Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 *
5 * Parts came from builtin-{top,stat,record}.c, see those files for further
6 * copyright notes.
7 */
8#include <api/fs/fs.h>
9#include <errno.h>
10#include <inttypes.h>
11#include <poll.h>
12#include "cpumap.h"
13#include "util/mmap.h"
14#include "thread_map.h"
15#include "target.h"
16#include "evlist.h"
17#include "evsel.h"
18#include "debug.h"
19#include "units.h"
20#include <internal/lib.h> // page_size
21#include "../perf.h"
22#include "asm/bug.h"
23#include "bpf-event.h"
24#include <signal.h>
25#include <unistd.h>
26#include <sched.h>
27#include <stdlib.h>
28
29#include "parse-events.h"
30#include <subcmd/parse-options.h>
31
32#include <fcntl.h>
33#include <sys/ioctl.h>
34#include <sys/mman.h>
35
36#include <linux/bitops.h>
37#include <linux/hash.h>
38#include <linux/log2.h>
39#include <linux/err.h>
40#include <linux/string.h>
41#include <linux/zalloc.h>
42#include <perf/evlist.h>
43#include <perf/evsel.h>
44#include <perf/cpumap.h>
45
46#include <internal/xyarray.h>
47
48#ifdef LACKS_SIGQUEUE_PROTOTYPE
49int sigqueue(pid_t pid, int sig, const union sigval value);
50#endif
51
52#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
53#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
54
55void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
56 struct perf_thread_map *threads)
57{
58 perf_evlist__init(&evlist->core);
59 perf_evlist__set_maps(&evlist->core, cpus, threads);
60 fdarray__init(&evlist->core.pollfd, 64);
61 evlist->workload.pid = -1;
62 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
63}
64
65struct evlist *evlist__new(void)
66{
67 struct evlist *evlist = zalloc(sizeof(*evlist));
68
69 if (evlist != NULL)
70 evlist__init(evlist, NULL, NULL);
71
72 return evlist;
73}
74
75struct evlist *perf_evlist__new_default(void)
76{
77 struct evlist *evlist = evlist__new();
78
79 if (evlist && perf_evlist__add_default(evlist)) {
80 evlist__delete(evlist);
81 evlist = NULL;
82 }
83
84 return evlist;
85}
86
87struct evlist *perf_evlist__new_dummy(void)
88{
89 struct evlist *evlist = evlist__new();
90
91 if (evlist && perf_evlist__add_dummy(evlist)) {
92 evlist__delete(evlist);
93 evlist = NULL;
94 }
95
96 return evlist;
97}
98
99/**
100 * perf_evlist__set_id_pos - set the positions of event ids.
101 * @evlist: selected event list
102 *
103 * Events with compatible sample types all have the same id_pos
104 * and is_pos. For convenience, put a copy on evlist.
105 */
106void perf_evlist__set_id_pos(struct evlist *evlist)
107{
108 struct evsel *first = evlist__first(evlist);
109
110 evlist->id_pos = first->id_pos;
111 evlist->is_pos = first->is_pos;
112}
113
114static void perf_evlist__update_id_pos(struct evlist *evlist)
115{
116 struct evsel *evsel;
117
118 evlist__for_each_entry(evlist, evsel)
119 perf_evsel__calc_id_pos(evsel);
120
121 perf_evlist__set_id_pos(evlist);
122}
123
124static void evlist__purge(struct evlist *evlist)
125{
126 struct evsel *pos, *n;
127
128 evlist__for_each_entry_safe(evlist, n, pos) {
129 list_del_init(&pos->core.node);
130 pos->evlist = NULL;
131 evsel__delete(pos);
132 }
133
134 evlist->core.nr_entries = 0;
135}
136
137void evlist__exit(struct evlist *evlist)
138{
139 zfree(&evlist->mmap);
140 zfree(&evlist->overwrite_mmap);
141 fdarray__exit(&evlist->core.pollfd);
142}
143
144void evlist__delete(struct evlist *evlist)
145{
146 if (evlist == NULL)
147 return;
148
149 evlist__munmap(evlist);
150 evlist__close(evlist);
151 perf_cpu_map__put(evlist->core.cpus);
152 perf_thread_map__put(evlist->core.threads);
153 evlist->core.cpus = NULL;
154 evlist->core.threads = NULL;
155 evlist__purge(evlist);
156 evlist__exit(evlist);
157 free(evlist);
158}
159
160void evlist__add(struct evlist *evlist, struct evsel *entry)
161{
162 entry->evlist = evlist;
163 entry->idx = evlist->core.nr_entries;
164 entry->tracking = !entry->idx;
165
166 perf_evlist__add(&evlist->core, &entry->core);
167
168 if (evlist->core.nr_entries == 1)
169 perf_evlist__set_id_pos(evlist);
170}
171
172void evlist__remove(struct evlist *evlist, struct evsel *evsel)
173{
174 evsel->evlist = NULL;
175 perf_evlist__remove(&evlist->core, &evsel->core);
176}
177
178void perf_evlist__splice_list_tail(struct evlist *evlist,
179 struct list_head *list)
180{
181 struct evsel *evsel, *temp;
182
183 __evlist__for_each_entry_safe(list, temp, evsel) {
184 list_del_init(&evsel->core.node);
185 evlist__add(evlist, evsel);
186 }
187}
188
189void __perf_evlist__set_leader(struct list_head *list)
190{
191 struct evsel *evsel, *leader;
192
193 leader = list_entry(list->next, struct evsel, core.node);
194 evsel = list_entry(list->prev, struct evsel, core.node);
195
196 leader->core.nr_members = evsel->idx - leader->idx + 1;
197
198 __evlist__for_each_entry(list, evsel) {
199 evsel->leader = leader;
200 }
201}
202
203void perf_evlist__set_leader(struct evlist *evlist)
204{
205 if (evlist->core.nr_entries) {
206 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
207 __perf_evlist__set_leader(&evlist->core.entries);
208 }
209}
210
211int __perf_evlist__add_default(struct evlist *evlist, bool precise)
212{
213 struct evsel *evsel = perf_evsel__new_cycles(precise);
214
215 if (evsel == NULL)
216 return -ENOMEM;
217
218 evlist__add(evlist, evsel);
219 return 0;
220}
221
222int perf_evlist__add_dummy(struct evlist *evlist)
223{
224 struct perf_event_attr attr = {
225 .type = PERF_TYPE_SOFTWARE,
226 .config = PERF_COUNT_SW_DUMMY,
227 .size = sizeof(attr), /* to capture ABI version */
228 };
229 struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
230
231 if (evsel == NULL)
232 return -ENOMEM;
233
234 evlist__add(evlist, evsel);
235 return 0;
236}
237
238static int evlist__add_attrs(struct evlist *evlist,
239 struct perf_event_attr *attrs, size_t nr_attrs)
240{
241 struct evsel *evsel, *n;
242 LIST_HEAD(head);
243 size_t i;
244
245 for (i = 0; i < nr_attrs; i++) {
246 evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
247 if (evsel == NULL)
248 goto out_delete_partial_list;
249 list_add_tail(&evsel->core.node, &head);
250 }
251
252 perf_evlist__splice_list_tail(evlist, &head);
253
254 return 0;
255
256out_delete_partial_list:
257 __evlist__for_each_entry_safe(&head, n, evsel)
258 evsel__delete(evsel);
259 return -1;
260}
261
262int __perf_evlist__add_default_attrs(struct evlist *evlist,
263 struct perf_event_attr *attrs, size_t nr_attrs)
264{
265 size_t i;
266
267 for (i = 0; i < nr_attrs; i++)
268 event_attr_init(attrs + i);
269
270 return evlist__add_attrs(evlist, attrs, nr_attrs);
271}
272
273struct evsel *
274perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
275{
276 struct evsel *evsel;
277
278 evlist__for_each_entry(evlist, evsel) {
279 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
280 (int)evsel->core.attr.config == id)
281 return evsel;
282 }
283
284 return NULL;
285}
286
287struct evsel *
288perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
289 const char *name)
290{
291 struct evsel *evsel;
292
293 evlist__for_each_entry(evlist, evsel) {
294 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
295 (strcmp(evsel->name, name) == 0))
296 return evsel;
297 }
298
299 return NULL;
300}
301
302int perf_evlist__add_newtp(struct evlist *evlist,
303 const char *sys, const char *name, void *handler)
304{
305 struct evsel *evsel = perf_evsel__newtp(sys, name);
306
307 if (IS_ERR(evsel))
308 return -1;
309
310 evsel->handler = handler;
311 evlist__add(evlist, evsel);
312 return 0;
313}
314
315static int perf_evlist__nr_threads(struct evlist *evlist,
316 struct evsel *evsel)
317{
318 if (evsel->core.system_wide)
319 return 1;
320 else
321 return perf_thread_map__nr(evlist->core.threads);
322}
323
324void evlist__disable(struct evlist *evlist)
325{
326 struct evsel *pos;
327
328 evlist__for_each_entry(evlist, pos) {
329 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->core.fd)
330 continue;
331 evsel__disable(pos);
332 }
333
334 evlist->enabled = false;
335}
336
337void evlist__enable(struct evlist *evlist)
338{
339 struct evsel *pos;
340
341 evlist__for_each_entry(evlist, pos) {
342 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
343 continue;
344 evsel__enable(pos);
345 }
346
347 evlist->enabled = true;
348}
349
350void perf_evlist__toggle_enable(struct evlist *evlist)
351{
352 (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
353}
354
355static int perf_evlist__enable_event_cpu(struct evlist *evlist,
356 struct evsel *evsel, int cpu)
357{
358 int thread;
359 int nr_threads = perf_evlist__nr_threads(evlist, evsel);
360
361 if (!evsel->core.fd)
362 return -EINVAL;
363
364 for (thread = 0; thread < nr_threads; thread++) {
365 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
366 if (err)
367 return err;
368 }
369 return 0;
370}
371
372static int perf_evlist__enable_event_thread(struct evlist *evlist,
373 struct evsel *evsel,
374 int thread)
375{
376 int cpu;
377 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
378
379 if (!evsel->core.fd)
380 return -EINVAL;
381
382 for (cpu = 0; cpu < nr_cpus; cpu++) {
383 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
384 if (err)
385 return err;
386 }
387 return 0;
388}
389
390int perf_evlist__enable_event_idx(struct evlist *evlist,
391 struct evsel *evsel, int idx)
392{
393 bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
394
395 if (per_cpu_mmaps)
396 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
397 else
398 return perf_evlist__enable_event_thread(evlist, evsel, idx);
399}
400
401int evlist__add_pollfd(struct evlist *evlist, int fd)
402{
403 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN);
404}
405
406static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
407 void *arg __maybe_unused)
408{
409 struct mmap *map = fda->priv[fd].ptr;
410
411 if (map)
412 perf_mmap__put(map);
413}
414
415int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
416{
417 return fdarray__filter(&evlist->core.pollfd, revents_and_mask,
418 perf_evlist__munmap_filtered, NULL);
419}
420
421int evlist__poll(struct evlist *evlist, int timeout)
422{
423 return perf_evlist__poll(&evlist->core, timeout);
424}
425
426static void perf_evlist__set_sid_idx(struct evlist *evlist,
427 struct evsel *evsel, int idx, int cpu,
428 int thread)
429{
430 struct perf_sample_id *sid = SID(evsel, cpu, thread);
431 sid->idx = idx;
432 if (evlist->core.cpus && cpu >= 0)
433 sid->cpu = evlist->core.cpus->map[cpu];
434 else
435 sid->cpu = -1;
436 if (!evsel->core.system_wide && evlist->core.threads && thread >= 0)
437 sid->tid = perf_thread_map__pid(evlist->core.threads, thread);
438 else
439 sid->tid = -1;
440}
441
442struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
443{
444 struct hlist_head *head;
445 struct perf_sample_id *sid;
446 int hash;
447
448 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
449 head = &evlist->core.heads[hash];
450
451 hlist_for_each_entry(sid, head, node)
452 if (sid->id == id)
453 return sid;
454
455 return NULL;
456}
457
458struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
459{
460 struct perf_sample_id *sid;
461
462 if (evlist->core.nr_entries == 1 || !id)
463 return evlist__first(evlist);
464
465 sid = perf_evlist__id2sid(evlist, id);
466 if (sid)
467 return container_of(sid->evsel, struct evsel, core);
468
469 if (!perf_evlist__sample_id_all(evlist))
470 return evlist__first(evlist);
471
472 return NULL;
473}
474
475struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
476 u64 id)
477{
478 struct perf_sample_id *sid;
479
480 if (!id)
481 return NULL;
482
483 sid = perf_evlist__id2sid(evlist, id);
484 if (sid)
485 return container_of(sid->evsel, struct evsel, core);
486
487 return NULL;
488}
489
490static int perf_evlist__event2id(struct evlist *evlist,
491 union perf_event *event, u64 *id)
492{
493 const __u64 *array = event->sample.array;
494 ssize_t n;
495
496 n = (event->header.size - sizeof(event->header)) >> 3;
497
498 if (event->header.type == PERF_RECORD_SAMPLE) {
499 if (evlist->id_pos >= n)
500 return -1;
501 *id = array[evlist->id_pos];
502 } else {
503 if (evlist->is_pos > n)
504 return -1;
505 n -= evlist->is_pos;
506 *id = array[n];
507 }
508 return 0;
509}
510
511struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
512 union perf_event *event)
513{
514 struct evsel *first = evlist__first(evlist);
515 struct hlist_head *head;
516 struct perf_sample_id *sid;
517 int hash;
518 u64 id;
519
520 if (evlist->core.nr_entries == 1)
521 return first;
522
523 if (!first->core.attr.sample_id_all &&
524 event->header.type != PERF_RECORD_SAMPLE)
525 return first;
526
527 if (perf_evlist__event2id(evlist, event, &id))
528 return NULL;
529
530 /* Synthesized events have an id of zero */
531 if (!id)
532 return first;
533
534 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
535 head = &evlist->core.heads[hash];
536
537 hlist_for_each_entry(sid, head, node) {
538 if (sid->id == id)
539 return container_of(sid->evsel, struct evsel, core);
540 }
541 return NULL;
542}
543
544static int perf_evlist__set_paused(struct evlist *evlist, bool value)
545{
546 int i;
547
548 if (!evlist->overwrite_mmap)
549 return 0;
550
551 for (i = 0; i < evlist->core.nr_mmaps; i++) {
552 int fd = evlist->overwrite_mmap[i].core.fd;
553 int err;
554
555 if (fd < 0)
556 continue;
557 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
558 if (err)
559 return err;
560 }
561 return 0;
562}
563
564static int perf_evlist__pause(struct evlist *evlist)
565{
566 return perf_evlist__set_paused(evlist, true);
567}
568
569static int perf_evlist__resume(struct evlist *evlist)
570{
571 return perf_evlist__set_paused(evlist, false);
572}
573
574static void evlist__munmap_nofree(struct evlist *evlist)
575{
576 int i;
577
578 if (evlist->mmap)
579 for (i = 0; i < evlist->core.nr_mmaps; i++)
580 perf_mmap__munmap(&evlist->mmap[i]);
581
582 if (evlist->overwrite_mmap)
583 for (i = 0; i < evlist->core.nr_mmaps; i++)
584 perf_mmap__munmap(&evlist->overwrite_mmap[i]);
585}
586
587void evlist__munmap(struct evlist *evlist)
588{
589 evlist__munmap_nofree(evlist);
590 zfree(&evlist->mmap);
591 zfree(&evlist->overwrite_mmap);
592}
593
594static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
595 bool overwrite)
596{
597 int i;
598 struct mmap *map;
599
600 evlist->core.nr_mmaps = perf_cpu_map__nr(evlist->core.cpus);
601 if (perf_cpu_map__empty(evlist->core.cpus))
602 evlist->core.nr_mmaps = perf_thread_map__nr(evlist->core.threads);
603 map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
604 if (!map)
605 return NULL;
606
607 for (i = 0; i < evlist->core.nr_mmaps; i++) {
608 map[i].core.fd = -1;
609 map[i].core.overwrite = overwrite;
610 /*
611 * When the perf_mmap() call is made we grab one refcount, plus
612 * one extra to let perf_mmap__consume() get the last
613 * events after all real references (perf_mmap__get()) are
614 * dropped.
615 *
616 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
617 * thus does perf_mmap__get() on it.
618 */
619 refcount_set(&map[i].core.refcnt, 0);
620 }
621 return map;
622}
623
624static bool
625perf_evlist__should_poll(struct evlist *evlist __maybe_unused,
626 struct evsel *evsel)
627{
628 if (evsel->core.attr.write_backward)
629 return false;
630 return true;
631}
632
633static int evlist__mmap_per_evsel(struct evlist *evlist, int idx,
634 struct mmap_params *mp, int cpu_idx,
635 int thread, int *_output, int *_output_overwrite)
636{
637 struct evsel *evsel;
638 int revent;
639 int evlist_cpu = cpu_map__cpu(evlist->core.cpus, cpu_idx);
640
641 evlist__for_each_entry(evlist, evsel) {
642 struct mmap *maps = evlist->mmap;
643 int *output = _output;
644 int fd;
645 int cpu;
646
647 mp->prot = PROT_READ | PROT_WRITE;
648 if (evsel->core.attr.write_backward) {
649 output = _output_overwrite;
650 maps = evlist->overwrite_mmap;
651
652 if (!maps) {
653 maps = evlist__alloc_mmap(evlist, true);
654 if (!maps)
655 return -1;
656 evlist->overwrite_mmap = maps;
657 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
658 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
659 }
660 mp->prot &= ~PROT_WRITE;
661 }
662
663 if (evsel->core.system_wide && thread)
664 continue;
665
666 cpu = perf_cpu_map__idx(evsel->core.cpus, evlist_cpu);
667 if (cpu == -1)
668 continue;
669
670 fd = FD(evsel, cpu, thread);
671
672 if (*output == -1) {
673 *output = fd;
674
675 if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
676 return -1;
677 } else {
678 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
679 return -1;
680
681 perf_mmap__get(&maps[idx]);
682 }
683
684 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
685
686 /*
687 * The system_wide flag causes a selected event to be opened
688 * always without a pid. Consequently it will never get a
689 * POLLHUP, but it is used for tracking in combination with
690 * other events, so it should not need to be polled anyway.
691 * Therefore don't add it for polling.
692 */
693 if (!evsel->core.system_wide &&
694 perf_evlist__add_pollfd(&evlist->core, fd, &maps[idx], revent) < 0) {
695 perf_mmap__put(&maps[idx]);
696 return -1;
697 }
698
699 if (evsel->core.attr.read_format & PERF_FORMAT_ID) {
700 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core, cpu, thread,
701 fd) < 0)
702 return -1;
703 perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
704 thread);
705 }
706 }
707
708 return 0;
709}
710
711static int evlist__mmap_per_cpu(struct evlist *evlist,
712 struct mmap_params *mp)
713{
714 int cpu, thread;
715 int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
716 int nr_threads = perf_thread_map__nr(evlist->core.threads);
717
718 pr_debug2("perf event ring buffer mmapped per cpu\n");
719 for (cpu = 0; cpu < nr_cpus; cpu++) {
720 int output = -1;
721 int output_overwrite = -1;
722
723 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
724 true);
725
726 for (thread = 0; thread < nr_threads; thread++) {
727 if (evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
728 thread, &output, &output_overwrite))
729 goto out_unmap;
730 }
731 }
732
733 return 0;
734
735out_unmap:
736 evlist__munmap_nofree(evlist);
737 return -1;
738}
739
740static int evlist__mmap_per_thread(struct evlist *evlist,
741 struct mmap_params *mp)
742{
743 int thread;
744 int nr_threads = perf_thread_map__nr(evlist->core.threads);
745
746 pr_debug2("perf event ring buffer mmapped per thread\n");
747 for (thread = 0; thread < nr_threads; thread++) {
748 int output = -1;
749 int output_overwrite = -1;
750
751 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
752 false);
753
754 if (evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
755 &output, &output_overwrite))
756 goto out_unmap;
757 }
758
759 return 0;
760
761out_unmap:
762 evlist__munmap_nofree(evlist);
763 return -1;
764}
765
766unsigned long perf_event_mlock_kb_in_pages(void)
767{
768 unsigned long pages;
769 int max;
770
771 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
772 /*
773 * Pick a once upon a time good value, i.e. things look
774 * strange since we can't read a sysctl value, but lets not
775 * die yet...
776 */
777 max = 512;
778 } else {
779 max -= (page_size / 1024);
780 }
781
782 pages = (max * 1024) / page_size;
783 if (!is_power_of_2(pages))
784 pages = rounddown_pow_of_two(pages);
785
786 return pages;
787}
788
789size_t evlist__mmap_size(unsigned long pages)
790{
791 if (pages == UINT_MAX)
792 pages = perf_event_mlock_kb_in_pages();
793 else if (!is_power_of_2(pages))
794 return 0;
795
796 return (pages + 1) * page_size;
797}
798
799static long parse_pages_arg(const char *str, unsigned long min,
800 unsigned long max)
801{
802 unsigned long pages, val;
803 static struct parse_tag tags[] = {
804 { .tag = 'B', .mult = 1 },
805 { .tag = 'K', .mult = 1 << 10 },
806 { .tag = 'M', .mult = 1 << 20 },
807 { .tag = 'G', .mult = 1 << 30 },
808 { .tag = 0 },
809 };
810
811 if (str == NULL)
812 return -EINVAL;
813
814 val = parse_tag_value(str, tags);
815 if (val != (unsigned long) -1) {
816 /* we got file size value */
817 pages = PERF_ALIGN(val, page_size) / page_size;
818 } else {
819 /* we got pages count value */
820 char *eptr;
821 pages = strtoul(str, &eptr, 10);
822 if (*eptr != '\0')
823 return -EINVAL;
824 }
825
826 if (pages == 0 && min == 0) {
827 /* leave number of pages at 0 */
828 } else if (!is_power_of_2(pages)) {
829 char buf[100];
830
831 /* round pages up to next power of 2 */
832 pages = roundup_pow_of_two(pages);
833 if (!pages)
834 return -EINVAL;
835
836 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
837 pr_info("rounding mmap pages size to %s (%lu pages)\n",
838 buf, pages);
839 }
840
841 if (pages > max)
842 return -EINVAL;
843
844 return pages;
845}
846
847int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
848{
849 unsigned long max = UINT_MAX;
850 long pages;
851
852 if (max > SIZE_MAX / page_size)
853 max = SIZE_MAX / page_size;
854
855 pages = parse_pages_arg(str, 1, max);
856 if (pages < 0) {
857 pr_err("Invalid argument for --mmap_pages/-m\n");
858 return -1;
859 }
860
861 *mmap_pages = pages;
862 return 0;
863}
864
865int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
866 int unset __maybe_unused)
867{
868 return __perf_evlist__parse_mmap_pages(opt->value, str);
869}
870
871/**
872 * evlist__mmap_ex - Create mmaps to receive events.
873 * @evlist: list of events
874 * @pages: map length in pages
875 * @overwrite: overwrite older events?
876 * @auxtrace_pages - auxtrace map length in pages
877 * @auxtrace_overwrite - overwrite older auxtrace data?
878 *
879 * If @overwrite is %false the user needs to signal event consumption using
880 * perf_mmap__write_tail(). Using evlist__mmap_read() does this
881 * automatically.
882 *
883 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
884 * consumption using auxtrace_mmap__write_tail().
885 *
886 * Return: %0 on success, negative error code otherwise.
887 */
888int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
889 unsigned int auxtrace_pages,
890 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
891 int comp_level)
892{
893 struct evsel *evsel;
894 const struct perf_cpu_map *cpus = evlist->core.cpus;
895 const struct perf_thread_map *threads = evlist->core.threads;
896 /*
897 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
898 * Its value is decided by evsel's write_backward.
899 * So &mp should not be passed through const pointer.
900 */
901 struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity, .flush = flush,
902 .comp_level = comp_level };
903
904 if (!evlist->mmap)
905 evlist->mmap = evlist__alloc_mmap(evlist, false);
906 if (!evlist->mmap)
907 return -ENOMEM;
908
909 if (evlist->core.pollfd.entries == NULL && perf_evlist__alloc_pollfd(&evlist->core) < 0)
910 return -ENOMEM;
911
912 evlist->core.mmap_len = evlist__mmap_size(pages);
913 pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
914 mp.mask = evlist->core.mmap_len - page_size - 1;
915
916 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
917 auxtrace_pages, auxtrace_overwrite);
918
919 evlist__for_each_entry(evlist, evsel) {
920 if ((evsel->core.attr.read_format & PERF_FORMAT_ID) &&
921 evsel->core.sample_id == NULL &&
922 perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr) < 0)
923 return -ENOMEM;
924 }
925
926 if (perf_cpu_map__empty(cpus))
927 return evlist__mmap_per_thread(evlist, &mp);
928
929 return evlist__mmap_per_cpu(evlist, &mp);
930}
931
932int evlist__mmap(struct evlist *evlist, unsigned int pages)
933{
934 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
935}
936
937int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
938{
939 bool all_threads = (target->per_thread && target->system_wide);
940 struct perf_cpu_map *cpus;
941 struct perf_thread_map *threads;
942
943 /*
944 * If specify '-a' and '--per-thread' to perf record, perf record
945 * will override '--per-thread'. target->per_thread = false and
946 * target->system_wide = true.
947 *
948 * If specify '--per-thread' only to perf record,
949 * target->per_thread = true and target->system_wide = false.
950 *
951 * So target->per_thread && target->system_wide is false.
952 * For perf record, thread_map__new_str doesn't call
953 * thread_map__new_all_cpus. That will keep perf record's
954 * current behavior.
955 *
956 * For perf stat, it allows the case that target->per_thread and
957 * target->system_wide are all true. It means to collect system-wide
958 * per-thread data. thread_map__new_str will call
959 * thread_map__new_all_cpus to enumerate all threads.
960 */
961 threads = thread_map__new_str(target->pid, target->tid, target->uid,
962 all_threads);
963
964 if (!threads)
965 return -1;
966
967 if (target__uses_dummy_map(target))
968 cpus = perf_cpu_map__dummy_new();
969 else
970 cpus = perf_cpu_map__new(target->cpu_list);
971
972 if (!cpus)
973 goto out_delete_threads;
974
975 evlist->core.has_user_cpus = !!target->cpu_list;
976
977 perf_evlist__set_maps(&evlist->core, cpus, threads);
978
979 return 0;
980
981out_delete_threads:
982 perf_thread_map__put(threads);
983 return -1;
984}
985
986void __perf_evlist__set_sample_bit(struct evlist *evlist,
987 enum perf_event_sample_format bit)
988{
989 struct evsel *evsel;
990
991 evlist__for_each_entry(evlist, evsel)
992 __perf_evsel__set_sample_bit(evsel, bit);
993}
994
995void __perf_evlist__reset_sample_bit(struct evlist *evlist,
996 enum perf_event_sample_format bit)
997{
998 struct evsel *evsel;
999
1000 evlist__for_each_entry(evlist, evsel)
1001 __perf_evsel__reset_sample_bit(evsel, bit);
1002}
1003
1004int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1005{
1006 struct evsel *evsel;
1007 int err = 0;
1008
1009 evlist__for_each_entry(evlist, evsel) {
1010 if (evsel->filter == NULL)
1011 continue;
1012
1013 /*
1014 * filters only work for tracepoint event, which doesn't have cpu limit.
1015 * So evlist and evsel should always be same.
1016 */
1017 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1018 if (err) {
1019 *err_evsel = evsel;
1020 break;
1021 }
1022 }
1023
1024 return err;
1025}
1026
1027int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1028{
1029 struct evsel *evsel;
1030 int err = 0;
1031
1032 evlist__for_each_entry(evlist, evsel) {
1033 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1034 continue;
1035
1036 err = perf_evsel__set_filter(evsel, filter);
1037 if (err)
1038 break;
1039 }
1040
1041 return err;
1042}
1043
1044int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1045{
1046 char *filter;
1047 int ret = -1;
1048 size_t i;
1049
1050 for (i = 0; i < npids; ++i) {
1051 if (i == 0) {
1052 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1053 return -1;
1054 } else {
1055 char *tmp;
1056
1057 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1058 goto out_free;
1059
1060 free(filter);
1061 filter = tmp;
1062 }
1063 }
1064
1065 ret = perf_evlist__set_tp_filter(evlist, filter);
1066out_free:
1067 free(filter);
1068 return ret;
1069}
1070
1071int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1072{
1073 return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1074}
1075
1076bool perf_evlist__valid_sample_type(struct evlist *evlist)
1077{
1078 struct evsel *pos;
1079
1080 if (evlist->core.nr_entries == 1)
1081 return true;
1082
1083 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1084 return false;
1085
1086 evlist__for_each_entry(evlist, pos) {
1087 if (pos->id_pos != evlist->id_pos ||
1088 pos->is_pos != evlist->is_pos)
1089 return false;
1090 }
1091
1092 return true;
1093}
1094
1095u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1096{
1097 struct evsel *evsel;
1098
1099 if (evlist->combined_sample_type)
1100 return evlist->combined_sample_type;
1101
1102 evlist__for_each_entry(evlist, evsel)
1103 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1104
1105 return evlist->combined_sample_type;
1106}
1107
1108u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1109{
1110 evlist->combined_sample_type = 0;
1111 return __perf_evlist__combined_sample_type(evlist);
1112}
1113
1114u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1115{
1116 struct evsel *evsel;
1117 u64 branch_type = 0;
1118
1119 evlist__for_each_entry(evlist, evsel)
1120 branch_type |= evsel->core.attr.branch_sample_type;
1121 return branch_type;
1122}
1123
1124bool perf_evlist__valid_read_format(struct evlist *evlist)
1125{
1126 struct evsel *first = evlist__first(evlist), *pos = first;
1127 u64 read_format = first->core.attr.read_format;
1128 u64 sample_type = first->core.attr.sample_type;
1129
1130 evlist__for_each_entry(evlist, pos) {
1131 if (read_format != pos->core.attr.read_format)
1132 return false;
1133 }
1134
1135 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1136 if ((sample_type & PERF_SAMPLE_READ) &&
1137 !(read_format & PERF_FORMAT_ID)) {
1138 return false;
1139 }
1140
1141 return true;
1142}
1143
1144u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1145{
1146 struct evsel *first = evlist__first(evlist);
1147 struct perf_sample *data;
1148 u64 sample_type;
1149 u16 size = 0;
1150
1151 if (!first->core.attr.sample_id_all)
1152 goto out;
1153
1154 sample_type = first->core.attr.sample_type;
1155
1156 if (sample_type & PERF_SAMPLE_TID)
1157 size += sizeof(data->tid) * 2;
1158
1159 if (sample_type & PERF_SAMPLE_TIME)
1160 size += sizeof(data->time);
1161
1162 if (sample_type & PERF_SAMPLE_ID)
1163 size += sizeof(data->id);
1164
1165 if (sample_type & PERF_SAMPLE_STREAM_ID)
1166 size += sizeof(data->stream_id);
1167
1168 if (sample_type & PERF_SAMPLE_CPU)
1169 size += sizeof(data->cpu) * 2;
1170
1171 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1172 size += sizeof(data->id);
1173out:
1174 return size;
1175}
1176
1177bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1178{
1179 struct evsel *first = evlist__first(evlist), *pos = first;
1180
1181 evlist__for_each_entry_continue(evlist, pos) {
1182 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1183 return false;
1184 }
1185
1186 return true;
1187}
1188
1189bool perf_evlist__sample_id_all(struct evlist *evlist)
1190{
1191 struct evsel *first = evlist__first(evlist);
1192 return first->core.attr.sample_id_all;
1193}
1194
1195void perf_evlist__set_selected(struct evlist *evlist,
1196 struct evsel *evsel)
1197{
1198 evlist->selected = evsel;
1199}
1200
1201void evlist__close(struct evlist *evlist)
1202{
1203 struct evsel *evsel;
1204
1205 evlist__for_each_entry_reverse(evlist, evsel)
1206 evsel__close(evsel);
1207}
1208
1209static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1210{
1211 struct perf_cpu_map *cpus;
1212 struct perf_thread_map *threads;
1213 int err = -ENOMEM;
1214
1215 /*
1216 * Try reading /sys/devices/system/cpu/online to get
1217 * an all cpus map.
1218 *
1219 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1220 * code needs an overhaul to properly forward the
1221 * error, and we may not want to do that fallback to a
1222 * default cpu identity map :-\
1223 */
1224 cpus = perf_cpu_map__new(NULL);
1225 if (!cpus)
1226 goto out;
1227
1228 threads = perf_thread_map__new_dummy();
1229 if (!threads)
1230 goto out_put;
1231
1232 perf_evlist__set_maps(&evlist->core, cpus, threads);
1233out:
1234 return err;
1235out_put:
1236 perf_cpu_map__put(cpus);
1237 goto out;
1238}
1239
1240int evlist__open(struct evlist *evlist)
1241{
1242 struct evsel *evsel;
1243 int err;
1244
1245 /*
1246 * Default: one fd per CPU, all threads, aka systemwide
1247 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1248 */
1249 if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1250 err = perf_evlist__create_syswide_maps(evlist);
1251 if (err < 0)
1252 goto out_err;
1253 }
1254
1255 perf_evlist__update_id_pos(evlist);
1256
1257 evlist__for_each_entry(evlist, evsel) {
1258 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1259 if (err < 0)
1260 goto out_err;
1261 }
1262
1263 return 0;
1264out_err:
1265 evlist__close(evlist);
1266 errno = -err;
1267 return err;
1268}
1269
1270int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1271 const char *argv[], bool pipe_output,
1272 void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1273{
1274 int child_ready_pipe[2], go_pipe[2];
1275 char bf;
1276
1277 if (pipe(child_ready_pipe) < 0) {
1278 perror("failed to create 'ready' pipe");
1279 return -1;
1280 }
1281
1282 if (pipe(go_pipe) < 0) {
1283 perror("failed to create 'go' pipe");
1284 goto out_close_ready_pipe;
1285 }
1286
1287 evlist->workload.pid = fork();
1288 if (evlist->workload.pid < 0) {
1289 perror("failed to fork");
1290 goto out_close_pipes;
1291 }
1292
1293 if (!evlist->workload.pid) {
1294 int ret;
1295
1296 if (pipe_output)
1297 dup2(2, 1);
1298
1299 signal(SIGTERM, SIG_DFL);
1300
1301 close(child_ready_pipe[0]);
1302 close(go_pipe[1]);
1303 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1304
1305 /*
1306 * Tell the parent we're ready to go
1307 */
1308 close(child_ready_pipe[1]);
1309
1310 /*
1311 * Wait until the parent tells us to go.
1312 */
1313 ret = read(go_pipe[0], &bf, 1);
1314 /*
1315 * The parent will ask for the execvp() to be performed by
1316 * writing exactly one byte, in workload.cork_fd, usually via
1317 * perf_evlist__start_workload().
1318 *
1319 * For cancelling the workload without actually running it,
1320 * the parent will just close workload.cork_fd, without writing
1321 * anything, i.e. read will return zero and we just exit()
1322 * here.
1323 */
1324 if (ret != 1) {
1325 if (ret == -1)
1326 perror("unable to read pipe");
1327 exit(ret);
1328 }
1329
1330 execvp(argv[0], (char **)argv);
1331
1332 if (exec_error) {
1333 union sigval val;
1334
1335 val.sival_int = errno;
1336 if (sigqueue(getppid(), SIGUSR1, val))
1337 perror(argv[0]);
1338 } else
1339 perror(argv[0]);
1340 exit(-1);
1341 }
1342
1343 if (exec_error) {
1344 struct sigaction act = {
1345 .sa_flags = SA_SIGINFO,
1346 .sa_sigaction = exec_error,
1347 };
1348 sigaction(SIGUSR1, &act, NULL);
1349 }
1350
1351 if (target__none(target)) {
1352 if (evlist->core.threads == NULL) {
1353 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1354 __func__, __LINE__);
1355 goto out_close_pipes;
1356 }
1357 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1358 }
1359
1360 close(child_ready_pipe[1]);
1361 close(go_pipe[0]);
1362 /*
1363 * wait for child to settle
1364 */
1365 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1366 perror("unable to read pipe");
1367 goto out_close_pipes;
1368 }
1369
1370 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1371 evlist->workload.cork_fd = go_pipe[1];
1372 close(child_ready_pipe[0]);
1373 return 0;
1374
1375out_close_pipes:
1376 close(go_pipe[0]);
1377 close(go_pipe[1]);
1378out_close_ready_pipe:
1379 close(child_ready_pipe[0]);
1380 close(child_ready_pipe[1]);
1381 return -1;
1382}
1383
1384int perf_evlist__start_workload(struct evlist *evlist)
1385{
1386 if (evlist->workload.cork_fd > 0) {
1387 char bf = 0;
1388 int ret;
1389 /*
1390 * Remove the cork, let it rip!
1391 */
1392 ret = write(evlist->workload.cork_fd, &bf, 1);
1393 if (ret < 0)
1394 perror("unable to write to pipe");
1395
1396 close(evlist->workload.cork_fd);
1397 return ret;
1398 }
1399
1400 return 0;
1401}
1402
1403int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1404 struct perf_sample *sample)
1405{
1406 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1407
1408 if (!evsel)
1409 return -EFAULT;
1410 return perf_evsel__parse_sample(evsel, event, sample);
1411}
1412
1413int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1414 union perf_event *event,
1415 u64 *timestamp)
1416{
1417 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1418
1419 if (!evsel)
1420 return -EFAULT;
1421 return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1422}
1423
1424int perf_evlist__strerror_open(struct evlist *evlist,
1425 int err, char *buf, size_t size)
1426{
1427 int printed, value;
1428 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1429
1430 switch (err) {
1431 case EACCES:
1432 case EPERM:
1433 printed = scnprintf(buf, size,
1434 "Error:\t%s.\n"
1435 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1436
1437 value = perf_event_paranoid();
1438
1439 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1440
1441 if (value >= 2) {
1442 printed += scnprintf(buf + printed, size - printed,
1443 "For your workloads it needs to be <= 1\nHint:\t");
1444 }
1445 printed += scnprintf(buf + printed, size - printed,
1446 "For system wide tracing it needs to be set to -1.\n");
1447
1448 printed += scnprintf(buf + printed, size - printed,
1449 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1450 "Hint:\tThe current value is %d.", value);
1451 break;
1452 case EINVAL: {
1453 struct evsel *first = evlist__first(evlist);
1454 int max_freq;
1455
1456 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1457 goto out_default;
1458
1459 if (first->core.attr.sample_freq < (u64)max_freq)
1460 goto out_default;
1461
1462 printed = scnprintf(buf, size,
1463 "Error:\t%s.\n"
1464 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1465 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1466 emsg, max_freq, first->core.attr.sample_freq);
1467 break;
1468 }
1469 default:
1470out_default:
1471 scnprintf(buf, size, "%s", emsg);
1472 break;
1473 }
1474
1475 return 0;
1476}
1477
1478int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1479{
1480 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1481 int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1482
1483 switch (err) {
1484 case EPERM:
1485 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1486 printed += scnprintf(buf + printed, size - printed,
1487 "Error:\t%s.\n"
1488 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1489 "Hint:\tTried using %zd kB.\n",
1490 emsg, pages_max_per_user, pages_attempted);
1491
1492 if (pages_attempted >= pages_max_per_user) {
1493 printed += scnprintf(buf + printed, size - printed,
1494 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1495 pages_max_per_user + pages_attempted);
1496 }
1497
1498 printed += scnprintf(buf + printed, size - printed,
1499 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1500 break;
1501 default:
1502 scnprintf(buf, size, "%s", emsg);
1503 break;
1504 }
1505
1506 return 0;
1507}
1508
1509void perf_evlist__to_front(struct evlist *evlist,
1510 struct evsel *move_evsel)
1511{
1512 struct evsel *evsel, *n;
1513 LIST_HEAD(move);
1514
1515 if (move_evsel == evlist__first(evlist))
1516 return;
1517
1518 evlist__for_each_entry_safe(evlist, n, evsel) {
1519 if (evsel->leader == move_evsel->leader)
1520 list_move_tail(&evsel->core.node, &move);
1521 }
1522
1523 list_splice(&move, &evlist->core.entries);
1524}
1525
1526void perf_evlist__set_tracking_event(struct evlist *evlist,
1527 struct evsel *tracking_evsel)
1528{
1529 struct evsel *evsel;
1530
1531 if (tracking_evsel->tracking)
1532 return;
1533
1534 evlist__for_each_entry(evlist, evsel) {
1535 if (evsel != tracking_evsel)
1536 evsel->tracking = false;
1537 }
1538
1539 tracking_evsel->tracking = true;
1540}
1541
1542struct evsel *
1543perf_evlist__find_evsel_by_str(struct evlist *evlist,
1544 const char *str)
1545{
1546 struct evsel *evsel;
1547
1548 evlist__for_each_entry(evlist, evsel) {
1549 if (!evsel->name)
1550 continue;
1551 if (strcmp(str, evsel->name) == 0)
1552 return evsel;
1553 }
1554
1555 return NULL;
1556}
1557
1558void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1559 enum bkw_mmap_state state)
1560{
1561 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1562 enum action {
1563 NONE,
1564 PAUSE,
1565 RESUME,
1566 } action = NONE;
1567
1568 if (!evlist->overwrite_mmap)
1569 return;
1570
1571 switch (old_state) {
1572 case BKW_MMAP_NOTREADY: {
1573 if (state != BKW_MMAP_RUNNING)
1574 goto state_err;
1575 break;
1576 }
1577 case BKW_MMAP_RUNNING: {
1578 if (state != BKW_MMAP_DATA_PENDING)
1579 goto state_err;
1580 action = PAUSE;
1581 break;
1582 }
1583 case BKW_MMAP_DATA_PENDING: {
1584 if (state != BKW_MMAP_EMPTY)
1585 goto state_err;
1586 break;
1587 }
1588 case BKW_MMAP_EMPTY: {
1589 if (state != BKW_MMAP_RUNNING)
1590 goto state_err;
1591 action = RESUME;
1592 break;
1593 }
1594 default:
1595 WARN_ONCE(1, "Shouldn't get there\n");
1596 }
1597
1598 evlist->bkw_mmap_state = state;
1599
1600 switch (action) {
1601 case PAUSE:
1602 perf_evlist__pause(evlist);
1603 break;
1604 case RESUME:
1605 perf_evlist__resume(evlist);
1606 break;
1607 case NONE:
1608 default:
1609 break;
1610 }
1611
1612state_err:
1613 return;
1614}
1615
1616bool perf_evlist__exclude_kernel(struct evlist *evlist)
1617{
1618 struct evsel *evsel;
1619
1620 evlist__for_each_entry(evlist, evsel) {
1621 if (!evsel->core.attr.exclude_kernel)
1622 return false;
1623 }
1624
1625 return true;
1626}
1627
1628/*
1629 * Events in data file are not collect in groups, but we still want
1630 * the group display. Set the artificial group and set the leader's
1631 * forced_leader flag to notify the display code.
1632 */
1633void perf_evlist__force_leader(struct evlist *evlist)
1634{
1635 if (!evlist->nr_groups) {
1636 struct evsel *leader = evlist__first(evlist);
1637
1638 perf_evlist__set_leader(evlist);
1639 leader->forced_leader = true;
1640 }
1641}
1642
1643struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1644 struct evsel *evsel)
1645{
1646 struct evsel *c2, *leader;
1647 bool is_open = true;
1648
1649 leader = evsel->leader;
1650 pr_debug("Weak group for %s/%d failed\n",
1651 leader->name, leader->core.nr_members);
1652
1653 /*
1654 * for_each_group_member doesn't work here because it doesn't
1655 * include the first entry.
1656 */
1657 evlist__for_each_entry(evsel_list, c2) {
1658 if (c2 == evsel)
1659 is_open = false;
1660 if (c2->leader == leader) {
1661 if (is_open)
1662 perf_evsel__close(&c2->core);
1663 c2->leader = c2;
1664 c2->core.nr_members = 0;
1665 }
1666 }
1667 return leader;
1668}
1669
1670int perf_evlist__add_sb_event(struct evlist **evlist,
1671 struct perf_event_attr *attr,
1672 perf_evsel__sb_cb_t cb,
1673 void *data)
1674{
1675 struct evsel *evsel;
1676 bool new_evlist = (*evlist) == NULL;
1677
1678 if (*evlist == NULL)
1679 *evlist = evlist__new();
1680 if (*evlist == NULL)
1681 return -1;
1682
1683 if (!attr->sample_id_all) {
1684 pr_warning("enabling sample_id_all for all side band events\n");
1685 attr->sample_id_all = 1;
1686 }
1687
1688 evsel = perf_evsel__new_idx(attr, (*evlist)->core.nr_entries);
1689 if (!evsel)
1690 goto out_err;
1691
1692 evsel->side_band.cb = cb;
1693 evsel->side_band.data = data;
1694 evlist__add(*evlist, evsel);
1695 return 0;
1696
1697out_err:
1698 if (new_evlist) {
1699 evlist__delete(*evlist);
1700 *evlist = NULL;
1701 }
1702 return -1;
1703}
1704
1705static void *perf_evlist__poll_thread(void *arg)
1706{
1707 struct evlist *evlist = arg;
1708 bool draining = false;
1709 int i, done = 0;
1710 /*
1711 * In order to read symbols from other namespaces perf to needs to call
1712 * setns(2). This isn't permitted if the struct_fs has multiple users.
1713 * unshare(2) the fs so that we may continue to setns into namespaces
1714 * that we're observing when, for instance, reading the build-ids at
1715 * the end of a 'perf record' session.
1716 */
1717 unshare(CLONE_FS);
1718
1719 while (!done) {
1720 bool got_data = false;
1721
1722 if (evlist->thread.done)
1723 draining = true;
1724
1725 if (!draining)
1726 evlist__poll(evlist, 1000);
1727
1728 for (i = 0; i < evlist->core.nr_mmaps; i++) {
1729 struct mmap *map = &evlist->mmap[i];
1730 union perf_event *event;
1731
1732 if (perf_mmap__read_init(map))
1733 continue;
1734 while ((event = perf_mmap__read_event(map)) != NULL) {
1735 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1736
1737 if (evsel && evsel->side_band.cb)
1738 evsel->side_band.cb(event, evsel->side_band.data);
1739 else
1740 pr_warning("cannot locate proper evsel for the side band event\n");
1741
1742 perf_mmap__consume(map);
1743 got_data = true;
1744 }
1745 perf_mmap__read_done(map);
1746 }
1747
1748 if (draining && !got_data)
1749 break;
1750 }
1751 return NULL;
1752}
1753
1754int perf_evlist__start_sb_thread(struct evlist *evlist,
1755 struct target *target)
1756{
1757 struct evsel *counter;
1758
1759 if (!evlist)
1760 return 0;
1761
1762 if (perf_evlist__create_maps(evlist, target))
1763 goto out_delete_evlist;
1764
1765 evlist__for_each_entry(evlist, counter) {
1766 if (evsel__open(counter, evlist->core.cpus,
1767 evlist->core.threads) < 0)
1768 goto out_delete_evlist;
1769 }
1770
1771 if (evlist__mmap(evlist, UINT_MAX))
1772 goto out_delete_evlist;
1773
1774 evlist__for_each_entry(evlist, counter) {
1775 if (evsel__enable(counter))
1776 goto out_delete_evlist;
1777 }
1778
1779 evlist->thread.done = 0;
1780 if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1781 goto out_delete_evlist;
1782
1783 return 0;
1784
1785out_delete_evlist:
1786 evlist__delete(evlist);
1787 evlist = NULL;
1788 return -1;
1789}
1790
1791void perf_evlist__stop_sb_thread(struct evlist *evlist)
1792{
1793 if (!evlist)
1794 return;
1795 evlist->thread.done = 1;
1796 pthread_join(evlist->thread.th, NULL);
1797 evlist__delete(evlist);
1798}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 *
5 * Parts came from builtin-{top,stat,record}.c, see those files for further
6 * copyright notes.
7 */
8#include <api/fs/fs.h>
9#include <errno.h>
10#include <inttypes.h>
11#include <poll.h>
12#include "cpumap.h"
13#include "util/mmap.h"
14#include "thread_map.h"
15#include "target.h"
16#include "evlist.h"
17#include "evsel.h"
18#include "record.h"
19#include "debug.h"
20#include "units.h"
21#include "bpf_counter.h"
22#include <internal/lib.h> // page_size
23#include "affinity.h"
24#include "../perf.h"
25#include "asm/bug.h"
26#include "bpf-event.h"
27#include "util/event.h"
28#include "util/string2.h"
29#include "util/perf_api_probe.h"
30#include "util/evsel_fprintf.h"
31#include "util/pmu.h"
32#include "util/sample.h"
33#include "util/bpf-filter.h"
34#include "util/stat.h"
35#include "util/util.h"
36#include <signal.h>
37#include <unistd.h>
38#include <sched.h>
39#include <stdlib.h>
40
41#include "parse-events.h"
42#include <subcmd/parse-options.h>
43
44#include <fcntl.h>
45#include <sys/ioctl.h>
46#include <sys/mman.h>
47#include <sys/prctl.h>
48#include <sys/timerfd.h>
49
50#include <linux/bitops.h>
51#include <linux/hash.h>
52#include <linux/log2.h>
53#include <linux/err.h>
54#include <linux/string.h>
55#include <linux/time64.h>
56#include <linux/zalloc.h>
57#include <perf/evlist.h>
58#include <perf/evsel.h>
59#include <perf/cpumap.h>
60#include <perf/mmap.h>
61
62#include <internal/xyarray.h>
63
64#ifdef LACKS_SIGQUEUE_PROTOTYPE
65int sigqueue(pid_t pid, int sig, const union sigval value);
66#endif
67
68#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
69#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
70
71void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
72 struct perf_thread_map *threads)
73{
74 perf_evlist__init(&evlist->core);
75 perf_evlist__set_maps(&evlist->core, cpus, threads);
76 evlist->workload.pid = -1;
77 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
78 evlist->ctl_fd.fd = -1;
79 evlist->ctl_fd.ack = -1;
80 evlist->ctl_fd.pos = -1;
81}
82
83struct evlist *evlist__new(void)
84{
85 struct evlist *evlist = zalloc(sizeof(*evlist));
86
87 if (evlist != NULL)
88 evlist__init(evlist, NULL, NULL);
89
90 return evlist;
91}
92
93struct evlist *evlist__new_default(void)
94{
95 struct evlist *evlist = evlist__new();
96 bool can_profile_kernel;
97 int err;
98
99 if (!evlist)
100 return NULL;
101
102 can_profile_kernel = perf_event_paranoid_check(1);
103 err = parse_event(evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu");
104 if (err) {
105 evlist__delete(evlist);
106 return NULL;
107 }
108
109 if (evlist->core.nr_entries > 1) {
110 struct evsel *evsel;
111
112 evlist__for_each_entry(evlist, evsel)
113 evsel__set_sample_id(evsel, /*can_sample_identifier=*/false);
114 }
115
116 return evlist;
117}
118
119struct evlist *evlist__new_dummy(void)
120{
121 struct evlist *evlist = evlist__new();
122
123 if (evlist && evlist__add_dummy(evlist)) {
124 evlist__delete(evlist);
125 evlist = NULL;
126 }
127
128 return evlist;
129}
130
131/**
132 * evlist__set_id_pos - set the positions of event ids.
133 * @evlist: selected event list
134 *
135 * Events with compatible sample types all have the same id_pos
136 * and is_pos. For convenience, put a copy on evlist.
137 */
138void evlist__set_id_pos(struct evlist *evlist)
139{
140 struct evsel *first = evlist__first(evlist);
141
142 evlist->id_pos = first->id_pos;
143 evlist->is_pos = first->is_pos;
144}
145
146static void evlist__update_id_pos(struct evlist *evlist)
147{
148 struct evsel *evsel;
149
150 evlist__for_each_entry(evlist, evsel)
151 evsel__calc_id_pos(evsel);
152
153 evlist__set_id_pos(evlist);
154}
155
156static void evlist__purge(struct evlist *evlist)
157{
158 struct evsel *pos, *n;
159
160 evlist__for_each_entry_safe(evlist, n, pos) {
161 list_del_init(&pos->core.node);
162 pos->evlist = NULL;
163 evsel__delete(pos);
164 }
165
166 evlist->core.nr_entries = 0;
167}
168
169void evlist__exit(struct evlist *evlist)
170{
171 event_enable_timer__exit(&evlist->eet);
172 zfree(&evlist->mmap);
173 zfree(&evlist->overwrite_mmap);
174 perf_evlist__exit(&evlist->core);
175}
176
177void evlist__delete(struct evlist *evlist)
178{
179 if (evlist == NULL)
180 return;
181
182 evlist__free_stats(evlist);
183 evlist__munmap(evlist);
184 evlist__close(evlist);
185 evlist__purge(evlist);
186 evlist__exit(evlist);
187 free(evlist);
188}
189
190void evlist__add(struct evlist *evlist, struct evsel *entry)
191{
192 perf_evlist__add(&evlist->core, &entry->core);
193 entry->evlist = evlist;
194 entry->tracking = !entry->core.idx;
195
196 if (evlist->core.nr_entries == 1)
197 evlist__set_id_pos(evlist);
198}
199
200void evlist__remove(struct evlist *evlist, struct evsel *evsel)
201{
202 evsel->evlist = NULL;
203 perf_evlist__remove(&evlist->core, &evsel->core);
204}
205
206void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list)
207{
208 while (!list_empty(list)) {
209 struct evsel *evsel, *temp, *leader = NULL;
210
211 __evlist__for_each_entry_safe(list, temp, evsel) {
212 list_del_init(&evsel->core.node);
213 evlist__add(evlist, evsel);
214 leader = evsel;
215 break;
216 }
217
218 __evlist__for_each_entry_safe(list, temp, evsel) {
219 if (evsel__has_leader(evsel, leader)) {
220 list_del_init(&evsel->core.node);
221 evlist__add(evlist, evsel);
222 }
223 }
224 }
225}
226
227int __evlist__set_tracepoints_handlers(struct evlist *evlist,
228 const struct evsel_str_handler *assocs, size_t nr_assocs)
229{
230 size_t i;
231 int err;
232
233 for (i = 0; i < nr_assocs; i++) {
234 // Adding a handler for an event not in this evlist, just ignore it.
235 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name);
236 if (evsel == NULL)
237 continue;
238
239 err = -EEXIST;
240 if (evsel->handler != NULL)
241 goto out;
242 evsel->handler = assocs[i].handler;
243 }
244
245 err = 0;
246out:
247 return err;
248}
249
250static void evlist__set_leader(struct evlist *evlist)
251{
252 perf_evlist__set_leader(&evlist->core);
253}
254
255static struct evsel *evlist__dummy_event(struct evlist *evlist)
256{
257 struct perf_event_attr attr = {
258 .type = PERF_TYPE_SOFTWARE,
259 .config = PERF_COUNT_SW_DUMMY,
260 .size = sizeof(attr), /* to capture ABI version */
261 /* Avoid frequency mode for dummy events to avoid associated timers. */
262 .freq = 0,
263 .sample_period = 1,
264 };
265
266 return evsel__new_idx(&attr, evlist->core.nr_entries);
267}
268
269int evlist__add_dummy(struct evlist *evlist)
270{
271 struct evsel *evsel = evlist__dummy_event(evlist);
272
273 if (evsel == NULL)
274 return -ENOMEM;
275
276 evlist__add(evlist, evsel);
277 return 0;
278}
279
280struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
281{
282 struct evsel *evsel = evlist__dummy_event(evlist);
283
284 if (!evsel)
285 return NULL;
286
287 evsel->core.attr.exclude_kernel = 1;
288 evsel->core.attr.exclude_guest = 1;
289 evsel->core.attr.exclude_hv = 1;
290 evsel->core.system_wide = system_wide;
291 evsel->no_aux_samples = true;
292 evsel->name = strdup("dummy:u");
293
294 evlist__add(evlist, evsel);
295 return evsel;
296}
297
298#ifdef HAVE_LIBTRACEEVENT
299struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide)
300{
301 struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0);
302
303 if (IS_ERR(evsel))
304 return evsel;
305
306 evsel__set_sample_bit(evsel, CPU);
307 evsel__set_sample_bit(evsel, TIME);
308
309 evsel->core.system_wide = system_wide;
310 evsel->no_aux_samples = true;
311
312 evlist__add(evlist, evsel);
313 return evsel;
314}
315#endif
316
317int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
318{
319 struct evsel *evsel, *n;
320 LIST_HEAD(head);
321 size_t i;
322
323 for (i = 0; i < nr_attrs; i++) {
324 evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
325 if (evsel == NULL)
326 goto out_delete_partial_list;
327 list_add_tail(&evsel->core.node, &head);
328 }
329
330 evlist__splice_list_tail(evlist, &head);
331
332 return 0;
333
334out_delete_partial_list:
335 __evlist__for_each_entry_safe(&head, n, evsel)
336 evsel__delete(evsel);
337 return -1;
338}
339
340int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
341{
342 size_t i;
343
344 for (i = 0; i < nr_attrs; i++)
345 event_attr_init(attrs + i);
346
347 return evlist__add_attrs(evlist, attrs, nr_attrs);
348}
349
350__weak int arch_evlist__add_default_attrs(struct evlist *evlist,
351 struct perf_event_attr *attrs,
352 size_t nr_attrs)
353{
354 if (!nr_attrs)
355 return 0;
356
357 return __evlist__add_default_attrs(evlist, attrs, nr_attrs);
358}
359
360struct evsel *evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
361{
362 struct evsel *evsel;
363
364 evlist__for_each_entry(evlist, evsel) {
365 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
366 (int)evsel->core.attr.config == id)
367 return evsel;
368 }
369
370 return NULL;
371}
372
373struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name)
374{
375 struct evsel *evsel;
376
377 evlist__for_each_entry(evlist, evsel) {
378 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
379 (strcmp(evsel->name, name) == 0))
380 return evsel;
381 }
382
383 return NULL;
384}
385
386#ifdef HAVE_LIBTRACEEVENT
387int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
388{
389 struct evsel *evsel = evsel__newtp(sys, name);
390
391 if (IS_ERR(evsel))
392 return -1;
393
394 evsel->handler = handler;
395 evlist__add(evlist, evsel);
396 return 0;
397}
398#endif
399
400struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity)
401{
402 struct evlist_cpu_iterator itr = {
403 .container = evlist,
404 .evsel = NULL,
405 .cpu_map_idx = 0,
406 .evlist_cpu_map_idx = 0,
407 .evlist_cpu_map_nr = perf_cpu_map__nr(evlist->core.all_cpus),
408 .cpu = (struct perf_cpu){ .cpu = -1},
409 .affinity = affinity,
410 };
411
412 if (evlist__empty(evlist)) {
413 /* Ensure the empty list doesn't iterate. */
414 itr.evlist_cpu_map_idx = itr.evlist_cpu_map_nr;
415 } else {
416 itr.evsel = evlist__first(evlist);
417 if (itr.affinity) {
418 itr.cpu = perf_cpu_map__cpu(evlist->core.all_cpus, 0);
419 affinity__set(itr.affinity, itr.cpu.cpu);
420 itr.cpu_map_idx = perf_cpu_map__idx(itr.evsel->core.cpus, itr.cpu);
421 /*
422 * If this CPU isn't in the evsel's cpu map then advance
423 * through the list.
424 */
425 if (itr.cpu_map_idx == -1)
426 evlist_cpu_iterator__next(&itr);
427 }
428 }
429 return itr;
430}
431
432void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr)
433{
434 while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) {
435 evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel);
436 evlist_cpu_itr->cpu_map_idx =
437 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
438 evlist_cpu_itr->cpu);
439 if (evlist_cpu_itr->cpu_map_idx != -1)
440 return;
441 }
442 evlist_cpu_itr->evlist_cpu_map_idx++;
443 if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) {
444 evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container);
445 evlist_cpu_itr->cpu =
446 perf_cpu_map__cpu(evlist_cpu_itr->container->core.all_cpus,
447 evlist_cpu_itr->evlist_cpu_map_idx);
448 if (evlist_cpu_itr->affinity)
449 affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu);
450 evlist_cpu_itr->cpu_map_idx =
451 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
452 evlist_cpu_itr->cpu);
453 /*
454 * If this CPU isn't in the evsel's cpu map then advance through
455 * the list.
456 */
457 if (evlist_cpu_itr->cpu_map_idx == -1)
458 evlist_cpu_iterator__next(evlist_cpu_itr);
459 }
460}
461
462bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr)
463{
464 return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr;
465}
466
467static int evsel__strcmp(struct evsel *pos, char *evsel_name)
468{
469 if (!evsel_name)
470 return 0;
471 if (evsel__is_dummy_event(pos))
472 return 1;
473 return !evsel__name_is(pos, evsel_name);
474}
475
476static int evlist__is_enabled(struct evlist *evlist)
477{
478 struct evsel *pos;
479
480 evlist__for_each_entry(evlist, pos) {
481 if (!evsel__is_group_leader(pos) || !pos->core.fd)
482 continue;
483 /* If at least one event is enabled, evlist is enabled. */
484 if (!pos->disabled)
485 return true;
486 }
487 return false;
488}
489
490static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
491{
492 struct evsel *pos;
493 struct evlist_cpu_iterator evlist_cpu_itr;
494 struct affinity saved_affinity, *affinity = NULL;
495 bool has_imm = false;
496
497 // See explanation in evlist__close()
498 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
499 if (affinity__setup(&saved_affinity) < 0)
500 return;
501 affinity = &saved_affinity;
502 }
503
504 /* Disable 'immediate' events last */
505 for (int imm = 0; imm <= 1; imm++) {
506 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
507 pos = evlist_cpu_itr.evsel;
508 if (evsel__strcmp(pos, evsel_name))
509 continue;
510 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
511 continue;
512 if (excl_dummy && evsel__is_dummy_event(pos))
513 continue;
514 if (pos->immediate)
515 has_imm = true;
516 if (pos->immediate != imm)
517 continue;
518 evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
519 }
520 if (!has_imm)
521 break;
522 }
523
524 affinity__cleanup(affinity);
525 evlist__for_each_entry(evlist, pos) {
526 if (evsel__strcmp(pos, evsel_name))
527 continue;
528 if (!evsel__is_group_leader(pos) || !pos->core.fd)
529 continue;
530 if (excl_dummy && evsel__is_dummy_event(pos))
531 continue;
532 pos->disabled = true;
533 }
534
535 /*
536 * If we disabled only single event, we need to check
537 * the enabled state of the evlist manually.
538 */
539 if (evsel_name)
540 evlist->enabled = evlist__is_enabled(evlist);
541 else
542 evlist->enabled = false;
543}
544
545void evlist__disable(struct evlist *evlist)
546{
547 __evlist__disable(evlist, NULL, false);
548}
549
550void evlist__disable_non_dummy(struct evlist *evlist)
551{
552 __evlist__disable(evlist, NULL, true);
553}
554
555void evlist__disable_evsel(struct evlist *evlist, char *evsel_name)
556{
557 __evlist__disable(evlist, evsel_name, false);
558}
559
560static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
561{
562 struct evsel *pos;
563 struct evlist_cpu_iterator evlist_cpu_itr;
564 struct affinity saved_affinity, *affinity = NULL;
565
566 // See explanation in evlist__close()
567 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
568 if (affinity__setup(&saved_affinity) < 0)
569 return;
570 affinity = &saved_affinity;
571 }
572
573 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
574 pos = evlist_cpu_itr.evsel;
575 if (evsel__strcmp(pos, evsel_name))
576 continue;
577 if (!evsel__is_group_leader(pos) || !pos->core.fd)
578 continue;
579 if (excl_dummy && evsel__is_dummy_event(pos))
580 continue;
581 evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
582 }
583 affinity__cleanup(affinity);
584 evlist__for_each_entry(evlist, pos) {
585 if (evsel__strcmp(pos, evsel_name))
586 continue;
587 if (!evsel__is_group_leader(pos) || !pos->core.fd)
588 continue;
589 if (excl_dummy && evsel__is_dummy_event(pos))
590 continue;
591 pos->disabled = false;
592 }
593
594 /*
595 * Even single event sets the 'enabled' for evlist,
596 * so the toggle can work properly and toggle to
597 * 'disabled' state.
598 */
599 evlist->enabled = true;
600}
601
602void evlist__enable(struct evlist *evlist)
603{
604 __evlist__enable(evlist, NULL, false);
605}
606
607void evlist__enable_non_dummy(struct evlist *evlist)
608{
609 __evlist__enable(evlist, NULL, true);
610}
611
612void evlist__enable_evsel(struct evlist *evlist, char *evsel_name)
613{
614 __evlist__enable(evlist, evsel_name, false);
615}
616
617void evlist__toggle_enable(struct evlist *evlist)
618{
619 (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
620}
621
622int evlist__add_pollfd(struct evlist *evlist, int fd)
623{
624 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
625}
626
627int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
628{
629 return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
630}
631
632#ifdef HAVE_EVENTFD_SUPPORT
633int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd)
634{
635 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
636 fdarray_flag__nonfilterable |
637 fdarray_flag__non_perf_event);
638}
639#endif
640
641int evlist__poll(struct evlist *evlist, int timeout)
642{
643 return perf_evlist__poll(&evlist->core, timeout);
644}
645
646struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id)
647{
648 struct hlist_head *head;
649 struct perf_sample_id *sid;
650 int hash;
651
652 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
653 head = &evlist->core.heads[hash];
654
655 hlist_for_each_entry(sid, head, node)
656 if (sid->id == id)
657 return sid;
658
659 return NULL;
660}
661
662struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id)
663{
664 struct perf_sample_id *sid;
665
666 if (evlist->core.nr_entries == 1 || !id)
667 return evlist__first(evlist);
668
669 sid = evlist__id2sid(evlist, id);
670 if (sid)
671 return container_of(sid->evsel, struct evsel, core);
672
673 if (!evlist__sample_id_all(evlist))
674 return evlist__first(evlist);
675
676 return NULL;
677}
678
679struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id)
680{
681 struct perf_sample_id *sid;
682
683 if (!id)
684 return NULL;
685
686 sid = evlist__id2sid(evlist, id);
687 if (sid)
688 return container_of(sid->evsel, struct evsel, core);
689
690 return NULL;
691}
692
693static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id)
694{
695 const __u64 *array = event->sample.array;
696 ssize_t n;
697
698 n = (event->header.size - sizeof(event->header)) >> 3;
699
700 if (event->header.type == PERF_RECORD_SAMPLE) {
701 if (evlist->id_pos >= n)
702 return -1;
703 *id = array[evlist->id_pos];
704 } else {
705 if (evlist->is_pos > n)
706 return -1;
707 n -= evlist->is_pos;
708 *id = array[n];
709 }
710 return 0;
711}
712
713struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event)
714{
715 struct evsel *first = evlist__first(evlist);
716 struct hlist_head *head;
717 struct perf_sample_id *sid;
718 int hash;
719 u64 id;
720
721 if (evlist->core.nr_entries == 1)
722 return first;
723
724 if (!first->core.attr.sample_id_all &&
725 event->header.type != PERF_RECORD_SAMPLE)
726 return first;
727
728 if (evlist__event2id(evlist, event, &id))
729 return NULL;
730
731 /* Synthesized events have an id of zero */
732 if (!id)
733 return first;
734
735 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
736 head = &evlist->core.heads[hash];
737
738 hlist_for_each_entry(sid, head, node) {
739 if (sid->id == id)
740 return container_of(sid->evsel, struct evsel, core);
741 }
742 return NULL;
743}
744
745static int evlist__set_paused(struct evlist *evlist, bool value)
746{
747 int i;
748
749 if (!evlist->overwrite_mmap)
750 return 0;
751
752 for (i = 0; i < evlist->core.nr_mmaps; i++) {
753 int fd = evlist->overwrite_mmap[i].core.fd;
754 int err;
755
756 if (fd < 0)
757 continue;
758 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
759 if (err)
760 return err;
761 }
762 return 0;
763}
764
765static int evlist__pause(struct evlist *evlist)
766{
767 return evlist__set_paused(evlist, true);
768}
769
770static int evlist__resume(struct evlist *evlist)
771{
772 return evlist__set_paused(evlist, false);
773}
774
775static void evlist__munmap_nofree(struct evlist *evlist)
776{
777 int i;
778
779 if (evlist->mmap)
780 for (i = 0; i < evlist->core.nr_mmaps; i++)
781 perf_mmap__munmap(&evlist->mmap[i].core);
782
783 if (evlist->overwrite_mmap)
784 for (i = 0; i < evlist->core.nr_mmaps; i++)
785 perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
786}
787
788void evlist__munmap(struct evlist *evlist)
789{
790 evlist__munmap_nofree(evlist);
791 zfree(&evlist->mmap);
792 zfree(&evlist->overwrite_mmap);
793}
794
795static void perf_mmap__unmap_cb(struct perf_mmap *map)
796{
797 struct mmap *m = container_of(map, struct mmap, core);
798
799 mmap__munmap(m);
800}
801
802static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
803 bool overwrite)
804{
805 int i;
806 struct mmap *map;
807
808 map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
809 if (!map)
810 return NULL;
811
812 for (i = 0; i < evlist->core.nr_mmaps; i++) {
813 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
814
815 /*
816 * When the perf_mmap() call is made we grab one refcount, plus
817 * one extra to let perf_mmap__consume() get the last
818 * events after all real references (perf_mmap__get()) are
819 * dropped.
820 *
821 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
822 * thus does perf_mmap__get() on it.
823 */
824 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
825 }
826
827 return map;
828}
829
830static void
831perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
832 struct perf_evsel *_evsel,
833 struct perf_mmap_param *_mp,
834 int idx)
835{
836 struct evlist *evlist = container_of(_evlist, struct evlist, core);
837 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
838 struct evsel *evsel = container_of(_evsel, struct evsel, core);
839
840 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx);
841}
842
843static struct perf_mmap*
844perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
845{
846 struct evlist *evlist = container_of(_evlist, struct evlist, core);
847 struct mmap *maps;
848
849 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
850
851 if (!maps) {
852 maps = evlist__alloc_mmap(evlist, overwrite);
853 if (!maps)
854 return NULL;
855
856 if (overwrite) {
857 evlist->overwrite_mmap = maps;
858 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
859 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
860 } else {
861 evlist->mmap = maps;
862 }
863 }
864
865 return &maps[idx].core;
866}
867
868static int
869perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
870 int output, struct perf_cpu cpu)
871{
872 struct mmap *map = container_of(_map, struct mmap, core);
873 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
874
875 return mmap__mmap(map, mp, output, cpu);
876}
877
878unsigned long perf_event_mlock_kb_in_pages(void)
879{
880 unsigned long pages;
881 int max;
882
883 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
884 /*
885 * Pick a once upon a time good value, i.e. things look
886 * strange since we can't read a sysctl value, but lets not
887 * die yet...
888 */
889 max = 512;
890 } else {
891 max -= (page_size / 1024);
892 }
893
894 pages = (max * 1024) / page_size;
895 if (!is_power_of_2(pages))
896 pages = rounddown_pow_of_two(pages);
897
898 return pages;
899}
900
901size_t evlist__mmap_size(unsigned long pages)
902{
903 if (pages == UINT_MAX)
904 pages = perf_event_mlock_kb_in_pages();
905 else if (!is_power_of_2(pages))
906 return 0;
907
908 return (pages + 1) * page_size;
909}
910
911static long parse_pages_arg(const char *str, unsigned long min,
912 unsigned long max)
913{
914 unsigned long pages, val;
915 static struct parse_tag tags[] = {
916 { .tag = 'B', .mult = 1 },
917 { .tag = 'K', .mult = 1 << 10 },
918 { .tag = 'M', .mult = 1 << 20 },
919 { .tag = 'G', .mult = 1 << 30 },
920 { .tag = 0 },
921 };
922
923 if (str == NULL)
924 return -EINVAL;
925
926 val = parse_tag_value(str, tags);
927 if (val != (unsigned long) -1) {
928 /* we got file size value */
929 pages = PERF_ALIGN(val, page_size) / page_size;
930 } else {
931 /* we got pages count value */
932 char *eptr;
933 pages = strtoul(str, &eptr, 10);
934 if (*eptr != '\0')
935 return -EINVAL;
936 }
937
938 if (pages == 0 && min == 0) {
939 /* leave number of pages at 0 */
940 } else if (!is_power_of_2(pages)) {
941 char buf[100];
942
943 /* round pages up to next power of 2 */
944 pages = roundup_pow_of_two(pages);
945 if (!pages)
946 return -EINVAL;
947
948 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
949 pr_info("rounding mmap pages size to %s (%lu pages)\n",
950 buf, pages);
951 }
952
953 if (pages > max)
954 return -EINVAL;
955
956 return pages;
957}
958
959int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
960{
961 unsigned long max = UINT_MAX;
962 long pages;
963
964 if (max > SIZE_MAX / page_size)
965 max = SIZE_MAX / page_size;
966
967 pages = parse_pages_arg(str, 1, max);
968 if (pages < 0) {
969 pr_err("Invalid argument for --mmap_pages/-m\n");
970 return -1;
971 }
972
973 *mmap_pages = pages;
974 return 0;
975}
976
977int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused)
978{
979 return __evlist__parse_mmap_pages(opt->value, str);
980}
981
982/**
983 * evlist__mmap_ex - Create mmaps to receive events.
984 * @evlist: list of events
985 * @pages: map length in pages
986 * @overwrite: overwrite older events?
987 * @auxtrace_pages - auxtrace map length in pages
988 * @auxtrace_overwrite - overwrite older auxtrace data?
989 *
990 * If @overwrite is %false the user needs to signal event consumption using
991 * perf_mmap__write_tail(). Using evlist__mmap_read() does this
992 * automatically.
993 *
994 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
995 * consumption using auxtrace_mmap__write_tail().
996 *
997 * Return: %0 on success, negative error code otherwise.
998 */
999int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
1000 unsigned int auxtrace_pages,
1001 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
1002 int comp_level)
1003{
1004 /*
1005 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1006 * Its value is decided by evsel's write_backward.
1007 * So &mp should not be passed through const pointer.
1008 */
1009 struct mmap_params mp = {
1010 .nr_cblocks = nr_cblocks,
1011 .affinity = affinity,
1012 .flush = flush,
1013 .comp_level = comp_level
1014 };
1015 struct perf_evlist_mmap_ops ops = {
1016 .idx = perf_evlist__mmap_cb_idx,
1017 .get = perf_evlist__mmap_cb_get,
1018 .mmap = perf_evlist__mmap_cb_mmap,
1019 };
1020
1021 evlist->core.mmap_len = evlist__mmap_size(pages);
1022 pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
1023
1024 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
1025 auxtrace_pages, auxtrace_overwrite);
1026
1027 return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
1028}
1029
1030int evlist__mmap(struct evlist *evlist, unsigned int pages)
1031{
1032 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1033}
1034
1035int evlist__create_maps(struct evlist *evlist, struct target *target)
1036{
1037 bool all_threads = (target->per_thread && target->system_wide);
1038 struct perf_cpu_map *cpus;
1039 struct perf_thread_map *threads;
1040
1041 /*
1042 * If specify '-a' and '--per-thread' to perf record, perf record
1043 * will override '--per-thread'. target->per_thread = false and
1044 * target->system_wide = true.
1045 *
1046 * If specify '--per-thread' only to perf record,
1047 * target->per_thread = true and target->system_wide = false.
1048 *
1049 * So target->per_thread && target->system_wide is false.
1050 * For perf record, thread_map__new_str doesn't call
1051 * thread_map__new_all_cpus. That will keep perf record's
1052 * current behavior.
1053 *
1054 * For perf stat, it allows the case that target->per_thread and
1055 * target->system_wide are all true. It means to collect system-wide
1056 * per-thread data. thread_map__new_str will call
1057 * thread_map__new_all_cpus to enumerate all threads.
1058 */
1059 threads = thread_map__new_str(target->pid, target->tid, target->uid,
1060 all_threads);
1061
1062 if (!threads)
1063 return -1;
1064
1065 if (target__uses_dummy_map(target))
1066 cpus = perf_cpu_map__new_any_cpu();
1067 else
1068 cpus = perf_cpu_map__new(target->cpu_list);
1069
1070 if (!cpus)
1071 goto out_delete_threads;
1072
1073 evlist->core.has_user_cpus = !!target->cpu_list;
1074
1075 perf_evlist__set_maps(&evlist->core, cpus, threads);
1076
1077 /* as evlist now has references, put count here */
1078 perf_cpu_map__put(cpus);
1079 perf_thread_map__put(threads);
1080
1081 return 0;
1082
1083out_delete_threads:
1084 perf_thread_map__put(threads);
1085 return -1;
1086}
1087
1088int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1089{
1090 struct evsel *evsel;
1091 int err = 0;
1092
1093 evlist__for_each_entry(evlist, evsel) {
1094 /*
1095 * filters only work for tracepoint event, which doesn't have cpu limit.
1096 * So evlist and evsel should always be same.
1097 */
1098 if (evsel->filter) {
1099 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1100 if (err) {
1101 *err_evsel = evsel;
1102 break;
1103 }
1104 }
1105
1106 /*
1107 * non-tracepoint events can have BPF filters.
1108 */
1109 if (!list_empty(&evsel->bpf_filters)) {
1110 err = perf_bpf_filter__prepare(evsel);
1111 if (err) {
1112 *err_evsel = evsel;
1113 break;
1114 }
1115 }
1116 }
1117
1118 return err;
1119}
1120
1121int evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1122{
1123 struct evsel *evsel;
1124 int err = 0;
1125
1126 if (filter == NULL)
1127 return -1;
1128
1129 evlist__for_each_entry(evlist, evsel) {
1130 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1131 continue;
1132
1133 err = evsel__set_filter(evsel, filter);
1134 if (err)
1135 break;
1136 }
1137
1138 return err;
1139}
1140
1141int evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1142{
1143 struct evsel *evsel;
1144 int err = 0;
1145
1146 if (filter == NULL)
1147 return -1;
1148
1149 evlist__for_each_entry(evlist, evsel) {
1150 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1151 continue;
1152
1153 err = evsel__append_tp_filter(evsel, filter);
1154 if (err)
1155 break;
1156 }
1157
1158 return err;
1159}
1160
1161char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1162{
1163 char *filter;
1164 size_t i;
1165
1166 for (i = 0; i < npids; ++i) {
1167 if (i == 0) {
1168 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1169 return NULL;
1170 } else {
1171 char *tmp;
1172
1173 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1174 goto out_free;
1175
1176 free(filter);
1177 filter = tmp;
1178 }
1179 }
1180
1181 return filter;
1182out_free:
1183 free(filter);
1184 return NULL;
1185}
1186
1187int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1188{
1189 char *filter = asprintf__tp_filter_pids(npids, pids);
1190 int ret = evlist__set_tp_filter(evlist, filter);
1191
1192 free(filter);
1193 return ret;
1194}
1195
1196int evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1197{
1198 return evlist__set_tp_filter_pids(evlist, 1, &pid);
1199}
1200
1201int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1202{
1203 char *filter = asprintf__tp_filter_pids(npids, pids);
1204 int ret = evlist__append_tp_filter(evlist, filter);
1205
1206 free(filter);
1207 return ret;
1208}
1209
1210int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1211{
1212 return evlist__append_tp_filter_pids(evlist, 1, &pid);
1213}
1214
1215bool evlist__valid_sample_type(struct evlist *evlist)
1216{
1217 struct evsel *pos;
1218
1219 if (evlist->core.nr_entries == 1)
1220 return true;
1221
1222 if (evlist->id_pos < 0 || evlist->is_pos < 0)
1223 return false;
1224
1225 evlist__for_each_entry(evlist, pos) {
1226 if (pos->id_pos != evlist->id_pos ||
1227 pos->is_pos != evlist->is_pos)
1228 return false;
1229 }
1230
1231 return true;
1232}
1233
1234u64 __evlist__combined_sample_type(struct evlist *evlist)
1235{
1236 struct evsel *evsel;
1237
1238 if (evlist->combined_sample_type)
1239 return evlist->combined_sample_type;
1240
1241 evlist__for_each_entry(evlist, evsel)
1242 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1243
1244 return evlist->combined_sample_type;
1245}
1246
1247u64 evlist__combined_sample_type(struct evlist *evlist)
1248{
1249 evlist->combined_sample_type = 0;
1250 return __evlist__combined_sample_type(evlist);
1251}
1252
1253u64 evlist__combined_branch_type(struct evlist *evlist)
1254{
1255 struct evsel *evsel;
1256 u64 branch_type = 0;
1257
1258 evlist__for_each_entry(evlist, evsel)
1259 branch_type |= evsel->core.attr.branch_sample_type;
1260 return branch_type;
1261}
1262
1263bool evlist__valid_read_format(struct evlist *evlist)
1264{
1265 struct evsel *first = evlist__first(evlist), *pos = first;
1266 u64 read_format = first->core.attr.read_format;
1267 u64 sample_type = first->core.attr.sample_type;
1268
1269 evlist__for_each_entry(evlist, pos) {
1270 if (read_format != pos->core.attr.read_format) {
1271 pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1272 read_format, (u64)pos->core.attr.read_format);
1273 }
1274 }
1275
1276 /* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */
1277 if ((sample_type & PERF_SAMPLE_READ) &&
1278 !(read_format & PERF_FORMAT_ID)) {
1279 return false;
1280 }
1281
1282 return true;
1283}
1284
1285u16 evlist__id_hdr_size(struct evlist *evlist)
1286{
1287 struct evsel *first = evlist__first(evlist);
1288
1289 return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0;
1290}
1291
1292bool evlist__valid_sample_id_all(struct evlist *evlist)
1293{
1294 struct evsel *first = evlist__first(evlist), *pos = first;
1295
1296 evlist__for_each_entry_continue(evlist, pos) {
1297 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1298 return false;
1299 }
1300
1301 return true;
1302}
1303
1304bool evlist__sample_id_all(struct evlist *evlist)
1305{
1306 struct evsel *first = evlist__first(evlist);
1307 return first->core.attr.sample_id_all;
1308}
1309
1310void evlist__set_selected(struct evlist *evlist, struct evsel *evsel)
1311{
1312 evlist->selected = evsel;
1313}
1314
1315void evlist__close(struct evlist *evlist)
1316{
1317 struct evsel *evsel;
1318 struct evlist_cpu_iterator evlist_cpu_itr;
1319 struct affinity affinity;
1320
1321 /*
1322 * With perf record core.user_requested_cpus is usually NULL.
1323 * Use the old method to handle this for now.
1324 */
1325 if (!evlist->core.user_requested_cpus ||
1326 cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
1327 evlist__for_each_entry_reverse(evlist, evsel)
1328 evsel__close(evsel);
1329 return;
1330 }
1331
1332 if (affinity__setup(&affinity) < 0)
1333 return;
1334
1335 evlist__for_each_cpu(evlist_cpu_itr, evlist, &affinity) {
1336 perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core,
1337 evlist_cpu_itr.cpu_map_idx);
1338 }
1339
1340 affinity__cleanup(&affinity);
1341 evlist__for_each_entry_reverse(evlist, evsel) {
1342 perf_evsel__free_fd(&evsel->core);
1343 perf_evsel__free_id(&evsel->core);
1344 }
1345 perf_evlist__reset_id_hash(&evlist->core);
1346}
1347
1348static int evlist__create_syswide_maps(struct evlist *evlist)
1349{
1350 struct perf_cpu_map *cpus;
1351 struct perf_thread_map *threads;
1352
1353 /*
1354 * Try reading /sys/devices/system/cpu/online to get
1355 * an all cpus map.
1356 *
1357 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1358 * code needs an overhaul to properly forward the
1359 * error, and we may not want to do that fallback to a
1360 * default cpu identity map :-\
1361 */
1362 cpus = perf_cpu_map__new_online_cpus();
1363 if (!cpus)
1364 goto out;
1365
1366 threads = perf_thread_map__new_dummy();
1367 if (!threads)
1368 goto out_put;
1369
1370 perf_evlist__set_maps(&evlist->core, cpus, threads);
1371
1372 perf_thread_map__put(threads);
1373out_put:
1374 perf_cpu_map__put(cpus);
1375out:
1376 return -ENOMEM;
1377}
1378
1379int evlist__open(struct evlist *evlist)
1380{
1381 struct evsel *evsel;
1382 int err;
1383
1384 /*
1385 * Default: one fd per CPU, all threads, aka systemwide
1386 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1387 */
1388 if (evlist->core.threads == NULL && evlist->core.user_requested_cpus == NULL) {
1389 err = evlist__create_syswide_maps(evlist);
1390 if (err < 0)
1391 goto out_err;
1392 }
1393
1394 evlist__update_id_pos(evlist);
1395
1396 evlist__for_each_entry(evlist, evsel) {
1397 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1398 if (err < 0)
1399 goto out_err;
1400 }
1401
1402 return 0;
1403out_err:
1404 evlist__close(evlist);
1405 errno = -err;
1406 return err;
1407}
1408
1409int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[],
1410 bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1411{
1412 int child_ready_pipe[2], go_pipe[2];
1413 char bf;
1414
1415 if (pipe(child_ready_pipe) < 0) {
1416 perror("failed to create 'ready' pipe");
1417 return -1;
1418 }
1419
1420 if (pipe(go_pipe) < 0) {
1421 perror("failed to create 'go' pipe");
1422 goto out_close_ready_pipe;
1423 }
1424
1425 evlist->workload.pid = fork();
1426 if (evlist->workload.pid < 0) {
1427 perror("failed to fork");
1428 goto out_close_pipes;
1429 }
1430
1431 if (!evlist->workload.pid) {
1432 int ret;
1433
1434 if (pipe_output)
1435 dup2(2, 1);
1436
1437 signal(SIGTERM, SIG_DFL);
1438
1439 close(child_ready_pipe[0]);
1440 close(go_pipe[1]);
1441 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1442
1443 /*
1444 * Change the name of this process not to confuse --exclude-perf users
1445 * that sees 'perf' in the window up to the execvp() and thinks that
1446 * perf samples are not being excluded.
1447 */
1448 prctl(PR_SET_NAME, "perf-exec");
1449
1450 /*
1451 * Tell the parent we're ready to go
1452 */
1453 close(child_ready_pipe[1]);
1454
1455 /*
1456 * Wait until the parent tells us to go.
1457 */
1458 ret = read(go_pipe[0], &bf, 1);
1459 /*
1460 * The parent will ask for the execvp() to be performed by
1461 * writing exactly one byte, in workload.cork_fd, usually via
1462 * evlist__start_workload().
1463 *
1464 * For cancelling the workload without actually running it,
1465 * the parent will just close workload.cork_fd, without writing
1466 * anything, i.e. read will return zero and we just exit()
1467 * here.
1468 */
1469 if (ret != 1) {
1470 if (ret == -1)
1471 perror("unable to read pipe");
1472 exit(ret);
1473 }
1474
1475 execvp(argv[0], (char **)argv);
1476
1477 if (exec_error) {
1478 union sigval val;
1479
1480 val.sival_int = errno;
1481 if (sigqueue(getppid(), SIGUSR1, val))
1482 perror(argv[0]);
1483 } else
1484 perror(argv[0]);
1485 exit(-1);
1486 }
1487
1488 if (exec_error) {
1489 struct sigaction act = {
1490 .sa_flags = SA_SIGINFO,
1491 .sa_sigaction = exec_error,
1492 };
1493 sigaction(SIGUSR1, &act, NULL);
1494 }
1495
1496 if (target__none(target)) {
1497 if (evlist->core.threads == NULL) {
1498 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1499 __func__, __LINE__);
1500 goto out_close_pipes;
1501 }
1502 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1503 }
1504
1505 close(child_ready_pipe[1]);
1506 close(go_pipe[0]);
1507 /*
1508 * wait for child to settle
1509 */
1510 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1511 perror("unable to read pipe");
1512 goto out_close_pipes;
1513 }
1514
1515 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1516 evlist->workload.cork_fd = go_pipe[1];
1517 close(child_ready_pipe[0]);
1518 return 0;
1519
1520out_close_pipes:
1521 close(go_pipe[0]);
1522 close(go_pipe[1]);
1523out_close_ready_pipe:
1524 close(child_ready_pipe[0]);
1525 close(child_ready_pipe[1]);
1526 return -1;
1527}
1528
1529int evlist__start_workload(struct evlist *evlist)
1530{
1531 if (evlist->workload.cork_fd > 0) {
1532 char bf = 0;
1533 int ret;
1534 /*
1535 * Remove the cork, let it rip!
1536 */
1537 ret = write(evlist->workload.cork_fd, &bf, 1);
1538 if (ret < 0)
1539 perror("unable to write to pipe");
1540
1541 close(evlist->workload.cork_fd);
1542 return ret;
1543 }
1544
1545 return 0;
1546}
1547
1548int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1549{
1550 struct evsel *evsel = evlist__event2evsel(evlist, event);
1551 int ret;
1552
1553 if (!evsel)
1554 return -EFAULT;
1555 ret = evsel__parse_sample(evsel, event, sample);
1556 if (ret)
1557 return ret;
1558 if (perf_guest && sample->id) {
1559 struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id);
1560
1561 if (sid) {
1562 sample->machine_pid = sid->machine_pid;
1563 sample->vcpu = sid->vcpu.cpu;
1564 }
1565 }
1566 return 0;
1567}
1568
1569int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp)
1570{
1571 struct evsel *evsel = evlist__event2evsel(evlist, event);
1572
1573 if (!evsel)
1574 return -EFAULT;
1575 return evsel__parse_sample_timestamp(evsel, event, timestamp);
1576}
1577
1578int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1579{
1580 int printed, value;
1581 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1582
1583 switch (err) {
1584 case EACCES:
1585 case EPERM:
1586 printed = scnprintf(buf, size,
1587 "Error:\t%s.\n"
1588 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1589
1590 value = perf_event_paranoid();
1591
1592 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1593
1594 if (value >= 2) {
1595 printed += scnprintf(buf + printed, size - printed,
1596 "For your workloads it needs to be <= 1\nHint:\t");
1597 }
1598 printed += scnprintf(buf + printed, size - printed,
1599 "For system wide tracing it needs to be set to -1.\n");
1600
1601 printed += scnprintf(buf + printed, size - printed,
1602 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1603 "Hint:\tThe current value is %d.", value);
1604 break;
1605 case EINVAL: {
1606 struct evsel *first = evlist__first(evlist);
1607 int max_freq;
1608
1609 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1610 goto out_default;
1611
1612 if (first->core.attr.sample_freq < (u64)max_freq)
1613 goto out_default;
1614
1615 printed = scnprintf(buf, size,
1616 "Error:\t%s.\n"
1617 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1618 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1619 emsg, max_freq, first->core.attr.sample_freq);
1620 break;
1621 }
1622 default:
1623out_default:
1624 scnprintf(buf, size, "%s", emsg);
1625 break;
1626 }
1627
1628 return 0;
1629}
1630
1631int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1632{
1633 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1634 int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1635
1636 switch (err) {
1637 case EPERM:
1638 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1639 printed += scnprintf(buf + printed, size - printed,
1640 "Error:\t%s.\n"
1641 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1642 "Hint:\tTried using %zd kB.\n",
1643 emsg, pages_max_per_user, pages_attempted);
1644
1645 if (pages_attempted >= pages_max_per_user) {
1646 printed += scnprintf(buf + printed, size - printed,
1647 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1648 pages_max_per_user + pages_attempted);
1649 }
1650
1651 printed += scnprintf(buf + printed, size - printed,
1652 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1653 break;
1654 default:
1655 scnprintf(buf, size, "%s", emsg);
1656 break;
1657 }
1658
1659 return 0;
1660}
1661
1662void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel)
1663{
1664 struct evsel *evsel, *n;
1665 LIST_HEAD(move);
1666
1667 if (move_evsel == evlist__first(evlist))
1668 return;
1669
1670 evlist__for_each_entry_safe(evlist, n, evsel) {
1671 if (evsel__leader(evsel) == evsel__leader(move_evsel))
1672 list_move_tail(&evsel->core.node, &move);
1673 }
1674
1675 list_splice(&move, &evlist->core.entries);
1676}
1677
1678struct evsel *evlist__get_tracking_event(struct evlist *evlist)
1679{
1680 struct evsel *evsel;
1681
1682 evlist__for_each_entry(evlist, evsel) {
1683 if (evsel->tracking)
1684 return evsel;
1685 }
1686
1687 return evlist__first(evlist);
1688}
1689
1690void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel)
1691{
1692 struct evsel *evsel;
1693
1694 if (tracking_evsel->tracking)
1695 return;
1696
1697 evlist__for_each_entry(evlist, evsel) {
1698 if (evsel != tracking_evsel)
1699 evsel->tracking = false;
1700 }
1701
1702 tracking_evsel->tracking = true;
1703}
1704
1705struct evsel *evlist__findnew_tracking_event(struct evlist *evlist, bool system_wide)
1706{
1707 struct evsel *evsel;
1708
1709 evsel = evlist__get_tracking_event(evlist);
1710 if (!evsel__is_dummy_event(evsel)) {
1711 evsel = evlist__add_aux_dummy(evlist, system_wide);
1712 if (!evsel)
1713 return NULL;
1714
1715 evlist__set_tracking_event(evlist, evsel);
1716 } else if (system_wide) {
1717 perf_evlist__go_system_wide(&evlist->core, &evsel->core);
1718 }
1719
1720 return evsel;
1721}
1722
1723struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str)
1724{
1725 struct evsel *evsel;
1726
1727 evlist__for_each_entry(evlist, evsel) {
1728 if (!evsel->name)
1729 continue;
1730 if (evsel__name_is(evsel, str))
1731 return evsel;
1732 }
1733
1734 return NULL;
1735}
1736
1737void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state)
1738{
1739 enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1740 enum action {
1741 NONE,
1742 PAUSE,
1743 RESUME,
1744 } action = NONE;
1745
1746 if (!evlist->overwrite_mmap)
1747 return;
1748
1749 switch (old_state) {
1750 case BKW_MMAP_NOTREADY: {
1751 if (state != BKW_MMAP_RUNNING)
1752 goto state_err;
1753 break;
1754 }
1755 case BKW_MMAP_RUNNING: {
1756 if (state != BKW_MMAP_DATA_PENDING)
1757 goto state_err;
1758 action = PAUSE;
1759 break;
1760 }
1761 case BKW_MMAP_DATA_PENDING: {
1762 if (state != BKW_MMAP_EMPTY)
1763 goto state_err;
1764 break;
1765 }
1766 case BKW_MMAP_EMPTY: {
1767 if (state != BKW_MMAP_RUNNING)
1768 goto state_err;
1769 action = RESUME;
1770 break;
1771 }
1772 default:
1773 WARN_ONCE(1, "Shouldn't get there\n");
1774 }
1775
1776 evlist->bkw_mmap_state = state;
1777
1778 switch (action) {
1779 case PAUSE:
1780 evlist__pause(evlist);
1781 break;
1782 case RESUME:
1783 evlist__resume(evlist);
1784 break;
1785 case NONE:
1786 default:
1787 break;
1788 }
1789
1790state_err:
1791 return;
1792}
1793
1794bool evlist__exclude_kernel(struct evlist *evlist)
1795{
1796 struct evsel *evsel;
1797
1798 evlist__for_each_entry(evlist, evsel) {
1799 if (!evsel->core.attr.exclude_kernel)
1800 return false;
1801 }
1802
1803 return true;
1804}
1805
1806/*
1807 * Events in data file are not collect in groups, but we still want
1808 * the group display. Set the artificial group and set the leader's
1809 * forced_leader flag to notify the display code.
1810 */
1811void evlist__force_leader(struct evlist *evlist)
1812{
1813 if (evlist__nr_groups(evlist) == 0) {
1814 struct evsel *leader = evlist__first(evlist);
1815
1816 evlist__set_leader(evlist);
1817 leader->forced_leader = true;
1818 }
1819}
1820
1821struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close)
1822{
1823 struct evsel *c2, *leader;
1824 bool is_open = true;
1825
1826 leader = evsel__leader(evsel);
1827
1828 pr_debug("Weak group for %s/%d failed\n",
1829 leader->name, leader->core.nr_members);
1830
1831 /*
1832 * for_each_group_member doesn't work here because it doesn't
1833 * include the first entry.
1834 */
1835 evlist__for_each_entry(evsel_list, c2) {
1836 if (c2 == evsel)
1837 is_open = false;
1838 if (evsel__has_leader(c2, leader)) {
1839 if (is_open && close)
1840 perf_evsel__close(&c2->core);
1841 /*
1842 * We want to close all members of the group and reopen
1843 * them. Some events, like Intel topdown, require being
1844 * in a group and so keep these in the group.
1845 */
1846 evsel__remove_from_group(c2, leader);
1847
1848 /*
1849 * Set this for all former members of the group
1850 * to indicate they get reopened.
1851 */
1852 c2->reset_group = true;
1853 }
1854 }
1855 /* Reset the leader count if all entries were removed. */
1856 if (leader->core.nr_members == 1)
1857 leader->core.nr_members = 0;
1858 return leader;
1859}
1860
1861static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1862{
1863 char *s, *p;
1864 int ret = 0, fd;
1865
1866 if (strncmp(str, "fifo:", 5))
1867 return -EINVAL;
1868
1869 str += 5;
1870 if (!*str || *str == ',')
1871 return -EINVAL;
1872
1873 s = strdup(str);
1874 if (!s)
1875 return -ENOMEM;
1876
1877 p = strchr(s, ',');
1878 if (p)
1879 *p = '\0';
1880
1881 /*
1882 * O_RDWR avoids POLLHUPs which is necessary to allow the other
1883 * end of a FIFO to be repeatedly opened and closed.
1884 */
1885 fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1886 if (fd < 0) {
1887 pr_err("Failed to open '%s'\n", s);
1888 ret = -errno;
1889 goto out_free;
1890 }
1891 *ctl_fd = fd;
1892 *ctl_fd_close = true;
1893
1894 if (p && *++p) {
1895 /* O_RDWR | O_NONBLOCK means the other end need not be open */
1896 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1897 if (fd < 0) {
1898 pr_err("Failed to open '%s'\n", p);
1899 ret = -errno;
1900 goto out_free;
1901 }
1902 *ctl_fd_ack = fd;
1903 }
1904
1905out_free:
1906 free(s);
1907 return ret;
1908}
1909
1910int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1911{
1912 char *comma = NULL, *endptr = NULL;
1913
1914 *ctl_fd_close = false;
1915
1916 if (strncmp(str, "fd:", 3))
1917 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1918
1919 *ctl_fd = strtoul(&str[3], &endptr, 0);
1920 if (endptr == &str[3])
1921 return -EINVAL;
1922
1923 comma = strchr(str, ',');
1924 if (comma) {
1925 if (endptr != comma)
1926 return -EINVAL;
1927
1928 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1929 if (endptr == comma + 1 || *endptr != '\0')
1930 return -EINVAL;
1931 }
1932
1933 return 0;
1934}
1935
1936void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1937{
1938 if (*ctl_fd_close) {
1939 *ctl_fd_close = false;
1940 close(ctl_fd);
1941 if (ctl_fd_ack >= 0)
1942 close(ctl_fd_ack);
1943 }
1944}
1945
1946int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1947{
1948 if (fd == -1) {
1949 pr_debug("Control descriptor is not initialized\n");
1950 return 0;
1951 }
1952
1953 evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1954 fdarray_flag__nonfilterable |
1955 fdarray_flag__non_perf_event);
1956 if (evlist->ctl_fd.pos < 0) {
1957 evlist->ctl_fd.pos = -1;
1958 pr_err("Failed to add ctl fd entry: %m\n");
1959 return -1;
1960 }
1961
1962 evlist->ctl_fd.fd = fd;
1963 evlist->ctl_fd.ack = ack;
1964
1965 return 0;
1966}
1967
1968bool evlist__ctlfd_initialized(struct evlist *evlist)
1969{
1970 return evlist->ctl_fd.pos >= 0;
1971}
1972
1973int evlist__finalize_ctlfd(struct evlist *evlist)
1974{
1975 struct pollfd *entries = evlist->core.pollfd.entries;
1976
1977 if (!evlist__ctlfd_initialized(evlist))
1978 return 0;
1979
1980 entries[evlist->ctl_fd.pos].fd = -1;
1981 entries[evlist->ctl_fd.pos].events = 0;
1982 entries[evlist->ctl_fd.pos].revents = 0;
1983
1984 evlist->ctl_fd.pos = -1;
1985 evlist->ctl_fd.ack = -1;
1986 evlist->ctl_fd.fd = -1;
1987
1988 return 0;
1989}
1990
1991static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
1992 char *cmd_data, size_t data_size)
1993{
1994 int err;
1995 char c;
1996 size_t bytes_read = 0;
1997
1998 *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1999 memset(cmd_data, 0, data_size);
2000 data_size--;
2001
2002 do {
2003 err = read(evlist->ctl_fd.fd, &c, 1);
2004 if (err > 0) {
2005 if (c == '\n' || c == '\0')
2006 break;
2007 cmd_data[bytes_read++] = c;
2008 if (bytes_read == data_size)
2009 break;
2010 continue;
2011 } else if (err == -1) {
2012 if (errno == EINTR)
2013 continue;
2014 if (errno == EAGAIN || errno == EWOULDBLOCK)
2015 err = 0;
2016 else
2017 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
2018 }
2019 break;
2020 } while (1);
2021
2022 pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
2023 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
2024
2025 if (bytes_read > 0) {
2026 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
2027 (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
2028 *cmd = EVLIST_CTL_CMD_ENABLE;
2029 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
2030 (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
2031 *cmd = EVLIST_CTL_CMD_DISABLE;
2032 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
2033 (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
2034 *cmd = EVLIST_CTL_CMD_SNAPSHOT;
2035 pr_debug("is snapshot\n");
2036 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG,
2037 (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) {
2038 *cmd = EVLIST_CTL_CMD_EVLIST;
2039 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG,
2040 (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) {
2041 *cmd = EVLIST_CTL_CMD_STOP;
2042 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG,
2043 (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) {
2044 *cmd = EVLIST_CTL_CMD_PING;
2045 }
2046 }
2047
2048 return bytes_read ? (int)bytes_read : err;
2049}
2050
2051int evlist__ctlfd_ack(struct evlist *evlist)
2052{
2053 int err;
2054
2055 if (evlist->ctl_fd.ack == -1)
2056 return 0;
2057
2058 err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
2059 sizeof(EVLIST_CTL_CMD_ACK_TAG));
2060 if (err == -1)
2061 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
2062
2063 return err;
2064}
2065
2066static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg)
2067{
2068 char *data = cmd_data + cmd_size;
2069
2070 /* no argument */
2071 if (!*data)
2072 return 0;
2073
2074 /* there's argument */
2075 if (*data == ' ') {
2076 *arg = data + 1;
2077 return 1;
2078 }
2079
2080 /* malformed */
2081 return -1;
2082}
2083
2084static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable)
2085{
2086 struct evsel *evsel;
2087 char *name;
2088 int err;
2089
2090 err = get_cmd_arg(cmd_data,
2091 enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 :
2092 sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1,
2093 &name);
2094 if (err < 0) {
2095 pr_info("failed: wrong command\n");
2096 return -1;
2097 }
2098
2099 if (err) {
2100 evsel = evlist__find_evsel_by_str(evlist, name);
2101 if (evsel) {
2102 if (enable)
2103 evlist__enable_evsel(evlist, name);
2104 else
2105 evlist__disable_evsel(evlist, name);
2106 pr_info("Event %s %s\n", evsel->name,
2107 enable ? "enabled" : "disabled");
2108 } else {
2109 pr_info("failed: can't find '%s' event\n", name);
2110 }
2111 } else {
2112 if (enable) {
2113 evlist__enable(evlist);
2114 pr_info(EVLIST_ENABLED_MSG);
2115 } else {
2116 evlist__disable(evlist);
2117 pr_info(EVLIST_DISABLED_MSG);
2118 }
2119 }
2120
2121 return 0;
2122}
2123
2124static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data)
2125{
2126 struct perf_attr_details details = { .verbose = false, };
2127 struct evsel *evsel;
2128 char *arg;
2129 int err;
2130
2131 err = get_cmd_arg(cmd_data,
2132 sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1,
2133 &arg);
2134 if (err < 0) {
2135 pr_info("failed: wrong command\n");
2136 return -1;
2137 }
2138
2139 if (err) {
2140 if (!strcmp(arg, "-v")) {
2141 details.verbose = true;
2142 } else if (!strcmp(arg, "-g")) {
2143 details.event_group = true;
2144 } else if (!strcmp(arg, "-F")) {
2145 details.freq = true;
2146 } else {
2147 pr_info("failed: wrong command\n");
2148 return -1;
2149 }
2150 }
2151
2152 evlist__for_each_entry(evlist, evsel)
2153 evsel__fprintf(evsel, &details, stderr);
2154
2155 return 0;
2156}
2157
2158int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
2159{
2160 int err = 0;
2161 char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
2162 int ctlfd_pos = evlist->ctl_fd.pos;
2163 struct pollfd *entries = evlist->core.pollfd.entries;
2164
2165 if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
2166 return 0;
2167
2168 if (entries[ctlfd_pos].revents & POLLIN) {
2169 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
2170 EVLIST_CTL_CMD_MAX_LEN);
2171 if (err > 0) {
2172 switch (*cmd) {
2173 case EVLIST_CTL_CMD_ENABLE:
2174 case EVLIST_CTL_CMD_DISABLE:
2175 err = evlist__ctlfd_enable(evlist, cmd_data,
2176 *cmd == EVLIST_CTL_CMD_ENABLE);
2177 break;
2178 case EVLIST_CTL_CMD_EVLIST:
2179 err = evlist__ctlfd_list(evlist, cmd_data);
2180 break;
2181 case EVLIST_CTL_CMD_SNAPSHOT:
2182 case EVLIST_CTL_CMD_STOP:
2183 case EVLIST_CTL_CMD_PING:
2184 break;
2185 case EVLIST_CTL_CMD_ACK:
2186 case EVLIST_CTL_CMD_UNSUPPORTED:
2187 default:
2188 pr_debug("ctlfd: unsupported %d\n", *cmd);
2189 break;
2190 }
2191 if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
2192 *cmd == EVLIST_CTL_CMD_SNAPSHOT))
2193 evlist__ctlfd_ack(evlist);
2194 }
2195 }
2196
2197 if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
2198 evlist__finalize_ctlfd(evlist);
2199 else
2200 entries[ctlfd_pos].revents = 0;
2201
2202 return err;
2203}
2204
2205/**
2206 * struct event_enable_time - perf record -D/--delay single time range.
2207 * @start: start of time range to enable events in milliseconds
2208 * @end: end of time range to enable events in milliseconds
2209 *
2210 * N.B. this structure is also accessed as an array of int.
2211 */
2212struct event_enable_time {
2213 int start;
2214 int end;
2215};
2216
2217static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first)
2218{
2219 const char *fmt = first ? "%u - %u %n" : " , %u - %u %n";
2220 int ret, start, end, n;
2221
2222 ret = sscanf(str, fmt, &start, &end, &n);
2223 if (ret != 2 || end <= start)
2224 return -EINVAL;
2225 if (range) {
2226 range->start = start;
2227 range->end = end;
2228 }
2229 return n;
2230}
2231
2232static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range)
2233{
2234 int incr = !!range;
2235 bool first = true;
2236 ssize_t ret, cnt;
2237
2238 for (cnt = 0; *str; cnt++) {
2239 ret = parse_event_enable_time(str, range, first);
2240 if (ret < 0)
2241 return ret;
2242 /* Check no overlap */
2243 if (!first && range && range->start <= range[-1].end)
2244 return -EINVAL;
2245 str += ret;
2246 range += incr;
2247 first = false;
2248 }
2249 return cnt;
2250}
2251
2252/**
2253 * struct event_enable_timer - control structure for perf record -D/--delay.
2254 * @evlist: event list
2255 * @times: time ranges that events are enabled (N.B. this is also accessed as an
2256 * array of int)
2257 * @times_cnt: number of time ranges
2258 * @timerfd: timer file descriptor
2259 * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray)
2260 * @times_step: current position in (int *)@times)[],
2261 * refer event_enable_timer__process()
2262 *
2263 * Note, this structure is only used when there are time ranges, not when there
2264 * is only an initial delay.
2265 */
2266struct event_enable_timer {
2267 struct evlist *evlist;
2268 struct event_enable_time *times;
2269 size_t times_cnt;
2270 int timerfd;
2271 int pollfd_pos;
2272 size_t times_step;
2273};
2274
2275static int str_to_delay(const char *str)
2276{
2277 char *endptr;
2278 long d;
2279
2280 d = strtol(str, &endptr, 10);
2281 if (*endptr || d > INT_MAX || d < -1)
2282 return 0;
2283 return d;
2284}
2285
2286int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts,
2287 const char *str, int unset)
2288{
2289 enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event;
2290 struct event_enable_timer *eet;
2291 ssize_t times_cnt;
2292 ssize_t ret;
2293 int err;
2294
2295 if (unset)
2296 return 0;
2297
2298 opts->target.initial_delay = str_to_delay(str);
2299 if (opts->target.initial_delay)
2300 return 0;
2301
2302 ret = parse_event_enable_times(str, NULL);
2303 if (ret < 0)
2304 return ret;
2305
2306 times_cnt = ret;
2307 if (times_cnt == 0)
2308 return -EINVAL;
2309
2310 eet = zalloc(sizeof(*eet));
2311 if (!eet)
2312 return -ENOMEM;
2313
2314 eet->times = calloc(times_cnt, sizeof(*eet->times));
2315 if (!eet->times) {
2316 err = -ENOMEM;
2317 goto free_eet;
2318 }
2319
2320 if (parse_event_enable_times(str, eet->times) != times_cnt) {
2321 err = -EINVAL;
2322 goto free_eet_times;
2323 }
2324
2325 eet->times_cnt = times_cnt;
2326
2327 eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC);
2328 if (eet->timerfd == -1) {
2329 err = -errno;
2330 pr_err("timerfd_create failed: %s\n", strerror(errno));
2331 goto free_eet_times;
2332 }
2333
2334 eet->pollfd_pos = perf_evlist__add_pollfd(&evlist->core, eet->timerfd, NULL, POLLIN, flags);
2335 if (eet->pollfd_pos < 0) {
2336 err = eet->pollfd_pos;
2337 goto close_timerfd;
2338 }
2339
2340 eet->evlist = evlist;
2341 evlist->eet = eet;
2342 opts->target.initial_delay = eet->times[0].start;
2343
2344 return 0;
2345
2346close_timerfd:
2347 close(eet->timerfd);
2348free_eet_times:
2349 zfree(&eet->times);
2350free_eet:
2351 free(eet);
2352 return err;
2353}
2354
2355static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms)
2356{
2357 struct itimerspec its = {
2358 .it_value.tv_sec = ms / MSEC_PER_SEC,
2359 .it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC,
2360 };
2361 int err = 0;
2362
2363 if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) {
2364 err = -errno;
2365 pr_err("timerfd_settime failed: %s\n", strerror(errno));
2366 }
2367 return err;
2368}
2369
2370int event_enable_timer__start(struct event_enable_timer *eet)
2371{
2372 int ms;
2373
2374 if (!eet)
2375 return 0;
2376
2377 ms = eet->times[0].end - eet->times[0].start;
2378 eet->times_step = 1;
2379
2380 return event_enable_timer__set_timer(eet, ms);
2381}
2382
2383int event_enable_timer__process(struct event_enable_timer *eet)
2384{
2385 struct pollfd *entries;
2386 short revents;
2387
2388 if (!eet)
2389 return 0;
2390
2391 entries = eet->evlist->core.pollfd.entries;
2392 revents = entries[eet->pollfd_pos].revents;
2393 entries[eet->pollfd_pos].revents = 0;
2394
2395 if (revents & POLLIN) {
2396 size_t step = eet->times_step;
2397 size_t pos = step / 2;
2398
2399 if (step & 1) {
2400 evlist__disable_non_dummy(eet->evlist);
2401 pr_info(EVLIST_DISABLED_MSG);
2402 if (pos >= eet->times_cnt - 1) {
2403 /* Disarm timer */
2404 event_enable_timer__set_timer(eet, 0);
2405 return 1; /* Stop */
2406 }
2407 } else {
2408 evlist__enable_non_dummy(eet->evlist);
2409 pr_info(EVLIST_ENABLED_MSG);
2410 }
2411
2412 step += 1;
2413 pos = step / 2;
2414
2415 if (pos < eet->times_cnt) {
2416 int *times = (int *)eet->times; /* Accessing 'times' as array of int */
2417 int ms = times[step] - times[step - 1];
2418
2419 eet->times_step = step;
2420 return event_enable_timer__set_timer(eet, ms);
2421 }
2422 }
2423
2424 return 0;
2425}
2426
2427void event_enable_timer__exit(struct event_enable_timer **ep)
2428{
2429 if (!ep || !*ep)
2430 return;
2431 zfree(&(*ep)->times);
2432 zfree(ep);
2433}
2434
2435struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
2436{
2437 struct evsel *evsel;
2438
2439 evlist__for_each_entry(evlist, evsel) {
2440 if (evsel->core.idx == idx)
2441 return evsel;
2442 }
2443 return NULL;
2444}
2445
2446int evlist__scnprintf_evsels(struct evlist *evlist, size_t size, char *bf)
2447{
2448 struct evsel *evsel;
2449 int printed = 0;
2450
2451 evlist__for_each_entry(evlist, evsel) {
2452 if (evsel__is_dummy_event(evsel))
2453 continue;
2454 if (size > (strlen(evsel__name(evsel)) + (printed ? 2 : 1))) {
2455 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "," : "", evsel__name(evsel));
2456 } else {
2457 printed += scnprintf(bf + printed, size - printed, "%s...", printed ? "," : "");
2458 break;
2459 }
2460 }
2461
2462 return printed;
2463}
2464
2465void evlist__check_mem_load_aux(struct evlist *evlist)
2466{
2467 struct evsel *leader, *evsel, *pos;
2468
2469 /*
2470 * For some platforms, the 'mem-loads' event is required to use
2471 * together with 'mem-loads-aux' within a group and 'mem-loads-aux'
2472 * must be the group leader. Now we disable this group before reporting
2473 * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
2474 * any valid memory load information.
2475 */
2476 evlist__for_each_entry(evlist, evsel) {
2477 leader = evsel__leader(evsel);
2478 if (leader == evsel)
2479 continue;
2480
2481 if (leader->name && strstr(leader->name, "mem-loads-aux")) {
2482 for_each_group_evsel(pos, leader) {
2483 evsel__set_leader(pos, pos);
2484 pos->core.nr_members = 0;
2485 }
2486 }
2487 }
2488}
2489
2490/**
2491 * evlist__warn_user_requested_cpus() - Check each evsel against requested CPUs
2492 * and warn if the user CPU list is inapplicable for the event's PMU's
2493 * CPUs. Not core PMUs list a CPU in sysfs, but this may be overwritten by a
2494 * user requested CPU and so any online CPU is applicable. Core PMUs handle
2495 * events on the CPUs in their list and otherwise the event isn't supported.
2496 * @evlist: The list of events being checked.
2497 * @cpu_list: The user provided list of CPUs.
2498 */
2499void evlist__warn_user_requested_cpus(struct evlist *evlist, const char *cpu_list)
2500{
2501 struct perf_cpu_map *user_requested_cpus;
2502 struct evsel *pos;
2503
2504 if (!cpu_list)
2505 return;
2506
2507 user_requested_cpus = perf_cpu_map__new(cpu_list);
2508 if (!user_requested_cpus)
2509 return;
2510
2511 evlist__for_each_entry(evlist, pos) {
2512 struct perf_cpu_map *intersect, *to_test;
2513 const struct perf_pmu *pmu = evsel__find_pmu(pos);
2514
2515 to_test = pmu && pmu->is_core ? pmu->cpus : cpu_map__online();
2516 intersect = perf_cpu_map__intersect(to_test, user_requested_cpus);
2517 if (!perf_cpu_map__equal(intersect, user_requested_cpus)) {
2518 char buf[128];
2519
2520 cpu_map__snprint(to_test, buf, sizeof(buf));
2521 pr_warning("WARNING: A requested CPU in '%s' is not supported by PMU '%s' (CPUs %s) for event '%s'\n",
2522 cpu_list, pmu ? pmu->name : "cpu", buf, evsel__name(pos));
2523 }
2524 perf_cpu_map__put(intersect);
2525 }
2526 perf_cpu_map__put(user_requested_cpus);
2527}
2528
2529void evlist__uniquify_name(struct evlist *evlist)
2530{
2531 char *new_name, empty_attributes[2] = ":", *attributes;
2532 struct evsel *pos;
2533
2534 if (perf_pmus__num_core_pmus() == 1)
2535 return;
2536
2537 evlist__for_each_entry(evlist, pos) {
2538 if (!evsel__is_hybrid(pos))
2539 continue;
2540
2541 if (strchr(pos->name, '/'))
2542 continue;
2543
2544 attributes = strchr(pos->name, ':');
2545 if (attributes)
2546 *attributes = '\0';
2547 else
2548 attributes = empty_attributes;
2549
2550 if (asprintf(&new_name, "%s/%s/%s", pos->pmu_name, pos->name, attributes + 1)) {
2551 free(pos->name);
2552 pos->name = new_name;
2553 } else {
2554 *attributes = ':';
2555 }
2556 }
2557}