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/tracing_path.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <linux/err.h>
17 #include <sys/resource.h>
19 #include "callchain.h"
25 #include "thread_map.h"
27 #include "perf_regs.h"
29 #include "trace-event.h"
40 } perf_missing_features;
42 static clockid_t clockid;
44 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
49 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
55 int (*init)(struct perf_evsel *evsel);
56 void (*fini)(struct perf_evsel *evsel);
57 } perf_evsel__object = {
58 .size = sizeof(struct perf_evsel),
59 .init = perf_evsel__no_extra_init,
60 .fini = perf_evsel__no_extra_fini,
63 int perf_evsel__object_config(size_t object_size,
64 int (*init)(struct perf_evsel *evsel),
65 void (*fini)(struct perf_evsel *evsel))
71 if (perf_evsel__object.size > object_size)
74 perf_evsel__object.size = object_size;
78 perf_evsel__object.init = init;
81 perf_evsel__object.fini = fini;
86 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
88 int __perf_evsel__sample_size(u64 sample_type)
90 u64 mask = sample_type & PERF_SAMPLE_MASK;
94 for (i = 0; i < 64; i++) {
95 if (mask & (1ULL << i))
105 * __perf_evsel__calc_id_pos - calculate id_pos.
106 * @sample_type: sample type
108 * This function returns the position of the event id (PERF_SAMPLE_ID or
109 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
112 static int __perf_evsel__calc_id_pos(u64 sample_type)
116 if (sample_type & PERF_SAMPLE_IDENTIFIER)
119 if (!(sample_type & PERF_SAMPLE_ID))
122 if (sample_type & PERF_SAMPLE_IP)
125 if (sample_type & PERF_SAMPLE_TID)
128 if (sample_type & PERF_SAMPLE_TIME)
131 if (sample_type & PERF_SAMPLE_ADDR)
138 * __perf_evsel__calc_is_pos - calculate is_pos.
139 * @sample_type: sample type
141 * This function returns the position (counting backwards) of the event id
142 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
143 * sample_id_all is used there is an id sample appended to non-sample events.
145 static int __perf_evsel__calc_is_pos(u64 sample_type)
149 if (sample_type & PERF_SAMPLE_IDENTIFIER)
152 if (!(sample_type & PERF_SAMPLE_ID))
155 if (sample_type & PERF_SAMPLE_CPU)
158 if (sample_type & PERF_SAMPLE_STREAM_ID)
164 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
166 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
167 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
170 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
171 enum perf_event_sample_format bit)
173 if (!(evsel->attr.sample_type & bit)) {
174 evsel->attr.sample_type |= bit;
175 evsel->sample_size += sizeof(u64);
176 perf_evsel__calc_id_pos(evsel);
180 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
181 enum perf_event_sample_format bit)
183 if (evsel->attr.sample_type & bit) {
184 evsel->attr.sample_type &= ~bit;
185 evsel->sample_size -= sizeof(u64);
186 perf_evsel__calc_id_pos(evsel);
190 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
191 bool can_sample_identifier)
193 if (can_sample_identifier) {
194 perf_evsel__reset_sample_bit(evsel, ID);
195 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
197 perf_evsel__set_sample_bit(evsel, ID);
199 evsel->attr.read_format |= PERF_FORMAT_ID;
202 void perf_evsel__init(struct perf_evsel *evsel,
203 struct perf_event_attr *attr, int idx)
206 evsel->tracking = !idx;
208 evsel->leader = evsel;
211 evsel->evlist = NULL;
213 INIT_LIST_HEAD(&evsel->node);
214 INIT_LIST_HEAD(&evsel->config_terms);
215 perf_evsel__object.init(evsel);
216 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
217 perf_evsel__calc_id_pos(evsel);
218 evsel->cmdline_group_boundary = false;
221 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
223 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
226 perf_evsel__init(evsel, attr, idx);
228 if (perf_evsel__is_bpf_output(evsel)) {
229 evsel->attr.sample_type |= PERF_SAMPLE_RAW;
230 evsel->attr.sample_period = 1;
237 * Returns pointer with encoded error via <linux/err.h> interface.
239 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
241 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
247 struct perf_event_attr attr = {
248 .type = PERF_TYPE_TRACEPOINT,
249 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
250 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
253 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
256 evsel->tp_format = trace_event__tp_format(sys, name);
257 if (IS_ERR(evsel->tp_format)) {
258 err = PTR_ERR(evsel->tp_format);
262 event_attr_init(&attr);
263 attr.config = evsel->tp_format->id;
264 attr.sample_period = 1;
265 perf_evsel__init(evsel, &attr, idx);
277 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
285 "stalled-cycles-frontend",
286 "stalled-cycles-backend",
290 static const char *__perf_evsel__hw_name(u64 config)
292 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
293 return perf_evsel__hw_names[config];
295 return "unknown-hardware";
298 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
300 int colon = 0, r = 0;
301 struct perf_event_attr *attr = &evsel->attr;
302 bool exclude_guest_default = false;
304 #define MOD_PRINT(context, mod) do { \
305 if (!attr->exclude_##context) { \
306 if (!colon) colon = ++r; \
307 r += scnprintf(bf + r, size - r, "%c", mod); \
310 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
311 MOD_PRINT(kernel, 'k');
312 MOD_PRINT(user, 'u');
314 exclude_guest_default = true;
317 if (attr->precise_ip) {
320 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
321 exclude_guest_default = true;
324 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
325 MOD_PRINT(host, 'H');
326 MOD_PRINT(guest, 'G');
334 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
336 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
337 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
340 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
353 static const char *__perf_evsel__sw_name(u64 config)
355 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
356 return perf_evsel__sw_names[config];
357 return "unknown-software";
360 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
362 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
363 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
366 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
370 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
372 if (type & HW_BREAKPOINT_R)
373 r += scnprintf(bf + r, size - r, "r");
375 if (type & HW_BREAKPOINT_W)
376 r += scnprintf(bf + r, size - r, "w");
378 if (type & HW_BREAKPOINT_X)
379 r += scnprintf(bf + r, size - r, "x");
384 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
386 struct perf_event_attr *attr = &evsel->attr;
387 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
388 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
391 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
392 [PERF_EVSEL__MAX_ALIASES] = {
393 { "L1-dcache", "l1-d", "l1d", "L1-data", },
394 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
396 { "dTLB", "d-tlb", "Data-TLB", },
397 { "iTLB", "i-tlb", "Instruction-TLB", },
398 { "branch", "branches", "bpu", "btb", "bpc", },
402 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
403 [PERF_EVSEL__MAX_ALIASES] = {
404 { "load", "loads", "read", },
405 { "store", "stores", "write", },
406 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
409 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
410 [PERF_EVSEL__MAX_ALIASES] = {
411 { "refs", "Reference", "ops", "access", },
412 { "misses", "miss", },
415 #define C(x) PERF_COUNT_HW_CACHE_##x
416 #define CACHE_READ (1 << C(OP_READ))
417 #define CACHE_WRITE (1 << C(OP_WRITE))
418 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
419 #define COP(x) (1 << x)
422 * cache operartion stat
423 * L1I : Read and prefetch only
424 * ITLB and BPU : Read-only
426 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
427 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
428 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
429 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
430 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
431 [C(ITLB)] = (CACHE_READ),
432 [C(BPU)] = (CACHE_READ),
433 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
436 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
438 if (perf_evsel__hw_cache_stat[type] & COP(op))
439 return true; /* valid */
441 return false; /* invalid */
444 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
445 char *bf, size_t size)
448 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
449 perf_evsel__hw_cache_op[op][0],
450 perf_evsel__hw_cache_result[result][0]);
453 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
454 perf_evsel__hw_cache_op[op][1]);
457 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
459 u8 op, result, type = (config >> 0) & 0xff;
460 const char *err = "unknown-ext-hardware-cache-type";
462 if (type > PERF_COUNT_HW_CACHE_MAX)
465 op = (config >> 8) & 0xff;
466 err = "unknown-ext-hardware-cache-op";
467 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
470 result = (config >> 16) & 0xff;
471 err = "unknown-ext-hardware-cache-result";
472 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
475 err = "invalid-cache";
476 if (!perf_evsel__is_cache_op_valid(type, op))
479 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
481 return scnprintf(bf, size, "%s", err);
484 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
486 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
487 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
490 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
492 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
493 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
496 const char *perf_evsel__name(struct perf_evsel *evsel)
503 switch (evsel->attr.type) {
505 perf_evsel__raw_name(evsel, bf, sizeof(bf));
508 case PERF_TYPE_HARDWARE:
509 perf_evsel__hw_name(evsel, bf, sizeof(bf));
512 case PERF_TYPE_HW_CACHE:
513 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
516 case PERF_TYPE_SOFTWARE:
517 perf_evsel__sw_name(evsel, bf, sizeof(bf));
520 case PERF_TYPE_TRACEPOINT:
521 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
524 case PERF_TYPE_BREAKPOINT:
525 perf_evsel__bp_name(evsel, bf, sizeof(bf));
529 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
534 evsel->name = strdup(bf);
536 return evsel->name ?: "unknown";
539 const char *perf_evsel__group_name(struct perf_evsel *evsel)
541 return evsel->group_name ?: "anon group";
544 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
547 struct perf_evsel *pos;
548 const char *group_name = perf_evsel__group_name(evsel);
550 ret = scnprintf(buf, size, "%s", group_name);
552 ret += scnprintf(buf + ret, size - ret, " { %s",
553 perf_evsel__name(evsel));
555 for_each_group_member(pos, evsel)
556 ret += scnprintf(buf + ret, size - ret, ", %s",
557 perf_evsel__name(pos));
559 ret += scnprintf(buf + ret, size - ret, " }");
565 perf_evsel__config_callgraph(struct perf_evsel *evsel,
566 struct record_opts *opts,
567 struct callchain_param *param)
569 bool function = perf_evsel__is_function_event(evsel);
570 struct perf_event_attr *attr = &evsel->attr;
572 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
574 if (param->record_mode == CALLCHAIN_LBR) {
575 if (!opts->branch_stack) {
576 if (attr->exclude_user) {
577 pr_warning("LBR callstack option is only available "
578 "to get user callchain information. "
579 "Falling back to framepointers.\n");
581 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
582 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
583 PERF_SAMPLE_BRANCH_CALL_STACK |
584 PERF_SAMPLE_BRANCH_NO_CYCLES |
585 PERF_SAMPLE_BRANCH_NO_FLAGS;
588 pr_warning("Cannot use LBR callstack with branch stack. "
589 "Falling back to framepointers.\n");
592 if (param->record_mode == CALLCHAIN_DWARF) {
594 perf_evsel__set_sample_bit(evsel, REGS_USER);
595 perf_evsel__set_sample_bit(evsel, STACK_USER);
596 attr->sample_regs_user = PERF_REGS_MASK;
597 attr->sample_stack_user = param->dump_size;
598 attr->exclude_callchain_user = 1;
600 pr_info("Cannot use DWARF unwind for function trace event,"
601 " falling back to framepointers.\n");
606 pr_info("Disabling user space callchains for function trace event.\n");
607 attr->exclude_callchain_user = 1;
612 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
613 struct callchain_param *param)
615 struct perf_event_attr *attr = &evsel->attr;
617 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
618 if (param->record_mode == CALLCHAIN_LBR) {
619 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
620 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
621 PERF_SAMPLE_BRANCH_CALL_STACK);
623 if (param->record_mode == CALLCHAIN_DWARF) {
624 perf_evsel__reset_sample_bit(evsel, REGS_USER);
625 perf_evsel__reset_sample_bit(evsel, STACK_USER);
629 static void apply_config_terms(struct perf_evsel *evsel,
630 struct record_opts *opts)
632 struct perf_evsel_config_term *term;
633 struct list_head *config_terms = &evsel->config_terms;
634 struct perf_event_attr *attr = &evsel->attr;
635 struct callchain_param param;
637 char *callgraph_buf = NULL;
639 /* callgraph default */
640 param.record_mode = callchain_param.record_mode;
642 list_for_each_entry(term, config_terms, list) {
643 switch (term->type) {
644 case PERF_EVSEL__CONFIG_TERM_PERIOD:
645 attr->sample_period = term->val.period;
648 case PERF_EVSEL__CONFIG_TERM_FREQ:
649 attr->sample_freq = term->val.freq;
652 case PERF_EVSEL__CONFIG_TERM_TIME:
654 perf_evsel__set_sample_bit(evsel, TIME);
656 perf_evsel__reset_sample_bit(evsel, TIME);
658 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
659 callgraph_buf = term->val.callgraph;
661 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
662 dump_size = term->val.stack_user;
664 case PERF_EVSEL__CONFIG_TERM_INHERIT:
666 * attr->inherit should has already been set by
667 * perf_evsel__config. If user explicitly set
668 * inherit using config terms, override global
669 * opt->no_inherit setting.
671 attr->inherit = term->val.inherit ? 1 : 0;
678 /* User explicitly set per-event callgraph, clear the old setting and reset. */
679 if ((callgraph_buf != NULL) || (dump_size > 0)) {
681 /* parse callgraph parameters */
682 if (callgraph_buf != NULL) {
683 if (!strcmp(callgraph_buf, "no")) {
684 param.enabled = false;
685 param.record_mode = CALLCHAIN_NONE;
687 param.enabled = true;
688 if (parse_callchain_record(callgraph_buf, ¶m)) {
689 pr_err("per-event callgraph setting for %s failed. "
690 "Apply callgraph global setting for it\n",
697 dump_size = round_up(dump_size, sizeof(u64));
698 param.dump_size = dump_size;
701 /* If global callgraph set, clear it */
702 if (callchain_param.enabled)
703 perf_evsel__reset_callgraph(evsel, &callchain_param);
705 /* set perf-event callgraph */
707 perf_evsel__config_callgraph(evsel, opts, ¶m);
712 * The enable_on_exec/disabled value strategy:
714 * 1) For any type of traced program:
715 * - all independent events and group leaders are disabled
716 * - all group members are enabled
718 * Group members are ruled by group leaders. They need to
719 * be enabled, because the group scheduling relies on that.
721 * 2) For traced programs executed by perf:
722 * - all independent events and group leaders have
724 * - we don't specifically enable or disable any event during
727 * Independent events and group leaders are initially disabled
728 * and get enabled by exec. Group members are ruled by group
729 * leaders as stated in 1).
731 * 3) For traced programs attached by perf (pid/tid):
732 * - we specifically enable or disable all events during
735 * When attaching events to already running traced we
736 * enable/disable events specifically, as there's no
737 * initial traced exec call.
739 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
741 struct perf_evsel *leader = evsel->leader;
742 struct perf_event_attr *attr = &evsel->attr;
743 int track = evsel->tracking;
744 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
746 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
747 attr->inherit = !opts->no_inherit;
749 perf_evsel__set_sample_bit(evsel, IP);
750 perf_evsel__set_sample_bit(evsel, TID);
752 if (evsel->sample_read) {
753 perf_evsel__set_sample_bit(evsel, READ);
756 * We need ID even in case of single event, because
757 * PERF_SAMPLE_READ process ID specific data.
759 perf_evsel__set_sample_id(evsel, false);
762 * Apply group format only if we belong to group
763 * with more than one members.
765 if (leader->nr_members > 1) {
766 attr->read_format |= PERF_FORMAT_GROUP;
772 * We default some events to have a default interval. But keep
773 * it a weak assumption overridable by the user.
775 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
776 opts->user_interval != ULLONG_MAX)) {
778 perf_evsel__set_sample_bit(evsel, PERIOD);
780 attr->sample_freq = opts->freq;
782 attr->sample_period = opts->default_interval;
787 * Disable sampling for all group members other
788 * than leader in case leader 'leads' the sampling.
790 if ((leader != evsel) && leader->sample_read) {
791 attr->sample_freq = 0;
792 attr->sample_period = 0;
795 if (opts->no_samples)
796 attr->sample_freq = 0;
798 if (opts->inherit_stat)
799 attr->inherit_stat = 1;
801 if (opts->sample_address) {
802 perf_evsel__set_sample_bit(evsel, ADDR);
803 attr->mmap_data = track;
807 * We don't allow user space callchains for function trace
808 * event, due to issues with page faults while tracing page
809 * fault handler and its overall trickiness nature.
811 if (perf_evsel__is_function_event(evsel))
812 evsel->attr.exclude_callchain_user = 1;
814 if (callchain_param.enabled && !evsel->no_aux_samples)
815 perf_evsel__config_callgraph(evsel, opts, &callchain_param);
817 if (opts->sample_intr_regs) {
818 attr->sample_regs_intr = opts->sample_intr_regs;
819 perf_evsel__set_sample_bit(evsel, REGS_INTR);
822 if (target__has_cpu(&opts->target))
823 perf_evsel__set_sample_bit(evsel, CPU);
826 perf_evsel__set_sample_bit(evsel, PERIOD);
829 * When the user explicitely disabled time don't force it here.
831 if (opts->sample_time &&
832 (!perf_missing_features.sample_id_all &&
833 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
834 opts->sample_time_set)))
835 perf_evsel__set_sample_bit(evsel, TIME);
837 if (opts->raw_samples && !evsel->no_aux_samples) {
838 perf_evsel__set_sample_bit(evsel, TIME);
839 perf_evsel__set_sample_bit(evsel, RAW);
840 perf_evsel__set_sample_bit(evsel, CPU);
843 if (opts->sample_address)
844 perf_evsel__set_sample_bit(evsel, DATA_SRC);
846 if (opts->no_buffering) {
848 attr->wakeup_events = 1;
850 if (opts->branch_stack && !evsel->no_aux_samples) {
851 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
852 attr->branch_sample_type = opts->branch_stack;
855 if (opts->sample_weight)
856 perf_evsel__set_sample_bit(evsel, WEIGHT);
860 attr->mmap2 = track && !perf_missing_features.mmap2;
863 if (opts->record_switch_events)
864 attr->context_switch = track;
866 if (opts->sample_transaction)
867 perf_evsel__set_sample_bit(evsel, TRANSACTION);
869 if (opts->running_time) {
870 evsel->attr.read_format |=
871 PERF_FORMAT_TOTAL_TIME_ENABLED |
872 PERF_FORMAT_TOTAL_TIME_RUNNING;
876 * XXX see the function comment above
878 * Disabling only independent events or group leaders,
879 * keeping group members enabled.
881 if (perf_evsel__is_group_leader(evsel))
885 * Setting enable_on_exec for independent events and
886 * group leaders for traced executed by perf.
888 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
889 !opts->initial_delay)
890 attr->enable_on_exec = 1;
892 if (evsel->immediate) {
894 attr->enable_on_exec = 0;
897 clockid = opts->clockid;
898 if (opts->use_clockid) {
899 attr->use_clockid = 1;
900 attr->clockid = opts->clockid;
903 if (evsel->precise_max)
904 perf_event_attr__set_max_precise_ip(attr);
906 if (opts->all_user) {
907 attr->exclude_kernel = 1;
908 attr->exclude_user = 0;
911 if (opts->all_kernel) {
912 attr->exclude_kernel = 0;
913 attr->exclude_user = 1;
917 * Apply event specific term settings,
918 * it overloads any global configuration.
920 apply_config_terms(evsel, opts);
923 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
927 if (evsel->system_wide)
930 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
933 for (cpu = 0; cpu < ncpus; cpu++) {
934 for (thread = 0; thread < nthreads; thread++) {
935 FD(evsel, cpu, thread) = -1;
940 return evsel->fd != NULL ? 0 : -ENOMEM;
943 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
948 if (evsel->system_wide)
951 for (cpu = 0; cpu < ncpus; cpu++) {
952 for (thread = 0; thread < nthreads; thread++) {
953 int fd = FD(evsel, cpu, thread),
954 err = ioctl(fd, ioc, arg);
964 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
967 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
968 PERF_EVENT_IOC_SET_FILTER,
972 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
974 char *new_filter = strdup(filter);
976 if (new_filter != NULL) {
978 evsel->filter = new_filter;
985 int perf_evsel__append_filter(struct perf_evsel *evsel,
986 const char *op, const char *filter)
990 if (evsel->filter == NULL)
991 return perf_evsel__set_filter(evsel, filter);
993 if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
995 evsel->filter = new_filter;
1002 int perf_evsel__enable(struct perf_evsel *evsel)
1004 int nthreads = thread_map__nr(evsel->threads);
1005 int ncpus = cpu_map__nr(evsel->cpus);
1007 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1008 PERF_EVENT_IOC_ENABLE,
1012 int perf_evsel__disable(struct perf_evsel *evsel)
1014 int nthreads = thread_map__nr(evsel->threads);
1015 int ncpus = cpu_map__nr(evsel->cpus);
1017 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1018 PERF_EVENT_IOC_DISABLE,
1022 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1024 if (ncpus == 0 || nthreads == 0)
1027 if (evsel->system_wide)
1030 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1031 if (evsel->sample_id == NULL)
1034 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1035 if (evsel->id == NULL) {
1036 xyarray__delete(evsel->sample_id);
1037 evsel->sample_id = NULL;
1044 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1046 xyarray__delete(evsel->fd);
1050 static void perf_evsel__free_id(struct perf_evsel *evsel)
1052 xyarray__delete(evsel->sample_id);
1053 evsel->sample_id = NULL;
1057 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1059 struct perf_evsel_config_term *term, *h;
1061 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1062 list_del(&term->list);
1067 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1071 if (evsel->system_wide)
1074 for (cpu = 0; cpu < ncpus; cpu++)
1075 for (thread = 0; thread < nthreads; ++thread) {
1076 close(FD(evsel, cpu, thread));
1077 FD(evsel, cpu, thread) = -1;
1081 void perf_evsel__exit(struct perf_evsel *evsel)
1083 assert(list_empty(&evsel->node));
1084 assert(evsel->evlist == NULL);
1085 perf_evsel__free_fd(evsel);
1086 perf_evsel__free_id(evsel);
1087 perf_evsel__free_config_terms(evsel);
1088 close_cgroup(evsel->cgrp);
1089 cpu_map__put(evsel->cpus);
1090 cpu_map__put(evsel->own_cpus);
1091 thread_map__put(evsel->threads);
1092 zfree(&evsel->group_name);
1093 zfree(&evsel->name);
1094 perf_evsel__object.fini(evsel);
1097 void perf_evsel__delete(struct perf_evsel *evsel)
1099 perf_evsel__exit(evsel);
1103 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1104 struct perf_counts_values *count)
1106 struct perf_counts_values tmp;
1108 if (!evsel->prev_raw_counts)
1112 tmp = evsel->prev_raw_counts->aggr;
1113 evsel->prev_raw_counts->aggr = *count;
1115 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1116 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1119 count->val = count->val - tmp.val;
1120 count->ena = count->ena - tmp.ena;
1121 count->run = count->run - tmp.run;
1124 void perf_counts_values__scale(struct perf_counts_values *count,
1125 bool scale, s8 *pscaled)
1130 if (count->run == 0) {
1133 } else if (count->run < count->ena) {
1135 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1138 count->ena = count->run = 0;
1144 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1145 struct perf_counts_values *count)
1147 memset(count, 0, sizeof(*count));
1149 if (FD(evsel, cpu, thread) < 0)
1152 if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1158 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1159 int cpu, int thread, bool scale)
1161 struct perf_counts_values count;
1162 size_t nv = scale ? 3 : 1;
1164 if (FD(evsel, cpu, thread) < 0)
1167 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1170 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1173 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1174 perf_counts_values__scale(&count, scale, NULL);
1175 *perf_counts(evsel->counts, cpu, thread) = count;
1179 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1181 struct perf_evsel *leader = evsel->leader;
1184 if (perf_evsel__is_group_leader(evsel))
1188 * Leader must be already processed/open,
1189 * if not it's a bug.
1191 BUG_ON(!leader->fd);
1193 fd = FD(leader, cpu, thread);
1204 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1206 bool first_bit = true;
1210 if (value & bits[i].bit) {
1211 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1214 } while (bits[++i].name != NULL);
1217 static void __p_sample_type(char *buf, size_t size, u64 value)
1219 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1220 struct bit_names bits[] = {
1221 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1222 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1223 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1224 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1225 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1230 __p_bits(buf, size, value, bits);
1233 static void __p_read_format(char *buf, size_t size, u64 value)
1235 #define bit_name(n) { PERF_FORMAT_##n, #n }
1236 struct bit_names bits[] = {
1237 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1238 bit_name(ID), bit_name(GROUP),
1242 __p_bits(buf, size, value, bits);
1245 #define BUF_SIZE 1024
1247 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1248 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1249 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1250 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1251 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1253 #define PRINT_ATTRn(_n, _f, _p) \
1257 ret += attr__fprintf(fp, _n, buf, priv);\
1261 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1263 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1264 attr__fprintf_f attr__fprintf, void *priv)
1269 PRINT_ATTRf(type, p_unsigned);
1270 PRINT_ATTRf(size, p_unsigned);
1271 PRINT_ATTRf(config, p_hex);
1272 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1273 PRINT_ATTRf(sample_type, p_sample_type);
1274 PRINT_ATTRf(read_format, p_read_format);
1276 PRINT_ATTRf(disabled, p_unsigned);
1277 PRINT_ATTRf(inherit, p_unsigned);
1278 PRINT_ATTRf(pinned, p_unsigned);
1279 PRINT_ATTRf(exclusive, p_unsigned);
1280 PRINT_ATTRf(exclude_user, p_unsigned);
1281 PRINT_ATTRf(exclude_kernel, p_unsigned);
1282 PRINT_ATTRf(exclude_hv, p_unsigned);
1283 PRINT_ATTRf(exclude_idle, p_unsigned);
1284 PRINT_ATTRf(mmap, p_unsigned);
1285 PRINT_ATTRf(comm, p_unsigned);
1286 PRINT_ATTRf(freq, p_unsigned);
1287 PRINT_ATTRf(inherit_stat, p_unsigned);
1288 PRINT_ATTRf(enable_on_exec, p_unsigned);
1289 PRINT_ATTRf(task, p_unsigned);
1290 PRINT_ATTRf(watermark, p_unsigned);
1291 PRINT_ATTRf(precise_ip, p_unsigned);
1292 PRINT_ATTRf(mmap_data, p_unsigned);
1293 PRINT_ATTRf(sample_id_all, p_unsigned);
1294 PRINT_ATTRf(exclude_host, p_unsigned);
1295 PRINT_ATTRf(exclude_guest, p_unsigned);
1296 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1297 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1298 PRINT_ATTRf(mmap2, p_unsigned);
1299 PRINT_ATTRf(comm_exec, p_unsigned);
1300 PRINT_ATTRf(use_clockid, p_unsigned);
1301 PRINT_ATTRf(context_switch, p_unsigned);
1303 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1304 PRINT_ATTRf(bp_type, p_unsigned);
1305 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1306 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1307 PRINT_ATTRf(branch_sample_type, p_unsigned);
1308 PRINT_ATTRf(sample_regs_user, p_hex);
1309 PRINT_ATTRf(sample_stack_user, p_unsigned);
1310 PRINT_ATTRf(clockid, p_signed);
1311 PRINT_ATTRf(sample_regs_intr, p_hex);
1312 PRINT_ATTRf(aux_watermark, p_unsigned);
1317 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1318 void *priv __attribute__((unused)))
1320 return fprintf(fp, " %-32s %s\n", name, val);
1323 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1324 struct thread_map *threads)
1326 int cpu, thread, nthreads;
1327 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1329 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1331 if (evsel->system_wide)
1334 nthreads = threads->nr;
1336 if (evsel->fd == NULL &&
1337 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1341 flags |= PERF_FLAG_PID_CGROUP;
1342 pid = evsel->cgrp->fd;
1345 fallback_missing_features:
1346 if (perf_missing_features.clockid_wrong)
1347 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1348 if (perf_missing_features.clockid) {
1349 evsel->attr.use_clockid = 0;
1350 evsel->attr.clockid = 0;
1352 if (perf_missing_features.cloexec)
1353 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1354 if (perf_missing_features.mmap2)
1355 evsel->attr.mmap2 = 0;
1356 if (perf_missing_features.exclude_guest)
1357 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1358 if (perf_missing_features.lbr_flags)
1359 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1360 PERF_SAMPLE_BRANCH_NO_CYCLES);
1362 if (perf_missing_features.sample_id_all)
1363 evsel->attr.sample_id_all = 0;
1366 fprintf(stderr, "%.60s\n", graph_dotted_line);
1367 fprintf(stderr, "perf_event_attr:\n");
1368 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1369 fprintf(stderr, "%.60s\n", graph_dotted_line);
1372 for (cpu = 0; cpu < cpus->nr; cpu++) {
1374 for (thread = 0; thread < nthreads; thread++) {
1377 if (!evsel->cgrp && !evsel->system_wide)
1378 pid = thread_map__pid(threads, thread);
1380 group_fd = get_group_fd(evsel, cpu, thread);
1382 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1383 pid, cpus->map[cpu], group_fd, flags);
1385 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1389 if (FD(evsel, cpu, thread) < 0) {
1391 pr_debug2("sys_perf_event_open failed, error %d\n",
1396 if (evsel->bpf_fd >= 0) {
1397 int evt_fd = FD(evsel, cpu, thread);
1398 int bpf_fd = evsel->bpf_fd;
1401 PERF_EVENT_IOC_SET_BPF,
1403 if (err && errno != EEXIST) {
1404 pr_err("failed to attach bpf fd %d: %s\n",
1405 bpf_fd, strerror(errno));
1411 set_rlimit = NO_CHANGE;
1414 * If we succeeded but had to kill clockid, fail and
1415 * have perf_evsel__open_strerror() print us a nice
1418 if (perf_missing_features.clockid ||
1419 perf_missing_features.clockid_wrong) {
1430 * perf stat needs between 5 and 22 fds per CPU. When we run out
1431 * of them try to increase the limits.
1433 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1435 int old_errno = errno;
1437 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1438 if (set_rlimit == NO_CHANGE)
1439 l.rlim_cur = l.rlim_max;
1441 l.rlim_cur = l.rlim_max + 1000;
1442 l.rlim_max = l.rlim_cur;
1444 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1453 if (err != -EINVAL || cpu > 0 || thread > 0)
1457 * Must probe features in the order they were added to the
1458 * perf_event_attr interface.
1460 if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1461 perf_missing_features.clockid_wrong = true;
1462 goto fallback_missing_features;
1463 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1464 perf_missing_features.clockid = true;
1465 goto fallback_missing_features;
1466 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1467 perf_missing_features.cloexec = true;
1468 goto fallback_missing_features;
1469 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1470 perf_missing_features.mmap2 = true;
1471 goto fallback_missing_features;
1472 } else if (!perf_missing_features.exclude_guest &&
1473 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1474 perf_missing_features.exclude_guest = true;
1475 goto fallback_missing_features;
1476 } else if (!perf_missing_features.sample_id_all) {
1477 perf_missing_features.sample_id_all = true;
1478 goto retry_sample_id;
1479 } else if (!perf_missing_features.lbr_flags &&
1480 (evsel->attr.branch_sample_type &
1481 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1482 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1483 perf_missing_features.lbr_flags = true;
1484 goto fallback_missing_features;
1489 while (--thread >= 0) {
1490 close(FD(evsel, cpu, thread));
1491 FD(evsel, cpu, thread) = -1;
1494 } while (--cpu >= 0);
1498 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1500 if (evsel->fd == NULL)
1503 perf_evsel__close_fd(evsel, ncpus, nthreads);
1504 perf_evsel__free_fd(evsel);
1516 struct thread_map map;
1518 } empty_thread_map = {
1523 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1524 struct thread_map *threads)
1527 /* Work around old compiler warnings about strict aliasing */
1528 cpus = &empty_cpu_map.map;
1531 if (threads == NULL)
1532 threads = &empty_thread_map.map;
1534 return __perf_evsel__open(evsel, cpus, threads);
1537 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1538 struct cpu_map *cpus)
1540 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1543 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1544 struct thread_map *threads)
1546 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1549 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1550 const union perf_event *event,
1551 struct perf_sample *sample)
1553 u64 type = evsel->attr.sample_type;
1554 const u64 *array = event->sample.array;
1555 bool swapped = evsel->needs_swap;
1558 array += ((event->header.size -
1559 sizeof(event->header)) / sizeof(u64)) - 1;
1561 if (type & PERF_SAMPLE_IDENTIFIER) {
1562 sample->id = *array;
1566 if (type & PERF_SAMPLE_CPU) {
1569 /* undo swap of u64, then swap on individual u32s */
1570 u.val64 = bswap_64(u.val64);
1571 u.val32[0] = bswap_32(u.val32[0]);
1574 sample->cpu = u.val32[0];
1578 if (type & PERF_SAMPLE_STREAM_ID) {
1579 sample->stream_id = *array;
1583 if (type & PERF_SAMPLE_ID) {
1584 sample->id = *array;
1588 if (type & PERF_SAMPLE_TIME) {
1589 sample->time = *array;
1593 if (type & PERF_SAMPLE_TID) {
1596 /* undo swap of u64, then swap on individual u32s */
1597 u.val64 = bswap_64(u.val64);
1598 u.val32[0] = bswap_32(u.val32[0]);
1599 u.val32[1] = bswap_32(u.val32[1]);
1602 sample->pid = u.val32[0];
1603 sample->tid = u.val32[1];
1610 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1613 return size > max_size || offset + size > endp;
1616 #define OVERFLOW_CHECK(offset, size, max_size) \
1618 if (overflow(endp, (max_size), (offset), (size))) \
1622 #define OVERFLOW_CHECK_u64(offset) \
1623 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1625 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1626 struct perf_sample *data)
1628 u64 type = evsel->attr.sample_type;
1629 bool swapped = evsel->needs_swap;
1631 u16 max_size = event->header.size;
1632 const void *endp = (void *)event + max_size;
1636 * used for cross-endian analysis. See git commit 65014ab3
1637 * for why this goofiness is needed.
1641 memset(data, 0, sizeof(*data));
1642 data->cpu = data->pid = data->tid = -1;
1643 data->stream_id = data->id = data->time = -1ULL;
1644 data->period = evsel->attr.sample_period;
1646 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1648 if (event->header.type != PERF_RECORD_SAMPLE) {
1649 if (!evsel->attr.sample_id_all)
1651 return perf_evsel__parse_id_sample(evsel, event, data);
1654 array = event->sample.array;
1657 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1658 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1659 * check the format does not go past the end of the event.
1661 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1665 if (type & PERF_SAMPLE_IDENTIFIER) {
1670 if (type & PERF_SAMPLE_IP) {
1675 if (type & PERF_SAMPLE_TID) {
1678 /* undo swap of u64, then swap on individual u32s */
1679 u.val64 = bswap_64(u.val64);
1680 u.val32[0] = bswap_32(u.val32[0]);
1681 u.val32[1] = bswap_32(u.val32[1]);
1684 data->pid = u.val32[0];
1685 data->tid = u.val32[1];
1689 if (type & PERF_SAMPLE_TIME) {
1690 data->time = *array;
1695 if (type & PERF_SAMPLE_ADDR) {
1696 data->addr = *array;
1700 if (type & PERF_SAMPLE_ID) {
1705 if (type & PERF_SAMPLE_STREAM_ID) {
1706 data->stream_id = *array;
1710 if (type & PERF_SAMPLE_CPU) {
1714 /* undo swap of u64, then swap on individual u32s */
1715 u.val64 = bswap_64(u.val64);
1716 u.val32[0] = bswap_32(u.val32[0]);
1719 data->cpu = u.val32[0];
1723 if (type & PERF_SAMPLE_PERIOD) {
1724 data->period = *array;
1728 if (type & PERF_SAMPLE_READ) {
1729 u64 read_format = evsel->attr.read_format;
1731 OVERFLOW_CHECK_u64(array);
1732 if (read_format & PERF_FORMAT_GROUP)
1733 data->read.group.nr = *array;
1735 data->read.one.value = *array;
1739 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1740 OVERFLOW_CHECK_u64(array);
1741 data->read.time_enabled = *array;
1745 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1746 OVERFLOW_CHECK_u64(array);
1747 data->read.time_running = *array;
1751 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1752 if (read_format & PERF_FORMAT_GROUP) {
1753 const u64 max_group_nr = UINT64_MAX /
1754 sizeof(struct sample_read_value);
1756 if (data->read.group.nr > max_group_nr)
1758 sz = data->read.group.nr *
1759 sizeof(struct sample_read_value);
1760 OVERFLOW_CHECK(array, sz, max_size);
1761 data->read.group.values =
1762 (struct sample_read_value *)array;
1763 array = (void *)array + sz;
1765 OVERFLOW_CHECK_u64(array);
1766 data->read.one.id = *array;
1771 if (type & PERF_SAMPLE_CALLCHAIN) {
1772 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1774 OVERFLOW_CHECK_u64(array);
1775 data->callchain = (struct ip_callchain *)array++;
1776 if (data->callchain->nr > max_callchain_nr)
1778 sz = data->callchain->nr * sizeof(u64);
1779 OVERFLOW_CHECK(array, sz, max_size);
1780 array = (void *)array + sz;
1783 if (type & PERF_SAMPLE_RAW) {
1784 OVERFLOW_CHECK_u64(array);
1786 if (WARN_ONCE(swapped,
1787 "Endianness of raw data not corrected!\n")) {
1788 /* undo swap of u64, then swap on individual u32s */
1789 u.val64 = bswap_64(u.val64);
1790 u.val32[0] = bswap_32(u.val32[0]);
1791 u.val32[1] = bswap_32(u.val32[1]);
1793 data->raw_size = u.val32[0];
1794 array = (void *)array + sizeof(u32);
1796 OVERFLOW_CHECK(array, data->raw_size, max_size);
1797 data->raw_data = (void *)array;
1798 array = (void *)array + data->raw_size;
1801 if (type & PERF_SAMPLE_BRANCH_STACK) {
1802 const u64 max_branch_nr = UINT64_MAX /
1803 sizeof(struct branch_entry);
1805 OVERFLOW_CHECK_u64(array);
1806 data->branch_stack = (struct branch_stack *)array++;
1808 if (data->branch_stack->nr > max_branch_nr)
1810 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1811 OVERFLOW_CHECK(array, sz, max_size);
1812 array = (void *)array + sz;
1815 if (type & PERF_SAMPLE_REGS_USER) {
1816 OVERFLOW_CHECK_u64(array);
1817 data->user_regs.abi = *array;
1820 if (data->user_regs.abi) {
1821 u64 mask = evsel->attr.sample_regs_user;
1823 sz = hweight_long(mask) * sizeof(u64);
1824 OVERFLOW_CHECK(array, sz, max_size);
1825 data->user_regs.mask = mask;
1826 data->user_regs.regs = (u64 *)array;
1827 array = (void *)array + sz;
1831 if (type & PERF_SAMPLE_STACK_USER) {
1832 OVERFLOW_CHECK_u64(array);
1835 data->user_stack.offset = ((char *)(array - 1)
1839 data->user_stack.size = 0;
1841 OVERFLOW_CHECK(array, sz, max_size);
1842 data->user_stack.data = (char *)array;
1843 array = (void *)array + sz;
1844 OVERFLOW_CHECK_u64(array);
1845 data->user_stack.size = *array++;
1846 if (WARN_ONCE(data->user_stack.size > sz,
1847 "user stack dump failure\n"))
1853 if (type & PERF_SAMPLE_WEIGHT) {
1854 OVERFLOW_CHECK_u64(array);
1855 data->weight = *array;
1859 data->data_src = PERF_MEM_DATA_SRC_NONE;
1860 if (type & PERF_SAMPLE_DATA_SRC) {
1861 OVERFLOW_CHECK_u64(array);
1862 data->data_src = *array;
1866 data->transaction = 0;
1867 if (type & PERF_SAMPLE_TRANSACTION) {
1868 OVERFLOW_CHECK_u64(array);
1869 data->transaction = *array;
1873 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1874 if (type & PERF_SAMPLE_REGS_INTR) {
1875 OVERFLOW_CHECK_u64(array);
1876 data->intr_regs.abi = *array;
1879 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1880 u64 mask = evsel->attr.sample_regs_intr;
1882 sz = hweight_long(mask) * sizeof(u64);
1883 OVERFLOW_CHECK(array, sz, max_size);
1884 data->intr_regs.mask = mask;
1885 data->intr_regs.regs = (u64 *)array;
1886 array = (void *)array + sz;
1893 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1896 size_t sz, result = sizeof(struct sample_event);
1898 if (type & PERF_SAMPLE_IDENTIFIER)
1899 result += sizeof(u64);
1901 if (type & PERF_SAMPLE_IP)
1902 result += sizeof(u64);
1904 if (type & PERF_SAMPLE_TID)
1905 result += sizeof(u64);
1907 if (type & PERF_SAMPLE_TIME)
1908 result += sizeof(u64);
1910 if (type & PERF_SAMPLE_ADDR)
1911 result += sizeof(u64);
1913 if (type & PERF_SAMPLE_ID)
1914 result += sizeof(u64);
1916 if (type & PERF_SAMPLE_STREAM_ID)
1917 result += sizeof(u64);
1919 if (type & PERF_SAMPLE_CPU)
1920 result += sizeof(u64);
1922 if (type & PERF_SAMPLE_PERIOD)
1923 result += sizeof(u64);
1925 if (type & PERF_SAMPLE_READ) {
1926 result += sizeof(u64);
1927 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1928 result += sizeof(u64);
1929 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1930 result += sizeof(u64);
1931 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1932 if (read_format & PERF_FORMAT_GROUP) {
1933 sz = sample->read.group.nr *
1934 sizeof(struct sample_read_value);
1937 result += sizeof(u64);
1941 if (type & PERF_SAMPLE_CALLCHAIN) {
1942 sz = (sample->callchain->nr + 1) * sizeof(u64);
1946 if (type & PERF_SAMPLE_RAW) {
1947 result += sizeof(u32);
1948 result += sample->raw_size;
1951 if (type & PERF_SAMPLE_BRANCH_STACK) {
1952 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1957 if (type & PERF_SAMPLE_REGS_USER) {
1958 if (sample->user_regs.abi) {
1959 result += sizeof(u64);
1960 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1963 result += sizeof(u64);
1967 if (type & PERF_SAMPLE_STACK_USER) {
1968 sz = sample->user_stack.size;
1969 result += sizeof(u64);
1972 result += sizeof(u64);
1976 if (type & PERF_SAMPLE_WEIGHT)
1977 result += sizeof(u64);
1979 if (type & PERF_SAMPLE_DATA_SRC)
1980 result += sizeof(u64);
1982 if (type & PERF_SAMPLE_TRANSACTION)
1983 result += sizeof(u64);
1985 if (type & PERF_SAMPLE_REGS_INTR) {
1986 if (sample->intr_regs.abi) {
1987 result += sizeof(u64);
1988 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1991 result += sizeof(u64);
1998 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2000 const struct perf_sample *sample,
2006 * used for cross-endian analysis. See git commit 65014ab3
2007 * for why this goofiness is needed.
2011 array = event->sample.array;
2013 if (type & PERF_SAMPLE_IDENTIFIER) {
2014 *array = sample->id;
2018 if (type & PERF_SAMPLE_IP) {
2019 *array = sample->ip;
2023 if (type & PERF_SAMPLE_TID) {
2024 u.val32[0] = sample->pid;
2025 u.val32[1] = sample->tid;
2028 * Inverse of what is done in perf_evsel__parse_sample
2030 u.val32[0] = bswap_32(u.val32[0]);
2031 u.val32[1] = bswap_32(u.val32[1]);
2032 u.val64 = bswap_64(u.val64);
2039 if (type & PERF_SAMPLE_TIME) {
2040 *array = sample->time;
2044 if (type & PERF_SAMPLE_ADDR) {
2045 *array = sample->addr;
2049 if (type & PERF_SAMPLE_ID) {
2050 *array = sample->id;
2054 if (type & PERF_SAMPLE_STREAM_ID) {
2055 *array = sample->stream_id;
2059 if (type & PERF_SAMPLE_CPU) {
2060 u.val32[0] = sample->cpu;
2063 * Inverse of what is done in perf_evsel__parse_sample
2065 u.val32[0] = bswap_32(u.val32[0]);
2066 u.val64 = bswap_64(u.val64);
2072 if (type & PERF_SAMPLE_PERIOD) {
2073 *array = sample->period;
2077 if (type & PERF_SAMPLE_READ) {
2078 if (read_format & PERF_FORMAT_GROUP)
2079 *array = sample->read.group.nr;
2081 *array = sample->read.one.value;
2084 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2085 *array = sample->read.time_enabled;
2089 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2090 *array = sample->read.time_running;
2094 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2095 if (read_format & PERF_FORMAT_GROUP) {
2096 sz = sample->read.group.nr *
2097 sizeof(struct sample_read_value);
2098 memcpy(array, sample->read.group.values, sz);
2099 array = (void *)array + sz;
2101 *array = sample->read.one.id;
2106 if (type & PERF_SAMPLE_CALLCHAIN) {
2107 sz = (sample->callchain->nr + 1) * sizeof(u64);
2108 memcpy(array, sample->callchain, sz);
2109 array = (void *)array + sz;
2112 if (type & PERF_SAMPLE_RAW) {
2113 u.val32[0] = sample->raw_size;
2114 if (WARN_ONCE(swapped,
2115 "Endianness of raw data not corrected!\n")) {
2117 * Inverse of what is done in perf_evsel__parse_sample
2119 u.val32[0] = bswap_32(u.val32[0]);
2120 u.val32[1] = bswap_32(u.val32[1]);
2121 u.val64 = bswap_64(u.val64);
2124 array = (void *)array + sizeof(u32);
2126 memcpy(array, sample->raw_data, sample->raw_size);
2127 array = (void *)array + sample->raw_size;
2130 if (type & PERF_SAMPLE_BRANCH_STACK) {
2131 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2133 memcpy(array, sample->branch_stack, sz);
2134 array = (void *)array + sz;
2137 if (type & PERF_SAMPLE_REGS_USER) {
2138 if (sample->user_regs.abi) {
2139 *array++ = sample->user_regs.abi;
2140 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2141 memcpy(array, sample->user_regs.regs, sz);
2142 array = (void *)array + sz;
2148 if (type & PERF_SAMPLE_STACK_USER) {
2149 sz = sample->user_stack.size;
2152 memcpy(array, sample->user_stack.data, sz);
2153 array = (void *)array + sz;
2158 if (type & PERF_SAMPLE_WEIGHT) {
2159 *array = sample->weight;
2163 if (type & PERF_SAMPLE_DATA_SRC) {
2164 *array = sample->data_src;
2168 if (type & PERF_SAMPLE_TRANSACTION) {
2169 *array = sample->transaction;
2173 if (type & PERF_SAMPLE_REGS_INTR) {
2174 if (sample->intr_regs.abi) {
2175 *array++ = sample->intr_regs.abi;
2176 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2177 memcpy(array, sample->intr_regs.regs, sz);
2178 array = (void *)array + sz;
2187 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2189 return pevent_find_field(evsel->tp_format, name);
2192 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2195 struct format_field *field = perf_evsel__field(evsel, name);
2201 offset = field->offset;
2203 if (field->flags & FIELD_IS_DYNAMIC) {
2204 offset = *(int *)(sample->raw_data + field->offset);
2208 return sample->raw_data + offset;
2211 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2214 struct format_field *field = perf_evsel__field(evsel, name);
2221 ptr = sample->raw_data + field->offset;
2223 switch (field->size) {
2227 value = *(u16 *)ptr;
2230 value = *(u32 *)ptr;
2233 memcpy(&value, ptr, sizeof(u64));
2239 if (!evsel->needs_swap)
2242 switch (field->size) {
2244 return bswap_16(value);
2246 return bswap_32(value);
2248 return bswap_64(value);
2256 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2262 ret += fprintf(fp, ",");
2264 ret += fprintf(fp, ":");
2268 va_start(args, fmt);
2269 ret += vfprintf(fp, fmt, args);
2274 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2276 return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2279 int perf_evsel__fprintf(struct perf_evsel *evsel,
2280 struct perf_attr_details *details, FILE *fp)
2285 if (details->event_group) {
2286 struct perf_evsel *pos;
2288 if (!perf_evsel__is_group_leader(evsel))
2291 if (evsel->nr_members > 1)
2292 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2294 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2295 for_each_group_member(pos, evsel)
2296 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2298 if (evsel->nr_members > 1)
2299 printed += fprintf(fp, "}");
2303 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2305 if (details->verbose) {
2306 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2307 __print_attr__fprintf, &first);
2308 } else if (details->freq) {
2309 const char *term = "sample_freq";
2311 if (!evsel->attr.freq)
2312 term = "sample_period";
2314 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2315 term, (u64)evsel->attr.sample_freq);
2318 if (details->trace_fields) {
2319 struct format_field *field;
2321 if (evsel->attr.type != PERF_TYPE_TRACEPOINT) {
2322 printed += comma_fprintf(fp, &first, " (not a tracepoint)");
2326 field = evsel->tp_format->format.fields;
2327 if (field == NULL) {
2328 printed += comma_fprintf(fp, &first, " (no trace field)");
2332 printed += comma_fprintf(fp, &first, " trace_fields: %s", field->name);
2334 field = field->next;
2336 printed += comma_fprintf(fp, &first, "%s", field->name);
2337 field = field->next;
2345 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2346 char *msg, size_t msgsize)
2350 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2351 evsel->attr.type == PERF_TYPE_HARDWARE &&
2352 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2354 * If it's cycles then fall back to hrtimer based
2355 * cpu-clock-tick sw counter, which is always available even if
2358 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2361 scnprintf(msg, msgsize, "%s",
2362 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2364 evsel->attr.type = PERF_TYPE_SOFTWARE;
2365 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2367 zfree(&evsel->name);
2369 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2370 (paranoid = perf_event_paranoid()) > 1) {
2371 const char *name = perf_evsel__name(evsel);
2374 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2379 evsel->name = new_name;
2380 scnprintf(msg, msgsize,
2381 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2382 evsel->attr.exclude_kernel = 1;
2390 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2391 int err, char *msg, size_t size)
2393 char sbuf[STRERR_BUFSIZE];
2398 return scnprintf(msg, size,
2399 "You may not have permission to collect %sstats.\n\n"
2400 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2401 "which controls use of the performance events system by\n"
2402 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2403 "The current value is %d:\n\n"
2404 " -1: Allow use of (almost) all events by all users\n"
2405 ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2406 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2407 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2408 target->system_wide ? "system-wide " : "",
2409 perf_event_paranoid());
2411 return scnprintf(msg, size, "The %s event is not supported.",
2412 perf_evsel__name(evsel));
2414 return scnprintf(msg, size, "%s",
2415 "Too many events are opened.\n"
2416 "Probably the maximum number of open file descriptors has been reached.\n"
2417 "Hint: Try again after reducing the number of events.\n"
2418 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2420 if (target->cpu_list)
2421 return scnprintf(msg, size, "%s",
2422 "No such device - did you specify an out-of-range profile CPU?\n");
2425 if (evsel->attr.precise_ip)
2426 return scnprintf(msg, size, "%s",
2427 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2428 #if defined(__i386__) || defined(__x86_64__)
2429 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2430 return scnprintf(msg, size, "%s",
2431 "No hardware sampling interrupt available.\n"
2432 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2436 if (find_process("oprofiled"))
2437 return scnprintf(msg, size,
2438 "The PMU counters are busy/taken by another profiler.\n"
2439 "We found oprofile daemon running, please stop it and try again.");
2442 if (perf_missing_features.clockid)
2443 return scnprintf(msg, size, "clockid feature not supported.");
2444 if (perf_missing_features.clockid_wrong)
2445 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2451 return scnprintf(msg, size,
2452 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2453 "/bin/dmesg may provide additional information.\n"
2454 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2455 err, strerror_r(err, sbuf, sizeof(sbuf)),
2456 perf_evsel__name(evsel));