Merge branch 'x86-seccomp-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[cascardo/linux.git] / tools / perf / util / evlist.c
1 /*
2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3  *
4  * Parts came from builtin-{top,stat,record}.c, see those files for further
5  * copyright notes.
6  *
7  * Released under the GPL v2. (and only v2, not any later version)
8  */
9 #include "util.h"
10 #include <api/fs/debugfs.h>
11 #include <poll.h>
12 #include "cpumap.h"
13 #include "thread_map.h"
14 #include "target.h"
15 #include "evlist.h"
16 #include "evsel.h"
17 #include "debug.h"
18 #include <unistd.h>
19
20 #include "parse-events.h"
21 #include "parse-options.h"
22
23 #include <sys/mman.h>
24
25 #include <linux/bitops.h>
26 #include <linux/hash.h>
27
28 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
29 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
30
31 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
32 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
33
34 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
35                        struct thread_map *threads)
36 {
37         int i;
38
39         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
40                 INIT_HLIST_HEAD(&evlist->heads[i]);
41         INIT_LIST_HEAD(&evlist->entries);
42         perf_evlist__set_maps(evlist, cpus, threads);
43         fdarray__init(&evlist->pollfd, 64);
44         evlist->workload.pid = -1;
45 }
46
47 struct perf_evlist *perf_evlist__new(void)
48 {
49         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
50
51         if (evlist != NULL)
52                 perf_evlist__init(evlist, NULL, NULL);
53
54         return evlist;
55 }
56
57 struct perf_evlist *perf_evlist__new_default(void)
58 {
59         struct perf_evlist *evlist = perf_evlist__new();
60
61         if (evlist && perf_evlist__add_default(evlist)) {
62                 perf_evlist__delete(evlist);
63                 evlist = NULL;
64         }
65
66         return evlist;
67 }
68
69 /**
70  * perf_evlist__set_id_pos - set the positions of event ids.
71  * @evlist: selected event list
72  *
73  * Events with compatible sample types all have the same id_pos
74  * and is_pos.  For convenience, put a copy on evlist.
75  */
76 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
77 {
78         struct perf_evsel *first = perf_evlist__first(evlist);
79
80         evlist->id_pos = first->id_pos;
81         evlist->is_pos = first->is_pos;
82 }
83
84 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
85 {
86         struct perf_evsel *evsel;
87
88         evlist__for_each(evlist, evsel)
89                 perf_evsel__calc_id_pos(evsel);
90
91         perf_evlist__set_id_pos(evlist);
92 }
93
94 static void perf_evlist__purge(struct perf_evlist *evlist)
95 {
96         struct perf_evsel *pos, *n;
97
98         evlist__for_each_safe(evlist, n, pos) {
99                 list_del_init(&pos->node);
100                 perf_evsel__delete(pos);
101         }
102
103         evlist->nr_entries = 0;
104 }
105
106 void perf_evlist__exit(struct perf_evlist *evlist)
107 {
108         zfree(&evlist->mmap);
109         fdarray__exit(&evlist->pollfd);
110 }
111
112 void perf_evlist__delete(struct perf_evlist *evlist)
113 {
114         perf_evlist__munmap(evlist);
115         perf_evlist__close(evlist);
116         cpu_map__delete(evlist->cpus);
117         thread_map__delete(evlist->threads);
118         evlist->cpus = NULL;
119         evlist->threads = NULL;
120         perf_evlist__purge(evlist);
121         perf_evlist__exit(evlist);
122         free(evlist);
123 }
124
125 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
126 {
127         list_add_tail(&entry->node, &evlist->entries);
128         entry->idx = evlist->nr_entries;
129         entry->tracking = !entry->idx;
130
131         if (!evlist->nr_entries++)
132                 perf_evlist__set_id_pos(evlist);
133 }
134
135 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
136                                    struct list_head *list,
137                                    int nr_entries)
138 {
139         bool set_id_pos = !evlist->nr_entries;
140
141         list_splice_tail(list, &evlist->entries);
142         evlist->nr_entries += nr_entries;
143         if (set_id_pos)
144                 perf_evlist__set_id_pos(evlist);
145 }
146
147 void __perf_evlist__set_leader(struct list_head *list)
148 {
149         struct perf_evsel *evsel, *leader;
150
151         leader = list_entry(list->next, struct perf_evsel, node);
152         evsel = list_entry(list->prev, struct perf_evsel, node);
153
154         leader->nr_members = evsel->idx - leader->idx + 1;
155
156         __evlist__for_each(list, evsel) {
157                 evsel->leader = leader;
158         }
159 }
160
161 void perf_evlist__set_leader(struct perf_evlist *evlist)
162 {
163         if (evlist->nr_entries) {
164                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
165                 __perf_evlist__set_leader(&evlist->entries);
166         }
167 }
168
169 int perf_evlist__add_default(struct perf_evlist *evlist)
170 {
171         struct perf_event_attr attr = {
172                 .type = PERF_TYPE_HARDWARE,
173                 .config = PERF_COUNT_HW_CPU_CYCLES,
174         };
175         struct perf_evsel *evsel;
176
177         event_attr_init(&attr);
178
179         evsel = perf_evsel__new(&attr);
180         if (evsel == NULL)
181                 goto error;
182
183         /* use strdup() because free(evsel) assumes name is allocated */
184         evsel->name = strdup("cycles");
185         if (!evsel->name)
186                 goto error_free;
187
188         perf_evlist__add(evlist, evsel);
189         return 0;
190 error_free:
191         perf_evsel__delete(evsel);
192 error:
193         return -ENOMEM;
194 }
195
196 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
197                                   struct perf_event_attr *attrs, size_t nr_attrs)
198 {
199         struct perf_evsel *evsel, *n;
200         LIST_HEAD(head);
201         size_t i;
202
203         for (i = 0; i < nr_attrs; i++) {
204                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
205                 if (evsel == NULL)
206                         goto out_delete_partial_list;
207                 list_add_tail(&evsel->node, &head);
208         }
209
210         perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
211
212         return 0;
213
214 out_delete_partial_list:
215         __evlist__for_each_safe(&head, n, evsel)
216                 perf_evsel__delete(evsel);
217         return -1;
218 }
219
220 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
221                                      struct perf_event_attr *attrs, size_t nr_attrs)
222 {
223         size_t i;
224
225         for (i = 0; i < nr_attrs; i++)
226                 event_attr_init(attrs + i);
227
228         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
229 }
230
231 struct perf_evsel *
232 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
233 {
234         struct perf_evsel *evsel;
235
236         evlist__for_each(evlist, evsel) {
237                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
238                     (int)evsel->attr.config == id)
239                         return evsel;
240         }
241
242         return NULL;
243 }
244
245 struct perf_evsel *
246 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
247                                      const char *name)
248 {
249         struct perf_evsel *evsel;
250
251         evlist__for_each(evlist, evsel) {
252                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
253                     (strcmp(evsel->name, name) == 0))
254                         return evsel;
255         }
256
257         return NULL;
258 }
259
260 int perf_evlist__add_newtp(struct perf_evlist *evlist,
261                            const char *sys, const char *name, void *handler)
262 {
263         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
264
265         if (evsel == NULL)
266                 return -1;
267
268         evsel->handler = handler;
269         perf_evlist__add(evlist, evsel);
270         return 0;
271 }
272
273 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
274                                    struct perf_evsel *evsel)
275 {
276         if (evsel->system_wide)
277                 return 1;
278         else
279                 return thread_map__nr(evlist->threads);
280 }
281
282 void perf_evlist__disable(struct perf_evlist *evlist)
283 {
284         int cpu, thread;
285         struct perf_evsel *pos;
286         int nr_cpus = cpu_map__nr(evlist->cpus);
287         int nr_threads;
288
289         for (cpu = 0; cpu < nr_cpus; cpu++) {
290                 evlist__for_each(evlist, pos) {
291                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
292                                 continue;
293                         nr_threads = perf_evlist__nr_threads(evlist, pos);
294                         for (thread = 0; thread < nr_threads; thread++)
295                                 ioctl(FD(pos, cpu, thread),
296                                       PERF_EVENT_IOC_DISABLE, 0);
297                 }
298         }
299 }
300
301 void perf_evlist__enable(struct perf_evlist *evlist)
302 {
303         int cpu, thread;
304         struct perf_evsel *pos;
305         int nr_cpus = cpu_map__nr(evlist->cpus);
306         int nr_threads;
307
308         for (cpu = 0; cpu < nr_cpus; cpu++) {
309                 evlist__for_each(evlist, pos) {
310                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
311                                 continue;
312                         nr_threads = perf_evlist__nr_threads(evlist, pos);
313                         for (thread = 0; thread < nr_threads; thread++)
314                                 ioctl(FD(pos, cpu, thread),
315                                       PERF_EVENT_IOC_ENABLE, 0);
316                 }
317         }
318 }
319
320 int perf_evlist__disable_event(struct perf_evlist *evlist,
321                                struct perf_evsel *evsel)
322 {
323         int cpu, thread, err;
324         int nr_cpus = cpu_map__nr(evlist->cpus);
325         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
326
327         if (!evsel->fd)
328                 return 0;
329
330         for (cpu = 0; cpu < nr_cpus; cpu++) {
331                 for (thread = 0; thread < nr_threads; thread++) {
332                         err = ioctl(FD(evsel, cpu, thread),
333                                     PERF_EVENT_IOC_DISABLE, 0);
334                         if (err)
335                                 return err;
336                 }
337         }
338         return 0;
339 }
340
341 int perf_evlist__enable_event(struct perf_evlist *evlist,
342                               struct perf_evsel *evsel)
343 {
344         int cpu, thread, err;
345         int nr_cpus = cpu_map__nr(evlist->cpus);
346         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
347
348         if (!evsel->fd)
349                 return -EINVAL;
350
351         for (cpu = 0; cpu < nr_cpus; cpu++) {
352                 for (thread = 0; thread < nr_threads; thread++) {
353                         err = ioctl(FD(evsel, cpu, thread),
354                                     PERF_EVENT_IOC_ENABLE, 0);
355                         if (err)
356                                 return err;
357                 }
358         }
359         return 0;
360 }
361
362 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
363                                          struct perf_evsel *evsel, int cpu)
364 {
365         int thread, err;
366         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
367
368         if (!evsel->fd)
369                 return -EINVAL;
370
371         for (thread = 0; thread < nr_threads; thread++) {
372                 err = ioctl(FD(evsel, cpu, thread),
373                             PERF_EVENT_IOC_ENABLE, 0);
374                 if (err)
375                         return err;
376         }
377         return 0;
378 }
379
380 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
381                                             struct perf_evsel *evsel,
382                                             int thread)
383 {
384         int cpu, err;
385         int nr_cpus = cpu_map__nr(evlist->cpus);
386
387         if (!evsel->fd)
388                 return -EINVAL;
389
390         for (cpu = 0; cpu < nr_cpus; cpu++) {
391                 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
392                 if (err)
393                         return err;
394         }
395         return 0;
396 }
397
398 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
399                                   struct perf_evsel *evsel, int idx)
400 {
401         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
402
403         if (per_cpu_mmaps)
404                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
405         else
406                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
407 }
408
409 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
410 {
411         int nr_cpus = cpu_map__nr(evlist->cpus);
412         int nr_threads = thread_map__nr(evlist->threads);
413         int nfds = 0;
414         struct perf_evsel *evsel;
415
416         list_for_each_entry(evsel, &evlist->entries, node) {
417                 if (evsel->system_wide)
418                         nfds += nr_cpus;
419                 else
420                         nfds += nr_cpus * nr_threads;
421         }
422
423         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
424             fdarray__grow(&evlist->pollfd, nfds) < 0)
425                 return -ENOMEM;
426
427         return 0;
428 }
429
430 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
431 {
432         int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
433         /*
434          * Save the idx so that when we filter out fds POLLHUP'ed we can
435          * close the associated evlist->mmap[] entry.
436          */
437         if (pos >= 0) {
438                 evlist->pollfd.priv[pos].idx = idx;
439
440                 fcntl(fd, F_SETFL, O_NONBLOCK);
441         }
442
443         return pos;
444 }
445
446 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
447 {
448         return __perf_evlist__add_pollfd(evlist, fd, -1);
449 }
450
451 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
452 {
453         struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
454
455         perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
456 }
457
458 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
459 {
460         return fdarray__filter(&evlist->pollfd, revents_and_mask,
461                                perf_evlist__munmap_filtered);
462 }
463
464 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
465 {
466         return fdarray__poll(&evlist->pollfd, timeout);
467 }
468
469 static void perf_evlist__id_hash(struct perf_evlist *evlist,
470                                  struct perf_evsel *evsel,
471                                  int cpu, int thread, u64 id)
472 {
473         int hash;
474         struct perf_sample_id *sid = SID(evsel, cpu, thread);
475
476         sid->id = id;
477         sid->evsel = evsel;
478         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
479         hlist_add_head(&sid->node, &evlist->heads[hash]);
480 }
481
482 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
483                          int cpu, int thread, u64 id)
484 {
485         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
486         evsel->id[evsel->ids++] = id;
487 }
488
489 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
490                                   struct perf_evsel *evsel,
491                                   int cpu, int thread, int fd)
492 {
493         u64 read_data[4] = { 0, };
494         int id_idx = 1; /* The first entry is the counter value */
495         u64 id;
496         int ret;
497
498         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
499         if (!ret)
500                 goto add;
501
502         if (errno != ENOTTY)
503                 return -1;
504
505         /* Legacy way to get event id.. All hail to old kernels! */
506
507         /*
508          * This way does not work with group format read, so bail
509          * out in that case.
510          */
511         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
512                 return -1;
513
514         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
515             read(fd, &read_data, sizeof(read_data)) == -1)
516                 return -1;
517
518         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
519                 ++id_idx;
520         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
521                 ++id_idx;
522
523         id = read_data[id_idx];
524
525  add:
526         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
527         return 0;
528 }
529
530 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
531 {
532         struct hlist_head *head;
533         struct perf_sample_id *sid;
534         int hash;
535
536         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
537         head = &evlist->heads[hash];
538
539         hlist_for_each_entry(sid, head, node)
540                 if (sid->id == id)
541                         return sid;
542
543         return NULL;
544 }
545
546 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
547 {
548         struct perf_sample_id *sid;
549
550         if (evlist->nr_entries == 1)
551                 return perf_evlist__first(evlist);
552
553         sid = perf_evlist__id2sid(evlist, id);
554         if (sid)
555                 return sid->evsel;
556
557         if (!perf_evlist__sample_id_all(evlist))
558                 return perf_evlist__first(evlist);
559
560         return NULL;
561 }
562
563 static int perf_evlist__event2id(struct perf_evlist *evlist,
564                                  union perf_event *event, u64 *id)
565 {
566         const u64 *array = event->sample.array;
567         ssize_t n;
568
569         n = (event->header.size - sizeof(event->header)) >> 3;
570
571         if (event->header.type == PERF_RECORD_SAMPLE) {
572                 if (evlist->id_pos >= n)
573                         return -1;
574                 *id = array[evlist->id_pos];
575         } else {
576                 if (evlist->is_pos > n)
577                         return -1;
578                 n -= evlist->is_pos;
579                 *id = array[n];
580         }
581         return 0;
582 }
583
584 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
585                                                    union perf_event *event)
586 {
587         struct perf_evsel *first = perf_evlist__first(evlist);
588         struct hlist_head *head;
589         struct perf_sample_id *sid;
590         int hash;
591         u64 id;
592
593         if (evlist->nr_entries == 1)
594                 return first;
595
596         if (!first->attr.sample_id_all &&
597             event->header.type != PERF_RECORD_SAMPLE)
598                 return first;
599
600         if (perf_evlist__event2id(evlist, event, &id))
601                 return NULL;
602
603         /* Synthesized events have an id of zero */
604         if (!id)
605                 return first;
606
607         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
608         head = &evlist->heads[hash];
609
610         hlist_for_each_entry(sid, head, node) {
611                 if (sid->id == id)
612                         return sid->evsel;
613         }
614         return NULL;
615 }
616
617 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
618 {
619         struct perf_mmap *md = &evlist->mmap[idx];
620         unsigned int head = perf_mmap__read_head(md);
621         unsigned int old = md->prev;
622         unsigned char *data = md->base + page_size;
623         union perf_event *event = NULL;
624
625         if (evlist->overwrite) {
626                 /*
627                  * If we're further behind than half the buffer, there's a chance
628                  * the writer will bite our tail and mess up the samples under us.
629                  *
630                  * If we somehow ended up ahead of the head, we got messed up.
631                  *
632                  * In either case, truncate and restart at head.
633                  */
634                 int diff = head - old;
635                 if (diff > md->mask / 2 || diff < 0) {
636                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
637
638                         /*
639                          * head points to a known good entry, start there.
640                          */
641                         old = head;
642                 }
643         }
644
645         if (old != head) {
646                 size_t size;
647
648                 event = (union perf_event *)&data[old & md->mask];
649                 size = event->header.size;
650
651                 /*
652                  * Event straddles the mmap boundary -- header should always
653                  * be inside due to u64 alignment of output.
654                  */
655                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
656                         unsigned int offset = old;
657                         unsigned int len = min(sizeof(*event), size), cpy;
658                         void *dst = md->event_copy;
659
660                         do {
661                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
662                                 memcpy(dst, &data[offset & md->mask], cpy);
663                                 offset += cpy;
664                                 dst += cpy;
665                                 len -= cpy;
666                         } while (len);
667
668                         event = (union perf_event *) md->event_copy;
669                 }
670
671                 old += size;
672         }
673
674         md->prev = old;
675
676         return event;
677 }
678
679 static bool perf_mmap__empty(struct perf_mmap *md)
680 {
681         return perf_mmap__read_head(md) != md->prev;
682 }
683
684 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
685 {
686         ++evlist->mmap[idx].refcnt;
687 }
688
689 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
690 {
691         BUG_ON(evlist->mmap[idx].refcnt == 0);
692
693         if (--evlist->mmap[idx].refcnt == 0)
694                 __perf_evlist__munmap(evlist, idx);
695 }
696
697 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
698 {
699         struct perf_mmap *md = &evlist->mmap[idx];
700
701         if (!evlist->overwrite) {
702                 unsigned int old = md->prev;
703
704                 perf_mmap__write_tail(md, old);
705         }
706
707         if (md->refcnt == 1 && perf_mmap__empty(md))
708                 perf_evlist__mmap_put(evlist, idx);
709 }
710
711 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
712 {
713         if (evlist->mmap[idx].base != NULL) {
714                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
715                 evlist->mmap[idx].base = NULL;
716                 evlist->mmap[idx].refcnt = 0;
717         }
718 }
719
720 void perf_evlist__munmap(struct perf_evlist *evlist)
721 {
722         int i;
723
724         if (evlist->mmap == NULL)
725                 return;
726
727         for (i = 0; i < evlist->nr_mmaps; i++)
728                 __perf_evlist__munmap(evlist, i);
729
730         zfree(&evlist->mmap);
731 }
732
733 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
734 {
735         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
736         if (cpu_map__empty(evlist->cpus))
737                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
738         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
739         return evlist->mmap != NULL ? 0 : -ENOMEM;
740 }
741
742 struct mmap_params {
743         int prot;
744         int mask;
745 };
746
747 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
748                                struct mmap_params *mp, int fd)
749 {
750         /*
751          * The last one will be done at perf_evlist__mmap_consume(), so that we
752          * make sure we don't prevent tools from consuming every last event in
753          * the ring buffer.
754          *
755          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
756          * anymore, but the last events for it are still in the ring buffer,
757          * waiting to be consumed.
758          *
759          * Tools can chose to ignore this at their own discretion, but the
760          * evlist layer can't just drop it when filtering events in
761          * perf_evlist__filter_pollfd().
762          */
763         evlist->mmap[idx].refcnt = 2;
764         evlist->mmap[idx].prev = 0;
765         evlist->mmap[idx].mask = mp->mask;
766         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
767                                       MAP_SHARED, fd, 0);
768         if (evlist->mmap[idx].base == MAP_FAILED) {
769                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
770                           errno);
771                 evlist->mmap[idx].base = NULL;
772                 return -1;
773         }
774
775         return 0;
776 }
777
778 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
779                                        struct mmap_params *mp, int cpu,
780                                        int thread, int *output)
781 {
782         struct perf_evsel *evsel;
783
784         evlist__for_each(evlist, evsel) {
785                 int fd;
786
787                 if (evsel->system_wide && thread)
788                         continue;
789
790                 fd = FD(evsel, cpu, thread);
791
792                 if (*output == -1) {
793                         *output = fd;
794                         if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
795                                 return -1;
796                 } else {
797                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
798                                 return -1;
799
800                         perf_evlist__mmap_get(evlist, idx);
801                 }
802
803                 if (__perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
804                         perf_evlist__mmap_put(evlist, idx);
805                         return -1;
806                 }
807
808                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
809                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
810                         return -1;
811         }
812
813         return 0;
814 }
815
816 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
817                                      struct mmap_params *mp)
818 {
819         int cpu, thread;
820         int nr_cpus = cpu_map__nr(evlist->cpus);
821         int nr_threads = thread_map__nr(evlist->threads);
822
823         pr_debug2("perf event ring buffer mmapped per cpu\n");
824         for (cpu = 0; cpu < nr_cpus; cpu++) {
825                 int output = -1;
826
827                 for (thread = 0; thread < nr_threads; thread++) {
828                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
829                                                         thread, &output))
830                                 goto out_unmap;
831                 }
832         }
833
834         return 0;
835
836 out_unmap:
837         for (cpu = 0; cpu < nr_cpus; cpu++)
838                 __perf_evlist__munmap(evlist, cpu);
839         return -1;
840 }
841
842 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
843                                         struct mmap_params *mp)
844 {
845         int thread;
846         int nr_threads = thread_map__nr(evlist->threads);
847
848         pr_debug2("perf event ring buffer mmapped per thread\n");
849         for (thread = 0; thread < nr_threads; thread++) {
850                 int output = -1;
851
852                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
853                                                 &output))
854                         goto out_unmap;
855         }
856
857         return 0;
858
859 out_unmap:
860         for (thread = 0; thread < nr_threads; thread++)
861                 __perf_evlist__munmap(evlist, thread);
862         return -1;
863 }
864
865 static size_t perf_evlist__mmap_size(unsigned long pages)
866 {
867         /* 512 kiB: default amount of unprivileged mlocked memory */
868         if (pages == UINT_MAX)
869                 pages = (512 * 1024) / page_size;
870         else if (!is_power_of_2(pages))
871                 return 0;
872
873         return (pages + 1) * page_size;
874 }
875
876 static long parse_pages_arg(const char *str, unsigned long min,
877                             unsigned long max)
878 {
879         unsigned long pages, val;
880         static struct parse_tag tags[] = {
881                 { .tag  = 'B', .mult = 1       },
882                 { .tag  = 'K', .mult = 1 << 10 },
883                 { .tag  = 'M', .mult = 1 << 20 },
884                 { .tag  = 'G', .mult = 1 << 30 },
885                 { .tag  = 0 },
886         };
887
888         if (str == NULL)
889                 return -EINVAL;
890
891         val = parse_tag_value(str, tags);
892         if (val != (unsigned long) -1) {
893                 /* we got file size value */
894                 pages = PERF_ALIGN(val, page_size) / page_size;
895         } else {
896                 /* we got pages count value */
897                 char *eptr;
898                 pages = strtoul(str, &eptr, 10);
899                 if (*eptr != '\0')
900                         return -EINVAL;
901         }
902
903         if (pages == 0 && min == 0) {
904                 /* leave number of pages at 0 */
905         } else if (!is_power_of_2(pages)) {
906                 /* round pages up to next power of 2 */
907                 pages = next_pow2_l(pages);
908                 if (!pages)
909                         return -EINVAL;
910                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
911                         pages * page_size, pages);
912         }
913
914         if (pages > max)
915                 return -EINVAL;
916
917         return pages;
918 }
919
920 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
921                                   int unset __maybe_unused)
922 {
923         unsigned int *mmap_pages = opt->value;
924         unsigned long max = UINT_MAX;
925         long pages;
926
927         if (max > SIZE_MAX / page_size)
928                 max = SIZE_MAX / page_size;
929
930         pages = parse_pages_arg(str, 1, max);
931         if (pages < 0) {
932                 pr_err("Invalid argument for --mmap_pages/-m\n");
933                 return -1;
934         }
935
936         *mmap_pages = pages;
937         return 0;
938 }
939
940 /**
941  * perf_evlist__mmap - Create mmaps to receive events.
942  * @evlist: list of events
943  * @pages: map length in pages
944  * @overwrite: overwrite older events?
945  *
946  * If @overwrite is %false the user needs to signal event consumption using
947  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
948  * automatically.
949  *
950  * Return: %0 on success, negative error code otherwise.
951  */
952 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
953                       bool overwrite)
954 {
955         struct perf_evsel *evsel;
956         const struct cpu_map *cpus = evlist->cpus;
957         const struct thread_map *threads = evlist->threads;
958         struct mmap_params mp = {
959                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
960         };
961
962         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
963                 return -ENOMEM;
964
965         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
966                 return -ENOMEM;
967
968         evlist->overwrite = overwrite;
969         evlist->mmap_len = perf_evlist__mmap_size(pages);
970         pr_debug("mmap size %zuB\n", evlist->mmap_len);
971         mp.mask = evlist->mmap_len - page_size - 1;
972
973         evlist__for_each(evlist, evsel) {
974                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
975                     evsel->sample_id == NULL &&
976                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
977                         return -ENOMEM;
978         }
979
980         if (cpu_map__empty(cpus))
981                 return perf_evlist__mmap_per_thread(evlist, &mp);
982
983         return perf_evlist__mmap_per_cpu(evlist, &mp);
984 }
985
986 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
987 {
988         evlist->threads = thread_map__new_str(target->pid, target->tid,
989                                               target->uid);
990
991         if (evlist->threads == NULL)
992                 return -1;
993
994         if (target__uses_dummy_map(target))
995                 evlist->cpus = cpu_map__dummy_new();
996         else
997                 evlist->cpus = cpu_map__new(target->cpu_list);
998
999         if (evlist->cpus == NULL)
1000                 goto out_delete_threads;
1001
1002         return 0;
1003
1004 out_delete_threads:
1005         thread_map__delete(evlist->threads);
1006         return -1;
1007 }
1008
1009 int perf_evlist__apply_filters(struct perf_evlist *evlist)
1010 {
1011         struct perf_evsel *evsel;
1012         int err = 0;
1013         const int ncpus = cpu_map__nr(evlist->cpus),
1014                   nthreads = thread_map__nr(evlist->threads);
1015
1016         evlist__for_each(evlist, evsel) {
1017                 if (evsel->filter == NULL)
1018                         continue;
1019
1020                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
1021                 if (err)
1022                         break;
1023         }
1024
1025         return err;
1026 }
1027
1028 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1029 {
1030         struct perf_evsel *evsel;
1031         int err = 0;
1032         const int ncpus = cpu_map__nr(evlist->cpus),
1033                   nthreads = thread_map__nr(evlist->threads);
1034
1035         evlist__for_each(evlist, evsel) {
1036                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
1037                 if (err)
1038                         break;
1039         }
1040
1041         return err;
1042 }
1043
1044 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1045 {
1046         struct perf_evsel *pos;
1047
1048         if (evlist->nr_entries == 1)
1049                 return true;
1050
1051         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1052                 return false;
1053
1054         evlist__for_each(evlist, pos) {
1055                 if (pos->id_pos != evlist->id_pos ||
1056                     pos->is_pos != evlist->is_pos)
1057                         return false;
1058         }
1059
1060         return true;
1061 }
1062
1063 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1064 {
1065         struct perf_evsel *evsel;
1066
1067         if (evlist->combined_sample_type)
1068                 return evlist->combined_sample_type;
1069
1070         evlist__for_each(evlist, evsel)
1071                 evlist->combined_sample_type |= evsel->attr.sample_type;
1072
1073         return evlist->combined_sample_type;
1074 }
1075
1076 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1077 {
1078         evlist->combined_sample_type = 0;
1079         return __perf_evlist__combined_sample_type(evlist);
1080 }
1081
1082 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1083 {
1084         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1085         u64 read_format = first->attr.read_format;
1086         u64 sample_type = first->attr.sample_type;
1087
1088         evlist__for_each(evlist, pos) {
1089                 if (read_format != pos->attr.read_format)
1090                         return false;
1091         }
1092
1093         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1094         if ((sample_type & PERF_SAMPLE_READ) &&
1095             !(read_format & PERF_FORMAT_ID)) {
1096                 return false;
1097         }
1098
1099         return true;
1100 }
1101
1102 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1103 {
1104         struct perf_evsel *first = perf_evlist__first(evlist);
1105         return first->attr.read_format;
1106 }
1107
1108 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1109 {
1110         struct perf_evsel *first = perf_evlist__first(evlist);
1111         struct perf_sample *data;
1112         u64 sample_type;
1113         u16 size = 0;
1114
1115         if (!first->attr.sample_id_all)
1116                 goto out;
1117
1118         sample_type = first->attr.sample_type;
1119
1120         if (sample_type & PERF_SAMPLE_TID)
1121                 size += sizeof(data->tid) * 2;
1122
1123        if (sample_type & PERF_SAMPLE_TIME)
1124                 size += sizeof(data->time);
1125
1126         if (sample_type & PERF_SAMPLE_ID)
1127                 size += sizeof(data->id);
1128
1129         if (sample_type & PERF_SAMPLE_STREAM_ID)
1130                 size += sizeof(data->stream_id);
1131
1132         if (sample_type & PERF_SAMPLE_CPU)
1133                 size += sizeof(data->cpu) * 2;
1134
1135         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1136                 size += sizeof(data->id);
1137 out:
1138         return size;
1139 }
1140
1141 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1142 {
1143         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1144
1145         evlist__for_each_continue(evlist, pos) {
1146                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1147                         return false;
1148         }
1149
1150         return true;
1151 }
1152
1153 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1154 {
1155         struct perf_evsel *first = perf_evlist__first(evlist);
1156         return first->attr.sample_id_all;
1157 }
1158
1159 void perf_evlist__set_selected(struct perf_evlist *evlist,
1160                                struct perf_evsel *evsel)
1161 {
1162         evlist->selected = evsel;
1163 }
1164
1165 void perf_evlist__close(struct perf_evlist *evlist)
1166 {
1167         struct perf_evsel *evsel;
1168         int ncpus = cpu_map__nr(evlist->cpus);
1169         int nthreads = thread_map__nr(evlist->threads);
1170         int n;
1171
1172         evlist__for_each_reverse(evlist, evsel) {
1173                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1174                 perf_evsel__close(evsel, n, nthreads);
1175         }
1176 }
1177
1178 int perf_evlist__open(struct perf_evlist *evlist)
1179 {
1180         struct perf_evsel *evsel;
1181         int err;
1182
1183         perf_evlist__update_id_pos(evlist);
1184
1185         evlist__for_each(evlist, evsel) {
1186                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1187                 if (err < 0)
1188                         goto out_err;
1189         }
1190
1191         return 0;
1192 out_err:
1193         perf_evlist__close(evlist);
1194         errno = -err;
1195         return err;
1196 }
1197
1198 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1199                                   const char *argv[], bool pipe_output,
1200                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1201 {
1202         int child_ready_pipe[2], go_pipe[2];
1203         char bf;
1204
1205         if (pipe(child_ready_pipe) < 0) {
1206                 perror("failed to create 'ready' pipe");
1207                 return -1;
1208         }
1209
1210         if (pipe(go_pipe) < 0) {
1211                 perror("failed to create 'go' pipe");
1212                 goto out_close_ready_pipe;
1213         }
1214
1215         evlist->workload.pid = fork();
1216         if (evlist->workload.pid < 0) {
1217                 perror("failed to fork");
1218                 goto out_close_pipes;
1219         }
1220
1221         if (!evlist->workload.pid) {
1222                 int ret;
1223
1224                 if (pipe_output)
1225                         dup2(2, 1);
1226
1227                 signal(SIGTERM, SIG_DFL);
1228
1229                 close(child_ready_pipe[0]);
1230                 close(go_pipe[1]);
1231                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1232
1233                 /*
1234                  * Tell the parent we're ready to go
1235                  */
1236                 close(child_ready_pipe[1]);
1237
1238                 /*
1239                  * Wait until the parent tells us to go.
1240                  */
1241                 ret = read(go_pipe[0], &bf, 1);
1242                 /*
1243                  * The parent will ask for the execvp() to be performed by
1244                  * writing exactly one byte, in workload.cork_fd, usually via
1245                  * perf_evlist__start_workload().
1246                  *
1247                  * For cancelling the workload without actuallin running it,
1248                  * the parent will just close workload.cork_fd, without writing
1249                  * anything, i.e. read will return zero and we just exit()
1250                  * here.
1251                  */
1252                 if (ret != 1) {
1253                         if (ret == -1)
1254                                 perror("unable to read pipe");
1255                         exit(ret);
1256                 }
1257
1258                 execvp(argv[0], (char **)argv);
1259
1260                 if (exec_error) {
1261                         union sigval val;
1262
1263                         val.sival_int = errno;
1264                         if (sigqueue(getppid(), SIGUSR1, val))
1265                                 perror(argv[0]);
1266                 } else
1267                         perror(argv[0]);
1268                 exit(-1);
1269         }
1270
1271         if (exec_error) {
1272                 struct sigaction act = {
1273                         .sa_flags     = SA_SIGINFO,
1274                         .sa_sigaction = exec_error,
1275                 };
1276                 sigaction(SIGUSR1, &act, NULL);
1277         }
1278
1279         if (target__none(target))
1280                 evlist->threads->map[0] = evlist->workload.pid;
1281
1282         close(child_ready_pipe[1]);
1283         close(go_pipe[0]);
1284         /*
1285          * wait for child to settle
1286          */
1287         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1288                 perror("unable to read pipe");
1289                 goto out_close_pipes;
1290         }
1291
1292         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1293         evlist->workload.cork_fd = go_pipe[1];
1294         close(child_ready_pipe[0]);
1295         return 0;
1296
1297 out_close_pipes:
1298         close(go_pipe[0]);
1299         close(go_pipe[1]);
1300 out_close_ready_pipe:
1301         close(child_ready_pipe[0]);
1302         close(child_ready_pipe[1]);
1303         return -1;
1304 }
1305
1306 int perf_evlist__start_workload(struct perf_evlist *evlist)
1307 {
1308         if (evlist->workload.cork_fd > 0) {
1309                 char bf = 0;
1310                 int ret;
1311                 /*
1312                  * Remove the cork, let it rip!
1313                  */
1314                 ret = write(evlist->workload.cork_fd, &bf, 1);
1315                 if (ret < 0)
1316                         perror("enable to write to pipe");
1317
1318                 close(evlist->workload.cork_fd);
1319                 return ret;
1320         }
1321
1322         return 0;
1323 }
1324
1325 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1326                               struct perf_sample *sample)
1327 {
1328         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1329
1330         if (!evsel)
1331                 return -EFAULT;
1332         return perf_evsel__parse_sample(evsel, event, sample);
1333 }
1334
1335 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1336 {
1337         struct perf_evsel *evsel;
1338         size_t printed = 0;
1339
1340         evlist__for_each(evlist, evsel) {
1341                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1342                                    perf_evsel__name(evsel));
1343         }
1344
1345         return printed + fprintf(fp, "\n");
1346 }
1347
1348 int perf_evlist__strerror_tp(struct perf_evlist *evlist __maybe_unused,
1349                              int err, char *buf, size_t size)
1350 {
1351         char sbuf[128];
1352
1353         switch (err) {
1354         case ENOENT:
1355                 scnprintf(buf, size, "%s",
1356                           "Error:\tUnable to find debugfs\n"
1357                           "Hint:\tWas your kernel was compiled with debugfs support?\n"
1358                           "Hint:\tIs the debugfs filesystem mounted?\n"
1359                           "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1360                 break;
1361         case EACCES:
1362                 scnprintf(buf, size,
1363                           "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1364                           "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1365                           debugfs_mountpoint, debugfs_mountpoint);
1366                 break;
1367         default:
1368                 scnprintf(buf, size, "%s", strerror_r(err, sbuf, sizeof(sbuf)));
1369                 break;
1370         }
1371
1372         return 0;
1373 }
1374
1375 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1376                                int err, char *buf, size_t size)
1377 {
1378         int printed, value;
1379         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1380
1381         switch (err) {
1382         case EACCES:
1383         case EPERM:
1384                 printed = scnprintf(buf, size,
1385                                     "Error:\t%s.\n"
1386                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1387
1388                 value = perf_event_paranoid();
1389
1390                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1391
1392                 if (value >= 2) {
1393                         printed += scnprintf(buf + printed, size - printed,
1394                                              "For your workloads it needs to be <= 1\nHint:\t");
1395                 }
1396                 printed += scnprintf(buf + printed, size - printed,
1397                                      "For system wide tracing it needs to be set to -1.\n");
1398
1399                 printed += scnprintf(buf + printed, size - printed,
1400                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1401                                     "Hint:\tThe current value is %d.", value);
1402                 break;
1403         default:
1404                 scnprintf(buf, size, "%s", emsg);
1405                 break;
1406         }
1407
1408         return 0;
1409 }
1410
1411 void perf_evlist__to_front(struct perf_evlist *evlist,
1412                            struct perf_evsel *move_evsel)
1413 {
1414         struct perf_evsel *evsel, *n;
1415         LIST_HEAD(move);
1416
1417         if (move_evsel == perf_evlist__first(evlist))
1418                 return;
1419
1420         evlist__for_each_safe(evlist, n, evsel) {
1421                 if (evsel->leader == move_evsel->leader)
1422                         list_move_tail(&evsel->node, &move);
1423         }
1424
1425         list_splice(&move, &evlist->entries);
1426 }
1427
1428 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1429                                      struct perf_evsel *tracking_evsel)
1430 {
1431         struct perf_evsel *evsel;
1432
1433         if (tracking_evsel->tracking)
1434                 return;
1435
1436         evlist__for_each(evlist, evsel) {
1437                 if (evsel != tracking_evsel)
1438                         evsel->tracking = false;
1439         }
1440
1441         tracking_evsel->tracking = true;
1442 }