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)
13 #include <linux/bitops.h>
14 #include <api/fs/fs.h>
15 #include <api/fs/tracing_path.h>
16 #include <traceevent/event-parse.h>
17 #include <linux/hw_breakpoint.h>
18 #include <linux/perf_event.h>
19 #include <linux/compiler.h>
20 #include <linux/err.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
26 #include "callchain.h"
33 #include "thread_map.h"
35 #include "perf_regs.h"
37 #include "trace-event.h"
39 #include "util/parse-branch-options.h"
41 #include "sane_ctype.h"
52 } perf_missing_features
;
54 static clockid_t clockid
;
56 static int perf_evsel__no_extra_init(struct perf_evsel
*evsel __maybe_unused
)
61 static void perf_evsel__no_extra_fini(struct perf_evsel
*evsel __maybe_unused
)
67 int (*init
)(struct perf_evsel
*evsel
);
68 void (*fini
)(struct perf_evsel
*evsel
);
69 } perf_evsel__object
= {
70 .size
= sizeof(struct perf_evsel
),
71 .init
= perf_evsel__no_extra_init
,
72 .fini
= perf_evsel__no_extra_fini
,
75 int perf_evsel__object_config(size_t object_size
,
76 int (*init
)(struct perf_evsel
*evsel
),
77 void (*fini
)(struct perf_evsel
*evsel
))
83 if (perf_evsel__object
.size
> object_size
)
86 perf_evsel__object
.size
= object_size
;
90 perf_evsel__object
.init
= init
;
93 perf_evsel__object
.fini
= fini
;
98 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
100 int __perf_evsel__sample_size(u64 sample_type
)
102 u64 mask
= sample_type
& PERF_SAMPLE_MASK
;
106 for (i
= 0; i
< 64; i
++) {
107 if (mask
& (1ULL << i
))
117 * __perf_evsel__calc_id_pos - calculate id_pos.
118 * @sample_type: sample type
120 * This function returns the position of the event id (PERF_SAMPLE_ID or
121 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
124 static int __perf_evsel__calc_id_pos(u64 sample_type
)
128 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
131 if (!(sample_type
& PERF_SAMPLE_ID
))
134 if (sample_type
& PERF_SAMPLE_IP
)
137 if (sample_type
& PERF_SAMPLE_TID
)
140 if (sample_type
& PERF_SAMPLE_TIME
)
143 if (sample_type
& PERF_SAMPLE_ADDR
)
150 * __perf_evsel__calc_is_pos - calculate is_pos.
151 * @sample_type: sample type
153 * This function returns the position (counting backwards) of the event id
154 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
155 * sample_id_all is used there is an id sample appended to non-sample events.
157 static int __perf_evsel__calc_is_pos(u64 sample_type
)
161 if (sample_type
& PERF_SAMPLE_IDENTIFIER
)
164 if (!(sample_type
& PERF_SAMPLE_ID
))
167 if (sample_type
& PERF_SAMPLE_CPU
)
170 if (sample_type
& PERF_SAMPLE_STREAM_ID
)
176 void perf_evsel__calc_id_pos(struct perf_evsel
*evsel
)
178 evsel
->id_pos
= __perf_evsel__calc_id_pos(evsel
->attr
.sample_type
);
179 evsel
->is_pos
= __perf_evsel__calc_is_pos(evsel
->attr
.sample_type
);
182 void __perf_evsel__set_sample_bit(struct perf_evsel
*evsel
,
183 enum perf_event_sample_format bit
)
185 if (!(evsel
->attr
.sample_type
& bit
)) {
186 evsel
->attr
.sample_type
|= bit
;
187 evsel
->sample_size
+= sizeof(u64
);
188 perf_evsel__calc_id_pos(evsel
);
192 void __perf_evsel__reset_sample_bit(struct perf_evsel
*evsel
,
193 enum perf_event_sample_format bit
)
195 if (evsel
->attr
.sample_type
& bit
) {
196 evsel
->attr
.sample_type
&= ~bit
;
197 evsel
->sample_size
-= sizeof(u64
);
198 perf_evsel__calc_id_pos(evsel
);
202 void perf_evsel__set_sample_id(struct perf_evsel
*evsel
,
203 bool can_sample_identifier
)
205 if (can_sample_identifier
) {
206 perf_evsel__reset_sample_bit(evsel
, ID
);
207 perf_evsel__set_sample_bit(evsel
, IDENTIFIER
);
209 perf_evsel__set_sample_bit(evsel
, ID
);
211 evsel
->attr
.read_format
|= PERF_FORMAT_ID
;
215 * perf_evsel__is_function_event - Return whether given evsel is a function
218 * @evsel - evsel selector to be tested
220 * Return %true if event is function trace event
222 bool perf_evsel__is_function_event(struct perf_evsel
*evsel
)
224 #define FUNCTION_EVENT "ftrace:function"
226 return evsel
->name
&&
227 !strncmp(FUNCTION_EVENT
, evsel
->name
, sizeof(FUNCTION_EVENT
));
229 #undef FUNCTION_EVENT
232 void perf_evsel__init(struct perf_evsel
*evsel
,
233 struct perf_event_attr
*attr
, int idx
)
236 evsel
->tracking
= !idx
;
238 evsel
->leader
= evsel
;
241 evsel
->evlist
= NULL
;
243 INIT_LIST_HEAD(&evsel
->node
);
244 INIT_LIST_HEAD(&evsel
->config_terms
);
245 perf_evsel__object
.init(evsel
);
246 evsel
->sample_size
= __perf_evsel__sample_size(attr
->sample_type
);
247 perf_evsel__calc_id_pos(evsel
);
248 evsel
->cmdline_group_boundary
= false;
249 evsel
->metric_expr
= NULL
;
250 evsel
->metric_name
= NULL
;
251 evsel
->metric_events
= NULL
;
252 evsel
->collect_stat
= false;
255 struct perf_evsel
*perf_evsel__new_idx(struct perf_event_attr
*attr
, int idx
)
257 struct perf_evsel
*evsel
= zalloc(perf_evsel__object
.size
);
260 perf_evsel__init(evsel
, attr
, idx
);
262 if (perf_evsel__is_bpf_output(evsel
)) {
263 evsel
->attr
.sample_type
|= (PERF_SAMPLE_RAW
| PERF_SAMPLE_TIME
|
264 PERF_SAMPLE_CPU
| PERF_SAMPLE_PERIOD
),
265 evsel
->attr
.sample_period
= 1;
271 struct perf_evsel
*perf_evsel__new_cycles(void)
273 struct perf_event_attr attr
= {
274 .type
= PERF_TYPE_HARDWARE
,
275 .config
= PERF_COUNT_HW_CPU_CYCLES
,
278 struct perf_evsel
*evsel
;
280 event_attr_init(&attr
);
282 * Unnamed union member, not supported as struct member named
283 * initializer in older compilers such as gcc 4.4.7
285 * Just for probing the precise_ip:
287 attr
.sample_period
= 1;
289 perf_event_attr__set_max_precise_ip(&attr
);
291 * Now let the usual logic to set up the perf_event_attr defaults
292 * to kick in when we return and before perf_evsel__open() is called.
294 attr
.sample_period
= 0;
296 evsel
= perf_evsel__new(&attr
);
300 /* use asprintf() because free(evsel) assumes name is allocated */
301 if (asprintf(&evsel
->name
, "cycles%.*s",
302 attr
.precise_ip
? attr
.precise_ip
+ 1 : 0, ":ppp") < 0)
307 perf_evsel__delete(evsel
);
313 * Returns pointer with encoded error via <linux/err.h> interface.
315 struct perf_evsel
*perf_evsel__newtp_idx(const char *sys
, const char *name
, int idx
)
317 struct perf_evsel
*evsel
= zalloc(perf_evsel__object
.size
);
323 struct perf_event_attr attr
= {
324 .type
= PERF_TYPE_TRACEPOINT
,
325 .sample_type
= (PERF_SAMPLE_RAW
| PERF_SAMPLE_TIME
|
326 PERF_SAMPLE_CPU
| PERF_SAMPLE_PERIOD
),
329 if (asprintf(&evsel
->name
, "%s:%s", sys
, name
) < 0)
332 evsel
->tp_format
= trace_event__tp_format(sys
, name
);
333 if (IS_ERR(evsel
->tp_format
)) {
334 err
= PTR_ERR(evsel
->tp_format
);
338 event_attr_init(&attr
);
339 attr
.config
= evsel
->tp_format
->id
;
340 attr
.sample_period
= 1;
341 perf_evsel__init(evsel
, &attr
, idx
);
353 const char *perf_evsel__hw_names
[PERF_COUNT_HW_MAX
] = {
361 "stalled-cycles-frontend",
362 "stalled-cycles-backend",
366 static const char *__perf_evsel__hw_name(u64 config
)
368 if (config
< PERF_COUNT_HW_MAX
&& perf_evsel__hw_names
[config
])
369 return perf_evsel__hw_names
[config
];
371 return "unknown-hardware";
374 static int perf_evsel__add_modifiers(struct perf_evsel
*evsel
, char *bf
, size_t size
)
376 int colon
= 0, r
= 0;
377 struct perf_event_attr
*attr
= &evsel
->attr
;
378 bool exclude_guest_default
= false;
380 #define MOD_PRINT(context, mod) do { \
381 if (!attr->exclude_##context) { \
382 if (!colon) colon = ++r; \
383 r += scnprintf(bf + r, size - r, "%c", mod); \
386 if (attr
->exclude_kernel
|| attr
->exclude_user
|| attr
->exclude_hv
) {
387 MOD_PRINT(kernel
, 'k');
388 MOD_PRINT(user
, 'u');
390 exclude_guest_default
= true;
393 if (attr
->precise_ip
) {
396 r
+= scnprintf(bf
+ r
, size
- r
, "%.*s", attr
->precise_ip
, "ppp");
397 exclude_guest_default
= true;
400 if (attr
->exclude_host
|| attr
->exclude_guest
== exclude_guest_default
) {
401 MOD_PRINT(host
, 'H');
402 MOD_PRINT(guest
, 'G');
410 static int perf_evsel__hw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
412 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__hw_name(evsel
->attr
.config
));
413 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
416 const char *perf_evsel__sw_names
[PERF_COUNT_SW_MAX
] = {
429 static const char *__perf_evsel__sw_name(u64 config
)
431 if (config
< PERF_COUNT_SW_MAX
&& perf_evsel__sw_names
[config
])
432 return perf_evsel__sw_names
[config
];
433 return "unknown-software";
436 static int perf_evsel__sw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
438 int r
= scnprintf(bf
, size
, "%s", __perf_evsel__sw_name(evsel
->attr
.config
));
439 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
442 static int __perf_evsel__bp_name(char *bf
, size_t size
, u64 addr
, u64 type
)
446 r
= scnprintf(bf
, size
, "mem:0x%" PRIx64
":", addr
);
448 if (type
& HW_BREAKPOINT_R
)
449 r
+= scnprintf(bf
+ r
, size
- r
, "r");
451 if (type
& HW_BREAKPOINT_W
)
452 r
+= scnprintf(bf
+ r
, size
- r
, "w");
454 if (type
& HW_BREAKPOINT_X
)
455 r
+= scnprintf(bf
+ r
, size
- r
, "x");
460 static int perf_evsel__bp_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
462 struct perf_event_attr
*attr
= &evsel
->attr
;
463 int r
= __perf_evsel__bp_name(bf
, size
, attr
->bp_addr
, attr
->bp_type
);
464 return r
+ perf_evsel__add_modifiers(evsel
, bf
+ r
, size
- r
);
467 const char *perf_evsel__hw_cache
[PERF_COUNT_HW_CACHE_MAX
]
468 [PERF_EVSEL__MAX_ALIASES
] = {
469 { "L1-dcache", "l1-d", "l1d", "L1-data", },
470 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
472 { "dTLB", "d-tlb", "Data-TLB", },
473 { "iTLB", "i-tlb", "Instruction-TLB", },
474 { "branch", "branches", "bpu", "btb", "bpc", },
478 const char *perf_evsel__hw_cache_op
[PERF_COUNT_HW_CACHE_OP_MAX
]
479 [PERF_EVSEL__MAX_ALIASES
] = {
480 { "load", "loads", "read", },
481 { "store", "stores", "write", },
482 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
485 const char *perf_evsel__hw_cache_result
[PERF_COUNT_HW_CACHE_RESULT_MAX
]
486 [PERF_EVSEL__MAX_ALIASES
] = {
487 { "refs", "Reference", "ops", "access", },
488 { "misses", "miss", },
491 #define C(x) PERF_COUNT_HW_CACHE_##x
492 #define CACHE_READ (1 << C(OP_READ))
493 #define CACHE_WRITE (1 << C(OP_WRITE))
494 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
495 #define COP(x) (1 << x)
498 * cache operartion stat
499 * L1I : Read and prefetch only
500 * ITLB and BPU : Read-only
502 static unsigned long perf_evsel__hw_cache_stat
[C(MAX
)] = {
503 [C(L1D
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
504 [C(L1I
)] = (CACHE_READ
| CACHE_PREFETCH
),
505 [C(LL
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
506 [C(DTLB
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
507 [C(ITLB
)] = (CACHE_READ
),
508 [C(BPU
)] = (CACHE_READ
),
509 [C(NODE
)] = (CACHE_READ
| CACHE_WRITE
| CACHE_PREFETCH
),
512 bool perf_evsel__is_cache_op_valid(u8 type
, u8 op
)
514 if (perf_evsel__hw_cache_stat
[type
] & COP(op
))
515 return true; /* valid */
517 return false; /* invalid */
520 int __perf_evsel__hw_cache_type_op_res_name(u8 type
, u8 op
, u8 result
,
521 char *bf
, size_t size
)
524 return scnprintf(bf
, size
, "%s-%s-%s", perf_evsel__hw_cache
[type
][0],
525 perf_evsel__hw_cache_op
[op
][0],
526 perf_evsel__hw_cache_result
[result
][0]);
529 return scnprintf(bf
, size
, "%s-%s", perf_evsel__hw_cache
[type
][0],
530 perf_evsel__hw_cache_op
[op
][1]);
533 static int __perf_evsel__hw_cache_name(u64 config
, char *bf
, size_t size
)
535 u8 op
, result
, type
= (config
>> 0) & 0xff;
536 const char *err
= "unknown-ext-hardware-cache-type";
538 if (type
>= PERF_COUNT_HW_CACHE_MAX
)
541 op
= (config
>> 8) & 0xff;
542 err
= "unknown-ext-hardware-cache-op";
543 if (op
>= PERF_COUNT_HW_CACHE_OP_MAX
)
546 result
= (config
>> 16) & 0xff;
547 err
= "unknown-ext-hardware-cache-result";
548 if (result
>= PERF_COUNT_HW_CACHE_RESULT_MAX
)
551 err
= "invalid-cache";
552 if (!perf_evsel__is_cache_op_valid(type
, op
))
555 return __perf_evsel__hw_cache_type_op_res_name(type
, op
, result
, bf
, size
);
557 return scnprintf(bf
, size
, "%s", err
);
560 static int perf_evsel__hw_cache_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
562 int ret
= __perf_evsel__hw_cache_name(evsel
->attr
.config
, bf
, size
);
563 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
566 static int perf_evsel__raw_name(struct perf_evsel
*evsel
, char *bf
, size_t size
)
568 int ret
= scnprintf(bf
, size
, "raw 0x%" PRIx64
, evsel
->attr
.config
);
569 return ret
+ perf_evsel__add_modifiers(evsel
, bf
+ ret
, size
- ret
);
572 const char *perf_evsel__name(struct perf_evsel
*evsel
)
579 switch (evsel
->attr
.type
) {
581 perf_evsel__raw_name(evsel
, bf
, sizeof(bf
));
584 case PERF_TYPE_HARDWARE
:
585 perf_evsel__hw_name(evsel
, bf
, sizeof(bf
));
588 case PERF_TYPE_HW_CACHE
:
589 perf_evsel__hw_cache_name(evsel
, bf
, sizeof(bf
));
592 case PERF_TYPE_SOFTWARE
:
593 perf_evsel__sw_name(evsel
, bf
, sizeof(bf
));
596 case PERF_TYPE_TRACEPOINT
:
597 scnprintf(bf
, sizeof(bf
), "%s", "unknown tracepoint");
600 case PERF_TYPE_BREAKPOINT
:
601 perf_evsel__bp_name(evsel
, bf
, sizeof(bf
));
605 scnprintf(bf
, sizeof(bf
), "unknown attr type: %d",
610 evsel
->name
= strdup(bf
);
612 return evsel
->name
?: "unknown";
615 const char *perf_evsel__group_name(struct perf_evsel
*evsel
)
617 return evsel
->group_name
?: "anon group";
620 int perf_evsel__group_desc(struct perf_evsel
*evsel
, char *buf
, size_t size
)
623 struct perf_evsel
*pos
;
624 const char *group_name
= perf_evsel__group_name(evsel
);
626 ret
= scnprintf(buf
, size
, "%s", group_name
);
628 ret
+= scnprintf(buf
+ ret
, size
- ret
, " { %s",
629 perf_evsel__name(evsel
));
631 for_each_group_member(pos
, evsel
)
632 ret
+= scnprintf(buf
+ ret
, size
- ret
, ", %s",
633 perf_evsel__name(pos
));
635 ret
+= scnprintf(buf
+ ret
, size
- ret
, " }");
640 void perf_evsel__config_callchain(struct perf_evsel
*evsel
,
641 struct record_opts
*opts
,
642 struct callchain_param
*param
)
644 bool function
= perf_evsel__is_function_event(evsel
);
645 struct perf_event_attr
*attr
= &evsel
->attr
;
647 perf_evsel__set_sample_bit(evsel
, CALLCHAIN
);
649 attr
->sample_max_stack
= param
->max_stack
;
651 if (param
->record_mode
== CALLCHAIN_LBR
) {
652 if (!opts
->branch_stack
) {
653 if (attr
->exclude_user
) {
654 pr_warning("LBR callstack option is only available "
655 "to get user callchain information. "
656 "Falling back to framepointers.\n");
658 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
659 attr
->branch_sample_type
= PERF_SAMPLE_BRANCH_USER
|
660 PERF_SAMPLE_BRANCH_CALL_STACK
|
661 PERF_SAMPLE_BRANCH_NO_CYCLES
|
662 PERF_SAMPLE_BRANCH_NO_FLAGS
;
665 pr_warning("Cannot use LBR callstack with branch stack. "
666 "Falling back to framepointers.\n");
669 if (param
->record_mode
== CALLCHAIN_DWARF
) {
671 perf_evsel__set_sample_bit(evsel
, REGS_USER
);
672 perf_evsel__set_sample_bit(evsel
, STACK_USER
);
673 attr
->sample_regs_user
= PERF_REGS_MASK
;
674 attr
->sample_stack_user
= param
->dump_size
;
675 attr
->exclude_callchain_user
= 1;
677 pr_info("Cannot use DWARF unwind for function trace event,"
678 " falling back to framepointers.\n");
683 pr_info("Disabling user space callchains for function trace event.\n");
684 attr
->exclude_callchain_user
= 1;
689 perf_evsel__reset_callgraph(struct perf_evsel
*evsel
,
690 struct callchain_param
*param
)
692 struct perf_event_attr
*attr
= &evsel
->attr
;
694 perf_evsel__reset_sample_bit(evsel
, CALLCHAIN
);
695 if (param
->record_mode
== CALLCHAIN_LBR
) {
696 perf_evsel__reset_sample_bit(evsel
, BRANCH_STACK
);
697 attr
->branch_sample_type
&= ~(PERF_SAMPLE_BRANCH_USER
|
698 PERF_SAMPLE_BRANCH_CALL_STACK
);
700 if (param
->record_mode
== CALLCHAIN_DWARF
) {
701 perf_evsel__reset_sample_bit(evsel
, REGS_USER
);
702 perf_evsel__reset_sample_bit(evsel
, STACK_USER
);
706 static void apply_config_terms(struct perf_evsel
*evsel
,
707 struct record_opts
*opts
)
709 struct perf_evsel_config_term
*term
;
710 struct list_head
*config_terms
= &evsel
->config_terms
;
711 struct perf_event_attr
*attr
= &evsel
->attr
;
712 struct callchain_param param
;
715 const char *callgraph_buf
= NULL
;
717 /* callgraph default */
718 param
.record_mode
= callchain_param
.record_mode
;
720 list_for_each_entry(term
, config_terms
, list
) {
721 switch (term
->type
) {
722 case PERF_EVSEL__CONFIG_TERM_PERIOD
:
723 attr
->sample_period
= term
->val
.period
;
726 case PERF_EVSEL__CONFIG_TERM_FREQ
:
727 attr
->sample_freq
= term
->val
.freq
;
730 case PERF_EVSEL__CONFIG_TERM_TIME
:
732 perf_evsel__set_sample_bit(evsel
, TIME
);
734 perf_evsel__reset_sample_bit(evsel
, TIME
);
736 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH
:
737 callgraph_buf
= term
->val
.callgraph
;
739 case PERF_EVSEL__CONFIG_TERM_BRANCH
:
740 if (term
->val
.branch
&& strcmp(term
->val
.branch
, "no")) {
741 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
742 parse_branch_str(term
->val
.branch
,
743 &attr
->branch_sample_type
);
745 perf_evsel__reset_sample_bit(evsel
, BRANCH_STACK
);
747 case PERF_EVSEL__CONFIG_TERM_STACK_USER
:
748 dump_size
= term
->val
.stack_user
;
750 case PERF_EVSEL__CONFIG_TERM_MAX_STACK
:
751 max_stack
= term
->val
.max_stack
;
753 case PERF_EVSEL__CONFIG_TERM_INHERIT
:
755 * attr->inherit should has already been set by
756 * perf_evsel__config. If user explicitly set
757 * inherit using config terms, override global
758 * opt->no_inherit setting.
760 attr
->inherit
= term
->val
.inherit
? 1 : 0;
762 case PERF_EVSEL__CONFIG_TERM_OVERWRITE
:
763 attr
->write_backward
= term
->val
.overwrite
? 1 : 0;
770 /* User explicitly set per-event callgraph, clear the old setting and reset. */
771 if ((callgraph_buf
!= NULL
) || (dump_size
> 0) || max_stack
) {
773 param
.max_stack
= max_stack
;
774 if (callgraph_buf
== NULL
)
775 callgraph_buf
= "fp";
778 /* parse callgraph parameters */
779 if (callgraph_buf
!= NULL
) {
780 if (!strcmp(callgraph_buf
, "no")) {
781 param
.enabled
= false;
782 param
.record_mode
= CALLCHAIN_NONE
;
784 param
.enabled
= true;
785 if (parse_callchain_record(callgraph_buf
, ¶m
)) {
786 pr_err("per-event callgraph setting for %s failed. "
787 "Apply callgraph global setting for it\n",
794 dump_size
= round_up(dump_size
, sizeof(u64
));
795 param
.dump_size
= dump_size
;
798 /* If global callgraph set, clear it */
799 if (callchain_param
.enabled
)
800 perf_evsel__reset_callgraph(evsel
, &callchain_param
);
802 /* set perf-event callgraph */
804 perf_evsel__config_callchain(evsel
, opts
, ¶m
);
809 * The enable_on_exec/disabled value strategy:
811 * 1) For any type of traced program:
812 * - all independent events and group leaders are disabled
813 * - all group members are enabled
815 * Group members are ruled by group leaders. They need to
816 * be enabled, because the group scheduling relies on that.
818 * 2) For traced programs executed by perf:
819 * - all independent events and group leaders have
821 * - we don't specifically enable or disable any event during
824 * Independent events and group leaders are initially disabled
825 * and get enabled by exec. Group members are ruled by group
826 * leaders as stated in 1).
828 * 3) For traced programs attached by perf (pid/tid):
829 * - we specifically enable or disable all events during
832 * When attaching events to already running traced we
833 * enable/disable events specifically, as there's no
834 * initial traced exec call.
836 void perf_evsel__config(struct perf_evsel
*evsel
, struct record_opts
*opts
,
837 struct callchain_param
*callchain
)
839 struct perf_evsel
*leader
= evsel
->leader
;
840 struct perf_event_attr
*attr
= &evsel
->attr
;
841 int track
= evsel
->tracking
;
842 bool per_cpu
= opts
->target
.default_per_cpu
&& !opts
->target
.per_thread
;
844 attr
->sample_id_all
= perf_missing_features
.sample_id_all
? 0 : 1;
845 attr
->inherit
= !opts
->no_inherit
;
846 attr
->write_backward
= opts
->overwrite
? 1 : 0;
848 perf_evsel__set_sample_bit(evsel
, IP
);
849 perf_evsel__set_sample_bit(evsel
, TID
);
851 if (evsel
->sample_read
) {
852 perf_evsel__set_sample_bit(evsel
, READ
);
855 * We need ID even in case of single event, because
856 * PERF_SAMPLE_READ process ID specific data.
858 perf_evsel__set_sample_id(evsel
, false);
861 * Apply group format only if we belong to group
862 * with more than one members.
864 if (leader
->nr_members
> 1) {
865 attr
->read_format
|= PERF_FORMAT_GROUP
;
871 * We default some events to have a default interval. But keep
872 * it a weak assumption overridable by the user.
874 if (!attr
->sample_period
|| (opts
->user_freq
!= UINT_MAX
||
875 opts
->user_interval
!= ULLONG_MAX
)) {
877 perf_evsel__set_sample_bit(evsel
, PERIOD
);
879 attr
->sample_freq
= opts
->freq
;
881 attr
->sample_period
= opts
->default_interval
;
886 * Disable sampling for all group members other
887 * than leader in case leader 'leads' the sampling.
889 if ((leader
!= evsel
) && leader
->sample_read
) {
890 attr
->sample_freq
= 0;
891 attr
->sample_period
= 0;
894 if (opts
->no_samples
)
895 attr
->sample_freq
= 0;
897 if (opts
->inherit_stat
)
898 attr
->inherit_stat
= 1;
900 if (opts
->sample_address
) {
901 perf_evsel__set_sample_bit(evsel
, ADDR
);
902 attr
->mmap_data
= track
;
906 * We don't allow user space callchains for function trace
907 * event, due to issues with page faults while tracing page
908 * fault handler and its overall trickiness nature.
910 if (perf_evsel__is_function_event(evsel
))
911 evsel
->attr
.exclude_callchain_user
= 1;
913 if (callchain
&& callchain
->enabled
&& !evsel
->no_aux_samples
)
914 perf_evsel__config_callchain(evsel
, opts
, callchain
);
916 if (opts
->sample_intr_regs
) {
917 attr
->sample_regs_intr
= opts
->sample_intr_regs
;
918 perf_evsel__set_sample_bit(evsel
, REGS_INTR
);
921 if (target__has_cpu(&opts
->target
) || opts
->sample_cpu
)
922 perf_evsel__set_sample_bit(evsel
, CPU
);
925 perf_evsel__set_sample_bit(evsel
, PERIOD
);
928 * When the user explicitly disabled time don't force it here.
930 if (opts
->sample_time
&&
931 (!perf_missing_features
.sample_id_all
&&
932 (!opts
->no_inherit
|| target__has_cpu(&opts
->target
) || per_cpu
||
933 opts
->sample_time_set
)))
934 perf_evsel__set_sample_bit(evsel
, TIME
);
936 if (opts
->raw_samples
&& !evsel
->no_aux_samples
) {
937 perf_evsel__set_sample_bit(evsel
, TIME
);
938 perf_evsel__set_sample_bit(evsel
, RAW
);
939 perf_evsel__set_sample_bit(evsel
, CPU
);
942 if (opts
->sample_address
)
943 perf_evsel__set_sample_bit(evsel
, DATA_SRC
);
945 if (opts
->no_buffering
) {
947 attr
->wakeup_events
= 1;
949 if (opts
->branch_stack
&& !evsel
->no_aux_samples
) {
950 perf_evsel__set_sample_bit(evsel
, BRANCH_STACK
);
951 attr
->branch_sample_type
= opts
->branch_stack
;
954 if (opts
->sample_weight
)
955 perf_evsel__set_sample_bit(evsel
, WEIGHT
);
959 attr
->mmap2
= track
&& !perf_missing_features
.mmap2
;
962 if (opts
->record_namespaces
)
963 attr
->namespaces
= track
;
965 if (opts
->record_switch_events
)
966 attr
->context_switch
= track
;
968 if (opts
->sample_transaction
)
969 perf_evsel__set_sample_bit(evsel
, TRANSACTION
);
971 if (opts
->running_time
) {
972 evsel
->attr
.read_format
|=
973 PERF_FORMAT_TOTAL_TIME_ENABLED
|
974 PERF_FORMAT_TOTAL_TIME_RUNNING
;
978 * XXX see the function comment above
980 * Disabling only independent events or group leaders,
981 * keeping group members enabled.
983 if (perf_evsel__is_group_leader(evsel
))
987 * Setting enable_on_exec for independent events and
988 * group leaders for traced executed by perf.
990 if (target__none(&opts
->target
) && perf_evsel__is_group_leader(evsel
) &&
991 !opts
->initial_delay
)
992 attr
->enable_on_exec
= 1;
994 if (evsel
->immediate
) {
996 attr
->enable_on_exec
= 0;
999 clockid
= opts
->clockid
;
1000 if (opts
->use_clockid
) {
1001 attr
->use_clockid
= 1;
1002 attr
->clockid
= opts
->clockid
;
1005 if (evsel
->precise_max
)
1006 perf_event_attr__set_max_precise_ip(attr
);
1008 if (opts
->all_user
) {
1009 attr
->exclude_kernel
= 1;
1010 attr
->exclude_user
= 0;
1013 if (opts
->all_kernel
) {
1014 attr
->exclude_kernel
= 0;
1015 attr
->exclude_user
= 1;
1019 * Apply event specific term settings,
1020 * it overloads any global configuration.
1022 apply_config_terms(evsel
, opts
);
1024 evsel
->ignore_missing_thread
= opts
->ignore_missing_thread
;
1027 static int perf_evsel__alloc_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1029 if (evsel
->system_wide
)
1032 evsel
->fd
= xyarray__new(ncpus
, nthreads
, sizeof(int));
1036 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
1037 for (thread
= 0; thread
< nthreads
; thread
++) {
1038 FD(evsel
, cpu
, thread
) = -1;
1043 return evsel
->fd
!= NULL
? 0 : -ENOMEM
;
1046 static int perf_evsel__run_ioctl(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
1051 if (evsel
->system_wide
)
1054 for (cpu
= 0; cpu
< ncpus
; cpu
++) {
1055 for (thread
= 0; thread
< nthreads
; thread
++) {
1056 int fd
= FD(evsel
, cpu
, thread
),
1057 err
= ioctl(fd
, ioc
, arg
);
1067 int perf_evsel__apply_filter(struct perf_evsel
*evsel
, int ncpus
, int nthreads
,
1070 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
1071 PERF_EVENT_IOC_SET_FILTER
,
1075 int perf_evsel__set_filter(struct perf_evsel
*evsel
, const char *filter
)
1077 char *new_filter
= strdup(filter
);
1079 if (new_filter
!= NULL
) {
1080 free(evsel
->filter
);
1081 evsel
->filter
= new_filter
;
1088 static int perf_evsel__append_filter(struct perf_evsel
*evsel
,
1089 const char *fmt
, const char *filter
)
1093 if (evsel
->filter
== NULL
)
1094 return perf_evsel__set_filter(evsel
, filter
);
1096 if (asprintf(&new_filter
, fmt
, evsel
->filter
, filter
) > 0) {
1097 free(evsel
->filter
);
1098 evsel
->filter
= new_filter
;
1105 int perf_evsel__append_tp_filter(struct perf_evsel
*evsel
, const char *filter
)
1107 return perf_evsel__append_filter(evsel
, "(%s) && (%s)", filter
);
1110 int perf_evsel__append_addr_filter(struct perf_evsel
*evsel
, const char *filter
)
1112 return perf_evsel__append_filter(evsel
, "%s,%s", filter
);
1115 int perf_evsel__enable(struct perf_evsel
*evsel
)
1117 int nthreads
= thread_map__nr(evsel
->threads
);
1118 int ncpus
= cpu_map__nr(evsel
->cpus
);
1120 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
1121 PERF_EVENT_IOC_ENABLE
,
1125 int perf_evsel__disable(struct perf_evsel
*evsel
)
1127 int nthreads
= thread_map__nr(evsel
->threads
);
1128 int ncpus
= cpu_map__nr(evsel
->cpus
);
1130 return perf_evsel__run_ioctl(evsel
, ncpus
, nthreads
,
1131 PERF_EVENT_IOC_DISABLE
,
1135 int perf_evsel__alloc_id(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1137 if (ncpus
== 0 || nthreads
== 0)
1140 if (evsel
->system_wide
)
1143 evsel
->sample_id
= xyarray__new(ncpus
, nthreads
, sizeof(struct perf_sample_id
));
1144 if (evsel
->sample_id
== NULL
)
1147 evsel
->id
= zalloc(ncpus
* nthreads
* sizeof(u64
));
1148 if (evsel
->id
== NULL
) {
1149 xyarray__delete(evsel
->sample_id
);
1150 evsel
->sample_id
= NULL
;
1157 static void perf_evsel__free_fd(struct perf_evsel
*evsel
)
1159 xyarray__delete(evsel
->fd
);
1163 static void perf_evsel__free_id(struct perf_evsel
*evsel
)
1165 xyarray__delete(evsel
->sample_id
);
1166 evsel
->sample_id
= NULL
;
1170 static void perf_evsel__free_config_terms(struct perf_evsel
*evsel
)
1172 struct perf_evsel_config_term
*term
, *h
;
1174 list_for_each_entry_safe(term
, h
, &evsel
->config_terms
, list
) {
1175 list_del(&term
->list
);
1180 void perf_evsel__close_fd(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1184 if (evsel
->system_wide
)
1187 for (cpu
= 0; cpu
< ncpus
; cpu
++)
1188 for (thread
= 0; thread
< nthreads
; ++thread
) {
1189 close(FD(evsel
, cpu
, thread
));
1190 FD(evsel
, cpu
, thread
) = -1;
1194 void perf_evsel__exit(struct perf_evsel
*evsel
)
1196 assert(list_empty(&evsel
->node
));
1197 assert(evsel
->evlist
== NULL
);
1198 perf_evsel__free_fd(evsel
);
1199 perf_evsel__free_id(evsel
);
1200 perf_evsel__free_config_terms(evsel
);
1201 close_cgroup(evsel
->cgrp
);
1202 cpu_map__put(evsel
->cpus
);
1203 cpu_map__put(evsel
->own_cpus
);
1204 thread_map__put(evsel
->threads
);
1205 zfree(&evsel
->group_name
);
1206 zfree(&evsel
->name
);
1207 perf_evsel__object
.fini(evsel
);
1210 void perf_evsel__delete(struct perf_evsel
*evsel
)
1212 perf_evsel__exit(evsel
);
1216 void perf_evsel__compute_deltas(struct perf_evsel
*evsel
, int cpu
, int thread
,
1217 struct perf_counts_values
*count
)
1219 struct perf_counts_values tmp
;
1221 if (!evsel
->prev_raw_counts
)
1225 tmp
= evsel
->prev_raw_counts
->aggr
;
1226 evsel
->prev_raw_counts
->aggr
= *count
;
1228 tmp
= *perf_counts(evsel
->prev_raw_counts
, cpu
, thread
);
1229 *perf_counts(evsel
->prev_raw_counts
, cpu
, thread
) = *count
;
1232 count
->val
= count
->val
- tmp
.val
;
1233 count
->ena
= count
->ena
- tmp
.ena
;
1234 count
->run
= count
->run
- tmp
.run
;
1237 void perf_counts_values__scale(struct perf_counts_values
*count
,
1238 bool scale
, s8
*pscaled
)
1243 if (count
->run
== 0) {
1246 } else if (count
->run
< count
->ena
) {
1248 count
->val
= (u64
)((double) count
->val
* count
->ena
/ count
->run
+ 0.5);
1251 count
->ena
= count
->run
= 0;
1257 int perf_evsel__read(struct perf_evsel
*evsel
, int cpu
, int thread
,
1258 struct perf_counts_values
*count
)
1260 memset(count
, 0, sizeof(*count
));
1262 if (FD(evsel
, cpu
, thread
) < 0)
1265 if (readn(FD(evsel
, cpu
, thread
), count
, sizeof(*count
)) <= 0)
1271 int __perf_evsel__read_on_cpu(struct perf_evsel
*evsel
,
1272 int cpu
, int thread
, bool scale
)
1274 struct perf_counts_values count
;
1275 size_t nv
= scale
? 3 : 1;
1277 if (FD(evsel
, cpu
, thread
) < 0)
1280 if (evsel
->counts
== NULL
&& perf_evsel__alloc_counts(evsel
, cpu
+ 1, thread
+ 1) < 0)
1283 if (readn(FD(evsel
, cpu
, thread
), &count
, nv
* sizeof(u64
)) <= 0)
1286 perf_evsel__compute_deltas(evsel
, cpu
, thread
, &count
);
1287 perf_counts_values__scale(&count
, scale
, NULL
);
1288 *perf_counts(evsel
->counts
, cpu
, thread
) = count
;
1292 static int get_group_fd(struct perf_evsel
*evsel
, int cpu
, int thread
)
1294 struct perf_evsel
*leader
= evsel
->leader
;
1297 if (perf_evsel__is_group_leader(evsel
))
1301 * Leader must be already processed/open,
1302 * if not it's a bug.
1304 BUG_ON(!leader
->fd
);
1306 fd
= FD(leader
, cpu
, thread
);
1317 static void __p_bits(char *buf
, size_t size
, u64 value
, struct bit_names
*bits
)
1319 bool first_bit
= true;
1323 if (value
& bits
[i
].bit
) {
1324 buf
+= scnprintf(buf
, size
, "%s%s", first_bit
? "" : "|", bits
[i
].name
);
1327 } while (bits
[++i
].name
!= NULL
);
1330 static void __p_sample_type(char *buf
, size_t size
, u64 value
)
1332 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1333 struct bit_names bits
[] = {
1334 bit_name(IP
), bit_name(TID
), bit_name(TIME
), bit_name(ADDR
),
1335 bit_name(READ
), bit_name(CALLCHAIN
), bit_name(ID
), bit_name(CPU
),
1336 bit_name(PERIOD
), bit_name(STREAM_ID
), bit_name(RAW
),
1337 bit_name(BRANCH_STACK
), bit_name(REGS_USER
), bit_name(STACK_USER
),
1338 bit_name(IDENTIFIER
), bit_name(REGS_INTR
), bit_name(DATA_SRC
),
1343 __p_bits(buf
, size
, value
, bits
);
1346 static void __p_branch_sample_type(char *buf
, size_t size
, u64 value
)
1348 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1349 struct bit_names bits
[] = {
1350 bit_name(USER
), bit_name(KERNEL
), bit_name(HV
), bit_name(ANY
),
1351 bit_name(ANY_CALL
), bit_name(ANY_RETURN
), bit_name(IND_CALL
),
1352 bit_name(ABORT_TX
), bit_name(IN_TX
), bit_name(NO_TX
),
1353 bit_name(COND
), bit_name(CALL_STACK
), bit_name(IND_JUMP
),
1354 bit_name(CALL
), bit_name(NO_FLAGS
), bit_name(NO_CYCLES
),
1358 __p_bits(buf
, size
, value
, bits
);
1361 static void __p_read_format(char *buf
, size_t size
, u64 value
)
1363 #define bit_name(n) { PERF_FORMAT_##n, #n }
1364 struct bit_names bits
[] = {
1365 bit_name(TOTAL_TIME_ENABLED
), bit_name(TOTAL_TIME_RUNNING
),
1366 bit_name(ID
), bit_name(GROUP
),
1370 __p_bits(buf
, size
, value
, bits
);
1373 #define BUF_SIZE 1024
1375 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1376 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1377 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1378 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1379 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1380 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1382 #define PRINT_ATTRn(_n, _f, _p) \
1386 ret += attr__fprintf(fp, _n, buf, priv);\
1390 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1392 int perf_event_attr__fprintf(FILE *fp
, struct perf_event_attr
*attr
,
1393 attr__fprintf_f attr__fprintf
, void *priv
)
1398 PRINT_ATTRf(type
, p_unsigned
);
1399 PRINT_ATTRf(size
, p_unsigned
);
1400 PRINT_ATTRf(config
, p_hex
);
1401 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period
, p_unsigned
);
1402 PRINT_ATTRf(sample_type
, p_sample_type
);
1403 PRINT_ATTRf(read_format
, p_read_format
);
1405 PRINT_ATTRf(disabled
, p_unsigned
);
1406 PRINT_ATTRf(inherit
, p_unsigned
);
1407 PRINT_ATTRf(pinned
, p_unsigned
);
1408 PRINT_ATTRf(exclusive
, p_unsigned
);
1409 PRINT_ATTRf(exclude_user
, p_unsigned
);
1410 PRINT_ATTRf(exclude_kernel
, p_unsigned
);
1411 PRINT_ATTRf(exclude_hv
, p_unsigned
);
1412 PRINT_ATTRf(exclude_idle
, p_unsigned
);
1413 PRINT_ATTRf(mmap
, p_unsigned
);
1414 PRINT_ATTRf(comm
, p_unsigned
);
1415 PRINT_ATTRf(freq
, p_unsigned
);
1416 PRINT_ATTRf(inherit_stat
, p_unsigned
);
1417 PRINT_ATTRf(enable_on_exec
, p_unsigned
);
1418 PRINT_ATTRf(task
, p_unsigned
);
1419 PRINT_ATTRf(watermark
, p_unsigned
);
1420 PRINT_ATTRf(precise_ip
, p_unsigned
);
1421 PRINT_ATTRf(mmap_data
, p_unsigned
);
1422 PRINT_ATTRf(sample_id_all
, p_unsigned
);
1423 PRINT_ATTRf(exclude_host
, p_unsigned
);
1424 PRINT_ATTRf(exclude_guest
, p_unsigned
);
1425 PRINT_ATTRf(exclude_callchain_kernel
, p_unsigned
);
1426 PRINT_ATTRf(exclude_callchain_user
, p_unsigned
);
1427 PRINT_ATTRf(mmap2
, p_unsigned
);
1428 PRINT_ATTRf(comm_exec
, p_unsigned
);
1429 PRINT_ATTRf(use_clockid
, p_unsigned
);
1430 PRINT_ATTRf(context_switch
, p_unsigned
);
1431 PRINT_ATTRf(write_backward
, p_unsigned
);
1433 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events
, p_unsigned
);
1434 PRINT_ATTRf(bp_type
, p_unsigned
);
1435 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr
, p_hex
);
1436 PRINT_ATTRn("{ bp_len, config2 }", bp_len
, p_hex
);
1437 PRINT_ATTRf(branch_sample_type
, p_branch_sample_type
);
1438 PRINT_ATTRf(sample_regs_user
, p_hex
);
1439 PRINT_ATTRf(sample_stack_user
, p_unsigned
);
1440 PRINT_ATTRf(clockid
, p_signed
);
1441 PRINT_ATTRf(sample_regs_intr
, p_hex
);
1442 PRINT_ATTRf(aux_watermark
, p_unsigned
);
1443 PRINT_ATTRf(sample_max_stack
, p_unsigned
);
1448 static int __open_attr__fprintf(FILE *fp
, const char *name
, const char *val
,
1449 void *priv __maybe_unused
)
1451 return fprintf(fp
, " %-32s %s\n", name
, val
);
1454 static bool ignore_missing_thread(struct perf_evsel
*evsel
,
1455 struct thread_map
*threads
,
1456 int thread
, int err
)
1458 if (!evsel
->ignore_missing_thread
)
1461 /* The system wide setup does not work with threads. */
1462 if (evsel
->system_wide
)
1465 /* The -ESRCH is perf event syscall errno for pid's not found. */
1469 /* If there's only one thread, let it fail. */
1470 if (threads
->nr
== 1)
1473 if (thread_map__remove(threads
, thread
))
1476 pr_warning("WARNING: Ignored open failure for pid %d\n",
1477 thread_map__pid(threads
, thread
));
1481 int perf_evsel__open(struct perf_evsel
*evsel
, struct cpu_map
*cpus
,
1482 struct thread_map
*threads
)
1484 int cpu
, thread
, nthreads
;
1485 unsigned long flags
= PERF_FLAG_FD_CLOEXEC
;
1487 enum { NO_CHANGE
, SET_TO_MAX
, INCREASED_MAX
} set_rlimit
= NO_CHANGE
;
1489 if (perf_missing_features
.write_backward
&& evsel
->attr
.write_backward
)
1493 static struct cpu_map
*empty_cpu_map
;
1495 if (empty_cpu_map
== NULL
) {
1496 empty_cpu_map
= cpu_map__dummy_new();
1497 if (empty_cpu_map
== NULL
)
1501 cpus
= empty_cpu_map
;
1504 if (threads
== NULL
) {
1505 static struct thread_map
*empty_thread_map
;
1507 if (empty_thread_map
== NULL
) {
1508 empty_thread_map
= thread_map__new_by_tid(-1);
1509 if (empty_thread_map
== NULL
)
1513 threads
= empty_thread_map
;
1516 if (evsel
->system_wide
)
1519 nthreads
= threads
->nr
;
1521 if (evsel
->fd
== NULL
&&
1522 perf_evsel__alloc_fd(evsel
, cpus
->nr
, nthreads
) < 0)
1526 flags
|= PERF_FLAG_PID_CGROUP
;
1527 pid
= evsel
->cgrp
->fd
;
1530 fallback_missing_features
:
1531 if (perf_missing_features
.clockid_wrong
)
1532 evsel
->attr
.clockid
= CLOCK_MONOTONIC
; /* should always work */
1533 if (perf_missing_features
.clockid
) {
1534 evsel
->attr
.use_clockid
= 0;
1535 evsel
->attr
.clockid
= 0;
1537 if (perf_missing_features
.cloexec
)
1538 flags
&= ~(unsigned long)PERF_FLAG_FD_CLOEXEC
;
1539 if (perf_missing_features
.mmap2
)
1540 evsel
->attr
.mmap2
= 0;
1541 if (perf_missing_features
.exclude_guest
)
1542 evsel
->attr
.exclude_guest
= evsel
->attr
.exclude_host
= 0;
1543 if (perf_missing_features
.lbr_flags
)
1544 evsel
->attr
.branch_sample_type
&= ~(PERF_SAMPLE_BRANCH_NO_FLAGS
|
1545 PERF_SAMPLE_BRANCH_NO_CYCLES
);
1547 if (perf_missing_features
.sample_id_all
)
1548 evsel
->attr
.sample_id_all
= 0;
1551 fprintf(stderr
, "%.60s\n", graph_dotted_line
);
1552 fprintf(stderr
, "perf_event_attr:\n");
1553 perf_event_attr__fprintf(stderr
, &evsel
->attr
, __open_attr__fprintf
, NULL
);
1554 fprintf(stderr
, "%.60s\n", graph_dotted_line
);
1557 for (cpu
= 0; cpu
< cpus
->nr
; cpu
++) {
1559 for (thread
= 0; thread
< nthreads
; thread
++) {
1562 if (!evsel
->cgrp
&& !evsel
->system_wide
)
1563 pid
= thread_map__pid(threads
, thread
);
1565 group_fd
= get_group_fd(evsel
, cpu
, thread
);
1567 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1568 pid
, cpus
->map
[cpu
], group_fd
, flags
);
1570 fd
= sys_perf_event_open(&evsel
->attr
, pid
, cpus
->map
[cpu
],
1573 FD(evsel
, cpu
, thread
) = fd
;
1578 if (ignore_missing_thread(evsel
, threads
, thread
, err
)) {
1580 * We just removed 1 thread, so take a step
1581 * back on thread index and lower the upper
1587 /* ... and pretend like nothing have happened. */
1592 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1597 pr_debug2(" = %d\n", fd
);
1599 if (evsel
->bpf_fd
>= 0) {
1601 int bpf_fd
= evsel
->bpf_fd
;
1604 PERF_EVENT_IOC_SET_BPF
,
1606 if (err
&& errno
!= EEXIST
) {
1607 pr_err("failed to attach bpf fd %d: %s\n",
1608 bpf_fd
, strerror(errno
));
1614 set_rlimit
= NO_CHANGE
;
1617 * If we succeeded but had to kill clockid, fail and
1618 * have perf_evsel__open_strerror() print us a nice
1621 if (perf_missing_features
.clockid
||
1622 perf_missing_features
.clockid_wrong
) {
1633 * perf stat needs between 5 and 22 fds per CPU. When we run out
1634 * of them try to increase the limits.
1636 if (err
== -EMFILE
&& set_rlimit
< INCREASED_MAX
) {
1638 int old_errno
= errno
;
1640 if (getrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1641 if (set_rlimit
== NO_CHANGE
)
1642 l
.rlim_cur
= l
.rlim_max
;
1644 l
.rlim_cur
= l
.rlim_max
+ 1000;
1645 l
.rlim_max
= l
.rlim_cur
;
1647 if (setrlimit(RLIMIT_NOFILE
, &l
) == 0) {
1656 if (err
!= -EINVAL
|| cpu
> 0 || thread
> 0)
1660 * Must probe features in the order they were added to the
1661 * perf_event_attr interface.
1663 if (!perf_missing_features
.write_backward
&& evsel
->attr
.write_backward
) {
1664 perf_missing_features
.write_backward
= true;
1666 } else if (!perf_missing_features
.clockid_wrong
&& evsel
->attr
.use_clockid
) {
1667 perf_missing_features
.clockid_wrong
= true;
1668 goto fallback_missing_features
;
1669 } else if (!perf_missing_features
.clockid
&& evsel
->attr
.use_clockid
) {
1670 perf_missing_features
.clockid
= true;
1671 goto fallback_missing_features
;
1672 } else if (!perf_missing_features
.cloexec
&& (flags
& PERF_FLAG_FD_CLOEXEC
)) {
1673 perf_missing_features
.cloexec
= true;
1674 goto fallback_missing_features
;
1675 } else if (!perf_missing_features
.mmap2
&& evsel
->attr
.mmap2
) {
1676 perf_missing_features
.mmap2
= true;
1677 goto fallback_missing_features
;
1678 } else if (!perf_missing_features
.exclude_guest
&&
1679 (evsel
->attr
.exclude_guest
|| evsel
->attr
.exclude_host
)) {
1680 perf_missing_features
.exclude_guest
= true;
1681 goto fallback_missing_features
;
1682 } else if (!perf_missing_features
.sample_id_all
) {
1683 perf_missing_features
.sample_id_all
= true;
1684 goto retry_sample_id
;
1685 } else if (!perf_missing_features
.lbr_flags
&&
1686 (evsel
->attr
.branch_sample_type
&
1687 (PERF_SAMPLE_BRANCH_NO_CYCLES
|
1688 PERF_SAMPLE_BRANCH_NO_FLAGS
))) {
1689 perf_missing_features
.lbr_flags
= true;
1690 goto fallback_missing_features
;
1694 while (--thread
>= 0) {
1695 close(FD(evsel
, cpu
, thread
));
1696 FD(evsel
, cpu
, thread
) = -1;
1699 } while (--cpu
>= 0);
1703 void perf_evsel__close(struct perf_evsel
*evsel
, int ncpus
, int nthreads
)
1705 if (evsel
->fd
== NULL
)
1708 perf_evsel__close_fd(evsel
, ncpus
, nthreads
);
1709 perf_evsel__free_fd(evsel
);
1712 int perf_evsel__open_per_cpu(struct perf_evsel
*evsel
,
1713 struct cpu_map
*cpus
)
1715 return perf_evsel__open(evsel
, cpus
, NULL
);
1718 int perf_evsel__open_per_thread(struct perf_evsel
*evsel
,
1719 struct thread_map
*threads
)
1721 return perf_evsel__open(evsel
, NULL
, threads
);
1724 static int perf_evsel__parse_id_sample(const struct perf_evsel
*evsel
,
1725 const union perf_event
*event
,
1726 struct perf_sample
*sample
)
1728 u64 type
= evsel
->attr
.sample_type
;
1729 const u64
*array
= event
->sample
.array
;
1730 bool swapped
= evsel
->needs_swap
;
1733 array
+= ((event
->header
.size
-
1734 sizeof(event
->header
)) / sizeof(u64
)) - 1;
1736 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1737 sample
->id
= *array
;
1741 if (type
& PERF_SAMPLE_CPU
) {
1744 /* undo swap of u64, then swap on individual u32s */
1745 u
.val64
= bswap_64(u
.val64
);
1746 u
.val32
[0] = bswap_32(u
.val32
[0]);
1749 sample
->cpu
= u
.val32
[0];
1753 if (type
& PERF_SAMPLE_STREAM_ID
) {
1754 sample
->stream_id
= *array
;
1758 if (type
& PERF_SAMPLE_ID
) {
1759 sample
->id
= *array
;
1763 if (type
& PERF_SAMPLE_TIME
) {
1764 sample
->time
= *array
;
1768 if (type
& PERF_SAMPLE_TID
) {
1771 /* undo swap of u64, then swap on individual u32s */
1772 u
.val64
= bswap_64(u
.val64
);
1773 u
.val32
[0] = bswap_32(u
.val32
[0]);
1774 u
.val32
[1] = bswap_32(u
.val32
[1]);
1777 sample
->pid
= u
.val32
[0];
1778 sample
->tid
= u
.val32
[1];
1785 static inline bool overflow(const void *endp
, u16 max_size
, const void *offset
,
1788 return size
> max_size
|| offset
+ size
> endp
;
1791 #define OVERFLOW_CHECK(offset, size, max_size) \
1793 if (overflow(endp, (max_size), (offset), (size))) \
1797 #define OVERFLOW_CHECK_u64(offset) \
1798 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1800 int perf_evsel__parse_sample(struct perf_evsel
*evsel
, union perf_event
*event
,
1801 struct perf_sample
*data
)
1803 u64 type
= evsel
->attr
.sample_type
;
1804 bool swapped
= evsel
->needs_swap
;
1806 u16 max_size
= event
->header
.size
;
1807 const void *endp
= (void *)event
+ max_size
;
1811 * used for cross-endian analysis. See git commit 65014ab3
1812 * for why this goofiness is needed.
1816 memset(data
, 0, sizeof(*data
));
1817 data
->cpu
= data
->pid
= data
->tid
= -1;
1818 data
->stream_id
= data
->id
= data
->time
= -1ULL;
1819 data
->period
= evsel
->attr
.sample_period
;
1820 data
->cpumode
= event
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
1822 if (event
->header
.type
!= PERF_RECORD_SAMPLE
) {
1823 if (!evsel
->attr
.sample_id_all
)
1825 return perf_evsel__parse_id_sample(evsel
, event
, data
);
1828 array
= event
->sample
.array
;
1831 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1832 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1833 * check the format does not go past the end of the event.
1835 if (evsel
->sample_size
+ sizeof(event
->header
) > event
->header
.size
)
1839 if (type
& PERF_SAMPLE_IDENTIFIER
) {
1844 if (type
& PERF_SAMPLE_IP
) {
1849 if (type
& PERF_SAMPLE_TID
) {
1852 /* undo swap of u64, then swap on individual u32s */
1853 u
.val64
= bswap_64(u
.val64
);
1854 u
.val32
[0] = bswap_32(u
.val32
[0]);
1855 u
.val32
[1] = bswap_32(u
.val32
[1]);
1858 data
->pid
= u
.val32
[0];
1859 data
->tid
= u
.val32
[1];
1863 if (type
& PERF_SAMPLE_TIME
) {
1864 data
->time
= *array
;
1869 if (type
& PERF_SAMPLE_ADDR
) {
1870 data
->addr
= *array
;
1874 if (type
& PERF_SAMPLE_ID
) {
1879 if (type
& PERF_SAMPLE_STREAM_ID
) {
1880 data
->stream_id
= *array
;
1884 if (type
& PERF_SAMPLE_CPU
) {
1888 /* undo swap of u64, then swap on individual u32s */
1889 u
.val64
= bswap_64(u
.val64
);
1890 u
.val32
[0] = bswap_32(u
.val32
[0]);
1893 data
->cpu
= u
.val32
[0];
1897 if (type
& PERF_SAMPLE_PERIOD
) {
1898 data
->period
= *array
;
1902 if (type
& PERF_SAMPLE_READ
) {
1903 u64 read_format
= evsel
->attr
.read_format
;
1905 OVERFLOW_CHECK_u64(array
);
1906 if (read_format
& PERF_FORMAT_GROUP
)
1907 data
->read
.group
.nr
= *array
;
1909 data
->read
.one
.value
= *array
;
1913 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
1914 OVERFLOW_CHECK_u64(array
);
1915 data
->read
.time_enabled
= *array
;
1919 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
1920 OVERFLOW_CHECK_u64(array
);
1921 data
->read
.time_running
= *array
;
1925 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1926 if (read_format
& PERF_FORMAT_GROUP
) {
1927 const u64 max_group_nr
= UINT64_MAX
/
1928 sizeof(struct sample_read_value
);
1930 if (data
->read
.group
.nr
> max_group_nr
)
1932 sz
= data
->read
.group
.nr
*
1933 sizeof(struct sample_read_value
);
1934 OVERFLOW_CHECK(array
, sz
, max_size
);
1935 data
->read
.group
.values
=
1936 (struct sample_read_value
*)array
;
1937 array
= (void *)array
+ sz
;
1939 OVERFLOW_CHECK_u64(array
);
1940 data
->read
.one
.id
= *array
;
1945 if (type
& PERF_SAMPLE_CALLCHAIN
) {
1946 const u64 max_callchain_nr
= UINT64_MAX
/ sizeof(u64
);
1948 OVERFLOW_CHECK_u64(array
);
1949 data
->callchain
= (struct ip_callchain
*)array
++;
1950 if (data
->callchain
->nr
> max_callchain_nr
)
1952 sz
= data
->callchain
->nr
* sizeof(u64
);
1953 OVERFLOW_CHECK(array
, sz
, max_size
);
1954 array
= (void *)array
+ sz
;
1957 if (type
& PERF_SAMPLE_RAW
) {
1958 OVERFLOW_CHECK_u64(array
);
1960 if (WARN_ONCE(swapped
,
1961 "Endianness of raw data not corrected!\n")) {
1962 /* undo swap of u64, then swap on individual u32s */
1963 u
.val64
= bswap_64(u
.val64
);
1964 u
.val32
[0] = bswap_32(u
.val32
[0]);
1965 u
.val32
[1] = bswap_32(u
.val32
[1]);
1967 data
->raw_size
= u
.val32
[0];
1968 array
= (void *)array
+ sizeof(u32
);
1970 OVERFLOW_CHECK(array
, data
->raw_size
, max_size
);
1971 data
->raw_data
= (void *)array
;
1972 array
= (void *)array
+ data
->raw_size
;
1975 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
1976 const u64 max_branch_nr
= UINT64_MAX
/
1977 sizeof(struct branch_entry
);
1979 OVERFLOW_CHECK_u64(array
);
1980 data
->branch_stack
= (struct branch_stack
*)array
++;
1982 if (data
->branch_stack
->nr
> max_branch_nr
)
1984 sz
= data
->branch_stack
->nr
* sizeof(struct branch_entry
);
1985 OVERFLOW_CHECK(array
, sz
, max_size
);
1986 array
= (void *)array
+ sz
;
1989 if (type
& PERF_SAMPLE_REGS_USER
) {
1990 OVERFLOW_CHECK_u64(array
);
1991 data
->user_regs
.abi
= *array
;
1994 if (data
->user_regs
.abi
) {
1995 u64 mask
= evsel
->attr
.sample_regs_user
;
1997 sz
= hweight_long(mask
) * sizeof(u64
);
1998 OVERFLOW_CHECK(array
, sz
, max_size
);
1999 data
->user_regs
.mask
= mask
;
2000 data
->user_regs
.regs
= (u64
*)array
;
2001 array
= (void *)array
+ sz
;
2005 if (type
& PERF_SAMPLE_STACK_USER
) {
2006 OVERFLOW_CHECK_u64(array
);
2009 data
->user_stack
.offset
= ((char *)(array
- 1)
2013 data
->user_stack
.size
= 0;
2015 OVERFLOW_CHECK(array
, sz
, max_size
);
2016 data
->user_stack
.data
= (char *)array
;
2017 array
= (void *)array
+ sz
;
2018 OVERFLOW_CHECK_u64(array
);
2019 data
->user_stack
.size
= *array
++;
2020 if (WARN_ONCE(data
->user_stack
.size
> sz
,
2021 "user stack dump failure\n"))
2026 if (type
& PERF_SAMPLE_WEIGHT
) {
2027 OVERFLOW_CHECK_u64(array
);
2028 data
->weight
= *array
;
2032 data
->data_src
= PERF_MEM_DATA_SRC_NONE
;
2033 if (type
& PERF_SAMPLE_DATA_SRC
) {
2034 OVERFLOW_CHECK_u64(array
);
2035 data
->data_src
= *array
;
2039 data
->transaction
= 0;
2040 if (type
& PERF_SAMPLE_TRANSACTION
) {
2041 OVERFLOW_CHECK_u64(array
);
2042 data
->transaction
= *array
;
2046 data
->intr_regs
.abi
= PERF_SAMPLE_REGS_ABI_NONE
;
2047 if (type
& PERF_SAMPLE_REGS_INTR
) {
2048 OVERFLOW_CHECK_u64(array
);
2049 data
->intr_regs
.abi
= *array
;
2052 if (data
->intr_regs
.abi
!= PERF_SAMPLE_REGS_ABI_NONE
) {
2053 u64 mask
= evsel
->attr
.sample_regs_intr
;
2055 sz
= hweight_long(mask
) * sizeof(u64
);
2056 OVERFLOW_CHECK(array
, sz
, max_size
);
2057 data
->intr_regs
.mask
= mask
;
2058 data
->intr_regs
.regs
= (u64
*)array
;
2059 array
= (void *)array
+ sz
;
2066 size_t perf_event__sample_event_size(const struct perf_sample
*sample
, u64 type
,
2069 size_t sz
, result
= sizeof(struct sample_event
);
2071 if (type
& PERF_SAMPLE_IDENTIFIER
)
2072 result
+= sizeof(u64
);
2074 if (type
& PERF_SAMPLE_IP
)
2075 result
+= sizeof(u64
);
2077 if (type
& PERF_SAMPLE_TID
)
2078 result
+= sizeof(u64
);
2080 if (type
& PERF_SAMPLE_TIME
)
2081 result
+= sizeof(u64
);
2083 if (type
& PERF_SAMPLE_ADDR
)
2084 result
+= sizeof(u64
);
2086 if (type
& PERF_SAMPLE_ID
)
2087 result
+= sizeof(u64
);
2089 if (type
& PERF_SAMPLE_STREAM_ID
)
2090 result
+= sizeof(u64
);
2092 if (type
& PERF_SAMPLE_CPU
)
2093 result
+= sizeof(u64
);
2095 if (type
& PERF_SAMPLE_PERIOD
)
2096 result
+= sizeof(u64
);
2098 if (type
& PERF_SAMPLE_READ
) {
2099 result
+= sizeof(u64
);
2100 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
)
2101 result
+= sizeof(u64
);
2102 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
)
2103 result
+= sizeof(u64
);
2104 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2105 if (read_format
& PERF_FORMAT_GROUP
) {
2106 sz
= sample
->read
.group
.nr
*
2107 sizeof(struct sample_read_value
);
2110 result
+= sizeof(u64
);
2114 if (type
& PERF_SAMPLE_CALLCHAIN
) {
2115 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
2119 if (type
& PERF_SAMPLE_RAW
) {
2120 result
+= sizeof(u32
);
2121 result
+= sample
->raw_size
;
2124 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
2125 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
2130 if (type
& PERF_SAMPLE_REGS_USER
) {
2131 if (sample
->user_regs
.abi
) {
2132 result
+= sizeof(u64
);
2133 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
2136 result
+= sizeof(u64
);
2140 if (type
& PERF_SAMPLE_STACK_USER
) {
2141 sz
= sample
->user_stack
.size
;
2142 result
+= sizeof(u64
);
2145 result
+= sizeof(u64
);
2149 if (type
& PERF_SAMPLE_WEIGHT
)
2150 result
+= sizeof(u64
);
2152 if (type
& PERF_SAMPLE_DATA_SRC
)
2153 result
+= sizeof(u64
);
2155 if (type
& PERF_SAMPLE_TRANSACTION
)
2156 result
+= sizeof(u64
);
2158 if (type
& PERF_SAMPLE_REGS_INTR
) {
2159 if (sample
->intr_regs
.abi
) {
2160 result
+= sizeof(u64
);
2161 sz
= hweight_long(sample
->intr_regs
.mask
) * sizeof(u64
);
2164 result
+= sizeof(u64
);
2171 int perf_event__synthesize_sample(union perf_event
*event
, u64 type
,
2173 const struct perf_sample
*sample
,
2179 * used for cross-endian analysis. See git commit 65014ab3
2180 * for why this goofiness is needed.
2184 array
= event
->sample
.array
;
2186 if (type
& PERF_SAMPLE_IDENTIFIER
) {
2187 *array
= sample
->id
;
2191 if (type
& PERF_SAMPLE_IP
) {
2192 *array
= sample
->ip
;
2196 if (type
& PERF_SAMPLE_TID
) {
2197 u
.val32
[0] = sample
->pid
;
2198 u
.val32
[1] = sample
->tid
;
2201 * Inverse of what is done in perf_evsel__parse_sample
2203 u
.val32
[0] = bswap_32(u
.val32
[0]);
2204 u
.val32
[1] = bswap_32(u
.val32
[1]);
2205 u
.val64
= bswap_64(u
.val64
);
2212 if (type
& PERF_SAMPLE_TIME
) {
2213 *array
= sample
->time
;
2217 if (type
& PERF_SAMPLE_ADDR
) {
2218 *array
= sample
->addr
;
2222 if (type
& PERF_SAMPLE_ID
) {
2223 *array
= sample
->id
;
2227 if (type
& PERF_SAMPLE_STREAM_ID
) {
2228 *array
= sample
->stream_id
;
2232 if (type
& PERF_SAMPLE_CPU
) {
2233 u
.val32
[0] = sample
->cpu
;
2236 * Inverse of what is done in perf_evsel__parse_sample
2238 u
.val32
[0] = bswap_32(u
.val32
[0]);
2239 u
.val64
= bswap_64(u
.val64
);
2245 if (type
& PERF_SAMPLE_PERIOD
) {
2246 *array
= sample
->period
;
2250 if (type
& PERF_SAMPLE_READ
) {
2251 if (read_format
& PERF_FORMAT_GROUP
)
2252 *array
= sample
->read
.group
.nr
;
2254 *array
= sample
->read
.one
.value
;
2257 if (read_format
& PERF_FORMAT_TOTAL_TIME_ENABLED
) {
2258 *array
= sample
->read
.time_enabled
;
2262 if (read_format
& PERF_FORMAT_TOTAL_TIME_RUNNING
) {
2263 *array
= sample
->read
.time_running
;
2267 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2268 if (read_format
& PERF_FORMAT_GROUP
) {
2269 sz
= sample
->read
.group
.nr
*
2270 sizeof(struct sample_read_value
);
2271 memcpy(array
, sample
->read
.group
.values
, sz
);
2272 array
= (void *)array
+ sz
;
2274 *array
= sample
->read
.one
.id
;
2279 if (type
& PERF_SAMPLE_CALLCHAIN
) {
2280 sz
= (sample
->callchain
->nr
+ 1) * sizeof(u64
);
2281 memcpy(array
, sample
->callchain
, sz
);
2282 array
= (void *)array
+ sz
;
2285 if (type
& PERF_SAMPLE_RAW
) {
2286 u
.val32
[0] = sample
->raw_size
;
2287 if (WARN_ONCE(swapped
,
2288 "Endianness of raw data not corrected!\n")) {
2290 * Inverse of what is done in perf_evsel__parse_sample
2292 u
.val32
[0] = bswap_32(u
.val32
[0]);
2293 u
.val32
[1] = bswap_32(u
.val32
[1]);
2294 u
.val64
= bswap_64(u
.val64
);
2297 array
= (void *)array
+ sizeof(u32
);
2299 memcpy(array
, sample
->raw_data
, sample
->raw_size
);
2300 array
= (void *)array
+ sample
->raw_size
;
2303 if (type
& PERF_SAMPLE_BRANCH_STACK
) {
2304 sz
= sample
->branch_stack
->nr
* sizeof(struct branch_entry
);
2306 memcpy(array
, sample
->branch_stack
, sz
);
2307 array
= (void *)array
+ sz
;
2310 if (type
& PERF_SAMPLE_REGS_USER
) {
2311 if (sample
->user_regs
.abi
) {
2312 *array
++ = sample
->user_regs
.abi
;
2313 sz
= hweight_long(sample
->user_regs
.mask
) * sizeof(u64
);
2314 memcpy(array
, sample
->user_regs
.regs
, sz
);
2315 array
= (void *)array
+ sz
;
2321 if (type
& PERF_SAMPLE_STACK_USER
) {
2322 sz
= sample
->user_stack
.size
;
2325 memcpy(array
, sample
->user_stack
.data
, sz
);
2326 array
= (void *)array
+ sz
;
2331 if (type
& PERF_SAMPLE_WEIGHT
) {
2332 *array
= sample
->weight
;
2336 if (type
& PERF_SAMPLE_DATA_SRC
) {
2337 *array
= sample
->data_src
;
2341 if (type
& PERF_SAMPLE_TRANSACTION
) {
2342 *array
= sample
->transaction
;
2346 if (type
& PERF_SAMPLE_REGS_INTR
) {
2347 if (sample
->intr_regs
.abi
) {
2348 *array
++ = sample
->intr_regs
.abi
;
2349 sz
= hweight_long(sample
->intr_regs
.mask
) * sizeof(u64
);
2350 memcpy(array
, sample
->intr_regs
.regs
, sz
);
2351 array
= (void *)array
+ sz
;
2360 struct format_field
*perf_evsel__field(struct perf_evsel
*evsel
, const char *name
)
2362 return pevent_find_field(evsel
->tp_format
, name
);
2365 void *perf_evsel__rawptr(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
2368 struct format_field
*field
= perf_evsel__field(evsel
, name
);
2374 offset
= field
->offset
;
2376 if (field
->flags
& FIELD_IS_DYNAMIC
) {
2377 offset
= *(int *)(sample
->raw_data
+ field
->offset
);
2381 return sample
->raw_data
+ offset
;
2384 u64
format_field__intval(struct format_field
*field
, struct perf_sample
*sample
,
2388 void *ptr
= sample
->raw_data
+ field
->offset
;
2390 switch (field
->size
) {
2394 value
= *(u16
*)ptr
;
2397 value
= *(u32
*)ptr
;
2400 memcpy(&value
, ptr
, sizeof(u64
));
2409 switch (field
->size
) {
2411 return bswap_16(value
);
2413 return bswap_32(value
);
2415 return bswap_64(value
);
2423 u64
perf_evsel__intval(struct perf_evsel
*evsel
, struct perf_sample
*sample
,
2426 struct format_field
*field
= perf_evsel__field(evsel
, name
);
2431 return field
? format_field__intval(field
, sample
, evsel
->needs_swap
) : 0;
2434 bool perf_evsel__fallback(struct perf_evsel
*evsel
, int err
,
2435 char *msg
, size_t msgsize
)
2439 if ((err
== ENOENT
|| err
== ENXIO
|| err
== ENODEV
) &&
2440 evsel
->attr
.type
== PERF_TYPE_HARDWARE
&&
2441 evsel
->attr
.config
== PERF_COUNT_HW_CPU_CYCLES
) {
2443 * If it's cycles then fall back to hrtimer based
2444 * cpu-clock-tick sw counter, which is always available even if
2447 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2450 scnprintf(msg
, msgsize
, "%s",
2451 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2453 evsel
->attr
.type
= PERF_TYPE_SOFTWARE
;
2454 evsel
->attr
.config
= PERF_COUNT_SW_CPU_CLOCK
;
2456 zfree(&evsel
->name
);
2458 } else if (err
== EACCES
&& !evsel
->attr
.exclude_kernel
&&
2459 (paranoid
= perf_event_paranoid()) > 1) {
2460 const char *name
= perf_evsel__name(evsel
);
2463 if (asprintf(&new_name
, "%s%su", name
, strchr(name
, ':') ? "" : ":") < 0)
2468 evsel
->name
= new_name
;
2469 scnprintf(msg
, msgsize
,
2470 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid
);
2471 evsel
->attr
.exclude_kernel
= 1;
2479 static bool find_process(const char *name
)
2481 size_t len
= strlen(name
);
2486 dir
= opendir(procfs__mountpoint());
2490 /* Walk through the directory. */
2491 while (ret
&& (d
= readdir(dir
)) != NULL
) {
2492 char path
[PATH_MAX
];
2496 if ((d
->d_type
!= DT_DIR
) ||
2497 !strcmp(".", d
->d_name
) ||
2498 !strcmp("..", d
->d_name
))
2501 scnprintf(path
, sizeof(path
), "%s/%s/comm",
2502 procfs__mountpoint(), d
->d_name
);
2504 if (filename__read_str(path
, &data
, &size
))
2507 ret
= strncmp(name
, data
, len
);
2512 return ret
? false : true;
2515 int perf_evsel__open_strerror(struct perf_evsel
*evsel
, struct target
*target
,
2516 int err
, char *msg
, size_t size
)
2518 char sbuf
[STRERR_BUFSIZE
];
2525 printed
= scnprintf(msg
, size
,
2526 "No permission to enable %s event.\n\n",
2527 perf_evsel__name(evsel
));
2529 return scnprintf(msg
+ printed
, size
- printed
,
2530 "You may not have permission to collect %sstats.\n\n"
2531 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2532 "which controls use of the performance events system by\n"
2533 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2534 "The current value is %d:\n\n"
2535 " -1: Allow use of (almost) all events by all users\n"
2536 ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2537 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2538 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2539 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2540 " kernel.perf_event_paranoid = -1\n" ,
2541 target
->system_wide
? "system-wide " : "",
2542 perf_event_paranoid());
2544 return scnprintf(msg
, size
, "The %s event is not supported.",
2545 perf_evsel__name(evsel
));
2547 return scnprintf(msg
, size
, "%s",
2548 "Too many events are opened.\n"
2549 "Probably the maximum number of open file descriptors has been reached.\n"
2550 "Hint: Try again after reducing the number of events.\n"
2551 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2553 if ((evsel
->attr
.sample_type
& PERF_SAMPLE_CALLCHAIN
) != 0 &&
2554 access("/proc/sys/kernel/perf_event_max_stack", F_OK
) == 0)
2555 return scnprintf(msg
, size
,
2556 "Not enough memory to setup event with callchain.\n"
2557 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2558 "Hint: Current value: %d", sysctl_perf_event_max_stack
);
2561 if (target
->cpu_list
)
2562 return scnprintf(msg
, size
, "%s",
2563 "No such device - did you specify an out-of-range profile CPU?");
2566 if (evsel
->attr
.sample_period
!= 0)
2567 return scnprintf(msg
, size
, "%s",
2568 "PMU Hardware doesn't support sampling/overflow-interrupts.");
2569 if (evsel
->attr
.precise_ip
)
2570 return scnprintf(msg
, size
, "%s",
2571 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2572 #if defined(__i386__) || defined(__x86_64__)
2573 if (evsel
->attr
.type
== PERF_TYPE_HARDWARE
)
2574 return scnprintf(msg
, size
, "%s",
2575 "No hardware sampling interrupt available.\n"
2576 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2580 if (find_process("oprofiled"))
2581 return scnprintf(msg
, size
,
2582 "The PMU counters are busy/taken by another profiler.\n"
2583 "We found oprofile daemon running, please stop it and try again.");
2586 if (evsel
->attr
.write_backward
&& perf_missing_features
.write_backward
)
2587 return scnprintf(msg
, size
, "Reading from overwrite event is not supported by this kernel.");
2588 if (perf_missing_features
.clockid
)
2589 return scnprintf(msg
, size
, "clockid feature not supported.");
2590 if (perf_missing_features
.clockid_wrong
)
2591 return scnprintf(msg
, size
, "wrong clockid (%d).", clockid
);
2597 return scnprintf(msg
, size
,
2598 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2599 "/bin/dmesg may provide additional information.\n"
2600 "No CONFIG_PERF_EVENTS=y kernel support configured?",
2601 err
, str_error_r(err
, sbuf
, sizeof(sbuf
)),
2602 perf_evsel__name(evsel
));
2605 char *perf_evsel__env_arch(struct perf_evsel
*evsel
)
2607 if (evsel
&& evsel
->evlist
&& evsel
->evlist
->env
)
2608 return evsel
->evlist
->env
->arch
;