perf tools: Add stat event synthesize function
[deliverable/linux.git] / tools / perf / util / event.c
1 #include <linux/types.h>
2 #include <sys/mman.h>
3 #include "event.h"
4 #include "debug.h"
5 #include "hist.h"
6 #include "machine.h"
7 #include "sort.h"
8 #include "string.h"
9 #include "strlist.h"
10 #include "thread.h"
11 #include "thread_map.h"
12 #include "symbol/kallsyms.h"
13 #include "asm/bug.h"
14 #include "stat.h"
15
16 static const char *perf_event__names[] = {
17 [0] = "TOTAL",
18 [PERF_RECORD_MMAP] = "MMAP",
19 [PERF_RECORD_MMAP2] = "MMAP2",
20 [PERF_RECORD_LOST] = "LOST",
21 [PERF_RECORD_COMM] = "COMM",
22 [PERF_RECORD_EXIT] = "EXIT",
23 [PERF_RECORD_THROTTLE] = "THROTTLE",
24 [PERF_RECORD_UNTHROTTLE] = "UNTHROTTLE",
25 [PERF_RECORD_FORK] = "FORK",
26 [PERF_RECORD_READ] = "READ",
27 [PERF_RECORD_SAMPLE] = "SAMPLE",
28 [PERF_RECORD_AUX] = "AUX",
29 [PERF_RECORD_ITRACE_START] = "ITRACE_START",
30 [PERF_RECORD_LOST_SAMPLES] = "LOST_SAMPLES",
31 [PERF_RECORD_SWITCH] = "SWITCH",
32 [PERF_RECORD_SWITCH_CPU_WIDE] = "SWITCH_CPU_WIDE",
33 [PERF_RECORD_HEADER_ATTR] = "ATTR",
34 [PERF_RECORD_HEADER_EVENT_TYPE] = "EVENT_TYPE",
35 [PERF_RECORD_HEADER_TRACING_DATA] = "TRACING_DATA",
36 [PERF_RECORD_HEADER_BUILD_ID] = "BUILD_ID",
37 [PERF_RECORD_FINISHED_ROUND] = "FINISHED_ROUND",
38 [PERF_RECORD_ID_INDEX] = "ID_INDEX",
39 [PERF_RECORD_AUXTRACE_INFO] = "AUXTRACE_INFO",
40 [PERF_RECORD_AUXTRACE] = "AUXTRACE",
41 [PERF_RECORD_AUXTRACE_ERROR] = "AUXTRACE_ERROR",
42 [PERF_RECORD_THREAD_MAP] = "THREAD_MAP",
43 [PERF_RECORD_CPU_MAP] = "CPU_MAP",
44 [PERF_RECORD_STAT_CONFIG] = "STAT_CONFIG",
45 [PERF_RECORD_STAT] = "STAT",
46 };
47
48 const char *perf_event__name(unsigned int id)
49 {
50 if (id >= ARRAY_SIZE(perf_event__names))
51 return "INVALID";
52 if (!perf_event__names[id])
53 return "UNKNOWN";
54 return perf_event__names[id];
55 }
56
57 static struct perf_sample synth_sample = {
58 .pid = -1,
59 .tid = -1,
60 .time = -1,
61 .stream_id = -1,
62 .cpu = -1,
63 .period = 1,
64 };
65
66 /*
67 * Assumes that the first 4095 bytes of /proc/pid/stat contains
68 * the comm, tgid and ppid.
69 */
70 static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len,
71 pid_t *tgid, pid_t *ppid)
72 {
73 char filename[PATH_MAX];
74 char bf[4096];
75 int fd;
76 size_t size = 0;
77 ssize_t n;
78 char *nl, *name, *tgids, *ppids;
79
80 *tgid = -1;
81 *ppid = -1;
82
83 snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
84
85 fd = open(filename, O_RDONLY);
86 if (fd < 0) {
87 pr_debug("couldn't open %s\n", filename);
88 return -1;
89 }
90
91 n = read(fd, bf, sizeof(bf) - 1);
92 close(fd);
93 if (n <= 0) {
94 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n",
95 pid);
96 return -1;
97 }
98 bf[n] = '\0';
99
100 name = strstr(bf, "Name:");
101 tgids = strstr(bf, "Tgid:");
102 ppids = strstr(bf, "PPid:");
103
104 if (name) {
105 name += 5; /* strlen("Name:") */
106
107 while (*name && isspace(*name))
108 ++name;
109
110 nl = strchr(name, '\n');
111 if (nl)
112 *nl = '\0';
113
114 size = strlen(name);
115 if (size >= len)
116 size = len - 1;
117 memcpy(comm, name, size);
118 comm[size] = '\0';
119 } else {
120 pr_debug("Name: string not found for pid %d\n", pid);
121 }
122
123 if (tgids) {
124 tgids += 5; /* strlen("Tgid:") */
125 *tgid = atoi(tgids);
126 } else {
127 pr_debug("Tgid: string not found for pid %d\n", pid);
128 }
129
130 if (ppids) {
131 ppids += 5; /* strlen("PPid:") */
132 *ppid = atoi(ppids);
133 } else {
134 pr_debug("PPid: string not found for pid %d\n", pid);
135 }
136
137 return 0;
138 }
139
140 static int perf_event__prepare_comm(union perf_event *event, pid_t pid,
141 struct machine *machine,
142 pid_t *tgid, pid_t *ppid)
143 {
144 size_t size;
145
146 *ppid = -1;
147
148 memset(&event->comm, 0, sizeof(event->comm));
149
150 if (machine__is_host(machine)) {
151 if (perf_event__get_comm_ids(pid, event->comm.comm,
152 sizeof(event->comm.comm),
153 tgid, ppid) != 0) {
154 return -1;
155 }
156 } else {
157 *tgid = machine->pid;
158 }
159
160 if (*tgid < 0)
161 return -1;
162
163 event->comm.pid = *tgid;
164 event->comm.header.type = PERF_RECORD_COMM;
165
166 size = strlen(event->comm.comm) + 1;
167 size = PERF_ALIGN(size, sizeof(u64));
168 memset(event->comm.comm + size, 0, machine->id_hdr_size);
169 event->comm.header.size = (sizeof(event->comm) -
170 (sizeof(event->comm.comm) - size) +
171 machine->id_hdr_size);
172 event->comm.tid = pid;
173
174 return 0;
175 }
176
177 pid_t perf_event__synthesize_comm(struct perf_tool *tool,
178 union perf_event *event, pid_t pid,
179 perf_event__handler_t process,
180 struct machine *machine)
181 {
182 pid_t tgid, ppid;
183
184 if (perf_event__prepare_comm(event, pid, machine, &tgid, &ppid) != 0)
185 return -1;
186
187 if (process(tool, event, &synth_sample, machine) != 0)
188 return -1;
189
190 return tgid;
191 }
192
193 static int perf_event__synthesize_fork(struct perf_tool *tool,
194 union perf_event *event,
195 pid_t pid, pid_t tgid, pid_t ppid,
196 perf_event__handler_t process,
197 struct machine *machine)
198 {
199 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size);
200
201 /*
202 * for main thread set parent to ppid from status file. For other
203 * threads set parent pid to main thread. ie., assume main thread
204 * spawns all threads in a process
205 */
206 if (tgid == pid) {
207 event->fork.ppid = ppid;
208 event->fork.ptid = ppid;
209 } else {
210 event->fork.ppid = tgid;
211 event->fork.ptid = tgid;
212 }
213 event->fork.pid = tgid;
214 event->fork.tid = pid;
215 event->fork.header.type = PERF_RECORD_FORK;
216
217 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size);
218
219 if (process(tool, event, &synth_sample, machine) != 0)
220 return -1;
221
222 return 0;
223 }
224
225 int perf_event__synthesize_mmap_events(struct perf_tool *tool,
226 union perf_event *event,
227 pid_t pid, pid_t tgid,
228 perf_event__handler_t process,
229 struct machine *machine,
230 bool mmap_data,
231 unsigned int proc_map_timeout)
232 {
233 char filename[PATH_MAX];
234 FILE *fp;
235 unsigned long long t;
236 bool truncation = false;
237 unsigned long long timeout = proc_map_timeout * 1000000ULL;
238 int rc = 0;
239
240 if (machine__is_default_guest(machine))
241 return 0;
242
243 snprintf(filename, sizeof(filename), "%s/proc/%d/maps",
244 machine->root_dir, pid);
245
246 fp = fopen(filename, "r");
247 if (fp == NULL) {
248 /*
249 * We raced with a task exiting - just return:
250 */
251 pr_debug("couldn't open %s\n", filename);
252 return -1;
253 }
254
255 event->header.type = PERF_RECORD_MMAP2;
256 t = rdclock();
257
258 while (1) {
259 char bf[BUFSIZ];
260 char prot[5];
261 char execname[PATH_MAX];
262 char anonstr[] = "//anon";
263 unsigned int ino;
264 size_t size;
265 ssize_t n;
266
267 if (fgets(bf, sizeof(bf), fp) == NULL)
268 break;
269
270 if ((rdclock() - t) > timeout) {
271 pr_warning("Reading %s time out. "
272 "You may want to increase "
273 "the time limit by --proc-map-timeout\n",
274 filename);
275 truncation = true;
276 goto out;
277 }
278
279 /* ensure null termination since stack will be reused. */
280 strcpy(execname, "");
281
282 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
283 n = sscanf(bf, "%"PRIx64"-%"PRIx64" %s %"PRIx64" %x:%x %u %s\n",
284 &event->mmap2.start, &event->mmap2.len, prot,
285 &event->mmap2.pgoff, &event->mmap2.maj,
286 &event->mmap2.min,
287 &ino, execname);
288
289 /*
290 * Anon maps don't have the execname.
291 */
292 if (n < 7)
293 continue;
294
295 event->mmap2.ino = (u64)ino;
296
297 /*
298 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c
299 */
300 if (machine__is_host(machine))
301 event->header.misc = PERF_RECORD_MISC_USER;
302 else
303 event->header.misc = PERF_RECORD_MISC_GUEST_USER;
304
305 /* map protection and flags bits */
306 event->mmap2.prot = 0;
307 event->mmap2.flags = 0;
308 if (prot[0] == 'r')
309 event->mmap2.prot |= PROT_READ;
310 if (prot[1] == 'w')
311 event->mmap2.prot |= PROT_WRITE;
312 if (prot[2] == 'x')
313 event->mmap2.prot |= PROT_EXEC;
314
315 if (prot[3] == 's')
316 event->mmap2.flags |= MAP_SHARED;
317 else
318 event->mmap2.flags |= MAP_PRIVATE;
319
320 if (prot[2] != 'x') {
321 if (!mmap_data || prot[0] != 'r')
322 continue;
323
324 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA;
325 }
326
327 out:
328 if (truncation)
329 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT;
330
331 if (!strcmp(execname, ""))
332 strcpy(execname, anonstr);
333
334 size = strlen(execname) + 1;
335 memcpy(event->mmap2.filename, execname, size);
336 size = PERF_ALIGN(size, sizeof(u64));
337 event->mmap2.len -= event->mmap.start;
338 event->mmap2.header.size = (sizeof(event->mmap2) -
339 (sizeof(event->mmap2.filename) - size));
340 memset(event->mmap2.filename + size, 0, machine->id_hdr_size);
341 event->mmap2.header.size += machine->id_hdr_size;
342 event->mmap2.pid = tgid;
343 event->mmap2.tid = pid;
344
345 if (process(tool, event, &synth_sample, machine) != 0) {
346 rc = -1;
347 break;
348 }
349
350 if (truncation)
351 break;
352 }
353
354 fclose(fp);
355 return rc;
356 }
357
358 int perf_event__synthesize_modules(struct perf_tool *tool,
359 perf_event__handler_t process,
360 struct machine *machine)
361 {
362 int rc = 0;
363 struct map *pos;
364 struct map_groups *kmaps = &machine->kmaps;
365 struct maps *maps = &kmaps->maps[MAP__FUNCTION];
366 union perf_event *event = zalloc((sizeof(event->mmap) +
367 machine->id_hdr_size));
368 if (event == NULL) {
369 pr_debug("Not enough memory synthesizing mmap event "
370 "for kernel modules\n");
371 return -1;
372 }
373
374 event->header.type = PERF_RECORD_MMAP;
375
376 /*
377 * kernel uses 0 for user space maps, see kernel/perf_event.c
378 * __perf_event_mmap
379 */
380 if (machine__is_host(machine))
381 event->header.misc = PERF_RECORD_MISC_KERNEL;
382 else
383 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
384
385 for (pos = maps__first(maps); pos; pos = map__next(pos)) {
386 size_t size;
387
388 if (__map__is_kernel(pos))
389 continue;
390
391 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64));
392 event->mmap.header.type = PERF_RECORD_MMAP;
393 event->mmap.header.size = (sizeof(event->mmap) -
394 (sizeof(event->mmap.filename) - size));
395 memset(event->mmap.filename + size, 0, machine->id_hdr_size);
396 event->mmap.header.size += machine->id_hdr_size;
397 event->mmap.start = pos->start;
398 event->mmap.len = pos->end - pos->start;
399 event->mmap.pid = machine->pid;
400
401 memcpy(event->mmap.filename, pos->dso->long_name,
402 pos->dso->long_name_len + 1);
403 if (process(tool, event, &synth_sample, machine) != 0) {
404 rc = -1;
405 break;
406 }
407 }
408
409 free(event);
410 return rc;
411 }
412
413 static int __event__synthesize_thread(union perf_event *comm_event,
414 union perf_event *mmap_event,
415 union perf_event *fork_event,
416 pid_t pid, int full,
417 perf_event__handler_t process,
418 struct perf_tool *tool,
419 struct machine *machine,
420 bool mmap_data,
421 unsigned int proc_map_timeout)
422 {
423 char filename[PATH_MAX];
424 DIR *tasks;
425 struct dirent dirent, *next;
426 pid_t tgid, ppid;
427 int rc = 0;
428
429 /* special case: only send one comm event using passed in pid */
430 if (!full) {
431 tgid = perf_event__synthesize_comm(tool, comm_event, pid,
432 process, machine);
433
434 if (tgid == -1)
435 return -1;
436
437 return perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
438 process, machine, mmap_data,
439 proc_map_timeout);
440 }
441
442 if (machine__is_default_guest(machine))
443 return 0;
444
445 snprintf(filename, sizeof(filename), "%s/proc/%d/task",
446 machine->root_dir, pid);
447
448 tasks = opendir(filename);
449 if (tasks == NULL) {
450 pr_debug("couldn't open %s\n", filename);
451 return 0;
452 }
453
454 while (!readdir_r(tasks, &dirent, &next) && next) {
455 char *end;
456 pid_t _pid;
457
458 _pid = strtol(dirent.d_name, &end, 10);
459 if (*end)
460 continue;
461
462 rc = -1;
463 if (perf_event__prepare_comm(comm_event, _pid, machine,
464 &tgid, &ppid) != 0)
465 break;
466
467 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid,
468 ppid, process, machine) < 0)
469 break;
470 /*
471 * Send the prepared comm event
472 */
473 if (process(tool, comm_event, &synth_sample, machine) != 0)
474 break;
475
476 rc = 0;
477 if (_pid == pid) {
478 /* process the parent's maps too */
479 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid,
480 process, machine, mmap_data, proc_map_timeout);
481 if (rc)
482 break;
483 }
484 }
485
486 closedir(tasks);
487 return rc;
488 }
489
490 int perf_event__synthesize_thread_map(struct perf_tool *tool,
491 struct thread_map *threads,
492 perf_event__handler_t process,
493 struct machine *machine,
494 bool mmap_data,
495 unsigned int proc_map_timeout)
496 {
497 union perf_event *comm_event, *mmap_event, *fork_event;
498 int err = -1, thread, j;
499
500 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
501 if (comm_event == NULL)
502 goto out;
503
504 mmap_event = malloc(sizeof(mmap_event->mmap) + machine->id_hdr_size);
505 if (mmap_event == NULL)
506 goto out_free_comm;
507
508 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
509 if (fork_event == NULL)
510 goto out_free_mmap;
511
512 err = 0;
513 for (thread = 0; thread < threads->nr; ++thread) {
514 if (__event__synthesize_thread(comm_event, mmap_event,
515 fork_event,
516 thread_map__pid(threads, thread), 0,
517 process, tool, machine,
518 mmap_data, proc_map_timeout)) {
519 err = -1;
520 break;
521 }
522
523 /*
524 * comm.pid is set to thread group id by
525 * perf_event__synthesize_comm
526 */
527 if ((int) comm_event->comm.pid != thread_map__pid(threads, thread)) {
528 bool need_leader = true;
529
530 /* is thread group leader in thread_map? */
531 for (j = 0; j < threads->nr; ++j) {
532 if ((int) comm_event->comm.pid == thread_map__pid(threads, j)) {
533 need_leader = false;
534 break;
535 }
536 }
537
538 /* if not, generate events for it */
539 if (need_leader &&
540 __event__synthesize_thread(comm_event, mmap_event,
541 fork_event,
542 comm_event->comm.pid, 0,
543 process, tool, machine,
544 mmap_data, proc_map_timeout)) {
545 err = -1;
546 break;
547 }
548 }
549 }
550 free(fork_event);
551 out_free_mmap:
552 free(mmap_event);
553 out_free_comm:
554 free(comm_event);
555 out:
556 return err;
557 }
558
559 int perf_event__synthesize_threads(struct perf_tool *tool,
560 perf_event__handler_t process,
561 struct machine *machine,
562 bool mmap_data,
563 unsigned int proc_map_timeout)
564 {
565 DIR *proc;
566 char proc_path[PATH_MAX];
567 struct dirent dirent, *next;
568 union perf_event *comm_event, *mmap_event, *fork_event;
569 int err = -1;
570
571 if (machine__is_default_guest(machine))
572 return 0;
573
574 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size);
575 if (comm_event == NULL)
576 goto out;
577
578 mmap_event = malloc(sizeof(mmap_event->mmap) + machine->id_hdr_size);
579 if (mmap_event == NULL)
580 goto out_free_comm;
581
582 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size);
583 if (fork_event == NULL)
584 goto out_free_mmap;
585
586 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir);
587 proc = opendir(proc_path);
588
589 if (proc == NULL)
590 goto out_free_fork;
591
592 while (!readdir_r(proc, &dirent, &next) && next) {
593 char *end;
594 pid_t pid = strtol(dirent.d_name, &end, 10);
595
596 if (*end) /* only interested in proper numerical dirents */
597 continue;
598 /*
599 * We may race with exiting thread, so don't stop just because
600 * one thread couldn't be synthesized.
601 */
602 __event__synthesize_thread(comm_event, mmap_event, fork_event, pid,
603 1, process, tool, machine, mmap_data,
604 proc_map_timeout);
605 }
606
607 err = 0;
608 closedir(proc);
609 out_free_fork:
610 free(fork_event);
611 out_free_mmap:
612 free(mmap_event);
613 out_free_comm:
614 free(comm_event);
615 out:
616 return err;
617 }
618
619 struct process_symbol_args {
620 const char *name;
621 u64 start;
622 };
623
624 static int find_symbol_cb(void *arg, const char *name, char type,
625 u64 start)
626 {
627 struct process_symbol_args *args = arg;
628
629 /*
630 * Must be a function or at least an alias, as in PARISC64, where "_text" is
631 * an 'A' to the same address as "_stext".
632 */
633 if (!(symbol_type__is_a(type, MAP__FUNCTION) ||
634 type == 'A') || strcmp(name, args->name))
635 return 0;
636
637 args->start = start;
638 return 1;
639 }
640
641 u64 kallsyms__get_function_start(const char *kallsyms_filename,
642 const char *symbol_name)
643 {
644 struct process_symbol_args args = { .name = symbol_name, };
645
646 if (kallsyms__parse(kallsyms_filename, &args, find_symbol_cb) <= 0)
647 return 0;
648
649 return args.start;
650 }
651
652 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool,
653 perf_event__handler_t process,
654 struct machine *machine)
655 {
656 size_t size;
657 const char *mmap_name;
658 char name_buff[PATH_MAX];
659 struct map *map = machine__kernel_map(machine);
660 struct kmap *kmap;
661 int err;
662 union perf_event *event;
663
664 if (map == NULL)
665 return -1;
666
667 /*
668 * We should get this from /sys/kernel/sections/.text, but till that is
669 * available use this, and after it is use this as a fallback for older
670 * kernels.
671 */
672 event = zalloc((sizeof(event->mmap) + machine->id_hdr_size));
673 if (event == NULL) {
674 pr_debug("Not enough memory synthesizing mmap event "
675 "for kernel modules\n");
676 return -1;
677 }
678
679 mmap_name = machine__mmap_name(machine, name_buff, sizeof(name_buff));
680 if (machine__is_host(machine)) {
681 /*
682 * kernel uses PERF_RECORD_MISC_USER for user space maps,
683 * see kernel/perf_event.c __perf_event_mmap
684 */
685 event->header.misc = PERF_RECORD_MISC_KERNEL;
686 } else {
687 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL;
688 }
689
690 kmap = map__kmap(map);
691 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename),
692 "%s%s", mmap_name, kmap->ref_reloc_sym->name) + 1;
693 size = PERF_ALIGN(size, sizeof(u64));
694 event->mmap.header.type = PERF_RECORD_MMAP;
695 event->mmap.header.size = (sizeof(event->mmap) -
696 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size);
697 event->mmap.pgoff = kmap->ref_reloc_sym->addr;
698 event->mmap.start = map->start;
699 event->mmap.len = map->end - event->mmap.start;
700 event->mmap.pid = machine->pid;
701
702 err = process(tool, event, &synth_sample, machine);
703 free(event);
704
705 return err;
706 }
707
708 int perf_event__synthesize_thread_map2(struct perf_tool *tool,
709 struct thread_map *threads,
710 perf_event__handler_t process,
711 struct machine *machine)
712 {
713 union perf_event *event;
714 int i, err, size;
715
716 size = sizeof(event->thread_map);
717 size += threads->nr * sizeof(event->thread_map.entries[0]);
718
719 event = zalloc(size);
720 if (!event)
721 return -ENOMEM;
722
723 event->header.type = PERF_RECORD_THREAD_MAP;
724 event->header.size = size;
725 event->thread_map.nr = threads->nr;
726
727 for (i = 0; i < threads->nr; i++) {
728 struct thread_map_event_entry *entry = &event->thread_map.entries[i];
729 char *comm = thread_map__comm(threads, i);
730
731 if (!comm)
732 comm = (char *) "";
733
734 entry->pid = thread_map__pid(threads, i);
735 strncpy((char *) &entry->comm, comm, sizeof(entry->comm));
736 }
737
738 err = process(tool, event, NULL, machine);
739
740 free(event);
741 return err;
742 }
743
744 static void synthesize_cpus(struct cpu_map_entries *cpus,
745 struct cpu_map *map)
746 {
747 int i;
748
749 cpus->nr = map->nr;
750
751 for (i = 0; i < map->nr; i++)
752 cpus->cpu[i] = map->map[i];
753 }
754
755 static void synthesize_mask(struct cpu_map_mask *mask,
756 struct cpu_map *map, int max)
757 {
758 int i;
759
760 mask->nr = BITS_TO_LONGS(max);
761 mask->long_size = sizeof(long);
762
763 for (i = 0; i < map->nr; i++)
764 set_bit(map->map[i], mask->mask);
765 }
766
767 static size_t cpus_size(struct cpu_map *map)
768 {
769 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16);
770 }
771
772 static size_t mask_size(struct cpu_map *map, int *max)
773 {
774 int i;
775
776 *max = 0;
777
778 for (i = 0; i < map->nr; i++) {
779 /* bit possition of the cpu is + 1 */
780 int bit = map->map[i] + 1;
781
782 if (bit > *max)
783 *max = bit;
784 }
785
786 return sizeof(struct cpu_map_mask) + BITS_TO_LONGS(*max) * sizeof(long);
787 }
788
789 void *cpu_map_data__alloc(struct cpu_map *map, size_t *size, u16 *type, int *max)
790 {
791 size_t size_cpus, size_mask;
792 bool is_dummy = cpu_map__empty(map);
793
794 /*
795 * Both array and mask data have variable size based
796 * on the number of cpus and their actual values.
797 * The size of the 'struct cpu_map_data' is:
798 *
799 * array = size of 'struct cpu_map_entries' +
800 * number of cpus * sizeof(u64)
801 *
802 * mask = size of 'struct cpu_map_mask' +
803 * maximum cpu bit converted to size of longs
804 *
805 * and finaly + the size of 'struct cpu_map_data'.
806 */
807 size_cpus = cpus_size(map);
808 size_mask = mask_size(map, max);
809
810 if (is_dummy || (size_cpus < size_mask)) {
811 *size += size_cpus;
812 *type = PERF_CPU_MAP__CPUS;
813 } else {
814 *size += size_mask;
815 *type = PERF_CPU_MAP__MASK;
816 }
817
818 *size += sizeof(struct cpu_map_data);
819 return zalloc(*size);
820 }
821
822 void cpu_map_data__synthesize(struct cpu_map_data *data, struct cpu_map *map,
823 u16 type, int max)
824 {
825 data->type = type;
826
827 switch (type) {
828 case PERF_CPU_MAP__CPUS:
829 synthesize_cpus((struct cpu_map_entries *) data->data, map);
830 break;
831 case PERF_CPU_MAP__MASK:
832 synthesize_mask((struct cpu_map_mask *) data->data, map, max);
833 default:
834 break;
835 };
836 }
837
838 static struct cpu_map_event* cpu_map_event__new(struct cpu_map *map)
839 {
840 size_t size = sizeof(struct cpu_map_event);
841 struct cpu_map_event *event;
842 int max;
843 u16 type;
844
845 event = cpu_map_data__alloc(map, &size, &type, &max);
846 if (!event)
847 return NULL;
848
849 event->header.type = PERF_RECORD_CPU_MAP;
850 event->header.size = size;
851 event->data.type = type;
852
853 cpu_map_data__synthesize(&event->data, map, type, max);
854 return event;
855 }
856
857 int perf_event__synthesize_cpu_map(struct perf_tool *tool,
858 struct cpu_map *map,
859 perf_event__handler_t process,
860 struct machine *machine)
861 {
862 struct cpu_map_event *event;
863 int err;
864
865 event = cpu_map_event__new(map);
866 if (!event)
867 return -ENOMEM;
868
869 err = process(tool, (union perf_event *) event, NULL, machine);
870
871 free(event);
872 return err;
873 }
874
875 int perf_event__synthesize_stat_config(struct perf_tool *tool,
876 struct perf_stat_config *config,
877 perf_event__handler_t process,
878 struct machine *machine)
879 {
880 struct stat_config_event *event;
881 int size, i = 0, err;
882
883 size = sizeof(*event);
884 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0]));
885
886 event = zalloc(size);
887 if (!event)
888 return -ENOMEM;
889
890 event->header.type = PERF_RECORD_STAT_CONFIG;
891 event->header.size = size;
892 event->nr = PERF_STAT_CONFIG_TERM__MAX;
893
894 #define ADD(__term, __val) \
895 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \
896 event->data[i].val = __val; \
897 i++;
898
899 ADD(AGGR_MODE, config->aggr_mode)
900 ADD(INTERVAL, config->interval)
901 ADD(SCALE, config->scale)
902
903 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX,
904 "stat config terms unbalanced\n");
905 #undef ADD
906
907 err = process(tool, (union perf_event *) event, NULL, machine);
908
909 free(event);
910 return err;
911 }
912
913 int perf_event__synthesize_stat(struct perf_tool *tool,
914 u32 cpu, u32 thread, u64 id,
915 struct perf_counts_values *count,
916 perf_event__handler_t process,
917 struct machine *machine)
918 {
919 struct stat_event event;
920
921 event.header.type = PERF_RECORD_STAT;
922 event.header.size = sizeof(event);
923 event.header.misc = 0;
924
925 event.id = id;
926 event.cpu = cpu;
927 event.thread = thread;
928 event.val = count->val;
929 event.ena = count->ena;
930 event.run = count->run;
931
932 return process(tool, (union perf_event *) &event, NULL, machine);
933 }
934
935 void perf_event__read_stat_config(struct perf_stat_config *config,
936 struct stat_config_event *event)
937 {
938 unsigned i;
939
940 for (i = 0; i < event->nr; i++) {
941
942 switch (event->data[i].tag) {
943 #define CASE(__term, __val) \
944 case PERF_STAT_CONFIG_TERM__##__term: \
945 config->__val = event->data[i].val; \
946 break;
947
948 CASE(AGGR_MODE, aggr_mode)
949 CASE(SCALE, scale)
950 CASE(INTERVAL, interval)
951 #undef CASE
952 default:
953 pr_warning("unknown stat config term %" PRIu64 "\n",
954 event->data[i].tag);
955 }
956 }
957 }
958
959 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp)
960 {
961 const char *s;
962
963 if (event->header.misc & PERF_RECORD_MISC_COMM_EXEC)
964 s = " exec";
965 else
966 s = "";
967
968 return fprintf(fp, "%s: %s:%d/%d\n", s, event->comm.comm, event->comm.pid, event->comm.tid);
969 }
970
971 int perf_event__process_comm(struct perf_tool *tool __maybe_unused,
972 union perf_event *event,
973 struct perf_sample *sample,
974 struct machine *machine)
975 {
976 return machine__process_comm_event(machine, event, sample);
977 }
978
979 int perf_event__process_lost(struct perf_tool *tool __maybe_unused,
980 union perf_event *event,
981 struct perf_sample *sample,
982 struct machine *machine)
983 {
984 return machine__process_lost_event(machine, event, sample);
985 }
986
987 int perf_event__process_aux(struct perf_tool *tool __maybe_unused,
988 union perf_event *event,
989 struct perf_sample *sample __maybe_unused,
990 struct machine *machine)
991 {
992 return machine__process_aux_event(machine, event);
993 }
994
995 int perf_event__process_itrace_start(struct perf_tool *tool __maybe_unused,
996 union perf_event *event,
997 struct perf_sample *sample __maybe_unused,
998 struct machine *machine)
999 {
1000 return machine__process_itrace_start_event(machine, event);
1001 }
1002
1003 int perf_event__process_lost_samples(struct perf_tool *tool __maybe_unused,
1004 union perf_event *event,
1005 struct perf_sample *sample,
1006 struct machine *machine)
1007 {
1008 return machine__process_lost_samples_event(machine, event, sample);
1009 }
1010
1011 int perf_event__process_switch(struct perf_tool *tool __maybe_unused,
1012 union perf_event *event,
1013 struct perf_sample *sample __maybe_unused,
1014 struct machine *machine)
1015 {
1016 return machine__process_switch_event(machine, event);
1017 }
1018
1019 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp)
1020 {
1021 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64 "]: %c %s\n",
1022 event->mmap.pid, event->mmap.tid, event->mmap.start,
1023 event->mmap.len, event->mmap.pgoff,
1024 (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) ? 'r' : 'x',
1025 event->mmap.filename);
1026 }
1027
1028 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp)
1029 {
1030 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64
1031 " %02x:%02x %"PRIu64" %"PRIu64"]: %c%c%c%c %s\n",
1032 event->mmap2.pid, event->mmap2.tid, event->mmap2.start,
1033 event->mmap2.len, event->mmap2.pgoff, event->mmap2.maj,
1034 event->mmap2.min, event->mmap2.ino,
1035 event->mmap2.ino_generation,
1036 (event->mmap2.prot & PROT_READ) ? 'r' : '-',
1037 (event->mmap2.prot & PROT_WRITE) ? 'w' : '-',
1038 (event->mmap2.prot & PROT_EXEC) ? 'x' : '-',
1039 (event->mmap2.flags & MAP_SHARED) ? 's' : 'p',
1040 event->mmap2.filename);
1041 }
1042
1043 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp)
1044 {
1045 struct thread_map *threads = thread_map__new_event(&event->thread_map);
1046 size_t ret;
1047
1048 ret = fprintf(fp, " nr: ");
1049
1050 if (threads)
1051 ret += thread_map__fprintf(threads, fp);
1052 else
1053 ret += fprintf(fp, "failed to get threads from event\n");
1054
1055 thread_map__put(threads);
1056 return ret;
1057 }
1058
1059 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp)
1060 {
1061 struct cpu_map *cpus = cpu_map__new_data(&event->cpu_map.data);
1062 size_t ret;
1063
1064 ret = fprintf(fp, " nr: ");
1065
1066 if (cpus)
1067 ret += cpu_map__fprintf(cpus, fp);
1068 else
1069 ret += fprintf(fp, "failed to get cpumap from event\n");
1070
1071 cpu_map__put(cpus);
1072 return ret;
1073 }
1074
1075 int perf_event__process_mmap(struct perf_tool *tool __maybe_unused,
1076 union perf_event *event,
1077 struct perf_sample *sample,
1078 struct machine *machine)
1079 {
1080 return machine__process_mmap_event(machine, event, sample);
1081 }
1082
1083 int perf_event__process_mmap2(struct perf_tool *tool __maybe_unused,
1084 union perf_event *event,
1085 struct perf_sample *sample,
1086 struct machine *machine)
1087 {
1088 return machine__process_mmap2_event(machine, event, sample);
1089 }
1090
1091 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp)
1092 {
1093 return fprintf(fp, "(%d:%d):(%d:%d)\n",
1094 event->fork.pid, event->fork.tid,
1095 event->fork.ppid, event->fork.ptid);
1096 }
1097
1098 int perf_event__process_fork(struct perf_tool *tool __maybe_unused,
1099 union perf_event *event,
1100 struct perf_sample *sample,
1101 struct machine *machine)
1102 {
1103 return machine__process_fork_event(machine, event, sample);
1104 }
1105
1106 int perf_event__process_exit(struct perf_tool *tool __maybe_unused,
1107 union perf_event *event,
1108 struct perf_sample *sample,
1109 struct machine *machine)
1110 {
1111 return machine__process_exit_event(machine, event, sample);
1112 }
1113
1114 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp)
1115 {
1116 return fprintf(fp, " offset: %#"PRIx64" size: %#"PRIx64" flags: %#"PRIx64" [%s%s]\n",
1117 event->aux.aux_offset, event->aux.aux_size,
1118 event->aux.flags,
1119 event->aux.flags & PERF_AUX_FLAG_TRUNCATED ? "T" : "",
1120 event->aux.flags & PERF_AUX_FLAG_OVERWRITE ? "O" : "");
1121 }
1122
1123 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp)
1124 {
1125 return fprintf(fp, " pid: %u tid: %u\n",
1126 event->itrace_start.pid, event->itrace_start.tid);
1127 }
1128
1129 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp)
1130 {
1131 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1132 const char *in_out = out ? "OUT" : "IN ";
1133
1134 if (event->header.type == PERF_RECORD_SWITCH)
1135 return fprintf(fp, " %s\n", in_out);
1136
1137 return fprintf(fp, " %s %s pid/tid: %5u/%-5u\n",
1138 in_out, out ? "next" : "prev",
1139 event->context_switch.next_prev_pid,
1140 event->context_switch.next_prev_tid);
1141 }
1142
1143 size_t perf_event__fprintf(union perf_event *event, FILE *fp)
1144 {
1145 size_t ret = fprintf(fp, "PERF_RECORD_%s",
1146 perf_event__name(event->header.type));
1147
1148 switch (event->header.type) {
1149 case PERF_RECORD_COMM:
1150 ret += perf_event__fprintf_comm(event, fp);
1151 break;
1152 case PERF_RECORD_FORK:
1153 case PERF_RECORD_EXIT:
1154 ret += perf_event__fprintf_task(event, fp);
1155 break;
1156 case PERF_RECORD_MMAP:
1157 ret += perf_event__fprintf_mmap(event, fp);
1158 break;
1159 case PERF_RECORD_MMAP2:
1160 ret += perf_event__fprintf_mmap2(event, fp);
1161 break;
1162 case PERF_RECORD_AUX:
1163 ret += perf_event__fprintf_aux(event, fp);
1164 break;
1165 case PERF_RECORD_ITRACE_START:
1166 ret += perf_event__fprintf_itrace_start(event, fp);
1167 break;
1168 case PERF_RECORD_SWITCH:
1169 case PERF_RECORD_SWITCH_CPU_WIDE:
1170 ret += perf_event__fprintf_switch(event, fp);
1171 break;
1172 default:
1173 ret += fprintf(fp, "\n");
1174 }
1175
1176 return ret;
1177 }
1178
1179 int perf_event__process(struct perf_tool *tool __maybe_unused,
1180 union perf_event *event,
1181 struct perf_sample *sample,
1182 struct machine *machine)
1183 {
1184 return machine__process_event(machine, event, sample);
1185 }
1186
1187 void thread__find_addr_map(struct thread *thread, u8 cpumode,
1188 enum map_type type, u64 addr,
1189 struct addr_location *al)
1190 {
1191 struct map_groups *mg = thread->mg;
1192 struct machine *machine = mg->machine;
1193 bool load_map = false;
1194
1195 al->machine = machine;
1196 al->thread = thread;
1197 al->addr = addr;
1198 al->cpumode = cpumode;
1199 al->filtered = 0;
1200
1201 if (machine == NULL) {
1202 al->map = NULL;
1203 return;
1204 }
1205
1206 if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) {
1207 al->level = 'k';
1208 mg = &machine->kmaps;
1209 load_map = true;
1210 } else if (cpumode == PERF_RECORD_MISC_USER && perf_host) {
1211 al->level = '.';
1212 } else if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) {
1213 al->level = 'g';
1214 mg = &machine->kmaps;
1215 load_map = true;
1216 } else if (cpumode == PERF_RECORD_MISC_GUEST_USER && perf_guest) {
1217 al->level = 'u';
1218 } else {
1219 al->level = 'H';
1220 al->map = NULL;
1221
1222 if ((cpumode == PERF_RECORD_MISC_GUEST_USER ||
1223 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) &&
1224 !perf_guest)
1225 al->filtered |= (1 << HIST_FILTER__GUEST);
1226 if ((cpumode == PERF_RECORD_MISC_USER ||
1227 cpumode == PERF_RECORD_MISC_KERNEL) &&
1228 !perf_host)
1229 al->filtered |= (1 << HIST_FILTER__HOST);
1230
1231 return;
1232 }
1233 try_again:
1234 al->map = map_groups__find(mg, type, al->addr);
1235 if (al->map == NULL) {
1236 /*
1237 * If this is outside of all known maps, and is a negative
1238 * address, try to look it up in the kernel dso, as it might be
1239 * a vsyscall or vdso (which executes in user-mode).
1240 *
1241 * XXX This is nasty, we should have a symbol list in the
1242 * "[vdso]" dso, but for now lets use the old trick of looking
1243 * in the whole kernel symbol list.
1244 */
1245 if (cpumode == PERF_RECORD_MISC_USER && machine &&
1246 mg != &machine->kmaps &&
1247 machine__kernel_ip(machine, al->addr)) {
1248 mg = &machine->kmaps;
1249 load_map = true;
1250 goto try_again;
1251 }
1252 } else {
1253 /*
1254 * Kernel maps might be changed when loading symbols so loading
1255 * must be done prior to using kernel maps.
1256 */
1257 if (load_map)
1258 map__load(al->map, machine->symbol_filter);
1259 al->addr = al->map->map_ip(al->map, al->addr);
1260 }
1261 }
1262
1263 void thread__find_addr_location(struct thread *thread,
1264 u8 cpumode, enum map_type type, u64 addr,
1265 struct addr_location *al)
1266 {
1267 thread__find_addr_map(thread, cpumode, type, addr, al);
1268 if (al->map != NULL)
1269 al->sym = map__find_symbol(al->map, al->addr,
1270 thread->mg->machine->symbol_filter);
1271 else
1272 al->sym = NULL;
1273 }
1274
1275 /*
1276 * Callers need to drop the reference to al->thread, obtained in
1277 * machine__findnew_thread()
1278 */
1279 int perf_event__preprocess_sample(const union perf_event *event,
1280 struct machine *machine,
1281 struct addr_location *al,
1282 struct perf_sample *sample)
1283 {
1284 u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1285 struct thread *thread = machine__findnew_thread(machine, sample->pid,
1286 sample->tid);
1287
1288 if (thread == NULL)
1289 return -1;
1290
1291 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
1292 /*
1293 * Have we already created the kernel maps for this machine?
1294 *
1295 * This should have happened earlier, when we processed the kernel MMAP
1296 * events, but for older perf.data files there was no such thing, so do
1297 * it now.
1298 */
1299 if (cpumode == PERF_RECORD_MISC_KERNEL &&
1300 machine__kernel_map(machine) == NULL)
1301 machine__create_kernel_maps(machine);
1302
1303 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, sample->ip, al);
1304 dump_printf(" ...... dso: %s\n",
1305 al->map ? al->map->dso->long_name :
1306 al->level == 'H' ? "[hypervisor]" : "<not found>");
1307
1308 if (thread__is_filtered(thread))
1309 al->filtered |= (1 << HIST_FILTER__THREAD);
1310
1311 al->sym = NULL;
1312 al->cpu = sample->cpu;
1313 al->socket = -1;
1314
1315 if (al->cpu >= 0) {
1316 struct perf_env *env = machine->env;
1317
1318 if (env && env->cpu)
1319 al->socket = env->cpu[al->cpu].socket_id;
1320 }
1321
1322 if (al->map) {
1323 struct dso *dso = al->map->dso;
1324
1325 if (symbol_conf.dso_list &&
1326 (!dso || !(strlist__has_entry(symbol_conf.dso_list,
1327 dso->short_name) ||
1328 (dso->short_name != dso->long_name &&
1329 strlist__has_entry(symbol_conf.dso_list,
1330 dso->long_name))))) {
1331 al->filtered |= (1 << HIST_FILTER__DSO);
1332 }
1333
1334 al->sym = map__find_symbol(al->map, al->addr,
1335 machine->symbol_filter);
1336 }
1337
1338 if (symbol_conf.sym_list &&
1339 (!al->sym || !strlist__has_entry(symbol_conf.sym_list,
1340 al->sym->name))) {
1341 al->filtered |= (1 << HIST_FILTER__SYMBOL);
1342 }
1343
1344 return 0;
1345 }
1346
1347 /*
1348 * The preprocess_sample method will return with reference counts for the
1349 * in it, when done using (and perhaps getting ref counts if needing to
1350 * keep a pointer to one of those entries) it must be paired with
1351 * addr_location__put(), so that the refcounts can be decremented.
1352 */
1353 void addr_location__put(struct addr_location *al)
1354 {
1355 thread__zput(al->thread);
1356 }
1357
1358 bool is_bts_event(struct perf_event_attr *attr)
1359 {
1360 return attr->type == PERF_TYPE_HARDWARE &&
1361 (attr->config & PERF_COUNT_HW_BRANCH_INSTRUCTIONS) &&
1362 attr->sample_period == 1;
1363 }
1364
1365 bool sample_addr_correlates_sym(struct perf_event_attr *attr)
1366 {
1367 if (attr->type == PERF_TYPE_SOFTWARE &&
1368 (attr->config == PERF_COUNT_SW_PAGE_FAULTS ||
1369 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
1370 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MAJ))
1371 return true;
1372
1373 if (is_bts_event(attr))
1374 return true;
1375
1376 return false;
1377 }
1378
1379 void perf_event__preprocess_sample_addr(union perf_event *event,
1380 struct perf_sample *sample,
1381 struct thread *thread,
1382 struct addr_location *al)
1383 {
1384 u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1385
1386 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, sample->addr, al);
1387 if (!al->map)
1388 thread__find_addr_map(thread, cpumode, MAP__VARIABLE,
1389 sample->addr, al);
1390
1391 al->cpu = sample->cpu;
1392 al->sym = NULL;
1393
1394 if (al->map)
1395 al->sym = map__find_symbol(al->map, al->addr, NULL);
1396 }
This page took 0.059794 seconds and 5 git commands to generate.