perf report: Display user/kernel differentiator
[deliverable/linux.git] / Documentation / perf_counter / builtin-record.c
CommitLineData
abaff32a 1/*
bf9e1876
IM
2 * builtin-record.c
3 *
4 * Builtin record command: Record the profile of a workload
5 * (or a CPU, or a PID) into the perf.data output file - for
6 * later analysis via perf report.
abaff32a 7 */
16f762a2 8#include "builtin.h"
bf9e1876
IM
9
10#include "perf.h"
11
6eda5838 12#include "util/util.h"
0e9b20b8 13#include "util/parse-options.h"
8ad8db37 14#include "util/parse-events.h"
a0055ae2 15#include "util/string.h"
6eda5838 16
97124d5e 17#include <unistd.h>
de9ac07b 18#include <sched.h>
de9ac07b 19
0e9b20b8
IM
20#define ALIGN(x, a) __ALIGN_MASK(x, (typeof(x))(a)-1)
21#define __ALIGN_MASK(x, mask) (((x)+(mask))&~(mask))
1a853e36 22
e61078a0
PZ
23static long default_interval = 100000;
24static long event_count[MAX_COUNTERS];
8ad8db37 25
de9ac07b 26static int fd[MAX_NR_CPUS][MAX_COUNTERS];
3cf165fc 27static int nr_cpus = 0;
de9ac07b 28static unsigned int page_size;
3cf165fc 29static unsigned int mmap_pages = 128;
cf1f4574 30static int freq = 0;
de9ac07b 31static int output;
23ac9cbe 32static const char *output_name = "perf.data";
de9ac07b 33static int group = 0;
16c8a109
PZ
34static unsigned int realtime_prio = 0;
35static int system_wide = 0;
1a853e36 36static pid_t target_pid = -1;
16c8a109 37static int inherit = 1;
97124d5e 38static int force = 0;
abaff32a 39static int append_file = 0;
de9ac07b
PZ
40
41const unsigned int default_count[] = {
42 1000000,
43 1000000,
44 10000,
45 10000,
46 1000000,
47 10000,
48};
49
de9ac07b
PZ
50struct mmap_data {
51 int counter;
52 void *base;
53 unsigned int mask;
54 unsigned int prev;
55};
56
57static unsigned int mmap_read_head(struct mmap_data *md)
58{
59 struct perf_counter_mmap_page *pc = md->base;
60 int head;
61
62 head = pc->data_head;
63 rmb();
64
65 return head;
66}
67
68static long events;
69static struct timeval last_read, this_read;
70
021e9f47
IM
71static __u64 bytes_written;
72
de9ac07b
PZ
73static void mmap_read(struct mmap_data *md)
74{
75 unsigned int head = mmap_read_head(md);
76 unsigned int old = md->prev;
77 unsigned char *data = md->base + page_size;
78 unsigned long size;
79 void *buf;
80 int diff;
81
82 gettimeofday(&this_read, NULL);
83
84 /*
85 * If we're further behind than half the buffer, there's a chance
86 * the writer will bite our tail and screw up the events under us.
87 *
88 * If we somehow ended up ahead of the head, we got messed up.
89 *
90 * In either case, truncate and restart at head.
91 */
92 diff = head - old;
93 if (diff > md->mask / 2 || diff < 0) {
94 struct timeval iv;
95 unsigned long msecs;
96
97 timersub(&this_read, &last_read, &iv);
98 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
99
100 fprintf(stderr, "WARNING: failed to keep up with mmap data."
101 " Last read %lu msecs ago.\n", msecs);
102
103 /*
104 * head points to a known good entry, start there.
105 */
106 old = head;
107 }
108
109 last_read = this_read;
110
111 if (old != head)
112 events++;
113
114 size = head - old;
115
116 if ((old & md->mask) + size != (head & md->mask)) {
117 buf = &data[old & md->mask];
118 size = md->mask + 1 - (old & md->mask);
119 old += size;
021e9f47 120
de9ac07b
PZ
121 while (size) {
122 int ret = write(output, buf, size);
021e9f47
IM
123
124 if (ret < 0)
125 die("failed to write");
126
de9ac07b
PZ
127 size -= ret;
128 buf += ret;
021e9f47
IM
129
130 bytes_written += ret;
de9ac07b
PZ
131 }
132 }
133
134 buf = &data[old & md->mask];
135 size = head - old;
136 old += size;
021e9f47 137
de9ac07b
PZ
138 while (size) {
139 int ret = write(output, buf, size);
021e9f47
IM
140
141 if (ret < 0)
142 die("failed to write");
143
de9ac07b
PZ
144 size -= ret;
145 buf += ret;
021e9f47
IM
146
147 bytes_written += ret;
de9ac07b
PZ
148 }
149
150 md->prev = old;
151}
152
153static volatile int done = 0;
154
16c8a109 155static void sig_handler(int sig)
de9ac07b 156{
16c8a109 157 done = 1;
de9ac07b
PZ
158}
159
16c8a109
PZ
160static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
161static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
162
163static int nr_poll;
164static int nr_cpu;
165
1a853e36 166struct mmap_event {
16f762a2
IM
167 struct perf_event_header header;
168 __u32 pid;
169 __u32 tid;
170 __u64 start;
171 __u64 len;
172 __u64 pgoff;
173 char filename[PATH_MAX];
1a853e36 174};
16f762a2 175
1a853e36 176struct comm_event {
16f762a2
IM
177 struct perf_event_header header;
178 __u32 pid;
179 __u32 tid;
180 char comm[16];
1a853e36
ACM
181};
182
f70e87d7 183static void pid_synthesize_comm_event(pid_t pid, int full)
1a853e36 184{
16f762a2 185 struct comm_event comm_ev;
1a853e36
ACM
186 char filename[PATH_MAX];
187 char bf[BUFSIZ];
a0055ae2 188 int fd, ret;
1a853e36 189 size_t size;
a0055ae2 190 char *field, *sep;
f70e87d7
PZ
191 DIR *tasks;
192 struct dirent dirent, *next;
1a853e36
ACM
193
194 snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);
195
196 fd = open(filename, O_RDONLY);
197 if (fd < 0) {
198 fprintf(stderr, "couldn't open %s\n", filename);
199 exit(EXIT_FAILURE);
200 }
201 if (read(fd, bf, sizeof(bf)) < 0) {
202 fprintf(stderr, "couldn't read %s\n", filename);
203 exit(EXIT_FAILURE);
204 }
205 close(fd);
206
a0055ae2 207 /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
1a853e36 208 memset(&comm_ev, 0, sizeof(comm_ev));
a0055ae2
ACM
209 field = strchr(bf, '(');
210 if (field == NULL)
211 goto out_failure;
212 sep = strchr(++field, ')');
213 if (sep == NULL)
214 goto out_failure;
215 size = sep - field;
216 memcpy(comm_ev.comm, field, size++);
f70e87d7
PZ
217
218 comm_ev.pid = pid;
1a853e36 219 comm_ev.header.type = PERF_EVENT_COMM;
1a853e36
ACM
220 size = ALIGN(size, sizeof(uint64_t));
221 comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
16f762a2 222
f70e87d7
PZ
223 if (!full) {
224 comm_ev.tid = pid;
225
226 ret = write(output, &comm_ev, comm_ev.header.size);
227 if (ret < 0) {
228 perror("failed to write");
229 exit(-1);
230 }
231 return;
232 }
233
234 snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
235
236 tasks = opendir(filename);
237 while (!readdir_r(tasks, &dirent, &next) && next) {
238 char *end;
239 pid = strtol(dirent.d_name, &end, 10);
240 if (*end)
241 continue;
242
243 comm_ev.tid = pid;
244
245 ret = write(output, &comm_ev, comm_ev.header.size);
246 if (ret < 0) {
247 perror("failed to write");
248 exit(-1);
249 }
1a853e36 250 }
f70e87d7
PZ
251 closedir(tasks);
252 return;
253
a0055ae2
ACM
254out_failure:
255 fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
256 filename);
257 exit(EXIT_FAILURE);
1a853e36
ACM
258}
259
f70e87d7 260static void pid_synthesize_mmap_events(pid_t pid)
1a853e36
ACM
261{
262 char filename[PATH_MAX];
263 FILE *fp;
264
265 snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
266
267 fp = fopen(filename, "r");
268 if (fp == NULL) {
269 fprintf(stderr, "couldn't open %s\n", filename);
270 exit(EXIT_FAILURE);
271 }
272 while (1) {
a0055ae2 273 char bf[BUFSIZ], *pbf = bf;
1a853e36
ACM
274 struct mmap_event mmap_ev = {
275 .header.type = PERF_EVENT_MMAP,
276 };
a0055ae2 277 int n;
1a853e36
ACM
278 size_t size;
279 if (fgets(bf, sizeof(bf), fp) == NULL)
280 break;
281
282 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
a0055ae2
ACM
283 n = hex2u64(pbf, &mmap_ev.start);
284 if (n < 0)
285 continue;
286 pbf += n + 1;
287 n = hex2u64(pbf, &mmap_ev.len);
288 if (n < 0)
289 continue;
290 pbf += n + 3;
291 if (*pbf == 'x') { /* vm_exec */
1a853e36
ACM
292 char *execname = strrchr(bf, ' ');
293
294 if (execname == NULL || execname[1] != '/')
295 continue;
296
297 execname += 1;
298 size = strlen(execname);
299 execname[size - 1] = '\0'; /* Remove \n */
300 memcpy(mmap_ev.filename, execname, size);
301 size = ALIGN(size, sizeof(uint64_t));
302 mmap_ev.len -= mmap_ev.start;
303 mmap_ev.header.size = (sizeof(mmap_ev) -
304 (sizeof(mmap_ev.filename) - size));
f70e87d7 305 mmap_ev.pid = pid;
1a853e36
ACM
306 mmap_ev.tid = pid;
307
308 if (write(output, &mmap_ev, mmap_ev.header.size) < 0) {
309 perror("failed to write");
310 exit(-1);
311 }
312 }
313 }
314
315 fclose(fp);
316}
317
f70e87d7
PZ
318static void synthesize_events(void)
319{
320 DIR *proc;
321 struct dirent dirent, *next;
322
323 proc = opendir("/proc");
324
325 while (!readdir_r(proc, &dirent, &next) && next) {
326 char *end;
327 pid_t pid;
328
329 pid = strtol(dirent.d_name, &end, 10);
330 if (*end) /* only interested in proper numerical dirents */
331 continue;
332
333 pid_synthesize_comm_event(pid, 1);
334 pid_synthesize_mmap_events(pid);
335 }
336
337 closedir(proc);
338}
339
f250c030
IM
340static int group_fd;
341
342static void create_counter(int counter, int cpu, pid_t pid)
de9ac07b 343{
c70975bc 344 struct perf_counter_attr attr;
16c8a109 345 int track = 1;
16c8a109 346
f250c030
IM
347 memset(&attr, 0, sizeof(attr));
348 attr.config = event_id[counter];
349 attr.sample_period = event_count[counter];
350 attr.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
cf1f4574 351 attr.freq = freq;
f250c030
IM
352 attr.mmap = track;
353 attr.comm = track;
cf1f4574 354 attr.inherit = (cpu < 0) && inherit;
16c8a109 355
f250c030 356 track = 0; /* only the first counter needs these */
16c8a109 357
f250c030 358 fd[nr_cpu][counter] = sys_perf_counter_open(&attr, pid, cpu, group_fd, 0);
16c8a109 359
f250c030
IM
360 if (fd[nr_cpu][counter] < 0) {
361 int err = errno;
16c8a109 362
f250c030
IM
363 error("syscall returned with %d (%s)\n",
364 fd[nr_cpu][counter], strerror(err));
365 if (err == EPERM)
366 printf("Are you root?\n");
367 exit(-1);
368 }
369 assert(fd[nr_cpu][counter] >= 0);
370 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
16c8a109 371
f250c030
IM
372 /*
373 * First counter acts as the group leader:
374 */
375 if (group && group_fd == -1)
376 group_fd = fd[nr_cpu][counter];
377
378 event_array[nr_poll].fd = fd[nr_cpu][counter];
379 event_array[nr_poll].events = POLLIN;
380 nr_poll++;
381
382 mmap_array[nr_cpu][counter].counter = counter;
383 mmap_array[nr_cpu][counter].prev = 0;
384 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
385 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
386 PROT_READ, MAP_SHARED, fd[nr_cpu][counter], 0);
387 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
388 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
389 exit(-1);
390 }
391}
f2521b6e 392
f250c030
IM
393static void open_counters(int cpu, pid_t pid)
394{
395 int counter;
16c8a109 396
f250c030
IM
397 if (pid > 0) {
398 pid_synthesize_comm_event(pid, 0);
399 pid_synthesize_mmap_events(pid);
16c8a109 400 }
f250c030
IM
401
402 group_fd = -1;
403 for (counter = 0; counter < nr_counters; counter++)
404 create_counter(counter, cpu, pid);
405
16c8a109
PZ
406 nr_cpu++;
407}
408
0e9b20b8 409static int __cmd_record(int argc, const char **argv)
16c8a109
PZ
410{
411 int i, counter;
abaff32a 412 struct stat st;
de9ac07b 413 pid_t pid;
abaff32a 414 int flags;
de9ac07b
PZ
415 int ret;
416
417 page_size = sysconf(_SC_PAGE_SIZE);
de9ac07b
PZ
418 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
419 assert(nr_cpus <= MAX_NR_CPUS);
420 assert(nr_cpus >= 0);
421
abaff32a
IM
422 if (!stat(output_name, &st) && !force && !append_file) {
423 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
97124d5e
PZ
424 output_name);
425 exit(-1);
426 }
427
abaff32a
IM
428 flags = O_CREAT|O_RDWR;
429 if (append_file)
430 flags |= O_APPEND;
431 else
432 flags |= O_TRUNC;
433
434 output = open(output_name, flags, S_IRUSR|S_IWUSR);
de9ac07b
PZ
435 if (output < 0) {
436 perror("failed to create output file");
437 exit(-1);
438 }
439
1a853e36 440 if (!system_wide) {
df97992c 441 open_counters(-1, target_pid != -1 ? target_pid : getpid());
1a853e36
ACM
442 } else for (i = 0; i < nr_cpus; i++)
443 open_counters(i, target_pid);
de9ac07b 444
16c8a109
PZ
445 signal(SIGCHLD, sig_handler);
446 signal(SIGINT, sig_handler);
de9ac07b 447
ef65b2a0 448 if (target_pid == -1 && argc) {
1a853e36
ACM
449 pid = fork();
450 if (pid < 0)
451 perror("failed to fork");
de9ac07b 452
1a853e36 453 if (!pid) {
0e9b20b8 454 if (execvp(argv[0], (char **)argv)) {
1a853e36
ACM
455 perror(argv[0]);
456 exit(-1);
457 }
de9ac07b
PZ
458 }
459 }
460
461 if (realtime_prio) {
462 struct sched_param param;
463
464 param.sched_priority = realtime_prio;
465 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
466 printf("Could not set realtime priority.\n");
467 exit(-1);
468 }
469 }
470
f70e87d7
PZ
471 if (system_wide)
472 synthesize_events();
de9ac07b
PZ
473
474 while (!done) {
475 int hits = events;
476
16c8a109 477 for (i = 0; i < nr_cpu; i++) {
de9ac07b
PZ
478 for (counter = 0; counter < nr_counters; counter++)
479 mmap_read(&mmap_array[i][counter]);
480 }
481
482 if (hits == events)
483 ret = poll(event_array, nr_poll, 100);
484 }
485
021e9f47
IM
486 /*
487 * Approximate RIP event size: 24 bytes.
488 */
489 fprintf(stderr,
490 "[ perf record: Captured and wrote %.3f MB %s (~%lld events) ]\n",
491 (double)bytes_written / 1024.0 / 1024.0,
492 output_name,
493 bytes_written / 24);
addc2785 494
de9ac07b
PZ
495 return 0;
496}
0e9b20b8 497
0e9b20b8 498static const char * const record_usage[] = {
9e096753
MG
499 "perf record [<options>] [<command>]",
500 "perf record [<options>] -- <command> [<options>]",
0e9b20b8
IM
501 NULL
502};
503
8ad8db37
IM
504static char events_help_msg[EVENTS_HELP_MAX];
505
5242519b 506static const struct option options[] = {
0e9b20b8 507 OPT_CALLBACK('e', "event", NULL, "event",
8ad8db37 508 events_help_msg, parse_events),
0e9b20b8
IM
509 OPT_INTEGER('p', "pid", &target_pid,
510 "record events on existing pid"),
511 OPT_INTEGER('r', "realtime", &realtime_prio,
512 "collect data with this RT SCHED_FIFO priority"),
513 OPT_BOOLEAN('a', "all-cpus", &system_wide,
514 "system-wide collection from all CPUs"),
abaff32a
IM
515 OPT_BOOLEAN('A', "append", &append_file,
516 "append to the output file to do incremental profiling"),
97124d5e
PZ
517 OPT_BOOLEAN('f', "force", &force,
518 "overwrite existing data file"),
e61078a0 519 OPT_LONG('c', "count", &default_interval,
abaff32a
IM
520 "event period to sample"),
521 OPT_STRING('o', "output", &output_name, "file",
522 "output file name"),
523 OPT_BOOLEAN('i', "inherit", &inherit,
524 "child tasks inherit counters"),
cf1f4574
IM
525 OPT_INTEGER('F', "freq", &freq,
526 "profile at this frequency"),
abaff32a
IM
527 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
528 "number of mmap data pages"),
0e9b20b8
IM
529 OPT_END()
530};
531
532int cmd_record(int argc, const char **argv, const char *prefix)
533{
534 int counter;
535
8ad8db37 536 create_events_help(events_help_msg);
0e9b20b8
IM
537
538 argc = parse_options(argc, argv, options, record_usage, 0);
ef65b2a0 539 if (!argc && target_pid == -1 && !system_wide)
0e9b20b8
IM
540 usage_with_options(record_usage, options);
541
542 if (!nr_counters) {
543 nr_counters = 1;
544 event_id[0] = 0;
545 }
546
cf1f4574
IM
547 if (freq) {
548 default_interval = freq;
549 freq = 1;
550 }
0e9b20b8
IM
551 for (counter = 0; counter < nr_counters; counter++) {
552 if (event_count[counter])
553 continue;
554
555 event_count[counter] = default_interval;
556 }
557
558 return __cmd_record(argc, argv);
559}
This page took 0.061515 seconds and 5 git commands to generate.