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
;
167 INIT_LIST_HEAD(&evsel
->node
);
168 hists__init(&evsel
->hists
);
169 evsel
->sample_size
= __perf_evsel__sample_size(attr
->sample_type
);
170 perf_evsel__calc_id_pos(evsel
);
173 struct perf_evsel
*perf_evsel__new_idx(struct perf_event_attr
*attr
, int idx
)
175 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
178 perf_evsel__init(evsel
, attr
, idx
);
183 struct event_format
*event_format__new(const char *sys
, const char *name
)
187 void *bf
= NULL
, *nbf
;
188 size_t size
= 0, alloc_size
= 0;
189 struct event_format
*format
= NULL
;
191 if (asprintf(&filename
, "%s/%s/%s/format", tracing_events_path
, sys
, name
) < 0)
194 fd
= open(filename
, O_RDONLY
);
196 goto out_free_filename
;
199 if (size
== alloc_size
) {
200 alloc_size
+= BUFSIZ
;
201 nbf
= realloc(bf
, alloc_size
);
207 n
= read(fd
, bf
+ size
, alloc_size
- size
);
213 pevent_parse_format(NULL
, &format
, bf
, size
, sys
);
224 struct perf_evsel
*perf_evsel__newtp_idx(const char *sys
, const char *name
, int idx
)
226 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
229 struct perf_event_attr attr
= {
230 .type
= PERF_TYPE_TRACEPOINT
,
231 .sample_type
= (PERF_SAMPLE_RAW
| PERF_SAMPLE_TIME
|
232 PERF_SAMPLE_CPU
| PERF_SAMPLE_PERIOD
),
235 if (asprintf(&evsel
->name
, "%s:%s", sys
, name
) < 0)
238 evsel
->tp_format
= event_format__new(sys
, name
);
239 if (evsel
->tp_format
== NULL
)
242 event_attr_init(&attr
);
243 attr
.config
= evsel
->tp_format
->id
;
244 attr
.sample_period
= 1;
245 perf_evsel__init(evsel
, &attr
, idx
);
256 const char *perf_evsel__hw_names
[PERF_COUNT_HW_MAX
] = {
264 "stalled-cycles-frontend",
265 "stalled-cycles-backend",
269 static const char *__perf_evsel__hw_name(u64 config
)
271 if (config
< PERF_COUNT_HW_MAX
&& perf_evsel__hw_names
[config
])
272 return perf_evsel__hw_names
[config
];
274 return "unknown-hardware";
277 static int perf_evsel__add_modifiers(struct perf_evsel
*evsel
, char *bf
, size_t size
)
279 int colon
= 0, r
= 0;
280 struct perf_event_attr
*attr
= &evsel
->attr
;
281 bool exclude_guest_default
= false;
283 #define MOD_PRINT(context, mod) do { \
284 if (!attr->exclude_##context) { \
285 if (!colon) colon = ++r; \
286 r += scnprintf(bf + r, size - r, "%c", mod); \
289 if (attr
->exclude_kernel
|| attr
->exclude_user
|| attr
->exclude_hv
) {
290 MOD_PRINT(kernel
, 'k');
291 MOD_PRINT(user
, 'u');
293 exclude_guest_default
= true;
296 if (attr
->precise_ip
) {
299 r
+= scnprintf(bf
+ r
, size
- r
, "%.*s", attr
->precise_ip
, "ppp");
300 exclude_guest_default
= true;
303 if (attr
->exclude_host
|| attr
->exclude_guest
== exclude_guest_default
) {
304 MOD_PRINT(host
, 'H');
305 MOD_PRINT(guest
, 'G');
313 static int perf_evsel__hw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
315 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__hw_name(evsel
->attr
.config
));
316 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
319 const char *perf_evsel__sw_names
[PERF_COUNT_SW_MAX
] = {
332 static const char *__perf_evsel__sw_name(u64 config
)
334 if (config
< PERF_COUNT_SW_MAX
&& perf_evsel__sw_names
[config
])
335 return perf_evsel__sw_names
[config
];
336 return "unknown-software";
339 static int perf_evsel__sw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
341 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__sw_name(evsel
->attr
.config
));
342 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
345 static int __perf_evsel__bp_name(char *bf
, size_t size
, u64 addr
, u64 type
)
349 r
= scnprintf(bf
, size
, "mem:0x%" PRIx64
":", addr
);
351 if (type
& HW_BREAKPOINT_R
)
352 r
+= scnprintf(bf
+ r
, size
- r
, "r");
354 if (type
& HW_BREAKPOINT_W
)
355 r
+= scnprintf(bf
+ r
, size
- r
, "w");
357 if (type
& HW_BREAKPOINT_X
)
358 r
+= scnprintf(bf
+ r
, size
- r
, "x");
363 static int perf_evsel__bp_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
365 struct perf_event_attr
*attr
= &evsel
->attr
;
366 int r
= __perf_evsel__bp_name(bf
, size
, attr
->bp_addr
, attr
->bp_type
);
367 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
370 const char *perf_evsel__hw_cache
[PERF_COUNT_HW_CACHE_MAX
]
371 [PERF_EVSEL__MAX_ALIASES
] = {
372 { "L1-dcache", "l1-d", "l1d", "L1-data", },
373 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
375 { "dTLB", "d-tlb", "Data-TLB", },
376 { "iTLB", "i-tlb", "Instruction-TLB", },
377 { "branch", "branches", "bpu", "btb", "bpc", },
381 const char *perf_evsel__hw_cache_op
[PERF_COUNT_HW_CACHE_OP_MAX
]
382 [PERF_EVSEL__MAX_ALIASES
] = {
383 { "load", "loads", "read", },
384 { "store", "stores", "write", },
385 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
388 const char *perf_evsel__hw_cache_result
[PERF_COUNT_HW_CACHE_RESULT_MAX
]
389 [PERF_EVSEL__MAX_ALIASES
] = {
390 { "refs", "Reference", "ops", "access", },
391 { "misses", "miss", },
394 #define C(x) PERF_COUNT_HW_CACHE_##x
395 #define CACHE_READ (1 << C(OP_READ))
396 #define CACHE_WRITE (1 << C(OP_WRITE))
397 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
398 #define COP(x) (1 << x)
401 * cache operartion stat
402 * L1I : Read and prefetch only
403 * ITLB and BPU : Read-only
405 static unsigned long perf_evsel__hw_cache_stat
[C(MAX
)] = {
406 [C(L1D
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
407 [C(L1I
)] = (CACHE_READ
| CACHE_PREFETCH
),
408 [C(LL
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
409 [C(DTLB
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
410 [C(ITLB
)] = (CACHE_READ
),
411 [C(BPU
)] = (CACHE_READ
),
412 [C(NODE
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
415 bool perf_evsel__is_cache_op_valid(u8 type
, u8 op
)
417 if (perf_evsel__hw_cache_stat
[type
] & COP(op
))
418 return true; /* valid */
420 return false; /* invalid */
423 int __perf_evsel__hw_cache_type_op_res_name(u8 type
, u8 op
, u8 result
,
424 char *bf
, size_t size
)
427 return scnprintf(bf
, size
, "%s-%s-%s", perf_evsel__hw_cache
[type
][0],
428 perf_evsel__hw_cache_op
[op
][0],
429 perf_evsel__hw_cache_result
[result
][0]);
432 return scnprintf(bf
, size
, "%s-%s", perf_evsel__hw_cache
[type
][0],
433 perf_evsel__hw_cache_op
[op
][1]);
436 static int __perf_evsel__hw_cache_name(u64 config
, char *bf
, size_t size
)
438 u8 op
, result
, type
= (config
>> 0) & 0xff;
439 const char *err
= "unknown-ext-hardware-cache-type";
441 if (type
> PERF_COUNT_HW_CACHE_MAX
)
444 op
= (config
>> 8) & 0xff;
445 err
= "unknown-ext-hardware-cache-op";
446 if (op
> PERF_COUNT_HW_CACHE_OP_MAX
)
449 result
= (config
>> 16) & 0xff;
450 err
= "unknown-ext-hardware-cache-result";
451 if (result
> PERF_COUNT_HW_CACHE_RESULT_MAX
)
454 err
= "invalid-cache";
455 if (!perf_evsel__is_cache_op_valid(type
, op
))
458 return __perf_evsel__hw_cache_type_op_res_name(type
, op
, result
, bf
, size
);
460 return scnprintf(bf
, size
, "%s", err
);
463 static int perf_evsel__hw_cache_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
465 int ret
= __perf_evsel__hw_cache_name(evsel
->attr
.config
, bf
, size
);
466 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
469 static int perf_evsel__raw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
471 int ret
= scnprintf(bf
, size
, "raw 0x%" PRIx64
, evsel
->attr
.config
);
472 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
475 const char *perf_evsel__name(struct perf_evsel
*evsel
)
482 switch (evsel
->attr
.type
) {
484 perf_evsel__raw_name(evsel
, bf
, sizeof(bf
));
487 case PERF_TYPE_HARDWARE
:
488 perf_evsel__hw_name(evsel
, bf
, sizeof(bf
));
491 case PERF_TYPE_HW_CACHE
:
492 perf_evsel__hw_cache_name(evsel
, bf
, sizeof(bf
));
495 case PERF_TYPE_SOFTWARE
:
496 perf_evsel__sw_name(evsel
, bf
, sizeof(bf
));
499 case PERF_TYPE_TRACEPOINT
:
500 scnprintf(bf
, sizeof(bf
), "%s", "unknown tracepoint");
503 case PERF_TYPE_BREAKPOINT
:
504 perf_evsel__bp_name(evsel
, bf
, sizeof(bf
));
508 scnprintf(bf
, sizeof(bf
), "unknown attr type: %d",
513 evsel
->name
= strdup(bf
);
515 return evsel
->name
?: "unknown";
518 const char *perf_evsel__group_name(struct perf_evsel
*evsel
)
520 return evsel
->group_name
?: "anon group";
523 int perf_evsel__group_desc(struct perf_evsel
*evsel
, char *buf
, size_t size
)
526 struct perf_evsel
*pos
;
527 const char *group_name
= perf_evsel__group_name(evsel
);
529 ret
= scnprintf(buf
, size
, "%s", group_name
);
531 ret
+= scnprintf(buf
+ ret
, size
- ret
, " { %s",
532 perf_evsel__name(evsel
));
534 for_each_group_member(pos
, evsel
)
535 ret
+= scnprintf(buf
+ ret
, size
- ret
, ", %s",
536 perf_evsel__name(pos
));
538 ret
+= scnprintf(buf
+ ret
, size
- ret
, " }");
544 * The enable_on_exec/disabled value strategy:
546 * 1) For any type of traced program:
547 * - all independent events and group leaders are disabled
548 * - all group members are enabled
550 * Group members are ruled by group leaders. They need to
551 * be enabled, because the group scheduling relies on that.
553 * 2) For traced programs executed by perf:
554 * - all independent events and group leaders have
556 * - we don't specifically enable or disable any event during
559 * Independent events and group leaders are initially disabled
560 * and get enabled by exec. Group members are ruled by group
561 * leaders as stated in 1).
563 * 3) For traced programs attached by perf (pid/tid):
564 * - we specifically enable or disable all events during
567 * When attaching events to already running traced we
568 * enable/disable events specifically, as there's no
569 * initial traced exec call.
571 void perf_evsel__config(struct perf_evsel
*evsel
,
572 struct perf_record_opts
*opts
)
574 struct perf_evsel
*leader
= evsel
->leader
;
575 struct perf_event_attr
*attr
= &evsel
->attr
;
576 int track
= !evsel
->idx
; /* only the first counter needs these */
577 bool per_cpu
= opts
->target
.default_per_cpu
&& !opts
->target
.per_thread
;
579 attr
->sample_id_all
= perf_missing_features
.sample_id_all
? 0 : 1;
580 attr
->inherit
= !opts
->no_inherit
;
582 perf_evsel__set_sample_bit(evsel
, IP
);
583 perf_evsel__set_sample_bit(evsel
, TID
);
585 if (evsel
->sample_read
) {
586 perf_evsel__set_sample_bit(evsel
, READ
);
589 * We need ID even in case of single event, because
590 * PERF_SAMPLE_READ process ID specific data.
592 perf_evsel__set_sample_id(evsel
, false);
595 * Apply group format only if we belong to group
596 * with more than one members.
598 if (leader
->nr_members
> 1) {
599 attr
->read_format
|= PERF_FORMAT_GROUP
;
605 * We default some events to a 1 default interval. But keep
606 * it a weak assumption overridable by the user.
608 if (!attr
->sample_period
|| (opts
->user_freq
!= UINT_MAX
&&
609 opts
->user_interval
!= ULLONG_MAX
)) {
611 perf_evsel__set_sample_bit(evsel
, PERIOD
);
613 attr
->sample_freq
= opts
->freq
;
615 attr
->sample_period
= opts
->default_interval
;
620 * Disable sampling for all group members other
621 * than leader in case leader 'leads' the sampling.
623 if ((leader
!= evsel
) && leader
->sample_read
) {
624 attr
->sample_freq
= 0;
625 attr
->sample_period
= 0;
628 if (opts
->no_samples
)
629 attr
->sample_freq
= 0;
631 if (opts
->inherit_stat
)
632 attr
->inherit_stat
= 1;
634 if (opts
->sample_address
) {
635 perf_evsel__set_sample_bit(evsel
, ADDR
);
636 attr
->mmap_data
= track
;
639 if (opts
->call_graph
) {
640 perf_evsel__set_sample_bit(evsel
, CALLCHAIN
);
642 if (opts
->call_graph
== CALLCHAIN_DWARF
) {
643 perf_evsel__set_sample_bit(evsel
, REGS_USER
);
644 perf_evsel__set_sample_bit(evsel
, STACK_USER
);
645 attr
->sample_regs_user
= PERF_REGS_MASK
;
646 attr
->sample_stack_user
= opts
->stack_dump_size
;
647 attr
->exclude_callchain_user
= 1;
651 if (target__has_cpu(&opts
->target
))
652 perf_evsel__set_sample_bit(evsel
, CPU
);
655 perf_evsel__set_sample_bit(evsel
, PERIOD
);
657 if (!perf_missing_features
.sample_id_all
&&
658 (opts
->sample_time
|| !opts
->no_inherit
||
659 target__has_cpu(&opts
->target
) || per_cpu
))
660 perf_evsel__set_sample_bit(evsel
, TIME
);
662 if (opts
->raw_samples
) {
663 perf_evsel__set_sample_bit(evsel
, TIME
);
664 perf_evsel__set_sample_bit(evsel
, RAW
);
665 perf_evsel__set_sample_bit(evsel
, CPU
);
668 if (opts
->sample_address
)
669 perf_evsel__set_sample_bit(evsel
, DATA_SRC
);
671 if (opts
->no_delay
) {
673 attr
->wakeup_events
= 1;
675 if (opts
->branch_stack
) {
676 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
677 attr
->branch_sample_type
= opts
->branch_stack
;
680 if (opts
->sample_weight
)
681 perf_evsel__set_sample_bit(evsel
, WEIGHT
);
686 if (opts
->sample_transaction
)
687 perf_evsel__set_sample_bit(evsel
, TRANSACTION
);
690 * XXX see the function comment above
692 * Disabling only independent events or group leaders,
693 * keeping group members enabled.
695 if (perf_evsel__is_group_leader(evsel
))
699 * Setting enable_on_exec for independent events and
700 * group leaders for traced executed by perf.
702 if (target__none(&opts
->target
) && perf_evsel__is_group_leader(evsel
))
703 attr
->enable_on_exec
= 1;
706 int perf_evsel__alloc_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
709 evsel
->fd
= xyarray__new(ncpus
, nthreads
, sizeof(int));
712 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
713 for (thread
= 0; thread
< nthreads
; thread
++) {
714 FD(evsel
, cpu
, thread
) = -1;
719 return evsel
->fd
!= NULL
? 0 : -ENOMEM
;
722 static int perf_evsel__run_ioctl(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
727 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
728 for (thread
= 0; thread
< nthreads
; thread
++) {
729 int fd
= FD(evsel
, cpu
, thread
),
730 err
= ioctl(fd
, ioc
, arg
);
740 int perf_evsel__set_filter(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
743 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
744 PERF_EVENT_IOC_SET_FILTER
,
748 int perf_evsel__enable(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
750 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
751 PERF_EVENT_IOC_ENABLE
,
755 int perf_evsel__alloc_id(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
757 evsel
->sample_id
= xyarray__new(ncpus
, nthreads
, sizeof(struct perf_sample_id
));
758 if (evsel
->sample_id
== NULL
)
761 evsel
->id
= zalloc(ncpus
* nthreads
* sizeof(u64
));
762 if (evsel
->id
== NULL
) {
763 xyarray__delete(evsel
->sample_id
);
764 evsel
->sample_id
= NULL
;
771 void perf_evsel__reset_counts(struct perf_evsel
*evsel
, int ncpus
)
773 memset(evsel
->counts
, 0, (sizeof(*evsel
->counts
) +
774 (ncpus
* sizeof(struct perf_counts_values
))));
777 int perf_evsel__alloc_counts(struct perf_evsel
*evsel
, int ncpus
)
779 evsel
->counts
= zalloc((sizeof(*evsel
->counts
) +
780 (ncpus
* sizeof(struct perf_counts_values
))));
781 return evsel
->counts
!= NULL
? 0 : -ENOMEM
;
784 void perf_evsel__free_fd(struct perf_evsel
*evsel
)
786 xyarray__delete(evsel
->fd
);
790 void perf_evsel__free_id(struct perf_evsel
*evsel
)
792 xyarray__delete(evsel
->sample_id
);
793 evsel
->sample_id
= NULL
;
798 void perf_evsel__close_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
802 for (cpu
= 0; cpu
< ncpus
; cpu
++)
803 for (thread
= 0; thread
< nthreads
; ++thread
) {
804 close(FD(evsel
, cpu
, thread
));
805 FD(evsel
, cpu
, thread
) = -1;
809 void perf_evsel__free_counts(struct perf_evsel
*evsel
)
814 void perf_evsel__exit(struct perf_evsel
*evsel
)
816 assert(list_empty(&evsel
->node
));
817 perf_evsel__free_fd(evsel
);
818 perf_evsel__free_id(evsel
);
821 void perf_evsel__delete(struct perf_evsel
*evsel
)
823 perf_evsel__exit(evsel
);
824 close_cgroup(evsel
->cgrp
);
825 free(evsel
->group_name
);
826 if (evsel
->tp_format
)
827 pevent_free_format(evsel
->tp_format
);
832 static inline void compute_deltas(struct perf_evsel
*evsel
,
834 struct perf_counts_values
*count
)
836 struct perf_counts_values tmp
;
838 if (!evsel
->prev_raw_counts
)
842 tmp
= evsel
->prev_raw_counts
->aggr
;
843 evsel
->prev_raw_counts
->aggr
= *count
;
845 tmp
= evsel
->prev_raw_counts
->cpu
[cpu
];
846 evsel
->prev_raw_counts
->cpu
[cpu
] = *count
;
849 count
->val
= count
->val
- tmp
.val
;
850 count
->ena
= count
->ena
- tmp
.ena
;
851 count
->run
= count
->run
- tmp
.run
;
854 int __perf_evsel__read_on_cpu(struct perf_evsel
*evsel
,
855 int cpu
, int thread
, bool scale
)
857 struct perf_counts_values count
;
858 size_t nv
= scale
? 3 : 1;
860 if (FD(evsel
, cpu
, thread
) < 0)
863 if (evsel
->counts
== NULL
&& perf_evsel__alloc_counts(evsel
, cpu
+ 1) < 0)
866 if (readn(FD(evsel
, cpu
, thread
), &count
, nv
* sizeof(u64
)) < 0)
869 compute_deltas(evsel
, cpu
, &count
);
874 else if (count
.run
< count
.ena
)
875 count
.val
= (u64
)((double)count
.val
* count
.ena
/ count
.run
+ 0.5);
877 count
.ena
= count
.run
= 0;
879 evsel
->counts
->cpu
[cpu
] = count
;
883 int __perf_evsel__read(struct perf_evsel
*evsel
,
884 int ncpus
, int nthreads
, bool scale
)
886 size_t nv
= scale
? 3 : 1;
888 struct perf_counts_values
*aggr
= &evsel
->counts
->aggr
, count
;
890 aggr
->val
= aggr
->ena
= aggr
->run
= 0;
892 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
893 for (thread
= 0; thread
< nthreads
; thread
++) {
894 if (FD(evsel
, cpu
, thread
) < 0)
897 if (readn(FD(evsel
, cpu
, thread
),
898 &count
, nv
* sizeof(u64
)) < 0)
901 aggr
->val
+= count
.val
;
903 aggr
->ena
+= count
.ena
;
904 aggr
->run
+= count
.run
;
909 compute_deltas(evsel
, -1, aggr
);
911 evsel
->counts
->scaled
= 0;
913 if (aggr
->run
== 0) {
914 evsel
->counts
->scaled
= -1;
919 if (aggr
->run
< aggr
->ena
) {
920 evsel
->counts
->scaled
= 1;
921 aggr
->val
= (u64
)((double)aggr
->val
* aggr
->ena
/ aggr
->run
+ 0.5);
924 aggr
->ena
= aggr
->run
= 0;
929 static int get_group_fd(struct perf_evsel
*evsel
, int cpu
, int thread
)
931 struct perf_evsel
*leader
= evsel
->leader
;
934 if (perf_evsel__is_group_leader(evsel
))
938 * Leader must be already processed/open,
943 fd
= FD(leader
, cpu
, thread
);
949 #define __PRINT_ATTR(fmt, cast, field) \
950 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
952 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
953 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
954 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
955 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
957 #define PRINT_ATTR2N(name1, field1, name2, field2) \
958 fprintf(fp, " %-19s %u %-19s %u\n", \
959 name1, attr->field1, name2, attr->field2)
961 #define PRINT_ATTR2(field1, field2) \
962 PRINT_ATTR2N(#field1, field1, #field2, field2)
964 static size_t perf_event_attr__fprintf(struct perf_event_attr
*attr
, FILE *fp
)
968 ret
+= fprintf(fp
, "%.60s\n", graph_dotted_line
);
969 ret
+= fprintf(fp
, "perf_event_attr:\n");
971 ret
+= PRINT_ATTR_U32(type
);
972 ret
+= PRINT_ATTR_U32(size
);
973 ret
+= PRINT_ATTR_X64(config
);
974 ret
+= PRINT_ATTR_U64(sample_period
);
975 ret
+= PRINT_ATTR_U64(sample_freq
);
976 ret
+= PRINT_ATTR_X64(sample_type
);
977 ret
+= PRINT_ATTR_X64(read_format
);
979 ret
+= PRINT_ATTR2(disabled
, inherit
);
980 ret
+= PRINT_ATTR2(pinned
, exclusive
);
981 ret
+= PRINT_ATTR2(exclude_user
, exclude_kernel
);
982 ret
+= PRINT_ATTR2(exclude_hv
, exclude_idle
);
983 ret
+= PRINT_ATTR2(mmap
, comm
);
984 ret
+= PRINT_ATTR2(freq
, inherit_stat
);
985 ret
+= PRINT_ATTR2(enable_on_exec
, task
);
986 ret
+= PRINT_ATTR2(watermark
, precise_ip
);
987 ret
+= PRINT_ATTR2(mmap_data
, sample_id_all
);
988 ret
+= PRINT_ATTR2(exclude_host
, exclude_guest
);
989 ret
+= PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel
,
990 "excl.callchain_user", exclude_callchain_user
);
991 ret
+= PRINT_ATTR_U32(mmap2
);
993 ret
+= PRINT_ATTR_U32(wakeup_events
);
994 ret
+= PRINT_ATTR_U32(wakeup_watermark
);
995 ret
+= PRINT_ATTR_X32(bp_type
);
996 ret
+= PRINT_ATTR_X64(bp_addr
);
997 ret
+= PRINT_ATTR_X64(config1
);
998 ret
+= PRINT_ATTR_U64(bp_len
);
999 ret
+= PRINT_ATTR_X64(config2
);
1000 ret
+= PRINT_ATTR_X64(branch_sample_type
);
1001 ret
+= PRINT_ATTR_X64(sample_regs_user
);
1002 ret
+= PRINT_ATTR_U32(sample_stack_user
);
1004 ret
+= fprintf(fp
, "%.60s\n", graph_dotted_line
);
1009 static int __perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1010 struct thread_map
*threads
)
1013 unsigned long flags
= 0;
1015 enum { NO_CHANGE
, SET_TO_MAX
, INCREASED_MAX
} set_rlimit
= NO_CHANGE
;
1017 if (evsel
->fd
== NULL
&&
1018 perf_evsel__alloc_fd(evsel
, cpus
->nr
, threads
->nr
) < 0)
1022 flags
= PERF_FLAG_PID_CGROUP
;
1023 pid
= evsel
->cgrp
->fd
;
1026 fallback_missing_features
:
1027 if (perf_missing_features
.mmap2
)
1028 evsel
->attr
.mmap2
= 0;
1029 if (perf_missing_features
.exclude_guest
)
1030 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
1032 if (perf_missing_features
.sample_id_all
)
1033 evsel
->attr
.sample_id_all
= 0;
1036 perf_event_attr__fprintf(&evsel
->attr
, stderr
);
1038 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
1040 for (thread
= 0; thread
< threads
->nr
; thread
++) {
1044 pid
= threads
->map
[thread
];
1046 group_fd
= get_group_fd(evsel
, cpu
, thread
);
1048 pr_debug2("perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1049 pid
, cpus
->map
[cpu
], group_fd
, flags
);
1051 FD(evsel
, cpu
, thread
) = sys_perf_event_open(&evsel
->attr
,
1055 if (FD(evsel
, cpu
, thread
) < 0) {
1057 pr_debug2("perf_event_open failed, error %d\n",
1061 set_rlimit
= NO_CHANGE
;
1069 * perf stat needs between 5 and 22 fds per CPU. When we run out
1070 * of them try to increase the limits.
1072 if (err
== -EMFILE
&& set_rlimit
< INCREASED_MAX
) {
1074 int old_errno
= errno
;
1076 if (getrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1077 if (set_rlimit
== NO_CHANGE
)
1078 l
.rlim_cur
= l
.rlim_max
;
1080 l
.rlim_cur
= l
.rlim_max
+ 1000;
1081 l
.rlim_max
= l
.rlim_cur
;
1083 if (setrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1092 if (err
!= -EINVAL
|| cpu
> 0 || thread
> 0)
1095 if (!perf_missing_features
.mmap2
&& evsel
->attr
.mmap2
) {
1096 perf_missing_features
.mmap2
= true;
1097 goto fallback_missing_features
;
1098 } else if (!perf_missing_features
.exclude_guest
&&
1099 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
1100 perf_missing_features
.exclude_guest
= true;
1101 goto fallback_missing_features
;
1102 } else if (!perf_missing_features
.sample_id_all
) {
1103 perf_missing_features
.sample_id_all
= true;
1104 goto retry_sample_id
;
1109 while (--thread
>= 0) {
1110 close(FD(evsel
, cpu
, thread
));
1111 FD(evsel
, cpu
, thread
) = -1;
1113 thread
= threads
->nr
;
1114 } while (--cpu
>= 0);
1118 void perf_evsel__close(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1120 if (evsel
->fd
== NULL
)
1123 perf_evsel__close_fd(evsel
, ncpus
, nthreads
);
1124 perf_evsel__free_fd(evsel
);
1137 struct thread_map map
;
1139 } empty_thread_map
= {
1144 int perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1145 struct thread_map
*threads
)
1148 /* Work around old compiler warnings about strict aliasing */
1149 cpus
= &empty_cpu_map
.map
;
1152 if (threads
== NULL
)
1153 threads
= &empty_thread_map
.map
;
1155 return __perf_evsel__open(evsel
, cpus
, threads
);
1158 int perf_evsel__open_per_cpu(struct perf_evsel
*evsel
,
1159 struct cpu_map
*cpus
)
1161 return __perf_evsel__open(evsel
, cpus
, &empty_thread_map
.map
);
1164 int perf_evsel__open_per_thread(struct perf_evsel
*evsel
,
1165 struct thread_map
*threads
)
1167 return __perf_evsel__open(evsel
, &empty_cpu_map
.map
, threads
);
1170 static int perf_evsel__parse_id_sample(const struct perf_evsel
*evsel
,
1171 const union perf_event
*event
,
1172 struct perf_sample
*sample
)
1174 u64 type
= evsel
->attr
.sample_type
;
1175 const u64
*array
= event
->sample
.array
;
1176 bool swapped
= evsel
->needs_swap
;
1179 array
+= ((event
->header
.size
-
1180 sizeof(event
->header
)) / sizeof(u64
)) - 1;
1182 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1183 sample
->id
= *array
;
1187 if (type
& PERF_SAMPLE_CPU
) {
1190 /* undo swap of u64, then swap on individual u32s */
1191 u
.val64
= bswap_64(u
.val64
);
1192 u
.val32
[0] = bswap_32(u
.val32
[0]);
1195 sample
->cpu
= u
.val32
[0];
1199 if (type
& PERF_SAMPLE_STREAM_ID
) {
1200 sample
->stream_id
= *array
;
1204 if (type
& PERF_SAMPLE_ID
) {
1205 sample
->id
= *array
;
1209 if (type
& PERF_SAMPLE_TIME
) {
1210 sample
->time
= *array
;
1214 if (type
& PERF_SAMPLE_TID
) {
1217 /* undo swap of u64, then swap on individual u32s */
1218 u
.val64
= bswap_64(u
.val64
);
1219 u
.val32
[0] = bswap_32(u
.val32
[0]);
1220 u
.val32
[1] = bswap_32(u
.val32
[1]);
1223 sample
->pid
= u
.val32
[0];
1224 sample
->tid
= u
.val32
[1];
1231 static inline bool overflow(const void *endp
, u16 max_size
, const void *offset
,
1234 return size
> max_size
|| offset
+ size
> endp
;
1237 #define OVERFLOW_CHECK(offset, size, max_size) \
1239 if (overflow(endp, (max_size), (offset), (size))) \
1243 #define OVERFLOW_CHECK_u64(offset) \
1244 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1246 int perf_evsel__parse_sample(struct perf_evsel
*evsel
, union perf_event
*event
,
1247 struct perf_sample
*data
)
1249 u64 type
= evsel
->attr
.sample_type
;
1250 bool swapped
= evsel
->needs_swap
;
1252 u16 max_size
= event
->header
.size
;
1253 const void *endp
= (void *)event
+ max_size
;
1257 * used for cross-endian analysis. See git commit 65014ab3
1258 * for why this goofiness is needed.
1262 memset(data
, 0, sizeof(*data
));
1263 data
->cpu
= data
->pid
= data
->tid
= -1;
1264 data
->stream_id
= data
->id
= data
->time
= -1ULL;
1268 if (event
->header
.type
!= PERF_RECORD_SAMPLE
) {
1269 if (!evsel
->attr
.sample_id_all
)
1271 return perf_evsel__parse_id_sample(evsel
, event
, data
);
1274 array
= event
->sample
.array
;
1277 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1278 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1279 * check the format does not go past the end of the event.
1281 if (evsel
->sample_size
+ sizeof(event
->header
) > event
->header
.size
)
1285 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1290 if (type
& PERF_SAMPLE_IP
) {
1295 if (type
& PERF_SAMPLE_TID
) {
1298 /* undo swap of u64, then swap on individual u32s */
1299 u
.val64
= bswap_64(u
.val64
);
1300 u
.val32
[0] = bswap_32(u
.val32
[0]);
1301 u
.val32
[1] = bswap_32(u
.val32
[1]);
1304 data
->pid
= u
.val32
[0];
1305 data
->tid
= u
.val32
[1];
1309 if (type
& PERF_SAMPLE_TIME
) {
1310 data
->time
= *array
;
1315 if (type
& PERF_SAMPLE_ADDR
) {
1316 data
->addr
= *array
;
1320 if (type
& PERF_SAMPLE_ID
) {
1325 if (type
& PERF_SAMPLE_STREAM_ID
) {
1326 data
->stream_id
= *array
;
1330 if (type
& PERF_SAMPLE_CPU
) {
1334 /* undo swap of u64, then swap on individual u32s */
1335 u
.val64
= bswap_64(u
.val64
);
1336 u
.val32
[0] = bswap_32(u
.val32
[0]);
1339 data
->cpu
= u
.val32
[0];
1343 if (type
& PERF_SAMPLE_PERIOD
) {
1344 data
->period
= *array
;
1348 if (type
& PERF_SAMPLE_READ
) {
1349 u64 read_format
= evsel
->attr
.read_format
;
1351 OVERFLOW_CHECK_u64(array
);
1352 if (read_format
& PERF_FORMAT_GROUP
)
1353 data
->read
.group
.nr
= *array
;
1355 data
->read
.one
.value
= *array
;
1359 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1360 OVERFLOW_CHECK_u64(array
);
1361 data
->read
.time_enabled
= *array
;
1365 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1366 OVERFLOW_CHECK_u64(array
);
1367 data
->read
.time_running
= *array
;
1371 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1372 if (read_format
& PERF_FORMAT_GROUP
) {
1373 const u64 max_group_nr
= UINT64_MAX
/
1374 sizeof(struct sample_read_value
);
1376 if (data
->read
.group
.nr
> max_group_nr
)
1378 sz
= data
->read
.group
.nr
*
1379 sizeof(struct sample_read_value
);
1380 OVERFLOW_CHECK(array
, sz
, max_size
);
1381 data
->read
.group
.values
=
1382 (struct sample_read_value
*)array
;
1383 array
= (void *)array
+ sz
;
1385 OVERFLOW_CHECK_u64(array
);
1386 data
->read
.one
.id
= *array
;
1391 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1392 const u64 max_callchain_nr
= UINT64_MAX
/ sizeof(u64
);
1394 OVERFLOW_CHECK_u64(array
);
1395 data
->callchain
= (struct ip_callchain
*)array
++;
1396 if (data
->callchain
->nr
> max_callchain_nr
)
1398 sz
= data
->callchain
->nr
* sizeof(u64
);
1399 OVERFLOW_CHECK(array
, sz
, max_size
);
1400 array
= (void *)array
+ sz
;
1403 if (type
& PERF_SAMPLE_RAW
) {
1404 OVERFLOW_CHECK_u64(array
);
1406 if (WARN_ONCE(swapped
,
1407 "Endianness of raw data not corrected!\n")) {
1408 /* undo swap of u64, then swap on individual u32s */
1409 u
.val64
= bswap_64(u
.val64
);
1410 u
.val32
[0] = bswap_32(u
.val32
[0]);
1411 u
.val32
[1] = bswap_32(u
.val32
[1]);
1413 data
->raw_size
= u
.val32
[0];
1414 array
= (void *)array
+ sizeof(u32
);
1416 OVERFLOW_CHECK(array
, data
->raw_size
, max_size
);
1417 data
->raw_data
= (void *)array
;
1418 array
= (void *)array
+ data
->raw_size
;
1421 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1422 const u64 max_branch_nr
= UINT64_MAX
/
1423 sizeof(struct branch_entry
);
1425 OVERFLOW_CHECK_u64(array
);
1426 data
->branch_stack
= (struct branch_stack
*)array
++;
1428 if (data
->branch_stack
->nr
> max_branch_nr
)
1430 sz
= data
->branch_stack
->nr
* sizeof(struct branch_entry
);
1431 OVERFLOW_CHECK(array
, sz
, max_size
);
1432 array
= (void *)array
+ sz
;
1435 if (type
& PERF_SAMPLE_REGS_USER
) {
1436 OVERFLOW_CHECK_u64(array
);
1437 data
->user_regs
.abi
= *array
;
1440 if (data
->user_regs
.abi
) {
1441 u64 regs_user
= evsel
->attr
.sample_regs_user
;
1443 sz
= hweight_long(regs_user
) * sizeof(u64
);
1444 OVERFLOW_CHECK(array
, sz
, max_size
);
1445 data
->user_regs
.regs
= (u64
*)array
;
1446 array
= (void *)array
+ sz
;
1450 if (type
& PERF_SAMPLE_STACK_USER
) {
1451 OVERFLOW_CHECK_u64(array
);
1454 data
->user_stack
.offset
= ((char *)(array
- 1)
1458 data
->user_stack
.size
= 0;
1460 OVERFLOW_CHECK(array
, sz
, max_size
);
1461 data
->user_stack
.data
= (char *)array
;
1462 array
= (void *)array
+ sz
;
1463 OVERFLOW_CHECK_u64(array
);
1464 data
->user_stack
.size
= *array
++;
1465 if (WARN_ONCE(data
->user_stack
.size
> sz
,
1466 "user stack dump failure\n"))
1472 if (type
& PERF_SAMPLE_WEIGHT
) {
1473 OVERFLOW_CHECK_u64(array
);
1474 data
->weight
= *array
;
1478 data
->data_src
= PERF_MEM_DATA_SRC_NONE
;
1479 if (type
& PERF_SAMPLE_DATA_SRC
) {
1480 OVERFLOW_CHECK_u64(array
);
1481 data
->data_src
= *array
;
1485 data
->transaction
= 0;
1486 if (type
& PERF_SAMPLE_TRANSACTION
) {
1487 OVERFLOW_CHECK_u64(array
);
1488 data
->transaction
= *array
;
1495 size_t perf_event__sample_event_size(const struct perf_sample
*sample
, u64 type
,
1496 u64 sample_regs_user
, u64 read_format
)
1498 size_t sz
, result
= sizeof(struct sample_event
);
1500 if (type
& PERF_SAMPLE_IDENTIFIER
)
1501 result
+= sizeof(u64
);
1503 if (type
& PERF_SAMPLE_IP
)
1504 result
+= sizeof(u64
);
1506 if (type
& PERF_SAMPLE_TID
)
1507 result
+= sizeof(u64
);
1509 if (type
& PERF_SAMPLE_TIME
)
1510 result
+= sizeof(u64
);
1512 if (type
& PERF_SAMPLE_ADDR
)
1513 result
+= sizeof(u64
);
1515 if (type
& PERF_SAMPLE_ID
)
1516 result
+= sizeof(u64
);
1518 if (type
& PERF_SAMPLE_STREAM_ID
)
1519 result
+= sizeof(u64
);
1521 if (type
& PERF_SAMPLE_CPU
)
1522 result
+= sizeof(u64
);
1524 if (type
& PERF_SAMPLE_PERIOD
)
1525 result
+= sizeof(u64
);
1527 if (type
& PERF_SAMPLE_READ
) {
1528 result
+= sizeof(u64
);
1529 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1530 result
+= sizeof(u64
);
1531 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1532 result
+= sizeof(u64
);
1533 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1534 if (read_format
& PERF_FORMAT_GROUP
) {
1535 sz
= sample
->read
.group
.nr
*
1536 sizeof(struct sample_read_value
);
1539 result
+= sizeof(u64
);
1543 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1544 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1548 if (type
& PERF_SAMPLE_RAW
) {
1549 result
+= sizeof(u32
);
1550 result
+= sample
->raw_size
;
1553 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1554 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1559 if (type
& PERF_SAMPLE_REGS_USER
) {
1560 if (sample
->user_regs
.abi
) {
1561 result
+= sizeof(u64
);
1562 sz
= hweight_long(sample_regs_user
) * sizeof(u64
);
1565 result
+= sizeof(u64
);
1569 if (type
& PERF_SAMPLE_STACK_USER
) {
1570 sz
= sample
->user_stack
.size
;
1571 result
+= sizeof(u64
);
1574 result
+= sizeof(u64
);
1578 if (type
& PERF_SAMPLE_WEIGHT
)
1579 result
+= sizeof(u64
);
1581 if (type
& PERF_SAMPLE_DATA_SRC
)
1582 result
+= sizeof(u64
);
1584 if (type
& PERF_SAMPLE_TRANSACTION
)
1585 result
+= sizeof(u64
);
1590 int perf_event__synthesize_sample(union perf_event
*event
, u64 type
,
1591 u64 sample_regs_user
, u64 read_format
,
1592 const struct perf_sample
*sample
,
1598 * used for cross-endian analysis. See git commit 65014ab3
1599 * for why this goofiness is needed.
1603 array
= event
->sample
.array
;
1605 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1606 *array
= sample
->id
;
1610 if (type
& PERF_SAMPLE_IP
) {
1611 *array
= sample
->ip
;
1615 if (type
& PERF_SAMPLE_TID
) {
1616 u
.val32
[0] = sample
->pid
;
1617 u
.val32
[1] = sample
->tid
;
1620 * Inverse of what is done in perf_evsel__parse_sample
1622 u
.val32
[0] = bswap_32(u
.val32
[0]);
1623 u
.val32
[1] = bswap_32(u
.val32
[1]);
1624 u
.val64
= bswap_64(u
.val64
);
1631 if (type
& PERF_SAMPLE_TIME
) {
1632 *array
= sample
->time
;
1636 if (type
& PERF_SAMPLE_ADDR
) {
1637 *array
= sample
->addr
;
1641 if (type
& PERF_SAMPLE_ID
) {
1642 *array
= sample
->id
;
1646 if (type
& PERF_SAMPLE_STREAM_ID
) {
1647 *array
= sample
->stream_id
;
1651 if (type
& PERF_SAMPLE_CPU
) {
1652 u
.val32
[0] = sample
->cpu
;
1655 * Inverse of what is done in perf_evsel__parse_sample
1657 u
.val32
[0] = bswap_32(u
.val32
[0]);
1658 u
.val64
= bswap_64(u
.val64
);
1664 if (type
& PERF_SAMPLE_PERIOD
) {
1665 *array
= sample
->period
;
1669 if (type
& PERF_SAMPLE_READ
) {
1670 if (read_format
& PERF_FORMAT_GROUP
)
1671 *array
= sample
->read
.group
.nr
;
1673 *array
= sample
->read
.one
.value
;
1676 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1677 *array
= sample
->read
.time_enabled
;
1681 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1682 *array
= sample
->read
.time_running
;
1686 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1687 if (read_format
& PERF_FORMAT_GROUP
) {
1688 sz
= sample
->read
.group
.nr
*
1689 sizeof(struct sample_read_value
);
1690 memcpy(array
, sample
->read
.group
.values
, sz
);
1691 array
= (void *)array
+ sz
;
1693 *array
= sample
->read
.one
.id
;
1698 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1699 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1700 memcpy(array
, sample
->callchain
, sz
);
1701 array
= (void *)array
+ sz
;
1704 if (type
& PERF_SAMPLE_RAW
) {
1705 u
.val32
[0] = sample
->raw_size
;
1706 if (WARN_ONCE(swapped
,
1707 "Endianness of raw data not corrected!\n")) {
1709 * Inverse of what is done in perf_evsel__parse_sample
1711 u
.val32
[0] = bswap_32(u
.val32
[0]);
1712 u
.val32
[1] = bswap_32(u
.val32
[1]);
1713 u
.val64
= bswap_64(u
.val64
);
1716 array
= (void *)array
+ sizeof(u32
);
1718 memcpy(array
, sample
->raw_data
, sample
->raw_size
);
1719 array
= (void *)array
+ sample
->raw_size
;
1722 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1723 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1725 memcpy(array
, sample
->branch_stack
, sz
);
1726 array
= (void *)array
+ sz
;
1729 if (type
& PERF_SAMPLE_REGS_USER
) {
1730 if (sample
->user_regs
.abi
) {
1731 *array
++ = sample
->user_regs
.abi
;
1732 sz
= hweight_long(sample_regs_user
) * sizeof(u64
);
1733 memcpy(array
, sample
->user_regs
.regs
, sz
);
1734 array
= (void *)array
+ sz
;
1740 if (type
& PERF_SAMPLE_STACK_USER
) {
1741 sz
= sample
->user_stack
.size
;
1744 memcpy(array
, sample
->user_stack
.data
, sz
);
1745 array
= (void *)array
+ sz
;
1750 if (type
& PERF_SAMPLE_WEIGHT
) {
1751 *array
= sample
->weight
;
1755 if (type
& PERF_SAMPLE_DATA_SRC
) {
1756 *array
= sample
->data_src
;
1760 if (type
& PERF_SAMPLE_TRANSACTION
) {
1761 *array
= sample
->transaction
;
1768 struct format_field
*perf_evsel__field(struct perf_evsel
*evsel
, const char *name
)
1770 return pevent_find_field(evsel
->tp_format
, name
);
1773 void *perf_evsel__rawptr(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
1776 struct format_field
*field
= perf_evsel__field(evsel
, name
);
1782 offset
= field
->offset
;
1784 if (field
->flags
& FIELD_IS_DYNAMIC
) {
1785 offset
= *(int *)(sample
->raw_data
+ field
->offset
);
1789 return sample
->raw_data
+ offset
;
1792 u64
perf_evsel__intval(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
1795 struct format_field
*field
= perf_evsel__field(evsel
, name
);
1802 ptr
= sample
->raw_data
+ field
->offset
;
1804 switch (field
->size
) {
1808 value
= *(u16
*)ptr
;
1811 value
= *(u32
*)ptr
;
1814 value
= *(u64
*)ptr
;
1820 if (!evsel
->needs_swap
)
1823 switch (field
->size
) {
1825 return bswap_16(value
);
1827 return bswap_32(value
);
1829 return bswap_64(value
);
1837 static int comma_fprintf(FILE *fp
, bool *first
, const char *fmt
, ...)
1843 ret
+= fprintf(fp
, ",");
1845 ret
+= fprintf(fp
, ":");
1849 va_start(args
, fmt
);
1850 ret
+= vfprintf(fp
, fmt
, args
);
1855 static int __if_fprintf(FILE *fp
, bool *first
, const char *field
, u64 value
)
1860 return comma_fprintf(fp
, first
, " %s: %" PRIu64
, field
, value
);
1863 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1870 static int bits__fprintf(FILE *fp
, const char *field
, u64 value
,
1871 struct bit_names
*bits
, bool *first
)
1873 int i
= 0, printed
= comma_fprintf(fp
, first
, " %s: ", field
);
1874 bool first_bit
= true;
1877 if (value
& bits
[i
].bit
) {
1878 printed
+= fprintf(fp
, "%s%s", first_bit
? "" : "|", bits
[i
].name
);
1881 } while (bits
[++i
].name
!= NULL
);
1886 static int sample_type__fprintf(FILE *fp
, bool *first
, u64 value
)
1888 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1889 struct bit_names bits
[] = {
1890 bit_name(IP
), bit_name(TID
), bit_name(TIME
), bit_name(ADDR
),
1891 bit_name(READ
), bit_name(CALLCHAIN
), bit_name(ID
), bit_name(CPU
),
1892 bit_name(PERIOD
), bit_name(STREAM_ID
), bit_name(RAW
),
1893 bit_name(BRANCH_STACK
), bit_name(REGS_USER
), bit_name(STACK_USER
),
1894 bit_name(IDENTIFIER
),
1898 return bits__fprintf(fp
, "sample_type", value
, bits
, first
);
1901 static int read_format__fprintf(FILE *fp
, bool *first
, u64 value
)
1903 #define bit_name(n) { PERF_FORMAT_##n, #n }
1904 struct bit_names bits
[] = {
1905 bit_name(TOTAL_TIME_ENABLED
), bit_name(TOTAL_TIME_RUNNING
),
1906 bit_name(ID
), bit_name(GROUP
),
1910 return bits__fprintf(fp
, "read_format", value
, bits
, first
);
1913 int perf_evsel__fprintf(struct perf_evsel
*evsel
,
1914 struct perf_attr_details
*details
, FILE *fp
)
1919 if (details
->event_group
) {
1920 struct perf_evsel
*pos
;
1922 if (!perf_evsel__is_group_leader(evsel
))
1925 if (evsel
->nr_members
> 1)
1926 printed
+= fprintf(fp
, "%s{", evsel
->group_name
?: "");
1928 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
1929 for_each_group_member(pos
, evsel
)
1930 printed
+= fprintf(fp
, ",%s", perf_evsel__name(pos
));
1932 if (evsel
->nr_members
> 1)
1933 printed
+= fprintf(fp
, "}");
1937 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
1939 if (details
->verbose
|| details
->freq
) {
1940 printed
+= comma_fprintf(fp
, &first
, " sample_freq=%" PRIu64
,
1941 (u64
)evsel
->attr
.sample_freq
);
1944 if (details
->verbose
) {
1950 printed
+= sample_type__fprintf(fp
, &first
, evsel
->attr
.sample_type
);
1951 if (evsel
->attr
.read_format
)
1952 printed
+= read_format__fprintf(fp
, &first
, evsel
->attr
.read_format
);
1956 if_print(exclusive
);
1957 if_print(exclude_user
);
1958 if_print(exclude_kernel
);
1959 if_print(exclude_hv
);
1960 if_print(exclude_idle
);
1965 if_print(inherit_stat
);
1966 if_print(enable_on_exec
);
1968 if_print(watermark
);
1969 if_print(precise_ip
);
1970 if_print(mmap_data
);
1971 if_print(sample_id_all
);
1972 if_print(exclude_host
);
1973 if_print(exclude_guest
);
1974 if_print(__reserved_1
);
1975 if_print(wakeup_events
);
1977 if_print(branch_sample_type
);
1984 bool perf_evsel__fallback(struct perf_evsel
*evsel
, int err
,
1985 char *msg
, size_t msgsize
)
1987 if ((err
== ENOENT
|| err
== ENXIO
|| err
== ENODEV
) &&
1988 evsel
->attr
.type
== PERF_TYPE_HARDWARE
&&
1989 evsel
->attr
.config
== PERF_COUNT_HW_CPU_CYCLES
) {
1991 * If it's cycles then fall back to hrtimer based
1992 * cpu-clock-tick sw counter, which is always available even if
1995 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1998 scnprintf(msg
, msgsize
, "%s",
1999 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2001 evsel
->attr
.type
= PERF_TYPE_SOFTWARE
;
2002 evsel
->attr
.config
= PERF_COUNT_SW_CPU_CLOCK
;
2012 int perf_evsel__open_strerror(struct perf_evsel
*evsel
, struct target
*target
,
2013 int err
, char *msg
, size_t size
)
2018 return scnprintf(msg
, size
,
2019 "You may not have permission to collect %sstats.\n"
2020 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2021 " -1 - Not paranoid at all\n"
2022 " 0 - Disallow raw tracepoint access for unpriv\n"
2023 " 1 - Disallow cpu events for unpriv\n"
2024 " 2 - Disallow kernel profiling for unpriv",
2025 target
->system_wide
? "system-wide " : "");
2027 return scnprintf(msg
, size
, "The %s event is not supported.",
2028 perf_evsel__name(evsel
));
2030 return scnprintf(msg
, size
, "%s",
2031 "Too many events are opened.\n"
2032 "Try again after reducing the number of events.");
2034 if (target
->cpu_list
)
2035 return scnprintf(msg
, size
, "%s",
2036 "No such device - did you specify an out-of-range profile CPU?\n");
2039 if (evsel
->attr
.precise_ip
)
2040 return scnprintf(msg
, size
, "%s",
2041 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2042 #if defined(__i386__) || defined(__x86_64__)
2043 if (evsel
->attr
.type
== PERF_TYPE_HARDWARE
)
2044 return scnprintf(msg
, size
, "%s",
2045 "No hardware sampling interrupt available.\n"
2046 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2053 return scnprintf(msg
, size
,
2054 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2055 "/bin/dmesg may provide additional information.\n"
2056 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2057 err
, strerror(err
), perf_evsel__name(evsel
));