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>
60 SDT_PROVIDER_DEFINE(lock);
61 SDT_PROBE_DEFINE1(lock, , , starvation, "u_int");
63 CTASSERT(LOCK_CLASS_MAX == 15);
65 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
67 &lock_class_mtx_sleep,
70 &lock_class_rm_sleepable,
76 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
77 const char *type, int flags)
81 /* Check for double-init and zero object. */
82 KASSERT(flags & LO_NEW || !lock_initialized(lock),
83 ("lock \"%s\" %p already initialized", name, lock));
85 /* Look up lock class to find its index. */
86 for (i = 0; i < LOCK_CLASS_MAX; i++)
87 if (lock_classes[i] == class) {
88 lock->lo_flags = i << LO_CLASSSHIFT;
91 KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
93 /* Initialize the lock object. */
95 lock->lo_flags |= flags | LO_INITIALIZED;
96 LOCK_LOG_INIT(lock, 0);
97 WITNESS_INIT(lock, (type != NULL) ? type : name);
101 lock_destroy(struct lock_object *lock)
104 KASSERT(lock_initialized(lock), ("lock %p is not initialized", lock));
105 WITNESS_DESTROY(lock);
106 LOCK_LOG_DESTROY(lock, 0);
107 lock->lo_flags &= ~LO_INITIALIZED;
110 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
112 static SYSCTL_NODE(_debug_lock, OID_AUTO, delay,
113 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
116 static u_int __read_mostly starvation_limit = 131072;
117 SYSCTL_INT(_debug_lock_delay, OID_AUTO, starvation_limit, CTLFLAG_RW,
118 &starvation_limit, 0, "");
120 static u_int __read_mostly restrict_starvation = 0;
121 SYSCTL_INT(_debug_lock_delay, OID_AUTO, restrict_starvation, CTLFLAG_RW,
122 &restrict_starvation, 0, "");
125 lock_delay(struct lock_delay_arg *la)
127 struct lock_delay_config *lc = la->config;
131 if (__predict_false(la->delay > lc->max))
134 for (i = la->delay; i > 0; i--)
137 la->spin_cnt += la->delay;
138 if (__predict_false(la->spin_cnt > starvation_limit)) {
139 SDT_PROBE1(lock, , , starvation, la->delay);
140 if (restrict_starvation)
141 la->delay = lc->base;
146 lock_roundup_2(u_int val)
150 for (res = 1; res <= val; res <<= 1)
157 lock_delay_default_init(struct lock_delay_config *lc)
161 lc->max = lock_roundup_2(mp_ncpus) * 256;
166 struct lock_delay_config __read_frequently locks_delay;
167 u_short __read_frequently locks_delay_retries;
168 u_short __read_frequently locks_delay_loops;
170 SYSCTL_U16(_debug_lock, OID_AUTO, delay_base, CTLFLAG_RW, &locks_delay.base,
172 SYSCTL_U16(_debug_lock, OID_AUTO, delay_max, CTLFLAG_RW, &locks_delay.max,
174 SYSCTL_U16(_debug_lock, OID_AUTO, delay_retries, CTLFLAG_RW, &locks_delay_retries,
176 SYSCTL_U16(_debug_lock, OID_AUTO, delay_loops, CTLFLAG_RW, &locks_delay_loops,
180 locks_delay_init(void *arg __unused)
183 lock_delay_default_init(&locks_delay);
184 locks_delay_retries = 10;
185 locks_delay_loops = max(10000, locks_delay.max);
187 LOCK_DELAY_SYSINIT(locks_delay_init);
190 DB_SHOW_COMMAND(lock, db_show_lock)
192 struct lock_object *lock;
193 struct lock_class *class;
197 lock = (struct lock_object *)addr;
198 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
199 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
202 class = LOCK_CLASS(lock);
203 db_printf(" class: %s\n", class->lc_name);
204 db_printf(" name: %s\n", lock->lo_name);
205 class->lc_ddb_show(lock);
209 #ifdef LOCK_PROFILING
212 * One object per-thread for each lock the thread owns. Tracks individual
215 struct lock_profile_object {
216 LIST_ENTRY(lock_profile_object) lpo_link;
217 struct lock_object *lpo_obj;
218 const char *lpo_file;
222 uint64_t lpo_acqtime;
223 uint64_t lpo_waittime;
224 u_int lpo_contest_locking;
228 * One lock_prof for each (file, line, lock object) triple.
231 SLIST_ENTRY(lock_prof) link;
232 struct lock_class *class;
237 uintmax_t cnt_wait_max;
242 uintmax_t cnt_contest_locking;
245 SLIST_HEAD(lphead, lock_prof);
247 #define LPROF_HASH_SIZE 4096
248 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
249 #define LPROF_CACHE_SIZE 4096
252 * Array of objects and profs for each type of object for each cpu. Spinlocks
253 * are handled separately because a thread may be preempted and acquire a
254 * spinlock while in the lock profiling code of a non-spinlock. In this way
255 * we only need a critical section to protect the per-cpu lists.
257 struct lock_prof_type {
258 struct lphead lpt_lpalloc;
259 struct lpohead lpt_lpoalloc;
260 struct lphead lpt_hash[LPROF_HASH_SIZE];
261 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
262 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
265 struct lock_prof_cpu {
266 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
269 DPCPU_DEFINE_STATIC(struct lock_prof_cpu, lp);
270 #define LP_CPU_SELF (DPCPU_PTR(lp))
271 #define LP_CPU(cpu) (DPCPU_ID_PTR((cpu), lp))
273 volatile int __read_mostly lock_prof_enable;
274 static volatile int lock_prof_resetting;
276 #define LPROF_SBUF_SIZE 256
278 static int lock_prof_rejected;
279 static int lock_prof_skipspin;
281 #ifndef USE_CPU_NANOSECONDS
289 /* From bintime2timespec */
290 ns = bt.sec * (uint64_t)1000000000;
291 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
297 lock_prof_init_type(struct lock_prof_type *type)
301 SLIST_INIT(&type->lpt_lpalloc);
302 LIST_INIT(&type->lpt_lpoalloc);
303 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
304 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
306 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
312 lock_prof_init(void *arg)
317 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[0]);
318 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[1]);
321 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
324 lock_prof_reset_wait(void)
328 * Spin relinquishing our cpu so that quiesce_all_cpus may
331 while (lock_prof_resetting)
332 sched_relinquish(curthread);
336 lock_prof_reset(void)
338 struct lock_prof_cpu *lpc;
342 * We not only race with acquiring and releasing locks but also
343 * thread exit. To be certain that threads exit without valid head
344 * pointers they must see resetting set before enabled is cleared.
345 * Otherwise a lock may not be removed from a per-thread list due
346 * to disabled being set but not wait for reset() to remove it below.
348 atomic_store_rel_int(&lock_prof_resetting, 1);
349 enabled = lock_prof_enable;
350 lock_prof_enable = 0;
352 * This both publishes lock_prof_enable as disabled and makes sure
353 * everyone else reads it if they are not far enough. We wait for the
356 cpus_fence_seq_cst();
357 quiesce_all_critical();
359 * Some objects may have migrated between CPUs. Clear all links
360 * before we zero the structures. Some items may still be linked
361 * into per-thread lists as well.
365 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
366 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
367 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
372 bzero(lpc, sizeof(*lpc));
373 lock_prof_init_type(&lpc->lpc_types[0]);
374 lock_prof_init_type(&lpc->lpc_types[1]);
377 * Paired with the fence from cpus_fence_seq_cst()
379 atomic_store_rel_int(&lock_prof_resetting, 0);
380 lock_prof_enable = enabled;
384 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
388 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
390 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
391 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
392 lp->cnt_wait / 1000, lp->cnt_cur,
393 lp->cnt_cur == 0 ? (uintmax_t)0 :
394 lp->cnt_tot / (lp->cnt_cur * 1000),
395 lp->cnt_cur == 0 ? (uintmax_t)0 :
396 lp->cnt_wait / (lp->cnt_cur * 1000),
397 (uintmax_t)0, lp->cnt_contest_locking,
398 p, lp->line, lp->class->lc_name, lp->name);
402 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
405 struct lock_prof_type *type;
409 dst->file = match->file;
410 dst->line = match->line;
411 dst->class = match->class;
412 dst->name = match->name;
415 type = &LP_CPU(cpu)->lpc_types[spin];
416 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
419 if (l->file != match->file || l->line != match->line ||
420 l->name != match->name)
423 if (l->cnt_max > dst->cnt_max)
424 dst->cnt_max = l->cnt_max;
425 if (l->cnt_wait_max > dst->cnt_wait_max)
426 dst->cnt_wait_max = l->cnt_wait_max;
427 dst->cnt_tot += l->cnt_tot;
428 dst->cnt_wait += l->cnt_wait;
429 dst->cnt_cur += l->cnt_cur;
430 dst->cnt_contest_locking += l->cnt_contest_locking;
436 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
442 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
443 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
444 struct lock_prof lp = {};
448 lock_prof_sum(l, &lp, i, spin, t);
449 lock_prof_output(&lp, sb);
455 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
461 error = sysctl_wire_old_buffer(req, 0);
464 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
465 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
466 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
467 enabled = lock_prof_enable;
468 lock_prof_enable = 0;
470 * See the comment in lock_prof_reset
472 cpus_fence_seq_cst();
473 quiesce_all_critical();
476 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[0], sb, 0, t);
477 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[1], sb, 1, t);
479 atomic_thread_fence_rel();
480 lock_prof_enable = enabled;
482 error = sbuf_finish(sb);
483 /* Output a trailing NUL. */
485 error = SYSCTL_OUT(req, "", 1);
491 enable_lock_prof(SYSCTL_HANDLER_ARGS)
495 v = lock_prof_enable;
496 error = sysctl_handle_int(oidp, &v, v, req);
499 if (req->newptr == NULL)
501 if (v == lock_prof_enable)
505 lock_prof_enable = !!v;
511 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
516 error = sysctl_handle_int(oidp, &v, 0, req);
519 if (req->newptr == NULL)
528 static struct lock_prof *
529 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
532 const char *unknown = "(unknown)";
533 struct lock_prof_type *type;
534 struct lock_prof *lp;
540 if (p == NULL || *p == '\0')
542 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
543 hash &= LPROF_HASH_MASK;
544 type = &LP_CPU_SELF->lpc_types[spin];
545 head = &type->lpt_hash[hash];
546 SLIST_FOREACH(lp, head, link) {
547 if (lp->line == line && lp->file == p &&
548 lp->name == lo->lo_name)
551 lp = SLIST_FIRST(&type->lpt_lpalloc);
553 lock_prof_rejected++;
556 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
559 lp->class = LOCK_CLASS(lo);
560 lp->name = lo->lo_name;
561 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
565 static struct lock_profile_object *
566 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
569 struct lock_profile_object *l;
570 struct lock_prof_type *type;
571 struct lpohead *head;
573 head = &curthread->td_lprof[spin];
574 LIST_FOREACH(l, head, lpo_link)
575 if (l->lpo_obj == lo && l->lpo_file == file &&
578 type = &LP_CPU_SELF->lpc_types[spin];
579 l = LIST_FIRST(&type->lpt_lpoalloc);
581 lock_prof_rejected++;
584 LIST_REMOVE(l, lpo_link);
589 LIST_INSERT_HEAD(head, l, lpo_link);
595 lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
596 uint64_t waittime, const char *file, int line)
598 struct lock_profile_object *l;
601 if (SCHEDULER_STOPPED())
604 /* don't reset the timer when/if recursing */
605 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
607 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
608 if (spin && lock_prof_skipspin == 1)
611 /* Recheck enabled now that we're in a critical section. */
612 if (lock_prof_enable == 0)
614 l = lock_profile_object_lookup(lo, spin, file, line);
618 if (++l->lpo_ref > 1)
620 l->lpo_contest_locking = contested;
621 l->lpo_acqtime = nanoseconds();
622 if (waittime && (l->lpo_acqtime > waittime))
623 l->lpo_waittime = l->lpo_acqtime - waittime;
628 * Paired with cpus_fence_seq_cst().
630 atomic_thread_fence_rel();
635 lock_profile_thread_exit(struct thread *td)
638 struct lock_profile_object *l;
640 MPASS(curthread->td_critnest == 0);
643 * If lock profiling was disabled we have to wait for reset to
644 * clear our pointers before we can exit safely.
646 lock_prof_reset_wait();
648 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
649 printf("thread still holds lock acquired at %s:%d\n",
650 l->lpo_file, l->lpo_line);
651 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
652 printf("thread still holds lock acquired at %s:%d\n",
653 l->lpo_file, l->lpo_line);
655 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
656 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
660 lock_profile_release_lock(struct lock_object *lo)
662 struct lock_profile_object *l;
663 struct lock_prof_type *type;
664 struct lock_prof *lp;
665 uint64_t curtime, holdtime;
666 struct lpohead *head;
669 if (SCHEDULER_STOPPED())
671 if (lo->lo_flags & LO_NOPROFILE)
673 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
674 head = &curthread->td_lprof[spin];
675 if (LIST_FIRST(head) == NULL)
678 /* Recheck enabled now that we're in a critical section. */
679 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
682 * If lock profiling is not enabled we still want to remove the
683 * lpo from our queue.
685 LIST_FOREACH(l, head, lpo_link)
686 if (l->lpo_obj == lo)
690 if (--l->lpo_ref > 0)
692 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
695 curtime = nanoseconds();
696 if (curtime < l->lpo_acqtime)
698 holdtime = curtime - l->lpo_acqtime;
701 * Record if the lock has been held longer now than ever
704 if (holdtime > lp->cnt_max)
705 lp->cnt_max = holdtime;
706 if (l->lpo_waittime > lp->cnt_wait_max)
707 lp->cnt_wait_max = l->lpo_waittime;
708 lp->cnt_tot += holdtime;
709 lp->cnt_wait += l->lpo_waittime;
710 lp->cnt_contest_locking += l->lpo_contest_locking;
711 lp->cnt_cur += l->lpo_cnt;
713 LIST_REMOVE(l, lpo_link);
714 type = &LP_CPU_SELF->lpc_types[spin];
715 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
718 * Paired with cpus_fence_seq_cst().
720 atomic_thread_fence_rel();
724 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof,
725 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
727 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
728 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
729 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
730 &lock_prof_rejected, 0, "Number of rejected profiling records");
731 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats,
732 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
733 dump_lock_prof_stats, "A",
734 "Lock profiling statistics");
735 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset,
736 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0,
737 reset_lock_prof_stats, "I",
738 "Reset lock profiling statistics");
739 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable,
740 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
741 enable_lock_prof, "I",
742 "Enable lock profiling");