2 * SPDX-License-Identifier: BSD-2-Clause
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>
35 #include "opt_mprof.h"
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
42 #include <sys/lock_profile.h>
43 #include <sys/malloc.h>
44 #include <sys/mutex.h>
48 #include <sys/sched.h>
50 #include <sys/sysctl.h>
56 #include <machine/cpufunc.h>
59 * Uncomment to validate that spin argument to acquire/release routines matches
60 * the flag in the lock
62 //#define LOCK_PROFILING_DEBUG_SPIN
64 SDT_PROVIDER_DEFINE(lock);
65 SDT_PROBE_DEFINE1(lock, , , starvation, "u_int");
67 CTASSERT(LOCK_CLASS_MAX == 15);
69 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
71 &lock_class_mtx_sleep,
74 &lock_class_rm_sleepable,
80 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
81 const char *type, int flags)
85 /* Check for double-init and zero object. */
86 KASSERT(flags & LO_NEW || !lock_initialized(lock),
87 ("lock \"%s\" %p already initialized", name, lock));
89 /* Look up lock class to find its index. */
90 for (i = 0; i < LOCK_CLASS_MAX; i++)
91 if (lock_classes[i] == class) {
92 lock->lo_flags = i << LO_CLASSSHIFT;
95 KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
97 /* Initialize the lock object. */
99 lock->lo_flags |= flags | LO_INITIALIZED;
100 LOCK_LOG_INIT(lock, 0);
101 WITNESS_INIT(lock, (type != NULL) ? type : name);
105 lock_destroy(struct lock_object *lock)
108 KASSERT(lock_initialized(lock), ("lock %p is not initialized", lock));
109 WITNESS_DESTROY(lock);
110 LOCK_LOG_DESTROY(lock, 0);
111 lock->lo_flags &= ~LO_INITIALIZED;
114 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
116 static SYSCTL_NODE(_debug_lock, OID_AUTO, delay,
117 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
120 static u_int __read_mostly starvation_limit = 131072;
121 SYSCTL_INT(_debug_lock_delay, OID_AUTO, starvation_limit, CTLFLAG_RW,
122 &starvation_limit, 0, "");
124 static u_int __read_mostly restrict_starvation = 0;
125 SYSCTL_INT(_debug_lock_delay, OID_AUTO, restrict_starvation, CTLFLAG_RW,
126 &restrict_starvation, 0, "");
129 lock_delay(struct lock_delay_arg *la)
131 struct lock_delay_config *lc = la->config;
135 if (__predict_false(la->delay > lc->max))
138 for (i = la->delay; i > 0; i--)
141 la->spin_cnt += la->delay;
142 if (__predict_false(la->spin_cnt > starvation_limit)) {
143 SDT_PROBE1(lock, , , starvation, la->delay);
144 if (restrict_starvation)
145 la->delay = lc->base;
150 lock_roundup_2(u_int val)
154 for (res = 1; res <= val; res <<= 1)
161 lock_delay_default_init(struct lock_delay_config *lc)
165 lc->max = lock_roundup_2(mp_ncpus) * 256;
170 struct lock_delay_config __read_frequently locks_delay;
171 u_short __read_frequently locks_delay_retries;
172 u_short __read_frequently locks_delay_loops;
174 SYSCTL_U16(_debug_lock, OID_AUTO, delay_base, CTLFLAG_RW, &locks_delay.base,
176 SYSCTL_U16(_debug_lock, OID_AUTO, delay_max, CTLFLAG_RW, &locks_delay.max,
178 SYSCTL_U16(_debug_lock, OID_AUTO, delay_retries, CTLFLAG_RW, &locks_delay_retries,
180 SYSCTL_U16(_debug_lock, OID_AUTO, delay_loops, CTLFLAG_RW, &locks_delay_loops,
184 locks_delay_init(void *arg __unused)
187 lock_delay_default_init(&locks_delay);
188 locks_delay_retries = 10;
189 locks_delay_loops = max(10000, locks_delay.max);
191 LOCK_DELAY_SYSINIT(locks_delay_init);
194 DB_SHOW_COMMAND(lock, db_show_lock)
196 struct lock_object *lock;
197 struct lock_class *class;
201 lock = (struct lock_object *)addr;
202 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
203 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
206 class = LOCK_CLASS(lock);
207 db_printf(" class: %s\n", class->lc_name);
208 db_printf(" name: %s\n", lock->lo_name);
209 class->lc_ddb_show(lock);
213 #ifdef LOCK_PROFILING
216 * One object per-thread for each lock the thread owns. Tracks individual
219 struct lock_profile_object {
220 LIST_ENTRY(lock_profile_object) lpo_link;
221 struct lock_object *lpo_obj;
222 const char *lpo_file;
226 uint64_t lpo_acqtime;
227 uint64_t lpo_waittime;
228 u_int lpo_contest_locking;
232 * One lock_prof for each (file, line, lock object) triple.
235 SLIST_ENTRY(lock_prof) link;
236 struct lock_class *class;
241 uintmax_t cnt_wait_max;
246 uintmax_t cnt_contest_locking;
249 SLIST_HEAD(lphead, lock_prof);
251 #define LPROF_HASH_SIZE 4096
252 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
253 #define LPROF_CACHE_SIZE 4096
256 * Array of objects and profs for each type of object for each cpu. Spinlocks
257 * are handled separately because a thread may be preempted and acquire a
258 * spinlock while in the lock profiling code of a non-spinlock. In this way
259 * we only need a critical section to protect the per-cpu lists.
261 struct lock_prof_type {
262 struct lphead lpt_lpalloc;
263 struct lpohead lpt_lpoalloc;
264 struct lphead lpt_hash[LPROF_HASH_SIZE];
265 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
266 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
269 struct lock_prof_cpu {
270 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
273 DPCPU_DEFINE_STATIC(struct lock_prof_cpu, lp);
274 #define LP_CPU_SELF (DPCPU_PTR(lp))
275 #define LP_CPU(cpu) (DPCPU_ID_PTR((cpu), lp))
277 volatile int __read_mostly lock_prof_enable;
278 int __read_mostly lock_contested_only;
279 static volatile int lock_prof_resetting;
281 #define LPROF_SBUF_SIZE 256
283 static int lock_prof_rejected;
284 static int lock_prof_skipspin;
286 #ifndef USE_CPU_NANOSECONDS
294 /* From bintime2timespec */
295 ns = bt.sec * (uint64_t)1000000000;
296 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
302 lock_prof_init_type(struct lock_prof_type *type)
306 SLIST_INIT(&type->lpt_lpalloc);
307 LIST_INIT(&type->lpt_lpoalloc);
308 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
309 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
311 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
317 lock_prof_init(void *arg)
322 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[0]);
323 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[1]);
326 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
329 lock_prof_reset_wait(void)
333 * Spin relinquishing our cpu so that quiesce_all_cpus may
336 while (lock_prof_resetting)
337 sched_relinquish(curthread);
341 lock_prof_reset(void)
343 struct lock_prof_cpu *lpc;
347 * We not only race with acquiring and releasing locks but also
348 * thread exit. To be certain that threads exit without valid head
349 * pointers they must see resetting set before enabled is cleared.
350 * Otherwise a lock may not be removed from a per-thread list due
351 * to disabled being set but not wait for reset() to remove it below.
353 atomic_store_rel_int(&lock_prof_resetting, 1);
354 enabled = lock_prof_enable;
355 lock_prof_enable = 0;
357 * This both publishes lock_prof_enable as disabled and makes sure
358 * everyone else reads it if they are not far enough. We wait for the
361 cpus_fence_seq_cst();
362 quiesce_all_critical();
364 * Some objects may have migrated between CPUs. Clear all links
365 * before we zero the structures. Some items may still be linked
366 * into per-thread lists as well.
370 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
371 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
372 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
377 bzero(lpc, sizeof(*lpc));
378 lock_prof_init_type(&lpc->lpc_types[0]);
379 lock_prof_init_type(&lpc->lpc_types[1]);
382 * Paired with the fence from cpus_fence_seq_cst()
384 atomic_store_rel_int(&lock_prof_resetting, 0);
385 lock_prof_enable = enabled;
389 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
393 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
395 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
396 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
397 lp->cnt_wait / 1000, lp->cnt_cur,
398 lp->cnt_cur == 0 ? (uintmax_t)0 :
399 lp->cnt_tot / (lp->cnt_cur * 1000),
400 lp->cnt_cur == 0 ? (uintmax_t)0 :
401 lp->cnt_wait / (lp->cnt_cur * 1000),
402 (uintmax_t)0, lp->cnt_contest_locking,
403 p, lp->line, lp->class->lc_name, lp->name);
407 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
410 struct lock_prof_type *type;
414 dst->file = match->file;
415 dst->line = match->line;
416 dst->class = match->class;
417 dst->name = match->name;
420 type = &LP_CPU(cpu)->lpc_types[spin];
421 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
424 if (l->file != match->file || l->line != match->line ||
425 l->name != match->name)
428 if (l->cnt_max > dst->cnt_max)
429 dst->cnt_max = l->cnt_max;
430 if (l->cnt_wait_max > dst->cnt_wait_max)
431 dst->cnt_wait_max = l->cnt_wait_max;
432 dst->cnt_tot += l->cnt_tot;
433 dst->cnt_wait += l->cnt_wait;
434 dst->cnt_cur += l->cnt_cur;
435 dst->cnt_contest_locking += l->cnt_contest_locking;
441 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
447 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
448 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
449 struct lock_prof lp = {};
453 lock_prof_sum(l, &lp, i, spin, t);
454 lock_prof_output(&lp, sb);
460 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
466 error = sysctl_wire_old_buffer(req, 0);
469 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
470 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
471 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
472 enabled = lock_prof_enable;
473 lock_prof_enable = 0;
475 * See the comment in lock_prof_reset
477 cpus_fence_seq_cst();
478 quiesce_all_critical();
481 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[0], sb, 0, t);
482 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[1], sb, 1, t);
484 atomic_thread_fence_rel();
485 lock_prof_enable = enabled;
487 error = sbuf_finish(sb);
488 /* Output a trailing NUL. */
490 error = SYSCTL_OUT(req, "", 1);
496 enable_lock_prof(SYSCTL_HANDLER_ARGS)
500 v = lock_prof_enable;
501 error = sysctl_handle_int(oidp, &v, v, req);
504 if (req->newptr == NULL)
506 if (v == lock_prof_enable)
510 lock_prof_enable = !!v;
516 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
521 error = sysctl_handle_int(oidp, &v, 0, req);
524 if (req->newptr == NULL)
533 static struct lock_prof *
534 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
537 const char *unknown = "(unknown)";
538 struct lock_prof_type *type;
539 struct lock_prof *lp;
545 if (p == NULL || *p == '\0')
547 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
548 hash &= LPROF_HASH_MASK;
549 type = &LP_CPU_SELF->lpc_types[spin];
550 head = &type->lpt_hash[hash];
551 SLIST_FOREACH(lp, head, link) {
552 if (lp->line == line && lp->file == p &&
553 lp->name == lo->lo_name)
556 lp = SLIST_FIRST(&type->lpt_lpalloc);
558 lock_prof_rejected++;
561 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
564 lp->class = LOCK_CLASS(lo);
565 lp->name = lo->lo_name;
566 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
570 static struct lock_profile_object *
571 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
574 struct lock_profile_object *l;
575 struct lock_prof_type *type;
576 struct lpohead *head;
578 head = &curthread->td_lprof[spin];
579 LIST_FOREACH(l, head, lpo_link)
580 if (l->lpo_obj == lo && l->lpo_file == file &&
583 type = &LP_CPU_SELF->lpc_types[spin];
584 l = LIST_FIRST(&type->lpt_lpoalloc);
586 lock_prof_rejected++;
589 LIST_REMOVE(l, lpo_link);
594 LIST_INSERT_HEAD(head, l, lpo_link);
600 lock_profile_obtain_lock_success(struct lock_object *lo, bool spin,
601 int contested, uint64_t waittime, const char *file, int line)
603 struct lock_profile_object *l;
605 #ifdef LOCK_PROFILING_DEBUG_SPIN
606 bool is_spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK);
607 if ((spin && !is_spin) || (!spin && is_spin))
608 printf("%s: lock %s spin mismatch (arg %d, flag %d)\n", __func__,
609 lo->lo_name, spin, is_spin);
612 /* don't reset the timer when/if recursing */
613 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
615 if (lock_contested_only && !contested)
617 if (spin && lock_prof_skipspin == 1)
620 if (SCHEDULER_STOPPED())
624 /* Recheck enabled now that we're in a critical section. */
625 if (lock_prof_enable == 0)
627 l = lock_profile_object_lookup(lo, spin, file, line);
631 if (++l->lpo_ref > 1)
633 l->lpo_contest_locking = contested;
634 l->lpo_acqtime = nanoseconds();
635 if (waittime && (l->lpo_acqtime > waittime))
636 l->lpo_waittime = l->lpo_acqtime - waittime;
641 * Paired with cpus_fence_seq_cst().
643 atomic_thread_fence_rel();
648 lock_profile_thread_exit(struct thread *td)
651 struct lock_profile_object *l;
653 MPASS(curthread->td_critnest == 0);
656 * If lock profiling was disabled we have to wait for reset to
657 * clear our pointers before we can exit safely.
659 lock_prof_reset_wait();
661 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
662 printf("thread still holds lock acquired at %s:%d\n",
663 l->lpo_file, l->lpo_line);
664 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
665 printf("thread still holds lock acquired at %s:%d\n",
666 l->lpo_file, l->lpo_line);
668 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
669 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
673 lock_profile_release_lock(struct lock_object *lo, bool spin)
675 struct lock_profile_object *l;
676 struct lock_prof_type *type;
677 struct lock_prof *lp;
678 uint64_t curtime, holdtime;
679 struct lpohead *head;
681 #ifdef LOCK_PROFILING_DEBUG_SPIN
682 bool is_spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK);
683 if ((spin && !is_spin) || (!spin && is_spin))
684 printf("%s: lock %s spin mismatch (arg %d, flag %d)\n", __func__,
685 lo->lo_name, spin, is_spin);
688 if (lo->lo_flags & LO_NOPROFILE)
690 head = &curthread->td_lprof[spin];
691 if (LIST_FIRST(head) == NULL)
693 if (SCHEDULER_STOPPED())
696 /* Recheck enabled now that we're in a critical section. */
697 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
700 * If lock profiling is not enabled we still want to remove the
701 * lpo from our queue.
703 LIST_FOREACH(l, head, lpo_link)
704 if (l->lpo_obj == lo)
708 if (--l->lpo_ref > 0)
710 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
713 curtime = nanoseconds();
714 if (curtime < l->lpo_acqtime)
716 holdtime = curtime - l->lpo_acqtime;
719 * Record if the lock has been held longer now than ever
722 if (holdtime > lp->cnt_max)
723 lp->cnt_max = holdtime;
724 if (l->lpo_waittime > lp->cnt_wait_max)
725 lp->cnt_wait_max = l->lpo_waittime;
726 lp->cnt_tot += holdtime;
727 lp->cnt_wait += l->lpo_waittime;
728 lp->cnt_contest_locking += l->lpo_contest_locking;
729 lp->cnt_cur += l->lpo_cnt;
731 LIST_REMOVE(l, lpo_link);
732 type = &LP_CPU_SELF->lpc_types[spin];
733 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
736 * Paired with cpus_fence_seq_cst().
738 atomic_thread_fence_rel();
742 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof,
743 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
745 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
746 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
747 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
748 &lock_prof_rejected, 0, "Number of rejected profiling records");
749 SYSCTL_INT(_debug_lock_prof, OID_AUTO, contested_only, CTLFLAG_RW,
750 &lock_contested_only, 0, "Only profile contested acquires");
751 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats,
752 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
753 dump_lock_prof_stats, "A",
754 "Lock profiling statistics");
755 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset,
756 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0,
757 reset_lock_prof_stats, "I",
758 "Reset lock profiling statistics");
759 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable,
760 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
761 enable_lock_prof, "I",
762 "Enable lock profiling");