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(!lock_initalized(lock), ("lock \"%s\" %p already initialized",
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_initalized(lock), ("lock %p is not initialized", lock));
101 WITNESS_DESTROY(lock);
102 LOCK_LOG_DESTROY(lock, 0);
103 lock->lo_flags &= ~LO_INITIALIZED;
107 DB_SHOW_COMMAND(lock, db_show_lock)
109 struct lock_object *lock;
110 struct lock_class *class;
114 lock = (struct lock_object *)addr;
115 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
116 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
119 class = LOCK_CLASS(lock);
120 db_printf(" class: %s\n", class->lc_name);
121 db_printf(" name: %s\n", lock->lo_name);
122 class->lc_ddb_show(lock);
126 #ifdef LOCK_PROFILING
129 * One object per-thread for each lock the thread owns. Tracks individual
132 struct lock_profile_object {
133 LIST_ENTRY(lock_profile_object) lpo_link;
134 struct lock_object *lpo_obj;
135 const char *lpo_file;
139 uint64_t lpo_acqtime;
140 uint64_t lpo_waittime;
141 u_int lpo_contest_locking;
145 * One lock_prof for each (file, line, lock object) triple.
148 SLIST_ENTRY(lock_prof) link;
149 struct lock_class *class;
154 uintmax_t cnt_wait_max;
159 uintmax_t cnt_contest_locking;
162 SLIST_HEAD(lphead, lock_prof);
164 #define LPROF_HASH_SIZE 4096
165 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
166 #define LPROF_CACHE_SIZE 4096
169 * Array of objects and profs for each type of object for each cpu. Spinlocks
170 * are handled separately because a thread may be preempted and acquire a
171 * spinlock while in the lock profiling code of a non-spinlock. In this way
172 * we only need a critical section to protect the per-cpu lists.
174 struct lock_prof_type {
175 struct lphead lpt_lpalloc;
176 struct lpohead lpt_lpoalloc;
177 struct lphead lpt_hash[LPROF_HASH_SIZE];
178 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
179 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
182 struct lock_prof_cpu {
183 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
186 struct lock_prof_cpu *lp_cpu[MAXCPU];
188 volatile int lock_prof_enable = 0;
189 static volatile int lock_prof_resetting;
191 #define LPROF_SBUF_SIZE 256
193 static int lock_prof_rejected;
194 static int lock_prof_skipspin;
195 static int lock_prof_skipcount;
197 #ifndef USE_CPU_NANOSECONDS
205 /* From bintime2timespec */
206 ns = bt.sec * (uint64_t)1000000000;
207 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
213 lock_prof_init_type(struct lock_prof_type *type)
217 SLIST_INIT(&type->lpt_lpalloc);
218 LIST_INIT(&type->lpt_lpoalloc);
219 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
220 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
222 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
228 lock_prof_init(void *arg)
232 for (cpu = 0; cpu <= mp_maxid; cpu++) {
233 lp_cpu[cpu] = malloc(sizeof(*lp_cpu[cpu]), M_DEVBUF,
235 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[0]);
236 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[1]);
239 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
242 * To be certain that lock profiling has idled on all cpus before we
243 * reset, we schedule the resetting thread on all active cpus. Since
244 * all operations happen within critical sections we can be sure that
245 * it is safe to zero the profiling structures.
263 lock_prof_reset_wait(void)
267 * Spin relinquishing our cpu so that lock_prof_idle may
270 while (lock_prof_resetting)
271 sched_relinquish(curthread);
275 lock_prof_reset(void)
277 struct lock_prof_cpu *lpc;
281 * We not only race with acquiring and releasing locks but also
282 * thread exit. To be certain that threads exit without valid head
283 * pointers they must see resetting set before enabled is cleared.
284 * Otherwise a lock may not be removed from a per-thread list due
285 * to disabled being set but not wait for reset() to remove it below.
287 atomic_store_rel_int(&lock_prof_resetting, 1);
288 enabled = lock_prof_enable;
289 lock_prof_enable = 0;
292 * Some objects may have migrated between CPUs. Clear all links
293 * before we zero the structures. Some items may still be linked
294 * into per-thread lists as well.
296 for (cpu = 0; cpu <= mp_maxid; cpu++) {
298 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
299 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
300 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
303 for (cpu = 0; cpu <= mp_maxid; cpu++) {
305 bzero(lpc, sizeof(*lpc));
306 lock_prof_init_type(&lpc->lpc_types[0]);
307 lock_prof_init_type(&lpc->lpc_types[1]);
309 atomic_store_rel_int(&lock_prof_resetting, 0);
310 lock_prof_enable = enabled;
314 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
318 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
320 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
321 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
322 lp->cnt_wait / 1000, lp->cnt_cur,
323 lp->cnt_cur == 0 ? (uintmax_t)0 :
324 lp->cnt_tot / (lp->cnt_cur * 1000),
325 lp->cnt_cur == 0 ? (uintmax_t)0 :
326 lp->cnt_wait / (lp->cnt_cur * 1000),
327 (uintmax_t)0, lp->cnt_contest_locking,
328 p, lp->line, lp->class->lc_name, lp->name);
332 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
335 struct lock_prof_type *type;
339 dst->file = match->file;
340 dst->line = match->line;
341 dst->class = match->class;
342 dst->name = match->name;
344 for (cpu = 0; cpu <= mp_maxid; cpu++) {
345 if (lp_cpu[cpu] == NULL)
347 type = &lp_cpu[cpu]->lpc_types[spin];
348 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
351 if (l->file != match->file || l->line != match->line ||
352 l->name != match->name)
355 if (l->cnt_max > dst->cnt_max)
356 dst->cnt_max = l->cnt_max;
357 if (l->cnt_wait_max > dst->cnt_wait_max)
358 dst->cnt_wait_max = l->cnt_wait_max;
359 dst->cnt_tot += l->cnt_tot;
360 dst->cnt_wait += l->cnt_wait;
361 dst->cnt_cur += l->cnt_cur;
362 dst->cnt_contest_locking += l->cnt_contest_locking;
369 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
375 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
376 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
377 struct lock_prof lp = {};
381 lock_prof_sum(l, &lp, i, spin, t);
382 lock_prof_output(&lp, sb);
388 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
394 error = sysctl_wire_old_buffer(req, 0);
397 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
398 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
399 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
400 enabled = lock_prof_enable;
401 lock_prof_enable = 0;
404 for (cpu = 0; cpu <= mp_maxid; cpu++) {
405 if (lp_cpu[cpu] == NULL)
407 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[0], sb, 0, t);
408 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[1], sb, 1, t);
410 lock_prof_enable = enabled;
412 error = sbuf_finish(sb);
413 /* Output a trailing NUL. */
415 error = SYSCTL_OUT(req, "", 1);
421 enable_lock_prof(SYSCTL_HANDLER_ARGS)
425 v = lock_prof_enable;
426 error = sysctl_handle_int(oidp, &v, v, req);
429 if (req->newptr == NULL)
431 if (v == lock_prof_enable)
435 lock_prof_enable = !!v;
441 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
446 error = sysctl_handle_int(oidp, &v, 0, req);
449 if (req->newptr == NULL)
458 static struct lock_prof *
459 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
462 const char *unknown = "(unknown)";
463 struct lock_prof_type *type;
464 struct lock_prof *lp;
470 if (p == NULL || *p == '\0')
472 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
473 hash &= LPROF_HASH_MASK;
474 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
475 head = &type->lpt_hash[hash];
476 SLIST_FOREACH(lp, head, link) {
477 if (lp->line == line && lp->file == p &&
478 lp->name == lo->lo_name)
482 lp = SLIST_FIRST(&type->lpt_lpalloc);
484 lock_prof_rejected++;
487 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
490 lp->class = LOCK_CLASS(lo);
491 lp->name = lo->lo_name;
492 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
496 static struct lock_profile_object *
497 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
500 struct lock_profile_object *l;
501 struct lock_prof_type *type;
502 struct lpohead *head;
504 head = &curthread->td_lprof[spin];
505 LIST_FOREACH(l, head, lpo_link)
506 if (l->lpo_obj == lo && l->lpo_file == file &&
509 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
510 l = LIST_FIRST(&type->lpt_lpoalloc);
512 lock_prof_rejected++;
515 LIST_REMOVE(l, lpo_link);
520 LIST_INSERT_HEAD(head, l, lpo_link);
526 lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
527 uint64_t waittime, const char *file, int line)
529 static int lock_prof_count;
530 struct lock_profile_object *l;
533 if (SCHEDULER_STOPPED())
536 /* don't reset the timer when/if recursing */
537 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
539 if (lock_prof_skipcount &&
540 (++lock_prof_count % lock_prof_skipcount) != 0)
542 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
543 if (spin && lock_prof_skipspin == 1)
546 /* Recheck enabled now that we're in a critical section. */
547 if (lock_prof_enable == 0)
549 l = lock_profile_object_lookup(lo, spin, file, line);
553 if (++l->lpo_ref > 1)
555 l->lpo_contest_locking = contested;
556 l->lpo_acqtime = nanoseconds();
557 if (waittime && (l->lpo_acqtime > waittime))
558 l->lpo_waittime = l->lpo_acqtime - waittime;
566 lock_profile_thread_exit(struct thread *td)
569 struct lock_profile_object *l;
571 MPASS(curthread->td_critnest == 0);
574 * If lock profiling was disabled we have to wait for reset to
575 * clear our pointers before we can exit safely.
577 lock_prof_reset_wait();
579 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
580 printf("thread still holds lock acquired at %s:%d\n",
581 l->lpo_file, l->lpo_line);
582 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
583 printf("thread still holds lock acquired at %s:%d\n",
584 l->lpo_file, l->lpo_line);
586 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
587 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
591 lock_profile_release_lock(struct lock_object *lo)
593 struct lock_profile_object *l;
594 struct lock_prof_type *type;
595 struct lock_prof *lp;
596 uint64_t curtime, holdtime;
597 struct lpohead *head;
600 if (SCHEDULER_STOPPED())
602 if (lo->lo_flags & LO_NOPROFILE)
604 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
605 head = &curthread->td_lprof[spin];
606 if (LIST_FIRST(head) == NULL)
609 /* Recheck enabled now that we're in a critical section. */
610 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
613 * If lock profiling is not enabled we still want to remove the
614 * lpo from our queue.
616 LIST_FOREACH(l, head, lpo_link)
617 if (l->lpo_obj == lo)
621 if (--l->lpo_ref > 0)
623 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
626 curtime = nanoseconds();
627 if (curtime < l->lpo_acqtime)
629 holdtime = curtime - l->lpo_acqtime;
632 * Record if the lock has been held longer now than ever
635 if (holdtime > lp->cnt_max)
636 lp->cnt_max = holdtime;
637 if (l->lpo_waittime > lp->cnt_wait_max)
638 lp->cnt_wait_max = l->lpo_waittime;
639 lp->cnt_tot += holdtime;
640 lp->cnt_wait += l->lpo_waittime;
641 lp->cnt_contest_locking += l->lpo_contest_locking;
642 lp->cnt_cur += l->lpo_cnt;
644 LIST_REMOVE(l, lpo_link);
645 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
646 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
651 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD, NULL, "lock debugging");
652 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL,
654 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
655 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
656 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
657 &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
658 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
659 &lock_prof_rejected, 0, "Number of rejected profiling records");
660 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
661 NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
662 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
663 NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
664 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
665 NULL, 0, enable_lock_prof, "I", "Enable lock profiling");