2 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * This module holds the global variables and functions used to maintain
29 * lock_object structures.
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
36 #include "opt_mprof.h"
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
43 #include <sys/lock_profile.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
49 #include <sys/sched.h>
51 #include <sys/sysctl.h>
57 #include <machine/cpufunc.h>
59 CTASSERT(LOCK_CLASS_MAX == 15);
61 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
63 &lock_class_mtx_sleep,
66 &lock_class_rm_sleepable,
72 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
73 const char *type, int flags)
77 /* Check for double-init and zero object. */
78 KASSERT(flags & LO_NEW || !lock_initialized(lock),
79 ("lock \"%s\" %p already initialized", name, lock));
81 /* Look up lock class to find its index. */
82 for (i = 0; i < LOCK_CLASS_MAX; i++)
83 if (lock_classes[i] == class) {
84 lock->lo_flags = i << LO_CLASSSHIFT;
87 KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
89 /* Initialize the lock object. */
91 lock->lo_flags |= flags | LO_INITIALIZED;
92 LOCK_LOG_INIT(lock, 0);
93 WITNESS_INIT(lock, (type != NULL) ? type : name);
97 lock_destroy(struct lock_object *lock)
100 KASSERT(lock_initialized(lock), ("lock %p is not initialized", lock));
101 WITNESS_DESTROY(lock);
102 LOCK_LOG_DESTROY(lock, 0);
103 lock->lo_flags &= ~LO_INITIALIZED;
107 lock_delay(struct lock_delay_arg *la)
109 u_int i, delay, backoff, min, max;
110 struct lock_delay_config *lc = la->config;
123 backoff = cpu_ticks() % delay;
127 for (i = 0; i < backoff; i++)
131 la->spin_cnt += backoff;
135 DB_SHOW_COMMAND(lock, db_show_lock)
137 struct lock_object *lock;
138 struct lock_class *class;
142 lock = (struct lock_object *)addr;
143 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
144 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
147 class = LOCK_CLASS(lock);
148 db_printf(" class: %s\n", class->lc_name);
149 db_printf(" name: %s\n", lock->lo_name);
150 class->lc_ddb_show(lock);
154 #ifdef LOCK_PROFILING
157 * One object per-thread for each lock the thread owns. Tracks individual
160 struct lock_profile_object {
161 LIST_ENTRY(lock_profile_object) lpo_link;
162 struct lock_object *lpo_obj;
163 const char *lpo_file;
167 uint64_t lpo_acqtime;
168 uint64_t lpo_waittime;
169 u_int lpo_contest_locking;
173 * One lock_prof for each (file, line, lock object) triple.
176 SLIST_ENTRY(lock_prof) link;
177 struct lock_class *class;
182 uintmax_t cnt_wait_max;
187 uintmax_t cnt_contest_locking;
190 SLIST_HEAD(lphead, lock_prof);
192 #define LPROF_HASH_SIZE 4096
193 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
194 #define LPROF_CACHE_SIZE 4096
197 * Array of objects and profs for each type of object for each cpu. Spinlocks
198 * are handled separately because a thread may be preempted and acquire a
199 * spinlock while in the lock profiling code of a non-spinlock. In this way
200 * we only need a critical section to protect the per-cpu lists.
202 struct lock_prof_type {
203 struct lphead lpt_lpalloc;
204 struct lpohead lpt_lpoalloc;
205 struct lphead lpt_hash[LPROF_HASH_SIZE];
206 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
207 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
210 struct lock_prof_cpu {
211 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
214 struct lock_prof_cpu *lp_cpu[MAXCPU];
216 volatile int lock_prof_enable = 0;
217 static volatile int lock_prof_resetting;
219 #define LPROF_SBUF_SIZE 256
221 static int lock_prof_rejected;
222 static int lock_prof_skipspin;
223 static int lock_prof_skipcount;
225 #ifndef USE_CPU_NANOSECONDS
233 /* From bintime2timespec */
234 ns = bt.sec * (uint64_t)1000000000;
235 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
241 lock_prof_init_type(struct lock_prof_type *type)
245 SLIST_INIT(&type->lpt_lpalloc);
246 LIST_INIT(&type->lpt_lpoalloc);
247 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
248 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
250 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
256 lock_prof_init(void *arg)
260 for (cpu = 0; cpu <= mp_maxid; cpu++) {
261 lp_cpu[cpu] = malloc(sizeof(*lp_cpu[cpu]), M_DEVBUF,
263 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[0]);
264 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[1]);
267 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
270 lock_prof_reset_wait(void)
274 * Spin relinquishing our cpu so that quiesce_all_cpus may
277 while (lock_prof_resetting)
278 sched_relinquish(curthread);
282 lock_prof_reset(void)
284 struct lock_prof_cpu *lpc;
288 * We not only race with acquiring and releasing locks but also
289 * thread exit. To be certain that threads exit without valid head
290 * pointers they must see resetting set before enabled is cleared.
291 * Otherwise a lock may not be removed from a per-thread list due
292 * to disabled being set but not wait for reset() to remove it below.
294 atomic_store_rel_int(&lock_prof_resetting, 1);
295 enabled = lock_prof_enable;
296 lock_prof_enable = 0;
297 quiesce_all_cpus("profreset", 0);
299 * Some objects may have migrated between CPUs. Clear all links
300 * before we zero the structures. Some items may still be linked
301 * into per-thread lists as well.
303 for (cpu = 0; cpu <= mp_maxid; cpu++) {
305 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
306 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
307 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
310 for (cpu = 0; cpu <= mp_maxid; cpu++) {
312 bzero(lpc, sizeof(*lpc));
313 lock_prof_init_type(&lpc->lpc_types[0]);
314 lock_prof_init_type(&lpc->lpc_types[1]);
316 atomic_store_rel_int(&lock_prof_resetting, 0);
317 lock_prof_enable = enabled;
321 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
325 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
327 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
328 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
329 lp->cnt_wait / 1000, lp->cnt_cur,
330 lp->cnt_cur == 0 ? (uintmax_t)0 :
331 lp->cnt_tot / (lp->cnt_cur * 1000),
332 lp->cnt_cur == 0 ? (uintmax_t)0 :
333 lp->cnt_wait / (lp->cnt_cur * 1000),
334 (uintmax_t)0, lp->cnt_contest_locking,
335 p, lp->line, lp->class->lc_name, lp->name);
339 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
342 struct lock_prof_type *type;
346 dst->file = match->file;
347 dst->line = match->line;
348 dst->class = match->class;
349 dst->name = match->name;
351 for (cpu = 0; cpu <= mp_maxid; cpu++) {
352 if (lp_cpu[cpu] == NULL)
354 type = &lp_cpu[cpu]->lpc_types[spin];
355 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
358 if (l->file != match->file || l->line != match->line ||
359 l->name != match->name)
362 if (l->cnt_max > dst->cnt_max)
363 dst->cnt_max = l->cnt_max;
364 if (l->cnt_wait_max > dst->cnt_wait_max)
365 dst->cnt_wait_max = l->cnt_wait_max;
366 dst->cnt_tot += l->cnt_tot;
367 dst->cnt_wait += l->cnt_wait;
368 dst->cnt_cur += l->cnt_cur;
369 dst->cnt_contest_locking += l->cnt_contest_locking;
376 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
382 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
383 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
384 struct lock_prof lp = {};
388 lock_prof_sum(l, &lp, i, spin, t);
389 lock_prof_output(&lp, sb);
395 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
401 error = sysctl_wire_old_buffer(req, 0);
404 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
405 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
406 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
407 enabled = lock_prof_enable;
408 lock_prof_enable = 0;
409 quiesce_all_cpus("profstat", 0);
411 for (cpu = 0; cpu <= mp_maxid; cpu++) {
412 if (lp_cpu[cpu] == NULL)
414 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[0], sb, 0, t);
415 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[1], sb, 1, t);
417 lock_prof_enable = enabled;
419 error = sbuf_finish(sb);
420 /* Output a trailing NUL. */
422 error = SYSCTL_OUT(req, "", 1);
428 enable_lock_prof(SYSCTL_HANDLER_ARGS)
432 v = lock_prof_enable;
433 error = sysctl_handle_int(oidp, &v, v, req);
436 if (req->newptr == NULL)
438 if (v == lock_prof_enable)
442 lock_prof_enable = !!v;
448 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
453 error = sysctl_handle_int(oidp, &v, 0, req);
456 if (req->newptr == NULL)
465 static struct lock_prof *
466 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
469 const char *unknown = "(unknown)";
470 struct lock_prof_type *type;
471 struct lock_prof *lp;
477 if (p == NULL || *p == '\0')
479 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
480 hash &= LPROF_HASH_MASK;
481 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
482 head = &type->lpt_hash[hash];
483 SLIST_FOREACH(lp, head, link) {
484 if (lp->line == line && lp->file == p &&
485 lp->name == lo->lo_name)
489 lp = SLIST_FIRST(&type->lpt_lpalloc);
491 lock_prof_rejected++;
494 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
497 lp->class = LOCK_CLASS(lo);
498 lp->name = lo->lo_name;
499 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
503 static struct lock_profile_object *
504 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
507 struct lock_profile_object *l;
508 struct lock_prof_type *type;
509 struct lpohead *head;
511 head = &curthread->td_lprof[spin];
512 LIST_FOREACH(l, head, lpo_link)
513 if (l->lpo_obj == lo && l->lpo_file == file &&
516 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
517 l = LIST_FIRST(&type->lpt_lpoalloc);
519 lock_prof_rejected++;
522 LIST_REMOVE(l, lpo_link);
527 LIST_INSERT_HEAD(head, l, lpo_link);
533 lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
534 uint64_t waittime, const char *file, int line)
536 static int lock_prof_count;
537 struct lock_profile_object *l;
540 if (SCHEDULER_STOPPED())
543 /* don't reset the timer when/if recursing */
544 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
546 if (lock_prof_skipcount &&
547 (++lock_prof_count % lock_prof_skipcount) != 0)
549 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
550 if (spin && lock_prof_skipspin == 1)
553 /* Recheck enabled now that we're in a critical section. */
554 if (lock_prof_enable == 0)
556 l = lock_profile_object_lookup(lo, spin, file, line);
560 if (++l->lpo_ref > 1)
562 l->lpo_contest_locking = contested;
563 l->lpo_acqtime = nanoseconds();
564 if (waittime && (l->lpo_acqtime > waittime))
565 l->lpo_waittime = l->lpo_acqtime - waittime;
573 lock_profile_thread_exit(struct thread *td)
576 struct lock_profile_object *l;
578 MPASS(curthread->td_critnest == 0);
581 * If lock profiling was disabled we have to wait for reset to
582 * clear our pointers before we can exit safely.
584 lock_prof_reset_wait();
586 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
587 printf("thread still holds lock acquired at %s:%d\n",
588 l->lpo_file, l->lpo_line);
589 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
590 printf("thread still holds lock acquired at %s:%d\n",
591 l->lpo_file, l->lpo_line);
593 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
594 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
598 lock_profile_release_lock(struct lock_object *lo)
600 struct lock_profile_object *l;
601 struct lock_prof_type *type;
602 struct lock_prof *lp;
603 uint64_t curtime, holdtime;
604 struct lpohead *head;
607 if (SCHEDULER_STOPPED())
609 if (lo->lo_flags & LO_NOPROFILE)
611 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
612 head = &curthread->td_lprof[spin];
613 if (LIST_FIRST(head) == NULL)
616 /* Recheck enabled now that we're in a critical section. */
617 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
620 * If lock profiling is not enabled we still want to remove the
621 * lpo from our queue.
623 LIST_FOREACH(l, head, lpo_link)
624 if (l->lpo_obj == lo)
628 if (--l->lpo_ref > 0)
630 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
633 curtime = nanoseconds();
634 if (curtime < l->lpo_acqtime)
636 holdtime = curtime - l->lpo_acqtime;
639 * Record if the lock has been held longer now than ever
642 if (holdtime > lp->cnt_max)
643 lp->cnt_max = holdtime;
644 if (l->lpo_waittime > lp->cnt_wait_max)
645 lp->cnt_wait_max = l->lpo_waittime;
646 lp->cnt_tot += holdtime;
647 lp->cnt_wait += l->lpo_waittime;
648 lp->cnt_contest_locking += l->lpo_contest_locking;
649 lp->cnt_cur += l->lpo_cnt;
651 LIST_REMOVE(l, lpo_link);
652 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
653 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
658 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD, NULL, "lock debugging");
659 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL,
661 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
662 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
663 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
664 &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
665 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
666 &lock_prof_rejected, 0, "Number of rejected profiling records");
667 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
668 NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
669 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
670 NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
671 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
672 NULL, 0, enable_lock_prof, "I", "Enable lock profiling");