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>
13 #include "thread_map.h"
20 #include "parse-events.h"
21 #include "parse-options.h"
25 #include <linux/bitops.h>
26 #include <linux/hash.h>
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
);
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)
34 void perf_evlist__init(struct perf_evlist
*evlist
, struct cpu_map
*cpus
,
35 struct thread_map
*threads
)
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;
47 struct perf_evlist
*perf_evlist__new(void)
49 struct perf_evlist
*evlist
= zalloc(sizeof(*evlist
));
52 perf_evlist__init(evlist
, NULL
, NULL
);
57 struct perf_evlist
*perf_evlist__new_default(void)
59 struct perf_evlist
*evlist
= perf_evlist__new();
61 if (evlist
&& perf_evlist__add_default(evlist
)) {
62 perf_evlist__delete(evlist
);
70 * perf_evlist__set_id_pos - set the positions of event ids.
71 * @evlist: selected event list
73 * Events with compatible sample types all have the same id_pos
74 * and is_pos. For convenience, put a copy on evlist.
76 void perf_evlist__set_id_pos(struct perf_evlist
*evlist
)
78 struct perf_evsel
*first
= perf_evlist__first(evlist
);
80 evlist
->id_pos
= first
->id_pos
;
81 evlist
->is_pos
= first
->is_pos
;
84 static void perf_evlist__update_id_pos(struct perf_evlist
*evlist
)
86 struct perf_evsel
*evsel
;
88 evlist__for_each(evlist
, evsel
)
89 perf_evsel__calc_id_pos(evsel
);
91 perf_evlist__set_id_pos(evlist
);
94 static void perf_evlist__purge(struct perf_evlist
*evlist
)
96 struct perf_evsel
*pos
, *n
;
98 evlist__for_each_safe(evlist
, n
, pos
) {
99 list_del_init(&pos
->node
);
100 perf_evsel__delete(pos
);
103 evlist
->nr_entries
= 0;
106 void perf_evlist__exit(struct perf_evlist
*evlist
)
108 zfree(&evlist
->mmap
);
109 fdarray__exit(&evlist
->pollfd
);
112 void perf_evlist__delete(struct perf_evlist
*evlist
)
114 perf_evlist__munmap(evlist
);
115 perf_evlist__close(evlist
);
116 cpu_map__delete(evlist
->cpus
);
117 thread_map__delete(evlist
->threads
);
119 evlist
->threads
= NULL
;
120 perf_evlist__purge(evlist
);
121 perf_evlist__exit(evlist
);
125 void perf_evlist__add(struct perf_evlist
*evlist
, struct perf_evsel
*entry
)
127 list_add_tail(&entry
->node
, &evlist
->entries
);
128 entry
->idx
= evlist
->nr_entries
;
129 entry
->tracking
= !entry
->idx
;
131 if (!evlist
->nr_entries
++)
132 perf_evlist__set_id_pos(evlist
);
135 void perf_evlist__splice_list_tail(struct perf_evlist
*evlist
,
136 struct list_head
*list
,
139 bool set_id_pos
= !evlist
->nr_entries
;
141 list_splice_tail(list
, &evlist
->entries
);
142 evlist
->nr_entries
+= nr_entries
;
144 perf_evlist__set_id_pos(evlist
);
147 void __perf_evlist__set_leader(struct list_head
*list
)
149 struct perf_evsel
*evsel
, *leader
;
151 leader
= list_entry(list
->next
, struct perf_evsel
, node
);
152 evsel
= list_entry(list
->prev
, struct perf_evsel
, node
);
154 leader
->nr_members
= evsel
->idx
- leader
->idx
+ 1;
156 __evlist__for_each(list
, evsel
) {
157 evsel
->leader
= leader
;
161 void perf_evlist__set_leader(struct perf_evlist
*evlist
)
163 if (evlist
->nr_entries
) {
164 evlist
->nr_groups
= evlist
->nr_entries
> 1 ? 1 : 0;
165 __perf_evlist__set_leader(&evlist
->entries
);
169 int perf_evlist__add_default(struct perf_evlist
*evlist
)
171 struct perf_event_attr attr
= {
172 .type
= PERF_TYPE_HARDWARE
,
173 .config
= PERF_COUNT_HW_CPU_CYCLES
,
175 struct perf_evsel
*evsel
;
177 event_attr_init(&attr
);
179 evsel
= perf_evsel__new(&attr
);
183 /* use strdup() because free(evsel) assumes name is allocated */
184 evsel
->name
= strdup("cycles");
188 perf_evlist__add(evlist
, evsel
);
191 perf_evsel__delete(evsel
);
196 static int perf_evlist__add_attrs(struct perf_evlist
*evlist
,
197 struct perf_event_attr
*attrs
, size_t nr_attrs
)
199 struct perf_evsel
*evsel
, *n
;
203 for (i
= 0; i
< nr_attrs
; i
++) {
204 evsel
= perf_evsel__new_idx(attrs
+ i
, evlist
->nr_entries
+ i
);
206 goto out_delete_partial_list
;
207 list_add_tail(&evsel
->node
, &head
);
210 perf_evlist__splice_list_tail(evlist
, &head
, nr_attrs
);
214 out_delete_partial_list
:
215 __evlist__for_each_safe(&head
, n
, evsel
)
216 perf_evsel__delete(evsel
);
220 int __perf_evlist__add_default_attrs(struct perf_evlist
*evlist
,
221 struct perf_event_attr
*attrs
, size_t nr_attrs
)
225 for (i
= 0; i
< nr_attrs
; i
++)
226 event_attr_init(attrs
+ i
);
228 return perf_evlist__add_attrs(evlist
, attrs
, nr_attrs
);
232 perf_evlist__find_tracepoint_by_id(struct perf_evlist
*evlist
, int id
)
234 struct perf_evsel
*evsel
;
236 evlist__for_each(evlist
, evsel
) {
237 if (evsel
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
238 (int)evsel
->attr
.config
== id
)
246 perf_evlist__find_tracepoint_by_name(struct perf_evlist
*evlist
,
249 struct perf_evsel
*evsel
;
251 evlist__for_each(evlist
, evsel
) {
252 if ((evsel
->attr
.type
== PERF_TYPE_TRACEPOINT
) &&
253 (strcmp(evsel
->name
, name
) == 0))
260 int perf_evlist__add_newtp(struct perf_evlist
*evlist
,
261 const char *sys
, const char *name
, void *handler
)
263 struct perf_evsel
*evsel
= perf_evsel__newtp(sys
, name
);
268 evsel
->handler
= handler
;
269 perf_evlist__add(evlist
, evsel
);
273 static int perf_evlist__nr_threads(struct perf_evlist
*evlist
,
274 struct perf_evsel
*evsel
)
276 if (evsel
->system_wide
)
279 return thread_map__nr(evlist
->threads
);
282 void perf_evlist__disable(struct perf_evlist
*evlist
)
285 struct perf_evsel
*pos
;
286 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
289 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
290 evlist__for_each(evlist
, pos
) {
291 if (!perf_evsel__is_group_leader(pos
) || !pos
->fd
)
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);
301 void perf_evlist__enable(struct perf_evlist
*evlist
)
304 struct perf_evsel
*pos
;
305 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
308 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
309 evlist__for_each(evlist
, pos
) {
310 if (!perf_evsel__is_group_leader(pos
) || !pos
->fd
)
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);
320 int perf_evlist__disable_event(struct perf_evlist
*evlist
,
321 struct perf_evsel
*evsel
)
323 int cpu
, thread
, err
;
324 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
325 int nr_threads
= perf_evlist__nr_threads(evlist
, evsel
);
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);
341 int perf_evlist__enable_event(struct perf_evlist
*evlist
,
342 struct perf_evsel
*evsel
)
344 int cpu
, thread
, err
;
345 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
346 int nr_threads
= perf_evlist__nr_threads(evlist
, evsel
);
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);
362 static int perf_evlist__enable_event_cpu(struct perf_evlist
*evlist
,
363 struct perf_evsel
*evsel
, int cpu
)
366 int nr_threads
= perf_evlist__nr_threads(evlist
, evsel
);
371 for (thread
= 0; thread
< nr_threads
; thread
++) {
372 err
= ioctl(FD(evsel
, cpu
, thread
),
373 PERF_EVENT_IOC_ENABLE
, 0);
380 static int perf_evlist__enable_event_thread(struct perf_evlist
*evlist
,
381 struct perf_evsel
*evsel
,
385 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
390 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
391 err
= ioctl(FD(evsel
, cpu
, thread
), PERF_EVENT_IOC_ENABLE
, 0);
398 int perf_evlist__enable_event_idx(struct perf_evlist
*evlist
,
399 struct perf_evsel
*evsel
, int idx
)
401 bool per_cpu_mmaps
= !cpu_map__empty(evlist
->cpus
);
404 return perf_evlist__enable_event_cpu(evlist
, evsel
, idx
);
406 return perf_evlist__enable_event_thread(evlist
, evsel
, idx
);
409 int perf_evlist__alloc_pollfd(struct perf_evlist
*evlist
)
411 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
412 int nr_threads
= thread_map__nr(evlist
->threads
);
414 struct perf_evsel
*evsel
;
416 evlist__for_each(evlist
, evsel
) {
417 if (evsel
->system_wide
)
420 nfds
+= nr_cpus
* nr_threads
;
423 if (fdarray__available_entries(&evlist
->pollfd
) < nfds
&&
424 fdarray__grow(&evlist
->pollfd
, nfds
) < 0)
430 static int __perf_evlist__add_pollfd(struct perf_evlist
*evlist
, int fd
, int idx
)
432 int pos
= fdarray__add(&evlist
->pollfd
, fd
, POLLIN
| POLLERR
| POLLHUP
);
434 * Save the idx so that when we filter out fds POLLHUP'ed we can
435 * close the associated evlist->mmap[] entry.
438 evlist
->pollfd
.priv
[pos
].idx
= idx
;
440 fcntl(fd
, F_SETFL
, O_NONBLOCK
);
446 int perf_evlist__add_pollfd(struct perf_evlist
*evlist
, int fd
)
448 return __perf_evlist__add_pollfd(evlist
, fd
, -1);
451 static void perf_evlist__munmap_filtered(struct fdarray
*fda
, int fd
)
453 struct perf_evlist
*evlist
= container_of(fda
, struct perf_evlist
, pollfd
);
455 perf_evlist__mmap_put(evlist
, fda
->priv
[fd
].idx
);
458 int perf_evlist__filter_pollfd(struct perf_evlist
*evlist
, short revents_and_mask
)
460 return fdarray__filter(&evlist
->pollfd
, revents_and_mask
,
461 perf_evlist__munmap_filtered
);
464 int perf_evlist__poll(struct perf_evlist
*evlist
, int timeout
)
466 return fdarray__poll(&evlist
->pollfd
, timeout
);
469 static void perf_evlist__id_hash(struct perf_evlist
*evlist
,
470 struct perf_evsel
*evsel
,
471 int cpu
, int thread
, u64 id
)
474 struct perf_sample_id
*sid
= SID(evsel
, cpu
, thread
);
478 hash
= hash_64(sid
->id
, PERF_EVLIST__HLIST_BITS
);
479 hlist_add_head(&sid
->node
, &evlist
->heads
[hash
]);
482 void perf_evlist__id_add(struct perf_evlist
*evlist
, struct perf_evsel
*evsel
,
483 int cpu
, int thread
, u64 id
)
485 perf_evlist__id_hash(evlist
, evsel
, cpu
, thread
, id
);
486 evsel
->id
[evsel
->ids
++] = id
;
489 static int perf_evlist__id_add_fd(struct perf_evlist
*evlist
,
490 struct perf_evsel
*evsel
,
491 int cpu
, int thread
, int fd
)
493 u64 read_data
[4] = { 0, };
494 int id_idx
= 1; /* The first entry is the counter value */
498 ret
= ioctl(fd
, PERF_EVENT_IOC_ID
, &id
);
505 /* Legacy way to get event id.. All hail to old kernels! */
508 * This way does not work with group format read, so bail
511 if (perf_evlist__read_format(evlist
) & PERF_FORMAT_GROUP
)
514 if (!(evsel
->attr
.read_format
& PERF_FORMAT_ID
) ||
515 read(fd
, &read_data
, sizeof(read_data
)) == -1)
518 if (evsel
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
520 if (evsel
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
523 id
= read_data
[id_idx
];
526 perf_evlist__id_add(evlist
, evsel
, cpu
, thread
, id
);
530 static void perf_evlist__set_sid_idx(struct perf_evlist
*evlist
,
531 struct perf_evsel
*evsel
, int idx
, int cpu
,
534 struct perf_sample_id
*sid
= SID(evsel
, cpu
, thread
);
536 if (evlist
->cpus
&& cpu
>= 0)
537 sid
->cpu
= evlist
->cpus
->map
[cpu
];
540 if (!evsel
->system_wide
&& evlist
->threads
&& thread
>= 0)
541 sid
->tid
= evlist
->threads
->map
[thread
];
546 struct perf_sample_id
*perf_evlist__id2sid(struct perf_evlist
*evlist
, u64 id
)
548 struct hlist_head
*head
;
549 struct perf_sample_id
*sid
;
552 hash
= hash_64(id
, PERF_EVLIST__HLIST_BITS
);
553 head
= &evlist
->heads
[hash
];
555 hlist_for_each_entry(sid
, head
, node
)
562 struct perf_evsel
*perf_evlist__id2evsel(struct perf_evlist
*evlist
, u64 id
)
564 struct perf_sample_id
*sid
;
566 if (evlist
->nr_entries
== 1)
567 return perf_evlist__first(evlist
);
569 sid
= perf_evlist__id2sid(evlist
, id
);
573 if (!perf_evlist__sample_id_all(evlist
))
574 return perf_evlist__first(evlist
);
579 static int perf_evlist__event2id(struct perf_evlist
*evlist
,
580 union perf_event
*event
, u64
*id
)
582 const u64
*array
= event
->sample
.array
;
585 n
= (event
->header
.size
- sizeof(event
->header
)) >> 3;
587 if (event
->header
.type
== PERF_RECORD_SAMPLE
) {
588 if (evlist
->id_pos
>= n
)
590 *id
= array
[evlist
->id_pos
];
592 if (evlist
->is_pos
> n
)
600 static struct perf_evsel
*perf_evlist__event2evsel(struct perf_evlist
*evlist
,
601 union perf_event
*event
)
603 struct perf_evsel
*first
= perf_evlist__first(evlist
);
604 struct hlist_head
*head
;
605 struct perf_sample_id
*sid
;
609 if (evlist
->nr_entries
== 1)
612 if (!first
->attr
.sample_id_all
&&
613 event
->header
.type
!= PERF_RECORD_SAMPLE
)
616 if (perf_evlist__event2id(evlist
, event
, &id
))
619 /* Synthesized events have an id of zero */
623 hash
= hash_64(id
, PERF_EVLIST__HLIST_BITS
);
624 head
= &evlist
->heads
[hash
];
626 hlist_for_each_entry(sid
, head
, node
) {
633 union perf_event
*perf_evlist__mmap_read(struct perf_evlist
*evlist
, int idx
)
635 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
636 unsigned int head
= perf_mmap__read_head(md
);
637 unsigned int old
= md
->prev
;
638 unsigned char *data
= md
->base
+ page_size
;
639 union perf_event
*event
= NULL
;
641 if (evlist
->overwrite
) {
643 * If we're further behind than half the buffer, there's a chance
644 * the writer will bite our tail and mess up the samples under us.
646 * If we somehow ended up ahead of the head, we got messed up.
648 * In either case, truncate and restart at head.
650 int diff
= head
- old
;
651 if (diff
> md
->mask
/ 2 || diff
< 0) {
652 fprintf(stderr
, "WARNING: failed to keep up with mmap data.\n");
655 * head points to a known good entry, start there.
664 event
= (union perf_event
*)&data
[old
& md
->mask
];
665 size
= event
->header
.size
;
668 * Event straddles the mmap boundary -- header should always
669 * be inside due to u64 alignment of output.
671 if ((old
& md
->mask
) + size
!= ((old
+ size
) & md
->mask
)) {
672 unsigned int offset
= old
;
673 unsigned int len
= min(sizeof(*event
), size
), cpy
;
674 void *dst
= md
->event_copy
;
677 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
678 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
684 event
= (union perf_event
*) md
->event_copy
;
695 static bool perf_mmap__empty(struct perf_mmap
*md
)
697 return perf_mmap__read_head(md
) != md
->prev
;
700 static void perf_evlist__mmap_get(struct perf_evlist
*evlist
, int idx
)
702 ++evlist
->mmap
[idx
].refcnt
;
705 static void perf_evlist__mmap_put(struct perf_evlist
*evlist
, int idx
)
707 BUG_ON(evlist
->mmap
[idx
].refcnt
== 0);
709 if (--evlist
->mmap
[idx
].refcnt
== 0)
710 __perf_evlist__munmap(evlist
, idx
);
713 void perf_evlist__mmap_consume(struct perf_evlist
*evlist
, int idx
)
715 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
717 if (!evlist
->overwrite
) {
718 unsigned int old
= md
->prev
;
720 perf_mmap__write_tail(md
, old
);
723 if (md
->refcnt
== 1 && perf_mmap__empty(md
))
724 perf_evlist__mmap_put(evlist
, idx
);
727 static void __perf_evlist__munmap(struct perf_evlist
*evlist
, int idx
)
729 if (evlist
->mmap
[idx
].base
!= NULL
) {
730 munmap(evlist
->mmap
[idx
].base
, evlist
->mmap_len
);
731 evlist
->mmap
[idx
].base
= NULL
;
732 evlist
->mmap
[idx
].refcnt
= 0;
736 void perf_evlist__munmap(struct perf_evlist
*evlist
)
740 if (evlist
->mmap
== NULL
)
743 for (i
= 0; i
< evlist
->nr_mmaps
; i
++)
744 __perf_evlist__munmap(evlist
, i
);
746 zfree(&evlist
->mmap
);
749 static int perf_evlist__alloc_mmap(struct perf_evlist
*evlist
)
751 evlist
->nr_mmaps
= cpu_map__nr(evlist
->cpus
);
752 if (cpu_map__empty(evlist
->cpus
))
753 evlist
->nr_mmaps
= thread_map__nr(evlist
->threads
);
754 evlist
->mmap
= zalloc(evlist
->nr_mmaps
* sizeof(struct perf_mmap
));
755 return evlist
->mmap
!= NULL
? 0 : -ENOMEM
;
763 static int __perf_evlist__mmap(struct perf_evlist
*evlist
, int idx
,
764 struct mmap_params
*mp
, int fd
)
767 * The last one will be done at perf_evlist__mmap_consume(), so that we
768 * make sure we don't prevent tools from consuming every last event in
771 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
772 * anymore, but the last events for it are still in the ring buffer,
773 * waiting to be consumed.
775 * Tools can chose to ignore this at their own discretion, but the
776 * evlist layer can't just drop it when filtering events in
777 * perf_evlist__filter_pollfd().
779 evlist
->mmap
[idx
].refcnt
= 2;
780 evlist
->mmap
[idx
].prev
= 0;
781 evlist
->mmap
[idx
].mask
= mp
->mask
;
782 evlist
->mmap
[idx
].base
= mmap(NULL
, evlist
->mmap_len
, mp
->prot
,
784 if (evlist
->mmap
[idx
].base
== MAP_FAILED
) {
785 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
787 evlist
->mmap
[idx
].base
= NULL
;
794 static int perf_evlist__mmap_per_evsel(struct perf_evlist
*evlist
, int idx
,
795 struct mmap_params
*mp
, int cpu
,
796 int thread
, int *output
)
798 struct perf_evsel
*evsel
;
800 evlist__for_each(evlist
, evsel
) {
803 if (evsel
->system_wide
&& thread
)
806 fd
= FD(evsel
, cpu
, thread
);
810 if (__perf_evlist__mmap(evlist
, idx
, mp
, *output
) < 0)
813 if (ioctl(fd
, PERF_EVENT_IOC_SET_OUTPUT
, *output
) != 0)
816 perf_evlist__mmap_get(evlist
, idx
);
819 if (__perf_evlist__add_pollfd(evlist
, fd
, idx
) < 0) {
820 perf_evlist__mmap_put(evlist
, idx
);
824 if (evsel
->attr
.read_format
& PERF_FORMAT_ID
) {
825 if (perf_evlist__id_add_fd(evlist
, evsel
, cpu
, thread
,
828 perf_evlist__set_sid_idx(evlist
, evsel
, idx
, cpu
,
836 static int perf_evlist__mmap_per_cpu(struct perf_evlist
*evlist
,
837 struct mmap_params
*mp
)
840 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
841 int nr_threads
= thread_map__nr(evlist
->threads
);
843 pr_debug2("perf event ring buffer mmapped per cpu\n");
844 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
847 for (thread
= 0; thread
< nr_threads
; thread
++) {
848 if (perf_evlist__mmap_per_evsel(evlist
, cpu
, mp
, cpu
,
857 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
858 __perf_evlist__munmap(evlist
, cpu
);
862 static int perf_evlist__mmap_per_thread(struct perf_evlist
*evlist
,
863 struct mmap_params
*mp
)
866 int nr_threads
= thread_map__nr(evlist
->threads
);
868 pr_debug2("perf event ring buffer mmapped per thread\n");
869 for (thread
= 0; thread
< nr_threads
; thread
++) {
872 if (perf_evlist__mmap_per_evsel(evlist
, thread
, mp
, 0, thread
,
880 for (thread
= 0; thread
< nr_threads
; thread
++)
881 __perf_evlist__munmap(evlist
, thread
);
885 static size_t perf_evlist__mmap_size(unsigned long pages
)
887 /* 512 kiB: default amount of unprivileged mlocked memory */
888 if (pages
== UINT_MAX
)
889 pages
= (512 * 1024) / page_size
;
890 else if (!is_power_of_2(pages
))
893 return (pages
+ 1) * page_size
;
896 static long parse_pages_arg(const char *str
, unsigned long min
,
899 unsigned long pages
, val
;
900 static struct parse_tag tags
[] = {
901 { .tag
= 'B', .mult
= 1 },
902 { .tag
= 'K', .mult
= 1 << 10 },
903 { .tag
= 'M', .mult
= 1 << 20 },
904 { .tag
= 'G', .mult
= 1 << 30 },
911 val
= parse_tag_value(str
, tags
);
912 if (val
!= (unsigned long) -1) {
913 /* we got file size value */
914 pages
= PERF_ALIGN(val
, page_size
) / page_size
;
916 /* we got pages count value */
918 pages
= strtoul(str
, &eptr
, 10);
923 if (pages
== 0 && min
== 0) {
924 /* leave number of pages at 0 */
925 } else if (!is_power_of_2(pages
)) {
926 /* round pages up to next power of 2 */
927 pages
= next_pow2_l(pages
);
930 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
931 pages
* page_size
, pages
);
940 int perf_evlist__parse_mmap_pages(const struct option
*opt
, const char *str
,
941 int unset __maybe_unused
)
943 unsigned int *mmap_pages
= opt
->value
;
944 unsigned long max
= UINT_MAX
;
947 if (max
> SIZE_MAX
/ page_size
)
948 max
= SIZE_MAX
/ page_size
;
950 pages
= parse_pages_arg(str
, 1, max
);
952 pr_err("Invalid argument for --mmap_pages/-m\n");
961 * perf_evlist__mmap - Create mmaps to receive events.
962 * @evlist: list of events
963 * @pages: map length in pages
964 * @overwrite: overwrite older events?
966 * If @overwrite is %false the user needs to signal event consumption using
967 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
970 * Return: %0 on success, negative error code otherwise.
972 int perf_evlist__mmap(struct perf_evlist
*evlist
, unsigned int pages
,
975 struct perf_evsel
*evsel
;
976 const struct cpu_map
*cpus
= evlist
->cpus
;
977 const struct thread_map
*threads
= evlist
->threads
;
978 struct mmap_params mp
= {
979 .prot
= PROT_READ
| (overwrite
? 0 : PROT_WRITE
),
982 if (evlist
->mmap
== NULL
&& perf_evlist__alloc_mmap(evlist
) < 0)
985 if (evlist
->pollfd
.entries
== NULL
&& perf_evlist__alloc_pollfd(evlist
) < 0)
988 evlist
->overwrite
= overwrite
;
989 evlist
->mmap_len
= perf_evlist__mmap_size(pages
);
990 pr_debug("mmap size %zuB\n", evlist
->mmap_len
);
991 mp
.mask
= evlist
->mmap_len
- page_size
- 1;
993 evlist__for_each(evlist
, evsel
) {
994 if ((evsel
->attr
.read_format
& PERF_FORMAT_ID
) &&
995 evsel
->sample_id
== NULL
&&
996 perf_evsel__alloc_id(evsel
, cpu_map__nr(cpus
), threads
->nr
) < 0)
1000 if (cpu_map__empty(cpus
))
1001 return perf_evlist__mmap_per_thread(evlist
, &mp
);
1003 return perf_evlist__mmap_per_cpu(evlist
, &mp
);
1006 int perf_evlist__create_maps(struct perf_evlist
*evlist
, struct target
*target
)
1008 evlist
->threads
= thread_map__new_str(target
->pid
, target
->tid
,
1011 if (evlist
->threads
== NULL
)
1014 if (target__uses_dummy_map(target
))
1015 evlist
->cpus
= cpu_map__dummy_new();
1017 evlist
->cpus
= cpu_map__new(target
->cpu_list
);
1019 if (evlist
->cpus
== NULL
)
1020 goto out_delete_threads
;
1025 thread_map__delete(evlist
->threads
);
1026 evlist
->threads
= NULL
;
1030 int perf_evlist__apply_filters(struct perf_evlist
*evlist
)
1032 struct perf_evsel
*evsel
;
1034 const int ncpus
= cpu_map__nr(evlist
->cpus
),
1035 nthreads
= thread_map__nr(evlist
->threads
);
1037 evlist__for_each(evlist
, evsel
) {
1038 if (evsel
->filter
== NULL
)
1041 err
= perf_evsel__set_filter(evsel
, ncpus
, nthreads
, evsel
->filter
);
1049 int perf_evlist__set_filter(struct perf_evlist
*evlist
, const char *filter
)
1051 struct perf_evsel
*evsel
;
1053 const int ncpus
= cpu_map__nr(evlist
->cpus
),
1054 nthreads
= thread_map__nr(evlist
->threads
);
1056 evlist__for_each(evlist
, evsel
) {
1057 err
= perf_evsel__set_filter(evsel
, ncpus
, nthreads
, filter
);
1065 bool perf_evlist__valid_sample_type(struct perf_evlist
*evlist
)
1067 struct perf_evsel
*pos
;
1069 if (evlist
->nr_entries
== 1)
1072 if (evlist
->id_pos
< 0 || evlist
->is_pos
< 0)
1075 evlist__for_each(evlist
, pos
) {
1076 if (pos
->id_pos
!= evlist
->id_pos
||
1077 pos
->is_pos
!= evlist
->is_pos
)
1084 u64
__perf_evlist__combined_sample_type(struct perf_evlist
*evlist
)
1086 struct perf_evsel
*evsel
;
1088 if (evlist
->combined_sample_type
)
1089 return evlist
->combined_sample_type
;
1091 evlist__for_each(evlist
, evsel
)
1092 evlist
->combined_sample_type
|= evsel
->attr
.sample_type
;
1094 return evlist
->combined_sample_type
;
1097 u64
perf_evlist__combined_sample_type(struct perf_evlist
*evlist
)
1099 evlist
->combined_sample_type
= 0;
1100 return __perf_evlist__combined_sample_type(evlist
);
1103 bool perf_evlist__valid_read_format(struct perf_evlist
*evlist
)
1105 struct perf_evsel
*first
= perf_evlist__first(evlist
), *pos
= first
;
1106 u64 read_format
= first
->attr
.read_format
;
1107 u64 sample_type
= first
->attr
.sample_type
;
1109 evlist__for_each(evlist
, pos
) {
1110 if (read_format
!= pos
->attr
.read_format
)
1114 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1115 if ((sample_type
& PERF_SAMPLE_READ
) &&
1116 !(read_format
& PERF_FORMAT_ID
)) {
1123 u64
perf_evlist__read_format(struct perf_evlist
*evlist
)
1125 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1126 return first
->attr
.read_format
;
1129 u16
perf_evlist__id_hdr_size(struct perf_evlist
*evlist
)
1131 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1132 struct perf_sample
*data
;
1136 if (!first
->attr
.sample_id_all
)
1139 sample_type
= first
->attr
.sample_type
;
1141 if (sample_type
& PERF_SAMPLE_TID
)
1142 size
+= sizeof(data
->tid
) * 2;
1144 if (sample_type
& PERF_SAMPLE_TIME
)
1145 size
+= sizeof(data
->time
);
1147 if (sample_type
& PERF_SAMPLE_ID
)
1148 size
+= sizeof(data
->id
);
1150 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1151 size
+= sizeof(data
->stream_id
);
1153 if (sample_type
& PERF_SAMPLE_CPU
)
1154 size
+= sizeof(data
->cpu
) * 2;
1156 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1157 size
+= sizeof(data
->id
);
1162 bool perf_evlist__valid_sample_id_all(struct perf_evlist
*evlist
)
1164 struct perf_evsel
*first
= perf_evlist__first(evlist
), *pos
= first
;
1166 evlist__for_each_continue(evlist
, pos
) {
1167 if (first
->attr
.sample_id_all
!= pos
->attr
.sample_id_all
)
1174 bool perf_evlist__sample_id_all(struct perf_evlist
*evlist
)
1176 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1177 return first
->attr
.sample_id_all
;
1180 void perf_evlist__set_selected(struct perf_evlist
*evlist
,
1181 struct perf_evsel
*evsel
)
1183 evlist
->selected
= evsel
;
1186 void perf_evlist__close(struct perf_evlist
*evlist
)
1188 struct perf_evsel
*evsel
;
1189 int ncpus
= cpu_map__nr(evlist
->cpus
);
1190 int nthreads
= thread_map__nr(evlist
->threads
);
1193 evlist__for_each_reverse(evlist
, evsel
) {
1194 n
= evsel
->cpus
? evsel
->cpus
->nr
: ncpus
;
1195 perf_evsel__close(evsel
, n
, nthreads
);
1199 static int perf_evlist__create_syswide_maps(struct perf_evlist
*evlist
)
1204 * Try reading /sys/devices/system/cpu/online to get
1207 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1208 * code needs an overhaul to properly forward the
1209 * error, and we may not want to do that fallback to a
1210 * default cpu identity map :-\
1212 evlist
->cpus
= cpu_map__new(NULL
);
1213 if (evlist
->cpus
== NULL
)
1216 evlist
->threads
= thread_map__new_dummy();
1217 if (evlist
->threads
== NULL
)
1224 cpu_map__delete(evlist
->cpus
);
1225 evlist
->cpus
= NULL
;
1229 int perf_evlist__open(struct perf_evlist
*evlist
)
1231 struct perf_evsel
*evsel
;
1235 * Default: one fd per CPU, all threads, aka systemwide
1236 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1238 if (evlist
->threads
== NULL
&& evlist
->cpus
== NULL
) {
1239 err
= perf_evlist__create_syswide_maps(evlist
);
1244 perf_evlist__update_id_pos(evlist
);
1246 evlist__for_each(evlist
, evsel
) {
1247 err
= perf_evsel__open(evsel
, evlist
->cpus
, evlist
->threads
);
1254 perf_evlist__close(evlist
);
1259 int perf_evlist__prepare_workload(struct perf_evlist
*evlist
, struct target
*target
,
1260 const char *argv
[], bool pipe_output
,
1261 void (*exec_error
)(int signo
, siginfo_t
*info
, void *ucontext
))
1263 int child_ready_pipe
[2], go_pipe
[2];
1266 if (pipe(child_ready_pipe
) < 0) {
1267 perror("failed to create 'ready' pipe");
1271 if (pipe(go_pipe
) < 0) {
1272 perror("failed to create 'go' pipe");
1273 goto out_close_ready_pipe
;
1276 evlist
->workload
.pid
= fork();
1277 if (evlist
->workload
.pid
< 0) {
1278 perror("failed to fork");
1279 goto out_close_pipes
;
1282 if (!evlist
->workload
.pid
) {
1288 signal(SIGTERM
, SIG_DFL
);
1290 close(child_ready_pipe
[0]);
1292 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
1295 * Tell the parent we're ready to go
1297 close(child_ready_pipe
[1]);
1300 * Wait until the parent tells us to go.
1302 ret
= read(go_pipe
[0], &bf
, 1);
1304 * The parent will ask for the execvp() to be performed by
1305 * writing exactly one byte, in workload.cork_fd, usually via
1306 * perf_evlist__start_workload().
1308 * For cancelling the workload without actuallin running it,
1309 * the parent will just close workload.cork_fd, without writing
1310 * anything, i.e. read will return zero and we just exit()
1315 perror("unable to read pipe");
1319 execvp(argv
[0], (char **)argv
);
1324 val
.sival_int
= errno
;
1325 if (sigqueue(getppid(), SIGUSR1
, val
))
1333 struct sigaction act
= {
1334 .sa_flags
= SA_SIGINFO
,
1335 .sa_sigaction
= exec_error
,
1337 sigaction(SIGUSR1
, &act
, NULL
);
1340 if (target__none(target
)) {
1341 if (evlist
->threads
== NULL
) {
1342 fprintf(stderr
, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1343 __func__
, __LINE__
);
1344 goto out_close_pipes
;
1346 evlist
->threads
->map
[0] = evlist
->workload
.pid
;
1349 close(child_ready_pipe
[1]);
1352 * wait for child to settle
1354 if (read(child_ready_pipe
[0], &bf
, 1) == -1) {
1355 perror("unable to read pipe");
1356 goto out_close_pipes
;
1359 fcntl(go_pipe
[1], F_SETFD
, FD_CLOEXEC
);
1360 evlist
->workload
.cork_fd
= go_pipe
[1];
1361 close(child_ready_pipe
[0]);
1367 out_close_ready_pipe
:
1368 close(child_ready_pipe
[0]);
1369 close(child_ready_pipe
[1]);
1373 int perf_evlist__start_workload(struct perf_evlist
*evlist
)
1375 if (evlist
->workload
.cork_fd
> 0) {
1379 * Remove the cork, let it rip!
1381 ret
= write(evlist
->workload
.cork_fd
, &bf
, 1);
1383 perror("enable to write to pipe");
1385 close(evlist
->workload
.cork_fd
);
1392 int perf_evlist__parse_sample(struct perf_evlist
*evlist
, union perf_event
*event
,
1393 struct perf_sample
*sample
)
1395 struct perf_evsel
*evsel
= perf_evlist__event2evsel(evlist
, event
);
1399 return perf_evsel__parse_sample(evsel
, event
, sample
);
1402 size_t perf_evlist__fprintf(struct perf_evlist
*evlist
, FILE *fp
)
1404 struct perf_evsel
*evsel
;
1407 evlist__for_each(evlist
, evsel
) {
1408 printed
+= fprintf(fp
, "%s%s", evsel
->idx
? ", " : "",
1409 perf_evsel__name(evsel
));
1412 return printed
+ fprintf(fp
, "\n");
1415 int perf_evlist__strerror_tp(struct perf_evlist
*evlist __maybe_unused
,
1416 int err
, char *buf
, size_t size
)
1422 scnprintf(buf
, size
, "%s",
1423 "Error:\tUnable to find debugfs\n"
1424 "Hint:\tWas your kernel was compiled with debugfs support?\n"
1425 "Hint:\tIs the debugfs filesystem mounted?\n"
1426 "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1429 scnprintf(buf
, size
,
1430 "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1431 "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1432 debugfs_mountpoint
, debugfs_mountpoint
);
1435 scnprintf(buf
, size
, "%s", strerror_r(err
, sbuf
, sizeof(sbuf
)));
1442 int perf_evlist__strerror_open(struct perf_evlist
*evlist __maybe_unused
,
1443 int err
, char *buf
, size_t size
)
1446 char sbuf
[STRERR_BUFSIZE
], *emsg
= strerror_r(err
, sbuf
, sizeof(sbuf
));
1451 printed
= scnprintf(buf
, size
,
1453 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg
);
1455 value
= perf_event_paranoid();
1457 printed
+= scnprintf(buf
+ printed
, size
- printed
, "\nHint:\t");
1460 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1461 "For your workloads it needs to be <= 1\nHint:\t");
1463 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1464 "For system wide tracing it needs to be set to -1.\n");
1466 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1467 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1468 "Hint:\tThe current value is %d.", value
);
1471 scnprintf(buf
, size
, "%s", emsg
);
1478 void perf_evlist__to_front(struct perf_evlist
*evlist
,
1479 struct perf_evsel
*move_evsel
)
1481 struct perf_evsel
*evsel
, *n
;
1484 if (move_evsel
== perf_evlist__first(evlist
))
1487 evlist__for_each_safe(evlist
, n
, evsel
) {
1488 if (evsel
->leader
== move_evsel
->leader
)
1489 list_move_tail(&evsel
->node
, &move
);
1492 list_splice(&move
, &evlist
->entries
);
1495 void perf_evlist__set_tracking_event(struct perf_evlist
*evlist
,
1496 struct perf_evsel
*tracking_evsel
)
1498 struct perf_evsel
*evsel
;
1500 if (tracking_evsel
->tracking
)
1503 evlist__for_each(evlist
, evsel
) {
1504 if (evsel
!= tracking_evsel
)
1505 evsel
->tracking
= false;
1508 tracking_evsel
->tracking
= true;