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>
44 #include <sys/lock_profile.h>
45 #include <sys/malloc.h>
46 #include <sys/mutex.h>
50 #include <sys/sched.h>
52 #include <sys/sysctl.h>
58 #include <machine/cpufunc.h>
61 * Uncomment to validate that spin argument to acquire/release routines matches
62 * the flag in the lock
64 //#define LOCK_PROFILING_DEBUG_SPIN
66 SDT_PROVIDER_DEFINE(lock);
67 SDT_PROBE_DEFINE1(lock, , , starvation, "u_int");
69 CTASSERT(LOCK_CLASS_MAX == 15);
71 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
73 &lock_class_mtx_sleep,
76 &lock_class_rm_sleepable,
82 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
83 const char *type, int flags)
87 /* Check for double-init and zero object. */
88 KASSERT(flags & LO_NEW || !lock_initialized(lock),
89 ("lock \"%s\" %p already initialized", name, lock));
91 /* Look up lock class to find its index. */
92 for (i = 0; i < LOCK_CLASS_MAX; i++)
93 if (lock_classes[i] == class) {
94 lock->lo_flags = i << LO_CLASSSHIFT;
97 KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
99 /* Initialize the lock object. */
100 lock->lo_name = name;
101 lock->lo_flags |= flags | LO_INITIALIZED;
102 LOCK_LOG_INIT(lock, 0);
103 WITNESS_INIT(lock, (type != NULL) ? type : name);
107 lock_destroy(struct lock_object *lock)
110 KASSERT(lock_initialized(lock), ("lock %p is not initialized", lock));
111 WITNESS_DESTROY(lock);
112 LOCK_LOG_DESTROY(lock, 0);
113 lock->lo_flags &= ~LO_INITIALIZED;
116 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
118 static SYSCTL_NODE(_debug_lock, OID_AUTO, delay,
119 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
122 static u_int __read_mostly starvation_limit = 131072;
123 SYSCTL_INT(_debug_lock_delay, OID_AUTO, starvation_limit, CTLFLAG_RW,
124 &starvation_limit, 0, "");
126 static u_int __read_mostly restrict_starvation = 0;
127 SYSCTL_INT(_debug_lock_delay, OID_AUTO, restrict_starvation, CTLFLAG_RW,
128 &restrict_starvation, 0, "");
131 lock_delay(struct lock_delay_arg *la)
133 struct lock_delay_config *lc = la->config;
137 if (__predict_false(la->delay > lc->max))
140 for (i = la->delay; i > 0; i--)
143 la->spin_cnt += la->delay;
144 if (__predict_false(la->spin_cnt > starvation_limit)) {
145 SDT_PROBE1(lock, , , starvation, la->delay);
146 if (restrict_starvation)
147 la->delay = lc->base;
152 lock_roundup_2(u_int val)
156 for (res = 1; res <= val; res <<= 1)
163 lock_delay_default_init(struct lock_delay_config *lc)
167 lc->max = lock_roundup_2(mp_ncpus) * 256;
172 struct lock_delay_config __read_frequently locks_delay;
173 u_short __read_frequently locks_delay_retries;
174 u_short __read_frequently locks_delay_loops;
176 SYSCTL_U16(_debug_lock, OID_AUTO, delay_base, CTLFLAG_RW, &locks_delay.base,
178 SYSCTL_U16(_debug_lock, OID_AUTO, delay_max, CTLFLAG_RW, &locks_delay.max,
180 SYSCTL_U16(_debug_lock, OID_AUTO, delay_retries, CTLFLAG_RW, &locks_delay_retries,
182 SYSCTL_U16(_debug_lock, OID_AUTO, delay_loops, CTLFLAG_RW, &locks_delay_loops,
186 locks_delay_init(void *arg __unused)
189 lock_delay_default_init(&locks_delay);
190 locks_delay_retries = 10;
191 locks_delay_loops = max(10000, locks_delay.max);
193 LOCK_DELAY_SYSINIT(locks_delay_init);
196 DB_SHOW_COMMAND(lock, db_show_lock)
198 struct lock_object *lock;
199 struct lock_class *class;
203 lock = (struct lock_object *)addr;
204 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
205 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
208 class = LOCK_CLASS(lock);
209 db_printf(" class: %s\n", class->lc_name);
210 db_printf(" name: %s\n", lock->lo_name);
211 class->lc_ddb_show(lock);
215 #ifdef LOCK_PROFILING
218 * One object per-thread for each lock the thread owns. Tracks individual
221 struct lock_profile_object {
222 LIST_ENTRY(lock_profile_object) lpo_link;
223 struct lock_object *lpo_obj;
224 const char *lpo_file;
228 uint64_t lpo_acqtime;
229 uint64_t lpo_waittime;
230 u_int lpo_contest_locking;
234 * One lock_prof for each (file, line, lock object) triple.
237 SLIST_ENTRY(lock_prof) link;
238 struct lock_class *class;
243 uintmax_t cnt_wait_max;
248 uintmax_t cnt_contest_locking;
251 SLIST_HEAD(lphead, lock_prof);
253 #define LPROF_HASH_SIZE 4096
254 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
255 #define LPROF_CACHE_SIZE 4096
258 * Array of objects and profs for each type of object for each cpu. Spinlocks
259 * are handled separately because a thread may be preempted and acquire a
260 * spinlock while in the lock profiling code of a non-spinlock. In this way
261 * we only need a critical section to protect the per-cpu lists.
263 struct lock_prof_type {
264 struct lphead lpt_lpalloc;
265 struct lpohead lpt_lpoalloc;
266 struct lphead lpt_hash[LPROF_HASH_SIZE];
267 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
268 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
271 struct lock_prof_cpu {
272 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
275 DPCPU_DEFINE_STATIC(struct lock_prof_cpu, lp);
276 #define LP_CPU_SELF (DPCPU_PTR(lp))
277 #define LP_CPU(cpu) (DPCPU_ID_PTR((cpu), lp))
279 volatile int __read_mostly lock_prof_enable;
280 int __read_mostly lock_contested_only;
281 static volatile int lock_prof_resetting;
283 #define LPROF_SBUF_SIZE 256
285 static int lock_prof_rejected;
286 static int lock_prof_skipspin;
288 #ifndef USE_CPU_NANOSECONDS
296 /* From bintime2timespec */
297 ns = bt.sec * (uint64_t)1000000000;
298 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
304 lock_prof_init_type(struct lock_prof_type *type)
308 SLIST_INIT(&type->lpt_lpalloc);
309 LIST_INIT(&type->lpt_lpoalloc);
310 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
311 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
313 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
319 lock_prof_init(void *arg)
324 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[0]);
325 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[1]);
328 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
331 lock_prof_reset_wait(void)
335 * Spin relinquishing our cpu so that quiesce_all_cpus may
338 while (lock_prof_resetting)
339 sched_relinquish(curthread);
343 lock_prof_reset(void)
345 struct lock_prof_cpu *lpc;
349 * We not only race with acquiring and releasing locks but also
350 * thread exit. To be certain that threads exit without valid head
351 * pointers they must see resetting set before enabled is cleared.
352 * Otherwise a lock may not be removed from a per-thread list due
353 * to disabled being set but not wait for reset() to remove it below.
355 atomic_store_rel_int(&lock_prof_resetting, 1);
356 enabled = lock_prof_enable;
357 lock_prof_enable = 0;
359 * This both publishes lock_prof_enable as disabled and makes sure
360 * everyone else reads it if they are not far enough. We wait for the
363 cpus_fence_seq_cst();
364 quiesce_all_critical();
366 * Some objects may have migrated between CPUs. Clear all links
367 * before we zero the structures. Some items may still be linked
368 * into per-thread lists as well.
372 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
373 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
374 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
379 bzero(lpc, sizeof(*lpc));
380 lock_prof_init_type(&lpc->lpc_types[0]);
381 lock_prof_init_type(&lpc->lpc_types[1]);
384 * Paired with the fence from cpus_fence_seq_cst()
386 atomic_store_rel_int(&lock_prof_resetting, 0);
387 lock_prof_enable = enabled;
391 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
395 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
397 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
398 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
399 lp->cnt_wait / 1000, lp->cnt_cur,
400 lp->cnt_cur == 0 ? (uintmax_t)0 :
401 lp->cnt_tot / (lp->cnt_cur * 1000),
402 lp->cnt_cur == 0 ? (uintmax_t)0 :
403 lp->cnt_wait / (lp->cnt_cur * 1000),
404 (uintmax_t)0, lp->cnt_contest_locking,
405 p, lp->line, lp->class->lc_name, lp->name);
409 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
412 struct lock_prof_type *type;
416 dst->file = match->file;
417 dst->line = match->line;
418 dst->class = match->class;
419 dst->name = match->name;
422 type = &LP_CPU(cpu)->lpc_types[spin];
423 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
426 if (l->file != match->file || l->line != match->line ||
427 l->name != match->name)
430 if (l->cnt_max > dst->cnt_max)
431 dst->cnt_max = l->cnt_max;
432 if (l->cnt_wait_max > dst->cnt_wait_max)
433 dst->cnt_wait_max = l->cnt_wait_max;
434 dst->cnt_tot += l->cnt_tot;
435 dst->cnt_wait += l->cnt_wait;
436 dst->cnt_cur += l->cnt_cur;
437 dst->cnt_contest_locking += l->cnt_contest_locking;
443 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
449 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
450 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
451 struct lock_prof lp = {};
455 lock_prof_sum(l, &lp, i, spin, t);
456 lock_prof_output(&lp, sb);
462 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
468 error = sysctl_wire_old_buffer(req, 0);
471 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
472 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
473 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
474 enabled = lock_prof_enable;
475 lock_prof_enable = 0;
477 * See the comment in lock_prof_reset
479 cpus_fence_seq_cst();
480 quiesce_all_critical();
483 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[0], sb, 0, t);
484 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[1], sb, 1, t);
486 atomic_thread_fence_rel();
487 lock_prof_enable = enabled;
489 error = sbuf_finish(sb);
490 /* Output a trailing NUL. */
492 error = SYSCTL_OUT(req, "", 1);
498 enable_lock_prof(SYSCTL_HANDLER_ARGS)
502 v = lock_prof_enable;
503 error = sysctl_handle_int(oidp, &v, v, req);
506 if (req->newptr == NULL)
508 if (v == lock_prof_enable)
512 lock_prof_enable = !!v;
518 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
523 error = sysctl_handle_int(oidp, &v, 0, req);
526 if (req->newptr == NULL)
535 static struct lock_prof *
536 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
539 const char *unknown = "(unknown)";
540 struct lock_prof_type *type;
541 struct lock_prof *lp;
547 if (p == NULL || *p == '\0')
549 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
550 hash &= LPROF_HASH_MASK;
551 type = &LP_CPU_SELF->lpc_types[spin];
552 head = &type->lpt_hash[hash];
553 SLIST_FOREACH(lp, head, link) {
554 if (lp->line == line && lp->file == p &&
555 lp->name == lo->lo_name)
558 lp = SLIST_FIRST(&type->lpt_lpalloc);
560 lock_prof_rejected++;
563 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
566 lp->class = LOCK_CLASS(lo);
567 lp->name = lo->lo_name;
568 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
572 static struct lock_profile_object *
573 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
576 struct lock_profile_object *l;
577 struct lock_prof_type *type;
578 struct lpohead *head;
580 head = &curthread->td_lprof[spin];
581 LIST_FOREACH(l, head, lpo_link)
582 if (l->lpo_obj == lo && l->lpo_file == file &&
585 type = &LP_CPU_SELF->lpc_types[spin];
586 l = LIST_FIRST(&type->lpt_lpoalloc);
588 lock_prof_rejected++;
591 LIST_REMOVE(l, lpo_link);
596 LIST_INSERT_HEAD(head, l, lpo_link);
602 lock_profile_obtain_lock_success(struct lock_object *lo, bool spin,
603 int contested, uint64_t waittime, const char *file, int line)
605 struct lock_profile_object *l;
607 #ifdef LOCK_PROFILING_DEBUG_SPIN
608 bool is_spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK);
609 if ((spin && !is_spin) || (!spin && is_spin))
610 printf("%s: lock %s spin mismatch (arg %d, flag %d)\n", __func__,
611 lo->lo_name, spin, is_spin);
614 /* don't reset the timer when/if recursing */
615 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
617 if (lock_contested_only && !contested)
619 if (spin && lock_prof_skipspin == 1)
622 if (SCHEDULER_STOPPED())
626 /* Recheck enabled now that we're in a critical section. */
627 if (lock_prof_enable == 0)
629 l = lock_profile_object_lookup(lo, spin, file, line);
633 if (++l->lpo_ref > 1)
635 l->lpo_contest_locking = contested;
636 l->lpo_acqtime = nanoseconds();
637 if (waittime && (l->lpo_acqtime > waittime))
638 l->lpo_waittime = l->lpo_acqtime - waittime;
643 * Paired with cpus_fence_seq_cst().
645 atomic_thread_fence_rel();
650 lock_profile_thread_exit(struct thread *td)
653 struct lock_profile_object *l;
655 MPASS(curthread->td_critnest == 0);
658 * If lock profiling was disabled we have to wait for reset to
659 * clear our pointers before we can exit safely.
661 lock_prof_reset_wait();
663 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
664 printf("thread still holds lock acquired at %s:%d\n",
665 l->lpo_file, l->lpo_line);
666 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
667 printf("thread still holds lock acquired at %s:%d\n",
668 l->lpo_file, l->lpo_line);
670 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
671 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
675 lock_profile_release_lock(struct lock_object *lo, bool spin)
677 struct lock_profile_object *l;
678 struct lock_prof_type *type;
679 struct lock_prof *lp;
680 uint64_t curtime, holdtime;
681 struct lpohead *head;
683 #ifdef LOCK_PROFILING_DEBUG_SPIN
684 bool is_spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK);
685 if ((spin && !is_spin) || (!spin && is_spin))
686 printf("%s: lock %s spin mismatch (arg %d, flag %d)\n", __func__,
687 lo->lo_name, spin, is_spin);
690 if (lo->lo_flags & LO_NOPROFILE)
692 head = &curthread->td_lprof[spin];
693 if (LIST_FIRST(head) == NULL)
695 if (SCHEDULER_STOPPED())
698 /* Recheck enabled now that we're in a critical section. */
699 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
702 * If lock profiling is not enabled we still want to remove the
703 * lpo from our queue.
705 LIST_FOREACH(l, head, lpo_link)
706 if (l->lpo_obj == lo)
710 if (--l->lpo_ref > 0)
712 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
715 curtime = nanoseconds();
716 if (curtime < l->lpo_acqtime)
718 holdtime = curtime - l->lpo_acqtime;
721 * Record if the lock has been held longer now than ever
724 if (holdtime > lp->cnt_max)
725 lp->cnt_max = holdtime;
726 if (l->lpo_waittime > lp->cnt_wait_max)
727 lp->cnt_wait_max = l->lpo_waittime;
728 lp->cnt_tot += holdtime;
729 lp->cnt_wait += l->lpo_waittime;
730 lp->cnt_contest_locking += l->lpo_contest_locking;
731 lp->cnt_cur += l->lpo_cnt;
733 LIST_REMOVE(l, lpo_link);
734 type = &LP_CPU_SELF->lpc_types[spin];
735 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
738 * Paired with cpus_fence_seq_cst().
740 atomic_thread_fence_rel();
744 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof,
745 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
747 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
748 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
749 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
750 &lock_prof_rejected, 0, "Number of rejected profiling records");
751 SYSCTL_INT(_debug_lock_prof, OID_AUTO, contested_only, CTLFLAG_RW,
752 &lock_contested_only, 0, "Only profile contested acquires");
753 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats,
754 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
755 dump_lock_prof_stats, "A",
756 "Lock profiling statistics");
757 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset,
758 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0,
759 reset_lock_prof_stats, "I",
760 "Reset lock profiling statistics");
761 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable,
762 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
763 enable_lock_prof, "I",
764 "Enable lock profiling");