subr_lock.c revision 175150
1/*-
2 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
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.
16 *
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
27 * SUCH DAMAGE.
28 */
29
30/*
31 * This module holds the global variables and functions used to maintain
32 * lock_object structures.
33 */
34
35#include <sys/cdefs.h>
36__FBSDID("$FreeBSD: head/sys/kern/subr_lock.c 175150 2008-01-08 01:11:40Z kris $");
37
38#include "opt_ddb.h"
39#include "opt_mprof.h"
40
41#include <sys/param.h>
42#include <sys/systm.h>
43#include <sys/kernel.h>
44#include <sys/ktr.h>
45#include <sys/linker_set.h>
46#include <sys/lock.h>
47#include <sys/lock_profile.h>
48#include <sys/malloc.h>
49#include <sys/pcpu.h>
50#include <sys/proc.h>
51#include <sys/sbuf.h>
52#include <sys/smp.h>
53#include <sys/sysctl.h>
54
55#ifdef DDB
56#include <ddb/ddb.h>
57#endif
58
59#include <machine/cpufunc.h>
60
61CTASSERT(LOCK_CLASS_MAX == 15);
62
63struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
64	&lock_class_mtx_spin,
65	&lock_class_mtx_sleep,
66	&lock_class_sx,
67	&lock_class_rm,
68	&lock_class_rw,
69	&lock_class_lockmgr,
70};
71
72void
73lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
74    const char *type, int flags)
75{
76	int i;
77
78	/* Check for double-init and zero object. */
79	KASSERT(!lock_initalized(lock), ("lock \"%s\" %p already initialized",
80	    name, lock));
81
82	/* Look up lock class to find its index. */
83	for (i = 0; i < LOCK_CLASS_MAX; i++)
84		if (lock_classes[i] == class) {
85			lock->lo_flags = i << LO_CLASSSHIFT;
86			break;
87		}
88	KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
89
90	/* Initialize the lock object. */
91	lock->lo_name = name;
92	lock->lo_type = type != NULL ? type : name;
93	lock->lo_flags |= flags | LO_INITIALIZED;
94	LOCK_LOG_INIT(lock, 0);
95	WITNESS_INIT(lock);
96}
97
98void
99lock_destroy(struct lock_object *lock)
100{
101
102	KASSERT(lock_initalized(lock), ("lock %p is not initialized", lock));
103	WITNESS_DESTROY(lock);
104	LOCK_LOG_DESTROY(lock, 0);
105	lock->lo_flags &= ~LO_INITIALIZED;
106}
107
108#ifdef DDB
109DB_SHOW_COMMAND(lock, db_show_lock)
110{
111	struct lock_object *lock;
112	struct lock_class *class;
113
114	if (!have_addr)
115		return;
116	lock = (struct lock_object *)addr;
117	if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
118		db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
119		return;
120	}
121	class = LOCK_CLASS(lock);
122	db_printf(" class: %s\n", class->lc_name);
123	db_printf(" name: %s\n", lock->lo_name);
124	if (lock->lo_type && lock->lo_type != lock->lo_name)
125		db_printf(" type: %s\n", lock->lo_type);
126	class->lc_ddb_show(lock);
127}
128#endif
129
130#ifdef LOCK_PROFILING
131
132/*
133 * One object per-thread for each lock the thread owns.  Tracks individual
134 * lock instances.
135 */
136struct lock_profile_object {
137	LIST_ENTRY(lock_profile_object) lpo_link;
138	struct lock_object *lpo_obj;
139	const char	*lpo_file;
140	int		lpo_line;
141	uint16_t	lpo_ref;
142	uint16_t	lpo_cnt;
143	u_int64_t	lpo_acqtime;
144	u_int64_t	lpo_waittime;
145	u_int		lpo_contest_locking;
146};
147
148/*
149 * One lock_prof for each (file, line, lock object) triple.
150 */
151struct lock_prof {
152	SLIST_ENTRY(lock_prof) link;
153	struct lock_class *class;
154	const char	*file;
155	const char	*name;
156	int		line;
157	int		ticks;
158	uintmax_t	cnt_max;
159	uintmax_t	cnt_tot;
160	uintmax_t	cnt_wait;
161	uintmax_t	cnt_cur;
162	uintmax_t	cnt_contest_locking;
163};
164
165SLIST_HEAD(lphead, lock_prof);
166
167#define	LPROF_HASH_SIZE		4096
168#define	LPROF_HASH_MASK		(LPROF_HASH_SIZE - 1)
169#define	LPROF_CACHE_SIZE	4096
170
171/*
172 * Array of objects and profs for each type of object for each cpu.  Spinlocks
173 * are handled seperately 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.
176 */
177struct 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];
183};
184
185struct lock_prof_cpu {
186	struct lock_prof_type	lpc_types[2]; /* One for spin one for other. */
187};
188
189struct lock_prof_cpu *lp_cpu[MAXCPU];
190
191int lock_prof_enable = 0;
192
193/* SWAG: sbuf size = avg stat. line size * number of locks */
194#define LPROF_SBUF_SIZE		256 * 400
195
196static int lock_prof_rejected;
197static int lock_prof_skipspin;
198static int lock_prof_skipcount;
199
200#ifndef USE_CPU_NANOSECONDS
201u_int64_t
202nanoseconds(void)
203{
204	struct bintime bt;
205	u_int64_t ns;
206
207	binuptime(&bt);
208	/* From bintime2timespec */
209	ns = bt.sec * (u_int64_t)1000000000;
210	ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32;
211	return (ns);
212}
213#endif
214
215static void
216lock_prof_init_type(struct lock_prof_type *type)
217{
218	int i;
219
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],
224		    link);
225		LIST_INSERT_HEAD(&type->lpt_lpoalloc, &type->lpt_objs[i],
226		    lpo_link);
227	}
228}
229
230static void
231lock_prof_init(void *arg)
232{
233	int cpu;
234
235	for (cpu = 0; cpu <= mp_maxid; cpu++) {
236		lp_cpu[cpu] = malloc(sizeof(*lp_cpu[cpu]), M_DEVBUF,
237		    M_WAITOK | M_ZERO);
238		lock_prof_init_type(&lp_cpu[cpu]->lpc_types[0]);
239		lock_prof_init_type(&lp_cpu[cpu]->lpc_types[1]);
240	}
241}
242SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
243
244static void
245lock_prof_reset(void)
246{
247	struct lock_prof_cpu *lpc;
248	int enabled, i, cpu;
249
250	enabled = lock_prof_enable;
251	lock_prof_enable = 0;
252	pause("lpreset", hz / 10);
253	for (cpu = 0; cpu <= mp_maxid; cpu++) {
254		lpc = lp_cpu[cpu];
255		for (i = 0; i < LPROF_CACHE_SIZE; i++) {
256			LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
257			LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
258		}
259		bzero(lpc, sizeof(*lpc));
260		lock_prof_init_type(&lpc->lpc_types[0]);
261		lock_prof_init_type(&lpc->lpc_types[1]);
262	}
263	lock_prof_enable = enabled;
264}
265
266static void
267lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
268{
269	const char *p;
270
271	for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
272	sbuf_printf(sb,
273	    "%6ju %12ju %12ju %11ju %5ju %5ju %12ju %12ju %s:%d (%s:%s)\n",
274	    lp->cnt_max / 1000, lp->cnt_tot / 1000,
275	    lp->cnt_wait / 1000, lp->cnt_cur,
276	    lp->cnt_cur == 0 ? (uintmax_t)0 :
277	    lp->cnt_tot / (lp->cnt_cur * 1000),
278	    lp->cnt_cur == 0 ? (uintmax_t)0 :
279	    lp->cnt_wait / (lp->cnt_cur * 1000),
280	    (uintmax_t)0, lp->cnt_contest_locking,
281	    p, lp->line, lp->class->lc_name, lp->name);
282}
283
284static void
285lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
286    int spin, int t)
287{
288	struct lock_prof_type *type;
289	struct lock_prof *l;
290	int cpu;
291
292	dst->file = match->file;
293	dst->line = match->line;
294	dst->class = match->class;
295	dst->name = match->name;
296
297	for (cpu = 0; cpu <= mp_maxid; cpu++) {
298		if (lp_cpu[cpu] == NULL)
299			continue;
300		type = &lp_cpu[cpu]->lpc_types[spin];
301		SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
302			if (l->ticks == t)
303				continue;
304			if (l->file != match->file || l->line != match->line ||
305			    l->name != match->name)
306				continue;
307			l->ticks = t;
308			if (l->cnt_max > dst->cnt_max)
309				dst->cnt_max = l->cnt_max;
310			dst->cnt_tot += l->cnt_tot;
311			dst->cnt_wait += l->cnt_wait;
312			dst->cnt_cur += l->cnt_cur;
313			dst->cnt_contest_locking += l->cnt_contest_locking;
314		}
315	}
316
317}
318
319static void
320lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
321    int t)
322{
323	struct lock_prof *l;
324	int i;
325
326	for (i = 0; i < LPROF_HASH_SIZE; ++i) {
327		SLIST_FOREACH(l, &type->lpt_hash[i], link) {
328			struct lock_prof lp = {};
329
330			if (l->ticks == t)
331				continue;
332			lock_prof_sum(l, &lp, i, spin, t);
333			lock_prof_output(&lp, sb);
334			if (sbuf_overflowed(sb))
335				return;
336		}
337	}
338}
339
340static int
341dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
342{
343	static int multiplier = 1;
344	struct sbuf *sb;
345	int error, cpu, t;
346	int enabled;
347
348retry_sbufops:
349	sb = sbuf_new(NULL, NULL, LPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN);
350	sbuf_printf(sb, "\n%6s %12s %12s %11s %5s %5s %12s %12s %s\n",
351	    "max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
352	enabled = lock_prof_enable;
353	lock_prof_enable = 0;
354	pause("lpreset", hz / 10);
355	t = ticks;
356	for (cpu = 0; cpu <= mp_maxid; cpu++) {
357		if (lp_cpu[cpu] == NULL)
358			continue;
359		lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[0], sb, 0, t);
360		lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[1], sb, 1, t);
361		if (sbuf_overflowed(sb)) {
362			sbuf_delete(sb);
363			multiplier++;
364			goto retry_sbufops;
365		}
366	}
367	lock_prof_enable = enabled;
368
369	sbuf_finish(sb);
370	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
371	sbuf_delete(sb);
372	return (error);
373}
374
375static int
376enable_lock_prof(SYSCTL_HANDLER_ARGS)
377{
378	int error, v;
379
380	v = lock_prof_enable;
381	error = sysctl_handle_int(oidp, &v, v, req);
382	if (error)
383		return (error);
384	if (req->newptr == NULL)
385		return (error);
386	if (v == lock_prof_enable)
387		return (0);
388	if (v == 1)
389		lock_prof_reset();
390	lock_prof_enable = !!v;
391
392	return (0);
393}
394
395static int
396reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
397{
398	int error, v;
399
400	v = 0;
401	error = sysctl_handle_int(oidp, &v, 0, req);
402	if (error)
403		return (error);
404	if (req->newptr == NULL)
405		return (error);
406	if (v == 0)
407		return (0);
408	lock_prof_reset();
409
410	return (0);
411}
412
413static struct lock_prof *
414lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
415    int line)
416{
417	const char *unknown = "(unknown)";
418	struct lock_prof_type *type;
419	struct lock_prof *lp;
420	struct lphead *head;
421	const char *p;
422	u_int hash;
423
424	p = file;
425	if (p == NULL || *p == '\0')
426		p = unknown;
427	hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
428	hash &= LPROF_HASH_MASK;
429	type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
430	head = &type->lpt_hash[hash];
431	SLIST_FOREACH(lp, head, link) {
432		if (lp->line == line && lp->file == p &&
433		    lp->name == lo->lo_name)
434			return (lp);
435
436	}
437	lp = SLIST_FIRST(&type->lpt_lpalloc);
438	if (lp == NULL) {
439		lock_prof_rejected++;
440		return (lp);
441	}
442	SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
443	lp->file = p;
444	lp->line = line;
445	lp->class = LOCK_CLASS(lo);
446	lp->name = lo->lo_name;
447	SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
448	return (lp);
449}
450
451static struct lock_profile_object *
452lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
453    int line)
454{
455	struct lock_profile_object *l;
456	struct lock_prof_type *type;
457	struct lpohead *head;
458
459	head = &curthread->td_lprof[spin];
460	LIST_FOREACH(l, head, lpo_link)
461		if (l->lpo_obj == lo && l->lpo_file == file &&
462		    l->lpo_line == line)
463			return (l);
464	critical_enter();
465	type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
466	l = LIST_FIRST(&type->lpt_lpoalloc);
467	if (l == NULL) {
468		lock_prof_rejected++;
469		critical_exit();
470		return (NULL);
471	}
472	LIST_REMOVE(l, lpo_link);
473	critical_exit();
474	l->lpo_obj = lo;
475	l->lpo_file = file;
476	l->lpo_line = line;
477	l->lpo_cnt = 0;
478	LIST_INSERT_HEAD(head, l, lpo_link);
479
480	return (l);
481}
482
483void
484lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
485    uint64_t waittime, const char *file, int line)
486{
487	static int lock_prof_count;
488	struct lock_profile_object *l;
489	int spin;
490
491	/* don't reset the timer when/if recursing */
492	if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
493		return;
494	if (lock_prof_skipcount &&
495	    (++lock_prof_count % lock_prof_skipcount) != 0)
496		return;
497	spin = LOCK_CLASS(lo) == &lock_class_mtx_spin;
498	if (spin && lock_prof_skipspin == 1)
499		return;
500	l = lock_profile_object_lookup(lo, spin, file, line);
501	if (l == NULL)
502		return;
503	l->lpo_cnt++;
504	if (++l->lpo_ref > 1)
505		return;
506	l->lpo_contest_locking = contested;
507	l->lpo_acqtime = nanoseconds();
508	if (waittime && (l->lpo_acqtime > waittime))
509		l->lpo_waittime = l->lpo_acqtime - waittime;
510	else
511		l->lpo_waittime = 0;
512}
513
514void
515lock_profile_release_lock(struct lock_object *lo)
516{
517	struct lock_profile_object *l;
518	struct lock_prof_type *type;
519	struct lock_prof *lp;
520	u_int64_t holdtime;
521	struct lpohead *head;
522	int spin;
523
524	if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
525		return;
526	spin = LOCK_CLASS(lo) == &lock_class_mtx_spin;
527	head = &curthread->td_lprof[spin];
528	critical_enter();
529	LIST_FOREACH(l, head, lpo_link)
530		if (l->lpo_obj == lo)
531			break;
532	if (l == NULL)
533		goto out;
534	if (--l->lpo_ref > 0)
535		goto out;
536	lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
537	if (lp == NULL)
538		goto release;
539	holdtime = nanoseconds() - l->lpo_acqtime;
540	if (holdtime < 0)
541		goto release;
542	/*
543	 * Record if the lock has been held longer now than ever
544	 * before.
545	 */
546	if (holdtime > lp->cnt_max)
547		lp->cnt_max = holdtime;
548	lp->cnt_tot += holdtime;
549	lp->cnt_wait += l->lpo_waittime;
550	lp->cnt_contest_locking += l->lpo_contest_locking;
551	lp->cnt_cur += l->lpo_cnt;
552release:
553	LIST_REMOVE(l, lpo_link);
554	type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
555	LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
556out:
557	critical_exit();
558}
559
560SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD, NULL, "lock debugging");
561SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL, "lock profiling");
562SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
563    &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
564SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
565    &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
566SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
567    &lock_prof_rejected, 0, "Number of rejected profiling records");
568SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
569    NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
570SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
571    NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
572SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
573    NULL, 0, enable_lock_prof, "I", "Enable lock profiling");
574
575#endif
576