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"
32 } perf_missing_features
;
34 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
36 int __perf_evsel__sample_size(u64 sample_type
)
38 u64 mask
= sample_type
& PERF_SAMPLE_MASK
;
42 for (i
= 0; i
< 64; i
++) {
43 if (mask
& (1ULL << i
))
53 * __perf_evsel__calc_id_pos - calculate id_pos.
54 * @sample_type: sample type
56 * This function returns the position of the event id (PERF_SAMPLE_ID or
57 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
60 static int __perf_evsel__calc_id_pos(u64 sample_type
)
64 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
67 if (!(sample_type
& PERF_SAMPLE_ID
))
70 if (sample_type
& PERF_SAMPLE_IP
)
73 if (sample_type
& PERF_SAMPLE_TID
)
76 if (sample_type
& PERF_SAMPLE_TIME
)
79 if (sample_type
& PERF_SAMPLE_ADDR
)
86 * __perf_evsel__calc_is_pos - calculate is_pos.
87 * @sample_type: sample type
89 * This function returns the position (counting backwards) of the event id
90 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
91 * sample_id_all is used there is an id sample appended to non-sample events.
93 static int __perf_evsel__calc_is_pos(u64 sample_type
)
97 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
100 if (!(sample_type
& PERF_SAMPLE_ID
))
103 if (sample_type
& PERF_SAMPLE_CPU
)
106 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
112 void perf_evsel__calc_id_pos(struct perf_evsel
*evsel
)
114 evsel
->id_pos
= __perf_evsel__calc_id_pos(evsel
->attr
.sample_type
);
115 evsel
->is_pos
= __perf_evsel__calc_is_pos(evsel
->attr
.sample_type
);
118 void hists__init(struct hists
*hists
)
120 memset(hists
, 0, sizeof(*hists
));
121 hists
->entries_in_array
[0] = hists
->entries_in_array
[1] = RB_ROOT
;
122 hists
->entries_in
= &hists
->entries_in_array
[0];
123 hists
->entries_collapsed
= RB_ROOT
;
124 hists
->entries
= RB_ROOT
;
125 pthread_mutex_init(&hists
->lock
, NULL
);
128 void __perf_evsel__set_sample_bit(struct perf_evsel
*evsel
,
129 enum perf_event_sample_format bit
)
131 if (!(evsel
->attr
.sample_type
& bit
)) {
132 evsel
->attr
.sample_type
|= bit
;
133 evsel
->sample_size
+= sizeof(u64
);
134 perf_evsel__calc_id_pos(evsel
);
138 void __perf_evsel__reset_sample_bit(struct perf_evsel
*evsel
,
139 enum perf_event_sample_format bit
)
141 if (evsel
->attr
.sample_type
& bit
) {
142 evsel
->attr
.sample_type
&= ~bit
;
143 evsel
->sample_size
-= sizeof(u64
);
144 perf_evsel__calc_id_pos(evsel
);
148 void perf_evsel__set_sample_id(struct perf_evsel
*evsel
,
149 bool can_sample_identifier
)
151 if (can_sample_identifier
) {
152 perf_evsel__reset_sample_bit(evsel
, ID
);
153 perf_evsel__set_sample_bit(evsel
, IDENTIFIER
);
155 perf_evsel__set_sample_bit(evsel
, ID
);
157 evsel
->attr
.read_format
|= PERF_FORMAT_ID
;
160 void perf_evsel__init(struct perf_evsel
*evsel
,
161 struct perf_event_attr
*attr
, int idx
)
165 evsel
->leader
= evsel
;
168 INIT_LIST_HEAD(&evsel
->node
);
169 hists__init(&evsel
->hists
);
170 evsel
->sample_size
= __perf_evsel__sample_size(attr
->sample_type
);
171 perf_evsel__calc_id_pos(evsel
);
174 struct perf_evsel
*perf_evsel__new_idx(struct perf_event_attr
*attr
, int idx
)
176 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
179 perf_evsel__init(evsel
, attr
, idx
);
184 struct perf_evsel
*perf_evsel__newtp_idx(const char *sys
, const char *name
, int idx
)
186 struct perf_evsel
*evsel
= zalloc(sizeof(*evsel
));
189 struct perf_event_attr attr
= {
190 .type
= PERF_TYPE_TRACEPOINT
,
191 .sample_type
= (PERF_SAMPLE_RAW
| PERF_SAMPLE_TIME
|
192 PERF_SAMPLE_CPU
| PERF_SAMPLE_PERIOD
),
195 if (asprintf(&evsel
->name
, "%s:%s", sys
, name
) < 0)
198 evsel
->tp_format
= trace_event__tp_format(sys
, name
);
199 if (evsel
->tp_format
== NULL
)
202 event_attr_init(&attr
);
203 attr
.config
= evsel
->tp_format
->id
;
204 attr
.sample_period
= 1;
205 perf_evsel__init(evsel
, &attr
, idx
);
216 const char *perf_evsel__hw_names
[PERF_COUNT_HW_MAX
] = {
224 "stalled-cycles-frontend",
225 "stalled-cycles-backend",
229 static const char *__perf_evsel__hw_name(u64 config
)
231 if (config
< PERF_COUNT_HW_MAX
&& perf_evsel__hw_names
[config
])
232 return perf_evsel__hw_names
[config
];
234 return "unknown-hardware";
237 static int perf_evsel__add_modifiers(struct perf_evsel
*evsel
, char *bf
, size_t size
)
239 int colon
= 0, r
= 0;
240 struct perf_event_attr
*attr
= &evsel
->attr
;
241 bool exclude_guest_default
= false;
243 #define MOD_PRINT(context, mod) do { \
244 if (!attr->exclude_##context) { \
245 if (!colon) colon = ++r; \
246 r += scnprintf(bf + r, size - r, "%c", mod); \
249 if (attr
->exclude_kernel
|| attr
->exclude_user
|| attr
->exclude_hv
) {
250 MOD_PRINT(kernel
, 'k');
251 MOD_PRINT(user
, 'u');
253 exclude_guest_default
= true;
256 if (attr
->precise_ip
) {
259 r
+= scnprintf(bf
+ r
, size
- r
, "%.*s", attr
->precise_ip
, "ppp");
260 exclude_guest_default
= true;
263 if (attr
->exclude_host
|| attr
->exclude_guest
== exclude_guest_default
) {
264 MOD_PRINT(host
, 'H');
265 MOD_PRINT(guest
, 'G');
273 static int perf_evsel__hw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
275 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__hw_name(evsel
->attr
.config
));
276 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
279 const char *perf_evsel__sw_names
[PERF_COUNT_SW_MAX
] = {
292 static const char *__perf_evsel__sw_name(u64 config
)
294 if (config
< PERF_COUNT_SW_MAX
&& perf_evsel__sw_names
[config
])
295 return perf_evsel__sw_names
[config
];
296 return "unknown-software";
299 static int perf_evsel__sw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
301 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__sw_name(evsel
->attr
.config
));
302 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
305 static int __perf_evsel__bp_name(char *bf
, size_t size
, u64 addr
, u64 type
)
309 r
= scnprintf(bf
, size
, "mem:0x%" PRIx64
":", addr
);
311 if (type
& HW_BREAKPOINT_R
)
312 r
+= scnprintf(bf
+ r
, size
- r
, "r");
314 if (type
& HW_BREAKPOINT_W
)
315 r
+= scnprintf(bf
+ r
, size
- r
, "w");
317 if (type
& HW_BREAKPOINT_X
)
318 r
+= scnprintf(bf
+ r
, size
- r
, "x");
323 static int perf_evsel__bp_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
325 struct perf_event_attr
*attr
= &evsel
->attr
;
326 int r
= __perf_evsel__bp_name(bf
, size
, attr
->bp_addr
, attr
->bp_type
);
327 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
330 const char *perf_evsel__hw_cache
[PERF_COUNT_HW_CACHE_MAX
]
331 [PERF_EVSEL__MAX_ALIASES
] = {
332 { "L1-dcache", "l1-d", "l1d", "L1-data", },
333 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
335 { "dTLB", "d-tlb", "Data-TLB", },
336 { "iTLB", "i-tlb", "Instruction-TLB", },
337 { "branch", "branches", "bpu", "btb", "bpc", },
341 const char *perf_evsel__hw_cache_op
[PERF_COUNT_HW_CACHE_OP_MAX
]
342 [PERF_EVSEL__MAX_ALIASES
] = {
343 { "load", "loads", "read", },
344 { "store", "stores", "write", },
345 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
348 const char *perf_evsel__hw_cache_result
[PERF_COUNT_HW_CACHE_RESULT_MAX
]
349 [PERF_EVSEL__MAX_ALIASES
] = {
350 { "refs", "Reference", "ops", "access", },
351 { "misses", "miss", },
354 #define C(x) PERF_COUNT_HW_CACHE_##x
355 #define CACHE_READ (1 << C(OP_READ))
356 #define CACHE_WRITE (1 << C(OP_WRITE))
357 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
358 #define COP(x) (1 << x)
361 * cache operartion stat
362 * L1I : Read and prefetch only
363 * ITLB and BPU : Read-only
365 static unsigned long perf_evsel__hw_cache_stat
[C(MAX
)] = {
366 [C(L1D
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
367 [C(L1I
)] = (CACHE_READ
| CACHE_PREFETCH
),
368 [C(LL
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
369 [C(DTLB
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
370 [C(ITLB
)] = (CACHE_READ
),
371 [C(BPU
)] = (CACHE_READ
),
372 [C(NODE
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
375 bool perf_evsel__is_cache_op_valid(u8 type
, u8 op
)
377 if (perf_evsel__hw_cache_stat
[type
] & COP(op
))
378 return true; /* valid */
380 return false; /* invalid */
383 int __perf_evsel__hw_cache_type_op_res_name(u8 type
, u8 op
, u8 result
,
384 char *bf
, size_t size
)
387 return scnprintf(bf
, size
, "%s-%s-%s", perf_evsel__hw_cache
[type
][0],
388 perf_evsel__hw_cache_op
[op
][0],
389 perf_evsel__hw_cache_result
[result
][0]);
392 return scnprintf(bf
, size
, "%s-%s", perf_evsel__hw_cache
[type
][0],
393 perf_evsel__hw_cache_op
[op
][1]);
396 static int __perf_evsel__hw_cache_name(u64 config
, char *bf
, size_t size
)
398 u8 op
, result
, type
= (config
>> 0) & 0xff;
399 const char *err
= "unknown-ext-hardware-cache-type";
401 if (type
> PERF_COUNT_HW_CACHE_MAX
)
404 op
= (config
>> 8) & 0xff;
405 err
= "unknown-ext-hardware-cache-op";
406 if (op
> PERF_COUNT_HW_CACHE_OP_MAX
)
409 result
= (config
>> 16) & 0xff;
410 err
= "unknown-ext-hardware-cache-result";
411 if (result
> PERF_COUNT_HW_CACHE_RESULT_MAX
)
414 err
= "invalid-cache";
415 if (!perf_evsel__is_cache_op_valid(type
, op
))
418 return __perf_evsel__hw_cache_type_op_res_name(type
, op
, result
, bf
, size
);
420 return scnprintf(bf
, size
, "%s", err
);
423 static int perf_evsel__hw_cache_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
425 int ret
= __perf_evsel__hw_cache_name(evsel
->attr
.config
, bf
, size
);
426 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
429 static int perf_evsel__raw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
431 int ret
= scnprintf(bf
, size
, "raw 0x%" PRIx64
, evsel
->attr
.config
);
432 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
435 const char *perf_evsel__name(struct perf_evsel
*evsel
)
442 switch (evsel
->attr
.type
) {
444 perf_evsel__raw_name(evsel
, bf
, sizeof(bf
));
447 case PERF_TYPE_HARDWARE
:
448 perf_evsel__hw_name(evsel
, bf
, sizeof(bf
));
451 case PERF_TYPE_HW_CACHE
:
452 perf_evsel__hw_cache_name(evsel
, bf
, sizeof(bf
));
455 case PERF_TYPE_SOFTWARE
:
456 perf_evsel__sw_name(evsel
, bf
, sizeof(bf
));
459 case PERF_TYPE_TRACEPOINT
:
460 scnprintf(bf
, sizeof(bf
), "%s", "unknown tracepoint");
463 case PERF_TYPE_BREAKPOINT
:
464 perf_evsel__bp_name(evsel
, bf
, sizeof(bf
));
468 scnprintf(bf
, sizeof(bf
), "unknown attr type: %d",
473 evsel
->name
= strdup(bf
);
475 return evsel
->name
?: "unknown";
478 const char *perf_evsel__group_name(struct perf_evsel
*evsel
)
480 return evsel
->group_name
?: "anon group";
483 int perf_evsel__group_desc(struct perf_evsel
*evsel
, char *buf
, size_t size
)
486 struct perf_evsel
*pos
;
487 const char *group_name
= perf_evsel__group_name(evsel
);
489 ret
= scnprintf(buf
, size
, "%s", group_name
);
491 ret
+= scnprintf(buf
+ ret
, size
- ret
, " { %s",
492 perf_evsel__name(evsel
));
494 for_each_group_member(pos
, evsel
)
495 ret
+= scnprintf(buf
+ ret
, size
- ret
, ", %s",
496 perf_evsel__name(pos
));
498 ret
+= scnprintf(buf
+ ret
, size
- ret
, " }");
504 perf_evsel__config_callgraph(struct perf_evsel
*evsel
,
505 struct record_opts
*opts
)
507 bool function
= perf_evsel__is_function_event(evsel
);
508 struct perf_event_attr
*attr
= &evsel
->attr
;
510 perf_evsel__set_sample_bit(evsel
, CALLCHAIN
);
512 if (opts
->call_graph
== CALLCHAIN_DWARF
) {
514 perf_evsel__set_sample_bit(evsel
, REGS_USER
);
515 perf_evsel__set_sample_bit(evsel
, STACK_USER
);
516 attr
->sample_regs_user
= PERF_REGS_MASK
;
517 attr
->sample_stack_user
= opts
->stack_dump_size
;
518 attr
->exclude_callchain_user
= 1;
520 pr_info("Cannot use DWARF unwind for function trace event,"
521 " falling back to framepointers.\n");
526 pr_info("Disabling user space callchains for function trace event.\n");
527 attr
->exclude_callchain_user
= 1;
532 * The enable_on_exec/disabled value strategy:
534 * 1) For any type of traced program:
535 * - all independent events and group leaders are disabled
536 * - all group members are enabled
538 * Group members are ruled by group leaders. They need to
539 * be enabled, because the group scheduling relies on that.
541 * 2) For traced programs executed by perf:
542 * - all independent events and group leaders have
544 * - we don't specifically enable or disable any event during
547 * Independent events and group leaders are initially disabled
548 * and get enabled by exec. Group members are ruled by group
549 * leaders as stated in 1).
551 * 3) For traced programs attached by perf (pid/tid):
552 * - we specifically enable or disable all events during
555 * When attaching events to already running traced we
556 * enable/disable events specifically, as there's no
557 * initial traced exec call.
559 void perf_evsel__config(struct perf_evsel
*evsel
, struct record_opts
*opts
)
561 struct perf_evsel
*leader
= evsel
->leader
;
562 struct perf_event_attr
*attr
= &evsel
->attr
;
563 int track
= !evsel
->idx
; /* only the first counter needs these */
564 bool per_cpu
= opts
->target
.default_per_cpu
&& !opts
->target
.per_thread
;
566 attr
->sample_id_all
= perf_missing_features
.sample_id_all
? 0 : 1;
567 attr
->inherit
= !opts
->no_inherit
;
569 perf_evsel__set_sample_bit(evsel
, IP
);
570 perf_evsel__set_sample_bit(evsel
, TID
);
572 if (evsel
->sample_read
) {
573 perf_evsel__set_sample_bit(evsel
, READ
);
576 * We need ID even in case of single event, because
577 * PERF_SAMPLE_READ process ID specific data.
579 perf_evsel__set_sample_id(evsel
, false);
582 * Apply group format only if we belong to group
583 * with more than one members.
585 if (leader
->nr_members
> 1) {
586 attr
->read_format
|= PERF_FORMAT_GROUP
;
592 * We default some events to have a default interval. But keep
593 * it a weak assumption overridable by the user.
595 if (!attr
->sample_period
|| (opts
->user_freq
!= UINT_MAX
||
596 opts
->user_interval
!= ULLONG_MAX
)) {
598 perf_evsel__set_sample_bit(evsel
, PERIOD
);
600 attr
->sample_freq
= opts
->freq
;
602 attr
->sample_period
= opts
->default_interval
;
607 * Disable sampling for all group members other
608 * than leader in case leader 'leads' the sampling.
610 if ((leader
!= evsel
) && leader
->sample_read
) {
611 attr
->sample_freq
= 0;
612 attr
->sample_period
= 0;
615 if (opts
->no_samples
)
616 attr
->sample_freq
= 0;
618 if (opts
->inherit_stat
)
619 attr
->inherit_stat
= 1;
621 if (opts
->sample_address
) {
622 perf_evsel__set_sample_bit(evsel
, ADDR
);
623 attr
->mmap_data
= track
;
626 if (opts
->call_graph_enabled
)
627 perf_evsel__config_callgraph(evsel
, opts
);
629 if (target__has_cpu(&opts
->target
))
630 perf_evsel__set_sample_bit(evsel
, CPU
);
633 perf_evsel__set_sample_bit(evsel
, PERIOD
);
635 if (!perf_missing_features
.sample_id_all
&&
636 (opts
->sample_time
|| !opts
->no_inherit
||
637 target__has_cpu(&opts
->target
) || per_cpu
))
638 perf_evsel__set_sample_bit(evsel
, TIME
);
640 if (opts
->raw_samples
) {
641 perf_evsel__set_sample_bit(evsel
, TIME
);
642 perf_evsel__set_sample_bit(evsel
, RAW
);
643 perf_evsel__set_sample_bit(evsel
, CPU
);
646 if (opts
->sample_address
)
647 perf_evsel__set_sample_bit(evsel
, DATA_SRC
);
649 if (opts
->no_buffering
) {
651 attr
->wakeup_events
= 1;
653 if (opts
->branch_stack
) {
654 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
655 attr
->branch_sample_type
= opts
->branch_stack
;
658 if (opts
->sample_weight
)
659 perf_evsel__set_sample_bit(evsel
, WEIGHT
);
662 attr
->mmap2
= track
&& !perf_missing_features
.mmap2
;
665 if (opts
->sample_transaction
)
666 perf_evsel__set_sample_bit(evsel
, TRANSACTION
);
669 * XXX see the function comment above
671 * Disabling only independent events or group leaders,
672 * keeping group members enabled.
674 if (perf_evsel__is_group_leader(evsel
))
678 * Setting enable_on_exec for independent events and
679 * group leaders for traced executed by perf.
681 if (target__none(&opts
->target
) && perf_evsel__is_group_leader(evsel
) &&
682 !opts
->initial_delay
)
683 attr
->enable_on_exec
= 1;
686 int perf_evsel__alloc_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
689 evsel
->fd
= xyarray__new(ncpus
, nthreads
, sizeof(int));
692 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
693 for (thread
= 0; thread
< nthreads
; thread
++) {
694 FD(evsel
, cpu
, thread
) = -1;
699 return evsel
->fd
!= NULL
? 0 : -ENOMEM
;
702 static int perf_evsel__run_ioctl(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
707 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
708 for (thread
= 0; thread
< nthreads
; thread
++) {
709 int fd
= FD(evsel
, cpu
, thread
),
710 err
= ioctl(fd
, ioc
, arg
);
720 int perf_evsel__set_filter(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
723 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
724 PERF_EVENT_IOC_SET_FILTER
,
728 int perf_evsel__enable(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
730 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
731 PERF_EVENT_IOC_ENABLE
,
735 int perf_evsel__alloc_id(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
737 evsel
->sample_id
= xyarray__new(ncpus
, nthreads
, sizeof(struct perf_sample_id
));
738 if (evsel
->sample_id
== NULL
)
741 evsel
->id
= zalloc(ncpus
* nthreads
* sizeof(u64
));
742 if (evsel
->id
== NULL
) {
743 xyarray__delete(evsel
->sample_id
);
744 evsel
->sample_id
= NULL
;
751 void perf_evsel__reset_counts(struct perf_evsel
*evsel
, int ncpus
)
753 memset(evsel
->counts
, 0, (sizeof(*evsel
->counts
) +
754 (ncpus
* sizeof(struct perf_counts_values
))));
757 int perf_evsel__alloc_counts(struct perf_evsel
*evsel
, int ncpus
)
759 evsel
->counts
= zalloc((sizeof(*evsel
->counts
) +
760 (ncpus
* sizeof(struct perf_counts_values
))));
761 return evsel
->counts
!= NULL
? 0 : -ENOMEM
;
764 void perf_evsel__free_fd(struct perf_evsel
*evsel
)
766 xyarray__delete(evsel
->fd
);
770 void perf_evsel__free_id(struct perf_evsel
*evsel
)
772 xyarray__delete(evsel
->sample_id
);
773 evsel
->sample_id
= NULL
;
777 void perf_evsel__close_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
781 for (cpu
= 0; cpu
< ncpus
; cpu
++)
782 for (thread
= 0; thread
< nthreads
; ++thread
) {
783 close(FD(evsel
, cpu
, thread
));
784 FD(evsel
, cpu
, thread
) = -1;
788 void perf_evsel__free_counts(struct perf_evsel
*evsel
)
790 zfree(&evsel
->counts
);
793 void perf_evsel__exit(struct perf_evsel
*evsel
)
795 assert(list_empty(&evsel
->node
));
796 perf_evsel__free_fd(evsel
);
797 perf_evsel__free_id(evsel
);
800 void perf_evsel__delete(struct perf_evsel
*evsel
)
802 perf_evsel__exit(evsel
);
803 close_cgroup(evsel
->cgrp
);
804 zfree(&evsel
->group_name
);
805 if (evsel
->tp_format
)
806 pevent_free_format(evsel
->tp_format
);
811 static inline void compute_deltas(struct perf_evsel
*evsel
,
813 struct perf_counts_values
*count
)
815 struct perf_counts_values tmp
;
817 if (!evsel
->prev_raw_counts
)
821 tmp
= evsel
->prev_raw_counts
->aggr
;
822 evsel
->prev_raw_counts
->aggr
= *count
;
824 tmp
= evsel
->prev_raw_counts
->cpu
[cpu
];
825 evsel
->prev_raw_counts
->cpu
[cpu
] = *count
;
828 count
->val
= count
->val
- tmp
.val
;
829 count
->ena
= count
->ena
- tmp
.ena
;
830 count
->run
= count
->run
- tmp
.run
;
833 int __perf_evsel__read_on_cpu(struct perf_evsel
*evsel
,
834 int cpu
, int thread
, bool scale
)
836 struct perf_counts_values count
;
837 size_t nv
= scale
? 3 : 1;
839 if (FD(evsel
, cpu
, thread
) < 0)
842 if (evsel
->counts
== NULL
&& perf_evsel__alloc_counts(evsel
, cpu
+ 1) < 0)
845 if (readn(FD(evsel
, cpu
, thread
), &count
, nv
* sizeof(u64
)) < 0)
848 compute_deltas(evsel
, cpu
, &count
);
853 else if (count
.run
< count
.ena
)
854 count
.val
= (u64
)((double)count
.val
* count
.ena
/ count
.run
+ 0.5);
856 count
.ena
= count
.run
= 0;
858 evsel
->counts
->cpu
[cpu
] = count
;
862 int __perf_evsel__read(struct perf_evsel
*evsel
,
863 int ncpus
, int nthreads
, bool scale
)
865 size_t nv
= scale
? 3 : 1;
867 struct perf_counts_values
*aggr
= &evsel
->counts
->aggr
, count
;
869 aggr
->val
= aggr
->ena
= aggr
->run
= 0;
871 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
872 for (thread
= 0; thread
< nthreads
; thread
++) {
873 if (FD(evsel
, cpu
, thread
) < 0)
876 if (readn(FD(evsel
, cpu
, thread
),
877 &count
, nv
* sizeof(u64
)) < 0)
880 aggr
->val
+= count
.val
;
882 aggr
->ena
+= count
.ena
;
883 aggr
->run
+= count
.run
;
888 compute_deltas(evsel
, -1, aggr
);
890 evsel
->counts
->scaled
= 0;
892 if (aggr
->run
== 0) {
893 evsel
->counts
->scaled
= -1;
898 if (aggr
->run
< aggr
->ena
) {
899 evsel
->counts
->scaled
= 1;
900 aggr
->val
= (u64
)((double)aggr
->val
* aggr
->ena
/ aggr
->run
+ 0.5);
903 aggr
->ena
= aggr
->run
= 0;
908 static int get_group_fd(struct perf_evsel
*evsel
, int cpu
, int thread
)
910 struct perf_evsel
*leader
= evsel
->leader
;
913 if (perf_evsel__is_group_leader(evsel
))
917 * Leader must be already processed/open,
922 fd
= FD(leader
, cpu
, thread
);
928 #define __PRINT_ATTR(fmt, cast, field) \
929 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
931 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
932 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
933 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
934 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
936 #define PRINT_ATTR2N(name1, field1, name2, field2) \
937 fprintf(fp, " %-19s %u %-19s %u\n", \
938 name1, attr->field1, name2, attr->field2)
940 #define PRINT_ATTR2(field1, field2) \
941 PRINT_ATTR2N(#field1, field1, #field2, field2)
943 static size_t perf_event_attr__fprintf(struct perf_event_attr
*attr
, FILE *fp
)
947 ret
+= fprintf(fp
, "%.60s\n", graph_dotted_line
);
948 ret
+= fprintf(fp
, "perf_event_attr:\n");
950 ret
+= PRINT_ATTR_U32(type
);
951 ret
+= PRINT_ATTR_U32(size
);
952 ret
+= PRINT_ATTR_X64(config
);
953 ret
+= PRINT_ATTR_U64(sample_period
);
954 ret
+= PRINT_ATTR_U64(sample_freq
);
955 ret
+= PRINT_ATTR_X64(sample_type
);
956 ret
+= PRINT_ATTR_X64(read_format
);
958 ret
+= PRINT_ATTR2(disabled
, inherit
);
959 ret
+= PRINT_ATTR2(pinned
, exclusive
);
960 ret
+= PRINT_ATTR2(exclude_user
, exclude_kernel
);
961 ret
+= PRINT_ATTR2(exclude_hv
, exclude_idle
);
962 ret
+= PRINT_ATTR2(mmap
, comm
);
963 ret
+= PRINT_ATTR2(freq
, inherit_stat
);
964 ret
+= PRINT_ATTR2(enable_on_exec
, task
);
965 ret
+= PRINT_ATTR2(watermark
, precise_ip
);
966 ret
+= PRINT_ATTR2(mmap_data
, sample_id_all
);
967 ret
+= PRINT_ATTR2(exclude_host
, exclude_guest
);
968 ret
+= PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel
,
969 "excl.callchain_user", exclude_callchain_user
);
970 ret
+= PRINT_ATTR_U32(mmap2
);
972 ret
+= PRINT_ATTR_U32(wakeup_events
);
973 ret
+= PRINT_ATTR_U32(wakeup_watermark
);
974 ret
+= PRINT_ATTR_X32(bp_type
);
975 ret
+= PRINT_ATTR_X64(bp_addr
);
976 ret
+= PRINT_ATTR_X64(config1
);
977 ret
+= PRINT_ATTR_U64(bp_len
);
978 ret
+= PRINT_ATTR_X64(config2
);
979 ret
+= PRINT_ATTR_X64(branch_sample_type
);
980 ret
+= PRINT_ATTR_X64(sample_regs_user
);
981 ret
+= PRINT_ATTR_U32(sample_stack_user
);
983 ret
+= fprintf(fp
, "%.60s\n", graph_dotted_line
);
988 static int __perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
989 struct thread_map
*threads
)
992 unsigned long flags
= 0;
994 enum { NO_CHANGE
, SET_TO_MAX
, INCREASED_MAX
} set_rlimit
= NO_CHANGE
;
996 if (evsel
->fd
== NULL
&&
997 perf_evsel__alloc_fd(evsel
, cpus
->nr
, threads
->nr
) < 0)
1001 flags
= PERF_FLAG_PID_CGROUP
;
1002 pid
= evsel
->cgrp
->fd
;
1005 fallback_missing_features
:
1006 if (perf_missing_features
.mmap2
)
1007 evsel
->attr
.mmap2
= 0;
1008 if (perf_missing_features
.exclude_guest
)
1009 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
1011 if (perf_missing_features
.sample_id_all
)
1012 evsel
->attr
.sample_id_all
= 0;
1015 perf_event_attr__fprintf(&evsel
->attr
, stderr
);
1017 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
1019 for (thread
= 0; thread
< threads
->nr
; thread
++) {
1023 pid
= threads
->map
[thread
];
1025 group_fd
= get_group_fd(evsel
, cpu
, thread
);
1027 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1028 pid
, cpus
->map
[cpu
], group_fd
, flags
);
1030 FD(evsel
, cpu
, thread
) = sys_perf_event_open(&evsel
->attr
,
1034 if (FD(evsel
, cpu
, thread
) < 0) {
1036 pr_debug2("sys_perf_event_open failed, error %d\n",
1040 set_rlimit
= NO_CHANGE
;
1048 * perf stat needs between 5 and 22 fds per CPU. When we run out
1049 * of them try to increase the limits.
1051 if (err
== -EMFILE
&& set_rlimit
< INCREASED_MAX
) {
1053 int old_errno
= errno
;
1055 if (getrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1056 if (set_rlimit
== NO_CHANGE
)
1057 l
.rlim_cur
= l
.rlim_max
;
1059 l
.rlim_cur
= l
.rlim_max
+ 1000;
1060 l
.rlim_max
= l
.rlim_cur
;
1062 if (setrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1071 if (err
!= -EINVAL
|| cpu
> 0 || thread
> 0)
1074 if (!perf_missing_features
.mmap2
&& evsel
->attr
.mmap2
) {
1075 perf_missing_features
.mmap2
= true;
1076 goto fallback_missing_features
;
1077 } else if (!perf_missing_features
.exclude_guest
&&
1078 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
1079 perf_missing_features
.exclude_guest
= true;
1080 goto fallback_missing_features
;
1081 } else if (!perf_missing_features
.sample_id_all
) {
1082 perf_missing_features
.sample_id_all
= true;
1083 goto retry_sample_id
;
1088 while (--thread
>= 0) {
1089 close(FD(evsel
, cpu
, thread
));
1090 FD(evsel
, cpu
, thread
) = -1;
1092 thread
= threads
->nr
;
1093 } while (--cpu
>= 0);
1097 void perf_evsel__close(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1099 if (evsel
->fd
== NULL
)
1102 perf_evsel__close_fd(evsel
, ncpus
, nthreads
);
1103 perf_evsel__free_fd(evsel
);
1115 struct thread_map map
;
1117 } empty_thread_map
= {
1122 int perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1123 struct thread_map
*threads
)
1126 /* Work around old compiler warnings about strict aliasing */
1127 cpus
= &empty_cpu_map
.map
;
1130 if (threads
== NULL
)
1131 threads
= &empty_thread_map
.map
;
1133 return __perf_evsel__open(evsel
, cpus
, threads
);
1136 int perf_evsel__open_per_cpu(struct perf_evsel
*evsel
,
1137 struct cpu_map
*cpus
)
1139 return __perf_evsel__open(evsel
, cpus
, &empty_thread_map
.map
);
1142 int perf_evsel__open_per_thread(struct perf_evsel
*evsel
,
1143 struct thread_map
*threads
)
1145 return __perf_evsel__open(evsel
, &empty_cpu_map
.map
, threads
);
1148 static int perf_evsel__parse_id_sample(const struct perf_evsel
*evsel
,
1149 const union perf_event
*event
,
1150 struct perf_sample
*sample
)
1152 u64 type
= evsel
->attr
.sample_type
;
1153 const u64
*array
= event
->sample
.array
;
1154 bool swapped
= evsel
->needs_swap
;
1157 array
+= ((event
->header
.size
-
1158 sizeof(event
->header
)) / sizeof(u64
)) - 1;
1160 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1161 sample
->id
= *array
;
1165 if (type
& PERF_SAMPLE_CPU
) {
1168 /* undo swap of u64, then swap on individual u32s */
1169 u
.val64
= bswap_64(u
.val64
);
1170 u
.val32
[0] = bswap_32(u
.val32
[0]);
1173 sample
->cpu
= u
.val32
[0];
1177 if (type
& PERF_SAMPLE_STREAM_ID
) {
1178 sample
->stream_id
= *array
;
1182 if (type
& PERF_SAMPLE_ID
) {
1183 sample
->id
= *array
;
1187 if (type
& PERF_SAMPLE_TIME
) {
1188 sample
->time
= *array
;
1192 if (type
& PERF_SAMPLE_TID
) {
1195 /* undo swap of u64, then swap on individual u32s */
1196 u
.val64
= bswap_64(u
.val64
);
1197 u
.val32
[0] = bswap_32(u
.val32
[0]);
1198 u
.val32
[1] = bswap_32(u
.val32
[1]);
1201 sample
->pid
= u
.val32
[0];
1202 sample
->tid
= u
.val32
[1];
1209 static inline bool overflow(const void *endp
, u16 max_size
, const void *offset
,
1212 return size
> max_size
|| offset
+ size
> endp
;
1215 #define OVERFLOW_CHECK(offset, size, max_size) \
1217 if (overflow(endp, (max_size), (offset), (size))) \
1221 #define OVERFLOW_CHECK_u64(offset) \
1222 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1224 int perf_evsel__parse_sample(struct perf_evsel
*evsel
, union perf_event
*event
,
1225 struct perf_sample
*data
)
1227 u64 type
= evsel
->attr
.sample_type
;
1228 bool swapped
= evsel
->needs_swap
;
1230 u16 max_size
= event
->header
.size
;
1231 const void *endp
= (void *)event
+ max_size
;
1235 * used for cross-endian analysis. See git commit 65014ab3
1236 * for why this goofiness is needed.
1240 memset(data
, 0, sizeof(*data
));
1241 data
->cpu
= data
->pid
= data
->tid
= -1;
1242 data
->stream_id
= data
->id
= data
->time
= -1ULL;
1243 data
->period
= evsel
->attr
.sample_period
;
1246 if (event
->header
.type
!= PERF_RECORD_SAMPLE
) {
1247 if (!evsel
->attr
.sample_id_all
)
1249 return perf_evsel__parse_id_sample(evsel
, event
, data
);
1252 array
= event
->sample
.array
;
1255 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1256 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1257 * check the format does not go past the end of the event.
1259 if (evsel
->sample_size
+ sizeof(event
->header
) > event
->header
.size
)
1263 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1268 if (type
& PERF_SAMPLE_IP
) {
1273 if (type
& PERF_SAMPLE_TID
) {
1276 /* undo swap of u64, then swap on individual u32s */
1277 u
.val64
= bswap_64(u
.val64
);
1278 u
.val32
[0] = bswap_32(u
.val32
[0]);
1279 u
.val32
[1] = bswap_32(u
.val32
[1]);
1282 data
->pid
= u
.val32
[0];
1283 data
->tid
= u
.val32
[1];
1287 if (type
& PERF_SAMPLE_TIME
) {
1288 data
->time
= *array
;
1293 if (type
& PERF_SAMPLE_ADDR
) {
1294 data
->addr
= *array
;
1298 if (type
& PERF_SAMPLE_ID
) {
1303 if (type
& PERF_SAMPLE_STREAM_ID
) {
1304 data
->stream_id
= *array
;
1308 if (type
& PERF_SAMPLE_CPU
) {
1312 /* undo swap of u64, then swap on individual u32s */
1313 u
.val64
= bswap_64(u
.val64
);
1314 u
.val32
[0] = bswap_32(u
.val32
[0]);
1317 data
->cpu
= u
.val32
[0];
1321 if (type
& PERF_SAMPLE_PERIOD
) {
1322 data
->period
= *array
;
1326 if (type
& PERF_SAMPLE_READ
) {
1327 u64 read_format
= evsel
->attr
.read_format
;
1329 OVERFLOW_CHECK_u64(array
);
1330 if (read_format
& PERF_FORMAT_GROUP
)
1331 data
->read
.group
.nr
= *array
;
1333 data
->read
.one
.value
= *array
;
1337 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1338 OVERFLOW_CHECK_u64(array
);
1339 data
->read
.time_enabled
= *array
;
1343 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1344 OVERFLOW_CHECK_u64(array
);
1345 data
->read
.time_running
= *array
;
1349 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1350 if (read_format
& PERF_FORMAT_GROUP
) {
1351 const u64 max_group_nr
= UINT64_MAX
/
1352 sizeof(struct sample_read_value
);
1354 if (data
->read
.group
.nr
> max_group_nr
)
1356 sz
= data
->read
.group
.nr
*
1357 sizeof(struct sample_read_value
);
1358 OVERFLOW_CHECK(array
, sz
, max_size
);
1359 data
->read
.group
.values
=
1360 (struct sample_read_value
*)array
;
1361 array
= (void *)array
+ sz
;
1363 OVERFLOW_CHECK_u64(array
);
1364 data
->read
.one
.id
= *array
;
1369 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1370 const u64 max_callchain_nr
= UINT64_MAX
/ sizeof(u64
);
1372 OVERFLOW_CHECK_u64(array
);
1373 data
->callchain
= (struct ip_callchain
*)array
++;
1374 if (data
->callchain
->nr
> max_callchain_nr
)
1376 sz
= data
->callchain
->nr
* sizeof(u64
);
1377 OVERFLOW_CHECK(array
, sz
, max_size
);
1378 array
= (void *)array
+ sz
;
1381 if (type
& PERF_SAMPLE_RAW
) {
1382 OVERFLOW_CHECK_u64(array
);
1384 if (WARN_ONCE(swapped
,
1385 "Endianness of raw data not corrected!\n")) {
1386 /* undo swap of u64, then swap on individual u32s */
1387 u
.val64
= bswap_64(u
.val64
);
1388 u
.val32
[0] = bswap_32(u
.val32
[0]);
1389 u
.val32
[1] = bswap_32(u
.val32
[1]);
1391 data
->raw_size
= u
.val32
[0];
1392 array
= (void *)array
+ sizeof(u32
);
1394 OVERFLOW_CHECK(array
, data
->raw_size
, max_size
);
1395 data
->raw_data
= (void *)array
;
1396 array
= (void *)array
+ data
->raw_size
;
1399 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1400 const u64 max_branch_nr
= UINT64_MAX
/
1401 sizeof(struct branch_entry
);
1403 OVERFLOW_CHECK_u64(array
);
1404 data
->branch_stack
= (struct branch_stack
*)array
++;
1406 if (data
->branch_stack
->nr
> max_branch_nr
)
1408 sz
= data
->branch_stack
->nr
* sizeof(struct branch_entry
);
1409 OVERFLOW_CHECK(array
, sz
, max_size
);
1410 array
= (void *)array
+ sz
;
1413 if (type
& PERF_SAMPLE_REGS_USER
) {
1414 OVERFLOW_CHECK_u64(array
);
1415 data
->user_regs
.abi
= *array
;
1418 if (data
->user_regs
.abi
) {
1419 u64 mask
= evsel
->attr
.sample_regs_user
;
1421 sz
= hweight_long(mask
) * sizeof(u64
);
1422 OVERFLOW_CHECK(array
, sz
, max_size
);
1423 data
->user_regs
.mask
= mask
;
1424 data
->user_regs
.regs
= (u64
*)array
;
1425 array
= (void *)array
+ sz
;
1429 if (type
& PERF_SAMPLE_STACK_USER
) {
1430 OVERFLOW_CHECK_u64(array
);
1433 data
->user_stack
.offset
= ((char *)(array
- 1)
1437 data
->user_stack
.size
= 0;
1439 OVERFLOW_CHECK(array
, sz
, max_size
);
1440 data
->user_stack
.data
= (char *)array
;
1441 array
= (void *)array
+ sz
;
1442 OVERFLOW_CHECK_u64(array
);
1443 data
->user_stack
.size
= *array
++;
1444 if (WARN_ONCE(data
->user_stack
.size
> sz
,
1445 "user stack dump failure\n"))
1451 if (type
& PERF_SAMPLE_WEIGHT
) {
1452 OVERFLOW_CHECK_u64(array
);
1453 data
->weight
= *array
;
1457 data
->data_src
= PERF_MEM_DATA_SRC_NONE
;
1458 if (type
& PERF_SAMPLE_DATA_SRC
) {
1459 OVERFLOW_CHECK_u64(array
);
1460 data
->data_src
= *array
;
1464 data
->transaction
= 0;
1465 if (type
& PERF_SAMPLE_TRANSACTION
) {
1466 OVERFLOW_CHECK_u64(array
);
1467 data
->transaction
= *array
;
1474 size_t perf_event__sample_event_size(const struct perf_sample
*sample
, u64 type
,
1477 size_t sz
, result
= sizeof(struct sample_event
);
1479 if (type
& PERF_SAMPLE_IDENTIFIER
)
1480 result
+= sizeof(u64
);
1482 if (type
& PERF_SAMPLE_IP
)
1483 result
+= sizeof(u64
);
1485 if (type
& PERF_SAMPLE_TID
)
1486 result
+= sizeof(u64
);
1488 if (type
& PERF_SAMPLE_TIME
)
1489 result
+= sizeof(u64
);
1491 if (type
& PERF_SAMPLE_ADDR
)
1492 result
+= sizeof(u64
);
1494 if (type
& PERF_SAMPLE_ID
)
1495 result
+= sizeof(u64
);
1497 if (type
& PERF_SAMPLE_STREAM_ID
)
1498 result
+= sizeof(u64
);
1500 if (type
& PERF_SAMPLE_CPU
)
1501 result
+= sizeof(u64
);
1503 if (type
& PERF_SAMPLE_PERIOD
)
1504 result
+= sizeof(u64
);
1506 if (type
& PERF_SAMPLE_READ
) {
1507 result
+= sizeof(u64
);
1508 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
1509 result
+= sizeof(u64
);
1510 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
1511 result
+= sizeof(u64
);
1512 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1513 if (read_format
& PERF_FORMAT_GROUP
) {
1514 sz
= sample
->read
.group
.nr
*
1515 sizeof(struct sample_read_value
);
1518 result
+= sizeof(u64
);
1522 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1523 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1527 if (type
& PERF_SAMPLE_RAW
) {
1528 result
+= sizeof(u32
);
1529 result
+= sample
->raw_size
;
1532 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1533 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1538 if (type
& PERF_SAMPLE_REGS_USER
) {
1539 if (sample
->user_regs
.abi
) {
1540 result
+= sizeof(u64
);
1541 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
1544 result
+= sizeof(u64
);
1548 if (type
& PERF_SAMPLE_STACK_USER
) {
1549 sz
= sample
->user_stack
.size
;
1550 result
+= sizeof(u64
);
1553 result
+= sizeof(u64
);
1557 if (type
& PERF_SAMPLE_WEIGHT
)
1558 result
+= sizeof(u64
);
1560 if (type
& PERF_SAMPLE_DATA_SRC
)
1561 result
+= sizeof(u64
);
1563 if (type
& PERF_SAMPLE_TRANSACTION
)
1564 result
+= sizeof(u64
);
1569 int perf_event__synthesize_sample(union perf_event
*event
, u64 type
,
1571 const struct perf_sample
*sample
,
1577 * used for cross-endian analysis. See git commit 65014ab3
1578 * for why this goofiness is needed.
1582 array
= event
->sample
.array
;
1584 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1585 *array
= sample
->id
;
1589 if (type
& PERF_SAMPLE_IP
) {
1590 *array
= sample
->ip
;
1594 if (type
& PERF_SAMPLE_TID
) {
1595 u
.val32
[0] = sample
->pid
;
1596 u
.val32
[1] = sample
->tid
;
1599 * Inverse of what is done in perf_evsel__parse_sample
1601 u
.val32
[0] = bswap_32(u
.val32
[0]);
1602 u
.val32
[1] = bswap_32(u
.val32
[1]);
1603 u
.val64
= bswap_64(u
.val64
);
1610 if (type
& PERF_SAMPLE_TIME
) {
1611 *array
= sample
->time
;
1615 if (type
& PERF_SAMPLE_ADDR
) {
1616 *array
= sample
->addr
;
1620 if (type
& PERF_SAMPLE_ID
) {
1621 *array
= sample
->id
;
1625 if (type
& PERF_SAMPLE_STREAM_ID
) {
1626 *array
= sample
->stream_id
;
1630 if (type
& PERF_SAMPLE_CPU
) {
1631 u
.val32
[0] = sample
->cpu
;
1634 * Inverse of what is done in perf_evsel__parse_sample
1636 u
.val32
[0] = bswap_32(u
.val32
[0]);
1637 u
.val64
= bswap_64(u
.val64
);
1643 if (type
& PERF_SAMPLE_PERIOD
) {
1644 *array
= sample
->period
;
1648 if (type
& PERF_SAMPLE_READ
) {
1649 if (read_format
& PERF_FORMAT_GROUP
)
1650 *array
= sample
->read
.group
.nr
;
1652 *array
= sample
->read
.one
.value
;
1655 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1656 *array
= sample
->read
.time_enabled
;
1660 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1661 *array
= sample
->read
.time_running
;
1665 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1666 if (read_format
& PERF_FORMAT_GROUP
) {
1667 sz
= sample
->read
.group
.nr
*
1668 sizeof(struct sample_read_value
);
1669 memcpy(array
, sample
->read
.group
.values
, sz
);
1670 array
= (void *)array
+ sz
;
1672 *array
= sample
->read
.one
.id
;
1677 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1678 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
1679 memcpy(array
, sample
->callchain
, sz
);
1680 array
= (void *)array
+ sz
;
1683 if (type
& PERF_SAMPLE_RAW
) {
1684 u
.val32
[0] = sample
->raw_size
;
1685 if (WARN_ONCE(swapped
,
1686 "Endianness of raw data not corrected!\n")) {
1688 * Inverse of what is done in perf_evsel__parse_sample
1690 u
.val32
[0] = bswap_32(u
.val32
[0]);
1691 u
.val32
[1] = bswap_32(u
.val32
[1]);
1692 u
.val64
= bswap_64(u
.val64
);
1695 array
= (void *)array
+ sizeof(u32
);
1697 memcpy(array
, sample
->raw_data
, sample
->raw_size
);
1698 array
= (void *)array
+ sample
->raw_size
;
1701 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1702 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
1704 memcpy(array
, sample
->branch_stack
, sz
);
1705 array
= (void *)array
+ sz
;
1708 if (type
& PERF_SAMPLE_REGS_USER
) {
1709 if (sample
->user_regs
.abi
) {
1710 *array
++ = sample
->user_regs
.abi
;
1711 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
1712 memcpy(array
, sample
->user_regs
.regs
, sz
);
1713 array
= (void *)array
+ sz
;
1719 if (type
& PERF_SAMPLE_STACK_USER
) {
1720 sz
= sample
->user_stack
.size
;
1723 memcpy(array
, sample
->user_stack
.data
, sz
);
1724 array
= (void *)array
+ sz
;
1729 if (type
& PERF_SAMPLE_WEIGHT
) {
1730 *array
= sample
->weight
;
1734 if (type
& PERF_SAMPLE_DATA_SRC
) {
1735 *array
= sample
->data_src
;
1739 if (type
& PERF_SAMPLE_TRANSACTION
) {
1740 *array
= sample
->transaction
;
1747 struct format_field
*perf_evsel__field(struct perf_evsel
*evsel
, const char *name
)
1749 return pevent_find_field(evsel
->tp_format
, name
);
1752 void *perf_evsel__rawptr(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
1755 struct format_field
*field
= perf_evsel__field(evsel
, name
);
1761 offset
= field
->offset
;
1763 if (field
->flags
& FIELD_IS_DYNAMIC
) {
1764 offset
= *(int *)(sample
->raw_data
+ field
->offset
);
1768 return sample
->raw_data
+ offset
;
1771 u64
perf_evsel__intval(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
1774 struct format_field
*field
= perf_evsel__field(evsel
, name
);
1781 ptr
= sample
->raw_data
+ field
->offset
;
1783 switch (field
->size
) {
1787 value
= *(u16
*)ptr
;
1790 value
= *(u32
*)ptr
;
1793 value
= *(u64
*)ptr
;
1799 if (!evsel
->needs_swap
)
1802 switch (field
->size
) {
1804 return bswap_16(value
);
1806 return bswap_32(value
);
1808 return bswap_64(value
);
1816 static int comma_fprintf(FILE *fp
, bool *first
, const char *fmt
, ...)
1822 ret
+= fprintf(fp
, ",");
1824 ret
+= fprintf(fp
, ":");
1828 va_start(args
, fmt
);
1829 ret
+= vfprintf(fp
, fmt
, args
);
1834 static int __if_fprintf(FILE *fp
, bool *first
, const char *field
, u64 value
)
1839 return comma_fprintf(fp
, first
, " %s: %" PRIu64
, field
, value
);
1842 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1849 static int bits__fprintf(FILE *fp
, const char *field
, u64 value
,
1850 struct bit_names
*bits
, bool *first
)
1852 int i
= 0, printed
= comma_fprintf(fp
, first
, " %s: ", field
);
1853 bool first_bit
= true;
1856 if (value
& bits
[i
].bit
) {
1857 printed
+= fprintf(fp
, "%s%s", first_bit
? "" : "|", bits
[i
].name
);
1860 } while (bits
[++i
].name
!= NULL
);
1865 static int sample_type__fprintf(FILE *fp
, bool *first
, u64 value
)
1867 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1868 struct bit_names bits
[] = {
1869 bit_name(IP
), bit_name(TID
), bit_name(TIME
), bit_name(ADDR
),
1870 bit_name(READ
), bit_name(CALLCHAIN
), bit_name(ID
), bit_name(CPU
),
1871 bit_name(PERIOD
), bit_name(STREAM_ID
), bit_name(RAW
),
1872 bit_name(BRANCH_STACK
), bit_name(REGS_USER
), bit_name(STACK_USER
),
1873 bit_name(IDENTIFIER
),
1877 return bits__fprintf(fp
, "sample_type", value
, bits
, first
);
1880 static int read_format__fprintf(FILE *fp
, bool *first
, u64 value
)
1882 #define bit_name(n) { PERF_FORMAT_##n, #n }
1883 struct bit_names bits
[] = {
1884 bit_name(TOTAL_TIME_ENABLED
), bit_name(TOTAL_TIME_RUNNING
),
1885 bit_name(ID
), bit_name(GROUP
),
1889 return bits__fprintf(fp
, "read_format", value
, bits
, first
);
1892 int perf_evsel__fprintf(struct perf_evsel
*evsel
,
1893 struct perf_attr_details
*details
, FILE *fp
)
1898 if (details
->event_group
) {
1899 struct perf_evsel
*pos
;
1901 if (!perf_evsel__is_group_leader(evsel
))
1904 if (evsel
->nr_members
> 1)
1905 printed
+= fprintf(fp
, "%s{", evsel
->group_name
?: "");
1907 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
1908 for_each_group_member(pos
, evsel
)
1909 printed
+= fprintf(fp
, ",%s", perf_evsel__name(pos
));
1911 if (evsel
->nr_members
> 1)
1912 printed
+= fprintf(fp
, "}");
1916 printed
+= fprintf(fp
, "%s", perf_evsel__name(evsel
));
1918 if (details
->verbose
|| details
->freq
) {
1919 printed
+= comma_fprintf(fp
, &first
, " sample_freq=%" PRIu64
,
1920 (u64
)evsel
->attr
.sample_freq
);
1923 if (details
->verbose
) {
1929 printed
+= sample_type__fprintf(fp
, &first
, evsel
->attr
.sample_type
);
1930 if (evsel
->attr
.read_format
)
1931 printed
+= read_format__fprintf(fp
, &first
, evsel
->attr
.read_format
);
1935 if_print(exclusive
);
1936 if_print(exclude_user
);
1937 if_print(exclude_kernel
);
1938 if_print(exclude_hv
);
1939 if_print(exclude_idle
);
1944 if_print(inherit_stat
);
1945 if_print(enable_on_exec
);
1947 if_print(watermark
);
1948 if_print(precise_ip
);
1949 if_print(mmap_data
);
1950 if_print(sample_id_all
);
1951 if_print(exclude_host
);
1952 if_print(exclude_guest
);
1953 if_print(__reserved_1
);
1954 if_print(wakeup_events
);
1956 if_print(branch_sample_type
);
1963 bool perf_evsel__fallback(struct perf_evsel
*evsel
, int err
,
1964 char *msg
, size_t msgsize
)
1966 if ((err
== ENOENT
|| err
== ENXIO
|| err
== ENODEV
) &&
1967 evsel
->attr
.type
== PERF_TYPE_HARDWARE
&&
1968 evsel
->attr
.config
== PERF_COUNT_HW_CPU_CYCLES
) {
1970 * If it's cycles then fall back to hrtimer based
1971 * cpu-clock-tick sw counter, which is always available even if
1974 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1977 scnprintf(msg
, msgsize
, "%s",
1978 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1980 evsel
->attr
.type
= PERF_TYPE_SOFTWARE
;
1981 evsel
->attr
.config
= PERF_COUNT_SW_CPU_CLOCK
;
1983 zfree(&evsel
->name
);
1990 int perf_evsel__open_strerror(struct perf_evsel
*evsel
, struct target
*target
,
1991 int err
, char *msg
, size_t size
)
1996 return scnprintf(msg
, size
,
1997 "You may not have permission to collect %sstats.\n"
1998 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
1999 " -1 - Not paranoid at all\n"
2000 " 0 - Disallow raw tracepoint access for unpriv\n"
2001 " 1 - Disallow cpu events for unpriv\n"
2002 " 2 - Disallow kernel profiling for unpriv",
2003 target
->system_wide
? "system-wide " : "");
2005 return scnprintf(msg
, size
, "The %s event is not supported.",
2006 perf_evsel__name(evsel
));
2008 return scnprintf(msg
, size
, "%s",
2009 "Too many events are opened.\n"
2010 "Try again after reducing the number of events.");
2012 if (target
->cpu_list
)
2013 return scnprintf(msg
, size
, "%s",
2014 "No such device - did you specify an out-of-range profile CPU?\n");
2017 if (evsel
->attr
.precise_ip
)
2018 return scnprintf(msg
, size
, "%s",
2019 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2020 #if defined(__i386__) || defined(__x86_64__)
2021 if (evsel
->attr
.type
== PERF_TYPE_HARDWARE
)
2022 return scnprintf(msg
, size
, "%s",
2023 "No hardware sampling interrupt available.\n"
2024 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2031 return scnprintf(msg
, size
,
2032 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2033 "/bin/dmesg may provide additional information.\n"
2034 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2035 err
, strerror(err
), perf_evsel__name(evsel
));