main.c revision 1.1 1 /* $NetBSD: main.c,v 1.1 2006/09/07 00:50:45 ad Exp $ */
2
3 /*-
4 * Copyright (c) 2006 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Andrew Doran.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the NetBSD
21 * Foundation, Inc. and its contributors.
22 * 4. Neither the name of The NetBSD Foundation nor the names of its
23 * contributors may be used to endorse or promote products derived
24 * from this software without specific prior written permission.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36 * POSSIBILITY OF SUCH DAMAGE.
37 */
38
39 /*
40 * TODO:
41 *
42 * - Need better analysis and tracking of events.
43 * - Should be binary format agnostic, but given that we're likely to be using
44 * ELF for quite a while that's not a big problem.
45 * - Shouldn't have to parse the namelist here. We should use something like
46 * FreeBSD's libelf.
47 * - The way the namelist is searched sucks, is it worth doing something
48 * better?
49 */
50
51 #include <sys/cdefs.h>
52 #ifndef lint
53 __RCSID("$NetBSD: main.c,v 1.1 2006/09/07 00:50:45 ad Exp $");
54 #endif /* not lint */
55
56 #include <sys/types.h>
57 #include <sys/param.h>
58 #include <sys/time.h>
59 #include <sys/fcntl.h>
60 #include <sys/ioctl.h>
61 #include <sys/wait.h>
62 #include <sys/signal.h>
63 #include <sys/lockstat.h>
64 #include <sys/sysctl.h>
65
66 #include <stdio.h>
67 #include <stdlib.h>
68 #include <string.h>
69 #include <limits.h>
70 #include <unistd.h>
71 #include <err.h>
72 #include <paths.h>
73 #include <util.h>
74 #include <ctype.h>
75 #include <errno.h>
76
77 #include "extern.h"
78
79 #define _PATH_DEV_LOCKSTAT "/dev/lockstat"
80
81 #define MILLI 1000.0
82 #define MICRO 1000000.0
83 #define NANO 1000000000.0
84 #define PICO 1000000000000.0
85
86 TAILQ_HEAD(lock_head, lockstruct);
87 typedef struct lock_head locklist_t;
88 TAILQ_HEAD(buf_head, lsbuf);
89 typedef struct buf_head buflist_t;
90
91 typedef struct lockstruct {
92 TAILQ_ENTRY(lockstruct) chain;
93 buflist_t bufs;
94 uintptr_t lock;
95 double times[LB_NEVENT];
96 uint32_t counts[LB_NEVENT];
97 u_int flags;
98 } lock_t;
99
100 typedef struct name {
101 const char *name;
102 int mask;
103 } name_t;
104
105 const name_t locknames[] = {
106 { "adaptive_mutex", LB_ADAPTIVE_MUTEX },
107 { "adaptive_rwlock", LB_ADAPTIVE_RWLOCK },
108 { "spin_mutex", LB_SPIN_MUTEX },
109 { "spin_rwlock", LB_SPIN_RWLOCK },
110 { "lockmgr", LB_LOCKMGR },
111 { NULL, 0 }
112 };
113
114 const name_t eventnames[] = {
115 { "spin", LB_SPIN },
116 { "sleep", LB_SLEEP },
117 { NULL, 0 },
118 };
119
120 const name_t alltypes[] = {
121 { "Adaptive mutex spin", LB_ADAPTIVE_MUTEX | LB_SPIN },
122 { "Adaptive mutex sleep", LB_ADAPTIVE_MUTEX | LB_SLEEP },
123 { "Adaptive RW lock spin", LB_ADAPTIVE_RWLOCK | LB_SPIN },
124 { "Adaptive RW lock sleep", LB_ADAPTIVE_RWLOCK | LB_SPIN },
125 { "Spin mutex spin", LB_SPIN_MUTEX | LB_SPIN },
126 { "Spin RW lock spin", LB_SPIN_RWLOCK | LB_SPIN },
127 { "lockmgr sleep", LB_LOCKMGR | LB_SLEEP },
128 { NULL, 0 }
129 };
130
131 locklist_t locklist[LB_NLOCK];
132
133 lsbuf_t *bufs;
134 lsdisable_t ld;
135 int lflag;
136 int nbufs;
137 int cflag;
138 int lsfd;
139 int displayed;
140 int bin64;
141 double tscale;
142 double cscale;
143 double cpuscale[sizeof(ld.ld_freq) / sizeof(ld.ld_freq[0])];
144 FILE *outfp;
145
146 void findsym(findsym_t, char *, uintptr_t *, uintptr_t *);
147 void spawn(int, char **);
148 void display(int, const char *name);
149 void listnames(const name_t *);
150 int matchname(const name_t *, const char *);
151 void makelists(void);
152 void nullsig(int);
153 void usage(void);
154 void resort(int, int);
155 int ncpu(void);
156
157 int
158 main(int argc, char **argv)
159 {
160 int eventtype, locktype, ch, nlfd, sflag, fd, i, pflag;
161 const char *nlistf, *outf;
162 char *lockname, *funcname;
163 const name_t *name;
164 lsenable_t le;
165 double ms;
166 char *p;
167
168 nlistf = NULL;
169 outf = NULL;
170 lockname = NULL;
171 funcname = NULL;
172 eventtype = -1;
173 locktype = -1;
174 nbufs = 0;
175 sflag = 0;
176 pflag = 0;
177
178 while ((ch = getopt(argc, argv, "E:F:L:M:N:T:b:ceflo:pst")) != -1)
179 switch (ch) {
180 case 'E':
181 eventtype = matchname(eventnames, optarg);
182 break;
183 case 'F':
184 funcname = optarg;
185 break;
186 case 'L':
187 lockname = optarg;
188 break;
189 case 'N':
190 nlistf = optarg;
191 break;
192 case 'T':
193 locktype = matchname(locknames, optarg);
194 break;
195 case 'b':
196 nbufs = (int)strtol(optarg, &p, 0);
197 if (!isdigit((u_int)*optarg) || *p != '\0')
198 usage();
199 break;
200 case 'c':
201 cflag = 1;
202 break;
203 case 'e':
204 listnames(eventnames);
205 break;
206 case 'l':
207 lflag = 1;
208 break;
209 case 'o':
210 outf = optarg;
211 break;
212 case 'p':
213 pflag = 1;
214 break;
215 case 's':
216 sflag = 1;
217 break;
218 case 't':
219 listnames(locknames);
220 break;
221 default:
222 usage();
223 }
224 argc -= optind;
225 argv += optind;
226
227 if (*argv == NULL)
228 usage();
229
230 if (outf) {
231 if ((fd = open(outf, O_WRONLY | O_CREAT, 0600)) == -1)
232 err(EXIT_FAILURE, "opening %s", outf);
233 outfp = fdopen(fd, "w");
234 } else
235 outfp = stdout;
236
237 /*
238 * Find the name list for resolving symbol names, and load it into
239 * memory.
240 */
241 if (nlistf == NULL) {
242 nlfd = open(_PATH_KSYMS, O_RDONLY);
243 nlistf = getbootfile();
244 } else
245 nlfd = -1;
246 if (nlfd == -1) {
247 if ((nlfd = open(nlistf, O_RDONLY)) < 0)
248 err(EXIT_FAILURE, "cannot open " _PATH_KSYMS " or %s",
249 nlistf);
250 }
251 if (loadsym32(nlfd) != 0) {
252 if (loadsym64(nlfd) != 0)
253 errx(EXIT_FAILURE, "unable to load symbol table");
254 bin64 = 1;
255 }
256 close(nlfd);
257
258 memset(&le, 0, sizeof(le));
259 le.le_nbufs = nbufs;
260
261 /*
262 * Set up initial filtering.
263 */
264 if (lockname != NULL) {
265 findsym(LOCK_BYNAME, lockname, &le.le_lock, NULL);
266 le.le_flags |= LE_ONE_LOCK;
267 }
268 if (!lflag)
269 le.le_flags |= LE_CALLSITE;
270 if (funcname != NULL) {
271 if (lflag)
272 usage();
273 findsym(FUNC_BYNAME, funcname, &le.le_csstart, &le.le_csend);
274 le.le_flags |= LE_ONE_CALLSITE;
275 }
276 le.le_mask = (eventtype & LB_EVENT_MASK) | (locktype & LB_LOCK_MASK);
277
278 /*
279 * Start tracing.
280 */
281 if ((lsfd = open(_PATH_DEV_LOCKSTAT, O_RDONLY)) < 0)
282 err(EXIT_FAILURE, "cannot open " _PATH_DEV_LOCKSTAT);
283 if (ioctl(lsfd, IOC_LOCKSTAT_GVERSION, &ch) < 0)
284 err(EXIT_FAILURE, "ioctl");
285 if (ch != LS_VERSION)
286 errx(EXIT_FAILURE, "incompatible lockstat interface version");
287 if (ioctl(lsfd, IOC_LOCKSTAT_ENABLE, &le))
288 err(EXIT_FAILURE, "cannot enable tracing");
289
290 /*
291 * Execute the traced program.
292 */
293 spawn(argc, argv);
294
295 /*
296 * Stop tracing, and read the trace buffers from the kernel.
297 */
298 if (ioctl(lsfd, IOC_LOCKSTAT_DISABLE, &ld) == -1) {
299 if (errno == EOVERFLOW) {
300 warnx("overflowed available kernel trace buffers");
301 exit(EXIT_FAILURE);
302 }
303 err(EXIT_FAILURE, "cannot disable tracing");
304 }
305 if ((bufs = malloc(ld.ld_size)) == NULL)
306 err(EXIT_FAILURE, "cannot allocate memory for user buffers");
307 if (read(lsfd, bufs, ld.ld_size) != ld.ld_size)
308 err(EXIT_FAILURE, "reading from " _PATH_DEV_LOCKSTAT);
309 if (close(lsfd))
310 err(EXIT_FAILURE, "close(" _PATH_DEV_LOCKSTAT ")");
311
312 /*
313 * Figure out how to scale the results, and build the lists. For
314 * internal use we convert all times from CPU frequency based to
315 * picoseconds, and values are eventually displayed in ms.
316 */
317 for (i = 0; i < sizeof(ld.ld_freq) / sizeof(ld.ld_freq[0]); i++)
318 if (ld.ld_freq[i] != 0)
319 cpuscale[i] = PICO / ld.ld_freq[i];
320 ms = ld.ld_time.tv_sec * MILLI + ld.ld_time.tv_nsec / MICRO;
321 if (pflag)
322 cscale = 1.0 / ncpu();
323 else
324 cscale = 1.0;
325 cscale *= (sflag ? MILLI / ms : 1.0);
326 tscale = cscale / NANO;
327 nbufs = (int)(ld.ld_size / sizeof(lsbuf_t));
328 makelists();
329
330 /*
331 * Display the results.
332 */
333 fprintf(outfp, "Elapsed time: %.2f seconds.", ms / MILLI);
334 if (sflag || pflag) {
335 fprintf(outfp, " Displaying ");
336 if (pflag)
337 fprintf(outfp, "per-CPU ");
338 if (sflag)
339 fprintf(outfp, "per-second ");
340 fprintf(outfp, "averages.");
341 }
342 putc('\n', outfp);
343
344 for (name = alltypes; name->name != NULL; name++) {
345 if (eventtype != -1 &&
346 (name->mask & LB_EVENT_MASK) != eventtype)
347 continue;
348 if (locktype != -1 &&
349 (name->mask & LB_LOCK_MASK) != locktype)
350 continue;
351
352 display(name->mask, name->name);
353 }
354
355 if (displayed == 0)
356 fprintf(outfp, "None of the selected events were recorded.\n");
357 exit(EXIT_SUCCESS);
358 }
359
360 void
361 usage(void)
362 {
363
364 fprintf(stderr,
365 "%s: usage:\n"
366 "%s [options] <command>\n\n"
367 "-F func\t\tlimit trace to one function\n"
368 "-E evt\t\tdisplay only one type of event\n"
369 "-L lock\t\tlimit trace to one lock (name, or address)\n"
370 "-N nlist\tspecify name list file\n"
371 "-T type\t\tdisplay only one type of lock\n"
372 "-b nbuf\t\tset number of event buffers to allocate\n"
373 "-c\t\treport percentage of total events by count, not time\n"
374 "-e\t\tlist event types\n"
375 "-l\t\ttrace only by lock\n"
376 "-o file\t\tsend output to named file, not stdout\n"
377 "-p\t\tshow average count/time per CPU, not total\n"
378 "-s\t\tshow average count/time per second, not total\n"
379 "-t\t\tlist lock types\n",
380 getprogname(), getprogname());
381
382 exit(EXIT_FAILURE);
383 }
384
385 void
386 nullsig(int junk)
387 {
388
389 (void)junk;
390 }
391
392 void
393 listnames(const name_t *name)
394 {
395
396 for (; name->name != NULL; name++)
397 printf("%s\n", name->name);
398
399 exit(EXIT_SUCCESS);
400 }
401
402 int
403 matchname(const name_t *name, const char *string)
404 {
405
406 for (; name->name != NULL; name++)
407 if (strcasecmp(name->name, string) == 0)
408 return name->mask;
409
410 warnx("unknown type `%s'", string);
411 usage();
412 return 0;
413 }
414
415 /*
416 * Return the number of CPUs in the running system.
417 */
418 int
419 ncpu(void)
420 {
421 int rv, mib[2];
422 size_t varlen;
423
424 mib[0] = CTL_HW;
425 mib[1] = HW_NCPU;
426 varlen = sizeof(rv);
427 if (sysctl(mib, 2, &rv, &varlen, NULL, (size_t)0) < 0)
428 rv = 1;
429
430 return (rv);
431 }
432
433 /*
434 * Call into the ELF parser and look up a symbol by name or by address.
435 */
436 void
437 findsym(findsym_t find, char *name, uintptr_t *start, uintptr_t *end)
438 {
439 uintptr_t tend;
440 char *p;
441 int rv;
442
443 if (end == NULL)
444 end = &tend;
445
446 if (find == LOCK_BYNAME) {
447 if (isdigit((u_int)name[0])) {
448 *start = (uintptr_t)strtoul(name, &p, 0);
449 if (*p == '\0')
450 return;
451 }
452 }
453
454 if (bin64)
455 rv = findsym64(find, name, start, end);
456 else
457 rv = findsym32(find, name, start, end);
458
459 if (find == FUNC_BYNAME || find == LOCK_BYNAME) {
460 if (rv == -1)
461 errx(EXIT_FAILURE, "unable to find symbol `%s'", name);
462 return;
463 }
464
465 if (rv == -1)
466 sprintf(name, "0x%016lx", (long)*start);
467 }
468
469 /*
470 * Fork off the child process and wait for it to complete. We trap SIGINT
471 * so that the caller can use Ctrl-C to stop tracing early and still get
472 * useful results.
473 */
474 void
475 spawn(int argc, char **argv)
476 {
477 pid_t pid;
478
479 switch (pid = fork()) {
480 case 0:
481 close(lsfd);
482 if (execvp(argv[0], argv) == -1)
483 err(EXIT_FAILURE, "cannot exec");
484 break;
485 case -1:
486 err(EXIT_FAILURE, "cannot fork to exec");
487 break;
488 default:
489 signal(SIGINT, nullsig);
490 wait(NULL);
491 signal(SIGINT, SIG_DFL);
492 break;
493 }
494 }
495
496 /*
497 * From the kernel supplied data, construct two dimensional lists of locks
498 * and event buffers, indexed by lock type.
499 */
500 void
501 makelists(void)
502 {
503 lsbuf_t *lb, *lb2, *max;
504 int i, type;
505 lock_t *l;
506
507 for (i = 0; i < LB_NLOCK; i++)
508 TAILQ_INIT(&locklist[i]);
509
510 for (lb = bufs, max = bufs + nbufs; lb < max; lb++) {
511 if (lb->lb_flags == 0)
512 continue;
513
514 /*
515 * Look for a record descibing this lock, and allocate a
516 * new one if needed.
517 */
518 type = ((lb->lb_flags & LB_LOCK_MASK) >> LB_LOCK_SHIFT) - 1;
519 TAILQ_FOREACH(l, &locklist[type], chain) {
520 if (l->lock == lb->lb_lock)
521 break;
522 }
523 if (l == NULL) {
524 l = (lock_t *)malloc(sizeof(*l));
525 l->flags = lb->lb_flags;
526 l->lock = lb->lb_lock;
527 memset(&l->counts, 0, sizeof(l->counts));
528 memset(&l->times, 0, sizeof(l->times));
529 TAILQ_INIT(&l->bufs);
530 TAILQ_INSERT_TAIL(&locklist[type], l, chain);
531 }
532
533 /*
534 * Scale the time values per buffer and summarise
535 * times+counts per lock.
536 */
537 for (i = 0; i < LB_NEVENT; i++) {
538 lb->lb_times[i] *= cpuscale[lb->lb_cpu];
539 l->counts[i] += lb->lb_counts[i];
540 l->times[i] += lb->lb_times[i];
541 }
542
543 /*
544 * Merge same lock+callsite pairs from multiple CPUs
545 * together.
546 */
547 TAILQ_FOREACH(lb2, &l->bufs, lb_chain.tailq) {
548 if (lb->lb_callsite == lb2->lb_callsite)
549 break;
550 }
551 if (lb2 != NULL) {
552 for (i = 0; i < LB_NEVENT; i++) {
553 lb2->lb_counts[i] += lb->lb_counts[i];
554 lb2->lb_times[i] += lb->lb_times[i];
555 }
556 } else
557 TAILQ_INSERT_HEAD(&l->bufs, lb, lb_chain.tailq);
558 }
559 }
560
561 /*
562 * Re-sort one list of locks / lock buffers by event type.
563 */
564 void
565 resort(int type, int event)
566 {
567 lsbuf_t *lb, *lb2;
568 locklist_t llist;
569 buflist_t blist;
570 lock_t *l, *l2;
571
572 TAILQ_INIT(&llist);
573 while ((l = TAILQ_FIRST(&locklist[type])) != NULL) {
574 TAILQ_REMOVE(&locklist[type], l, chain);
575
576 /*
577 * Sort the buffers into the per-lock list.
578 */
579 TAILQ_INIT(&blist);
580 while ((lb = TAILQ_FIRST(&l->bufs)) != NULL) {
581 TAILQ_REMOVE(&l->bufs, lb, lb_chain.tailq);
582
583 lb2 = TAILQ_FIRST(&blist);
584 while (lb2 != NULL) {
585 if (cflag) {
586 if (lb->lb_counts[event] >
587 lb2->lb_counts[event])
588 break;
589 } else if (lb->lb_times[event] >
590 lb2->lb_times[event])
591 break;
592 lb2 = TAILQ_NEXT(lb2, lb_chain.tailq);
593 }
594 if (lb2 == NULL)
595 TAILQ_INSERT_TAIL(&blist, lb, lb_chain.tailq);
596 else
597 TAILQ_INSERT_BEFORE(lb2, lb, lb_chain.tailq);
598 }
599 l->bufs = blist;
600
601 /*
602 * Sort this lock into the per-type list, based on the
603 * totals per lock.
604 */
605 l2 = TAILQ_FIRST(&llist);
606 while (l2 != NULL) {
607 if (cflag) {
608 if (l->counts[event] > l2->counts[event])
609 break;
610 } else if (l->times[event] > l2->times[event])
611 break;
612 l2 = TAILQ_NEXT(l2, chain);
613 }
614 if (l2 == NULL)
615 TAILQ_INSERT_TAIL(&llist, l, chain);
616 else
617 TAILQ_INSERT_BEFORE(l2, l, chain);
618 }
619 locklist[type] = llist;
620 }
621
622 /*
623 * Display a summary table for one lock type / event type pair.
624 */
625 void
626 display(int mask, const char *name)
627 {
628 lock_t *l;
629 lsbuf_t *lb;
630 int event, type;
631 double pcscale, metric;
632 char lname[256], fname[256];
633
634 type = ((mask & LB_LOCK_MASK) >> LB_LOCK_SHIFT) - 1;
635 if (TAILQ_FIRST(&locklist[type]) == NULL)
636 return;
637
638 event = (mask & LB_EVENT_MASK) - 1;
639 resort(type, event);
640
641 fprintf(outfp, "\n-- %s\n\n"
642 "Total%% Count Time/ms Lock Caller\n"
643 "------ ------- --------- ------------------ ----------------------------------\n",
644 name);
645
646 /*
647 * Sum up all events for this type of lock + event.
648 */
649 pcscale = 0;
650 TAILQ_FOREACH(l, &locklist[type], chain) {
651 if (cflag)
652 pcscale += l->counts[event];
653 else
654 pcscale += l->times[event];
655 displayed++;
656 }
657 if (pcscale == 0)
658 pcscale = 100;
659 else
660 pcscale = (100.0 / pcscale);
661
662 /*
663 * For each lock, print a summary total, followed by a breakdown by
664 * caller.
665 */
666 TAILQ_FOREACH(l, &locklist[type], chain) {
667 if (cflag)
668 metric = l->counts[event];
669 else
670 metric = l->times[event];
671 metric *= pcscale;
672
673 findsym(LOCK_BYADDR, lname, &l->lock, NULL);
674
675 fprintf(outfp, "%6.2f %7d %9.2f %-18s <all>\n", metric,
676 (int)(l->counts[event] * cscale),
677 l->times[event] * tscale, lname);
678
679 if (lflag)
680 continue;
681
682 TAILQ_FOREACH(lb, &l->bufs, lb_chain.tailq) {
683 if (cflag)
684 metric = lb->lb_counts[event];
685 else
686 metric = lb->lb_times[event];
687 metric *= pcscale;
688
689 findsym(LOCK_BYADDR, lname, &lb->lb_lock, NULL);
690 findsym(FUNC_BYADDR, fname, &lb->lb_callsite, NULL);
691 fprintf(outfp, "%6.2f %7d %9.2f %-18s %s\n", metric,
692 (int)(lb->lb_counts[event] * cscale),
693 lb->lb_times[event] * tscale,
694 lname, fname);
695 }
696 }
697 }
698