xref: /freebsd/sys/kern/subr_lock.c (revision 9336e0699bda8a301cd2bfa37106b6ec5e32012e)
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("$FreeBSD$");
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 
61 CTASSERT(LOCK_CLASS_MAX == 15);
62 
63 struct 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 
72 void
73 lock_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 
98 void
99 lock_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
109 DB_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  */
136 struct 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  */
151 struct lock_prof {
152 	SLIST_ENTRY(lock_prof) link;
153 	const char	*file;
154 	const char	*name;
155 	int		line;
156 	int		ticks;
157 	const char	*type;
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 
165 SLIST_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  */
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];
183 };
184 
185 struct lock_prof_cpu {
186 	struct lock_prof_type	lpc_types[2]; /* One for spin one for other. */
187 };
188 
189 struct lock_prof_cpu *lp_cpu[MAXCPU];
190 
191 int lock_prof_enable = 0;
192 
193 /* SWAG: sbuf size = avg stat. line size * number of locks */
194 #define LPROF_SBUF_SIZE		256 * 400
195 
196 static int lock_prof_rejected;
197 static int lock_prof_skipspin;
198 static int lock_prof_skipcount;
199 
200 #ifndef USE_CPU_NANOSECONDS
201 u_int64_t
202 nanoseconds(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 
215 static void
216 lock_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 
230 static void
231 lock_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 }
242 SYSINIT(lockprof, SI_SUB_SMP, SI_ORDER_ANY, lock_prof_init, NULL);
243 
244 static void
245 lock_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 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
253 		lpc = lp_cpu[cpu];
254 		for (i = 0; i < LPROF_CACHE_SIZE; i++) {
255 			LIST_REMOVE(&lpc->lpc_types[0].lpt_objs[i], lpo_link);
256 			LIST_REMOVE(&lpc->lpc_types[1].lpt_objs[i], lpo_link);
257 		}
258 		bzero(lpc, sizeof(*lpc));
259 		lock_prof_init_type(&lpc->lpc_types[0]);
260 		lock_prof_init_type(&lpc->lpc_types[1]);
261 	}
262 	lock_prof_enable = enabled;
263 }
264 
265 static void
266 lock_prof_output(struct lock_prof *lp, struct sbuf *sb)
267 {
268 	const char *p;
269 
270 	for (p = lp->file; p != NULL && strncmp(p, "../", 3) == 0; p += 3);
271 	sbuf_printf(sb,
272 	    "%6ju %12ju %12ju %11ju %5ju %5ju %12ju %12ju %s:%d (%s:%s)\n",
273 	    lp->cnt_max / 1000, lp->cnt_tot / 1000,
274 	    lp->cnt_wait / 1000, lp->cnt_cur,
275 	    lp->cnt_cur == 0 ? (uintmax_t)0 :
276 	    lp->cnt_tot / (lp->cnt_cur * 1000),
277 	    lp->cnt_cur == 0 ? (uintmax_t)0 :
278 	    lp->cnt_wait / (lp->cnt_cur * 1000),
279 	    (uintmax_t)0, lp->cnt_contest_locking,
280 	    p, lp->line, lp->type, lp->name);
281 }
282 
283 static void
284 lock_prof_sum(struct lock_prof *match, struct lock_prof *dst, int hash,
285     int spin, int t)
286 {
287 	struct lock_prof_type *type;
288 	struct lock_prof *l;
289 	int cpu;
290 
291 	dst->file = match->file;
292 	dst->line = match->line;
293 	dst->type = match->type;
294 	dst->name = match->name;
295 
296 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
297 		if (lp_cpu[cpu] == NULL)
298 			continue;
299 		type = &lp_cpu[cpu]->lpc_types[spin];
300 		SLIST_FOREACH(l, &type->lpt_hash[hash], link) {
301 			if (l->ticks == t)
302 				continue;
303 			if (l->file != match->file || l->line != match->line ||
304 			    l->name != match->name || l->type != match->type)
305 				continue;
306 			l->ticks = t;
307 			if (l->cnt_max > dst->cnt_max)
308 				dst->cnt_max = l->cnt_max;
309 			dst->cnt_tot += l->cnt_tot;
310 			dst->cnt_wait += l->cnt_wait;
311 			dst->cnt_cur += l->cnt_cur;
312 			dst->cnt_contest_locking += l->cnt_contest_locking;
313 		}
314 	}
315 
316 }
317 
318 static void
319 lock_prof_type_stats(struct lock_prof_type *type, struct sbuf *sb, int spin,
320     int t)
321 {
322 	struct lock_prof *l;
323 	int i;
324 
325 	for (i = 0; i < LPROF_HASH_SIZE; ++i) {
326 		SLIST_FOREACH(l, &type->lpt_hash[i], link) {
327 			struct lock_prof lp = {};
328 
329 			if (l->ticks == t)
330 				continue;
331 			lock_prof_sum(l, &lp, i, spin, t);
332 			lock_prof_output(&lp, sb);
333 			if (sbuf_overflowed(sb))
334 				return;
335 		}
336 	}
337 }
338 
339 static int
340 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
341 {
342 	static int multiplier = 1;
343 	struct sbuf *sb;
344 	int error, cpu, t;
345 
346 retry_sbufops:
347 	sb = sbuf_new(NULL, NULL, LPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN);
348 	sbuf_printf(sb, "\n%6s %12s %12s %11s %5s %5s %12s %12s %s\n",
349 	    "max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
350 	t = ticks;
351 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
352 		if (lp_cpu[cpu] == NULL)
353 			continue;
354 		lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[0], sb, 0, t);
355 		lock_prof_type_stats(&lp_cpu[cpu]->lpc_types[1], sb, 1, t);
356 		if (sbuf_overflowed(sb)) {
357 			sbuf_delete(sb);
358 			multiplier++;
359 			goto retry_sbufops;
360 		}
361 	}
362 
363 	sbuf_finish(sb);
364 	error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
365 	sbuf_delete(sb);
366 	return (error);
367 }
368 
369 static int
370 enable_lock_prof(SYSCTL_HANDLER_ARGS)
371 {
372 	int error, v;
373 
374 	v = lock_prof_enable;
375 	error = sysctl_handle_int(oidp, &v, v, req);
376 	if (error)
377 		return (error);
378 	if (req->newptr == NULL)
379 		return (error);
380 	if (v == lock_prof_enable)
381 		return (0);
382 	if (v == 1)
383 		lock_prof_reset();
384 	lock_prof_enable = !!v;
385 
386 	return (0);
387 }
388 
389 static int
390 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
391 {
392 	int error, v;
393 
394 	v = 0;
395 	error = sysctl_handle_int(oidp, &v, 0, req);
396 	if (error)
397 		return (error);
398 	if (req->newptr == NULL)
399 		return (error);
400 	if (v == 0)
401 		return (0);
402 	lock_prof_reset();
403 
404 	return (0);
405 }
406 
407 static struct lock_prof *
408 lock_profile_lookup(struct lock_object *lo, int spin, const char *file,
409     int line)
410 {
411 	const char *unknown = "(unknown)";
412 	struct lock_prof_type *type;
413 	struct lock_prof *lp;
414 	struct lphead *head;
415 	const char *p;
416 	u_int hash;
417 
418 	p = file;
419 	if (p == NULL || *p == '\0')
420 		p = unknown;
421 	hash = (uintptr_t)lo->lo_name * 31 + (uintptr_t)p * 31 + line;
422 	hash &= LPROF_HASH_MASK;
423 	type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
424 	head = &type->lpt_hash[hash];
425 	SLIST_FOREACH(lp, head, link) {
426 		if (lp->line == line && lp->file == p &&
427 		    lp->name == lo->lo_name)
428 			return (lp);
429 
430 	}
431 	lp = SLIST_FIRST(&type->lpt_lpalloc);
432 	if (lp == NULL) {
433 		lock_prof_rejected++;
434 		return (lp);
435 	}
436 	SLIST_REMOVE_HEAD(&type->lpt_lpalloc, link);
437 	lp->file = p;
438 	lp->line = line;
439 	lp->type = lo->lo_type;
440 	lp->name = lo->lo_name;
441 	SLIST_INSERT_HEAD(&type->lpt_hash[hash], lp, link);
442 	return (lp);
443 }
444 
445 static struct lock_profile_object *
446 lock_profile_object_lookup(struct lock_object *lo, int spin, const char *file,
447     int line)
448 {
449 	struct lock_profile_object *l;
450 	struct lock_prof_type *type;
451 	struct lpohead *head;
452 
453 	head = &curthread->td_lprof[spin];
454 	LIST_FOREACH(l, head, lpo_link)
455 		if (l->lpo_obj == lo && l->lpo_file == file &&
456 		    l->lpo_line == line)
457 			return (l);
458 	critical_enter();
459 	type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
460 	l = LIST_FIRST(&type->lpt_lpoalloc);
461 	if (l == NULL) {
462 		lock_prof_rejected++;
463 		critical_exit();
464 		return (NULL);
465 	}
466 	LIST_REMOVE(l, lpo_link);
467 	critical_exit();
468 	l->lpo_obj = lo;
469 	l->lpo_file = file;
470 	l->lpo_line = line;
471 	l->lpo_cnt = 0;
472 	LIST_INSERT_HEAD(head, l, lpo_link);
473 
474 	return (l);
475 }
476 
477 void
478 lock_profile_obtain_lock_success(struct lock_object *lo, int contested,
479     uint64_t waittime, const char *file, int line)
480 {
481 	static int lock_prof_count;
482 	struct lock_profile_object *l;
483 	int spin;
484 
485 	/* don't reset the timer when/if recursing */
486 	if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
487 		return;
488 	if (lock_prof_skipcount &&
489 	    (++lock_prof_count % lock_prof_skipcount) == 0)
490 		return;
491 	spin = LOCK_CLASS(lo) == &lock_class_mtx_spin;
492 	if (spin && lock_prof_skipspin == 1)
493 		return;
494 	l = lock_profile_object_lookup(lo, spin, file, line);
495 	if (l == NULL)
496 		return;
497 	l->lpo_cnt++;
498 	if (++l->lpo_ref > 1)
499 		return;
500 	l->lpo_contest_locking = contested;
501 	l->lpo_acqtime = nanoseconds();
502 	if (waittime && (l->lpo_acqtime > waittime))
503 		l->lpo_waittime = l->lpo_acqtime - waittime;
504 	else
505 		l->lpo_waittime = 0;
506 }
507 
508 void
509 lock_profile_release_lock(struct lock_object *lo)
510 {
511 	struct lock_profile_object *l;
512 	struct lock_prof_type *type;
513 	struct lock_prof *lp;
514 	u_int64_t holdtime;
515 	struct lpohead *head;
516 	int spin;
517 
518 	if (!lock_prof_enable || (lo->lo_flags & LO_NOPROFILE))
519 		return;
520 	spin = LOCK_CLASS(lo) == &lock_class_mtx_spin;
521 	head = &curthread->td_lprof[spin];
522 	critical_enter();
523 	LIST_FOREACH(l, head, lpo_link)
524 		if (l->lpo_obj == lo)
525 			break;
526 	if (l == NULL)
527 		goto out;
528 	if (--l->lpo_ref > 0)
529 		goto out;
530 	lp = lock_profile_lookup(lo, spin, l->lpo_file, l->lpo_line);
531 	if (lp == NULL)
532 		goto release;
533 	holdtime = nanoseconds() - l->lpo_acqtime;
534 	if (holdtime < 0)
535 		goto release;
536 	/*
537 	 * Record if the lock has been held longer now than ever
538 	 * before.
539 	 */
540 	if (holdtime > lp->cnt_max)
541 		lp->cnt_max = holdtime;
542 	lp->cnt_tot += holdtime;
543 	lp->cnt_wait += l->lpo_waittime;
544 	lp->cnt_contest_locking += l->lpo_contest_locking;
545 	lp->cnt_cur += l->lpo_cnt;
546 release:
547 	LIST_REMOVE(l, lpo_link);
548 	type = &lp_cpu[PCPU_GET(cpuid)]->lpc_types[spin];
549 	LIST_INSERT_HEAD(&type->lpt_lpoalloc, l, lpo_link);
550 out:
551 	critical_exit();
552 }
553 
554 SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD, NULL, "lock debugging");
555 SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL, "lock profiling");
556 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipspin, CTLFLAG_RW,
557     &lock_prof_skipspin, 0, "Skip profiling on spinlocks.");
558 SYSCTL_INT(_debug_lock_prof, OID_AUTO, skipcount, CTLFLAG_RW,
559     &lock_prof_skipcount, 0, "Sample approximately every N lock acquisitions.");
560 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
561     &lock_prof_rejected, 0, "Number of rejected profiling records");
562 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
563     NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
564 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
565     NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
566 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RW,
567     NULL, 0, enable_lock_prof, "I", "Enable lock profiling");
568 
569 #endif
570