,
187 locks_delay_init(void *arg __unused)
190 lock_delay_default_init(&locks_delay);
191 locks_delay_retries = 10;
192 locks_delay_loops = max(10000, locks_delay.max);
194 LOCK_DELAY_SYSINIT(locks_delay_init);
197 DB_SHOW_COMMAND(lock, db_show_lock)
199 struct lock_object *lock;
200 struct lock_class *class;
204 lock = (struct lock_object *)addr;
205 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
206 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
209 class = LOCK_CLASS(lock);
210 db_printf(" class: %s\n", class->lc_name);
211 db_printf(" name: %s\n", lock->lo_name);
212 class->lc_ddb_show(lock);
216 #ifdef LOCK_PROFILING
219 * One object per-thread for each lock the thread owns. Tracks individual
222 struct lock_profile_object {
223 LIST_ENTRY(lock_profile_object) lpo_link;
224 struct lock_object *lpo_obj;
225 const char *lpo_file;
229 uint64_t lpo_acqtime;
230 uint64_t lpo_waittime;
231 u_int lpo_contest_locking;
235 * One lock_prof for each (file, line, lock object) triple.
238 SLIST_ENTRY(lock_prof) link;
239 struct lock_class *class;
244 uintmax_t cnt_wait_max;
249 uintmax_t cnt_contest_locking;
252 SLIST_HEAD(lphead, lock_prof);
254 #define LPROF_HASH_SIZE 4096
255 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
256 #define LPROF_CACHE_SIZE 4096
259 * Array of objects and profs for each type of object for each cpu. Spinlocks
260 * are handled separately because a thread may be preempted and acquire a
261 * spinlock while in the lock profiling code of a non-spinlock. In this way
262 * we only need a critical section to protect the per-cpu lists.
264 struct lock_prof_type {
265 struct lphead lpt_lpalloc;
266 struct lpohead lpt_lpoalloc;
267 struct lphead lpt_hash[LPROF_HASH_SIZE];
268 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
269 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
272 struct lock_prof_cpu {
273 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
276 DPCPU_DEFINE_STATIC(struct lock_prof_cpu, lp);
277 #define LP_CPU_SELF (DPCPU_PTR(lp))
278 #define LP_CPU(cpu) (DPCPU_ID_PTR((cpu), lp))
280 volatile int __read_mostly lock_prof_enable;
281 int __read_mostly lock_contested_only;
282 static volatile int lock_prof_resetting;
284 #define LPROF_SBUF_SIZE 256
286 static int lock_prof_rejected;
287 static int lock_prof_skipspin;
289 #ifndef USE_CPU_NANOSECONDS
297 /* From bintime2timespec */
298 ns = bt.sec * (uint64_t)1000000000;
299 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
305 lock_prof_init_type(struct lock_prof_type *type)
309 SLIST_INIT(&type->lpt_lpalloc);
310 LIST_INIT(&type->lpt_lpoalloc);
311 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
312 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
314 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
320 lock_prof_init(void *arg)
325 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[0]);
326 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[1]);
329 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
332 lock_prof_reset_wait(void)
336 * Spin relinquishing our cpu so that quiesce_all_cpus may
339 while (lock_prof_resetting)
340 sched_relinquish(curthread);
344 lock_prof_reset(void)
346 struct lock_prof_cpu *lpc;
350 * We not only race with acquiring and releasing locks but also
351 * thread exit. To be certain that threads exit without valid head
352 * pointers they must see resetting set before enabled is cleared.
353 * Otherwise a lock may not be removed from a per-thread list due
354 * to disabled being set but not wait for reset() to remove it below.
356 atomic_store_rel_int(&lock_prof_resetting, 1);
357 enabled = lock_prof_enable;
358 lock_prof_enable = 0;
360 * This both publishes lock_prof_enable as disabled and makes sure
361 * everyone else reads it if they are not far enough. We wait for the
364 cpus_fence_seq_cst();
365 quiesce_all_critical();
367 * Some objects may have migrated between CPUs. Clear all links
368 * before we zero the structures. Some items may still be linked
369 * into per-thread lists as well.
373 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
374 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
375 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
380 bzero(lpc, sizeof(*lpc));
381 lock_prof_init_type(&lpc->lpc_types[0]);
382 lock_prof_init_type(&lpc->lpc_types[1]);
385 * Paired with the fence from cpus_fence_seq_cst()
387 atomic_store_rel_int(&lock_prof_resetting, 0);
388 lock_prof_enable = enabled;
392 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
396 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
398 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
399 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
400 lp->cnt_wait / 1000, lp->cnt_cur,
401 lp->cnt_cur == 0 ? (uintmax_t)0 :
402 lp->cnt_tot / (lp->cnt_cur * 1000),
403 lp->cnt_cur == 0 ? (uintmax_t)0 :
404 lp->cnt_wait / (lp->cnt_cur * 1000),
405 (uintmax_t)0, lp->cnt_contest_locking,
406 p, lp->line, lp->class->lc_name, lp->name);
410 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
413 struct lock_prof_type *type;
417 dst->file = match->file;
418 dst->line = match->line;
419 dst->class = match->class;
420 dst->name = match->name;
423 type = &LP_CPU(cpu)->lpc_types[spin];
424 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
427 if (l->file != match->file || l->line != match->line ||
428 l->name != match->name)
431 if (l->cnt_max > dst->cnt_max)
432 dst->cnt_max = l->cnt_max;
433 if (l->cnt_wait_max > dst->cnt_wait_max)
434 dst->cnt_wait_max = l->cnt_wait_max;
435 dst->cnt_tot += l->cnt_tot;
436 dst->cnt_wait += l->cnt_wait;
437 dst->cnt_cur += l->cnt_cur;
438 dst->cnt_contest_locking += l->cnt_contest_locking;
444 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
450 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
451 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
452 struct lock_prof lp = {};
456 lock_prof_sum(l, &lp, i, spin, t);
457 lock_prof_output(&lp, sb);
463 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
469 error = sysctl_wire_old_buffer(req, 0);
472 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
473 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
474 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
475 enabled = lock_prof_enable;
476 lock_prof_enable = 0;
478 * See the comment in lock_prof_reset
480 cpus_fence_seq_cst();
481 quiesce_all_critical();
484 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[0], sb, 0, t);
485 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[1], sb, 1, t);
487 atomic_thread_fence_rel();
488 lock_prof_enable = enabled;
490 error = sbuf_finish(sb);
491 /* Output a trailing NUL. */
493 error = SYSCTL_OUT(req, "", 1);
499 enable_lock_prof(SYSCTL_HANDLER_ARGS)
503 v = lock_prof_enable;
504 error = sysctl_handle_int(oidp, &v, v, req);
507 if (req->newptr == NULL)
509 if (v == lock_prof_enable)
513 lock_prof_enable = !!v;
519 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
524 error = sysctl_handle_int(oidp, &v, 0, req);
527 if (req->newptr == NULL)
536 static struct lock_prof *
537 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
540 const char *unknown = "(unknown)";
541 struct lock_prof_type *type;
542 struct lock_prof *lp;
548 if (p == NULL || *p == '\0')
550 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
551 hash &= LPROF_HASH_MASK;
552 type = &LP_CPU_SELF->lpc_types[spin];
553 head = &type->lpt_hash[hash];
554 SLIST_FOREACH(lp, head, link) {
555 if (lp->line == line && lp->file == p &&
556 lp->name == lo->lo_name)
559 lp = SLIST_FIRST(&type->lpt_lpalloc);
561 lock_prof_rejected++;
564 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
567 lp->class = LOCK_CLASS(lo);
568 lp->name = lo->lo_name;
569 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
573 static struct lock_profile_object *
574 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
577 struct lock_profile_object *l;
578 struct lock_prof_type *type;
579 struct lpohead *head;
581 head = &curthread->td_lprof[spin];
582 LIST_FOREACH(l, head, lpo_link)
583 if (l->lpo_obj == lo && l->lpo_file == file &&
586 type = &LP_CPU_SELF->lpc_types[spin];
587 l = LIST_FIRST(&type->lpt_lpoalloc);
589 lock_prof_rejected++;
592 LIST_REMOVE(l, lpo_link);
597 LIST_INSERT_HEAD(head, l, lpo_link);
603 lock_profile_obtain_lock_success(struct lock_object *lo, bool spin,
604 int contested, uint64_t waittime, const char *file, int line)
606 struct lock_profile_object *l;
608 #ifdef LOCK_PROFILING_DEBUG_SPIN
609 bool is_spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK);
610 if ((spin && !is_spin) || (!spin && is_spin))
611 printf("%s: lock %s spin mismatch (arg %d, flag %d)\n", __func__,
612 lo->lo_name, spin, is_spin);
615 /* don't reset the timer when/if recursing */
616 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
618 if (lock_contested_only && !contested)
620 if (spin && lock_prof_skipspin == 1)
623 if (SCHEDULER_STOPPED())
627 /* Recheck enabled now that we're in a critical section. */
628 if (lock_prof_enable == 0)
630 l = lock_profile_object_lookup(lo, spin, file, line);
634 if (++l->lpo_ref > 1)
636 l->lpo_contest_locking = contested;
637 l->lpo_acqtime = nanoseconds();
638 if (waittime && (l->lpo_acqtime > waittime))
639 l->lpo_waittime = l->lpo_acqtime - waittime;
644 * Paired with cpus_fence_seq_cst().
646 atomic_thread_fence_rel();
651 lock_profile_thread_exit(struct thread *td)
654 struct lock_profile_object *l;
656 MPASS(curthread->td_critnest == 0);
659 * If lock profiling was disabled we have to wait for reset to
660 * clear our pointers before we can exit safely.
662 lock_prof_reset_wait();
664 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
665 printf("thread still holds lock acquired at %s:%d\n",
666 l->lpo_file, l->lpo_line);
667 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
668 printf("thread still holds lock acquired at %s:%d\n",
669 l->lpo_file, l->lpo_line);
671 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
672 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
676 lock_profile_release_lock(struct lock_object *lo, bool spin)
678 struct lock_profile_object *l;
679 struct lock_prof_type *type;
680 struct lock_prof *lp;
681 uint64_t curtime, holdtime;
682 struct lpohead *head;
684 #ifdef LOCK_PROFILING_DEBUG_SPIN
685 bool is_spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK);
686 if ((spin && !is_spin) || (!spin && is_spin))
687 printf("%s: lock %s spin mismatch (arg %d, flag %d)\n", __func__,
688 lo->lo_name, spin, is_spin);
691 if (lo->lo_flags & LO_NOPROFILE)
693 head = &curthread->td_lprof[spin];
694 if (LIST_FIRST(head) == NULL)
696 if (SCHEDULER_STOPPED())
699 /* Recheck enabled now that we're in a critical section. */
700 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
703 * If lock profiling is not enabled we still want to remove the
704 * lpo from our queue.
706 LIST_FOREACH(l, head, lpo_link)
707 if (l->lpo_obj == lo)
711 if (--l->lpo_ref > 0)
713 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
716 curtime = nanoseconds();
717 if (curtime < l->lpo_acqtime)
719 holdtime = curtime - l->lpo_acqtime;
722 * Record if the lock has been held longer now than ever
725 if (holdtime > lp->cnt_max)
726 lp->cnt_max = holdtime;
727 if (l->lpo_waittime > lp->cnt_wait_max)
728 lp->cnt_wait_max = l->lpo_waittime;
729 lp->cnt_tot += holdtime;
730 lp->cnt_wait += l->lpo_waittime;
731 lp->cnt_contest_locking += l->lpo_contest_locking;
732 lp->cnt_cur += l->lpo_cnt;
734 LIST_REMOVE(l, lpo_link);
735 type = &LP_CPU_SELF->lpc_types[spin];
736 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
739 * Paired with cpus_fence_seq_cst().
741 atomic_thread_fence_rel();
745 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof,
746 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
748 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
749 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
750 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
751 &lock_prof_rejected, 0, "Number of rejected profiling records");
752 SYSCTL_INT(_debug_lock_prof, OID_AUTO, contested_only, CTLFLAG_RW,
753 &lock_contested_only, 0, "Only profile contested acquires");
754 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats,
755 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
756 dump_lock_prof_stats, "A",
757 "Lock profiling statistics");
758 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset,
759 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0,
760 reset_lock_prof_stats, "I",
761 "Reset lock profiling statistics");
762 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable,
763 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
764 enable_lock_prof, "I",
765 "Enable lock profiling");