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
,
549 struct callchain_param
*param
)
551 bool function
= perf_evsel__is_function_event(evsel
);
552 struct perf_event_attr
*attr
= &evsel
->attr
;
554 perf_evsel__set_sample_bit(evsel
, CALLCHAIN
);
556 if (param
->record_mode
== CALLCHAIN_LBR
) {
557 if (!opts
->branch_stack
) {
558 if (attr
->exclude_user
) {
559 pr_warning("LBR callstack option is only available "
560 "to get user callchain information. "
561 "Falling back to framepointers.\n");
563 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
564 attr
->branch_sample_type
= PERF_SAMPLE_BRANCH_USER
|
565 PERF_SAMPLE_BRANCH_CALL_STACK
;
568 pr_warning("Cannot use LBR callstack with branch stack. "
569 "Falling back to framepointers.\n");
572 if (param
->record_mode
== CALLCHAIN_DWARF
) {
574 perf_evsel__set_sample_bit(evsel
, REGS_USER
);
575 perf_evsel__set_sample_bit(evsel
, STACK_USER
);
576 attr
->sample_regs_user
= PERF_REGS_MASK
;
577 attr
->sample_stack_user
= param
->dump_size
;
578 attr
->exclude_callchain_user
= 1;
580 pr_info("Cannot use DWARF unwind for function trace event,"
581 " falling back to framepointers.\n");
586 pr_info("Disabling user space callchains for function trace event.\n");
587 attr
->exclude_callchain_user
= 1;
591 static void apply_config_terms(struct perf_evsel
*evsel
)
593 struct perf_evsel_config_term
*term
;
594 struct list_head
*config_terms
= &evsel
->config_terms
;
595 struct perf_event_attr
*attr
= &evsel
->attr
;
597 list_for_each_entry(term
, config_terms
, list
) {
598 switch (term
->type
) {
599 case PERF_EVSEL__CONFIG_TERM_PERIOD
:
600 attr
->sample_period
= term
->val
.period
;
602 case PERF_EVSEL__CONFIG_TERM_TIME
:
604 perf_evsel__set_sample_bit(evsel
, TIME
);
606 perf_evsel__reset_sample_bit(evsel
, TIME
);
615 * The enable_on_exec/disabled value strategy:
617 * 1) For any type of traced program:
618 * - all independent events and group leaders are disabled
619 * - all group members are enabled
621 * Group members are ruled by group leaders. They need to
622 * be enabled, because the group scheduling relies on that.
624 * 2) For traced programs executed by perf:
625 * - all independent events and group leaders have
627 * - we don't specifically enable or disable any event during
630 * Independent events and group leaders are initially disabled
631 * and get enabled by exec. Group members are ruled by group
632 * leaders as stated in 1).
634 * 3) For traced programs attached by perf (pid/tid):
635 * - we specifically enable or disable all events during
638 * When attaching events to already running traced we
639 * enable/disable events specifically, as there's no
640 * initial traced exec call.
642 void perf_evsel__config(struct perf_evsel
*evsel
, struct record_opts
*opts
)
644 struct perf_evsel
*leader
= evsel
->leader
;
645 struct perf_event_attr
*attr
= &evsel
->attr
;
646 int track
= evsel
->tracking
;
647 bool per_cpu
= opts
->target
.default_per_cpu
&& !opts
->target
.per_thread
;
649 attr
->sample_id_all
= perf_missing_features
.sample_id_all
? 0 : 1;
650 attr
->inherit
= !opts
->no_inherit
;
652 perf_evsel__set_sample_bit(evsel
, IP
);
653 perf_evsel__set_sample_bit(evsel
, TID
);
655 if (evsel
->sample_read
) {
656 perf_evsel__set_sample_bit(evsel
, READ
);
659 * We need ID even in case of single event, because
660 * PERF_SAMPLE_READ process ID specific data.
662 perf_evsel__set_sample_id(evsel
, false);
665 * Apply group format only if we belong to group
666 * with more than one members.
668 if (leader
->nr_members
> 1) {
669 attr
->read_format
|= PERF_FORMAT_GROUP
;
675 * We default some events to have a default interval. But keep
676 * it a weak assumption overridable by the user.
678 if (!attr
->sample_period
|| (opts
->user_freq
!= UINT_MAX
||
679 opts
->user_interval
!= ULLONG_MAX
)) {
681 perf_evsel__set_sample_bit(evsel
, PERIOD
);
683 attr
->sample_freq
= opts
->freq
;
685 attr
->sample_period
= opts
->default_interval
;
690 * Disable sampling for all group members other
691 * than leader in case leader 'leads' the sampling.
693 if ((leader
!= evsel
) && leader
->sample_read
) {
694 attr
->sample_freq
= 0;
695 attr
->sample_period
= 0;
698 if (opts
->no_samples
)
699 attr
->sample_freq
= 0;
701 if (opts
->inherit_stat
)
702 attr
->inherit_stat
= 1;
704 if (opts
->sample_address
) {
705 perf_evsel__set_sample_bit(evsel
, ADDR
);
706 attr
->mmap_data
= track
;
710 * We don't allow user space callchains for function trace
711 * event, due to issues with page faults while tracing page
712 * fault handler and its overall trickiness nature.
714 if (perf_evsel__is_function_event(evsel
))
715 evsel
->attr
.exclude_callchain_user
= 1;
717 if (callchain_param
.enabled
&& !evsel
->no_aux_samples
)
718 perf_evsel__config_callgraph(evsel
, opts
, &callchain_param
);
720 if (opts
->sample_intr_regs
) {
721 attr
->sample_regs_intr
= PERF_REGS_MASK
;
722 perf_evsel__set_sample_bit(evsel
, REGS_INTR
);
725 if (target__has_cpu(&opts
->target
))
726 perf_evsel__set_sample_bit(evsel
, CPU
);
729 perf_evsel__set_sample_bit(evsel
, PERIOD
);
732 * When the user explicitely disabled time don't force it here.
734 if (opts
->sample_time
&&
735 (!perf_missing_features
.sample_id_all
&&
736 (!opts
->no_inherit
|| target__has_cpu(&opts
->target
) || per_cpu
||
737 opts
->sample_time_set
)))
738 perf_evsel__set_sample_bit(evsel
, TIME
);
740 if (opts
->raw_samples
&& !evsel
->no_aux_samples
) {
741 perf_evsel__set_sample_bit(evsel
, TIME
);
742 perf_evsel__set_sample_bit(evsel
, RAW
);
743 perf_evsel__set_sample_bit(evsel
, CPU
);
746 if (opts
->sample_address
)
747 perf_evsel__set_sample_bit(evsel
, DATA_SRC
);
749 if (opts
->no_buffering
) {
751 attr
->wakeup_events
= 1;
753 if (opts
->branch_stack
&& !evsel
->no_aux_samples
) {
754 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
755 attr
->branch_sample_type
= opts
->branch_stack
;
758 if (opts
->sample_weight
)
759 perf_evsel__set_sample_bit(evsel
, WEIGHT
);
763 attr
->mmap2
= track
&& !perf_missing_features
.mmap2
;
766 if (opts
->record_switch_events
)
767 attr
->context_switch
= track
;
769 if (opts
->sample_transaction
)
770 perf_evsel__set_sample_bit(evsel
, TRANSACTION
);
772 if (opts
->running_time
) {
773 evsel
->attr
.read_format
|=
774 PERF_FORMAT_TOTAL_TIME_ENABLED
|
775 PERF_FORMAT_TOTAL_TIME_RUNNING
;
779 * XXX see the function comment above
781 * Disabling only independent events or group leaders,
782 * keeping group members enabled.
784 if (perf_evsel__is_group_leader(evsel
))
788 * Setting enable_on_exec for independent events and
789 * group leaders for traced executed by perf.
791 if (target__none(&opts
->target
) && perf_evsel__is_group_leader(evsel
) &&
792 !opts
->initial_delay
)
793 attr
->enable_on_exec
= 1;
795 if (evsel
->immediate
) {
797 attr
->enable_on_exec
= 0;
800 clockid
= opts
->clockid
;
801 if (opts
->use_clockid
) {
802 attr
->use_clockid
= 1;
803 attr
->clockid
= opts
->clockid
;
807 * Apply event specific term settings,
808 * it overloads any global configuration.
810 apply_config_terms(evsel
);
813 static int perf_evsel__alloc_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
817 if (evsel
->system_wide
)
820 evsel
->fd
= xyarray__new(ncpus
, nthreads
, sizeof(int));
823 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
824 for (thread
= 0; thread
< nthreads
; thread
++) {
825 FD(evsel
, cpu
, thread
) = -1;
830 return evsel
->fd
!= NULL
? 0 : -ENOMEM
;
833 static int perf_evsel__run_ioctl(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
838 if (evsel
->system_wide
)
841 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
842 for (thread
= 0; thread
< nthreads
; thread
++) {
843 int fd
= FD(evsel
, cpu
, thread
),
844 err
= ioctl(fd
, ioc
, arg
);
854 int perf_evsel__apply_filter(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
857 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
858 PERF_EVENT_IOC_SET_FILTER
,
862 int perf_evsel__set_filter(struct perf_evsel
*evsel
, const char *filter
)
864 char *new_filter
= strdup(filter
);
866 if (new_filter
!= NULL
) {
868 evsel
->filter
= new_filter
;
875 int perf_evsel__append_filter(struct perf_evsel
*evsel
,
876 const char *op
, const char *filter
)
880 if (evsel
->filter
== NULL
)
881 return perf_evsel__set_filter(evsel
, filter
);
883 if (asprintf(&new_filter
,"(%s) %s (%s)", evsel
->filter
, op
, filter
) > 0) {
885 evsel
->filter
= new_filter
;
892 int perf_evsel__enable(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
894 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
895 PERF_EVENT_IOC_ENABLE
,
899 int perf_evsel__alloc_id(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
901 if (ncpus
== 0 || nthreads
== 0)
904 if (evsel
->system_wide
)
907 evsel
->sample_id
= xyarray__new(ncpus
, nthreads
, sizeof(struct perf_sample_id
));
908 if (evsel
->sample_id
== NULL
)
911 evsel
->id
= zalloc(ncpus
* nthreads
* sizeof(u64
));
912 if (evsel
->id
== NULL
) {
913 xyarray__delete(evsel
->sample_id
);
914 evsel
->sample_id
= NULL
;
921 static void perf_evsel__free_fd(struct perf_evsel
*evsel
)
923 xyarray__delete(evsel
->fd
);
927 static void perf_evsel__free_id(struct perf_evsel
*evsel
)
929 xyarray__delete(evsel
->sample_id
);
930 evsel
->sample_id
= NULL
;
934 static void perf_evsel__free_config_terms(struct perf_evsel
*evsel
)
936 struct perf_evsel_config_term
*term
, *h
;
938 list_for_each_entry_safe(term
, h
, &evsel
->config_terms
, list
) {
939 list_del(&term
->list
);
944 void perf_evsel__close_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
948 if (evsel
->system_wide
)
951 for (cpu
= 0; cpu
< ncpus
; cpu
++)
952 for (thread
= 0; thread
< nthreads
; ++thread
) {
953 close(FD(evsel
, cpu
, thread
));
954 FD(evsel
, cpu
, thread
) = -1;
958 void perf_evsel__exit(struct perf_evsel
*evsel
)
960 assert(list_empty(&evsel
->node
));
961 perf_evsel__free_fd(evsel
);
962 perf_evsel__free_id(evsel
);
963 perf_evsel__free_config_terms(evsel
);
964 close_cgroup(evsel
->cgrp
);
965 cpu_map__put(evsel
->cpus
);
966 thread_map__put(evsel
->threads
);
967 zfree(&evsel
->group_name
);
969 perf_evsel__object
.fini(evsel
);
972 void perf_evsel__delete(struct perf_evsel
*evsel
)
974 perf_evsel__exit(evsel
);
978 void perf_evsel__compute_deltas(struct perf_evsel
*evsel
, int cpu
, int thread
,
979 struct perf_counts_values
*count
)
981 struct perf_counts_values tmp
;
983 if (!evsel
->prev_raw_counts
)
987 tmp
= evsel
->prev_raw_counts
->aggr
;
988 evsel
->prev_raw_counts
->aggr
= *count
;
990 tmp
= *perf_counts(evsel
->prev_raw_counts
, cpu
, thread
);
991 *perf_counts(evsel
->prev_raw_counts
, cpu
, thread
) = *count
;
994 count
->val
= count
->val
- tmp
.val
;
995 count
->ena
= count
->ena
- tmp
.ena
;
996 count
->run
= count
->run
- tmp
.run
;
999 void perf_counts_values__scale(struct perf_counts_values
*count
,
1000 bool scale
, s8
*pscaled
)
1005 if (count
->run
== 0) {
1008 } else if (count
->run
< count
->ena
) {
1010 count
->val
= (u64
)((double) count
->val
* count
->ena
/ count
->run
+ 0.5);
1013 count
->ena
= count
->run
= 0;
1019 int perf_evsel__read(struct perf_evsel
*evsel
, int cpu
, int thread
,
1020 struct perf_counts_values
*count
)
1022 memset(count
, 0, sizeof(*count
));
1024 if (FD(evsel
, cpu
, thread
) < 0)
1027 if (readn(FD(evsel
, cpu
, thread
), count
, sizeof(*count
)) < 0)
1033 int __perf_evsel__read_on_cpu(struct perf_evsel
*evsel
,
1034 int cpu
, int thread
, bool scale
)
1036 struct perf_counts_values count
;
1037 size_t nv
= scale
? 3 : 1;
1039 if (FD(evsel
, cpu
, thread
) < 0)
1042 if (evsel
->counts
== NULL
&& perf_evsel__alloc_counts(evsel
, cpu
+ 1, thread
+ 1) < 0)
1045 if (readn(FD(evsel
, cpu
, thread
), &count
, nv
* sizeof(u64
)) < 0)
1048 perf_evsel__compute_deltas(evsel
, cpu
, thread
, &count
);
1049 perf_counts_values__scale(&count
, scale
, NULL
);
1050 *perf_counts(evsel
->counts
, cpu
, thread
) = count
;
1054 static int get_group_fd(struct perf_evsel
*evsel
, int cpu
, int thread
)
1056 struct perf_evsel
*leader
= evsel
->leader
;
1059 if (perf_evsel__is_group_leader(evsel
))
1063 * Leader must be already processed/open,
1064 * if not it's a bug.
1066 BUG_ON(!leader
->fd
);
1068 fd
= FD(leader
, cpu
, thread
);
1079 static void __p_bits(char *buf
, size_t size
, u64 value
, struct bit_names
*bits
)
1081 bool first_bit
= true;
1085 if (value
& bits
[i
].bit
) {
1086 buf
+= scnprintf(buf
, size
, "%s%s", first_bit
? "" : "|", bits
[i
].name
);
1089 } while (bits
[++i
].name
!= NULL
);
1092 static void __p_sample_type(char *buf
, size_t size
, u64 value
)
1094 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1095 struct bit_names bits
[] = {
1096 bit_name(IP
), bit_name(TID
), bit_name(TIME
), bit_name(ADDR
),
1097 bit_name(READ
), bit_name(CALLCHAIN
), bit_name(ID
), bit_name(CPU
),
1098 bit_name(PERIOD
), bit_name(STREAM_ID
), bit_name(RAW
),
1099 bit_name(BRANCH_STACK
), bit_name(REGS_USER
), bit_name(STACK_USER
),
1100 bit_name(IDENTIFIER
), bit_name(REGS_INTR
),
1104 __p_bits(buf
, size
, value
, bits
);
1107 static void __p_read_format(char *buf
, size_t size
, u64 value
)
1109 #define bit_name(n) { PERF_FORMAT_##n, #n }
1110 struct bit_names bits
[] = {
1111 bit_name(TOTAL_TIME_ENABLED
), bit_name(TOTAL_TIME_RUNNING
),
1112 bit_name(ID
), bit_name(GROUP
),
1116 __p_bits(buf
, size
, value
, bits
);
1119 #define BUF_SIZE 1024
1121 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1122 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1123 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1124 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1125 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1127 #define PRINT_ATTRn(_n, _f, _p) \
1131 ret += attr__fprintf(fp, _n, buf, priv);\
1135 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1137 int perf_event_attr__fprintf(FILE *fp
, struct perf_event_attr
*attr
,
1138 attr__fprintf_f attr__fprintf
, void *priv
)
1143 PRINT_ATTRf(type
, p_unsigned
);
1144 PRINT_ATTRf(size
, p_unsigned
);
1145 PRINT_ATTRf(config
, p_hex
);
1146 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period
, p_unsigned
);
1147 PRINT_ATTRf(sample_type
, p_sample_type
);
1148 PRINT_ATTRf(read_format
, p_read_format
);
1150 PRINT_ATTRf(disabled
, p_unsigned
);
1151 PRINT_ATTRf(inherit
, p_unsigned
);
1152 PRINT_ATTRf(pinned
, p_unsigned
);
1153 PRINT_ATTRf(exclusive
, p_unsigned
);
1154 PRINT_ATTRf(exclude_user
, p_unsigned
);
1155 PRINT_ATTRf(exclude_kernel
, p_unsigned
);
1156 PRINT_ATTRf(exclude_hv
, p_unsigned
);
1157 PRINT_ATTRf(exclude_idle
, p_unsigned
);
1158 PRINT_ATTRf(mmap
, p_unsigned
);
1159 PRINT_ATTRf(comm
, p_unsigned
);
1160 PRINT_ATTRf(freq
, p_unsigned
);
1161 PRINT_ATTRf(inherit_stat
, p_unsigned
);
1162 PRINT_ATTRf(enable_on_exec
, p_unsigned
);
1163 PRINT_ATTRf(task
, p_unsigned
);
1164 PRINT_ATTRf(watermark
, p_unsigned
);
1165 PRINT_ATTRf(precise_ip
, p_unsigned
);
1166 PRINT_ATTRf(mmap_data
, p_unsigned
);
1167 PRINT_ATTRf(sample_id_all
, p_unsigned
);
1168 PRINT_ATTRf(exclude_host
, p_unsigned
);
1169 PRINT_ATTRf(exclude_guest
, p_unsigned
);
1170 PRINT_ATTRf(exclude_callchain_kernel
, p_unsigned
);
1171 PRINT_ATTRf(exclude_callchain_user
, p_unsigned
);
1172 PRINT_ATTRf(mmap2
, p_unsigned
);
1173 PRINT_ATTRf(comm_exec
, p_unsigned
);
1174 PRINT_ATTRf(use_clockid
, p_unsigned
);
1175 PRINT_ATTRf(context_switch
, p_unsigned
);
1177 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events
, p_unsigned
);
1178 PRINT_ATTRf(bp_type
, p_unsigned
);
1179 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr
, p_hex
);
1180 PRINT_ATTRn("{ bp_len, config2 }", bp_len
, p_hex
);
1181 PRINT_ATTRf(sample_regs_user
, p_hex
);
1182 PRINT_ATTRf(sample_stack_user
, p_unsigned
);
1183 PRINT_ATTRf(clockid
, p_signed
);
1184 PRINT_ATTRf(sample_regs_intr
, p_hex
);
1185 PRINT_ATTRf(aux_watermark
, p_unsigned
);
1190 static int __open_attr__fprintf(FILE *fp
, const char *name
, const char *val
,
1191 void *priv
__attribute__((unused
)))
1193 return fprintf(fp
, " %-32s %s\n", name
, val
);
1196 static int __perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1197 struct thread_map
*threads
)
1199 int cpu
, thread
, nthreads
;
1200 unsigned long flags
= PERF_FLAG_FD_CLOEXEC
;
1202 enum { NO_CHANGE
, SET_TO_MAX
, INCREASED_MAX
} set_rlimit
= NO_CHANGE
;
1204 if (evsel
->system_wide
)
1207 nthreads
= threads
->nr
;
1209 if (evsel
->fd
== NULL
&&
1210 perf_evsel__alloc_fd(evsel
, cpus
->nr
, nthreads
) < 0)
1214 flags
|= PERF_FLAG_PID_CGROUP
;
1215 pid
= evsel
->cgrp
->fd
;
1218 fallback_missing_features
:
1219 if (perf_missing_features
.clockid_wrong
)
1220 evsel
->attr
.clockid
= CLOCK_MONOTONIC
; /* should always work */
1221 if (perf_missing_features
.clockid
) {
1222 evsel
->attr
.use_clockid
= 0;
1223 evsel
->attr
.clockid
= 0;
1225 if (perf_missing_features
.cloexec
)
1226 flags
&= ~(unsigned long)PERF_FLAG_FD_CLOEXEC
;
1227 if (perf_missing_features
.mmap2
)
1228 evsel
->attr
.mmap2
= 0;
1229 if (perf_missing_features
.exclude_guest
)
1230 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
1232 if (perf_missing_features
.sample_id_all
)
1233 evsel
->attr
.sample_id_all
= 0;
1236 fprintf(stderr
, "%.60s\n", graph_dotted_line
);
1237 fprintf(stderr
, "perf_event_attr:\n");
1238 perf_event_attr__fprintf(stderr
, &evsel
->attr
, __open_attr__fprintf
, NULL
);
1239 fprintf(stderr
, "%.60s\n", graph_dotted_line
);
1242 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
1244 for (thread
= 0; thread
< nthreads
; thread
++) {
1247 if (!evsel
->cgrp
&& !evsel
->system_wide
)
1248 pid
= thread_map__pid(threads
, thread
);
1250 group_fd
= get_group_fd(evsel
, cpu
, thread
);
1252 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1253 pid
, cpus
->map
[cpu
], group_fd
, flags
);
1255 FD(evsel
, cpu
, thread
) = sys_perf_event_open(&evsel
->attr
,
1259 if (FD(evsel
, cpu
, thread
) < 0) {
1261 pr_debug2("sys_perf_event_open failed, error %d\n",
1265 set_rlimit
= NO_CHANGE
;
1268 * If we succeeded but had to kill clockid, fail and
1269 * have perf_evsel__open_strerror() print us a nice
1272 if (perf_missing_features
.clockid
||
1273 perf_missing_features
.clockid_wrong
) {
1284 * perf stat needs between 5 and 22 fds per CPU. When we run out
1285 * of them try to increase the limits.
1287 if (err
== -EMFILE
&& set_rlimit
< INCREASED_MAX
) {
1289 int old_errno
= errno
;
1291 if (getrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1292 if (set_rlimit
== NO_CHANGE
)
1293 l
.rlim_cur
= l
.rlim_max
;
1295 l
.rlim_cur
= l
.rlim_max
+ 1000;
1296 l
.rlim_max
= l
.rlim_cur
;
1298 if (setrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1307 if (err
!= -EINVAL
|| cpu
> 0 || thread
> 0)
1311 * Must probe features in the order they were added to the
1312 * perf_event_attr interface.
1314 if (!perf_missing_features
.clockid_wrong
&& evsel
->attr
.use_clockid
) {
1315 perf_missing_features
.clockid_wrong
= true;
1316 goto fallback_missing_features
;
1317 } else if (!perf_missing_features
.clockid
&& evsel
->attr
.use_clockid
) {
1318 perf_missing_features
.clockid
= true;
1319 goto fallback_missing_features
;
1320 } else if (!perf_missing_features
.cloexec
&& (flags
& PERF_FLAG_FD_CLOEXEC
)) {
1321 perf_missing_features
.cloexec
= true;
1322 goto fallback_missing_features
;
1323 } else if (!perf_missing_features
.mmap2
&& evsel
->attr
.mmap2
) {
1324 perf_missing_features
.mmap2
= true;
1325 goto fallback_missing_features
;
1326 } else if (!perf_missing_features
.exclude_guest
&&
1327 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
1328 perf_missing_features
.exclude_guest
= true;
1329 goto fallback_missing_features
;
1330 } else if (!perf_missing_features
.sample_id_all
) {
1331 perf_missing_features
.sample_id_all
= true;
1332 goto retry_sample_id
;
1337 while (--thread
>= 0) {
1338 close(FD(evsel
, cpu
, thread
));
1339 FD(evsel
, cpu
, thread
) = -1;
1342 } while (--cpu
>= 0);
1346 void perf_evsel__close(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1348 if (evsel
->fd
== NULL
)
1351 perf_evsel__close_fd(evsel
, ncpus
, nthreads
);
1352 perf_evsel__free_fd(evsel
);
1364 struct thread_map map
;
1366 } empty_thread_map
= {
1371 int perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1372 struct thread_map
*threads
)
1375 /* Work around old compiler warnings about strict aliasing */
1376 cpus
= &empty_cpu_map
.map
;
1379 if (threads
== NULL
)
1380 threads
= &empty_thread_map
.map
;
1382 return __perf_evsel__open(evsel
, cpus
, threads
);
1385 int perf_evsel__open_per_cpu(struct perf_evsel
*evsel
,
1386 struct cpu_map
*cpus
)
1388 return __perf_evsel__open(evsel
, cpus
, &empty_thread_map
.map
);
1391 int perf_evsel__open_per_thread(struct perf_evsel
*evsel
,
1392 struct thread_map
*threads
)
1394 return __perf_evsel__open(evsel
, &empty_cpu_map
.map
, threads
);
1397 static int perf_evsel__parse_id_sample(const struct perf_evsel
*evsel
,
1398 const union perf_event
*event
,
1399 struct perf_sample
*sample
)
1401 u64 type
= evsel
->attr
.sample_type
;
1402 const u64
*array
= event
->sample
.array
;
1403 bool swapped
= evsel
->needs_swap
;
1406 array
+= ((event
->header
.size
-
1407 sizeof(event
->header
)) / sizeof(u64
)) - 1;
1409 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1410 sample
->id
= *array
;
1414 if (type
& PERF_SAMPLE_CPU
) {
1417 /* undo swap of u64, then swap on individual u32s */
1418 u
.val64
= bswap_64(u
.val64
);
1419 u
.val32
[0] = bswap_32(u
.val32
[0]);
1422 sample
->cpu
= u
.val32
[0];
1426 if (type
& PERF_SAMPLE_STREAM_ID
) {
1427 sample
->stream_id
= *array
;
1431 if (type
& PERF_SAMPLE_ID
) {
1432 sample
->id
= *array
;
1436 if (type
& PERF_SAMPLE_TIME
) {
1437 sample
->time
= *array
;
1441 if (type
& PERF_SAMPLE_TID
) {
1444 /* undo swap of u64, then swap on individual u32s */
1445 u
.val64
= bswap_64(u
.val64
);
1446 u
.val32
[0] = bswap_32(u
.val32
[0]);
1447 u
.val32
[1] = bswap_32(u
.val32
[1]);
1450 sample
->pid
= u
.val32
[0];
1451 sample
->tid
= u
.val32
[1];
1458 static inline bool overflow(const void *endp
, u16 max_size
, const void *offset
,
1461 return size
> max_size
|| offset
+ size
> endp
;
1464 #define OVERFLOW_CHECK(offset, size, max_size) \
1466 if (overflow(endp, (max_size), (offset), (size))) \
1470 #define OVERFLOW_CHECK_u64(offset) \
1471 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1473 int perf_evsel__parse_sample(struct perf_evsel
*evsel
, union perf_event
*event
,
1474 struct perf_sample
*data
)
1476 u64 type
= evsel
->attr
.sample_type
;
1477 bool swapped
= evsel
->needs_swap
;
1479 u16 max_size
= event
->header
.size
;
1480 const void *endp
= (void *)event
+ max_size
;
1484 * used for cross-endian analysis. See git commit 65014ab3
1485 * for why this goofiness is needed.
1489 memset(data
, 0, sizeof(*data
));
1490 data
->cpu
= data
->pid
= data
->tid
= -1;
1491 data
->stream_id
= data
->id
= data
->time
= -1ULL;
1492 data
->period
= evsel
->attr
.sample_period
;
1495 if (event
->header
.type
!= PERF_RECORD_SAMPLE
) {
1496 if (!evsel
->attr
.sample_id_all
)
1498 return perf_evsel__parse_id_sample(evsel
, event
, data
);
1501 array
= event
->sample
.array
;
1504 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1505 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1506 * check the format does not go past the end of the event.
1508 if (evsel
->sample_size
+ sizeof(event
->header
) > event
->header
.size
)
1512 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1517 if (type
& PERF_SAMPLE_IP
) {
1522 if (type
& PERF_SAMPLE_TID
) {
1525 /* undo swap of u64, then swap on individual u32s */
1526 u
.val64
= bswap_64(u
.val64
);
1527 u
.val32
[0] = bswap_32(u
.val32
[0]);
1528 u
.val32
[1] = bswap_32(u
.val32
[1]);
1531 data
->pid
= u
.val32
[0];
1532 data
->tid
= u
.val32
[1];
1536 if (type
& PERF_SAMPLE_TIME
) {
1537 data
->time
= *array
;
1542 if (type
& PERF_SAMPLE_ADDR
) {
1543 data
->addr
= *array
;
1547 if (type
& PERF_SAMPLE_ID
) {
1552 if (type
& PERF_SAMPLE_STREAM_ID
) {
1553 data
->stream_id
= *array
;
1557 if (type
& PERF_SAMPLE_CPU
) {
1561 /* undo swap of u64, then swap on individual u32s */
1562 u
.val64
= bswap_64(u
.val64
);
1563 u
.val32
[0] = bswap_32(u
.val32
[0]);
1566 data
->cpu
= u
.val32
[0];
1570 if (type
& PERF_SAMPLE_PERIOD
) {
1571 data
->period
= *array
;
1575 if (type
& PERF_SAMPLE_READ
) {
1576 u64 read_format
= evsel
->attr
.read_format
;
1578 OVERFLOW_CHECK_u64(array
);
1579 if (read_format
& PERF_FORMAT_GROUP
)
1580 data
->read
.group
.nr
= *array
;
1582 data
->read
.one
.value
= *array
;
1586 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1587 OVERFLOW_CHECK_u64(array
);
1588 data
->read
.time_enabled
= *array
;
1592 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1593 OVERFLOW_CHECK_u64(array
);
1594 data
->read
.time_running
= *array
;
1598 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1599 if (read_format
& PERF_FORMAT_GROUP
) {
1600 const u64 max_group_nr
= UINT64_MAX
/
1601 sizeof(struct sample_read_value
);
1603 if (data
->read
.group
.nr
> max_group_nr
)
1605 sz
= data
->read
.group
.nr
*
1606 sizeof(struct sample_read_value
);
1607 OVERFLOW_CHECK(array
, sz
, max_size
);
1608 data
->read
.group
.values
=
1609 (struct sample_read_value
*)array
;
1610 array
= (void *)array
+ sz
;
1612 OVERFLOW_CHECK_u64(array
);
1613 data
->read
.one
.id
= *array
;
1618 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1619 const u64 max_callchain_nr
= UINT64_MAX
/ sizeof(u64
);
1621 OVERFLOW_CHECK_u64(array
);
1622 data
->callchain
= (struct ip_callchain
*)array
++;
1623 if (data
->callchain
->nr
> max_callchain_nr
)
1625 sz
= data
->callchain
->nr
* sizeof(u64
);
1626 OVERFLOW_CHECK(array
, sz
, max_size
);
1627 array
= (void *)array
+ sz
;
1630 if (type
& PERF_SAMPLE_RAW
) {
1631 OVERFLOW_CHECK_u64(array
);
1633 if (WARN_ONCE(swapped
,
1634 "Endianness of raw data not corrected!\n")) {
1635 /* undo swap of u64, then swap on individual u32s */
1636 u
.val64
= bswap_64(u
.val64
);
1637 u
.val32
[0] = bswap_32(u
.val32
[0]);
1638 u
.val32
[1] = bswap_32(u
.val32
[1]);
1640 data
->raw_size
= u
.val32
[0];
1641 array
= (void *)array
+ sizeof(u32
);
1643 OVERFLOW_CHECK(array
, data
->raw_size
, max_size
);
1644 data
->raw_data
= (void *)array
;
1645 array
= (void *)array
+ data
->raw_size
;
1648 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1649 const u64 max_branch_nr
= UINT64_MAX
/
1650 sizeof(struct branch_entry
);
1652 OVERFLOW_CHECK_u64(array
);
1653 data
->branch_stack
= (struct branch_stack
*)array
++;
1655 if (data
->branch_stack
->nr
> max_branch_nr
)
1657 sz
= data
->branch_stack
->nr
* sizeof(struct branch_entry
);
1658 OVERFLOW_CHECK(array
, sz
, max_size
);
1659 array
= (void *)array
+ sz
;
1662 if (type
& PERF_SAMPLE_REGS_USER
) {
1663 OVERFLOW_CHECK_u64(array
);
1664 data
->user_regs
.abi
= *array
;
1667 if (data
->user_regs
.abi
) {
1668 u64 mask
= evsel
->attr
.sample_regs_user
;
1670 sz
= hweight_long(mask
) * sizeof(u64
);
1671 OVERFLOW_CHECK(array
, sz
, max_size
);
1672 data
->user_regs
.mask
= mask
;
1673 data
->user_regs
.regs
= (u64
*)array
;
1674 array
= (void *)array
+ sz
;
1678 if (type
& PERF_SAMPLE_STACK_USER
) {
1679 OVERFLOW_CHECK_u64(array
);
1682 data
->user_stack
.offset
= ((char *)(array
- 1)
1686 data
->user_stack
.size
= 0;
1688 OVERFLOW_CHECK(array
, sz
, max_size
);
1689 data
->user_stack
.data
= (char *)array
;
1690 array
= (void *)array
+ sz
;
1691 OVERFLOW_CHECK_u64(array
);
1692 data
->user_stack
.size
= *array
++;
1693 if (WARN_ONCE(data
->user_stack
.size
> sz
,
1694 "user stack dump failure\n"))
1700 if (type
& PERF_SAMPLE_WEIGHT
) {
1701 OVERFLOW_CHECK_u64(array
);
1702 data
->weight
= *array
;
1706 data
->data_src
= PERF_MEM_DATA_SRC_NONE
;
1707 if (type
& PERF_SAMPLE_DATA_SRC
) {
1708 OVERFLOW_CHECK_u64(array
);
1709 data
->data_src
= *array
;
1713 data
->transaction
= 0;
1714 if (type
& PERF_SAMPLE_TRANSACTION
) {
1715 OVERFLOW_CHECK_u64(array
);
1716 data
->transaction
= *array
;
1720 data
->intr_regs
.abi
= PERF_SAMPLE_REGS_ABI_NONE
;
1721 if (type
& PERF_SAMPLE_REGS_INTR
) {
1722 OVERFLOW_CHECK_u64(array
);
1723 data
->intr_regs
.abi
= *array
;
1726 if (data
->intr_regs
.abi
!= PERF_SAMPLE_REGS_ABI_NONE
) {
1727 u64 mask
= evsel
->attr
.sample_regs_intr
;
1729 sz
= hweight_long(mask
) * sizeof(u64
);
1730 OVERFLOW_CHECK(array
, sz
, max_size
);
1731 data
->intr_regs
.mask
= mask
;
1732 data
->intr_regs
.regs
= (u64
*)array
;
1733 array
= (void *)array
+ sz
;
1740 size_t perf_event__sample_event_size(const struct perf_sample
*sample
, u64 type
,
1743 size_t sz
, result
= sizeof(struct sample_event
);
1745 if (type
& PERF_SAMPLE_IDENTIFIER
)
1746 result
+= sizeof(u64
);
1748 if (type
& PERF_SAMPLE_IP
)
1749 result
+= sizeof(u64
);
1751 if (type
& PERF_SAMPLE_TID
)
1752 result
+= sizeof(u64
);
1754 if (type
& PERF_SAMPLE_TIME
)
1755 result
+= sizeof(u64
);
1757 if (type
& PERF_SAMPLE_ADDR
)
1758 result
+= sizeof(u64
);
1760 if (type
& PERF_SAMPLE_ID
)
1761 result
+= sizeof(u64
);
1763 if (type
& PERF_SAMPLE_STREAM_ID
)
1764 result
+= sizeof(u64
);
1766 if (type
& PERF_SAMPLE_CPU
)
1767 result
+= sizeof(u64
);
1769 if (type
& PERF_SAMPLE_PERIOD
)
1770 result
+= sizeof(u64
);
1772 if (type
& PERF_SAMPLE_READ
) {
1773 result
+= sizeof(u64
);
1774 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1775 result
+= sizeof(u64
);
1776 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1777 result
+= sizeof(u64
);
1778 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1779 if (read_format
& PERF_FORMAT_GROUP
) {
1780 sz
= sample
->read
.group
.nr
*
1781 sizeof(struct sample_read_value
);
1784 result
+= sizeof(u64
);
1788 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1789 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1793 if (type
& PERF_SAMPLE_RAW
) {
1794 result
+= sizeof(u32
);
1795 result
+= sample
->raw_size
;
1798 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1799 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1804 if (type
& PERF_SAMPLE_REGS_USER
) {
1805 if (sample
->user_regs
.abi
) {
1806 result
+= sizeof(u64
);
1807 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
1810 result
+= sizeof(u64
);
1814 if (type
& PERF_SAMPLE_STACK_USER
) {
1815 sz
= sample
->user_stack
.size
;
1816 result
+= sizeof(u64
);
1819 result
+= sizeof(u64
);
1823 if (type
& PERF_SAMPLE_WEIGHT
)
1824 result
+= sizeof(u64
);
1826 if (type
& PERF_SAMPLE_DATA_SRC
)
1827 result
+= sizeof(u64
);
1829 if (type
& PERF_SAMPLE_TRANSACTION
)
1830 result
+= sizeof(u64
);
1832 if (type
& PERF_SAMPLE_REGS_INTR
) {
1833 if (sample
->intr_regs
.abi
) {
1834 result
+= sizeof(u64
);
1835 sz
= hweight_long(sample
->intr_regs
.mask
) * sizeof(u64
);
1838 result
+= sizeof(u64
);
1845 int perf_event__synthesize_sample(union perf_event
*event
, u64 type
,
1847 const struct perf_sample
*sample
,
1853 * used for cross-endian analysis. See git commit 65014ab3
1854 * for why this goofiness is needed.
1858 array
= event
->sample
.array
;
1860 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1861 *array
= sample
->id
;
1865 if (type
& PERF_SAMPLE_IP
) {
1866 *array
= sample
->ip
;
1870 if (type
& PERF_SAMPLE_TID
) {
1871 u
.val32
[0] = sample
->pid
;
1872 u
.val32
[1] = sample
->tid
;
1875 * Inverse of what is done in perf_evsel__parse_sample
1877 u
.val32
[0] = bswap_32(u
.val32
[0]);
1878 u
.val32
[1] = bswap_32(u
.val32
[1]);
1879 u
.val64
= bswap_64(u
.val64
);
1886 if (type
& PERF_SAMPLE_TIME
) {
1887 *array
= sample
->time
;
1891 if (type
& PERF_SAMPLE_ADDR
) {
1892 *array
= sample
->addr
;
1896 if (type
& PERF_SAMPLE_ID
) {
1897 *array
= sample
->id
;
1901 if (type
& PERF_SAMPLE_STREAM_ID
) {
1902 *array
= sample
->stream_id
;
1906 if (type
& PERF_SAMPLE_CPU
) {
1907 u
.val32
[0] = sample
->cpu
;
1910 * Inverse of what is done in perf_evsel__parse_sample
1912 u
.val32
[0] = bswap_32(u
.val32
[0]);
1913 u
.val64
= bswap_64(u
.val64
);
1919 if (type
& PERF_SAMPLE_PERIOD
) {
1920 *array
= sample
->period
;
1924 if (type
& PERF_SAMPLE_READ
) {
1925 if (read_format
& PERF_FORMAT_GROUP
)
1926 *array
= sample
->read
.group
.nr
;
1928 *array
= sample
->read
.one
.value
;
1931 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1932 *array
= sample
->read
.time_enabled
;
1936 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1937 *array
= sample
->read
.time_running
;
1941 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1942 if (read_format
& PERF_FORMAT_GROUP
) {
1943 sz
= sample
->read
.group
.nr
*
1944 sizeof(struct sample_read_value
);
1945 memcpy(array
, sample
->read
.group
.values
, sz
);
1946 array
= (void *)array
+ sz
;
1948 *array
= sample
->read
.one
.id
;
1953 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1954 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1955 memcpy(array
, sample
->callchain
, sz
);
1956 array
= (void *)array
+ sz
;
1959 if (type
& PERF_SAMPLE_RAW
) {
1960 u
.val32
[0] = sample
->raw_size
;
1961 if (WARN_ONCE(swapped
,
1962 "Endianness of raw data not corrected!\n")) {
1964 * Inverse of what is done in perf_evsel__parse_sample
1966 u
.val32
[0] = bswap_32(u
.val32
[0]);
1967 u
.val32
[1] = bswap_32(u
.val32
[1]);
1968 u
.val64
= bswap_64(u
.val64
);
1971 array
= (void *)array
+ sizeof(u32
);
1973 memcpy(array
, sample
->raw_data
, sample
->raw_size
);
1974 array
= (void *)array
+ sample
->raw_size
;
1977 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1978 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1980 memcpy(array
, sample
->branch_stack
, sz
);
1981 array
= (void *)array
+ sz
;
1984 if (type
& PERF_SAMPLE_REGS_USER
) {
1985 if (sample
->user_regs
.abi
) {
1986 *array
++ = sample
->user_regs
.abi
;
1987 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
1988 memcpy(array
, sample
->user_regs
.regs
, sz
);
1989 array
= (void *)array
+ sz
;
1995 if (type
& PERF_SAMPLE_STACK_USER
) {
1996 sz
= sample
->user_stack
.size
;
1999 memcpy(array
, sample
->user_stack
.data
, sz
);
2000 array
= (void *)array
+ sz
;
2005 if (type
& PERF_SAMPLE_WEIGHT
) {
2006 *array
= sample
->weight
;
2010 if (type
& PERF_SAMPLE_DATA_SRC
) {
2011 *array
= sample
->data_src
;
2015 if (type
& PERF_SAMPLE_TRANSACTION
) {
2016 *array
= sample
->transaction
;
2020 if (type
& PERF_SAMPLE_REGS_INTR
) {
2021 if (sample
->intr_regs
.abi
) {
2022 *array
++ = sample
->intr_regs
.abi
;
2023 sz
= hweight_long(sample
->intr_regs
.mask
) * sizeof(u64
);
2024 memcpy(array
, sample
->intr_regs
.regs
, sz
);
2025 array
= (void *)array
+ sz
;
2034 struct format_field
*perf_evsel__field(struct perf_evsel
*evsel
, const char *name
)
2036 return pevent_find_field(evsel
->tp_format
, name
);
2039 void *perf_evsel__rawptr(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
2042 struct format_field
*field
= perf_evsel__field(evsel
, name
);
2048 offset
= field
->offset
;
2050 if (field
->flags
& FIELD_IS_DYNAMIC
) {
2051 offset
= *(int *)(sample
->raw_data
+ field
->offset
);
2055 return sample
->raw_data
+ offset
;
2058 u64
perf_evsel__intval(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
2061 struct format_field
*field
= perf_evsel__field(evsel
, name
);
2068 ptr
= sample
->raw_data
+ field
->offset
;
2070 switch (field
->size
) {
2074 value
= *(u16
*)ptr
;
2077 value
= *(u32
*)ptr
;
2080 memcpy(&value
, ptr
, sizeof(u64
));
2086 if (!evsel
->needs_swap
)
2089 switch (field
->size
) {
2091 return bswap_16(value
);
2093 return bswap_32(value
);
2095 return bswap_64(value
);
2103 static int comma_fprintf(FILE *fp
, bool *first
, const char *fmt
, ...)
2109 ret
+= fprintf(fp
, ",");
2111 ret
+= fprintf(fp
, ":");
2115 va_start(args
, fmt
);
2116 ret
+= vfprintf(fp
, fmt
, args
);
2121 static int __print_attr__fprintf(FILE *fp
, const char *name
, const char *val
, void *priv
)
2123 return comma_fprintf(fp
, (bool *)priv
, " %s: %s", name
, val
);
2126 int perf_evsel__fprintf(struct perf_evsel
*evsel
,
2127 struct perf_attr_details
*details
, FILE *fp
)
2132 if (details
->event_group
) {
2133 struct perf_evsel
*pos
;
2135 if (!perf_evsel__is_group_leader(evsel
))
2138 if (evsel
->nr_members
> 1)
2139 printed
+= fprintf(fp
, "%s{", evsel
->group_name
?: "");
2141 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
2142 for_each_group_member(pos
, evsel
)
2143 printed
+= fprintf(fp
, ",%s", perf_evsel__name(pos
));
2145 if (evsel
->nr_members
> 1)
2146 printed
+= fprintf(fp
, "}");
2150 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
2152 if (details
->verbose
) {
2153 printed
+= perf_event_attr__fprintf(fp
, &evsel
->attr
,
2154 __print_attr__fprintf
, &first
);
2155 } else if (details
->freq
) {
2156 printed
+= comma_fprintf(fp
, &first
, " sample_freq=%" PRIu64
,
2157 (u64
)evsel
->attr
.sample_freq
);
2164 bool perf_evsel__fallback(struct perf_evsel
*evsel
, int err
,
2165 char *msg
, size_t msgsize
)
2167 if ((err
== ENOENT
|| err
== ENXIO
|| err
== ENODEV
) &&
2168 evsel
->attr
.type
== PERF_TYPE_HARDWARE
&&
2169 evsel
->attr
.config
== PERF_COUNT_HW_CPU_CYCLES
) {
2171 * If it's cycles then fall back to hrtimer based
2172 * cpu-clock-tick sw counter, which is always available even if
2175 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2178 scnprintf(msg
, msgsize
, "%s",
2179 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2181 evsel
->attr
.type
= PERF_TYPE_SOFTWARE
;
2182 evsel
->attr
.config
= PERF_COUNT_SW_CPU_CLOCK
;
2184 zfree(&evsel
->name
);
2191 int perf_evsel__open_strerror(struct perf_evsel
*evsel
, struct target
*target
,
2192 int err
, char *msg
, size_t size
)
2194 char sbuf
[STRERR_BUFSIZE
];
2199 return scnprintf(msg
, size
,
2200 "You may not have permission to collect %sstats.\n"
2201 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2202 " -1 - Not paranoid at all\n"
2203 " 0 - Disallow raw tracepoint access for unpriv\n"
2204 " 1 - Disallow cpu events for unpriv\n"
2205 " 2 - Disallow kernel profiling for unpriv",
2206 target
->system_wide
? "system-wide " : "");
2208 return scnprintf(msg
, size
, "The %s event is not supported.",
2209 perf_evsel__name(evsel
));
2211 return scnprintf(msg
, size
, "%s",
2212 "Too many events are opened.\n"
2213 "Probably the maximum number of open file descriptors has been reached.\n"
2214 "Hint: Try again after reducing the number of events.\n"
2215 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2217 if (target
->cpu_list
)
2218 return scnprintf(msg
, size
, "%s",
2219 "No such device - did you specify an out-of-range profile CPU?\n");
2222 if (evsel
->attr
.precise_ip
)
2223 return scnprintf(msg
, size
, "%s",
2224 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2225 #if defined(__i386__) || defined(__x86_64__)
2226 if (evsel
->attr
.type
== PERF_TYPE_HARDWARE
)
2227 return scnprintf(msg
, size
, "%s",
2228 "No hardware sampling interrupt available.\n"
2229 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2233 if (find_process("oprofiled"))
2234 return scnprintf(msg
, size
,
2235 "The PMU counters are busy/taken by another profiler.\n"
2236 "We found oprofile daemon running, please stop it and try again.");
2239 if (perf_missing_features
.clockid
)
2240 return scnprintf(msg
, size
, "clockid feature not supported.");
2241 if (perf_missing_features
.clockid_wrong
)
2242 return scnprintf(msg
, size
, "wrong clockid (%d).", clockid
);
2248 return scnprintf(msg
, size
,
2249 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2250 "/bin/dmesg may provide additional information.\n"
2251 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2252 err
, strerror_r(err
, sbuf
, sizeof(sbuf
)),
2253 perf_evsel__name(evsel
));