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"
30 } perf_missing_features;
32 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
34 int __perf_evsel__sample_size(u64 sample_type)
36 u64 mask = sample_type & PERF_SAMPLE_MASK;
40 for (i = 0; i < 64; i++) {
41 if (mask & (1ULL << i))
51 * __perf_evsel__calc_id_pos - calculate id_pos.
52 * @sample_type: sample type
54 * This function returns the position of the event id (PERF_SAMPLE_ID or
55 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
58 static int __perf_evsel__calc_id_pos(u64 sample_type)
62 if (sample_type & PERF_SAMPLE_IDENTIFIER)
65 if (!(sample_type & PERF_SAMPLE_ID))
68 if (sample_type & PERF_SAMPLE_IP)
71 if (sample_type & PERF_SAMPLE_TID)
74 if (sample_type & PERF_SAMPLE_TIME)
77 if (sample_type & PERF_SAMPLE_ADDR)
84 * __perf_evsel__calc_is_pos - calculate is_pos.
85 * @sample_type: sample type
87 * This function returns the position (counting backwards) of the event id
88 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
89 * sample_id_all is used there is an id sample appended to non-sample events.
91 static int __perf_evsel__calc_is_pos(u64 sample_type)
95 if (sample_type & PERF_SAMPLE_IDENTIFIER)
98 if (!(sample_type & PERF_SAMPLE_ID))
101 if (sample_type & PERF_SAMPLE_CPU)
104 if (sample_type & PERF_SAMPLE_STREAM_ID)
110 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
112 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
113 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
116 void hists__init(struct hists *hists)
118 memset(hists, 0, sizeof(*hists));
119 hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
120 hists->entries_in = &hists->entries_in_array[0];
121 hists->entries_collapsed = RB_ROOT;
122 hists->entries = RB_ROOT;
123 pthread_mutex_init(&hists->lock, NULL);
126 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
127 enum perf_event_sample_format bit)
129 if (!(evsel->attr.sample_type & bit)) {
130 evsel->attr.sample_type |= bit;
131 evsel->sample_size += sizeof(u64);
132 perf_evsel__calc_id_pos(evsel);
136 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
137 enum perf_event_sample_format bit)
139 if (evsel->attr.sample_type & bit) {
140 evsel->attr.sample_type &= ~bit;
141 evsel->sample_size -= sizeof(u64);
142 perf_evsel__calc_id_pos(evsel);
146 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
147 bool can_sample_identifier)
149 if (can_sample_identifier) {
150 perf_evsel__reset_sample_bit(evsel, ID);
151 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
153 perf_evsel__set_sample_bit(evsel, ID);
155 evsel->attr.read_format |= PERF_FORMAT_ID;
158 void perf_evsel__init(struct perf_evsel *evsel,
159 struct perf_event_attr *attr, int idx)
163 evsel->leader = evsel;
164 INIT_LIST_HEAD(&evsel->node);
165 hists__init(&evsel->hists);
166 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
167 perf_evsel__calc_id_pos(evsel);
170 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
172 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
175 perf_evsel__init(evsel, attr, idx);
180 struct event_format *event_format__new(const char *sys, const char *name)
184 void *bf = NULL, *nbf;
185 size_t size = 0, alloc_size = 0;
186 struct event_format *format = NULL;
188 if (asprintf(&filename, "%s/%s/%s/format", tracing_events_path, sys, name) < 0)
191 fd = open(filename, O_RDONLY);
193 goto out_free_filename;
196 if (size == alloc_size) {
197 alloc_size += BUFSIZ;
198 nbf = realloc(bf, alloc_size);
204 n = read(fd, bf + size, alloc_size - size);
210 pevent_parse_format(&format, bf, size, sys);
221 struct perf_evsel *perf_evsel__newtp(const char *sys, const char *name, int idx)
223 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
226 struct perf_event_attr attr = {
227 .type = PERF_TYPE_TRACEPOINT,
228 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
229 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
232 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
235 evsel->tp_format = event_format__new(sys, name);
236 if (evsel->tp_format == NULL)
239 event_attr_init(&attr);
240 attr.config = evsel->tp_format->id;
241 attr.sample_period = 1;
242 perf_evsel__init(evsel, &attr, idx);
253 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
261 "stalled-cycles-frontend",
262 "stalled-cycles-backend",
266 static const char *__perf_evsel__hw_name(u64 config)
268 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
269 return perf_evsel__hw_names[config];
271 return "unknown-hardware";
274 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
276 int colon = 0, r = 0;
277 struct perf_event_attr *attr = &evsel->attr;
278 bool exclude_guest_default = false;
280 #define MOD_PRINT(context, mod) do { \
281 if (!attr->exclude_##context) { \
282 if (!colon) colon = ++r; \
283 r += scnprintf(bf + r, size - r, "%c", mod); \
286 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
287 MOD_PRINT(kernel, 'k');
288 MOD_PRINT(user, 'u');
290 exclude_guest_default = true;
293 if (attr->precise_ip) {
296 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
297 exclude_guest_default = true;
300 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
301 MOD_PRINT(host, 'H');
302 MOD_PRINT(guest, 'G');
310 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
312 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
313 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
316 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
329 static const char *__perf_evsel__sw_name(u64 config)
331 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
332 return perf_evsel__sw_names[config];
333 return "unknown-software";
336 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
338 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
339 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
342 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
346 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
348 if (type & HW_BREAKPOINT_R)
349 r += scnprintf(bf + r, size - r, "r");
351 if (type & HW_BREAKPOINT_W)
352 r += scnprintf(bf + r, size - r, "w");
354 if (type & HW_BREAKPOINT_X)
355 r += scnprintf(bf + r, size - r, "x");
360 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
362 struct perf_event_attr *attr = &evsel->attr;
363 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
364 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
367 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
368 [PERF_EVSEL__MAX_ALIASES] = {
369 { "L1-dcache", "l1-d", "l1d", "L1-data", },
370 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
372 { "dTLB", "d-tlb", "Data-TLB", },
373 { "iTLB", "i-tlb", "Instruction-TLB", },
374 { "branch", "branches", "bpu", "btb", "bpc", },
378 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
379 [PERF_EVSEL__MAX_ALIASES] = {
380 { "load", "loads", "read", },
381 { "store", "stores", "write", },
382 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
385 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
386 [PERF_EVSEL__MAX_ALIASES] = {
387 { "refs", "Reference", "ops", "access", },
388 { "misses", "miss", },
391 #define C(x) PERF_COUNT_HW_CACHE_##x
392 #define CACHE_READ (1 << C(OP_READ))
393 #define CACHE_WRITE (1 << C(OP_WRITE))
394 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
395 #define COP(x) (1 << x)
398 * cache operartion stat
399 * L1I : Read and prefetch only
400 * ITLB and BPU : Read-only
402 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
403 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
404 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
405 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
406 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
407 [C(ITLB)] = (CACHE_READ),
408 [C(BPU)] = (CACHE_READ),
409 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
412 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
414 if (perf_evsel__hw_cache_stat[type] & COP(op))
415 return true; /* valid */
417 return false; /* invalid */
420 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
421 char *bf, size_t size)
424 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
425 perf_evsel__hw_cache_op[op][0],
426 perf_evsel__hw_cache_result[result][0]);
429 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
430 perf_evsel__hw_cache_op[op][1]);
433 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
435 u8 op, result, type = (config >> 0) & 0xff;
436 const char *err = "unknown-ext-hardware-cache-type";
438 if (type > PERF_COUNT_HW_CACHE_MAX)
441 op = (config >> 8) & 0xff;
442 err = "unknown-ext-hardware-cache-op";
443 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
446 result = (config >> 16) & 0xff;
447 err = "unknown-ext-hardware-cache-result";
448 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
451 err = "invalid-cache";
452 if (!perf_evsel__is_cache_op_valid(type, op))
455 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
457 return scnprintf(bf, size, "%s", err);
460 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
462 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
463 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
466 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
468 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
469 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
472 const char *perf_evsel__name(struct perf_evsel *evsel)
479 switch (evsel->attr.type) {
481 perf_evsel__raw_name(evsel, bf, sizeof(bf));
484 case PERF_TYPE_HARDWARE:
485 perf_evsel__hw_name(evsel, bf, sizeof(bf));
488 case PERF_TYPE_HW_CACHE:
489 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
492 case PERF_TYPE_SOFTWARE:
493 perf_evsel__sw_name(evsel, bf, sizeof(bf));
496 case PERF_TYPE_TRACEPOINT:
497 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
500 case PERF_TYPE_BREAKPOINT:
501 perf_evsel__bp_name(evsel, bf, sizeof(bf));
505 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
510 evsel->name = strdup(bf);
512 return evsel->name ?: "unknown";
515 const char *perf_evsel__group_name(struct perf_evsel *evsel)
517 return evsel->group_name ?: "anon group";
520 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
523 struct perf_evsel *pos;
524 const char *group_name = perf_evsel__group_name(evsel);
526 ret = scnprintf(buf, size, "%s", group_name);
528 ret += scnprintf(buf + ret, size - ret, " { %s",
529 perf_evsel__name(evsel));
531 for_each_group_member(pos, evsel)
532 ret += scnprintf(buf + ret, size - ret, ", %s",
533 perf_evsel__name(pos));
535 ret += scnprintf(buf + ret, size - ret, " }");
541 * The enable_on_exec/disabled value strategy:
543 * 1) For any type of traced program:
544 * - all independent events and group leaders are disabled
545 * - all group members are enabled
547 * Group members are ruled by group leaders. They need to
548 * be enabled, because the group scheduling relies on that.
550 * 2) For traced programs executed by perf:
551 * - all independent events and group leaders have
553 * - we don't specifically enable or disable any event during
556 * Independent events and group leaders are initially disabled
557 * and get enabled by exec. Group members are ruled by group
558 * leaders as stated in 1).
560 * 3) For traced programs attached by perf (pid/tid):
561 * - we specifically enable or disable all events during
564 * When attaching events to already running traced we
565 * enable/disable events specifically, as there's no
566 * initial traced exec call.
568 void perf_evsel__config(struct perf_evsel *evsel,
569 struct perf_record_opts *opts)
571 struct perf_evsel *leader = evsel->leader;
572 struct perf_event_attr *attr = &evsel->attr;
573 int track = !evsel->idx; /* only the first counter needs these */
575 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
576 attr->inherit = !opts->no_inherit;
578 perf_evsel__set_sample_bit(evsel, IP);
579 perf_evsel__set_sample_bit(evsel, TID);
581 if (evsel->sample_read) {
582 perf_evsel__set_sample_bit(evsel, READ);
585 * We need ID even in case of single event, because
586 * PERF_SAMPLE_READ process ID specific data.
588 perf_evsel__set_sample_id(evsel, false);
591 * Apply group format only if we belong to group
592 * with more than one members.
594 if (leader->nr_members > 1) {
595 attr->read_format |= PERF_FORMAT_GROUP;
601 * We default some events to a 1 default interval. But keep
602 * it a weak assumption overridable by the user.
604 if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
605 opts->user_interval != ULLONG_MAX)) {
607 perf_evsel__set_sample_bit(evsel, PERIOD);
609 attr->sample_freq = opts->freq;
611 attr->sample_period = opts->default_interval;
616 * Disable sampling for all group members other
617 * than leader in case leader 'leads' the sampling.
619 if ((leader != evsel) && leader->sample_read) {
620 attr->sample_freq = 0;
621 attr->sample_period = 0;
624 if (opts->no_samples)
625 attr->sample_freq = 0;
627 if (opts->inherit_stat)
628 attr->inherit_stat = 1;
630 if (opts->sample_address) {
631 perf_evsel__set_sample_bit(evsel, ADDR);
632 attr->mmap_data = track;
635 if (opts->call_graph) {
636 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
638 if (opts->call_graph == CALLCHAIN_DWARF) {
639 perf_evsel__set_sample_bit(evsel, REGS_USER);
640 perf_evsel__set_sample_bit(evsel, STACK_USER);
641 attr->sample_regs_user = PERF_REGS_MASK;
642 attr->sample_stack_user = opts->stack_dump_size;
643 attr->exclude_callchain_user = 1;
647 if (perf_target__has_cpu(&opts->target))
648 perf_evsel__set_sample_bit(evsel, CPU);
651 perf_evsel__set_sample_bit(evsel, PERIOD);
653 if (!perf_missing_features.sample_id_all &&
654 (opts->sample_time || !opts->no_inherit ||
655 perf_target__has_cpu(&opts->target)))
656 perf_evsel__set_sample_bit(evsel, TIME);
658 if (opts->raw_samples) {
659 perf_evsel__set_sample_bit(evsel, TIME);
660 perf_evsel__set_sample_bit(evsel, RAW);
661 perf_evsel__set_sample_bit(evsel, CPU);
664 if (opts->sample_address)
665 attr->sample_type |= PERF_SAMPLE_DATA_SRC;
667 if (opts->no_delay) {
669 attr->wakeup_events = 1;
671 if (opts->branch_stack) {
672 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
673 attr->branch_sample_type = opts->branch_stack;
676 if (opts->sample_weight)
677 attr->sample_type |= PERF_SAMPLE_WEIGHT;
683 * XXX see the function comment above
685 * Disabling only independent events or group leaders,
686 * keeping group members enabled.
688 if (perf_evsel__is_group_leader(evsel))
692 * Setting enable_on_exec for independent events and
693 * group leaders for traced executed by perf.
695 if (perf_target__none(&opts->target) && perf_evsel__is_group_leader(evsel))
696 attr->enable_on_exec = 1;
699 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
702 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
705 for (cpu = 0; cpu < ncpus; cpu++) {
706 for (thread = 0; thread < nthreads; thread++) {
707 FD(evsel, cpu, thread) = -1;
712 return evsel->fd != NULL ? 0 : -ENOMEM;
715 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
720 for (cpu = 0; cpu < ncpus; cpu++) {
721 for (thread = 0; thread < nthreads; thread++) {
722 int fd = FD(evsel, cpu, thread),
723 err = ioctl(fd, ioc, arg);
733 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
736 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
737 PERF_EVENT_IOC_SET_FILTER,
741 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
743 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
744 PERF_EVENT_IOC_ENABLE,
748 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
750 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
751 if (evsel->sample_id == NULL)
754 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
755 if (evsel->id == NULL) {
756 xyarray__delete(evsel->sample_id);
757 evsel->sample_id = NULL;
764 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
766 memset(evsel->counts, 0, (sizeof(*evsel->counts) +
767 (ncpus * sizeof(struct perf_counts_values))));
770 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
772 evsel->counts = zalloc((sizeof(*evsel->counts) +
773 (ncpus * sizeof(struct perf_counts_values))));
774 return evsel->counts != NULL ? 0 : -ENOMEM;
777 void perf_evsel__free_fd(struct perf_evsel *evsel)
779 xyarray__delete(evsel->fd);
783 void perf_evsel__free_id(struct perf_evsel *evsel)
785 xyarray__delete(evsel->sample_id);
786 evsel->sample_id = NULL;
791 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
795 for (cpu = 0; cpu < ncpus; cpu++)
796 for (thread = 0; thread < nthreads; ++thread) {
797 close(FD(evsel, cpu, thread));
798 FD(evsel, cpu, thread) = -1;
802 void perf_evsel__free_counts(struct perf_evsel *evsel)
807 void perf_evsel__exit(struct perf_evsel *evsel)
809 assert(list_empty(&evsel->node));
810 perf_evsel__free_fd(evsel);
811 perf_evsel__free_id(evsel);
814 void perf_evsel__delete(struct perf_evsel *evsel)
816 perf_evsel__exit(evsel);
817 close_cgroup(evsel->cgrp);
818 free(evsel->group_name);
819 if (evsel->tp_format)
820 pevent_free_format(evsel->tp_format);
825 static inline void compute_deltas(struct perf_evsel *evsel,
827 struct perf_counts_values *count)
829 struct perf_counts_values tmp;
831 if (!evsel->prev_raw_counts)
835 tmp = evsel->prev_raw_counts->aggr;
836 evsel->prev_raw_counts->aggr = *count;
838 tmp = evsel->prev_raw_counts->cpu[cpu];
839 evsel->prev_raw_counts->cpu[cpu] = *count;
842 count->val = count->val - tmp.val;
843 count->ena = count->ena - tmp.ena;
844 count->run = count->run - tmp.run;
847 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
848 int cpu, int thread, bool scale)
850 struct perf_counts_values count;
851 size_t nv = scale ? 3 : 1;
853 if (FD(evsel, cpu, thread) < 0)
856 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
859 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
862 compute_deltas(evsel, cpu, &count);
867 else if (count.run < count.ena)
868 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
870 count.ena = count.run = 0;
872 evsel->counts->cpu[cpu] = count;
876 int __perf_evsel__read(struct perf_evsel *evsel,
877 int ncpus, int nthreads, bool scale)
879 size_t nv = scale ? 3 : 1;
881 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
883 aggr->val = aggr->ena = aggr->run = 0;
885 for (cpu = 0; cpu < ncpus; cpu++) {
886 for (thread = 0; thread < nthreads; thread++) {
887 if (FD(evsel, cpu, thread) < 0)
890 if (readn(FD(evsel, cpu, thread),
891 &count, nv * sizeof(u64)) < 0)
894 aggr->val += count.val;
896 aggr->ena += count.ena;
897 aggr->run += count.run;
902 compute_deltas(evsel, -1, aggr);
904 evsel->counts->scaled = 0;
906 if (aggr->run == 0) {
907 evsel->counts->scaled = -1;
912 if (aggr->run < aggr->ena) {
913 evsel->counts->scaled = 1;
914 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
917 aggr->ena = aggr->run = 0;
922 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
924 struct perf_evsel *leader = evsel->leader;
927 if (perf_evsel__is_group_leader(evsel))
931 * Leader must be already processed/open,
936 fd = FD(leader, cpu, thread);
942 #define __PRINT_ATTR(fmt, cast, field) \
943 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
945 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
946 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
947 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
948 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
950 #define PRINT_ATTR2N(name1, field1, name2, field2) \
951 fprintf(fp, " %-19s %u %-19s %u\n", \
952 name1, attr->field1, name2, attr->field2)
954 #define PRINT_ATTR2(field1, field2) \
955 PRINT_ATTR2N(#field1, field1, #field2, field2)
957 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
961 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
962 ret += fprintf(fp, "perf_event_attr:\n");
964 ret += PRINT_ATTR_U32(type);
965 ret += PRINT_ATTR_U32(size);
966 ret += PRINT_ATTR_X64(config);
967 ret += PRINT_ATTR_U64(sample_period);
968 ret += PRINT_ATTR_U64(sample_freq);
969 ret += PRINT_ATTR_X64(sample_type);
970 ret += PRINT_ATTR_X64(read_format);
972 ret += PRINT_ATTR2(disabled, inherit);
973 ret += PRINT_ATTR2(pinned, exclusive);
974 ret += PRINT_ATTR2(exclude_user, exclude_kernel);
975 ret += PRINT_ATTR2(exclude_hv, exclude_idle);
976 ret += PRINT_ATTR2(mmap, comm);
977 ret += PRINT_ATTR2(freq, inherit_stat);
978 ret += PRINT_ATTR2(enable_on_exec, task);
979 ret += PRINT_ATTR2(watermark, precise_ip);
980 ret += PRINT_ATTR2(mmap_data, sample_id_all);
981 ret += PRINT_ATTR2(exclude_host, exclude_guest);
982 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
983 "excl.callchain_user", exclude_callchain_user);
985 ret += PRINT_ATTR_U32(wakeup_events);
986 ret += PRINT_ATTR_U32(wakeup_watermark);
987 ret += PRINT_ATTR_X32(bp_type);
988 ret += PRINT_ATTR_X64(bp_addr);
989 ret += PRINT_ATTR_X64(config1);
990 ret += PRINT_ATTR_U64(bp_len);
991 ret += PRINT_ATTR_X64(config2);
992 ret += PRINT_ATTR_X64(branch_sample_type);
993 ret += PRINT_ATTR_X64(sample_regs_user);
994 ret += PRINT_ATTR_U32(sample_stack_user);
996 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1001 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1002 struct thread_map *threads)
1005 unsigned long flags = 0;
1007 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1009 if (evsel->fd == NULL &&
1010 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
1014 flags = PERF_FLAG_PID_CGROUP;
1015 pid = evsel->cgrp->fd;
1018 fallback_missing_features:
1019 if (perf_missing_features.exclude_guest)
1020 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1022 if (perf_missing_features.sample_id_all)
1023 evsel->attr.sample_id_all = 0;
1026 perf_event_attr__fprintf(&evsel->attr, stderr);
1028 for (cpu = 0; cpu < cpus->nr; cpu++) {
1030 for (thread = 0; thread < threads->nr; thread++) {
1034 pid = threads->map[thread];
1036 group_fd = get_group_fd(evsel, cpu, thread);
1038 pr_debug2("perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1039 pid, cpus->map[cpu], group_fd, flags);
1041 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1045 if (FD(evsel, cpu, thread) < 0) {
1049 set_rlimit = NO_CHANGE;
1057 * perf stat needs between 5 and 22 fds per CPU. When we run out
1058 * of them try to increase the limits.
1060 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1062 int old_errno = errno;
1064 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1065 if (set_rlimit == NO_CHANGE)
1066 l.rlim_cur = l.rlim_max;
1068 l.rlim_cur = l.rlim_max + 1000;
1069 l.rlim_max = l.rlim_cur;
1071 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1080 if (err != -EINVAL || cpu > 0 || thread > 0)
1083 if (!perf_missing_features.exclude_guest &&
1084 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1085 perf_missing_features.exclude_guest = true;
1086 goto fallback_missing_features;
1087 } else if (!perf_missing_features.sample_id_all) {
1088 perf_missing_features.sample_id_all = true;
1089 goto retry_sample_id;
1094 while (--thread >= 0) {
1095 close(FD(evsel, cpu, thread));
1096 FD(evsel, cpu, thread) = -1;
1098 thread = threads->nr;
1099 } while (--cpu >= 0);
1103 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1105 if (evsel->fd == NULL)
1108 perf_evsel__close_fd(evsel, ncpus, nthreads);
1109 perf_evsel__free_fd(evsel);
1122 struct thread_map map;
1124 } empty_thread_map = {
1129 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1130 struct thread_map *threads)
1133 /* Work around old compiler warnings about strict aliasing */
1134 cpus = &empty_cpu_map.map;
1137 if (threads == NULL)
1138 threads = &empty_thread_map.map;
1140 return __perf_evsel__open(evsel, cpus, threads);
1143 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1144 struct cpu_map *cpus)
1146 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1149 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1150 struct thread_map *threads)
1152 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1155 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1156 const union perf_event *event,
1157 struct perf_sample *sample)
1159 u64 type = evsel->attr.sample_type;
1160 const u64 *array = event->sample.array;
1161 bool swapped = evsel->needs_swap;
1164 array += ((event->header.size -
1165 sizeof(event->header)) / sizeof(u64)) - 1;
1167 if (type & PERF_SAMPLE_IDENTIFIER) {
1168 sample->id = *array;
1172 if (type & PERF_SAMPLE_CPU) {
1175 /* undo swap of u64, then swap on individual u32s */
1176 u.val64 = bswap_64(u.val64);
1177 u.val32[0] = bswap_32(u.val32[0]);
1180 sample->cpu = u.val32[0];
1184 if (type & PERF_SAMPLE_STREAM_ID) {
1185 sample->stream_id = *array;
1189 if (type & PERF_SAMPLE_ID) {
1190 sample->id = *array;
1194 if (type & PERF_SAMPLE_TIME) {
1195 sample->time = *array;
1199 if (type & PERF_SAMPLE_TID) {
1202 /* undo swap of u64, then swap on individual u32s */
1203 u.val64 = bswap_64(u.val64);
1204 u.val32[0] = bswap_32(u.val32[0]);
1205 u.val32[1] = bswap_32(u.val32[1]);
1208 sample->pid = u.val32[0];
1209 sample->tid = u.val32[1];
1215 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1218 return size > max_size || offset + size > endp;
1221 #define OVERFLOW_CHECK(offset, size, max_size) \
1223 if (overflow(endp, (max_size), (offset), (size))) \
1227 #define OVERFLOW_CHECK_u64(offset) \
1228 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1230 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1231 struct perf_sample *data)
1233 u64 type = evsel->attr.sample_type;
1234 bool swapped = evsel->needs_swap;
1236 u16 max_size = event->header.size;
1237 const void *endp = (void *)event + max_size;
1241 * used for cross-endian analysis. See git commit 65014ab3
1242 * for why this goofiness is needed.
1246 memset(data, 0, sizeof(*data));
1247 data->cpu = data->pid = data->tid = -1;
1248 data->stream_id = data->id = data->time = -1ULL;
1252 if (event->header.type != PERF_RECORD_SAMPLE) {
1253 if (!evsel->attr.sample_id_all)
1255 return perf_evsel__parse_id_sample(evsel, event, data);
1258 array = event->sample.array;
1261 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1262 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1263 * check the format does not go past the end of the event.
1265 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1269 if (type & PERF_SAMPLE_IDENTIFIER) {
1274 if (type & PERF_SAMPLE_IP) {
1279 if (type & PERF_SAMPLE_TID) {
1282 /* undo swap of u64, then swap on individual u32s */
1283 u.val64 = bswap_64(u.val64);
1284 u.val32[0] = bswap_32(u.val32[0]);
1285 u.val32[1] = bswap_32(u.val32[1]);
1288 data->pid = u.val32[0];
1289 data->tid = u.val32[1];
1293 if (type & PERF_SAMPLE_TIME) {
1294 data->time = *array;
1299 if (type & PERF_SAMPLE_ADDR) {
1300 data->addr = *array;
1304 if (type & PERF_SAMPLE_ID) {
1309 if (type & PERF_SAMPLE_STREAM_ID) {
1310 data->stream_id = *array;
1314 if (type & PERF_SAMPLE_CPU) {
1318 /* undo swap of u64, then swap on individual u32s */
1319 u.val64 = bswap_64(u.val64);
1320 u.val32[0] = bswap_32(u.val32[0]);
1323 data->cpu = u.val32[0];
1327 if (type & PERF_SAMPLE_PERIOD) {
1328 data->period = *array;
1332 if (type & PERF_SAMPLE_READ) {
1333 u64 read_format = evsel->attr.read_format;
1335 OVERFLOW_CHECK_u64(array);
1336 if (read_format & PERF_FORMAT_GROUP)
1337 data->read.group.nr = *array;
1339 data->read.one.value = *array;
1343 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1344 OVERFLOW_CHECK_u64(array);
1345 data->read.time_enabled = *array;
1349 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1350 OVERFLOW_CHECK_u64(array);
1351 data->read.time_running = *array;
1355 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1356 if (read_format & PERF_FORMAT_GROUP) {
1357 const u64 max_group_nr = UINT64_MAX /
1358 sizeof(struct sample_read_value);
1360 if (data->read.group.nr > max_group_nr)
1362 sz = data->read.group.nr *
1363 sizeof(struct sample_read_value);
1364 OVERFLOW_CHECK(array, sz, max_size);
1365 data->read.group.values =
1366 (struct sample_read_value *)array;
1367 array = (void *)array + sz;
1369 OVERFLOW_CHECK_u64(array);
1370 data->read.one.id = *array;
1375 if (type & PERF_SAMPLE_CALLCHAIN) {
1376 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1378 OVERFLOW_CHECK_u64(array);
1379 data->callchain = (struct ip_callchain *)array++;
1380 if (data->callchain->nr > max_callchain_nr)
1382 sz = data->callchain->nr * sizeof(u64);
1383 OVERFLOW_CHECK(array, sz, max_size);
1384 array = (void *)array + sz;
1387 if (type & PERF_SAMPLE_RAW) {
1388 OVERFLOW_CHECK_u64(array);
1390 if (WARN_ONCE(swapped,
1391 "Endianness of raw data not corrected!\n")) {
1392 /* undo swap of u64, then swap on individual u32s */
1393 u.val64 = bswap_64(u.val64);
1394 u.val32[0] = bswap_32(u.val32[0]);
1395 u.val32[1] = bswap_32(u.val32[1]);
1397 data->raw_size = u.val32[0];
1398 array = (void *)array + sizeof(u32);
1400 OVERFLOW_CHECK(array, data->raw_size, max_size);
1401 data->raw_data = (void *)array;
1402 array = (void *)array + data->raw_size;
1405 if (type & PERF_SAMPLE_BRANCH_STACK) {
1406 const u64 max_branch_nr = UINT64_MAX /
1407 sizeof(struct branch_entry);
1409 OVERFLOW_CHECK_u64(array);
1410 data->branch_stack = (struct branch_stack *)array++;
1412 if (data->branch_stack->nr > max_branch_nr)
1414 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1415 OVERFLOW_CHECK(array, sz, max_size);
1416 array = (void *)array + sz;
1419 if (type & PERF_SAMPLE_REGS_USER) {
1420 OVERFLOW_CHECK_u64(array);
1421 data->user_regs.abi = *array;
1424 if (data->user_regs.abi) {
1425 u64 regs_user = evsel->attr.sample_regs_user;
1427 sz = hweight_long(regs_user) * sizeof(u64);
1428 OVERFLOW_CHECK(array, sz, max_size);
1429 data->user_regs.regs = (u64 *)array;
1430 array = (void *)array + sz;
1434 if (type & PERF_SAMPLE_STACK_USER) {
1435 OVERFLOW_CHECK_u64(array);
1438 data->user_stack.offset = ((char *)(array - 1)
1442 data->user_stack.size = 0;
1444 OVERFLOW_CHECK(array, sz, max_size);
1445 data->user_stack.data = (char *)array;
1446 array = (void *)array + sz;
1447 OVERFLOW_CHECK_u64(array);
1448 data->user_stack.size = *array++;
1453 if (type & PERF_SAMPLE_WEIGHT) {
1454 OVERFLOW_CHECK_u64(array);
1455 data->weight = *array;
1459 data->data_src = PERF_MEM_DATA_SRC_NONE;
1460 if (type & PERF_SAMPLE_DATA_SRC) {
1461 OVERFLOW_CHECK_u64(array);
1462 data->data_src = *array;
1469 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1470 u64 sample_regs_user, u64 read_format)
1472 size_t sz, result = sizeof(struct sample_event);
1474 if (type & PERF_SAMPLE_IDENTIFIER)
1475 result += sizeof(u64);
1477 if (type & PERF_SAMPLE_IP)
1478 result += sizeof(u64);
1480 if (type & PERF_SAMPLE_TID)
1481 result += sizeof(u64);
1483 if (type & PERF_SAMPLE_TIME)
1484 result += sizeof(u64);
1486 if (type & PERF_SAMPLE_ADDR)
1487 result += sizeof(u64);
1489 if (type & PERF_SAMPLE_ID)
1490 result += sizeof(u64);
1492 if (type & PERF_SAMPLE_STREAM_ID)
1493 result += sizeof(u64);
1495 if (type & PERF_SAMPLE_CPU)
1496 result += sizeof(u64);
1498 if (type & PERF_SAMPLE_PERIOD)
1499 result += sizeof(u64);
1501 if (type & PERF_SAMPLE_READ) {
1502 result += sizeof(u64);
1503 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1504 result += sizeof(u64);
1505 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1506 result += sizeof(u64);
1507 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1508 if (read_format & PERF_FORMAT_GROUP) {
1509 sz = sample->read.group.nr *
1510 sizeof(struct sample_read_value);
1513 result += sizeof(u64);
1517 if (type & PERF_SAMPLE_CALLCHAIN) {
1518 sz = (sample->callchain->nr + 1) * sizeof(u64);
1522 if (type & PERF_SAMPLE_RAW) {
1523 result += sizeof(u32);
1524 result += sample->raw_size;
1527 if (type & PERF_SAMPLE_BRANCH_STACK) {
1528 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1533 if (type & PERF_SAMPLE_REGS_USER) {
1534 if (sample->user_regs.abi) {
1535 result += sizeof(u64);
1536 sz = hweight_long(sample_regs_user) * sizeof(u64);
1539 result += sizeof(u64);
1543 if (type & PERF_SAMPLE_STACK_USER) {
1544 sz = sample->user_stack.size;
1545 result += sizeof(u64);
1548 result += sizeof(u64);
1552 if (type & PERF_SAMPLE_WEIGHT)
1553 result += sizeof(u64);
1555 if (type & PERF_SAMPLE_DATA_SRC)
1556 result += sizeof(u64);
1561 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1562 u64 sample_regs_user, u64 read_format,
1563 const struct perf_sample *sample,
1569 * used for cross-endian analysis. See git commit 65014ab3
1570 * for why this goofiness is needed.
1574 array = event->sample.array;
1576 if (type & PERF_SAMPLE_IDENTIFIER) {
1577 *array = sample->id;
1581 if (type & PERF_SAMPLE_IP) {
1582 *array = sample->ip;
1586 if (type & PERF_SAMPLE_TID) {
1587 u.val32[0] = sample->pid;
1588 u.val32[1] = sample->tid;
1591 * Inverse of what is done in perf_evsel__parse_sample
1593 u.val32[0] = bswap_32(u.val32[0]);
1594 u.val32[1] = bswap_32(u.val32[1]);
1595 u.val64 = bswap_64(u.val64);
1602 if (type & PERF_SAMPLE_TIME) {
1603 *array = sample->time;
1607 if (type & PERF_SAMPLE_ADDR) {
1608 *array = sample->addr;
1612 if (type & PERF_SAMPLE_ID) {
1613 *array = sample->id;
1617 if (type & PERF_SAMPLE_STREAM_ID) {
1618 *array = sample->stream_id;
1622 if (type & PERF_SAMPLE_CPU) {
1623 u.val32[0] = sample->cpu;
1626 * Inverse of what is done in perf_evsel__parse_sample
1628 u.val32[0] = bswap_32(u.val32[0]);
1629 u.val64 = bswap_64(u.val64);
1635 if (type & PERF_SAMPLE_PERIOD) {
1636 *array = sample->period;
1640 if (type & PERF_SAMPLE_READ) {
1641 if (read_format & PERF_FORMAT_GROUP)
1642 *array = sample->read.group.nr;
1644 *array = sample->read.one.value;
1647 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1648 *array = sample->read.time_enabled;
1652 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1653 *array = sample->read.time_running;
1657 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1658 if (read_format & PERF_FORMAT_GROUP) {
1659 sz = sample->read.group.nr *
1660 sizeof(struct sample_read_value);
1661 memcpy(array, sample->read.group.values, sz);
1662 array = (void *)array + sz;
1664 *array = sample->read.one.id;
1669 if (type & PERF_SAMPLE_CALLCHAIN) {
1670 sz = (sample->callchain->nr + 1) * sizeof(u64);
1671 memcpy(array, sample->callchain, sz);
1672 array = (void *)array + sz;
1675 if (type & PERF_SAMPLE_RAW) {
1676 u.val32[0] = sample->raw_size;
1677 if (WARN_ONCE(swapped,
1678 "Endianness of raw data not corrected!\n")) {
1680 * Inverse of what is done in perf_evsel__parse_sample
1682 u.val32[0] = bswap_32(u.val32[0]);
1683 u.val32[1] = bswap_32(u.val32[1]);
1684 u.val64 = bswap_64(u.val64);
1687 array = (void *)array + sizeof(u32);
1689 memcpy(array, sample->raw_data, sample->raw_size);
1690 array = (void *)array + sample->raw_size;
1693 if (type & PERF_SAMPLE_BRANCH_STACK) {
1694 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1696 memcpy(array, sample->branch_stack, sz);
1697 array = (void *)array + sz;
1700 if (type & PERF_SAMPLE_REGS_USER) {
1701 if (sample->user_regs.abi) {
1702 *array++ = sample->user_regs.abi;
1703 sz = hweight_long(sample_regs_user) * sizeof(u64);
1704 memcpy(array, sample->user_regs.regs, sz);
1705 array = (void *)array + sz;
1711 if (type & PERF_SAMPLE_STACK_USER) {
1712 sz = sample->user_stack.size;
1715 memcpy(array, sample->user_stack.data, sz);
1716 array = (void *)array + sz;
1721 if (type & PERF_SAMPLE_WEIGHT) {
1722 *array = sample->weight;
1726 if (type & PERF_SAMPLE_DATA_SRC) {
1727 *array = sample->data_src;
1734 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1736 return pevent_find_field(evsel->tp_format, name);
1739 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1742 struct format_field *field = perf_evsel__field(evsel, name);
1748 offset = field->offset;
1750 if (field->flags & FIELD_IS_DYNAMIC) {
1751 offset = *(int *)(sample->raw_data + field->offset);
1755 return sample->raw_data + offset;
1758 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1761 struct format_field *field = perf_evsel__field(evsel, name);
1768 ptr = sample->raw_data + field->offset;
1770 switch (field->size) {
1774 value = *(u16 *)ptr;
1777 value = *(u32 *)ptr;
1780 value = *(u64 *)ptr;
1786 if (!evsel->needs_swap)
1789 switch (field->size) {
1791 return bswap_16(value);
1793 return bswap_32(value);
1795 return bswap_64(value);
1803 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1809 ret += fprintf(fp, ",");
1811 ret += fprintf(fp, ":");
1815 va_start(args, fmt);
1816 ret += vfprintf(fp, fmt, args);
1821 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1826 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1829 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1836 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1837 struct bit_names *bits, bool *first)
1839 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1840 bool first_bit = true;
1843 if (value & bits[i].bit) {
1844 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1847 } while (bits[++i].name != NULL);
1852 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1854 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1855 struct bit_names bits[] = {
1856 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1857 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1858 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1859 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1860 bit_name(IDENTIFIER),
1864 return bits__fprintf(fp, "sample_type", value, bits, first);
1867 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1869 #define bit_name(n) { PERF_FORMAT_##n, #n }
1870 struct bit_names bits[] = {
1871 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1872 bit_name(ID), bit_name(GROUP),
1876 return bits__fprintf(fp, "read_format", value, bits, first);
1879 int perf_evsel__fprintf(struct perf_evsel *evsel,
1880 struct perf_attr_details *details, FILE *fp)
1885 if (details->event_group) {
1886 struct perf_evsel *pos;
1888 if (!perf_evsel__is_group_leader(evsel))
1891 if (evsel->nr_members > 1)
1892 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1894 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1895 for_each_group_member(pos, evsel)
1896 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1898 if (evsel->nr_members > 1)
1899 printed += fprintf(fp, "}");
1903 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1905 if (details->verbose || details->freq) {
1906 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1907 (u64)evsel->attr.sample_freq);
1910 if (details->verbose) {
1916 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1917 if (evsel->attr.read_format)
1918 printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1922 if_print(exclusive);
1923 if_print(exclude_user);
1924 if_print(exclude_kernel);
1925 if_print(exclude_hv);
1926 if_print(exclude_idle);
1930 if_print(inherit_stat);
1931 if_print(enable_on_exec);
1933 if_print(watermark);
1934 if_print(precise_ip);
1935 if_print(mmap_data);
1936 if_print(sample_id_all);
1937 if_print(exclude_host);
1938 if_print(exclude_guest);
1939 if_print(__reserved_1);
1940 if_print(wakeup_events);
1942 if_print(branch_sample_type);
1949 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1950 char *msg, size_t msgsize)
1952 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
1953 evsel->attr.type == PERF_TYPE_HARDWARE &&
1954 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1956 * If it's cycles then fall back to hrtimer based
1957 * cpu-clock-tick sw counter, which is always available even if
1960 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1963 scnprintf(msg, msgsize, "%s",
1964 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1966 evsel->attr.type = PERF_TYPE_SOFTWARE;
1967 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
1977 int perf_evsel__open_strerror(struct perf_evsel *evsel,
1978 struct perf_target *target,
1979 int err, char *msg, size_t size)
1984 return scnprintf(msg, size,
1985 "You may not have permission to collect %sstats.\n"
1986 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
1987 " -1 - Not paranoid at all\n"
1988 " 0 - Disallow raw tracepoint access for unpriv\n"
1989 " 1 - Disallow cpu events for unpriv\n"
1990 " 2 - Disallow kernel profiling for unpriv",
1991 target->system_wide ? "system-wide " : "");
1993 return scnprintf(msg, size, "The %s event is not supported.",
1994 perf_evsel__name(evsel));
1996 return scnprintf(msg, size, "%s",
1997 "Too many events are opened.\n"
1998 "Try again after reducing the number of events.");
2000 if (target->cpu_list)
2001 return scnprintf(msg, size, "%s",
2002 "No such device - did you specify an out-of-range profile CPU?\n");
2005 if (evsel->attr.precise_ip)
2006 return scnprintf(msg, size, "%s",
2007 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2008 #if defined(__i386__) || defined(__x86_64__)
2009 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2010 return scnprintf(msg, size, "%s",
2011 "No hardware sampling interrupt available.\n"
2012 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2019 return scnprintf(msg, size,
2020 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2021 "/bin/dmesg may provide additional information.\n"
2022 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2023 err, strerror(err), perf_evsel__name(evsel));