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, NULL, "lock debugging");
111 static SYSCTL_NODE(_debug_lock, OID_AUTO, delay, CTLFLAG_RD, NULL,
114 static u_int __read_mostly starvation_limit = 131072;
115 SYSCTL_INT(_debug_lock_delay, OID_AUTO, starvation_limit, CTLFLAG_RW,
116 &starvation_limit, 0, "");
118 static u_int __read_mostly restrict_starvation = 0;
119 SYSCTL_INT(_debug_lock_delay, OID_AUTO, restrict_starvation, CTLFLAG_RW,
120 &restrict_starvation, 0, "");
123 lock_delay(struct lock_delay_arg *la)
125 struct lock_delay_config *lc = la->config;
129 if (__predict_false(la->delay > lc->max))
132 for (i = la->delay; i > 0; i--)
135 la->spin_cnt += la->delay;
136 if (__predict_false(la->spin_cnt > starvation_limit)) {
137 SDT_PROBE1(lock, , , starvation, la->delay);
138 if (restrict_starvation)
139 la->delay = lc->base;
144 lock_roundup_2(u_int val)
148 for (res = 1; res <= val; res <<= 1)
155 lock_delay_default_init(struct lock_delay_config *lc)
159 lc->max = lock_roundup_2(mp_ncpus) * 256;
164 struct lock_delay_config __read_frequently locks_delay;
165 u_short __read_frequently locks_delay_retries;
166 u_short __read_frequently locks_delay_loops;
168 SYSCTL_U16(_debug_lock, OID_AUTO, delay_base, CTLFLAG_RW, &locks_delay.base,
170 SYSCTL_U16(_debug_lock, OID_AUTO, delay_max, CTLFLAG_RW, &locks_delay.max,
172 SYSCTL_U16(_debug_lock, OID_AUTO, delay_retries, CTLFLAG_RW, &locks_delay_retries,
174 SYSCTL_U16(_debug_lock, OID_AUTO, delay_loops, CTLFLAG_RW, &locks_delay_loops,
178 locks_delay_init(void *arg __unused)
181 lock_delay_default_init(&locks_delay);
182 locks_delay_retries = 10;
183 locks_delay_loops = max(10000, locks_delay.max);
185 LOCK_DELAY_SYSINIT(locks_delay_init);
188 DB_SHOW_COMMAND(lock, db_show_lock)
190 struct lock_object *lock;
191 struct lock_class *class;
195 lock = (struct lock_object *)addr;
196 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
197 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
200 class = LOCK_CLASS(lock);
201 db_printf(" class: %s\n", class->lc_name);
202 db_printf(" name: %s\n", lock->lo_name);
203 class->lc_ddb_show(lock);
207 #ifdef LOCK_PROFILING
210 * One object per-thread for each lock the thread owns. Tracks individual
213 struct lock_profile_object {
214 LIST_ENTRY(lock_profile_object) lpo_link;
215 struct lock_object *lpo_obj;
216 const char *lpo_file;
220 uint64_t lpo_acqtime;
221 uint64_t lpo_waittime;
222 u_int lpo_contest_locking;
226 * One lock_prof for each (file, line, lock object) triple.
229 SLIST_ENTRY(lock_prof) link;
230 struct lock_class *class;
235 uintmax_t cnt_wait_max;
240 uintmax_t cnt_contest_locking;
243 SLIST_HEAD(lphead, lock_prof);
245 #define LPROF_HASH_SIZE 4096
246 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
247 #define LPROF_CACHE_SIZE 4096
250 * Array of objects and profs for each type of object for each cpu. Spinlocks
251 * are handled separately because a thread may be preempted and acquire a
252 * spinlock while in the lock profiling code of a non-spinlock. In this way
253 * we only need a critical section to protect the per-cpu lists.
255 struct lock_prof_type {
256 struct lphead lpt_lpalloc;
257 struct lpohead lpt_lpoalloc;
258 struct lphead lpt_hash[LPROF_HASH_SIZE];
259 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
260 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
263 struct lock_prof_cpu {
264 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
267 DPCPU_DEFINE_STATIC(struct lock_prof_cpu, lp);
268 #define LP_CPU_SELF (DPCPU_PTR(lp))
269 #define LP_CPU(cpu) (DPCPU_ID_PTR((cpu), lp))
271 volatile int __read_mostly lock_prof_enable;
272 static volatile int lock_prof_resetting;
274 #define LPROF_SBUF_SIZE 256
276 static int lock_prof_rejected;
277 static int lock_prof_skipspin;
278 static int lock_prof_skipcount;
280 #ifndef USE_CPU_NANOSECONDS
288 /* From bintime2timespec */
289 ns = bt.sec * (uint64_t)1000000000;
290 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
296 lock_prof_init_type(struct lock_prof_type *type)
300 SLIST_INIT(&type->lpt_lpalloc);
301 LIST_INIT(&type->lpt_lpoalloc);
302 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
303 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
305 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
311 lock_prof_init(void *arg)
316 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[0]);
317 lock_prof_init_type(&LP_CPU(cpu)->lpc_types[1]);
320 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
323 lock_prof_reset_wait(void)
327 * Spin relinquishing our cpu so that quiesce_all_cpus may
330 while (lock_prof_resetting)
331 sched_relinquish(curthread);
335 lock_prof_reset(void)
337 struct lock_prof_cpu *lpc;
341 * We not only race with acquiring and releasing locks but also
342 * thread exit. To be certain that threads exit without valid head
343 * pointers they must see resetting set before enabled is cleared.
344 * Otherwise a lock may not be removed from a per-thread list due
345 * to disabled being set but not wait for reset() to remove it below.
347 atomic_store_rel_int(&lock_prof_resetting, 1);
348 enabled = lock_prof_enable;
349 lock_prof_enable = 0;
351 * This both publishes lock_prof_enable as disabled and makes sure
352 * everyone else reads it if they are not far enough. We wait for the
355 cpus_fence_seq_cst();
356 quiesce_all_critical();
358 * Some objects may have migrated between CPUs. Clear all links
359 * before we zero the structures. Some items may still be linked
360 * into per-thread lists as well.
364 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
365 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
366 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
371 bzero(lpc, sizeof(*lpc));
372 lock_prof_init_type(&lpc->lpc_types[0]);
373 lock_prof_init_type(&lpc->lpc_types[1]);
376 * Paired with the fence from cpus_fence_seq_cst()
378 atomic_store_rel_int(&lock_prof_resetting, 0);
379 lock_prof_enable = enabled;
383 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
387 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
389 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
390 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
391 lp->cnt_wait / 1000, lp->cnt_cur,
392 lp->cnt_cur == 0 ? (uintmax_t)0 :
393 lp->cnt_tot / (lp->cnt_cur * 1000),
394 lp->cnt_cur == 0 ? (uintmax_t)0 :
395 lp->cnt_wait / (lp->cnt_cur * 1000),
396 (uintmax_t)0, lp->cnt_contest_locking,
397 p, lp->line, lp->class->lc_name, lp->name);
401 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
404 struct lock_prof_type *type;
408 dst->file = match->file;
409 dst->line = match->line;
410 dst->class = match->class;
411 dst->name = match->name;
414 type = &LP_CPU(cpu)->lpc_types[spin];
415 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
418 if (l->file != match->file || l->line != match->line ||
419 l->name != match->name)
422 if (l->cnt_max > dst->cnt_max)
423 dst->cnt_max = l->cnt_max;
424 if (l->cnt_wait_max > dst->cnt_wait_max)
425 dst->cnt_wait_max = l->cnt_wait_max;
426 dst->cnt_tot += l->cnt_tot;
427 dst->cnt_wait += l->cnt_wait;
428 dst->cnt_cur += l->cnt_cur;
429 dst->cnt_contest_locking += l->cnt_contest_locking;
435 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
441 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
442 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
443 struct lock_prof lp = {};
447 lock_prof_sum(l, &lp, i, spin, t);
448 lock_prof_output(&lp, sb);
454 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
460 error = sysctl_wire_old_buffer(req, 0);
463 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
464 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
465 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
466 enabled = lock_prof_enable;
467 lock_prof_enable = 0;
469 * See the comment in lock_prof_reset
471 cpus_fence_seq_cst();
472 quiesce_all_critical();
475 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[0], sb, 0, t);
476 lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[1], sb, 1, t);
478 atomic_thread_fence_rel();
479 lock_prof_enable = enabled;
481 error = sbuf_finish(sb);
482 /* Output a trailing NUL. */
484 error = SYSCTL_OUT(req, "", 1);
490 enable_lock_prof(SYSCTL_HANDLER_ARGS)
494 v = lock_prof_enable;
495 error = sysctl_handle_int(oidp, &v, v, req);
498 if (req->newptr == NULL)
500 if (v == lock_prof_enable)
504 lock_prof_enable = !!v;
510 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
515 error = sysctl_handle_int(oidp, &v, 0, req);
518 if (req->newptr == NULL)
527 static struct lock_prof *
528 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
531 const char *unknown = "(unknown)";
532 struct lock_prof_type *type;
533 struct lock_prof *lp;
539 if (p == NULL || *p == '\0')
541 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
542 hash &= LPROF_HASH_MASK;
543 type = &LP_CPU_SELF->lpc_types[spin];
544 head = &type->lpt_hash[hash];
545 SLIST_FOREACH(lp, head, link) {
546 if (lp->line == line && lp->file == p &&
547 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 static int lock_prof_count;
599 struct lock_profile_object *l;
602 if (SCHEDULER_STOPPED())
605 /* don't reset the timer when/if recursing */
606 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
608 if (lock_prof_skipcount &&
609 (++lock_prof_count % lock_prof_skipcount) != 0)
611 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
612 if (spin && lock_prof_skipspin == 1)
615 /* Recheck enabled now that we're in a critical section. */
616 if (lock_prof_enable == 0)
618 l = lock_profile_object_lookup(lo, spin, file, line);
622 if (++l->lpo_ref > 1)
624 l->lpo_contest_locking = contested;
625 l->lpo_acqtime = nanoseconds();
626 if (waittime && (l->lpo_acqtime > waittime))
627 l->lpo_waittime = l->lpo_acqtime - waittime;
632 * Paired with cpus_fence_seq_cst().
634 atomic_thread_fence_rel();
639 lock_profile_thread_exit(struct thread *td)
642 struct lock_profile_object *l;
644 MPASS(curthread->td_critnest == 0);
647 * If lock profiling was disabled we have to wait for reset to
648 * clear our pointers before we can exit safely.
650 lock_prof_reset_wait();
652 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
653 printf("thread still holds lock acquired at %s:%d\n",
654 l->lpo_file, l->lpo_line);
655 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
656 printf("thread still holds lock acquired at %s:%d\n",
657 l->lpo_file, l->lpo_line);
659 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
660 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
664 lock_profile_release_lock(struct lock_object *lo)
666 struct lock_profile_object *l;
667 struct lock_prof_type *type;
668 struct lock_prof *lp;
669 uint64_t curtime, holdtime;
670 struct lpohead *head;
673 if (SCHEDULER_STOPPED())
675 if (lo->lo_flags & LO_NOPROFILE)
677 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
678 head = &curthread->td_lprof[spin];
679 if (LIST_FIRST(head) == NULL)
682 /* Recheck enabled now that we're in a critical section. */
683 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
686 * If lock profiling is not enabled we still want to remove the
687 * lpo from our queue.
689 LIST_FOREACH(l, head, lpo_link)
690 if (l->lpo_obj == lo)
694 if (--l->lpo_ref > 0)
696 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
699 curtime = nanoseconds();
700 if (curtime < l->lpo_acqtime)
702 holdtime = curtime - l->lpo_acqtime;
705 * Record if the lock has been held longer now than ever
708 if (holdtime > lp->cnt_max)
709 lp->cnt_max = holdtime;
710 if (l->lpo_waittime > lp->cnt_wait_max)
711 lp->cnt_wait_max = l->lpo_waittime;
712 lp->cnt_tot += holdtime;
713 lp->cnt_wait += l->lpo_waittime;
714 lp->cnt_contest_locking += l->lpo_contest_locking;
715 lp->cnt_cur += l->lpo_cnt;
717 LIST_REMOVE(l, lpo_link);
718 type = &LP_CPU_SELF->lpc_types[spin];
719 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
722 * Paired with cpus_fence_seq_cst().
724 atomic_thread_fence_rel();
728 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL,
730 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
731 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
732 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
733 &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
734 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
735 &lock_prof_rejected, 0, "Number of rejected profiling records");
736 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
737 NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
738 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
739 NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
740 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
741 NULL, 0, enable_lock_prof, "I", "Enable lock profiling");