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/debugfs.h>
11 #include <api/fs/fs.h>
14 #include "thread_map.h"
21 #include "parse-events.h"
22 #include "parse-options.h"
26 #include <linux/bitops.h>
27 #include <linux/hash.h>
28 #include <linux/log2.h>
30 static void perf_evlist__mmap_put(struct perf_evlist
*evlist
, int idx
);
31 static void __perf_evlist__munmap(struct perf_evlist
*evlist
, int idx
);
33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
34 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
36 void perf_evlist__init(struct perf_evlist
*evlist
, struct cpu_map
*cpus
,
37 struct thread_map
*threads
)
41 for (i
= 0; i
< PERF_EVLIST__HLIST_SIZE
; ++i
)
42 INIT_HLIST_HEAD(&evlist
->heads
[i
]);
43 INIT_LIST_HEAD(&evlist
->entries
);
44 perf_evlist__set_maps(evlist
, cpus
, threads
);
45 fdarray__init(&evlist
->pollfd
, 64);
46 evlist
->workload
.pid
= -1;
49 struct perf_evlist
*perf_evlist__new(void)
51 struct perf_evlist
*evlist
= zalloc(sizeof(*evlist
));
54 perf_evlist__init(evlist
, NULL
, NULL
);
59 struct perf_evlist
*perf_evlist__new_default(void)
61 struct perf_evlist
*evlist
= perf_evlist__new();
63 if (evlist
&& perf_evlist__add_default(evlist
)) {
64 perf_evlist__delete(evlist
);
72 * perf_evlist__set_id_pos - set the positions of event ids.
73 * @evlist: selected event list
75 * Events with compatible sample types all have the same id_pos
76 * and is_pos. For convenience, put a copy on evlist.
78 void perf_evlist__set_id_pos(struct perf_evlist
*evlist
)
80 struct perf_evsel
*first
= perf_evlist__first(evlist
);
82 evlist
->id_pos
= first
->id_pos
;
83 evlist
->is_pos
= first
->is_pos
;
86 static void perf_evlist__update_id_pos(struct perf_evlist
*evlist
)
88 struct perf_evsel
*evsel
;
90 evlist__for_each(evlist
, evsel
)
91 perf_evsel__calc_id_pos(evsel
);
93 perf_evlist__set_id_pos(evlist
);
96 static void perf_evlist__purge(struct perf_evlist
*evlist
)
98 struct perf_evsel
*pos
, *n
;
100 evlist__for_each_safe(evlist
, n
, pos
) {
101 list_del_init(&pos
->node
);
102 perf_evsel__delete(pos
);
105 evlist
->nr_entries
= 0;
108 void perf_evlist__exit(struct perf_evlist
*evlist
)
110 zfree(&evlist
->mmap
);
111 fdarray__exit(&evlist
->pollfd
);
114 void perf_evlist__delete(struct perf_evlist
*evlist
)
116 perf_evlist__munmap(evlist
);
117 perf_evlist__close(evlist
);
118 cpu_map__delete(evlist
->cpus
);
119 thread_map__delete(evlist
->threads
);
121 evlist
->threads
= NULL
;
122 perf_evlist__purge(evlist
);
123 perf_evlist__exit(evlist
);
127 void perf_evlist__add(struct perf_evlist
*evlist
, struct perf_evsel
*entry
)
129 list_add_tail(&entry
->node
, &evlist
->entries
);
130 entry
->idx
= evlist
->nr_entries
;
131 entry
->tracking
= !entry
->idx
;
133 if (!evlist
->nr_entries
++)
134 perf_evlist__set_id_pos(evlist
);
137 void perf_evlist__splice_list_tail(struct perf_evlist
*evlist
,
138 struct list_head
*list
,
141 bool set_id_pos
= !evlist
->nr_entries
;
143 list_splice_tail(list
, &evlist
->entries
);
144 evlist
->nr_entries
+= nr_entries
;
146 perf_evlist__set_id_pos(evlist
);
149 void __perf_evlist__set_leader(struct list_head
*list
)
151 struct perf_evsel
*evsel
, *leader
;
153 leader
= list_entry(list
->next
, struct perf_evsel
, node
);
154 evsel
= list_entry(list
->prev
, struct perf_evsel
, node
);
156 leader
->nr_members
= evsel
->idx
- leader
->idx
+ 1;
158 __evlist__for_each(list
, evsel
) {
159 evsel
->leader
= leader
;
163 void perf_evlist__set_leader(struct perf_evlist
*evlist
)
165 if (evlist
->nr_entries
) {
166 evlist
->nr_groups
= evlist
->nr_entries
> 1 ? 1 : 0;
167 __perf_evlist__set_leader(&evlist
->entries
);
171 int perf_evlist__add_default(struct perf_evlist
*evlist
)
173 struct perf_event_attr attr
= {
174 .type
= PERF_TYPE_HARDWARE
,
175 .config
= PERF_COUNT_HW_CPU_CYCLES
,
177 struct perf_evsel
*evsel
;
179 event_attr_init(&attr
);
181 evsel
= perf_evsel__new(&attr
);
185 /* use strdup() because free(evsel) assumes name is allocated */
186 evsel
->name
= strdup("cycles");
190 perf_evlist__add(evlist
, evsel
);
193 perf_evsel__delete(evsel
);
198 static int perf_evlist__add_attrs(struct perf_evlist
*evlist
,
199 struct perf_event_attr
*attrs
, size_t nr_attrs
)
201 struct perf_evsel
*evsel
, *n
;
205 for (i
= 0; i
< nr_attrs
; i
++) {
206 evsel
= perf_evsel__new_idx(attrs
+ i
, evlist
->nr_entries
+ i
);
208 goto out_delete_partial_list
;
209 list_add_tail(&evsel
->node
, &head
);
212 perf_evlist__splice_list_tail(evlist
, &head
, nr_attrs
);
216 out_delete_partial_list
:
217 __evlist__for_each_safe(&head
, n
, evsel
)
218 perf_evsel__delete(evsel
);
222 int __perf_evlist__add_default_attrs(struct perf_evlist
*evlist
,
223 struct perf_event_attr
*attrs
, size_t nr_attrs
)
227 for (i
= 0; i
< nr_attrs
; i
++)
228 event_attr_init(attrs
+ i
);
230 return perf_evlist__add_attrs(evlist
, attrs
, nr_attrs
);
234 perf_evlist__find_tracepoint_by_id(struct perf_evlist
*evlist
, int id
)
236 struct perf_evsel
*evsel
;
238 evlist__for_each(evlist
, evsel
) {
239 if (evsel
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
240 (int)evsel
->attr
.config
== id
)
248 perf_evlist__find_tracepoint_by_name(struct perf_evlist
*evlist
,
251 struct perf_evsel
*evsel
;
253 evlist__for_each(evlist
, evsel
) {
254 if ((evsel
->attr
.type
== PERF_TYPE_TRACEPOINT
) &&
255 (strcmp(evsel
->name
, name
) == 0))
262 int perf_evlist__add_newtp(struct perf_evlist
*evlist
,
263 const char *sys
, const char *name
, void *handler
)
265 struct perf_evsel
*evsel
= perf_evsel__newtp(sys
, name
);
270 evsel
->handler
= handler
;
271 perf_evlist__add(evlist
, evsel
);
275 static int perf_evlist__nr_threads(struct perf_evlist
*evlist
,
276 struct perf_evsel
*evsel
)
278 if (evsel
->system_wide
)
281 return thread_map__nr(evlist
->threads
);
284 void perf_evlist__disable(struct perf_evlist
*evlist
)
287 struct perf_evsel
*pos
;
288 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
291 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
292 evlist__for_each(evlist
, pos
) {
293 if (!perf_evsel__is_group_leader(pos
) || !pos
->fd
)
295 nr_threads
= perf_evlist__nr_threads(evlist
, pos
);
296 for (thread
= 0; thread
< nr_threads
; thread
++)
297 ioctl(FD(pos
, cpu
, thread
),
298 PERF_EVENT_IOC_DISABLE
, 0);
303 void perf_evlist__enable(struct perf_evlist
*evlist
)
306 struct perf_evsel
*pos
;
307 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
310 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
311 evlist__for_each(evlist
, pos
) {
312 if (!perf_evsel__is_group_leader(pos
) || !pos
->fd
)
314 nr_threads
= perf_evlist__nr_threads(evlist
, pos
);
315 for (thread
= 0; thread
< nr_threads
; thread
++)
316 ioctl(FD(pos
, cpu
, thread
),
317 PERF_EVENT_IOC_ENABLE
, 0);
322 int perf_evlist__disable_event(struct perf_evlist
*evlist
,
323 struct perf_evsel
*evsel
)
325 int cpu
, thread
, err
;
326 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
327 int nr_threads
= perf_evlist__nr_threads(evlist
, evsel
);
332 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
333 for (thread
= 0; thread
< nr_threads
; thread
++) {
334 err
= ioctl(FD(evsel
, cpu
, thread
),
335 PERF_EVENT_IOC_DISABLE
, 0);
343 int perf_evlist__enable_event(struct perf_evlist
*evlist
,
344 struct perf_evsel
*evsel
)
346 int cpu
, thread
, err
;
347 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
348 int nr_threads
= perf_evlist__nr_threads(evlist
, evsel
);
353 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
354 for (thread
= 0; thread
< nr_threads
; thread
++) {
355 err
= ioctl(FD(evsel
, cpu
, thread
),
356 PERF_EVENT_IOC_ENABLE
, 0);
364 static int perf_evlist__enable_event_cpu(struct perf_evlist
*evlist
,
365 struct perf_evsel
*evsel
, int cpu
)
368 int nr_threads
= perf_evlist__nr_threads(evlist
, evsel
);
373 for (thread
= 0; thread
< nr_threads
; thread
++) {
374 err
= ioctl(FD(evsel
, cpu
, thread
),
375 PERF_EVENT_IOC_ENABLE
, 0);
382 static int perf_evlist__enable_event_thread(struct perf_evlist
*evlist
,
383 struct perf_evsel
*evsel
,
387 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
392 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
393 err
= ioctl(FD(evsel
, cpu
, thread
), PERF_EVENT_IOC_ENABLE
, 0);
400 int perf_evlist__enable_event_idx(struct perf_evlist
*evlist
,
401 struct perf_evsel
*evsel
, int idx
)
403 bool per_cpu_mmaps
= !cpu_map__empty(evlist
->cpus
);
406 return perf_evlist__enable_event_cpu(evlist
, evsel
, idx
);
408 return perf_evlist__enable_event_thread(evlist
, evsel
, idx
);
411 int perf_evlist__alloc_pollfd(struct perf_evlist
*evlist
)
413 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
414 int nr_threads
= thread_map__nr(evlist
->threads
);
416 struct perf_evsel
*evsel
;
418 evlist__for_each(evlist
, evsel
) {
419 if (evsel
->system_wide
)
422 nfds
+= nr_cpus
* nr_threads
;
425 if (fdarray__available_entries(&evlist
->pollfd
) < nfds
&&
426 fdarray__grow(&evlist
->pollfd
, nfds
) < 0)
432 static int __perf_evlist__add_pollfd(struct perf_evlist
*evlist
, int fd
, int idx
)
434 int pos
= fdarray__add(&evlist
->pollfd
, fd
, POLLIN
| POLLERR
| POLLHUP
);
436 * Save the idx so that when we filter out fds POLLHUP'ed we can
437 * close the associated evlist->mmap[] entry.
440 evlist
->pollfd
.priv
[pos
].idx
= idx
;
442 fcntl(fd
, F_SETFL
, O_NONBLOCK
);
448 int perf_evlist__add_pollfd(struct perf_evlist
*evlist
, int fd
)
450 return __perf_evlist__add_pollfd(evlist
, fd
, -1);
453 static void perf_evlist__munmap_filtered(struct fdarray
*fda
, int fd
)
455 struct perf_evlist
*evlist
= container_of(fda
, struct perf_evlist
, pollfd
);
457 perf_evlist__mmap_put(evlist
, fda
->priv
[fd
].idx
);
460 int perf_evlist__filter_pollfd(struct perf_evlist
*evlist
, short revents_and_mask
)
462 return fdarray__filter(&evlist
->pollfd
, revents_and_mask
,
463 perf_evlist__munmap_filtered
);
466 int perf_evlist__poll(struct perf_evlist
*evlist
, int timeout
)
468 return fdarray__poll(&evlist
->pollfd
, timeout
);
471 static void perf_evlist__id_hash(struct perf_evlist
*evlist
,
472 struct perf_evsel
*evsel
,
473 int cpu
, int thread
, u64 id
)
476 struct perf_sample_id
*sid
= SID(evsel
, cpu
, thread
);
480 hash
= hash_64(sid
->id
, PERF_EVLIST__HLIST_BITS
);
481 hlist_add_head(&sid
->node
, &evlist
->heads
[hash
]);
484 void perf_evlist__id_add(struct perf_evlist
*evlist
, struct perf_evsel
*evsel
,
485 int cpu
, int thread
, u64 id
)
487 perf_evlist__id_hash(evlist
, evsel
, cpu
, thread
, id
);
488 evsel
->id
[evsel
->ids
++] = id
;
491 static int perf_evlist__id_add_fd(struct perf_evlist
*evlist
,
492 struct perf_evsel
*evsel
,
493 int cpu
, int thread
, int fd
)
495 u64 read_data
[4] = { 0, };
496 int id_idx
= 1; /* The first entry is the counter value */
500 ret
= ioctl(fd
, PERF_EVENT_IOC_ID
, &id
);
507 /* Legacy way to get event id.. All hail to old kernels! */
510 * This way does not work with group format read, so bail
513 if (perf_evlist__read_format(evlist
) & PERF_FORMAT_GROUP
)
516 if (!(evsel
->attr
.read_format
& PERF_FORMAT_ID
) ||
517 read(fd
, &read_data
, sizeof(read_data
)) == -1)
520 if (evsel
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
522 if (evsel
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
525 id
= read_data
[id_idx
];
528 perf_evlist__id_add(evlist
, evsel
, cpu
, thread
, id
);
532 static void perf_evlist__set_sid_idx(struct perf_evlist
*evlist
,
533 struct perf_evsel
*evsel
, int idx
, int cpu
,
536 struct perf_sample_id
*sid
= SID(evsel
, cpu
, thread
);
538 if (evlist
->cpus
&& cpu
>= 0)
539 sid
->cpu
= evlist
->cpus
->map
[cpu
];
542 if (!evsel
->system_wide
&& evlist
->threads
&& thread
>= 0)
543 sid
->tid
= evlist
->threads
->map
[thread
];
548 struct perf_sample_id
*perf_evlist__id2sid(struct perf_evlist
*evlist
, u64 id
)
550 struct hlist_head
*head
;
551 struct perf_sample_id
*sid
;
554 hash
= hash_64(id
, PERF_EVLIST__HLIST_BITS
);
555 head
= &evlist
->heads
[hash
];
557 hlist_for_each_entry(sid
, head
, node
)
564 struct perf_evsel
*perf_evlist__id2evsel(struct perf_evlist
*evlist
, u64 id
)
566 struct perf_sample_id
*sid
;
568 if (evlist
->nr_entries
== 1)
569 return perf_evlist__first(evlist
);
571 sid
= perf_evlist__id2sid(evlist
, id
);
575 if (!perf_evlist__sample_id_all(evlist
))
576 return perf_evlist__first(evlist
);
581 static int perf_evlist__event2id(struct perf_evlist
*evlist
,
582 union perf_event
*event
, u64
*id
)
584 const u64
*array
= event
->sample
.array
;
587 n
= (event
->header
.size
- sizeof(event
->header
)) >> 3;
589 if (event
->header
.type
== PERF_RECORD_SAMPLE
) {
590 if (evlist
->id_pos
>= n
)
592 *id
= array
[evlist
->id_pos
];
594 if (evlist
->is_pos
> n
)
602 static struct perf_evsel
*perf_evlist__event2evsel(struct perf_evlist
*evlist
,
603 union perf_event
*event
)
605 struct perf_evsel
*first
= perf_evlist__first(evlist
);
606 struct hlist_head
*head
;
607 struct perf_sample_id
*sid
;
611 if (evlist
->nr_entries
== 1)
614 if (!first
->attr
.sample_id_all
&&
615 event
->header
.type
!= PERF_RECORD_SAMPLE
)
618 if (perf_evlist__event2id(evlist
, event
, &id
))
621 /* Synthesized events have an id of zero */
625 hash
= hash_64(id
, PERF_EVLIST__HLIST_BITS
);
626 head
= &evlist
->heads
[hash
];
628 hlist_for_each_entry(sid
, head
, node
) {
635 union perf_event
*perf_evlist__mmap_read(struct perf_evlist
*evlist
, int idx
)
637 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
638 unsigned int head
= perf_mmap__read_head(md
);
639 unsigned int old
= md
->prev
;
640 unsigned char *data
= md
->base
+ page_size
;
641 union perf_event
*event
= NULL
;
643 if (evlist
->overwrite
) {
645 * If we're further behind than half the buffer, there's a chance
646 * the writer will bite our tail and mess up the samples under us.
648 * If we somehow ended up ahead of the head, we got messed up.
650 * In either case, truncate and restart at head.
652 int diff
= head
- old
;
653 if (diff
> md
->mask
/ 2 || diff
< 0) {
654 fprintf(stderr
, "WARNING: failed to keep up with mmap data.\n");
657 * head points to a known good entry, start there.
666 event
= (union perf_event
*)&data
[old
& md
->mask
];
667 size
= event
->header
.size
;
670 * Event straddles the mmap boundary -- header should always
671 * be inside due to u64 alignment of output.
673 if ((old
& md
->mask
) + size
!= ((old
+ size
) & md
->mask
)) {
674 unsigned int offset
= old
;
675 unsigned int len
= min(sizeof(*event
), size
), cpy
;
676 void *dst
= md
->event_copy
;
679 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
680 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
686 event
= (union perf_event
*) md
->event_copy
;
697 static bool perf_mmap__empty(struct perf_mmap
*md
)
699 return perf_mmap__read_head(md
) != md
->prev
;
702 static void perf_evlist__mmap_get(struct perf_evlist
*evlist
, int idx
)
704 ++evlist
->mmap
[idx
].refcnt
;
707 static void perf_evlist__mmap_put(struct perf_evlist
*evlist
, int idx
)
709 BUG_ON(evlist
->mmap
[idx
].refcnt
== 0);
711 if (--evlist
->mmap
[idx
].refcnt
== 0)
712 __perf_evlist__munmap(evlist
, idx
);
715 void perf_evlist__mmap_consume(struct perf_evlist
*evlist
, int idx
)
717 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
719 if (!evlist
->overwrite
) {
720 unsigned int old
= md
->prev
;
722 perf_mmap__write_tail(md
, old
);
725 if (md
->refcnt
== 1 && perf_mmap__empty(md
))
726 perf_evlist__mmap_put(evlist
, idx
);
729 static void __perf_evlist__munmap(struct perf_evlist
*evlist
, int idx
)
731 if (evlist
->mmap
[idx
].base
!= NULL
) {
732 munmap(evlist
->mmap
[idx
].base
, evlist
->mmap_len
);
733 evlist
->mmap
[idx
].base
= NULL
;
734 evlist
->mmap
[idx
].refcnt
= 0;
738 void perf_evlist__munmap(struct perf_evlist
*evlist
)
742 if (evlist
->mmap
== NULL
)
745 for (i
= 0; i
< evlist
->nr_mmaps
; i
++)
746 __perf_evlist__munmap(evlist
, i
);
748 zfree(&evlist
->mmap
);
751 static int perf_evlist__alloc_mmap(struct perf_evlist
*evlist
)
753 evlist
->nr_mmaps
= cpu_map__nr(evlist
->cpus
);
754 if (cpu_map__empty(evlist
->cpus
))
755 evlist
->nr_mmaps
= thread_map__nr(evlist
->threads
);
756 evlist
->mmap
= zalloc(evlist
->nr_mmaps
* sizeof(struct perf_mmap
));
757 return evlist
->mmap
!= NULL
? 0 : -ENOMEM
;
765 static int __perf_evlist__mmap(struct perf_evlist
*evlist
, int idx
,
766 struct mmap_params
*mp
, int fd
)
769 * The last one will be done at perf_evlist__mmap_consume(), so that we
770 * make sure we don't prevent tools from consuming every last event in
773 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
774 * anymore, but the last events for it are still in the ring buffer,
775 * waiting to be consumed.
777 * Tools can chose to ignore this at their own discretion, but the
778 * evlist layer can't just drop it when filtering events in
779 * perf_evlist__filter_pollfd().
781 evlist
->mmap
[idx
].refcnt
= 2;
782 evlist
->mmap
[idx
].prev
= 0;
783 evlist
->mmap
[idx
].mask
= mp
->mask
;
784 evlist
->mmap
[idx
].base
= mmap(NULL
, evlist
->mmap_len
, mp
->prot
,
786 if (evlist
->mmap
[idx
].base
== MAP_FAILED
) {
787 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
789 evlist
->mmap
[idx
].base
= NULL
;
796 static int perf_evlist__mmap_per_evsel(struct perf_evlist
*evlist
, int idx
,
797 struct mmap_params
*mp
, int cpu
,
798 int thread
, int *output
)
800 struct perf_evsel
*evsel
;
802 evlist__for_each(evlist
, evsel
) {
805 if (evsel
->system_wide
&& thread
)
808 fd
= FD(evsel
, cpu
, thread
);
812 if (__perf_evlist__mmap(evlist
, idx
, mp
, *output
) < 0)
815 if (ioctl(fd
, PERF_EVENT_IOC_SET_OUTPUT
, *output
) != 0)
818 perf_evlist__mmap_get(evlist
, idx
);
822 * The system_wide flag causes a selected event to be opened
823 * always without a pid. Consequently it will never get a
824 * POLLHUP, but it is used for tracking in combination with
825 * other events, so it should not need to be polled anyway.
826 * Therefore don't add it for polling.
828 if (!evsel
->system_wide
&&
829 __perf_evlist__add_pollfd(evlist
, fd
, idx
) < 0) {
830 perf_evlist__mmap_put(evlist
, idx
);
834 if (evsel
->attr
.read_format
& PERF_FORMAT_ID
) {
835 if (perf_evlist__id_add_fd(evlist
, evsel
, cpu
, thread
,
838 perf_evlist__set_sid_idx(evlist
, evsel
, idx
, cpu
,
846 static int perf_evlist__mmap_per_cpu(struct perf_evlist
*evlist
,
847 struct mmap_params
*mp
)
850 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
851 int nr_threads
= thread_map__nr(evlist
->threads
);
853 pr_debug2("perf event ring buffer mmapped per cpu\n");
854 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
857 for (thread
= 0; thread
< nr_threads
; thread
++) {
858 if (perf_evlist__mmap_per_evsel(evlist
, cpu
, mp
, cpu
,
867 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
868 __perf_evlist__munmap(evlist
, cpu
);
872 static int perf_evlist__mmap_per_thread(struct perf_evlist
*evlist
,
873 struct mmap_params
*mp
)
876 int nr_threads
= thread_map__nr(evlist
->threads
);
878 pr_debug2("perf event ring buffer mmapped per thread\n");
879 for (thread
= 0; thread
< nr_threads
; thread
++) {
882 if (perf_evlist__mmap_per_evsel(evlist
, thread
, mp
, 0, thread
,
890 for (thread
= 0; thread
< nr_threads
; thread
++)
891 __perf_evlist__munmap(evlist
, thread
);
895 static size_t perf_evlist__mmap_size(unsigned long pages
)
897 if (pages
== UINT_MAX
) {
900 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max
) < 0) {
902 * Pick a once upon a time good value, i.e. things look
903 * strange since we can't read a sysctl value, but lets not
908 max
-= (page_size
/ 1024);
911 pages
= (max
* 1024) / page_size
;
912 if (!is_power_of_2(pages
))
913 pages
= rounddown_pow_of_two(pages
);
914 } else if (!is_power_of_2(pages
))
917 return (pages
+ 1) * page_size
;
920 static long parse_pages_arg(const char *str
, unsigned long min
,
923 unsigned long pages
, val
;
924 static struct parse_tag tags
[] = {
925 { .tag
= 'B', .mult
= 1 },
926 { .tag
= 'K', .mult
= 1 << 10 },
927 { .tag
= 'M', .mult
= 1 << 20 },
928 { .tag
= 'G', .mult
= 1 << 30 },
935 val
= parse_tag_value(str
, tags
);
936 if (val
!= (unsigned long) -1) {
937 /* we got file size value */
938 pages
= PERF_ALIGN(val
, page_size
) / page_size
;
940 /* we got pages count value */
942 pages
= strtoul(str
, &eptr
, 10);
947 if (pages
== 0 && min
== 0) {
948 /* leave number of pages at 0 */
949 } else if (!is_power_of_2(pages
)) {
950 /* round pages up to next power of 2 */
951 pages
= roundup_pow_of_two(pages
);
954 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
955 pages
* page_size
, pages
);
964 int perf_evlist__parse_mmap_pages(const struct option
*opt
, const char *str
,
965 int unset __maybe_unused
)
967 unsigned int *mmap_pages
= opt
->value
;
968 unsigned long max
= UINT_MAX
;
971 if (max
> SIZE_MAX
/ page_size
)
972 max
= SIZE_MAX
/ page_size
;
974 pages
= parse_pages_arg(str
, 1, max
);
976 pr_err("Invalid argument for --mmap_pages/-m\n");
985 * perf_evlist__mmap - Create mmaps to receive events.
986 * @evlist: list of events
987 * @pages: map length in pages
988 * @overwrite: overwrite older events?
990 * If @overwrite is %false the user needs to signal event consumption using
991 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
994 * Return: %0 on success, negative error code otherwise.
996 int perf_evlist__mmap(struct perf_evlist
*evlist
, unsigned int pages
,
999 struct perf_evsel
*evsel
;
1000 const struct cpu_map
*cpus
= evlist
->cpus
;
1001 const struct thread_map
*threads
= evlist
->threads
;
1002 struct mmap_params mp
= {
1003 .prot
= PROT_READ
| (overwrite
? 0 : PROT_WRITE
),
1006 if (evlist
->mmap
== NULL
&& perf_evlist__alloc_mmap(evlist
) < 0)
1009 if (evlist
->pollfd
.entries
== NULL
&& perf_evlist__alloc_pollfd(evlist
) < 0)
1012 evlist
->overwrite
= overwrite
;
1013 evlist
->mmap_len
= perf_evlist__mmap_size(pages
);
1014 pr_debug("mmap size %zuB\n", evlist
->mmap_len
);
1015 mp
.mask
= evlist
->mmap_len
- page_size
- 1;
1017 evlist__for_each(evlist
, evsel
) {
1018 if ((evsel
->attr
.read_format
& PERF_FORMAT_ID
) &&
1019 evsel
->sample_id
== NULL
&&
1020 perf_evsel__alloc_id(evsel
, cpu_map__nr(cpus
), threads
->nr
) < 0)
1024 if (cpu_map__empty(cpus
))
1025 return perf_evlist__mmap_per_thread(evlist
, &mp
);
1027 return perf_evlist__mmap_per_cpu(evlist
, &mp
);
1030 int perf_evlist__create_maps(struct perf_evlist
*evlist
, struct target
*target
)
1032 evlist
->threads
= thread_map__new_str(target
->pid
, target
->tid
,
1035 if (evlist
->threads
== NULL
)
1038 if (target__uses_dummy_map(target
))
1039 evlist
->cpus
= cpu_map__dummy_new();
1041 evlist
->cpus
= cpu_map__new(target
->cpu_list
);
1043 if (evlist
->cpus
== NULL
)
1044 goto out_delete_threads
;
1049 thread_map__delete(evlist
->threads
);
1050 evlist
->threads
= NULL
;
1054 int perf_evlist__apply_filters(struct perf_evlist
*evlist
)
1056 struct perf_evsel
*evsel
;
1058 const int ncpus
= cpu_map__nr(evlist
->cpus
),
1059 nthreads
= thread_map__nr(evlist
->threads
);
1061 evlist__for_each(evlist
, evsel
) {
1062 if (evsel
->filter
== NULL
)
1065 err
= perf_evsel__set_filter(evsel
, ncpus
, nthreads
, evsel
->filter
);
1073 int perf_evlist__set_filter(struct perf_evlist
*evlist
, const char *filter
)
1075 struct perf_evsel
*evsel
;
1077 const int ncpus
= cpu_map__nr(evlist
->cpus
),
1078 nthreads
= thread_map__nr(evlist
->threads
);
1080 evlist__for_each(evlist
, evsel
) {
1081 err
= perf_evsel__set_filter(evsel
, ncpus
, nthreads
, filter
);
1089 bool perf_evlist__valid_sample_type(struct perf_evlist
*evlist
)
1091 struct perf_evsel
*pos
;
1093 if (evlist
->nr_entries
== 1)
1096 if (evlist
->id_pos
< 0 || evlist
->is_pos
< 0)
1099 evlist__for_each(evlist
, pos
) {
1100 if (pos
->id_pos
!= evlist
->id_pos
||
1101 pos
->is_pos
!= evlist
->is_pos
)
1108 u64
__perf_evlist__combined_sample_type(struct perf_evlist
*evlist
)
1110 struct perf_evsel
*evsel
;
1112 if (evlist
->combined_sample_type
)
1113 return evlist
->combined_sample_type
;
1115 evlist__for_each(evlist
, evsel
)
1116 evlist
->combined_sample_type
|= evsel
->attr
.sample_type
;
1118 return evlist
->combined_sample_type
;
1121 u64
perf_evlist__combined_sample_type(struct perf_evlist
*evlist
)
1123 evlist
->combined_sample_type
= 0;
1124 return __perf_evlist__combined_sample_type(evlist
);
1127 bool perf_evlist__valid_read_format(struct perf_evlist
*evlist
)
1129 struct perf_evsel
*first
= perf_evlist__first(evlist
), *pos
= first
;
1130 u64 read_format
= first
->attr
.read_format
;
1131 u64 sample_type
= first
->attr
.sample_type
;
1133 evlist__for_each(evlist
, pos
) {
1134 if (read_format
!= pos
->attr
.read_format
)
1138 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1139 if ((sample_type
& PERF_SAMPLE_READ
) &&
1140 !(read_format
& PERF_FORMAT_ID
)) {
1147 u64
perf_evlist__read_format(struct perf_evlist
*evlist
)
1149 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1150 return first
->attr
.read_format
;
1153 u16
perf_evlist__id_hdr_size(struct perf_evlist
*evlist
)
1155 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1156 struct perf_sample
*data
;
1160 if (!first
->attr
.sample_id_all
)
1163 sample_type
= first
->attr
.sample_type
;
1165 if (sample_type
& PERF_SAMPLE_TID
)
1166 size
+= sizeof(data
->tid
) * 2;
1168 if (sample_type
& PERF_SAMPLE_TIME
)
1169 size
+= sizeof(data
->time
);
1171 if (sample_type
& PERF_SAMPLE_ID
)
1172 size
+= sizeof(data
->id
);
1174 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1175 size
+= sizeof(data
->stream_id
);
1177 if (sample_type
& PERF_SAMPLE_CPU
)
1178 size
+= sizeof(data
->cpu
) * 2;
1180 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1181 size
+= sizeof(data
->id
);
1186 bool perf_evlist__valid_sample_id_all(struct perf_evlist
*evlist
)
1188 struct perf_evsel
*first
= perf_evlist__first(evlist
), *pos
= first
;
1190 evlist__for_each_continue(evlist
, pos
) {
1191 if (first
->attr
.sample_id_all
!= pos
->attr
.sample_id_all
)
1198 bool perf_evlist__sample_id_all(struct perf_evlist
*evlist
)
1200 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1201 return first
->attr
.sample_id_all
;
1204 void perf_evlist__set_selected(struct perf_evlist
*evlist
,
1205 struct perf_evsel
*evsel
)
1207 evlist
->selected
= evsel
;
1210 void perf_evlist__close(struct perf_evlist
*evlist
)
1212 struct perf_evsel
*evsel
;
1213 int ncpus
= cpu_map__nr(evlist
->cpus
);
1214 int nthreads
= thread_map__nr(evlist
->threads
);
1217 evlist__for_each_reverse(evlist
, evsel
) {
1218 n
= evsel
->cpus
? evsel
->cpus
->nr
: ncpus
;
1219 perf_evsel__close(evsel
, n
, nthreads
);
1223 static int perf_evlist__create_syswide_maps(struct perf_evlist
*evlist
)
1228 * Try reading /sys/devices/system/cpu/online to get
1231 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1232 * code needs an overhaul to properly forward the
1233 * error, and we may not want to do that fallback to a
1234 * default cpu identity map :-\
1236 evlist
->cpus
= cpu_map__new(NULL
);
1237 if (evlist
->cpus
== NULL
)
1240 evlist
->threads
= thread_map__new_dummy();
1241 if (evlist
->threads
== NULL
)
1248 cpu_map__delete(evlist
->cpus
);
1249 evlist
->cpus
= NULL
;
1253 int perf_evlist__open(struct perf_evlist
*evlist
)
1255 struct perf_evsel
*evsel
;
1259 * Default: one fd per CPU, all threads, aka systemwide
1260 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1262 if (evlist
->threads
== NULL
&& evlist
->cpus
== NULL
) {
1263 err
= perf_evlist__create_syswide_maps(evlist
);
1268 perf_evlist__update_id_pos(evlist
);
1270 evlist__for_each(evlist
, evsel
) {
1271 err
= perf_evsel__open(evsel
, evlist
->cpus
, evlist
->threads
);
1278 perf_evlist__close(evlist
);
1283 int perf_evlist__prepare_workload(struct perf_evlist
*evlist
, struct target
*target
,
1284 const char *argv
[], bool pipe_output
,
1285 void (*exec_error
)(int signo
, siginfo_t
*info
, void *ucontext
))
1287 int child_ready_pipe
[2], go_pipe
[2];
1290 if (pipe(child_ready_pipe
) < 0) {
1291 perror("failed to create 'ready' pipe");
1295 if (pipe(go_pipe
) < 0) {
1296 perror("failed to create 'go' pipe");
1297 goto out_close_ready_pipe
;
1300 evlist
->workload
.pid
= fork();
1301 if (evlist
->workload
.pid
< 0) {
1302 perror("failed to fork");
1303 goto out_close_pipes
;
1306 if (!evlist
->workload
.pid
) {
1312 signal(SIGTERM
, SIG_DFL
);
1314 close(child_ready_pipe
[0]);
1316 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
1319 * Tell the parent we're ready to go
1321 close(child_ready_pipe
[1]);
1324 * Wait until the parent tells us to go.
1326 ret
= read(go_pipe
[0], &bf
, 1);
1328 * The parent will ask for the execvp() to be performed by
1329 * writing exactly one byte, in workload.cork_fd, usually via
1330 * perf_evlist__start_workload().
1332 * For cancelling the workload without actuallin running it,
1333 * the parent will just close workload.cork_fd, without writing
1334 * anything, i.e. read will return zero and we just exit()
1339 perror("unable to read pipe");
1343 execvp(argv
[0], (char **)argv
);
1348 val
.sival_int
= errno
;
1349 if (sigqueue(getppid(), SIGUSR1
, val
))
1357 struct sigaction act
= {
1358 .sa_flags
= SA_SIGINFO
,
1359 .sa_sigaction
= exec_error
,
1361 sigaction(SIGUSR1
, &act
, NULL
);
1364 if (target__none(target
)) {
1365 if (evlist
->threads
== NULL
) {
1366 fprintf(stderr
, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1367 __func__
, __LINE__
);
1368 goto out_close_pipes
;
1370 evlist
->threads
->map
[0] = evlist
->workload
.pid
;
1373 close(child_ready_pipe
[1]);
1376 * wait for child to settle
1378 if (read(child_ready_pipe
[0], &bf
, 1) == -1) {
1379 perror("unable to read pipe");
1380 goto out_close_pipes
;
1383 fcntl(go_pipe
[1], F_SETFD
, FD_CLOEXEC
);
1384 evlist
->workload
.cork_fd
= go_pipe
[1];
1385 close(child_ready_pipe
[0]);
1391 out_close_ready_pipe
:
1392 close(child_ready_pipe
[0]);
1393 close(child_ready_pipe
[1]);
1397 int perf_evlist__start_workload(struct perf_evlist
*evlist
)
1399 if (evlist
->workload
.cork_fd
> 0) {
1403 * Remove the cork, let it rip!
1405 ret
= write(evlist
->workload
.cork_fd
, &bf
, 1);
1407 perror("enable to write to pipe");
1409 close(evlist
->workload
.cork_fd
);
1416 int perf_evlist__parse_sample(struct perf_evlist
*evlist
, union perf_event
*event
,
1417 struct perf_sample
*sample
)
1419 struct perf_evsel
*evsel
= perf_evlist__event2evsel(evlist
, event
);
1423 return perf_evsel__parse_sample(evsel
, event
, sample
);
1426 size_t perf_evlist__fprintf(struct perf_evlist
*evlist
, FILE *fp
)
1428 struct perf_evsel
*evsel
;
1431 evlist__for_each(evlist
, evsel
) {
1432 printed
+= fprintf(fp
, "%s%s", evsel
->idx
? ", " : "",
1433 perf_evsel__name(evsel
));
1436 return printed
+ fprintf(fp
, "\n");
1439 int perf_evlist__strerror_tp(struct perf_evlist
*evlist __maybe_unused
,
1440 int err
, char *buf
, size_t size
)
1446 scnprintf(buf
, size
, "%s",
1447 "Error:\tUnable to find debugfs\n"
1448 "Hint:\tWas your kernel was compiled with debugfs support?\n"
1449 "Hint:\tIs the debugfs filesystem mounted?\n"
1450 "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1453 scnprintf(buf
, size
,
1454 "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1455 "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1456 debugfs_mountpoint
, debugfs_mountpoint
);
1459 scnprintf(buf
, size
, "%s", strerror_r(err
, sbuf
, sizeof(sbuf
)));
1466 int perf_evlist__strerror_open(struct perf_evlist
*evlist __maybe_unused
,
1467 int err
, char *buf
, size_t size
)
1470 char sbuf
[STRERR_BUFSIZE
], *emsg
= strerror_r(err
, sbuf
, sizeof(sbuf
));
1475 printed
= scnprintf(buf
, size
,
1477 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg
);
1479 value
= perf_event_paranoid();
1481 printed
+= scnprintf(buf
+ printed
, size
- printed
, "\nHint:\t");
1484 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1485 "For your workloads it needs to be <= 1\nHint:\t");
1487 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1488 "For system wide tracing it needs to be set to -1.\n");
1490 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1491 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1492 "Hint:\tThe current value is %d.", value
);
1495 scnprintf(buf
, size
, "%s", emsg
);
1502 int perf_evlist__strerror_mmap(struct perf_evlist
*evlist
, int err
, char *buf
, size_t size
)
1504 char sbuf
[STRERR_BUFSIZE
], *emsg
= strerror_r(err
, sbuf
, sizeof(sbuf
));
1505 int pages_attempted
= evlist
->mmap_len
/ 1024, pages_max_per_user
, printed
= 0;
1509 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user
);
1510 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1512 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1513 "Hint:\tTried using %zd kB.\n",
1514 emsg
, pages_max_per_user
, pages_attempted
);
1516 if (pages_attempted
>= pages_max_per_user
) {
1517 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1518 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1519 pages_max_per_user
+ pages_attempted
);
1522 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1523 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1526 scnprintf(buf
, size
, "%s", emsg
);
1533 void perf_evlist__to_front(struct perf_evlist
*evlist
,
1534 struct perf_evsel
*move_evsel
)
1536 struct perf_evsel
*evsel
, *n
;
1539 if (move_evsel
== perf_evlist__first(evlist
))
1542 evlist__for_each_safe(evlist
, n
, evsel
) {
1543 if (evsel
->leader
== move_evsel
->leader
)
1544 list_move_tail(&evsel
->node
, &move
);
1547 list_splice(&move
, &evlist
->entries
);
1550 void perf_evlist__set_tracking_event(struct perf_evlist
*evlist
,
1551 struct perf_evsel
*tracking_evsel
)
1553 struct perf_evsel
*evsel
;
1555 if (tracking_evsel
->tracking
)
1558 evlist__for_each(evlist
, evsel
) {
1559 if (evsel
!= tracking_evsel
)
1560 evsel
->tracking
= false;
1563 tracking_evsel
->tracking
= true;