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 <api/fs/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>
18 #include "callchain.h"
24 #include "thread_map.h"
26 #include "perf_regs.h"
28 #include "trace-event.h"
38 } perf_missing_features
;
40 static clockid_t clockid
;
42 static int perf_evsel__no_extra_init(struct perf_evsel
*evsel __maybe_unused
)
47 static void perf_evsel__no_extra_fini(struct perf_evsel
*evsel __maybe_unused
)
53 int (*init
)(struct perf_evsel
*evsel
);
54 void (*fini
)(struct perf_evsel
*evsel
);
55 } perf_evsel__object
= {
56 .size
= sizeof(struct perf_evsel
),
57 .init
= perf_evsel__no_extra_init
,
58 .fini
= perf_evsel__no_extra_fini
,
61 int perf_evsel__object_config(size_t object_size
,
62 int (*init
)(struct perf_evsel
*evsel
),
63 void (*fini
)(struct perf_evsel
*evsel
))
69 if (perf_evsel__object
.size
> object_size
)
72 perf_evsel__object
.size
= object_size
;
76 perf_evsel__object
.init
= init
;
79 perf_evsel__object
.fini
= fini
;
84 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
86 int __perf_evsel__sample_size(u64 sample_type
)
88 u64 mask
= sample_type
& PERF_SAMPLE_MASK
;
92 for (i
= 0; i
< 64; i
++) {
93 if (mask
& (1ULL << i
))
103 * __perf_evsel__calc_id_pos - calculate id_pos.
104 * @sample_type: sample type
106 * This function returns the position of the event id (PERF_SAMPLE_ID or
107 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
110 static int __perf_evsel__calc_id_pos(u64 sample_type
)
114 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
117 if (!(sample_type
& PERF_SAMPLE_ID
))
120 if (sample_type
& PERF_SAMPLE_IP
)
123 if (sample_type
& PERF_SAMPLE_TID
)
126 if (sample_type
& PERF_SAMPLE_TIME
)
129 if (sample_type
& PERF_SAMPLE_ADDR
)
136 * __perf_evsel__calc_is_pos - calculate is_pos.
137 * @sample_type: sample type
139 * This function returns the position (counting backwards) of the event id
140 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
141 * sample_id_all is used there is an id sample appended to non-sample events.
143 static int __perf_evsel__calc_is_pos(u64 sample_type
)
147 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
150 if (!(sample_type
& PERF_SAMPLE_ID
))
153 if (sample_type
& PERF_SAMPLE_CPU
)
156 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
162 void perf_evsel__calc_id_pos(struct perf_evsel
*evsel
)
164 evsel
->id_pos
= __perf_evsel__calc_id_pos(evsel
->attr
.sample_type
);
165 evsel
->is_pos
= __perf_evsel__calc_is_pos(evsel
->attr
.sample_type
);
168 void __perf_evsel__set_sample_bit(struct perf_evsel
*evsel
,
169 enum perf_event_sample_format bit
)
171 if (!(evsel
->attr
.sample_type
& bit
)) {
172 evsel
->attr
.sample_type
|= bit
;
173 evsel
->sample_size
+= sizeof(u64
);
174 perf_evsel__calc_id_pos(evsel
);
178 void __perf_evsel__reset_sample_bit(struct perf_evsel
*evsel
,
179 enum perf_event_sample_format bit
)
181 if (evsel
->attr
.sample_type
& bit
) {
182 evsel
->attr
.sample_type
&= ~bit
;
183 evsel
->sample_size
-= sizeof(u64
);
184 perf_evsel__calc_id_pos(evsel
);
188 void perf_evsel__set_sample_id(struct perf_evsel
*evsel
,
189 bool can_sample_identifier
)
191 if (can_sample_identifier
) {
192 perf_evsel__reset_sample_bit(evsel
, ID
);
193 perf_evsel__set_sample_bit(evsel
, IDENTIFIER
);
195 perf_evsel__set_sample_bit(evsel
, ID
);
197 evsel
->attr
.read_format
|= PERF_FORMAT_ID
;
200 void perf_evsel__init(struct perf_evsel
*evsel
,
201 struct perf_event_attr
*attr
, int idx
)
204 evsel
->tracking
= !idx
;
206 evsel
->leader
= evsel
;
209 INIT_LIST_HEAD(&evsel
->node
);
210 INIT_LIST_HEAD(&evsel
->config_terms
);
211 perf_evsel__object
.init(evsel
);
212 evsel
->sample_size
= __perf_evsel__sample_size(attr
->sample_type
);
213 perf_evsel__calc_id_pos(evsel
);
214 evsel
->cmdline_group_boundary
= false;
217 struct perf_evsel
*perf_evsel__new_idx(struct perf_event_attr
*attr
, int idx
)
219 struct perf_evsel
*evsel
= zalloc(perf_evsel__object
.size
);
222 perf_evsel__init(evsel
, attr
, idx
);
227 struct perf_evsel
*perf_evsel__newtp_idx(const char *sys
, const char *name
, int idx
)
229 struct perf_evsel
*evsel
= zalloc(perf_evsel__object
.size
);
232 struct perf_event_attr attr
= {
233 .type
= PERF_TYPE_TRACEPOINT
,
234 .sample_type
= (PERF_SAMPLE_RAW
| PERF_SAMPLE_TIME
|
235 PERF_SAMPLE_CPU
| PERF_SAMPLE_PERIOD
),
238 if (asprintf(&evsel
->name
, "%s:%s", sys
, name
) < 0)
241 evsel
->tp_format
= trace_event__tp_format(sys
, name
);
242 if (evsel
->tp_format
== NULL
)
245 event_attr_init(&attr
);
246 attr
.config
= evsel
->tp_format
->id
;
247 attr
.sample_period
= 1;
248 perf_evsel__init(evsel
, &attr
, idx
);
259 const char *perf_evsel__hw_names
[PERF_COUNT_HW_MAX
] = {
267 "stalled-cycles-frontend",
268 "stalled-cycles-backend",
272 static const char *__perf_evsel__hw_name(u64 config
)
274 if (config
< PERF_COUNT_HW_MAX
&& perf_evsel__hw_names
[config
])
275 return perf_evsel__hw_names
[config
];
277 return "unknown-hardware";
280 static int perf_evsel__add_modifiers(struct perf_evsel
*evsel
, char *bf
, size_t size
)
282 int colon
= 0, r
= 0;
283 struct perf_event_attr
*attr
= &evsel
->attr
;
284 bool exclude_guest_default
= false;
286 #define MOD_PRINT(context, mod) do { \
287 if (!attr->exclude_##context) { \
288 if (!colon) colon = ++r; \
289 r += scnprintf(bf + r, size - r, "%c", mod); \
292 if (attr
->exclude_kernel
|| attr
->exclude_user
|| attr
->exclude_hv
) {
293 MOD_PRINT(kernel
, 'k');
294 MOD_PRINT(user
, 'u');
296 exclude_guest_default
= true;
299 if (attr
->precise_ip
) {
302 r
+= scnprintf(bf
+ r
, size
- r
, "%.*s", attr
->precise_ip
, "ppp");
303 exclude_guest_default
= true;
306 if (attr
->exclude_host
|| attr
->exclude_guest
== exclude_guest_default
) {
307 MOD_PRINT(host
, 'H');
308 MOD_PRINT(guest
, 'G');
316 static int perf_evsel__hw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
318 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__hw_name(evsel
->attr
.config
));
319 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
322 const char *perf_evsel__sw_names
[PERF_COUNT_SW_MAX
] = {
335 static const char *__perf_evsel__sw_name(u64 config
)
337 if (config
< PERF_COUNT_SW_MAX
&& perf_evsel__sw_names
[config
])
338 return perf_evsel__sw_names
[config
];
339 return "unknown-software";
342 static int perf_evsel__sw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
344 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__sw_name(evsel
->attr
.config
));
345 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
348 static int __perf_evsel__bp_name(char *bf
, size_t size
, u64 addr
, u64 type
)
352 r
= scnprintf(bf
, size
, "mem:0x%" PRIx64
":", addr
);
354 if (type
& HW_BREAKPOINT_R
)
355 r
+= scnprintf(bf
+ r
, size
- r
, "r");
357 if (type
& HW_BREAKPOINT_W
)
358 r
+= scnprintf(bf
+ r
, size
- r
, "w");
360 if (type
& HW_BREAKPOINT_X
)
361 r
+= scnprintf(bf
+ r
, size
- r
, "x");
366 static int perf_evsel__bp_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
368 struct perf_event_attr
*attr
= &evsel
->attr
;
369 int r
= __perf_evsel__bp_name(bf
, size
, attr
->bp_addr
, attr
->bp_type
);
370 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
373 const char *perf_evsel__hw_cache
[PERF_COUNT_HW_CACHE_MAX
]
374 [PERF_EVSEL__MAX_ALIASES
] = {
375 { "L1-dcache", "l1-d", "l1d", "L1-data", },
376 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
378 { "dTLB", "d-tlb", "Data-TLB", },
379 { "iTLB", "i-tlb", "Instruction-TLB", },
380 { "branch", "branches", "bpu", "btb", "bpc", },
384 const char *perf_evsel__hw_cache_op
[PERF_COUNT_HW_CACHE_OP_MAX
]
385 [PERF_EVSEL__MAX_ALIASES
] = {
386 { "load", "loads", "read", },
387 { "store", "stores", "write", },
388 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
391 const char *perf_evsel__hw_cache_result
[PERF_COUNT_HW_CACHE_RESULT_MAX
]
392 [PERF_EVSEL__MAX_ALIASES
] = {
393 { "refs", "Reference", "ops", "access", },
394 { "misses", "miss", },
397 #define C(x) PERF_COUNT_HW_CACHE_##x
398 #define CACHE_READ (1 << C(OP_READ))
399 #define CACHE_WRITE (1 << C(OP_WRITE))
400 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
401 #define COP(x) (1 << x)
404 * cache operartion stat
405 * L1I : Read and prefetch only
406 * ITLB and BPU : Read-only
408 static unsigned long perf_evsel__hw_cache_stat
[C(MAX
)] = {
409 [C(L1D
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
410 [C(L1I
)] = (CACHE_READ
| CACHE_PREFETCH
),
411 [C(LL
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
412 [C(DTLB
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
413 [C(ITLB
)] = (CACHE_READ
),
414 [C(BPU
)] = (CACHE_READ
),
415 [C(NODE
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
418 bool perf_evsel__is_cache_op_valid(u8 type
, u8 op
)
420 if (perf_evsel__hw_cache_stat
[type
] & COP(op
))
421 return true; /* valid */
423 return false; /* invalid */
426 int __perf_evsel__hw_cache_type_op_res_name(u8 type
, u8 op
, u8 result
,
427 char *bf
, size_t size
)
430 return scnprintf(bf
, size
, "%s-%s-%s", perf_evsel__hw_cache
[type
][0],
431 perf_evsel__hw_cache_op
[op
][0],
432 perf_evsel__hw_cache_result
[result
][0]);
435 return scnprintf(bf
, size
, "%s-%s", perf_evsel__hw_cache
[type
][0],
436 perf_evsel__hw_cache_op
[op
][1]);
439 static int __perf_evsel__hw_cache_name(u64 config
, char *bf
, size_t size
)
441 u8 op
, result
, type
= (config
>> 0) & 0xff;
442 const char *err
= "unknown-ext-hardware-cache-type";
444 if (type
> PERF_COUNT_HW_CACHE_MAX
)
447 op
= (config
>> 8) & 0xff;
448 err
= "unknown-ext-hardware-cache-op";
449 if (op
> PERF_COUNT_HW_CACHE_OP_MAX
)
452 result
= (config
>> 16) & 0xff;
453 err
= "unknown-ext-hardware-cache-result";
454 if (result
> PERF_COUNT_HW_CACHE_RESULT_MAX
)
457 err
= "invalid-cache";
458 if (!perf_evsel__is_cache_op_valid(type
, op
))
461 return __perf_evsel__hw_cache_type_op_res_name(type
, op
, result
, bf
, size
);
463 return scnprintf(bf
, size
, "%s", err
);
466 static int perf_evsel__hw_cache_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
468 int ret
= __perf_evsel__hw_cache_name(evsel
->attr
.config
, bf
, size
);
469 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
472 static int perf_evsel__raw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
474 int ret
= scnprintf(bf
, size
, "raw 0x%" PRIx64
, evsel
->attr
.config
);
475 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
478 const char *perf_evsel__name(struct perf_evsel
*evsel
)
485 switch (evsel
->attr
.type
) {
487 perf_evsel__raw_name(evsel
, bf
, sizeof(bf
));
490 case PERF_TYPE_HARDWARE
:
491 perf_evsel__hw_name(evsel
, bf
, sizeof(bf
));
494 case PERF_TYPE_HW_CACHE
:
495 perf_evsel__hw_cache_name(evsel
, bf
, sizeof(bf
));
498 case PERF_TYPE_SOFTWARE
:
499 perf_evsel__sw_name(evsel
, bf
, sizeof(bf
));
502 case PERF_TYPE_TRACEPOINT
:
503 scnprintf(bf
, sizeof(bf
), "%s", "unknown tracepoint");
506 case PERF_TYPE_BREAKPOINT
:
507 perf_evsel__bp_name(evsel
, bf
, sizeof(bf
));
511 scnprintf(bf
, sizeof(bf
), "unknown attr type: %d",
516 evsel
->name
= strdup(bf
);
518 return evsel
->name
?: "unknown";
521 const char *perf_evsel__group_name(struct perf_evsel
*evsel
)
523 return evsel
->group_name
?: "anon group";
526 int perf_evsel__group_desc(struct perf_evsel
*evsel
, char *buf
, size_t size
)
529 struct perf_evsel
*pos
;
530 const char *group_name
= perf_evsel__group_name(evsel
);
532 ret
= scnprintf(buf
, size
, "%s", group_name
);
534 ret
+= scnprintf(buf
+ ret
, size
- ret
, " { %s",
535 perf_evsel__name(evsel
));
537 for_each_group_member(pos
, evsel
)
538 ret
+= scnprintf(buf
+ ret
, size
- ret
, ", %s",
539 perf_evsel__name(pos
));
541 ret
+= scnprintf(buf
+ ret
, size
- ret
, " }");
547 perf_evsel__config_callgraph(struct perf_evsel
*evsel
,
548 struct record_opts
*opts
)
550 bool function
= perf_evsel__is_function_event(evsel
);
551 struct perf_event_attr
*attr
= &evsel
->attr
;
553 perf_evsel__set_sample_bit(evsel
, CALLCHAIN
);
555 if (callchain_param
.record_mode
== CALLCHAIN_LBR
) {
556 if (!opts
->branch_stack
) {
557 if (attr
->exclude_user
) {
558 pr_warning("LBR callstack option is only available "
559 "to get user callchain information. "
560 "Falling back to framepointers.\n");
562 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
563 attr
->branch_sample_type
= PERF_SAMPLE_BRANCH_USER
|
564 PERF_SAMPLE_BRANCH_CALL_STACK
;
567 pr_warning("Cannot use LBR callstack with branch stack. "
568 "Falling back to framepointers.\n");
571 if (callchain_param
.record_mode
== CALLCHAIN_DWARF
) {
573 perf_evsel__set_sample_bit(evsel
, REGS_USER
);
574 perf_evsel__set_sample_bit(evsel
, STACK_USER
);
575 attr
->sample_regs_user
= PERF_REGS_MASK
;
576 attr
->sample_stack_user
= callchain_param
.dump_size
;
577 attr
->exclude_callchain_user
= 1;
579 pr_info("Cannot use DWARF unwind for function trace event,"
580 " falling back to framepointers.\n");
585 pr_info("Disabling user space callchains for function trace event.\n");
586 attr
->exclude_callchain_user
= 1;
590 static void apply_config_terms(struct perf_event_attr
*attr __maybe_unused
,
591 struct list_head
*config_terms
)
593 struct perf_evsel_config_term
*term
;
595 list_for_each_entry(term
, config_terms
, list
) {
596 switch (term
->type
) {
597 case PERF_EVSEL__CONFIG_TERM_PERIOD
:
598 attr
->sample_period
= term
->val
.period
;
606 * The enable_on_exec/disabled value strategy:
608 * 1) For any type of traced program:
609 * - all independent events and group leaders are disabled
610 * - all group members are enabled
612 * Group members are ruled by group leaders. They need to
613 * be enabled, because the group scheduling relies on that.
615 * 2) For traced programs executed by perf:
616 * - all independent events and group leaders have
618 * - we don't specifically enable or disable any event during
621 * Independent events and group leaders are initially disabled
622 * and get enabled by exec. Group members are ruled by group
623 * leaders as stated in 1).
625 * 3) For traced programs attached by perf (pid/tid):
626 * - we specifically enable or disable all events during
629 * When attaching events to already running traced we
630 * enable/disable events specifically, as there's no
631 * initial traced exec call.
633 void perf_evsel__config(struct perf_evsel
*evsel
, struct record_opts
*opts
)
635 struct perf_evsel
*leader
= evsel
->leader
;
636 struct perf_event_attr
*attr
= &evsel
->attr
;
637 int track
= evsel
->tracking
;
638 bool per_cpu
= opts
->target
.default_per_cpu
&& !opts
->target
.per_thread
;
640 attr
->sample_id_all
= perf_missing_features
.sample_id_all
? 0 : 1;
641 attr
->inherit
= !opts
->no_inherit
;
643 perf_evsel__set_sample_bit(evsel
, IP
);
644 perf_evsel__set_sample_bit(evsel
, TID
);
646 if (evsel
->sample_read
) {
647 perf_evsel__set_sample_bit(evsel
, READ
);
650 * We need ID even in case of single event, because
651 * PERF_SAMPLE_READ process ID specific data.
653 perf_evsel__set_sample_id(evsel
, false);
656 * Apply group format only if we belong to group
657 * with more than one members.
659 if (leader
->nr_members
> 1) {
660 attr
->read_format
|= PERF_FORMAT_GROUP
;
666 * We default some events to have a default interval. But keep
667 * it a weak assumption overridable by the user.
669 if (!attr
->sample_period
|| (opts
->user_freq
!= UINT_MAX
||
670 opts
->user_interval
!= ULLONG_MAX
)) {
672 perf_evsel__set_sample_bit(evsel
, PERIOD
);
674 attr
->sample_freq
= opts
->freq
;
676 attr
->sample_period
= opts
->default_interval
;
681 * Disable sampling for all group members other
682 * than leader in case leader 'leads' the sampling.
684 if ((leader
!= evsel
) && leader
->sample_read
) {
685 attr
->sample_freq
= 0;
686 attr
->sample_period
= 0;
689 if (opts
->no_samples
)
690 attr
->sample_freq
= 0;
692 if (opts
->inherit_stat
)
693 attr
->inherit_stat
= 1;
695 if (opts
->sample_address
) {
696 perf_evsel__set_sample_bit(evsel
, ADDR
);
697 attr
->mmap_data
= track
;
701 * We don't allow user space callchains for function trace
702 * event, due to issues with page faults while tracing page
703 * fault handler and its overall trickiness nature.
705 if (perf_evsel__is_function_event(evsel
))
706 evsel
->attr
.exclude_callchain_user
= 1;
708 if (callchain_param
.enabled
&& !evsel
->no_aux_samples
)
709 perf_evsel__config_callgraph(evsel
, opts
);
711 if (opts
->sample_intr_regs
) {
712 attr
->sample_regs_intr
= PERF_REGS_MASK
;
713 perf_evsel__set_sample_bit(evsel
, REGS_INTR
);
716 if (target__has_cpu(&opts
->target
))
717 perf_evsel__set_sample_bit(evsel
, CPU
);
720 perf_evsel__set_sample_bit(evsel
, PERIOD
);
723 * When the user explicitely disabled time don't force it here.
725 if (opts
->sample_time
&&
726 (!perf_missing_features
.sample_id_all
&&
727 (!opts
->no_inherit
|| target__has_cpu(&opts
->target
) || per_cpu
||
728 opts
->sample_time_set
)))
729 perf_evsel__set_sample_bit(evsel
, TIME
);
731 if (opts
->raw_samples
&& !evsel
->no_aux_samples
) {
732 perf_evsel__set_sample_bit(evsel
, TIME
);
733 perf_evsel__set_sample_bit(evsel
, RAW
);
734 perf_evsel__set_sample_bit(evsel
, CPU
);
737 if (opts
->sample_address
)
738 perf_evsel__set_sample_bit(evsel
, DATA_SRC
);
740 if (opts
->no_buffering
) {
742 attr
->wakeup_events
= 1;
744 if (opts
->branch_stack
&& !evsel
->no_aux_samples
) {
745 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
746 attr
->branch_sample_type
= opts
->branch_stack
;
749 if (opts
->sample_weight
)
750 perf_evsel__set_sample_bit(evsel
, WEIGHT
);
754 attr
->mmap2
= track
&& !perf_missing_features
.mmap2
;
757 if (opts
->record_switch_events
)
758 attr
->context_switch
= track
;
760 if (opts
->sample_transaction
)
761 perf_evsel__set_sample_bit(evsel
, TRANSACTION
);
763 if (opts
->running_time
) {
764 evsel
->attr
.read_format
|=
765 PERF_FORMAT_TOTAL_TIME_ENABLED
|
766 PERF_FORMAT_TOTAL_TIME_RUNNING
;
770 * XXX see the function comment above
772 * Disabling only independent events or group leaders,
773 * keeping group members enabled.
775 if (perf_evsel__is_group_leader(evsel
))
779 * Setting enable_on_exec for independent events and
780 * group leaders for traced executed by perf.
782 if (target__none(&opts
->target
) && perf_evsel__is_group_leader(evsel
) &&
783 !opts
->initial_delay
)
784 attr
->enable_on_exec
= 1;
786 if (evsel
->immediate
) {
788 attr
->enable_on_exec
= 0;
791 clockid
= opts
->clockid
;
792 if (opts
->use_clockid
) {
793 attr
->use_clockid
= 1;
794 attr
->clockid
= opts
->clockid
;
798 * Apply event specific term settings,
799 * it overloads any global configuration.
801 apply_config_terms(attr
, &evsel
->config_terms
);
804 static int perf_evsel__alloc_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
808 if (evsel
->system_wide
)
811 evsel
->fd
= xyarray__new(ncpus
, nthreads
, sizeof(int));
814 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
815 for (thread
= 0; thread
< nthreads
; thread
++) {
816 FD(evsel
, cpu
, thread
) = -1;
821 return evsel
->fd
!= NULL
? 0 : -ENOMEM
;
824 static int perf_evsel__run_ioctl(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
829 if (evsel
->system_wide
)
832 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
833 for (thread
= 0; thread
< nthreads
; thread
++) {
834 int fd
= FD(evsel
, cpu
, thread
),
835 err
= ioctl(fd
, ioc
, arg
);
845 int perf_evsel__apply_filter(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
848 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
849 PERF_EVENT_IOC_SET_FILTER
,
853 int perf_evsel__set_filter(struct perf_evsel
*evsel
, const char *filter
)
855 char *new_filter
= strdup(filter
);
857 if (new_filter
!= NULL
) {
859 evsel
->filter
= new_filter
;
866 int perf_evsel__append_filter(struct perf_evsel
*evsel
,
867 const char *op
, const char *filter
)
871 if (evsel
->filter
== NULL
)
872 return perf_evsel__set_filter(evsel
, filter
);
874 if (asprintf(&new_filter
,"(%s) %s (%s)", evsel
->filter
, op
, filter
) > 0) {
876 evsel
->filter
= new_filter
;
883 int perf_evsel__enable(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
885 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
886 PERF_EVENT_IOC_ENABLE
,
890 int perf_evsel__alloc_id(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
892 if (ncpus
== 0 || nthreads
== 0)
895 if (evsel
->system_wide
)
898 evsel
->sample_id
= xyarray__new(ncpus
, nthreads
, sizeof(struct perf_sample_id
));
899 if (evsel
->sample_id
== NULL
)
902 evsel
->id
= zalloc(ncpus
* nthreads
* sizeof(u64
));
903 if (evsel
->id
== NULL
) {
904 xyarray__delete(evsel
->sample_id
);
905 evsel
->sample_id
= NULL
;
912 static void perf_evsel__free_fd(struct perf_evsel
*evsel
)
914 xyarray__delete(evsel
->fd
);
918 static void perf_evsel__free_id(struct perf_evsel
*evsel
)
920 xyarray__delete(evsel
->sample_id
);
921 evsel
->sample_id
= NULL
;
925 static void perf_evsel__free_config_terms(struct perf_evsel
*evsel
)
927 struct perf_evsel_config_term
*term
, *h
;
929 list_for_each_entry_safe(term
, h
, &evsel
->config_terms
, list
) {
930 list_del(&term
->list
);
935 void perf_evsel__close_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
939 if (evsel
->system_wide
)
942 for (cpu
= 0; cpu
< ncpus
; cpu
++)
943 for (thread
= 0; thread
< nthreads
; ++thread
) {
944 close(FD(evsel
, cpu
, thread
));
945 FD(evsel
, cpu
, thread
) = -1;
949 void perf_evsel__exit(struct perf_evsel
*evsel
)
951 assert(list_empty(&evsel
->node
));
952 perf_evsel__free_fd(evsel
);
953 perf_evsel__free_id(evsel
);
954 perf_evsel__free_config_terms(evsel
);
955 close_cgroup(evsel
->cgrp
);
956 cpu_map__put(evsel
->cpus
);
957 thread_map__put(evsel
->threads
);
958 zfree(&evsel
->group_name
);
960 perf_evsel__object
.fini(evsel
);
963 void perf_evsel__delete(struct perf_evsel
*evsel
)
965 perf_evsel__exit(evsel
);
969 void perf_evsel__compute_deltas(struct perf_evsel
*evsel
, int cpu
, int thread
,
970 struct perf_counts_values
*count
)
972 struct perf_counts_values tmp
;
974 if (!evsel
->prev_raw_counts
)
978 tmp
= evsel
->prev_raw_counts
->aggr
;
979 evsel
->prev_raw_counts
->aggr
= *count
;
981 tmp
= *perf_counts(evsel
->prev_raw_counts
, cpu
, thread
);
982 *perf_counts(evsel
->prev_raw_counts
, cpu
, thread
) = *count
;
985 count
->val
= count
->val
- tmp
.val
;
986 count
->ena
= count
->ena
- tmp
.ena
;
987 count
->run
= count
->run
- tmp
.run
;
990 void perf_counts_values__scale(struct perf_counts_values
*count
,
991 bool scale
, s8
*pscaled
)
996 if (count
->run
== 0) {
999 } else if (count
->run
< count
->ena
) {
1001 count
->val
= (u64
)((double) count
->val
* count
->ena
/ count
->run
+ 0.5);
1004 count
->ena
= count
->run
= 0;
1010 int perf_evsel__read(struct perf_evsel
*evsel
, int cpu
, int thread
,
1011 struct perf_counts_values
*count
)
1013 memset(count
, 0, sizeof(*count
));
1015 if (FD(evsel
, cpu
, thread
) < 0)
1018 if (readn(FD(evsel
, cpu
, thread
), count
, sizeof(*count
)) < 0)
1024 int __perf_evsel__read_on_cpu(struct perf_evsel
*evsel
,
1025 int cpu
, int thread
, bool scale
)
1027 struct perf_counts_values count
;
1028 size_t nv
= scale
? 3 : 1;
1030 if (FD(evsel
, cpu
, thread
) < 0)
1033 if (evsel
->counts
== NULL
&& perf_evsel__alloc_counts(evsel
, cpu
+ 1, thread
+ 1) < 0)
1036 if (readn(FD(evsel
, cpu
, thread
), &count
, nv
* sizeof(u64
)) < 0)
1039 perf_evsel__compute_deltas(evsel
, cpu
, thread
, &count
);
1040 perf_counts_values__scale(&count
, scale
, NULL
);
1041 *perf_counts(evsel
->counts
, cpu
, thread
) = count
;
1045 static int get_group_fd(struct perf_evsel
*evsel
, int cpu
, int thread
)
1047 struct perf_evsel
*leader
= evsel
->leader
;
1050 if (perf_evsel__is_group_leader(evsel
))
1054 * Leader must be already processed/open,
1055 * if not it's a bug.
1057 BUG_ON(!leader
->fd
);
1059 fd
= FD(leader
, cpu
, thread
);
1070 static void __p_bits(char *buf
, size_t size
, u64 value
, struct bit_names
*bits
)
1072 bool first_bit
= true;
1076 if (value
& bits
[i
].bit
) {
1077 buf
+= scnprintf(buf
, size
, "%s%s", first_bit
? "" : "|", bits
[i
].name
);
1080 } while (bits
[++i
].name
!= NULL
);
1083 static void __p_sample_type(char *buf
, size_t size
, u64 value
)
1085 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1086 struct bit_names bits
[] = {
1087 bit_name(IP
), bit_name(TID
), bit_name(TIME
), bit_name(ADDR
),
1088 bit_name(READ
), bit_name(CALLCHAIN
), bit_name(ID
), bit_name(CPU
),
1089 bit_name(PERIOD
), bit_name(STREAM_ID
), bit_name(RAW
),
1090 bit_name(BRANCH_STACK
), bit_name(REGS_USER
), bit_name(STACK_USER
),
1091 bit_name(IDENTIFIER
), bit_name(REGS_INTR
),
1095 __p_bits(buf
, size
, value
, bits
);
1098 static void __p_read_format(char *buf
, size_t size
, u64 value
)
1100 #define bit_name(n) { PERF_FORMAT_##n, #n }
1101 struct bit_names bits
[] = {
1102 bit_name(TOTAL_TIME_ENABLED
), bit_name(TOTAL_TIME_RUNNING
),
1103 bit_name(ID
), bit_name(GROUP
),
1107 __p_bits(buf
, size
, value
, bits
);
1110 #define BUF_SIZE 1024
1112 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1113 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1114 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1115 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1116 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1118 #define PRINT_ATTRn(_n, _f, _p) \
1122 ret += attr__fprintf(fp, _n, buf, priv);\
1126 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1128 int perf_event_attr__fprintf(FILE *fp
, struct perf_event_attr
*attr
,
1129 attr__fprintf_f attr__fprintf
, void *priv
)
1134 PRINT_ATTRf(type
, p_unsigned
);
1135 PRINT_ATTRf(size
, p_unsigned
);
1136 PRINT_ATTRf(config
, p_hex
);
1137 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period
, p_unsigned
);
1138 PRINT_ATTRf(sample_type
, p_sample_type
);
1139 PRINT_ATTRf(read_format
, p_read_format
);
1141 PRINT_ATTRf(disabled
, p_unsigned
);
1142 PRINT_ATTRf(inherit
, p_unsigned
);
1143 PRINT_ATTRf(pinned
, p_unsigned
);
1144 PRINT_ATTRf(exclusive
, p_unsigned
);
1145 PRINT_ATTRf(exclude_user
, p_unsigned
);
1146 PRINT_ATTRf(exclude_kernel
, p_unsigned
);
1147 PRINT_ATTRf(exclude_hv
, p_unsigned
);
1148 PRINT_ATTRf(exclude_idle
, p_unsigned
);
1149 PRINT_ATTRf(mmap
, p_unsigned
);
1150 PRINT_ATTRf(comm
, p_unsigned
);
1151 PRINT_ATTRf(freq
, p_unsigned
);
1152 PRINT_ATTRf(inherit_stat
, p_unsigned
);
1153 PRINT_ATTRf(enable_on_exec
, p_unsigned
);
1154 PRINT_ATTRf(task
, p_unsigned
);
1155 PRINT_ATTRf(watermark
, p_unsigned
);
1156 PRINT_ATTRf(precise_ip
, p_unsigned
);
1157 PRINT_ATTRf(mmap_data
, p_unsigned
);
1158 PRINT_ATTRf(sample_id_all
, p_unsigned
);
1159 PRINT_ATTRf(exclude_host
, p_unsigned
);
1160 PRINT_ATTRf(exclude_guest
, p_unsigned
);
1161 PRINT_ATTRf(exclude_callchain_kernel
, p_unsigned
);
1162 PRINT_ATTRf(exclude_callchain_user
, p_unsigned
);
1163 PRINT_ATTRf(mmap2
, p_unsigned
);
1164 PRINT_ATTRf(comm_exec
, p_unsigned
);
1165 PRINT_ATTRf(use_clockid
, p_unsigned
);
1166 PRINT_ATTRf(context_switch
, p_unsigned
);
1168 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events
, p_unsigned
);
1169 PRINT_ATTRf(bp_type
, p_unsigned
);
1170 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr
, p_hex
);
1171 PRINT_ATTRn("{ bp_len, config2 }", bp_len
, p_hex
);
1172 PRINT_ATTRf(sample_regs_user
, p_hex
);
1173 PRINT_ATTRf(sample_stack_user
, p_unsigned
);
1174 PRINT_ATTRf(clockid
, p_signed
);
1175 PRINT_ATTRf(sample_regs_intr
, p_hex
);
1176 PRINT_ATTRf(aux_watermark
, p_unsigned
);
1181 static int __open_attr__fprintf(FILE *fp
, const char *name
, const char *val
,
1182 void *priv
__attribute__((unused
)))
1184 return fprintf(fp
, " %-32s %s\n", name
, val
);
1187 static int __perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1188 struct thread_map
*threads
)
1190 int cpu
, thread
, nthreads
;
1191 unsigned long flags
= PERF_FLAG_FD_CLOEXEC
;
1193 enum { NO_CHANGE
, SET_TO_MAX
, INCREASED_MAX
} set_rlimit
= NO_CHANGE
;
1195 if (evsel
->system_wide
)
1198 nthreads
= threads
->nr
;
1200 if (evsel
->fd
== NULL
&&
1201 perf_evsel__alloc_fd(evsel
, cpus
->nr
, nthreads
) < 0)
1205 flags
|= PERF_FLAG_PID_CGROUP
;
1206 pid
= evsel
->cgrp
->fd
;
1209 fallback_missing_features
:
1210 if (perf_missing_features
.clockid_wrong
)
1211 evsel
->attr
.clockid
= CLOCK_MONOTONIC
; /* should always work */
1212 if (perf_missing_features
.clockid
) {
1213 evsel
->attr
.use_clockid
= 0;
1214 evsel
->attr
.clockid
= 0;
1216 if (perf_missing_features
.cloexec
)
1217 flags
&= ~(unsigned long)PERF_FLAG_FD_CLOEXEC
;
1218 if (perf_missing_features
.mmap2
)
1219 evsel
->attr
.mmap2
= 0;
1220 if (perf_missing_features
.exclude_guest
)
1221 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
1223 if (perf_missing_features
.sample_id_all
)
1224 evsel
->attr
.sample_id_all
= 0;
1227 fprintf(stderr
, "%.60s\n", graph_dotted_line
);
1228 fprintf(stderr
, "perf_event_attr:\n");
1229 perf_event_attr__fprintf(stderr
, &evsel
->attr
, __open_attr__fprintf
, NULL
);
1230 fprintf(stderr
, "%.60s\n", graph_dotted_line
);
1233 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
1235 for (thread
= 0; thread
< nthreads
; thread
++) {
1238 if (!evsel
->cgrp
&& !evsel
->system_wide
)
1239 pid
= thread_map__pid(threads
, thread
);
1241 group_fd
= get_group_fd(evsel
, cpu
, thread
);
1243 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1244 pid
, cpus
->map
[cpu
], group_fd
, flags
);
1246 FD(evsel
, cpu
, thread
) = sys_perf_event_open(&evsel
->attr
,
1250 if (FD(evsel
, cpu
, thread
) < 0) {
1252 pr_debug2("sys_perf_event_open failed, error %d\n",
1256 set_rlimit
= NO_CHANGE
;
1259 * If we succeeded but had to kill clockid, fail and
1260 * have perf_evsel__open_strerror() print us a nice
1263 if (perf_missing_features
.clockid
||
1264 perf_missing_features
.clockid_wrong
) {
1275 * perf stat needs between 5 and 22 fds per CPU. When we run out
1276 * of them try to increase the limits.
1278 if (err
== -EMFILE
&& set_rlimit
< INCREASED_MAX
) {
1280 int old_errno
= errno
;
1282 if (getrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1283 if (set_rlimit
== NO_CHANGE
)
1284 l
.rlim_cur
= l
.rlim_max
;
1286 l
.rlim_cur
= l
.rlim_max
+ 1000;
1287 l
.rlim_max
= l
.rlim_cur
;
1289 if (setrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1298 if (err
!= -EINVAL
|| cpu
> 0 || thread
> 0)
1302 * Must probe features in the order they were added to the
1303 * perf_event_attr interface.
1305 if (!perf_missing_features
.clockid_wrong
&& evsel
->attr
.use_clockid
) {
1306 perf_missing_features
.clockid_wrong
= true;
1307 goto fallback_missing_features
;
1308 } else if (!perf_missing_features
.clockid
&& evsel
->attr
.use_clockid
) {
1309 perf_missing_features
.clockid
= true;
1310 goto fallback_missing_features
;
1311 } else if (!perf_missing_features
.cloexec
&& (flags
& PERF_FLAG_FD_CLOEXEC
)) {
1312 perf_missing_features
.cloexec
= true;
1313 goto fallback_missing_features
;
1314 } else if (!perf_missing_features
.mmap2
&& evsel
->attr
.mmap2
) {
1315 perf_missing_features
.mmap2
= true;
1316 goto fallback_missing_features
;
1317 } else if (!perf_missing_features
.exclude_guest
&&
1318 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
1319 perf_missing_features
.exclude_guest
= true;
1320 goto fallback_missing_features
;
1321 } else if (!perf_missing_features
.sample_id_all
) {
1322 perf_missing_features
.sample_id_all
= true;
1323 goto retry_sample_id
;
1328 while (--thread
>= 0) {
1329 close(FD(evsel
, cpu
, thread
));
1330 FD(evsel
, cpu
, thread
) = -1;
1333 } while (--cpu
>= 0);
1337 void perf_evsel__close(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1339 if (evsel
->fd
== NULL
)
1342 perf_evsel__close_fd(evsel
, ncpus
, nthreads
);
1343 perf_evsel__free_fd(evsel
);
1355 struct thread_map map
;
1357 } empty_thread_map
= {
1362 int perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1363 struct thread_map
*threads
)
1366 /* Work around old compiler warnings about strict aliasing */
1367 cpus
= &empty_cpu_map
.map
;
1370 if (threads
== NULL
)
1371 threads
= &empty_thread_map
.map
;
1373 return __perf_evsel__open(evsel
, cpus
, threads
);
1376 int perf_evsel__open_per_cpu(struct perf_evsel
*evsel
,
1377 struct cpu_map
*cpus
)
1379 return __perf_evsel__open(evsel
, cpus
, &empty_thread_map
.map
);
1382 int perf_evsel__open_per_thread(struct perf_evsel
*evsel
,
1383 struct thread_map
*threads
)
1385 return __perf_evsel__open(evsel
, &empty_cpu_map
.map
, threads
);
1388 static int perf_evsel__parse_id_sample(const struct perf_evsel
*evsel
,
1389 const union perf_event
*event
,
1390 struct perf_sample
*sample
)
1392 u64 type
= evsel
->attr
.sample_type
;
1393 const u64
*array
= event
->sample
.array
;
1394 bool swapped
= evsel
->needs_swap
;
1397 array
+= ((event
->header
.size
-
1398 sizeof(event
->header
)) / sizeof(u64
)) - 1;
1400 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1401 sample
->id
= *array
;
1405 if (type
& PERF_SAMPLE_CPU
) {
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]);
1413 sample
->cpu
= u
.val32
[0];
1417 if (type
& PERF_SAMPLE_STREAM_ID
) {
1418 sample
->stream_id
= *array
;
1422 if (type
& PERF_SAMPLE_ID
) {
1423 sample
->id
= *array
;
1427 if (type
& PERF_SAMPLE_TIME
) {
1428 sample
->time
= *array
;
1432 if (type
& PERF_SAMPLE_TID
) {
1435 /* undo swap of u64, then swap on individual u32s */
1436 u
.val64
= bswap_64(u
.val64
);
1437 u
.val32
[0] = bswap_32(u
.val32
[0]);
1438 u
.val32
[1] = bswap_32(u
.val32
[1]);
1441 sample
->pid
= u
.val32
[0];
1442 sample
->tid
= u
.val32
[1];
1449 static inline bool overflow(const void *endp
, u16 max_size
, const void *offset
,
1452 return size
> max_size
|| offset
+ size
> endp
;
1455 #define OVERFLOW_CHECK(offset, size, max_size) \
1457 if (overflow(endp, (max_size), (offset), (size))) \
1461 #define OVERFLOW_CHECK_u64(offset) \
1462 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1464 int perf_evsel__parse_sample(struct perf_evsel
*evsel
, union perf_event
*event
,
1465 struct perf_sample
*data
)
1467 u64 type
= evsel
->attr
.sample_type
;
1468 bool swapped
= evsel
->needs_swap
;
1470 u16 max_size
= event
->header
.size
;
1471 const void *endp
= (void *)event
+ max_size
;
1475 * used for cross-endian analysis. See git commit 65014ab3
1476 * for why this goofiness is needed.
1480 memset(data
, 0, sizeof(*data
));
1481 data
->cpu
= data
->pid
= data
->tid
= -1;
1482 data
->stream_id
= data
->id
= data
->time
= -1ULL;
1483 data
->period
= evsel
->attr
.sample_period
;
1486 if (event
->header
.type
!= PERF_RECORD_SAMPLE
) {
1487 if (!evsel
->attr
.sample_id_all
)
1489 return perf_evsel__parse_id_sample(evsel
, event
, data
);
1492 array
= event
->sample
.array
;
1495 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1496 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1497 * check the format does not go past the end of the event.
1499 if (evsel
->sample_size
+ sizeof(event
->header
) > event
->header
.size
)
1503 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1508 if (type
& PERF_SAMPLE_IP
) {
1513 if (type
& PERF_SAMPLE_TID
) {
1516 /* undo swap of u64, then swap on individual u32s */
1517 u
.val64
= bswap_64(u
.val64
);
1518 u
.val32
[0] = bswap_32(u
.val32
[0]);
1519 u
.val32
[1] = bswap_32(u
.val32
[1]);
1522 data
->pid
= u
.val32
[0];
1523 data
->tid
= u
.val32
[1];
1527 if (type
& PERF_SAMPLE_TIME
) {
1528 data
->time
= *array
;
1533 if (type
& PERF_SAMPLE_ADDR
) {
1534 data
->addr
= *array
;
1538 if (type
& PERF_SAMPLE_ID
) {
1543 if (type
& PERF_SAMPLE_STREAM_ID
) {
1544 data
->stream_id
= *array
;
1548 if (type
& PERF_SAMPLE_CPU
) {
1552 /* undo swap of u64, then swap on individual u32s */
1553 u
.val64
= bswap_64(u
.val64
);
1554 u
.val32
[0] = bswap_32(u
.val32
[0]);
1557 data
->cpu
= u
.val32
[0];
1561 if (type
& PERF_SAMPLE_PERIOD
) {
1562 data
->period
= *array
;
1566 if (type
& PERF_SAMPLE_READ
) {
1567 u64 read_format
= evsel
->attr
.read_format
;
1569 OVERFLOW_CHECK_u64(array
);
1570 if (read_format
& PERF_FORMAT_GROUP
)
1571 data
->read
.group
.nr
= *array
;
1573 data
->read
.one
.value
= *array
;
1577 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1578 OVERFLOW_CHECK_u64(array
);
1579 data
->read
.time_enabled
= *array
;
1583 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1584 OVERFLOW_CHECK_u64(array
);
1585 data
->read
.time_running
= *array
;
1589 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1590 if (read_format
& PERF_FORMAT_GROUP
) {
1591 const u64 max_group_nr
= UINT64_MAX
/
1592 sizeof(struct sample_read_value
);
1594 if (data
->read
.group
.nr
> max_group_nr
)
1596 sz
= data
->read
.group
.nr
*
1597 sizeof(struct sample_read_value
);
1598 OVERFLOW_CHECK(array
, sz
, max_size
);
1599 data
->read
.group
.values
=
1600 (struct sample_read_value
*)array
;
1601 array
= (void *)array
+ sz
;
1603 OVERFLOW_CHECK_u64(array
);
1604 data
->read
.one
.id
= *array
;
1609 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1610 const u64 max_callchain_nr
= UINT64_MAX
/ sizeof(u64
);
1612 OVERFLOW_CHECK_u64(array
);
1613 data
->callchain
= (struct ip_callchain
*)array
++;
1614 if (data
->callchain
->nr
> max_callchain_nr
)
1616 sz
= data
->callchain
->nr
* sizeof(u64
);
1617 OVERFLOW_CHECK(array
, sz
, max_size
);
1618 array
= (void *)array
+ sz
;
1621 if (type
& PERF_SAMPLE_RAW
) {
1622 OVERFLOW_CHECK_u64(array
);
1624 if (WARN_ONCE(swapped
,
1625 "Endianness of raw data not corrected!\n")) {
1626 /* undo swap of u64, then swap on individual u32s */
1627 u
.val64
= bswap_64(u
.val64
);
1628 u
.val32
[0] = bswap_32(u
.val32
[0]);
1629 u
.val32
[1] = bswap_32(u
.val32
[1]);
1631 data
->raw_size
= u
.val32
[0];
1632 array
= (void *)array
+ sizeof(u32
);
1634 OVERFLOW_CHECK(array
, data
->raw_size
, max_size
);
1635 data
->raw_data
= (void *)array
;
1636 array
= (void *)array
+ data
->raw_size
;
1639 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1640 const u64 max_branch_nr
= UINT64_MAX
/
1641 sizeof(struct branch_entry
);
1643 OVERFLOW_CHECK_u64(array
);
1644 data
->branch_stack
= (struct branch_stack
*)array
++;
1646 if (data
->branch_stack
->nr
> max_branch_nr
)
1648 sz
= data
->branch_stack
->nr
* sizeof(struct branch_entry
);
1649 OVERFLOW_CHECK(array
, sz
, max_size
);
1650 array
= (void *)array
+ sz
;
1653 if (type
& PERF_SAMPLE_REGS_USER
) {
1654 OVERFLOW_CHECK_u64(array
);
1655 data
->user_regs
.abi
= *array
;
1658 if (data
->user_regs
.abi
) {
1659 u64 mask
= evsel
->attr
.sample_regs_user
;
1661 sz
= hweight_long(mask
) * sizeof(u64
);
1662 OVERFLOW_CHECK(array
, sz
, max_size
);
1663 data
->user_regs
.mask
= mask
;
1664 data
->user_regs
.regs
= (u64
*)array
;
1665 array
= (void *)array
+ sz
;
1669 if (type
& PERF_SAMPLE_STACK_USER
) {
1670 OVERFLOW_CHECK_u64(array
);
1673 data
->user_stack
.offset
= ((char *)(array
- 1)
1677 data
->user_stack
.size
= 0;
1679 OVERFLOW_CHECK(array
, sz
, max_size
);
1680 data
->user_stack
.data
= (char *)array
;
1681 array
= (void *)array
+ sz
;
1682 OVERFLOW_CHECK_u64(array
);
1683 data
->user_stack
.size
= *array
++;
1684 if (WARN_ONCE(data
->user_stack
.size
> sz
,
1685 "user stack dump failure\n"))
1691 if (type
& PERF_SAMPLE_WEIGHT
) {
1692 OVERFLOW_CHECK_u64(array
);
1693 data
->weight
= *array
;
1697 data
->data_src
= PERF_MEM_DATA_SRC_NONE
;
1698 if (type
& PERF_SAMPLE_DATA_SRC
) {
1699 OVERFLOW_CHECK_u64(array
);
1700 data
->data_src
= *array
;
1704 data
->transaction
= 0;
1705 if (type
& PERF_SAMPLE_TRANSACTION
) {
1706 OVERFLOW_CHECK_u64(array
);
1707 data
->transaction
= *array
;
1711 data
->intr_regs
.abi
= PERF_SAMPLE_REGS_ABI_NONE
;
1712 if (type
& PERF_SAMPLE_REGS_INTR
) {
1713 OVERFLOW_CHECK_u64(array
);
1714 data
->intr_regs
.abi
= *array
;
1717 if (data
->intr_regs
.abi
!= PERF_SAMPLE_REGS_ABI_NONE
) {
1718 u64 mask
= evsel
->attr
.sample_regs_intr
;
1720 sz
= hweight_long(mask
) * sizeof(u64
);
1721 OVERFLOW_CHECK(array
, sz
, max_size
);
1722 data
->intr_regs
.mask
= mask
;
1723 data
->intr_regs
.regs
= (u64
*)array
;
1724 array
= (void *)array
+ sz
;
1731 size_t perf_event__sample_event_size(const struct perf_sample
*sample
, u64 type
,
1734 size_t sz
, result
= sizeof(struct sample_event
);
1736 if (type
& PERF_SAMPLE_IDENTIFIER
)
1737 result
+= sizeof(u64
);
1739 if (type
& PERF_SAMPLE_IP
)
1740 result
+= sizeof(u64
);
1742 if (type
& PERF_SAMPLE_TID
)
1743 result
+= sizeof(u64
);
1745 if (type
& PERF_SAMPLE_TIME
)
1746 result
+= sizeof(u64
);
1748 if (type
& PERF_SAMPLE_ADDR
)
1749 result
+= sizeof(u64
);
1751 if (type
& PERF_SAMPLE_ID
)
1752 result
+= sizeof(u64
);
1754 if (type
& PERF_SAMPLE_STREAM_ID
)
1755 result
+= sizeof(u64
);
1757 if (type
& PERF_SAMPLE_CPU
)
1758 result
+= sizeof(u64
);
1760 if (type
& PERF_SAMPLE_PERIOD
)
1761 result
+= sizeof(u64
);
1763 if (type
& PERF_SAMPLE_READ
) {
1764 result
+= sizeof(u64
);
1765 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1766 result
+= sizeof(u64
);
1767 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1768 result
+= sizeof(u64
);
1769 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1770 if (read_format
& PERF_FORMAT_GROUP
) {
1771 sz
= sample
->read
.group
.nr
*
1772 sizeof(struct sample_read_value
);
1775 result
+= sizeof(u64
);
1779 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1780 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1784 if (type
& PERF_SAMPLE_RAW
) {
1785 result
+= sizeof(u32
);
1786 result
+= sample
->raw_size
;
1789 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1790 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1795 if (type
& PERF_SAMPLE_REGS_USER
) {
1796 if (sample
->user_regs
.abi
) {
1797 result
+= sizeof(u64
);
1798 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
1801 result
+= sizeof(u64
);
1805 if (type
& PERF_SAMPLE_STACK_USER
) {
1806 sz
= sample
->user_stack
.size
;
1807 result
+= sizeof(u64
);
1810 result
+= sizeof(u64
);
1814 if (type
& PERF_SAMPLE_WEIGHT
)
1815 result
+= sizeof(u64
);
1817 if (type
& PERF_SAMPLE_DATA_SRC
)
1818 result
+= sizeof(u64
);
1820 if (type
& PERF_SAMPLE_TRANSACTION
)
1821 result
+= sizeof(u64
);
1823 if (type
& PERF_SAMPLE_REGS_INTR
) {
1824 if (sample
->intr_regs
.abi
) {
1825 result
+= sizeof(u64
);
1826 sz
= hweight_long(sample
->intr_regs
.mask
) * sizeof(u64
);
1829 result
+= sizeof(u64
);
1836 int perf_event__synthesize_sample(union perf_event
*event
, u64 type
,
1838 const struct perf_sample
*sample
,
1844 * used for cross-endian analysis. See git commit 65014ab3
1845 * for why this goofiness is needed.
1849 array
= event
->sample
.array
;
1851 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1852 *array
= sample
->id
;
1856 if (type
& PERF_SAMPLE_IP
) {
1857 *array
= sample
->ip
;
1861 if (type
& PERF_SAMPLE_TID
) {
1862 u
.val32
[0] = sample
->pid
;
1863 u
.val32
[1] = sample
->tid
;
1866 * Inverse of what is done in perf_evsel__parse_sample
1868 u
.val32
[0] = bswap_32(u
.val32
[0]);
1869 u
.val32
[1] = bswap_32(u
.val32
[1]);
1870 u
.val64
= bswap_64(u
.val64
);
1877 if (type
& PERF_SAMPLE_TIME
) {
1878 *array
= sample
->time
;
1882 if (type
& PERF_SAMPLE_ADDR
) {
1883 *array
= sample
->addr
;
1887 if (type
& PERF_SAMPLE_ID
) {
1888 *array
= sample
->id
;
1892 if (type
& PERF_SAMPLE_STREAM_ID
) {
1893 *array
= sample
->stream_id
;
1897 if (type
& PERF_SAMPLE_CPU
) {
1898 u
.val32
[0] = sample
->cpu
;
1901 * Inverse of what is done in perf_evsel__parse_sample
1903 u
.val32
[0] = bswap_32(u
.val32
[0]);
1904 u
.val64
= bswap_64(u
.val64
);
1910 if (type
& PERF_SAMPLE_PERIOD
) {
1911 *array
= sample
->period
;
1915 if (type
& PERF_SAMPLE_READ
) {
1916 if (read_format
& PERF_FORMAT_GROUP
)
1917 *array
= sample
->read
.group
.nr
;
1919 *array
= sample
->read
.one
.value
;
1922 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1923 *array
= sample
->read
.time_enabled
;
1927 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1928 *array
= sample
->read
.time_running
;
1932 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1933 if (read_format
& PERF_FORMAT_GROUP
) {
1934 sz
= sample
->read
.group
.nr
*
1935 sizeof(struct sample_read_value
);
1936 memcpy(array
, sample
->read
.group
.values
, sz
);
1937 array
= (void *)array
+ sz
;
1939 *array
= sample
->read
.one
.id
;
1944 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1945 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1946 memcpy(array
, sample
->callchain
, sz
);
1947 array
= (void *)array
+ sz
;
1950 if (type
& PERF_SAMPLE_RAW
) {
1951 u
.val32
[0] = sample
->raw_size
;
1952 if (WARN_ONCE(swapped
,
1953 "Endianness of raw data not corrected!\n")) {
1955 * Inverse of what is done in perf_evsel__parse_sample
1957 u
.val32
[0] = bswap_32(u
.val32
[0]);
1958 u
.val32
[1] = bswap_32(u
.val32
[1]);
1959 u
.val64
= bswap_64(u
.val64
);
1962 array
= (void *)array
+ sizeof(u32
);
1964 memcpy(array
, sample
->raw_data
, sample
->raw_size
);
1965 array
= (void *)array
+ sample
->raw_size
;
1968 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1969 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1971 memcpy(array
, sample
->branch_stack
, sz
);
1972 array
= (void *)array
+ sz
;
1975 if (type
& PERF_SAMPLE_REGS_USER
) {
1976 if (sample
->user_regs
.abi
) {
1977 *array
++ = sample
->user_regs
.abi
;
1978 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
1979 memcpy(array
, sample
->user_regs
.regs
, sz
);
1980 array
= (void *)array
+ sz
;
1986 if (type
& PERF_SAMPLE_STACK_USER
) {
1987 sz
= sample
->user_stack
.size
;
1990 memcpy(array
, sample
->user_stack
.data
, sz
);
1991 array
= (void *)array
+ sz
;
1996 if (type
& PERF_SAMPLE_WEIGHT
) {
1997 *array
= sample
->weight
;
2001 if (type
& PERF_SAMPLE_DATA_SRC
) {
2002 *array
= sample
->data_src
;
2006 if (type
& PERF_SAMPLE_TRANSACTION
) {
2007 *array
= sample
->transaction
;
2011 if (type
& PERF_SAMPLE_REGS_INTR
) {
2012 if (sample
->intr_regs
.abi
) {
2013 *array
++ = sample
->intr_regs
.abi
;
2014 sz
= hweight_long(sample
->intr_regs
.mask
) * sizeof(u64
);
2015 memcpy(array
, sample
->intr_regs
.regs
, sz
);
2016 array
= (void *)array
+ sz
;
2025 struct format_field
*perf_evsel__field(struct perf_evsel
*evsel
, const char *name
)
2027 return pevent_find_field(evsel
->tp_format
, name
);
2030 void *perf_evsel__rawptr(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
2033 struct format_field
*field
= perf_evsel__field(evsel
, name
);
2039 offset
= field
->offset
;
2041 if (field
->flags
& FIELD_IS_DYNAMIC
) {
2042 offset
= *(int *)(sample
->raw_data
+ field
->offset
);
2046 return sample
->raw_data
+ offset
;
2049 u64
perf_evsel__intval(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
2052 struct format_field
*field
= perf_evsel__field(evsel
, name
);
2059 ptr
= sample
->raw_data
+ field
->offset
;
2061 switch (field
->size
) {
2065 value
= *(u16
*)ptr
;
2068 value
= *(u32
*)ptr
;
2071 memcpy(&value
, ptr
, sizeof(u64
));
2077 if (!evsel
->needs_swap
)
2080 switch (field
->size
) {
2082 return bswap_16(value
);
2084 return bswap_32(value
);
2086 return bswap_64(value
);
2094 static int comma_fprintf(FILE *fp
, bool *first
, const char *fmt
, ...)
2100 ret
+= fprintf(fp
, ",");
2102 ret
+= fprintf(fp
, ":");
2106 va_start(args
, fmt
);
2107 ret
+= vfprintf(fp
, fmt
, args
);
2112 static int __print_attr__fprintf(FILE *fp
, const char *name
, const char *val
, void *priv
)
2114 return comma_fprintf(fp
, (bool *)priv
, " %s: %s", name
, val
);
2117 int perf_evsel__fprintf(struct perf_evsel
*evsel
,
2118 struct perf_attr_details
*details
, FILE *fp
)
2123 if (details
->event_group
) {
2124 struct perf_evsel
*pos
;
2126 if (!perf_evsel__is_group_leader(evsel
))
2129 if (evsel
->nr_members
> 1)
2130 printed
+= fprintf(fp
, "%s{", evsel
->group_name
?: "");
2132 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
2133 for_each_group_member(pos
, evsel
)
2134 printed
+= fprintf(fp
, ",%s", perf_evsel__name(pos
));
2136 if (evsel
->nr_members
> 1)
2137 printed
+= fprintf(fp
, "}");
2141 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
2143 if (details
->verbose
) {
2144 printed
+= perf_event_attr__fprintf(fp
, &evsel
->attr
,
2145 __print_attr__fprintf
, &first
);
2146 } else if (details
->freq
) {
2147 printed
+= comma_fprintf(fp
, &first
, " sample_freq=%" PRIu64
,
2148 (u64
)evsel
->attr
.sample_freq
);
2155 bool perf_evsel__fallback(struct perf_evsel
*evsel
, int err
,
2156 char *msg
, size_t msgsize
)
2158 if ((err
== ENOENT
|| err
== ENXIO
|| err
== ENODEV
) &&
2159 evsel
->attr
.type
== PERF_TYPE_HARDWARE
&&
2160 evsel
->attr
.config
== PERF_COUNT_HW_CPU_CYCLES
) {
2162 * If it's cycles then fall back to hrtimer based
2163 * cpu-clock-tick sw counter, which is always available even if
2166 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2169 scnprintf(msg
, msgsize
, "%s",
2170 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2172 evsel
->attr
.type
= PERF_TYPE_SOFTWARE
;
2173 evsel
->attr
.config
= PERF_COUNT_SW_CPU_CLOCK
;
2175 zfree(&evsel
->name
);
2182 int perf_evsel__open_strerror(struct perf_evsel
*evsel
, struct target
*target
,
2183 int err
, char *msg
, size_t size
)
2185 char sbuf
[STRERR_BUFSIZE
];
2190 return scnprintf(msg
, size
,
2191 "You may not have permission to collect %sstats.\n"
2192 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2193 " -1 - Not paranoid at all\n"
2194 " 0 - Disallow raw tracepoint access for unpriv\n"
2195 " 1 - Disallow cpu events for unpriv\n"
2196 " 2 - Disallow kernel profiling for unpriv",
2197 target
->system_wide
? "system-wide " : "");
2199 return scnprintf(msg
, size
, "The %s event is not supported.",
2200 perf_evsel__name(evsel
));
2202 return scnprintf(msg
, size
, "%s",
2203 "Too many events are opened.\n"
2204 "Probably the maximum number of open file descriptors has been reached.\n"
2205 "Hint: Try again after reducing the number of events.\n"
2206 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2208 if (target
->cpu_list
)
2209 return scnprintf(msg
, size
, "%s",
2210 "No such device - did you specify an out-of-range profile CPU?\n");
2213 if (evsel
->attr
.precise_ip
)
2214 return scnprintf(msg
, size
, "%s",
2215 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2216 #if defined(__i386__) || defined(__x86_64__)
2217 if (evsel
->attr
.type
== PERF_TYPE_HARDWARE
)
2218 return scnprintf(msg
, size
, "%s",
2219 "No hardware sampling interrupt available.\n"
2220 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2224 if (find_process("oprofiled"))
2225 return scnprintf(msg
, size
,
2226 "The PMU counters are busy/taken by another profiler.\n"
2227 "We found oprofile daemon running, please stop it and try again.");
2230 if (perf_missing_features
.clockid
)
2231 return scnprintf(msg
, size
, "clockid feature not supported.");
2232 if (perf_missing_features
.clockid_wrong
)
2233 return scnprintf(msg
, size
, "wrong clockid (%d).", clockid
);
2239 return scnprintf(msg
, size
,
2240 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2241 "/bin/dmesg may provide additional information.\n"
2242 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2243 err
, strerror_r(err
, sbuf
, sizeof(sbuf
)),
2244 perf_evsel__name(evsel
));