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"
35 } perf_missing_features;
37 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
42 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
48 int (*init)(struct perf_evsel *evsel);
49 void (*fini)(struct perf_evsel *evsel);
50 } perf_evsel__object = {
51 .size = sizeof(struct perf_evsel),
52 .init = perf_evsel__no_extra_init,
53 .fini = perf_evsel__no_extra_fini,
56 int perf_evsel__object_config(size_t object_size,
57 int (*init)(struct perf_evsel *evsel),
58 void (*fini)(struct perf_evsel *evsel))
64 if (perf_evsel__object.size > object_size)
67 perf_evsel__object.size = object_size;
71 perf_evsel__object.init = init;
74 perf_evsel__object.fini = fini;
79 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
81 int __perf_evsel__sample_size(u64 sample_type)
83 u64 mask = sample_type & PERF_SAMPLE_MASK;
87 for (i = 0; i < 64; i++) {
88 if (mask & (1ULL << i))
98 * __perf_evsel__calc_id_pos - calculate id_pos.
99 * @sample_type: sample type
101 * This function returns the position of the event id (PERF_SAMPLE_ID or
102 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
105 static int __perf_evsel__calc_id_pos(u64 sample_type)
109 if (sample_type & PERF_SAMPLE_IDENTIFIER)
112 if (!(sample_type & PERF_SAMPLE_ID))
115 if (sample_type & PERF_SAMPLE_IP)
118 if (sample_type & PERF_SAMPLE_TID)
121 if (sample_type & PERF_SAMPLE_TIME)
124 if (sample_type & PERF_SAMPLE_ADDR)
131 * __perf_evsel__calc_is_pos - calculate is_pos.
132 * @sample_type: sample type
134 * This function returns the position (counting backwards) of the event id
135 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
136 * sample_id_all is used there is an id sample appended to non-sample events.
138 static int __perf_evsel__calc_is_pos(u64 sample_type)
142 if (sample_type & PERF_SAMPLE_IDENTIFIER)
145 if (!(sample_type & PERF_SAMPLE_ID))
148 if (sample_type & PERF_SAMPLE_CPU)
151 if (sample_type & PERF_SAMPLE_STREAM_ID)
157 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
159 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
160 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
163 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
164 enum perf_event_sample_format bit)
166 if (!(evsel->attr.sample_type & bit)) {
167 evsel->attr.sample_type |= bit;
168 evsel->sample_size += sizeof(u64);
169 perf_evsel__calc_id_pos(evsel);
173 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
174 enum perf_event_sample_format bit)
176 if (evsel->attr.sample_type & bit) {
177 evsel->attr.sample_type &= ~bit;
178 evsel->sample_size -= sizeof(u64);
179 perf_evsel__calc_id_pos(evsel);
183 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
184 bool can_sample_identifier)
186 if (can_sample_identifier) {
187 perf_evsel__reset_sample_bit(evsel, ID);
188 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
190 perf_evsel__set_sample_bit(evsel, ID);
192 evsel->attr.read_format |= PERF_FORMAT_ID;
195 void perf_evsel__init(struct perf_evsel *evsel,
196 struct perf_event_attr *attr, int idx)
199 evsel->tracking = !idx;
201 evsel->leader = evsel;
204 INIT_LIST_HEAD(&evsel->node);
205 perf_evsel__object.init(evsel);
206 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
207 perf_evsel__calc_id_pos(evsel);
210 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
212 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
215 perf_evsel__init(evsel, attr, idx);
220 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
222 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
225 struct perf_event_attr attr = {
226 .type = PERF_TYPE_TRACEPOINT,
227 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
228 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
231 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
234 evsel->tp_format = trace_event__tp_format(sys, name);
235 if (evsel->tp_format == NULL)
238 event_attr_init(&attr);
239 attr.config = evsel->tp_format->id;
240 attr.sample_period = 1;
241 perf_evsel__init(evsel, &attr, idx);
252 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
260 "stalled-cycles-frontend",
261 "stalled-cycles-backend",
265 static const char *__perf_evsel__hw_name(u64 config)
267 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
268 return perf_evsel__hw_names[config];
270 return "unknown-hardware";
273 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
275 int colon = 0, r = 0;
276 struct perf_event_attr *attr = &evsel->attr;
277 bool exclude_guest_default = false;
279 #define MOD_PRINT(context, mod) do { \
280 if (!attr->exclude_##context) { \
281 if (!colon) colon = ++r; \
282 r += scnprintf(bf + r, size - r, "%c", mod); \
285 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
286 MOD_PRINT(kernel, 'k');
287 MOD_PRINT(user, 'u');
289 exclude_guest_default = true;
292 if (attr->precise_ip) {
295 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
296 exclude_guest_default = true;
299 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
300 MOD_PRINT(host, 'H');
301 MOD_PRINT(guest, 'G');
309 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
311 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
312 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
315 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
328 static const char *__perf_evsel__sw_name(u64 config)
330 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
331 return perf_evsel__sw_names[config];
332 return "unknown-software";
335 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
337 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
338 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
341 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
345 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
347 if (type & HW_BREAKPOINT_R)
348 r += scnprintf(bf + r, size - r, "r");
350 if (type & HW_BREAKPOINT_W)
351 r += scnprintf(bf + r, size - r, "w");
353 if (type & HW_BREAKPOINT_X)
354 r += scnprintf(bf + r, size - r, "x");
359 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
361 struct perf_event_attr *attr = &evsel->attr;
362 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
363 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
366 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
367 [PERF_EVSEL__MAX_ALIASES] = {
368 { "L1-dcache", "l1-d", "l1d", "L1-data", },
369 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
371 { "dTLB", "d-tlb", "Data-TLB", },
372 { "iTLB", "i-tlb", "Instruction-TLB", },
373 { "branch", "branches", "bpu", "btb", "bpc", },
377 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
378 [PERF_EVSEL__MAX_ALIASES] = {
379 { "load", "loads", "read", },
380 { "store", "stores", "write", },
381 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
384 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
385 [PERF_EVSEL__MAX_ALIASES] = {
386 { "refs", "Reference", "ops", "access", },
387 { "misses", "miss", },
390 #define C(x) PERF_COUNT_HW_CACHE_##x
391 #define CACHE_READ (1 << C(OP_READ))
392 #define CACHE_WRITE (1 << C(OP_WRITE))
393 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
394 #define COP(x) (1 << x)
397 * cache operartion stat
398 * L1I : Read and prefetch only
399 * ITLB and BPU : Read-only
401 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
402 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
403 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
404 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
405 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
406 [C(ITLB)] = (CACHE_READ),
407 [C(BPU)] = (CACHE_READ),
408 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
411 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
413 if (perf_evsel__hw_cache_stat[type] & COP(op))
414 return true; /* valid */
416 return false; /* invalid */
419 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
420 char *bf, size_t size)
423 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
424 perf_evsel__hw_cache_op[op][0],
425 perf_evsel__hw_cache_result[result][0]);
428 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
429 perf_evsel__hw_cache_op[op][1]);
432 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
434 u8 op, result, type = (config >> 0) & 0xff;
435 const char *err = "unknown-ext-hardware-cache-type";
437 if (type > PERF_COUNT_HW_CACHE_MAX)
440 op = (config >> 8) & 0xff;
441 err = "unknown-ext-hardware-cache-op";
442 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
445 result = (config >> 16) & 0xff;
446 err = "unknown-ext-hardware-cache-result";
447 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
450 err = "invalid-cache";
451 if (!perf_evsel__is_cache_op_valid(type, op))
454 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
456 return scnprintf(bf, size, "%s", err);
459 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
461 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
462 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
465 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
467 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
468 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
471 const char *perf_evsel__name(struct perf_evsel *evsel)
478 switch (evsel->attr.type) {
480 perf_evsel__raw_name(evsel, bf, sizeof(bf));
483 case PERF_TYPE_HARDWARE:
484 perf_evsel__hw_name(evsel, bf, sizeof(bf));
487 case PERF_TYPE_HW_CACHE:
488 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
491 case PERF_TYPE_SOFTWARE:
492 perf_evsel__sw_name(evsel, bf, sizeof(bf));
495 case PERF_TYPE_TRACEPOINT:
496 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
499 case PERF_TYPE_BREAKPOINT:
500 perf_evsel__bp_name(evsel, bf, sizeof(bf));
504 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
509 evsel->name = strdup(bf);
511 return evsel->name ?: "unknown";
514 const char *perf_evsel__group_name(struct perf_evsel *evsel)
516 return evsel->group_name ?: "anon group";
519 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
522 struct perf_evsel *pos;
523 const char *group_name = perf_evsel__group_name(evsel);
525 ret = scnprintf(buf, size, "%s", group_name);
527 ret += scnprintf(buf + ret, size - ret, " { %s",
528 perf_evsel__name(evsel));
530 for_each_group_member(pos, evsel)
531 ret += scnprintf(buf + ret, size - ret, ", %s",
532 perf_evsel__name(pos));
534 ret += scnprintf(buf + ret, size - ret, " }");
540 perf_evsel__config_callgraph(struct perf_evsel *evsel)
542 bool function = perf_evsel__is_function_event(evsel);
543 struct perf_event_attr *attr = &evsel->attr;
545 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
547 if (callchain_param.record_mode == CALLCHAIN_DWARF) {
549 perf_evsel__set_sample_bit(evsel, REGS_USER);
550 perf_evsel__set_sample_bit(evsel, STACK_USER);
551 attr->sample_regs_user = PERF_REGS_MASK;
552 attr->sample_stack_user = callchain_param.dump_size;
553 attr->exclude_callchain_user = 1;
555 pr_info("Cannot use DWARF unwind for function trace event,"
556 " falling back to framepointers.\n");
561 pr_info("Disabling user space callchains for function trace event.\n");
562 attr->exclude_callchain_user = 1;
567 * The enable_on_exec/disabled value strategy:
569 * 1) For any type of traced program:
570 * - all independent events and group leaders are disabled
571 * - all group members are enabled
573 * Group members are ruled by group leaders. They need to
574 * be enabled, because the group scheduling relies on that.
576 * 2) For traced programs executed by perf:
577 * - all independent events and group leaders have
579 * - we don't specifically enable or disable any event during
582 * Independent events and group leaders are initially disabled
583 * and get enabled by exec. Group members are ruled by group
584 * leaders as stated in 1).
586 * 3) For traced programs attached by perf (pid/tid):
587 * - we specifically enable or disable all events during
590 * When attaching events to already running traced we
591 * enable/disable events specifically, as there's no
592 * initial traced exec call.
594 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
596 struct perf_evsel *leader = evsel->leader;
597 struct perf_event_attr *attr = &evsel->attr;
598 int track = evsel->tracking;
599 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
601 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
602 attr->inherit = !opts->no_inherit;
604 perf_evsel__set_sample_bit(evsel, IP);
605 perf_evsel__set_sample_bit(evsel, TID);
607 if (evsel->sample_read) {
608 perf_evsel__set_sample_bit(evsel, READ);
611 * We need ID even in case of single event, because
612 * PERF_SAMPLE_READ process ID specific data.
614 perf_evsel__set_sample_id(evsel, false);
617 * Apply group format only if we belong to group
618 * with more than one members.
620 if (leader->nr_members > 1) {
621 attr->read_format |= PERF_FORMAT_GROUP;
627 * We default some events to have a default interval. But keep
628 * it a weak assumption overridable by the user.
630 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
631 opts->user_interval != ULLONG_MAX)) {
633 perf_evsel__set_sample_bit(evsel, PERIOD);
635 attr->sample_freq = opts->freq;
637 attr->sample_period = opts->default_interval;
642 * Disable sampling for all group members other
643 * than leader in case leader 'leads' the sampling.
645 if ((leader != evsel) && leader->sample_read) {
646 attr->sample_freq = 0;
647 attr->sample_period = 0;
650 if (opts->no_samples)
651 attr->sample_freq = 0;
653 if (opts->inherit_stat)
654 attr->inherit_stat = 1;
656 if (opts->sample_address) {
657 perf_evsel__set_sample_bit(evsel, ADDR);
658 attr->mmap_data = track;
662 * We don't allow user space callchains for function trace
663 * event, due to issues with page faults while tracing page
664 * fault handler and its overall trickiness nature.
666 if (perf_evsel__is_function_event(evsel))
667 evsel->attr.exclude_callchain_user = 1;
669 if (callchain_param.enabled && !evsel->no_aux_samples)
670 perf_evsel__config_callgraph(evsel);
672 if (opts->sample_intr_regs) {
673 attr->sample_regs_intr = PERF_REGS_MASK;
674 perf_evsel__set_sample_bit(evsel, REGS_INTR);
677 if (target__has_cpu(&opts->target))
678 perf_evsel__set_sample_bit(evsel, CPU);
681 perf_evsel__set_sample_bit(evsel, PERIOD);
684 * When the user explicitely disabled time don't force it here.
686 if (opts->sample_time &&
687 (!perf_missing_features.sample_id_all &&
688 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu)))
689 perf_evsel__set_sample_bit(evsel, TIME);
691 if (opts->raw_samples && !evsel->no_aux_samples) {
692 perf_evsel__set_sample_bit(evsel, TIME);
693 perf_evsel__set_sample_bit(evsel, RAW);
694 perf_evsel__set_sample_bit(evsel, CPU);
697 if (opts->sample_address)
698 perf_evsel__set_sample_bit(evsel, DATA_SRC);
700 if (opts->no_buffering) {
702 attr->wakeup_events = 1;
704 if (opts->branch_stack && !evsel->no_aux_samples) {
705 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
706 attr->branch_sample_type = opts->branch_stack;
709 if (opts->sample_weight)
710 perf_evsel__set_sample_bit(evsel, WEIGHT);
713 attr->mmap2 = track && !perf_missing_features.mmap2;
716 if (opts->sample_transaction)
717 perf_evsel__set_sample_bit(evsel, TRANSACTION);
720 * XXX see the function comment above
722 * Disabling only independent events or group leaders,
723 * keeping group members enabled.
725 if (perf_evsel__is_group_leader(evsel))
729 * Setting enable_on_exec for independent events and
730 * group leaders for traced executed by perf.
732 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
733 !opts->initial_delay)
734 attr->enable_on_exec = 1;
736 if (evsel->immediate) {
738 attr->enable_on_exec = 0;
742 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
746 if (evsel->system_wide)
749 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
752 for (cpu = 0; cpu < ncpus; cpu++) {
753 for (thread = 0; thread < nthreads; thread++) {
754 FD(evsel, cpu, thread) = -1;
759 return evsel->fd != NULL ? 0 : -ENOMEM;
762 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
767 if (evsel->system_wide)
770 for (cpu = 0; cpu < ncpus; cpu++) {
771 for (thread = 0; thread < nthreads; thread++) {
772 int fd = FD(evsel, cpu, thread),
773 err = ioctl(fd, ioc, arg);
783 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
786 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
787 PERF_EVENT_IOC_SET_FILTER,
791 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
793 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
794 PERF_EVENT_IOC_ENABLE,
798 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
800 if (evsel->system_wide)
803 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
804 if (evsel->sample_id == NULL)
807 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
808 if (evsel->id == NULL) {
809 xyarray__delete(evsel->sample_id);
810 evsel->sample_id = NULL;
817 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
819 memset(evsel->counts, 0, (sizeof(*evsel->counts) +
820 (ncpus * sizeof(struct perf_counts_values))));
823 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
825 evsel->counts = zalloc((sizeof(*evsel->counts) +
826 (ncpus * sizeof(struct perf_counts_values))));
827 return evsel->counts != NULL ? 0 : -ENOMEM;
830 static void perf_evsel__free_fd(struct perf_evsel *evsel)
832 xyarray__delete(evsel->fd);
836 static void perf_evsel__free_id(struct perf_evsel *evsel)
838 xyarray__delete(evsel->sample_id);
839 evsel->sample_id = NULL;
843 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
847 if (evsel->system_wide)
850 for (cpu = 0; cpu < ncpus; cpu++)
851 for (thread = 0; thread < nthreads; ++thread) {
852 close(FD(evsel, cpu, thread));
853 FD(evsel, cpu, thread) = -1;
857 void perf_evsel__free_counts(struct perf_evsel *evsel)
859 zfree(&evsel->counts);
862 void perf_evsel__exit(struct perf_evsel *evsel)
864 assert(list_empty(&evsel->node));
865 perf_evsel__free_fd(evsel);
866 perf_evsel__free_id(evsel);
867 close_cgroup(evsel->cgrp);
868 zfree(&evsel->group_name);
870 perf_evsel__object.fini(evsel);
873 void perf_evsel__delete(struct perf_evsel *evsel)
875 perf_evsel__exit(evsel);
879 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu,
880 struct perf_counts_values *count)
882 struct perf_counts_values tmp;
884 if (!evsel->prev_raw_counts)
888 tmp = evsel->prev_raw_counts->aggr;
889 evsel->prev_raw_counts->aggr = *count;
891 tmp = evsel->prev_raw_counts->cpu[cpu];
892 evsel->prev_raw_counts->cpu[cpu] = *count;
895 count->val = count->val - tmp.val;
896 count->ena = count->ena - tmp.ena;
897 count->run = count->run - tmp.run;
900 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
901 int cpu, int thread, bool scale)
903 struct perf_counts_values count;
904 size_t nv = scale ? 3 : 1;
906 if (FD(evsel, cpu, thread) < 0)
909 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
912 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
915 perf_evsel__compute_deltas(evsel, cpu, &count);
920 else if (count.run < count.ena)
921 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
923 count.ena = count.run = 0;
925 evsel->counts->cpu[cpu] = count;
929 int __perf_evsel__read(struct perf_evsel *evsel,
930 int ncpus, int nthreads, bool scale)
932 size_t nv = scale ? 3 : 1;
934 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
936 if (evsel->system_wide)
939 aggr->val = aggr->ena = aggr->run = 0;
941 for (cpu = 0; cpu < ncpus; cpu++) {
942 for (thread = 0; thread < nthreads; thread++) {
943 if (FD(evsel, cpu, thread) < 0)
946 if (readn(FD(evsel, cpu, thread),
947 &count, nv * sizeof(u64)) < 0)
950 aggr->val += count.val;
952 aggr->ena += count.ena;
953 aggr->run += count.run;
958 perf_evsel__compute_deltas(evsel, -1, aggr);
960 evsel->counts->scaled = 0;
962 if (aggr->run == 0) {
963 evsel->counts->scaled = -1;
968 if (aggr->run < aggr->ena) {
969 evsel->counts->scaled = 1;
970 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
973 aggr->ena = aggr->run = 0;
978 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
980 struct perf_evsel *leader = evsel->leader;
983 if (perf_evsel__is_group_leader(evsel))
987 * Leader must be already processed/open,
992 fd = FD(leader, cpu, thread);
998 #define __PRINT_ATTR(fmt, cast, field) \
999 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
1001 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
1002 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
1003 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
1004 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
1006 #define PRINT_ATTR2N(name1, field1, name2, field2) \
1007 fprintf(fp, " %-19s %u %-19s %u\n", \
1008 name1, attr->field1, name2, attr->field2)
1010 #define PRINT_ATTR2(field1, field2) \
1011 PRINT_ATTR2N(#field1, field1, #field2, field2)
1013 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
1017 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1018 ret += fprintf(fp, "perf_event_attr:\n");
1020 ret += PRINT_ATTR_U32(type);
1021 ret += PRINT_ATTR_U32(size);
1022 ret += PRINT_ATTR_X64(config);
1023 ret += PRINT_ATTR_U64(sample_period);
1024 ret += PRINT_ATTR_U64(sample_freq);
1025 ret += PRINT_ATTR_X64(sample_type);
1026 ret += PRINT_ATTR_X64(read_format);
1028 ret += PRINT_ATTR2(disabled, inherit);
1029 ret += PRINT_ATTR2(pinned, exclusive);
1030 ret += PRINT_ATTR2(exclude_user, exclude_kernel);
1031 ret += PRINT_ATTR2(exclude_hv, exclude_idle);
1032 ret += PRINT_ATTR2(mmap, comm);
1033 ret += PRINT_ATTR2(mmap2, comm_exec);
1034 ret += PRINT_ATTR2(freq, inherit_stat);
1035 ret += PRINT_ATTR2(enable_on_exec, task);
1036 ret += PRINT_ATTR2(watermark, precise_ip);
1037 ret += PRINT_ATTR2(mmap_data, sample_id_all);
1038 ret += PRINT_ATTR2(exclude_host, exclude_guest);
1039 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
1040 "excl.callchain_user", exclude_callchain_user);
1042 ret += PRINT_ATTR_U32(wakeup_events);
1043 ret += PRINT_ATTR_U32(wakeup_watermark);
1044 ret += PRINT_ATTR_X32(bp_type);
1045 ret += PRINT_ATTR_X64(bp_addr);
1046 ret += PRINT_ATTR_X64(config1);
1047 ret += PRINT_ATTR_U64(bp_len);
1048 ret += PRINT_ATTR_X64(config2);
1049 ret += PRINT_ATTR_X64(branch_sample_type);
1050 ret += PRINT_ATTR_X64(sample_regs_user);
1051 ret += PRINT_ATTR_U32(sample_stack_user);
1052 ret += PRINT_ATTR_X64(sample_regs_intr);
1054 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1059 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1060 struct thread_map *threads)
1062 int cpu, thread, nthreads;
1063 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1065 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1067 if (evsel->system_wide)
1070 nthreads = threads->nr;
1072 if (evsel->fd == NULL &&
1073 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1077 flags |= PERF_FLAG_PID_CGROUP;
1078 pid = evsel->cgrp->fd;
1081 fallback_missing_features:
1082 if (perf_missing_features.cloexec)
1083 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1084 if (perf_missing_features.mmap2)
1085 evsel->attr.mmap2 = 0;
1086 if (perf_missing_features.exclude_guest)
1087 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1089 if (perf_missing_features.sample_id_all)
1090 evsel->attr.sample_id_all = 0;
1093 perf_event_attr__fprintf(&evsel->attr, stderr);
1095 for (cpu = 0; cpu < cpus->nr; cpu++) {
1097 for (thread = 0; thread < nthreads; thread++) {
1100 if (!evsel->cgrp && !evsel->system_wide)
1101 pid = threads->map[thread];
1103 group_fd = get_group_fd(evsel, cpu, thread);
1105 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1106 pid, cpus->map[cpu], group_fd, flags);
1108 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1112 if (FD(evsel, cpu, thread) < 0) {
1114 pr_debug2("sys_perf_event_open failed, error %d\n",
1118 set_rlimit = NO_CHANGE;
1126 * perf stat needs between 5 and 22 fds per CPU. When we run out
1127 * of them try to increase the limits.
1129 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1131 int old_errno = errno;
1133 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1134 if (set_rlimit == NO_CHANGE)
1135 l.rlim_cur = l.rlim_max;
1137 l.rlim_cur = l.rlim_max + 1000;
1138 l.rlim_max = l.rlim_cur;
1140 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1149 if (err != -EINVAL || cpu > 0 || thread > 0)
1152 if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1153 perf_missing_features.cloexec = true;
1154 goto fallback_missing_features;
1155 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1156 perf_missing_features.mmap2 = true;
1157 goto fallback_missing_features;
1158 } else if (!perf_missing_features.exclude_guest &&
1159 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1160 perf_missing_features.exclude_guest = true;
1161 goto fallback_missing_features;
1162 } else if (!perf_missing_features.sample_id_all) {
1163 perf_missing_features.sample_id_all = true;
1164 goto retry_sample_id;
1169 while (--thread >= 0) {
1170 close(FD(evsel, cpu, thread));
1171 FD(evsel, cpu, thread) = -1;
1174 } while (--cpu >= 0);
1178 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1180 if (evsel->fd == NULL)
1183 perf_evsel__close_fd(evsel, ncpus, nthreads);
1184 perf_evsel__free_fd(evsel);
1196 struct thread_map map;
1198 } empty_thread_map = {
1203 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1204 struct thread_map *threads)
1207 /* Work around old compiler warnings about strict aliasing */
1208 cpus = &empty_cpu_map.map;
1211 if (threads == NULL)
1212 threads = &empty_thread_map.map;
1214 return __perf_evsel__open(evsel, cpus, threads);
1217 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1218 struct cpu_map *cpus)
1220 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1223 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1224 struct thread_map *threads)
1226 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1229 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1230 const union perf_event *event,
1231 struct perf_sample *sample)
1233 u64 type = evsel->attr.sample_type;
1234 const u64 *array = event->sample.array;
1235 bool swapped = evsel->needs_swap;
1238 array += ((event->header.size -
1239 sizeof(event->header)) / sizeof(u64)) - 1;
1241 if (type & PERF_SAMPLE_IDENTIFIER) {
1242 sample->id = *array;
1246 if (type & PERF_SAMPLE_CPU) {
1249 /* undo swap of u64, then swap on individual u32s */
1250 u.val64 = bswap_64(u.val64);
1251 u.val32[0] = bswap_32(u.val32[0]);
1254 sample->cpu = u.val32[0];
1258 if (type & PERF_SAMPLE_STREAM_ID) {
1259 sample->stream_id = *array;
1263 if (type & PERF_SAMPLE_ID) {
1264 sample->id = *array;
1268 if (type & PERF_SAMPLE_TIME) {
1269 sample->time = *array;
1273 if (type & PERF_SAMPLE_TID) {
1276 /* undo swap of u64, then swap on individual u32s */
1277 u.val64 = bswap_64(u.val64);
1278 u.val32[0] = bswap_32(u.val32[0]);
1279 u.val32[1] = bswap_32(u.val32[1]);
1282 sample->pid = u.val32[0];
1283 sample->tid = u.val32[1];
1290 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1293 return size > max_size || offset + size > endp;
1296 #define OVERFLOW_CHECK(offset, size, max_size) \
1298 if (overflow(endp, (max_size), (offset), (size))) \
1302 #define OVERFLOW_CHECK_u64(offset) \
1303 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1305 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1306 struct perf_sample *data)
1308 u64 type = evsel->attr.sample_type;
1309 bool swapped = evsel->needs_swap;
1311 u16 max_size = event->header.size;
1312 const void *endp = (void *)event + max_size;
1316 * used for cross-endian analysis. See git commit 65014ab3
1317 * for why this goofiness is needed.
1321 memset(data, 0, sizeof(*data));
1322 data->cpu = data->pid = data->tid = -1;
1323 data->stream_id = data->id = data->time = -1ULL;
1324 data->period = evsel->attr.sample_period;
1327 if (event->header.type != PERF_RECORD_SAMPLE) {
1328 if (!evsel->attr.sample_id_all)
1330 return perf_evsel__parse_id_sample(evsel, event, data);
1333 array = event->sample.array;
1336 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1337 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1338 * check the format does not go past the end of the event.
1340 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1344 if (type & PERF_SAMPLE_IDENTIFIER) {
1349 if (type & PERF_SAMPLE_IP) {
1354 if (type & PERF_SAMPLE_TID) {
1357 /* undo swap of u64, then swap on individual u32s */
1358 u.val64 = bswap_64(u.val64);
1359 u.val32[0] = bswap_32(u.val32[0]);
1360 u.val32[1] = bswap_32(u.val32[1]);
1363 data->pid = u.val32[0];
1364 data->tid = u.val32[1];
1368 if (type & PERF_SAMPLE_TIME) {
1369 data->time = *array;
1374 if (type & PERF_SAMPLE_ADDR) {
1375 data->addr = *array;
1379 if (type & PERF_SAMPLE_ID) {
1384 if (type & PERF_SAMPLE_STREAM_ID) {
1385 data->stream_id = *array;
1389 if (type & PERF_SAMPLE_CPU) {
1393 /* undo swap of u64, then swap on individual u32s */
1394 u.val64 = bswap_64(u.val64);
1395 u.val32[0] = bswap_32(u.val32[0]);
1398 data->cpu = u.val32[0];
1402 if (type & PERF_SAMPLE_PERIOD) {
1403 data->period = *array;
1407 if (type & PERF_SAMPLE_READ) {
1408 u64 read_format = evsel->attr.read_format;
1410 OVERFLOW_CHECK_u64(array);
1411 if (read_format & PERF_FORMAT_GROUP)
1412 data->read.group.nr = *array;
1414 data->read.one.value = *array;
1418 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1419 OVERFLOW_CHECK_u64(array);
1420 data->read.time_enabled = *array;
1424 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1425 OVERFLOW_CHECK_u64(array);
1426 data->read.time_running = *array;
1430 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1431 if (read_format & PERF_FORMAT_GROUP) {
1432 const u64 max_group_nr = UINT64_MAX /
1433 sizeof(struct sample_read_value);
1435 if (data->read.group.nr > max_group_nr)
1437 sz = data->read.group.nr *
1438 sizeof(struct sample_read_value);
1439 OVERFLOW_CHECK(array, sz, max_size);
1440 data->read.group.values =
1441 (struct sample_read_value *)array;
1442 array = (void *)array + sz;
1444 OVERFLOW_CHECK_u64(array);
1445 data->read.one.id = *array;
1450 if (type & PERF_SAMPLE_CALLCHAIN) {
1451 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1453 OVERFLOW_CHECK_u64(array);
1454 data->callchain = (struct ip_callchain *)array++;
1455 if (data->callchain->nr > max_callchain_nr)
1457 sz = data->callchain->nr * sizeof(u64);
1458 OVERFLOW_CHECK(array, sz, max_size);
1459 array = (void *)array + sz;
1462 if (type & PERF_SAMPLE_RAW) {
1463 OVERFLOW_CHECK_u64(array);
1465 if (WARN_ONCE(swapped,
1466 "Endianness of raw data not corrected!\n")) {
1467 /* undo swap of u64, then swap on individual u32s */
1468 u.val64 = bswap_64(u.val64);
1469 u.val32[0] = bswap_32(u.val32[0]);
1470 u.val32[1] = bswap_32(u.val32[1]);
1472 data->raw_size = u.val32[0];
1473 array = (void *)array + sizeof(u32);
1475 OVERFLOW_CHECK(array, data->raw_size, max_size);
1476 data->raw_data = (void *)array;
1477 array = (void *)array + data->raw_size;
1480 if (type & PERF_SAMPLE_BRANCH_STACK) {
1481 const u64 max_branch_nr = UINT64_MAX /
1482 sizeof(struct branch_entry);
1484 OVERFLOW_CHECK_u64(array);
1485 data->branch_stack = (struct branch_stack *)array++;
1487 if (data->branch_stack->nr > max_branch_nr)
1489 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1490 OVERFLOW_CHECK(array, sz, max_size);
1491 array = (void *)array + sz;
1494 if (type & PERF_SAMPLE_REGS_USER) {
1495 OVERFLOW_CHECK_u64(array);
1496 data->user_regs.abi = *array;
1499 if (data->user_regs.abi) {
1500 u64 mask = evsel->attr.sample_regs_user;
1502 sz = hweight_long(mask) * sizeof(u64);
1503 OVERFLOW_CHECK(array, sz, max_size);
1504 data->user_regs.mask = mask;
1505 data->user_regs.regs = (u64 *)array;
1506 array = (void *)array + sz;
1510 if (type & PERF_SAMPLE_STACK_USER) {
1511 OVERFLOW_CHECK_u64(array);
1514 data->user_stack.offset = ((char *)(array - 1)
1518 data->user_stack.size = 0;
1520 OVERFLOW_CHECK(array, sz, max_size);
1521 data->user_stack.data = (char *)array;
1522 array = (void *)array + sz;
1523 OVERFLOW_CHECK_u64(array);
1524 data->user_stack.size = *array++;
1525 if (WARN_ONCE(data->user_stack.size > sz,
1526 "user stack dump failure\n"))
1532 if (type & PERF_SAMPLE_WEIGHT) {
1533 OVERFLOW_CHECK_u64(array);
1534 data->weight = *array;
1538 data->data_src = PERF_MEM_DATA_SRC_NONE;
1539 if (type & PERF_SAMPLE_DATA_SRC) {
1540 OVERFLOW_CHECK_u64(array);
1541 data->data_src = *array;
1545 data->transaction = 0;
1546 if (type & PERF_SAMPLE_TRANSACTION) {
1547 OVERFLOW_CHECK_u64(array);
1548 data->transaction = *array;
1552 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1553 if (type & PERF_SAMPLE_REGS_INTR) {
1554 OVERFLOW_CHECK_u64(array);
1555 data->intr_regs.abi = *array;
1558 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1559 u64 mask = evsel->attr.sample_regs_intr;
1561 sz = hweight_long(mask) * sizeof(u64);
1562 OVERFLOW_CHECK(array, sz, max_size);
1563 data->intr_regs.mask = mask;
1564 data->intr_regs.regs = (u64 *)array;
1565 array = (void *)array + sz;
1572 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1575 size_t sz, result = sizeof(struct sample_event);
1577 if (type & PERF_SAMPLE_IDENTIFIER)
1578 result += sizeof(u64);
1580 if (type & PERF_SAMPLE_IP)
1581 result += sizeof(u64);
1583 if (type & PERF_SAMPLE_TID)
1584 result += sizeof(u64);
1586 if (type & PERF_SAMPLE_TIME)
1587 result += sizeof(u64);
1589 if (type & PERF_SAMPLE_ADDR)
1590 result += sizeof(u64);
1592 if (type & PERF_SAMPLE_ID)
1593 result += sizeof(u64);
1595 if (type & PERF_SAMPLE_STREAM_ID)
1596 result += sizeof(u64);
1598 if (type & PERF_SAMPLE_CPU)
1599 result += sizeof(u64);
1601 if (type & PERF_SAMPLE_PERIOD)
1602 result += sizeof(u64);
1604 if (type & PERF_SAMPLE_READ) {
1605 result += sizeof(u64);
1606 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1607 result += sizeof(u64);
1608 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1609 result += sizeof(u64);
1610 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1611 if (read_format & PERF_FORMAT_GROUP) {
1612 sz = sample->read.group.nr *
1613 sizeof(struct sample_read_value);
1616 result += sizeof(u64);
1620 if (type & PERF_SAMPLE_CALLCHAIN) {
1621 sz = (sample->callchain->nr + 1) * sizeof(u64);
1625 if (type & PERF_SAMPLE_RAW) {
1626 result += sizeof(u32);
1627 result += sample->raw_size;
1630 if (type & PERF_SAMPLE_BRANCH_STACK) {
1631 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1636 if (type & PERF_SAMPLE_REGS_USER) {
1637 if (sample->user_regs.abi) {
1638 result += sizeof(u64);
1639 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1642 result += sizeof(u64);
1646 if (type & PERF_SAMPLE_STACK_USER) {
1647 sz = sample->user_stack.size;
1648 result += sizeof(u64);
1651 result += sizeof(u64);
1655 if (type & PERF_SAMPLE_WEIGHT)
1656 result += sizeof(u64);
1658 if (type & PERF_SAMPLE_DATA_SRC)
1659 result += sizeof(u64);
1661 if (type & PERF_SAMPLE_TRANSACTION)
1662 result += sizeof(u64);
1664 if (type & PERF_SAMPLE_REGS_INTR) {
1665 if (sample->intr_regs.abi) {
1666 result += sizeof(u64);
1667 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1670 result += sizeof(u64);
1677 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1679 const struct perf_sample *sample,
1685 * used for cross-endian analysis. See git commit 65014ab3
1686 * for why this goofiness is needed.
1690 array = event->sample.array;
1692 if (type & PERF_SAMPLE_IDENTIFIER) {
1693 *array = sample->id;
1697 if (type & PERF_SAMPLE_IP) {
1698 *array = sample->ip;
1702 if (type & PERF_SAMPLE_TID) {
1703 u.val32[0] = sample->pid;
1704 u.val32[1] = sample->tid;
1707 * Inverse of what is done in perf_evsel__parse_sample
1709 u.val32[0] = bswap_32(u.val32[0]);
1710 u.val32[1] = bswap_32(u.val32[1]);
1711 u.val64 = bswap_64(u.val64);
1718 if (type & PERF_SAMPLE_TIME) {
1719 *array = sample->time;
1723 if (type & PERF_SAMPLE_ADDR) {
1724 *array = sample->addr;
1728 if (type & PERF_SAMPLE_ID) {
1729 *array = sample->id;
1733 if (type & PERF_SAMPLE_STREAM_ID) {
1734 *array = sample->stream_id;
1738 if (type & PERF_SAMPLE_CPU) {
1739 u.val32[0] = sample->cpu;
1742 * Inverse of what is done in perf_evsel__parse_sample
1744 u.val32[0] = bswap_32(u.val32[0]);
1745 u.val64 = bswap_64(u.val64);
1751 if (type & PERF_SAMPLE_PERIOD) {
1752 *array = sample->period;
1756 if (type & PERF_SAMPLE_READ) {
1757 if (read_format & PERF_FORMAT_GROUP)
1758 *array = sample->read.group.nr;
1760 *array = sample->read.one.value;
1763 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1764 *array = sample->read.time_enabled;
1768 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1769 *array = sample->read.time_running;
1773 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1774 if (read_format & PERF_FORMAT_GROUP) {
1775 sz = sample->read.group.nr *
1776 sizeof(struct sample_read_value);
1777 memcpy(array, sample->read.group.values, sz);
1778 array = (void *)array + sz;
1780 *array = sample->read.one.id;
1785 if (type & PERF_SAMPLE_CALLCHAIN) {
1786 sz = (sample->callchain->nr + 1) * sizeof(u64);
1787 memcpy(array, sample->callchain, sz);
1788 array = (void *)array + sz;
1791 if (type & PERF_SAMPLE_RAW) {
1792 u.val32[0] = sample->raw_size;
1793 if (WARN_ONCE(swapped,
1794 "Endianness of raw data not corrected!\n")) {
1796 * Inverse of what is done in perf_evsel__parse_sample
1798 u.val32[0] = bswap_32(u.val32[0]);
1799 u.val32[1] = bswap_32(u.val32[1]);
1800 u.val64 = bswap_64(u.val64);
1803 array = (void *)array + sizeof(u32);
1805 memcpy(array, sample->raw_data, sample->raw_size);
1806 array = (void *)array + sample->raw_size;
1809 if (type & PERF_SAMPLE_BRANCH_STACK) {
1810 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1812 memcpy(array, sample->branch_stack, sz);
1813 array = (void *)array + sz;
1816 if (type & PERF_SAMPLE_REGS_USER) {
1817 if (sample->user_regs.abi) {
1818 *array++ = sample->user_regs.abi;
1819 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1820 memcpy(array, sample->user_regs.regs, sz);
1821 array = (void *)array + sz;
1827 if (type & PERF_SAMPLE_STACK_USER) {
1828 sz = sample->user_stack.size;
1831 memcpy(array, sample->user_stack.data, sz);
1832 array = (void *)array + sz;
1837 if (type & PERF_SAMPLE_WEIGHT) {
1838 *array = sample->weight;
1842 if (type & PERF_SAMPLE_DATA_SRC) {
1843 *array = sample->data_src;
1847 if (type & PERF_SAMPLE_TRANSACTION) {
1848 *array = sample->transaction;
1852 if (type & PERF_SAMPLE_REGS_INTR) {
1853 if (sample->intr_regs.abi) {
1854 *array++ = sample->intr_regs.abi;
1855 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1856 memcpy(array, sample->intr_regs.regs, sz);
1857 array = (void *)array + sz;
1866 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1868 return pevent_find_field(evsel->tp_format, name);
1871 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1874 struct format_field *field = perf_evsel__field(evsel, name);
1880 offset = field->offset;
1882 if (field->flags & FIELD_IS_DYNAMIC) {
1883 offset = *(int *)(sample->raw_data + field->offset);
1887 return sample->raw_data + offset;
1890 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1893 struct format_field *field = perf_evsel__field(evsel, name);
1900 ptr = sample->raw_data + field->offset;
1902 switch (field->size) {
1906 value = *(u16 *)ptr;
1909 value = *(u32 *)ptr;
1912 value = *(u64 *)ptr;
1918 if (!evsel->needs_swap)
1921 switch (field->size) {
1923 return bswap_16(value);
1925 return bswap_32(value);
1927 return bswap_64(value);
1935 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1941 ret += fprintf(fp, ",");
1943 ret += fprintf(fp, ":");
1947 va_start(args, fmt);
1948 ret += vfprintf(fp, fmt, args);
1953 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1958 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1961 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1968 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1969 struct bit_names *bits, bool *first)
1971 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1972 bool first_bit = true;
1975 if (value & bits[i].bit) {
1976 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1979 } while (bits[++i].name != NULL);
1984 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1986 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1987 struct bit_names bits[] = {
1988 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1989 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1990 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1991 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1992 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1996 return bits__fprintf(fp, "sample_type", value, bits, first);
1999 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
2001 #define bit_name(n) { PERF_FORMAT_##n, #n }
2002 struct bit_names bits[] = {
2003 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
2004 bit_name(ID), bit_name(GROUP),
2008 return bits__fprintf(fp, "read_format", value, bits, first);
2011 int perf_evsel__fprintf(struct perf_evsel *evsel,
2012 struct perf_attr_details *details, FILE *fp)
2017 if (details->event_group) {
2018 struct perf_evsel *pos;
2020 if (!perf_evsel__is_group_leader(evsel))
2023 if (evsel->nr_members > 1)
2024 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2026 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2027 for_each_group_member(pos, evsel)
2028 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2030 if (evsel->nr_members > 1)
2031 printed += fprintf(fp, "}");
2035 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2037 if (details->verbose || details->freq) {
2038 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2039 (u64)evsel->attr.sample_freq);
2042 if (details->verbose) {
2048 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
2049 if (evsel->attr.read_format)
2050 printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
2054 if_print(exclusive);
2055 if_print(exclude_user);
2056 if_print(exclude_kernel);
2057 if_print(exclude_hv);
2058 if_print(exclude_idle);
2062 if_print(comm_exec);
2064 if_print(inherit_stat);
2065 if_print(enable_on_exec);
2067 if_print(watermark);
2068 if_print(precise_ip);
2069 if_print(mmap_data);
2070 if_print(sample_id_all);
2071 if_print(exclude_host);
2072 if_print(exclude_guest);
2073 if_print(__reserved_1);
2074 if_print(wakeup_events);
2076 if_print(branch_sample_type);
2083 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2084 char *msg, size_t msgsize)
2086 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2087 evsel->attr.type == PERF_TYPE_HARDWARE &&
2088 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2090 * If it's cycles then fall back to hrtimer based
2091 * cpu-clock-tick sw counter, which is always available even if
2094 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2097 scnprintf(msg, msgsize, "%s",
2098 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2100 evsel->attr.type = PERF_TYPE_SOFTWARE;
2101 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2103 zfree(&evsel->name);
2110 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2111 int err, char *msg, size_t size)
2113 char sbuf[STRERR_BUFSIZE];
2118 return scnprintf(msg, size,
2119 "You may not have permission to collect %sstats.\n"
2120 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2121 " -1 - Not paranoid at all\n"
2122 " 0 - Disallow raw tracepoint access for unpriv\n"
2123 " 1 - Disallow cpu events for unpriv\n"
2124 " 2 - Disallow kernel profiling for unpriv",
2125 target->system_wide ? "system-wide " : "");
2127 return scnprintf(msg, size, "The %s event is not supported.",
2128 perf_evsel__name(evsel));
2130 return scnprintf(msg, size, "%s",
2131 "Too many events are opened.\n"
2132 "Try again after reducing the number of events.");
2134 if (target->cpu_list)
2135 return scnprintf(msg, size, "%s",
2136 "No such device - did you specify an out-of-range profile CPU?\n");
2139 if (evsel->attr.precise_ip)
2140 return scnprintf(msg, size, "%s",
2141 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2142 #if defined(__i386__) || defined(__x86_64__)
2143 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2144 return scnprintf(msg, size, "%s",
2145 "No hardware sampling interrupt available.\n"
2146 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2150 if (find_process("oprofiled"))
2151 return scnprintf(msg, size,
2152 "The PMU counters are busy/taken by another profiler.\n"
2153 "We found oprofile daemon running, please stop it and try again.");
2159 return scnprintf(msg, size,
2160 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2161 "/bin/dmesg may provide additional information.\n"
2162 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2163 err, strerror_r(err, sbuf, sizeof(sbuf)),
2164 perf_evsel__name(evsel));