xref: /freebsd/sys/kern/subr_lock.c (revision f0a75d274af375d15b97b830966b99a02b7db911)
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/ktr.h>
44 #include <sys/linker_set.h>
45 #include <sys/lock.h>
46 #include <sys/sbuf.h>
47 #include <sys/sysctl.h>
48 #include <sys/lock_profile.h>
49 
50 #ifdef DDB
51 #include <ddb/ddb.h>
52 #endif
53 
54 CTASSERT(LOCK_CLASS_MAX == 15);
55 
56 struct lock_class *lock_classes[LOCK_CLASS_MAX + 1] = {
57 	&lock_class_mtx_spin,
58 	&lock_class_mtx_sleep,
59 	&lock_class_sx,
60 	&lock_class_rw,
61 	&lock_class_lockmgr,
62 };
63 
64 #ifdef LOCK_PROFILING
65 #include <machine/cpufunc.h>
66 
67 SYSCTL_NODE(_debug, OID_AUTO, lock, CTLFLAG_RD, NULL, "lock debugging");
68 SYSCTL_NODE(_debug_lock, OID_AUTO, prof, CTLFLAG_RD, NULL, "lock profiling");
69 int lock_prof_enable = 0;
70 SYSCTL_INT(_debug_lock_prof, OID_AUTO, enable, CTLFLAG_RW,
71     &lock_prof_enable, 0, "Enable lock profiling");
72 
73 /*
74  * lprof_buf is a static pool of profiling records to avoid possible
75  * reentrance of the memory allocation functions.
76  *
77  * Note: NUM_LPROF_BUFFERS must be smaller than LPROF_HASH_SIZE.
78  */
79 struct lock_prof lprof_buf[LPROF_HASH_SIZE];
80 static int allocated_lprof_buf;
81 struct mtx lprof_locks[LPROF_LOCK_SIZE];
82 
83 
84 /* SWAG: sbuf size = avg stat. line size * number of locks */
85 #define LPROF_SBUF_SIZE		256 * 400
86 
87 static int lock_prof_acquisitions;
88 SYSCTL_INT(_debug_lock_prof, OID_AUTO, acquisitions, CTLFLAG_RD,
89     &lock_prof_acquisitions, 0, "Number of lock acquistions recorded");
90 static int lock_prof_records;
91 SYSCTL_INT(_debug_lock_prof, OID_AUTO, records, CTLFLAG_RD,
92     &lock_prof_records, 0, "Number of profiling records");
93 static int lock_prof_maxrecords = LPROF_HASH_SIZE;
94 SYSCTL_INT(_debug_lock_prof, OID_AUTO, maxrecords, CTLFLAG_RD,
95     &lock_prof_maxrecords, 0, "Maximum number of profiling records");
96 static int lock_prof_rejected;
97 SYSCTL_INT(_debug_lock_prof, OID_AUTO, rejected, CTLFLAG_RD,
98     &lock_prof_rejected, 0, "Number of rejected profiling records");
99 static int lock_prof_hashsize = LPROF_HASH_SIZE;
100 SYSCTL_INT(_debug_lock_prof, OID_AUTO, hashsize, CTLFLAG_RD,
101     &lock_prof_hashsize, 0, "Hash size");
102 static int lock_prof_collisions = 0;
103 SYSCTL_INT(_debug_lock_prof, OID_AUTO, collisions, CTLFLAG_RD,
104     &lock_prof_collisions, 0, "Number of hash collisions");
105 
106 #ifndef USE_CPU_NANOSECONDS
107 u_int64_t
108 nanoseconds(void)
109 {
110 	struct timespec tv;
111 
112 	nanotime(&tv);
113 	return (tv.tv_sec * (u_int64_t)1000000000 + tv.tv_nsec);
114 }
115 #endif
116 
117 static int
118 dump_lock_prof_stats(SYSCTL_HANDLER_ARGS)
119 {
120         struct sbuf *sb;
121         int error, i;
122         static int multiplier = 1;
123         const char *p;
124 
125         if (allocated_lprof_buf == 0)
126                 return (SYSCTL_OUT(req, "No locking recorded",
127                     sizeof("No locking recorded")));
128 
129 retry_sbufops:
130         sb = sbuf_new(NULL, NULL, LPROF_SBUF_SIZE * multiplier, SBUF_FIXEDLEN);
131         sbuf_printf(sb, "\n%6s %12s %12s %11s %5s %5s %12s %12s %s\n",
132             "max", "total", "wait_total", "count", "avg", "wait_avg", "cnt_hold", "cnt_lock", "name");
133         for (i = 0; i < LPROF_HASH_SIZE; ++i) {
134                 if (lprof_buf[i].name == NULL)
135                         continue;
136                 for (p = lprof_buf[i].file;
137                         p != NULL && strncmp(p, "../", 3) == 0; p += 3)
138                                 /* nothing */ ;
139                 sbuf_printf(sb, "%6ju %12ju %12ju %11ju %5ju %5ju %12ju %12ju %s:%d (%s:%s)\n",
140                     lprof_buf[i].cnt_max / 1000,
141                     lprof_buf[i].cnt_tot / 1000,
142                     lprof_buf[i].cnt_wait / 1000,
143                     lprof_buf[i].cnt_cur,
144                     lprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 :
145                         lprof_buf[i].cnt_tot / (lprof_buf[i].cnt_cur * 1000),
146                     lprof_buf[i].cnt_cur == 0 ? (uintmax_t)0 :
147                         lprof_buf[i].cnt_wait / (lprof_buf[i].cnt_cur * 1000),
148                     lprof_buf[i].cnt_contest_holding,
149                     lprof_buf[i].cnt_contest_locking,
150                     p, lprof_buf[i].line,
151 			    lprof_buf[i].type,
152 			    lprof_buf[i].name);
153                 if (sbuf_overflowed(sb)) {
154                         sbuf_delete(sb);
155                         multiplier++;
156                         goto retry_sbufops;
157                 }
158         }
159 
160         sbuf_finish(sb);
161         error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
162         sbuf_delete(sb);
163         return (error);
164 }
165 static int
166 reset_lock_prof_stats(SYSCTL_HANDLER_ARGS)
167 {
168         int error, v;
169 
170         if (allocated_lprof_buf == 0)
171                 return (0);
172 
173         v = 0;
174         error = sysctl_handle_int(oidp, &v, 0, req);
175         if (error)
176                 return (error);
177         if (req->newptr == NULL)
178                 return (error);
179         if (v == 0)
180                 return (0);
181 
182         bzero(lprof_buf, LPROF_HASH_SIZE*sizeof(*lprof_buf));
183         allocated_lprof_buf = 0;
184         return (0);
185 }
186 
187 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, stats, CTLTYPE_STRING | CTLFLAG_RD,
188     NULL, 0, dump_lock_prof_stats, "A", "Lock profiling statistics");
189 
190 SYSCTL_PROC(_debug_lock_prof, OID_AUTO, reset, CTLTYPE_INT | CTLFLAG_RW,
191     NULL, 0, reset_lock_prof_stats, "I", "Reset lock profiling statistics");
192 #endif
193 
194 void
195 lock_init(struct lock_object *lock, struct lock_class *class, const char *name,
196     const char *type, int flags)
197 {
198 	int i;
199 
200 	/* Check for double-init and zero object. */
201 	KASSERT(!lock_initalized(lock), ("lock \"%s\" %p already initialized",
202 	    name, lock));
203 
204 	/* Look up lock class to find its index. */
205 	for (i = 0; i < LOCK_CLASS_MAX; i++)
206 		if (lock_classes[i] == class) {
207 			lock->lo_flags = i << LO_CLASSSHIFT;
208 			break;
209 		}
210 	KASSERT(i < LOCK_CLASS_MAX, ("unknown lock class %p", class));
211 
212 	/* Initialize the lock object. */
213 	lock->lo_name = name;
214 	lock->lo_type = type != NULL ? type : name;
215 	lock->lo_flags |= flags | LO_INITIALIZED;
216 	LOCK_LOG_INIT(lock, 0);
217 	WITNESS_INIT(lock);
218 }
219 
220 void
221 lock_destroy(struct lock_object *lock)
222 {
223 
224 	KASSERT(lock_initalized(lock), ("lock %p is not initialized", lock));
225 	WITNESS_DESTROY(lock);
226 	LOCK_LOG_DESTROY(lock, 0);
227 	lock->lo_flags &= ~LO_INITIALIZED;
228 }
229 
230 #ifdef DDB
231 DB_SHOW_COMMAND(lock, db_show_lock)
232 {
233 	struct lock_object *lock;
234 	struct lock_class *class;
235 
236 	if (!have_addr)
237 		return;
238 	lock = (struct lock_object *)addr;
239 	if (LO_CLASSINDEX(lock) > LOCK_CLASS_MAX) {
240 		db_printf("Unknown lock class: %d\n", LO_CLASSINDEX(lock));
241 		return;
242 	}
243 	class = LOCK_CLASS(lock);
244 	db_printf(" class: %s\n", class->lc_name);
245 	db_printf(" name: %s\n", lock->lo_name);
246 	if (lock->lo_type && lock->lo_type != lock->lo_name)
247 		db_printf(" type: %s\n", lock->lo_type);
248 	class->lc_ddb_show(lock);
249 }
250 #endif
251 
252 #ifdef LOCK_PROFILING
253 void _lock_profile_obtain_lock_success(struct lock_object *lo, int contested, uint64_t waittime, const char *file, int line)
254 {
255         struct lock_profile_object *l = &lo->lo_profile_obj;
256 
257 	lo->lo_profile_obj.lpo_contest_holding = 0;
258 
259 	if (contested)
260 		lo->lo_profile_obj.lpo_contest_locking++;
261 
262 	l->lpo_filename = file;
263 	l->lpo_lineno = line;
264 	l->lpo_acqtime = nanoseconds();
265 	if (waittime && (l->lpo_acqtime > waittime))
266 		l->lpo_waittime = l->lpo_acqtime - waittime;
267 	else
268 		l->lpo_waittime = 0;
269 }
270 
271 void _lock_profile_release_lock(struct lock_object *lo)
272 {
273         struct lock_profile_object *l = &lo->lo_profile_obj;
274 
275         if (l->lpo_acqtime && !(lo->lo_flags & LO_NOPROFILE)) {
276                 const char *unknown = "(unknown)";
277                 u_int64_t acqtime, now, waittime;
278                 struct lock_prof *mpp;
279                 u_int hash;
280                 const char *p = l->lpo_filename;
281                 int collision = 0;
282 
283                 now = nanoseconds();
284                 acqtime = l->lpo_acqtime;
285                 waittime = l->lpo_waittime;
286                 if (now <= acqtime)
287                         return;
288                 if (p == NULL || *p == '\0')
289                         p = unknown;
290                 hash = (l->lpo_namehash * 31 * 31 + (uintptr_t)p * 31 + l->lpo_lineno) & LPROF_HASH_MASK;
291                 CTR5(KTR_SPARE1, "Hashing %s(%x) %s:%d to %d", l->lpo_filename,
292 		     l->lpo_namehash, p, l->lpo_lineno, hash);
293                 mpp = &lprof_buf[hash];
294                 while (mpp->name != NULL) {
295                         if (mpp->line == l->lpo_lineno &&
296                           mpp->file == p &&
297                           mpp->namehash == l->lpo_namehash)
298                                 break;
299                         /* If the lprof_hash entry is allocated to someone
300 			 * else, try the next one
301 			 */
302                         collision = 1;
303                         CTR4(KTR_SPARE1, "Hash collision, %s:%d %s(%x)", mpp->file,
304 			     mpp->line, mpp->name, mpp->namehash);
305                         hash = (hash + 1) & LPROF_HASH_MASK;
306                         mpp = &lprof_buf[hash];
307                 }
308                 if (mpp->name == NULL) {
309                         int buf;
310 
311                         buf = atomic_fetchadd_int(&allocated_lprof_buf, 1);
312                         /* Just exit if we cannot get a trace buffer */
313                         if (buf >= LPROF_HASH_SIZE) {
314                                 ++lock_prof_rejected;
315                                 return;
316                         }
317 			mpp->file = p;
318 			mpp->line = l->lpo_lineno;
319 			mpp->namehash = l->lpo_namehash;
320 			mpp->type = l->lpo_type;
321 			mpp->name = lo->lo_name;
322 
323 			if (collision)
324 				++lock_prof_collisions;
325 
326                         /*
327 			 * We might have raced someone else but who cares,
328 			 * they'll try again next time
329 			 */
330                         ++lock_prof_records;
331                 }
332                 LPROF_LOCK(hash);
333                 /*
334                  * Record if the lock has been held longer now than ever
335                  * before.
336                  */
337                 if (now - acqtime > mpp->cnt_max)
338                         mpp->cnt_max = now - acqtime;
339                 mpp->cnt_tot += now - acqtime;
340                 mpp->cnt_wait += waittime;
341                 mpp->cnt_cur++;
342                 /*
343                  * There's a small race, really we should cmpxchg
344                  * 0 with the current value, but that would bill
345                  * the contention to the wrong lock instance if
346                  * it followed this also.
347                  */
348                 mpp->cnt_contest_holding += l->lpo_contest_holding;
349                 mpp->cnt_contest_locking += l->lpo_contest_locking;
350                 LPROF_UNLOCK(hash);
351 
352         }
353         l->lpo_acqtime = 0;
354         l->lpo_waittime = 0;
355         l->lpo_contest_locking = 0;
356         l->lpo_contest_holding = 0;
357 }
358 #endif
359