2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * This module holds the global variables and functions used to maintain
31 * lock_object structures.
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
38 #include "opt_mprof.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.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>
61 SDT_PROVIDER_DEFINE(lock);
62 SDT_PROBE_DEFINE1(lock, , , starvation, "u_int");
64 CTASSERT(LOCK_CLASS_MAX == 15);
66 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
68 &lock_class_mtx_sleep,
71 &lock_class_rm_sleepable,
77 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
78 const char *type, int flags)
82 /* Check for double-init and zero object. */
83 KASSERT(flags & LO_NEW || !lock_initialized(lock),
84 ("lock \"%s\" %p already initialized", name, lock));
86 /* Look up lock class to find its index. */
87 for (i = 0; i < LOCK_CLASS_MAX; i++)
88 if (lock_classes[i] == class) {
89 lock->lo_flags = i << LO_CLASSSHIFT;
92 KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
94 /* Initialize the lock object. */
96 lock->lo_flags |= flags | LO_INITIALIZED;
97 LOCK_LOG_INIT(lock, 0);
98 WITNESS_INIT(lock, (type != NULL) ? type : name);
102 lock_destroy(struct lock_object *lock)
105 KASSERT(lock_initialized(lock), ("lock %p is not initialized", lock));
106 WITNESS_DESTROY(lock);
107 LOCK_LOG_DESTROY(lock, 0);
108 lock->lo_flags &= ~LO_INITIALIZED;
111 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD, NULL, "lock debugging");
112 static SYSCTL_NODE(_debug_lock, OID_AUTO, delay, CTLFLAG_RD, NULL,
115 static u_int __read_mostly starvation_limit = 131072;
116 SYSCTL_INT(_debug_lock_delay, OID_AUTO, starvation_limit, CTLFLAG_RW,
117 &starvation_limit, 0, "");
119 static u_int __read_mostly restrict_starvation = 0;
120 SYSCTL_INT(_debug_lock_delay, OID_AUTO, restrict_starvation, CTLFLAG_RW,
121 &restrict_starvation, 0, "");
124 lock_delay(struct lock_delay_arg *la)
126 struct lock_delay_config *lc = la->config;
130 if (__predict_false(la->delay > lc->max))
133 for (i = la->delay; i > 0; i--)
136 la->spin_cnt += la->delay;
137 if (__predict_false(la->spin_cnt > starvation_limit)) {
138 SDT_PROBE1(lock, , , starvation, la->delay);
139 if (restrict_starvation)
140 la->delay = lc->base;
145 lock_roundup_2(u_int val)
149 for (res = 1; res <= val; res <<= 1)
156 lock_delay_default_init(struct lock_delay_config *lc)
160 lc->max = lock_roundup_2(mp_ncpus) * 256;
166 DB_SHOW_COMMAND(lock, db_show_lock)
168 struct lock_object *lock;
169 struct lock_class *class;
173 lock = (struct lock_object *)addr;
174 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
175 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
178 class = LOCK_CLASS(lock);
179 db_printf(" class: %s\n", class->lc_name);
180 db_printf(" name: %s\n", lock->lo_name);
181 class->lc_ddb_show(lock);
185 #ifdef LOCK_PROFILING
188 * One object per-thread for each lock the thread owns. Tracks individual
191 struct lock_profile_object {
192 LIST_ENTRY(lock_profile_object) lpo_link;
193 struct lock_object *lpo_obj;
194 const char *lpo_file;
198 uint64_t lpo_acqtime;
199 uint64_t lpo_waittime;
200 u_int lpo_contest_locking;
204 * One lock_prof for each (file, line, lock object) triple.
207 SLIST_ENTRY(lock_prof) link;
208 struct lock_class *class;
213 uintmax_t cnt_wait_max;
218 uintmax_t cnt_contest_locking;
221 SLIST_HEAD(lphead, lock_prof);
223 #define LPROF_HASH_SIZE 4096
224 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
225 #define LPROF_CACHE_SIZE 4096
228 * Array of objects and profs for each type of object for each cpu. Spinlocks
229 * are handled separately because a thread may be preempted and acquire a
230 * spinlock while in the lock profiling code of a non-spinlock. In this way
231 * we only need a critical section to protect the per-cpu lists.
233 struct lock_prof_type {
234 struct lphead lpt_lpalloc;
235 struct lpohead lpt_lpoalloc;
236 struct lphead lpt_hash[LPROF_HASH_SIZE];
237 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
238 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
241 struct lock_prof_cpu {
242 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
245 struct lock_prof_cpu *lp_cpu[MAXCPU];
247 volatile int __read_mostly lock_prof_enable;
248 static volatile int lock_prof_resetting;
250 #define LPROF_SBUF_SIZE 256
252 static int lock_prof_rejected;
253 static int lock_prof_skipspin;
254 static int lock_prof_skipcount;
256 #ifndef USE_CPU_NANOSECONDS
264 /* From bintime2timespec */
265 ns = bt.sec * (uint64_t)1000000000;
266 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
272 lock_prof_init_type(struct lock_prof_type *type)
276 SLIST_INIT(&type->lpt_lpalloc);
277 LIST_INIT(&type->lpt_lpoalloc);
278 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
279 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
281 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
287 lock_prof_init(void *arg)
291 for (cpu = 0; cpu <= mp_maxid; cpu++) {
292 lp_cpu[cpu] = malloc(sizeof(*lp_cpu[cpu]), M_DEVBUF,
294 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[0]);
295 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[1]);
298 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
301 lock_prof_reset_wait(void)
305 * Spin relinquishing our cpu so that quiesce_all_cpus may
308 while (lock_prof_resetting)
309 sched_relinquish(curthread);
313 lock_prof_reset(void)
315 struct lock_prof_cpu *lpc;
319 * We not only race with acquiring and releasing locks but also
320 * thread exit. To be certain that threads exit without valid head
321 * pointers they must see resetting set before enabled is cleared.
322 * Otherwise a lock may not be removed from a per-thread list due
323 * to disabled being set but not wait for reset() to remove it below.
325 atomic_store_rel_int(&lock_prof_resetting, 1);
326 enabled = lock_prof_enable;
327 lock_prof_enable = 0;
328 quiesce_all_cpus("profreset", 0);
330 * Some objects may have migrated between CPUs. Clear all links
331 * before we zero the structures. Some items may still be linked
332 * into per-thread lists as well.
334 for (cpu = 0; cpu <= mp_maxid; cpu++) {
336 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
337 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
338 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
341 for (cpu = 0; cpu <= mp_maxid; cpu++) {
343 bzero(lpc, sizeof(*lpc));
344 lock_prof_init_type(&lpc->lpc_types[0]);
345 lock_prof_init_type(&lpc->lpc_types[1]);
347 atomic_store_rel_int(&lock_prof_resetting, 0);
348 lock_prof_enable = enabled;
352 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
356 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
358 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
359 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
360 lp->cnt_wait / 1000, lp->cnt_cur,
361 lp->cnt_cur == 0 ? (uintmax_t)0 :
362 lp->cnt_tot / (lp->cnt_cur * 1000),
363 lp->cnt_cur == 0 ? (uintmax_t)0 :
364 lp->cnt_wait / (lp->cnt_cur * 1000),
365 (uintmax_t)0, lp->cnt_contest_locking,
366 p, lp->line, lp->class->lc_name, lp->name);
370 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
373 struct lock_prof_type *type;
377 dst->file = match->file;
378 dst->line = match->line;
379 dst->class = match->class;
380 dst->name = match->name;
382 for (cpu = 0; cpu <= mp_maxid; cpu++) {
383 if (lp_cpu[cpu] == NULL)
385 type = &lp_cpu[cpu]->lpc_types[spin];
386 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
389 if (l->file != match->file || l->line != match->line ||
390 l->name != match->name)
393 if (l->cnt_max > dst->cnt_max)
394 dst->cnt_max = l->cnt_max;
395 if (l->cnt_wait_max > dst->cnt_wait_max)
396 dst->cnt_wait_max = l->cnt_wait_max;
397 dst->cnt_tot += l->cnt_tot;
398 dst->cnt_wait += l->cnt_wait;
399 dst->cnt_cur += l->cnt_cur;
400 dst->cnt_contest_locking += l->cnt_contest_locking;
407 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
413 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
414 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
415 struct lock_prof lp = {};
419 lock_prof_sum(l, &lp, i, spin, t);
420 lock_prof_output(&lp, sb);
426 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
432 error = sysctl_wire_old_buffer(req, 0);
435 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
436 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
437 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
438 enabled = lock_prof_enable;
439 lock_prof_enable = 0;
440 quiesce_all_cpus("profstat", 0);
442 for (cpu = 0; cpu <= mp_maxid; cpu++) {
443 if (lp_cpu[cpu] == NULL)
445 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[0], sb, 0, t);
446 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[1], sb, 1, t);
448 lock_prof_enable = enabled;
450 error = sbuf_finish(sb);
451 /* Output a trailing NUL. */
453 error = SYSCTL_OUT(req, "", 1);
459 enable_lock_prof(SYSCTL_HANDLER_ARGS)
463 v = lock_prof_enable;
464 error = sysctl_handle_int(oidp, &v, v, req);
467 if (req->newptr == NULL)
469 if (v == lock_prof_enable)
473 lock_prof_enable = !!v;
479 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
484 error = sysctl_handle_int(oidp, &v, 0, req);
487 if (req->newptr == NULL)
496 static struct lock_prof *
497 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
500 const char *unknown = "(unknown)";
501 struct lock_prof_type *type;
502 struct lock_prof *lp;
508 if (p == NULL || *p == '\0')
510 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
511 hash &= LPROF_HASH_MASK;
512 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
513 head = &type->lpt_hash[hash];
514 SLIST_FOREACH(lp, head, link) {
515 if (lp->line == line && lp->file == p &&
516 lp->name == lo->lo_name)
520 lp = SLIST_FIRST(&type->lpt_lpalloc);
522 lock_prof_rejected++;
525 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
528 lp->class = LOCK_CLASS(lo);
529 lp->name = lo->lo_name;
530 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
534 static struct lock_profile_object *
535 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
538 struct lock_profile_object *l;
539 struct lock_prof_type *type;
540 struct lpohead *head;
542 head = &curthread->td_lprof[spin];
543 LIST_FOREACH(l, head, lpo_link)
544 if (l->lpo_obj == lo && l->lpo_file == file &&
547 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
548 l = LIST_FIRST(&type->lpt_lpoalloc);
550 lock_prof_rejected++;
553 LIST_REMOVE(l, lpo_link);
558 LIST_INSERT_HEAD(head, l, lpo_link);
564 lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
565 uint64_t waittime, const char *file, int line)
567 static int lock_prof_count;
568 struct lock_profile_object *l;
571 if (SCHEDULER_STOPPED())
574 /* don't reset the timer when/if recursing */
575 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
577 if (lock_prof_skipcount &&
578 (++lock_prof_count % lock_prof_skipcount) != 0)
580 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
581 if (spin && lock_prof_skipspin == 1)
584 /* Recheck enabled now that we're in a critical section. */
585 if (lock_prof_enable == 0)
587 l = lock_profile_object_lookup(lo, spin, file, line);
591 if (++l->lpo_ref > 1)
593 l->lpo_contest_locking = contested;
594 l->lpo_acqtime = nanoseconds();
595 if (waittime && (l->lpo_acqtime > waittime))
596 l->lpo_waittime = l->lpo_acqtime - waittime;
604 lock_profile_thread_exit(struct thread *td)
607 struct lock_profile_object *l;
609 MPASS(curthread->td_critnest == 0);
612 * If lock profiling was disabled we have to wait for reset to
613 * clear our pointers before we can exit safely.
615 lock_prof_reset_wait();
617 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
618 printf("thread still holds lock acquired at %s:%d\n",
619 l->lpo_file, l->lpo_line);
620 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
621 printf("thread still holds lock acquired at %s:%d\n",
622 l->lpo_file, l->lpo_line);
624 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
625 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
629 lock_profile_release_lock(struct lock_object *lo)
631 struct lock_profile_object *l;
632 struct lock_prof_type *type;
633 struct lock_prof *lp;
634 uint64_t curtime, holdtime;
635 struct lpohead *head;
638 if (SCHEDULER_STOPPED())
640 if (lo->lo_flags & LO_NOPROFILE)
642 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
643 head = &curthread->td_lprof[spin];
644 if (LIST_FIRST(head) == NULL)
647 /* Recheck enabled now that we're in a critical section. */
648 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
651 * If lock profiling is not enabled we still want to remove the
652 * lpo from our queue.
654 LIST_FOREACH(l, head, lpo_link)
655 if (l->lpo_obj == lo)
659 if (--l->lpo_ref > 0)
661 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
664 curtime = nanoseconds();
665 if (curtime < l->lpo_acqtime)
667 holdtime = curtime - l->lpo_acqtime;
670 * Record if the lock has been held longer now than ever
673 if (holdtime > lp->cnt_max)
674 lp->cnt_max = holdtime;
675 if (l->lpo_waittime > lp->cnt_wait_max)
676 lp->cnt_wait_max = l->lpo_waittime;
677 lp->cnt_tot += holdtime;
678 lp->cnt_wait += l->lpo_waittime;
679 lp->cnt_contest_locking += l->lpo_contest_locking;
680 lp->cnt_cur += l->lpo_cnt;
682 LIST_REMOVE(l, lpo_link);
683 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
684 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
689 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL,
691 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
692 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
693 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
694 &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
695 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
696 &lock_prof_rejected, 0, "Number of rejected profiling records");
697 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
698 NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
699 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
700 NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
701 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
702 NULL, 0, enable_lock_prof, "I", "Enable lock profiling");