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)
16 #include "thread_map.h"
18 #include "../../include/linux/perf_event.h"
20 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
21 #define GROUP_FD(group_fd, cpu) (*(int *)xyarray__entry(group_fd, cpu, 0))
23 int __perf_evsel__sample_size(u64 sample_type
)
25 u64 mask
= sample_type
& PERF_SAMPLE_MASK
;
29 for (i
= 0; i
< 64; i
++) {
30 if (mask
& (1ULL << i
))
39 void hists__init(struct hists
*hists
)
41 memset(hists
, 0, sizeof(*hists
));
42 hists
->entries_in_array
[0] = hists
->entries_in_array
[1] = RB_ROOT
;
43 hists
->entries_in
= &hists
->entries_in_array
[0];
44 hists
->entries_collapsed
= RB_ROOT
;
45 hists
->entries
= RB_ROOT
;
46 pthread_mutex_init(&hists
->lock
, NULL
);
49 void perf_evsel__init(struct perf_evsel
*evsel
,
50 struct perf_event_attr
*attr
, int idx
)
54 INIT_LIST_HEAD(&evsel
->node
);
55 hists__init(&evsel
->hists
);
58 struct perf_evsel
*perf_evsel__new(struct perf_event_attr
*attr
, int idx
)
60 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
63 perf_evsel__init(evsel
, attr
, idx
);
68 static const char *perf_evsel__hw_names
[PERF_COUNT_HW_MAX
] = {
76 "stalled-cycles-frontend",
77 "stalled-cycles-backend",
81 const char *__perf_evsel__hw_name(u64 config
)
83 if (config
< PERF_COUNT_HW_MAX
&& perf_evsel__hw_names
[config
])
84 return perf_evsel__hw_names
[config
];
86 return "unknown-hardware";
89 static int perf_evsel__hw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
92 struct perf_event_attr
*attr
= &evsel
->attr
;
93 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__hw_name(attr
->config
));
94 bool exclude_guest_default
= false;
96 #define MOD_PRINT(context, mod) do { \
97 if (!attr->exclude_##context) { \
98 if (!colon) colon = r++; \
99 r += scnprintf(bf + r, size - r, "%c", mod); \
102 if (attr
->exclude_kernel
|| attr
->exclude_user
|| attr
->exclude_hv
) {
103 MOD_PRINT(kernel
, 'k');
104 MOD_PRINT(user
, 'u');
106 exclude_guest_default
= true;
109 if (attr
->precise_ip
) {
112 r
+= scnprintf(bf
+ r
, size
- r
, "%.*s", attr
->precise_ip
, "ppp");
113 exclude_guest_default
= true;
116 if (attr
->exclude_host
|| attr
->exclude_guest
== exclude_guest_default
) {
117 MOD_PRINT(host
, 'H');
118 MOD_PRINT(guest
, 'G');
126 int perf_evsel__name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
130 switch (evsel
->attr
.type
) {
132 ret
= scnprintf(bf
, size
, "raw 0x%" PRIx64
, evsel
->attr
.config
);
135 case PERF_TYPE_HARDWARE
:
136 ret
= perf_evsel__hw_name(evsel
, bf
, size
);
142 * This is the minimal perf_evsel__name so that we can
143 * reconstruct event names taking into account event modifiers.
145 * The old event_name uses it now for raw anr hw events, so that
146 * we don't drag all the parsing stuff into the python binding.
148 * On the next devel cycle the rest of the event naming will be
157 void perf_evsel__config(struct perf_evsel
*evsel
, struct perf_record_opts
*opts
,
158 struct perf_evsel
*first
)
160 struct perf_event_attr
*attr
= &evsel
->attr
;
161 int track
= !evsel
->idx
; /* only the first counter needs these */
164 attr
->sample_id_all
= opts
->sample_id_all_missing
? 0 : 1;
165 attr
->inherit
= !opts
->no_inherit
;
166 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
167 PERF_FORMAT_TOTAL_TIME_RUNNING
|
170 attr
->sample_type
|= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
173 * We default some events to a 1 default interval. But keep
174 * it a weak assumption overridable by the user.
176 if (!attr
->sample_period
|| (opts
->user_freq
!= UINT_MAX
&&
177 opts
->user_interval
!= ULLONG_MAX
)) {
179 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
181 attr
->sample_freq
= opts
->freq
;
183 attr
->sample_period
= opts
->default_interval
;
187 if (opts
->no_samples
)
188 attr
->sample_freq
= 0;
190 if (opts
->inherit_stat
)
191 attr
->inherit_stat
= 1;
193 if (opts
->sample_address
) {
194 attr
->sample_type
|= PERF_SAMPLE_ADDR
;
195 attr
->mmap_data
= track
;
198 if (opts
->call_graph
)
199 attr
->sample_type
|= PERF_SAMPLE_CALLCHAIN
;
201 if (perf_target__has_cpu(&opts
->target
))
202 attr
->sample_type
|= PERF_SAMPLE_CPU
;
205 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
207 if (!opts
->sample_id_all_missing
&&
208 (opts
->sample_time
|| !opts
->no_inherit
||
209 perf_target__has_cpu(&opts
->target
)))
210 attr
->sample_type
|= PERF_SAMPLE_TIME
;
212 if (opts
->raw_samples
) {
213 attr
->sample_type
|= PERF_SAMPLE_TIME
;
214 attr
->sample_type
|= PERF_SAMPLE_RAW
;
215 attr
->sample_type
|= PERF_SAMPLE_CPU
;
218 if (opts
->no_delay
) {
220 attr
->wakeup_events
= 1;
222 if (opts
->branch_stack
) {
223 attr
->sample_type
|= PERF_SAMPLE_BRANCH_STACK
;
224 attr
->branch_sample_type
= opts
->branch_stack
;
230 if (perf_target__none(&opts
->target
) &&
231 (!opts
->group
|| evsel
== first
)) {
232 attr
->enable_on_exec
= 1;
236 int perf_evsel__alloc_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
239 evsel
->fd
= xyarray__new(ncpus
, nthreads
, sizeof(int));
242 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
243 for (thread
= 0; thread
< nthreads
; thread
++) {
244 FD(evsel
, cpu
, thread
) = -1;
249 return evsel
->fd
!= NULL
? 0 : -ENOMEM
;
252 int perf_evsel__alloc_id(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
254 evsel
->sample_id
= xyarray__new(ncpus
, nthreads
, sizeof(struct perf_sample_id
));
255 if (evsel
->sample_id
== NULL
)
258 evsel
->id
= zalloc(ncpus
* nthreads
* sizeof(u64
));
259 if (evsel
->id
== NULL
) {
260 xyarray__delete(evsel
->sample_id
);
261 evsel
->sample_id
= NULL
;
268 int perf_evsel__alloc_counts(struct perf_evsel
*evsel
, int ncpus
)
270 evsel
->counts
= zalloc((sizeof(*evsel
->counts
) +
271 (ncpus
* sizeof(struct perf_counts_values
))));
272 return evsel
->counts
!= NULL
? 0 : -ENOMEM
;
275 void perf_evsel__free_fd(struct perf_evsel
*evsel
)
277 xyarray__delete(evsel
->fd
);
281 void perf_evsel__free_id(struct perf_evsel
*evsel
)
283 xyarray__delete(evsel
->sample_id
);
284 evsel
->sample_id
= NULL
;
289 void perf_evsel__close_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
293 for (cpu
= 0; cpu
< ncpus
; cpu
++)
294 for (thread
= 0; thread
< nthreads
; ++thread
) {
295 close(FD(evsel
, cpu
, thread
));
296 FD(evsel
, cpu
, thread
) = -1;
300 void perf_evsel__exit(struct perf_evsel
*evsel
)
302 assert(list_empty(&evsel
->node
));
303 xyarray__delete(evsel
->fd
);
304 xyarray__delete(evsel
->sample_id
);
308 void perf_evsel__delete(struct perf_evsel
*evsel
)
310 perf_evsel__exit(evsel
);
311 close_cgroup(evsel
->cgrp
);
316 int __perf_evsel__read_on_cpu(struct perf_evsel
*evsel
,
317 int cpu
, int thread
, bool scale
)
319 struct perf_counts_values count
;
320 size_t nv
= scale
? 3 : 1;
322 if (FD(evsel
, cpu
, thread
) < 0)
325 if (evsel
->counts
== NULL
&& perf_evsel__alloc_counts(evsel
, cpu
+ 1) < 0)
328 if (readn(FD(evsel
, cpu
, thread
), &count
, nv
* sizeof(u64
)) < 0)
334 else if (count
.run
< count
.ena
)
335 count
.val
= (u64
)((double)count
.val
* count
.ena
/ count
.run
+ 0.5);
337 count
.ena
= count
.run
= 0;
339 evsel
->counts
->cpu
[cpu
] = count
;
343 int __perf_evsel__read(struct perf_evsel
*evsel
,
344 int ncpus
, int nthreads
, bool scale
)
346 size_t nv
= scale
? 3 : 1;
348 struct perf_counts_values
*aggr
= &evsel
->counts
->aggr
, count
;
350 aggr
->val
= aggr
->ena
= aggr
->run
= 0;
352 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
353 for (thread
= 0; thread
< nthreads
; thread
++) {
354 if (FD(evsel
, cpu
, thread
) < 0)
357 if (readn(FD(evsel
, cpu
, thread
),
358 &count
, nv
* sizeof(u64
)) < 0)
361 aggr
->val
+= count
.val
;
363 aggr
->ena
+= count
.ena
;
364 aggr
->run
+= count
.run
;
369 evsel
->counts
->scaled
= 0;
371 if (aggr
->run
== 0) {
372 evsel
->counts
->scaled
= -1;
377 if (aggr
->run
< aggr
->ena
) {
378 evsel
->counts
->scaled
= 1;
379 aggr
->val
= (u64
)((double)aggr
->val
* aggr
->ena
/ aggr
->run
+ 0.5);
382 aggr
->ena
= aggr
->run
= 0;
387 static int __perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
388 struct thread_map
*threads
, bool group
,
389 struct xyarray
*group_fds
)
392 unsigned long flags
= 0;
395 if (evsel
->fd
== NULL
&&
396 perf_evsel__alloc_fd(evsel
, cpus
->nr
, threads
->nr
) < 0)
400 flags
= PERF_FLAG_PID_CGROUP
;
401 pid
= evsel
->cgrp
->fd
;
404 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
405 int group_fd
= group_fds
? GROUP_FD(group_fds
, cpu
) : -1;
407 for (thread
= 0; thread
< threads
->nr
; thread
++) {
410 pid
= threads
->map
[thread
];
412 FD(evsel
, cpu
, thread
) = sys_perf_event_open(&evsel
->attr
,
416 if (FD(evsel
, cpu
, thread
) < 0) {
421 if (group
&& group_fd
== -1)
422 group_fd
= FD(evsel
, cpu
, thread
);
430 while (--thread
>= 0) {
431 close(FD(evsel
, cpu
, thread
));
432 FD(evsel
, cpu
, thread
) = -1;
434 thread
= threads
->nr
;
435 } while (--cpu
>= 0);
439 void perf_evsel__close(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
441 if (evsel
->fd
== NULL
)
444 perf_evsel__close_fd(evsel
, ncpus
, nthreads
);
445 perf_evsel__free_fd(evsel
);
458 struct thread_map map
;
460 } empty_thread_map
= {
465 int perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
466 struct thread_map
*threads
, bool group
,
467 struct xyarray
*group_fd
)
470 /* Work around old compiler warnings about strict aliasing */
471 cpus
= &empty_cpu_map
.map
;
475 threads
= &empty_thread_map
.map
;
477 return __perf_evsel__open(evsel
, cpus
, threads
, group
, group_fd
);
480 int perf_evsel__open_per_cpu(struct perf_evsel
*evsel
,
481 struct cpu_map
*cpus
, bool group
,
482 struct xyarray
*group_fd
)
484 return __perf_evsel__open(evsel
, cpus
, &empty_thread_map
.map
, group
,
488 int perf_evsel__open_per_thread(struct perf_evsel
*evsel
,
489 struct thread_map
*threads
, bool group
,
490 struct xyarray
*group_fd
)
492 return __perf_evsel__open(evsel
, &empty_cpu_map
.map
, threads
, group
,
496 static int perf_event__parse_id_sample(const union perf_event
*event
, u64 type
,
497 struct perf_sample
*sample
,
500 const u64
*array
= event
->sample
.array
;
503 array
+= ((event
->header
.size
-
504 sizeof(event
->header
)) / sizeof(u64
)) - 1;
506 if (type
& PERF_SAMPLE_CPU
) {
509 /* undo swap of u64, then swap on individual u32s */
510 u
.val64
= bswap_64(u
.val64
);
511 u
.val32
[0] = bswap_32(u
.val32
[0]);
514 sample
->cpu
= u
.val32
[0];
518 if (type
& PERF_SAMPLE_STREAM_ID
) {
519 sample
->stream_id
= *array
;
523 if (type
& PERF_SAMPLE_ID
) {
528 if (type
& PERF_SAMPLE_TIME
) {
529 sample
->time
= *array
;
533 if (type
& PERF_SAMPLE_TID
) {
536 /* undo swap of u64, then swap on individual u32s */
537 u
.val64
= bswap_64(u
.val64
);
538 u
.val32
[0] = bswap_32(u
.val32
[0]);
539 u
.val32
[1] = bswap_32(u
.val32
[1]);
542 sample
->pid
= u
.val32
[0];
543 sample
->tid
= u
.val32
[1];
549 static bool sample_overlap(const union perf_event
*event
,
550 const void *offset
, u64 size
)
552 const void *base
= event
;
554 if (offset
+ size
> base
+ event
->header
.size
)
560 int perf_event__parse_sample(const union perf_event
*event
, u64 type
,
561 int sample_size
, bool sample_id_all
,
562 struct perf_sample
*data
, bool swapped
)
567 * used for cross-endian analysis. See git commit 65014ab3
568 * for why this goofiness is needed.
572 memset(data
, 0, sizeof(*data
));
573 data
->cpu
= data
->pid
= data
->tid
= -1;
574 data
->stream_id
= data
->id
= data
->time
= -1ULL;
577 if (event
->header
.type
!= PERF_RECORD_SAMPLE
) {
580 return perf_event__parse_id_sample(event
, type
, data
, swapped
);
583 array
= event
->sample
.array
;
585 if (sample_size
+ sizeof(event
->header
) > event
->header
.size
)
588 if (type
& PERF_SAMPLE_IP
) {
589 data
->ip
= event
->ip
.ip
;
593 if (type
& PERF_SAMPLE_TID
) {
596 /* undo swap of u64, then swap on individual u32s */
597 u
.val64
= bswap_64(u
.val64
);
598 u
.val32
[0] = bswap_32(u
.val32
[0]);
599 u
.val32
[1] = bswap_32(u
.val32
[1]);
602 data
->pid
= u
.val32
[0];
603 data
->tid
= u
.val32
[1];
607 if (type
& PERF_SAMPLE_TIME
) {
613 if (type
& PERF_SAMPLE_ADDR
) {
619 if (type
& PERF_SAMPLE_ID
) {
624 if (type
& PERF_SAMPLE_STREAM_ID
) {
625 data
->stream_id
= *array
;
629 if (type
& PERF_SAMPLE_CPU
) {
633 /* undo swap of u64, then swap on individual u32s */
634 u
.val64
= bswap_64(u
.val64
);
635 u
.val32
[0] = bswap_32(u
.val32
[0]);
638 data
->cpu
= u
.val32
[0];
642 if (type
& PERF_SAMPLE_PERIOD
) {
643 data
->period
= *array
;
647 if (type
& PERF_SAMPLE_READ
) {
648 fprintf(stderr
, "PERF_SAMPLE_READ is unsupported for now\n");
652 if (type
& PERF_SAMPLE_CALLCHAIN
) {
653 if (sample_overlap(event
, array
, sizeof(data
->callchain
->nr
)))
656 data
->callchain
= (struct ip_callchain
*)array
;
658 if (sample_overlap(event
, array
, data
->callchain
->nr
))
661 array
+= 1 + data
->callchain
->nr
;
664 if (type
& PERF_SAMPLE_RAW
) {
668 if (WARN_ONCE(swapped
,
669 "Endianness of raw data not corrected!\n")) {
670 /* undo swap of u64, then swap on individual u32s */
671 u
.val64
= bswap_64(u
.val64
);
672 u
.val32
[0] = bswap_32(u
.val32
[0]);
673 u
.val32
[1] = bswap_32(u
.val32
[1]);
676 if (sample_overlap(event
, array
, sizeof(u32
)))
679 data
->raw_size
= u
.val32
[0];
680 pdata
= (void *) array
+ sizeof(u32
);
682 if (sample_overlap(event
, pdata
, data
->raw_size
))
685 data
->raw_data
= (void *) pdata
;
687 array
= (void *)array
+ data
->raw_size
+ sizeof(u32
);
690 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
693 data
->branch_stack
= (struct branch_stack
*)array
;
696 sz
= data
->branch_stack
->nr
* sizeof(struct branch_entry
);
703 int perf_event__synthesize_sample(union perf_event
*event
, u64 type
,
704 const struct perf_sample
*sample
,
710 * used for cross-endian analysis. See git commit 65014ab3
711 * for why this goofiness is needed.
715 array
= event
->sample
.array
;
717 if (type
& PERF_SAMPLE_IP
) {
718 event
->ip
.ip
= sample
->ip
;
722 if (type
& PERF_SAMPLE_TID
) {
723 u
.val32
[0] = sample
->pid
;
724 u
.val32
[1] = sample
->tid
;
727 * Inverse of what is done in perf_event__parse_sample
729 u
.val32
[0] = bswap_32(u
.val32
[0]);
730 u
.val32
[1] = bswap_32(u
.val32
[1]);
731 u
.val64
= bswap_64(u
.val64
);
738 if (type
& PERF_SAMPLE_TIME
) {
739 *array
= sample
->time
;
743 if (type
& PERF_SAMPLE_ADDR
) {
744 *array
= sample
->addr
;
748 if (type
& PERF_SAMPLE_ID
) {
753 if (type
& PERF_SAMPLE_STREAM_ID
) {
754 *array
= sample
->stream_id
;
758 if (type
& PERF_SAMPLE_CPU
) {
759 u
.val32
[0] = sample
->cpu
;
762 * Inverse of what is done in perf_event__parse_sample
764 u
.val32
[0] = bswap_32(u
.val32
[0]);
765 u
.val64
= bswap_64(u
.val64
);
771 if (type
& PERF_SAMPLE_PERIOD
) {
772 *array
= sample
->period
;