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/tracing_path.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <linux/err.h>
17 #include <sys/resource.h>
19 #include "callchain.h"
25 #include "thread_map.h"
27 #include "perf_regs.h"
29 #include "trace-event.h"
39 } perf_missing_features;
41 static clockid_t clockid;
43 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
48 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
54 int (*init)(struct perf_evsel *evsel);
55 void (*fini)(struct perf_evsel *evsel);
56 } perf_evsel__object = {
57 .size = sizeof(struct perf_evsel),
58 .init = perf_evsel__no_extra_init,
59 .fini = perf_evsel__no_extra_fini,
62 int perf_evsel__object_config(size_t object_size,
63 int (*init)(struct perf_evsel *evsel),
64 void (*fini)(struct perf_evsel *evsel))
70 if (perf_evsel__object.size > object_size)
73 perf_evsel__object.size = object_size;
77 perf_evsel__object.init = init;
80 perf_evsel__object.fini = fini;
85 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
87 int __perf_evsel__sample_size(u64 sample_type)
89 u64 mask = sample_type & PERF_SAMPLE_MASK;
93 for (i = 0; i < 64; i++) {
94 if (mask & (1ULL << i))
104 * __perf_evsel__calc_id_pos - calculate id_pos.
105 * @sample_type: sample type
107 * This function returns the position of the event id (PERF_SAMPLE_ID or
108 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
111 static int __perf_evsel__calc_id_pos(u64 sample_type)
115 if (sample_type & PERF_SAMPLE_IDENTIFIER)
118 if (!(sample_type & PERF_SAMPLE_ID))
121 if (sample_type & PERF_SAMPLE_IP)
124 if (sample_type & PERF_SAMPLE_TID)
127 if (sample_type & PERF_SAMPLE_TIME)
130 if (sample_type & PERF_SAMPLE_ADDR)
137 * __perf_evsel__calc_is_pos - calculate is_pos.
138 * @sample_type: sample type
140 * This function returns the position (counting backwards) of the event id
141 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
142 * sample_id_all is used there is an id sample appended to non-sample events.
144 static int __perf_evsel__calc_is_pos(u64 sample_type)
148 if (sample_type & PERF_SAMPLE_IDENTIFIER)
151 if (!(sample_type & PERF_SAMPLE_ID))
154 if (sample_type & PERF_SAMPLE_CPU)
157 if (sample_type & PERF_SAMPLE_STREAM_ID)
163 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
165 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
166 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
169 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
170 enum perf_event_sample_format bit)
172 if (!(evsel->attr.sample_type & bit)) {
173 evsel->attr.sample_type |= bit;
174 evsel->sample_size += sizeof(u64);
175 perf_evsel__calc_id_pos(evsel);
179 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
180 enum perf_event_sample_format bit)
182 if (evsel->attr.sample_type & bit) {
183 evsel->attr.sample_type &= ~bit;
184 evsel->sample_size -= sizeof(u64);
185 perf_evsel__calc_id_pos(evsel);
189 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
190 bool can_sample_identifier)
192 if (can_sample_identifier) {
193 perf_evsel__reset_sample_bit(evsel, ID);
194 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
196 perf_evsel__set_sample_bit(evsel, ID);
198 evsel->attr.read_format |= PERF_FORMAT_ID;
201 void perf_evsel__init(struct perf_evsel *evsel,
202 struct perf_event_attr *attr, int idx)
205 evsel->tracking = !idx;
207 evsel->leader = evsel;
210 evsel->evlist = NULL;
211 INIT_LIST_HEAD(&evsel->node);
212 INIT_LIST_HEAD(&evsel->config_terms);
213 perf_evsel__object.init(evsel);
214 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
215 perf_evsel__calc_id_pos(evsel);
216 evsel->cmdline_group_boundary = false;
219 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
221 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
224 perf_evsel__init(evsel, attr, idx);
230 * Returns pointer with encoded error via <linux/err.h> interface.
232 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
234 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
240 struct perf_event_attr attr = {
241 .type = PERF_TYPE_TRACEPOINT,
242 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
243 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
246 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
249 evsel->tp_format = trace_event__tp_format(sys, name);
250 if (IS_ERR(evsel->tp_format)) {
251 err = PTR_ERR(evsel->tp_format);
255 event_attr_init(&attr);
256 attr.config = evsel->tp_format->id;
257 attr.sample_period = 1;
258 perf_evsel__init(evsel, &attr, idx);
270 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
278 "stalled-cycles-frontend",
279 "stalled-cycles-backend",
283 static const char *__perf_evsel__hw_name(u64 config)
285 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
286 return perf_evsel__hw_names[config];
288 return "unknown-hardware";
291 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
293 int colon = 0, r = 0;
294 struct perf_event_attr *attr = &evsel->attr;
295 bool exclude_guest_default = false;
297 #define MOD_PRINT(context, mod) do { \
298 if (!attr->exclude_##context) { \
299 if (!colon) colon = ++r; \
300 r += scnprintf(bf + r, size - r, "%c", mod); \
303 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
304 MOD_PRINT(kernel, 'k');
305 MOD_PRINT(user, 'u');
307 exclude_guest_default = true;
310 if (attr->precise_ip) {
313 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
314 exclude_guest_default = true;
317 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
318 MOD_PRINT(host, 'H');
319 MOD_PRINT(guest, 'G');
327 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
329 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
330 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
333 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
346 static const char *__perf_evsel__sw_name(u64 config)
348 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
349 return perf_evsel__sw_names[config];
350 return "unknown-software";
353 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
355 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
356 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
359 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
363 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
365 if (type & HW_BREAKPOINT_R)
366 r += scnprintf(bf + r, size - r, "r");
368 if (type & HW_BREAKPOINT_W)
369 r += scnprintf(bf + r, size - r, "w");
371 if (type & HW_BREAKPOINT_X)
372 r += scnprintf(bf + r, size - r, "x");
377 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
379 struct perf_event_attr *attr = &evsel->attr;
380 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
381 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
384 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
385 [PERF_EVSEL__MAX_ALIASES] = {
386 { "L1-dcache", "l1-d", "l1d", "L1-data", },
387 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
389 { "dTLB", "d-tlb", "Data-TLB", },
390 { "iTLB", "i-tlb", "Instruction-TLB", },
391 { "branch", "branches", "bpu", "btb", "bpc", },
395 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
396 [PERF_EVSEL__MAX_ALIASES] = {
397 { "load", "loads", "read", },
398 { "store", "stores", "write", },
399 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
402 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
403 [PERF_EVSEL__MAX_ALIASES] = {
404 { "refs", "Reference", "ops", "access", },
405 { "misses", "miss", },
408 #define C(x) PERF_COUNT_HW_CACHE_##x
409 #define CACHE_READ (1 << C(OP_READ))
410 #define CACHE_WRITE (1 << C(OP_WRITE))
411 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
412 #define COP(x) (1 << x)
415 * cache operartion stat
416 * L1I : Read and prefetch only
417 * ITLB and BPU : Read-only
419 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
420 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
421 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
422 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
423 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
424 [C(ITLB)] = (CACHE_READ),
425 [C(BPU)] = (CACHE_READ),
426 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
429 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
431 if (perf_evsel__hw_cache_stat[type] & COP(op))
432 return true; /* valid */
434 return false; /* invalid */
437 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
438 char *bf, size_t size)
441 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
442 perf_evsel__hw_cache_op[op][0],
443 perf_evsel__hw_cache_result[result][0]);
446 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
447 perf_evsel__hw_cache_op[op][1]);
450 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
452 u8 op, result, type = (config >> 0) & 0xff;
453 const char *err = "unknown-ext-hardware-cache-type";
455 if (type > PERF_COUNT_HW_CACHE_MAX)
458 op = (config >> 8) & 0xff;
459 err = "unknown-ext-hardware-cache-op";
460 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
463 result = (config >> 16) & 0xff;
464 err = "unknown-ext-hardware-cache-result";
465 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
468 err = "invalid-cache";
469 if (!perf_evsel__is_cache_op_valid(type, op))
472 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
474 return scnprintf(bf, size, "%s", err);
477 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
479 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
480 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
483 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
485 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
486 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
489 const char *perf_evsel__name(struct perf_evsel *evsel)
496 switch (evsel->attr.type) {
498 perf_evsel__raw_name(evsel, bf, sizeof(bf));
501 case PERF_TYPE_HARDWARE:
502 perf_evsel__hw_name(evsel, bf, sizeof(bf));
505 case PERF_TYPE_HW_CACHE:
506 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
509 case PERF_TYPE_SOFTWARE:
510 perf_evsel__sw_name(evsel, bf, sizeof(bf));
513 case PERF_TYPE_TRACEPOINT:
514 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
517 case PERF_TYPE_BREAKPOINT:
518 perf_evsel__bp_name(evsel, bf, sizeof(bf));
522 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
527 evsel->name = strdup(bf);
529 return evsel->name ?: "unknown";
532 const char *perf_evsel__group_name(struct perf_evsel *evsel)
534 return evsel->group_name ?: "anon group";
537 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
540 struct perf_evsel *pos;
541 const char *group_name = perf_evsel__group_name(evsel);
543 ret = scnprintf(buf, size, "%s", group_name);
545 ret += scnprintf(buf + ret, size - ret, " { %s",
546 perf_evsel__name(evsel));
548 for_each_group_member(pos, evsel)
549 ret += scnprintf(buf + ret, size - ret, ", %s",
550 perf_evsel__name(pos));
552 ret += scnprintf(buf + ret, size - ret, " }");
558 perf_evsel__config_callgraph(struct perf_evsel *evsel,
559 struct record_opts *opts,
560 struct callchain_param *param)
562 bool function = perf_evsel__is_function_event(evsel);
563 struct perf_event_attr *attr = &evsel->attr;
565 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
567 if (param->record_mode == CALLCHAIN_LBR) {
568 if (!opts->branch_stack) {
569 if (attr->exclude_user) {
570 pr_warning("LBR callstack option is only available "
571 "to get user callchain information. "
572 "Falling back to framepointers.\n");
574 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
575 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
576 PERF_SAMPLE_BRANCH_CALL_STACK;
579 pr_warning("Cannot use LBR callstack with branch stack. "
580 "Falling back to framepointers.\n");
583 if (param->record_mode == CALLCHAIN_DWARF) {
585 perf_evsel__set_sample_bit(evsel, REGS_USER);
586 perf_evsel__set_sample_bit(evsel, STACK_USER);
587 attr->sample_regs_user = PERF_REGS_MASK;
588 attr->sample_stack_user = param->dump_size;
589 attr->exclude_callchain_user = 1;
591 pr_info("Cannot use DWARF unwind for function trace event,"
592 " falling back to framepointers.\n");
597 pr_info("Disabling user space callchains for function trace event.\n");
598 attr->exclude_callchain_user = 1;
603 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
604 struct callchain_param *param)
606 struct perf_event_attr *attr = &evsel->attr;
608 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
609 if (param->record_mode == CALLCHAIN_LBR) {
610 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
611 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
612 PERF_SAMPLE_BRANCH_CALL_STACK);
614 if (param->record_mode == CALLCHAIN_DWARF) {
615 perf_evsel__reset_sample_bit(evsel, REGS_USER);
616 perf_evsel__reset_sample_bit(evsel, STACK_USER);
620 static void apply_config_terms(struct perf_evsel *evsel,
621 struct record_opts *opts)
623 struct perf_evsel_config_term *term;
624 struct list_head *config_terms = &evsel->config_terms;
625 struct perf_event_attr *attr = &evsel->attr;
626 struct callchain_param param;
628 char *callgraph_buf = NULL;
630 /* callgraph default */
631 param.record_mode = callchain_param.record_mode;
633 list_for_each_entry(term, config_terms, list) {
634 switch (term->type) {
635 case PERF_EVSEL__CONFIG_TERM_PERIOD:
636 attr->sample_period = term->val.period;
639 case PERF_EVSEL__CONFIG_TERM_FREQ:
640 attr->sample_freq = term->val.freq;
643 case PERF_EVSEL__CONFIG_TERM_TIME:
645 perf_evsel__set_sample_bit(evsel, TIME);
647 perf_evsel__reset_sample_bit(evsel, TIME);
649 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
650 callgraph_buf = term->val.callgraph;
652 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
653 dump_size = term->val.stack_user;
660 /* User explicitly set per-event callgraph, clear the old setting and reset. */
661 if ((callgraph_buf != NULL) || (dump_size > 0)) {
663 /* parse callgraph parameters */
664 if (callgraph_buf != NULL) {
665 if (!strcmp(callgraph_buf, "no")) {
666 param.enabled = false;
667 param.record_mode = CALLCHAIN_NONE;
669 param.enabled = true;
670 if (parse_callchain_record(callgraph_buf, ¶m)) {
671 pr_err("per-event callgraph setting for %s failed. "
672 "Apply callgraph global setting for it\n",
679 dump_size = round_up(dump_size, sizeof(u64));
680 param.dump_size = dump_size;
683 /* If global callgraph set, clear it */
684 if (callchain_param.enabled)
685 perf_evsel__reset_callgraph(evsel, &callchain_param);
687 /* set perf-event callgraph */
689 perf_evsel__config_callgraph(evsel, opts, ¶m);
694 * The enable_on_exec/disabled value strategy:
696 * 1) For any type of traced program:
697 * - all independent events and group leaders are disabled
698 * - all group members are enabled
700 * Group members are ruled by group leaders. They need to
701 * be enabled, because the group scheduling relies on that.
703 * 2) For traced programs executed by perf:
704 * - all independent events and group leaders have
706 * - we don't specifically enable or disable any event during
709 * Independent events and group leaders are initially disabled
710 * and get enabled by exec. Group members are ruled by group
711 * leaders as stated in 1).
713 * 3) For traced programs attached by perf (pid/tid):
714 * - we specifically enable or disable all events during
717 * When attaching events to already running traced we
718 * enable/disable events specifically, as there's no
719 * initial traced exec call.
721 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
723 struct perf_evsel *leader = evsel->leader;
724 struct perf_event_attr *attr = &evsel->attr;
725 int track = evsel->tracking;
726 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
728 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
729 attr->inherit = !opts->no_inherit;
731 perf_evsel__set_sample_bit(evsel, IP);
732 perf_evsel__set_sample_bit(evsel, TID);
734 if (evsel->sample_read) {
735 perf_evsel__set_sample_bit(evsel, READ);
738 * We need ID even in case of single event, because
739 * PERF_SAMPLE_READ process ID specific data.
741 perf_evsel__set_sample_id(evsel, false);
744 * Apply group format only if we belong to group
745 * with more than one members.
747 if (leader->nr_members > 1) {
748 attr->read_format |= PERF_FORMAT_GROUP;
754 * We default some events to have a default interval. But keep
755 * it a weak assumption overridable by the user.
757 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
758 opts->user_interval != ULLONG_MAX)) {
760 perf_evsel__set_sample_bit(evsel, PERIOD);
762 attr->sample_freq = opts->freq;
764 attr->sample_period = opts->default_interval;
769 * Disable sampling for all group members other
770 * than leader in case leader 'leads' the sampling.
772 if ((leader != evsel) && leader->sample_read) {
773 attr->sample_freq = 0;
774 attr->sample_period = 0;
777 if (opts->no_samples)
778 attr->sample_freq = 0;
780 if (opts->inherit_stat)
781 attr->inherit_stat = 1;
783 if (opts->sample_address) {
784 perf_evsel__set_sample_bit(evsel, ADDR);
785 attr->mmap_data = track;
789 * We don't allow user space callchains for function trace
790 * event, due to issues with page faults while tracing page
791 * fault handler and its overall trickiness nature.
793 if (perf_evsel__is_function_event(evsel))
794 evsel->attr.exclude_callchain_user = 1;
796 if (callchain_param.enabled && !evsel->no_aux_samples)
797 perf_evsel__config_callgraph(evsel, opts, &callchain_param);
799 if (opts->sample_intr_regs) {
800 attr->sample_regs_intr = opts->sample_intr_regs;
801 perf_evsel__set_sample_bit(evsel, REGS_INTR);
804 if (target__has_cpu(&opts->target))
805 perf_evsel__set_sample_bit(evsel, CPU);
808 perf_evsel__set_sample_bit(evsel, PERIOD);
811 * When the user explicitely disabled time don't force it here.
813 if (opts->sample_time &&
814 (!perf_missing_features.sample_id_all &&
815 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
816 opts->sample_time_set)))
817 perf_evsel__set_sample_bit(evsel, TIME);
819 if (opts->raw_samples && !evsel->no_aux_samples) {
820 perf_evsel__set_sample_bit(evsel, TIME);
821 perf_evsel__set_sample_bit(evsel, RAW);
822 perf_evsel__set_sample_bit(evsel, CPU);
825 if (opts->sample_address)
826 perf_evsel__set_sample_bit(evsel, DATA_SRC);
828 if (opts->no_buffering) {
830 attr->wakeup_events = 1;
832 if (opts->branch_stack && !evsel->no_aux_samples) {
833 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
834 attr->branch_sample_type = opts->branch_stack;
837 if (opts->sample_weight)
838 perf_evsel__set_sample_bit(evsel, WEIGHT);
842 attr->mmap2 = track && !perf_missing_features.mmap2;
845 if (opts->record_switch_events)
846 attr->context_switch = track;
848 if (opts->sample_transaction)
849 perf_evsel__set_sample_bit(evsel, TRANSACTION);
851 if (opts->running_time) {
852 evsel->attr.read_format |=
853 PERF_FORMAT_TOTAL_TIME_ENABLED |
854 PERF_FORMAT_TOTAL_TIME_RUNNING;
858 * XXX see the function comment above
860 * Disabling only independent events or group leaders,
861 * keeping group members enabled.
863 if (perf_evsel__is_group_leader(evsel))
867 * Setting enable_on_exec for independent events and
868 * group leaders for traced executed by perf.
870 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
871 !opts->initial_delay)
872 attr->enable_on_exec = 1;
874 if (evsel->immediate) {
876 attr->enable_on_exec = 0;
879 clockid = opts->clockid;
880 if (opts->use_clockid) {
881 attr->use_clockid = 1;
882 attr->clockid = opts->clockid;
886 * Apply event specific term settings,
887 * it overloads any global configuration.
889 apply_config_terms(evsel, opts);
892 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
896 if (evsel->system_wide)
899 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
902 for (cpu = 0; cpu < ncpus; cpu++) {
903 for (thread = 0; thread < nthreads; thread++) {
904 FD(evsel, cpu, thread) = -1;
909 return evsel->fd != NULL ? 0 : -ENOMEM;
912 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
917 if (evsel->system_wide)
920 for (cpu = 0; cpu < ncpus; cpu++) {
921 for (thread = 0; thread < nthreads; thread++) {
922 int fd = FD(evsel, cpu, thread),
923 err = ioctl(fd, ioc, arg);
933 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
936 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
937 PERF_EVENT_IOC_SET_FILTER,
941 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
943 char *new_filter = strdup(filter);
945 if (new_filter != NULL) {
947 evsel->filter = new_filter;
954 int perf_evsel__append_filter(struct perf_evsel *evsel,
955 const char *op, const char *filter)
959 if (evsel->filter == NULL)
960 return perf_evsel__set_filter(evsel, filter);
962 if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
964 evsel->filter = new_filter;
971 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
973 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
974 PERF_EVENT_IOC_ENABLE,
978 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
980 if (ncpus == 0 || nthreads == 0)
983 if (evsel->system_wide)
986 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
987 if (evsel->sample_id == NULL)
990 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
991 if (evsel->id == NULL) {
992 xyarray__delete(evsel->sample_id);
993 evsel->sample_id = NULL;
1000 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1002 xyarray__delete(evsel->fd);
1006 static void perf_evsel__free_id(struct perf_evsel *evsel)
1008 xyarray__delete(evsel->sample_id);
1009 evsel->sample_id = NULL;
1013 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1015 struct perf_evsel_config_term *term, *h;
1017 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1018 list_del(&term->list);
1023 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1027 if (evsel->system_wide)
1030 for (cpu = 0; cpu < ncpus; cpu++)
1031 for (thread = 0; thread < nthreads; ++thread) {
1032 close(FD(evsel, cpu, thread));
1033 FD(evsel, cpu, thread) = -1;
1037 void perf_evsel__exit(struct perf_evsel *evsel)
1039 assert(list_empty(&evsel->node));
1040 assert(evsel->evlist == NULL);
1041 perf_evsel__free_fd(evsel);
1042 perf_evsel__free_id(evsel);
1043 perf_evsel__free_config_terms(evsel);
1044 close_cgroup(evsel->cgrp);
1045 cpu_map__put(evsel->cpus);
1046 cpu_map__put(evsel->own_cpus);
1047 thread_map__put(evsel->threads);
1048 zfree(&evsel->group_name);
1049 zfree(&evsel->name);
1050 perf_evsel__object.fini(evsel);
1053 void perf_evsel__delete(struct perf_evsel *evsel)
1055 perf_evsel__exit(evsel);
1059 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1060 struct perf_counts_values *count)
1062 struct perf_counts_values tmp;
1064 if (!evsel->prev_raw_counts)
1068 tmp = evsel->prev_raw_counts->aggr;
1069 evsel->prev_raw_counts->aggr = *count;
1071 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1072 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1075 count->val = count->val - tmp.val;
1076 count->ena = count->ena - tmp.ena;
1077 count->run = count->run - tmp.run;
1080 void perf_counts_values__scale(struct perf_counts_values *count,
1081 bool scale, s8 *pscaled)
1086 if (count->run == 0) {
1089 } else if (count->run < count->ena) {
1091 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1094 count->ena = count->run = 0;
1100 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1101 struct perf_counts_values *count)
1103 memset(count, 0, sizeof(*count));
1105 if (FD(evsel, cpu, thread) < 0)
1108 if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1114 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1115 int cpu, int thread, bool scale)
1117 struct perf_counts_values count;
1118 size_t nv = scale ? 3 : 1;
1120 if (FD(evsel, cpu, thread) < 0)
1123 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1126 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1129 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1130 perf_counts_values__scale(&count, scale, NULL);
1131 *perf_counts(evsel->counts, cpu, thread) = count;
1135 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1137 struct perf_evsel *leader = evsel->leader;
1140 if (perf_evsel__is_group_leader(evsel))
1144 * Leader must be already processed/open,
1145 * if not it's a bug.
1147 BUG_ON(!leader->fd);
1149 fd = FD(leader, cpu, thread);
1160 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1162 bool first_bit = true;
1166 if (value & bits[i].bit) {
1167 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1170 } while (bits[++i].name != NULL);
1173 static void __p_sample_type(char *buf, size_t size, u64 value)
1175 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1176 struct bit_names bits[] = {
1177 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1178 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1179 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1180 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1181 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1185 __p_bits(buf, size, value, bits);
1188 static void __p_read_format(char *buf, size_t size, u64 value)
1190 #define bit_name(n) { PERF_FORMAT_##n, #n }
1191 struct bit_names bits[] = {
1192 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1193 bit_name(ID), bit_name(GROUP),
1197 __p_bits(buf, size, value, bits);
1200 #define BUF_SIZE 1024
1202 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1203 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1204 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1205 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1206 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1208 #define PRINT_ATTRn(_n, _f, _p) \
1212 ret += attr__fprintf(fp, _n, buf, priv);\
1216 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1218 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1219 attr__fprintf_f attr__fprintf, void *priv)
1224 PRINT_ATTRf(type, p_unsigned);
1225 PRINT_ATTRf(size, p_unsigned);
1226 PRINT_ATTRf(config, p_hex);
1227 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1228 PRINT_ATTRf(sample_type, p_sample_type);
1229 PRINT_ATTRf(read_format, p_read_format);
1231 PRINT_ATTRf(disabled, p_unsigned);
1232 PRINT_ATTRf(inherit, p_unsigned);
1233 PRINT_ATTRf(pinned, p_unsigned);
1234 PRINT_ATTRf(exclusive, p_unsigned);
1235 PRINT_ATTRf(exclude_user, p_unsigned);
1236 PRINT_ATTRf(exclude_kernel, p_unsigned);
1237 PRINT_ATTRf(exclude_hv, p_unsigned);
1238 PRINT_ATTRf(exclude_idle, p_unsigned);
1239 PRINT_ATTRf(mmap, p_unsigned);
1240 PRINT_ATTRf(comm, p_unsigned);
1241 PRINT_ATTRf(freq, p_unsigned);
1242 PRINT_ATTRf(inherit_stat, p_unsigned);
1243 PRINT_ATTRf(enable_on_exec, p_unsigned);
1244 PRINT_ATTRf(task, p_unsigned);
1245 PRINT_ATTRf(watermark, p_unsigned);
1246 PRINT_ATTRf(precise_ip, p_unsigned);
1247 PRINT_ATTRf(mmap_data, p_unsigned);
1248 PRINT_ATTRf(sample_id_all, p_unsigned);
1249 PRINT_ATTRf(exclude_host, p_unsigned);
1250 PRINT_ATTRf(exclude_guest, p_unsigned);
1251 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1252 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1253 PRINT_ATTRf(mmap2, p_unsigned);
1254 PRINT_ATTRf(comm_exec, p_unsigned);
1255 PRINT_ATTRf(use_clockid, p_unsigned);
1256 PRINT_ATTRf(context_switch, p_unsigned);
1258 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1259 PRINT_ATTRf(bp_type, p_unsigned);
1260 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1261 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1262 PRINT_ATTRf(sample_regs_user, p_hex);
1263 PRINT_ATTRf(sample_stack_user, p_unsigned);
1264 PRINT_ATTRf(clockid, p_signed);
1265 PRINT_ATTRf(sample_regs_intr, p_hex);
1266 PRINT_ATTRf(aux_watermark, p_unsigned);
1271 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1272 void *priv __attribute__((unused)))
1274 return fprintf(fp, " %-32s %s\n", name, val);
1277 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1278 struct thread_map *threads)
1280 int cpu, thread, nthreads;
1281 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1283 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1285 if (evsel->system_wide)
1288 nthreads = threads->nr;
1290 if (evsel->fd == NULL &&
1291 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1295 flags |= PERF_FLAG_PID_CGROUP;
1296 pid = evsel->cgrp->fd;
1299 fallback_missing_features:
1300 if (perf_missing_features.clockid_wrong)
1301 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1302 if (perf_missing_features.clockid) {
1303 evsel->attr.use_clockid = 0;
1304 evsel->attr.clockid = 0;
1306 if (perf_missing_features.cloexec)
1307 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1308 if (perf_missing_features.mmap2)
1309 evsel->attr.mmap2 = 0;
1310 if (perf_missing_features.exclude_guest)
1311 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1313 if (perf_missing_features.sample_id_all)
1314 evsel->attr.sample_id_all = 0;
1317 fprintf(stderr, "%.60s\n", graph_dotted_line);
1318 fprintf(stderr, "perf_event_attr:\n");
1319 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1320 fprintf(stderr, "%.60s\n", graph_dotted_line);
1323 for (cpu = 0; cpu < cpus->nr; cpu++) {
1325 for (thread = 0; thread < nthreads; thread++) {
1328 if (!evsel->cgrp && !evsel->system_wide)
1329 pid = thread_map__pid(threads, thread);
1331 group_fd = get_group_fd(evsel, cpu, thread);
1333 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1334 pid, cpus->map[cpu], group_fd, flags);
1336 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1340 if (FD(evsel, cpu, thread) < 0) {
1342 pr_debug2("sys_perf_event_open failed, error %d\n",
1346 set_rlimit = NO_CHANGE;
1349 * If we succeeded but had to kill clockid, fail and
1350 * have perf_evsel__open_strerror() print us a nice
1353 if (perf_missing_features.clockid ||
1354 perf_missing_features.clockid_wrong) {
1365 * perf stat needs between 5 and 22 fds per CPU. When we run out
1366 * of them try to increase the limits.
1368 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1370 int old_errno = errno;
1372 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1373 if (set_rlimit == NO_CHANGE)
1374 l.rlim_cur = l.rlim_max;
1376 l.rlim_cur = l.rlim_max + 1000;
1377 l.rlim_max = l.rlim_cur;
1379 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1388 if (err != -EINVAL || cpu > 0 || thread > 0)
1392 * Must probe features in the order they were added to the
1393 * perf_event_attr interface.
1395 if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1396 perf_missing_features.clockid_wrong = true;
1397 goto fallback_missing_features;
1398 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1399 perf_missing_features.clockid = true;
1400 goto fallback_missing_features;
1401 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1402 perf_missing_features.cloexec = true;
1403 goto fallback_missing_features;
1404 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1405 perf_missing_features.mmap2 = true;
1406 goto fallback_missing_features;
1407 } else if (!perf_missing_features.exclude_guest &&
1408 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1409 perf_missing_features.exclude_guest = true;
1410 goto fallback_missing_features;
1411 } else if (!perf_missing_features.sample_id_all) {
1412 perf_missing_features.sample_id_all = true;
1413 goto retry_sample_id;
1418 while (--thread >= 0) {
1419 close(FD(evsel, cpu, thread));
1420 FD(evsel, cpu, thread) = -1;
1423 } while (--cpu >= 0);
1427 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1429 if (evsel->fd == NULL)
1432 perf_evsel__close_fd(evsel, ncpus, nthreads);
1433 perf_evsel__free_fd(evsel);
1445 struct thread_map map;
1447 } empty_thread_map = {
1452 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1453 struct thread_map *threads)
1456 /* Work around old compiler warnings about strict aliasing */
1457 cpus = &empty_cpu_map.map;
1460 if (threads == NULL)
1461 threads = &empty_thread_map.map;
1463 return __perf_evsel__open(evsel, cpus, threads);
1466 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1467 struct cpu_map *cpus)
1469 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1472 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1473 struct thread_map *threads)
1475 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1478 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1479 const union perf_event *event,
1480 struct perf_sample *sample)
1482 u64 type = evsel->attr.sample_type;
1483 const u64 *array = event->sample.array;
1484 bool swapped = evsel->needs_swap;
1487 array += ((event->header.size -
1488 sizeof(event->header)) / sizeof(u64)) - 1;
1490 if (type & PERF_SAMPLE_IDENTIFIER) {
1491 sample->id = *array;
1495 if (type & PERF_SAMPLE_CPU) {
1498 /* undo swap of u64, then swap on individual u32s */
1499 u.val64 = bswap_64(u.val64);
1500 u.val32[0] = bswap_32(u.val32[0]);
1503 sample->cpu = u.val32[0];
1507 if (type & PERF_SAMPLE_STREAM_ID) {
1508 sample->stream_id = *array;
1512 if (type & PERF_SAMPLE_ID) {
1513 sample->id = *array;
1517 if (type & PERF_SAMPLE_TIME) {
1518 sample->time = *array;
1522 if (type & PERF_SAMPLE_TID) {
1525 /* undo swap of u64, then swap on individual u32s */
1526 u.val64 = bswap_64(u.val64);
1527 u.val32[0] = bswap_32(u.val32[0]);
1528 u.val32[1] = bswap_32(u.val32[1]);
1531 sample->pid = u.val32[0];
1532 sample->tid = u.val32[1];
1539 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1542 return size > max_size || offset + size > endp;
1545 #define OVERFLOW_CHECK(offset, size, max_size) \
1547 if (overflow(endp, (max_size), (offset), (size))) \
1551 #define OVERFLOW_CHECK_u64(offset) \
1552 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1554 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1555 struct perf_sample *data)
1557 u64 type = evsel->attr.sample_type;
1558 bool swapped = evsel->needs_swap;
1560 u16 max_size = event->header.size;
1561 const void *endp = (void *)event + max_size;
1565 * used for cross-endian analysis. See git commit 65014ab3
1566 * for why this goofiness is needed.
1570 memset(data, 0, sizeof(*data));
1571 data->cpu = data->pid = data->tid = -1;
1572 data->stream_id = data->id = data->time = -1ULL;
1573 data->period = evsel->attr.sample_period;
1576 if (event->header.type != PERF_RECORD_SAMPLE) {
1577 if (!evsel->attr.sample_id_all)
1579 return perf_evsel__parse_id_sample(evsel, event, data);
1582 array = event->sample.array;
1585 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1586 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1587 * check the format does not go past the end of the event.
1589 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1593 if (type & PERF_SAMPLE_IDENTIFIER) {
1598 if (type & PERF_SAMPLE_IP) {
1603 if (type & PERF_SAMPLE_TID) {
1606 /* undo swap of u64, then swap on individual u32s */
1607 u.val64 = bswap_64(u.val64);
1608 u.val32[0] = bswap_32(u.val32[0]);
1609 u.val32[1] = bswap_32(u.val32[1]);
1612 data->pid = u.val32[0];
1613 data->tid = u.val32[1];
1617 if (type & PERF_SAMPLE_TIME) {
1618 data->time = *array;
1623 if (type & PERF_SAMPLE_ADDR) {
1624 data->addr = *array;
1628 if (type & PERF_SAMPLE_ID) {
1633 if (type & PERF_SAMPLE_STREAM_ID) {
1634 data->stream_id = *array;
1638 if (type & PERF_SAMPLE_CPU) {
1642 /* undo swap of u64, then swap on individual u32s */
1643 u.val64 = bswap_64(u.val64);
1644 u.val32[0] = bswap_32(u.val32[0]);
1647 data->cpu = u.val32[0];
1651 if (type & PERF_SAMPLE_PERIOD) {
1652 data->period = *array;
1656 if (type & PERF_SAMPLE_READ) {
1657 u64 read_format = evsel->attr.read_format;
1659 OVERFLOW_CHECK_u64(array);
1660 if (read_format & PERF_FORMAT_GROUP)
1661 data->read.group.nr = *array;
1663 data->read.one.value = *array;
1667 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1668 OVERFLOW_CHECK_u64(array);
1669 data->read.time_enabled = *array;
1673 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1674 OVERFLOW_CHECK_u64(array);
1675 data->read.time_running = *array;
1679 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1680 if (read_format & PERF_FORMAT_GROUP) {
1681 const u64 max_group_nr = UINT64_MAX /
1682 sizeof(struct sample_read_value);
1684 if (data->read.group.nr > max_group_nr)
1686 sz = data->read.group.nr *
1687 sizeof(struct sample_read_value);
1688 OVERFLOW_CHECK(array, sz, max_size);
1689 data->read.group.values =
1690 (struct sample_read_value *)array;
1691 array = (void *)array + sz;
1693 OVERFLOW_CHECK_u64(array);
1694 data->read.one.id = *array;
1699 if (type & PERF_SAMPLE_CALLCHAIN) {
1700 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1702 OVERFLOW_CHECK_u64(array);
1703 data->callchain = (struct ip_callchain *)array++;
1704 if (data->callchain->nr > max_callchain_nr)
1706 sz = data->callchain->nr * sizeof(u64);
1707 OVERFLOW_CHECK(array, sz, max_size);
1708 array = (void *)array + sz;
1711 if (type & PERF_SAMPLE_RAW) {
1712 OVERFLOW_CHECK_u64(array);
1714 if (WARN_ONCE(swapped,
1715 "Endianness of raw data not corrected!\n")) {
1716 /* undo swap of u64, then swap on individual u32s */
1717 u.val64 = bswap_64(u.val64);
1718 u.val32[0] = bswap_32(u.val32[0]);
1719 u.val32[1] = bswap_32(u.val32[1]);
1721 data->raw_size = u.val32[0];
1722 array = (void *)array + sizeof(u32);
1724 OVERFLOW_CHECK(array, data->raw_size, max_size);
1725 data->raw_data = (void *)array;
1726 array = (void *)array + data->raw_size;
1729 if (type & PERF_SAMPLE_BRANCH_STACK) {
1730 const u64 max_branch_nr = UINT64_MAX /
1731 sizeof(struct branch_entry);
1733 OVERFLOW_CHECK_u64(array);
1734 data->branch_stack = (struct branch_stack *)array++;
1736 if (data->branch_stack->nr > max_branch_nr)
1738 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1739 OVERFLOW_CHECK(array, sz, max_size);
1740 array = (void *)array + sz;
1743 if (type & PERF_SAMPLE_REGS_USER) {
1744 OVERFLOW_CHECK_u64(array);
1745 data->user_regs.abi = *array;
1748 if (data->user_regs.abi) {
1749 u64 mask = evsel->attr.sample_regs_user;
1751 sz = hweight_long(mask) * sizeof(u64);
1752 OVERFLOW_CHECK(array, sz, max_size);
1753 data->user_regs.mask = mask;
1754 data->user_regs.regs = (u64 *)array;
1755 array = (void *)array + sz;
1759 if (type & PERF_SAMPLE_STACK_USER) {
1760 OVERFLOW_CHECK_u64(array);
1763 data->user_stack.offset = ((char *)(array - 1)
1767 data->user_stack.size = 0;
1769 OVERFLOW_CHECK(array, sz, max_size);
1770 data->user_stack.data = (char *)array;
1771 array = (void *)array + sz;
1772 OVERFLOW_CHECK_u64(array);
1773 data->user_stack.size = *array++;
1774 if (WARN_ONCE(data->user_stack.size > sz,
1775 "user stack dump failure\n"))
1781 if (type & PERF_SAMPLE_WEIGHT) {
1782 OVERFLOW_CHECK_u64(array);
1783 data->weight = *array;
1787 data->data_src = PERF_MEM_DATA_SRC_NONE;
1788 if (type & PERF_SAMPLE_DATA_SRC) {
1789 OVERFLOW_CHECK_u64(array);
1790 data->data_src = *array;
1794 data->transaction = 0;
1795 if (type & PERF_SAMPLE_TRANSACTION) {
1796 OVERFLOW_CHECK_u64(array);
1797 data->transaction = *array;
1801 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1802 if (type & PERF_SAMPLE_REGS_INTR) {
1803 OVERFLOW_CHECK_u64(array);
1804 data->intr_regs.abi = *array;
1807 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1808 u64 mask = evsel->attr.sample_regs_intr;
1810 sz = hweight_long(mask) * sizeof(u64);
1811 OVERFLOW_CHECK(array, sz, max_size);
1812 data->intr_regs.mask = mask;
1813 data->intr_regs.regs = (u64 *)array;
1814 array = (void *)array + sz;
1821 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1824 size_t sz, result = sizeof(struct sample_event);
1826 if (type & PERF_SAMPLE_IDENTIFIER)
1827 result += sizeof(u64);
1829 if (type & PERF_SAMPLE_IP)
1830 result += sizeof(u64);
1832 if (type & PERF_SAMPLE_TID)
1833 result += sizeof(u64);
1835 if (type & PERF_SAMPLE_TIME)
1836 result += sizeof(u64);
1838 if (type & PERF_SAMPLE_ADDR)
1839 result += sizeof(u64);
1841 if (type & PERF_SAMPLE_ID)
1842 result += sizeof(u64);
1844 if (type & PERF_SAMPLE_STREAM_ID)
1845 result += sizeof(u64);
1847 if (type & PERF_SAMPLE_CPU)
1848 result += sizeof(u64);
1850 if (type & PERF_SAMPLE_PERIOD)
1851 result += sizeof(u64);
1853 if (type & PERF_SAMPLE_READ) {
1854 result += sizeof(u64);
1855 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1856 result += sizeof(u64);
1857 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1858 result += sizeof(u64);
1859 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1860 if (read_format & PERF_FORMAT_GROUP) {
1861 sz = sample->read.group.nr *
1862 sizeof(struct sample_read_value);
1865 result += sizeof(u64);
1869 if (type & PERF_SAMPLE_CALLCHAIN) {
1870 sz = (sample->callchain->nr + 1) * sizeof(u64);
1874 if (type & PERF_SAMPLE_RAW) {
1875 result += sizeof(u32);
1876 result += sample->raw_size;
1879 if (type & PERF_SAMPLE_BRANCH_STACK) {
1880 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1885 if (type & PERF_SAMPLE_REGS_USER) {
1886 if (sample->user_regs.abi) {
1887 result += sizeof(u64);
1888 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1891 result += sizeof(u64);
1895 if (type & PERF_SAMPLE_STACK_USER) {
1896 sz = sample->user_stack.size;
1897 result += sizeof(u64);
1900 result += sizeof(u64);
1904 if (type & PERF_SAMPLE_WEIGHT)
1905 result += sizeof(u64);
1907 if (type & PERF_SAMPLE_DATA_SRC)
1908 result += sizeof(u64);
1910 if (type & PERF_SAMPLE_TRANSACTION)
1911 result += sizeof(u64);
1913 if (type & PERF_SAMPLE_REGS_INTR) {
1914 if (sample->intr_regs.abi) {
1915 result += sizeof(u64);
1916 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1919 result += sizeof(u64);
1926 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1928 const struct perf_sample *sample,
1934 * used for cross-endian analysis. See git commit 65014ab3
1935 * for why this goofiness is needed.
1939 array = event->sample.array;
1941 if (type & PERF_SAMPLE_IDENTIFIER) {
1942 *array = sample->id;
1946 if (type & PERF_SAMPLE_IP) {
1947 *array = sample->ip;
1951 if (type & PERF_SAMPLE_TID) {
1952 u.val32[0] = sample->pid;
1953 u.val32[1] = sample->tid;
1956 * Inverse of what is done in perf_evsel__parse_sample
1958 u.val32[0] = bswap_32(u.val32[0]);
1959 u.val32[1] = bswap_32(u.val32[1]);
1960 u.val64 = bswap_64(u.val64);
1967 if (type & PERF_SAMPLE_TIME) {
1968 *array = sample->time;
1972 if (type & PERF_SAMPLE_ADDR) {
1973 *array = sample->addr;
1977 if (type & PERF_SAMPLE_ID) {
1978 *array = sample->id;
1982 if (type & PERF_SAMPLE_STREAM_ID) {
1983 *array = sample->stream_id;
1987 if (type & PERF_SAMPLE_CPU) {
1988 u.val32[0] = sample->cpu;
1991 * Inverse of what is done in perf_evsel__parse_sample
1993 u.val32[0] = bswap_32(u.val32[0]);
1994 u.val64 = bswap_64(u.val64);
2000 if (type & PERF_SAMPLE_PERIOD) {
2001 *array = sample->period;
2005 if (type & PERF_SAMPLE_READ) {
2006 if (read_format & PERF_FORMAT_GROUP)
2007 *array = sample->read.group.nr;
2009 *array = sample->read.one.value;
2012 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2013 *array = sample->read.time_enabled;
2017 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2018 *array = sample->read.time_running;
2022 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2023 if (read_format & PERF_FORMAT_GROUP) {
2024 sz = sample->read.group.nr *
2025 sizeof(struct sample_read_value);
2026 memcpy(array, sample->read.group.values, sz);
2027 array = (void *)array + sz;
2029 *array = sample->read.one.id;
2034 if (type & PERF_SAMPLE_CALLCHAIN) {
2035 sz = (sample->callchain->nr + 1) * sizeof(u64);
2036 memcpy(array, sample->callchain, sz);
2037 array = (void *)array + sz;
2040 if (type & PERF_SAMPLE_RAW) {
2041 u.val32[0] = sample->raw_size;
2042 if (WARN_ONCE(swapped,
2043 "Endianness of raw data not corrected!\n")) {
2045 * Inverse of what is done in perf_evsel__parse_sample
2047 u.val32[0] = bswap_32(u.val32[0]);
2048 u.val32[1] = bswap_32(u.val32[1]);
2049 u.val64 = bswap_64(u.val64);
2052 array = (void *)array + sizeof(u32);
2054 memcpy(array, sample->raw_data, sample->raw_size);
2055 array = (void *)array + sample->raw_size;
2058 if (type & PERF_SAMPLE_BRANCH_STACK) {
2059 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2061 memcpy(array, sample->branch_stack, sz);
2062 array = (void *)array + sz;
2065 if (type & PERF_SAMPLE_REGS_USER) {
2066 if (sample->user_regs.abi) {
2067 *array++ = sample->user_regs.abi;
2068 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2069 memcpy(array, sample->user_regs.regs, sz);
2070 array = (void *)array + sz;
2076 if (type & PERF_SAMPLE_STACK_USER) {
2077 sz = sample->user_stack.size;
2080 memcpy(array, sample->user_stack.data, sz);
2081 array = (void *)array + sz;
2086 if (type & PERF_SAMPLE_WEIGHT) {
2087 *array = sample->weight;
2091 if (type & PERF_SAMPLE_DATA_SRC) {
2092 *array = sample->data_src;
2096 if (type & PERF_SAMPLE_TRANSACTION) {
2097 *array = sample->transaction;
2101 if (type & PERF_SAMPLE_REGS_INTR) {
2102 if (sample->intr_regs.abi) {
2103 *array++ = sample->intr_regs.abi;
2104 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2105 memcpy(array, sample->intr_regs.regs, sz);
2106 array = (void *)array + sz;
2115 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2117 return pevent_find_field(evsel->tp_format, name);
2120 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2123 struct format_field *field = perf_evsel__field(evsel, name);
2129 offset = field->offset;
2131 if (field->flags & FIELD_IS_DYNAMIC) {
2132 offset = *(int *)(sample->raw_data + field->offset);
2136 return sample->raw_data + offset;
2139 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2142 struct format_field *field = perf_evsel__field(evsel, name);
2149 ptr = sample->raw_data + field->offset;
2151 switch (field->size) {
2155 value = *(u16 *)ptr;
2158 value = *(u32 *)ptr;
2161 memcpy(&value, ptr, sizeof(u64));
2167 if (!evsel->needs_swap)
2170 switch (field->size) {
2172 return bswap_16(value);
2174 return bswap_32(value);
2176 return bswap_64(value);
2184 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2190 ret += fprintf(fp, ",");
2192 ret += fprintf(fp, ":");
2196 va_start(args, fmt);
2197 ret += vfprintf(fp, fmt, args);
2202 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2204 return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2207 int perf_evsel__fprintf(struct perf_evsel *evsel,
2208 struct perf_attr_details *details, FILE *fp)
2213 if (details->event_group) {
2214 struct perf_evsel *pos;
2216 if (!perf_evsel__is_group_leader(evsel))
2219 if (evsel->nr_members > 1)
2220 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2222 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2223 for_each_group_member(pos, evsel)
2224 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2226 if (evsel->nr_members > 1)
2227 printed += fprintf(fp, "}");
2231 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2233 if (details->verbose) {
2234 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2235 __print_attr__fprintf, &first);
2236 } else if (details->freq) {
2237 const char *term = "sample_freq";
2239 if (!evsel->attr.freq)
2240 term = "sample_period";
2242 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2243 term, (u64)evsel->attr.sample_freq);
2250 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2251 char *msg, size_t msgsize)
2253 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2254 evsel->attr.type == PERF_TYPE_HARDWARE &&
2255 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2257 * If it's cycles then fall back to hrtimer based
2258 * cpu-clock-tick sw counter, which is always available even if
2261 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2264 scnprintf(msg, msgsize, "%s",
2265 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2267 evsel->attr.type = PERF_TYPE_SOFTWARE;
2268 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2270 zfree(&evsel->name);
2277 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2278 int err, char *msg, size_t size)
2280 char sbuf[STRERR_BUFSIZE];
2285 return scnprintf(msg, size,
2286 "You may not have permission to collect %sstats.\n"
2287 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2288 " -1 - Not paranoid at all\n"
2289 " 0 - Disallow raw tracepoint access for unpriv\n"
2290 " 1 - Disallow cpu events for unpriv\n"
2291 " 2 - Disallow kernel profiling for unpriv",
2292 target->system_wide ? "system-wide " : "");
2294 return scnprintf(msg, size, "The %s event is not supported.",
2295 perf_evsel__name(evsel));
2297 return scnprintf(msg, size, "%s",
2298 "Too many events are opened.\n"
2299 "Probably the maximum number of open file descriptors has been reached.\n"
2300 "Hint: Try again after reducing the number of events.\n"
2301 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2303 if (target->cpu_list)
2304 return scnprintf(msg, size, "%s",
2305 "No such device - did you specify an out-of-range profile CPU?\n");
2308 if (evsel->attr.precise_ip)
2309 return scnprintf(msg, size, "%s",
2310 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2311 #if defined(__i386__) || defined(__x86_64__)
2312 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2313 return scnprintf(msg, size, "%s",
2314 "No hardware sampling interrupt available.\n"
2315 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2319 if (find_process("oprofiled"))
2320 return scnprintf(msg, size,
2321 "The PMU counters are busy/taken by another profiler.\n"
2322 "We found oprofile daemon running, please stop it and try again.");
2325 if (perf_missing_features.clockid)
2326 return scnprintf(msg, size, "clockid feature not supported.");
2327 if (perf_missing_features.clockid_wrong)
2328 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2334 return scnprintf(msg, size,
2335 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2336 "/bin/dmesg may provide additional information.\n"
2337 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2338 err, strerror_r(err, sbuf, sizeof(sbuf)),
2339 perf_evsel__name(evsel));