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.
13 * 3. Neither the name of the author nor the names of any co-contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * This module holds the global variables and functions used to maintain
32 * lock_object structures.
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
39 #include "opt_mprof.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
46 #include <sys/lock_profile.h>
47 #include <sys/malloc.h>
48 #include <sys/mutex.h>
52 #include <sys/sched.h>
54 #include <sys/sysctl.h>
60 #include <machine/cpufunc.h>
62 CTASSERT(LOCK_CLASS_MAX == 15);
64 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
66 &lock_class_mtx_sleep,
69 &lock_class_rm_sleepable,
75 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
76 const char *type, int flags)
80 /* Check for double-init and zero object. */
81 KASSERT(!lock_initialized(lock), ("lock \"%s\" %p already initialized",
84 /* Look up lock class to find its index. */
85 for (i = 0; i < LOCK_CLASS_MAX; i++)
86 if (lock_classes[i] == class) {
87 lock->lo_flags = i << LO_CLASSSHIFT;
90 KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
92 /* Initialize the lock object. */
94 lock->lo_flags |= flags | LO_INITIALIZED;
95 LOCK_LOG_INIT(lock, 0);
96 WITNESS_INIT(lock, (type != NULL) ? type : name);
100 lock_destroy(struct lock_object *lock)
103 KASSERT(lock_initialized(lock), ("lock %p is not initialized", lock));
104 WITNESS_DESTROY(lock);
105 LOCK_LOG_DESTROY(lock, 0);
106 lock->lo_flags &= ~LO_INITIALIZED;
110 DB_SHOW_COMMAND(lock, db_show_lock)
112 struct lock_object *lock;
113 struct lock_class *class;
117 lock = (struct lock_object *)addr;
118 if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
119 db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
122 class = LOCK_CLASS(lock);
123 db_printf(" class: %s\n", class->lc_name);
124 db_printf(" name: %s\n", lock->lo_name);
125 class->lc_ddb_show(lock);
129 #ifdef LOCK_PROFILING
132 * One object per-thread for each lock the thread owns. Tracks individual
135 struct lock_profile_object {
136 LIST_ENTRY(lock_profile_object) lpo_link;
137 struct lock_object *lpo_obj;
138 const char *lpo_file;
142 uint64_t lpo_acqtime;
143 uint64_t lpo_waittime;
144 u_int lpo_contest_locking;
148 * One lock_prof for each (file, line, lock object) triple.
151 SLIST_ENTRY(lock_prof) link;
152 struct lock_class *class;
157 uintmax_t cnt_wait_max;
162 uintmax_t cnt_contest_locking;
165 SLIST_HEAD(lphead, lock_prof);
167 #define LPROF_HASH_SIZE 4096
168 #define LPROF_HASH_MASK (LPROF_HASH_SIZE - 1)
169 #define LPROF_CACHE_SIZE 4096
172 * Array of objects and profs for each type of object for each cpu. Spinlocks
173 * are handled separately because a thread may be preempted and acquire a
174 * spinlock while in the lock profiling code of a non-spinlock. In this way
175 * we only need a critical section to protect the per-cpu lists.
177 struct lock_prof_type {
178 struct lphead lpt_lpalloc;
179 struct lpohead lpt_lpoalloc;
180 struct lphead lpt_hash[LPROF_HASH_SIZE];
181 struct lock_prof lpt_prof[LPROF_CACHE_SIZE];
182 struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
185 struct lock_prof_cpu {
186 struct lock_prof_type lpc_types[2]; /* One for spin one for other. */
189 struct lock_prof_cpu *lp_cpu[MAXCPU];
191 volatile int lock_prof_enable = 0;
192 static volatile int lock_prof_resetting;
194 #define LPROF_SBUF_SIZE 256
196 static int lock_prof_rejected;
197 static int lock_prof_skipspin;
198 static int lock_prof_skipcount;
200 #ifndef USE_CPU_NANOSECONDS
208 /* From bintime2timespec */
209 ns = bt.sec * (uint64_t)1000000000;
210 ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
216 lock_prof_init_type(struct lock_prof_type *type)
220 SLIST_INIT(&type->lpt_lpalloc);
221 LIST_INIT(&type->lpt_lpoalloc);
222 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
223 SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
225 LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
231 lock_prof_init(void *arg)
235 for (cpu = 0; cpu <= mp_maxid; cpu++) {
236 lp_cpu[cpu] = malloc(sizeof(*lp_cpu[cpu]), M_DEVBUF,
238 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[0]);
239 lock_prof_init_type(&lp_cpu[cpu]->lpc_types[1]);
242 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
245 lock_prof_reset_wait(void)
249 * Spin relinquishing our cpu so that quiesce_all_cpus may
252 while (lock_prof_resetting)
253 sched_relinquish(curthread);
257 lock_prof_reset(void)
259 struct lock_prof_cpu *lpc;
263 * We not only race with acquiring and releasing locks but also
264 * thread exit. To be certain that threads exit without valid head
265 * pointers they must see resetting set before enabled is cleared.
266 * Otherwise a lock may not be removed from a per-thread list due
267 * to disabled being set but not wait for reset() to remove it below.
269 atomic_store_rel_int(&lock_prof_resetting, 1);
270 enabled = lock_prof_enable;
271 lock_prof_enable = 0;
272 quiesce_all_cpus("profreset", 0);
274 * Some objects may have migrated between CPUs. Clear all links
275 * before we zero the structures. Some items may still be linked
276 * into per-thread lists as well.
278 for (cpu = 0; cpu <= mp_maxid; cpu++) {
280 for (i = 0; i < LPROF_CACHE_SIZE; i++) {
281 LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
282 LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
285 for (cpu = 0; cpu <= mp_maxid; cpu++) {
287 bzero(lpc, sizeof(*lpc));
288 lock_prof_init_type(&lpc->lpc_types[0]);
289 lock_prof_init_type(&lpc->lpc_types[1]);
291 atomic_store_rel_int(&lock_prof_resetting, 0);
292 lock_prof_enable = enabled;
296 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
300 for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
302 "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
303 lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
304 lp->cnt_wait / 1000, lp->cnt_cur,
305 lp->cnt_cur == 0 ? (uintmax_t)0 :
306 lp->cnt_tot / (lp->cnt_cur * 1000),
307 lp->cnt_cur == 0 ? (uintmax_t)0 :
308 lp->cnt_wait / (lp->cnt_cur * 1000),
309 (uintmax_t)0, lp->cnt_contest_locking,
310 p, lp->line, lp->class->lc_name, lp->name);
314 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
317 struct lock_prof_type *type;
321 dst->file = match->file;
322 dst->line = match->line;
323 dst->class = match->class;
324 dst->name = match->name;
326 for (cpu = 0; cpu <= mp_maxid; cpu++) {
327 if (lp_cpu[cpu] == NULL)
329 type = &lp_cpu[cpu]->lpc_types[spin];
330 SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
333 if (l->file != match->file || l->line != match->line ||
334 l->name != match->name)
337 if (l->cnt_max > dst->cnt_max)
338 dst->cnt_max = l->cnt_max;
339 if (l->cnt_wait_max > dst->cnt_wait_max)
340 dst->cnt_wait_max = l->cnt_wait_max;
341 dst->cnt_tot += l->cnt_tot;
342 dst->cnt_wait += l->cnt_wait;
343 dst->cnt_cur += l->cnt_cur;
344 dst->cnt_contest_locking += l->cnt_contest_locking;
351 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
357 for (i = 0; i < LPROF_HASH_SIZE; ++i) {
358 SLIST_FOREACH(l, &type->lpt_hash[i], link) {
359 struct lock_prof lp = {};
363 lock_prof_sum(l, &lp, i, spin, t);
364 lock_prof_output(&lp, sb);
370 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
376 error = sysctl_wire_old_buffer(req, 0);
379 sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
380 sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
381 "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
382 enabled = lock_prof_enable;
383 lock_prof_enable = 0;
384 quiesce_all_cpus("profstat", 0);
386 for (cpu = 0; cpu <= mp_maxid; cpu++) {
387 if (lp_cpu[cpu] == NULL)
389 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[0], sb, 0, t);
390 lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[1], sb, 1, t);
392 lock_prof_enable = enabled;
394 error = sbuf_finish(sb);
395 /* Output a trailing NUL. */
397 error = SYSCTL_OUT(req, "", 1);
403 enable_lock_prof(SYSCTL_HANDLER_ARGS)
407 v = lock_prof_enable;
408 error = sysctl_handle_int(oidp, &v, v, req);
411 if (req->newptr == NULL)
413 if (v == lock_prof_enable)
417 lock_prof_enable = !!v;
423 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
428 error = sysctl_handle_int(oidp, &v, 0, req);
431 if (req->newptr == NULL)
440 static struct lock_prof *
441 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
444 const char *unknown = "(unknown)";
445 struct lock_prof_type *type;
446 struct lock_prof *lp;
452 if (p == NULL || *p == '\0')
454 hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
455 hash &= LPROF_HASH_MASK;
456 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
457 head = &type->lpt_hash[hash];
458 SLIST_FOREACH(lp, head, link) {
459 if (lp->line == line && lp->file == p &&
460 lp->name == lo->lo_name)
464 lp = SLIST_FIRST(&type->lpt_lpalloc);
466 lock_prof_rejected++;
469 SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
472 lp->class = LOCK_CLASS(lo);
473 lp->name = lo->lo_name;
474 SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
478 static struct lock_profile_object *
479 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
482 struct lock_profile_object *l;
483 struct lock_prof_type *type;
484 struct lpohead *head;
486 head = &curthread->td_lprof[spin];
487 LIST_FOREACH(l, head, lpo_link)
488 if (l->lpo_obj == lo && l->lpo_file == file &&
491 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
492 l = LIST_FIRST(&type->lpt_lpoalloc);
494 lock_prof_rejected++;
497 LIST_REMOVE(l, lpo_link);
502 LIST_INSERT_HEAD(head, l, lpo_link);
508 lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
509 uint64_t waittime, const char *file, int line)
511 static int lock_prof_count;
512 struct lock_profile_object *l;
515 if (SCHEDULER_STOPPED())
518 /* don't reset the timer when/if recursing */
519 if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
521 if (lock_prof_skipcount &&
522 (++lock_prof_count % lock_prof_skipcount) != 0)
524 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
525 if (spin && lock_prof_skipspin == 1)
528 /* Recheck enabled now that we're in a critical section. */
529 if (lock_prof_enable == 0)
531 l = lock_profile_object_lookup(lo, spin, file, line);
535 if (++l->lpo_ref > 1)
537 l->lpo_contest_locking = contested;
538 l->lpo_acqtime = nanoseconds();
539 if (waittime && (l->lpo_acqtime > waittime))
540 l->lpo_waittime = l->lpo_acqtime - waittime;
548 lock_profile_thread_exit(struct thread *td)
551 struct lock_profile_object *l;
553 MPASS(curthread->td_critnest == 0);
556 * If lock profiling was disabled we have to wait for reset to
557 * clear our pointers before we can exit safely.
559 lock_prof_reset_wait();
561 LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
562 printf("thread still holds lock acquired at %s:%d\n",
563 l->lpo_file, l->lpo_line);
564 LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
565 printf("thread still holds lock acquired at %s:%d\n",
566 l->lpo_file, l->lpo_line);
568 MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
569 MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
573 lock_profile_release_lock(struct lock_object *lo)
575 struct lock_profile_object *l;
576 struct lock_prof_type *type;
577 struct lock_prof *lp;
578 uint64_t curtime, holdtime;
579 struct lpohead *head;
582 if (SCHEDULER_STOPPED())
584 if (lo->lo_flags & LO_NOPROFILE)
586 spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
587 head = &curthread->td_lprof[spin];
588 if (LIST_FIRST(head) == NULL)
591 /* Recheck enabled now that we're in a critical section. */
592 if (lock_prof_enable == 0 && lock_prof_resetting == 1)
595 * If lock profiling is not enabled we still want to remove the
596 * lpo from our queue.
598 LIST_FOREACH(l, head, lpo_link)
599 if (l->lpo_obj == lo)
603 if (--l->lpo_ref > 0)
605 lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
608 curtime = nanoseconds();
609 if (curtime < l->lpo_acqtime)
611 holdtime = curtime - l->lpo_acqtime;
614 * Record if the lock has been held longer now than ever
617 if (holdtime > lp->cnt_max)
618 lp->cnt_max = holdtime;
619 if (l->lpo_waittime > lp->cnt_wait_max)
620 lp->cnt_wait_max = l->lpo_waittime;
621 lp->cnt_tot += holdtime;
622 lp->cnt_wait += l->lpo_waittime;
623 lp->cnt_contest_locking += l->lpo_contest_locking;
624 lp->cnt_cur += l->lpo_cnt;
626 LIST_REMOVE(l, lpo_link);
627 type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
628 LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
633 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD, NULL, "lock debugging");
634 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL,
636 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
637 &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
638 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
639 &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
640 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
641 &lock_prof_rejected, 0, "Number of rejected profiling records");
642 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
643 NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
644 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
645 NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
646 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
647 NULL, 0, enable_lock_prof, "I", "Enable lock profiling");