xref: /kernel/linux/linux-5.10/tools/perf/util/evlist.c (revision 8c2ecf20)
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 "affinity.h"
22#include "../perf.h"
23#include "asm/bug.h"
24#include "bpf-event.h"
25#include "util/string2.h"
26#include "util/perf_api_probe.h"
27#include <signal.h>
28#include <unistd.h>
29#include <sched.h>
30#include <stdlib.h>
31
32#include "parse-events.h"
33#include <subcmd/parse-options.h>
34
35#include <fcntl.h>
36#include <sys/ioctl.h>
37#include <sys/mman.h>
38
39#include <linux/bitops.h>
40#include <linux/hash.h>
41#include <linux/log2.h>
42#include <linux/err.h>
43#include <linux/string.h>
44#include <linux/zalloc.h>
45#include <perf/evlist.h>
46#include <perf/evsel.h>
47#include <perf/cpumap.h>
48#include <perf/mmap.h>
49
50#include <internal/xyarray.h>
51
52#ifdef LACKS_SIGQUEUE_PROTOTYPE
53int sigqueue(pid_t pid, int sig, const union sigval value);
54#endif
55
56#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
57#define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
58
59void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
60		  struct perf_thread_map *threads)
61{
62	perf_evlist__init(&evlist->core);
63	perf_evlist__set_maps(&evlist->core, cpus, threads);
64	evlist->workload.pid = -1;
65	evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
66	evlist->ctl_fd.fd = -1;
67	evlist->ctl_fd.ack = -1;
68	evlist->ctl_fd.pos = -1;
69}
70
71struct evlist *evlist__new(void)
72{
73	struct evlist *evlist = zalloc(sizeof(*evlist));
74
75	if (evlist != NULL)
76		evlist__init(evlist, NULL, NULL);
77
78	return evlist;
79}
80
81struct evlist *perf_evlist__new_default(void)
82{
83	struct evlist *evlist = evlist__new();
84
85	if (evlist && evlist__add_default(evlist)) {
86		evlist__delete(evlist);
87		evlist = NULL;
88	}
89
90	return evlist;
91}
92
93struct evlist *perf_evlist__new_dummy(void)
94{
95	struct evlist *evlist = evlist__new();
96
97	if (evlist && evlist__add_dummy(evlist)) {
98		evlist__delete(evlist);
99		evlist = NULL;
100	}
101
102	return evlist;
103}
104
105/**
106 * perf_evlist__set_id_pos - set the positions of event ids.
107 * @evlist: selected event list
108 *
109 * Events with compatible sample types all have the same id_pos
110 * and is_pos.  For convenience, put a copy on evlist.
111 */
112void perf_evlist__set_id_pos(struct evlist *evlist)
113{
114	struct evsel *first = evlist__first(evlist);
115
116	evlist->id_pos = first->id_pos;
117	evlist->is_pos = first->is_pos;
118}
119
120static void perf_evlist__update_id_pos(struct evlist *evlist)
121{
122	struct evsel *evsel;
123
124	evlist__for_each_entry(evlist, evsel)
125		evsel__calc_id_pos(evsel);
126
127	perf_evlist__set_id_pos(evlist);
128}
129
130static void evlist__purge(struct evlist *evlist)
131{
132	struct evsel *pos, *n;
133
134	evlist__for_each_entry_safe(evlist, n, pos) {
135		list_del_init(&pos->core.node);
136		pos->evlist = NULL;
137		evsel__delete(pos);
138	}
139
140	evlist->core.nr_entries = 0;
141}
142
143void evlist__exit(struct evlist *evlist)
144{
145	zfree(&evlist->mmap);
146	zfree(&evlist->overwrite_mmap);
147	perf_evlist__exit(&evlist->core);
148}
149
150void evlist__delete(struct evlist *evlist)
151{
152	if (evlist == NULL)
153		return;
154
155	evlist__munmap(evlist);
156	evlist__close(evlist);
157	evlist__purge(evlist);
158	evlist__exit(evlist);
159	free(evlist);
160}
161
162void evlist__add(struct evlist *evlist, struct evsel *entry)
163{
164	entry->evlist = evlist;
165	entry->idx = evlist->core.nr_entries;
166	entry->tracking = !entry->idx;
167
168	perf_evlist__add(&evlist->core, &entry->core);
169
170	if (evlist->core.nr_entries == 1)
171		perf_evlist__set_id_pos(evlist);
172}
173
174void evlist__remove(struct evlist *evlist, struct evsel *evsel)
175{
176	evsel->evlist = NULL;
177	perf_evlist__remove(&evlist->core, &evsel->core);
178}
179
180void perf_evlist__splice_list_tail(struct evlist *evlist,
181				   struct list_head *list)
182{
183	struct evsel *evsel, *temp;
184
185	__evlist__for_each_entry_safe(list, temp, evsel) {
186		list_del_init(&evsel->core.node);
187		evlist__add(evlist, evsel);
188	}
189}
190
191int __evlist__set_tracepoints_handlers(struct evlist *evlist,
192				       const struct evsel_str_handler *assocs, size_t nr_assocs)
193{
194	struct evsel *evsel;
195	size_t i;
196	int err;
197
198	for (i = 0; i < nr_assocs; i++) {
199		// Adding a handler for an event not in this evlist, just ignore it.
200		evsel = perf_evlist__find_tracepoint_by_name(evlist, assocs[i].name);
201		if (evsel == NULL)
202			continue;
203
204		err = -EEXIST;
205		if (evsel->handler != NULL)
206			goto out;
207		evsel->handler = assocs[i].handler;
208	}
209
210	err = 0;
211out:
212	return err;
213}
214
215void __perf_evlist__set_leader(struct list_head *list)
216{
217	struct evsel *evsel, *leader;
218
219	leader = list_entry(list->next, struct evsel, core.node);
220	evsel = list_entry(list->prev, struct evsel, core.node);
221
222	leader->core.nr_members = evsel->idx - leader->idx + 1;
223
224	__evlist__for_each_entry(list, evsel) {
225		evsel->leader = leader;
226	}
227}
228
229void perf_evlist__set_leader(struct evlist *evlist)
230{
231	if (evlist->core.nr_entries) {
232		evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
233		__perf_evlist__set_leader(&evlist->core.entries);
234	}
235}
236
237int __evlist__add_default(struct evlist *evlist, bool precise)
238{
239	struct evsel *evsel = evsel__new_cycles(precise);
240
241	if (evsel == NULL)
242		return -ENOMEM;
243
244	evlist__add(evlist, evsel);
245	return 0;
246}
247
248int evlist__add_dummy(struct evlist *evlist)
249{
250	struct perf_event_attr attr = {
251		.type	= PERF_TYPE_SOFTWARE,
252		.config = PERF_COUNT_SW_DUMMY,
253		.size	= sizeof(attr), /* to capture ABI version */
254	};
255	struct evsel *evsel = evsel__new_idx(&attr, evlist->core.nr_entries);
256
257	if (evsel == NULL)
258		return -ENOMEM;
259
260	evlist__add(evlist, evsel);
261	return 0;
262}
263
264static int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
265{
266	struct evsel *evsel, *n;
267	LIST_HEAD(head);
268	size_t i;
269
270	for (i = 0; i < nr_attrs; i++) {
271		evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
272		if (evsel == NULL)
273			goto out_delete_partial_list;
274		list_add_tail(&evsel->core.node, &head);
275	}
276
277	perf_evlist__splice_list_tail(evlist, &head);
278
279	return 0;
280
281out_delete_partial_list:
282	__evlist__for_each_entry_safe(&head, n, evsel)
283		evsel__delete(evsel);
284	return -1;
285}
286
287int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
288{
289	size_t i;
290
291	for (i = 0; i < nr_attrs; i++)
292		event_attr_init(attrs + i);
293
294	return evlist__add_attrs(evlist, attrs, nr_attrs);
295}
296
297struct evsel *
298perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
299{
300	struct evsel *evsel;
301
302	evlist__for_each_entry(evlist, evsel) {
303		if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
304		    (int)evsel->core.attr.config == id)
305			return evsel;
306	}
307
308	return NULL;
309}
310
311struct evsel *
312perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
313				     const char *name)
314{
315	struct evsel *evsel;
316
317	evlist__for_each_entry(evlist, evsel) {
318		if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
319		    (strcmp(evsel->name, name) == 0))
320			return evsel;
321	}
322
323	return NULL;
324}
325
326int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
327{
328	struct evsel *evsel = evsel__newtp(sys, name);
329
330	if (IS_ERR(evsel))
331		return -1;
332
333	evsel->handler = handler;
334	evlist__add(evlist, evsel);
335	return 0;
336}
337
338static int perf_evlist__nr_threads(struct evlist *evlist,
339				   struct evsel *evsel)
340{
341	if (evsel->core.system_wide)
342		return 1;
343	else
344		return perf_thread_map__nr(evlist->core.threads);
345}
346
347void evlist__cpu_iter_start(struct evlist *evlist)
348{
349	struct evsel *pos;
350
351	/*
352	 * Reset the per evsel cpu_iter. This is needed because
353	 * each evsel's cpumap may have a different index space,
354	 * and some operations need the index to modify
355	 * the FD xyarray (e.g. open, close)
356	 */
357	evlist__for_each_entry(evlist, pos)
358		pos->cpu_iter = 0;
359}
360
361bool evsel__cpu_iter_skip_no_inc(struct evsel *ev, int cpu)
362{
363	if (ev->cpu_iter >= ev->core.cpus->nr)
364		return true;
365	if (cpu >= 0 && ev->core.cpus->map[ev->cpu_iter] != cpu)
366		return true;
367	return false;
368}
369
370bool evsel__cpu_iter_skip(struct evsel *ev, int cpu)
371{
372	if (!evsel__cpu_iter_skip_no_inc(ev, cpu)) {
373		ev->cpu_iter++;
374		return false;
375	}
376	return true;
377}
378
379void evlist__disable(struct evlist *evlist)
380{
381	struct evsel *pos;
382	struct affinity affinity;
383	int cpu, i, imm = 0;
384	bool has_imm = false;
385
386	if (affinity__setup(&affinity) < 0)
387		return;
388
389	/* Disable 'immediate' events last */
390	for (imm = 0; imm <= 1; imm++) {
391		evlist__for_each_cpu(evlist, i, cpu) {
392			affinity__set(&affinity, cpu);
393
394			evlist__for_each_entry(evlist, pos) {
395				if (evsel__cpu_iter_skip(pos, cpu))
396					continue;
397				if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
398					continue;
399				if (pos->immediate)
400					has_imm = true;
401				if (pos->immediate != imm)
402					continue;
403				evsel__disable_cpu(pos, pos->cpu_iter - 1);
404			}
405		}
406		if (!has_imm)
407			break;
408	}
409
410	affinity__cleanup(&affinity);
411	evlist__for_each_entry(evlist, pos) {
412		if (!evsel__is_group_leader(pos) || !pos->core.fd)
413			continue;
414		pos->disabled = true;
415	}
416
417	evlist->enabled = false;
418}
419
420void evlist__enable(struct evlist *evlist)
421{
422	struct evsel *pos;
423	struct affinity affinity;
424	int cpu, i;
425
426	if (affinity__setup(&affinity) < 0)
427		return;
428
429	evlist__for_each_cpu(evlist, i, cpu) {
430		affinity__set(&affinity, cpu);
431
432		evlist__for_each_entry(evlist, pos) {
433			if (evsel__cpu_iter_skip(pos, cpu))
434				continue;
435			if (!evsel__is_group_leader(pos) || !pos->core.fd)
436				continue;
437			evsel__enable_cpu(pos, pos->cpu_iter - 1);
438		}
439	}
440	affinity__cleanup(&affinity);
441	evlist__for_each_entry(evlist, pos) {
442		if (!evsel__is_group_leader(pos) || !pos->core.fd)
443			continue;
444		pos->disabled = false;
445	}
446
447	evlist->enabled = true;
448}
449
450void perf_evlist__toggle_enable(struct evlist *evlist)
451{
452	(evlist->enabled ? evlist__disable : evlist__enable)(evlist);
453}
454
455static int perf_evlist__enable_event_cpu(struct evlist *evlist,
456					 struct evsel *evsel, int cpu)
457{
458	int thread;
459	int nr_threads = perf_evlist__nr_threads(evlist, evsel);
460
461	if (!evsel->core.fd)
462		return -EINVAL;
463
464	for (thread = 0; thread < nr_threads; thread++) {
465		int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
466		if (err)
467			return err;
468	}
469	return 0;
470}
471
472static int perf_evlist__enable_event_thread(struct evlist *evlist,
473					    struct evsel *evsel,
474					    int thread)
475{
476	int cpu;
477	int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
478
479	if (!evsel->core.fd)
480		return -EINVAL;
481
482	for (cpu = 0; cpu < nr_cpus; cpu++) {
483		int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
484		if (err)
485			return err;
486	}
487	return 0;
488}
489
490int perf_evlist__enable_event_idx(struct evlist *evlist,
491				  struct evsel *evsel, int idx)
492{
493	bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
494
495	if (per_cpu_mmaps)
496		return perf_evlist__enable_event_cpu(evlist, evsel, idx);
497	else
498		return perf_evlist__enable_event_thread(evlist, evsel, idx);
499}
500
501int evlist__add_pollfd(struct evlist *evlist, int fd)
502{
503	return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
504}
505
506int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
507{
508	return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
509}
510
511#ifdef HAVE_EVENTFD_SUPPORT
512int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd)
513{
514	return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
515				       fdarray_flag__nonfilterable);
516}
517#endif
518
519int evlist__poll(struct evlist *evlist, int timeout)
520{
521	return perf_evlist__poll(&evlist->core, timeout);
522}
523
524struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
525{
526	struct hlist_head *head;
527	struct perf_sample_id *sid;
528	int hash;
529
530	hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
531	head = &evlist->core.heads[hash];
532
533	hlist_for_each_entry(sid, head, node)
534		if (sid->id == id)
535			return sid;
536
537	return NULL;
538}
539
540struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
541{
542	struct perf_sample_id *sid;
543
544	if (evlist->core.nr_entries == 1 || !id)
545		return evlist__first(evlist);
546
547	sid = perf_evlist__id2sid(evlist, id);
548	if (sid)
549		return container_of(sid->evsel, struct evsel, core);
550
551	if (!evlist__sample_id_all(evlist))
552		return evlist__first(evlist);
553
554	return NULL;
555}
556
557struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
558						u64 id)
559{
560	struct perf_sample_id *sid;
561
562	if (!id)
563		return NULL;
564
565	sid = perf_evlist__id2sid(evlist, id);
566	if (sid)
567		return container_of(sid->evsel, struct evsel, core);
568
569	return NULL;
570}
571
572static int perf_evlist__event2id(struct evlist *evlist,
573				 union perf_event *event, u64 *id)
574{
575	const __u64 *array = event->sample.array;
576	ssize_t n;
577
578	n = (event->header.size - sizeof(event->header)) >> 3;
579
580	if (event->header.type == PERF_RECORD_SAMPLE) {
581		if (evlist->id_pos >= n)
582			return -1;
583		*id = array[evlist->id_pos];
584	} else {
585		if (evlist->is_pos > n)
586			return -1;
587		n -= evlist->is_pos;
588		*id = array[n];
589	}
590	return 0;
591}
592
593struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
594					    union perf_event *event)
595{
596	struct evsel *first = evlist__first(evlist);
597	struct hlist_head *head;
598	struct perf_sample_id *sid;
599	int hash;
600	u64 id;
601
602	if (evlist->core.nr_entries == 1)
603		return first;
604
605	if (!first->core.attr.sample_id_all &&
606	    event->header.type != PERF_RECORD_SAMPLE)
607		return first;
608
609	if (perf_evlist__event2id(evlist, event, &id))
610		return NULL;
611
612	/* Synthesized events have an id of zero */
613	if (!id)
614		return first;
615
616	hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
617	head = &evlist->core.heads[hash];
618
619	hlist_for_each_entry(sid, head, node) {
620		if (sid->id == id)
621			return container_of(sid->evsel, struct evsel, core);
622	}
623	return NULL;
624}
625
626static int perf_evlist__set_paused(struct evlist *evlist, bool value)
627{
628	int i;
629
630	if (!evlist->overwrite_mmap)
631		return 0;
632
633	for (i = 0; i < evlist->core.nr_mmaps; i++) {
634		int fd = evlist->overwrite_mmap[i].core.fd;
635		int err;
636
637		if (fd < 0)
638			continue;
639		err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
640		if (err)
641			return err;
642	}
643	return 0;
644}
645
646static int perf_evlist__pause(struct evlist *evlist)
647{
648	return perf_evlist__set_paused(evlist, true);
649}
650
651static int perf_evlist__resume(struct evlist *evlist)
652{
653	return perf_evlist__set_paused(evlist, false);
654}
655
656static void evlist__munmap_nofree(struct evlist *evlist)
657{
658	int i;
659
660	if (evlist->mmap)
661		for (i = 0; i < evlist->core.nr_mmaps; i++)
662			perf_mmap__munmap(&evlist->mmap[i].core);
663
664	if (evlist->overwrite_mmap)
665		for (i = 0; i < evlist->core.nr_mmaps; i++)
666			perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
667}
668
669void evlist__munmap(struct evlist *evlist)
670{
671	evlist__munmap_nofree(evlist);
672	zfree(&evlist->mmap);
673	zfree(&evlist->overwrite_mmap);
674}
675
676static void perf_mmap__unmap_cb(struct perf_mmap *map)
677{
678	struct mmap *m = container_of(map, struct mmap, core);
679
680	mmap__munmap(m);
681}
682
683static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
684				       bool overwrite)
685{
686	int i;
687	struct mmap *map;
688
689	map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
690	if (!map)
691		return NULL;
692
693	for (i = 0; i < evlist->core.nr_mmaps; i++) {
694		struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
695
696		/*
697		 * When the perf_mmap() call is made we grab one refcount, plus
698		 * one extra to let perf_mmap__consume() get the last
699		 * events after all real references (perf_mmap__get()) are
700		 * dropped.
701		 *
702		 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
703		 * thus does perf_mmap__get() on it.
704		 */
705		perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
706	}
707
708	return map;
709}
710
711static void
712perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
713			 struct perf_mmap_param *_mp,
714			 int idx, bool per_cpu)
715{
716	struct evlist *evlist = container_of(_evlist, struct evlist, core);
717	struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
718
719	auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, idx, per_cpu);
720}
721
722static struct perf_mmap*
723perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
724{
725	struct evlist *evlist = container_of(_evlist, struct evlist, core);
726	struct mmap *maps;
727
728	maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
729
730	if (!maps) {
731		maps = evlist__alloc_mmap(evlist, overwrite);
732		if (!maps)
733			return NULL;
734
735		if (overwrite) {
736			evlist->overwrite_mmap = maps;
737			if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
738				perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
739		} else {
740			evlist->mmap = maps;
741		}
742	}
743
744	return &maps[idx].core;
745}
746
747static int
748perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
749			  int output, int cpu)
750{
751	struct mmap *map = container_of(_map, struct mmap, core);
752	struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
753
754	return mmap__mmap(map, mp, output, cpu);
755}
756
757unsigned long perf_event_mlock_kb_in_pages(void)
758{
759	unsigned long pages;
760	int max;
761
762	if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
763		/*
764		 * Pick a once upon a time good value, i.e. things look
765		 * strange since we can't read a sysctl value, but lets not
766		 * die yet...
767		 */
768		max = 512;
769	} else {
770		max -= (page_size / 1024);
771	}
772
773	pages = (max * 1024) / page_size;
774	if (!is_power_of_2(pages))
775		pages = rounddown_pow_of_two(pages);
776
777	return pages;
778}
779
780size_t evlist__mmap_size(unsigned long pages)
781{
782	if (pages == UINT_MAX)
783		pages = perf_event_mlock_kb_in_pages();
784	else if (!is_power_of_2(pages))
785		return 0;
786
787	return (pages + 1) * page_size;
788}
789
790static long parse_pages_arg(const char *str, unsigned long min,
791			    unsigned long max)
792{
793	unsigned long pages, val;
794	static struct parse_tag tags[] = {
795		{ .tag  = 'B', .mult = 1       },
796		{ .tag  = 'K', .mult = 1 << 10 },
797		{ .tag  = 'M', .mult = 1 << 20 },
798		{ .tag  = 'G', .mult = 1 << 30 },
799		{ .tag  = 0 },
800	};
801
802	if (str == NULL)
803		return -EINVAL;
804
805	val = parse_tag_value(str, tags);
806	if (val != (unsigned long) -1) {
807		/* we got file size value */
808		pages = PERF_ALIGN(val, page_size) / page_size;
809	} else {
810		/* we got pages count value */
811		char *eptr;
812		pages = strtoul(str, &eptr, 10);
813		if (*eptr != '\0')
814			return -EINVAL;
815	}
816
817	if (pages == 0 && min == 0) {
818		/* leave number of pages at 0 */
819	} else if (!is_power_of_2(pages)) {
820		char buf[100];
821
822		/* round pages up to next power of 2 */
823		pages = roundup_pow_of_two(pages);
824		if (!pages)
825			return -EINVAL;
826
827		unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
828		pr_info("rounding mmap pages size to %s (%lu pages)\n",
829			buf, pages);
830	}
831
832	if (pages > max)
833		return -EINVAL;
834
835	return pages;
836}
837
838int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
839{
840	unsigned long max = UINT_MAX;
841	long pages;
842
843	if (max > SIZE_MAX / page_size)
844		max = SIZE_MAX / page_size;
845
846	pages = parse_pages_arg(str, 1, max);
847	if (pages < 0) {
848		pr_err("Invalid argument for --mmap_pages/-m\n");
849		return -1;
850	}
851
852	*mmap_pages = pages;
853	return 0;
854}
855
856int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
857				  int unset __maybe_unused)
858{
859	return __perf_evlist__parse_mmap_pages(opt->value, str);
860}
861
862/**
863 * evlist__mmap_ex - Create mmaps to receive events.
864 * @evlist: list of events
865 * @pages: map length in pages
866 * @overwrite: overwrite older events?
867 * @auxtrace_pages - auxtrace map length in pages
868 * @auxtrace_overwrite - overwrite older auxtrace data?
869 *
870 * If @overwrite is %false the user needs to signal event consumption using
871 * perf_mmap__write_tail().  Using evlist__mmap_read() does this
872 * automatically.
873 *
874 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
875 * consumption using auxtrace_mmap__write_tail().
876 *
877 * Return: %0 on success, negative error code otherwise.
878 */
879int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
880			 unsigned int auxtrace_pages,
881			 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
882			 int comp_level)
883{
884	/*
885	 * Delay setting mp.prot: set it before calling perf_mmap__mmap.
886	 * Its value is decided by evsel's write_backward.
887	 * So &mp should not be passed through const pointer.
888	 */
889	struct mmap_params mp = {
890		.nr_cblocks	= nr_cblocks,
891		.affinity	= affinity,
892		.flush		= flush,
893		.comp_level	= comp_level
894	};
895	struct perf_evlist_mmap_ops ops = {
896		.idx  = perf_evlist__mmap_cb_idx,
897		.get  = perf_evlist__mmap_cb_get,
898		.mmap = perf_evlist__mmap_cb_mmap,
899	};
900
901	evlist->core.mmap_len = evlist__mmap_size(pages);
902	pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
903
904	auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
905				   auxtrace_pages, auxtrace_overwrite);
906
907	return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
908}
909
910int evlist__mmap(struct evlist *evlist, unsigned int pages)
911{
912	return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
913}
914
915int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
916{
917	bool all_threads = (target->per_thread && target->system_wide);
918	struct perf_cpu_map *cpus;
919	struct perf_thread_map *threads;
920
921	/*
922	 * If specify '-a' and '--per-thread' to perf record, perf record
923	 * will override '--per-thread'. target->per_thread = false and
924	 * target->system_wide = true.
925	 *
926	 * If specify '--per-thread' only to perf record,
927	 * target->per_thread = true and target->system_wide = false.
928	 *
929	 * So target->per_thread && target->system_wide is false.
930	 * For perf record, thread_map__new_str doesn't call
931	 * thread_map__new_all_cpus. That will keep perf record's
932	 * current behavior.
933	 *
934	 * For perf stat, it allows the case that target->per_thread and
935	 * target->system_wide are all true. It means to collect system-wide
936	 * per-thread data. thread_map__new_str will call
937	 * thread_map__new_all_cpus to enumerate all threads.
938	 */
939	threads = thread_map__new_str(target->pid, target->tid, target->uid,
940				      all_threads);
941
942	if (!threads)
943		return -1;
944
945	if (target__uses_dummy_map(target))
946		cpus = perf_cpu_map__dummy_new();
947	else
948		cpus = perf_cpu_map__new(target->cpu_list);
949
950	if (!cpus)
951		goto out_delete_threads;
952
953	evlist->core.has_user_cpus = !!target->cpu_list;
954
955	perf_evlist__set_maps(&evlist->core, cpus, threads);
956
957	/* as evlist now has references, put count here */
958	perf_cpu_map__put(cpus);
959	perf_thread_map__put(threads);
960
961	return 0;
962
963out_delete_threads:
964	perf_thread_map__put(threads);
965	return -1;
966}
967
968void __perf_evlist__set_sample_bit(struct evlist *evlist,
969				   enum perf_event_sample_format bit)
970{
971	struct evsel *evsel;
972
973	evlist__for_each_entry(evlist, evsel)
974		__evsel__set_sample_bit(evsel, bit);
975}
976
977void __perf_evlist__reset_sample_bit(struct evlist *evlist,
978				     enum perf_event_sample_format bit)
979{
980	struct evsel *evsel;
981
982	evlist__for_each_entry(evlist, evsel)
983		__evsel__reset_sample_bit(evsel, bit);
984}
985
986int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
987{
988	struct evsel *evsel;
989	int err = 0;
990
991	evlist__for_each_entry(evlist, evsel) {
992		if (evsel->filter == NULL)
993			continue;
994
995		/*
996		 * filters only work for tracepoint event, which doesn't have cpu limit.
997		 * So evlist and evsel should always be same.
998		 */
999		err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1000		if (err) {
1001			*err_evsel = evsel;
1002			break;
1003		}
1004	}
1005
1006	return err;
1007}
1008
1009int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1010{
1011	struct evsel *evsel;
1012	int err = 0;
1013
1014	if (filter == NULL)
1015		return -1;
1016
1017	evlist__for_each_entry(evlist, evsel) {
1018		if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1019			continue;
1020
1021		err = evsel__set_filter(evsel, filter);
1022		if (err)
1023			break;
1024	}
1025
1026	return err;
1027}
1028
1029int perf_evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1030{
1031	struct evsel *evsel;
1032	int err = 0;
1033
1034	if (filter == NULL)
1035		return -1;
1036
1037	evlist__for_each_entry(evlist, evsel) {
1038		if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1039			continue;
1040
1041		err = evsel__append_tp_filter(evsel, filter);
1042		if (err)
1043			break;
1044	}
1045
1046	return err;
1047}
1048
1049char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1050{
1051	char *filter;
1052	size_t i;
1053
1054	for (i = 0; i < npids; ++i) {
1055		if (i == 0) {
1056			if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1057				return NULL;
1058		} else {
1059			char *tmp;
1060
1061			if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1062				goto out_free;
1063
1064			free(filter);
1065			filter = tmp;
1066		}
1067	}
1068
1069	return filter;
1070out_free:
1071	free(filter);
1072	return NULL;
1073}
1074
1075int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1076{
1077	char *filter = asprintf__tp_filter_pids(npids, pids);
1078	int ret = perf_evlist__set_tp_filter(evlist, filter);
1079
1080	free(filter);
1081	return ret;
1082}
1083
1084int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1085{
1086	return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1087}
1088
1089int perf_evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1090{
1091	char *filter = asprintf__tp_filter_pids(npids, pids);
1092	int ret = perf_evlist__append_tp_filter(evlist, filter);
1093
1094	free(filter);
1095	return ret;
1096}
1097
1098int perf_evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1099{
1100	return perf_evlist__append_tp_filter_pids(evlist, 1, &pid);
1101}
1102
1103bool evlist__valid_sample_type(struct evlist *evlist)
1104{
1105	struct evsel *pos;
1106
1107	if (evlist->core.nr_entries == 1)
1108		return true;
1109
1110	if (evlist->id_pos < 0 || evlist->is_pos < 0)
1111		return false;
1112
1113	evlist__for_each_entry(evlist, pos) {
1114		if (pos->id_pos != evlist->id_pos ||
1115		    pos->is_pos != evlist->is_pos)
1116			return false;
1117	}
1118
1119	return true;
1120}
1121
1122u64 __evlist__combined_sample_type(struct evlist *evlist)
1123{
1124	struct evsel *evsel;
1125
1126	if (evlist->combined_sample_type)
1127		return evlist->combined_sample_type;
1128
1129	evlist__for_each_entry(evlist, evsel)
1130		evlist->combined_sample_type |= evsel->core.attr.sample_type;
1131
1132	return evlist->combined_sample_type;
1133}
1134
1135u64 evlist__combined_sample_type(struct evlist *evlist)
1136{
1137	evlist->combined_sample_type = 0;
1138	return __evlist__combined_sample_type(evlist);
1139}
1140
1141u64 evlist__combined_branch_type(struct evlist *evlist)
1142{
1143	struct evsel *evsel;
1144	u64 branch_type = 0;
1145
1146	evlist__for_each_entry(evlist, evsel)
1147		branch_type |= evsel->core.attr.branch_sample_type;
1148	return branch_type;
1149}
1150
1151bool perf_evlist__valid_read_format(struct evlist *evlist)
1152{
1153	struct evsel *first = evlist__first(evlist), *pos = first;
1154	u64 read_format = first->core.attr.read_format;
1155	u64 sample_type = first->core.attr.sample_type;
1156
1157	evlist__for_each_entry(evlist, pos) {
1158		if (read_format != pos->core.attr.read_format) {
1159			pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1160				 read_format, (u64)pos->core.attr.read_format);
1161		}
1162	}
1163
1164	/* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1165	if ((sample_type & PERF_SAMPLE_READ) &&
1166	    !(read_format & PERF_FORMAT_ID)) {
1167		return false;
1168	}
1169
1170	return true;
1171}
1172
1173u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1174{
1175	struct evsel *first = evlist__first(evlist);
1176	struct perf_sample *data;
1177	u64 sample_type;
1178	u16 size = 0;
1179
1180	if (!first->core.attr.sample_id_all)
1181		goto out;
1182
1183	sample_type = first->core.attr.sample_type;
1184
1185	if (sample_type & PERF_SAMPLE_TID)
1186		size += sizeof(data->tid) * 2;
1187
1188       if (sample_type & PERF_SAMPLE_TIME)
1189		size += sizeof(data->time);
1190
1191	if (sample_type & PERF_SAMPLE_ID)
1192		size += sizeof(data->id);
1193
1194	if (sample_type & PERF_SAMPLE_STREAM_ID)
1195		size += sizeof(data->stream_id);
1196
1197	if (sample_type & PERF_SAMPLE_CPU)
1198		size += sizeof(data->cpu) * 2;
1199
1200	if (sample_type & PERF_SAMPLE_IDENTIFIER)
1201		size += sizeof(data->id);
1202out:
1203	return size;
1204}
1205
1206bool evlist__valid_sample_id_all(struct evlist *evlist)
1207{
1208	struct evsel *first = evlist__first(evlist), *pos = first;
1209
1210	evlist__for_each_entry_continue(evlist, pos) {
1211		if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1212			return false;
1213	}
1214
1215	return true;
1216}
1217
1218bool evlist__sample_id_all(struct evlist *evlist)
1219{
1220	struct evsel *first = evlist__first(evlist);
1221	return first->core.attr.sample_id_all;
1222}
1223
1224void perf_evlist__set_selected(struct evlist *evlist,
1225			       struct evsel *evsel)
1226{
1227	evlist->selected = evsel;
1228}
1229
1230void evlist__close(struct evlist *evlist)
1231{
1232	struct evsel *evsel;
1233	struct affinity affinity;
1234	int cpu, i;
1235
1236	/*
1237	 * With perf record core.cpus is usually NULL.
1238	 * Use the old method to handle this for now.
1239	 */
1240	if (!evlist->core.cpus) {
1241		evlist__for_each_entry_reverse(evlist, evsel)
1242			evsel__close(evsel);
1243		return;
1244	}
1245
1246	if (affinity__setup(&affinity) < 0)
1247		return;
1248	evlist__for_each_cpu(evlist, i, cpu) {
1249		affinity__set(&affinity, cpu);
1250
1251		evlist__for_each_entry_reverse(evlist, evsel) {
1252			if (evsel__cpu_iter_skip(evsel, cpu))
1253			    continue;
1254			perf_evsel__close_cpu(&evsel->core, evsel->cpu_iter - 1);
1255		}
1256	}
1257	affinity__cleanup(&affinity);
1258	evlist__for_each_entry_reverse(evlist, evsel) {
1259		perf_evsel__free_fd(&evsel->core);
1260		perf_evsel__free_id(&evsel->core);
1261	}
1262}
1263
1264static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1265{
1266	struct perf_cpu_map *cpus;
1267	struct perf_thread_map *threads;
1268	int err = -ENOMEM;
1269
1270	/*
1271	 * Try reading /sys/devices/system/cpu/online to get
1272	 * an all cpus map.
1273	 *
1274	 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1275	 * code needs an overhaul to properly forward the
1276	 * error, and we may not want to do that fallback to a
1277	 * default cpu identity map :-\
1278	 */
1279	cpus = perf_cpu_map__new(NULL);
1280	if (!cpus)
1281		goto out;
1282
1283	threads = perf_thread_map__new_dummy();
1284	if (!threads)
1285		goto out_put;
1286
1287	perf_evlist__set_maps(&evlist->core, cpus, threads);
1288
1289	perf_thread_map__put(threads);
1290out_put:
1291	perf_cpu_map__put(cpus);
1292out:
1293	return err;
1294}
1295
1296int evlist__open(struct evlist *evlist)
1297{
1298	struct evsel *evsel;
1299	int err;
1300
1301	/*
1302	 * Default: one fd per CPU, all threads, aka systemwide
1303	 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1304	 */
1305	if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1306		err = perf_evlist__create_syswide_maps(evlist);
1307		if (err < 0)
1308			goto out_err;
1309	}
1310
1311	perf_evlist__update_id_pos(evlist);
1312
1313	evlist__for_each_entry(evlist, evsel) {
1314		err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1315		if (err < 0)
1316			goto out_err;
1317	}
1318
1319	return 0;
1320out_err:
1321	evlist__close(evlist);
1322	errno = -err;
1323	return err;
1324}
1325
1326int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1327				  const char *argv[], bool pipe_output,
1328				  void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1329{
1330	int child_ready_pipe[2], go_pipe[2];
1331	char bf;
1332
1333	if (pipe(child_ready_pipe) < 0) {
1334		perror("failed to create 'ready' pipe");
1335		return -1;
1336	}
1337
1338	if (pipe(go_pipe) < 0) {
1339		perror("failed to create 'go' pipe");
1340		goto out_close_ready_pipe;
1341	}
1342
1343	evlist->workload.pid = fork();
1344	if (evlist->workload.pid < 0) {
1345		perror("failed to fork");
1346		goto out_close_pipes;
1347	}
1348
1349	if (!evlist->workload.pid) {
1350		int ret;
1351
1352		if (pipe_output)
1353			dup2(2, 1);
1354
1355		signal(SIGTERM, SIG_DFL);
1356
1357		close(child_ready_pipe[0]);
1358		close(go_pipe[1]);
1359		fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1360
1361		/*
1362		 * Tell the parent we're ready to go
1363		 */
1364		close(child_ready_pipe[1]);
1365
1366		/*
1367		 * Wait until the parent tells us to go.
1368		 */
1369		ret = read(go_pipe[0], &bf, 1);
1370		/*
1371		 * The parent will ask for the execvp() to be performed by
1372		 * writing exactly one byte, in workload.cork_fd, usually via
1373		 * perf_evlist__start_workload().
1374		 *
1375		 * For cancelling the workload without actually running it,
1376		 * the parent will just close workload.cork_fd, without writing
1377		 * anything, i.e. read will return zero and we just exit()
1378		 * here.
1379		 */
1380		if (ret != 1) {
1381			if (ret == -1)
1382				perror("unable to read pipe");
1383			exit(ret);
1384		}
1385
1386		execvp(argv[0], (char **)argv);
1387
1388		if (exec_error) {
1389			union sigval val;
1390
1391			val.sival_int = errno;
1392			if (sigqueue(getppid(), SIGUSR1, val))
1393				perror(argv[0]);
1394		} else
1395			perror(argv[0]);
1396		exit(-1);
1397	}
1398
1399	if (exec_error) {
1400		struct sigaction act = {
1401			.sa_flags     = SA_SIGINFO,
1402			.sa_sigaction = exec_error,
1403		};
1404		sigaction(SIGUSR1, &act, NULL);
1405	}
1406
1407	if (target__none(target)) {
1408		if (evlist->core.threads == NULL) {
1409			fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1410				__func__, __LINE__);
1411			goto out_close_pipes;
1412		}
1413		perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1414	}
1415
1416	close(child_ready_pipe[1]);
1417	close(go_pipe[0]);
1418	/*
1419	 * wait for child to settle
1420	 */
1421	if (read(child_ready_pipe[0], &bf, 1) == -1) {
1422		perror("unable to read pipe");
1423		goto out_close_pipes;
1424	}
1425
1426	fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1427	evlist->workload.cork_fd = go_pipe[1];
1428	close(child_ready_pipe[0]);
1429	return 0;
1430
1431out_close_pipes:
1432	close(go_pipe[0]);
1433	close(go_pipe[1]);
1434out_close_ready_pipe:
1435	close(child_ready_pipe[0]);
1436	close(child_ready_pipe[1]);
1437	return -1;
1438}
1439
1440int perf_evlist__start_workload(struct evlist *evlist)
1441{
1442	if (evlist->workload.cork_fd > 0) {
1443		char bf = 0;
1444		int ret;
1445		/*
1446		 * Remove the cork, let it rip!
1447		 */
1448		ret = write(evlist->workload.cork_fd, &bf, 1);
1449		if (ret < 0)
1450			perror("unable to write to pipe");
1451
1452		close(evlist->workload.cork_fd);
1453		return ret;
1454	}
1455
1456	return 0;
1457}
1458
1459int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1460			      struct perf_sample *sample)
1461{
1462	struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1463
1464	if (!evsel)
1465		return -EFAULT;
1466	return evsel__parse_sample(evsel, event, sample);
1467}
1468
1469int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1470					union perf_event *event,
1471					u64 *timestamp)
1472{
1473	struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1474
1475	if (!evsel)
1476		return -EFAULT;
1477	return evsel__parse_sample_timestamp(evsel, event, timestamp);
1478}
1479
1480int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1481{
1482	int printed, value;
1483	char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1484
1485	switch (err) {
1486	case EACCES:
1487	case EPERM:
1488		printed = scnprintf(buf, size,
1489				    "Error:\t%s.\n"
1490				    "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1491
1492		value = perf_event_paranoid();
1493
1494		printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1495
1496		if (value >= 2) {
1497			printed += scnprintf(buf + printed, size - printed,
1498					     "For your workloads it needs to be <= 1\nHint:\t");
1499		}
1500		printed += scnprintf(buf + printed, size - printed,
1501				     "For system wide tracing it needs to be set to -1.\n");
1502
1503		printed += scnprintf(buf + printed, size - printed,
1504				    "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1505				    "Hint:\tThe current value is %d.", value);
1506		break;
1507	case EINVAL: {
1508		struct evsel *first = evlist__first(evlist);
1509		int max_freq;
1510
1511		if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1512			goto out_default;
1513
1514		if (first->core.attr.sample_freq < (u64)max_freq)
1515			goto out_default;
1516
1517		printed = scnprintf(buf, size,
1518				    "Error:\t%s.\n"
1519				    "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1520				    "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1521				    emsg, max_freq, first->core.attr.sample_freq);
1522		break;
1523	}
1524	default:
1525out_default:
1526		scnprintf(buf, size, "%s", emsg);
1527		break;
1528	}
1529
1530	return 0;
1531}
1532
1533int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1534{
1535	char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1536	int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1537
1538	switch (err) {
1539	case EPERM:
1540		sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1541		printed += scnprintf(buf + printed, size - printed,
1542				     "Error:\t%s.\n"
1543				     "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1544				     "Hint:\tTried using %zd kB.\n",
1545				     emsg, pages_max_per_user, pages_attempted);
1546
1547		if (pages_attempted >= pages_max_per_user) {
1548			printed += scnprintf(buf + printed, size - printed,
1549					     "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1550					     pages_max_per_user + pages_attempted);
1551		}
1552
1553		printed += scnprintf(buf + printed, size - printed,
1554				     "Hint:\tTry using a smaller -m/--mmap-pages value.");
1555		break;
1556	default:
1557		scnprintf(buf, size, "%s", emsg);
1558		break;
1559	}
1560
1561	return 0;
1562}
1563
1564void perf_evlist__to_front(struct evlist *evlist,
1565			   struct evsel *move_evsel)
1566{
1567	struct evsel *evsel, *n;
1568	LIST_HEAD(move);
1569
1570	if (move_evsel == evlist__first(evlist))
1571		return;
1572
1573	evlist__for_each_entry_safe(evlist, n, evsel) {
1574		if (evsel->leader == move_evsel->leader)
1575			list_move_tail(&evsel->core.node, &move);
1576	}
1577
1578	list_splice(&move, &evlist->core.entries);
1579}
1580
1581struct evsel *perf_evlist__get_tracking_event(struct evlist *evlist)
1582{
1583	struct evsel *evsel;
1584
1585	evlist__for_each_entry(evlist, evsel) {
1586		if (evsel->tracking)
1587			return evsel;
1588	}
1589
1590	return evlist__first(evlist);
1591}
1592
1593void perf_evlist__set_tracking_event(struct evlist *evlist,
1594				     struct evsel *tracking_evsel)
1595{
1596	struct evsel *evsel;
1597
1598	if (tracking_evsel->tracking)
1599		return;
1600
1601	evlist__for_each_entry(evlist, evsel) {
1602		if (evsel != tracking_evsel)
1603			evsel->tracking = false;
1604	}
1605
1606	tracking_evsel->tracking = true;
1607}
1608
1609struct evsel *
1610perf_evlist__find_evsel_by_str(struct evlist *evlist,
1611			       const char *str)
1612{
1613	struct evsel *evsel;
1614
1615	evlist__for_each_entry(evlist, evsel) {
1616		if (!evsel->name)
1617			continue;
1618		if (strcmp(str, evsel->name) == 0)
1619			return evsel;
1620	}
1621
1622	return NULL;
1623}
1624
1625void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1626				  enum bkw_mmap_state state)
1627{
1628	enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1629	enum action {
1630		NONE,
1631		PAUSE,
1632		RESUME,
1633	} action = NONE;
1634
1635	if (!evlist->overwrite_mmap)
1636		return;
1637
1638	switch (old_state) {
1639	case BKW_MMAP_NOTREADY: {
1640		if (state != BKW_MMAP_RUNNING)
1641			goto state_err;
1642		break;
1643	}
1644	case BKW_MMAP_RUNNING: {
1645		if (state != BKW_MMAP_DATA_PENDING)
1646			goto state_err;
1647		action = PAUSE;
1648		break;
1649	}
1650	case BKW_MMAP_DATA_PENDING: {
1651		if (state != BKW_MMAP_EMPTY)
1652			goto state_err;
1653		break;
1654	}
1655	case BKW_MMAP_EMPTY: {
1656		if (state != BKW_MMAP_RUNNING)
1657			goto state_err;
1658		action = RESUME;
1659		break;
1660	}
1661	default:
1662		WARN_ONCE(1, "Shouldn't get there\n");
1663	}
1664
1665	evlist->bkw_mmap_state = state;
1666
1667	switch (action) {
1668	case PAUSE:
1669		perf_evlist__pause(evlist);
1670		break;
1671	case RESUME:
1672		perf_evlist__resume(evlist);
1673		break;
1674	case NONE:
1675	default:
1676		break;
1677	}
1678
1679state_err:
1680	return;
1681}
1682
1683bool perf_evlist__exclude_kernel(struct evlist *evlist)
1684{
1685	struct evsel *evsel;
1686
1687	evlist__for_each_entry(evlist, evsel) {
1688		if (!evsel->core.attr.exclude_kernel)
1689			return false;
1690	}
1691
1692	return true;
1693}
1694
1695/*
1696 * Events in data file are not collect in groups, but we still want
1697 * the group display. Set the artificial group and set the leader's
1698 * forced_leader flag to notify the display code.
1699 */
1700void perf_evlist__force_leader(struct evlist *evlist)
1701{
1702	if (!evlist->nr_groups) {
1703		struct evsel *leader = evlist__first(evlist);
1704
1705		perf_evlist__set_leader(evlist);
1706		leader->forced_leader = true;
1707	}
1708}
1709
1710struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1711						 struct evsel *evsel,
1712						bool close)
1713{
1714	struct evsel *c2, *leader;
1715	bool is_open = true;
1716
1717	leader = evsel->leader;
1718	pr_debug("Weak group for %s/%d failed\n",
1719			leader->name, leader->core.nr_members);
1720
1721	/*
1722	 * for_each_group_member doesn't work here because it doesn't
1723	 * include the first entry.
1724	 */
1725	evlist__for_each_entry(evsel_list, c2) {
1726		if (c2 == evsel)
1727			is_open = false;
1728		if (c2->leader == leader) {
1729			if (is_open && close)
1730				perf_evsel__close(&c2->core);
1731			c2->leader = c2;
1732			c2->core.nr_members = 0;
1733			/*
1734			 * Set this for all former members of the group
1735			 * to indicate they get reopened.
1736			 */
1737			c2->reset_group = true;
1738		}
1739	}
1740	return leader;
1741}
1742
1743static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1744{
1745	char *s, *p;
1746	int ret = 0, fd;
1747
1748	if (strncmp(str, "fifo:", 5))
1749		return -EINVAL;
1750
1751	str += 5;
1752	if (!*str || *str == ',')
1753		return -EINVAL;
1754
1755	s = strdup(str);
1756	if (!s)
1757		return -ENOMEM;
1758
1759	p = strchr(s, ',');
1760	if (p)
1761		*p = '\0';
1762
1763	/*
1764	 * O_RDWR avoids POLLHUPs which is necessary to allow the other
1765	 * end of a FIFO to be repeatedly opened and closed.
1766	 */
1767	fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1768	if (fd < 0) {
1769		pr_err("Failed to open '%s'\n", s);
1770		ret = -errno;
1771		goto out_free;
1772	}
1773	*ctl_fd = fd;
1774	*ctl_fd_close = true;
1775
1776	if (p && *++p) {
1777		/* O_RDWR | O_NONBLOCK means the other end need not be open */
1778		fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1779		if (fd < 0) {
1780			pr_err("Failed to open '%s'\n", p);
1781			ret = -errno;
1782			goto out_free;
1783		}
1784		*ctl_fd_ack = fd;
1785	}
1786
1787out_free:
1788	free(s);
1789	return ret;
1790}
1791
1792int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1793{
1794	char *comma = NULL, *endptr = NULL;
1795
1796	*ctl_fd_close = false;
1797
1798	if (strncmp(str, "fd:", 3))
1799		return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1800
1801	*ctl_fd = strtoul(&str[3], &endptr, 0);
1802	if (endptr == &str[3])
1803		return -EINVAL;
1804
1805	comma = strchr(str, ',');
1806	if (comma) {
1807		if (endptr != comma)
1808			return -EINVAL;
1809
1810		*ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1811		if (endptr == comma + 1 || *endptr != '\0')
1812			return -EINVAL;
1813	}
1814
1815	return 0;
1816}
1817
1818void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1819{
1820	if (*ctl_fd_close) {
1821		*ctl_fd_close = false;
1822		close(ctl_fd);
1823		if (ctl_fd_ack >= 0)
1824			close(ctl_fd_ack);
1825	}
1826}
1827
1828int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1829{
1830	if (fd == -1) {
1831		pr_debug("Control descriptor is not initialized\n");
1832		return 0;
1833	}
1834
1835	evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1836						     fdarray_flag__nonfilterable);
1837	if (evlist->ctl_fd.pos < 0) {
1838		evlist->ctl_fd.pos = -1;
1839		pr_err("Failed to add ctl fd entry: %m\n");
1840		return -1;
1841	}
1842
1843	evlist->ctl_fd.fd = fd;
1844	evlist->ctl_fd.ack = ack;
1845
1846	return 0;
1847}
1848
1849bool evlist__ctlfd_initialized(struct evlist *evlist)
1850{
1851	return evlist->ctl_fd.pos >= 0;
1852}
1853
1854int evlist__finalize_ctlfd(struct evlist *evlist)
1855{
1856	struct pollfd *entries = evlist->core.pollfd.entries;
1857
1858	if (!evlist__ctlfd_initialized(evlist))
1859		return 0;
1860
1861	entries[evlist->ctl_fd.pos].fd = -1;
1862	entries[evlist->ctl_fd.pos].events = 0;
1863	entries[evlist->ctl_fd.pos].revents = 0;
1864
1865	evlist->ctl_fd.pos = -1;
1866	evlist->ctl_fd.ack = -1;
1867	evlist->ctl_fd.fd = -1;
1868
1869	return 0;
1870}
1871
1872static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
1873			      char *cmd_data, size_t data_size)
1874{
1875	int err;
1876	char c;
1877	size_t bytes_read = 0;
1878
1879	*cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1880	memset(cmd_data, 0, data_size);
1881	data_size--;
1882
1883	do {
1884		err = read(evlist->ctl_fd.fd, &c, 1);
1885		if (err > 0) {
1886			if (c == '\n' || c == '\0')
1887				break;
1888			cmd_data[bytes_read++] = c;
1889			if (bytes_read == data_size)
1890				break;
1891			continue;
1892		} else if (err == -1) {
1893			if (errno == EINTR)
1894				continue;
1895			if (errno == EAGAIN || errno == EWOULDBLOCK)
1896				err = 0;
1897			else
1898				pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
1899		}
1900		break;
1901	} while (1);
1902
1903	pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
1904		 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
1905
1906	if (bytes_read > 0) {
1907		if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
1908			     (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
1909			*cmd = EVLIST_CTL_CMD_ENABLE;
1910		} else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
1911				    (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
1912			*cmd = EVLIST_CTL_CMD_DISABLE;
1913		} else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
1914				    (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
1915			*cmd = EVLIST_CTL_CMD_SNAPSHOT;
1916			pr_debug("is snapshot\n");
1917		}
1918	}
1919
1920	return bytes_read ? (int)bytes_read : err;
1921}
1922
1923int evlist__ctlfd_ack(struct evlist *evlist)
1924{
1925	int err;
1926
1927	if (evlist->ctl_fd.ack == -1)
1928		return 0;
1929
1930	err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
1931		    sizeof(EVLIST_CTL_CMD_ACK_TAG));
1932	if (err == -1)
1933		pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
1934
1935	return err;
1936}
1937
1938int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
1939{
1940	int err = 0;
1941	char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
1942	int ctlfd_pos = evlist->ctl_fd.pos;
1943	struct pollfd *entries = evlist->core.pollfd.entries;
1944
1945	if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
1946		return 0;
1947
1948	if (entries[ctlfd_pos].revents & POLLIN) {
1949		err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
1950					 EVLIST_CTL_CMD_MAX_LEN);
1951		if (err > 0) {
1952			switch (*cmd) {
1953			case EVLIST_CTL_CMD_ENABLE:
1954				evlist__enable(evlist);
1955				break;
1956			case EVLIST_CTL_CMD_DISABLE:
1957				evlist__disable(evlist);
1958				break;
1959			case EVLIST_CTL_CMD_SNAPSHOT:
1960				break;
1961			case EVLIST_CTL_CMD_ACK:
1962			case EVLIST_CTL_CMD_UNSUPPORTED:
1963			default:
1964				pr_debug("ctlfd: unsupported %d\n", *cmd);
1965				break;
1966			}
1967			if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
1968			      *cmd == EVLIST_CTL_CMD_SNAPSHOT))
1969				evlist__ctlfd_ack(evlist);
1970		}
1971	}
1972
1973	if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
1974		evlist__finalize_ctlfd(evlist);
1975	else
1976		entries[ctlfd_pos].revents = 0;
1977
1978	return err;
1979}
1980
1981struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
1982{
1983	struct evsel *evsel;
1984
1985	evlist__for_each_entry(evlist, evsel) {
1986		if (evsel->idx == idx)
1987			return evsel;
1988	}
1989	return NULL;
1990}
1991