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 <lk/debugfs.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <sys/resource.h>
22 #include "thread_map.h"
24 #include "perf_regs.h"
31 } perf_missing_features;
33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
35 int __perf_evsel__sample_size(u64 sample_type)
37 u64 mask = sample_type & PERF_SAMPLE_MASK;
41 for (i = 0; i < 64; i++) {
42 if (mask & (1ULL << i))
52 * __perf_evsel__calc_id_pos - calculate id_pos.
53 * @sample_type: sample type
55 * This function returns the position of the event id (PERF_SAMPLE_ID or
56 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
59 static int __perf_evsel__calc_id_pos(u64 sample_type)
63 if (sample_type & PERF_SAMPLE_IDENTIFIER)
66 if (!(sample_type & PERF_SAMPLE_ID))
69 if (sample_type & PERF_SAMPLE_IP)
72 if (sample_type & PERF_SAMPLE_TID)
75 if (sample_type & PERF_SAMPLE_TIME)
78 if (sample_type & PERF_SAMPLE_ADDR)
85 * __perf_evsel__calc_is_pos - calculate is_pos.
86 * @sample_type: sample type
88 * This function returns the position (counting backwards) of the event id
89 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
90 * sample_id_all is used there is an id sample appended to non-sample events.
92 static int __perf_evsel__calc_is_pos(u64 sample_type)
96 if (sample_type & PERF_SAMPLE_IDENTIFIER)
99 if (!(sample_type & PERF_SAMPLE_ID))
102 if (sample_type & PERF_SAMPLE_CPU)
105 if (sample_type & PERF_SAMPLE_STREAM_ID)
111 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
113 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
114 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
117 void hists__init(struct hists *hists)
119 memset(hists, 0, sizeof(*hists));
120 hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
121 hists->entries_in = &hists->entries_in_array[0];
122 hists->entries_collapsed = RB_ROOT;
123 hists->entries = RB_ROOT;
124 pthread_mutex_init(&hists->lock, NULL);
127 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
128 enum perf_event_sample_format bit)
130 if (!(evsel->attr.sample_type & bit)) {
131 evsel->attr.sample_type |= bit;
132 evsel->sample_size += sizeof(u64);
133 perf_evsel__calc_id_pos(evsel);
137 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
138 enum perf_event_sample_format bit)
140 if (evsel->attr.sample_type & bit) {
141 evsel->attr.sample_type &= ~bit;
142 evsel->sample_size -= sizeof(u64);
143 perf_evsel__calc_id_pos(evsel);
147 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
148 bool can_sample_identifier)
150 if (can_sample_identifier) {
151 perf_evsel__reset_sample_bit(evsel, ID);
152 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
154 perf_evsel__set_sample_bit(evsel, ID);
156 evsel->attr.read_format |= PERF_FORMAT_ID;
159 void perf_evsel__init(struct perf_evsel *evsel,
160 struct perf_event_attr *attr, int idx)
164 evsel->leader = evsel;
167 INIT_LIST_HEAD(&evsel->node);
168 hists__init(&evsel->hists);
169 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
170 perf_evsel__calc_id_pos(evsel);
173 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
175 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
178 perf_evsel__init(evsel, attr, idx);
183 struct event_format *event_format__new(const char *sys, const char *name)
187 void *bf = NULL, *nbf;
188 size_t size = 0, alloc_size = 0;
189 struct event_format *format = NULL;
191 if (asprintf(&filename, "%s/%s/%s/format", tracing_events_path, sys, name) < 0)
194 fd = open(filename, O_RDONLY);
196 goto out_free_filename;
199 if (size == alloc_size) {
200 alloc_size += BUFSIZ;
201 nbf = realloc(bf, alloc_size);
207 n = read(fd, bf + size, alloc_size - size);
213 pevent_parse_format(&format, bf, size, sys);
224 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
226 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
229 struct perf_event_attr attr = {
230 .type = PERF_TYPE_TRACEPOINT,
231 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
232 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
235 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
238 evsel->tp_format = event_format__new(sys, name);
239 if (evsel->tp_format == NULL)
242 event_attr_init(&attr);
243 attr.config = evsel->tp_format->id;
244 attr.sample_period = 1;
245 perf_evsel__init(evsel, &attr, idx);
256 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
264 "stalled-cycles-frontend",
265 "stalled-cycles-backend",
269 static const char *__perf_evsel__hw_name(u64 config)
271 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
272 return perf_evsel__hw_names[config];
274 return "unknown-hardware";
277 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
279 int colon = 0, r = 0;
280 struct perf_event_attr *attr = &evsel->attr;
281 bool exclude_guest_default = false;
283 #define MOD_PRINT(context, mod) do { \
284 if (!attr->exclude_##context) { \
285 if (!colon) colon = ++r; \
286 r += scnprintf(bf + r, size - r, "%c", mod); \
289 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
290 MOD_PRINT(kernel, 'k');
291 MOD_PRINT(user, 'u');
293 exclude_guest_default = true;
296 if (attr->precise_ip) {
299 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
300 exclude_guest_default = true;
303 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
304 MOD_PRINT(host, 'H');
305 MOD_PRINT(guest, 'G');
313 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
315 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
316 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
319 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
332 static const char *__perf_evsel__sw_name(u64 config)
334 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
335 return perf_evsel__sw_names[config];
336 return "unknown-software";
339 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
341 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
342 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
345 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
349 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
351 if (type & HW_BREAKPOINT_R)
352 r += scnprintf(bf + r, size - r, "r");
354 if (type & HW_BREAKPOINT_W)
355 r += scnprintf(bf + r, size - r, "w");
357 if (type & HW_BREAKPOINT_X)
358 r += scnprintf(bf + r, size - r, "x");
363 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
365 struct perf_event_attr *attr = &evsel->attr;
366 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
367 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
370 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
371 [PERF_EVSEL__MAX_ALIASES] = {
372 { "L1-dcache", "l1-d", "l1d", "L1-data", },
373 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
375 { "dTLB", "d-tlb", "Data-TLB", },
376 { "iTLB", "i-tlb", "Instruction-TLB", },
377 { "branch", "branches", "bpu", "btb", "bpc", },
381 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
382 [PERF_EVSEL__MAX_ALIASES] = {
383 { "load", "loads", "read", },
384 { "store", "stores", "write", },
385 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
388 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
389 [PERF_EVSEL__MAX_ALIASES] = {
390 { "refs", "Reference", "ops", "access", },
391 { "misses", "miss", },
394 #define C(x) PERF_COUNT_HW_CACHE_##x
395 #define CACHE_READ (1 << C(OP_READ))
396 #define CACHE_WRITE (1 << C(OP_WRITE))
397 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
398 #define COP(x) (1 << x)
401 * cache operartion stat
402 * L1I : Read and prefetch only
403 * ITLB and BPU : Read-only
405 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
406 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
407 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
408 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
409 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
410 [C(ITLB)] = (CACHE_READ),
411 [C(BPU)] = (CACHE_READ),
412 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
415 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
417 if (perf_evsel__hw_cache_stat[type] & COP(op))
418 return true; /* valid */
420 return false; /* invalid */
423 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
424 char *bf, size_t size)
427 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
428 perf_evsel__hw_cache_op[op][0],
429 perf_evsel__hw_cache_result[result][0]);
432 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
433 perf_evsel__hw_cache_op[op][1]);
436 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
438 u8 op, result, type = (config >> 0) & 0xff;
439 const char *err = "unknown-ext-hardware-cache-type";
441 if (type > PERF_COUNT_HW_CACHE_MAX)
444 op = (config >> 8) & 0xff;
445 err = "unknown-ext-hardware-cache-op";
446 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
449 result = (config >> 16) & 0xff;
450 err = "unknown-ext-hardware-cache-result";
451 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
454 err = "invalid-cache";
455 if (!perf_evsel__is_cache_op_valid(type, op))
458 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
460 return scnprintf(bf, size, "%s", err);
463 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
465 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
466 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
469 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
471 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
472 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
475 const char *perf_evsel__name(struct perf_evsel *evsel)
482 switch (evsel->attr.type) {
484 perf_evsel__raw_name(evsel, bf, sizeof(bf));
487 case PERF_TYPE_HARDWARE:
488 perf_evsel__hw_name(evsel, bf, sizeof(bf));
491 case PERF_TYPE_HW_CACHE:
492 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
495 case PERF_TYPE_SOFTWARE:
496 perf_evsel__sw_name(evsel, bf, sizeof(bf));
499 case PERF_TYPE_TRACEPOINT:
500 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
503 case PERF_TYPE_BREAKPOINT:
504 perf_evsel__bp_name(evsel, bf, sizeof(bf));
508 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
513 evsel->name = strdup(bf);
515 return evsel->name ?: "unknown";
518 const char *perf_evsel__group_name(struct perf_evsel *evsel)
520 return evsel->group_name ?: "anon group";
523 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
526 struct perf_evsel *pos;
527 const char *group_name = perf_evsel__group_name(evsel);
529 ret = scnprintf(buf, size, "%s", group_name);
531 ret += scnprintf(buf + ret, size - ret, " { %s",
532 perf_evsel__name(evsel));
534 for_each_group_member(pos, evsel)
535 ret += scnprintf(buf + ret, size - ret, ", %s",
536 perf_evsel__name(pos));
538 ret += scnprintf(buf + ret, size - ret, " }");
544 * The enable_on_exec/disabled value strategy:
546 * 1) For any type of traced program:
547 * - all independent events and group leaders are disabled
548 * - all group members are enabled
550 * Group members are ruled by group leaders. They need to
551 * be enabled, because the group scheduling relies on that.
553 * 2) For traced programs executed by perf:
554 * - all independent events and group leaders have
556 * - we don't specifically enable or disable any event during
559 * Independent events and group leaders are initially disabled
560 * and get enabled by exec. Group members are ruled by group
561 * leaders as stated in 1).
563 * 3) For traced programs attached by perf (pid/tid):
564 * - we specifically enable or disable all events during
567 * When attaching events to already running traced we
568 * enable/disable events specifically, as there's no
569 * initial traced exec call.
571 void perf_evsel__config(struct perf_evsel *evsel,
572 struct perf_record_opts *opts)
574 struct perf_evsel *leader = evsel->leader;
575 struct perf_event_attr *attr = &evsel->attr;
576 int track = !evsel->idx; /* only the first counter needs these */
578 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
579 attr->inherit = !opts->no_inherit;
581 perf_evsel__set_sample_bit(evsel, IP);
582 perf_evsel__set_sample_bit(evsel, TID);
584 if (evsel->sample_read) {
585 perf_evsel__set_sample_bit(evsel, READ);
588 * We need ID even in case of single event, because
589 * PERF_SAMPLE_READ process ID specific data.
591 perf_evsel__set_sample_id(evsel, false);
594 * Apply group format only if we belong to group
595 * with more than one members.
597 if (leader->nr_members > 1) {
598 attr->read_format |= PERF_FORMAT_GROUP;
604 * We default some events to a 1 default interval. But keep
605 * it a weak assumption overridable by the user.
607 if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
608 opts->user_interval != ULLONG_MAX)) {
610 perf_evsel__set_sample_bit(evsel, PERIOD);
612 attr->sample_freq = opts->freq;
614 attr->sample_period = opts->default_interval;
619 * Disable sampling for all group members other
620 * than leader in case leader 'leads' the sampling.
622 if ((leader != evsel) && leader->sample_read) {
623 attr->sample_freq = 0;
624 attr->sample_period = 0;
627 if (opts->no_samples)
628 attr->sample_freq = 0;
630 if (opts->inherit_stat)
631 attr->inherit_stat = 1;
633 if (opts->sample_address) {
634 perf_evsel__set_sample_bit(evsel, ADDR);
635 attr->mmap_data = track;
638 if (opts->call_graph) {
639 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
641 if (opts->call_graph == CALLCHAIN_DWARF) {
642 perf_evsel__set_sample_bit(evsel, REGS_USER);
643 perf_evsel__set_sample_bit(evsel, STACK_USER);
644 attr->sample_regs_user = PERF_REGS_MASK;
645 attr->sample_stack_user = opts->stack_dump_size;
646 attr->exclude_callchain_user = 1;
650 if (target__has_cpu(&opts->target) || opts->target.force_per_cpu)
651 perf_evsel__set_sample_bit(evsel, CPU);
654 perf_evsel__set_sample_bit(evsel, PERIOD);
656 if (!perf_missing_features.sample_id_all &&
657 (opts->sample_time || !opts->no_inherit ||
658 target__has_cpu(&opts->target) || opts->target.force_per_cpu))
659 perf_evsel__set_sample_bit(evsel, TIME);
661 if (opts->raw_samples) {
662 perf_evsel__set_sample_bit(evsel, TIME);
663 perf_evsel__set_sample_bit(evsel, RAW);
664 perf_evsel__set_sample_bit(evsel, CPU);
667 if (opts->sample_address)
668 perf_evsel__set_sample_bit(evsel, DATA_SRC);
670 if (opts->no_delay) {
672 attr->wakeup_events = 1;
674 if (opts->branch_stack) {
675 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
676 attr->branch_sample_type = opts->branch_stack;
679 if (opts->sample_weight)
680 perf_evsel__set_sample_bit(evsel, WEIGHT);
685 if (opts->sample_transaction)
686 perf_evsel__set_sample_bit(evsel, TRANSACTION);
689 * XXX see the function comment above
691 * Disabling only independent events or group leaders,
692 * keeping group members enabled.
694 if (perf_evsel__is_group_leader(evsel))
698 * Setting enable_on_exec for independent events and
699 * group leaders for traced executed by perf.
701 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel))
702 attr->enable_on_exec = 1;
705 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
708 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
711 for (cpu = 0; cpu < ncpus; cpu++) {
712 for (thread = 0; thread < nthreads; thread++) {
713 FD(evsel, cpu, thread) = -1;
718 return evsel->fd != NULL ? 0 : -ENOMEM;
721 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
726 for (cpu = 0; cpu < ncpus; cpu++) {
727 for (thread = 0; thread < nthreads; thread++) {
728 int fd = FD(evsel, cpu, thread),
729 err = ioctl(fd, ioc, arg);
739 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
742 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
743 PERF_EVENT_IOC_SET_FILTER,
747 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
749 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
750 PERF_EVENT_IOC_ENABLE,
754 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
756 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
757 if (evsel->sample_id == NULL)
760 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
761 if (evsel->id == NULL) {
762 xyarray__delete(evsel->sample_id);
763 evsel->sample_id = NULL;
770 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
772 memset(evsel->counts, 0, (sizeof(*evsel->counts) +
773 (ncpus * sizeof(struct perf_counts_values))));
776 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
778 evsel->counts = zalloc((sizeof(*evsel->counts) +
779 (ncpus * sizeof(struct perf_counts_values))));
780 return evsel->counts != NULL ? 0 : -ENOMEM;
783 void perf_evsel__free_fd(struct perf_evsel *evsel)
785 xyarray__delete(evsel->fd);
789 void perf_evsel__free_id(struct perf_evsel *evsel)
791 xyarray__delete(evsel->sample_id);
792 evsel->sample_id = NULL;
797 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
801 for (cpu = 0; cpu < ncpus; cpu++)
802 for (thread = 0; thread < nthreads; ++thread) {
803 close(FD(evsel, cpu, thread));
804 FD(evsel, cpu, thread) = -1;
808 void perf_evsel__free_counts(struct perf_evsel *evsel)
813 void perf_evsel__exit(struct perf_evsel *evsel)
815 assert(list_empty(&evsel->node));
816 perf_evsel__free_fd(evsel);
817 perf_evsel__free_id(evsel);
820 void perf_evsel__delete(struct perf_evsel *evsel)
822 perf_evsel__exit(evsel);
823 close_cgroup(evsel->cgrp);
824 free(evsel->group_name);
825 if (evsel->tp_format)
826 pevent_free_format(evsel->tp_format);
831 static inline void compute_deltas(struct perf_evsel *evsel,
833 struct perf_counts_values *count)
835 struct perf_counts_values tmp;
837 if (!evsel->prev_raw_counts)
841 tmp = evsel->prev_raw_counts->aggr;
842 evsel->prev_raw_counts->aggr = *count;
844 tmp = evsel->prev_raw_counts->cpu[cpu];
845 evsel->prev_raw_counts->cpu[cpu] = *count;
848 count->val = count->val - tmp.val;
849 count->ena = count->ena - tmp.ena;
850 count->run = count->run - tmp.run;
853 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
854 int cpu, int thread, bool scale)
856 struct perf_counts_values count;
857 size_t nv = scale ? 3 : 1;
859 if (FD(evsel, cpu, thread) < 0)
862 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
865 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
868 compute_deltas(evsel, cpu, &count);
873 else if (count.run < count.ena)
874 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
876 count.ena = count.run = 0;
878 evsel->counts->cpu[cpu] = count;
882 int __perf_evsel__read(struct perf_evsel *evsel,
883 int ncpus, int nthreads, bool scale)
885 size_t nv = scale ? 3 : 1;
887 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
889 aggr->val = aggr->ena = aggr->run = 0;
891 for (cpu = 0; cpu < ncpus; cpu++) {
892 for (thread = 0; thread < nthreads; thread++) {
893 if (FD(evsel, cpu, thread) < 0)
896 if (readn(FD(evsel, cpu, thread),
897 &count, nv * sizeof(u64)) < 0)
900 aggr->val += count.val;
902 aggr->ena += count.ena;
903 aggr->run += count.run;
908 compute_deltas(evsel, -1, aggr);
910 evsel->counts->scaled = 0;
912 if (aggr->run == 0) {
913 evsel->counts->scaled = -1;
918 if (aggr->run < aggr->ena) {
919 evsel->counts->scaled = 1;
920 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
923 aggr->ena = aggr->run = 0;
928 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
930 struct perf_evsel *leader = evsel->leader;
933 if (perf_evsel__is_group_leader(evsel))
937 * Leader must be already processed/open,
942 fd = FD(leader, cpu, thread);
948 #define __PRINT_ATTR(fmt, cast, field) \
949 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
951 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
952 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
953 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
954 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
956 #define PRINT_ATTR2N(name1, field1, name2, field2) \
957 fprintf(fp, " %-19s %u %-19s %u\n", \
958 name1, attr->field1, name2, attr->field2)
960 #define PRINT_ATTR2(field1, field2) \
961 PRINT_ATTR2N(#field1, field1, #field2, field2)
963 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
967 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
968 ret += fprintf(fp, "perf_event_attr:\n");
970 ret += PRINT_ATTR_U32(type);
971 ret += PRINT_ATTR_U32(size);
972 ret += PRINT_ATTR_X64(config);
973 ret += PRINT_ATTR_U64(sample_period);
974 ret += PRINT_ATTR_U64(sample_freq);
975 ret += PRINT_ATTR_X64(sample_type);
976 ret += PRINT_ATTR_X64(read_format);
978 ret += PRINT_ATTR2(disabled, inherit);
979 ret += PRINT_ATTR2(pinned, exclusive);
980 ret += PRINT_ATTR2(exclude_user, exclude_kernel);
981 ret += PRINT_ATTR2(exclude_hv, exclude_idle);
982 ret += PRINT_ATTR2(mmap, comm);
983 ret += PRINT_ATTR2(freq, inherit_stat);
984 ret += PRINT_ATTR2(enable_on_exec, task);
985 ret += PRINT_ATTR2(watermark, precise_ip);
986 ret += PRINT_ATTR2(mmap_data, sample_id_all);
987 ret += PRINT_ATTR2(exclude_host, exclude_guest);
988 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
989 "excl.callchain_user", exclude_callchain_user);
990 ret += PRINT_ATTR_U32(mmap2);
992 ret += PRINT_ATTR_U32(wakeup_events);
993 ret += PRINT_ATTR_U32(wakeup_watermark);
994 ret += PRINT_ATTR_X32(bp_type);
995 ret += PRINT_ATTR_X64(bp_addr);
996 ret += PRINT_ATTR_X64(config1);
997 ret += PRINT_ATTR_U64(bp_len);
998 ret += PRINT_ATTR_X64(config2);
999 ret += PRINT_ATTR_X64(branch_sample_type);
1000 ret += PRINT_ATTR_X64(sample_regs_user);
1001 ret += PRINT_ATTR_U32(sample_stack_user);
1003 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1008 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1009 struct thread_map *threads)
1012 unsigned long flags = 0;
1014 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1016 if (evsel->fd == NULL &&
1017 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
1021 flags = PERF_FLAG_PID_CGROUP;
1022 pid = evsel->cgrp->fd;
1025 fallback_missing_features:
1026 if (perf_missing_features.mmap2)
1027 evsel->attr.mmap2 = 0;
1028 if (perf_missing_features.exclude_guest)
1029 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1031 if (perf_missing_features.sample_id_all)
1032 evsel->attr.sample_id_all = 0;
1035 perf_event_attr__fprintf(&evsel->attr, stderr);
1037 for (cpu = 0; cpu < cpus->nr; cpu++) {
1039 for (thread = 0; thread < threads->nr; thread++) {
1043 pid = threads->map[thread];
1045 group_fd = get_group_fd(evsel, cpu, thread);
1047 pr_debug2("perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1048 pid, cpus->map[cpu], group_fd, flags);
1050 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1054 if (FD(evsel, cpu, thread) < 0) {
1056 pr_debug2("perf_event_open failed, error %d\n",
1060 set_rlimit = NO_CHANGE;
1068 * perf stat needs between 5 and 22 fds per CPU. When we run out
1069 * of them try to increase the limits.
1071 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1073 int old_errno = errno;
1075 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1076 if (set_rlimit == NO_CHANGE)
1077 l.rlim_cur = l.rlim_max;
1079 l.rlim_cur = l.rlim_max + 1000;
1080 l.rlim_max = l.rlim_cur;
1082 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1091 if (err != -EINVAL || cpu > 0 || thread > 0)
1094 if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1095 perf_missing_features.mmap2 = true;
1096 goto fallback_missing_features;
1097 } else if (!perf_missing_features.exclude_guest &&
1098 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1099 perf_missing_features.exclude_guest = true;
1100 goto fallback_missing_features;
1101 } else if (!perf_missing_features.sample_id_all) {
1102 perf_missing_features.sample_id_all = true;
1103 goto retry_sample_id;
1108 while (--thread >= 0) {
1109 close(FD(evsel, cpu, thread));
1110 FD(evsel, cpu, thread) = -1;
1112 thread = threads->nr;
1113 } while (--cpu >= 0);
1117 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1119 if (evsel->fd == NULL)
1122 perf_evsel__close_fd(evsel, ncpus, nthreads);
1123 perf_evsel__free_fd(evsel);
1136 struct thread_map map;
1138 } empty_thread_map = {
1143 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1144 struct thread_map *threads)
1147 /* Work around old compiler warnings about strict aliasing */
1148 cpus = &empty_cpu_map.map;
1151 if (threads == NULL)
1152 threads = &empty_thread_map.map;
1154 return __perf_evsel__open(evsel, cpus, threads);
1157 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1158 struct cpu_map *cpus)
1160 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1163 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1164 struct thread_map *threads)
1166 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1169 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1170 const union perf_event *event,
1171 struct perf_sample *sample)
1173 u64 type = evsel->attr.sample_type;
1174 const u64 *array = event->sample.array;
1175 bool swapped = evsel->needs_swap;
1178 array += ((event->header.size -
1179 sizeof(event->header)) / sizeof(u64)) - 1;
1181 if (type & PERF_SAMPLE_IDENTIFIER) {
1182 sample->id = *array;
1186 if (type & PERF_SAMPLE_CPU) {
1189 /* undo swap of u64, then swap on individual u32s */
1190 u.val64 = bswap_64(u.val64);
1191 u.val32[0] = bswap_32(u.val32[0]);
1194 sample->cpu = u.val32[0];
1198 if (type & PERF_SAMPLE_STREAM_ID) {
1199 sample->stream_id = *array;
1203 if (type & PERF_SAMPLE_ID) {
1204 sample->id = *array;
1208 if (type & PERF_SAMPLE_TIME) {
1209 sample->time = *array;
1213 if (type & PERF_SAMPLE_TID) {
1216 /* undo swap of u64, then swap on individual u32s */
1217 u.val64 = bswap_64(u.val64);
1218 u.val32[0] = bswap_32(u.val32[0]);
1219 u.val32[1] = bswap_32(u.val32[1]);
1222 sample->pid = u.val32[0];
1223 sample->tid = u.val32[1];
1230 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1233 return size > max_size || offset + size > endp;
1236 #define OVERFLOW_CHECK(offset, size, max_size) \
1238 if (overflow(endp, (max_size), (offset), (size))) \
1242 #define OVERFLOW_CHECK_u64(offset) \
1243 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1245 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1246 struct perf_sample *data)
1248 u64 type = evsel->attr.sample_type;
1249 bool swapped = evsel->needs_swap;
1251 u16 max_size = event->header.size;
1252 const void *endp = (void *)event + max_size;
1256 * used for cross-endian analysis. See git commit 65014ab3
1257 * for why this goofiness is needed.
1261 memset(data, 0, sizeof(*data));
1262 data->cpu = data->pid = data->tid = -1;
1263 data->stream_id = data->id = data->time = -1ULL;
1267 if (event->header.type != PERF_RECORD_SAMPLE) {
1268 if (!evsel->attr.sample_id_all)
1270 return perf_evsel__parse_id_sample(evsel, event, data);
1273 array = event->sample.array;
1276 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1277 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1278 * check the format does not go past the end of the event.
1280 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1284 if (type & PERF_SAMPLE_IDENTIFIER) {
1289 if (type & PERF_SAMPLE_IP) {
1294 if (type & PERF_SAMPLE_TID) {
1297 /* undo swap of u64, then swap on individual u32s */
1298 u.val64 = bswap_64(u.val64);
1299 u.val32[0] = bswap_32(u.val32[0]);
1300 u.val32[1] = bswap_32(u.val32[1]);
1303 data->pid = u.val32[0];
1304 data->tid = u.val32[1];
1308 if (type & PERF_SAMPLE_TIME) {
1309 data->time = *array;
1314 if (type & PERF_SAMPLE_ADDR) {
1315 data->addr = *array;
1319 if (type & PERF_SAMPLE_ID) {
1324 if (type & PERF_SAMPLE_STREAM_ID) {
1325 data->stream_id = *array;
1329 if (type & PERF_SAMPLE_CPU) {
1333 /* undo swap of u64, then swap on individual u32s */
1334 u.val64 = bswap_64(u.val64);
1335 u.val32[0] = bswap_32(u.val32[0]);
1338 data->cpu = u.val32[0];
1342 if (type & PERF_SAMPLE_PERIOD) {
1343 data->period = *array;
1347 if (type & PERF_SAMPLE_READ) {
1348 u64 read_format = evsel->attr.read_format;
1350 OVERFLOW_CHECK_u64(array);
1351 if (read_format & PERF_FORMAT_GROUP)
1352 data->read.group.nr = *array;
1354 data->read.one.value = *array;
1358 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1359 OVERFLOW_CHECK_u64(array);
1360 data->read.time_enabled = *array;
1364 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1365 OVERFLOW_CHECK_u64(array);
1366 data->read.time_running = *array;
1370 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1371 if (read_format & PERF_FORMAT_GROUP) {
1372 const u64 max_group_nr = UINT64_MAX /
1373 sizeof(struct sample_read_value);
1375 if (data->read.group.nr > max_group_nr)
1377 sz = data->read.group.nr *
1378 sizeof(struct sample_read_value);
1379 OVERFLOW_CHECK(array, sz, max_size);
1380 data->read.group.values =
1381 (struct sample_read_value *)array;
1382 array = (void *)array + sz;
1384 OVERFLOW_CHECK_u64(array);
1385 data->read.one.id = *array;
1390 if (type & PERF_SAMPLE_CALLCHAIN) {
1391 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1393 OVERFLOW_CHECK_u64(array);
1394 data->callchain = (struct ip_callchain *)array++;
1395 if (data->callchain->nr > max_callchain_nr)
1397 sz = data->callchain->nr * sizeof(u64);
1398 OVERFLOW_CHECK(array, sz, max_size);
1399 array = (void *)array + sz;
1402 if (type & PERF_SAMPLE_RAW) {
1403 OVERFLOW_CHECK_u64(array);
1405 if (WARN_ONCE(swapped,
1406 "Endianness of raw data not corrected!\n")) {
1407 /* undo swap of u64, then swap on individual u32s */
1408 u.val64 = bswap_64(u.val64);
1409 u.val32[0] = bswap_32(u.val32[0]);
1410 u.val32[1] = bswap_32(u.val32[1]);
1412 data->raw_size = u.val32[0];
1413 array = (void *)array + sizeof(u32);
1415 OVERFLOW_CHECK(array, data->raw_size, max_size);
1416 data->raw_data = (void *)array;
1417 array = (void *)array + data->raw_size;
1420 if (type & PERF_SAMPLE_BRANCH_STACK) {
1421 const u64 max_branch_nr = UINT64_MAX /
1422 sizeof(struct branch_entry);
1424 OVERFLOW_CHECK_u64(array);
1425 data->branch_stack = (struct branch_stack *)array++;
1427 if (data->branch_stack->nr > max_branch_nr)
1429 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1430 OVERFLOW_CHECK(array, sz, max_size);
1431 array = (void *)array + sz;
1434 if (type & PERF_SAMPLE_REGS_USER) {
1435 OVERFLOW_CHECK_u64(array);
1436 data->user_regs.abi = *array;
1439 if (data->user_regs.abi) {
1440 u64 regs_user = evsel->attr.sample_regs_user;
1442 sz = hweight_long(regs_user) * sizeof(u64);
1443 OVERFLOW_CHECK(array, sz, max_size);
1444 data->user_regs.regs = (u64 *)array;
1445 array = (void *)array + sz;
1449 if (type & PERF_SAMPLE_STACK_USER) {
1450 OVERFLOW_CHECK_u64(array);
1453 data->user_stack.offset = ((char *)(array - 1)
1457 data->user_stack.size = 0;
1459 OVERFLOW_CHECK(array, sz, max_size);
1460 data->user_stack.data = (char *)array;
1461 array = (void *)array + sz;
1462 OVERFLOW_CHECK_u64(array);
1463 data->user_stack.size = *array++;
1464 if (WARN_ONCE(data->user_stack.size > sz,
1465 "user stack dump failure\n"))
1471 if (type & PERF_SAMPLE_WEIGHT) {
1472 OVERFLOW_CHECK_u64(array);
1473 data->weight = *array;
1477 data->data_src = PERF_MEM_DATA_SRC_NONE;
1478 if (type & PERF_SAMPLE_DATA_SRC) {
1479 OVERFLOW_CHECK_u64(array);
1480 data->data_src = *array;
1484 data->transaction = 0;
1485 if (type & PERF_SAMPLE_TRANSACTION) {
1486 OVERFLOW_CHECK_u64(array);
1487 data->transaction = *array;
1494 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1495 u64 sample_regs_user, u64 read_format)
1497 size_t sz, result = sizeof(struct sample_event);
1499 if (type & PERF_SAMPLE_IDENTIFIER)
1500 result += sizeof(u64);
1502 if (type & PERF_SAMPLE_IP)
1503 result += sizeof(u64);
1505 if (type & PERF_SAMPLE_TID)
1506 result += sizeof(u64);
1508 if (type & PERF_SAMPLE_TIME)
1509 result += sizeof(u64);
1511 if (type & PERF_SAMPLE_ADDR)
1512 result += sizeof(u64);
1514 if (type & PERF_SAMPLE_ID)
1515 result += sizeof(u64);
1517 if (type & PERF_SAMPLE_STREAM_ID)
1518 result += sizeof(u64);
1520 if (type & PERF_SAMPLE_CPU)
1521 result += sizeof(u64);
1523 if (type & PERF_SAMPLE_PERIOD)
1524 result += sizeof(u64);
1526 if (type & PERF_SAMPLE_READ) {
1527 result += sizeof(u64);
1528 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1529 result += sizeof(u64);
1530 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1531 result += sizeof(u64);
1532 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1533 if (read_format & PERF_FORMAT_GROUP) {
1534 sz = sample->read.group.nr *
1535 sizeof(struct sample_read_value);
1538 result += sizeof(u64);
1542 if (type & PERF_SAMPLE_CALLCHAIN) {
1543 sz = (sample->callchain->nr + 1) * sizeof(u64);
1547 if (type & PERF_SAMPLE_RAW) {
1548 result += sizeof(u32);
1549 result += sample->raw_size;
1552 if (type & PERF_SAMPLE_BRANCH_STACK) {
1553 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1558 if (type & PERF_SAMPLE_REGS_USER) {
1559 if (sample->user_regs.abi) {
1560 result += sizeof(u64);
1561 sz = hweight_long(sample_regs_user) * sizeof(u64);
1564 result += sizeof(u64);
1568 if (type & PERF_SAMPLE_STACK_USER) {
1569 sz = sample->user_stack.size;
1570 result += sizeof(u64);
1573 result += sizeof(u64);
1577 if (type & PERF_SAMPLE_WEIGHT)
1578 result += sizeof(u64);
1580 if (type & PERF_SAMPLE_DATA_SRC)
1581 result += sizeof(u64);
1583 if (type & PERF_SAMPLE_TRANSACTION)
1584 result += sizeof(u64);
1589 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1590 u64 sample_regs_user, u64 read_format,
1591 const struct perf_sample *sample,
1597 * used for cross-endian analysis. See git commit 65014ab3
1598 * for why this goofiness is needed.
1602 array = event->sample.array;
1604 if (type & PERF_SAMPLE_IDENTIFIER) {
1605 *array = sample->id;
1609 if (type & PERF_SAMPLE_IP) {
1610 *array = sample->ip;
1614 if (type & PERF_SAMPLE_TID) {
1615 u.val32[0] = sample->pid;
1616 u.val32[1] = sample->tid;
1619 * Inverse of what is done in perf_evsel__parse_sample
1621 u.val32[0] = bswap_32(u.val32[0]);
1622 u.val32[1] = bswap_32(u.val32[1]);
1623 u.val64 = bswap_64(u.val64);
1630 if (type & PERF_SAMPLE_TIME) {
1631 *array = sample->time;
1635 if (type & PERF_SAMPLE_ADDR) {
1636 *array = sample->addr;
1640 if (type & PERF_SAMPLE_ID) {
1641 *array = sample->id;
1645 if (type & PERF_SAMPLE_STREAM_ID) {
1646 *array = sample->stream_id;
1650 if (type & PERF_SAMPLE_CPU) {
1651 u.val32[0] = sample->cpu;
1654 * Inverse of what is done in perf_evsel__parse_sample
1656 u.val32[0] = bswap_32(u.val32[0]);
1657 u.val64 = bswap_64(u.val64);
1663 if (type & PERF_SAMPLE_PERIOD) {
1664 *array = sample->period;
1668 if (type & PERF_SAMPLE_READ) {
1669 if (read_format & PERF_FORMAT_GROUP)
1670 *array = sample->read.group.nr;
1672 *array = sample->read.one.value;
1675 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1676 *array = sample->read.time_enabled;
1680 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1681 *array = sample->read.time_running;
1685 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1686 if (read_format & PERF_FORMAT_GROUP) {
1687 sz = sample->read.group.nr *
1688 sizeof(struct sample_read_value);
1689 memcpy(array, sample->read.group.values, sz);
1690 array = (void *)array + sz;
1692 *array = sample->read.one.id;
1697 if (type & PERF_SAMPLE_CALLCHAIN) {
1698 sz = (sample->callchain->nr + 1) * sizeof(u64);
1699 memcpy(array, sample->callchain, sz);
1700 array = (void *)array + sz;
1703 if (type & PERF_SAMPLE_RAW) {
1704 u.val32[0] = sample->raw_size;
1705 if (WARN_ONCE(swapped,
1706 "Endianness of raw data not corrected!\n")) {
1708 * Inverse of what is done in perf_evsel__parse_sample
1710 u.val32[0] = bswap_32(u.val32[0]);
1711 u.val32[1] = bswap_32(u.val32[1]);
1712 u.val64 = bswap_64(u.val64);
1715 array = (void *)array + sizeof(u32);
1717 memcpy(array, sample->raw_data, sample->raw_size);
1718 array = (void *)array + sample->raw_size;
1721 if (type & PERF_SAMPLE_BRANCH_STACK) {
1722 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1724 memcpy(array, sample->branch_stack, sz);
1725 array = (void *)array + sz;
1728 if (type & PERF_SAMPLE_REGS_USER) {
1729 if (sample->user_regs.abi) {
1730 *array++ = sample->user_regs.abi;
1731 sz = hweight_long(sample_regs_user) * sizeof(u64);
1732 memcpy(array, sample->user_regs.regs, sz);
1733 array = (void *)array + sz;
1739 if (type & PERF_SAMPLE_STACK_USER) {
1740 sz = sample->user_stack.size;
1743 memcpy(array, sample->user_stack.data, sz);
1744 array = (void *)array + sz;
1749 if (type & PERF_SAMPLE_WEIGHT) {
1750 *array = sample->weight;
1754 if (type & PERF_SAMPLE_DATA_SRC) {
1755 *array = sample->data_src;
1759 if (type & PERF_SAMPLE_TRANSACTION) {
1760 *array = sample->transaction;
1767 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1769 return pevent_find_field(evsel->tp_format, name);
1772 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1775 struct format_field *field = perf_evsel__field(evsel, name);
1781 offset = field->offset;
1783 if (field->flags & FIELD_IS_DYNAMIC) {
1784 offset = *(int *)(sample->raw_data + field->offset);
1788 return sample->raw_data + offset;
1791 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1794 struct format_field *field = perf_evsel__field(evsel, name);
1801 ptr = sample->raw_data + field->offset;
1803 switch (field->size) {
1807 value = *(u16 *)ptr;
1810 value = *(u32 *)ptr;
1813 value = *(u64 *)ptr;
1819 if (!evsel->needs_swap)
1822 switch (field->size) {
1824 return bswap_16(value);
1826 return bswap_32(value);
1828 return bswap_64(value);
1836 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1842 ret += fprintf(fp, ",");
1844 ret += fprintf(fp, ":");
1848 va_start(args, fmt);
1849 ret += vfprintf(fp, fmt, args);
1854 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1859 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1862 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1869 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1870 struct bit_names *bits, bool *first)
1872 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1873 bool first_bit = true;
1876 if (value & bits[i].bit) {
1877 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1880 } while (bits[++i].name != NULL);
1885 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1887 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1888 struct bit_names bits[] = {
1889 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1890 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1891 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1892 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1893 bit_name(IDENTIFIER),
1897 return bits__fprintf(fp, "sample_type", value, bits, first);
1900 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1902 #define bit_name(n) { PERF_FORMAT_##n, #n }
1903 struct bit_names bits[] = {
1904 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1905 bit_name(ID), bit_name(GROUP),
1909 return bits__fprintf(fp, "read_format", value, bits, first);
1912 int perf_evsel__fprintf(struct perf_evsel *evsel,
1913 struct perf_attr_details *details, FILE *fp)
1918 if (details->event_group) {
1919 struct perf_evsel *pos;
1921 if (!perf_evsel__is_group_leader(evsel))
1924 if (evsel->nr_members > 1)
1925 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1927 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1928 for_each_group_member(pos, evsel)
1929 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1931 if (evsel->nr_members > 1)
1932 printed += fprintf(fp, "}");
1936 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1938 if (details->verbose || details->freq) {
1939 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1940 (u64)evsel->attr.sample_freq);
1943 if (details->verbose) {
1949 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1950 if (evsel->attr.read_format)
1951 printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1955 if_print(exclusive);
1956 if_print(exclude_user);
1957 if_print(exclude_kernel);
1958 if_print(exclude_hv);
1959 if_print(exclude_idle);
1964 if_print(inherit_stat);
1965 if_print(enable_on_exec);
1967 if_print(watermark);
1968 if_print(precise_ip);
1969 if_print(mmap_data);
1970 if_print(sample_id_all);
1971 if_print(exclude_host);
1972 if_print(exclude_guest);
1973 if_print(__reserved_1);
1974 if_print(wakeup_events);
1976 if_print(branch_sample_type);
1983 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1984 char *msg, size_t msgsize)
1986 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
1987 evsel->attr.type == PERF_TYPE_HARDWARE &&
1988 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1990 * If it's cycles then fall back to hrtimer based
1991 * cpu-clock-tick sw counter, which is always available even if
1994 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1997 scnprintf(msg, msgsize, "%s",
1998 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2000 evsel->attr.type = PERF_TYPE_SOFTWARE;
2001 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2011 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2012 int err, char *msg, size_t size)
2017 return scnprintf(msg, size,
2018 "You may not have permission to collect %sstats.\n"
2019 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2020 " -1 - Not paranoid at all\n"
2021 " 0 - Disallow raw tracepoint access for unpriv\n"
2022 " 1 - Disallow cpu events for unpriv\n"
2023 " 2 - Disallow kernel profiling for unpriv",
2024 target->system_wide ? "system-wide " : "");
2026 return scnprintf(msg, size, "The %s event is not supported.",
2027 perf_evsel__name(evsel));
2029 return scnprintf(msg, size, "%s",
2030 "Too many events are opened.\n"
2031 "Try again after reducing the number of events.");
2033 if (target->cpu_list)
2034 return scnprintf(msg, size, "%s",
2035 "No such device - did you specify an out-of-range profile CPU?\n");
2038 if (evsel->attr.precise_ip)
2039 return scnprintf(msg, size, "%s",
2040 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2041 #if defined(__i386__) || defined(__x86_64__)
2042 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2043 return scnprintf(msg, size, "%s",
2044 "No hardware sampling interrupt available.\n"
2045 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2052 return scnprintf(msg, size,
2053 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2054 "/bin/dmesg may provide additional information.\n"
2055 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2056 err, strerror(err), perf_evsel__name(evsel));