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"
20 #include "parse-events.h"
21 #include <subcmd/parse-options.h>
25 #include <linux/bitops.h>
26 #include <linux/hash.h>
27 #include <linux/log2.h>
28 #include <linux/err.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;
47 evlist
->backward
= false;
50 struct perf_evlist
*perf_evlist__new(void)
52 struct perf_evlist
*evlist
= zalloc(sizeof(*evlist
));
55 perf_evlist__init(evlist
, NULL
, NULL
);
60 struct perf_evlist
*perf_evlist__new_default(void)
62 struct perf_evlist
*evlist
= perf_evlist__new();
64 if (evlist
&& perf_evlist__add_default(evlist
)) {
65 perf_evlist__delete(evlist
);
72 struct perf_evlist
*perf_evlist__new_dummy(void)
74 struct perf_evlist
*evlist
= perf_evlist__new();
76 if (evlist
&& perf_evlist__add_dummy(evlist
)) {
77 perf_evlist__delete(evlist
);
85 * perf_evlist__set_id_pos - set the positions of event ids.
86 * @evlist: selected event list
88 * Events with compatible sample types all have the same id_pos
89 * and is_pos. For convenience, put a copy on evlist.
91 void perf_evlist__set_id_pos(struct perf_evlist
*evlist
)
93 struct perf_evsel
*first
= perf_evlist__first(evlist
);
95 evlist
->id_pos
= first
->id_pos
;
96 evlist
->is_pos
= first
->is_pos
;
99 static void perf_evlist__update_id_pos(struct perf_evlist
*evlist
)
101 struct perf_evsel
*evsel
;
103 evlist__for_each(evlist
, evsel
)
104 perf_evsel__calc_id_pos(evsel
);
106 perf_evlist__set_id_pos(evlist
);
109 static void perf_evlist__purge(struct perf_evlist
*evlist
)
111 struct perf_evsel
*pos
, *n
;
113 evlist__for_each_safe(evlist
, n
, pos
) {
114 list_del_init(&pos
->node
);
116 perf_evsel__delete(pos
);
119 evlist
->nr_entries
= 0;
122 void perf_evlist__exit(struct perf_evlist
*evlist
)
124 zfree(&evlist
->mmap
);
125 fdarray__exit(&evlist
->pollfd
);
128 void perf_evlist__delete(struct perf_evlist
*evlist
)
130 perf_evlist__munmap(evlist
);
131 perf_evlist__close(evlist
);
132 cpu_map__put(evlist
->cpus
);
133 thread_map__put(evlist
->threads
);
135 evlist
->threads
= NULL
;
136 perf_evlist__purge(evlist
);
137 perf_evlist__exit(evlist
);
141 static void __perf_evlist__propagate_maps(struct perf_evlist
*evlist
,
142 struct perf_evsel
*evsel
)
145 * We already have cpus for evsel (via PMU sysfs) so
146 * keep it, if there's no target cpu list defined.
148 if (!evsel
->own_cpus
|| evlist
->has_user_cpus
) {
149 cpu_map__put(evsel
->cpus
);
150 evsel
->cpus
= cpu_map__get(evlist
->cpus
);
151 } else if (evsel
->cpus
!= evsel
->own_cpus
) {
152 cpu_map__put(evsel
->cpus
);
153 evsel
->cpus
= cpu_map__get(evsel
->own_cpus
);
156 thread_map__put(evsel
->threads
);
157 evsel
->threads
= thread_map__get(evlist
->threads
);
160 static void perf_evlist__propagate_maps(struct perf_evlist
*evlist
)
162 struct perf_evsel
*evsel
;
164 evlist__for_each(evlist
, evsel
)
165 __perf_evlist__propagate_maps(evlist
, evsel
);
168 void perf_evlist__add(struct perf_evlist
*evlist
, struct perf_evsel
*entry
)
170 entry
->evlist
= evlist
;
171 list_add_tail(&entry
->node
, &evlist
->entries
);
172 entry
->idx
= evlist
->nr_entries
;
173 entry
->tracking
= !entry
->idx
;
175 if (!evlist
->nr_entries
++)
176 perf_evlist__set_id_pos(evlist
);
178 __perf_evlist__propagate_maps(evlist
, entry
);
181 void perf_evlist__remove(struct perf_evlist
*evlist
, struct perf_evsel
*evsel
)
183 evsel
->evlist
= NULL
;
184 list_del_init(&evsel
->node
);
185 evlist
->nr_entries
-= 1;
188 void perf_evlist__splice_list_tail(struct perf_evlist
*evlist
,
189 struct list_head
*list
)
191 struct perf_evsel
*evsel
, *temp
;
193 __evlist__for_each_safe(list
, temp
, evsel
) {
194 list_del_init(&evsel
->node
);
195 perf_evlist__add(evlist
, evsel
);
199 void __perf_evlist__set_leader(struct list_head
*list
)
201 struct perf_evsel
*evsel
, *leader
;
203 leader
= list_entry(list
->next
, struct perf_evsel
, node
);
204 evsel
= list_entry(list
->prev
, struct perf_evsel
, node
);
206 leader
->nr_members
= evsel
->idx
- leader
->idx
+ 1;
208 __evlist__for_each(list
, evsel
) {
209 evsel
->leader
= leader
;
213 void perf_evlist__set_leader(struct perf_evlist
*evlist
)
215 if (evlist
->nr_entries
) {
216 evlist
->nr_groups
= evlist
->nr_entries
> 1 ? 1 : 0;
217 __perf_evlist__set_leader(&evlist
->entries
);
221 void perf_event_attr__set_max_precise_ip(struct perf_event_attr
*attr
)
223 attr
->precise_ip
= 3;
225 while (attr
->precise_ip
!= 0) {
226 int fd
= sys_perf_event_open(attr
, 0, -1, -1, 0);
235 int perf_evlist__add_default(struct perf_evlist
*evlist
)
237 struct perf_event_attr attr
= {
238 .type
= PERF_TYPE_HARDWARE
,
239 .config
= PERF_COUNT_HW_CPU_CYCLES
,
241 struct perf_evsel
*evsel
;
243 event_attr_init(&attr
);
245 perf_event_attr__set_max_precise_ip(&attr
);
247 evsel
= perf_evsel__new(&attr
);
251 /* use asprintf() because free(evsel) assumes name is allocated */
252 if (asprintf(&evsel
->name
, "cycles%.*s",
253 attr
.precise_ip
? attr
.precise_ip
+ 1 : 0, ":ppp") < 0)
256 perf_evlist__add(evlist
, evsel
);
259 perf_evsel__delete(evsel
);
264 int perf_evlist__add_dummy(struct perf_evlist
*evlist
)
266 struct perf_event_attr attr
= {
267 .type
= PERF_TYPE_SOFTWARE
,
268 .config
= PERF_COUNT_SW_DUMMY
,
269 .size
= sizeof(attr
), /* to capture ABI version */
271 struct perf_evsel
*evsel
= perf_evsel__new(&attr
);
276 perf_evlist__add(evlist
, evsel
);
280 static int perf_evlist__add_attrs(struct perf_evlist
*evlist
,
281 struct perf_event_attr
*attrs
, size_t nr_attrs
)
283 struct perf_evsel
*evsel
, *n
;
287 for (i
= 0; i
< nr_attrs
; i
++) {
288 evsel
= perf_evsel__new_idx(attrs
+ i
, evlist
->nr_entries
+ i
);
290 goto out_delete_partial_list
;
291 list_add_tail(&evsel
->node
, &head
);
294 perf_evlist__splice_list_tail(evlist
, &head
);
298 out_delete_partial_list
:
299 __evlist__for_each_safe(&head
, n
, evsel
)
300 perf_evsel__delete(evsel
);
304 int __perf_evlist__add_default_attrs(struct perf_evlist
*evlist
,
305 struct perf_event_attr
*attrs
, size_t nr_attrs
)
309 for (i
= 0; i
< nr_attrs
; i
++)
310 event_attr_init(attrs
+ i
);
312 return perf_evlist__add_attrs(evlist
, attrs
, nr_attrs
);
316 perf_evlist__find_tracepoint_by_id(struct perf_evlist
*evlist
, int id
)
318 struct perf_evsel
*evsel
;
320 evlist__for_each(evlist
, evsel
) {
321 if (evsel
->attr
.type
== PERF_TYPE_TRACEPOINT
&&
322 (int)evsel
->attr
.config
== id
)
330 perf_evlist__find_tracepoint_by_name(struct perf_evlist
*evlist
,
333 struct perf_evsel
*evsel
;
335 evlist__for_each(evlist
, evsel
) {
336 if ((evsel
->attr
.type
== PERF_TYPE_TRACEPOINT
) &&
337 (strcmp(evsel
->name
, name
) == 0))
344 int perf_evlist__add_newtp(struct perf_evlist
*evlist
,
345 const char *sys
, const char *name
, void *handler
)
347 struct perf_evsel
*evsel
= perf_evsel__newtp(sys
, name
);
352 evsel
->handler
= handler
;
353 perf_evlist__add(evlist
, evsel
);
357 static int perf_evlist__nr_threads(struct perf_evlist
*evlist
,
358 struct perf_evsel
*evsel
)
360 if (evsel
->system_wide
)
363 return thread_map__nr(evlist
->threads
);
366 void perf_evlist__disable(struct perf_evlist
*evlist
)
368 struct perf_evsel
*pos
;
370 evlist__for_each(evlist
, pos
) {
371 if (!perf_evsel__is_group_leader(pos
) || !pos
->fd
)
373 perf_evsel__disable(pos
);
376 evlist
->enabled
= false;
379 void perf_evlist__enable(struct perf_evlist
*evlist
)
381 struct perf_evsel
*pos
;
383 evlist__for_each(evlist
, pos
) {
384 if (!perf_evsel__is_group_leader(pos
) || !pos
->fd
)
386 perf_evsel__enable(pos
);
389 evlist
->enabled
= true;
392 void perf_evlist__toggle_enable(struct perf_evlist
*evlist
)
394 (evlist
->enabled
? perf_evlist__disable
: perf_evlist__enable
)(evlist
);
397 static int perf_evlist__enable_event_cpu(struct perf_evlist
*evlist
,
398 struct perf_evsel
*evsel
, int cpu
)
401 int nr_threads
= perf_evlist__nr_threads(evlist
, evsel
);
406 for (thread
= 0; thread
< nr_threads
; thread
++) {
407 err
= ioctl(FD(evsel
, cpu
, thread
),
408 PERF_EVENT_IOC_ENABLE
, 0);
415 static int perf_evlist__enable_event_thread(struct perf_evlist
*evlist
,
416 struct perf_evsel
*evsel
,
420 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
425 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
426 err
= ioctl(FD(evsel
, cpu
, thread
), PERF_EVENT_IOC_ENABLE
, 0);
433 int perf_evlist__enable_event_idx(struct perf_evlist
*evlist
,
434 struct perf_evsel
*evsel
, int idx
)
436 bool per_cpu_mmaps
= !cpu_map__empty(evlist
->cpus
);
439 return perf_evlist__enable_event_cpu(evlist
, evsel
, idx
);
441 return perf_evlist__enable_event_thread(evlist
, evsel
, idx
);
444 int perf_evlist__alloc_pollfd(struct perf_evlist
*evlist
)
446 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
447 int nr_threads
= thread_map__nr(evlist
->threads
);
449 struct perf_evsel
*evsel
;
451 evlist__for_each(evlist
, evsel
) {
452 if (evsel
->system_wide
)
455 nfds
+= nr_cpus
* nr_threads
;
458 if (fdarray__available_entries(&evlist
->pollfd
) < nfds
&&
459 fdarray__grow(&evlist
->pollfd
, nfds
) < 0)
465 static int __perf_evlist__add_pollfd(struct perf_evlist
*evlist
, int fd
, int idx
)
467 int pos
= fdarray__add(&evlist
->pollfd
, fd
, POLLIN
| POLLERR
| POLLHUP
);
469 * Save the idx so that when we filter out fds POLLHUP'ed we can
470 * close the associated evlist->mmap[] entry.
473 evlist
->pollfd
.priv
[pos
].idx
= idx
;
475 fcntl(fd
, F_SETFL
, O_NONBLOCK
);
481 int perf_evlist__add_pollfd(struct perf_evlist
*evlist
, int fd
)
483 return __perf_evlist__add_pollfd(evlist
, fd
, -1);
486 static void perf_evlist__munmap_filtered(struct fdarray
*fda
, int fd
)
488 struct perf_evlist
*evlist
= container_of(fda
, struct perf_evlist
, pollfd
);
490 perf_evlist__mmap_put(evlist
, fda
->priv
[fd
].idx
);
493 int perf_evlist__filter_pollfd(struct perf_evlist
*evlist
, short revents_and_mask
)
495 return fdarray__filter(&evlist
->pollfd
, revents_and_mask
,
496 perf_evlist__munmap_filtered
);
499 int perf_evlist__poll(struct perf_evlist
*evlist
, int timeout
)
501 return fdarray__poll(&evlist
->pollfd
, timeout
);
504 static void perf_evlist__id_hash(struct perf_evlist
*evlist
,
505 struct perf_evsel
*evsel
,
506 int cpu
, int thread
, u64 id
)
509 struct perf_sample_id
*sid
= SID(evsel
, cpu
, thread
);
513 hash
= hash_64(sid
->id
, PERF_EVLIST__HLIST_BITS
);
514 hlist_add_head(&sid
->node
, &evlist
->heads
[hash
]);
517 void perf_evlist__id_add(struct perf_evlist
*evlist
, struct perf_evsel
*evsel
,
518 int cpu
, int thread
, u64 id
)
520 perf_evlist__id_hash(evlist
, evsel
, cpu
, thread
, id
);
521 evsel
->id
[evsel
->ids
++] = id
;
524 int perf_evlist__id_add_fd(struct perf_evlist
*evlist
,
525 struct perf_evsel
*evsel
,
526 int cpu
, int thread
, int fd
)
528 u64 read_data
[4] = { 0, };
529 int id_idx
= 1; /* The first entry is the counter value */
533 ret
= ioctl(fd
, PERF_EVENT_IOC_ID
, &id
);
540 /* Legacy way to get event id.. All hail to old kernels! */
543 * This way does not work with group format read, so bail
546 if (perf_evlist__read_format(evlist
) & PERF_FORMAT_GROUP
)
549 if (!(evsel
->attr
.read_format
& PERF_FORMAT_ID
) ||
550 read(fd
, &read_data
, sizeof(read_data
)) == -1)
553 if (evsel
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
555 if (evsel
->attr
.read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
558 id
= read_data
[id_idx
];
561 perf_evlist__id_add(evlist
, evsel
, cpu
, thread
, id
);
565 static void perf_evlist__set_sid_idx(struct perf_evlist
*evlist
,
566 struct perf_evsel
*evsel
, int idx
, int cpu
,
569 struct perf_sample_id
*sid
= SID(evsel
, cpu
, thread
);
571 if (evlist
->cpus
&& cpu
>= 0)
572 sid
->cpu
= evlist
->cpus
->map
[cpu
];
575 if (!evsel
->system_wide
&& evlist
->threads
&& thread
>= 0)
576 sid
->tid
= thread_map__pid(evlist
->threads
, thread
);
581 struct perf_sample_id
*perf_evlist__id2sid(struct perf_evlist
*evlist
, u64 id
)
583 struct hlist_head
*head
;
584 struct perf_sample_id
*sid
;
587 hash
= hash_64(id
, PERF_EVLIST__HLIST_BITS
);
588 head
= &evlist
->heads
[hash
];
590 hlist_for_each_entry(sid
, head
, node
)
597 struct perf_evsel
*perf_evlist__id2evsel(struct perf_evlist
*evlist
, u64 id
)
599 struct perf_sample_id
*sid
;
601 if (evlist
->nr_entries
== 1 || !id
)
602 return perf_evlist__first(evlist
);
604 sid
= perf_evlist__id2sid(evlist
, id
);
608 if (!perf_evlist__sample_id_all(evlist
))
609 return perf_evlist__first(evlist
);
614 struct perf_evsel
*perf_evlist__id2evsel_strict(struct perf_evlist
*evlist
,
617 struct perf_sample_id
*sid
;
622 sid
= perf_evlist__id2sid(evlist
, id
);
629 static int perf_evlist__event2id(struct perf_evlist
*evlist
,
630 union perf_event
*event
, u64
*id
)
632 const u64
*array
= event
->sample
.array
;
635 n
= (event
->header
.size
- sizeof(event
->header
)) >> 3;
637 if (event
->header
.type
== PERF_RECORD_SAMPLE
) {
638 if (evlist
->id_pos
>= n
)
640 *id
= array
[evlist
->id_pos
];
642 if (evlist
->is_pos
> n
)
650 static struct perf_evsel
*perf_evlist__event2evsel(struct perf_evlist
*evlist
,
651 union perf_event
*event
)
653 struct perf_evsel
*first
= perf_evlist__first(evlist
);
654 struct hlist_head
*head
;
655 struct perf_sample_id
*sid
;
659 if (evlist
->nr_entries
== 1)
662 if (!first
->attr
.sample_id_all
&&
663 event
->header
.type
!= PERF_RECORD_SAMPLE
)
666 if (perf_evlist__event2id(evlist
, event
, &id
))
669 /* Synthesized events have an id of zero */
673 hash
= hash_64(id
, PERF_EVLIST__HLIST_BITS
);
674 head
= &evlist
->heads
[hash
];
676 hlist_for_each_entry(sid
, head
, node
) {
683 static int perf_evlist__set_paused(struct perf_evlist
*evlist
, bool value
)
687 for (i
= 0; i
< evlist
->nr_mmaps
; i
++) {
688 int fd
= evlist
->mmap
[i
].fd
;
693 err
= ioctl(fd
, PERF_EVENT_IOC_PAUSE_OUTPUT
, value
? 1 : 0);
700 int perf_evlist__pause(struct perf_evlist
*evlist
)
702 return perf_evlist__set_paused(evlist
, true);
705 int perf_evlist__resume(struct perf_evlist
*evlist
)
707 return perf_evlist__set_paused(evlist
, false);
710 /* When check_messup is true, 'end' must points to a good entry */
711 static union perf_event
*
712 perf_mmap__read(struct perf_mmap
*md
, bool check_messup
, u64 start
,
715 unsigned char *data
= md
->base
+ page_size
;
716 union perf_event
*event
= NULL
;
717 int diff
= end
- start
;
721 * If we're further behind than half the buffer, there's a chance
722 * the writer will bite our tail and mess up the samples under us.
724 * If we somehow ended up ahead of the 'end', we got messed up.
726 * In either case, truncate and restart at 'end'.
728 if (diff
> md
->mask
/ 2 || diff
< 0) {
729 fprintf(stderr
, "WARNING: failed to keep up with mmap data.\n");
732 * 'end' points to a known good entry, start there.
739 if (diff
>= (int)sizeof(event
->header
)) {
742 event
= (union perf_event
*)&data
[start
& md
->mask
];
743 size
= event
->header
.size
;
745 if (size
< sizeof(event
->header
) || diff
< (int)size
) {
751 * Event straddles the mmap boundary -- header should always
752 * be inside due to u64 alignment of output.
754 if ((start
& md
->mask
) + size
!= ((start
+ size
) & md
->mask
)) {
755 unsigned int offset
= start
;
756 unsigned int len
= min(sizeof(*event
), size
), cpy
;
757 void *dst
= md
->event_copy
;
760 cpy
= min(md
->mask
+ 1 - (offset
& md
->mask
), len
);
761 memcpy(dst
, &data
[offset
& md
->mask
], cpy
);
767 event
= (union perf_event
*) md
->event_copy
;
780 union perf_event
*perf_evlist__mmap_read(struct perf_evlist
*evlist
, int idx
)
782 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
787 * Check if event was unmapped due to a POLLHUP/POLLERR.
789 if (!atomic_read(&md
->refcnt
))
792 head
= perf_mmap__read_head(md
);
794 return perf_mmap__read(md
, evlist
->overwrite
, old
, head
, &md
->prev
);
798 perf_evlist__mmap_read_backward(struct perf_evlist
*evlist
, int idx
)
800 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
802 u64 start
= md
->prev
;
805 * Check if event was unmapped due to a POLLHUP/POLLERR.
807 if (!atomic_read(&md
->refcnt
))
810 head
= perf_mmap__read_head(md
);
815 * 'head' pointer starts from 0. Kernel minus sizeof(record) form
816 * it each time when kernel writes to it, so in fact 'head' is
817 * negative. 'end' pointer is made manually by adding the size of
818 * the ring buffer to 'head' pointer, means the validate data can
819 * read is the whole ring buffer. If 'end' is positive, the ring
820 * buffer has not fully filled, so we must adjust 'end' to 0.
822 * However, since both 'head' and 'end' is unsigned, we can't
823 * simply compare 'end' against 0. Here we compare '-head' and
824 * the size of the ring buffer, where -head is the number of bytes
825 * kernel write to the ring buffer.
827 if (-head
< (u64
)(md
->mask
+ 1))
830 end
= head
+ md
->mask
+ 1;
832 return perf_mmap__read(md
, false, start
, end
, &md
->prev
);
835 void perf_evlist__mmap_read_catchup(struct perf_evlist
*evlist
, int idx
)
837 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
840 if (!atomic_read(&md
->refcnt
))
843 head
= perf_mmap__read_head(md
);
847 static bool perf_mmap__empty(struct perf_mmap
*md
)
849 return perf_mmap__read_head(md
) == md
->prev
&& !md
->auxtrace_mmap
.base
;
852 static void perf_evlist__mmap_get(struct perf_evlist
*evlist
, int idx
)
854 atomic_inc(&evlist
->mmap
[idx
].refcnt
);
857 static void perf_evlist__mmap_put(struct perf_evlist
*evlist
, int idx
)
859 BUG_ON(atomic_read(&evlist
->mmap
[idx
].refcnt
) == 0);
861 if (atomic_dec_and_test(&evlist
->mmap
[idx
].refcnt
))
862 __perf_evlist__munmap(evlist
, idx
);
865 void perf_evlist__mmap_consume(struct perf_evlist
*evlist
, int idx
)
867 struct perf_mmap
*md
= &evlist
->mmap
[idx
];
869 if (!evlist
->overwrite
) {
872 perf_mmap__write_tail(md
, old
);
875 if (atomic_read(&md
->refcnt
) == 1 && perf_mmap__empty(md
))
876 perf_evlist__mmap_put(evlist
, idx
);
879 int __weak
auxtrace_mmap__mmap(struct auxtrace_mmap
*mm __maybe_unused
,
880 struct auxtrace_mmap_params
*mp __maybe_unused
,
881 void *userpg __maybe_unused
,
882 int fd __maybe_unused
)
887 void __weak
auxtrace_mmap__munmap(struct auxtrace_mmap
*mm __maybe_unused
)
891 void __weak
auxtrace_mmap_params__init(
892 struct auxtrace_mmap_params
*mp __maybe_unused
,
893 off_t auxtrace_offset __maybe_unused
,
894 unsigned int auxtrace_pages __maybe_unused
,
895 bool auxtrace_overwrite __maybe_unused
)
899 void __weak
auxtrace_mmap_params__set_idx(
900 struct auxtrace_mmap_params
*mp __maybe_unused
,
901 struct perf_evlist
*evlist __maybe_unused
,
902 int idx __maybe_unused
,
903 bool per_cpu __maybe_unused
)
907 static void __perf_evlist__munmap(struct perf_evlist
*evlist
, int idx
)
909 if (evlist
->mmap
[idx
].base
!= NULL
) {
910 munmap(evlist
->mmap
[idx
].base
, evlist
->mmap_len
);
911 evlist
->mmap
[idx
].base
= NULL
;
912 evlist
->mmap
[idx
].fd
= -1;
913 atomic_set(&evlist
->mmap
[idx
].refcnt
, 0);
915 auxtrace_mmap__munmap(&evlist
->mmap
[idx
].auxtrace_mmap
);
918 void perf_evlist__munmap(struct perf_evlist
*evlist
)
922 if (evlist
->mmap
== NULL
)
925 for (i
= 0; i
< evlist
->nr_mmaps
; i
++)
926 __perf_evlist__munmap(evlist
, i
);
928 zfree(&evlist
->mmap
);
931 static int perf_evlist__alloc_mmap(struct perf_evlist
*evlist
)
935 evlist
->nr_mmaps
= cpu_map__nr(evlist
->cpus
);
936 if (cpu_map__empty(evlist
->cpus
))
937 evlist
->nr_mmaps
= thread_map__nr(evlist
->threads
);
938 evlist
->mmap
= zalloc(evlist
->nr_mmaps
* sizeof(struct perf_mmap
));
939 for (i
= 0; i
< evlist
->nr_mmaps
; i
++)
940 evlist
->mmap
[i
].fd
= -1;
941 return evlist
->mmap
!= NULL
? 0 : -ENOMEM
;
947 struct auxtrace_mmap_params auxtrace_mp
;
950 static int __perf_evlist__mmap(struct perf_evlist
*evlist
, int idx
,
951 struct mmap_params
*mp
, int fd
)
954 * The last one will be done at perf_evlist__mmap_consume(), so that we
955 * make sure we don't prevent tools from consuming every last event in
958 * I.e. we can get the POLLHUP meaning that the fd doesn't exist
959 * anymore, but the last events for it are still in the ring buffer,
960 * waiting to be consumed.
962 * Tools can chose to ignore this at their own discretion, but the
963 * evlist layer can't just drop it when filtering events in
964 * perf_evlist__filter_pollfd().
966 atomic_set(&evlist
->mmap
[idx
].refcnt
, 2);
967 evlist
->mmap
[idx
].prev
= 0;
968 evlist
->mmap
[idx
].mask
= mp
->mask
;
969 evlist
->mmap
[idx
].base
= mmap(NULL
, evlist
->mmap_len
, mp
->prot
,
971 if (evlist
->mmap
[idx
].base
== MAP_FAILED
) {
972 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
974 evlist
->mmap
[idx
].base
= NULL
;
977 evlist
->mmap
[idx
].fd
= fd
;
979 if (auxtrace_mmap__mmap(&evlist
->mmap
[idx
].auxtrace_mmap
,
980 &mp
->auxtrace_mp
, evlist
->mmap
[idx
].base
, fd
))
986 static int perf_evlist__mmap_per_evsel(struct perf_evlist
*evlist
, int idx
,
987 struct mmap_params
*mp
, int cpu
,
988 int thread
, int *output
)
990 struct perf_evsel
*evsel
;
992 evlist__for_each(evlist
, evsel
) {
995 if (evsel
->system_wide
&& thread
)
998 fd
= FD(evsel
, cpu
, thread
);
1000 if (*output
== -1) {
1002 if (__perf_evlist__mmap(evlist
, idx
, mp
, *output
) < 0)
1005 if (ioctl(fd
, PERF_EVENT_IOC_SET_OUTPUT
, *output
) != 0)
1008 perf_evlist__mmap_get(evlist
, idx
);
1012 * The system_wide flag causes a selected event to be opened
1013 * always without a pid. Consequently it will never get a
1014 * POLLHUP, but it is used for tracking in combination with
1015 * other events, so it should not need to be polled anyway.
1016 * Therefore don't add it for polling.
1018 if (!evsel
->system_wide
&&
1019 __perf_evlist__add_pollfd(evlist
, fd
, idx
) < 0) {
1020 perf_evlist__mmap_put(evlist
, idx
);
1024 if (evsel
->attr
.read_format
& PERF_FORMAT_ID
) {
1025 if (perf_evlist__id_add_fd(evlist
, evsel
, cpu
, thread
,
1028 perf_evlist__set_sid_idx(evlist
, evsel
, idx
, cpu
,
1036 static int perf_evlist__mmap_per_cpu(struct perf_evlist
*evlist
,
1037 struct mmap_params
*mp
)
1040 int nr_cpus
= cpu_map__nr(evlist
->cpus
);
1041 int nr_threads
= thread_map__nr(evlist
->threads
);
1043 pr_debug2("perf event ring buffer mmapped per cpu\n");
1044 for (cpu
= 0; cpu
< nr_cpus
; cpu
++) {
1047 auxtrace_mmap_params__set_idx(&mp
->auxtrace_mp
, evlist
, cpu
,
1050 for (thread
= 0; thread
< nr_threads
; thread
++) {
1051 if (perf_evlist__mmap_per_evsel(evlist
, cpu
, mp
, cpu
,
1060 for (cpu
= 0; cpu
< nr_cpus
; cpu
++)
1061 __perf_evlist__munmap(evlist
, cpu
);
1065 static int perf_evlist__mmap_per_thread(struct perf_evlist
*evlist
,
1066 struct mmap_params
*mp
)
1069 int nr_threads
= thread_map__nr(evlist
->threads
);
1071 pr_debug2("perf event ring buffer mmapped per thread\n");
1072 for (thread
= 0; thread
< nr_threads
; thread
++) {
1075 auxtrace_mmap_params__set_idx(&mp
->auxtrace_mp
, evlist
, thread
,
1078 if (perf_evlist__mmap_per_evsel(evlist
, thread
, mp
, 0, thread
,
1086 for (thread
= 0; thread
< nr_threads
; thread
++)
1087 __perf_evlist__munmap(evlist
, thread
);
1091 unsigned long perf_event_mlock_kb_in_pages(void)
1093 unsigned long pages
;
1096 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max
) < 0) {
1098 * Pick a once upon a time good value, i.e. things look
1099 * strange since we can't read a sysctl value, but lets not
1104 max
-= (page_size
/ 1024);
1107 pages
= (max
* 1024) / page_size
;
1108 if (!is_power_of_2(pages
))
1109 pages
= rounddown_pow_of_two(pages
);
1114 static size_t perf_evlist__mmap_size(unsigned long pages
)
1116 if (pages
== UINT_MAX
)
1117 pages
= perf_event_mlock_kb_in_pages();
1118 else if (!is_power_of_2(pages
))
1121 return (pages
+ 1) * page_size
;
1124 static long parse_pages_arg(const char *str
, unsigned long min
,
1127 unsigned long pages
, val
;
1128 static struct parse_tag tags
[] = {
1129 { .tag
= 'B', .mult
= 1 },
1130 { .tag
= 'K', .mult
= 1 << 10 },
1131 { .tag
= 'M', .mult
= 1 << 20 },
1132 { .tag
= 'G', .mult
= 1 << 30 },
1139 val
= parse_tag_value(str
, tags
);
1140 if (val
!= (unsigned long) -1) {
1141 /* we got file size value */
1142 pages
= PERF_ALIGN(val
, page_size
) / page_size
;
1144 /* we got pages count value */
1146 pages
= strtoul(str
, &eptr
, 10);
1151 if (pages
== 0 && min
== 0) {
1152 /* leave number of pages at 0 */
1153 } else if (!is_power_of_2(pages
)) {
1154 /* round pages up to next power of 2 */
1155 pages
= roundup_pow_of_two(pages
);
1158 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1159 pages
* page_size
, pages
);
1168 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages
, const char *str
)
1170 unsigned long max
= UINT_MAX
;
1173 if (max
> SIZE_MAX
/ page_size
)
1174 max
= SIZE_MAX
/ page_size
;
1176 pages
= parse_pages_arg(str
, 1, max
);
1178 pr_err("Invalid argument for --mmap_pages/-m\n");
1182 *mmap_pages
= pages
;
1186 int perf_evlist__parse_mmap_pages(const struct option
*opt
, const char *str
,
1187 int unset __maybe_unused
)
1189 return __perf_evlist__parse_mmap_pages(opt
->value
, str
);
1193 * perf_evlist__mmap_ex - Create mmaps to receive events.
1194 * @evlist: list of events
1195 * @pages: map length in pages
1196 * @overwrite: overwrite older events?
1197 * @auxtrace_pages - auxtrace map length in pages
1198 * @auxtrace_overwrite - overwrite older auxtrace data?
1200 * If @overwrite is %false the user needs to signal event consumption using
1201 * perf_mmap__write_tail(). Using perf_evlist__mmap_read() does this
1204 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1205 * consumption using auxtrace_mmap__write_tail().
1207 * Return: %0 on success, negative error code otherwise.
1209 int perf_evlist__mmap_ex(struct perf_evlist
*evlist
, unsigned int pages
,
1210 bool overwrite
, unsigned int auxtrace_pages
,
1211 bool auxtrace_overwrite
)
1213 struct perf_evsel
*evsel
;
1214 const struct cpu_map
*cpus
= evlist
->cpus
;
1215 const struct thread_map
*threads
= evlist
->threads
;
1216 struct mmap_params mp
= {
1217 .prot
= PROT_READ
| (overwrite
? 0 : PROT_WRITE
),
1220 if (evlist
->mmap
== NULL
&& perf_evlist__alloc_mmap(evlist
) < 0)
1223 if (evlist
->pollfd
.entries
== NULL
&& perf_evlist__alloc_pollfd(evlist
) < 0)
1226 evlist
->overwrite
= overwrite
;
1227 evlist
->mmap_len
= perf_evlist__mmap_size(pages
);
1228 pr_debug("mmap size %zuB\n", evlist
->mmap_len
);
1229 mp
.mask
= evlist
->mmap_len
- page_size
- 1;
1231 auxtrace_mmap_params__init(&mp
.auxtrace_mp
, evlist
->mmap_len
,
1232 auxtrace_pages
, auxtrace_overwrite
);
1234 evlist__for_each(evlist
, evsel
) {
1235 if ((evsel
->attr
.read_format
& PERF_FORMAT_ID
) &&
1236 evsel
->sample_id
== NULL
&&
1237 perf_evsel__alloc_id(evsel
, cpu_map__nr(cpus
), threads
->nr
) < 0)
1241 if (cpu_map__empty(cpus
))
1242 return perf_evlist__mmap_per_thread(evlist
, &mp
);
1244 return perf_evlist__mmap_per_cpu(evlist
, &mp
);
1247 int perf_evlist__mmap(struct perf_evlist
*evlist
, unsigned int pages
,
1250 return perf_evlist__mmap_ex(evlist
, pages
, overwrite
, 0, false);
1253 int perf_evlist__create_maps(struct perf_evlist
*evlist
, struct target
*target
)
1255 struct cpu_map
*cpus
;
1256 struct thread_map
*threads
;
1258 threads
= thread_map__new_str(target
->pid
, target
->tid
, target
->uid
);
1263 if (target__uses_dummy_map(target
))
1264 cpus
= cpu_map__dummy_new();
1266 cpus
= cpu_map__new(target
->cpu_list
);
1269 goto out_delete_threads
;
1271 evlist
->has_user_cpus
= !!target
->cpu_list
;
1273 perf_evlist__set_maps(evlist
, cpus
, threads
);
1278 thread_map__put(threads
);
1282 void perf_evlist__set_maps(struct perf_evlist
*evlist
, struct cpu_map
*cpus
,
1283 struct thread_map
*threads
)
1286 * Allow for the possibility that one or another of the maps isn't being
1287 * changed i.e. don't put it. Note we are assuming the maps that are
1288 * being applied are brand new and evlist is taking ownership of the
1289 * original reference count of 1. If that is not the case it is up to
1290 * the caller to increase the reference count.
1292 if (cpus
!= evlist
->cpus
) {
1293 cpu_map__put(evlist
->cpus
);
1294 evlist
->cpus
= cpu_map__get(cpus
);
1297 if (threads
!= evlist
->threads
) {
1298 thread_map__put(evlist
->threads
);
1299 evlist
->threads
= thread_map__get(threads
);
1302 perf_evlist__propagate_maps(evlist
);
1305 void __perf_evlist__set_sample_bit(struct perf_evlist
*evlist
,
1306 enum perf_event_sample_format bit
)
1308 struct perf_evsel
*evsel
;
1310 evlist__for_each(evlist
, evsel
)
1311 __perf_evsel__set_sample_bit(evsel
, bit
);
1314 void __perf_evlist__reset_sample_bit(struct perf_evlist
*evlist
,
1315 enum perf_event_sample_format bit
)
1317 struct perf_evsel
*evsel
;
1319 evlist__for_each(evlist
, evsel
)
1320 __perf_evsel__reset_sample_bit(evsel
, bit
);
1323 int perf_evlist__apply_filters(struct perf_evlist
*evlist
, struct perf_evsel
**err_evsel
)
1325 struct perf_evsel
*evsel
;
1327 const int ncpus
= cpu_map__nr(evlist
->cpus
),
1328 nthreads
= thread_map__nr(evlist
->threads
);
1330 evlist__for_each(evlist
, evsel
) {
1331 if (evsel
->filter
== NULL
)
1335 * filters only work for tracepoint event, which doesn't have cpu limit.
1336 * So evlist and evsel should always be same.
1338 err
= perf_evsel__apply_filter(evsel
, ncpus
, nthreads
, evsel
->filter
);
1348 int perf_evlist__set_filter(struct perf_evlist
*evlist
, const char *filter
)
1350 struct perf_evsel
*evsel
;
1353 evlist__for_each(evlist
, evsel
) {
1354 if (evsel
->attr
.type
!= PERF_TYPE_TRACEPOINT
)
1357 err
= perf_evsel__set_filter(evsel
, filter
);
1365 int perf_evlist__set_filter_pids(struct perf_evlist
*evlist
, size_t npids
, pid_t
*pids
)
1371 for (i
= 0; i
< npids
; ++i
) {
1373 if (asprintf(&filter
, "common_pid != %d", pids
[i
]) < 0)
1378 if (asprintf(&tmp
, "%s && common_pid != %d", filter
, pids
[i
]) < 0)
1386 ret
= perf_evlist__set_filter(evlist
, filter
);
1392 int perf_evlist__set_filter_pid(struct perf_evlist
*evlist
, pid_t pid
)
1394 return perf_evlist__set_filter_pids(evlist
, 1, &pid
);
1397 bool perf_evlist__valid_sample_type(struct perf_evlist
*evlist
)
1399 struct perf_evsel
*pos
;
1401 if (evlist
->nr_entries
== 1)
1404 if (evlist
->id_pos
< 0 || evlist
->is_pos
< 0)
1407 evlist__for_each(evlist
, pos
) {
1408 if (pos
->id_pos
!= evlist
->id_pos
||
1409 pos
->is_pos
!= evlist
->is_pos
)
1416 u64
__perf_evlist__combined_sample_type(struct perf_evlist
*evlist
)
1418 struct perf_evsel
*evsel
;
1420 if (evlist
->combined_sample_type
)
1421 return evlist
->combined_sample_type
;
1423 evlist__for_each(evlist
, evsel
)
1424 evlist
->combined_sample_type
|= evsel
->attr
.sample_type
;
1426 return evlist
->combined_sample_type
;
1429 u64
perf_evlist__combined_sample_type(struct perf_evlist
*evlist
)
1431 evlist
->combined_sample_type
= 0;
1432 return __perf_evlist__combined_sample_type(evlist
);
1435 u64
perf_evlist__combined_branch_type(struct perf_evlist
*evlist
)
1437 struct perf_evsel
*evsel
;
1438 u64 branch_type
= 0;
1440 evlist__for_each(evlist
, evsel
)
1441 branch_type
|= evsel
->attr
.branch_sample_type
;
1445 bool perf_evlist__valid_read_format(struct perf_evlist
*evlist
)
1447 struct perf_evsel
*first
= perf_evlist__first(evlist
), *pos
= first
;
1448 u64 read_format
= first
->attr
.read_format
;
1449 u64 sample_type
= first
->attr
.sample_type
;
1451 evlist__for_each(evlist
, pos
) {
1452 if (read_format
!= pos
->attr
.read_format
)
1456 /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1457 if ((sample_type
& PERF_SAMPLE_READ
) &&
1458 !(read_format
& PERF_FORMAT_ID
)) {
1465 u64
perf_evlist__read_format(struct perf_evlist
*evlist
)
1467 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1468 return first
->attr
.read_format
;
1471 u16
perf_evlist__id_hdr_size(struct perf_evlist
*evlist
)
1473 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1474 struct perf_sample
*data
;
1478 if (!first
->attr
.sample_id_all
)
1481 sample_type
= first
->attr
.sample_type
;
1483 if (sample_type
& PERF_SAMPLE_TID
)
1484 size
+= sizeof(data
->tid
) * 2;
1486 if (sample_type
& PERF_SAMPLE_TIME
)
1487 size
+= sizeof(data
->time
);
1489 if (sample_type
& PERF_SAMPLE_ID
)
1490 size
+= sizeof(data
->id
);
1492 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
1493 size
+= sizeof(data
->stream_id
);
1495 if (sample_type
& PERF_SAMPLE_CPU
)
1496 size
+= sizeof(data
->cpu
) * 2;
1498 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
1499 size
+= sizeof(data
->id
);
1504 bool perf_evlist__valid_sample_id_all(struct perf_evlist
*evlist
)
1506 struct perf_evsel
*first
= perf_evlist__first(evlist
), *pos
= first
;
1508 evlist__for_each_continue(evlist
, pos
) {
1509 if (first
->attr
.sample_id_all
!= pos
->attr
.sample_id_all
)
1516 bool perf_evlist__sample_id_all(struct perf_evlist
*evlist
)
1518 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1519 return first
->attr
.sample_id_all
;
1522 void perf_evlist__set_selected(struct perf_evlist
*evlist
,
1523 struct perf_evsel
*evsel
)
1525 evlist
->selected
= evsel
;
1528 void perf_evlist__close(struct perf_evlist
*evlist
)
1530 struct perf_evsel
*evsel
;
1531 int ncpus
= cpu_map__nr(evlist
->cpus
);
1532 int nthreads
= thread_map__nr(evlist
->threads
);
1535 evlist__for_each_reverse(evlist
, evsel
) {
1536 n
= evsel
->cpus
? evsel
->cpus
->nr
: ncpus
;
1537 perf_evsel__close(evsel
, n
, nthreads
);
1541 static int perf_evlist__create_syswide_maps(struct perf_evlist
*evlist
)
1543 struct cpu_map
*cpus
;
1544 struct thread_map
*threads
;
1548 * Try reading /sys/devices/system/cpu/online to get
1551 * FIXME: -ENOMEM is the best we can do here, the cpu_map
1552 * code needs an overhaul to properly forward the
1553 * error, and we may not want to do that fallback to a
1554 * default cpu identity map :-\
1556 cpus
= cpu_map__new(NULL
);
1560 threads
= thread_map__new_dummy();
1564 perf_evlist__set_maps(evlist
, cpus
, threads
);
1572 int perf_evlist__open(struct perf_evlist
*evlist
)
1574 struct perf_evsel
*evsel
;
1578 * Default: one fd per CPU, all threads, aka systemwide
1579 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1581 if (evlist
->threads
== NULL
&& evlist
->cpus
== NULL
) {
1582 err
= perf_evlist__create_syswide_maps(evlist
);
1587 perf_evlist__update_id_pos(evlist
);
1589 evlist__for_each(evlist
, evsel
) {
1590 err
= perf_evsel__open(evsel
, evsel
->cpus
, evsel
->threads
);
1597 perf_evlist__close(evlist
);
1602 int perf_evlist__prepare_workload(struct perf_evlist
*evlist
, struct target
*target
,
1603 const char *argv
[], bool pipe_output
,
1604 void (*exec_error
)(int signo
, siginfo_t
*info
, void *ucontext
))
1606 int child_ready_pipe
[2], go_pipe
[2];
1609 if (pipe(child_ready_pipe
) < 0) {
1610 perror("failed to create 'ready' pipe");
1614 if (pipe(go_pipe
) < 0) {
1615 perror("failed to create 'go' pipe");
1616 goto out_close_ready_pipe
;
1619 evlist
->workload
.pid
= fork();
1620 if (evlist
->workload
.pid
< 0) {
1621 perror("failed to fork");
1622 goto out_close_pipes
;
1625 if (!evlist
->workload
.pid
) {
1631 signal(SIGTERM
, SIG_DFL
);
1633 close(child_ready_pipe
[0]);
1635 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
1638 * Tell the parent we're ready to go
1640 close(child_ready_pipe
[1]);
1643 * Wait until the parent tells us to go.
1645 ret
= read(go_pipe
[0], &bf
, 1);
1647 * The parent will ask for the execvp() to be performed by
1648 * writing exactly one byte, in workload.cork_fd, usually via
1649 * perf_evlist__start_workload().
1651 * For cancelling the workload without actually running it,
1652 * the parent will just close workload.cork_fd, without writing
1653 * anything, i.e. read will return zero and we just exit()
1658 perror("unable to read pipe");
1662 execvp(argv
[0], (char **)argv
);
1667 val
.sival_int
= errno
;
1668 if (sigqueue(getppid(), SIGUSR1
, val
))
1676 struct sigaction act
= {
1677 .sa_flags
= SA_SIGINFO
,
1678 .sa_sigaction
= exec_error
,
1680 sigaction(SIGUSR1
, &act
, NULL
);
1683 if (target__none(target
)) {
1684 if (evlist
->threads
== NULL
) {
1685 fprintf(stderr
, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1686 __func__
, __LINE__
);
1687 goto out_close_pipes
;
1689 thread_map__set_pid(evlist
->threads
, 0, evlist
->workload
.pid
);
1692 close(child_ready_pipe
[1]);
1695 * wait for child to settle
1697 if (read(child_ready_pipe
[0], &bf
, 1) == -1) {
1698 perror("unable to read pipe");
1699 goto out_close_pipes
;
1702 fcntl(go_pipe
[1], F_SETFD
, FD_CLOEXEC
);
1703 evlist
->workload
.cork_fd
= go_pipe
[1];
1704 close(child_ready_pipe
[0]);
1710 out_close_ready_pipe
:
1711 close(child_ready_pipe
[0]);
1712 close(child_ready_pipe
[1]);
1716 int perf_evlist__start_workload(struct perf_evlist
*evlist
)
1718 if (evlist
->workload
.cork_fd
> 0) {
1722 * Remove the cork, let it rip!
1724 ret
= write(evlist
->workload
.cork_fd
, &bf
, 1);
1726 perror("enable to write to pipe");
1728 close(evlist
->workload
.cork_fd
);
1735 int perf_evlist__parse_sample(struct perf_evlist
*evlist
, union perf_event
*event
,
1736 struct perf_sample
*sample
)
1738 struct perf_evsel
*evsel
= perf_evlist__event2evsel(evlist
, event
);
1742 return perf_evsel__parse_sample(evsel
, event
, sample
);
1745 size_t perf_evlist__fprintf(struct perf_evlist
*evlist
, FILE *fp
)
1747 struct perf_evsel
*evsel
;
1750 evlist__for_each(evlist
, evsel
) {
1751 printed
+= fprintf(fp
, "%s%s", evsel
->idx
? ", " : "",
1752 perf_evsel__name(evsel
));
1755 return printed
+ fprintf(fp
, "\n");
1758 int perf_evlist__strerror_open(struct perf_evlist
*evlist
,
1759 int err
, char *buf
, size_t size
)
1762 char sbuf
[STRERR_BUFSIZE
], *emsg
= strerror_r(err
, sbuf
, sizeof(sbuf
));
1767 printed
= scnprintf(buf
, size
,
1769 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg
);
1771 value
= perf_event_paranoid();
1773 printed
+= scnprintf(buf
+ printed
, size
- printed
, "\nHint:\t");
1776 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1777 "For your workloads it needs to be <= 1\nHint:\t");
1779 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1780 "For system wide tracing it needs to be set to -1.\n");
1782 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1783 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1784 "Hint:\tThe current value is %d.", value
);
1787 struct perf_evsel
*first
= perf_evlist__first(evlist
);
1790 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq
) < 0)
1793 if (first
->attr
.sample_freq
< (u64
)max_freq
)
1796 printed
= scnprintf(buf
, size
,
1798 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1799 "Hint:\tThe current value is %d and %" PRIu64
" is being requested.",
1800 emsg
, max_freq
, first
->attr
.sample_freq
);
1805 scnprintf(buf
, size
, "%s", emsg
);
1812 int perf_evlist__strerror_mmap(struct perf_evlist
*evlist
, int err
, char *buf
, size_t size
)
1814 char sbuf
[STRERR_BUFSIZE
], *emsg
= strerror_r(err
, sbuf
, sizeof(sbuf
));
1815 int pages_attempted
= evlist
->mmap_len
/ 1024, pages_max_per_user
, printed
= 0;
1819 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user
);
1820 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1822 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1823 "Hint:\tTried using %zd kB.\n",
1824 emsg
, pages_max_per_user
, pages_attempted
);
1826 if (pages_attempted
>= pages_max_per_user
) {
1827 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1828 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1829 pages_max_per_user
+ pages_attempted
);
1832 printed
+= scnprintf(buf
+ printed
, size
- printed
,
1833 "Hint:\tTry using a smaller -m/--mmap-pages value.");
1836 scnprintf(buf
, size
, "%s", emsg
);
1843 void perf_evlist__to_front(struct perf_evlist
*evlist
,
1844 struct perf_evsel
*move_evsel
)
1846 struct perf_evsel
*evsel
, *n
;
1849 if (move_evsel
== perf_evlist__first(evlist
))
1852 evlist__for_each_safe(evlist
, n
, evsel
) {
1853 if (evsel
->leader
== move_evsel
->leader
)
1854 list_move_tail(&evsel
->node
, &move
);
1857 list_splice(&move
, &evlist
->entries
);
1860 void perf_evlist__set_tracking_event(struct perf_evlist
*evlist
,
1861 struct perf_evsel
*tracking_evsel
)
1863 struct perf_evsel
*evsel
;
1865 if (tracking_evsel
->tracking
)
1868 evlist__for_each(evlist
, evsel
) {
1869 if (evsel
!= tracking_evsel
)
1870 evsel
->tracking
= false;
1873 tracking_evsel
->tracking
= true;
1877 perf_evlist__find_evsel_by_str(struct perf_evlist
*evlist
,
1880 struct perf_evsel
*evsel
;
1882 evlist__for_each(evlist
, evsel
) {
1885 if (strcmp(str
, evsel
->name
) == 0)