projects/linux·blob master · refs

2798 lines · 64.6 KiB

1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <[email protected]>
4 *
5 * Parts came from builtin-{top,stat,record}.c, see those files for further
6 * copyright notes.
7 */
8#include "evlist.h"
9
10#include <errno.h>
11#include <inttypes.h>
12#include <signal.h>
13#include <stdlib.h>
14#include <string.h>
15
16#include <fcntl.h>
17#include <linux/bitops.h>
18#include <linux/err.h>
19#include <linux/hash.h>
20#include <linux/log2.h>
21#include <linux/string.h>
22#include <linux/time64.h>
23#include <linux/zalloc.h>
24#include <poll.h>
25#include <sched.h>
26#include <sys/ioctl.h>
27#include <sys/mman.h>
28#include <sys/prctl.h>
29#include <sys/timerfd.h>
30#include <sys/wait.h>
31#include <unistd.h>
32
33#include <api/fs/fs.h>
34#include <internal/lib.h> // page_size
35#include <internal/rc_check.h>
36#include <internal/xyarray.h>
37#include <perf/cpumap.h>
38#include <perf/evlist.h>
39#include <perf/evsel.h>
40#include <perf/mmap.h>
41#include <subcmd/parse-options.h>
42
43#include "../perf.h"
44#include "affinity.h"
45#include "asm/bug.h"
46#include "bpf-event.h"
47#include "bpf-filter.h"
48#include "bpf_counter.h"
49#include "cpumap.h"
50#include "debug.h"
51#include "dwarf-regs.h"
52#include "env.h"
53#include "event.h"
54#include "evsel.h"
55#include "evsel_fprintf.h"
56#include "intel-tpebs.h"
57#include "metricgroup.h"
58#include "mmap.h"
59#include "parse-events.h"
60#include "perf_api_probe.h"
61#include "pmu.h"
62#include "pmus.h"
63#include "record.h"
64#include "sample.h"
65#include "stat.h"
66#include "strbuf.h"
67#include "string2.h"
68#include "target.h"
69#include "thread_map.h"
70#include "units.h"
71#include "util.h"
72
73#ifdef LACKS_SIGQUEUE_PROTOTYPE
74int sigqueue(pid_t pid, int sig, const union sigval value);
75#endif
76
77#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
78#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
79
80static void event_enable_timer__exit(struct event_enable_timer **ep);
81
82struct evlist *evlist__new(void)
83{
84 struct evlist *result;
85 RC_STRUCT(evlist) *evlist;
86
87 evlist = zalloc(sizeof(*evlist));
88 if (ADD_RC_CHK(result, evlist)) {
89 perf_evlist__init(evlist__core(result));
90 perf_evlist__set_maps(evlist__core(result), /*cpus=*/NULL, /*threads=*/NULL);
91 evlist__set_workload_pid(result, -1);
92 evlist__set_bkw_mmap_state(result, BKW_MMAP_NOTREADY);
93 evlist__set_ctl_fd_fd(result, -1);
94 evlist__set_ctl_fd_ack(result, -1);
95 evlist__set_ctl_fd_pos(result, -1);
96 evlist__set_nr_br_cntr(result, -1);
97 metricgroup__rblist_init(evlist__metric_events(result));
98 INIT_LIST_HEAD(&evlist->deferred_samples);
99 refcount_set(evlist__refcnt(result), 1);
100 } else {
101 free(evlist);
102 result = NULL;
103 }
104 return result;
105}
106
107struct evlist *evlist__new_default(const struct target *target, bool sample_callchains)
108{
109 struct evlist *evlist = evlist__new();
110 bool can_profile_kernel;
111 struct perf_pmu *pmu = NULL;
112 char buf[256];
113 int err;
114
115 if (!evlist)
116 return NULL;
117
118 can_profile_kernel = perf_event_paranoid_check(1);
119
120 if (EM_HOST == EM_S390 && sample_callchains) {
121 snprintf(buf, sizeof(buf), "software/%s/%s",
122 target__has_cpu(target) ? "cpu-clock" : "task-clock",
123 can_profile_kernel ? "P" : "Pu");
124 err = parse_event(evlist, buf);
125 if (err)
126 goto out_err;
127 } else {
128 while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
129 snprintf(buf, sizeof(buf), "%s/cycles/%s", pmu->name,
130 can_profile_kernel ? "P" : "Pu");
131 err = parse_event(evlist, buf);
132 if (err)
133 goto out_err;
134 }
135 }
136
137 /* If there is only 1 event a sample identifier isn't necessary. */
138 if (evlist__nr_entries(evlist) > 1) {
139 struct evsel *evsel;
140
141 evlist__for_each_entry(evlist, evsel)
142 evsel__set_sample_id(evsel, /*can_sample_identifier=*/false);
143 }
144
145 return evlist;
146out_err:
147 evlist__put(evlist);
148 return NULL;
149}
150
151struct evlist *evlist__new_dummy(void)
152{
153 struct evlist *evlist = evlist__new();
154
155 if (evlist && evlist__add_dummy(evlist)) {
156 evlist__put(evlist);
157 evlist = NULL;
158 }
159
160 return evlist;
161}
162
163struct evlist *evlist__get(struct evlist *evlist)
164{
165 struct evlist *result;
166
167 if (RC_CHK_GET(result, evlist))
168 refcount_inc(evlist__refcnt(evlist));
169
170 return result;
171}
172
173/**
174 * evlist__set_id_pos - set the positions of event ids.
175 * @evlist: selected event list
176 *
177 * Events with compatible sample types all have the same id_pos
178 * and is_pos. For convenience, put a copy on evlist.
179 */
180void evlist__set_id_pos(struct evlist *evlist)
181{
182 struct evsel *first = evlist__first(evlist);
183
184 RC_CHK_ACCESS(evlist)->id_pos = first->id_pos;
185 RC_CHK_ACCESS(evlist)->is_pos = first->is_pos;
186}
187
188static void evlist__update_id_pos(struct evlist *evlist)
189{
190 struct evsel *evsel;
191
192 evlist__for_each_entry(evlist, evsel)
193 evsel__calc_id_pos(evsel);
194
195 evlist__set_id_pos(evlist);
196}
197
198static void evlist__purge(struct evlist *evlist)
199{
200 struct evsel *pos, *n;
201
202 evlist__for_each_entry_safe(evlist, n, pos) {
203 list_del_init(&pos->core.node);
204 if (pos->evlist) {
205 if (!RC_CHK_EQUAL(pos->evlist, evlist)) {
206 evlist__put(pos->evlist);
207 } else {
208 refcount_dec_and_test(evlist__refcnt(pos->evlist));
209 RC_CHK_PUT(pos->evlist);
210 }
211 }
212 pos->evlist = NULL;
213 evsel__put(pos);
214 }
215
216 evlist__core(evlist)->nr_entries = 0;
217}
218
219static void evlist__exit(struct evlist *evlist)
220{
221 metricgroup__rblist_exit(evlist__metric_events(evlist));
222 event_enable_timer__exit(&RC_CHK_ACCESS(evlist)->eet);
223 free(evlist__mmap(evlist));
224 free(evlist__overwrite_mmap(evlist));
225 perf_evlist__exit(evlist__core(evlist));
226}
227
228void evlist__put(struct evlist *evlist)
229{
230 struct evsel *evsel;
231 unsigned int count, ref_cnt;
232
233 if (evlist == NULL)
234 return;
235
236 if (refcount_dec_and_test(evlist__refcnt(evlist)))
237 goto out_delete;
238
239retry:
240 count = refcount_read(evlist__refcnt(evlist));
241 ref_cnt = count;
242 evlist__for_each_entry(evlist, evsel) {
243 if (RC_CHK_EQUAL(evsel->evlist, evlist) && count &&
244 refcount_read(&evsel->refcnt) == 1)
245 count--;
246 }
247 if (refcount_read(evlist__refcnt(evlist)) != ref_cnt)
248 goto retry;
249
250 if (count != 0) {
251 /*
252 * Not the last reference except for back references from
253 * evsels.
254 */
255 RC_CHK_PUT(evlist);
256 return;
257 }
258out_delete:
259 evlist__free_stats(evlist);
260 evlist__do_munmap(evlist);
261 evlist__close(evlist);
262 evlist__purge(evlist);
263 evlist__exit(evlist);
264 RC_CHK_FREE(evlist);
265}
266
267void evlist__add(struct evlist *evlist, struct evsel *entry)
268{
269 perf_evlist__add(evlist__core(evlist), &entry->core);
270 evlist__put(entry->evlist);
271 entry->evlist = evlist__get(evlist);
272 entry->tracking = !entry->core.idx;
273
274 if (evlist__nr_entries(evlist) == 1)
275 evlist__set_id_pos(evlist);
276}
277
278void evlist__remove(struct evlist *evlist, struct evsel *evsel)
279{
280 perf_evlist__remove(evlist__core(evlist), &evsel->core);
281 evlist__put(evsel->evlist);
282 evsel->evlist = NULL;
283}
284
285void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list)
286{
287 while (!list_empty(list)) {
288 struct evsel *evsel, *temp, *leader = NULL;
289
290 __evlist__for_each_entry_safe(list, temp, evsel) {
291 list_del_init(&evsel->core.node);
292 evlist__add(evlist, evsel);
293 leader = evsel;
294 break;
295 }
296
297 __evlist__for_each_entry_safe(list, temp, evsel) {
298 if (evsel__has_leader(evsel, leader)) {
299 list_del_init(&evsel->core.node);
300 evlist__add(evlist, evsel);
301 }
302 }
303 }
304}
305
306int __evlist__set_tracepoints_handlers(struct evlist *evlist,
307 const struct evsel_str_handler *assocs, size_t nr_assocs)
308{
309 size_t i;
310 int err;
311
312 for (i = 0; i < nr_assocs; i++) {
313 // Adding a handler for an event not in this evlist, just ignore it.
314 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name);
315 if (evsel == NULL)
316 continue;
317
318 err = -EEXIST;
319 if (evsel->handler != NULL)
320 goto out;
321 evsel->handler = assocs[i].handler;
322 }
323
324 err = 0;
325out:
326 return err;
327}
328
329static void evlist__set_leader(struct evlist *evlist)
330{
331 perf_evlist__set_leader(evlist__core(evlist));
332}
333
334static struct evsel *evlist__dummy_event(struct evlist *evlist)
335{
336 struct perf_event_attr attr = {
337 .type = PERF_TYPE_SOFTWARE,
338 .config = PERF_COUNT_SW_DUMMY,
339 .size = sizeof(attr), /* to capture ABI version */
340 /* Avoid frequency mode for dummy events to avoid associated timers. */
341 .freq = 0,
342 .sample_period = 1,
343 };
344
345 return evsel__new_idx(&attr, evlist__nr_entries(evlist));
346}
347
348int evlist__add_dummy(struct evlist *evlist)
349{
350 struct evsel *evsel = evlist__dummy_event(evlist);
351
352 if (evsel == NULL)
353 return -ENOMEM;
354
355 evlist__add(evlist, evsel);
356 return 0;
357}
358
359struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
360{
361 struct evsel *evsel = evlist__dummy_event(evlist);
362
363 if (!evsel)
364 return NULL;
365
366 evsel->core.attr.exclude_kernel = 1;
367 evsel->core.attr.exclude_guest = 1;
368 evsel->core.attr.exclude_hv = 1;
369 evsel->core.system_wide = system_wide;
370 evsel->no_aux_samples = true;
371 evsel->name = strdup("dummy:u");
372
373 evlist__add(evlist, evsel);
374 return evsel;
375}
376
377#ifdef HAVE_LIBTRACEEVENT
378struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide)
379{
380 struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0,
381 /*format=*/true);
382
383 if (IS_ERR(evsel))
384 return evsel;
385
386 evsel__set_sample_bit(evsel, CPU);
387 evsel__set_sample_bit(evsel, TIME);
388
389 evsel->core.system_wide = system_wide;
390 evsel->no_aux_samples = true;
391
392 evlist__add(evlist, evsel);
393 return evsel;
394}
395#endif
396
397struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name)
398{
399 struct evsel *evsel;
400
401 evlist__for_each_entry(evlist, evsel) {
402 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
403 (strcmp(evsel->name, name) == 0))
404 return evsel;
405 }
406
407 return NULL;
408}
409
410#ifdef HAVE_LIBTRACEEVENT
411int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
412{
413 struct evsel *evsel = evsel__newtp(sys, name);
414
415 if (IS_ERR(evsel))
416 return -1;
417
418 evsel->handler = handler;
419 evlist__add(evlist, evsel);
420 return 0;
421}
422#endif
423
424/*
425 * Should sched_setaffinity be used with evlist__for_each_cpu? Determine if
426 * migrating the thread will avoid possibly numerous IPIs.
427 */
428static bool evlist__use_affinity(struct evlist *evlist)
429{
430 struct evsel *pos;
431 struct perf_cpu_map *used_cpus = NULL;
432 bool ret = false;
433
434 if (evlist__no_affinity(evlist) || !evlist__core(evlist)->user_requested_cpus ||
435 cpu_map__is_dummy(evlist__core(evlist)->user_requested_cpus))
436 return false;
437
438 evlist__for_each_entry(evlist, pos) {
439 struct perf_cpu_map *intersect;
440
441 if (!perf_pmu__benefits_from_affinity(pos->pmu))
442 continue;
443
444 if (evsel__is_dummy_event(pos)) {
445 /*
446 * The dummy event is opened on all CPUs so assume >1
447 * event with shared CPUs.
448 */
449 ret = true;
450 break;
451 }
452 if (evsel__is_retire_lat(pos)) {
453 /*
454 * Retirement latency events are similar to tool ones in
455 * their implementation, and so don't require affinity.
456 */
457 continue;
458 }
459 if (perf_cpu_map__is_empty(used_cpus)) {
460 /* First benefitting event, we want >1 on a common CPU. */
461 used_cpus = perf_cpu_map__get(pos->core.cpus);
462 continue;
463 }
464 if ((pos->core.attr.read_format & PERF_FORMAT_GROUP) &&
465 evsel__leader(pos) != pos) {
466 /* Skip members of the same sample group. */
467 continue;
468 }
469 intersect = perf_cpu_map__intersect(used_cpus, pos->core.cpus);
470 if (!perf_cpu_map__is_empty(intersect)) {
471 /* >1 event with shared CPUs. */
472 perf_cpu_map__put(intersect);
473 ret = true;
474 break;
475 }
476 perf_cpu_map__put(intersect);
477 perf_cpu_map__merge(&used_cpus, pos->core.cpus);
478 }
479 perf_cpu_map__put(used_cpus);
480 return ret;
481}
482
483void evlist_cpu_iterator__init(struct evlist_cpu_iterator *itr, struct evlist *evlist)
484{
485 *itr = (struct evlist_cpu_iterator){
486 .container = evlist,
487 .evsel = NULL,
488 .cpu_map_idx = 0,
489 .evlist_cpu_map_idx = 0,
490 .evlist_cpu_map_nr = perf_cpu_map__nr(evlist__core(evlist)->all_cpus),
491 .cpu = (struct perf_cpu){ .cpu = -1},
492 .affinity = NULL,
493 };
494
495 if (evlist__empty(evlist)) {
496 /* Ensure the empty list doesn't iterate. */
497 itr->evlist_cpu_map_idx = itr->evlist_cpu_map_nr;
498 return;
499 }
500
501 if (evlist__use_affinity(evlist)) {
502 if (affinity__setup(&itr->saved_affinity) == 0)
503 itr->affinity = &itr->saved_affinity;
504 }
505 itr->evsel = evlist__first(evlist);
506 itr->cpu = perf_cpu_map__cpu(evlist__core(evlist)->all_cpus, 0);
507 if (itr->affinity)
508 affinity__set(itr->affinity, itr->cpu.cpu);
509 itr->cpu_map_idx = perf_cpu_map__idx(itr->evsel->core.cpus, itr->cpu);
510 /*
511 * If this CPU isn't in the evsel's cpu map then advance
512 * through the list.
513 */
514 if (itr->cpu_map_idx == -1)
515 evlist_cpu_iterator__next(itr);
516}
517
518void evlist_cpu_iterator__exit(struct evlist_cpu_iterator *itr)
519{
520 if (!itr->affinity)
521 return;
522
523 affinity__cleanup(itr->affinity);
524 itr->affinity = NULL;
525}
526
527void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr)
528{
529 while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) {
530 evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel);
531 evlist_cpu_itr->cpu_map_idx =
532 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
533 evlist_cpu_itr->cpu);
534 if (evlist_cpu_itr->cpu_map_idx != -1)
535 return;
536 }
537 evlist_cpu_itr->evlist_cpu_map_idx++;
538 if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) {
539 evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container);
540 evlist_cpu_itr->cpu =
541 perf_cpu_map__cpu(evlist__core(evlist_cpu_itr->container)->all_cpus,
542 evlist_cpu_itr->evlist_cpu_map_idx);
543 if (evlist_cpu_itr->affinity)
544 affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu);
545 evlist_cpu_itr->cpu_map_idx =
546 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
547 evlist_cpu_itr->cpu);
548 /*
549 * If this CPU isn't in the evsel's cpu map then advance through
550 * the list.
551 */
552 if (evlist_cpu_itr->cpu_map_idx == -1)
553 evlist_cpu_iterator__next(evlist_cpu_itr);
554 } else {
555 evlist_cpu_iterator__exit(evlist_cpu_itr);
556 }
557}
558
559static int evsel__strcmp(struct evsel *pos, char *evsel_name)
560{
561 if (!evsel_name)
562 return 0;
563 if (evsel__is_dummy_event(pos))
564 return 1;
565 return !evsel__name_is(pos, evsel_name);
566}
567
568static bool evlist__is_enabled(struct evlist *evlist)
569{
570 struct evsel *pos;
571
572 evlist__for_each_entry(evlist, pos) {
573 if (!evsel__is_group_leader(pos) || !pos->core.fd)
574 continue;
575 /* If at least one event is enabled, evlist is enabled. */
576 if (!pos->disabled)
577 return true;
578 }
579 return false;
580}
581
582static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
583{
584 struct evsel *pos, *member;
585 struct evlist_cpu_iterator evlist_cpu_itr;
586 bool has_imm = false;
587
588 /* Disable 'immediate' events last */
589 for (int imm = 0; imm <= 1; imm++) {
590 evlist__for_each_cpu(evlist_cpu_itr, evlist) {
591 pos = evlist_cpu_itr.evsel;
592 if (evsel__strcmp(pos, evsel_name))
593 continue;
594 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
595 continue;
596 if (excl_dummy && evsel__is_dummy_event(pos))
597 continue;
598 if (pos->immediate)
599 has_imm = true;
600 if (pos->immediate != imm)
601 continue;
602 evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
603 }
604 if (!has_imm)
605 break;
606 }
607
608 evlist__for_each_entry(evlist, pos) {
609 if (evsel__strcmp(pos, evsel_name))
610 continue;
611 if (!evsel__is_group_leader(pos) || !pos->core.fd)
612 continue;
613 if (excl_dummy && evsel__is_dummy_event(pos))
614 continue;
615 pos->disabled = true;
616
617 for_each_group_member(member, pos)
618 member->disabled = true;
619 }
620
621 /*
622 * If we disabled only single event, we need to check
623 * the enabled state of the evlist manually.
624 */
625 evlist__set_enabled(evlist, evsel_name ? evlist__is_enabled(evlist) : false);
626}
627
628void evlist__disable(struct evlist *evlist)
629{
630 __evlist__disable(evlist, NULL, false);
631}
632
633void evlist__disable_non_dummy(struct evlist *evlist)
634{
635 __evlist__disable(evlist, NULL, true);
636}
637
638void evlist__disable_evsel(struct evlist *evlist, char *evsel_name)
639{
640 __evlist__disable(evlist, evsel_name, false);
641}
642
643static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
644{
645 struct evsel *pos, *member;
646 struct evlist_cpu_iterator evlist_cpu_itr;
647
648 evlist__for_each_cpu(evlist_cpu_itr, evlist) {
649 pos = evlist_cpu_itr.evsel;
650 if (evsel__strcmp(pos, evsel_name))
651 continue;
652 if (!evsel__is_group_leader(pos) || !pos->core.fd)
653 continue;
654 if (excl_dummy && evsel__is_dummy_event(pos))
655 continue;
656 evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
657 }
658 evlist__for_each_entry(evlist, pos) {
659 if (evsel__strcmp(pos, evsel_name))
660 continue;
661 if (!evsel__is_group_leader(pos) || !pos->core.fd)
662 continue;
663 if (excl_dummy && evsel__is_dummy_event(pos))
664 continue;
665 pos->disabled = false;
666
667 for_each_group_member(member, pos)
668 member->disabled = false;
669 }
670
671 /*
672 * Even single event sets the 'enabled' for evlist,
673 * so the toggle can work properly and toggle to
674 * 'disabled' state.
675 */
676 evlist__set_enabled(evlist, true);
677}
678
679void evlist__enable(struct evlist *evlist)
680{
681 __evlist__enable(evlist, NULL, false);
682}
683
684void evlist__enable_non_dummy(struct evlist *evlist)
685{
686 __evlist__enable(evlist, NULL, true);
687}
688
689void evlist__enable_evsel(struct evlist *evlist, char *evsel_name)
690{
691 __evlist__enable(evlist, evsel_name, false);
692}
693
694void evlist__toggle_enable(struct evlist *evlist)
695{
696 (evlist__enabled(evlist) ? evlist__disable : evlist__enable)(evlist);
697}
698
699int evlist__add_pollfd(struct evlist *evlist, int fd)
700{
701 return perf_evlist__add_pollfd(evlist__core(evlist), fd, NULL, POLLIN,
702 fdarray_flag__default);
703}
704
705int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
706{
707 return perf_evlist__filter_pollfd(evlist__core(evlist), revents_and_mask);
708}
709
710#ifdef HAVE_EVENTFD_SUPPORT
711int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd)
712{
713 return perf_evlist__add_pollfd(evlist__core(evlist), fd, NULL, POLLIN,
714 fdarray_flag__nonfilterable |
715 fdarray_flag__non_perf_event);
716}
717#endif
718
719int evlist__poll(struct evlist *evlist, int timeout)
720{
721 return perf_evlist__poll(evlist__core(evlist), timeout);
722}
723
724struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id)
725{
726 struct hlist_head *head;
727 struct perf_sample_id *sid;
728 int hash;
729
730 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
731 head = &evlist__core(evlist)->heads[hash];
732
733 hlist_for_each_entry(sid, head, node)
734 if (sid->id == id)
735 return sid;
736
737 return NULL;
738}
739
740struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id)
741{
742 struct perf_sample_id *sid;
743
744 if (evlist__nr_entries(evlist) == 1 || !id)
745 return evlist__first(evlist);
746
747 sid = evlist__id2sid(evlist, id);
748 if (sid)
749 return container_of(sid->evsel, struct evsel, core);
750
751 if (!evlist__sample_id_all(evlist))
752 return evlist__first(evlist);
753
754 return NULL;
755}
756
757struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id)
758{
759 struct perf_sample_id *sid;
760
761 if (!id)
762 return NULL;
763
764 sid = evlist__id2sid(evlist, id);
765 if (sid)
766 return container_of(sid->evsel, struct evsel, core);
767
768 return NULL;
769}
770
771static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id)
772{
773 const __u64 *array = event->sample.array;
774 ssize_t n;
775
776 n = (event->header.size - sizeof(event->header)) >> 3;
777
778 if (event->header.type == PERF_RECORD_SAMPLE) {
779 if (evlist__id_pos(evlist) >= n)
780 return -1;
781 *id = array[evlist__id_pos(evlist)];
782 } else {
783 if (evlist__is_pos(evlist) > n)
784 return -1;
785 n -= evlist__is_pos(evlist);
786 *id = array[n];
787 }
788 return 0;
789}
790
791struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event)
792{
793 struct evsel *first = evlist__first(evlist);
794 struct hlist_head *head;
795 struct perf_sample_id *sid;
796 int hash;
797 u64 id;
798
799 if (evlist__nr_entries(evlist) == 1)
800 return first;
801
802 if (!first->core.attr.sample_id_all &&
803 event->header.type != PERF_RECORD_SAMPLE)
804 return first;
805
806 if (evlist__event2id(evlist, event, &id))
807 return NULL;
808
809 /* Synthesized events have an id of zero */
810 if (!id)
811 return first;
812
813 hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
814 head = &evlist__core(evlist)->heads[hash];
815
816 hlist_for_each_entry(sid, head, node) {
817 if (sid->id == id)
818 return container_of(sid->evsel, struct evsel, core);
819 }
820 return NULL;
821}
822
823static int evlist__set_paused(struct evlist *evlist, bool value)
824{
825 int i;
826
827 if (!evlist__overwrite_mmap(evlist))
828 return 0;
829
830 for (i = 0; i < evlist__core(evlist)->nr_mmaps; i++) {
831 int fd = evlist__overwrite_mmap(evlist)[i].core.fd;
832 int err;
833
834 if (fd < 0)
835 continue;
836 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
837 if (err)
838 return err;
839 }
840 return 0;
841}
842
843static int evlist__pause(struct evlist *evlist)
844{
845 return evlist__set_paused(evlist, true);
846}
847
848static int evlist__resume(struct evlist *evlist)
849{
850 return evlist__set_paused(evlist, false);
851}
852
853static void evlist__munmap_nofree(struct evlist *evlist)
854{
855 int i;
856
857 if (evlist__mmap(evlist))
858 for (i = 0; i < evlist__core(evlist)->nr_mmaps; i++)
859 perf_mmap__munmap(&evlist__mmap(evlist)[i].core);
860
861 if (evlist__overwrite_mmap(evlist))
862 for (i = 0; i < evlist__core(evlist)->nr_mmaps; i++)
863 perf_mmap__munmap(&evlist__overwrite_mmap(evlist)[i].core);
864}
865
866void evlist__do_munmap(struct evlist *evlist)
867{
868 evlist__munmap_nofree(evlist);
869 zfree(&RC_CHK_ACCESS(evlist)->mmap);
870 zfree(&RC_CHK_ACCESS(evlist)->overwrite_mmap);
871}
872
873static void perf_mmap__unmap_cb(struct perf_mmap *map)
874{
875 struct mmap *m = container_of(map, struct mmap, core);
876
877 mmap__munmap(m);
878}
879
880static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
881 bool overwrite)
882{
883 int i;
884 struct mmap *map = calloc(evlist__core(evlist)->nr_mmaps, sizeof(struct mmap));
885
886 if (!map)
887 return NULL;
888
889 for (i = 0; i < evlist__core(evlist)->nr_mmaps; i++) {
890 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
891
892 /*
893 * When the perf_mmap() call is made we grab one refcount, plus
894 * one extra to let perf_mmap__consume() get the last
895 * events after all real references (perf_mmap__get()) are
896 * dropped.
897 *
898 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
899 * thus does perf_mmap__get() on it.
900 */
901 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
902 }
903
904 return map;
905}
906
907static struct evlist *from_list_start(struct perf_evlist *core)
908{
909#ifdef REFCNT_CHECKING
910 RC_STRUCT(evlist) *core_evlist = container_of(core, RC_STRUCT(evlist), core);
911 struct evlist *evlist;
912
913 if (ADD_RC_CHK(evlist, core_evlist))
914 refcount_inc(evlist__refcnt(evlist));
915
916 return evlist;
917#else
918 return container_of(core, struct evlist, core);
919#endif
920}
921
922static void from_list_end(struct evlist *evlist __maybe_unused)
923{
924#ifdef REFCNT_CHECKING
925 evlist__put(evlist);
926#endif
927}
928
929static void
930perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
931 struct perf_evsel *_evsel,
932 struct perf_mmap_param *_mp,
933 int idx)
934{
935 struct evlist *evlist = from_list_start(_evlist);
936 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
937 struct evsel *evsel = container_of(_evsel, struct evsel, core);
938
939 if (!evlist)
940 return;
941
942 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx);
943
944 from_list_end(evlist);
945}
946
947static struct perf_mmap*
948perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
949{
950 struct evlist *evlist = from_list_start(_evlist);
951 struct mmap *maps;
952
953 if (!evlist)
954 return NULL;
955
956 maps = overwrite ? evlist__overwrite_mmap(evlist) : evlist__mmap(evlist);
957
958 if (!maps) {
959 maps = evlist__alloc_mmap(evlist, overwrite);
960 if (!maps) {
961 from_list_end(evlist);
962 return NULL;
963 }
964
965 if (overwrite) {
966 RC_CHK_ACCESS(evlist)->overwrite_mmap = maps;
967 if (evlist__bkw_mmap_state(evlist) == BKW_MMAP_NOTREADY)
968 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
969 } else {
970 RC_CHK_ACCESS(evlist)->mmap = maps;
971 }
972 }
973 from_list_end(evlist);
974 return &maps[idx].core;
975}
976
977static int
978perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
979 int output, struct perf_cpu cpu)
980{
981 struct mmap *map = container_of(_map, struct mmap, core);
982 struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
983
984 return mmap__mmap(map, mp, output, cpu);
985}
986
987unsigned long perf_event_mlock_kb_in_pages(void)
988{
989 unsigned long pages;
990 int max;
991
992 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
993 /*
994 * Pick a once upon a time good value, i.e. things look
995 * strange since we can't read a sysctl value, but lets not
996 * die yet...
997 */
998 max = 512;
999 } else {
1000 max -= (page_size / 1024);
1001 }
1002
1003 pages = (max * 1024) / page_size;
1004 if (!is_power_of_2(pages))
1005 pages = rounddown_pow_of_two(pages);
1006
1007 return pages;
1008}
1009
1010size_t evlist__mmap_size(unsigned long pages)
1011{
1012 if (pages == UINT_MAX)
1013 pages = perf_event_mlock_kb_in_pages();
1014 else if (!is_power_of_2(pages))
1015 return 0;
1016
1017 return (pages + 1) * page_size;
1018}
1019
1020static long parse_pages_arg(const char *str, unsigned long min,
1021 unsigned long max)
1022{
1023 unsigned long pages, val;
1024 static struct parse_tag tags[] = {
1025 { .tag = 'B', .mult = 1 },
1026 { .tag = 'K', .mult = 1 << 10 },
1027 { .tag = 'M', .mult = 1 << 20 },
1028 { .tag = 'G', .mult = 1 << 30 },
1029 { .tag = 0 },
1030 };
1031
1032 if (str == NULL)
1033 return -EINVAL;
1034
1035 val = parse_tag_value(str, tags);
1036 if (val != (unsigned long) -1) {
1037 /* we got file size value */
1038 pages = PERF_ALIGN(val, page_size) / page_size;
1039 } else {
1040 /* we got pages count value */
1041 char *eptr;
1042 pages = strtoul(str, &eptr, 10);
1043 if (*eptr != '\0')
1044 return -EINVAL;
1045 }
1046
1047 if (pages == 0 && min == 0) {
1048 /* leave number of pages at 0 */
1049 } else if (!is_power_of_2(pages)) {
1050 char buf[100];
1051
1052 /* round pages up to next power of 2 */
1053 pages = roundup_pow_of_two(pages);
1054 if (!pages)
1055 return -EINVAL;
1056
1057 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
1058 pr_info("rounding mmap pages size to %s (%lu pages)\n",
1059 buf, pages);
1060 }
1061
1062 if (pages > max)
1063 return -EINVAL;
1064
1065 return pages;
1066}
1067
1068int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1069{
1070 unsigned long max = UINT_MAX;
1071 long pages;
1072
1073 if (max > SIZE_MAX / page_size)
1074 max = SIZE_MAX / page_size;
1075
1076 pages = parse_pages_arg(str, 1, max);
1077 if (pages < 0) {
1078 pr_err("Invalid argument for --mmap_pages/-m\n");
1079 return -1;
1080 }
1081
1082 *mmap_pages = pages;
1083 return 0;
1084}
1085
1086int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused)
1087{
1088 return __evlist__parse_mmap_pages(opt->value, str);
1089}
1090
1091/**
1092 * evlist__mmap_ex - Create mmaps to receive events.
1093 * @evlist: list of events
1094 * @pages: map length in pages
1095 * @overwrite: overwrite older events?
1096 * @auxtrace_pages - auxtrace map length in pages
1097 * @auxtrace_overwrite - overwrite older auxtrace data?
1098 *
1099 * If @overwrite is %false the user needs to signal event consumption using
1100 * perf_mmap__write_tail(). Using evlist__mmap_read() does this
1101 * automatically.
1102 *
1103 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1104 * consumption using auxtrace_mmap__write_tail().
1105 *
1106 * Return: %0 on success, negative error code otherwise.
1107 */
1108int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
1109 unsigned int auxtrace_pages,
1110 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
1111 int comp_level)
1112{
1113 /*
1114 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1115 * Its value is decided by evsel's write_backward.
1116 * So &mp should not be passed through const pointer.
1117 */
1118 struct mmap_params mp = {
1119 .nr_cblocks = nr_cblocks,
1120 .affinity = affinity,
1121 .flush = flush,
1122 .comp_level = comp_level
1123 };
1124 struct perf_evlist_mmap_ops ops = {
1125 .idx = perf_evlist__mmap_cb_idx,
1126 .get = perf_evlist__mmap_cb_get,
1127 .mmap = perf_evlist__mmap_cb_mmap,
1128 };
1129
1130 evlist__core(evlist)->mmap_len = evlist__mmap_size(pages);
1131 pr_debug("mmap size %zuB\n", evlist__core(evlist)->mmap_len);
1132
1133 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist__core(evlist)->mmap_len,
1134 auxtrace_pages, auxtrace_overwrite);
1135
1136 return perf_evlist__mmap_ops(evlist__core(evlist), &ops, &mp.core);
1137}
1138
1139int evlist__do_mmap(struct evlist *evlist, unsigned int pages)
1140{
1141 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1142}
1143
1144int evlist__create_maps(struct evlist *evlist, struct target *target)
1145{
1146 bool all_threads = (target->per_thread && target->system_wide);
1147 struct perf_cpu_map *cpus;
1148 struct perf_thread_map *threads;
1149
1150 /*
1151 * If specify '-a' and '--per-thread' to perf record, perf record
1152 * will override '--per-thread'. target->per_thread = false and
1153 * target->system_wide = true.
1154 *
1155 * If specify '--per-thread' only to perf record,
1156 * target->per_thread = true and target->system_wide = false.
1157 *
1158 * So target->per_thread && target->system_wide is false.
1159 * For perf record, thread_map__new_str doesn't call
1160 * thread_map__new_all_cpus. That will keep perf record's
1161 * current behavior.
1162 *
1163 * For perf stat, it allows the case that target->per_thread and
1164 * target->system_wide are all true. It means to collect system-wide
1165 * per-thread data. thread_map__new_str will call
1166 * thread_map__new_all_cpus to enumerate all threads.
1167 */
1168 threads = thread_map__new_str(target->pid, target->tid, all_threads);
1169
1170 if (!threads)
1171 return -1;
1172
1173 if (target__uses_dummy_map(target) && !evlist__has_bpf_output(evlist))
1174 cpus = perf_cpu_map__new_any_cpu();
1175 else
1176 cpus = perf_cpu_map__new(target->cpu_list);
1177
1178 if (!cpus)
1179 goto out_delete_threads;
1180
1181 evlist__core(evlist)->has_user_cpus = !!target->cpu_list;
1182
1183 perf_evlist__set_maps(evlist__core(evlist), cpus, threads);
1184
1185 /* as evlist now has references, put count here */
1186 perf_cpu_map__put(cpus);
1187 perf_thread_map__put(threads);
1188
1189 return 0;
1190
1191out_delete_threads:
1192 perf_thread_map__put(threads);
1193 return -1;
1194}
1195
1196int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel,
1197 struct target *target)
1198{
1199 struct evsel *evsel;
1200 int err = 0;
1201
1202 evlist__for_each_entry(evlist, evsel) {
1203 /*
1204 * filters only work for tracepoint event, which doesn't have cpu limit.
1205 * So evlist and evsel should always be same.
1206 */
1207 if (evsel->filter) {
1208 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1209 if (err) {
1210 *err_evsel = evsel;
1211 break;
1212 }
1213 }
1214
1215 /*
1216 * non-tracepoint events can have BPF filters.
1217 */
1218 if (!list_empty(&evsel->bpf_filters)) {
1219 err = perf_bpf_filter__prepare(evsel, target);
1220 if (err) {
1221 *err_evsel = evsel;
1222 break;
1223 }
1224 }
1225 }
1226
1227 return err;
1228}
1229
1230int evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1231{
1232 struct evsel *evsel;
1233 int err = 0;
1234
1235 if (filter == NULL)
1236 return -1;
1237
1238 evlist__for_each_entry(evlist, evsel) {
1239 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1240 continue;
1241
1242 err = evsel__set_filter(evsel, filter);
1243 if (err)
1244 break;
1245 }
1246
1247 return err;
1248}
1249
1250int evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1251{
1252 struct evsel *evsel;
1253 int err = 0;
1254
1255 if (filter == NULL)
1256 return -1;
1257
1258 evlist__for_each_entry(evlist, evsel) {
1259 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1260 continue;
1261
1262 err = evsel__append_tp_filter(evsel, filter);
1263 if (err)
1264 break;
1265 }
1266
1267 return err;
1268}
1269
1270char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1271{
1272 char *filter;
1273 size_t i;
1274
1275 for (i = 0; i < npids; ++i) {
1276 if (i == 0) {
1277 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1278 return NULL;
1279 } else {
1280 char *tmp;
1281
1282 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1283 goto out_free;
1284
1285 free(filter);
1286 filter = tmp;
1287 }
1288 }
1289
1290 return filter;
1291out_free:
1292 free(filter);
1293 return NULL;
1294}
1295
1296int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1297{
1298 char *filter = asprintf__tp_filter_pids(npids, pids);
1299 int ret = evlist__set_tp_filter(evlist, filter);
1300
1301 free(filter);
1302 return ret;
1303}
1304
1305int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1306{
1307 char *filter = asprintf__tp_filter_pids(npids, pids);
1308 int ret = evlist__append_tp_filter(evlist, filter);
1309
1310 free(filter);
1311 return ret;
1312}
1313
1314int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1315{
1316 return evlist__append_tp_filter_pids(evlist, 1, &pid);
1317}
1318
1319bool evlist__valid_sample_type(struct evlist *evlist)
1320{
1321 struct evsel *pos;
1322
1323 if (evlist__nr_entries(evlist) == 1)
1324 return true;
1325
1326 if (evlist__id_pos(evlist) < 0 || evlist__is_pos(evlist) < 0)
1327 return false;
1328
1329 evlist__for_each_entry(evlist, pos) {
1330 if (pos->id_pos != evlist__id_pos(evlist) ||
1331 pos->is_pos != evlist__is_pos(evlist))
1332 return false;
1333 }
1334
1335 return true;
1336}
1337
1338u64 __evlist__combined_sample_type(struct evlist *evlist)
1339{
1340 struct evsel *evsel;
1341
1342 if (RC_CHK_ACCESS(evlist)->combined_sample_type)
1343 return RC_CHK_ACCESS(evlist)->combined_sample_type;
1344
1345 evlist__for_each_entry(evlist, evsel)
1346 RC_CHK_ACCESS(evlist)->combined_sample_type |= evsel->core.attr.sample_type;
1347
1348 return RC_CHK_ACCESS(evlist)->combined_sample_type;
1349}
1350
1351u64 evlist__combined_sample_type(struct evlist *evlist)
1352{
1353 RC_CHK_ACCESS(evlist)->combined_sample_type = 0;
1354 return __evlist__combined_sample_type(evlist);
1355}
1356
1357u64 evlist__combined_branch_type(struct evlist *evlist)
1358{
1359 struct evsel *evsel;
1360 u64 branch_type = 0;
1361
1362 evlist__for_each_entry(evlist, evsel)
1363 branch_type |= evsel->core.attr.branch_sample_type;
1364 return branch_type;
1365}
1366
1367static struct evsel *
1368evlist__find_dup_event_from_prev(struct evlist *evlist, struct evsel *event)
1369{
1370 struct evsel *pos;
1371
1372 evlist__for_each_entry(evlist, pos) {
1373 if (event == pos)
1374 break;
1375 if ((pos->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) &&
1376 !strcmp(pos->name, event->name))
1377 return pos;
1378 }
1379 return NULL;
1380}
1381
1382#define MAX_NR_ABBR_NAME (26 * 11)
1383
1384/*
1385 * The abbr name is from A to Z9. If the number of event
1386 * which requires the branch counter > MAX_NR_ABBR_NAME,
1387 * return NA.
1388 */
1389static void evlist__new_abbr_name(char *name)
1390{
1391 static int idx;
1392 int i = idx / 26;
1393
1394 if (idx >= MAX_NR_ABBR_NAME) {
1395 name[0] = 'N';
1396 name[1] = 'A';
1397 name[2] = '\0';
1398 return;
1399 }
1400
1401 name[0] = 'A' + (idx % 26);
1402
1403 if (!i)
1404 name[1] = '\0';
1405 else {
1406 name[1] = '0' + i - 1;
1407 name[2] = '\0';
1408 }
1409
1410 idx++;
1411}
1412
1413void evlist__update_br_cntr(struct evlist *evlist)
1414{
1415 struct evsel *evsel, *dup;
1416 int i = 0;
1417
1418 evlist__for_each_entry(evlist, evsel) {
1419 if (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) {
1420 evsel->br_cntr_idx = i++;
1421 evsel__leader(evsel)->br_cntr_nr++;
1422
1423 dup = evlist__find_dup_event_from_prev(evlist, evsel);
1424 if (dup)
1425 memcpy(evsel->abbr_name, dup->abbr_name, 3 * sizeof(char));
1426 else
1427 evlist__new_abbr_name(evsel->abbr_name);
1428 }
1429 }
1430 evlist__set_nr_br_cntr(evlist, i);
1431}
1432
1433bool evlist__valid_read_format(struct evlist *evlist)
1434{
1435 struct evsel *first = evlist__first(evlist), *pos = first;
1436 u64 read_format = first->core.attr.read_format;
1437 u64 sample_type = first->core.attr.sample_type;
1438
1439 evlist__for_each_entry(evlist, pos) {
1440 if (read_format != pos->core.attr.read_format) {
1441 pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1442 read_format, (u64)pos->core.attr.read_format);
1443 }
1444 }
1445
1446 /* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */
1447 if ((sample_type & PERF_SAMPLE_READ) &&
1448 !(read_format & PERF_FORMAT_ID)) {
1449 return false;
1450 }
1451
1452 return true;
1453}
1454
1455u16 evlist__id_hdr_size(struct evlist *evlist)
1456{
1457 struct evsel *first = evlist__first(evlist);
1458
1459 return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0;
1460}
1461
1462bool evlist__valid_sample_id_all(struct evlist *evlist)
1463{
1464 struct evsel *first = evlist__first(evlist), *pos = first;
1465
1466 evlist__for_each_entry_continue(evlist, pos) {
1467 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1468 return false;
1469 }
1470
1471 return true;
1472}
1473
1474bool evlist__sample_id_all(struct evlist *evlist)
1475{
1476 struct evsel *first = evlist__first(evlist);
1477 return first->core.attr.sample_id_all;
1478}
1479
1480void evlist__close(struct evlist *evlist)
1481{
1482 struct evsel *evsel;
1483 struct evlist_cpu_iterator evlist_cpu_itr;
1484
1485 evlist__for_each_cpu(evlist_cpu_itr, evlist) {
1486 if (evlist_cpu_itr.cpu_map_idx == 0 && evsel__is_retire_lat(evlist_cpu_itr.evsel))
1487 evsel__tpebs_close(evlist_cpu_itr.evsel);
1488 perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core,
1489 evlist_cpu_itr.cpu_map_idx);
1490 }
1491
1492 evlist__for_each_entry_reverse(evlist, evsel) {
1493 perf_evsel__free_fd(&evsel->core);
1494 perf_evsel__free_id(&evsel->core);
1495 }
1496 perf_evlist__reset_id_hash(evlist__core(evlist));
1497}
1498
1499static int evlist__create_syswide_maps(struct evlist *evlist)
1500{
1501 struct perf_cpu_map *cpus;
1502 struct perf_thread_map *threads;
1503
1504 /*
1505 * Try reading /sys/devices/system/cpu/online to get
1506 * an all cpus map.
1507 *
1508 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1509 * code needs an overhaul to properly forward the
1510 * error, and we may not want to do that fallback to a
1511 * default cpu identity map :-\
1512 */
1513 cpus = perf_cpu_map__new_online_cpus();
1514 if (!cpus)
1515 return -ENOMEM;
1516
1517 threads = perf_thread_map__new_dummy();
1518 if (!threads) {
1519 perf_cpu_map__put(cpus);
1520 return -ENOMEM;
1521 }
1522
1523 perf_evlist__set_maps(evlist__core(evlist), cpus, threads);
1524 perf_thread_map__put(threads);
1525 perf_cpu_map__put(cpus);
1526 return 0;
1527}
1528
1529int evlist__open(struct evlist *evlist)
1530{
1531 struct evsel *evsel;
1532 int err;
1533
1534 /*
1535 * Default: one fd per CPU, all threads, aka systemwide
1536 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1537 */
1538 if (evlist__core(evlist)->threads == NULL &&
1539 evlist__core(evlist)->user_requested_cpus == NULL) {
1540 err = evlist__create_syswide_maps(evlist);
1541 if (err < 0)
1542 goto out_err;
1543 }
1544
1545 evlist__update_id_pos(evlist);
1546
1547 evlist__for_each_entry(evlist, evsel) {
1548 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1549 if (err < 0)
1550 goto out_err;
1551 }
1552
1553 return 0;
1554out_err:
1555 evlist__close(evlist);
1556 errno = -err;
1557 return err;
1558}
1559
1560int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[],
1561 bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1562{
1563 int child_ready_pipe[2], go_pipe[2];
1564 char bf;
1565
1566 if (target__none(target)) {
1567 const char *cmd = strrchr(argv[0], '/');
1568
1569 if (cmd)
1570 cmd++;
1571 else
1572 cmd = argv[0];
1573
1574 if (!strcmp(cmd, "sleep")) {
1575 ui__warning("workload '%s' specified without the system-wide (-a) option\n",
1576 cmd);
1577 }
1578 }
1579
1580 evlist__set_workload_cork_fd(evlist, -1);
1581
1582 if (pipe(child_ready_pipe) < 0) {
1583 perror("failed to create 'ready' pipe");
1584 return -1;
1585 }
1586
1587 if (pipe(go_pipe) < 0) {
1588 perror("failed to create 'go' pipe");
1589 goto out_close_ready_pipe;
1590 }
1591
1592 evlist__set_workload_pid(evlist, fork());
1593 if (evlist__workload_pid(evlist) < 0) {
1594 perror("failed to fork");
1595 goto out_close_pipes;
1596 }
1597
1598 if (!evlist__workload_pid(evlist)) {
1599 int ret;
1600
1601 if (pipe_output)
1602 dup2(2, 1);
1603
1604 signal(SIGTERM, SIG_DFL);
1605
1606 close(child_ready_pipe[0]);
1607 close(go_pipe[1]);
1608 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1609
1610 /*
1611 * Change the name of this process not to confuse --exclude-perf users
1612 * that sees 'perf' in the window up to the execvp() and thinks that
1613 * perf samples are not being excluded.
1614 */
1615 prctl(PR_SET_NAME, "perf-exec");
1616
1617 /*
1618 * Tell the parent we're ready to go
1619 */
1620 close(child_ready_pipe[1]);
1621
1622 /*
1623 * Wait until the parent tells us to go.
1624 */
1625 ret = read(go_pipe[0], &bf, 1);
1626 /*
1627 * The parent will ask for the execvp() to be performed by
1628 * writing exactly one byte, in workload.cork_fd, usually via
1629 * evlist__start_workload().
1630 *
1631 * For cancelling the workload without actually running it,
1632 * the parent will just close workload.cork_fd, without writing
1633 * anything, i.e. read will return zero and we just exit()
1634 * here (See evlist__cancel_workload()).
1635 */
1636 if (ret != 1) {
1637 if (ret == -1)
1638 perror("unable to read pipe");
1639 exit(ret);
1640 }
1641
1642 execvp(argv[0], (char **)argv);
1643
1644 if (exec_error) {
1645 union sigval val;
1646
1647 val.sival_int = errno;
1648 if (sigqueue(getppid(), SIGUSR1, val))
1649 perror(argv[0]);
1650 } else
1651 perror(argv[0]);
1652 exit(-1);
1653 }
1654
1655 if (exec_error) {
1656 struct sigaction act = {
1657 .sa_flags = SA_SIGINFO,
1658 .sa_sigaction = exec_error,
1659 };
1660 sigaction(SIGUSR1, &act, NULL);
1661 }
1662
1663 if (target__none(target)) {
1664 if (evlist__core(evlist)->threads == NULL) {
1665 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1666 __func__, __LINE__);
1667 goto out_close_pipes;
1668 }
1669 perf_thread_map__set_pid(evlist__core(evlist)->threads, 0,
1670 evlist__workload_pid(evlist));
1671 }
1672
1673 close(child_ready_pipe[1]);
1674 close(go_pipe[0]);
1675 /*
1676 * wait for child to settle
1677 */
1678 if (read(child_ready_pipe[0], &bf, 1) == -1) {
1679 perror("unable to read pipe");
1680 goto out_close_pipes;
1681 }
1682
1683 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1684 evlist__set_workload_cork_fd(evlist, go_pipe[1]);
1685 close(child_ready_pipe[0]);
1686 return 0;
1687
1688out_close_pipes:
1689 close(go_pipe[0]);
1690 close(go_pipe[1]);
1691out_close_ready_pipe:
1692 close(child_ready_pipe[0]);
1693 close(child_ready_pipe[1]);
1694 return -1;
1695}
1696
1697int evlist__start_workload(struct evlist *evlist)
1698{
1699 if (evlist__workload_cork_fd(evlist) >= 0) {
1700 char bf = 0;
1701 int ret;
1702 /*
1703 * Remove the cork, let it rip!
1704 */
1705 ret = write(evlist__workload_cork_fd(evlist), &bf, 1);
1706 if (ret < 0)
1707 perror("unable to write to pipe");
1708
1709 close(evlist__workload_cork_fd(evlist));
1710 evlist__set_workload_cork_fd(evlist, -1);
1711 return ret;
1712 }
1713
1714 return 0;
1715}
1716
1717void evlist__cancel_workload(struct evlist *evlist)
1718{
1719 int status;
1720
1721 if (evlist__workload_cork_fd(evlist) >= 0) {
1722 close(evlist__workload_cork_fd(evlist));
1723 evlist__set_workload_cork_fd(evlist, -1);
1724 waitpid(evlist__workload_pid(evlist), &status, WNOHANG);
1725 }
1726}
1727
1728int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1729{
1730 struct evsel *evsel = evlist__event2evsel(evlist, event);
1731 int ret;
1732
1733 if (!evsel) {
1734 /* Ensure the sample is okay for perf_sample__exit. */
1735 perf_sample__init(sample, /*all=*/false);
1736 return -EFAULT;
1737 }
1738 ret = evsel__parse_sample(evsel, event, sample);
1739 if (ret)
1740 return ret;
1741 if (perf_guest && sample->id) {
1742 struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id);
1743
1744 if (sid) {
1745 sample->machine_pid = sid->machine_pid;
1746 sample->vcpu = sid->vcpu.cpu;
1747 }
1748 }
1749 return 0;
1750}
1751
1752int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp)
1753{
1754 struct evsel *evsel = evlist__event2evsel(evlist, event);
1755
1756 if (!evsel)
1757 return -EFAULT;
1758 return evsel__parse_sample_timestamp(evsel, event, timestamp);
1759}
1760
1761int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1762{
1763 int printed, value;
1764
1765 switch (err) {
1766 case EACCES:
1767 case EPERM:
1768 errno = err;
1769 printed = scnprintf(buf, size,
1770 "Error:\t%m.\n"
1771 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.");
1772
1773 value = perf_event_paranoid();
1774
1775 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1776
1777 if (value >= 2) {
1778 printed += scnprintf(buf + printed, size - printed,
1779 "For your workloads it needs to be <= 1\nHint:\t");
1780 }
1781 printed += scnprintf(buf + printed, size - printed,
1782 "For system wide tracing it needs to be set to -1.\n");
1783
1784 printed += scnprintf(buf + printed, size - printed,
1785 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1786 "Hint:\tThe current value is %d.", value);
1787 break;
1788 case EINVAL: {
1789 struct evsel *first = evlist__first(evlist);
1790 int max_freq;
1791
1792 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1793 goto out_default;
1794
1795 if (first->core.attr.sample_freq < (u64)max_freq)
1796 goto out_default;
1797
1798 errno = err;
1799 printed = scnprintf(buf, size,
1800 "Error:\t%m.\n"
1801 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1802 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1803 max_freq, first->core.attr.sample_freq);
1804 break;
1805 }
1806 default:
1807out_default:
1808 errno = err;
1809 scnprintf(buf, size, "%m");
1810 break;
1811 }
1812
1813 return 0;
1814}
1815
1816int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1817{
1818 int pages_attempted = evlist__core(evlist)->mmap_len / 1024;
1819 int pages_max_per_user, printed = 0;
1820
1821 switch (err) {
1822 case EPERM:
1823 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1824 errno = err;
1825 printed += scnprintf(buf + printed, size - printed,
1826 "Error:\t%m.\n"
1827 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1828 "Hint:\tTried using %zd kB.\n",
1829 pages_max_per_user, pages_attempted);
1830
1831 if (pages_attempted >= pages_max_per_user) {
1832 printed += scnprintf(buf + printed, size - printed,
1833 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1834 pages_max_per_user + pages_attempted);
1835 }
1836
1837 printed += scnprintf(buf + printed, size - printed,
1838 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1839 break;
1840 default:
1841 errno = err;
1842 scnprintf(buf, size, "%m");
1843 break;
1844 }
1845
1846 return 0;
1847}
1848
1849void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel)
1850{
1851 struct evsel *evsel, *n;
1852 LIST_HEAD(move);
1853
1854 if (move_evsel == evlist__first(evlist))
1855 return;
1856
1857 evlist__for_each_entry_safe(evlist, n, evsel) {
1858 if (evsel__leader(evsel) == evsel__leader(move_evsel))
1859 list_move_tail(&evsel->core.node, &move);
1860 }
1861
1862 list_splice(&move, &evlist__core(evlist)->entries);
1863}
1864
1865struct evsel *evlist__get_tracking_event(struct evlist *evlist)
1866{
1867 struct evsel *evsel;
1868
1869 evlist__for_each_entry(evlist, evsel) {
1870 if (evsel->tracking)
1871 return evsel;
1872 }
1873
1874 return evlist__first(evlist);
1875}
1876
1877void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel)
1878{
1879 struct evsel *evsel;
1880
1881 if (tracking_evsel->tracking)
1882 return;
1883
1884 evlist__for_each_entry(evlist, evsel) {
1885 if (evsel != tracking_evsel)
1886 evsel->tracking = false;
1887 }
1888
1889 tracking_evsel->tracking = true;
1890}
1891
1892struct evsel *evlist__findnew_tracking_event(struct evlist *evlist, bool system_wide)
1893{
1894 struct evsel *evsel;
1895
1896 evsel = evlist__get_tracking_event(evlist);
1897 if (!evsel__is_dummy_event(evsel)) {
1898 evsel = evlist__add_aux_dummy(evlist, system_wide);
1899 if (!evsel)
1900 return NULL;
1901
1902 evlist__set_tracking_event(evlist, evsel);
1903 } else if (system_wide) {
1904 perf_evlist__go_system_wide(evlist__core(evlist), &evsel->core);
1905 }
1906
1907 return evsel;
1908}
1909
1910struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str)
1911{
1912 struct evsel *evsel;
1913
1914 evlist__for_each_entry(evlist, evsel) {
1915 if (!evsel->name)
1916 continue;
1917 if (evsel__name_is(evsel, str))
1918 return evsel;
1919 }
1920
1921 return NULL;
1922}
1923
1924void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state)
1925{
1926 enum bkw_mmap_state old_state = evlist__bkw_mmap_state(evlist);
1927 enum action {
1928 NONE,
1929 PAUSE,
1930 RESUME,
1931 } action = NONE;
1932
1933 if (!evlist__overwrite_mmap(evlist))
1934 return;
1935
1936 switch (old_state) {
1937 case BKW_MMAP_NOTREADY: {
1938 if (state != BKW_MMAP_RUNNING)
1939 goto state_err;
1940 break;
1941 }
1942 case BKW_MMAP_RUNNING: {
1943 if (state != BKW_MMAP_DATA_PENDING)
1944 goto state_err;
1945 action = PAUSE;
1946 break;
1947 }
1948 case BKW_MMAP_DATA_PENDING: {
1949 if (state != BKW_MMAP_EMPTY)
1950 goto state_err;
1951 break;
1952 }
1953 case BKW_MMAP_EMPTY: {
1954 if (state != BKW_MMAP_RUNNING)
1955 goto state_err;
1956 action = RESUME;
1957 break;
1958 }
1959 default:
1960 WARN_ONCE(1, "Shouldn't get there\n");
1961 }
1962
1963 evlist__set_bkw_mmap_state(evlist, state);
1964
1965 switch (action) {
1966 case PAUSE:
1967 evlist__pause(evlist);
1968 break;
1969 case RESUME:
1970 evlist__resume(evlist);
1971 break;
1972 case NONE:
1973 default:
1974 break;
1975 }
1976
1977state_err:
1978 return;
1979}
1980
1981bool evlist__exclude_kernel(struct evlist *evlist)
1982{
1983 struct evsel *evsel;
1984
1985 evlist__for_each_entry(evlist, evsel) {
1986 if (!evsel->core.attr.exclude_kernel)
1987 return false;
1988 }
1989
1990 return true;
1991}
1992
1993/*
1994 * Events in data file are not collect in groups, but we still want
1995 * the group display. Set the artificial group and set the leader's
1996 * forced_leader flag to notify the display code.
1997 */
1998void evlist__force_leader(struct evlist *evlist)
1999{
2000 if (evlist__nr_groups(evlist) == 0) {
2001 struct evsel *leader = evlist__first(evlist);
2002
2003 evlist__set_leader(evlist);
2004 leader->forced_leader = true;
2005 }
2006}
2007
2008struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close)
2009{
2010 struct evsel *c2, *leader;
2011 bool is_open = true;
2012
2013 leader = evsel__leader(evsel);
2014
2015 pr_debug("Weak group for %s/%d failed\n",
2016 leader->name, leader->core.nr_members);
2017
2018 /*
2019 * for_each_group_member doesn't work here because it doesn't
2020 * include the first entry.
2021 */
2022 evlist__for_each_entry(evsel_list, c2) {
2023 if (c2 == evsel)
2024 is_open = false;
2025 if (evsel__has_leader(c2, leader)) {
2026 if (is_open && close)
2027 perf_evsel__close(&c2->core);
2028 /*
2029 * We want to close all members of the group and reopen
2030 * them. Some events, like Intel topdown, require being
2031 * in a group and so keep these in the group.
2032 */
2033 evsel__remove_from_group(c2, leader);
2034
2035 /*
2036 * Set this for all former members of the group
2037 * to indicate they get reopened.
2038 */
2039 c2->reset_group = true;
2040 }
2041 }
2042 /* Reset the leader count if all entries were removed. */
2043 if (leader->core.nr_members == 1)
2044 leader->core.nr_members = 0;
2045 return leader;
2046}
2047
2048static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
2049{
2050 char *s, *p;
2051 int ret = 0, fd;
2052
2053 if (strncmp(str, "fifo:", 5))
2054 return -EINVAL;
2055
2056 str += 5;
2057 if (!*str || *str == ',')
2058 return -EINVAL;
2059
2060 s = strdup(str);
2061 if (!s)
2062 return -ENOMEM;
2063
2064 p = strchr(s, ',');
2065 if (p)
2066 *p = '\0';
2067
2068 /*
2069 * O_RDWR avoids POLLHUPs which is necessary to allow the other
2070 * end of a FIFO to be repeatedly opened and closed.
2071 */
2072 fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
2073 if (fd < 0) {
2074 ret = -errno;
2075 pr_err("Failed to open '%s': %m\n", s);
2076 goto out_free;
2077 }
2078 *ctl_fd = fd;
2079 *ctl_fd_close = true;
2080
2081 if (p && *++p) {
2082 /* O_RDWR | O_NONBLOCK means the other end need not be open */
2083 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
2084 if (fd < 0) {
2085 pr_err("Failed to open '%s': %m\n", p);
2086 ret = -errno;
2087 goto out_free;
2088 }
2089 *ctl_fd_ack = fd;
2090 }
2091
2092out_free:
2093 free(s);
2094 return ret;
2095}
2096
2097int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
2098{
2099 const char *comma = NULL;
2100 char *endptr = NULL;
2101
2102 *ctl_fd_close = false;
2103
2104 if (strncmp(str, "fd:", 3))
2105 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
2106
2107 *ctl_fd = strtoul(&str[3], &endptr, 0);
2108 if (endptr == &str[3])
2109 return -EINVAL;
2110
2111 comma = strchr(str, ',');
2112 if (comma) {
2113 if (endptr != comma)
2114 return -EINVAL;
2115
2116 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
2117 if (endptr == comma + 1 || *endptr != '\0')
2118 return -EINVAL;
2119 }
2120
2121 return 0;
2122}
2123
2124void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
2125{
2126 if (*ctl_fd_close) {
2127 *ctl_fd_close = false;
2128 close(ctl_fd);
2129 if (ctl_fd_ack >= 0)
2130 close(ctl_fd_ack);
2131 }
2132}
2133
2134int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
2135{
2136 if (fd == -1) {
2137 pr_debug("Control descriptor is not initialized\n");
2138 return 0;
2139 }
2140
2141 evlist__set_ctl_fd_pos(evlist,
2142 perf_evlist__add_pollfd(evlist__core(evlist), fd, NULL, POLLIN,
2143 fdarray_flag__nonfilterable |
2144 fdarray_flag__non_perf_event));
2145 if (evlist__ctl_fd_pos(evlist) < 0) {
2146 evlist__set_ctl_fd_pos(evlist, -1);
2147 pr_err("Failed to add ctl fd entry: %m\n");
2148 return -1;
2149 }
2150
2151 evlist__set_ctl_fd_fd(evlist, fd);
2152 evlist__set_ctl_fd_ack(evlist, ack);
2153
2154 return 0;
2155}
2156
2157bool evlist__ctlfd_initialized(struct evlist *evlist)
2158{
2159 return evlist__ctl_fd_pos(evlist) >= 0;
2160}
2161
2162int evlist__finalize_ctlfd(struct evlist *evlist)
2163{
2164 struct pollfd *entries = evlist__core(evlist)->pollfd.entries;
2165
2166 if (!evlist__ctlfd_initialized(evlist))
2167 return 0;
2168
2169 entries[evlist__ctl_fd_pos(evlist)].fd = -1;
2170 entries[evlist__ctl_fd_pos(evlist)].events = 0;
2171 entries[evlist__ctl_fd_pos(evlist)].revents = 0;
2172
2173 evlist__set_ctl_fd_pos(evlist, -1);
2174 evlist__set_ctl_fd_ack(evlist, -1);
2175 evlist__set_ctl_fd_fd(evlist, -1);
2176
2177 return 0;
2178}
2179
2180static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
2181 char *cmd_data, size_t data_size)
2182{
2183 int err;
2184 char c;
2185 size_t bytes_read = 0;
2186
2187 *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
2188 memset(cmd_data, 0, data_size);
2189 data_size--;
2190
2191 do {
2192 err = read(evlist__ctl_fd_fd(evlist), &c, 1);
2193 if (err > 0) {
2194 if (c == '\n' || c == '\0')
2195 break;
2196 cmd_data[bytes_read++] = c;
2197 if (bytes_read == data_size)
2198 break;
2199 continue;
2200 } else if (err == -1) {
2201 if (errno == EINTR)
2202 continue;
2203 if (errno == EAGAIN || errno == EWOULDBLOCK)
2204 err = 0;
2205 else
2206 pr_err("Failed to read from ctlfd %d: %m\n",
2207 evlist__ctl_fd_fd(evlist));
2208 }
2209 break;
2210 } while (1);
2211
2212 pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
2213 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
2214
2215 if (bytes_read > 0) {
2216 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
2217 (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
2218 *cmd = EVLIST_CTL_CMD_ENABLE;
2219 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
2220 (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
2221 *cmd = EVLIST_CTL_CMD_DISABLE;
2222 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
2223 (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
2224 *cmd = EVLIST_CTL_CMD_SNAPSHOT;
2225 pr_debug("is snapshot\n");
2226 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG,
2227 (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) {
2228 *cmd = EVLIST_CTL_CMD_EVLIST;
2229 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG,
2230 (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) {
2231 *cmd = EVLIST_CTL_CMD_STOP;
2232 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG,
2233 (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) {
2234 *cmd = EVLIST_CTL_CMD_PING;
2235 }
2236 }
2237
2238 return bytes_read ? (int)bytes_read : err;
2239}
2240
2241int evlist__ctlfd_ack(struct evlist *evlist)
2242{
2243 int err;
2244
2245 if (evlist__ctl_fd_ack(evlist) == -1)
2246 return 0;
2247
2248 err = write(evlist__ctl_fd_ack(evlist), EVLIST_CTL_CMD_ACK_TAG,
2249 sizeof(EVLIST_CTL_CMD_ACK_TAG));
2250 if (err == -1)
2251 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist__ctl_fd_ack(evlist));
2252
2253 return err;
2254}
2255
2256static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg)
2257{
2258 char *data = cmd_data + cmd_size;
2259
2260 /* no argument */
2261 if (!*data)
2262 return 0;
2263
2264 /* there's argument */
2265 if (*data == ' ') {
2266 *arg = data + 1;
2267 return 1;
2268 }
2269
2270 /* malformed */
2271 return -1;
2272}
2273
2274static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable)
2275{
2276 struct evsel *evsel;
2277 char *name;
2278 int err;
2279
2280 err = get_cmd_arg(cmd_data,
2281 enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 :
2282 sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1,
2283 &name);
2284 if (err < 0) {
2285 pr_info("failed: wrong command\n");
2286 return -1;
2287 }
2288
2289 if (err) {
2290 evsel = evlist__find_evsel_by_str(evlist, name);
2291 if (evsel) {
2292 if (enable)
2293 evlist__enable_evsel(evlist, name);
2294 else
2295 evlist__disable_evsel(evlist, name);
2296 pr_info("Event %s %s\n", evsel->name,
2297 enable ? "enabled" : "disabled");
2298 } else {
2299 pr_info("failed: can't find '%s' event\n", name);
2300 }
2301 } else {
2302 if (enable) {
2303 evlist__enable(evlist);
2304 pr_info(EVLIST_ENABLED_MSG);
2305 } else {
2306 evlist__disable(evlist);
2307 pr_info(EVLIST_DISABLED_MSG);
2308 }
2309 }
2310
2311 return 0;
2312}
2313
2314static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data)
2315{
2316 struct perf_attr_details details = { .verbose = false, };
2317 struct evsel *evsel;
2318 char *arg;
2319 int err;
2320
2321 err = get_cmd_arg(cmd_data,
2322 sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1,
2323 &arg);
2324 if (err < 0) {
2325 pr_info("failed: wrong command\n");
2326 return -1;
2327 }
2328
2329 if (err) {
2330 if (!strcmp(arg, "-v")) {
2331 details.verbose = true;
2332 } else if (!strcmp(arg, "-g")) {
2333 details.event_group = true;
2334 } else if (!strcmp(arg, "-F")) {
2335 details.freq = true;
2336 } else {
2337 pr_info("failed: wrong command\n");
2338 return -1;
2339 }
2340 }
2341
2342 evlist__for_each_entry(evlist, evsel)
2343 evsel__fprintf(evsel, &details, stderr);
2344
2345 return 0;
2346}
2347
2348int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
2349{
2350 int err = 0;
2351 char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
2352 int ctlfd_pos = evlist__ctl_fd_pos(evlist);
2353 struct pollfd *entries = evlist__core(evlist)->pollfd.entries;
2354
2355 if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
2356 return 0;
2357
2358 if (entries[ctlfd_pos].revents & POLLIN) {
2359 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
2360 EVLIST_CTL_CMD_MAX_LEN);
2361 if (err > 0) {
2362 switch (*cmd) {
2363 case EVLIST_CTL_CMD_ENABLE:
2364 case EVLIST_CTL_CMD_DISABLE:
2365 err = evlist__ctlfd_enable(evlist, cmd_data,
2366 *cmd == EVLIST_CTL_CMD_ENABLE);
2367 break;
2368 case EVLIST_CTL_CMD_EVLIST:
2369 err = evlist__ctlfd_list(evlist, cmd_data);
2370 break;
2371 case EVLIST_CTL_CMD_SNAPSHOT:
2372 case EVLIST_CTL_CMD_STOP:
2373 case EVLIST_CTL_CMD_PING:
2374 break;
2375 case EVLIST_CTL_CMD_ACK:
2376 case EVLIST_CTL_CMD_UNSUPPORTED:
2377 default:
2378 pr_debug("ctlfd: unsupported %d\n", *cmd);
2379 break;
2380 }
2381 if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
2382 *cmd == EVLIST_CTL_CMD_SNAPSHOT))
2383 evlist__ctlfd_ack(evlist);
2384 }
2385 }
2386
2387 if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
2388 evlist__finalize_ctlfd(evlist);
2389 else
2390 entries[ctlfd_pos].revents = 0;
2391
2392 return err;
2393}
2394
2395/**
2396 * struct event_enable_time - perf record -D/--delay single time range.
2397 * @start: start of time range to enable events in milliseconds
2398 * @end: end of time range to enable events in milliseconds
2399 *
2400 * N.B. this structure is also accessed as an array of int.
2401 */
2402struct event_enable_time {
2403 int start;
2404 int end;
2405};
2406
2407static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first)
2408{
2409 const char *fmt = first ? "%u - %u %n" : " , %u - %u %n";
2410 int ret, start, end, n;
2411
2412 ret = sscanf(str, fmt, &start, &end, &n);
2413 if (ret != 2 || end <= start)
2414 return -EINVAL;
2415 if (range) {
2416 range->start = start;
2417 range->end = end;
2418 }
2419 return n;
2420}
2421
2422static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range)
2423{
2424 int incr = !!range;
2425 bool first = true;
2426 ssize_t ret, cnt;
2427
2428 for (cnt = 0; *str; cnt++) {
2429 ret = parse_event_enable_time(str, range, first);
2430 if (ret < 0)
2431 return ret;
2432 /* Check no overlap */
2433 if (!first && range && range->start <= range[-1].end)
2434 return -EINVAL;
2435 str += ret;
2436 range += incr;
2437 first = false;
2438 }
2439 return cnt;
2440}
2441
2442/**
2443 * struct event_enable_timer - control structure for perf record -D/--delay.
2444 * @evlist: event list
2445 * @times: time ranges that events are enabled (N.B. this is also accessed as an
2446 * array of int)
2447 * @times_cnt: number of time ranges
2448 * @timerfd: timer file descriptor
2449 * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray)
2450 * @times_step: current position in (int *)@times)[],
2451 * refer event_enable_timer__process()
2452 *
2453 * Note, this structure is only used when there are time ranges, not when there
2454 * is only an initial delay.
2455 */
2456struct event_enable_timer {
2457 struct evlist *evlist;
2458 struct event_enable_time *times;
2459 size_t times_cnt;
2460 int timerfd;
2461 int pollfd_pos;
2462 size_t times_step;
2463};
2464
2465static int str_to_delay(const char *str)
2466{
2467 char *endptr;
2468 long d;
2469
2470 d = strtol(str, &endptr, 10);
2471 if (*endptr || d > INT_MAX || d < -1)
2472 return 0;
2473 return d;
2474}
2475
2476int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts,
2477 const char *str, int unset)
2478{
2479 enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event;
2480 struct event_enable_timer *eet;
2481 ssize_t times_cnt;
2482 ssize_t ret;
2483 int err;
2484
2485 if (unset)
2486 return 0;
2487
2488 opts->target.initial_delay = str_to_delay(str);
2489 if (opts->target.initial_delay)
2490 return 0;
2491
2492 ret = parse_event_enable_times(str, NULL);
2493 if (ret < 0)
2494 return ret;
2495
2496 times_cnt = ret;
2497 if (times_cnt == 0)
2498 return -EINVAL;
2499
2500 eet = zalloc(sizeof(*eet));
2501 if (!eet)
2502 return -ENOMEM;
2503
2504 eet->times = calloc(times_cnt, sizeof(*eet->times));
2505 if (!eet->times) {
2506 err = -ENOMEM;
2507 goto free_eet;
2508 }
2509
2510 if (parse_event_enable_times(str, eet->times) != times_cnt) {
2511 err = -EINVAL;
2512 goto free_eet_times;
2513 }
2514
2515 eet->times_cnt = times_cnt;
2516
2517 eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC);
2518 if (eet->timerfd == -1) {
2519 err = -errno;
2520 pr_err("timerfd_create failed: %m\n");
2521 goto free_eet_times;
2522 }
2523
2524 eet->pollfd_pos = perf_evlist__add_pollfd(evlist__core(evlist), eet->timerfd,
2525 NULL, POLLIN, flags);
2526 if (eet->pollfd_pos < 0) {
2527 err = eet->pollfd_pos;
2528 goto close_timerfd;
2529 }
2530
2531 eet->evlist = evlist;
2532 RC_CHK_ACCESS(evlist)->eet = eet;
2533 opts->target.initial_delay = eet->times[0].start;
2534
2535 return 0;
2536
2537close_timerfd:
2538 close(eet->timerfd);
2539free_eet_times:
2540 zfree(&eet->times);
2541free_eet:
2542 free(eet);
2543 return err;
2544}
2545
2546static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms)
2547{
2548 struct itimerspec its = {
2549 .it_value.tv_sec = ms / MSEC_PER_SEC,
2550 .it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC,
2551 };
2552 int err = 0;
2553
2554 if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) {
2555 err = -errno;
2556 pr_err("timerfd_settime failed: %m\n");
2557 }
2558 return err;
2559}
2560
2561int event_enable_timer__start(struct event_enable_timer *eet)
2562{
2563 int ms;
2564
2565 if (!eet)
2566 return 0;
2567
2568 ms = eet->times[0].end - eet->times[0].start;
2569 eet->times_step = 1;
2570
2571 return event_enable_timer__set_timer(eet, ms);
2572}
2573
2574int event_enable_timer__process(struct event_enable_timer *eet)
2575{
2576 struct pollfd *entries;
2577 short revents;
2578
2579 if (!eet)
2580 return 0;
2581
2582 entries = evlist__core(eet->evlist)->pollfd.entries;
2583 revents = entries[eet->pollfd_pos].revents;
2584 entries[eet->pollfd_pos].revents = 0;
2585
2586 if (revents & POLLIN) {
2587 size_t step = eet->times_step;
2588 size_t pos = step / 2;
2589
2590 if (step & 1) {
2591 evlist__disable_non_dummy(eet->evlist);
2592 pr_info(EVLIST_DISABLED_MSG);
2593 if (pos >= eet->times_cnt - 1) {
2594 /* Disarm timer */
2595 event_enable_timer__set_timer(eet, 0);
2596 return 1; /* Stop */
2597 }
2598 } else {
2599 evlist__enable_non_dummy(eet->evlist);
2600 pr_info(EVLIST_ENABLED_MSG);
2601 }
2602
2603 step += 1;
2604 pos = step / 2;
2605
2606 if (pos < eet->times_cnt) {
2607 int *times = (int *)eet->times; /* Accessing 'times' as array of int */
2608 int ms = times[step] - times[step - 1];
2609
2610 eet->times_step = step;
2611 return event_enable_timer__set_timer(eet, ms);
2612 }
2613 }
2614
2615 return 0;
2616}
2617
2618static void event_enable_timer__exit(struct event_enable_timer **ep)
2619{
2620 if (!ep || !*ep)
2621 return;
2622 zfree(&(*ep)->times);
2623 zfree(ep);
2624}
2625
2626struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
2627{
2628 struct evsel *evsel;
2629
2630 evlist__for_each_entry(evlist, evsel) {
2631 if (evsel->core.idx == idx)
2632 return evsel;
2633 }
2634 return NULL;
2635}
2636
2637void evlist__format_evsels(struct evlist *evlist, struct strbuf *sb, size_t max_length)
2638{
2639 struct evsel *evsel, *leader = NULL;
2640 bool first = true;
2641
2642 evlist__for_each_entry(evlist, evsel) {
2643 struct evsel *new_leader = evsel__leader(evsel);
2644
2645 if (evsel__is_dummy_event(evsel))
2646 continue;
2647
2648 if (leader != new_leader && leader && leader->core.nr_members > 1)
2649 strbuf_addch(sb, '}');
2650
2651 if (!first)
2652 strbuf_addch(sb, ',');
2653
2654 if (sb->len > max_length) {
2655 strbuf_addstr(sb, "...");
2656 return;
2657 }
2658 if (leader != new_leader && new_leader->core.nr_members > 1)
2659 strbuf_addch(sb, '{');
2660
2661 strbuf_addstr(sb, evsel__name(evsel));
2662 first = false;
2663 leader = new_leader;
2664 }
2665 if (leader && leader->core.nr_members > 1)
2666 strbuf_addch(sb, '}');
2667}
2668
2669void evlist__check_mem_load_aux(struct evlist *evlist)
2670{
2671 struct evsel *leader, *evsel, *pos;
2672
2673 /*
2674 * For some platforms, the 'mem-loads' event is required to use
2675 * together with 'mem-loads-aux' within a group and 'mem-loads-aux'
2676 * must be the group leader. Now we disable this group before reporting
2677 * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
2678 * any valid memory load information.
2679 */
2680 evlist__for_each_entry(evlist, evsel) {
2681 leader = evsel__leader(evsel);
2682 if (leader == evsel)
2683 continue;
2684
2685 if (leader->name && strstr(leader->name, "mem-loads-aux")) {
2686 for_each_group_evsel(pos, leader) {
2687 evsel__set_leader(pos, pos);
2688 pos->core.nr_members = 0;
2689 }
2690 }
2691 }
2692}
2693
2694/**
2695 * evlist__warn_user_requested_cpus() - Check each evsel against requested CPUs
2696 * and warn if the user CPU list is inapplicable for the event's PMU's
2697 * CPUs. Not core PMUs list a CPU in sysfs, but this may be overwritten by a
2698 * user requested CPU and so any online CPU is applicable. Core PMUs handle
2699 * events on the CPUs in their list and otherwise the event isn't supported.
2700 * @evlist: The list of events being checked.
2701 * @cpu_list: The user provided list of CPUs.
2702 */
2703void evlist__warn_user_requested_cpus(struct evlist *evlist, const char *cpu_list)
2704{
2705 struct perf_cpu_map *user_requested_cpus;
2706 struct evsel *pos;
2707
2708 if (!cpu_list)
2709 return;
2710
2711 user_requested_cpus = perf_cpu_map__new(cpu_list);
2712 if (!user_requested_cpus)
2713 return;
2714
2715 evlist__for_each_entry(evlist, pos) {
2716 evsel__warn_user_requested_cpus(pos, user_requested_cpus);
2717 }
2718 perf_cpu_map__put(user_requested_cpus);
2719}
2720
2721/* Should uniquify be disabled for the evlist? */
2722static bool evlist__disable_uniquify(struct evlist *evlist)
2723{
2724 struct evsel *counter;
2725 struct perf_pmu *last_pmu = NULL;
2726 bool first = true;
2727
2728 evlist__for_each_entry(evlist, counter) {
2729 /* If PMUs vary then uniquify can be useful. */
2730 if (!first && counter->pmu != last_pmu)
2731 return false;
2732 first = false;
2733 if (counter->pmu) {
2734 /* Allow uniquify for uncore PMUs. */
2735 if (!counter->pmu->is_core)
2736 return false;
2737 /* Keep hybrid event names uniquified for clarity. */
2738 if (perf_pmus__num_core_pmus() > 1)
2739 return false;
2740 }
2741 last_pmu = counter->pmu;
2742 }
2743 return true;
2744}
2745
2746static bool evlist__set_needs_uniquify(struct evlist *evlist, const struct perf_stat_config *config)
2747{
2748 struct evsel *counter;
2749 bool needs_uniquify = false;
2750
2751 if (evlist__disable_uniquify(evlist)) {
2752 evlist__for_each_entry(evlist, counter)
2753 counter->uniquified_name = true;
2754 return false;
2755 }
2756
2757 evlist__for_each_entry(evlist, counter) {
2758 if (evsel__set_needs_uniquify(counter, config))
2759 needs_uniquify = true;
2760 }
2761 return needs_uniquify;
2762}
2763
2764void evlist__uniquify_evsel_names(struct evlist *evlist, const struct perf_stat_config *config)
2765{
2766 if (evlist__set_needs_uniquify(evlist, config)) {
2767 struct evsel *pos;
2768
2769 evlist__for_each_entry(evlist, pos)
2770 evsel__uniquify_counter(pos);
2771 }
2772}
2773
2774bool evlist__has_bpf_output(struct evlist *evlist)
2775{
2776 struct evsel *evsel;
2777
2778 evlist__for_each_entry(evlist, evsel) {
2779 if (evsel__is_bpf_output(evsel))
2780 return true;
2781 }
2782
2783 return false;
2784}
2785
2786bool evlist__needs_bpf_sb_event(struct evlist *evlist)
2787{
2788 struct evsel *evsel;
2789
2790 evlist__for_each_entry(evlist, evsel) {
2791 if (evsel__is_dummy_event(evsel))
2792 continue;
2793 if (!evsel->core.attr.exclude_kernel)
2794 return true;
2795 }
2796
2797 return false;
2798}