2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
7 * Released under the GPL v2. (and only v2, not any later version)
11 #include <linux/bitops.h>
12 #include <api/fs/debugfs.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <sys/resource.h>
18 #include "callchain.h"
24 #include "thread_map.h"
26 #include "perf_regs.h"
28 #include "trace-event.h"
38 } perf_missing_features;
40 static clockid_t clockid;
42 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
47 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
53 int (*init)(struct perf_evsel *evsel);
54 void (*fini)(struct perf_evsel *evsel);
55 } perf_evsel__object = {
56 .size = sizeof(struct perf_evsel),
57 .init = perf_evsel__no_extra_init,
58 .fini = perf_evsel__no_extra_fini,
61 int perf_evsel__object_config(size_t object_size,
62 int (*init)(struct perf_evsel *evsel),
63 void (*fini)(struct perf_evsel *evsel))
69 if (perf_evsel__object.size > object_size)
72 perf_evsel__object.size = object_size;
76 perf_evsel__object.init = init;
79 perf_evsel__object.fini = fini;
84 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
86 int __perf_evsel__sample_size(u64 sample_type)
88 u64 mask = sample_type & PERF_SAMPLE_MASK;
92 for (i = 0; i < 64; i++) {
93 if (mask & (1ULL << i))
103 * __perf_evsel__calc_id_pos - calculate id_pos.
104 * @sample_type: sample type
106 * This function returns the position of the event id (PERF_SAMPLE_ID or
107 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
110 static int __perf_evsel__calc_id_pos(u64 sample_type)
114 if (sample_type & PERF_SAMPLE_IDENTIFIER)
117 if (!(sample_type & PERF_SAMPLE_ID))
120 if (sample_type & PERF_SAMPLE_IP)
123 if (sample_type & PERF_SAMPLE_TID)
126 if (sample_type & PERF_SAMPLE_TIME)
129 if (sample_type & PERF_SAMPLE_ADDR)
136 * __perf_evsel__calc_is_pos - calculate is_pos.
137 * @sample_type: sample type
139 * This function returns the position (counting backwards) of the event id
140 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
141 * sample_id_all is used there is an id sample appended to non-sample events.
143 static int __perf_evsel__calc_is_pos(u64 sample_type)
147 if (sample_type & PERF_SAMPLE_IDENTIFIER)
150 if (!(sample_type & PERF_SAMPLE_ID))
153 if (sample_type & PERF_SAMPLE_CPU)
156 if (sample_type & PERF_SAMPLE_STREAM_ID)
162 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
164 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
165 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
168 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
169 enum perf_event_sample_format bit)
171 if (!(evsel->attr.sample_type & bit)) {
172 evsel->attr.sample_type |= bit;
173 evsel->sample_size += sizeof(u64);
174 perf_evsel__calc_id_pos(evsel);
178 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
179 enum perf_event_sample_format bit)
181 if (evsel->attr.sample_type & bit) {
182 evsel->attr.sample_type &= ~bit;
183 evsel->sample_size -= sizeof(u64);
184 perf_evsel__calc_id_pos(evsel);
188 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
189 bool can_sample_identifier)
191 if (can_sample_identifier) {
192 perf_evsel__reset_sample_bit(evsel, ID);
193 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
195 perf_evsel__set_sample_bit(evsel, ID);
197 evsel->attr.read_format |= PERF_FORMAT_ID;
200 void perf_evsel__init(struct perf_evsel *evsel,
201 struct perf_event_attr *attr, int idx)
204 evsel->tracking = !idx;
206 evsel->leader = evsel;
209 INIT_LIST_HEAD(&evsel->node);
210 perf_evsel__object.init(evsel);
211 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
212 perf_evsel__calc_id_pos(evsel);
215 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
217 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
220 perf_evsel__init(evsel, attr, idx);
225 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
227 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
230 struct perf_event_attr attr = {
231 .type = PERF_TYPE_TRACEPOINT,
232 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
233 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
236 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
239 evsel->tp_format = trace_event__tp_format(sys, name);
240 if (evsel->tp_format == NULL)
243 event_attr_init(&attr);
244 attr.config = evsel->tp_format->id;
245 attr.sample_period = 1;
246 perf_evsel__init(evsel, &attr, idx);
257 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
265 "stalled-cycles-frontend",
266 "stalled-cycles-backend",
270 static const char *__perf_evsel__hw_name(u64 config)
272 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
273 return perf_evsel__hw_names[config];
275 return "unknown-hardware";
278 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
280 int colon = 0, r = 0;
281 struct perf_event_attr *attr = &evsel->attr;
282 bool exclude_guest_default = false;
284 #define MOD_PRINT(context, mod) do { \
285 if (!attr->exclude_##context) { \
286 if (!colon) colon = ++r; \
287 r += scnprintf(bf + r, size - r, "%c", mod); \
290 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
291 MOD_PRINT(kernel, 'k');
292 MOD_PRINT(user, 'u');
294 exclude_guest_default = true;
297 if (attr->precise_ip) {
300 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
301 exclude_guest_default = true;
304 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
305 MOD_PRINT(host, 'H');
306 MOD_PRINT(guest, 'G');
314 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
316 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
317 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
320 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
333 static const char *__perf_evsel__sw_name(u64 config)
335 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
336 return perf_evsel__sw_names[config];
337 return "unknown-software";
340 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
342 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
343 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
346 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
350 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
352 if (type & HW_BREAKPOINT_R)
353 r += scnprintf(bf + r, size - r, "r");
355 if (type & HW_BREAKPOINT_W)
356 r += scnprintf(bf + r, size - r, "w");
358 if (type & HW_BREAKPOINT_X)
359 r += scnprintf(bf + r, size - r, "x");
364 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
366 struct perf_event_attr *attr = &evsel->attr;
367 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
368 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
371 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
372 [PERF_EVSEL__MAX_ALIASES] = {
373 { "L1-dcache", "l1-d", "l1d", "L1-data", },
374 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
376 { "dTLB", "d-tlb", "Data-TLB", },
377 { "iTLB", "i-tlb", "Instruction-TLB", },
378 { "branch", "branches", "bpu", "btb", "bpc", },
382 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
383 [PERF_EVSEL__MAX_ALIASES] = {
384 { "load", "loads", "read", },
385 { "store", "stores", "write", },
386 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
389 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
390 [PERF_EVSEL__MAX_ALIASES] = {
391 { "refs", "Reference", "ops", "access", },
392 { "misses", "miss", },
395 #define C(x) PERF_COUNT_HW_CACHE_##x
396 #define CACHE_READ (1 << C(OP_READ))
397 #define CACHE_WRITE (1 << C(OP_WRITE))
398 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
399 #define COP(x) (1 << x)
402 * cache operartion stat
403 * L1I : Read and prefetch only
404 * ITLB and BPU : Read-only
406 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
407 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
408 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
409 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
410 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
411 [C(ITLB)] = (CACHE_READ),
412 [C(BPU)] = (CACHE_READ),
413 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
416 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
418 if (perf_evsel__hw_cache_stat[type] & COP(op))
419 return true; /* valid */
421 return false; /* invalid */
424 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
425 char *bf, size_t size)
428 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
429 perf_evsel__hw_cache_op[op][0],
430 perf_evsel__hw_cache_result[result][0]);
433 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
434 perf_evsel__hw_cache_op[op][1]);
437 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
439 u8 op, result, type = (config >> 0) & 0xff;
440 const char *err = "unknown-ext-hardware-cache-type";
442 if (type > PERF_COUNT_HW_CACHE_MAX)
445 op = (config >> 8) & 0xff;
446 err = "unknown-ext-hardware-cache-op";
447 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
450 result = (config >> 16) & 0xff;
451 err = "unknown-ext-hardware-cache-result";
452 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
455 err = "invalid-cache";
456 if (!perf_evsel__is_cache_op_valid(type, op))
459 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
461 return scnprintf(bf, size, "%s", err);
464 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
466 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
467 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
470 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
472 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
473 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
476 const char *perf_evsel__name(struct perf_evsel *evsel)
483 switch (evsel->attr.type) {
485 perf_evsel__raw_name(evsel, bf, sizeof(bf));
488 case PERF_TYPE_HARDWARE:
489 perf_evsel__hw_name(evsel, bf, sizeof(bf));
492 case PERF_TYPE_HW_CACHE:
493 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
496 case PERF_TYPE_SOFTWARE:
497 perf_evsel__sw_name(evsel, bf, sizeof(bf));
500 case PERF_TYPE_TRACEPOINT:
501 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
504 case PERF_TYPE_BREAKPOINT:
505 perf_evsel__bp_name(evsel, bf, sizeof(bf));
509 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
514 evsel->name = strdup(bf);
516 return evsel->name ?: "unknown";
519 const char *perf_evsel__group_name(struct perf_evsel *evsel)
521 return evsel->group_name ?: "anon group";
524 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
527 struct perf_evsel *pos;
528 const char *group_name = perf_evsel__group_name(evsel);
530 ret = scnprintf(buf, size, "%s", group_name);
532 ret += scnprintf(buf + ret, size - ret, " { %s",
533 perf_evsel__name(evsel));
535 for_each_group_member(pos, evsel)
536 ret += scnprintf(buf + ret, size - ret, ", %s",
537 perf_evsel__name(pos));
539 ret += scnprintf(buf + ret, size - ret, " }");
545 perf_evsel__config_callgraph(struct perf_evsel *evsel,
546 struct record_opts *opts)
548 bool function = perf_evsel__is_function_event(evsel);
549 struct perf_event_attr *attr = &evsel->attr;
551 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
553 if (callchain_param.record_mode == CALLCHAIN_LBR) {
554 if (!opts->branch_stack) {
555 if (attr->exclude_user) {
556 pr_warning("LBR callstack option is only available "
557 "to get user callchain information. "
558 "Falling back to framepointers.\n");
560 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
561 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
562 PERF_SAMPLE_BRANCH_CALL_STACK;
565 pr_warning("Cannot use LBR callstack with branch stack. "
566 "Falling back to framepointers.\n");
569 if (callchain_param.record_mode == CALLCHAIN_DWARF) {
571 perf_evsel__set_sample_bit(evsel, REGS_USER);
572 perf_evsel__set_sample_bit(evsel, STACK_USER);
573 attr->sample_regs_user = PERF_REGS_MASK;
574 attr->sample_stack_user = callchain_param.dump_size;
575 attr->exclude_callchain_user = 1;
577 pr_info("Cannot use DWARF unwind for function trace event,"
578 " falling back to framepointers.\n");
583 pr_info("Disabling user space callchains for function trace event.\n");
584 attr->exclude_callchain_user = 1;
589 * The enable_on_exec/disabled value strategy:
591 * 1) For any type of traced program:
592 * - all independent events and group leaders are disabled
593 * - all group members are enabled
595 * Group members are ruled by group leaders. They need to
596 * be enabled, because the group scheduling relies on that.
598 * 2) For traced programs executed by perf:
599 * - all independent events and group leaders have
601 * - we don't specifically enable or disable any event during
604 * Independent events and group leaders are initially disabled
605 * and get enabled by exec. Group members are ruled by group
606 * leaders as stated in 1).
608 * 3) For traced programs attached by perf (pid/tid):
609 * - we specifically enable or disable all events during
612 * When attaching events to already running traced we
613 * enable/disable events specifically, as there's no
614 * initial traced exec call.
616 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
618 struct perf_evsel *leader = evsel->leader;
619 struct perf_event_attr *attr = &evsel->attr;
620 int track = evsel->tracking;
621 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
623 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
624 attr->inherit = !opts->no_inherit;
626 perf_evsel__set_sample_bit(evsel, IP);
627 perf_evsel__set_sample_bit(evsel, TID);
629 if (evsel->sample_read) {
630 perf_evsel__set_sample_bit(evsel, READ);
633 * We need ID even in case of single event, because
634 * PERF_SAMPLE_READ process ID specific data.
636 perf_evsel__set_sample_id(evsel, false);
639 * Apply group format only if we belong to group
640 * with more than one members.
642 if (leader->nr_members > 1) {
643 attr->read_format |= PERF_FORMAT_GROUP;
649 * We default some events to have a default interval. But keep
650 * it a weak assumption overridable by the user.
652 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
653 opts->user_interval != ULLONG_MAX)) {
655 perf_evsel__set_sample_bit(evsel, PERIOD);
657 attr->sample_freq = opts->freq;
659 attr->sample_period = opts->default_interval;
664 * Disable sampling for all group members other
665 * than leader in case leader 'leads' the sampling.
667 if ((leader != evsel) && leader->sample_read) {
668 attr->sample_freq = 0;
669 attr->sample_period = 0;
672 if (opts->no_samples)
673 attr->sample_freq = 0;
675 if (opts->inherit_stat)
676 attr->inherit_stat = 1;
678 if (opts->sample_address) {
679 perf_evsel__set_sample_bit(evsel, ADDR);
680 attr->mmap_data = track;
684 * We don't allow user space callchains for function trace
685 * event, due to issues with page faults while tracing page
686 * fault handler and its overall trickiness nature.
688 if (perf_evsel__is_function_event(evsel))
689 evsel->attr.exclude_callchain_user = 1;
691 if (callchain_param.enabled && !evsel->no_aux_samples)
692 perf_evsel__config_callgraph(evsel, opts);
694 if (opts->sample_intr_regs) {
695 attr->sample_regs_intr = PERF_REGS_MASK;
696 perf_evsel__set_sample_bit(evsel, REGS_INTR);
699 if (target__has_cpu(&opts->target))
700 perf_evsel__set_sample_bit(evsel, CPU);
703 perf_evsel__set_sample_bit(evsel, PERIOD);
706 * When the user explicitely disabled time don't force it here.
708 if (opts->sample_time &&
709 (!perf_missing_features.sample_id_all &&
710 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu)))
711 perf_evsel__set_sample_bit(evsel, TIME);
713 if (opts->raw_samples && !evsel->no_aux_samples) {
714 perf_evsel__set_sample_bit(evsel, TIME);
715 perf_evsel__set_sample_bit(evsel, RAW);
716 perf_evsel__set_sample_bit(evsel, CPU);
719 if (opts->sample_address)
720 perf_evsel__set_sample_bit(evsel, DATA_SRC);
722 if (opts->no_buffering) {
724 attr->wakeup_events = 1;
726 if (opts->branch_stack && !evsel->no_aux_samples) {
727 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
728 attr->branch_sample_type = opts->branch_stack;
731 if (opts->sample_weight)
732 perf_evsel__set_sample_bit(evsel, WEIGHT);
736 attr->mmap2 = track && !perf_missing_features.mmap2;
739 if (opts->sample_transaction)
740 perf_evsel__set_sample_bit(evsel, TRANSACTION);
742 if (opts->running_time) {
743 evsel->attr.read_format |=
744 PERF_FORMAT_TOTAL_TIME_ENABLED |
745 PERF_FORMAT_TOTAL_TIME_RUNNING;
749 * XXX see the function comment above
751 * Disabling only independent events or group leaders,
752 * keeping group members enabled.
754 if (perf_evsel__is_group_leader(evsel))
758 * Setting enable_on_exec for independent events and
759 * group leaders for traced executed by perf.
761 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
762 !opts->initial_delay)
763 attr->enable_on_exec = 1;
765 if (evsel->immediate) {
767 attr->enable_on_exec = 0;
770 clockid = opts->clockid;
771 if (opts->use_clockid) {
772 attr->use_clockid = 1;
773 attr->clockid = opts->clockid;
777 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
781 if (evsel->system_wide)
784 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
787 for (cpu = 0; cpu < ncpus; cpu++) {
788 for (thread = 0; thread < nthreads; thread++) {
789 FD(evsel, cpu, thread) = -1;
794 return evsel->fd != NULL ? 0 : -ENOMEM;
797 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
802 if (evsel->system_wide)
805 for (cpu = 0; cpu < ncpus; cpu++) {
806 for (thread = 0; thread < nthreads; thread++) {
807 int fd = FD(evsel, cpu, thread),
808 err = ioctl(fd, ioc, arg);
818 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
821 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
822 PERF_EVENT_IOC_SET_FILTER,
826 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
828 char *new_filter = strdup(filter);
830 if (new_filter != NULL) {
832 evsel->filter = new_filter;
839 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
841 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
842 PERF_EVENT_IOC_ENABLE,
846 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
848 if (ncpus == 0 || nthreads == 0)
851 if (evsel->system_wide)
854 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
855 if (evsel->sample_id == NULL)
858 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
859 if (evsel->id == NULL) {
860 xyarray__delete(evsel->sample_id);
861 evsel->sample_id = NULL;
868 static void perf_evsel__free_fd(struct perf_evsel *evsel)
870 xyarray__delete(evsel->fd);
874 static void perf_evsel__free_id(struct perf_evsel *evsel)
876 xyarray__delete(evsel->sample_id);
877 evsel->sample_id = NULL;
881 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
885 if (evsel->system_wide)
888 for (cpu = 0; cpu < ncpus; cpu++)
889 for (thread = 0; thread < nthreads; ++thread) {
890 close(FD(evsel, cpu, thread));
891 FD(evsel, cpu, thread) = -1;
895 void perf_evsel__exit(struct perf_evsel *evsel)
897 assert(list_empty(&evsel->node));
898 perf_evsel__free_fd(evsel);
899 perf_evsel__free_id(evsel);
900 close_cgroup(evsel->cgrp);
901 cpu_map__put(evsel->cpus);
902 thread_map__put(evsel->threads);
903 zfree(&evsel->group_name);
905 perf_evsel__object.fini(evsel);
908 void perf_evsel__delete(struct perf_evsel *evsel)
910 perf_evsel__exit(evsel);
914 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
915 struct perf_counts_values *count)
917 struct perf_counts_values tmp;
919 if (!evsel->prev_raw_counts)
923 tmp = evsel->prev_raw_counts->aggr;
924 evsel->prev_raw_counts->aggr = *count;
926 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
927 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
930 count->val = count->val - tmp.val;
931 count->ena = count->ena - tmp.ena;
932 count->run = count->run - tmp.run;
935 void perf_counts_values__scale(struct perf_counts_values *count,
936 bool scale, s8 *pscaled)
941 if (count->run == 0) {
944 } else if (count->run < count->ena) {
946 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
949 count->ena = count->run = 0;
955 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
956 struct perf_counts_values *count)
958 memset(count, 0, sizeof(*count));
960 if (FD(evsel, cpu, thread) < 0)
963 if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
969 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
970 int cpu, int thread, bool scale)
972 struct perf_counts_values count;
973 size_t nv = scale ? 3 : 1;
975 if (FD(evsel, cpu, thread) < 0)
978 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
981 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
984 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
985 perf_counts_values__scale(&count, scale, NULL);
986 *perf_counts(evsel->counts, cpu, thread) = count;
990 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
992 struct perf_evsel *leader = evsel->leader;
995 if (perf_evsel__is_group_leader(evsel))
999 * Leader must be already processed/open,
1000 * if not it's a bug.
1002 BUG_ON(!leader->fd);
1004 fd = FD(leader, cpu, thread);
1015 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1017 bool first_bit = true;
1021 if (value & bits[i].bit) {
1022 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1025 } while (bits[++i].name != NULL);
1028 static void __p_sample_type(char *buf, size_t size, u64 value)
1030 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1031 struct bit_names bits[] = {
1032 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1033 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1034 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1035 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1036 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1040 __p_bits(buf, size, value, bits);
1043 static void __p_read_format(char *buf, size_t size, u64 value)
1045 #define bit_name(n) { PERF_FORMAT_##n, #n }
1046 struct bit_names bits[] = {
1047 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1048 bit_name(ID), bit_name(GROUP),
1052 __p_bits(buf, size, value, bits);
1055 #define BUF_SIZE 1024
1057 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1058 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1059 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1060 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1061 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1063 #define PRINT_ATTRn(_n, _f, _p) \
1067 ret += attr__fprintf(fp, _n, buf, priv);\
1071 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1073 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1074 attr__fprintf_f attr__fprintf, void *priv)
1079 PRINT_ATTRf(type, p_unsigned);
1080 PRINT_ATTRf(size, p_unsigned);
1081 PRINT_ATTRf(config, p_hex);
1082 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1083 PRINT_ATTRf(sample_type, p_sample_type);
1084 PRINT_ATTRf(read_format, p_read_format);
1086 PRINT_ATTRf(disabled, p_unsigned);
1087 PRINT_ATTRf(inherit, p_unsigned);
1088 PRINT_ATTRf(pinned, p_unsigned);
1089 PRINT_ATTRf(exclusive, p_unsigned);
1090 PRINT_ATTRf(exclude_user, p_unsigned);
1091 PRINT_ATTRf(exclude_kernel, p_unsigned);
1092 PRINT_ATTRf(exclude_hv, p_unsigned);
1093 PRINT_ATTRf(exclude_idle, p_unsigned);
1094 PRINT_ATTRf(mmap, p_unsigned);
1095 PRINT_ATTRf(comm, p_unsigned);
1096 PRINT_ATTRf(freq, p_unsigned);
1097 PRINT_ATTRf(inherit_stat, p_unsigned);
1098 PRINT_ATTRf(enable_on_exec, p_unsigned);
1099 PRINT_ATTRf(task, p_unsigned);
1100 PRINT_ATTRf(watermark, p_unsigned);
1101 PRINT_ATTRf(precise_ip, p_unsigned);
1102 PRINT_ATTRf(mmap_data, p_unsigned);
1103 PRINT_ATTRf(sample_id_all, p_unsigned);
1104 PRINT_ATTRf(exclude_host, p_unsigned);
1105 PRINT_ATTRf(exclude_guest, p_unsigned);
1106 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1107 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1108 PRINT_ATTRf(mmap2, p_unsigned);
1109 PRINT_ATTRf(comm_exec, p_unsigned);
1110 PRINT_ATTRf(use_clockid, p_unsigned);
1112 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1113 PRINT_ATTRf(bp_type, p_unsigned);
1114 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1115 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1116 PRINT_ATTRf(sample_regs_user, p_hex);
1117 PRINT_ATTRf(sample_stack_user, p_unsigned);
1118 PRINT_ATTRf(clockid, p_signed);
1119 PRINT_ATTRf(sample_regs_intr, p_hex);
1120 PRINT_ATTRf(aux_watermark, p_unsigned);
1125 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1126 void *priv __attribute__((unused)))
1128 return fprintf(fp, " %-32s %s\n", name, val);
1131 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1132 struct thread_map *threads)
1134 int cpu, thread, nthreads;
1135 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1137 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1139 if (evsel->system_wide)
1142 nthreads = threads->nr;
1144 if (evsel->fd == NULL &&
1145 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1149 flags |= PERF_FLAG_PID_CGROUP;
1150 pid = evsel->cgrp->fd;
1153 fallback_missing_features:
1154 if (perf_missing_features.clockid_wrong)
1155 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1156 if (perf_missing_features.clockid) {
1157 evsel->attr.use_clockid = 0;
1158 evsel->attr.clockid = 0;
1160 if (perf_missing_features.cloexec)
1161 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1162 if (perf_missing_features.mmap2)
1163 evsel->attr.mmap2 = 0;
1164 if (perf_missing_features.exclude_guest)
1165 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1167 if (perf_missing_features.sample_id_all)
1168 evsel->attr.sample_id_all = 0;
1171 fprintf(stderr, "%.60s\n", graph_dotted_line);
1172 fprintf(stderr, "perf_event_attr:\n");
1173 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1174 fprintf(stderr, "%.60s\n", graph_dotted_line);
1177 for (cpu = 0; cpu < cpus->nr; cpu++) {
1179 for (thread = 0; thread < nthreads; thread++) {
1182 if (!evsel->cgrp && !evsel->system_wide)
1183 pid = thread_map__pid(threads, thread);
1185 group_fd = get_group_fd(evsel, cpu, thread);
1187 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1188 pid, cpus->map[cpu], group_fd, flags);
1190 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1194 if (FD(evsel, cpu, thread) < 0) {
1196 pr_debug2("sys_perf_event_open failed, error %d\n",
1200 set_rlimit = NO_CHANGE;
1203 * If we succeeded but had to kill clockid, fail and
1204 * have perf_evsel__open_strerror() print us a nice
1207 if (perf_missing_features.clockid ||
1208 perf_missing_features.clockid_wrong) {
1219 * perf stat needs between 5 and 22 fds per CPU. When we run out
1220 * of them try to increase the limits.
1222 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1224 int old_errno = errno;
1226 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1227 if (set_rlimit == NO_CHANGE)
1228 l.rlim_cur = l.rlim_max;
1230 l.rlim_cur = l.rlim_max + 1000;
1231 l.rlim_max = l.rlim_cur;
1233 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1242 if (err != -EINVAL || cpu > 0 || thread > 0)
1246 * Must probe features in the order they were added to the
1247 * perf_event_attr interface.
1249 if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1250 perf_missing_features.clockid_wrong = true;
1251 goto fallback_missing_features;
1252 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1253 perf_missing_features.clockid = true;
1254 goto fallback_missing_features;
1255 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1256 perf_missing_features.cloexec = true;
1257 goto fallback_missing_features;
1258 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1259 perf_missing_features.mmap2 = true;
1260 goto fallback_missing_features;
1261 } else if (!perf_missing_features.exclude_guest &&
1262 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1263 perf_missing_features.exclude_guest = true;
1264 goto fallback_missing_features;
1265 } else if (!perf_missing_features.sample_id_all) {
1266 perf_missing_features.sample_id_all = true;
1267 goto retry_sample_id;
1272 while (--thread >= 0) {
1273 close(FD(evsel, cpu, thread));
1274 FD(evsel, cpu, thread) = -1;
1277 } while (--cpu >= 0);
1281 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1283 if (evsel->fd == NULL)
1286 perf_evsel__close_fd(evsel, ncpus, nthreads);
1287 perf_evsel__free_fd(evsel);
1299 struct thread_map map;
1301 } empty_thread_map = {
1306 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1307 struct thread_map *threads)
1310 /* Work around old compiler warnings about strict aliasing */
1311 cpus = &empty_cpu_map.map;
1314 if (threads == NULL)
1315 threads = &empty_thread_map.map;
1317 return __perf_evsel__open(evsel, cpus, threads);
1320 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1321 struct cpu_map *cpus)
1323 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1326 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1327 struct thread_map *threads)
1329 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1332 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1333 const union perf_event *event,
1334 struct perf_sample *sample)
1336 u64 type = evsel->attr.sample_type;
1337 const u64 *array = event->sample.array;
1338 bool swapped = evsel->needs_swap;
1341 array += ((event->header.size -
1342 sizeof(event->header)) / sizeof(u64)) - 1;
1344 if (type & PERF_SAMPLE_IDENTIFIER) {
1345 sample->id = *array;
1349 if (type & PERF_SAMPLE_CPU) {
1352 /* undo swap of u64, then swap on individual u32s */
1353 u.val64 = bswap_64(u.val64);
1354 u.val32[0] = bswap_32(u.val32[0]);
1357 sample->cpu = u.val32[0];
1361 if (type & PERF_SAMPLE_STREAM_ID) {
1362 sample->stream_id = *array;
1366 if (type & PERF_SAMPLE_ID) {
1367 sample->id = *array;
1371 if (type & PERF_SAMPLE_TIME) {
1372 sample->time = *array;
1376 if (type & PERF_SAMPLE_TID) {
1379 /* undo swap of u64, then swap on individual u32s */
1380 u.val64 = bswap_64(u.val64);
1381 u.val32[0] = bswap_32(u.val32[0]);
1382 u.val32[1] = bswap_32(u.val32[1]);
1385 sample->pid = u.val32[0];
1386 sample->tid = u.val32[1];
1393 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1396 return size > max_size || offset + size > endp;
1399 #define OVERFLOW_CHECK(offset, size, max_size) \
1401 if (overflow(endp, (max_size), (offset), (size))) \
1405 #define OVERFLOW_CHECK_u64(offset) \
1406 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1408 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1409 struct perf_sample *data)
1411 u64 type = evsel->attr.sample_type;
1412 bool swapped = evsel->needs_swap;
1414 u16 max_size = event->header.size;
1415 const void *endp = (void *)event + max_size;
1419 * used for cross-endian analysis. See git commit 65014ab3
1420 * for why this goofiness is needed.
1424 memset(data, 0, sizeof(*data));
1425 data->cpu = data->pid = data->tid = -1;
1426 data->stream_id = data->id = data->time = -1ULL;
1427 data->period = evsel->attr.sample_period;
1430 if (event->header.type != PERF_RECORD_SAMPLE) {
1431 if (!evsel->attr.sample_id_all)
1433 return perf_evsel__parse_id_sample(evsel, event, data);
1436 array = event->sample.array;
1439 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1440 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1441 * check the format does not go past the end of the event.
1443 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1447 if (type & PERF_SAMPLE_IDENTIFIER) {
1452 if (type & PERF_SAMPLE_IP) {
1457 if (type & PERF_SAMPLE_TID) {
1460 /* undo swap of u64, then swap on individual u32s */
1461 u.val64 = bswap_64(u.val64);
1462 u.val32[0] = bswap_32(u.val32[0]);
1463 u.val32[1] = bswap_32(u.val32[1]);
1466 data->pid = u.val32[0];
1467 data->tid = u.val32[1];
1471 if (type & PERF_SAMPLE_TIME) {
1472 data->time = *array;
1477 if (type & PERF_SAMPLE_ADDR) {
1478 data->addr = *array;
1482 if (type & PERF_SAMPLE_ID) {
1487 if (type & PERF_SAMPLE_STREAM_ID) {
1488 data->stream_id = *array;
1492 if (type & PERF_SAMPLE_CPU) {
1496 /* undo swap of u64, then swap on individual u32s */
1497 u.val64 = bswap_64(u.val64);
1498 u.val32[0] = bswap_32(u.val32[0]);
1501 data->cpu = u.val32[0];
1505 if (type & PERF_SAMPLE_PERIOD) {
1506 data->period = *array;
1510 if (type & PERF_SAMPLE_READ) {
1511 u64 read_format = evsel->attr.read_format;
1513 OVERFLOW_CHECK_u64(array);
1514 if (read_format & PERF_FORMAT_GROUP)
1515 data->read.group.nr = *array;
1517 data->read.one.value = *array;
1521 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1522 OVERFLOW_CHECK_u64(array);
1523 data->read.time_enabled = *array;
1527 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1528 OVERFLOW_CHECK_u64(array);
1529 data->read.time_running = *array;
1533 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1534 if (read_format & PERF_FORMAT_GROUP) {
1535 const u64 max_group_nr = UINT64_MAX /
1536 sizeof(struct sample_read_value);
1538 if (data->read.group.nr > max_group_nr)
1540 sz = data->read.group.nr *
1541 sizeof(struct sample_read_value);
1542 OVERFLOW_CHECK(array, sz, max_size);
1543 data->read.group.values =
1544 (struct sample_read_value *)array;
1545 array = (void *)array + sz;
1547 OVERFLOW_CHECK_u64(array);
1548 data->read.one.id = *array;
1553 if (type & PERF_SAMPLE_CALLCHAIN) {
1554 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1556 OVERFLOW_CHECK_u64(array);
1557 data->callchain = (struct ip_callchain *)array++;
1558 if (data->callchain->nr > max_callchain_nr)
1560 sz = data->callchain->nr * sizeof(u64);
1561 OVERFLOW_CHECK(array, sz, max_size);
1562 array = (void *)array + sz;
1565 if (type & PERF_SAMPLE_RAW) {
1566 OVERFLOW_CHECK_u64(array);
1568 if (WARN_ONCE(swapped,
1569 "Endianness of raw data not corrected!\n")) {
1570 /* undo swap of u64, then swap on individual u32s */
1571 u.val64 = bswap_64(u.val64);
1572 u.val32[0] = bswap_32(u.val32[0]);
1573 u.val32[1] = bswap_32(u.val32[1]);
1575 data->raw_size = u.val32[0];
1576 array = (void *)array + sizeof(u32);
1578 OVERFLOW_CHECK(array, data->raw_size, max_size);
1579 data->raw_data = (void *)array;
1580 array = (void *)array + data->raw_size;
1583 if (type & PERF_SAMPLE_BRANCH_STACK) {
1584 const u64 max_branch_nr = UINT64_MAX /
1585 sizeof(struct branch_entry);
1587 OVERFLOW_CHECK_u64(array);
1588 data->branch_stack = (struct branch_stack *)array++;
1590 if (data->branch_stack->nr > max_branch_nr)
1592 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1593 OVERFLOW_CHECK(array, sz, max_size);
1594 array = (void *)array + sz;
1597 if (type & PERF_SAMPLE_REGS_USER) {
1598 OVERFLOW_CHECK_u64(array);
1599 data->user_regs.abi = *array;
1602 if (data->user_regs.abi) {
1603 u64 mask = evsel->attr.sample_regs_user;
1605 sz = hweight_long(mask) * sizeof(u64);
1606 OVERFLOW_CHECK(array, sz, max_size);
1607 data->user_regs.mask = mask;
1608 data->user_regs.regs = (u64 *)array;
1609 array = (void *)array + sz;
1613 if (type & PERF_SAMPLE_STACK_USER) {
1614 OVERFLOW_CHECK_u64(array);
1617 data->user_stack.offset = ((char *)(array - 1)
1621 data->user_stack.size = 0;
1623 OVERFLOW_CHECK(array, sz, max_size);
1624 data->user_stack.data = (char *)array;
1625 array = (void *)array + sz;
1626 OVERFLOW_CHECK_u64(array);
1627 data->user_stack.size = *array++;
1628 if (WARN_ONCE(data->user_stack.size > sz,
1629 "user stack dump failure\n"))
1635 if (type & PERF_SAMPLE_WEIGHT) {
1636 OVERFLOW_CHECK_u64(array);
1637 data->weight = *array;
1641 data->data_src = PERF_MEM_DATA_SRC_NONE;
1642 if (type & PERF_SAMPLE_DATA_SRC) {
1643 OVERFLOW_CHECK_u64(array);
1644 data->data_src = *array;
1648 data->transaction = 0;
1649 if (type & PERF_SAMPLE_TRANSACTION) {
1650 OVERFLOW_CHECK_u64(array);
1651 data->transaction = *array;
1655 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1656 if (type & PERF_SAMPLE_REGS_INTR) {
1657 OVERFLOW_CHECK_u64(array);
1658 data->intr_regs.abi = *array;
1661 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1662 u64 mask = evsel->attr.sample_regs_intr;
1664 sz = hweight_long(mask) * sizeof(u64);
1665 OVERFLOW_CHECK(array, sz, max_size);
1666 data->intr_regs.mask = mask;
1667 data->intr_regs.regs = (u64 *)array;
1668 array = (void *)array + sz;
1675 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1678 size_t sz, result = sizeof(struct sample_event);
1680 if (type & PERF_SAMPLE_IDENTIFIER)
1681 result += sizeof(u64);
1683 if (type & PERF_SAMPLE_IP)
1684 result += sizeof(u64);
1686 if (type & PERF_SAMPLE_TID)
1687 result += sizeof(u64);
1689 if (type & PERF_SAMPLE_TIME)
1690 result += sizeof(u64);
1692 if (type & PERF_SAMPLE_ADDR)
1693 result += sizeof(u64);
1695 if (type & PERF_SAMPLE_ID)
1696 result += sizeof(u64);
1698 if (type & PERF_SAMPLE_STREAM_ID)
1699 result += sizeof(u64);
1701 if (type & PERF_SAMPLE_CPU)
1702 result += sizeof(u64);
1704 if (type & PERF_SAMPLE_PERIOD)
1705 result += sizeof(u64);
1707 if (type & PERF_SAMPLE_READ) {
1708 result += sizeof(u64);
1709 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1710 result += sizeof(u64);
1711 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1712 result += sizeof(u64);
1713 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1714 if (read_format & PERF_FORMAT_GROUP) {
1715 sz = sample->read.group.nr *
1716 sizeof(struct sample_read_value);
1719 result += sizeof(u64);
1723 if (type & PERF_SAMPLE_CALLCHAIN) {
1724 sz = (sample->callchain->nr + 1) * sizeof(u64);
1728 if (type & PERF_SAMPLE_RAW) {
1729 result += sizeof(u32);
1730 result += sample->raw_size;
1733 if (type & PERF_SAMPLE_BRANCH_STACK) {
1734 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1739 if (type & PERF_SAMPLE_REGS_USER) {
1740 if (sample->user_regs.abi) {
1741 result += sizeof(u64);
1742 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1745 result += sizeof(u64);
1749 if (type & PERF_SAMPLE_STACK_USER) {
1750 sz = sample->user_stack.size;
1751 result += sizeof(u64);
1754 result += sizeof(u64);
1758 if (type & PERF_SAMPLE_WEIGHT)
1759 result += sizeof(u64);
1761 if (type & PERF_SAMPLE_DATA_SRC)
1762 result += sizeof(u64);
1764 if (type & PERF_SAMPLE_TRANSACTION)
1765 result += sizeof(u64);
1767 if (type & PERF_SAMPLE_REGS_INTR) {
1768 if (sample->intr_regs.abi) {
1769 result += sizeof(u64);
1770 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1773 result += sizeof(u64);
1780 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1782 const struct perf_sample *sample,
1788 * used for cross-endian analysis. See git commit 65014ab3
1789 * for why this goofiness is needed.
1793 array = event->sample.array;
1795 if (type & PERF_SAMPLE_IDENTIFIER) {
1796 *array = sample->id;
1800 if (type & PERF_SAMPLE_IP) {
1801 *array = sample->ip;
1805 if (type & PERF_SAMPLE_TID) {
1806 u.val32[0] = sample->pid;
1807 u.val32[1] = sample->tid;
1810 * Inverse of what is done in perf_evsel__parse_sample
1812 u.val32[0] = bswap_32(u.val32[0]);
1813 u.val32[1] = bswap_32(u.val32[1]);
1814 u.val64 = bswap_64(u.val64);
1821 if (type & PERF_SAMPLE_TIME) {
1822 *array = sample->time;
1826 if (type & PERF_SAMPLE_ADDR) {
1827 *array = sample->addr;
1831 if (type & PERF_SAMPLE_ID) {
1832 *array = sample->id;
1836 if (type & PERF_SAMPLE_STREAM_ID) {
1837 *array = sample->stream_id;
1841 if (type & PERF_SAMPLE_CPU) {
1842 u.val32[0] = sample->cpu;
1845 * Inverse of what is done in perf_evsel__parse_sample
1847 u.val32[0] = bswap_32(u.val32[0]);
1848 u.val64 = bswap_64(u.val64);
1854 if (type & PERF_SAMPLE_PERIOD) {
1855 *array = sample->period;
1859 if (type & PERF_SAMPLE_READ) {
1860 if (read_format & PERF_FORMAT_GROUP)
1861 *array = sample->read.group.nr;
1863 *array = sample->read.one.value;
1866 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1867 *array = sample->read.time_enabled;
1871 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1872 *array = sample->read.time_running;
1876 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1877 if (read_format & PERF_FORMAT_GROUP) {
1878 sz = sample->read.group.nr *
1879 sizeof(struct sample_read_value);
1880 memcpy(array, sample->read.group.values, sz);
1881 array = (void *)array + sz;
1883 *array = sample->read.one.id;
1888 if (type & PERF_SAMPLE_CALLCHAIN) {
1889 sz = (sample->callchain->nr + 1) * sizeof(u64);
1890 memcpy(array, sample->callchain, sz);
1891 array = (void *)array + sz;
1894 if (type & PERF_SAMPLE_RAW) {
1895 u.val32[0] = sample->raw_size;
1896 if (WARN_ONCE(swapped,
1897 "Endianness of raw data not corrected!\n")) {
1899 * Inverse of what is done in perf_evsel__parse_sample
1901 u.val32[0] = bswap_32(u.val32[0]);
1902 u.val32[1] = bswap_32(u.val32[1]);
1903 u.val64 = bswap_64(u.val64);
1906 array = (void *)array + sizeof(u32);
1908 memcpy(array, sample->raw_data, sample->raw_size);
1909 array = (void *)array + sample->raw_size;
1912 if (type & PERF_SAMPLE_BRANCH_STACK) {
1913 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1915 memcpy(array, sample->branch_stack, sz);
1916 array = (void *)array + sz;
1919 if (type & PERF_SAMPLE_REGS_USER) {
1920 if (sample->user_regs.abi) {
1921 *array++ = sample->user_regs.abi;
1922 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1923 memcpy(array, sample->user_regs.regs, sz);
1924 array = (void *)array + sz;
1930 if (type & PERF_SAMPLE_STACK_USER) {
1931 sz = sample->user_stack.size;
1934 memcpy(array, sample->user_stack.data, sz);
1935 array = (void *)array + sz;
1940 if (type & PERF_SAMPLE_WEIGHT) {
1941 *array = sample->weight;
1945 if (type & PERF_SAMPLE_DATA_SRC) {
1946 *array = sample->data_src;
1950 if (type & PERF_SAMPLE_TRANSACTION) {
1951 *array = sample->transaction;
1955 if (type & PERF_SAMPLE_REGS_INTR) {
1956 if (sample->intr_regs.abi) {
1957 *array++ = sample->intr_regs.abi;
1958 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1959 memcpy(array, sample->intr_regs.regs, sz);
1960 array = (void *)array + sz;
1969 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1971 return pevent_find_field(evsel->tp_format, name);
1974 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1977 struct format_field *field = perf_evsel__field(evsel, name);
1983 offset = field->offset;
1985 if (field->flags & FIELD_IS_DYNAMIC) {
1986 offset = *(int *)(sample->raw_data + field->offset);
1990 return sample->raw_data + offset;
1993 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1996 struct format_field *field = perf_evsel__field(evsel, name);
2003 ptr = sample->raw_data + field->offset;
2005 switch (field->size) {
2009 value = *(u16 *)ptr;
2012 value = *(u32 *)ptr;
2015 memcpy(&value, ptr, sizeof(u64));
2021 if (!evsel->needs_swap)
2024 switch (field->size) {
2026 return bswap_16(value);
2028 return bswap_32(value);
2030 return bswap_64(value);
2038 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2044 ret += fprintf(fp, ",");
2046 ret += fprintf(fp, ":");
2050 va_start(args, fmt);
2051 ret += vfprintf(fp, fmt, args);
2056 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2058 return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2061 int perf_evsel__fprintf(struct perf_evsel *evsel,
2062 struct perf_attr_details *details, FILE *fp)
2067 if (details->event_group) {
2068 struct perf_evsel *pos;
2070 if (!perf_evsel__is_group_leader(evsel))
2073 if (evsel->nr_members > 1)
2074 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2076 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2077 for_each_group_member(pos, evsel)
2078 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2080 if (evsel->nr_members > 1)
2081 printed += fprintf(fp, "}");
2085 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2087 if (details->verbose) {
2088 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2089 __print_attr__fprintf, &first);
2090 } else if (details->freq) {
2091 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2092 (u64)evsel->attr.sample_freq);
2099 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2100 char *msg, size_t msgsize)
2102 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2103 evsel->attr.type == PERF_TYPE_HARDWARE &&
2104 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2106 * If it's cycles then fall back to hrtimer based
2107 * cpu-clock-tick sw counter, which is always available even if
2110 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2113 scnprintf(msg, msgsize, "%s",
2114 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2116 evsel->attr.type = PERF_TYPE_SOFTWARE;
2117 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2119 zfree(&evsel->name);
2126 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2127 int err, char *msg, size_t size)
2129 char sbuf[STRERR_BUFSIZE];
2134 return scnprintf(msg, size,
2135 "You may not have permission to collect %sstats.\n"
2136 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2137 " -1 - Not paranoid at all\n"
2138 " 0 - Disallow raw tracepoint access for unpriv\n"
2139 " 1 - Disallow cpu events for unpriv\n"
2140 " 2 - Disallow kernel profiling for unpriv",
2141 target->system_wide ? "system-wide " : "");
2143 return scnprintf(msg, size, "The %s event is not supported.",
2144 perf_evsel__name(evsel));
2146 return scnprintf(msg, size, "%s",
2147 "Too many events are opened.\n"
2148 "Probably the maximum number of open file descriptors has been reached.\n"
2149 "Hint: Try again after reducing the number of events.\n"
2150 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2152 if (target->cpu_list)
2153 return scnprintf(msg, size, "%s",
2154 "No such device - did you specify an out-of-range profile CPU?\n");
2157 if (evsel->attr.precise_ip)
2158 return scnprintf(msg, size, "%s",
2159 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2160 #if defined(__i386__) || defined(__x86_64__)
2161 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2162 return scnprintf(msg, size, "%s",
2163 "No hardware sampling interrupt available.\n"
2164 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2168 if (find_process("oprofiled"))
2169 return scnprintf(msg, size,
2170 "The PMU counters are busy/taken by another profiler.\n"
2171 "We found oprofile daemon running, please stop it and try again.");
2174 if (perf_missing_features.clockid)
2175 return scnprintf(msg, size, "clockid feature not supported.");
2176 if (perf_missing_features.clockid_wrong)
2177 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2183 return scnprintf(msg, size,
2184 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2185 "/bin/dmesg may provide additional information.\n"
2186 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2187 err, strerror_r(err, sbuf, sizeof(sbuf)),
2188 perf_evsel__name(evsel));