2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
7 * Released under the GPL v2. (and only v2, not any later version)
10 #include <api/fs/fs.h>
13 #include "thread_map.h"
21 #include "parse-events.h"
22 #include <subcmd/parse-options.h>
26 #include <linux/bitops.h>
27 #include <linux/hash.h>
28 #include <linux/log2.h>
29 #include <linux/err.h>
31 static void perf_mmap__munmap(struct perf_mmap
*map
);
32 static void perf_mmap__put(struct perf_mmap
*map
);
34 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
35 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
37 void perf_evlist__init(struct perf_evlist
*evlist
, struct cpu_map
*cpus
,
38 struct thread_map
*threads
)
42 for (i
= 0; i
< PERF_EVLIST__HLIST_SIZE
; ++i
)
43 INIT_HLIST_HEAD(&evlist
->heads
[i
]);
44 INIT_LIST_HEAD(&evlist
->entries
);
45 perf_evlist__set_maps(evlist
, cpus
, threads
);
46 fdarray__init(&evlist
->pollfd
, 64);
47 evlist
->workload
.pid
= -1;
48 evlist
->bkw_mmap_state
= BKW_MMAP_NOTREADY
;
51 struct perf_evlist
*perf_evlist__new(void)
53 struct perf_evlist
*evlist
= zalloc(sizeof(*evlist
));
56 perf_evlist__init(evlist
, NULL
, NULL
);
61 struct perf_evlist
*perf_evlist__new_default(void)
63 struct perf_evlist
*evlist
= perf_evlist__new();
65 if (evlist
&& perf_evlist__add_default(evlist
)) {
66 perf_evlist__delete(evlist
);
73 struct perf_evlist
*perf_evlist__new_dummy(void)
75 struct perf_evlist
*evlist
= perf_evlist__new();
77 if (evlist
&& perf_evlist__add_dummy(evlist
)) {
78 perf_evlist__delete(evlist
);
86 * perf_evlist__set_id_pos - set the positions of event ids.
87 * @evlist: selected event list
89 * Events with compatible sample types all have the same id_pos
90 * and is_pos. For convenience, put a copy on evlist.
92 void perf_evlist__set_id_pos(struct perf_evlist
*evlist
)
94 struct perf_evsel
*first
= perf_evlist__first(evlist
);
96 evlist
->id_pos
= first
->id_pos
;
97 evlist
->is_pos
= first
->is_pos
;
100 static void perf_evlist__update_id_pos(struct perf_evlist
*evlist
)
102 struct perf_evsel
*evsel
;
104 evlist__for_each_entry(evlist
, evsel
)
105 perf_evsel__calc_id_pos(evsel
);
107 perf_evlist__set_id_pos(evlist
);
110 static void perf_evlist__purge(struct perf_evlist
*evlist
)
112 struct perf_evsel
*pos
, *n
;
114 evlist__for_each_entry_safe(evlist
, n
, pos
) {
115 list_del_init(&pos
->node
);
117 perf_evsel__delete(pos
);
120 evlist
->nr_entries
= 0;
123 void perf_evlist__exit(struct perf_evlist
*evlist
)
125 zfree(&evlist
->mmap
);
126 zfree(&evlist
->backward_mmap
);
127 fdarray__exit(&evlist
->pollfd
);
130 void perf_evlist__delete(struct perf_evlist
*evlist
)
135 perf_evlist__munmap(evlist
);
136 perf_evlist__close(evlist
);
137 cpu_map__put(evlist
->cpus
);
138 thread_map__put(evlist
->threads
);
140 evlist
->threads
= NULL
;
141 perf_evlist__purge(evlist
);
142 perf_evlist__exit(evlist
);
146 static void __perf_evlist__propagate_maps(struct perf_evlist
*evlist
,
147 struct perf_evsel
*evsel
)
150 * We already have cpus for evsel (via PMU sysfs) so
151 * keep it, if there's no target cpu list defined.
153 if (!evsel
->own_cpus
|| evlist
->has_user_cpus
) {
154 cpu_map__put(evsel
->cpus
);
155 evsel
->cpus
= cpu_map__get(evlist
->cpus
);
156 } else if (evsel
->cpus
!= evsel
->own_cpus
) {
157 cpu_map__put(evsel
->cpus
);
158 evsel
->cpus
= cpu_map__get(evsel
->own_cpus
);
161 thread_map__put(evsel
->threads
);
162 evsel
->threads
= thread_map__get(evlist
->threads
);
165 static void perf_evlist__propagate_maps(struct perf_evlist
*evlist
)
167 struct perf_evsel
*evsel
;
169 evlist__for_each_entry(evlist
, evsel
)
170 __perf_evlist__propagate_maps(evlist
, evsel
);
173 void perf_evlist__add(struct perf_evlist
*evlist
, struct perf_evsel
*entry
)
175 entry
->evlist
= evlist
;
176 list_add_tail(&entry
->node
, &evlist
->entries
);
177 entry
->idx
= evlist
->nr_entries
;
178 entry
->tracking
= !entry
->idx
;
180 if (!evlist
->nr_entries
++)
181 perf_evlist__set_id_pos(evlist
);
183 __perf_evlist__propagate_maps(evlist
, entry
);
186 void perf_evlist__remove(struct perf_evlist
*evlist
, struct perf_evsel
*evsel
)
188 evsel
->evlist
= NULL
;
189 list_del_init(&evsel
->node
);
190 evlist
->nr_entries
-= 1;
193 void perf_evlist__splice_list_tail(struct perf_evlist
*evlist
,
194 struct list_head
*list
)
196 struct perf_evsel
*evsel
, *temp
;
198 __evlist__for_each_entry_safe(list
, temp
, evsel
) {
199 list_del_init(&evsel
->node
);
200 perf_evlist__add(evlist
, evsel
);
204 void __perf_evlist__set_leader(struct list_head
*list
)
206 struct perf_evsel
*evsel
, *leader
;
208 leader
= list_entry(list
->next
, struct perf_evsel
, node
);
209 evsel
= list_entry(list
->prev
, struct perf_evsel
, node
);
211 leader
->nr_members
= evsel
->idx
- leader
->idx
+ 1;
213 __evlist__for_each_entry(list
, evsel
) {
214 evsel
->leader
= leader
;
218 void perf_evlist__set_leader(struct perf_evlist
*evlist
)
220 if (evlist
->nr_entries
) {
221 evlist
->nr_groups
= evlist
->nr_entries
> 1 ? 1 : 0;
222 __perf_evlist__set_leader(&evlist
->entries
);
226 void perf_event_attr__set_max_precise_ip(struct perf_event_attr
*attr
)
228 attr
->precise_ip
= 3;
230 while (attr
->precise_ip
!= 0) {
231 int fd
= sys_perf_event_open(attr
, 0, -1, -1, 0);
240 int perf_evlist__add_default(struct perf_evlist
*evlist
)
242 struct perf_evsel
*evsel
= perf_evsel__new_cycles();
247 perf_evlist__add(evlist
, evsel
);
251 int perf_evlist__add_dummy(struct perf_evlist
*evlist
)
253 struct perf_event_attr attr
= {
254 .type
= PERF_TYPE_SOFTWARE
,
255 .config
= PERF_COUNT_SW_DUMMY
,
256 .size
= sizeof(attr
), /* to capture ABI version */
258 struct perf_evsel
*evsel
= perf_evsel__new(&attr
);
263 perf_evlist__add(evlist
, evsel
);
267 static int perf_evlist__add_attrs(struct perf_evlist
*evlist
,
268 struct perf_event_attr
*attrs
, size_t nr_attrs
)
270 struct perf_evsel
*evsel
, *n
;
274 for (i
= 0; i
< nr_attrs
; i
++) {
275 evsel
= perf_evsel__new_idx(attrs
+ i
, evlist
->nr_entries
+ i
);
277 goto out_delete_partial_list
;
278 list_add_tail(&evsel
->node
, &head
);
281 perf_evlist__splice_list_tail(evlist
, &head
);
285 out_delete_partial_list
:
286 __evlist__for_each_entry_safe(&head
, n
, evsel
)
287 perf_evsel__delete(evsel
);
291 int __perf_evlist__add_default_attrs(struct perf_evlist
*evlist
,
292 struct perf_event_attr
*attrs
, size_t nr_attrs
)
296 for (i
= 0; i
< nr_attrs
; i
++)
297 event_attr_init(attrs
+ i
);
299 return perf_evlist__add_attrs(evlist
, attrs
, nr_attrs
);
303 perf_evlist__find_tracepoint_by_id(struct perf_evlist
*evlist
, int id
)
305 struct perf_evsel
*evsel
;
307 evlist__for_each_entry(evlist
, evsel
) {
308 if (evsel
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
309 (int)evsel
->attr
.config
== id
)
317 perf_evlist__find_tracepoint_by_name(struct perf_evlist
*evlist
,
320 struct perf_evsel
*evsel
;
322 evlist__for_each_entry(evlist
, evsel
) {
323 if ((evsel
->attr
.type
== PERF_TYPE_TRACEPOINT
) &&
324 (strcmp(evsel
->name
, name
) == 0))
331 int perf_evlist__add_newtp(struct perf_evlist
*evlist
,
332 const char *sys
, const char *name
, void *handler
)
334 struct perf_evsel
*evsel
= perf_evsel__newtp(sys
, name
);
339 evsel
->handler
= handler
;
340 perf_evlist__add(evlist
, evsel
);
344 static int perf_evlist__nr_threads(struct perf_evlist
*evlist
,
345 struct perf_evsel
*evsel
)
347 if (evsel
->system_wide
)
350 return thread_map__nr(evlist
->threads
);
353 void perf_evlist__disable(struct perf_evlist
*evlist
)
355 struct perf_evsel
*pos
;
357 evlist__for_each_entry(evlist
, pos
) {
358 if (!perf_evsel__is_group_leader(pos
) || !pos
->fd
)
360 perf_evsel__disable(pos
);
363 evlist
->enabled
= false;
366 void perf_evlist__enable(struct perf_evlist
*evlist
)
368 struct perf_evsel
*pos
;
370 evlist__for_each_entry(evlist
, pos
) {
371 if (!perf_evsel__is_group_leader(pos
) || !pos
->fd
)
373 perf_evsel__enable(pos
);
376 evlist
->enabled
= true;
379 void perf_evlist__toggle_enable(struct perf_evlist
*evlist
)
381 (evlist
->enabled
? perf_evlist__disable
: perf_evlist__enable
)(evlist
);
384 static int perf_evlist__enable_event_cpu(struct perf_evlist
*evlist
,
385 struct perf_evsel
*evsel
, int cpu
)
388 int nr_threads
= perf_evlist__nr_threads(evlist
, evsel
);
393 for (thread
= 0; thread
< nr_threads
; thread
++) {
394 int err
= ioctl(FD(evsel
, cpu
, thread
), PERF_EVENT_IOC_ENABLE
, 0);
401 static int perf_evlist__enable_event_thread(struct perf_evlist
*evlist
,
402 struct perf_evsel
*evsel
,
406 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
411 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
412 int err
= ioctl(FD(evsel
, cpu
, thread
), PERF_EVENT_IOC_ENABLE
, 0);
419 int perf_evlist__enable_event_idx(struct perf_evlist
*evlist
,
420 struct perf_evsel
*evsel
, int idx
)
422 bool per_cpu_mmaps
= !cpu_map__empty(evlist
->cpus
);
425 return perf_evlist__enable_event_cpu(evlist
, evsel
, idx
);
427 return perf_evlist__enable_event_thread(evlist
, evsel
, idx
);
430 int perf_evlist__alloc_pollfd(struct perf_evlist
*evlist
)
432 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
433 int nr_threads
= thread_map__nr(evlist
->threads
);
435 struct perf_evsel
*evsel
;
437 evlist__for_each_entry(evlist
, evsel
) {
438 if (evsel
->system_wide
)
441 nfds
+= nr_cpus
* nr_threads
;
444 if (fdarray__available_entries(&evlist
->pollfd
) < nfds
&&
445 fdarray__grow(&evlist
->pollfd
, nfds
) < 0)
451 static int __perf_evlist__add_pollfd(struct perf_evlist
*evlist
, int fd
,
452 struct perf_mmap
*map
, short revent
)
454 int pos
= fdarray__add(&evlist
->pollfd
, fd
, revent
| POLLERR
| POLLHUP
);
456 * Save the idx so that when we filter out fds POLLHUP'ed we can
457 * close the associated evlist->mmap[] entry.
460 evlist
->pollfd
.priv
[pos
].ptr
= map
;
462 fcntl(fd
, F_SETFL
, O_NONBLOCK
);
468 int perf_evlist__add_pollfd(struct perf_evlist
*evlist
, int fd
)
470 return __perf_evlist__add_pollfd(evlist
, fd
, NULL
, POLLIN
);
473 static void perf_evlist__munmap_filtered(struct fdarray
*fda
, int fd
,
474 void *arg __maybe_unused
)
476 struct perf_mmap
*map
= fda
->priv
[fd
].ptr
;
482 int perf_evlist__filter_pollfd(struct perf_evlist
*evlist
, short revents_and_mask
)
484 return fdarray__filter(&evlist
->pollfd
, revents_and_mask
,
485 perf_evlist__munmap_filtered
, NULL
);
488 int perf_evlist__poll(struct perf_evlist
*evlist
, int timeout
)
490 return fdarray__poll(&evlist
->pollfd
, timeout
);
493 static void perf_evlist__id_hash(struct perf_evlist
*evlist
,
494 struct perf_evsel
*evsel
,
495 int cpu
, int thread
, u64 id
)
498 struct perf_sample_id
*sid
= SID(evsel
, cpu
, thread
);
502 hash
= hash_64(sid
->id
, PERF_EVLIST__HLIST_BITS
);
503 hlist_add_head(&sid
->node
, &evlist
->heads
[hash
]);
506 void perf_evlist__id_add(struct perf_evlist
*evlist
, struct perf_evsel
*evsel
,
507 int cpu
, int thread
, u64 id
)
509 perf_evlist__id_hash(evlist
, evsel
, cpu
, thread
, id
);
510 evsel
->id
[evsel
->ids
++] = id
;
513 int perf_evlist__id_add_fd(struct perf_evlist
*evlist
,
514 struct perf_evsel
*evsel
,
515 int cpu
, int thread
, int fd
)
517 u64 read_data
[4] = { 0, };
518 int id_idx
= 1; /* The first entry is the counter value */
522 ret
= ioctl(fd
, PERF_EVENT_IOC_ID
, &id
);
529 /* Legacy way to get event id.. All hail to old kernels! */
532 * This way does not work with group format read, so bail
535 if (perf_evlist__read_format(evlist
) & PERF_FORMAT_GROUP
)
538 if (!(evsel
->attr
.read_format
& PERF_FORMAT_ID
) ||
539 read(fd
, &read_data
, sizeof(read_data
)) == -1)
542 if (evsel
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
544 if (evsel
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
547 id
= read_data
[id_idx
];
550 perf_evlist__id_add(evlist
, evsel
, cpu
, thread
, id
);
554 static void perf_evlist__set_sid_idx(struct perf_evlist
*evlist
,
555 struct perf_evsel
*evsel
, int idx
, int cpu
,
558 struct perf_sample_id
*sid
= SID(evsel
, cpu
, thread
);
560 if (evlist
->cpus
&& cpu
>= 0)
561 sid
->cpu
= evlist
->cpus
->map
[cpu
];
564 if (!evsel
->system_wide
&& evlist
->threads
&& thread
>= 0)
565 sid
->tid
= thread_map__pid(evlist
->threads
, thread
);
570 struct perf_sample_id
*perf_evlist__id2sid(struct perf_evlist
*evlist
, u64 id
)
572 struct hlist_head
*head
;
573 struct perf_sample_id
*sid
;
576 hash
= hash_64(id
, PERF_EVLIST__HLIST_BITS
);
577 head
= &evlist
->heads
[hash
];
579 hlist_for_each_entry(sid
, head
, node
)
586 struct perf_evsel
*perf_evlist__id2evsel(struct perf_evlist
*evlist
, u64 id
)
588 struct perf_sample_id
*sid
;
590 if (evlist
->nr_entries
== 1 || !id
)
591 return perf_evlist__first(evlist
);
593 sid
= perf_evlist__id2sid(evlist
, id
);
597 if (!perf_evlist__sample_id_all(evlist
))
598 return perf_evlist__first(evlist
);
603 struct perf_evsel
*perf_evlist__id2evsel_strict(struct perf_evlist
*evlist
,
606 struct perf_sample_id
*sid
;
611 sid
= perf_evlist__id2sid(evlist
, id
);
618 static int perf_evlist__event2id(struct perf_evlist
*evlist
,
619 union perf_event
*event
, u64
*id
)
621 const u64
*array
= event
->sample
.array
;
624 n
= (event
->header
.size
- sizeof(event
->header
)) >> 3;
626 if (event
->header
.type
== PERF_RECORD_SAMPLE
) {
627 if (evlist
->id_pos
>= n
)
629 *id
= array
[evlist
->id_pos
];
631 if (evlist
->is_pos
> n
)
639 struct perf_evsel
*perf_evlist__event2evsel(struct perf_evlist
*evlist
,
640 union perf_event
*event
)
642 struct perf_evsel
*first
= perf_evlist__first(evlist
);
643 struct hlist_head
*head
;
644 struct perf_sample_id
*sid
;
648 if (evlist
->nr_entries
== 1)
651 if (!first
->attr
.sample_id_all
&&
652 event
->header
.type
!= PERF_RECORD_SAMPLE
)
655 if (perf_evlist__event2id(evlist
, event
, &id
))
658 /* Synthesized events have an id of zero */
662 hash
= hash_64(id
, PERF_EVLIST__HLIST_BITS
);
663 head
= &evlist
->heads
[hash
];
665 hlist_for_each_entry(sid
, head
, node
) {
672 static int perf_evlist__set_paused(struct perf_evlist
*evlist
, bool value
)
676 if (!evlist
->backward_mmap
)
679 for (i
= 0; i
< evlist
->nr_mmaps
; i
++) {
680 int fd
= evlist
->backward_mmap
[i
].fd
;
685 err
= ioctl(fd
, PERF_EVENT_IOC_PAUSE_OUTPUT
, value
? 1 : 0);
692 static int perf_evlist__pause(struct perf_evlist
*evlist
)
694 return perf_evlist__set_paused(evlist
, true);
697 static int perf_evlist__resume(struct perf_evlist
*evlist
)
699 return perf_evlist__set_paused(evlist
, false);
702 /* When check_messup is true, 'end' must points to a good entry */
703 static union perf_event
*
704 perf_mmap__read(struct perf_mmap
*md
, bool check_messup
, u64 start
,
707 unsigned char *data
= md
->base
+ page_size
;
708 union perf_event
*event
= NULL
;
709 int diff
= end
- start
;
713 * If we're further behind than half the buffer, there's a chance
714 * the writer will bite our tail and mess up the samples under us.
716 * If we somehow ended up ahead of the 'end', we got messed up.
718 * In either case, truncate and restart at 'end'.
720 if (diff
> md
->mask
/ 2 || diff
< 0) {
721 fprintf(stderr
, "WARNING: failed to keep up with mmap data.\n");
724 * 'end' points to a known good entry, start there.
731 if (diff
>= (int)sizeof(event
->header
)) {
734 event
= (union perf_event
*)&data
[start
& md
->mask
];
735 size
= event
->header
.size
;
737 if (size
< sizeof(event
->header
) || diff
< (int)size
) {
743 * Event straddles the mmap boundary -- header should always
744 * be inside due to u64 alignment of output.
746 if ((start
& md
->mask
) + size
!= ((start
+ size
) & md
->mask
)) {
747 unsigned int offset
= start
;
748 unsigned int len
= min(sizeof(*event
), size
), cpy
;
749 void *dst
= md
->event_copy
;
752 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
753 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
759 event
= (union perf_event
*) md
->event_copy
;
772 union perf_event
*perf_mmap__read_forward(struct perf_mmap
*md
, bool check_messup
)
778 * Check if event was unmapped due to a POLLHUP/POLLERR.
780 if (!atomic_read(&md
->refcnt
))
783 head
= perf_mmap__read_head(md
);
785 return perf_mmap__read(md
, check_messup
, old
, head
, &md
->prev
);
789 perf_mmap__read_backward(struct perf_mmap
*md
)
792 u64 start
= md
->prev
;
795 * Check if event was unmapped due to a POLLHUP/POLLERR.
797 if (!atomic_read(&md
->refcnt
))
800 head
= perf_mmap__read_head(md
);
805 * 'head' pointer starts from 0. Kernel minus sizeof(record) form
806 * it each time when kernel writes to it, so in fact 'head' is
807 * negative. 'end' pointer is made manually by adding the size of
808 * the ring buffer to 'head' pointer, means the validate data can
809 * read is the whole ring buffer. If 'end' is positive, the ring
810 * buffer has not fully filled, so we must adjust 'end' to 0.
812 * However, since both 'head' and 'end' is unsigned, we can't
813 * simply compare 'end' against 0. Here we compare '-head' and
814 * the size of the ring buffer, where -head is the number of bytes
815 * kernel write to the ring buffer.
817 if (-head
< (u64
)(md
->mask
+ 1))
820 end
= head
+ md
->mask
+ 1;
822 return perf_mmap__read(md
, false, start
, end
, &md
->prev
);
825 union perf_event
*perf_evlist__mmap_read_forward(struct perf_evlist
*evlist
, int idx
)
827 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
830 * Check messup is required for forward overwritable ring buffer:
831 * memory pointed by md->prev can be overwritten in this case.
832 * No need for read-write ring buffer: kernel stop outputting when
833 * it hit md->prev (perf_mmap__consume()).
835 return perf_mmap__read_forward(md
, evlist
->overwrite
);
838 union perf_event
*perf_evlist__mmap_read_backward(struct perf_evlist
*evlist
, int idx
)
840 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
843 * No need to check messup for backward ring buffer:
844 * We can always read arbitrary long data from a backward
845 * ring buffer unless we forget to pause it before reading.
847 return perf_mmap__read_backward(md
);
850 union perf_event
*perf_evlist__mmap_read(struct perf_evlist
*evlist
, int idx
)
852 return perf_evlist__mmap_read_forward(evlist
, idx
);
855 void perf_mmap__read_catchup(struct perf_mmap
*md
)
859 if (!atomic_read(&md
->refcnt
))
862 head
= perf_mmap__read_head(md
);
866 void perf_evlist__mmap_read_catchup(struct perf_evlist
*evlist
, int idx
)
868 perf_mmap__read_catchup(&evlist
->mmap
[idx
]);
871 static bool perf_mmap__empty(struct perf_mmap
*md
)
873 return perf_mmap__read_head(md
) == md
->prev
&& !md
->auxtrace_mmap
.base
;
876 static void perf_mmap__get(struct perf_mmap
*map
)
878 atomic_inc(&map
->refcnt
);
881 static void perf_mmap__put(struct perf_mmap
*md
)
883 BUG_ON(md
->base
&& atomic_read(&md
->refcnt
) == 0);
885 if (atomic_dec_and_test(&md
->refcnt
))
886 perf_mmap__munmap(md
);
889 void perf_mmap__consume(struct perf_mmap
*md
, bool overwrite
)
894 perf_mmap__write_tail(md
, old
);
897 if (atomic_read(&md
->refcnt
) == 1 && perf_mmap__empty(md
))
901 void perf_evlist__mmap_consume(struct perf_evlist
*evlist
, int idx
)
903 perf_mmap__consume(&evlist
->mmap
[idx
], evlist
->overwrite
);
906 int __weak
auxtrace_mmap__mmap(struct auxtrace_mmap
*mm __maybe_unused
,
907 struct auxtrace_mmap_params
*mp __maybe_unused
,
908 void *userpg __maybe_unused
,
909 int fd __maybe_unused
)
914 void __weak
auxtrace_mmap__munmap(struct auxtrace_mmap
*mm __maybe_unused
)
918 void __weak
auxtrace_mmap_params__init(
919 struct auxtrace_mmap_params
*mp __maybe_unused
,
920 off_t auxtrace_offset __maybe_unused
,
921 unsigned int auxtrace_pages __maybe_unused
,
922 bool auxtrace_overwrite __maybe_unused
)
926 void __weak
auxtrace_mmap_params__set_idx(
927 struct auxtrace_mmap_params
*mp __maybe_unused
,
928 struct perf_evlist
*evlist __maybe_unused
,
929 int idx __maybe_unused
,
930 bool per_cpu __maybe_unused
)
934 static void perf_mmap__munmap(struct perf_mmap
*map
)
936 if (map
->base
!= NULL
) {
937 munmap(map
->base
, perf_mmap__mmap_len(map
));
940 atomic_set(&map
->refcnt
, 0);
942 auxtrace_mmap__munmap(&map
->auxtrace_mmap
);
945 static void perf_evlist__munmap_nofree(struct perf_evlist
*evlist
)
950 for (i
= 0; i
< evlist
->nr_mmaps
; i
++)
951 perf_mmap__munmap(&evlist
->mmap
[i
]);
953 if (evlist
->backward_mmap
)
954 for (i
= 0; i
< evlist
->nr_mmaps
; i
++)
955 perf_mmap__munmap(&evlist
->backward_mmap
[i
]);
958 void perf_evlist__munmap(struct perf_evlist
*evlist
)
960 perf_evlist__munmap_nofree(evlist
);
961 zfree(&evlist
->mmap
);
962 zfree(&evlist
->backward_mmap
);
965 static struct perf_mmap
*perf_evlist__alloc_mmap(struct perf_evlist
*evlist
)
968 struct perf_mmap
*map
;
970 evlist
->nr_mmaps
= cpu_map__nr(evlist
->cpus
);
971 if (cpu_map__empty(evlist
->cpus
))
972 evlist
->nr_mmaps
= thread_map__nr(evlist
->threads
);
973 map
= zalloc(evlist
->nr_mmaps
* sizeof(struct perf_mmap
));
977 for (i
= 0; i
< evlist
->nr_mmaps
; i
++)
985 struct auxtrace_mmap_params auxtrace_mp
;
988 static int perf_mmap__mmap(struct perf_mmap
*map
,
989 struct mmap_params
*mp
, int fd
)
992 * The last one will be done at perf_evlist__mmap_consume(), so that we
993 * make sure we don't prevent tools from consuming every last event in
996 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
997 * anymore, but the last events for it are still in the ring buffer,
998 * waiting to be consumed.
1000 * Tools can chose to ignore this at their own discretion, but the
1001 * evlist layer can't just drop it when filtering events in
1002 * perf_evlist__filter_pollfd().
1004 atomic_set(&map
->refcnt
, 2);
1006 map
->mask
= mp
->mask
;
1007 map
->base
= mmap(NULL
, perf_mmap__mmap_len(map
), mp
->prot
,
1009 if (map
->base
== MAP_FAILED
) {
1010 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
1017 if (auxtrace_mmap__mmap(&map
->auxtrace_mmap
,
1018 &mp
->auxtrace_mp
, map
->base
, fd
))
1025 perf_evlist__should_poll(struct perf_evlist
*evlist __maybe_unused
,
1026 struct perf_evsel
*evsel
)
1028 if (evsel
->attr
.write_backward
)
1033 static int perf_evlist__mmap_per_evsel(struct perf_evlist
*evlist
, int idx
,
1034 struct mmap_params
*mp
, int cpu_idx
,
1035 int thread
, int *_output
, int *_output_backward
)
1037 struct perf_evsel
*evsel
;
1039 int evlist_cpu
= cpu_map__cpu(evlist
->cpus
, cpu_idx
);
1041 evlist__for_each_entry(evlist
, evsel
) {
1042 struct perf_mmap
*maps
= evlist
->mmap
;
1043 int *output
= _output
;
1047 if (evsel
->attr
.write_backward
) {
1048 output
= _output_backward
;
1049 maps
= evlist
->backward_mmap
;
1052 maps
= perf_evlist__alloc_mmap(evlist
);
1055 evlist
->backward_mmap
= maps
;
1056 if (evlist
->bkw_mmap_state
== BKW_MMAP_NOTREADY
)
1057 perf_evlist__toggle_bkw_mmap(evlist
, BKW_MMAP_RUNNING
);
1061 if (evsel
->system_wide
&& thread
)
1064 cpu
= cpu_map__idx(evsel
->cpus
, evlist_cpu
);
1068 fd
= FD(evsel
, cpu
, thread
);
1070 if (*output
== -1) {
1073 if (perf_mmap__mmap(&maps
[idx
], mp
, *output
) < 0)
1076 if (ioctl(fd
, PERF_EVENT_IOC_SET_OUTPUT
, *output
) != 0)
1079 perf_mmap__get(&maps
[idx
]);
1082 revent
= perf_evlist__should_poll(evlist
, evsel
) ? POLLIN
: 0;
1085 * The system_wide flag causes a selected event to be opened
1086 * always without a pid. Consequently it will never get a
1087 * POLLHUP, but it is used for tracking in combination with
1088 * other events, so it should not need to be polled anyway.
1089 * Therefore don't add it for polling.
1091 if (!evsel
->system_wide
&&
1092 __perf_evlist__add_pollfd(evlist
, fd
, &maps
[idx
], revent
) < 0) {
1093 perf_mmap__put(&maps
[idx
]);
1097 if (evsel
->attr
.read_format
& PERF_FORMAT_ID
) {
1098 if (perf_evlist__id_add_fd(evlist
, evsel
, cpu
, thread
,
1101 perf_evlist__set_sid_idx(evlist
, evsel
, idx
, cpu
,
1109 static int perf_evlist__mmap_per_cpu(struct perf_evlist
*evlist
,
1110 struct mmap_params
*mp
)
1113 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
1114 int nr_threads
= thread_map__nr(evlist
->threads
);
1116 pr_debug2("perf event ring buffer mmapped per cpu\n");
1117 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
1119 int output_backward
= -1;
1121 auxtrace_mmap_params__set_idx(&mp
->auxtrace_mp
, evlist
, cpu
,
1124 for (thread
= 0; thread
< nr_threads
; thread
++) {
1125 if (perf_evlist__mmap_per_evsel(evlist
, cpu
, mp
, cpu
,
1126 thread
, &output
, &output_backward
))
1134 perf_evlist__munmap_nofree(evlist
);
1138 static int perf_evlist__mmap_per_thread(struct perf_evlist
*evlist
,
1139 struct mmap_params
*mp
)
1142 int nr_threads
= thread_map__nr(evlist
->threads
);
1144 pr_debug2("perf event ring buffer mmapped per thread\n");
1145 for (thread
= 0; thread
< nr_threads
; thread
++) {
1147 int output_backward
= -1;
1149 auxtrace_mmap_params__set_idx(&mp
->auxtrace_mp
, evlist
, thread
,
1152 if (perf_evlist__mmap_per_evsel(evlist
, thread
, mp
, 0, thread
,
1153 &output
, &output_backward
))
1160 perf_evlist__munmap_nofree(evlist
);
1164 unsigned long perf_event_mlock_kb_in_pages(void)
1166 unsigned long pages
;
1169 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max
) < 0) {
1171 * Pick a once upon a time good value, i.e. things look
1172 * strange since we can't read a sysctl value, but lets not
1177 max
-= (page_size
/ 1024);
1180 pages
= (max
* 1024) / page_size
;
1181 if (!is_power_of_2(pages
))
1182 pages
= rounddown_pow_of_two(pages
);
1187 size_t perf_evlist__mmap_size(unsigned long pages
)
1189 if (pages
== UINT_MAX
)
1190 pages
= perf_event_mlock_kb_in_pages();
1191 else if (!is_power_of_2(pages
))
1194 return (pages
+ 1) * page_size
;
1197 static long parse_pages_arg(const char *str
, unsigned long min
,
1200 unsigned long pages
, val
;
1201 static struct parse_tag tags
[] = {
1202 { .tag
= 'B', .mult
= 1 },
1203 { .tag
= 'K', .mult
= 1 << 10 },
1204 { .tag
= 'M', .mult
= 1 << 20 },
1205 { .tag
= 'G', .mult
= 1 << 30 },
1212 val
= parse_tag_value(str
, tags
);
1213 if (val
!= (unsigned long) -1) {
1214 /* we got file size value */
1215 pages
= PERF_ALIGN(val
, page_size
) / page_size
;
1217 /* we got pages count value */
1219 pages
= strtoul(str
, &eptr
, 10);
1224 if (pages
== 0 && min
== 0) {
1225 /* leave number of pages at 0 */
1226 } else if (!is_power_of_2(pages
)) {
1229 /* round pages up to next power of 2 */
1230 pages
= roundup_pow_of_two(pages
);
1234 unit_number__scnprintf(buf
, sizeof(buf
), pages
* page_size
);
1235 pr_info("rounding mmap pages size to %s (%lu pages)\n",
1245 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages
, const char *str
)
1247 unsigned long max
= UINT_MAX
;
1250 if (max
> SIZE_MAX
/ page_size
)
1251 max
= SIZE_MAX
/ page_size
;
1253 pages
= parse_pages_arg(str
, 1, max
);
1255 pr_err("Invalid argument for --mmap_pages/-m\n");
1259 *mmap_pages
= pages
;
1263 int perf_evlist__parse_mmap_pages(const struct option
*opt
, const char *str
,
1264 int unset __maybe_unused
)
1266 return __perf_evlist__parse_mmap_pages(opt
->value
, str
);
1270 * perf_evlist__mmap_ex - Create mmaps to receive events.
1271 * @evlist: list of events
1272 * @pages: map length in pages
1273 * @overwrite: overwrite older events?
1274 * @auxtrace_pages - auxtrace map length in pages
1275 * @auxtrace_overwrite - overwrite older auxtrace data?
1277 * If @overwrite is %false the user needs to signal event consumption using
1278 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
1281 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1282 * consumption using auxtrace_mmap__write_tail().
1284 * Return: %0 on success, negative error code otherwise.
1286 int perf_evlist__mmap_ex(struct perf_evlist
*evlist
, unsigned int pages
,
1287 bool overwrite
, unsigned int auxtrace_pages
,
1288 bool auxtrace_overwrite
)
1290 struct perf_evsel
*evsel
;
1291 const struct cpu_map
*cpus
= evlist
->cpus
;
1292 const struct thread_map
*threads
= evlist
->threads
;
1293 struct mmap_params mp
= {
1294 .prot
= PROT_READ
| (overwrite
? 0 : PROT_WRITE
),
1298 evlist
->mmap
= perf_evlist__alloc_mmap(evlist
);
1302 if (evlist
->pollfd
.entries
== NULL
&& perf_evlist__alloc_pollfd(evlist
) < 0)
1305 evlist
->overwrite
= overwrite
;
1306 evlist
->mmap_len
= perf_evlist__mmap_size(pages
);
1307 pr_debug("mmap size %zuB\n", evlist
->mmap_len
);
1308 mp
.mask
= evlist
->mmap_len
- page_size
- 1;
1310 auxtrace_mmap_params__init(&mp
.auxtrace_mp
, evlist
->mmap_len
,
1311 auxtrace_pages
, auxtrace_overwrite
);
1313 evlist__for_each_entry(evlist
, evsel
) {
1314 if ((evsel
->attr
.read_format
& PERF_FORMAT_ID
) &&
1315 evsel
->sample_id
== NULL
&&
1316 perf_evsel__alloc_id(evsel
, cpu_map__nr(cpus
), threads
->nr
) < 0)
1320 if (cpu_map__empty(cpus
))
1321 return perf_evlist__mmap_per_thread(evlist
, &mp
);
1323 return perf_evlist__mmap_per_cpu(evlist
, &mp
);
1326 int perf_evlist__mmap(struct perf_evlist
*evlist
, unsigned int pages
,
1329 return perf_evlist__mmap_ex(evlist
, pages
, overwrite
, 0, false);
1332 int perf_evlist__create_maps(struct perf_evlist
*evlist
, struct target
*target
)
1334 struct cpu_map
*cpus
;
1335 struct thread_map
*threads
;
1337 threads
= thread_map__new_str(target
->pid
, target
->tid
, target
->uid
);
1342 if (target__uses_dummy_map(target
))
1343 cpus
= cpu_map__dummy_new();
1345 cpus
= cpu_map__new(target
->cpu_list
);
1348 goto out_delete_threads
;
1350 evlist
->has_user_cpus
= !!target
->cpu_list
;
1352 perf_evlist__set_maps(evlist
, cpus
, threads
);
1357 thread_map__put(threads
);
1361 void perf_evlist__set_maps(struct perf_evlist
*evlist
, struct cpu_map
*cpus
,
1362 struct thread_map
*threads
)
1365 * Allow for the possibility that one or another of the maps isn't being
1366 * changed i.e. don't put it. Note we are assuming the maps that are
1367 * being applied are brand new and evlist is taking ownership of the
1368 * original reference count of 1. If that is not the case it is up to
1369 * the caller to increase the reference count.
1371 if (cpus
!= evlist
->cpus
) {
1372 cpu_map__put(evlist
->cpus
);
1373 evlist
->cpus
= cpu_map__get(cpus
);
1376 if (threads
!= evlist
->threads
) {
1377 thread_map__put(evlist
->threads
);
1378 evlist
->threads
= thread_map__get(threads
);
1381 perf_evlist__propagate_maps(evlist
);
1384 void __perf_evlist__set_sample_bit(struct perf_evlist
*evlist
,
1385 enum perf_event_sample_format bit
)
1387 struct perf_evsel
*evsel
;
1389 evlist__for_each_entry(evlist
, evsel
)
1390 __perf_evsel__set_sample_bit(evsel
, bit
);
1393 void __perf_evlist__reset_sample_bit(struct perf_evlist
*evlist
,
1394 enum perf_event_sample_format bit
)
1396 struct perf_evsel
*evsel
;
1398 evlist__for_each_entry(evlist
, evsel
)
1399 __perf_evsel__reset_sample_bit(evsel
, bit
);
1402 int perf_evlist__apply_filters(struct perf_evlist
*evlist
, struct perf_evsel
**err_evsel
)
1404 struct perf_evsel
*evsel
;
1406 const int ncpus
= cpu_map__nr(evlist
->cpus
),
1407 nthreads
= thread_map__nr(evlist
->threads
);
1409 evlist__for_each_entry(evlist
, evsel
) {
1410 if (evsel
->filter
== NULL
)
1414 * filters only work for tracepoint event, which doesn't have cpu limit.
1415 * So evlist and evsel should always be same.
1417 err
= perf_evsel__apply_filter(evsel
, ncpus
, nthreads
, evsel
->filter
);
1427 int perf_evlist__set_filter(struct perf_evlist
*evlist
, const char *filter
)
1429 struct perf_evsel
*evsel
;
1432 evlist__for_each_entry(evlist
, evsel
) {
1433 if (evsel
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
1436 err
= perf_evsel__set_filter(evsel
, filter
);
1444 int perf_evlist__set_filter_pids(struct perf_evlist
*evlist
, size_t npids
, pid_t
*pids
)
1450 for (i
= 0; i
< npids
; ++i
) {
1452 if (asprintf(&filter
, "common_pid != %d", pids
[i
]) < 0)
1457 if (asprintf(&tmp
, "%s && common_pid != %d", filter
, pids
[i
]) < 0)
1465 ret
= perf_evlist__set_filter(evlist
, filter
);
1471 int perf_evlist__set_filter_pid(struct perf_evlist
*evlist
, pid_t pid
)
1473 return perf_evlist__set_filter_pids(evlist
, 1, &pid
);
1476 bool perf_evlist__valid_sample_type(struct perf_evlist
*evlist
)
1478 struct perf_evsel
*pos
;
1480 if (evlist
->nr_entries
== 1)
1483 if (evlist
->id_pos
< 0 || evlist
->is_pos
< 0)
1486 evlist__for_each_entry(evlist
, pos
) {
1487 if (pos
->id_pos
!= evlist
->id_pos
||
1488 pos
->is_pos
!= evlist
->is_pos
)
1495 u64
__perf_evlist__combined_sample_type(struct perf_evlist
*evlist
)
1497 struct perf_evsel
*evsel
;
1499 if (evlist
->combined_sample_type
)
1500 return evlist
->combined_sample_type
;
1502 evlist__for_each_entry(evlist
, evsel
)
1503 evlist
->combined_sample_type
|= evsel
->attr
.sample_type
;
1505 return evlist
->combined_sample_type
;
1508 u64
perf_evlist__combined_sample_type(struct perf_evlist
*evlist
)
1510 evlist
->combined_sample_type
= 0;
1511 return __perf_evlist__combined_sample_type(evlist
);
1514 u64
perf_evlist__combined_branch_type(struct perf_evlist
*evlist
)
1516 struct perf_evsel
*evsel
;
1517 u64 branch_type
= 0;
1519 evlist__for_each_entry(evlist
, evsel
)
1520 branch_type
|= evsel
->attr
.branch_sample_type
;
1524 bool perf_evlist__valid_read_format(struct perf_evlist
*evlist
)
1526 struct perf_evsel
*first
= perf_evlist__first(evlist
), *pos
= first
;
1527 u64 read_format
= first
->attr
.read_format
;
1528 u64 sample_type
= first
->attr
.sample_type
;
1530 evlist__for_each_entry(evlist
, pos
) {
1531 if (read_format
!= pos
->attr
.read_format
)
1535 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1536 if ((sample_type
& PERF_SAMPLE_READ
) &&
1537 !(read_format
& PERF_FORMAT_ID
)) {
1544 u64
perf_evlist__read_format(struct perf_evlist
*evlist
)
1546 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1547 return first
->attr
.read_format
;
1550 u16
perf_evlist__id_hdr_size(struct perf_evlist
*evlist
)
1552 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1553 struct perf_sample
*data
;
1557 if (!first
->attr
.sample_id_all
)
1560 sample_type
= first
->attr
.sample_type
;
1562 if (sample_type
& PERF_SAMPLE_TID
)
1563 size
+= sizeof(data
->tid
) * 2;
1565 if (sample_type
& PERF_SAMPLE_TIME
)
1566 size
+= sizeof(data
->time
);
1568 if (sample_type
& PERF_SAMPLE_ID
)
1569 size
+= sizeof(data
->id
);
1571 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1572 size
+= sizeof(data
->stream_id
);
1574 if (sample_type
& PERF_SAMPLE_CPU
)
1575 size
+= sizeof(data
->cpu
) * 2;
1577 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1578 size
+= sizeof(data
->id
);
1583 bool perf_evlist__valid_sample_id_all(struct perf_evlist
*evlist
)
1585 struct perf_evsel
*first
= perf_evlist__first(evlist
), *pos
= first
;
1587 evlist__for_each_entry_continue(evlist
, pos
) {
1588 if (first
->attr
.sample_id_all
!= pos
->attr
.sample_id_all
)
1595 bool perf_evlist__sample_id_all(struct perf_evlist
*evlist
)
1597 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1598 return first
->attr
.sample_id_all
;
1601 void perf_evlist__set_selected(struct perf_evlist
*evlist
,
1602 struct perf_evsel
*evsel
)
1604 evlist
->selected
= evsel
;
1607 void perf_evlist__close(struct perf_evlist
*evlist
)
1609 struct perf_evsel
*evsel
;
1610 int ncpus
= cpu_map__nr(evlist
->cpus
);
1611 int nthreads
= thread_map__nr(evlist
->threads
);
1613 evlist__for_each_entry_reverse(evlist
, evsel
) {
1614 int n
= evsel
->cpus
? evsel
->cpus
->nr
: ncpus
;
1615 perf_evsel__close(evsel
, n
, nthreads
);
1619 static int perf_evlist__create_syswide_maps(struct perf_evlist
*evlist
)
1621 struct cpu_map
*cpus
;
1622 struct thread_map
*threads
;
1626 * Try reading /sys/devices/system/cpu/online to get
1629 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1630 * code needs an overhaul to properly forward the
1631 * error, and we may not want to do that fallback to a
1632 * default cpu identity map :-\
1634 cpus
= cpu_map__new(NULL
);
1638 threads
= thread_map__new_dummy();
1642 perf_evlist__set_maps(evlist
, cpus
, threads
);
1650 int perf_evlist__open(struct perf_evlist
*evlist
)
1652 struct perf_evsel
*evsel
;
1656 * Default: one fd per CPU, all threads, aka systemwide
1657 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1659 if (evlist
->threads
== NULL
&& evlist
->cpus
== NULL
) {
1660 err
= perf_evlist__create_syswide_maps(evlist
);
1665 perf_evlist__update_id_pos(evlist
);
1667 evlist__for_each_entry(evlist
, evsel
) {
1668 err
= perf_evsel__open(evsel
, evsel
->cpus
, evsel
->threads
);
1675 perf_evlist__close(evlist
);
1680 int perf_evlist__prepare_workload(struct perf_evlist
*evlist
, struct target
*target
,
1681 const char *argv
[], bool pipe_output
,
1682 void (*exec_error
)(int signo
, siginfo_t
*info
, void *ucontext
))
1684 int child_ready_pipe
[2], go_pipe
[2];
1687 if (pipe(child_ready_pipe
) < 0) {
1688 perror("failed to create 'ready' pipe");
1692 if (pipe(go_pipe
) < 0) {
1693 perror("failed to create 'go' pipe");
1694 goto out_close_ready_pipe
;
1697 evlist
->workload
.pid
= fork();
1698 if (evlist
->workload
.pid
< 0) {
1699 perror("failed to fork");
1700 goto out_close_pipes
;
1703 if (!evlist
->workload
.pid
) {
1709 signal(SIGTERM
, SIG_DFL
);
1711 close(child_ready_pipe
[0]);
1713 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
1716 * Tell the parent we're ready to go
1718 close(child_ready_pipe
[1]);
1721 * Wait until the parent tells us to go.
1723 ret
= read(go_pipe
[0], &bf
, 1);
1725 * The parent will ask for the execvp() to be performed by
1726 * writing exactly one byte, in workload.cork_fd, usually via
1727 * perf_evlist__start_workload().
1729 * For cancelling the workload without actually running it,
1730 * the parent will just close workload.cork_fd, without writing
1731 * anything, i.e. read will return zero and we just exit()
1736 perror("unable to read pipe");
1740 execvp(argv
[0], (char **)argv
);
1745 val
.sival_int
= errno
;
1746 if (sigqueue(getppid(), SIGUSR1
, val
))
1754 struct sigaction act
= {
1755 .sa_flags
= SA_SIGINFO
,
1756 .sa_sigaction
= exec_error
,
1758 sigaction(SIGUSR1
, &act
, NULL
);
1761 if (target__none(target
)) {
1762 if (evlist
->threads
== NULL
) {
1763 fprintf(stderr
, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1764 __func__
, __LINE__
);
1765 goto out_close_pipes
;
1767 thread_map__set_pid(evlist
->threads
, 0, evlist
->workload
.pid
);
1770 close(child_ready_pipe
[1]);
1773 * wait for child to settle
1775 if (read(child_ready_pipe
[0], &bf
, 1) == -1) {
1776 perror("unable to read pipe");
1777 goto out_close_pipes
;
1780 fcntl(go_pipe
[1], F_SETFD
, FD_CLOEXEC
);
1781 evlist
->workload
.cork_fd
= go_pipe
[1];
1782 close(child_ready_pipe
[0]);
1788 out_close_ready_pipe
:
1789 close(child_ready_pipe
[0]);
1790 close(child_ready_pipe
[1]);
1794 int perf_evlist__start_workload(struct perf_evlist
*evlist
)
1796 if (evlist
->workload
.cork_fd
> 0) {
1800 * Remove the cork, let it rip!
1802 ret
= write(evlist
->workload
.cork_fd
, &bf
, 1);
1804 perror("unable to write to pipe");
1806 close(evlist
->workload
.cork_fd
);
1813 int perf_evlist__parse_sample(struct perf_evlist
*evlist
, union perf_event
*event
,
1814 struct perf_sample
*sample
)
1816 struct perf_evsel
*evsel
= perf_evlist__event2evsel(evlist
, event
);
1820 return perf_evsel__parse_sample(evsel
, event
, sample
);
1823 size_t perf_evlist__fprintf(struct perf_evlist
*evlist
, FILE *fp
)
1825 struct perf_evsel
*evsel
;
1828 evlist__for_each_entry(evlist
, evsel
) {
1829 printed
+= fprintf(fp
, "%s%s", evsel
->idx
? ", " : "",
1830 perf_evsel__name(evsel
));
1833 return printed
+ fprintf(fp
, "\n");
1836 int perf_evlist__strerror_open(struct perf_evlist
*evlist
,
1837 int err
, char *buf
, size_t size
)
1840 char sbuf
[STRERR_BUFSIZE
], *emsg
= str_error_r(err
, sbuf
, sizeof(sbuf
));
1845 printed
= scnprintf(buf
, size
,
1847 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg
);
1849 value
= perf_event_paranoid();
1851 printed
+= scnprintf(buf
+ printed
, size
- printed
, "\nHint:\t");
1854 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1855 "For your workloads it needs to be <= 1\nHint:\t");
1857 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1858 "For system wide tracing it needs to be set to -1.\n");
1860 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1861 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1862 "Hint:\tThe current value is %d.", value
);
1865 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1868 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq
) < 0)
1871 if (first
->attr
.sample_freq
< (u64
)max_freq
)
1874 printed
= scnprintf(buf
, size
,
1876 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1877 "Hint:\tThe current value is %d and %" PRIu64
" is being requested.",
1878 emsg
, max_freq
, first
->attr
.sample_freq
);
1883 scnprintf(buf
, size
, "%s", emsg
);
1890 int perf_evlist__strerror_mmap(struct perf_evlist
*evlist
, int err
, char *buf
, size_t size
)
1892 char sbuf
[STRERR_BUFSIZE
], *emsg
= str_error_r(err
, sbuf
, sizeof(sbuf
));
1893 int pages_attempted
= evlist
->mmap_len
/ 1024, pages_max_per_user
, printed
= 0;
1897 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user
);
1898 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1900 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1901 "Hint:\tTried using %zd kB.\n",
1902 emsg
, pages_max_per_user
, pages_attempted
);
1904 if (pages_attempted
>= pages_max_per_user
) {
1905 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1906 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1907 pages_max_per_user
+ pages_attempted
);
1910 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1911 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1914 scnprintf(buf
, size
, "%s", emsg
);
1921 void perf_evlist__to_front(struct perf_evlist
*evlist
,
1922 struct perf_evsel
*move_evsel
)
1924 struct perf_evsel
*evsel
, *n
;
1927 if (move_evsel
== perf_evlist__first(evlist
))
1930 evlist__for_each_entry_safe(evlist
, n
, evsel
) {
1931 if (evsel
->leader
== move_evsel
->leader
)
1932 list_move_tail(&evsel
->node
, &move
);
1935 list_splice(&move
, &evlist
->entries
);
1938 void perf_evlist__set_tracking_event(struct perf_evlist
*evlist
,
1939 struct perf_evsel
*tracking_evsel
)
1941 struct perf_evsel
*evsel
;
1943 if (tracking_evsel
->tracking
)
1946 evlist__for_each_entry(evlist
, evsel
) {
1947 if (evsel
!= tracking_evsel
)
1948 evsel
->tracking
= false;
1951 tracking_evsel
->tracking
= true;
1955 perf_evlist__find_evsel_by_str(struct perf_evlist
*evlist
,
1958 struct perf_evsel
*evsel
;
1960 evlist__for_each_entry(evlist
, evsel
) {
1963 if (strcmp(str
, evsel
->name
) == 0)
1970 void perf_evlist__toggle_bkw_mmap(struct perf_evlist
*evlist
,
1971 enum bkw_mmap_state state
)
1973 enum bkw_mmap_state old_state
= evlist
->bkw_mmap_state
;
1980 if (!evlist
->backward_mmap
)
1983 switch (old_state
) {
1984 case BKW_MMAP_NOTREADY
: {
1985 if (state
!= BKW_MMAP_RUNNING
)
1989 case BKW_MMAP_RUNNING
: {
1990 if (state
!= BKW_MMAP_DATA_PENDING
)
1995 case BKW_MMAP_DATA_PENDING
: {
1996 if (state
!= BKW_MMAP_EMPTY
)
2000 case BKW_MMAP_EMPTY
: {
2001 if (state
!= BKW_MMAP_RUNNING
)
2007 WARN_ONCE(1, "Shouldn't get there\n");
2010 evlist
->bkw_mmap_state
= state
;
2014 perf_evlist__pause(evlist
);
2017 perf_evlist__resume(evlist
);