2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
7 * Released under the GPL v2. (and only v2, not any later version)
11 #include <linux/bitops.h>
12 #include <api/fs/debugfs.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <sys/resource.h>
18 #include "callchain.h"
24 #include "thread_map.h"
26 #include "perf_regs.h"
28 #include "trace-event.h"
38 } perf_missing_features;
40 static clockid_t clockid;
42 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
47 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
53 int (*init)(struct perf_evsel *evsel);
54 void (*fini)(struct perf_evsel *evsel);
55 } perf_evsel__object = {
56 .size = sizeof(struct perf_evsel),
57 .init = perf_evsel__no_extra_init,
58 .fini = perf_evsel__no_extra_fini,
61 int perf_evsel__object_config(size_t object_size,
62 int (*init)(struct perf_evsel *evsel),
63 void (*fini)(struct perf_evsel *evsel))
69 if (perf_evsel__object.size > object_size)
72 perf_evsel__object.size = object_size;
76 perf_evsel__object.init = init;
79 perf_evsel__object.fini = fini;
84 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
86 int __perf_evsel__sample_size(u64 sample_type)
88 u64 mask = sample_type & PERF_SAMPLE_MASK;
92 for (i = 0; i < 64; i++) {
93 if (mask & (1ULL << i))
103 * __perf_evsel__calc_id_pos - calculate id_pos.
104 * @sample_type: sample type
106 * This function returns the position of the event id (PERF_SAMPLE_ID or
107 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
110 static int __perf_evsel__calc_id_pos(u64 sample_type)
114 if (sample_type & PERF_SAMPLE_IDENTIFIER)
117 if (!(sample_type & PERF_SAMPLE_ID))
120 if (sample_type & PERF_SAMPLE_IP)
123 if (sample_type & PERF_SAMPLE_TID)
126 if (sample_type & PERF_SAMPLE_TIME)
129 if (sample_type & PERF_SAMPLE_ADDR)
136 * __perf_evsel__calc_is_pos - calculate is_pos.
137 * @sample_type: sample type
139 * This function returns the position (counting backwards) of the event id
140 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
141 * sample_id_all is used there is an id sample appended to non-sample events.
143 static int __perf_evsel__calc_is_pos(u64 sample_type)
147 if (sample_type & PERF_SAMPLE_IDENTIFIER)
150 if (!(sample_type & PERF_SAMPLE_ID))
153 if (sample_type & PERF_SAMPLE_CPU)
156 if (sample_type & PERF_SAMPLE_STREAM_ID)
162 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
164 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
165 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
168 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
169 enum perf_event_sample_format bit)
171 if (!(evsel->attr.sample_type & bit)) {
172 evsel->attr.sample_type |= bit;
173 evsel->sample_size += sizeof(u64);
174 perf_evsel__calc_id_pos(evsel);
178 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
179 enum perf_event_sample_format bit)
181 if (evsel->attr.sample_type & bit) {
182 evsel->attr.sample_type &= ~bit;
183 evsel->sample_size -= sizeof(u64);
184 perf_evsel__calc_id_pos(evsel);
188 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
189 bool can_sample_identifier)
191 if (can_sample_identifier) {
192 perf_evsel__reset_sample_bit(evsel, ID);
193 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
195 perf_evsel__set_sample_bit(evsel, ID);
197 evsel->attr.read_format |= PERF_FORMAT_ID;
200 void perf_evsel__init(struct perf_evsel *evsel,
201 struct perf_event_attr *attr, int idx)
204 evsel->tracking = !idx;
206 evsel->leader = evsel;
209 evsel->evlist = NULL;
210 INIT_LIST_HEAD(&evsel->node);
211 INIT_LIST_HEAD(&evsel->config_terms);
212 perf_evsel__object.init(evsel);
213 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
214 perf_evsel__calc_id_pos(evsel);
215 evsel->cmdline_group_boundary = false;
218 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
220 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
223 perf_evsel__init(evsel, attr, idx);
228 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
230 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
233 struct perf_event_attr attr = {
234 .type = PERF_TYPE_TRACEPOINT,
235 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
236 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
239 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
242 evsel->tp_format = trace_event__tp_format(sys, name);
243 if (evsel->tp_format == NULL)
246 event_attr_init(&attr);
247 attr.config = evsel->tp_format->id;
248 attr.sample_period = 1;
249 perf_evsel__init(evsel, &attr, idx);
260 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
268 "stalled-cycles-frontend",
269 "stalled-cycles-backend",
273 static const char *__perf_evsel__hw_name(u64 config)
275 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
276 return perf_evsel__hw_names[config];
278 return "unknown-hardware";
281 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
283 int colon = 0, r = 0;
284 struct perf_event_attr *attr = &evsel->attr;
285 bool exclude_guest_default = false;
287 #define MOD_PRINT(context, mod) do { \
288 if (!attr->exclude_##context) { \
289 if (!colon) colon = ++r; \
290 r += scnprintf(bf + r, size - r, "%c", mod); \
293 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
294 MOD_PRINT(kernel, 'k');
295 MOD_PRINT(user, 'u');
297 exclude_guest_default = true;
300 if (attr->precise_ip) {
303 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
304 exclude_guest_default = true;
307 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
308 MOD_PRINT(host, 'H');
309 MOD_PRINT(guest, 'G');
317 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
319 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
320 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
323 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
336 static const char *__perf_evsel__sw_name(u64 config)
338 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
339 return perf_evsel__sw_names[config];
340 return "unknown-software";
343 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
345 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
346 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
349 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
353 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
355 if (type & HW_BREAKPOINT_R)
356 r += scnprintf(bf + r, size - r, "r");
358 if (type & HW_BREAKPOINT_W)
359 r += scnprintf(bf + r, size - r, "w");
361 if (type & HW_BREAKPOINT_X)
362 r += scnprintf(bf + r, size - r, "x");
367 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
369 struct perf_event_attr *attr = &evsel->attr;
370 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
371 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
374 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
375 [PERF_EVSEL__MAX_ALIASES] = {
376 { "L1-dcache", "l1-d", "l1d", "L1-data", },
377 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
379 { "dTLB", "d-tlb", "Data-TLB", },
380 { "iTLB", "i-tlb", "Instruction-TLB", },
381 { "branch", "branches", "bpu", "btb", "bpc", },
385 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
386 [PERF_EVSEL__MAX_ALIASES] = {
387 { "load", "loads", "read", },
388 { "store", "stores", "write", },
389 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
392 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
393 [PERF_EVSEL__MAX_ALIASES] = {
394 { "refs", "Reference", "ops", "access", },
395 { "misses", "miss", },
398 #define C(x) PERF_COUNT_HW_CACHE_##x
399 #define CACHE_READ (1 << C(OP_READ))
400 #define CACHE_WRITE (1 << C(OP_WRITE))
401 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
402 #define COP(x) (1 << x)
405 * cache operartion stat
406 * L1I : Read and prefetch only
407 * ITLB and BPU : Read-only
409 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
410 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
411 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
412 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
413 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
414 [C(ITLB)] = (CACHE_READ),
415 [C(BPU)] = (CACHE_READ),
416 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
419 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
421 if (perf_evsel__hw_cache_stat[type] & COP(op))
422 return true; /* valid */
424 return false; /* invalid */
427 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
428 char *bf, size_t size)
431 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
432 perf_evsel__hw_cache_op[op][0],
433 perf_evsel__hw_cache_result[result][0]);
436 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
437 perf_evsel__hw_cache_op[op][1]);
440 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
442 u8 op, result, type = (config >> 0) & 0xff;
443 const char *err = "unknown-ext-hardware-cache-type";
445 if (type > PERF_COUNT_HW_CACHE_MAX)
448 op = (config >> 8) & 0xff;
449 err = "unknown-ext-hardware-cache-op";
450 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
453 result = (config >> 16) & 0xff;
454 err = "unknown-ext-hardware-cache-result";
455 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
458 err = "invalid-cache";
459 if (!perf_evsel__is_cache_op_valid(type, op))
462 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
464 return scnprintf(bf, size, "%s", err);
467 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
469 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
470 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
473 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
475 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
476 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
479 const char *perf_evsel__name(struct perf_evsel *evsel)
486 switch (evsel->attr.type) {
488 perf_evsel__raw_name(evsel, bf, sizeof(bf));
491 case PERF_TYPE_HARDWARE:
492 perf_evsel__hw_name(evsel, bf, sizeof(bf));
495 case PERF_TYPE_HW_CACHE:
496 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
499 case PERF_TYPE_SOFTWARE:
500 perf_evsel__sw_name(evsel, bf, sizeof(bf));
503 case PERF_TYPE_TRACEPOINT:
504 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
507 case PERF_TYPE_BREAKPOINT:
508 perf_evsel__bp_name(evsel, bf, sizeof(bf));
512 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
517 evsel->name = strdup(bf);
519 return evsel->name ?: "unknown";
522 const char *perf_evsel__group_name(struct perf_evsel *evsel)
524 return evsel->group_name ?: "anon group";
527 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
530 struct perf_evsel *pos;
531 const char *group_name = perf_evsel__group_name(evsel);
533 ret = scnprintf(buf, size, "%s", group_name);
535 ret += scnprintf(buf + ret, size - ret, " { %s",
536 perf_evsel__name(evsel));
538 for_each_group_member(pos, evsel)
539 ret += scnprintf(buf + ret, size - ret, ", %s",
540 perf_evsel__name(pos));
542 ret += scnprintf(buf + ret, size - ret, " }");
548 perf_evsel__config_callgraph(struct perf_evsel *evsel,
549 struct record_opts *opts,
550 struct callchain_param *param)
552 bool function = perf_evsel__is_function_event(evsel);
553 struct perf_event_attr *attr = &evsel->attr;
555 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
557 if (param->record_mode == CALLCHAIN_LBR) {
558 if (!opts->branch_stack) {
559 if (attr->exclude_user) {
560 pr_warning("LBR callstack option is only available "
561 "to get user callchain information. "
562 "Falling back to framepointers.\n");
564 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
565 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
566 PERF_SAMPLE_BRANCH_CALL_STACK;
569 pr_warning("Cannot use LBR callstack with branch stack. "
570 "Falling back to framepointers.\n");
573 if (param->record_mode == CALLCHAIN_DWARF) {
575 perf_evsel__set_sample_bit(evsel, REGS_USER);
576 perf_evsel__set_sample_bit(evsel, STACK_USER);
577 attr->sample_regs_user = PERF_REGS_MASK;
578 attr->sample_stack_user = param->dump_size;
579 attr->exclude_callchain_user = 1;
581 pr_info("Cannot use DWARF unwind for function trace event,"
582 " falling back to framepointers.\n");
587 pr_info("Disabling user space callchains for function trace event.\n");
588 attr->exclude_callchain_user = 1;
593 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
594 struct callchain_param *param)
596 struct perf_event_attr *attr = &evsel->attr;
598 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
599 if (param->record_mode == CALLCHAIN_LBR) {
600 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
601 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
602 PERF_SAMPLE_BRANCH_CALL_STACK);
604 if (param->record_mode == CALLCHAIN_DWARF) {
605 perf_evsel__reset_sample_bit(evsel, REGS_USER);
606 perf_evsel__reset_sample_bit(evsel, STACK_USER);
610 static void apply_config_terms(struct perf_evsel *evsel,
611 struct record_opts *opts)
613 struct perf_evsel_config_term *term;
614 struct list_head *config_terms = &evsel->config_terms;
615 struct perf_event_attr *attr = &evsel->attr;
616 struct callchain_param param;
618 char *callgraph_buf = NULL;
620 /* callgraph default */
621 param.record_mode = callchain_param.record_mode;
623 list_for_each_entry(term, config_terms, list) {
624 switch (term->type) {
625 case PERF_EVSEL__CONFIG_TERM_PERIOD:
626 attr->sample_period = term->val.period;
629 case PERF_EVSEL__CONFIG_TERM_FREQ:
630 attr->sample_freq = term->val.freq;
633 case PERF_EVSEL__CONFIG_TERM_TIME:
635 perf_evsel__set_sample_bit(evsel, TIME);
637 perf_evsel__reset_sample_bit(evsel, TIME);
639 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
640 callgraph_buf = term->val.callgraph;
642 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
643 dump_size = term->val.stack_user;
650 /* User explicitly set per-event callgraph, clear the old setting and reset. */
651 if ((callgraph_buf != NULL) || (dump_size > 0)) {
653 /* parse callgraph parameters */
654 if (callgraph_buf != NULL) {
655 if (!strcmp(callgraph_buf, "no")) {
656 param.enabled = false;
657 param.record_mode = CALLCHAIN_NONE;
659 param.enabled = true;
660 if (parse_callchain_record(callgraph_buf, ¶m)) {
661 pr_err("per-event callgraph setting for %s failed. "
662 "Apply callgraph global setting for it\n",
669 dump_size = round_up(dump_size, sizeof(u64));
670 param.dump_size = dump_size;
673 /* If global callgraph set, clear it */
674 if (callchain_param.enabled)
675 perf_evsel__reset_callgraph(evsel, &callchain_param);
677 /* set perf-event callgraph */
679 perf_evsel__config_callgraph(evsel, opts, ¶m);
684 * The enable_on_exec/disabled value strategy:
686 * 1) For any type of traced program:
687 * - all independent events and group leaders are disabled
688 * - all group members are enabled
690 * Group members are ruled by group leaders. They need to
691 * be enabled, because the group scheduling relies on that.
693 * 2) For traced programs executed by perf:
694 * - all independent events and group leaders have
696 * - we don't specifically enable or disable any event during
699 * Independent events and group leaders are initially disabled
700 * and get enabled by exec. Group members are ruled by group
701 * leaders as stated in 1).
703 * 3) For traced programs attached by perf (pid/tid):
704 * - we specifically enable or disable all events during
707 * When attaching events to already running traced we
708 * enable/disable events specifically, as there's no
709 * initial traced exec call.
711 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
713 struct perf_evsel *leader = evsel->leader;
714 struct perf_event_attr *attr = &evsel->attr;
715 int track = evsel->tracking;
716 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
718 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
719 attr->inherit = !opts->no_inherit;
721 perf_evsel__set_sample_bit(evsel, IP);
722 perf_evsel__set_sample_bit(evsel, TID);
724 if (evsel->sample_read) {
725 perf_evsel__set_sample_bit(evsel, READ);
728 * We need ID even in case of single event, because
729 * PERF_SAMPLE_READ process ID specific data.
731 perf_evsel__set_sample_id(evsel, false);
734 * Apply group format only if we belong to group
735 * with more than one members.
737 if (leader->nr_members > 1) {
738 attr->read_format |= PERF_FORMAT_GROUP;
744 * We default some events to have a default interval. But keep
745 * it a weak assumption overridable by the user.
747 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
748 opts->user_interval != ULLONG_MAX)) {
750 perf_evsel__set_sample_bit(evsel, PERIOD);
752 attr->sample_freq = opts->freq;
754 attr->sample_period = opts->default_interval;
759 * Disable sampling for all group members other
760 * than leader in case leader 'leads' the sampling.
762 if ((leader != evsel) && leader->sample_read) {
763 attr->sample_freq = 0;
764 attr->sample_period = 0;
767 if (opts->no_samples)
768 attr->sample_freq = 0;
770 if (opts->inherit_stat)
771 attr->inherit_stat = 1;
773 if (opts->sample_address) {
774 perf_evsel__set_sample_bit(evsel, ADDR);
775 attr->mmap_data = track;
779 * We don't allow user space callchains for function trace
780 * event, due to issues with page faults while tracing page
781 * fault handler and its overall trickiness nature.
783 if (perf_evsel__is_function_event(evsel))
784 evsel->attr.exclude_callchain_user = 1;
786 if (callchain_param.enabled && !evsel->no_aux_samples)
787 perf_evsel__config_callgraph(evsel, opts, &callchain_param);
789 if (opts->sample_intr_regs) {
790 attr->sample_regs_intr = opts->sample_intr_regs;
791 perf_evsel__set_sample_bit(evsel, REGS_INTR);
794 if (target__has_cpu(&opts->target))
795 perf_evsel__set_sample_bit(evsel, CPU);
798 perf_evsel__set_sample_bit(evsel, PERIOD);
801 * When the user explicitely disabled time don't force it here.
803 if (opts->sample_time &&
804 (!perf_missing_features.sample_id_all &&
805 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
806 opts->sample_time_set)))
807 perf_evsel__set_sample_bit(evsel, TIME);
809 if (opts->raw_samples && !evsel->no_aux_samples) {
810 perf_evsel__set_sample_bit(evsel, TIME);
811 perf_evsel__set_sample_bit(evsel, RAW);
812 perf_evsel__set_sample_bit(evsel, CPU);
815 if (opts->sample_address)
816 perf_evsel__set_sample_bit(evsel, DATA_SRC);
818 if (opts->no_buffering) {
820 attr->wakeup_events = 1;
822 if (opts->branch_stack && !evsel->no_aux_samples) {
823 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
824 attr->branch_sample_type = opts->branch_stack;
827 if (opts->sample_weight)
828 perf_evsel__set_sample_bit(evsel, WEIGHT);
832 attr->mmap2 = track && !perf_missing_features.mmap2;
835 if (opts->record_switch_events)
836 attr->context_switch = track;
838 if (opts->sample_transaction)
839 perf_evsel__set_sample_bit(evsel, TRANSACTION);
841 if (opts->running_time) {
842 evsel->attr.read_format |=
843 PERF_FORMAT_TOTAL_TIME_ENABLED |
844 PERF_FORMAT_TOTAL_TIME_RUNNING;
848 * XXX see the function comment above
850 * Disabling only independent events or group leaders,
851 * keeping group members enabled.
853 if (perf_evsel__is_group_leader(evsel))
857 * Setting enable_on_exec for independent events and
858 * group leaders for traced executed by perf.
860 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
861 !opts->initial_delay)
862 attr->enable_on_exec = 1;
864 if (evsel->immediate) {
866 attr->enable_on_exec = 0;
869 clockid = opts->clockid;
870 if (opts->use_clockid) {
871 attr->use_clockid = 1;
872 attr->clockid = opts->clockid;
876 * Apply event specific term settings,
877 * it overloads any global configuration.
879 apply_config_terms(evsel, opts);
882 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
886 if (evsel->system_wide)
889 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
892 for (cpu = 0; cpu < ncpus; cpu++) {
893 for (thread = 0; thread < nthreads; thread++) {
894 FD(evsel, cpu, thread) = -1;
899 return evsel->fd != NULL ? 0 : -ENOMEM;
902 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
907 if (evsel->system_wide)
910 for (cpu = 0; cpu < ncpus; cpu++) {
911 for (thread = 0; thread < nthreads; thread++) {
912 int fd = FD(evsel, cpu, thread),
913 err = ioctl(fd, ioc, arg);
923 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
926 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
927 PERF_EVENT_IOC_SET_FILTER,
931 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
933 char *new_filter = strdup(filter);
935 if (new_filter != NULL) {
937 evsel->filter = new_filter;
944 int perf_evsel__append_filter(struct perf_evsel *evsel,
945 const char *op, const char *filter)
949 if (evsel->filter == NULL)
950 return perf_evsel__set_filter(evsel, filter);
952 if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
954 evsel->filter = new_filter;
961 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
963 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
964 PERF_EVENT_IOC_ENABLE,
968 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
970 if (ncpus == 0 || nthreads == 0)
973 if (evsel->system_wide)
976 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
977 if (evsel->sample_id == NULL)
980 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
981 if (evsel->id == NULL) {
982 xyarray__delete(evsel->sample_id);
983 evsel->sample_id = NULL;
990 static void perf_evsel__free_fd(struct perf_evsel *evsel)
992 xyarray__delete(evsel->fd);
996 static void perf_evsel__free_id(struct perf_evsel *evsel)
998 xyarray__delete(evsel->sample_id);
999 evsel->sample_id = NULL;
1003 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1005 struct perf_evsel_config_term *term, *h;
1007 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1008 list_del(&term->list);
1013 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1017 if (evsel->system_wide)
1020 for (cpu = 0; cpu < ncpus; cpu++)
1021 for (thread = 0; thread < nthreads; ++thread) {
1022 close(FD(evsel, cpu, thread));
1023 FD(evsel, cpu, thread) = -1;
1027 void perf_evsel__exit(struct perf_evsel *evsel)
1029 assert(list_empty(&evsel->node));
1030 assert(evsel->evlist == NULL);
1031 perf_evsel__free_fd(evsel);
1032 perf_evsel__free_id(evsel);
1033 perf_evsel__free_config_terms(evsel);
1034 close_cgroup(evsel->cgrp);
1035 cpu_map__put(evsel->cpus);
1036 cpu_map__put(evsel->own_cpus);
1037 thread_map__put(evsel->threads);
1038 zfree(&evsel->group_name);
1039 zfree(&evsel->name);
1040 perf_evsel__object.fini(evsel);
1043 void perf_evsel__delete(struct perf_evsel *evsel)
1045 perf_evsel__exit(evsel);
1049 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1050 struct perf_counts_values *count)
1052 struct perf_counts_values tmp;
1054 if (!evsel->prev_raw_counts)
1058 tmp = evsel->prev_raw_counts->aggr;
1059 evsel->prev_raw_counts->aggr = *count;
1061 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1062 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1065 count->val = count->val - tmp.val;
1066 count->ena = count->ena - tmp.ena;
1067 count->run = count->run - tmp.run;
1070 void perf_counts_values__scale(struct perf_counts_values *count,
1071 bool scale, s8 *pscaled)
1076 if (count->run == 0) {
1079 } else if (count->run < count->ena) {
1081 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1084 count->ena = count->run = 0;
1090 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1091 struct perf_counts_values *count)
1093 memset(count, 0, sizeof(*count));
1095 if (FD(evsel, cpu, thread) < 0)
1098 if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1104 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1105 int cpu, int thread, bool scale)
1107 struct perf_counts_values count;
1108 size_t nv = scale ? 3 : 1;
1110 if (FD(evsel, cpu, thread) < 0)
1113 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1116 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1119 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1120 perf_counts_values__scale(&count, scale, NULL);
1121 *perf_counts(evsel->counts, cpu, thread) = count;
1125 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1127 struct perf_evsel *leader = evsel->leader;
1130 if (perf_evsel__is_group_leader(evsel))
1134 * Leader must be already processed/open,
1135 * if not it's a bug.
1137 BUG_ON(!leader->fd);
1139 fd = FD(leader, cpu, thread);
1150 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1152 bool first_bit = true;
1156 if (value & bits[i].bit) {
1157 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1160 } while (bits[++i].name != NULL);
1163 static void __p_sample_type(char *buf, size_t size, u64 value)
1165 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1166 struct bit_names bits[] = {
1167 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1168 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1169 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1170 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1171 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1175 __p_bits(buf, size, value, bits);
1178 static void __p_read_format(char *buf, size_t size, u64 value)
1180 #define bit_name(n) { PERF_FORMAT_##n, #n }
1181 struct bit_names bits[] = {
1182 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1183 bit_name(ID), bit_name(GROUP),
1187 __p_bits(buf, size, value, bits);
1190 #define BUF_SIZE 1024
1192 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1193 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1194 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1195 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1196 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1198 #define PRINT_ATTRn(_n, _f, _p) \
1202 ret += attr__fprintf(fp, _n, buf, priv);\
1206 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1208 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1209 attr__fprintf_f attr__fprintf, void *priv)
1214 PRINT_ATTRf(type, p_unsigned);
1215 PRINT_ATTRf(size, p_unsigned);
1216 PRINT_ATTRf(config, p_hex);
1217 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1218 PRINT_ATTRf(sample_type, p_sample_type);
1219 PRINT_ATTRf(read_format, p_read_format);
1221 PRINT_ATTRf(disabled, p_unsigned);
1222 PRINT_ATTRf(inherit, p_unsigned);
1223 PRINT_ATTRf(pinned, p_unsigned);
1224 PRINT_ATTRf(exclusive, p_unsigned);
1225 PRINT_ATTRf(exclude_user, p_unsigned);
1226 PRINT_ATTRf(exclude_kernel, p_unsigned);
1227 PRINT_ATTRf(exclude_hv, p_unsigned);
1228 PRINT_ATTRf(exclude_idle, p_unsigned);
1229 PRINT_ATTRf(mmap, p_unsigned);
1230 PRINT_ATTRf(comm, p_unsigned);
1231 PRINT_ATTRf(freq, p_unsigned);
1232 PRINT_ATTRf(inherit_stat, p_unsigned);
1233 PRINT_ATTRf(enable_on_exec, p_unsigned);
1234 PRINT_ATTRf(task, p_unsigned);
1235 PRINT_ATTRf(watermark, p_unsigned);
1236 PRINT_ATTRf(precise_ip, p_unsigned);
1237 PRINT_ATTRf(mmap_data, p_unsigned);
1238 PRINT_ATTRf(sample_id_all, p_unsigned);
1239 PRINT_ATTRf(exclude_host, p_unsigned);
1240 PRINT_ATTRf(exclude_guest, p_unsigned);
1241 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1242 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1243 PRINT_ATTRf(mmap2, p_unsigned);
1244 PRINT_ATTRf(comm_exec, p_unsigned);
1245 PRINT_ATTRf(use_clockid, p_unsigned);
1246 PRINT_ATTRf(context_switch, p_unsigned);
1248 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1249 PRINT_ATTRf(bp_type, p_unsigned);
1250 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1251 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1252 PRINT_ATTRf(sample_regs_user, p_hex);
1253 PRINT_ATTRf(sample_stack_user, p_unsigned);
1254 PRINT_ATTRf(clockid, p_signed);
1255 PRINT_ATTRf(sample_regs_intr, p_hex);
1256 PRINT_ATTRf(aux_watermark, p_unsigned);
1261 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1262 void *priv __attribute__((unused)))
1264 return fprintf(fp, " %-32s %s\n", name, val);
1267 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1268 struct thread_map *threads)
1270 int cpu, thread, nthreads;
1271 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1273 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1275 if (evsel->system_wide)
1278 nthreads = threads->nr;
1280 if (evsel->fd == NULL &&
1281 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1285 flags |= PERF_FLAG_PID_CGROUP;
1286 pid = evsel->cgrp->fd;
1289 fallback_missing_features:
1290 if (perf_missing_features.clockid_wrong)
1291 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1292 if (perf_missing_features.clockid) {
1293 evsel->attr.use_clockid = 0;
1294 evsel->attr.clockid = 0;
1296 if (perf_missing_features.cloexec)
1297 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1298 if (perf_missing_features.mmap2)
1299 evsel->attr.mmap2 = 0;
1300 if (perf_missing_features.exclude_guest)
1301 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1303 if (perf_missing_features.sample_id_all)
1304 evsel->attr.sample_id_all = 0;
1307 fprintf(stderr, "%.60s\n", graph_dotted_line);
1308 fprintf(stderr, "perf_event_attr:\n");
1309 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1310 fprintf(stderr, "%.60s\n", graph_dotted_line);
1313 for (cpu = 0; cpu < cpus->nr; cpu++) {
1315 for (thread = 0; thread < nthreads; thread++) {
1318 if (!evsel->cgrp && !evsel->system_wide)
1319 pid = thread_map__pid(threads, thread);
1321 group_fd = get_group_fd(evsel, cpu, thread);
1323 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1324 pid, cpus->map[cpu], group_fd, flags);
1326 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1330 if (FD(evsel, cpu, thread) < 0) {
1332 pr_debug2("sys_perf_event_open failed, error %d\n",
1336 set_rlimit = NO_CHANGE;
1339 * If we succeeded but had to kill clockid, fail and
1340 * have perf_evsel__open_strerror() print us a nice
1343 if (perf_missing_features.clockid ||
1344 perf_missing_features.clockid_wrong) {
1355 * perf stat needs between 5 and 22 fds per CPU. When we run out
1356 * of them try to increase the limits.
1358 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1360 int old_errno = errno;
1362 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1363 if (set_rlimit == NO_CHANGE)
1364 l.rlim_cur = l.rlim_max;
1366 l.rlim_cur = l.rlim_max + 1000;
1367 l.rlim_max = l.rlim_cur;
1369 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1378 if (err != -EINVAL || cpu > 0 || thread > 0)
1382 * Must probe features in the order they were added to the
1383 * perf_event_attr interface.
1385 if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1386 perf_missing_features.clockid_wrong = true;
1387 goto fallback_missing_features;
1388 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1389 perf_missing_features.clockid = true;
1390 goto fallback_missing_features;
1391 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1392 perf_missing_features.cloexec = true;
1393 goto fallback_missing_features;
1394 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1395 perf_missing_features.mmap2 = true;
1396 goto fallback_missing_features;
1397 } else if (!perf_missing_features.exclude_guest &&
1398 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1399 perf_missing_features.exclude_guest = true;
1400 goto fallback_missing_features;
1401 } else if (!perf_missing_features.sample_id_all) {
1402 perf_missing_features.sample_id_all = true;
1403 goto retry_sample_id;
1408 while (--thread >= 0) {
1409 close(FD(evsel, cpu, thread));
1410 FD(evsel, cpu, thread) = -1;
1413 } while (--cpu >= 0);
1417 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1419 if (evsel->fd == NULL)
1422 perf_evsel__close_fd(evsel, ncpus, nthreads);
1423 perf_evsel__free_fd(evsel);
1435 struct thread_map map;
1437 } empty_thread_map = {
1442 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1443 struct thread_map *threads)
1446 /* Work around old compiler warnings about strict aliasing */
1447 cpus = &empty_cpu_map.map;
1450 if (threads == NULL)
1451 threads = &empty_thread_map.map;
1453 return __perf_evsel__open(evsel, cpus, threads);
1456 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1457 struct cpu_map *cpus)
1459 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1462 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1463 struct thread_map *threads)
1465 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1468 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1469 const union perf_event *event,
1470 struct perf_sample *sample)
1472 u64 type = evsel->attr.sample_type;
1473 const u64 *array = event->sample.array;
1474 bool swapped = evsel->needs_swap;
1477 array += ((event->header.size -
1478 sizeof(event->header)) / sizeof(u64)) - 1;
1480 if (type & PERF_SAMPLE_IDENTIFIER) {
1481 sample->id = *array;
1485 if (type & PERF_SAMPLE_CPU) {
1488 /* undo swap of u64, then swap on individual u32s */
1489 u.val64 = bswap_64(u.val64);
1490 u.val32[0] = bswap_32(u.val32[0]);
1493 sample->cpu = u.val32[0];
1497 if (type & PERF_SAMPLE_STREAM_ID) {
1498 sample->stream_id = *array;
1502 if (type & PERF_SAMPLE_ID) {
1503 sample->id = *array;
1507 if (type & PERF_SAMPLE_TIME) {
1508 sample->time = *array;
1512 if (type & PERF_SAMPLE_TID) {
1515 /* undo swap of u64, then swap on individual u32s */
1516 u.val64 = bswap_64(u.val64);
1517 u.val32[0] = bswap_32(u.val32[0]);
1518 u.val32[1] = bswap_32(u.val32[1]);
1521 sample->pid = u.val32[0];
1522 sample->tid = u.val32[1];
1529 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1532 return size > max_size || offset + size > endp;
1535 #define OVERFLOW_CHECK(offset, size, max_size) \
1537 if (overflow(endp, (max_size), (offset), (size))) \
1541 #define OVERFLOW_CHECK_u64(offset) \
1542 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1544 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1545 struct perf_sample *data)
1547 u64 type = evsel->attr.sample_type;
1548 bool swapped = evsel->needs_swap;
1550 u16 max_size = event->header.size;
1551 const void *endp = (void *)event + max_size;
1555 * used for cross-endian analysis. See git commit 65014ab3
1556 * for why this goofiness is needed.
1560 memset(data, 0, sizeof(*data));
1561 data->cpu = data->pid = data->tid = -1;
1562 data->stream_id = data->id = data->time = -1ULL;
1563 data->period = evsel->attr.sample_period;
1566 if (event->header.type != PERF_RECORD_SAMPLE) {
1567 if (!evsel->attr.sample_id_all)
1569 return perf_evsel__parse_id_sample(evsel, event, data);
1572 array = event->sample.array;
1575 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1576 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1577 * check the format does not go past the end of the event.
1579 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1583 if (type & PERF_SAMPLE_IDENTIFIER) {
1588 if (type & PERF_SAMPLE_IP) {
1593 if (type & PERF_SAMPLE_TID) {
1596 /* undo swap of u64, then swap on individual u32s */
1597 u.val64 = bswap_64(u.val64);
1598 u.val32[0] = bswap_32(u.val32[0]);
1599 u.val32[1] = bswap_32(u.val32[1]);
1602 data->pid = u.val32[0];
1603 data->tid = u.val32[1];
1607 if (type & PERF_SAMPLE_TIME) {
1608 data->time = *array;
1613 if (type & PERF_SAMPLE_ADDR) {
1614 data->addr = *array;
1618 if (type & PERF_SAMPLE_ID) {
1623 if (type & PERF_SAMPLE_STREAM_ID) {
1624 data->stream_id = *array;
1628 if (type & PERF_SAMPLE_CPU) {
1632 /* undo swap of u64, then swap on individual u32s */
1633 u.val64 = bswap_64(u.val64);
1634 u.val32[0] = bswap_32(u.val32[0]);
1637 data->cpu = u.val32[0];
1641 if (type & PERF_SAMPLE_PERIOD) {
1642 data->period = *array;
1646 if (type & PERF_SAMPLE_READ) {
1647 u64 read_format = evsel->attr.read_format;
1649 OVERFLOW_CHECK_u64(array);
1650 if (read_format & PERF_FORMAT_GROUP)
1651 data->read.group.nr = *array;
1653 data->read.one.value = *array;
1657 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1658 OVERFLOW_CHECK_u64(array);
1659 data->read.time_enabled = *array;
1663 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1664 OVERFLOW_CHECK_u64(array);
1665 data->read.time_running = *array;
1669 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1670 if (read_format & PERF_FORMAT_GROUP) {
1671 const u64 max_group_nr = UINT64_MAX /
1672 sizeof(struct sample_read_value);
1674 if (data->read.group.nr > max_group_nr)
1676 sz = data->read.group.nr *
1677 sizeof(struct sample_read_value);
1678 OVERFLOW_CHECK(array, sz, max_size);
1679 data->read.group.values =
1680 (struct sample_read_value *)array;
1681 array = (void *)array + sz;
1683 OVERFLOW_CHECK_u64(array);
1684 data->read.one.id = *array;
1689 if (type & PERF_SAMPLE_CALLCHAIN) {
1690 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1692 OVERFLOW_CHECK_u64(array);
1693 data->callchain = (struct ip_callchain *)array++;
1694 if (data->callchain->nr > max_callchain_nr)
1696 sz = data->callchain->nr * sizeof(u64);
1697 OVERFLOW_CHECK(array, sz, max_size);
1698 array = (void *)array + sz;
1701 if (type & PERF_SAMPLE_RAW) {
1702 OVERFLOW_CHECK_u64(array);
1704 if (WARN_ONCE(swapped,
1705 "Endianness of raw data not corrected!\n")) {
1706 /* undo swap of u64, then swap on individual u32s */
1707 u.val64 = bswap_64(u.val64);
1708 u.val32[0] = bswap_32(u.val32[0]);
1709 u.val32[1] = bswap_32(u.val32[1]);
1711 data->raw_size = u.val32[0];
1712 array = (void *)array + sizeof(u32);
1714 OVERFLOW_CHECK(array, data->raw_size, max_size);
1715 data->raw_data = (void *)array;
1716 array = (void *)array + data->raw_size;
1719 if (type & PERF_SAMPLE_BRANCH_STACK) {
1720 const u64 max_branch_nr = UINT64_MAX /
1721 sizeof(struct branch_entry);
1723 OVERFLOW_CHECK_u64(array);
1724 data->branch_stack = (struct branch_stack *)array++;
1726 if (data->branch_stack->nr > max_branch_nr)
1728 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1729 OVERFLOW_CHECK(array, sz, max_size);
1730 array = (void *)array + sz;
1733 if (type & PERF_SAMPLE_REGS_USER) {
1734 OVERFLOW_CHECK_u64(array);
1735 data->user_regs.abi = *array;
1738 if (data->user_regs.abi) {
1739 u64 mask = evsel->attr.sample_regs_user;
1741 sz = hweight_long(mask) * sizeof(u64);
1742 OVERFLOW_CHECK(array, sz, max_size);
1743 data->user_regs.mask = mask;
1744 data->user_regs.regs = (u64 *)array;
1745 array = (void *)array + sz;
1749 if (type & PERF_SAMPLE_STACK_USER) {
1750 OVERFLOW_CHECK_u64(array);
1753 data->user_stack.offset = ((char *)(array - 1)
1757 data->user_stack.size = 0;
1759 OVERFLOW_CHECK(array, sz, max_size);
1760 data->user_stack.data = (char *)array;
1761 array = (void *)array + sz;
1762 OVERFLOW_CHECK_u64(array);
1763 data->user_stack.size = *array++;
1764 if (WARN_ONCE(data->user_stack.size > sz,
1765 "user stack dump failure\n"))
1771 if (type & PERF_SAMPLE_WEIGHT) {
1772 OVERFLOW_CHECK_u64(array);
1773 data->weight = *array;
1777 data->data_src = PERF_MEM_DATA_SRC_NONE;
1778 if (type & PERF_SAMPLE_DATA_SRC) {
1779 OVERFLOW_CHECK_u64(array);
1780 data->data_src = *array;
1784 data->transaction = 0;
1785 if (type & PERF_SAMPLE_TRANSACTION) {
1786 OVERFLOW_CHECK_u64(array);
1787 data->transaction = *array;
1791 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1792 if (type & PERF_SAMPLE_REGS_INTR) {
1793 OVERFLOW_CHECK_u64(array);
1794 data->intr_regs.abi = *array;
1797 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1798 u64 mask = evsel->attr.sample_regs_intr;
1800 sz = hweight_long(mask) * sizeof(u64);
1801 OVERFLOW_CHECK(array, sz, max_size);
1802 data->intr_regs.mask = mask;
1803 data->intr_regs.regs = (u64 *)array;
1804 array = (void *)array + sz;
1811 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1814 size_t sz, result = sizeof(struct sample_event);
1816 if (type & PERF_SAMPLE_IDENTIFIER)
1817 result += sizeof(u64);
1819 if (type & PERF_SAMPLE_IP)
1820 result += sizeof(u64);
1822 if (type & PERF_SAMPLE_TID)
1823 result += sizeof(u64);
1825 if (type & PERF_SAMPLE_TIME)
1826 result += sizeof(u64);
1828 if (type & PERF_SAMPLE_ADDR)
1829 result += sizeof(u64);
1831 if (type & PERF_SAMPLE_ID)
1832 result += sizeof(u64);
1834 if (type & PERF_SAMPLE_STREAM_ID)
1835 result += sizeof(u64);
1837 if (type & PERF_SAMPLE_CPU)
1838 result += sizeof(u64);
1840 if (type & PERF_SAMPLE_PERIOD)
1841 result += sizeof(u64);
1843 if (type & PERF_SAMPLE_READ) {
1844 result += sizeof(u64);
1845 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1846 result += sizeof(u64);
1847 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1848 result += sizeof(u64);
1849 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1850 if (read_format & PERF_FORMAT_GROUP) {
1851 sz = sample->read.group.nr *
1852 sizeof(struct sample_read_value);
1855 result += sizeof(u64);
1859 if (type & PERF_SAMPLE_CALLCHAIN) {
1860 sz = (sample->callchain->nr + 1) * sizeof(u64);
1864 if (type & PERF_SAMPLE_RAW) {
1865 result += sizeof(u32);
1866 result += sample->raw_size;
1869 if (type & PERF_SAMPLE_BRANCH_STACK) {
1870 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1875 if (type & PERF_SAMPLE_REGS_USER) {
1876 if (sample->user_regs.abi) {
1877 result += sizeof(u64);
1878 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1881 result += sizeof(u64);
1885 if (type & PERF_SAMPLE_STACK_USER) {
1886 sz = sample->user_stack.size;
1887 result += sizeof(u64);
1890 result += sizeof(u64);
1894 if (type & PERF_SAMPLE_WEIGHT)
1895 result += sizeof(u64);
1897 if (type & PERF_SAMPLE_DATA_SRC)
1898 result += sizeof(u64);
1900 if (type & PERF_SAMPLE_TRANSACTION)
1901 result += sizeof(u64);
1903 if (type & PERF_SAMPLE_REGS_INTR) {
1904 if (sample->intr_regs.abi) {
1905 result += sizeof(u64);
1906 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1909 result += sizeof(u64);
1916 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1918 const struct perf_sample *sample,
1924 * used for cross-endian analysis. See git commit 65014ab3
1925 * for why this goofiness is needed.
1929 array = event->sample.array;
1931 if (type & PERF_SAMPLE_IDENTIFIER) {
1932 *array = sample->id;
1936 if (type & PERF_SAMPLE_IP) {
1937 *array = sample->ip;
1941 if (type & PERF_SAMPLE_TID) {
1942 u.val32[0] = sample->pid;
1943 u.val32[1] = sample->tid;
1946 * Inverse of what is done in perf_evsel__parse_sample
1948 u.val32[0] = bswap_32(u.val32[0]);
1949 u.val32[1] = bswap_32(u.val32[1]);
1950 u.val64 = bswap_64(u.val64);
1957 if (type & PERF_SAMPLE_TIME) {
1958 *array = sample->time;
1962 if (type & PERF_SAMPLE_ADDR) {
1963 *array = sample->addr;
1967 if (type & PERF_SAMPLE_ID) {
1968 *array = sample->id;
1972 if (type & PERF_SAMPLE_STREAM_ID) {
1973 *array = sample->stream_id;
1977 if (type & PERF_SAMPLE_CPU) {
1978 u.val32[0] = sample->cpu;
1981 * Inverse of what is done in perf_evsel__parse_sample
1983 u.val32[0] = bswap_32(u.val32[0]);
1984 u.val64 = bswap_64(u.val64);
1990 if (type & PERF_SAMPLE_PERIOD) {
1991 *array = sample->period;
1995 if (type & PERF_SAMPLE_READ) {
1996 if (read_format & PERF_FORMAT_GROUP)
1997 *array = sample->read.group.nr;
1999 *array = sample->read.one.value;
2002 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2003 *array = sample->read.time_enabled;
2007 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2008 *array = sample->read.time_running;
2012 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2013 if (read_format & PERF_FORMAT_GROUP) {
2014 sz = sample->read.group.nr *
2015 sizeof(struct sample_read_value);
2016 memcpy(array, sample->read.group.values, sz);
2017 array = (void *)array + sz;
2019 *array = sample->read.one.id;
2024 if (type & PERF_SAMPLE_CALLCHAIN) {
2025 sz = (sample->callchain->nr + 1) * sizeof(u64);
2026 memcpy(array, sample->callchain, sz);
2027 array = (void *)array + sz;
2030 if (type & PERF_SAMPLE_RAW) {
2031 u.val32[0] = sample->raw_size;
2032 if (WARN_ONCE(swapped,
2033 "Endianness of raw data not corrected!\n")) {
2035 * Inverse of what is done in perf_evsel__parse_sample
2037 u.val32[0] = bswap_32(u.val32[0]);
2038 u.val32[1] = bswap_32(u.val32[1]);
2039 u.val64 = bswap_64(u.val64);
2042 array = (void *)array + sizeof(u32);
2044 memcpy(array, sample->raw_data, sample->raw_size);
2045 array = (void *)array + sample->raw_size;
2048 if (type & PERF_SAMPLE_BRANCH_STACK) {
2049 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2051 memcpy(array, sample->branch_stack, sz);
2052 array = (void *)array + sz;
2055 if (type & PERF_SAMPLE_REGS_USER) {
2056 if (sample->user_regs.abi) {
2057 *array++ = sample->user_regs.abi;
2058 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2059 memcpy(array, sample->user_regs.regs, sz);
2060 array = (void *)array + sz;
2066 if (type & PERF_SAMPLE_STACK_USER) {
2067 sz = sample->user_stack.size;
2070 memcpy(array, sample->user_stack.data, sz);
2071 array = (void *)array + sz;
2076 if (type & PERF_SAMPLE_WEIGHT) {
2077 *array = sample->weight;
2081 if (type & PERF_SAMPLE_DATA_SRC) {
2082 *array = sample->data_src;
2086 if (type & PERF_SAMPLE_TRANSACTION) {
2087 *array = sample->transaction;
2091 if (type & PERF_SAMPLE_REGS_INTR) {
2092 if (sample->intr_regs.abi) {
2093 *array++ = sample->intr_regs.abi;
2094 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2095 memcpy(array, sample->intr_regs.regs, sz);
2096 array = (void *)array + sz;
2105 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2107 return pevent_find_field(evsel->tp_format, name);
2110 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2113 struct format_field *field = perf_evsel__field(evsel, name);
2119 offset = field->offset;
2121 if (field->flags & FIELD_IS_DYNAMIC) {
2122 offset = *(int *)(sample->raw_data + field->offset);
2126 return sample->raw_data + offset;
2129 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2132 struct format_field *field = perf_evsel__field(evsel, name);
2139 ptr = sample->raw_data + field->offset;
2141 switch (field->size) {
2145 value = *(u16 *)ptr;
2148 value = *(u32 *)ptr;
2151 memcpy(&value, ptr, sizeof(u64));
2157 if (!evsel->needs_swap)
2160 switch (field->size) {
2162 return bswap_16(value);
2164 return bswap_32(value);
2166 return bswap_64(value);
2174 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2180 ret += fprintf(fp, ",");
2182 ret += fprintf(fp, ":");
2186 va_start(args, fmt);
2187 ret += vfprintf(fp, fmt, args);
2192 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2194 return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2197 int perf_evsel__fprintf(struct perf_evsel *evsel,
2198 struct perf_attr_details *details, FILE *fp)
2203 if (details->event_group) {
2204 struct perf_evsel *pos;
2206 if (!perf_evsel__is_group_leader(evsel))
2209 if (evsel->nr_members > 1)
2210 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2212 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2213 for_each_group_member(pos, evsel)
2214 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2216 if (evsel->nr_members > 1)
2217 printed += fprintf(fp, "}");
2221 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2223 if (details->verbose) {
2224 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2225 __print_attr__fprintf, &first);
2226 } else if (details->freq) {
2227 const char *term = "sample_freq";
2229 if (!evsel->attr.freq)
2230 term = "sample_period";
2232 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2233 term, (u64)evsel->attr.sample_freq);
2240 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2241 char *msg, size_t msgsize)
2243 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2244 evsel->attr.type == PERF_TYPE_HARDWARE &&
2245 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2247 * If it's cycles then fall back to hrtimer based
2248 * cpu-clock-tick sw counter, which is always available even if
2251 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2254 scnprintf(msg, msgsize, "%s",
2255 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2257 evsel->attr.type = PERF_TYPE_SOFTWARE;
2258 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2260 zfree(&evsel->name);
2267 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2268 int err, char *msg, size_t size)
2270 char sbuf[STRERR_BUFSIZE];
2275 return scnprintf(msg, size,
2276 "You may not have permission to collect %sstats.\n"
2277 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2278 " -1 - Not paranoid at all\n"
2279 " 0 - Disallow raw tracepoint access for unpriv\n"
2280 " 1 - Disallow cpu events for unpriv\n"
2281 " 2 - Disallow kernel profiling for unpriv",
2282 target->system_wide ? "system-wide " : "");
2284 return scnprintf(msg, size, "The %s event is not supported.",
2285 perf_evsel__name(evsel));
2287 return scnprintf(msg, size, "%s",
2288 "Too many events are opened.\n"
2289 "Probably the maximum number of open file descriptors has been reached.\n"
2290 "Hint: Try again after reducing the number of events.\n"
2291 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2293 if (target->cpu_list)
2294 return scnprintf(msg, size, "%s",
2295 "No such device - did you specify an out-of-range profile CPU?\n");
2298 if (evsel->attr.precise_ip)
2299 return scnprintf(msg, size, "%s",
2300 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2301 #if defined(__i386__) || defined(__x86_64__)
2302 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2303 return scnprintf(msg, size, "%s",
2304 "No hardware sampling interrupt available.\n"
2305 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2309 if (find_process("oprofiled"))
2310 return scnprintf(msg, size,
2311 "The PMU counters are busy/taken by another profiler.\n"
2312 "We found oprofile daemon running, please stop it and try again.");
2315 if (perf_missing_features.clockid)
2316 return scnprintf(msg, size, "clockid feature not supported.");
2317 if (perf_missing_features.clockid_wrong)
2318 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2324 return scnprintf(msg, size,
2325 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2326 "/bin/dmesg may provide additional information.\n"
2327 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2328 err, strerror_r(err, sbuf, sizeof(sbuf)),
2329 perf_evsel__name(evsel));