xref: /freebsd/sys/kern/subr_lock.c (revision 3cbb4cc200f8a0ad7ed08233425ea54524a21f1c)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
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.
14  *
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
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * This module holds the global variables and functions used to maintain
30  * lock_object structures.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include "opt_ddb.h"
37 #include "opt_mprof.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/ktr.h>
43 #include <sys/lock.h>
44 #include <sys/lock_profile.h>
45 #include <sys/malloc.h>
46 #include <sys/mutex.h>
47 #include <sys/pcpu.h>
48 #include <sys/proc.h>
49 #include <sys/sbuf.h>
50 #include <sys/sched.h>
51 #include <sys/smp.h>
52 #include <sys/sysctl.h>
53 
54 #ifdef DDB
55 #include <ddb/ddb.h>
56 #endif
57 
58 #include <machine/cpufunc.h>
59 
60 SDT_PROVIDER_DEFINE(lock);
61 SDT_PROBE_DEFINE1(lock, , , starvation, "u_int");
62 
63 CTASSERT(LOCK_CLASS_MAX == 15);
64 
65 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
66 	&lock_class_mtx_spin,
67 	&lock_class_mtx_sleep,
68 	&lock_class_sx,
69 	&lock_class_rm,
70 	&lock_class_rm_sleepable,
71 	&lock_class_rw,
72 	&lock_class_lockmgr,
73 };
74 
75 void
76 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
77     const char *type, int flags)
78 {
79 	int i;
80 
81 	/* Check for double-init and zero object. */
82 	KASSERT(flags & LO_NEW || !lock_initialized(lock),
83 	    ("lock \"%s\" %p already initialized", name, lock));
84 
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;
89 			break;
90 		}
91 	KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
92 
93 	/* Initialize the lock object. */
94 	lock->lo_name = name;
95 	lock->lo_flags |= flags | LO_INITIALIZED;
96 	LOCK_LOG_INIT(lock, 0);
97 	WITNESS_INIT(lock, (type != NULL) ? type : name);
98 }
99 
100 void
101 lock_destroy(struct lock_object *lock)
102 {
103 
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;
108 }
109 
110 static SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
111     "lock debugging");
112 static SYSCTL_NODE(_debug_lock, OID_AUTO, delay,
113     CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
114     "lock delay");
115 
116 static u_int __read_mostly starvation_limit = 131072;
117 SYSCTL_INT(_debug_lock_delay, OID_AUTO, starvation_limit, CTLFLAG_RW,
118     &starvation_limit, 0, "");
119 
120 static u_int __read_mostly restrict_starvation = 0;
121 SYSCTL_INT(_debug_lock_delay, OID_AUTO, restrict_starvation, CTLFLAG_RW,
122     &restrict_starvation, 0, "");
123 
124 void
125 lock_delay(struct lock_delay_arg *la)
126 {
127 	struct lock_delay_config *lc = la->config;
128 	u_short i;
129 
130 	la->delay <<= 1;
131 	if (__predict_false(la->delay > lc->max))
132 		la->delay = lc->max;
133 
134 	for (i = la->delay; i > 0; i--)
135 		cpu_spinwait();
136 
137 	la->spin_cnt += la->delay;
138 	if (__predict_false(la->spin_cnt > starvation_limit)) {
139 		SDT_PROBE1(lock, , , starvation, la->delay);
140 		if (restrict_starvation)
141 			la->delay = lc->base;
142 	}
143 }
144 
145 static u_int
146 lock_roundup_2(u_int val)
147 {
148 	u_int res;
149 
150 	for (res = 1; res <= val; res <<= 1)
151 		continue;
152 
153 	return (res);
154 }
155 
156 void
157 lock_delay_default_init(struct lock_delay_config *lc)
158 {
159 
160 	lc->base = 1;
161 	lc->max = lock_roundup_2(mp_ncpus) * 256;
162 	if (lc->max > 32678)
163 		lc->max = 32678;
164 }
165 
166 struct lock_delay_config __read_frequently locks_delay;
167 u_short __read_frequently locks_delay_retries;
168 u_short __read_frequently locks_delay_loops;
169 
170 SYSCTL_U16(_debug_lock, OID_AUTO, delay_base, CTLFLAG_RW, &locks_delay.base,
171     0, "");
172 SYSCTL_U16(_debug_lock, OID_AUTO, delay_max, CTLFLAG_RW, &locks_delay.max,
173     0, "");
174 SYSCTL_U16(_debug_lock, OID_AUTO, delay_retries, CTLFLAG_RW, &locks_delay_retries,
175     0, "");
176 SYSCTL_U16(_debug_lock, OID_AUTO, delay_loops, CTLFLAG_RW, &locks_delay_loops,
177     0, "");
178 
179 static void
180 locks_delay_init(void *arg __unused)
181 {
182 
183 	lock_delay_default_init(&locks_delay);
184 	locks_delay_retries = 10;
185 	locks_delay_loops = max(10000, locks_delay.max);
186 }
187 LOCK_DELAY_SYSINIT(locks_delay_init);
188 
189 #ifdef DDB
190 DB_SHOW_COMMAND(lock, db_show_lock)
191 {
192 	struct lock_object *lock;
193 	struct lock_class *class;
194 
195 	if (!have_addr)
196 		return;
197 	lock = (struct lock_object *)addr;
198 	if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
199 		db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
200 		return;
201 	}
202 	class = LOCK_CLASS(lock);
203 	db_printf(" class: %s\n", class->lc_name);
204 	db_printf(" name: %s\n", lock->lo_name);
205 	class->lc_ddb_show(lock);
206 }
207 #endif
208 
209 #ifdef LOCK_PROFILING
210 
211 /*
212  * One object per-thread for each lock the thread owns.  Tracks individual
213  * lock instances.
214  */
215 struct lock_profile_object {
216 	LIST_ENTRY(lock_profile_object) lpo_link;
217 	struct lock_object *lpo_obj;
218 	const char	*lpo_file;
219 	int		lpo_line;
220 	uint16_t	lpo_ref;
221 	uint16_t	lpo_cnt;
222 	uint64_t	lpo_acqtime;
223 	uint64_t	lpo_waittime;
224 	u_int		lpo_contest_locking;
225 };
226 
227 /*
228  * One lock_prof for each (file, line, lock object) triple.
229  */
230 struct lock_prof {
231 	SLIST_ENTRY(lock_prof) link;
232 	struct lock_class *class;
233 	const char	*file;
234 	const char	*name;
235 	int		line;
236 	int		ticks;
237 	uintmax_t	cnt_wait_max;
238 	uintmax_t	cnt_max;
239 	uintmax_t	cnt_tot;
240 	uintmax_t	cnt_wait;
241 	uintmax_t	cnt_cur;
242 	uintmax_t	cnt_contest_locking;
243 };
244 
245 SLIST_HEAD(lphead, lock_prof);
246 
247 #define	LPROF_HASH_SIZE		4096
248 #define	LPROF_HASH_MASK		(LPROF_HASH_SIZE - 1)
249 #define	LPROF_CACHE_SIZE	4096
250 
251 /*
252  * Array of objects and profs for each type of object for each cpu.  Spinlocks
253  * are handled separately because a thread may be preempted and acquire a
254  * spinlock while in the lock profiling code of a non-spinlock.  In this way
255  * we only need a critical section to protect the per-cpu lists.
256  */
257 struct lock_prof_type {
258 	struct lphead		lpt_lpalloc;
259 	struct lpohead		lpt_lpoalloc;
260 	struct lphead		lpt_hash[LPROF_HASH_SIZE];
261 	struct lock_prof	lpt_prof[LPROF_CACHE_SIZE];
262 	struct lock_profile_object lpt_objs[LPROF_CACHE_SIZE];
263 };
264 
265 struct lock_prof_cpu {
266 	struct lock_prof_type	lpc_types[2]; /* One for spin one for other. */
267 };
268 
269 DPCPU_DEFINE_STATIC(struct lock_prof_cpu, lp);
270 #define	LP_CPU_SELF	(DPCPU_PTR(lp))
271 #define	LP_CPU(cpu)	(DPCPU_ID_PTR((cpu), lp))
272 
273 volatile int __read_mostly lock_prof_enable;
274 static volatile int lock_prof_resetting;
275 
276 #define LPROF_SBUF_SIZE		256
277 
278 static int lock_prof_rejected;
279 static int lock_prof_skipspin;
280 static int lock_prof_skipcount;
281 
282 #ifndef USE_CPU_NANOSECONDS
283 uint64_t
284 nanoseconds(void)
285 {
286 	struct bintime bt;
287 	uint64_t ns;
288 
289 	binuptime(&bt);
290 	/* From bintime2timespec */
291 	ns = bt.sec * (uint64_t)1000000000;
292 	ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
293 	return (ns);
294 }
295 #endif
296 
297 static void
298 lock_prof_init_type(struct lock_prof_type *type)
299 {
300 	int i;
301 
302 	SLIST_INIT(&type->lpt_lpalloc);
303 	LIST_INIT(&type->lpt_lpoalloc);
304 	for (i = 0; i < LPROF_CACHE_SIZE; i++) {
305 		SLIST_INSERT_HEAD(&type->lpt_lpalloc, &type->lpt_prof[i],
306 		    link);
307 		LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
308 		    lpo_link);
309 	}
310 }
311 
312 static void
313 lock_prof_init(void *arg)
314 {
315 	int cpu;
316 
317 	CPU_FOREACH(cpu) {
318 		lock_prof_init_type(&LP_CPU(cpu)->lpc_types[0]);
319 		lock_prof_init_type(&LP_CPU(cpu)->lpc_types[1]);
320 	}
321 }
322 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
323 
324 static void
325 lock_prof_reset_wait(void)
326 {
327 
328 	/*
329 	 * Spin relinquishing our cpu so that quiesce_all_cpus may
330 	 * complete.
331 	 */
332 	while (lock_prof_resetting)
333 		sched_relinquish(curthread);
334 }
335 
336 static void
337 lock_prof_reset(void)
338 {
339 	struct lock_prof_cpu *lpc;
340 	int enabled, i, cpu;
341 
342 	/*
343 	 * We not only race with acquiring and releasing locks but also
344 	 * thread exit.  To be certain that threads exit without valid head
345 	 * pointers they must see resetting set before enabled is cleared.
346 	 * Otherwise a lock may not be removed from a per-thread list due
347 	 * to disabled being set but not wait for reset() to remove it below.
348 	 */
349 	atomic_store_rel_int(&lock_prof_resetting, 1);
350 	enabled = lock_prof_enable;
351 	lock_prof_enable = 0;
352 	/*
353 	 * This both publishes lock_prof_enable as disabled and makes sure
354 	 * everyone else reads it if they are not far enough. We wait for the
355 	 * rest down below.
356 	 */
357 	cpus_fence_seq_cst();
358 	quiesce_all_critical();
359 	/*
360 	 * Some objects may have migrated between CPUs.  Clear all links
361 	 * before we zero the structures.  Some items may still be linked
362 	 * into per-thread lists as well.
363 	 */
364 	CPU_FOREACH(cpu) {
365 		lpc = LP_CPU(cpu);
366 		for (i = 0; i < LPROF_CACHE_SIZE; i++) {
367 			LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
368 			LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
369 		}
370 	}
371 	CPU_FOREACH(cpu) {
372 		lpc = LP_CPU(cpu);
373 		bzero(lpc, sizeof(*lpc));
374 		lock_prof_init_type(&lpc->lpc_types[0]);
375 		lock_prof_init_type(&lpc->lpc_types[1]);
376 	}
377 	/*
378 	 * Paired with the fence from cpus_fence_seq_cst()
379 	 */
380 	atomic_store_rel_int(&lock_prof_resetting, 0);
381 	lock_prof_enable = enabled;
382 }
383 
384 static void
385 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
386 {
387 	const char *p;
388 
389 	for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
390 	sbuf_printf(sb,
391 	    "%8ju %9ju %11ju %11ju %11ju %6ju %6ju %2ju %6ju %s:%d (%s:%s)\n",
392 	    lp->cnt_max / 1000, lp->cnt_wait_max / 1000, lp->cnt_tot / 1000,
393 	    lp->cnt_wait / 1000, lp->cnt_cur,
394 	    lp->cnt_cur == 0 ? (uintmax_t)0 :
395 	    lp->cnt_tot / (lp->cnt_cur * 1000),
396 	    lp->cnt_cur == 0 ? (uintmax_t)0 :
397 	    lp->cnt_wait / (lp->cnt_cur * 1000),
398 	    (uintmax_t)0, lp->cnt_contest_locking,
399 	    p, lp->line, lp->class->lc_name, lp->name);
400 }
401 
402 static void
403 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
404     int spin, int t)
405 {
406 	struct lock_prof_type *type;
407 	struct lock_prof *l;
408 	int cpu;
409 
410 	dst->file = match->file;
411 	dst->line = match->line;
412 	dst->class = match->class;
413 	dst->name = match->name;
414 
415 	CPU_FOREACH(cpu) {
416 		type = &LP_CPU(cpu)->lpc_types[spin];
417 		SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
418 			if (l->ticks == t)
419 				continue;
420 			if (l->file != match->file || l->line != match->line ||
421 			    l->name != match->name)
422 				continue;
423 			l->ticks = t;
424 			if (l->cnt_max > dst->cnt_max)
425 				dst->cnt_max = l->cnt_max;
426 			if (l->cnt_wait_max > dst->cnt_wait_max)
427 				dst->cnt_wait_max = l->cnt_wait_max;
428 			dst->cnt_tot += l->cnt_tot;
429 			dst->cnt_wait += l->cnt_wait;
430 			dst->cnt_cur += l->cnt_cur;
431 			dst->cnt_contest_locking += l->cnt_contest_locking;
432 		}
433 	}
434 }
435 
436 static void
437 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
438     int t)
439 {
440 	struct lock_prof *l;
441 	int i;
442 
443 	for (i = 0; i < LPROF_HASH_SIZE; ++i) {
444 		SLIST_FOREACH(l, &type->lpt_hash[i], link) {
445 			struct lock_prof lp = {};
446 
447 			if (l->ticks == t)
448 				continue;
449 			lock_prof_sum(l, &lp, i, spin, t);
450 			lock_prof_output(&lp, sb);
451 		}
452 	}
453 }
454 
455 static int
456 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
457 {
458 	struct sbuf *sb;
459 	int error, cpu, t;
460 	int enabled;
461 
462 	error = sysctl_wire_old_buffer(req, 0);
463 	if (error != 0)
464 		return (error);
465 	sb = sbuf_new_for_sysctl(NULL, NULL, LPROF_SBUF_SIZE, req);
466 	sbuf_printf(sb, "\n%8s %9s %11s %11s %11s %6s %6s %2s %6s %s\n",
467 	    "max", "wait_max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
468 	enabled = lock_prof_enable;
469 	lock_prof_enable = 0;
470 	/*
471 	 * See the comment in lock_prof_reset
472 	 */
473 	cpus_fence_seq_cst();
474 	quiesce_all_critical();
475 	t = ticks;
476 	CPU_FOREACH(cpu) {
477 		lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[0], sb, 0, t);
478 		lock_prof_type_stats(&LP_CPU(cpu)->lpc_types[1], sb, 1, t);
479 	}
480 	atomic_thread_fence_rel();
481 	lock_prof_enable = enabled;
482 
483 	error = sbuf_finish(sb);
484 	/* Output a trailing NUL. */
485 	if (error == 0)
486 		error = SYSCTL_OUT(req, "", 1);
487 	sbuf_delete(sb);
488 	return (error);
489 }
490 
491 static int
492 enable_lock_prof(SYSCTL_HANDLER_ARGS)
493 {
494 	int error, v;
495 
496 	v = lock_prof_enable;
497 	error = sysctl_handle_int(oidp, &v, v, req);
498 	if (error)
499 		return (error);
500 	if (req->newptr == NULL)
501 		return (error);
502 	if (v == lock_prof_enable)
503 		return (0);
504 	if (v == 1)
505 		lock_prof_reset();
506 	lock_prof_enable = !!v;
507 
508 	return (0);
509 }
510 
511 static int
512 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
513 {
514 	int error, v;
515 
516 	v = 0;
517 	error = sysctl_handle_int(oidp, &v, 0, req);
518 	if (error)
519 		return (error);
520 	if (req->newptr == NULL)
521 		return (error);
522 	if (v == 0)
523 		return (0);
524 	lock_prof_reset();
525 
526 	return (0);
527 }
528 
529 static struct lock_prof *
530 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
531     int line)
532 {
533 	const char *unknown = "(unknown)";
534 	struct lock_prof_type *type;
535 	struct lock_prof *lp;
536 	struct lphead *head;
537 	const char *p;
538 	u_int hash;
539 
540 	p = file;
541 	if (p == NULL || *p == '\0')
542 		p = unknown;
543 	hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
544 	hash &= LPROF_HASH_MASK;
545 	type = &LP_CPU_SELF->lpc_types[spin];
546 	head = &type->lpt_hash[hash];
547 	SLIST_FOREACH(lp, head, link) {
548 		if (lp->line == line && lp->file == p &&
549 		    lp->name == lo->lo_name)
550 			return (lp);
551 	}
552 	lp = SLIST_FIRST(&type->lpt_lpalloc);
553 	if (lp == NULL) {
554 		lock_prof_rejected++;
555 		return (lp);
556 	}
557 	SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
558 	lp->file = p;
559 	lp->line = line;
560 	lp->class = LOCK_CLASS(lo);
561 	lp->name = lo->lo_name;
562 	SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
563 	return (lp);
564 }
565 
566 static struct lock_profile_object *
567 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
568     int line)
569 {
570 	struct lock_profile_object *l;
571 	struct lock_prof_type *type;
572 	struct lpohead *head;
573 
574 	head = &curthread->td_lprof[spin];
575 	LIST_FOREACH(l, head, lpo_link)
576 		if (l->lpo_obj == lo && l->lpo_file == file &&
577 		    l->lpo_line == line)
578 			return (l);
579 	type = &LP_CPU_SELF->lpc_types[spin];
580 	l = LIST_FIRST(&type->lpt_lpoalloc);
581 	if (l == NULL) {
582 		lock_prof_rejected++;
583 		return (NULL);
584 	}
585 	LIST_REMOVE(l, lpo_link);
586 	l->lpo_obj = lo;
587 	l->lpo_file = file;
588 	l->lpo_line = line;
589 	l->lpo_cnt = 0;
590 	LIST_INSERT_HEAD(head, l, lpo_link);
591 
592 	return (l);
593 }
594 
595 void
596 lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
597     uint64_t waittime, const char *file, int line)
598 {
599 	static int lock_prof_count;
600 	struct lock_profile_object *l;
601 	int spin;
602 
603 	if (SCHEDULER_STOPPED())
604 		return;
605 
606 	/* don't reset the timer when/if recursing */
607 	if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
608 		return;
609 	if (lock_prof_skipcount &&
610 	    (++lock_prof_count % lock_prof_skipcount) != 0)
611 		return;
612 	spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
613 	if (spin && lock_prof_skipspin == 1)
614 		return;
615 	critical_enter();
616 	/* Recheck enabled now that we're in a critical section. */
617 	if (lock_prof_enable == 0)
618 		goto out;
619 	l = lock_profile_object_lookup(lo, spin, file, line);
620 	if (l == NULL)
621 		goto out;
622 	l->lpo_cnt++;
623 	if (++l->lpo_ref > 1)
624 		goto out;
625 	l->lpo_contest_locking = contested;
626 	l->lpo_acqtime = nanoseconds();
627 	if (waittime && (l->lpo_acqtime > waittime))
628 		l->lpo_waittime = l->lpo_acqtime - waittime;
629 	else
630 		l->lpo_waittime = 0;
631 out:
632 	/*
633 	 * Paired with cpus_fence_seq_cst().
634 	 */
635 	atomic_thread_fence_rel();
636 	critical_exit();
637 }
638 
639 void
640 lock_profile_thread_exit(struct thread *td)
641 {
642 #ifdef INVARIANTS
643 	struct lock_profile_object *l;
644 
645 	MPASS(curthread->td_critnest == 0);
646 #endif
647 	/*
648 	 * If lock profiling was disabled we have to wait for reset to
649 	 * clear our pointers before we can exit safely.
650 	 */
651 	lock_prof_reset_wait();
652 #ifdef INVARIANTS
653 	LIST_FOREACH(l, &td->td_lprof[0], lpo_link)
654 		printf("thread still holds lock acquired at %s:%d\n",
655 		    l->lpo_file, l->lpo_line);
656 	LIST_FOREACH(l, &td->td_lprof[1], lpo_link)
657 		printf("thread still holds lock acquired at %s:%d\n",
658 		    l->lpo_file, l->lpo_line);
659 #endif
660 	MPASS(LIST_FIRST(&td->td_lprof[0]) == NULL);
661 	MPASS(LIST_FIRST(&td->td_lprof[1]) == NULL);
662 }
663 
664 void
665 lock_profile_release_lock(struct lock_object *lo)
666 {
667 	struct lock_profile_object *l;
668 	struct lock_prof_type *type;
669 	struct lock_prof *lp;
670 	uint64_t curtime, holdtime;
671 	struct lpohead *head;
672 	int spin;
673 
674 	if (SCHEDULER_STOPPED())
675 		return;
676 	if (lo->lo_flags & LO_NOPROFILE)
677 		return;
678 	spin = (LOCK_CLASS(lo)->lc_flags & LC_SPINLOCK) ? 1 : 0;
679 	head = &curthread->td_lprof[spin];
680 	if (LIST_FIRST(head) == NULL)
681 		return;
682 	critical_enter();
683 	/* Recheck enabled now that we're in a critical section. */
684 	if (lock_prof_enable == 0 && lock_prof_resetting == 1)
685 		goto out;
686 	/*
687 	 * If lock profiling is not enabled we still want to remove the
688 	 * lpo from our queue.
689 	 */
690 	LIST_FOREACH(l, head, lpo_link)
691 		if (l->lpo_obj == lo)
692 			break;
693 	if (l == NULL)
694 		goto out;
695 	if (--l->lpo_ref > 0)
696 		goto out;
697 	lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
698 	if (lp == NULL)
699 		goto release;
700 	curtime = nanoseconds();
701 	if (curtime < l->lpo_acqtime)
702 		goto release;
703 	holdtime = curtime - l->lpo_acqtime;
704 
705 	/*
706 	 * Record if the lock has been held longer now than ever
707 	 * before.
708 	 */
709 	if (holdtime > lp->cnt_max)
710 		lp->cnt_max = holdtime;
711 	if (l->lpo_waittime > lp->cnt_wait_max)
712 		lp->cnt_wait_max = l->lpo_waittime;
713 	lp->cnt_tot += holdtime;
714 	lp->cnt_wait += l->lpo_waittime;
715 	lp->cnt_contest_locking += l->lpo_contest_locking;
716 	lp->cnt_cur += l->lpo_cnt;
717 release:
718 	LIST_REMOVE(l, lpo_link);
719 	type = &LP_CPU_SELF->lpc_types[spin];
720 	LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
721 out:
722 	/*
723 	 * Paired with cpus_fence_seq_cst().
724 	 */
725 	atomic_thread_fence_rel();
726 	critical_exit();
727 }
728 
729 static SYSCTL_NODE(_debug_lock, OID_AUTO, prof,
730     CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
731     "lock profiling");
732 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
733     &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
734 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
735     &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
736 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
737     &lock_prof_rejected, 0, "Number of rejected profiling records");
738 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats,
739     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
740     dump_lock_prof_stats, "A",
741     "Lock profiling statistics");
742 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset,
743     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0,
744     reset_lock_prof_stats, "I",
745     "Reset lock profiling statistics");
746 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable,
747     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
748     enable_lock_prof, "I",
749     "Enable lock profiling");
750 
751 #endif
752