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>
22 #include "thread_map.h"
24 #include "perf_regs.h"
26 #include "trace-event.h"
33 } perf_missing_features
;
35 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
37 int __perf_evsel__sample_size(u64 sample_type
)
39 u64 mask
= sample_type
& PERF_SAMPLE_MASK
;
43 for (i
= 0; i
< 64; i
++) {
44 if (mask
& (1ULL << i
))
54 * __perf_evsel__calc_id_pos - calculate id_pos.
55 * @sample_type: sample type
57 * This function returns the position of the event id (PERF_SAMPLE_ID or
58 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
61 static int __perf_evsel__calc_id_pos(u64 sample_type
)
65 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
68 if (!(sample_type
& PERF_SAMPLE_ID
))
71 if (sample_type
& PERF_SAMPLE_IP
)
74 if (sample_type
& PERF_SAMPLE_TID
)
77 if (sample_type
& PERF_SAMPLE_TIME
)
80 if (sample_type
& PERF_SAMPLE_ADDR
)
87 * __perf_evsel__calc_is_pos - calculate is_pos.
88 * @sample_type: sample type
90 * This function returns the position (counting backwards) of the event id
91 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
92 * sample_id_all is used there is an id sample appended to non-sample events.
94 static int __perf_evsel__calc_is_pos(u64 sample_type
)
98 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
101 if (!(sample_type
& PERF_SAMPLE_ID
))
104 if (sample_type
& PERF_SAMPLE_CPU
)
107 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
113 void perf_evsel__calc_id_pos(struct perf_evsel
*evsel
)
115 evsel
->id_pos
= __perf_evsel__calc_id_pos(evsel
->attr
.sample_type
);
116 evsel
->is_pos
= __perf_evsel__calc_is_pos(evsel
->attr
.sample_type
);
119 void hists__init(struct hists
*hists
)
121 memset(hists
, 0, sizeof(*hists
));
122 hists
->entries_in_array
[0] = hists
->entries_in_array
[1] = RB_ROOT
;
123 hists
->entries_in
= &hists
->entries_in_array
[0];
124 hists
->entries_collapsed
= RB_ROOT
;
125 hists
->entries
= RB_ROOT
;
126 pthread_mutex_init(&hists
->lock
, NULL
);
129 void __perf_evsel__set_sample_bit(struct perf_evsel
*evsel
,
130 enum perf_event_sample_format bit
)
132 if (!(evsel
->attr
.sample_type
& bit
)) {
133 evsel
->attr
.sample_type
|= bit
;
134 evsel
->sample_size
+= sizeof(u64
);
135 perf_evsel__calc_id_pos(evsel
);
139 void __perf_evsel__reset_sample_bit(struct perf_evsel
*evsel
,
140 enum perf_event_sample_format bit
)
142 if (evsel
->attr
.sample_type
& bit
) {
143 evsel
->attr
.sample_type
&= ~bit
;
144 evsel
->sample_size
-= sizeof(u64
);
145 perf_evsel__calc_id_pos(evsel
);
149 void perf_evsel__set_sample_id(struct perf_evsel
*evsel
,
150 bool can_sample_identifier
)
152 if (can_sample_identifier
) {
153 perf_evsel__reset_sample_bit(evsel
, ID
);
154 perf_evsel__set_sample_bit(evsel
, IDENTIFIER
);
156 perf_evsel__set_sample_bit(evsel
, ID
);
158 evsel
->attr
.read_format
|= PERF_FORMAT_ID
;
161 void perf_evsel__init(struct perf_evsel
*evsel
,
162 struct perf_event_attr
*attr
, int idx
)
166 evsel
->leader
= evsel
;
169 INIT_LIST_HEAD(&evsel
->node
);
170 hists__init(&evsel
->hists
);
171 evsel
->sample_size
= __perf_evsel__sample_size(attr
->sample_type
);
172 perf_evsel__calc_id_pos(evsel
);
175 struct perf_evsel
*perf_evsel__new_idx(struct perf_event_attr
*attr
, int idx
)
177 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
180 perf_evsel__init(evsel
, attr
, idx
);
185 struct perf_evsel
*perf_evsel__newtp_idx(const char *sys
, const char *name
, int idx
)
187 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
190 struct perf_event_attr attr
= {
191 .type
= PERF_TYPE_TRACEPOINT
,
192 .sample_type
= (PERF_SAMPLE_RAW
| PERF_SAMPLE_TIME
|
193 PERF_SAMPLE_CPU
| PERF_SAMPLE_PERIOD
),
196 if (asprintf(&evsel
->name
, "%s:%s", sys
, name
) < 0)
199 evsel
->tp_format
= trace_event__tp_format(sys
, name
);
200 if (evsel
->tp_format
== NULL
)
203 event_attr_init(&attr
);
204 attr
.config
= evsel
->tp_format
->id
;
205 attr
.sample_period
= 1;
206 perf_evsel__init(evsel
, &attr
, idx
);
217 const char *perf_evsel__hw_names
[PERF_COUNT_HW_MAX
] = {
225 "stalled-cycles-frontend",
226 "stalled-cycles-backend",
230 static const char *__perf_evsel__hw_name(u64 config
)
232 if (config
< PERF_COUNT_HW_MAX
&& perf_evsel__hw_names
[config
])
233 return perf_evsel__hw_names
[config
];
235 return "unknown-hardware";
238 static int perf_evsel__add_modifiers(struct perf_evsel
*evsel
, char *bf
, size_t size
)
240 int colon
= 0, r
= 0;
241 struct perf_event_attr
*attr
= &evsel
->attr
;
242 bool exclude_guest_default
= false;
244 #define MOD_PRINT(context, mod) do { \
245 if (!attr->exclude_##context) { \
246 if (!colon) colon = ++r; \
247 r += scnprintf(bf + r, size - r, "%c", mod); \
250 if (attr
->exclude_kernel
|| attr
->exclude_user
|| attr
->exclude_hv
) {
251 MOD_PRINT(kernel
, 'k');
252 MOD_PRINT(user
, 'u');
254 exclude_guest_default
= true;
257 if (attr
->precise_ip
) {
260 r
+= scnprintf(bf
+ r
, size
- r
, "%.*s", attr
->precise_ip
, "ppp");
261 exclude_guest_default
= true;
264 if (attr
->exclude_host
|| attr
->exclude_guest
== exclude_guest_default
) {
265 MOD_PRINT(host
, 'H');
266 MOD_PRINT(guest
, 'G');
274 static int perf_evsel__hw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
276 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__hw_name(evsel
->attr
.config
));
277 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
280 const char *perf_evsel__sw_names
[PERF_COUNT_SW_MAX
] = {
293 static const char *__perf_evsel__sw_name(u64 config
)
295 if (config
< PERF_COUNT_SW_MAX
&& perf_evsel__sw_names
[config
])
296 return perf_evsel__sw_names
[config
];
297 return "unknown-software";
300 static int perf_evsel__sw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
302 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__sw_name(evsel
->attr
.config
));
303 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
306 static int __perf_evsel__bp_name(char *bf
, size_t size
, u64 addr
, u64 type
)
310 r
= scnprintf(bf
, size
, "mem:0x%" PRIx64
":", addr
);
312 if (type
& HW_BREAKPOINT_R
)
313 r
+= scnprintf(bf
+ r
, size
- r
, "r");
315 if (type
& HW_BREAKPOINT_W
)
316 r
+= scnprintf(bf
+ r
, size
- r
, "w");
318 if (type
& HW_BREAKPOINT_X
)
319 r
+= scnprintf(bf
+ r
, size
- r
, "x");
324 static int perf_evsel__bp_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
326 struct perf_event_attr
*attr
= &evsel
->attr
;
327 int r
= __perf_evsel__bp_name(bf
, size
, attr
->bp_addr
, attr
->bp_type
);
328 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
331 const char *perf_evsel__hw_cache
[PERF_COUNT_HW_CACHE_MAX
]
332 [PERF_EVSEL__MAX_ALIASES
] = {
333 { "L1-dcache", "l1-d", "l1d", "L1-data", },
334 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
336 { "dTLB", "d-tlb", "Data-TLB", },
337 { "iTLB", "i-tlb", "Instruction-TLB", },
338 { "branch", "branches", "bpu", "btb", "bpc", },
342 const char *perf_evsel__hw_cache_op
[PERF_COUNT_HW_CACHE_OP_MAX
]
343 [PERF_EVSEL__MAX_ALIASES
] = {
344 { "load", "loads", "read", },
345 { "store", "stores", "write", },
346 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
349 const char *perf_evsel__hw_cache_result
[PERF_COUNT_HW_CACHE_RESULT_MAX
]
350 [PERF_EVSEL__MAX_ALIASES
] = {
351 { "refs", "Reference", "ops", "access", },
352 { "misses", "miss", },
355 #define C(x) PERF_COUNT_HW_CACHE_##x
356 #define CACHE_READ (1 << C(OP_READ))
357 #define CACHE_WRITE (1 << C(OP_WRITE))
358 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
359 #define COP(x) (1 << x)
362 * cache operartion stat
363 * L1I : Read and prefetch only
364 * ITLB and BPU : Read-only
366 static unsigned long perf_evsel__hw_cache_stat
[C(MAX
)] = {
367 [C(L1D
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
368 [C(L1I
)] = (CACHE_READ
| CACHE_PREFETCH
),
369 [C(LL
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
370 [C(DTLB
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
371 [C(ITLB
)] = (CACHE_READ
),
372 [C(BPU
)] = (CACHE_READ
),
373 [C(NODE
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
376 bool perf_evsel__is_cache_op_valid(u8 type
, u8 op
)
378 if (perf_evsel__hw_cache_stat
[type
] & COP(op
))
379 return true; /* valid */
381 return false; /* invalid */
384 int __perf_evsel__hw_cache_type_op_res_name(u8 type
, u8 op
, u8 result
,
385 char *bf
, size_t size
)
388 return scnprintf(bf
, size
, "%s-%s-%s", perf_evsel__hw_cache
[type
][0],
389 perf_evsel__hw_cache_op
[op
][0],
390 perf_evsel__hw_cache_result
[result
][0]);
393 return scnprintf(bf
, size
, "%s-%s", perf_evsel__hw_cache
[type
][0],
394 perf_evsel__hw_cache_op
[op
][1]);
397 static int __perf_evsel__hw_cache_name(u64 config
, char *bf
, size_t size
)
399 u8 op
, result
, type
= (config
>> 0) & 0xff;
400 const char *err
= "unknown-ext-hardware-cache-type";
402 if (type
> PERF_COUNT_HW_CACHE_MAX
)
405 op
= (config
>> 8) & 0xff;
406 err
= "unknown-ext-hardware-cache-op";
407 if (op
> PERF_COUNT_HW_CACHE_OP_MAX
)
410 result
= (config
>> 16) & 0xff;
411 err
= "unknown-ext-hardware-cache-result";
412 if (result
> PERF_COUNT_HW_CACHE_RESULT_MAX
)
415 err
= "invalid-cache";
416 if (!perf_evsel__is_cache_op_valid(type
, op
))
419 return __perf_evsel__hw_cache_type_op_res_name(type
, op
, result
, bf
, size
);
421 return scnprintf(bf
, size
, "%s", err
);
424 static int perf_evsel__hw_cache_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
426 int ret
= __perf_evsel__hw_cache_name(evsel
->attr
.config
, bf
, size
);
427 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
430 static int perf_evsel__raw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
432 int ret
= scnprintf(bf
, size
, "raw 0x%" PRIx64
, evsel
->attr
.config
);
433 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
436 const char *perf_evsel__name(struct perf_evsel
*evsel
)
443 switch (evsel
->attr
.type
) {
445 perf_evsel__raw_name(evsel
, bf
, sizeof(bf
));
448 case PERF_TYPE_HARDWARE
:
449 perf_evsel__hw_name(evsel
, bf
, sizeof(bf
));
452 case PERF_TYPE_HW_CACHE
:
453 perf_evsel__hw_cache_name(evsel
, bf
, sizeof(bf
));
456 case PERF_TYPE_SOFTWARE
:
457 perf_evsel__sw_name(evsel
, bf
, sizeof(bf
));
460 case PERF_TYPE_TRACEPOINT
:
461 scnprintf(bf
, sizeof(bf
), "%s", "unknown tracepoint");
464 case PERF_TYPE_BREAKPOINT
:
465 perf_evsel__bp_name(evsel
, bf
, sizeof(bf
));
469 scnprintf(bf
, sizeof(bf
), "unknown attr type: %d",
474 evsel
->name
= strdup(bf
);
476 return evsel
->name
?: "unknown";
479 const char *perf_evsel__group_name(struct perf_evsel
*evsel
)
481 return evsel
->group_name
?: "anon group";
484 int perf_evsel__group_desc(struct perf_evsel
*evsel
, char *buf
, size_t size
)
487 struct perf_evsel
*pos
;
488 const char *group_name
= perf_evsel__group_name(evsel
);
490 ret
= scnprintf(buf
, size
, "%s", group_name
);
492 ret
+= scnprintf(buf
+ ret
, size
- ret
, " { %s",
493 perf_evsel__name(evsel
));
495 for_each_group_member(pos
, evsel
)
496 ret
+= scnprintf(buf
+ ret
, size
- ret
, ", %s",
497 perf_evsel__name(pos
));
499 ret
+= scnprintf(buf
+ ret
, size
- ret
, " }");
505 perf_evsel__config_callgraph(struct perf_evsel
*evsel
,
506 struct record_opts
*opts
)
508 bool function
= perf_evsel__is_function_event(evsel
);
509 struct perf_event_attr
*attr
= &evsel
->attr
;
511 perf_evsel__set_sample_bit(evsel
, CALLCHAIN
);
513 if (opts
->call_graph
== CALLCHAIN_DWARF
) {
515 perf_evsel__set_sample_bit(evsel
, REGS_USER
);
516 perf_evsel__set_sample_bit(evsel
, STACK_USER
);
517 attr
->sample_regs_user
= PERF_REGS_MASK
;
518 attr
->sample_stack_user
= opts
->stack_dump_size
;
519 attr
->exclude_callchain_user
= 1;
521 pr_info("Cannot use DWARF unwind for function trace event,"
522 " falling back to framepointers.\n");
527 pr_info("Disabling user space callchains for function trace event.\n");
528 attr
->exclude_callchain_user
= 1;
533 * The enable_on_exec/disabled value strategy:
535 * 1) For any type of traced program:
536 * - all independent events and group leaders are disabled
537 * - all group members are enabled
539 * Group members are ruled by group leaders. They need to
540 * be enabled, because the group scheduling relies on that.
542 * 2) For traced programs executed by perf:
543 * - all independent events and group leaders have
545 * - we don't specifically enable or disable any event during
548 * Independent events and group leaders are initially disabled
549 * and get enabled by exec. Group members are ruled by group
550 * leaders as stated in 1).
552 * 3) For traced programs attached by perf (pid/tid):
553 * - we specifically enable or disable all events during
556 * When attaching events to already running traced we
557 * enable/disable events specifically, as there's no
558 * initial traced exec call.
560 void perf_evsel__config(struct perf_evsel
*evsel
, struct record_opts
*opts
)
562 struct perf_evsel
*leader
= evsel
->leader
;
563 struct perf_event_attr
*attr
= &evsel
->attr
;
564 int track
= !evsel
->idx
; /* only the first counter needs these */
565 bool per_cpu
= opts
->target
.default_per_cpu
&& !opts
->target
.per_thread
;
567 attr
->sample_id_all
= perf_missing_features
.sample_id_all
? 0 : 1;
568 attr
->inherit
= !opts
->no_inherit
;
570 perf_evsel__set_sample_bit(evsel
, IP
);
571 perf_evsel__set_sample_bit(evsel
, TID
);
573 if (evsel
->sample_read
) {
574 perf_evsel__set_sample_bit(evsel
, READ
);
577 * We need ID even in case of single event, because
578 * PERF_SAMPLE_READ process ID specific data.
580 perf_evsel__set_sample_id(evsel
, false);
583 * Apply group format only if we belong to group
584 * with more than one members.
586 if (leader
->nr_members
> 1) {
587 attr
->read_format
|= PERF_FORMAT_GROUP
;
593 * We default some events to have a default interval. But keep
594 * it a weak assumption overridable by the user.
596 if (!attr
->sample_period
|| (opts
->user_freq
!= UINT_MAX
||
597 opts
->user_interval
!= ULLONG_MAX
)) {
599 perf_evsel__set_sample_bit(evsel
, PERIOD
);
601 attr
->sample_freq
= opts
->freq
;
603 attr
->sample_period
= opts
->default_interval
;
608 * Disable sampling for all group members other
609 * than leader in case leader 'leads' the sampling.
611 if ((leader
!= evsel
) && leader
->sample_read
) {
612 attr
->sample_freq
= 0;
613 attr
->sample_period
= 0;
616 if (opts
->no_samples
)
617 attr
->sample_freq
= 0;
619 if (opts
->inherit_stat
)
620 attr
->inherit_stat
= 1;
622 if (opts
->sample_address
) {
623 perf_evsel__set_sample_bit(evsel
, ADDR
);
624 attr
->mmap_data
= track
;
627 if (opts
->call_graph_enabled
&& !evsel
->no_aux_samples
)
628 perf_evsel__config_callgraph(evsel
, opts
);
630 if (target__has_cpu(&opts
->target
))
631 perf_evsel__set_sample_bit(evsel
, CPU
);
634 perf_evsel__set_sample_bit(evsel
, PERIOD
);
637 * When the user explicitely disabled time don't force it here.
639 if (opts
->sample_time
&&
640 (!perf_missing_features
.sample_id_all
&&
641 (!opts
->no_inherit
|| target__has_cpu(&opts
->target
) || per_cpu
)))
642 perf_evsel__set_sample_bit(evsel
, TIME
);
644 if (opts
->raw_samples
&& !evsel
->no_aux_samples
) {
645 perf_evsel__set_sample_bit(evsel
, TIME
);
646 perf_evsel__set_sample_bit(evsel
, RAW
);
647 perf_evsel__set_sample_bit(evsel
, CPU
);
650 if (opts
->sample_address
)
651 perf_evsel__set_sample_bit(evsel
, DATA_SRC
);
653 if (opts
->no_buffering
) {
655 attr
->wakeup_events
= 1;
657 if (opts
->branch_stack
&& !evsel
->no_aux_samples
) {
658 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
659 attr
->branch_sample_type
= opts
->branch_stack
;
662 if (opts
->sample_weight
)
663 perf_evsel__set_sample_bit(evsel
, WEIGHT
);
666 attr
->mmap2
= track
&& !perf_missing_features
.mmap2
;
669 if (opts
->sample_transaction
)
670 perf_evsel__set_sample_bit(evsel
, TRANSACTION
);
673 * XXX see the function comment above
675 * Disabling only independent events or group leaders,
676 * keeping group members enabled.
678 if (perf_evsel__is_group_leader(evsel
))
682 * Setting enable_on_exec for independent events and
683 * group leaders for traced executed by perf.
685 if (target__none(&opts
->target
) && perf_evsel__is_group_leader(evsel
) &&
686 !opts
->initial_delay
)
687 attr
->enable_on_exec
= 1;
689 if (evsel
->immediate
) {
691 attr
->enable_on_exec
= 0;
695 int perf_evsel__alloc_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
698 evsel
->fd
= xyarray__new(ncpus
, nthreads
, sizeof(int));
701 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
702 for (thread
= 0; thread
< nthreads
; thread
++) {
703 FD(evsel
, cpu
, thread
) = -1;
708 return evsel
->fd
!= NULL
? 0 : -ENOMEM
;
711 static int perf_evsel__run_ioctl(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
716 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
717 for (thread
= 0; thread
< nthreads
; thread
++) {
718 int fd
= FD(evsel
, cpu
, thread
),
719 err
= ioctl(fd
, ioc
, arg
);
729 int perf_evsel__set_filter(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
732 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
733 PERF_EVENT_IOC_SET_FILTER
,
737 int perf_evsel__enable(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
739 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
740 PERF_EVENT_IOC_ENABLE
,
744 int perf_evsel__alloc_id(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
746 evsel
->sample_id
= xyarray__new(ncpus
, nthreads
, sizeof(struct perf_sample_id
));
747 if (evsel
->sample_id
== NULL
)
750 evsel
->id
= zalloc(ncpus
* nthreads
* sizeof(u64
));
751 if (evsel
->id
== NULL
) {
752 xyarray__delete(evsel
->sample_id
);
753 evsel
->sample_id
= NULL
;
760 void perf_evsel__reset_counts(struct perf_evsel
*evsel
, int ncpus
)
762 memset(evsel
->counts
, 0, (sizeof(*evsel
->counts
) +
763 (ncpus
* sizeof(struct perf_counts_values
))));
766 int perf_evsel__alloc_counts(struct perf_evsel
*evsel
, int ncpus
)
768 evsel
->counts
= zalloc((sizeof(*evsel
->counts
) +
769 (ncpus
* sizeof(struct perf_counts_values
))));
770 return evsel
->counts
!= NULL
? 0 : -ENOMEM
;
773 void perf_evsel__free_fd(struct perf_evsel
*evsel
)
775 xyarray__delete(evsel
->fd
);
779 void perf_evsel__free_id(struct perf_evsel
*evsel
)
781 xyarray__delete(evsel
->sample_id
);
782 evsel
->sample_id
= NULL
;
786 void perf_evsel__close_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
790 for (cpu
= 0; cpu
< ncpus
; cpu
++)
791 for (thread
= 0; thread
< nthreads
; ++thread
) {
792 close(FD(evsel
, cpu
, thread
));
793 FD(evsel
, cpu
, thread
) = -1;
797 void perf_evsel__free_counts(struct perf_evsel
*evsel
)
799 zfree(&evsel
->counts
);
802 void perf_evsel__exit(struct perf_evsel
*evsel
)
804 assert(list_empty(&evsel
->node
));
805 perf_evsel__free_fd(evsel
);
806 perf_evsel__free_id(evsel
);
809 void perf_evsel__delete(struct perf_evsel
*evsel
)
811 perf_evsel__exit(evsel
);
812 close_cgroup(evsel
->cgrp
);
813 zfree(&evsel
->group_name
);
814 if (evsel
->tp_format
)
815 pevent_free_format(evsel
->tp_format
);
820 static inline void compute_deltas(struct perf_evsel
*evsel
,
822 struct perf_counts_values
*count
)
824 struct perf_counts_values tmp
;
826 if (!evsel
->prev_raw_counts
)
830 tmp
= evsel
->prev_raw_counts
->aggr
;
831 evsel
->prev_raw_counts
->aggr
= *count
;
833 tmp
= evsel
->prev_raw_counts
->cpu
[cpu
];
834 evsel
->prev_raw_counts
->cpu
[cpu
] = *count
;
837 count
->val
= count
->val
- tmp
.val
;
838 count
->ena
= count
->ena
- tmp
.ena
;
839 count
->run
= count
->run
- tmp
.run
;
842 int __perf_evsel__read_on_cpu(struct perf_evsel
*evsel
,
843 int cpu
, int thread
, bool scale
)
845 struct perf_counts_values count
;
846 size_t nv
= scale
? 3 : 1;
848 if (FD(evsel
, cpu
, thread
) < 0)
851 if (evsel
->counts
== NULL
&& perf_evsel__alloc_counts(evsel
, cpu
+ 1) < 0)
854 if (readn(FD(evsel
, cpu
, thread
), &count
, nv
* sizeof(u64
)) < 0)
857 compute_deltas(evsel
, cpu
, &count
);
862 else if (count
.run
< count
.ena
)
863 count
.val
= (u64
)((double)count
.val
* count
.ena
/ count
.run
+ 0.5);
865 count
.ena
= count
.run
= 0;
867 evsel
->counts
->cpu
[cpu
] = count
;
871 int __perf_evsel__read(struct perf_evsel
*evsel
,
872 int ncpus
, int nthreads
, bool scale
)
874 size_t nv
= scale
? 3 : 1;
876 struct perf_counts_values
*aggr
= &evsel
->counts
->aggr
, count
;
878 aggr
->val
= aggr
->ena
= aggr
->run
= 0;
880 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
881 for (thread
= 0; thread
< nthreads
; thread
++) {
882 if (FD(evsel
, cpu
, thread
) < 0)
885 if (readn(FD(evsel
, cpu
, thread
),
886 &count
, nv
* sizeof(u64
)) < 0)
889 aggr
->val
+= count
.val
;
891 aggr
->ena
+= count
.ena
;
892 aggr
->run
+= count
.run
;
897 compute_deltas(evsel
, -1, aggr
);
899 evsel
->counts
->scaled
= 0;
901 if (aggr
->run
== 0) {
902 evsel
->counts
->scaled
= -1;
907 if (aggr
->run
< aggr
->ena
) {
908 evsel
->counts
->scaled
= 1;
909 aggr
->val
= (u64
)((double)aggr
->val
* aggr
->ena
/ aggr
->run
+ 0.5);
912 aggr
->ena
= aggr
->run
= 0;
917 static int get_group_fd(struct perf_evsel
*evsel
, int cpu
, int thread
)
919 struct perf_evsel
*leader
= evsel
->leader
;
922 if (perf_evsel__is_group_leader(evsel
))
926 * Leader must be already processed/open,
931 fd
= FD(leader
, cpu
, thread
);
937 #define __PRINT_ATTR(fmt, cast, field) \
938 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
940 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
941 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
942 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
943 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
945 #define PRINT_ATTR2N(name1, field1, name2, field2) \
946 fprintf(fp, " %-19s %u %-19s %u\n", \
947 name1, attr->field1, name2, attr->field2)
949 #define PRINT_ATTR2(field1, field2) \
950 PRINT_ATTR2N(#field1, field1, #field2, field2)
952 static size_t perf_event_attr__fprintf(struct perf_event_attr
*attr
, FILE *fp
)
956 ret
+= fprintf(fp
, "%.60s\n", graph_dotted_line
);
957 ret
+= fprintf(fp
, "perf_event_attr:\n");
959 ret
+= PRINT_ATTR_U32(type
);
960 ret
+= PRINT_ATTR_U32(size
);
961 ret
+= PRINT_ATTR_X64(config
);
962 ret
+= PRINT_ATTR_U64(sample_period
);
963 ret
+= PRINT_ATTR_U64(sample_freq
);
964 ret
+= PRINT_ATTR_X64(sample_type
);
965 ret
+= PRINT_ATTR_X64(read_format
);
967 ret
+= PRINT_ATTR2(disabled
, inherit
);
968 ret
+= PRINT_ATTR2(pinned
, exclusive
);
969 ret
+= PRINT_ATTR2(exclude_user
, exclude_kernel
);
970 ret
+= PRINT_ATTR2(exclude_hv
, exclude_idle
);
971 ret
+= PRINT_ATTR2(mmap
, comm
);
972 ret
+= PRINT_ATTR2(mmap2
, comm_exec
);
973 ret
+= PRINT_ATTR2(freq
, inherit_stat
);
974 ret
+= PRINT_ATTR2(enable_on_exec
, task
);
975 ret
+= PRINT_ATTR2(watermark
, precise_ip
);
976 ret
+= PRINT_ATTR2(mmap_data
, sample_id_all
);
977 ret
+= PRINT_ATTR2(exclude_host
, exclude_guest
);
978 ret
+= PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel
,
979 "excl.callchain_user", exclude_callchain_user
);
981 ret
+= PRINT_ATTR_U32(wakeup_events
);
982 ret
+= PRINT_ATTR_U32(wakeup_watermark
);
983 ret
+= PRINT_ATTR_X32(bp_type
);
984 ret
+= PRINT_ATTR_X64(bp_addr
);
985 ret
+= PRINT_ATTR_X64(config1
);
986 ret
+= PRINT_ATTR_U64(bp_len
);
987 ret
+= PRINT_ATTR_X64(config2
);
988 ret
+= PRINT_ATTR_X64(branch_sample_type
);
989 ret
+= PRINT_ATTR_X64(sample_regs_user
);
990 ret
+= PRINT_ATTR_U32(sample_stack_user
);
992 ret
+= fprintf(fp
, "%.60s\n", graph_dotted_line
);
997 static int __perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
998 struct thread_map
*threads
)
1001 unsigned long flags
= PERF_FLAG_FD_CLOEXEC
;
1003 enum { NO_CHANGE
, SET_TO_MAX
, INCREASED_MAX
} set_rlimit
= NO_CHANGE
;
1005 if (evsel
->fd
== NULL
&&
1006 perf_evsel__alloc_fd(evsel
, cpus
->nr
, threads
->nr
) < 0)
1010 flags
|= PERF_FLAG_PID_CGROUP
;
1011 pid
= evsel
->cgrp
->fd
;
1014 fallback_missing_features
:
1015 if (perf_missing_features
.cloexec
)
1016 flags
&= ~(unsigned long)PERF_FLAG_FD_CLOEXEC
;
1017 if (perf_missing_features
.mmap2
)
1018 evsel
->attr
.mmap2
= 0;
1019 if (perf_missing_features
.exclude_guest
)
1020 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
1022 if (perf_missing_features
.sample_id_all
)
1023 evsel
->attr
.sample_id_all
= 0;
1026 perf_event_attr__fprintf(&evsel
->attr
, stderr
);
1028 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
1030 for (thread
= 0; thread
< threads
->nr
; thread
++) {
1034 pid
= threads
->map
[thread
];
1036 group_fd
= get_group_fd(evsel
, cpu
, thread
);
1038 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1039 pid
, cpus
->map
[cpu
], group_fd
, flags
);
1041 FD(evsel
, cpu
, thread
) = sys_perf_event_open(&evsel
->attr
,
1045 if (FD(evsel
, cpu
, thread
) < 0) {
1047 pr_debug2("sys_perf_event_open failed, error %d\n",
1051 set_rlimit
= NO_CHANGE
;
1059 * perf stat needs between 5 and 22 fds per CPU. When we run out
1060 * of them try to increase the limits.
1062 if (err
== -EMFILE
&& set_rlimit
< INCREASED_MAX
) {
1064 int old_errno
= errno
;
1066 if (getrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1067 if (set_rlimit
== NO_CHANGE
)
1068 l
.rlim_cur
= l
.rlim_max
;
1070 l
.rlim_cur
= l
.rlim_max
+ 1000;
1071 l
.rlim_max
= l
.rlim_cur
;
1073 if (setrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1082 if (err
!= -EINVAL
|| cpu
> 0 || thread
> 0)
1085 if (!perf_missing_features
.cloexec
&& (flags
& PERF_FLAG_FD_CLOEXEC
)) {
1086 perf_missing_features
.cloexec
= true;
1087 goto fallback_missing_features
;
1088 } else if (!perf_missing_features
.mmap2
&& evsel
->attr
.mmap2
) {
1089 perf_missing_features
.mmap2
= true;
1090 goto fallback_missing_features
;
1091 } else if (!perf_missing_features
.exclude_guest
&&
1092 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
1093 perf_missing_features
.exclude_guest
= true;
1094 goto fallback_missing_features
;
1095 } else if (!perf_missing_features
.sample_id_all
) {
1096 perf_missing_features
.sample_id_all
= true;
1097 goto retry_sample_id
;
1102 while (--thread
>= 0) {
1103 close(FD(evsel
, cpu
, thread
));
1104 FD(evsel
, cpu
, thread
) = -1;
1106 thread
= threads
->nr
;
1107 } while (--cpu
>= 0);
1111 void perf_evsel__close(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1113 if (evsel
->fd
== NULL
)
1116 perf_evsel__close_fd(evsel
, ncpus
, nthreads
);
1117 perf_evsel__free_fd(evsel
);
1129 struct thread_map map
;
1131 } empty_thread_map
= {
1136 int perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1137 struct thread_map
*threads
)
1140 /* Work around old compiler warnings about strict aliasing */
1141 cpus
= &empty_cpu_map
.map
;
1144 if (threads
== NULL
)
1145 threads
= &empty_thread_map
.map
;
1147 return __perf_evsel__open(evsel
, cpus
, threads
);
1150 int perf_evsel__open_per_cpu(struct perf_evsel
*evsel
,
1151 struct cpu_map
*cpus
)
1153 return __perf_evsel__open(evsel
, cpus
, &empty_thread_map
.map
);
1156 int perf_evsel__open_per_thread(struct perf_evsel
*evsel
,
1157 struct thread_map
*threads
)
1159 return __perf_evsel__open(evsel
, &empty_cpu_map
.map
, threads
);
1162 static int perf_evsel__parse_id_sample(const struct perf_evsel
*evsel
,
1163 const union perf_event
*event
,
1164 struct perf_sample
*sample
)
1166 u64 type
= evsel
->attr
.sample_type
;
1167 const u64
*array
= event
->sample
.array
;
1168 bool swapped
= evsel
->needs_swap
;
1171 array
+= ((event
->header
.size
-
1172 sizeof(event
->header
)) / sizeof(u64
)) - 1;
1174 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1175 sample
->id
= *array
;
1179 if (type
& PERF_SAMPLE_CPU
) {
1182 /* undo swap of u64, then swap on individual u32s */
1183 u
.val64
= bswap_64(u
.val64
);
1184 u
.val32
[0] = bswap_32(u
.val32
[0]);
1187 sample
->cpu
= u
.val32
[0];
1191 if (type
& PERF_SAMPLE_STREAM_ID
) {
1192 sample
->stream_id
= *array
;
1196 if (type
& PERF_SAMPLE_ID
) {
1197 sample
->id
= *array
;
1201 if (type
& PERF_SAMPLE_TIME
) {
1202 sample
->time
= *array
;
1206 if (type
& PERF_SAMPLE_TID
) {
1209 /* undo swap of u64, then swap on individual u32s */
1210 u
.val64
= bswap_64(u
.val64
);
1211 u
.val32
[0] = bswap_32(u
.val32
[0]);
1212 u
.val32
[1] = bswap_32(u
.val32
[1]);
1215 sample
->pid
= u
.val32
[0];
1216 sample
->tid
= u
.val32
[1];
1223 static inline bool overflow(const void *endp
, u16 max_size
, const void *offset
,
1226 return size
> max_size
|| offset
+ size
> endp
;
1229 #define OVERFLOW_CHECK(offset, size, max_size) \
1231 if (overflow(endp, (max_size), (offset), (size))) \
1235 #define OVERFLOW_CHECK_u64(offset) \
1236 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1238 int perf_evsel__parse_sample(struct perf_evsel
*evsel
, union perf_event
*event
,
1239 struct perf_sample
*data
)
1241 u64 type
= evsel
->attr
.sample_type
;
1242 bool swapped
= evsel
->needs_swap
;
1244 u16 max_size
= event
->header
.size
;
1245 const void *endp
= (void *)event
+ max_size
;
1249 * used for cross-endian analysis. See git commit 65014ab3
1250 * for why this goofiness is needed.
1254 memset(data
, 0, sizeof(*data
));
1255 data
->cpu
= data
->pid
= data
->tid
= -1;
1256 data
->stream_id
= data
->id
= data
->time
= -1ULL;
1257 data
->period
= evsel
->attr
.sample_period
;
1260 if (event
->header
.type
!= PERF_RECORD_SAMPLE
) {
1261 if (!evsel
->attr
.sample_id_all
)
1263 return perf_evsel__parse_id_sample(evsel
, event
, data
);
1266 array
= event
->sample
.array
;
1269 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1270 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1271 * check the format does not go past the end of the event.
1273 if (evsel
->sample_size
+ sizeof(event
->header
) > event
->header
.size
)
1277 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1282 if (type
& PERF_SAMPLE_IP
) {
1287 if (type
& PERF_SAMPLE_TID
) {
1290 /* undo swap of u64, then swap on individual u32s */
1291 u
.val64
= bswap_64(u
.val64
);
1292 u
.val32
[0] = bswap_32(u
.val32
[0]);
1293 u
.val32
[1] = bswap_32(u
.val32
[1]);
1296 data
->pid
= u
.val32
[0];
1297 data
->tid
= u
.val32
[1];
1301 if (type
& PERF_SAMPLE_TIME
) {
1302 data
->time
= *array
;
1307 if (type
& PERF_SAMPLE_ADDR
) {
1308 data
->addr
= *array
;
1312 if (type
& PERF_SAMPLE_ID
) {
1317 if (type
& PERF_SAMPLE_STREAM_ID
) {
1318 data
->stream_id
= *array
;
1322 if (type
& PERF_SAMPLE_CPU
) {
1326 /* undo swap of u64, then swap on individual u32s */
1327 u
.val64
= bswap_64(u
.val64
);
1328 u
.val32
[0] = bswap_32(u
.val32
[0]);
1331 data
->cpu
= u
.val32
[0];
1335 if (type
& PERF_SAMPLE_PERIOD
) {
1336 data
->period
= *array
;
1340 if (type
& PERF_SAMPLE_READ
) {
1341 u64 read_format
= evsel
->attr
.read_format
;
1343 OVERFLOW_CHECK_u64(array
);
1344 if (read_format
& PERF_FORMAT_GROUP
)
1345 data
->read
.group
.nr
= *array
;
1347 data
->read
.one
.value
= *array
;
1351 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1352 OVERFLOW_CHECK_u64(array
);
1353 data
->read
.time_enabled
= *array
;
1357 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1358 OVERFLOW_CHECK_u64(array
);
1359 data
->read
.time_running
= *array
;
1363 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1364 if (read_format
& PERF_FORMAT_GROUP
) {
1365 const u64 max_group_nr
= UINT64_MAX
/
1366 sizeof(struct sample_read_value
);
1368 if (data
->read
.group
.nr
> max_group_nr
)
1370 sz
= data
->read
.group
.nr
*
1371 sizeof(struct sample_read_value
);
1372 OVERFLOW_CHECK(array
, sz
, max_size
);
1373 data
->read
.group
.values
=
1374 (struct sample_read_value
*)array
;
1375 array
= (void *)array
+ sz
;
1377 OVERFLOW_CHECK_u64(array
);
1378 data
->read
.one
.id
= *array
;
1383 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1384 const u64 max_callchain_nr
= UINT64_MAX
/ sizeof(u64
);
1386 OVERFLOW_CHECK_u64(array
);
1387 data
->callchain
= (struct ip_callchain
*)array
++;
1388 if (data
->callchain
->nr
> max_callchain_nr
)
1390 sz
= data
->callchain
->nr
* sizeof(u64
);
1391 OVERFLOW_CHECK(array
, sz
, max_size
);
1392 array
= (void *)array
+ sz
;
1395 if (type
& PERF_SAMPLE_RAW
) {
1396 OVERFLOW_CHECK_u64(array
);
1398 if (WARN_ONCE(swapped
,
1399 "Endianness of raw data not corrected!\n")) {
1400 /* undo swap of u64, then swap on individual u32s */
1401 u
.val64
= bswap_64(u
.val64
);
1402 u
.val32
[0] = bswap_32(u
.val32
[0]);
1403 u
.val32
[1] = bswap_32(u
.val32
[1]);
1405 data
->raw_size
= u
.val32
[0];
1406 array
= (void *)array
+ sizeof(u32
);
1408 OVERFLOW_CHECK(array
, data
->raw_size
, max_size
);
1409 data
->raw_data
= (void *)array
;
1410 array
= (void *)array
+ data
->raw_size
;
1413 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1414 const u64 max_branch_nr
= UINT64_MAX
/
1415 sizeof(struct branch_entry
);
1417 OVERFLOW_CHECK_u64(array
);
1418 data
->branch_stack
= (struct branch_stack
*)array
++;
1420 if (data
->branch_stack
->nr
> max_branch_nr
)
1422 sz
= data
->branch_stack
->nr
* sizeof(struct branch_entry
);
1423 OVERFLOW_CHECK(array
, sz
, max_size
);
1424 array
= (void *)array
+ sz
;
1427 if (type
& PERF_SAMPLE_REGS_USER
) {
1428 OVERFLOW_CHECK_u64(array
);
1429 data
->user_regs
.abi
= *array
;
1432 if (data
->user_regs
.abi
) {
1433 u64 mask
= evsel
->attr
.sample_regs_user
;
1435 sz
= hweight_long(mask
) * sizeof(u64
);
1436 OVERFLOW_CHECK(array
, sz
, max_size
);
1437 data
->user_regs
.mask
= mask
;
1438 data
->user_regs
.regs
= (u64
*)array
;
1439 array
= (void *)array
+ sz
;
1443 if (type
& PERF_SAMPLE_STACK_USER
) {
1444 OVERFLOW_CHECK_u64(array
);
1447 data
->user_stack
.offset
= ((char *)(array
- 1)
1451 data
->user_stack
.size
= 0;
1453 OVERFLOW_CHECK(array
, sz
, max_size
);
1454 data
->user_stack
.data
= (char *)array
;
1455 array
= (void *)array
+ sz
;
1456 OVERFLOW_CHECK_u64(array
);
1457 data
->user_stack
.size
= *array
++;
1458 if (WARN_ONCE(data
->user_stack
.size
> sz
,
1459 "user stack dump failure\n"))
1465 if (type
& PERF_SAMPLE_WEIGHT
) {
1466 OVERFLOW_CHECK_u64(array
);
1467 data
->weight
= *array
;
1471 data
->data_src
= PERF_MEM_DATA_SRC_NONE
;
1472 if (type
& PERF_SAMPLE_DATA_SRC
) {
1473 OVERFLOW_CHECK_u64(array
);
1474 data
->data_src
= *array
;
1478 data
->transaction
= 0;
1479 if (type
& PERF_SAMPLE_TRANSACTION
) {
1480 OVERFLOW_CHECK_u64(array
);
1481 data
->transaction
= *array
;
1488 size_t perf_event__sample_event_size(const struct perf_sample
*sample
, u64 type
,
1491 size_t sz
, result
= sizeof(struct sample_event
);
1493 if (type
& PERF_SAMPLE_IDENTIFIER
)
1494 result
+= sizeof(u64
);
1496 if (type
& PERF_SAMPLE_IP
)
1497 result
+= sizeof(u64
);
1499 if (type
& PERF_SAMPLE_TID
)
1500 result
+= sizeof(u64
);
1502 if (type
& PERF_SAMPLE_TIME
)
1503 result
+= sizeof(u64
);
1505 if (type
& PERF_SAMPLE_ADDR
)
1506 result
+= sizeof(u64
);
1508 if (type
& PERF_SAMPLE_ID
)
1509 result
+= sizeof(u64
);
1511 if (type
& PERF_SAMPLE_STREAM_ID
)
1512 result
+= sizeof(u64
);
1514 if (type
& PERF_SAMPLE_CPU
)
1515 result
+= sizeof(u64
);
1517 if (type
& PERF_SAMPLE_PERIOD
)
1518 result
+= sizeof(u64
);
1520 if (type
& PERF_SAMPLE_READ
) {
1521 result
+= sizeof(u64
);
1522 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1523 result
+= sizeof(u64
);
1524 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1525 result
+= sizeof(u64
);
1526 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1527 if (read_format
& PERF_FORMAT_GROUP
) {
1528 sz
= sample
->read
.group
.nr
*
1529 sizeof(struct sample_read_value
);
1532 result
+= sizeof(u64
);
1536 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1537 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1541 if (type
& PERF_SAMPLE_RAW
) {
1542 result
+= sizeof(u32
);
1543 result
+= sample
->raw_size
;
1546 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1547 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1552 if (type
& PERF_SAMPLE_REGS_USER
) {
1553 if (sample
->user_regs
.abi
) {
1554 result
+= sizeof(u64
);
1555 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
1558 result
+= sizeof(u64
);
1562 if (type
& PERF_SAMPLE_STACK_USER
) {
1563 sz
= sample
->user_stack
.size
;
1564 result
+= sizeof(u64
);
1567 result
+= sizeof(u64
);
1571 if (type
& PERF_SAMPLE_WEIGHT
)
1572 result
+= sizeof(u64
);
1574 if (type
& PERF_SAMPLE_DATA_SRC
)
1575 result
+= sizeof(u64
);
1577 if (type
& PERF_SAMPLE_TRANSACTION
)
1578 result
+= sizeof(u64
);
1583 int perf_event__synthesize_sample(union perf_event
*event
, u64 type
,
1585 const struct perf_sample
*sample
,
1591 * used for cross-endian analysis. See git commit 65014ab3
1592 * for why this goofiness is needed.
1596 array
= event
->sample
.array
;
1598 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1599 *array
= sample
->id
;
1603 if (type
& PERF_SAMPLE_IP
) {
1604 *array
= sample
->ip
;
1608 if (type
& PERF_SAMPLE_TID
) {
1609 u
.val32
[0] = sample
->pid
;
1610 u
.val32
[1] = sample
->tid
;
1613 * Inverse of what is done in perf_evsel__parse_sample
1615 u
.val32
[0] = bswap_32(u
.val32
[0]);
1616 u
.val32
[1] = bswap_32(u
.val32
[1]);
1617 u
.val64
= bswap_64(u
.val64
);
1624 if (type
& PERF_SAMPLE_TIME
) {
1625 *array
= sample
->time
;
1629 if (type
& PERF_SAMPLE_ADDR
) {
1630 *array
= sample
->addr
;
1634 if (type
& PERF_SAMPLE_ID
) {
1635 *array
= sample
->id
;
1639 if (type
& PERF_SAMPLE_STREAM_ID
) {
1640 *array
= sample
->stream_id
;
1644 if (type
& PERF_SAMPLE_CPU
) {
1645 u
.val32
[0] = sample
->cpu
;
1648 * Inverse of what is done in perf_evsel__parse_sample
1650 u
.val32
[0] = bswap_32(u
.val32
[0]);
1651 u
.val64
= bswap_64(u
.val64
);
1657 if (type
& PERF_SAMPLE_PERIOD
) {
1658 *array
= sample
->period
;
1662 if (type
& PERF_SAMPLE_READ
) {
1663 if (read_format
& PERF_FORMAT_GROUP
)
1664 *array
= sample
->read
.group
.nr
;
1666 *array
= sample
->read
.one
.value
;
1669 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1670 *array
= sample
->read
.time_enabled
;
1674 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1675 *array
= sample
->read
.time_running
;
1679 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1680 if (read_format
& PERF_FORMAT_GROUP
) {
1681 sz
= sample
->read
.group
.nr
*
1682 sizeof(struct sample_read_value
);
1683 memcpy(array
, sample
->read
.group
.values
, sz
);
1684 array
= (void *)array
+ sz
;
1686 *array
= sample
->read
.one
.id
;
1691 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1692 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1693 memcpy(array
, sample
->callchain
, sz
);
1694 array
= (void *)array
+ sz
;
1697 if (type
& PERF_SAMPLE_RAW
) {
1698 u
.val32
[0] = sample
->raw_size
;
1699 if (WARN_ONCE(swapped
,
1700 "Endianness of raw data not corrected!\n")) {
1702 * Inverse of what is done in perf_evsel__parse_sample
1704 u
.val32
[0] = bswap_32(u
.val32
[0]);
1705 u
.val32
[1] = bswap_32(u
.val32
[1]);
1706 u
.val64
= bswap_64(u
.val64
);
1709 array
= (void *)array
+ sizeof(u32
);
1711 memcpy(array
, sample
->raw_data
, sample
->raw_size
);
1712 array
= (void *)array
+ sample
->raw_size
;
1715 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1716 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1718 memcpy(array
, sample
->branch_stack
, sz
);
1719 array
= (void *)array
+ sz
;
1722 if (type
& PERF_SAMPLE_REGS_USER
) {
1723 if (sample
->user_regs
.abi
) {
1724 *array
++ = sample
->user_regs
.abi
;
1725 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
1726 memcpy(array
, sample
->user_regs
.regs
, sz
);
1727 array
= (void *)array
+ sz
;
1733 if (type
& PERF_SAMPLE_STACK_USER
) {
1734 sz
= sample
->user_stack
.size
;
1737 memcpy(array
, sample
->user_stack
.data
, sz
);
1738 array
= (void *)array
+ sz
;
1743 if (type
& PERF_SAMPLE_WEIGHT
) {
1744 *array
= sample
->weight
;
1748 if (type
& PERF_SAMPLE_DATA_SRC
) {
1749 *array
= sample
->data_src
;
1753 if (type
& PERF_SAMPLE_TRANSACTION
) {
1754 *array
= sample
->transaction
;
1761 struct format_field
*perf_evsel__field(struct perf_evsel
*evsel
, const char *name
)
1763 return pevent_find_field(evsel
->tp_format
, name
);
1766 void *perf_evsel__rawptr(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
1769 struct format_field
*field
= perf_evsel__field(evsel
, name
);
1775 offset
= field
->offset
;
1777 if (field
->flags
& FIELD_IS_DYNAMIC
) {
1778 offset
= *(int *)(sample
->raw_data
+ field
->offset
);
1782 return sample
->raw_data
+ offset
;
1785 u64
perf_evsel__intval(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
1788 struct format_field
*field
= perf_evsel__field(evsel
, name
);
1795 ptr
= sample
->raw_data
+ field
->offset
;
1797 switch (field
->size
) {
1801 value
= *(u16
*)ptr
;
1804 value
= *(u32
*)ptr
;
1807 value
= *(u64
*)ptr
;
1813 if (!evsel
->needs_swap
)
1816 switch (field
->size
) {
1818 return bswap_16(value
);
1820 return bswap_32(value
);
1822 return bswap_64(value
);
1830 static int comma_fprintf(FILE *fp
, bool *first
, const char *fmt
, ...)
1836 ret
+= fprintf(fp
, ",");
1838 ret
+= fprintf(fp
, ":");
1842 va_start(args
, fmt
);
1843 ret
+= vfprintf(fp
, fmt
, args
);
1848 static int __if_fprintf(FILE *fp
, bool *first
, const char *field
, u64 value
)
1853 return comma_fprintf(fp
, first
, " %s: %" PRIu64
, field
, value
);
1856 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1863 static int bits__fprintf(FILE *fp
, const char *field
, u64 value
,
1864 struct bit_names
*bits
, bool *first
)
1866 int i
= 0, printed
= comma_fprintf(fp
, first
, " %s: ", field
);
1867 bool first_bit
= true;
1870 if (value
& bits
[i
].bit
) {
1871 printed
+= fprintf(fp
, "%s%s", first_bit
? "" : "|", bits
[i
].name
);
1874 } while (bits
[++i
].name
!= NULL
);
1879 static int sample_type__fprintf(FILE *fp
, bool *first
, u64 value
)
1881 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1882 struct bit_names bits
[] = {
1883 bit_name(IP
), bit_name(TID
), bit_name(TIME
), bit_name(ADDR
),
1884 bit_name(READ
), bit_name(CALLCHAIN
), bit_name(ID
), bit_name(CPU
),
1885 bit_name(PERIOD
), bit_name(STREAM_ID
), bit_name(RAW
),
1886 bit_name(BRANCH_STACK
), bit_name(REGS_USER
), bit_name(STACK_USER
),
1887 bit_name(IDENTIFIER
),
1891 return bits__fprintf(fp
, "sample_type", value
, bits
, first
);
1894 static int read_format__fprintf(FILE *fp
, bool *first
, u64 value
)
1896 #define bit_name(n) { PERF_FORMAT_##n, #n }
1897 struct bit_names bits
[] = {
1898 bit_name(TOTAL_TIME_ENABLED
), bit_name(TOTAL_TIME_RUNNING
),
1899 bit_name(ID
), bit_name(GROUP
),
1903 return bits__fprintf(fp
, "read_format", value
, bits
, first
);
1906 int perf_evsel__fprintf(struct perf_evsel
*evsel
,
1907 struct perf_attr_details
*details
, FILE *fp
)
1912 if (details
->event_group
) {
1913 struct perf_evsel
*pos
;
1915 if (!perf_evsel__is_group_leader(evsel
))
1918 if (evsel
->nr_members
> 1)
1919 printed
+= fprintf(fp
, "%s{", evsel
->group_name
?: "");
1921 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
1922 for_each_group_member(pos
, evsel
)
1923 printed
+= fprintf(fp
, ",%s", perf_evsel__name(pos
));
1925 if (evsel
->nr_members
> 1)
1926 printed
+= fprintf(fp
, "}");
1930 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
1932 if (details
->verbose
|| details
->freq
) {
1933 printed
+= comma_fprintf(fp
, &first
, " sample_freq=%" PRIu64
,
1934 (u64
)evsel
->attr
.sample_freq
);
1937 if (details
->verbose
) {
1943 printed
+= sample_type__fprintf(fp
, &first
, evsel
->attr
.sample_type
);
1944 if (evsel
->attr
.read_format
)
1945 printed
+= read_format__fprintf(fp
, &first
, evsel
->attr
.read_format
);
1949 if_print(exclusive
);
1950 if_print(exclude_user
);
1951 if_print(exclude_kernel
);
1952 if_print(exclude_hv
);
1953 if_print(exclude_idle
);
1957 if_print(comm_exec
);
1959 if_print(inherit_stat
);
1960 if_print(enable_on_exec
);
1962 if_print(watermark
);
1963 if_print(precise_ip
);
1964 if_print(mmap_data
);
1965 if_print(sample_id_all
);
1966 if_print(exclude_host
);
1967 if_print(exclude_guest
);
1968 if_print(__reserved_1
);
1969 if_print(wakeup_events
);
1971 if_print(branch_sample_type
);
1978 bool perf_evsel__fallback(struct perf_evsel
*evsel
, int err
,
1979 char *msg
, size_t msgsize
)
1981 if ((err
== ENOENT
|| err
== ENXIO
|| err
== ENODEV
) &&
1982 evsel
->attr
.type
== PERF_TYPE_HARDWARE
&&
1983 evsel
->attr
.config
== PERF_COUNT_HW_CPU_CYCLES
) {
1985 * If it's cycles then fall back to hrtimer based
1986 * cpu-clock-tick sw counter, which is always available even if
1989 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1992 scnprintf(msg
, msgsize
, "%s",
1993 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1995 evsel
->attr
.type
= PERF_TYPE_SOFTWARE
;
1996 evsel
->attr
.config
= PERF_COUNT_SW_CPU_CLOCK
;
1998 zfree(&evsel
->name
);
2005 int perf_evsel__open_strerror(struct perf_evsel
*evsel
, struct target
*target
,
2006 int err
, char *msg
, size_t size
)
2011 return scnprintf(msg
, size
,
2012 "You may not have permission to collect %sstats.\n"
2013 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2014 " -1 - Not paranoid at all\n"
2015 " 0 - Disallow raw tracepoint access for unpriv\n"
2016 " 1 - Disallow cpu events for unpriv\n"
2017 " 2 - Disallow kernel profiling for unpriv",
2018 target
->system_wide
? "system-wide " : "");
2020 return scnprintf(msg
, size
, "The %s event is not supported.",
2021 perf_evsel__name(evsel
));
2023 return scnprintf(msg
, size
, "%s",
2024 "Too many events are opened.\n"
2025 "Try again after reducing the number of events.");
2027 if (target
->cpu_list
)
2028 return scnprintf(msg
, size
, "%s",
2029 "No such device - did you specify an out-of-range profile CPU?\n");
2032 if (evsel
->attr
.precise_ip
)
2033 return scnprintf(msg
, size
, "%s",
2034 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2035 #if defined(__i386__) || defined(__x86_64__)
2036 if (evsel
->attr
.type
== PERF_TYPE_HARDWARE
)
2037 return scnprintf(msg
, size
, "%s",
2038 "No hardware sampling interrupt available.\n"
2039 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2043 if (find_process("oprofiled"))
2044 return scnprintf(msg
, size
,
2045 "The PMU counters are busy/taken by another profiler.\n"
2046 "We found oprofile daemon running, please stop it and try again.");
2052 return scnprintf(msg
, size
,
2053 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2054 "/bin/dmesg may provide additional information.\n"
2055 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2056 err
, strerror(err
), perf_evsel__name(evsel
));