Commit | Line | Data |
---|---|---|
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 | */ |
b8f46c5a XG |
8 | #define _FILE_OFFSET_BITS 64 |
9 | ||
16f762a2 | 10 | #include "builtin.h" |
bf9e1876 IM |
11 | |
12 | #include "perf.h" | |
13 | ||
6122e4e4 | 14 | #include "util/build-id.h" |
6eda5838 | 15 | #include "util/util.h" |
0e9b20b8 | 16 | #include "util/parse-options.h" |
8ad8db37 | 17 | #include "util/parse-events.h" |
6eda5838 | 18 | |
7c6a1c65 | 19 | #include "util/header.h" |
66e274f3 | 20 | #include "util/event.h" |
8f28827a | 21 | #include "util/debug.h" |
94c744b6 | 22 | #include "util/session.h" |
8d06367f | 23 | #include "util/symbol.h" |
a12b51c4 | 24 | #include "util/cpumap.h" |
7c6a1c65 | 25 | |
97124d5e | 26 | #include <unistd.h> |
de9ac07b | 27 | #include <sched.h> |
de9ac07b | 28 | |
d6d901c2 | 29 | static int *fd[MAX_NR_CPUS][MAX_COUNTERS]; |
a21ca2ca | 30 | |
7e4ff9e3 | 31 | static long default_interval = 0; |
a21ca2ca | 32 | |
42e59d7d | 33 | static int nr_cpus = 0; |
de9ac07b | 34 | static unsigned int page_size; |
42e59d7d IM |
35 | static unsigned int mmap_pages = 128; |
36 | static int freq = 1000; | |
de9ac07b | 37 | static int output; |
529870e3 | 38 | static int pipe_output = 0; |
23ac9cbe | 39 | static const char *output_name = "perf.data"; |
42e59d7d IM |
40 | static int group = 0; |
41 | static unsigned int realtime_prio = 0; | |
c0555642 IM |
42 | static bool raw_samples = false; |
43 | static bool system_wide = false; | |
42e59d7d IM |
44 | static int profile_cpu = -1; |
45 | static pid_t target_pid = -1; | |
d6d901c2 ZY |
46 | static pid_t target_tid = -1; |
47 | static pid_t *all_tids = NULL; | |
48 | static int thread_num = 0; | |
42e59d7d | 49 | static pid_t child_pid = -1; |
c0555642 IM |
50 | static bool inherit = true; |
51 | static bool force = false; | |
52 | static bool append_file = false; | |
53 | static bool call_graph = false; | |
54 | static bool inherit_stat = false; | |
55 | static bool no_samples = false; | |
56 | static bool sample_address = false; | |
57 | static bool multiplex = false; | |
42e59d7d IM |
58 | static int multiplex_fd = -1; |
59 | ||
60 | static long samples = 0; | |
a21ca2ca IM |
61 | static struct timeval last_read; |
62 | static struct timeval this_read; | |
63 | ||
42e59d7d | 64 | static u64 bytes_written = 0; |
a21ca2ca | 65 | |
d6d901c2 | 66 | static struct pollfd *event_array; |
a21ca2ca | 67 | |
42e59d7d IM |
68 | static int nr_poll = 0; |
69 | static int nr_cpu = 0; | |
a21ca2ca | 70 | |
42e59d7d | 71 | static int file_new = 1; |
6122e4e4 | 72 | static off_t post_processing_offset; |
7c6a1c65 | 73 | |
94c744b6 | 74 | static struct perf_session *session; |
f5970550 | 75 | |
de9ac07b | 76 | struct mmap_data { |
a21ca2ca IM |
77 | int counter; |
78 | void *base; | |
79 | unsigned int mask; | |
80 | unsigned int prev; | |
de9ac07b PZ |
81 | }; |
82 | ||
d6d901c2 | 83 | static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS]; |
a21ca2ca | 84 | |
9d91a6f7 | 85 | static unsigned long mmap_read_head(struct mmap_data *md) |
de9ac07b | 86 | { |
cdd6c482 | 87 | struct perf_event_mmap_page *pc = md->base; |
9d91a6f7 | 88 | long head; |
de9ac07b PZ |
89 | |
90 | head = pc->data_head; | |
91 | rmb(); | |
92 | ||
93 | return head; | |
94 | } | |
95 | ||
9d91a6f7 PZ |
96 | static void mmap_write_tail(struct mmap_data *md, unsigned long tail) |
97 | { | |
cdd6c482 | 98 | struct perf_event_mmap_page *pc = md->base; |
9d91a6f7 PZ |
99 | |
100 | /* | |
101 | * ensure all reads are done before we write the tail out. | |
102 | */ | |
103 | /* mb(); */ | |
104 | pc->data_tail = tail; | |
105 | } | |
106 | ||
9215545e TZ |
107 | static void advance_output(size_t size) |
108 | { | |
109 | bytes_written += size; | |
110 | } | |
111 | ||
f5970550 PZ |
112 | static void write_output(void *buf, size_t size) |
113 | { | |
114 | while (size) { | |
115 | int ret = write(output, buf, size); | |
116 | ||
117 | if (ret < 0) | |
118 | die("failed to write"); | |
119 | ||
120 | size -= ret; | |
121 | buf += ret; | |
122 | ||
123 | bytes_written += ret; | |
124 | } | |
125 | } | |
126 | ||
d8f66248 ACM |
127 | static int process_synthesized_event(event_t *event, |
128 | struct perf_session *self __used) | |
234fbbf5 | 129 | { |
6122e4e4 | 130 | write_output(event, event->header.size); |
234fbbf5 ACM |
131 | return 0; |
132 | } | |
133 | ||
de9ac07b PZ |
134 | static void mmap_read(struct mmap_data *md) |
135 | { | |
136 | unsigned int head = mmap_read_head(md); | |
137 | unsigned int old = md->prev; | |
138 | unsigned char *data = md->base + page_size; | |
139 | unsigned long size; | |
140 | void *buf; | |
141 | int diff; | |
142 | ||
143 | gettimeofday(&this_read, NULL); | |
144 | ||
145 | /* | |
146 | * If we're further behind than half the buffer, there's a chance | |
2debbc83 | 147 | * the writer will bite our tail and mess up the samples under us. |
de9ac07b PZ |
148 | * |
149 | * If we somehow ended up ahead of the head, we got messed up. | |
150 | * | |
151 | * In either case, truncate and restart at head. | |
152 | */ | |
153 | diff = head - old; | |
9d91a6f7 | 154 | if (diff < 0) { |
de9ac07b PZ |
155 | struct timeval iv; |
156 | unsigned long msecs; | |
157 | ||
158 | timersub(&this_read, &last_read, &iv); | |
159 | msecs = iv.tv_sec*1000 + iv.tv_usec/1000; | |
160 | ||
161 | fprintf(stderr, "WARNING: failed to keep up with mmap data." | |
162 | " Last read %lu msecs ago.\n", msecs); | |
163 | ||
164 | /* | |
165 | * head points to a known good entry, start there. | |
166 | */ | |
167 | old = head; | |
168 | } | |
169 | ||
170 | last_read = this_read; | |
171 | ||
172 | if (old != head) | |
2debbc83 | 173 | samples++; |
de9ac07b PZ |
174 | |
175 | size = head - old; | |
176 | ||
177 | if ((old & md->mask) + size != (head & md->mask)) { | |
178 | buf = &data[old & md->mask]; | |
179 | size = md->mask + 1 - (old & md->mask); | |
180 | old += size; | |
021e9f47 | 181 | |
6122e4e4 | 182 | write_output(buf, size); |
de9ac07b PZ |
183 | } |
184 | ||
185 | buf = &data[old & md->mask]; | |
186 | size = head - old; | |
187 | old += size; | |
021e9f47 | 188 | |
6122e4e4 | 189 | write_output(buf, size); |
de9ac07b PZ |
190 | |
191 | md->prev = old; | |
9d91a6f7 | 192 | mmap_write_tail(md, old); |
de9ac07b PZ |
193 | } |
194 | ||
195 | static volatile int done = 0; | |
f7b7c26e | 196 | static volatile int signr = -1; |
de9ac07b | 197 | |
16c8a109 | 198 | static void sig_handler(int sig) |
de9ac07b | 199 | { |
16c8a109 | 200 | done = 1; |
f7b7c26e PZ |
201 | signr = sig; |
202 | } | |
203 | ||
204 | static void sig_atexit(void) | |
205 | { | |
933da83a CW |
206 | if (child_pid != -1) |
207 | kill(child_pid, SIGTERM); | |
208 | ||
f7b7c26e PZ |
209 | if (signr == -1) |
210 | return; | |
211 | ||
212 | signal(signr, SIG_DFL); | |
213 | kill(getpid(), signr); | |
de9ac07b PZ |
214 | } |
215 | ||
f250c030 IM |
216 | static int group_fd; |
217 | ||
cdd6c482 | 218 | static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr) |
7c6a1c65 PZ |
219 | { |
220 | struct perf_header_attr *h_attr; | |
221 | ||
94c744b6 ACM |
222 | if (nr < session->header.attrs) { |
223 | h_attr = session->header.attr[nr]; | |
7c6a1c65 PZ |
224 | } else { |
225 | h_attr = perf_header_attr__new(a); | |
dc79c0fc | 226 | if (h_attr != NULL) |
94c744b6 | 227 | if (perf_header__add_attr(&session->header, h_attr) < 0) { |
11deb1f9 ACM |
228 | perf_header_attr__delete(h_attr); |
229 | h_attr = NULL; | |
230 | } | |
7c6a1c65 PZ |
231 | } |
232 | ||
233 | return h_attr; | |
234 | } | |
235 | ||
d6d901c2 | 236 | static void create_counter(int counter, int cpu) |
de9ac07b | 237 | { |
c171b552 | 238 | char *filter = filters[counter]; |
cdd6c482 | 239 | struct perf_event_attr *attr = attrs + counter; |
7c6a1c65 PZ |
240 | struct perf_header_attr *h_attr; |
241 | int track = !counter; /* only the first counter needs these */ | |
d6d901c2 | 242 | int thread_index; |
c171b552 | 243 | int ret; |
7c6a1c65 PZ |
244 | struct { |
245 | u64 count; | |
246 | u64 time_enabled; | |
247 | u64 time_running; | |
248 | u64 id; | |
249 | } read_data; | |
250 | ||
251 | attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | | |
252 | PERF_FORMAT_TOTAL_TIME_RUNNING | | |
253 | PERF_FORMAT_ID; | |
16c8a109 | 254 | |
3a9f131f | 255 | attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID; |
3efa1cc9 | 256 | |
8907fd60 EM |
257 | if (nr_counters > 1) |
258 | attr->sample_type |= PERF_SAMPLE_ID; | |
259 | ||
1dba15e7 | 260 | if (freq) { |
ea1900e5 | 261 | attr->sample_type |= PERF_SAMPLE_PERIOD; |
a21ca2ca IM |
262 | attr->freq = 1; |
263 | attr->sample_freq = freq; | |
1dba15e7 | 264 | } |
3efa1cc9 | 265 | |
649c48a9 PZ |
266 | if (no_samples) |
267 | attr->sample_freq = 0; | |
268 | ||
269 | if (inherit_stat) | |
270 | attr->inherit_stat = 1; | |
271 | ||
4bba828d AB |
272 | if (sample_address) |
273 | attr->sample_type |= PERF_SAMPLE_ADDR; | |
274 | ||
3efa1cc9 IM |
275 | if (call_graph) |
276 | attr->sample_type |= PERF_SAMPLE_CALLCHAIN; | |
277 | ||
cd6feeea | 278 | if (raw_samples) { |
6ddf259d | 279 | attr->sample_type |= PERF_SAMPLE_TIME; |
daac07b2 | 280 | attr->sample_type |= PERF_SAMPLE_RAW; |
cd6feeea IM |
281 | attr->sample_type |= PERF_SAMPLE_CPU; |
282 | } | |
f413cdb8 | 283 | |
a21ca2ca IM |
284 | attr->mmap = track; |
285 | attr->comm = track; | |
60ab2716 | 286 | attr->inherit = inherit; |
46be604b ZY |
287 | if (target_pid == -1 && !system_wide) { |
288 | attr->disabled = 1; | |
bedbfdea | 289 | attr->enable_on_exec = 1; |
46be604b | 290 | } |
bedbfdea | 291 | |
d6d901c2 | 292 | for (thread_index = 0; thread_index < thread_num; thread_index++) { |
3da297a6 | 293 | try_again: |
d6d901c2 ZY |
294 | fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr, |
295 | all_tids[thread_index], cpu, group_fd, 0); | |
296 | ||
297 | if (fd[nr_cpu][counter][thread_index] < 0) { | |
298 | int err = errno; | |
299 | ||
300 | if (err == EPERM || err == EACCES) | |
301 | die("Permission error - are you root?\n" | |
302 | "\t Consider tweaking" | |
303 | " /proc/sys/kernel/perf_event_paranoid.\n"); | |
304 | else if (err == ENODEV && profile_cpu != -1) { | |
305 | die("No such device - did you specify" | |
306 | " an out-of-range profile CPU?\n"); | |
307 | } | |
3da297a6 | 308 | |
d6d901c2 ZY |
309 | /* |
310 | * If it's cycles then fall back to hrtimer | |
311 | * based cpu-clock-tick sw counter, which | |
312 | * is always available even if no PMU support: | |
313 | */ | |
314 | if (attr->type == PERF_TYPE_HARDWARE | |
315 | && attr->config == PERF_COUNT_HW_CPU_CYCLES) { | |
316 | ||
317 | if (verbose) | |
318 | warning(" ... trying to fall back to cpu-clock-ticks\n"); | |
319 | attr->type = PERF_TYPE_SOFTWARE; | |
320 | attr->config = PERF_COUNT_SW_CPU_CLOCK; | |
321 | goto try_again; | |
322 | } | |
323 | printf("\n"); | |
324 | error("perfcounter syscall returned with %d (%s)\n", | |
325 | fd[nr_cpu][counter][thread_index], strerror(err)); | |
bfd45118 SK |
326 | |
327 | #if defined(__i386__) || defined(__x86_64__) | |
d6d901c2 ZY |
328 | if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP) |
329 | die("No hardware sampling interrupt available." | |
330 | " No APIC? If so then you can boot the kernel" | |
331 | " with the \"lapic\" boot parameter to" | |
332 | " force-enable it.\n"); | |
bfd45118 SK |
333 | #endif |
334 | ||
d6d901c2 ZY |
335 | die("No CONFIG_PERF_EVENTS=y kernel support configured?\n"); |
336 | exit(-1); | |
337 | } | |
3da297a6 | 338 | |
d6d901c2 ZY |
339 | h_attr = get_header_attr(attr, counter); |
340 | if (h_attr == NULL) | |
341 | die("nomem\n"); | |
7c6a1c65 | 342 | |
d6d901c2 ZY |
343 | if (!file_new) { |
344 | if (memcmp(&h_attr->attr, attr, sizeof(*attr))) { | |
345 | fprintf(stderr, "incompatible append\n"); | |
346 | exit(-1); | |
347 | } | |
7c6a1c65 | 348 | } |
7c6a1c65 | 349 | |
d6d901c2 ZY |
350 | if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) { |
351 | perror("Unable to read perf file descriptor\n"); | |
352 | exit(-1); | |
353 | } | |
7c6a1c65 | 354 | |
d6d901c2 ZY |
355 | if (perf_header_attr__add_id(h_attr, read_data.id) < 0) { |
356 | pr_warning("Not enough memory to add id\n"); | |
357 | exit(-1); | |
358 | } | |
7c6a1c65 | 359 | |
d6d901c2 ZY |
360 | assert(fd[nr_cpu][counter][thread_index] >= 0); |
361 | fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK); | |
16c8a109 | 362 | |
d6d901c2 ZY |
363 | /* |
364 | * First counter acts as the group leader: | |
365 | */ | |
366 | if (group && group_fd == -1) | |
367 | group_fd = fd[nr_cpu][counter][thread_index]; | |
368 | if (multiplex && multiplex_fd == -1) | |
369 | multiplex_fd = fd[nr_cpu][counter][thread_index]; | |
f250c030 | 370 | |
d6d901c2 | 371 | if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) { |
4502d77c | 372 | |
d6d901c2 ZY |
373 | ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd); |
374 | assert(ret != -1); | |
375 | } else { | |
376 | event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index]; | |
377 | event_array[nr_poll].events = POLLIN; | |
378 | nr_poll++; | |
379 | ||
380 | mmap_array[nr_cpu][counter][thread_index].counter = counter; | |
381 | mmap_array[nr_cpu][counter][thread_index].prev = 0; | |
382 | mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1; | |
383 | mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size, | |
384 | PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0); | |
385 | if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) { | |
386 | error("failed to mmap with %d (%s)\n", errno, strerror(errno)); | |
387 | exit(-1); | |
388 | } | |
ea57c4f5 | 389 | } |
d1302522 | 390 | |
d6d901c2 ZY |
391 | if (filter != NULL) { |
392 | ret = ioctl(fd[nr_cpu][counter][thread_index], | |
393 | PERF_EVENT_IOC_SET_FILTER, filter); | |
394 | if (ret) { | |
395 | error("failed to set filter with %d (%s)\n", errno, | |
396 | strerror(errno)); | |
397 | exit(-1); | |
398 | } | |
c171b552 LZ |
399 | } |
400 | } | |
f250c030 | 401 | } |
f2521b6e | 402 | |
d6d901c2 | 403 | static void open_counters(int cpu) |
f250c030 IM |
404 | { |
405 | int counter; | |
16c8a109 | 406 | |
f250c030 IM |
407 | group_fd = -1; |
408 | for (counter = 0; counter < nr_counters; counter++) | |
d6d901c2 | 409 | create_counter(counter, cpu); |
f250c030 | 410 | |
16c8a109 PZ |
411 | nr_cpu++; |
412 | } | |
413 | ||
6122e4e4 ACM |
414 | static int process_buildids(void) |
415 | { | |
416 | u64 size = lseek(output, 0, SEEK_CUR); | |
417 | ||
9f591fd7 ACM |
418 | if (size == 0) |
419 | return 0; | |
420 | ||
6122e4e4 ACM |
421 | session->fd = output; |
422 | return __perf_session__process_events(session, post_processing_offset, | |
423 | size - post_processing_offset, | |
424 | size, &build_id__mark_dso_hit_ops); | |
425 | } | |
426 | ||
f5970550 PZ |
427 | static void atexit_header(void) |
428 | { | |
c7929e47 TZ |
429 | if (!pipe_output) { |
430 | session->header.data_size += bytes_written; | |
f5970550 | 431 | |
c7929e47 TZ |
432 | process_buildids(); |
433 | perf_header__write(&session->header, output, true); | |
434 | } else { | |
435 | int err; | |
436 | ||
437 | err = event__synthesize_build_ids(process_synthesized_event, | |
438 | session); | |
439 | if (err < 0) | |
440 | pr_err("Couldn't synthesize build ids.\n"); | |
441 | } | |
f5970550 PZ |
442 | } |
443 | ||
d4db3f16 | 444 | static int __cmd_record(int argc, const char **argv) |
16c8a109 PZ |
445 | { |
446 | int i, counter; | |
abaff32a | 447 | struct stat st; |
7c6a1c65 | 448 | pid_t pid = 0; |
abaff32a | 449 | int flags; |
4dc0a04b | 450 | int err; |
8b412664 | 451 | unsigned long waking = 0; |
856e9660 | 452 | int child_ready_pipe[2], go_pipe[2]; |
46be604b | 453 | const bool forks = argc > 0; |
856e9660 | 454 | char buf; |
de9ac07b PZ |
455 | |
456 | page_size = sysconf(_SC_PAGE_SIZE); | |
de9ac07b | 457 | |
f5970550 PZ |
458 | atexit(sig_atexit); |
459 | signal(SIGCHLD, sig_handler); | |
460 | signal(SIGINT, sig_handler); | |
461 | ||
d4db3f16 | 462 | if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) { |
856e9660 PZ |
463 | perror("failed to create pipes"); |
464 | exit(-1); | |
465 | } | |
466 | ||
529870e3 TZ |
467 | if (!strcmp(output_name, "-")) |
468 | pipe_output = 1; | |
469 | else if (!stat(output_name, &st) && st.st_size) { | |
b38d3464 ACM |
470 | if (!force) { |
471 | if (!append_file) { | |
472 | pr_err("Error, output file %s exists, use -A " | |
473 | "to append or -f to overwrite.\n", | |
474 | output_name); | |
475 | exit(-1); | |
476 | } | |
477 | } else { | |
478 | char oldname[PATH_MAX]; | |
479 | snprintf(oldname, sizeof(oldname), "%s.old", | |
480 | output_name); | |
481 | unlink(oldname); | |
482 | rename(output_name, oldname); | |
266e0e21 PH |
483 | } |
484 | } else { | |
c0555642 | 485 | append_file = false; |
97124d5e PZ |
486 | } |
487 | ||
f887f301 | 488 | flags = O_CREAT|O_RDWR; |
abaff32a | 489 | if (append_file) |
f5970550 | 490 | file_new = 0; |
abaff32a IM |
491 | else |
492 | flags |= O_TRUNC; | |
493 | ||
529870e3 TZ |
494 | if (pipe_output) |
495 | output = STDOUT_FILENO; | |
496 | else | |
497 | output = open(output_name, flags, S_IRUSR | S_IWUSR); | |
de9ac07b PZ |
498 | if (output < 0) { |
499 | perror("failed to create output file"); | |
500 | exit(-1); | |
501 | } | |
502 | ||
75be6cf4 | 503 | session = perf_session__new(output_name, O_WRONLY, force); |
94c744b6 | 504 | if (session == NULL) { |
a9a70bbc ACM |
505 | pr_err("Not enough memory for reading perf file header\n"); |
506 | return -1; | |
507 | } | |
508 | ||
4dc0a04b | 509 | if (!file_new) { |
8dc58101 | 510 | err = perf_header__read(session, output); |
4dc0a04b ACM |
511 | if (err < 0) |
512 | return err; | |
513 | } | |
514 | ||
9df37ddd | 515 | if (raw_samples) { |
94c744b6 | 516 | perf_header__set_feat(&session->header, HEADER_TRACE_INFO); |
9df37ddd FW |
517 | } else { |
518 | for (i = 0; i < nr_counters; i++) { | |
519 | if (attrs[i].sample_type & PERF_SAMPLE_RAW) { | |
94c744b6 | 520 | perf_header__set_feat(&session->header, HEADER_TRACE_INFO); |
9df37ddd FW |
521 | break; |
522 | } | |
523 | } | |
524 | } | |
03456a15 | 525 | |
f5970550 PZ |
526 | atexit(atexit_header); |
527 | ||
d4db3f16 | 528 | if (forks) { |
46be604b | 529 | child_pid = fork(); |
856e9660 PZ |
530 | if (pid < 0) { |
531 | perror("failed to fork"); | |
532 | exit(-1); | |
533 | } | |
7c6a1c65 | 534 | |
46be604b | 535 | if (!child_pid) { |
529870e3 TZ |
536 | if (pipe_output) |
537 | dup2(2, 1); | |
856e9660 PZ |
538 | close(child_ready_pipe[0]); |
539 | close(go_pipe[1]); | |
540 | fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); | |
541 | ||
542 | /* | |
543 | * Do a dummy execvp to get the PLT entry resolved, | |
544 | * so we avoid the resolver overhead on the real | |
545 | * execvp call. | |
546 | */ | |
547 | execvp("", (char **)argv); | |
548 | ||
549 | /* | |
550 | * Tell the parent we're ready to go | |
551 | */ | |
552 | close(child_ready_pipe[1]); | |
553 | ||
554 | /* | |
555 | * Wait until the parent tells us to go. | |
556 | */ | |
557 | if (read(go_pipe[0], &buf, 1) == -1) | |
558 | perror("unable to read pipe"); | |
559 | ||
560 | execvp(argv[0], (char **)argv); | |
561 | ||
562 | perror(argv[0]); | |
563 | exit(-1); | |
0a5ac846 | 564 | } |
856e9660 | 565 | |
d6d901c2 ZY |
566 | if (!system_wide && target_tid == -1 && target_pid == -1) |
567 | all_tids[0] = child_pid; | |
568 | ||
856e9660 PZ |
569 | close(child_ready_pipe[1]); |
570 | close(go_pipe[0]); | |
571 | /* | |
572 | * wait for child to settle | |
573 | */ | |
574 | if (read(child_ready_pipe[0], &buf, 1) == -1) { | |
575 | perror("unable to read pipe"); | |
576 | exit(-1); | |
577 | } | |
578 | close(child_ready_pipe[0]); | |
579 | } | |
580 | ||
60ab2716 | 581 | if ((!system_wide && !inherit) || profile_cpu != -1) { |
d6d901c2 | 582 | open_counters(profile_cpu); |
856e9660 | 583 | } else { |
a12b51c4 | 584 | nr_cpus = read_cpu_map(); |
856e9660 | 585 | for (i = 0; i < nr_cpus; i++) |
d6d901c2 | 586 | open_counters(cpumap[i]); |
0a5ac846 | 587 | } |
de9ac07b | 588 | |
529870e3 TZ |
589 | if (pipe_output) { |
590 | err = perf_header__write_pipe(output); | |
591 | if (err < 0) | |
592 | return err; | |
593 | } else if (file_new) { | |
94c744b6 | 594 | err = perf_header__write(&session->header, output, false); |
d5eed904 ACM |
595 | if (err < 0) |
596 | return err; | |
56b03f3c ACM |
597 | } |
598 | ||
6122e4e4 ACM |
599 | post_processing_offset = lseek(output, 0, SEEK_CUR); |
600 | ||
2c46dbb5 TZ |
601 | if (pipe_output) { |
602 | err = event__synthesize_attrs(&session->header, | |
603 | process_synthesized_event, | |
604 | session); | |
605 | if (err < 0) { | |
606 | pr_err("Couldn't synthesize attrs.\n"); | |
607 | return err; | |
608 | } | |
cd19a035 TZ |
609 | |
610 | err = event__synthesize_event_types(process_synthesized_event, | |
611 | session); | |
612 | if (err < 0) { | |
613 | pr_err("Couldn't synthesize event_types.\n"); | |
614 | return err; | |
615 | } | |
9215545e TZ |
616 | |
617 | err = event__synthesize_tracing_data(output, attrs, | |
618 | nr_counters, | |
619 | process_synthesized_event, | |
620 | session); | |
621 | if (err <= 0) { | |
622 | pr_err("Couldn't record tracing data.\n"); | |
623 | return err; | |
624 | } | |
625 | ||
626 | advance_output(err); | |
2c46dbb5 TZ |
627 | } |
628 | ||
56b03f3c ACM |
629 | err = event__synthesize_kernel_mmap(process_synthesized_event, |
630 | session, "_text"); | |
70162138 ACM |
631 | if (err < 0) |
632 | err = event__synthesize_kernel_mmap(process_synthesized_event, | |
633 | session, "_stext"); | |
56b03f3c ACM |
634 | if (err < 0) { |
635 | pr_err("Couldn't record kernel reference relocation symbol.\n"); | |
636 | return err; | |
b7cece76 ACM |
637 | } |
638 | ||
639 | err = event__synthesize_modules(process_synthesized_event, session); | |
640 | if (err < 0) { | |
641 | pr_err("Couldn't record kernel reference relocation symbol.\n"); | |
642 | return err; | |
d5eed904 | 643 | } |
7c6a1c65 | 644 | |
d4db3f16 | 645 | if (!system_wide && profile_cpu == -1) |
d6d901c2 | 646 | event__synthesize_thread(target_tid, process_synthesized_event, |
d8f66248 | 647 | session); |
234fbbf5 | 648 | else |
d8f66248 | 649 | event__synthesize_threads(process_synthesized_event, session); |
7c6a1c65 | 650 | |
de9ac07b PZ |
651 | if (realtime_prio) { |
652 | struct sched_param param; | |
653 | ||
654 | param.sched_priority = realtime_prio; | |
655 | if (sched_setscheduler(0, SCHED_FIFO, ¶m)) { | |
6beba7ad | 656 | pr_err("Could not set realtime priority.\n"); |
de9ac07b PZ |
657 | exit(-1); |
658 | } | |
659 | } | |
660 | ||
856e9660 PZ |
661 | /* |
662 | * Let the child rip | |
663 | */ | |
d4db3f16 ACM |
664 | if (forks) |
665 | close(go_pipe[1]); | |
856e9660 | 666 | |
649c48a9 | 667 | for (;;) { |
2debbc83 | 668 | int hits = samples; |
d6d901c2 | 669 | int thread; |
de9ac07b | 670 | |
16c8a109 | 671 | for (i = 0; i < nr_cpu; i++) { |
ea57c4f5 | 672 | for (counter = 0; counter < nr_counters; counter++) { |
d6d901c2 ZY |
673 | for (thread = 0; |
674 | thread < thread_num; thread++) { | |
675 | if (mmap_array[i][counter][thread].base) | |
676 | mmap_read(&mmap_array[i][counter][thread]); | |
677 | } | |
678 | ||
ea57c4f5 | 679 | } |
de9ac07b PZ |
680 | } |
681 | ||
649c48a9 PZ |
682 | if (hits == samples) { |
683 | if (done) | |
684 | break; | |
4dc0a04b | 685 | err = poll(event_array, nr_poll, -1); |
8b412664 PZ |
686 | waking++; |
687 | } | |
688 | ||
689 | if (done) { | |
690 | for (i = 0; i < nr_cpu; i++) { | |
d6d901c2 ZY |
691 | for (counter = 0; |
692 | counter < nr_counters; | |
693 | counter++) { | |
694 | for (thread = 0; | |
695 | thread < thread_num; | |
696 | thread++) | |
697 | ioctl(fd[i][counter][thread], | |
698 | PERF_EVENT_IOC_DISABLE); | |
699 | } | |
8b412664 | 700 | } |
649c48a9 | 701 | } |
de9ac07b PZ |
702 | } |
703 | ||
8b412664 PZ |
704 | fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking); |
705 | ||
021e9f47 IM |
706 | /* |
707 | * Approximate RIP event size: 24 bytes. | |
708 | */ | |
709 | fprintf(stderr, | |
2debbc83 | 710 | "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n", |
021e9f47 IM |
711 | (double)bytes_written / 1024.0 / 1024.0, |
712 | output_name, | |
713 | bytes_written / 24); | |
addc2785 | 714 | |
de9ac07b PZ |
715 | return 0; |
716 | } | |
0e9b20b8 | 717 | |
0e9b20b8 | 718 | static const char * const record_usage[] = { |
9e096753 MG |
719 | "perf record [<options>] [<command>]", |
720 | "perf record [<options>] -- <command> [<options>]", | |
0e9b20b8 IM |
721 | NULL |
722 | }; | |
723 | ||
5242519b | 724 | static const struct option options[] = { |
0e9b20b8 | 725 | OPT_CALLBACK('e', "event", NULL, "event", |
86847b62 TG |
726 | "event selector. use 'perf list' to list available events", |
727 | parse_events), | |
c171b552 LZ |
728 | OPT_CALLBACK(0, "filter", NULL, "filter", |
729 | "event filter", parse_filter), | |
0e9b20b8 | 730 | OPT_INTEGER('p', "pid", &target_pid, |
d6d901c2 ZY |
731 | "record events on existing process id"), |
732 | OPT_INTEGER('t', "tid", &target_tid, | |
733 | "record events on existing thread id"), | |
0e9b20b8 IM |
734 | OPT_INTEGER('r', "realtime", &realtime_prio, |
735 | "collect data with this RT SCHED_FIFO priority"), | |
daac07b2 FW |
736 | OPT_BOOLEAN('R', "raw-samples", &raw_samples, |
737 | "collect raw sample records from all opened counters"), | |
0e9b20b8 IM |
738 | OPT_BOOLEAN('a', "all-cpus", &system_wide, |
739 | "system-wide collection from all CPUs"), | |
abaff32a IM |
740 | OPT_BOOLEAN('A', "append", &append_file, |
741 | "append to the output file to do incremental profiling"), | |
0a5ac846 JA |
742 | OPT_INTEGER('C', "profile_cpu", &profile_cpu, |
743 | "CPU to profile on"), | |
97124d5e PZ |
744 | OPT_BOOLEAN('f', "force", &force, |
745 | "overwrite existing data file"), | |
e61078a0 | 746 | OPT_LONG('c', "count", &default_interval, |
abaff32a IM |
747 | "event period to sample"), |
748 | OPT_STRING('o', "output", &output_name, "file", | |
749 | "output file name"), | |
750 | OPT_BOOLEAN('i', "inherit", &inherit, | |
751 | "child tasks inherit counters"), | |
cf1f4574 IM |
752 | OPT_INTEGER('F', "freq", &freq, |
753 | "profile at this frequency"), | |
abaff32a IM |
754 | OPT_INTEGER('m', "mmap-pages", &mmap_pages, |
755 | "number of mmap data pages"), | |
3efa1cc9 IM |
756 | OPT_BOOLEAN('g', "call-graph", &call_graph, |
757 | "do call-graph (stack chain/backtrace) recording"), | |
c0555642 | 758 | OPT_INCR('v', "verbose", &verbose, |
3da297a6 | 759 | "be more verbose (show counter open errors, etc)"), |
649c48a9 PZ |
760 | OPT_BOOLEAN('s', "stat", &inherit_stat, |
761 | "per thread counts"), | |
4bba828d AB |
762 | OPT_BOOLEAN('d', "data", &sample_address, |
763 | "Sample addresses"), | |
649c48a9 PZ |
764 | OPT_BOOLEAN('n', "no-samples", &no_samples, |
765 | "don't sample"), | |
d1302522 FW |
766 | OPT_BOOLEAN('M', "multiplex", &multiplex, |
767 | "multiplex counter output in a single channel"), | |
0e9b20b8 IM |
768 | OPT_END() |
769 | }; | |
770 | ||
f37a291c | 771 | int cmd_record(int argc, const char **argv, const char *prefix __used) |
0e9b20b8 IM |
772 | { |
773 | int counter; | |
d6d901c2 | 774 | int i,j; |
0e9b20b8 | 775 | |
a0541234 | 776 | argc = parse_options(argc, argv, options, record_usage, |
655000e7 | 777 | PARSE_OPT_STOP_AT_NON_OPTION); |
d6d901c2 ZY |
778 | if (!argc && target_pid == -1 && target_tid == -1 && |
779 | !system_wide && profile_cpu == -1) | |
0e9b20b8 IM |
780 | usage_with_options(record_usage, options); |
781 | ||
655000e7 ACM |
782 | symbol__init(); |
783 | ||
bbd36e5e PZ |
784 | if (!nr_counters) { |
785 | nr_counters = 1; | |
786 | attrs[0].type = PERF_TYPE_HARDWARE; | |
787 | attrs[0].config = PERF_COUNT_HW_CPU_CYCLES; | |
788 | } | |
0e9b20b8 | 789 | |
d6d901c2 ZY |
790 | if (target_pid != -1) { |
791 | target_tid = target_pid; | |
792 | thread_num = find_all_tid(target_pid, &all_tids); | |
793 | if (thread_num <= 0) { | |
794 | fprintf(stderr, "Can't find all threads of pid %d\n", | |
795 | target_pid); | |
796 | usage_with_options(record_usage, options); | |
797 | } | |
798 | } else { | |
799 | all_tids=malloc(sizeof(pid_t)); | |
800 | if (!all_tids) | |
801 | return -ENOMEM; | |
802 | ||
803 | all_tids[0] = target_tid; | |
804 | thread_num = 1; | |
805 | } | |
806 | ||
807 | for (i = 0; i < MAX_NR_CPUS; i++) { | |
808 | for (j = 0; j < MAX_COUNTERS; j++) { | |
809 | fd[i][j] = malloc(sizeof(int)*thread_num); | |
5a103174 | 810 | mmap_array[i][j] = zalloc( |
d6d901c2 ZY |
811 | sizeof(struct mmap_data)*thread_num); |
812 | if (!fd[i][j] || !mmap_array[i][j]) | |
813 | return -ENOMEM; | |
814 | } | |
815 | } | |
816 | event_array = malloc( | |
817 | sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num); | |
818 | if (!event_array) | |
819 | return -ENOMEM; | |
820 | ||
7e4ff9e3 MG |
821 | /* |
822 | * User specified count overrides default frequency. | |
823 | */ | |
824 | if (default_interval) | |
825 | freq = 0; | |
826 | else if (freq) { | |
827 | default_interval = freq; | |
828 | } else { | |
829 | fprintf(stderr, "frequency and count are zero, aborting\n"); | |
830 | exit(EXIT_FAILURE); | |
831 | } | |
832 | ||
0e9b20b8 | 833 | for (counter = 0; counter < nr_counters; counter++) { |
a21ca2ca | 834 | if (attrs[counter].sample_period) |
0e9b20b8 IM |
835 | continue; |
836 | ||
a21ca2ca | 837 | attrs[counter].sample_period = default_interval; |
0e9b20b8 IM |
838 | } |
839 | ||
840 | return __cmd_record(argc, argv); | |
841 | } |