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)
11 #include <linux/bitops.h>
12 #include <lk/debugfs.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <sys/resource.h>
22 #include "thread_map.h"
24 #include "perf_regs.h"
31 } perf_missing_features
;
33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
35 int __perf_evsel__sample_size(u64 sample_type
)
37 u64 mask
= sample_type
& PERF_SAMPLE_MASK
;
41 for (i
= 0; i
< 64; i
++) {
42 if (mask
& (1ULL << i
))
52 * __perf_evsel__calc_id_pos - calculate id_pos.
53 * @sample_type: sample type
55 * This function returns the position of the event id (PERF_SAMPLE_ID or
56 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
59 static int __perf_evsel__calc_id_pos(u64 sample_type
)
63 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
66 if (!(sample_type
& PERF_SAMPLE_ID
))
69 if (sample_type
& PERF_SAMPLE_IP
)
72 if (sample_type
& PERF_SAMPLE_TID
)
75 if (sample_type
& PERF_SAMPLE_TIME
)
78 if (sample_type
& PERF_SAMPLE_ADDR
)
85 * __perf_evsel__calc_is_pos - calculate is_pos.
86 * @sample_type: sample type
88 * This function returns the position (counting backwards) of the event id
89 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
90 * sample_id_all is used there is an id sample appended to non-sample events.
92 static int __perf_evsel__calc_is_pos(u64 sample_type
)
96 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
99 if (!(sample_type
& PERF_SAMPLE_ID
))
102 if (sample_type
& PERF_SAMPLE_CPU
)
105 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
111 void perf_evsel__calc_id_pos(struct perf_evsel
*evsel
)
113 evsel
->id_pos
= __perf_evsel__calc_id_pos(evsel
->attr
.sample_type
);
114 evsel
->is_pos
= __perf_evsel__calc_is_pos(evsel
->attr
.sample_type
);
117 void hists__init(struct hists
*hists
)
119 memset(hists
, 0, sizeof(*hists
));
120 hists
->entries_in_array
[0] = hists
->entries_in_array
[1] = RB_ROOT
;
121 hists
->entries_in
= &hists
->entries_in_array
[0];
122 hists
->entries_collapsed
= RB_ROOT
;
123 hists
->entries
= RB_ROOT
;
124 pthread_mutex_init(&hists
->lock
, NULL
);
127 void __perf_evsel__set_sample_bit(struct perf_evsel
*evsel
,
128 enum perf_event_sample_format bit
)
130 if (!(evsel
->attr
.sample_type
& bit
)) {
131 evsel
->attr
.sample_type
|= bit
;
132 evsel
->sample_size
+= sizeof(u64
);
133 perf_evsel__calc_id_pos(evsel
);
137 void __perf_evsel__reset_sample_bit(struct perf_evsel
*evsel
,
138 enum perf_event_sample_format bit
)
140 if (evsel
->attr
.sample_type
& bit
) {
141 evsel
->attr
.sample_type
&= ~bit
;
142 evsel
->sample_size
-= sizeof(u64
);
143 perf_evsel__calc_id_pos(evsel
);
147 void perf_evsel__set_sample_id(struct perf_evsel
*evsel
,
148 bool can_sample_identifier
)
150 if (can_sample_identifier
) {
151 perf_evsel__reset_sample_bit(evsel
, ID
);
152 perf_evsel__set_sample_bit(evsel
, IDENTIFIER
);
154 perf_evsel__set_sample_bit(evsel
, ID
);
156 evsel
->attr
.read_format
|= PERF_FORMAT_ID
;
159 void perf_evsel__init(struct perf_evsel
*evsel
,
160 struct perf_event_attr
*attr
, int idx
)
164 evsel
->leader
= evsel
;
165 INIT_LIST_HEAD(&evsel
->node
);
166 hists__init(&evsel
->hists
);
167 evsel
->sample_size
= __perf_evsel__sample_size(attr
->sample_type
);
168 perf_evsel__calc_id_pos(evsel
);
171 struct perf_evsel
*perf_evsel__new(struct perf_event_attr
*attr
, int idx
)
173 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
176 perf_evsel__init(evsel
, attr
, idx
);
181 struct event_format
*event_format__new(const char *sys
, const char *name
)
185 void *bf
= NULL
, *nbf
;
186 size_t size
= 0, alloc_size
= 0;
187 struct event_format
*format
= NULL
;
189 if (asprintf(&filename
, "%s/%s/%s/format", tracing_events_path
, sys
, name
) < 0)
192 fd
= open(filename
, O_RDONLY
);
194 goto out_free_filename
;
197 if (size
== alloc_size
) {
198 alloc_size
+= BUFSIZ
;
199 nbf
= realloc(bf
, alloc_size
);
205 n
= read(fd
, bf
+ size
, alloc_size
- size
);
211 pevent_parse_format(&format
, bf
, size
, sys
);
222 struct perf_evsel
*perf_evsel__newtp(const char *sys
, const char *name
, int idx
)
224 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
227 struct perf_event_attr attr
= {
228 .type
= PERF_TYPE_TRACEPOINT
,
229 .sample_type
= (PERF_SAMPLE_RAW
| PERF_SAMPLE_TIME
|
230 PERF_SAMPLE_CPU
| PERF_SAMPLE_PERIOD
),
233 if (asprintf(&evsel
->name
, "%s:%s", sys
, name
) < 0)
236 evsel
->tp_format
= event_format__new(sys
, name
);
237 if (evsel
->tp_format
== NULL
)
240 event_attr_init(&attr
);
241 attr
.config
= evsel
->tp_format
->id
;
242 attr
.sample_period
= 1;
243 perf_evsel__init(evsel
, &attr
, idx
);
254 const char *perf_evsel__hw_names
[PERF_COUNT_HW_MAX
] = {
262 "stalled-cycles-frontend",
263 "stalled-cycles-backend",
267 static const char *__perf_evsel__hw_name(u64 config
)
269 if (config
< PERF_COUNT_HW_MAX
&& perf_evsel__hw_names
[config
])
270 return perf_evsel__hw_names
[config
];
272 return "unknown-hardware";
275 static int perf_evsel__add_modifiers(struct perf_evsel
*evsel
, char *bf
, size_t size
)
277 int colon
= 0, r
= 0;
278 struct perf_event_attr
*attr
= &evsel
->attr
;
279 bool exclude_guest_default
= false;
281 #define MOD_PRINT(context, mod) do { \
282 if (!attr->exclude_##context) { \
283 if (!colon) colon = ++r; \
284 r += scnprintf(bf + r, size - r, "%c", mod); \
287 if (attr
->exclude_kernel
|| attr
->exclude_user
|| attr
->exclude_hv
) {
288 MOD_PRINT(kernel
, 'k');
289 MOD_PRINT(user
, 'u');
291 exclude_guest_default
= true;
294 if (attr
->precise_ip
) {
297 r
+= scnprintf(bf
+ r
, size
- r
, "%.*s", attr
->precise_ip
, "ppp");
298 exclude_guest_default
= true;
301 if (attr
->exclude_host
|| attr
->exclude_guest
== exclude_guest_default
) {
302 MOD_PRINT(host
, 'H');
303 MOD_PRINT(guest
, 'G');
311 static int perf_evsel__hw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
313 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__hw_name(evsel
->attr
.config
));
314 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
317 const char *perf_evsel__sw_names
[PERF_COUNT_SW_MAX
] = {
330 static const char *__perf_evsel__sw_name(u64 config
)
332 if (config
< PERF_COUNT_SW_MAX
&& perf_evsel__sw_names
[config
])
333 return perf_evsel__sw_names
[config
];
334 return "unknown-software";
337 static int perf_evsel__sw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
339 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__sw_name(evsel
->attr
.config
));
340 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
343 static int __perf_evsel__bp_name(char *bf
, size_t size
, u64 addr
, u64 type
)
347 r
= scnprintf(bf
, size
, "mem:0x%" PRIx64
":", addr
);
349 if (type
& HW_BREAKPOINT_R
)
350 r
+= scnprintf(bf
+ r
, size
- r
, "r");
352 if (type
& HW_BREAKPOINT_W
)
353 r
+= scnprintf(bf
+ r
, size
- r
, "w");
355 if (type
& HW_BREAKPOINT_X
)
356 r
+= scnprintf(bf
+ r
, size
- r
, "x");
361 static int perf_evsel__bp_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
363 struct perf_event_attr
*attr
= &evsel
->attr
;
364 int r
= __perf_evsel__bp_name(bf
, size
, attr
->bp_addr
, attr
->bp_type
);
365 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
368 const char *perf_evsel__hw_cache
[PERF_COUNT_HW_CACHE_MAX
]
369 [PERF_EVSEL__MAX_ALIASES
] = {
370 { "L1-dcache", "l1-d", "l1d", "L1-data", },
371 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
373 { "dTLB", "d-tlb", "Data-TLB", },
374 { "iTLB", "i-tlb", "Instruction-TLB", },
375 { "branch", "branches", "bpu", "btb", "bpc", },
379 const char *perf_evsel__hw_cache_op
[PERF_COUNT_HW_CACHE_OP_MAX
]
380 [PERF_EVSEL__MAX_ALIASES
] = {
381 { "load", "loads", "read", },
382 { "store", "stores", "write", },
383 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
386 const char *perf_evsel__hw_cache_result
[PERF_COUNT_HW_CACHE_RESULT_MAX
]
387 [PERF_EVSEL__MAX_ALIASES
] = {
388 { "refs", "Reference", "ops", "access", },
389 { "misses", "miss", },
392 #define C(x) PERF_COUNT_HW_CACHE_##x
393 #define CACHE_READ (1 << C(OP_READ))
394 #define CACHE_WRITE (1 << C(OP_WRITE))
395 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
396 #define COP(x) (1 << x)
399 * cache operartion stat
400 * L1I : Read and prefetch only
401 * ITLB and BPU : Read-only
403 static unsigned long perf_evsel__hw_cache_stat
[C(MAX
)] = {
404 [C(L1D
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
405 [C(L1I
)] = (CACHE_READ
| CACHE_PREFETCH
),
406 [C(LL
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
407 [C(DTLB
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
408 [C(ITLB
)] = (CACHE_READ
),
409 [C(BPU
)] = (CACHE_READ
),
410 [C(NODE
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
413 bool perf_evsel__is_cache_op_valid(u8 type
, u8 op
)
415 if (perf_evsel__hw_cache_stat
[type
] & COP(op
))
416 return true; /* valid */
418 return false; /* invalid */
421 int __perf_evsel__hw_cache_type_op_res_name(u8 type
, u8 op
, u8 result
,
422 char *bf
, size_t size
)
425 return scnprintf(bf
, size
, "%s-%s-%s", perf_evsel__hw_cache
[type
][0],
426 perf_evsel__hw_cache_op
[op
][0],
427 perf_evsel__hw_cache_result
[result
][0]);
430 return scnprintf(bf
, size
, "%s-%s", perf_evsel__hw_cache
[type
][0],
431 perf_evsel__hw_cache_op
[op
][1]);
434 static int __perf_evsel__hw_cache_name(u64 config
, char *bf
, size_t size
)
436 u8 op
, result
, type
= (config
>> 0) & 0xff;
437 const char *err
= "unknown-ext-hardware-cache-type";
439 if (type
> PERF_COUNT_HW_CACHE_MAX
)
442 op
= (config
>> 8) & 0xff;
443 err
= "unknown-ext-hardware-cache-op";
444 if (op
> PERF_COUNT_HW_CACHE_OP_MAX
)
447 result
= (config
>> 16) & 0xff;
448 err
= "unknown-ext-hardware-cache-result";
449 if (result
> PERF_COUNT_HW_CACHE_RESULT_MAX
)
452 err
= "invalid-cache";
453 if (!perf_evsel__is_cache_op_valid(type
, op
))
456 return __perf_evsel__hw_cache_type_op_res_name(type
, op
, result
, bf
, size
);
458 return scnprintf(bf
, size
, "%s", err
);
461 static int perf_evsel__hw_cache_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
463 int ret
= __perf_evsel__hw_cache_name(evsel
->attr
.config
, bf
, size
);
464 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
467 static int perf_evsel__raw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
469 int ret
= scnprintf(bf
, size
, "raw 0x%" PRIx64
, evsel
->attr
.config
);
470 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
473 const char *perf_evsel__name(struct perf_evsel
*evsel
)
480 switch (evsel
->attr
.type
) {
482 perf_evsel__raw_name(evsel
, bf
, sizeof(bf
));
485 case PERF_TYPE_HARDWARE
:
486 perf_evsel__hw_name(evsel
, bf
, sizeof(bf
));
489 case PERF_TYPE_HW_CACHE
:
490 perf_evsel__hw_cache_name(evsel
, bf
, sizeof(bf
));
493 case PERF_TYPE_SOFTWARE
:
494 perf_evsel__sw_name(evsel
, bf
, sizeof(bf
));
497 case PERF_TYPE_TRACEPOINT
:
498 scnprintf(bf
, sizeof(bf
), "%s", "unknown tracepoint");
501 case PERF_TYPE_BREAKPOINT
:
502 perf_evsel__bp_name(evsel
, bf
, sizeof(bf
));
506 scnprintf(bf
, sizeof(bf
), "unknown attr type: %d",
511 evsel
->name
= strdup(bf
);
513 return evsel
->name
?: "unknown";
516 const char *perf_evsel__group_name(struct perf_evsel
*evsel
)
518 return evsel
->group_name
?: "anon group";
521 int perf_evsel__group_desc(struct perf_evsel
*evsel
, char *buf
, size_t size
)
524 struct perf_evsel
*pos
;
525 const char *group_name
= perf_evsel__group_name(evsel
);
527 ret
= scnprintf(buf
, size
, "%s", group_name
);
529 ret
+= scnprintf(buf
+ ret
, size
- ret
, " { %s",
530 perf_evsel__name(evsel
));
532 for_each_group_member(pos
, evsel
)
533 ret
+= scnprintf(buf
+ ret
, size
- ret
, ", %s",
534 perf_evsel__name(pos
));
536 ret
+= scnprintf(buf
+ ret
, size
- ret
, " }");
542 * The enable_on_exec/disabled value strategy:
544 * 1) For any type of traced program:
545 * - all independent events and group leaders are disabled
546 * - all group members are enabled
548 * Group members are ruled by group leaders. They need to
549 * be enabled, because the group scheduling relies on that.
551 * 2) For traced programs executed by perf:
552 * - all independent events and group leaders have
554 * - we don't specifically enable or disable any event during
557 * Independent events and group leaders are initially disabled
558 * and get enabled by exec. Group members are ruled by group
559 * leaders as stated in 1).
561 * 3) For traced programs attached by perf (pid/tid):
562 * - we specifically enable or disable all events during
565 * When attaching events to already running traced we
566 * enable/disable events specifically, as there's no
567 * initial traced exec call.
569 void perf_evsel__config(struct perf_evsel
*evsel
,
570 struct perf_record_opts
*opts
)
572 struct perf_evsel
*leader
= evsel
->leader
;
573 struct perf_event_attr
*attr
= &evsel
->attr
;
574 int track
= !evsel
->idx
; /* only the first counter needs these */
576 attr
->sample_id_all
= perf_missing_features
.sample_id_all
? 0 : 1;
577 attr
->inherit
= !opts
->no_inherit
;
579 perf_evsel__set_sample_bit(evsel
, IP
);
580 perf_evsel__set_sample_bit(evsel
, TID
);
582 if (evsel
->sample_read
) {
583 perf_evsel__set_sample_bit(evsel
, READ
);
586 * We need ID even in case of single event, because
587 * PERF_SAMPLE_READ process ID specific data.
589 perf_evsel__set_sample_id(evsel
, false);
592 * Apply group format only if we belong to group
593 * with more than one members.
595 if (leader
->nr_members
> 1) {
596 attr
->read_format
|= PERF_FORMAT_GROUP
;
602 * We default some events to a 1 default interval. But keep
603 * it a weak assumption overridable by the user.
605 if (!attr
->sample_period
|| (opts
->user_freq
!= UINT_MAX
&&
606 opts
->user_interval
!= ULLONG_MAX
)) {
608 perf_evsel__set_sample_bit(evsel
, PERIOD
);
610 attr
->sample_freq
= opts
->freq
;
612 attr
->sample_period
= opts
->default_interval
;
617 * Disable sampling for all group members other
618 * than leader in case leader 'leads' the sampling.
620 if ((leader
!= evsel
) && leader
->sample_read
) {
621 attr
->sample_freq
= 0;
622 attr
->sample_period
= 0;
625 if (opts
->no_samples
)
626 attr
->sample_freq
= 0;
628 if (opts
->inherit_stat
)
629 attr
->inherit_stat
= 1;
631 if (opts
->sample_address
) {
632 perf_evsel__set_sample_bit(evsel
, ADDR
);
633 attr
->mmap_data
= track
;
636 if (opts
->call_graph
) {
637 perf_evsel__set_sample_bit(evsel
, CALLCHAIN
);
639 if (opts
->call_graph
== CALLCHAIN_DWARF
) {
640 perf_evsel__set_sample_bit(evsel
, REGS_USER
);
641 perf_evsel__set_sample_bit(evsel
, STACK_USER
);
642 attr
->sample_regs_user
= PERF_REGS_MASK
;
643 attr
->sample_stack_user
= opts
->stack_dump_size
;
644 attr
->exclude_callchain_user
= 1;
648 if (perf_target__has_cpu(&opts
->target
))
649 perf_evsel__set_sample_bit(evsel
, CPU
);
652 perf_evsel__set_sample_bit(evsel
, PERIOD
);
654 if (!perf_missing_features
.sample_id_all
&&
655 (opts
->sample_time
|| !opts
->no_inherit
||
656 perf_target__has_cpu(&opts
->target
)))
657 perf_evsel__set_sample_bit(evsel
, TIME
);
659 if (opts
->raw_samples
) {
660 perf_evsel__set_sample_bit(evsel
, TIME
);
661 perf_evsel__set_sample_bit(evsel
, RAW
);
662 perf_evsel__set_sample_bit(evsel
, CPU
);
665 if (opts
->sample_address
)
666 attr
->sample_type
|= PERF_SAMPLE_DATA_SRC
;
668 if (opts
->no_delay
) {
670 attr
->wakeup_events
= 1;
672 if (opts
->branch_stack
) {
673 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
674 attr
->branch_sample_type
= opts
->branch_stack
;
677 if (opts
->sample_weight
)
678 attr
->sample_type
|= PERF_SAMPLE_WEIGHT
;
681 attr
->mmap2
= track
&& !perf_missing_features
.mmap2
;
685 * XXX see the function comment above
687 * Disabling only independent events or group leaders,
688 * keeping group members enabled.
690 if (perf_evsel__is_group_leader(evsel
))
694 * Setting enable_on_exec for independent events and
695 * group leaders for traced executed by perf.
697 if (perf_target__none(&opts
->target
) && perf_evsel__is_group_leader(evsel
))
698 attr
->enable_on_exec
= 1;
701 int perf_evsel__alloc_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
704 evsel
->fd
= xyarray__new(ncpus
, nthreads
, sizeof(int));
707 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
708 for (thread
= 0; thread
< nthreads
; thread
++) {
709 FD(evsel
, cpu
, thread
) = -1;
714 return evsel
->fd
!= NULL
? 0 : -ENOMEM
;
717 static int perf_evsel__run_ioctl(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
722 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
723 for (thread
= 0; thread
< nthreads
; thread
++) {
724 int fd
= FD(evsel
, cpu
, thread
),
725 err
= ioctl(fd
, ioc
, arg
);
735 int perf_evsel__set_filter(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
738 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
739 PERF_EVENT_IOC_SET_FILTER
,
743 int perf_evsel__enable(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
745 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
746 PERF_EVENT_IOC_ENABLE
,
750 int perf_evsel__alloc_id(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
752 evsel
->sample_id
= xyarray__new(ncpus
, nthreads
, sizeof(struct perf_sample_id
));
753 if (evsel
->sample_id
== NULL
)
756 evsel
->id
= zalloc(ncpus
* nthreads
* sizeof(u64
));
757 if (evsel
->id
== NULL
) {
758 xyarray__delete(evsel
->sample_id
);
759 evsel
->sample_id
= NULL
;
766 void perf_evsel__reset_counts(struct perf_evsel
*evsel
, int ncpus
)
768 memset(evsel
->counts
, 0, (sizeof(*evsel
->counts
) +
769 (ncpus
* sizeof(struct perf_counts_values
))));
772 int perf_evsel__alloc_counts(struct perf_evsel
*evsel
, int ncpus
)
774 evsel
->counts
= zalloc((sizeof(*evsel
->counts
) +
775 (ncpus
* sizeof(struct perf_counts_values
))));
776 return evsel
->counts
!= NULL
? 0 : -ENOMEM
;
779 void perf_evsel__free_fd(struct perf_evsel
*evsel
)
781 xyarray__delete(evsel
->fd
);
785 void perf_evsel__free_id(struct perf_evsel
*evsel
)
787 xyarray__delete(evsel
->sample_id
);
788 evsel
->sample_id
= NULL
;
793 void perf_evsel__close_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
797 for (cpu
= 0; cpu
< ncpus
; cpu
++)
798 for (thread
= 0; thread
< nthreads
; ++thread
) {
799 close(FD(evsel
, cpu
, thread
));
800 FD(evsel
, cpu
, thread
) = -1;
804 void perf_evsel__free_counts(struct perf_evsel
*evsel
)
809 void perf_evsel__exit(struct perf_evsel
*evsel
)
811 assert(list_empty(&evsel
->node
));
812 perf_evsel__free_fd(evsel
);
813 perf_evsel__free_id(evsel
);
816 void perf_evsel__delete(struct perf_evsel
*evsel
)
818 perf_evsel__exit(evsel
);
819 close_cgroup(evsel
->cgrp
);
820 free(evsel
->group_name
);
821 if (evsel
->tp_format
)
822 pevent_free_format(evsel
->tp_format
);
827 static inline void compute_deltas(struct perf_evsel
*evsel
,
829 struct perf_counts_values
*count
)
831 struct perf_counts_values tmp
;
833 if (!evsel
->prev_raw_counts
)
837 tmp
= evsel
->prev_raw_counts
->aggr
;
838 evsel
->prev_raw_counts
->aggr
= *count
;
840 tmp
= evsel
->prev_raw_counts
->cpu
[cpu
];
841 evsel
->prev_raw_counts
->cpu
[cpu
] = *count
;
844 count
->val
= count
->val
- tmp
.val
;
845 count
->ena
= count
->ena
- tmp
.ena
;
846 count
->run
= count
->run
- tmp
.run
;
849 int __perf_evsel__read_on_cpu(struct perf_evsel
*evsel
,
850 int cpu
, int thread
, bool scale
)
852 struct perf_counts_values count
;
853 size_t nv
= scale
? 3 : 1;
855 if (FD(evsel
, cpu
, thread
) < 0)
858 if (evsel
->counts
== NULL
&& perf_evsel__alloc_counts(evsel
, cpu
+ 1) < 0)
861 if (readn(FD(evsel
, cpu
, thread
), &count
, nv
* sizeof(u64
)) < 0)
864 compute_deltas(evsel
, cpu
, &count
);
869 else if (count
.run
< count
.ena
)
870 count
.val
= (u64
)((double)count
.val
* count
.ena
/ count
.run
+ 0.5);
872 count
.ena
= count
.run
= 0;
874 evsel
->counts
->cpu
[cpu
] = count
;
878 int __perf_evsel__read(struct perf_evsel
*evsel
,
879 int ncpus
, int nthreads
, bool scale
)
881 size_t nv
= scale
? 3 : 1;
883 struct perf_counts_values
*aggr
= &evsel
->counts
->aggr
, count
;
885 aggr
->val
= aggr
->ena
= aggr
->run
= 0;
887 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
888 for (thread
= 0; thread
< nthreads
; thread
++) {
889 if (FD(evsel
, cpu
, thread
) < 0)
892 if (readn(FD(evsel
, cpu
, thread
),
893 &count
, nv
* sizeof(u64
)) < 0)
896 aggr
->val
+= count
.val
;
898 aggr
->ena
+= count
.ena
;
899 aggr
->run
+= count
.run
;
904 compute_deltas(evsel
, -1, aggr
);
906 evsel
->counts
->scaled
= 0;
908 if (aggr
->run
== 0) {
909 evsel
->counts
->scaled
= -1;
914 if (aggr
->run
< aggr
->ena
) {
915 evsel
->counts
->scaled
= 1;
916 aggr
->val
= (u64
)((double)aggr
->val
* aggr
->ena
/ aggr
->run
+ 0.5);
919 aggr
->ena
= aggr
->run
= 0;
924 static int get_group_fd(struct perf_evsel
*evsel
, int cpu
, int thread
)
926 struct perf_evsel
*leader
= evsel
->leader
;
929 if (perf_evsel__is_group_leader(evsel
))
933 * Leader must be already processed/open,
938 fd
= FD(leader
, cpu
, thread
);
944 #define __PRINT_ATTR(fmt, cast, field) \
945 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
947 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
948 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
949 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
950 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
952 #define PRINT_ATTR2N(name1, field1, name2, field2) \
953 fprintf(fp, " %-19s %u %-19s %u\n", \
954 name1, attr->field1, name2, attr->field2)
956 #define PRINT_ATTR2(field1, field2) \
957 PRINT_ATTR2N(#field1, field1, #field2, field2)
959 static size_t perf_event_attr__fprintf(struct perf_event_attr
*attr
, FILE *fp
)
963 ret
+= fprintf(fp
, "%.60s\n", graph_dotted_line
);
964 ret
+= fprintf(fp
, "perf_event_attr:\n");
966 ret
+= PRINT_ATTR_U32(type
);
967 ret
+= PRINT_ATTR_U32(size
);
968 ret
+= PRINT_ATTR_X64(config
);
969 ret
+= PRINT_ATTR_U64(sample_period
);
970 ret
+= PRINT_ATTR_U64(sample_freq
);
971 ret
+= PRINT_ATTR_X64(sample_type
);
972 ret
+= PRINT_ATTR_X64(read_format
);
974 ret
+= PRINT_ATTR2(disabled
, inherit
);
975 ret
+= PRINT_ATTR2(pinned
, exclusive
);
976 ret
+= PRINT_ATTR2(exclude_user
, exclude_kernel
);
977 ret
+= PRINT_ATTR2(exclude_hv
, exclude_idle
);
978 ret
+= PRINT_ATTR2(mmap
, comm
);
979 ret
+= PRINT_ATTR2(freq
, inherit_stat
);
980 ret
+= PRINT_ATTR2(enable_on_exec
, task
);
981 ret
+= PRINT_ATTR2(watermark
, precise_ip
);
982 ret
+= PRINT_ATTR2(mmap_data
, sample_id_all
);
983 ret
+= PRINT_ATTR2(exclude_host
, exclude_guest
);
984 ret
+= PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel
,
985 "excl.callchain_user", exclude_callchain_user
);
987 ret
+= PRINT_ATTR_U32(wakeup_events
);
988 ret
+= PRINT_ATTR_U32(wakeup_watermark
);
989 ret
+= PRINT_ATTR_X32(bp_type
);
990 ret
+= PRINT_ATTR_X64(bp_addr
);
991 ret
+= PRINT_ATTR_X64(config1
);
992 ret
+= PRINT_ATTR_U64(bp_len
);
993 ret
+= PRINT_ATTR_X64(config2
);
994 ret
+= PRINT_ATTR_X64(branch_sample_type
);
995 ret
+= PRINT_ATTR_X64(sample_regs_user
);
996 ret
+= PRINT_ATTR_U32(sample_stack_user
);
998 ret
+= fprintf(fp
, "%.60s\n", graph_dotted_line
);
1003 static int __perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1004 struct thread_map
*threads
)
1007 unsigned long flags
= 0;
1009 enum { NO_CHANGE
, SET_TO_MAX
, INCREASED_MAX
} set_rlimit
= NO_CHANGE
;
1011 if (evsel
->fd
== NULL
&&
1012 perf_evsel__alloc_fd(evsel
, cpus
->nr
, threads
->nr
) < 0)
1016 flags
= PERF_FLAG_PID_CGROUP
;
1017 pid
= evsel
->cgrp
->fd
;
1020 fallback_missing_features
:
1021 if (perf_missing_features
.mmap2
)
1022 evsel
->attr
.mmap2
= 0;
1023 if (perf_missing_features
.exclude_guest
)
1024 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
1026 if (perf_missing_features
.sample_id_all
)
1027 evsel
->attr
.sample_id_all
= 0;
1030 perf_event_attr__fprintf(&evsel
->attr
, stderr
);
1032 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
1034 for (thread
= 0; thread
< threads
->nr
; thread
++) {
1038 pid
= threads
->map
[thread
];
1040 group_fd
= get_group_fd(evsel
, cpu
, thread
);
1042 pr_debug2("perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1043 pid
, cpus
->map
[cpu
], group_fd
, flags
);
1045 FD(evsel
, cpu
, thread
) = sys_perf_event_open(&evsel
->attr
,
1049 if (FD(evsel
, cpu
, thread
) < 0) {
1053 set_rlimit
= NO_CHANGE
;
1061 * perf stat needs between 5 and 22 fds per CPU. When we run out
1062 * of them try to increase the limits.
1064 if (err
== -EMFILE
&& set_rlimit
< INCREASED_MAX
) {
1066 int old_errno
= errno
;
1068 if (getrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1069 if (set_rlimit
== NO_CHANGE
)
1070 l
.rlim_cur
= l
.rlim_max
;
1072 l
.rlim_cur
= l
.rlim_max
+ 1000;
1073 l
.rlim_max
= l
.rlim_cur
;
1075 if (setrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1084 if (err
!= -EINVAL
|| cpu
> 0 || thread
> 0)
1087 if (!perf_missing_features
.mmap2
&& evsel
->attr
.mmap2
) {
1088 perf_missing_features
.mmap2
= true;
1089 goto fallback_missing_features
;
1090 } else if (!perf_missing_features
.exclude_guest
&&
1091 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
1092 perf_missing_features
.exclude_guest
= true;
1093 goto fallback_missing_features
;
1094 } else if (!perf_missing_features
.sample_id_all
) {
1095 perf_missing_features
.sample_id_all
= true;
1096 goto retry_sample_id
;
1101 while (--thread
>= 0) {
1102 close(FD(evsel
, cpu
, thread
));
1103 FD(evsel
, cpu
, thread
) = -1;
1105 thread
= threads
->nr
;
1106 } while (--cpu
>= 0);
1110 void perf_evsel__close(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1112 if (evsel
->fd
== NULL
)
1115 perf_evsel__close_fd(evsel
, ncpus
, nthreads
);
1116 perf_evsel__free_fd(evsel
);
1129 struct thread_map map
;
1131 } empty_thread_map
= {
1136 int perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1137 struct thread_map
*threads
)
1140 /* Work around old compiler warnings about strict aliasing */
1141 cpus
= &empty_cpu_map
.map
;
1144 if (threads
== NULL
)
1145 threads
= &empty_thread_map
.map
;
1147 return __perf_evsel__open(evsel
, cpus
, threads
);
1150 int perf_evsel__open_per_cpu(struct perf_evsel
*evsel
,
1151 struct cpu_map
*cpus
)
1153 return __perf_evsel__open(evsel
, cpus
, &empty_thread_map
.map
);
1156 int perf_evsel__open_per_thread(struct perf_evsel
*evsel
,
1157 struct thread_map
*threads
)
1159 return __perf_evsel__open(evsel
, &empty_cpu_map
.map
, threads
);
1162 static int perf_evsel__parse_id_sample(const struct perf_evsel
*evsel
,
1163 const union perf_event
*event
,
1164 struct perf_sample
*sample
)
1166 u64 type
= evsel
->attr
.sample_type
;
1167 const u64
*array
= event
->sample
.array
;
1168 bool swapped
= evsel
->needs_swap
;
1171 array
+= ((event
->header
.size
-
1172 sizeof(event
->header
)) / sizeof(u64
)) - 1;
1174 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1175 sample
->id
= *array
;
1179 if (type
& PERF_SAMPLE_CPU
) {
1182 /* undo swap of u64, then swap on individual u32s */
1183 u
.val64
= bswap_64(u
.val64
);
1184 u
.val32
[0] = bswap_32(u
.val32
[0]);
1187 sample
->cpu
= u
.val32
[0];
1191 if (type
& PERF_SAMPLE_STREAM_ID
) {
1192 sample
->stream_id
= *array
;
1196 if (type
& PERF_SAMPLE_ID
) {
1197 sample
->id
= *array
;
1201 if (type
& PERF_SAMPLE_TIME
) {
1202 sample
->time
= *array
;
1206 if (type
& PERF_SAMPLE_TID
) {
1209 /* undo swap of u64, then swap on individual u32s */
1210 u
.val64
= bswap_64(u
.val64
);
1211 u
.val32
[0] = bswap_32(u
.val32
[0]);
1212 u
.val32
[1] = bswap_32(u
.val32
[1]);
1215 sample
->pid
= u
.val32
[0];
1216 sample
->tid
= u
.val32
[1];
1222 static inline bool overflow(const void *endp
, u16 max_size
, const void *offset
,
1225 return size
> max_size
|| offset
+ size
> endp
;
1228 #define OVERFLOW_CHECK(offset, size, max_size) \
1230 if (overflow(endp, (max_size), (offset), (size))) \
1234 #define OVERFLOW_CHECK_u64(offset) \
1235 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1237 int perf_evsel__parse_sample(struct perf_evsel
*evsel
, union perf_event
*event
,
1238 struct perf_sample
*data
)
1240 u64 type
= evsel
->attr
.sample_type
;
1241 bool swapped
= evsel
->needs_swap
;
1243 u16 max_size
= event
->header
.size
;
1244 const void *endp
= (void *)event
+ max_size
;
1248 * used for cross-endian analysis. See git commit 65014ab3
1249 * for why this goofiness is needed.
1253 memset(data
, 0, sizeof(*data
));
1254 data
->cpu
= data
->pid
= data
->tid
= -1;
1255 data
->stream_id
= data
->id
= data
->time
= -1ULL;
1259 if (event
->header
.type
!= PERF_RECORD_SAMPLE
) {
1260 if (!evsel
->attr
.sample_id_all
)
1262 return perf_evsel__parse_id_sample(evsel
, event
, data
);
1265 array
= event
->sample
.array
;
1268 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1269 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1270 * check the format does not go past the end of the event.
1272 if (evsel
->sample_size
+ sizeof(event
->header
) > event
->header
.size
)
1276 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1281 if (type
& PERF_SAMPLE_IP
) {
1286 if (type
& PERF_SAMPLE_TID
) {
1289 /* undo swap of u64, then swap on individual u32s */
1290 u
.val64
= bswap_64(u
.val64
);
1291 u
.val32
[0] = bswap_32(u
.val32
[0]);
1292 u
.val32
[1] = bswap_32(u
.val32
[1]);
1295 data
->pid
= u
.val32
[0];
1296 data
->tid
= u
.val32
[1];
1300 if (type
& PERF_SAMPLE_TIME
) {
1301 data
->time
= *array
;
1306 if (type
& PERF_SAMPLE_ADDR
) {
1307 data
->addr
= *array
;
1311 if (type
& PERF_SAMPLE_ID
) {
1316 if (type
& PERF_SAMPLE_STREAM_ID
) {
1317 data
->stream_id
= *array
;
1321 if (type
& PERF_SAMPLE_CPU
) {
1325 /* undo swap of u64, then swap on individual u32s */
1326 u
.val64
= bswap_64(u
.val64
);
1327 u
.val32
[0] = bswap_32(u
.val32
[0]);
1330 data
->cpu
= u
.val32
[0];
1334 if (type
& PERF_SAMPLE_PERIOD
) {
1335 data
->period
= *array
;
1339 if (type
& PERF_SAMPLE_READ
) {
1340 u64 read_format
= evsel
->attr
.read_format
;
1342 OVERFLOW_CHECK_u64(array
);
1343 if (read_format
& PERF_FORMAT_GROUP
)
1344 data
->read
.group
.nr
= *array
;
1346 data
->read
.one
.value
= *array
;
1350 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1351 OVERFLOW_CHECK_u64(array
);
1352 data
->read
.time_enabled
= *array
;
1356 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1357 OVERFLOW_CHECK_u64(array
);
1358 data
->read
.time_running
= *array
;
1362 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1363 if (read_format
& PERF_FORMAT_GROUP
) {
1364 const u64 max_group_nr
= UINT64_MAX
/
1365 sizeof(struct sample_read_value
);
1367 if (data
->read
.group
.nr
> max_group_nr
)
1369 sz
= data
->read
.group
.nr
*
1370 sizeof(struct sample_read_value
);
1371 OVERFLOW_CHECK(array
, sz
, max_size
);
1372 data
->read
.group
.values
=
1373 (struct sample_read_value
*)array
;
1374 array
= (void *)array
+ sz
;
1376 OVERFLOW_CHECK_u64(array
);
1377 data
->read
.one
.id
= *array
;
1382 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1383 const u64 max_callchain_nr
= UINT64_MAX
/ sizeof(u64
);
1385 OVERFLOW_CHECK_u64(array
);
1386 data
->callchain
= (struct ip_callchain
*)array
++;
1387 if (data
->callchain
->nr
> max_callchain_nr
)
1389 sz
= data
->callchain
->nr
* sizeof(u64
);
1390 OVERFLOW_CHECK(array
, sz
, max_size
);
1391 array
= (void *)array
+ sz
;
1394 if (type
& PERF_SAMPLE_RAW
) {
1395 OVERFLOW_CHECK_u64(array
);
1397 if (WARN_ONCE(swapped
,
1398 "Endianness of raw data not corrected!\n")) {
1399 /* undo swap of u64, then swap on individual u32s */
1400 u
.val64
= bswap_64(u
.val64
);
1401 u
.val32
[0] = bswap_32(u
.val32
[0]);
1402 u
.val32
[1] = bswap_32(u
.val32
[1]);
1404 data
->raw_size
= u
.val32
[0];
1405 array
= (void *)array
+ sizeof(u32
);
1407 OVERFLOW_CHECK(array
, data
->raw_size
, max_size
);
1408 data
->raw_data
= (void *)array
;
1409 array
= (void *)array
+ data
->raw_size
;
1412 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1413 const u64 max_branch_nr
= UINT64_MAX
/
1414 sizeof(struct branch_entry
);
1416 OVERFLOW_CHECK_u64(array
);
1417 data
->branch_stack
= (struct branch_stack
*)array
++;
1419 if (data
->branch_stack
->nr
> max_branch_nr
)
1421 sz
= data
->branch_stack
->nr
* sizeof(struct branch_entry
);
1422 OVERFLOW_CHECK(array
, sz
, max_size
);
1423 array
= (void *)array
+ sz
;
1426 if (type
& PERF_SAMPLE_REGS_USER
) {
1427 OVERFLOW_CHECK_u64(array
);
1428 data
->user_regs
.abi
= *array
;
1431 if (data
->user_regs
.abi
) {
1432 u64 regs_user
= evsel
->attr
.sample_regs_user
;
1434 sz
= hweight_long(regs_user
) * sizeof(u64
);
1435 OVERFLOW_CHECK(array
, sz
, max_size
);
1436 data
->user_regs
.regs
= (u64
*)array
;
1437 array
= (void *)array
+ sz
;
1441 if (type
& PERF_SAMPLE_STACK_USER
) {
1442 OVERFLOW_CHECK_u64(array
);
1445 data
->user_stack
.offset
= ((char *)(array
- 1)
1449 data
->user_stack
.size
= 0;
1451 OVERFLOW_CHECK(array
, sz
, max_size
);
1452 data
->user_stack
.data
= (char *)array
;
1453 array
= (void *)array
+ sz
;
1454 OVERFLOW_CHECK_u64(array
);
1455 data
->user_stack
.size
= *array
++;
1460 if (type
& PERF_SAMPLE_WEIGHT
) {
1461 OVERFLOW_CHECK_u64(array
);
1462 data
->weight
= *array
;
1466 data
->data_src
= PERF_MEM_DATA_SRC_NONE
;
1467 if (type
& PERF_SAMPLE_DATA_SRC
) {
1468 OVERFLOW_CHECK_u64(array
);
1469 data
->data_src
= *array
;
1476 size_t perf_event__sample_event_size(const struct perf_sample
*sample
, u64 type
,
1477 u64 sample_regs_user
, u64 read_format
)
1479 size_t sz
, result
= sizeof(struct sample_event
);
1481 if (type
& PERF_SAMPLE_IDENTIFIER
)
1482 result
+= sizeof(u64
);
1484 if (type
& PERF_SAMPLE_IP
)
1485 result
+= sizeof(u64
);
1487 if (type
& PERF_SAMPLE_TID
)
1488 result
+= sizeof(u64
);
1490 if (type
& PERF_SAMPLE_TIME
)
1491 result
+= sizeof(u64
);
1493 if (type
& PERF_SAMPLE_ADDR
)
1494 result
+= sizeof(u64
);
1496 if (type
& PERF_SAMPLE_ID
)
1497 result
+= sizeof(u64
);
1499 if (type
& PERF_SAMPLE_STREAM_ID
)
1500 result
+= sizeof(u64
);
1502 if (type
& PERF_SAMPLE_CPU
)
1503 result
+= sizeof(u64
);
1505 if (type
& PERF_SAMPLE_PERIOD
)
1506 result
+= sizeof(u64
);
1508 if (type
& PERF_SAMPLE_READ
) {
1509 result
+= sizeof(u64
);
1510 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1511 result
+= sizeof(u64
);
1512 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1513 result
+= sizeof(u64
);
1514 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1515 if (read_format
& PERF_FORMAT_GROUP
) {
1516 sz
= sample
->read
.group
.nr
*
1517 sizeof(struct sample_read_value
);
1520 result
+= sizeof(u64
);
1524 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1525 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1529 if (type
& PERF_SAMPLE_RAW
) {
1530 result
+= sizeof(u32
);
1531 result
+= sample
->raw_size
;
1534 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1535 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1540 if (type
& PERF_SAMPLE_REGS_USER
) {
1541 if (sample
->user_regs
.abi
) {
1542 result
+= sizeof(u64
);
1543 sz
= hweight_long(sample_regs_user
) * sizeof(u64
);
1546 result
+= sizeof(u64
);
1550 if (type
& PERF_SAMPLE_STACK_USER
) {
1551 sz
= sample
->user_stack
.size
;
1552 result
+= sizeof(u64
);
1555 result
+= sizeof(u64
);
1559 if (type
& PERF_SAMPLE_WEIGHT
)
1560 result
+= sizeof(u64
);
1562 if (type
& PERF_SAMPLE_DATA_SRC
)
1563 result
+= sizeof(u64
);
1568 int perf_event__synthesize_sample(union perf_event
*event
, u64 type
,
1569 u64 sample_regs_user
, u64 read_format
,
1570 const struct perf_sample
*sample
,
1576 * used for cross-endian analysis. See git commit 65014ab3
1577 * for why this goofiness is needed.
1581 array
= event
->sample
.array
;
1583 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1584 *array
= sample
->id
;
1588 if (type
& PERF_SAMPLE_IP
) {
1589 *array
= sample
->ip
;
1593 if (type
& PERF_SAMPLE_TID
) {
1594 u
.val32
[0] = sample
->pid
;
1595 u
.val32
[1] = sample
->tid
;
1598 * Inverse of what is done in perf_evsel__parse_sample
1600 u
.val32
[0] = bswap_32(u
.val32
[0]);
1601 u
.val32
[1] = bswap_32(u
.val32
[1]);
1602 u
.val64
= bswap_64(u
.val64
);
1609 if (type
& PERF_SAMPLE_TIME
) {
1610 *array
= sample
->time
;
1614 if (type
& PERF_SAMPLE_ADDR
) {
1615 *array
= sample
->addr
;
1619 if (type
& PERF_SAMPLE_ID
) {
1620 *array
= sample
->id
;
1624 if (type
& PERF_SAMPLE_STREAM_ID
) {
1625 *array
= sample
->stream_id
;
1629 if (type
& PERF_SAMPLE_CPU
) {
1630 u
.val32
[0] = sample
->cpu
;
1633 * Inverse of what is done in perf_evsel__parse_sample
1635 u
.val32
[0] = bswap_32(u
.val32
[0]);
1636 u
.val64
= bswap_64(u
.val64
);
1642 if (type
& PERF_SAMPLE_PERIOD
) {
1643 *array
= sample
->period
;
1647 if (type
& PERF_SAMPLE_READ
) {
1648 if (read_format
& PERF_FORMAT_GROUP
)
1649 *array
= sample
->read
.group
.nr
;
1651 *array
= sample
->read
.one
.value
;
1654 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1655 *array
= sample
->read
.time_enabled
;
1659 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1660 *array
= sample
->read
.time_running
;
1664 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1665 if (read_format
& PERF_FORMAT_GROUP
) {
1666 sz
= sample
->read
.group
.nr
*
1667 sizeof(struct sample_read_value
);
1668 memcpy(array
, sample
->read
.group
.values
, sz
);
1669 array
= (void *)array
+ sz
;
1671 *array
= sample
->read
.one
.id
;
1676 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1677 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1678 memcpy(array
, sample
->callchain
, sz
);
1679 array
= (void *)array
+ sz
;
1682 if (type
& PERF_SAMPLE_RAW
) {
1683 u
.val32
[0] = sample
->raw_size
;
1684 if (WARN_ONCE(swapped
,
1685 "Endianness of raw data not corrected!\n")) {
1687 * Inverse of what is done in perf_evsel__parse_sample
1689 u
.val32
[0] = bswap_32(u
.val32
[0]);
1690 u
.val32
[1] = bswap_32(u
.val32
[1]);
1691 u
.val64
= bswap_64(u
.val64
);
1694 array
= (void *)array
+ sizeof(u32
);
1696 memcpy(array
, sample
->raw_data
, sample
->raw_size
);
1697 array
= (void *)array
+ sample
->raw_size
;
1700 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1701 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1703 memcpy(array
, sample
->branch_stack
, sz
);
1704 array
= (void *)array
+ sz
;
1707 if (type
& PERF_SAMPLE_REGS_USER
) {
1708 if (sample
->user_regs
.abi
) {
1709 *array
++ = sample
->user_regs
.abi
;
1710 sz
= hweight_long(sample_regs_user
) * sizeof(u64
);
1711 memcpy(array
, sample
->user_regs
.regs
, sz
);
1712 array
= (void *)array
+ sz
;
1718 if (type
& PERF_SAMPLE_STACK_USER
) {
1719 sz
= sample
->user_stack
.size
;
1722 memcpy(array
, sample
->user_stack
.data
, sz
);
1723 array
= (void *)array
+ sz
;
1728 if (type
& PERF_SAMPLE_WEIGHT
) {
1729 *array
= sample
->weight
;
1733 if (type
& PERF_SAMPLE_DATA_SRC
) {
1734 *array
= sample
->data_src
;
1741 struct format_field
*perf_evsel__field(struct perf_evsel
*evsel
, const char *name
)
1743 return pevent_find_field(evsel
->tp_format
, name
);
1746 void *perf_evsel__rawptr(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
1749 struct format_field
*field
= perf_evsel__field(evsel
, name
);
1755 offset
= field
->offset
;
1757 if (field
->flags
& FIELD_IS_DYNAMIC
) {
1758 offset
= *(int *)(sample
->raw_data
+ field
->offset
);
1762 return sample
->raw_data
+ offset
;
1765 u64
perf_evsel__intval(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
1768 struct format_field
*field
= perf_evsel__field(evsel
, name
);
1775 ptr
= sample
->raw_data
+ field
->offset
;
1777 switch (field
->size
) {
1781 value
= *(u16
*)ptr
;
1784 value
= *(u32
*)ptr
;
1787 value
= *(u64
*)ptr
;
1793 if (!evsel
->needs_swap
)
1796 switch (field
->size
) {
1798 return bswap_16(value
);
1800 return bswap_32(value
);
1802 return bswap_64(value
);
1810 static int comma_fprintf(FILE *fp
, bool *first
, const char *fmt
, ...)
1816 ret
+= fprintf(fp
, ",");
1818 ret
+= fprintf(fp
, ":");
1822 va_start(args
, fmt
);
1823 ret
+= vfprintf(fp
, fmt
, args
);
1828 static int __if_fprintf(FILE *fp
, bool *first
, const char *field
, u64 value
)
1833 return comma_fprintf(fp
, first
, " %s: %" PRIu64
, field
, value
);
1836 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1843 static int bits__fprintf(FILE *fp
, const char *field
, u64 value
,
1844 struct bit_names
*bits
, bool *first
)
1846 int i
= 0, printed
= comma_fprintf(fp
, first
, " %s: ", field
);
1847 bool first_bit
= true;
1850 if (value
& bits
[i
].bit
) {
1851 printed
+= fprintf(fp
, "%s%s", first_bit
? "" : "|", bits
[i
].name
);
1854 } while (bits
[++i
].name
!= NULL
);
1859 static int sample_type__fprintf(FILE *fp
, bool *first
, u64 value
)
1861 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1862 struct bit_names bits
[] = {
1863 bit_name(IP
), bit_name(TID
), bit_name(TIME
), bit_name(ADDR
),
1864 bit_name(READ
), bit_name(CALLCHAIN
), bit_name(ID
), bit_name(CPU
),
1865 bit_name(PERIOD
), bit_name(STREAM_ID
), bit_name(RAW
),
1866 bit_name(BRANCH_STACK
), bit_name(REGS_USER
), bit_name(STACK_USER
),
1867 bit_name(IDENTIFIER
),
1871 return bits__fprintf(fp
, "sample_type", value
, bits
, first
);
1874 static int read_format__fprintf(FILE *fp
, bool *first
, u64 value
)
1876 #define bit_name(n) { PERF_FORMAT_##n, #n }
1877 struct bit_names bits
[] = {
1878 bit_name(TOTAL_TIME_ENABLED
), bit_name(TOTAL_TIME_RUNNING
),
1879 bit_name(ID
), bit_name(GROUP
),
1883 return bits__fprintf(fp
, "read_format", value
, bits
, first
);
1886 int perf_evsel__fprintf(struct perf_evsel
*evsel
,
1887 struct perf_attr_details
*details
, FILE *fp
)
1892 if (details
->event_group
) {
1893 struct perf_evsel
*pos
;
1895 if (!perf_evsel__is_group_leader(evsel
))
1898 if (evsel
->nr_members
> 1)
1899 printed
+= fprintf(fp
, "%s{", evsel
->group_name
?: "");
1901 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
1902 for_each_group_member(pos
, evsel
)
1903 printed
+= fprintf(fp
, ",%s", perf_evsel__name(pos
));
1905 if (evsel
->nr_members
> 1)
1906 printed
+= fprintf(fp
, "}");
1910 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
1912 if (details
->verbose
|| details
->freq
) {
1913 printed
+= comma_fprintf(fp
, &first
, " sample_freq=%" PRIu64
,
1914 (u64
)evsel
->attr
.sample_freq
);
1917 if (details
->verbose
) {
1923 printed
+= sample_type__fprintf(fp
, &first
, evsel
->attr
.sample_type
);
1924 if (evsel
->attr
.read_format
)
1925 printed
+= read_format__fprintf(fp
, &first
, evsel
->attr
.read_format
);
1929 if_print(exclusive
);
1930 if_print(exclude_user
);
1931 if_print(exclude_kernel
);
1932 if_print(exclude_hv
);
1933 if_print(exclude_idle
);
1938 if_print(inherit_stat
);
1939 if_print(enable_on_exec
);
1941 if_print(watermark
);
1942 if_print(precise_ip
);
1943 if_print(mmap_data
);
1944 if_print(sample_id_all
);
1945 if_print(exclude_host
);
1946 if_print(exclude_guest
);
1947 if_print(__reserved_1
);
1948 if_print(wakeup_events
);
1950 if_print(branch_sample_type
);
1957 bool perf_evsel__fallback(struct perf_evsel
*evsel
, int err
,
1958 char *msg
, size_t msgsize
)
1960 if ((err
== ENOENT
|| err
== ENXIO
|| err
== ENODEV
) &&
1961 evsel
->attr
.type
== PERF_TYPE_HARDWARE
&&
1962 evsel
->attr
.config
== PERF_COUNT_HW_CPU_CYCLES
) {
1964 * If it's cycles then fall back to hrtimer based
1965 * cpu-clock-tick sw counter, which is always available even if
1968 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1971 scnprintf(msg
, msgsize
, "%s",
1972 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1974 evsel
->attr
.type
= PERF_TYPE_SOFTWARE
;
1975 evsel
->attr
.config
= PERF_COUNT_SW_CPU_CLOCK
;
1985 int perf_evsel__open_strerror(struct perf_evsel
*evsel
,
1986 struct perf_target
*target
,
1987 int err
, char *msg
, size_t size
)
1992 return scnprintf(msg
, size
,
1993 "You may not have permission to collect %sstats.\n"
1994 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
1995 " -1 - Not paranoid at all\n"
1996 " 0 - Disallow raw tracepoint access for unpriv\n"
1997 " 1 - Disallow cpu events for unpriv\n"
1998 " 2 - Disallow kernel profiling for unpriv",
1999 target
->system_wide
? "system-wide " : "");
2001 return scnprintf(msg
, size
, "The %s event is not supported.",
2002 perf_evsel__name(evsel
));
2004 return scnprintf(msg
, size
, "%s",
2005 "Too many events are opened.\n"
2006 "Try again after reducing the number of events.");
2008 if (target
->cpu_list
)
2009 return scnprintf(msg
, size
, "%s",
2010 "No such device - did you specify an out-of-range profile CPU?\n");
2013 if (evsel
->attr
.precise_ip
)
2014 return scnprintf(msg
, size
, "%s",
2015 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2016 #if defined(__i386__) || defined(__x86_64__)
2017 if (evsel
->attr
.type
== PERF_TYPE_HARDWARE
)
2018 return scnprintf(msg
, size
, "%s",
2019 "No hardware sampling interrupt available.\n"
2020 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2027 return scnprintf(msg
, size
,
2028 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2029 "/bin/dmesg may provide additional information.\n"
2030 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2031 err
, strerror(err
), perf_evsel__name(evsel
));