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"
31 #include "util/parse-branch-options.h"
42 } perf_missing_features;
44 static clockid_t clockid;
46 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
51 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
57 int (*init)(struct perf_evsel *evsel);
58 void (*fini)(struct perf_evsel *evsel);
59 } perf_evsel__object = {
60 .size = sizeof(struct perf_evsel),
61 .init = perf_evsel__no_extra_init,
62 .fini = perf_evsel__no_extra_fini,
65 int perf_evsel__object_config(size_t object_size,
66 int (*init)(struct perf_evsel *evsel),
67 void (*fini)(struct perf_evsel *evsel))
73 if (perf_evsel__object.size > object_size)
76 perf_evsel__object.size = object_size;
80 perf_evsel__object.init = init;
83 perf_evsel__object.fini = fini;
88 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
90 int __perf_evsel__sample_size(u64 sample_type)
92 u64 mask = sample_type & PERF_SAMPLE_MASK;
96 for (i = 0; i < 64; i++) {
97 if (mask & (1ULL << i))
107 * __perf_evsel__calc_id_pos - calculate id_pos.
108 * @sample_type: sample type
110 * This function returns the position of the event id (PERF_SAMPLE_ID or
111 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
114 static int __perf_evsel__calc_id_pos(u64 sample_type)
118 if (sample_type & PERF_SAMPLE_IDENTIFIER)
121 if (!(sample_type & PERF_SAMPLE_ID))
124 if (sample_type & PERF_SAMPLE_IP)
127 if (sample_type & PERF_SAMPLE_TID)
130 if (sample_type & PERF_SAMPLE_TIME)
133 if (sample_type & PERF_SAMPLE_ADDR)
140 * __perf_evsel__calc_is_pos - calculate is_pos.
141 * @sample_type: sample type
143 * This function returns the position (counting backwards) of the event id
144 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
145 * sample_id_all is used there is an id sample appended to non-sample events.
147 static int __perf_evsel__calc_is_pos(u64 sample_type)
151 if (sample_type & PERF_SAMPLE_IDENTIFIER)
154 if (!(sample_type & PERF_SAMPLE_ID))
157 if (sample_type & PERF_SAMPLE_CPU)
160 if (sample_type & PERF_SAMPLE_STREAM_ID)
166 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
168 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
169 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
172 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
173 enum perf_event_sample_format bit)
175 if (!(evsel->attr.sample_type & bit)) {
176 evsel->attr.sample_type |= bit;
177 evsel->sample_size += sizeof(u64);
178 perf_evsel__calc_id_pos(evsel);
182 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
183 enum perf_event_sample_format bit)
185 if (evsel->attr.sample_type & bit) {
186 evsel->attr.sample_type &= ~bit;
187 evsel->sample_size -= sizeof(u64);
188 perf_evsel__calc_id_pos(evsel);
192 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
193 bool can_sample_identifier)
195 if (can_sample_identifier) {
196 perf_evsel__reset_sample_bit(evsel, ID);
197 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
199 perf_evsel__set_sample_bit(evsel, ID);
201 evsel->attr.read_format |= PERF_FORMAT_ID;
205 * perf_evsel__is_function_event - Return whether given evsel is a function
208 * @evsel - evsel selector to be tested
210 * Return %true if event is function trace event
212 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
214 #define FUNCTION_EVENT "ftrace:function"
216 return evsel->name &&
217 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
219 #undef FUNCTION_EVENT
222 void perf_evsel__init(struct perf_evsel *evsel,
223 struct perf_event_attr *attr, int idx)
226 evsel->tracking = !idx;
228 evsel->leader = evsel;
231 evsel->evlist = NULL;
233 INIT_LIST_HEAD(&evsel->node);
234 INIT_LIST_HEAD(&evsel->config_terms);
235 perf_evsel__object.init(evsel);
236 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
237 perf_evsel__calc_id_pos(evsel);
238 evsel->cmdline_group_boundary = false;
241 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
243 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
246 perf_evsel__init(evsel, attr, idx);
248 if (perf_evsel__is_bpf_output(evsel)) {
249 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
250 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
251 evsel->attr.sample_period = 1;
257 struct perf_evsel *perf_evsel__new_cycles(void)
259 struct perf_event_attr attr = {
260 .type = PERF_TYPE_HARDWARE,
261 .config = PERF_COUNT_HW_CPU_CYCLES,
263 struct perf_evsel *evsel;
265 event_attr_init(&attr);
267 perf_event_attr__set_max_precise_ip(&attr);
269 evsel = perf_evsel__new(&attr);
273 /* use asprintf() because free(evsel) assumes name is allocated */
274 if (asprintf(&evsel->name, "cycles%.*s",
275 attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
280 perf_evsel__delete(evsel);
286 * Returns pointer with encoded error via <linux/err.h> interface.
288 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
290 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
296 struct perf_event_attr attr = {
297 .type = PERF_TYPE_TRACEPOINT,
298 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
299 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
302 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
305 evsel->tp_format = trace_event__tp_format(sys, name);
306 if (IS_ERR(evsel->tp_format)) {
307 err = PTR_ERR(evsel->tp_format);
311 event_attr_init(&attr);
312 attr.config = evsel->tp_format->id;
313 attr.sample_period = 1;
314 perf_evsel__init(evsel, &attr, idx);
326 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
334 "stalled-cycles-frontend",
335 "stalled-cycles-backend",
339 static const char *__perf_evsel__hw_name(u64 config)
341 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
342 return perf_evsel__hw_names[config];
344 return "unknown-hardware";
347 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
349 int colon = 0, r = 0;
350 struct perf_event_attr *attr = &evsel->attr;
351 bool exclude_guest_default = false;
353 #define MOD_PRINT(context, mod) do { \
354 if (!attr->exclude_##context) { \
355 if (!colon) colon = ++r; \
356 r += scnprintf(bf + r, size - r, "%c", mod); \
359 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
360 MOD_PRINT(kernel, 'k');
361 MOD_PRINT(user, 'u');
363 exclude_guest_default = true;
366 if (attr->precise_ip) {
369 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
370 exclude_guest_default = true;
373 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
374 MOD_PRINT(host, 'H');
375 MOD_PRINT(guest, 'G');
383 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
385 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
386 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
389 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
402 static const char *__perf_evsel__sw_name(u64 config)
404 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
405 return perf_evsel__sw_names[config];
406 return "unknown-software";
409 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
411 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
412 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
415 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
419 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
421 if (type & HW_BREAKPOINT_R)
422 r += scnprintf(bf + r, size - r, "r");
424 if (type & HW_BREAKPOINT_W)
425 r += scnprintf(bf + r, size - r, "w");
427 if (type & HW_BREAKPOINT_X)
428 r += scnprintf(bf + r, size - r, "x");
433 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
435 struct perf_event_attr *attr = &evsel->attr;
436 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
437 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
440 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
441 [PERF_EVSEL__MAX_ALIASES] = {
442 { "L1-dcache", "l1-d", "l1d", "L1-data", },
443 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
445 { "dTLB", "d-tlb", "Data-TLB", },
446 { "iTLB", "i-tlb", "Instruction-TLB", },
447 { "branch", "branches", "bpu", "btb", "bpc", },
451 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
452 [PERF_EVSEL__MAX_ALIASES] = {
453 { "load", "loads", "read", },
454 { "store", "stores", "write", },
455 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
458 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
459 [PERF_EVSEL__MAX_ALIASES] = {
460 { "refs", "Reference", "ops", "access", },
461 { "misses", "miss", },
464 #define C(x) PERF_COUNT_HW_CACHE_##x
465 #define CACHE_READ (1 << C(OP_READ))
466 #define CACHE_WRITE (1 << C(OP_WRITE))
467 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
468 #define COP(x) (1 << x)
471 * cache operartion stat
472 * L1I : Read and prefetch only
473 * ITLB and BPU : Read-only
475 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
476 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
477 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
478 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
479 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
480 [C(ITLB)] = (CACHE_READ),
481 [C(BPU)] = (CACHE_READ),
482 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
485 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
487 if (perf_evsel__hw_cache_stat[type] & COP(op))
488 return true; /* valid */
490 return false; /* invalid */
493 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
494 char *bf, size_t size)
497 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
498 perf_evsel__hw_cache_op[op][0],
499 perf_evsel__hw_cache_result[result][0]);
502 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
503 perf_evsel__hw_cache_op[op][1]);
506 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
508 u8 op, result, type = (config >> 0) & 0xff;
509 const char *err = "unknown-ext-hardware-cache-type";
511 if (type >= PERF_COUNT_HW_CACHE_MAX)
514 op = (config >> 8) & 0xff;
515 err = "unknown-ext-hardware-cache-op";
516 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
519 result = (config >> 16) & 0xff;
520 err = "unknown-ext-hardware-cache-result";
521 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
524 err = "invalid-cache";
525 if (!perf_evsel__is_cache_op_valid(type, op))
528 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
530 return scnprintf(bf, size, "%s", err);
533 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
535 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
536 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
539 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
541 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
542 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
545 const char *perf_evsel__name(struct perf_evsel *evsel)
552 switch (evsel->attr.type) {
554 perf_evsel__raw_name(evsel, bf, sizeof(bf));
557 case PERF_TYPE_HARDWARE:
558 perf_evsel__hw_name(evsel, bf, sizeof(bf));
561 case PERF_TYPE_HW_CACHE:
562 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
565 case PERF_TYPE_SOFTWARE:
566 perf_evsel__sw_name(evsel, bf, sizeof(bf));
569 case PERF_TYPE_TRACEPOINT:
570 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
573 case PERF_TYPE_BREAKPOINT:
574 perf_evsel__bp_name(evsel, bf, sizeof(bf));
578 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
583 evsel->name = strdup(bf);
585 return evsel->name ?: "unknown";
588 const char *perf_evsel__group_name(struct perf_evsel *evsel)
590 return evsel->group_name ?: "anon group";
593 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
596 struct perf_evsel *pos;
597 const char *group_name = perf_evsel__group_name(evsel);
599 ret = scnprintf(buf, size, "%s", group_name);
601 ret += scnprintf(buf + ret, size - ret, " { %s",
602 perf_evsel__name(evsel));
604 for_each_group_member(pos, evsel)
605 ret += scnprintf(buf + ret, size - ret, ", %s",
606 perf_evsel__name(pos));
608 ret += scnprintf(buf + ret, size - ret, " }");
613 void perf_evsel__config_callchain(struct perf_evsel *evsel,
614 struct record_opts *opts,
615 struct callchain_param *param)
617 bool function = perf_evsel__is_function_event(evsel);
618 struct perf_event_attr *attr = &evsel->attr;
620 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
622 attr->sample_max_stack = param->max_stack;
624 if (param->record_mode == CALLCHAIN_LBR) {
625 if (!opts->branch_stack) {
626 if (attr->exclude_user) {
627 pr_warning("LBR callstack option is only available "
628 "to get user callchain information. "
629 "Falling back to framepointers.\n");
631 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
632 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
633 PERF_SAMPLE_BRANCH_CALL_STACK |
634 PERF_SAMPLE_BRANCH_NO_CYCLES |
635 PERF_SAMPLE_BRANCH_NO_FLAGS;
638 pr_warning("Cannot use LBR callstack with branch stack. "
639 "Falling back to framepointers.\n");
642 if (param->record_mode == CALLCHAIN_DWARF) {
644 perf_evsel__set_sample_bit(evsel, REGS_USER);
645 perf_evsel__set_sample_bit(evsel, STACK_USER);
646 attr->sample_regs_user = PERF_REGS_MASK;
647 attr->sample_stack_user = param->dump_size;
648 attr->exclude_callchain_user = 1;
650 pr_info("Cannot use DWARF unwind for function trace event,"
651 " falling back to framepointers.\n");
656 pr_info("Disabling user space callchains for function trace event.\n");
657 attr->exclude_callchain_user = 1;
662 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
663 struct callchain_param *param)
665 struct perf_event_attr *attr = &evsel->attr;
667 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
668 if (param->record_mode == CALLCHAIN_LBR) {
669 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
670 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
671 PERF_SAMPLE_BRANCH_CALL_STACK);
673 if (param->record_mode == CALLCHAIN_DWARF) {
674 perf_evsel__reset_sample_bit(evsel, REGS_USER);
675 perf_evsel__reset_sample_bit(evsel, STACK_USER);
679 static void apply_config_terms(struct perf_evsel *evsel,
680 struct record_opts *opts)
682 struct perf_evsel_config_term *term;
683 struct list_head *config_terms = &evsel->config_terms;
684 struct perf_event_attr *attr = &evsel->attr;
685 struct callchain_param param;
688 const char *callgraph_buf = NULL;
690 /* callgraph default */
691 param.record_mode = callchain_param.record_mode;
693 list_for_each_entry(term, config_terms, list) {
694 switch (term->type) {
695 case PERF_EVSEL__CONFIG_TERM_PERIOD:
696 attr->sample_period = term->val.period;
699 case PERF_EVSEL__CONFIG_TERM_FREQ:
700 attr->sample_freq = term->val.freq;
703 case PERF_EVSEL__CONFIG_TERM_TIME:
705 perf_evsel__set_sample_bit(evsel, TIME);
707 perf_evsel__reset_sample_bit(evsel, TIME);
709 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
710 callgraph_buf = term->val.callgraph;
712 case PERF_EVSEL__CONFIG_TERM_BRANCH:
713 if (term->val.branch && strcmp(term->val.branch, "no")) {
714 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
715 parse_branch_str(term->val.branch,
716 &attr->branch_sample_type);
718 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
720 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
721 dump_size = term->val.stack_user;
723 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
724 max_stack = term->val.max_stack;
726 case PERF_EVSEL__CONFIG_TERM_INHERIT:
728 * attr->inherit should has already been set by
729 * perf_evsel__config. If user explicitly set
730 * inherit using config terms, override global
731 * opt->no_inherit setting.
733 attr->inherit = term->val.inherit ? 1 : 0;
735 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
736 attr->write_backward = term->val.overwrite ? 1 : 0;
743 /* User explicitly set per-event callgraph, clear the old setting and reset. */
744 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
746 param.max_stack = max_stack;
747 if (callgraph_buf == NULL)
748 callgraph_buf = "fp";
751 /* parse callgraph parameters */
752 if (callgraph_buf != NULL) {
753 if (!strcmp(callgraph_buf, "no")) {
754 param.enabled = false;
755 param.record_mode = CALLCHAIN_NONE;
757 param.enabled = true;
758 if (parse_callchain_record(callgraph_buf, ¶m)) {
759 pr_err("per-event callgraph setting for %s failed. "
760 "Apply callgraph global setting for it\n",
767 dump_size = round_up(dump_size, sizeof(u64));
768 param.dump_size = dump_size;
771 /* If global callgraph set, clear it */
772 if (callchain_param.enabled)
773 perf_evsel__reset_callgraph(evsel, &callchain_param);
775 /* set perf-event callgraph */
777 perf_evsel__config_callchain(evsel, opts, ¶m);
782 * The enable_on_exec/disabled value strategy:
784 * 1) For any type of traced program:
785 * - all independent events and group leaders are disabled
786 * - all group members are enabled
788 * Group members are ruled by group leaders. They need to
789 * be enabled, because the group scheduling relies on that.
791 * 2) For traced programs executed by perf:
792 * - all independent events and group leaders have
794 * - we don't specifically enable or disable any event during
797 * Independent events and group leaders are initially disabled
798 * and get enabled by exec. Group members are ruled by group
799 * leaders as stated in 1).
801 * 3) For traced programs attached by perf (pid/tid):
802 * - we specifically enable or disable all events during
805 * When attaching events to already running traced we
806 * enable/disable events specifically, as there's no
807 * initial traced exec call.
809 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
810 struct callchain_param *callchain)
812 struct perf_evsel *leader = evsel->leader;
813 struct perf_event_attr *attr = &evsel->attr;
814 int track = evsel->tracking;
815 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
817 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
818 attr->inherit = !opts->no_inherit;
819 attr->write_backward = opts->overwrite ? 1 : 0;
821 perf_evsel__set_sample_bit(evsel, IP);
822 perf_evsel__set_sample_bit(evsel, TID);
824 if (evsel->sample_read) {
825 perf_evsel__set_sample_bit(evsel, READ);
828 * We need ID even in case of single event, because
829 * PERF_SAMPLE_READ process ID specific data.
831 perf_evsel__set_sample_id(evsel, false);
834 * Apply group format only if we belong to group
835 * with more than one members.
837 if (leader->nr_members > 1) {
838 attr->read_format |= PERF_FORMAT_GROUP;
844 * We default some events to have a default interval. But keep
845 * it a weak assumption overridable by the user.
847 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
848 opts->user_interval != ULLONG_MAX)) {
850 perf_evsel__set_sample_bit(evsel, PERIOD);
852 attr->sample_freq = opts->freq;
854 attr->sample_period = opts->default_interval;
859 * Disable sampling for all group members other
860 * than leader in case leader 'leads' the sampling.
862 if ((leader != evsel) && leader->sample_read) {
863 attr->sample_freq = 0;
864 attr->sample_period = 0;
867 if (opts->no_samples)
868 attr->sample_freq = 0;
870 if (opts->inherit_stat)
871 attr->inherit_stat = 1;
873 if (opts->sample_address) {
874 perf_evsel__set_sample_bit(evsel, ADDR);
875 attr->mmap_data = track;
879 * We don't allow user space callchains for function trace
880 * event, due to issues with page faults while tracing page
881 * fault handler and its overall trickiness nature.
883 if (perf_evsel__is_function_event(evsel))
884 evsel->attr.exclude_callchain_user = 1;
886 if (callchain && callchain->enabled && !evsel->no_aux_samples)
887 perf_evsel__config_callchain(evsel, opts, callchain);
889 if (opts->sample_intr_regs) {
890 attr->sample_regs_intr = opts->sample_intr_regs;
891 perf_evsel__set_sample_bit(evsel, REGS_INTR);
894 if (target__has_cpu(&opts->target) || opts->sample_cpu)
895 perf_evsel__set_sample_bit(evsel, CPU);
898 perf_evsel__set_sample_bit(evsel, PERIOD);
901 * When the user explicitly disabled time don't force it here.
903 if (opts->sample_time &&
904 (!perf_missing_features.sample_id_all &&
905 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
906 opts->sample_time_set)))
907 perf_evsel__set_sample_bit(evsel, TIME);
909 if (opts->raw_samples && !evsel->no_aux_samples) {
910 perf_evsel__set_sample_bit(evsel, TIME);
911 perf_evsel__set_sample_bit(evsel, RAW);
912 perf_evsel__set_sample_bit(evsel, CPU);
915 if (opts->sample_address)
916 perf_evsel__set_sample_bit(evsel, DATA_SRC);
918 if (opts->no_buffering) {
920 attr->wakeup_events = 1;
922 if (opts->branch_stack && !evsel->no_aux_samples) {
923 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
924 attr->branch_sample_type = opts->branch_stack;
927 if (opts->sample_weight)
928 perf_evsel__set_sample_bit(evsel, WEIGHT);
932 attr->mmap2 = track && !perf_missing_features.mmap2;
935 if (opts->record_switch_events)
936 attr->context_switch = track;
938 if (opts->sample_transaction)
939 perf_evsel__set_sample_bit(evsel, TRANSACTION);
941 if (opts->running_time) {
942 evsel->attr.read_format |=
943 PERF_FORMAT_TOTAL_TIME_ENABLED |
944 PERF_FORMAT_TOTAL_TIME_RUNNING;
948 * XXX see the function comment above
950 * Disabling only independent events or group leaders,
951 * keeping group members enabled.
953 if (perf_evsel__is_group_leader(evsel))
957 * Setting enable_on_exec for independent events and
958 * group leaders for traced executed by perf.
960 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
961 !opts->initial_delay)
962 attr->enable_on_exec = 1;
964 if (evsel->immediate) {
966 attr->enable_on_exec = 0;
969 clockid = opts->clockid;
970 if (opts->use_clockid) {
971 attr->use_clockid = 1;
972 attr->clockid = opts->clockid;
975 if (evsel->precise_max)
976 perf_event_attr__set_max_precise_ip(attr);
978 if (opts->all_user) {
979 attr->exclude_kernel = 1;
980 attr->exclude_user = 0;
983 if (opts->all_kernel) {
984 attr->exclude_kernel = 0;
985 attr->exclude_user = 1;
989 * Apply event specific term settings,
990 * it overloads any global configuration.
992 apply_config_terms(evsel, opts);
995 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
997 if (evsel->system_wide)
1000 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1004 for (cpu = 0; cpu < ncpus; cpu++) {
1005 for (thread = 0; thread < nthreads; thread++) {
1006 FD(evsel, cpu, thread) = -1;
1011 return evsel->fd != NULL ? 0 : -ENOMEM;
1014 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
1019 if (evsel->system_wide)
1022 for (cpu = 0; cpu < ncpus; cpu++) {
1023 for (thread = 0; thread < nthreads; thread++) {
1024 int fd = FD(evsel, cpu, thread),
1025 err = ioctl(fd, ioc, arg);
1035 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
1038 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1039 PERF_EVENT_IOC_SET_FILTER,
1043 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1045 char *new_filter = strdup(filter);
1047 if (new_filter != NULL) {
1048 free(evsel->filter);
1049 evsel->filter = new_filter;
1056 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1057 const char *fmt, const char *filter)
1061 if (evsel->filter == NULL)
1062 return perf_evsel__set_filter(evsel, filter);
1064 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1065 free(evsel->filter);
1066 evsel->filter = new_filter;
1073 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1075 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1078 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1080 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1083 int perf_evsel__enable(struct perf_evsel *evsel)
1085 int nthreads = thread_map__nr(evsel->threads);
1086 int ncpus = cpu_map__nr(evsel->cpus);
1088 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1089 PERF_EVENT_IOC_ENABLE,
1093 int perf_evsel__disable(struct perf_evsel *evsel)
1095 int nthreads = thread_map__nr(evsel->threads);
1096 int ncpus = cpu_map__nr(evsel->cpus);
1098 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1099 PERF_EVENT_IOC_DISABLE,
1103 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1105 if (ncpus == 0 || nthreads == 0)
1108 if (evsel->system_wide)
1111 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1112 if (evsel->sample_id == NULL)
1115 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1116 if (evsel->id == NULL) {
1117 xyarray__delete(evsel->sample_id);
1118 evsel->sample_id = NULL;
1125 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1127 xyarray__delete(evsel->fd);
1131 static void perf_evsel__free_id(struct perf_evsel *evsel)
1133 xyarray__delete(evsel->sample_id);
1134 evsel->sample_id = NULL;
1138 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1140 struct perf_evsel_config_term *term, *h;
1142 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1143 list_del(&term->list);
1148 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1152 if (evsel->system_wide)
1155 for (cpu = 0; cpu < ncpus; cpu++)
1156 for (thread = 0; thread < nthreads; ++thread) {
1157 close(FD(evsel, cpu, thread));
1158 FD(evsel, cpu, thread) = -1;
1162 void perf_evsel__exit(struct perf_evsel *evsel)
1164 assert(list_empty(&evsel->node));
1165 assert(evsel->evlist == NULL);
1166 perf_evsel__free_fd(evsel);
1167 perf_evsel__free_id(evsel);
1168 perf_evsel__free_config_terms(evsel);
1169 close_cgroup(evsel->cgrp);
1170 cpu_map__put(evsel->cpus);
1171 cpu_map__put(evsel->own_cpus);
1172 thread_map__put(evsel->threads);
1173 zfree(&evsel->group_name);
1174 zfree(&evsel->name);
1175 perf_evsel__object.fini(evsel);
1178 void perf_evsel__delete(struct perf_evsel *evsel)
1180 perf_evsel__exit(evsel);
1184 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1185 struct perf_counts_values *count)
1187 struct perf_counts_values tmp;
1189 if (!evsel->prev_raw_counts)
1193 tmp = evsel->prev_raw_counts->aggr;
1194 evsel->prev_raw_counts->aggr = *count;
1196 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1197 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1200 count->val = count->val - tmp.val;
1201 count->ena = count->ena - tmp.ena;
1202 count->run = count->run - tmp.run;
1205 void perf_counts_values__scale(struct perf_counts_values *count,
1206 bool scale, s8 *pscaled)
1211 if (count->run == 0) {
1214 } else if (count->run < count->ena) {
1216 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1219 count->ena = count->run = 0;
1225 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1226 struct perf_counts_values *count)
1228 memset(count, 0, sizeof(*count));
1230 if (FD(evsel, cpu, thread) < 0)
1233 if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1239 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1240 int cpu, int thread, bool scale)
1242 struct perf_counts_values count;
1243 size_t nv = scale ? 3 : 1;
1245 if (FD(evsel, cpu, thread) < 0)
1248 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1251 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1254 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1255 perf_counts_values__scale(&count, scale, NULL);
1256 *perf_counts(evsel->counts, cpu, thread) = count;
1260 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1262 struct perf_evsel *leader = evsel->leader;
1265 if (perf_evsel__is_group_leader(evsel))
1269 * Leader must be already processed/open,
1270 * if not it's a bug.
1272 BUG_ON(!leader->fd);
1274 fd = FD(leader, cpu, thread);
1285 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1287 bool first_bit = true;
1291 if (value & bits[i].bit) {
1292 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1295 } while (bits[++i].name != NULL);
1298 static void __p_sample_type(char *buf, size_t size, u64 value)
1300 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1301 struct bit_names bits[] = {
1302 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1303 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1304 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1305 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1306 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1311 __p_bits(buf, size, value, bits);
1314 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1316 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1317 struct bit_names bits[] = {
1318 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1319 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1320 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1321 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1322 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1326 __p_bits(buf, size, value, bits);
1329 static void __p_read_format(char *buf, size_t size, u64 value)
1331 #define bit_name(n) { PERF_FORMAT_##n, #n }
1332 struct bit_names bits[] = {
1333 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1334 bit_name(ID), bit_name(GROUP),
1338 __p_bits(buf, size, value, bits);
1341 #define BUF_SIZE 1024
1343 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1344 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1345 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1346 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1347 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1348 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1350 #define PRINT_ATTRn(_n, _f, _p) \
1354 ret += attr__fprintf(fp, _n, buf, priv);\
1358 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1360 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1361 attr__fprintf_f attr__fprintf, void *priv)
1366 PRINT_ATTRf(type, p_unsigned);
1367 PRINT_ATTRf(size, p_unsigned);
1368 PRINT_ATTRf(config, p_hex);
1369 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1370 PRINT_ATTRf(sample_type, p_sample_type);
1371 PRINT_ATTRf(read_format, p_read_format);
1373 PRINT_ATTRf(disabled, p_unsigned);
1374 PRINT_ATTRf(inherit, p_unsigned);
1375 PRINT_ATTRf(pinned, p_unsigned);
1376 PRINT_ATTRf(exclusive, p_unsigned);
1377 PRINT_ATTRf(exclude_user, p_unsigned);
1378 PRINT_ATTRf(exclude_kernel, p_unsigned);
1379 PRINT_ATTRf(exclude_hv, p_unsigned);
1380 PRINT_ATTRf(exclude_idle, p_unsigned);
1381 PRINT_ATTRf(mmap, p_unsigned);
1382 PRINT_ATTRf(comm, p_unsigned);
1383 PRINT_ATTRf(freq, p_unsigned);
1384 PRINT_ATTRf(inherit_stat, p_unsigned);
1385 PRINT_ATTRf(enable_on_exec, p_unsigned);
1386 PRINT_ATTRf(task, p_unsigned);
1387 PRINT_ATTRf(watermark, p_unsigned);
1388 PRINT_ATTRf(precise_ip, p_unsigned);
1389 PRINT_ATTRf(mmap_data, p_unsigned);
1390 PRINT_ATTRf(sample_id_all, p_unsigned);
1391 PRINT_ATTRf(exclude_host, p_unsigned);
1392 PRINT_ATTRf(exclude_guest, p_unsigned);
1393 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1394 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1395 PRINT_ATTRf(mmap2, p_unsigned);
1396 PRINT_ATTRf(comm_exec, p_unsigned);
1397 PRINT_ATTRf(use_clockid, p_unsigned);
1398 PRINT_ATTRf(context_switch, p_unsigned);
1399 PRINT_ATTRf(write_backward, p_unsigned);
1401 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1402 PRINT_ATTRf(bp_type, p_unsigned);
1403 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1404 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1405 PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1406 PRINT_ATTRf(sample_regs_user, p_hex);
1407 PRINT_ATTRf(sample_stack_user, p_unsigned);
1408 PRINT_ATTRf(clockid, p_signed);
1409 PRINT_ATTRf(sample_regs_intr, p_hex);
1410 PRINT_ATTRf(aux_watermark, p_unsigned);
1411 PRINT_ATTRf(sample_max_stack, p_unsigned);
1416 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1417 void *priv __attribute__((unused)))
1419 return fprintf(fp, " %-32s %s\n", name, val);
1422 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1423 struct thread_map *threads)
1425 int cpu, thread, nthreads;
1426 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1428 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1430 if (perf_missing_features.write_backward && evsel->attr.write_backward)
1433 if (evsel->system_wide)
1436 nthreads = threads->nr;
1438 if (evsel->fd == NULL &&
1439 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1443 flags |= PERF_FLAG_PID_CGROUP;
1444 pid = evsel->cgrp->fd;
1447 fallback_missing_features:
1448 if (perf_missing_features.clockid_wrong)
1449 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1450 if (perf_missing_features.clockid) {
1451 evsel->attr.use_clockid = 0;
1452 evsel->attr.clockid = 0;
1454 if (perf_missing_features.cloexec)
1455 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1456 if (perf_missing_features.mmap2)
1457 evsel->attr.mmap2 = 0;
1458 if (perf_missing_features.exclude_guest)
1459 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1460 if (perf_missing_features.lbr_flags)
1461 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1462 PERF_SAMPLE_BRANCH_NO_CYCLES);
1464 if (perf_missing_features.sample_id_all)
1465 evsel->attr.sample_id_all = 0;
1468 fprintf(stderr, "%.60s\n", graph_dotted_line);
1469 fprintf(stderr, "perf_event_attr:\n");
1470 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1471 fprintf(stderr, "%.60s\n", graph_dotted_line);
1474 for (cpu = 0; cpu < cpus->nr; cpu++) {
1476 for (thread = 0; thread < nthreads; thread++) {
1479 if (!evsel->cgrp && !evsel->system_wide)
1480 pid = thread_map__pid(threads, thread);
1482 group_fd = get_group_fd(evsel, cpu, thread);
1484 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1485 pid, cpus->map[cpu], group_fd, flags);
1487 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1491 if (FD(evsel, cpu, thread) < 0) {
1493 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1498 pr_debug2(" = %d\n", FD(evsel, cpu, thread));
1500 if (evsel->bpf_fd >= 0) {
1501 int evt_fd = FD(evsel, cpu, thread);
1502 int bpf_fd = evsel->bpf_fd;
1505 PERF_EVENT_IOC_SET_BPF,
1507 if (err && errno != EEXIST) {
1508 pr_err("failed to attach bpf fd %d: %s\n",
1509 bpf_fd, strerror(errno));
1515 set_rlimit = NO_CHANGE;
1518 * If we succeeded but had to kill clockid, fail and
1519 * have perf_evsel__open_strerror() print us a nice
1522 if (perf_missing_features.clockid ||
1523 perf_missing_features.clockid_wrong) {
1534 * perf stat needs between 5 and 22 fds per CPU. When we run out
1535 * of them try to increase the limits.
1537 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1539 int old_errno = errno;
1541 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1542 if (set_rlimit == NO_CHANGE)
1543 l.rlim_cur = l.rlim_max;
1545 l.rlim_cur = l.rlim_max + 1000;
1546 l.rlim_max = l.rlim_cur;
1548 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1557 if (err != -EINVAL || cpu > 0 || thread > 0)
1561 * Must probe features in the order they were added to the
1562 * perf_event_attr interface.
1564 if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1565 perf_missing_features.write_backward = true;
1567 } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1568 perf_missing_features.clockid_wrong = true;
1569 goto fallback_missing_features;
1570 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1571 perf_missing_features.clockid = true;
1572 goto fallback_missing_features;
1573 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1574 perf_missing_features.cloexec = true;
1575 goto fallback_missing_features;
1576 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1577 perf_missing_features.mmap2 = true;
1578 goto fallback_missing_features;
1579 } else if (!perf_missing_features.exclude_guest &&
1580 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1581 perf_missing_features.exclude_guest = true;
1582 goto fallback_missing_features;
1583 } else if (!perf_missing_features.sample_id_all) {
1584 perf_missing_features.sample_id_all = true;
1585 goto retry_sample_id;
1586 } else if (!perf_missing_features.lbr_flags &&
1587 (evsel->attr.branch_sample_type &
1588 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1589 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1590 perf_missing_features.lbr_flags = true;
1591 goto fallback_missing_features;
1595 while (--thread >= 0) {
1596 close(FD(evsel, cpu, thread));
1597 FD(evsel, cpu, thread) = -1;
1600 } while (--cpu >= 0);
1604 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1606 if (evsel->fd == NULL)
1609 perf_evsel__close_fd(evsel, ncpus, nthreads);
1610 perf_evsel__free_fd(evsel);
1622 struct thread_map map;
1624 } empty_thread_map = {
1629 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1630 struct thread_map *threads)
1633 /* Work around old compiler warnings about strict aliasing */
1634 cpus = &empty_cpu_map.map;
1637 if (threads == NULL)
1638 threads = &empty_thread_map.map;
1640 return __perf_evsel__open(evsel, cpus, threads);
1643 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1644 struct cpu_map *cpus)
1646 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1649 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1650 struct thread_map *threads)
1652 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1655 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1656 const union perf_event *event,
1657 struct perf_sample *sample)
1659 u64 type = evsel->attr.sample_type;
1660 const u64 *array = event->sample.array;
1661 bool swapped = evsel->needs_swap;
1664 array += ((event->header.size -
1665 sizeof(event->header)) / sizeof(u64)) - 1;
1667 if (type & PERF_SAMPLE_IDENTIFIER) {
1668 sample->id = *array;
1672 if (type & PERF_SAMPLE_CPU) {
1675 /* undo swap of u64, then swap on individual u32s */
1676 u.val64 = bswap_64(u.val64);
1677 u.val32[0] = bswap_32(u.val32[0]);
1680 sample->cpu = u.val32[0];
1684 if (type & PERF_SAMPLE_STREAM_ID) {
1685 sample->stream_id = *array;
1689 if (type & PERF_SAMPLE_ID) {
1690 sample->id = *array;
1694 if (type & PERF_SAMPLE_TIME) {
1695 sample->time = *array;
1699 if (type & PERF_SAMPLE_TID) {
1702 /* undo swap of u64, then swap on individual u32s */
1703 u.val64 = bswap_64(u.val64);
1704 u.val32[0] = bswap_32(u.val32[0]);
1705 u.val32[1] = bswap_32(u.val32[1]);
1708 sample->pid = u.val32[0];
1709 sample->tid = u.val32[1];
1716 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1719 return size > max_size || offset + size > endp;
1722 #define OVERFLOW_CHECK(offset, size, max_size) \
1724 if (overflow(endp, (max_size), (offset), (size))) \
1728 #define OVERFLOW_CHECK_u64(offset) \
1729 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1731 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1732 struct perf_sample *data)
1734 u64 type = evsel->attr.sample_type;
1735 bool swapped = evsel->needs_swap;
1737 u16 max_size = event->header.size;
1738 const void *endp = (void *)event + max_size;
1742 * used for cross-endian analysis. See git commit 65014ab3
1743 * for why this goofiness is needed.
1747 memset(data, 0, sizeof(*data));
1748 data->cpu = data->pid = data->tid = -1;
1749 data->stream_id = data->id = data->time = -1ULL;
1750 data->period = evsel->attr.sample_period;
1751 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1753 if (event->header.type != PERF_RECORD_SAMPLE) {
1754 if (!evsel->attr.sample_id_all)
1756 return perf_evsel__parse_id_sample(evsel, event, data);
1759 array = event->sample.array;
1762 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1763 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1764 * check the format does not go past the end of the event.
1766 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1770 if (type & PERF_SAMPLE_IDENTIFIER) {
1775 if (type & PERF_SAMPLE_IP) {
1780 if (type & PERF_SAMPLE_TID) {
1783 /* undo swap of u64, then swap on individual u32s */
1784 u.val64 = bswap_64(u.val64);
1785 u.val32[0] = bswap_32(u.val32[0]);
1786 u.val32[1] = bswap_32(u.val32[1]);
1789 data->pid = u.val32[0];
1790 data->tid = u.val32[1];
1794 if (type & PERF_SAMPLE_TIME) {
1795 data->time = *array;
1800 if (type & PERF_SAMPLE_ADDR) {
1801 data->addr = *array;
1805 if (type & PERF_SAMPLE_ID) {
1810 if (type & PERF_SAMPLE_STREAM_ID) {
1811 data->stream_id = *array;
1815 if (type & PERF_SAMPLE_CPU) {
1819 /* undo swap of u64, then swap on individual u32s */
1820 u.val64 = bswap_64(u.val64);
1821 u.val32[0] = bswap_32(u.val32[0]);
1824 data->cpu = u.val32[0];
1828 if (type & PERF_SAMPLE_PERIOD) {
1829 data->period = *array;
1833 if (type & PERF_SAMPLE_READ) {
1834 u64 read_format = evsel->attr.read_format;
1836 OVERFLOW_CHECK_u64(array);
1837 if (read_format & PERF_FORMAT_GROUP)
1838 data->read.group.nr = *array;
1840 data->read.one.value = *array;
1844 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1845 OVERFLOW_CHECK_u64(array);
1846 data->read.time_enabled = *array;
1850 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1851 OVERFLOW_CHECK_u64(array);
1852 data->read.time_running = *array;
1856 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1857 if (read_format & PERF_FORMAT_GROUP) {
1858 const u64 max_group_nr = UINT64_MAX /
1859 sizeof(struct sample_read_value);
1861 if (data->read.group.nr > max_group_nr)
1863 sz = data->read.group.nr *
1864 sizeof(struct sample_read_value);
1865 OVERFLOW_CHECK(array, sz, max_size);
1866 data->read.group.values =
1867 (struct sample_read_value *)array;
1868 array = (void *)array + sz;
1870 OVERFLOW_CHECK_u64(array);
1871 data->read.one.id = *array;
1876 if (type & PERF_SAMPLE_CALLCHAIN) {
1877 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1879 OVERFLOW_CHECK_u64(array);
1880 data->callchain = (struct ip_callchain *)array++;
1881 if (data->callchain->nr > max_callchain_nr)
1883 sz = data->callchain->nr * sizeof(u64);
1884 OVERFLOW_CHECK(array, sz, max_size);
1885 array = (void *)array + sz;
1888 if (type & PERF_SAMPLE_RAW) {
1889 OVERFLOW_CHECK_u64(array);
1891 if (WARN_ONCE(swapped,
1892 "Endianness of raw data not corrected!\n")) {
1893 /* undo swap of u64, then swap on individual u32s */
1894 u.val64 = bswap_64(u.val64);
1895 u.val32[0] = bswap_32(u.val32[0]);
1896 u.val32[1] = bswap_32(u.val32[1]);
1898 data->raw_size = u.val32[0];
1899 array = (void *)array + sizeof(u32);
1901 OVERFLOW_CHECK(array, data->raw_size, max_size);
1902 data->raw_data = (void *)array;
1903 array = (void *)array + data->raw_size;
1906 if (type & PERF_SAMPLE_BRANCH_STACK) {
1907 const u64 max_branch_nr = UINT64_MAX /
1908 sizeof(struct branch_entry);
1910 OVERFLOW_CHECK_u64(array);
1911 data->branch_stack = (struct branch_stack *)array++;
1913 if (data->branch_stack->nr > max_branch_nr)
1915 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1916 OVERFLOW_CHECK(array, sz, max_size);
1917 array = (void *)array + sz;
1920 if (type & PERF_SAMPLE_REGS_USER) {
1921 OVERFLOW_CHECK_u64(array);
1922 data->user_regs.abi = *array;
1925 if (data->user_regs.abi) {
1926 u64 mask = evsel->attr.sample_regs_user;
1928 sz = hweight_long(mask) * sizeof(u64);
1929 OVERFLOW_CHECK(array, sz, max_size);
1930 data->user_regs.mask = mask;
1931 data->user_regs.regs = (u64 *)array;
1932 array = (void *)array + sz;
1936 if (type & PERF_SAMPLE_STACK_USER) {
1937 OVERFLOW_CHECK_u64(array);
1940 data->user_stack.offset = ((char *)(array - 1)
1944 data->user_stack.size = 0;
1946 OVERFLOW_CHECK(array, sz, max_size);
1947 data->user_stack.data = (char *)array;
1948 array = (void *)array + sz;
1949 OVERFLOW_CHECK_u64(array);
1950 data->user_stack.size = *array++;
1951 if (WARN_ONCE(data->user_stack.size > sz,
1952 "user stack dump failure\n"))
1957 if (type & PERF_SAMPLE_WEIGHT) {
1958 OVERFLOW_CHECK_u64(array);
1959 data->weight = *array;
1963 data->data_src = PERF_MEM_DATA_SRC_NONE;
1964 if (type & PERF_SAMPLE_DATA_SRC) {
1965 OVERFLOW_CHECK_u64(array);
1966 data->data_src = *array;
1970 data->transaction = 0;
1971 if (type & PERF_SAMPLE_TRANSACTION) {
1972 OVERFLOW_CHECK_u64(array);
1973 data->transaction = *array;
1977 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1978 if (type & PERF_SAMPLE_REGS_INTR) {
1979 OVERFLOW_CHECK_u64(array);
1980 data->intr_regs.abi = *array;
1983 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1984 u64 mask = evsel->attr.sample_regs_intr;
1986 sz = hweight_long(mask) * sizeof(u64);
1987 OVERFLOW_CHECK(array, sz, max_size);
1988 data->intr_regs.mask = mask;
1989 data->intr_regs.regs = (u64 *)array;
1990 array = (void *)array + sz;
1997 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2000 size_t sz, result = sizeof(struct sample_event);
2002 if (type & PERF_SAMPLE_IDENTIFIER)
2003 result += sizeof(u64);
2005 if (type & PERF_SAMPLE_IP)
2006 result += sizeof(u64);
2008 if (type & PERF_SAMPLE_TID)
2009 result += sizeof(u64);
2011 if (type & PERF_SAMPLE_TIME)
2012 result += sizeof(u64);
2014 if (type & PERF_SAMPLE_ADDR)
2015 result += sizeof(u64);
2017 if (type & PERF_SAMPLE_ID)
2018 result += sizeof(u64);
2020 if (type & PERF_SAMPLE_STREAM_ID)
2021 result += sizeof(u64);
2023 if (type & PERF_SAMPLE_CPU)
2024 result += sizeof(u64);
2026 if (type & PERF_SAMPLE_PERIOD)
2027 result += sizeof(u64);
2029 if (type & PERF_SAMPLE_READ) {
2030 result += sizeof(u64);
2031 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2032 result += sizeof(u64);
2033 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2034 result += sizeof(u64);
2035 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2036 if (read_format & PERF_FORMAT_GROUP) {
2037 sz = sample->read.group.nr *
2038 sizeof(struct sample_read_value);
2041 result += sizeof(u64);
2045 if (type & PERF_SAMPLE_CALLCHAIN) {
2046 sz = (sample->callchain->nr + 1) * sizeof(u64);
2050 if (type & PERF_SAMPLE_RAW) {
2051 result += sizeof(u32);
2052 result += sample->raw_size;
2055 if (type & PERF_SAMPLE_BRANCH_STACK) {
2056 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2061 if (type & PERF_SAMPLE_REGS_USER) {
2062 if (sample->user_regs.abi) {
2063 result += sizeof(u64);
2064 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2067 result += sizeof(u64);
2071 if (type & PERF_SAMPLE_STACK_USER) {
2072 sz = sample->user_stack.size;
2073 result += sizeof(u64);
2076 result += sizeof(u64);
2080 if (type & PERF_SAMPLE_WEIGHT)
2081 result += sizeof(u64);
2083 if (type & PERF_SAMPLE_DATA_SRC)
2084 result += sizeof(u64);
2086 if (type & PERF_SAMPLE_TRANSACTION)
2087 result += sizeof(u64);
2089 if (type & PERF_SAMPLE_REGS_INTR) {
2090 if (sample->intr_regs.abi) {
2091 result += sizeof(u64);
2092 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2095 result += sizeof(u64);
2102 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2104 const struct perf_sample *sample,
2110 * used for cross-endian analysis. See git commit 65014ab3
2111 * for why this goofiness is needed.
2115 array = event->sample.array;
2117 if (type & PERF_SAMPLE_IDENTIFIER) {
2118 *array = sample->id;
2122 if (type & PERF_SAMPLE_IP) {
2123 *array = sample->ip;
2127 if (type & PERF_SAMPLE_TID) {
2128 u.val32[0] = sample->pid;
2129 u.val32[1] = sample->tid;
2132 * Inverse of what is done in perf_evsel__parse_sample
2134 u.val32[0] = bswap_32(u.val32[0]);
2135 u.val32[1] = bswap_32(u.val32[1]);
2136 u.val64 = bswap_64(u.val64);
2143 if (type & PERF_SAMPLE_TIME) {
2144 *array = sample->time;
2148 if (type & PERF_SAMPLE_ADDR) {
2149 *array = sample->addr;
2153 if (type & PERF_SAMPLE_ID) {
2154 *array = sample->id;
2158 if (type & PERF_SAMPLE_STREAM_ID) {
2159 *array = sample->stream_id;
2163 if (type & PERF_SAMPLE_CPU) {
2164 u.val32[0] = sample->cpu;
2167 * Inverse of what is done in perf_evsel__parse_sample
2169 u.val32[0] = bswap_32(u.val32[0]);
2170 u.val64 = bswap_64(u.val64);
2176 if (type & PERF_SAMPLE_PERIOD) {
2177 *array = sample->period;
2181 if (type & PERF_SAMPLE_READ) {
2182 if (read_format & PERF_FORMAT_GROUP)
2183 *array = sample->read.group.nr;
2185 *array = sample->read.one.value;
2188 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2189 *array = sample->read.time_enabled;
2193 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2194 *array = sample->read.time_running;
2198 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2199 if (read_format & PERF_FORMAT_GROUP) {
2200 sz = sample->read.group.nr *
2201 sizeof(struct sample_read_value);
2202 memcpy(array, sample->read.group.values, sz);
2203 array = (void *)array + sz;
2205 *array = sample->read.one.id;
2210 if (type & PERF_SAMPLE_CALLCHAIN) {
2211 sz = (sample->callchain->nr + 1) * sizeof(u64);
2212 memcpy(array, sample->callchain, sz);
2213 array = (void *)array + sz;
2216 if (type & PERF_SAMPLE_RAW) {
2217 u.val32[0] = sample->raw_size;
2218 if (WARN_ONCE(swapped,
2219 "Endianness of raw data not corrected!\n")) {
2221 * Inverse of what is done in perf_evsel__parse_sample
2223 u.val32[0] = bswap_32(u.val32[0]);
2224 u.val32[1] = bswap_32(u.val32[1]);
2225 u.val64 = bswap_64(u.val64);
2228 array = (void *)array + sizeof(u32);
2230 memcpy(array, sample->raw_data, sample->raw_size);
2231 array = (void *)array + sample->raw_size;
2234 if (type & PERF_SAMPLE_BRANCH_STACK) {
2235 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2237 memcpy(array, sample->branch_stack, sz);
2238 array = (void *)array + sz;
2241 if (type & PERF_SAMPLE_REGS_USER) {
2242 if (sample->user_regs.abi) {
2243 *array++ = sample->user_regs.abi;
2244 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2245 memcpy(array, sample->user_regs.regs, sz);
2246 array = (void *)array + sz;
2252 if (type & PERF_SAMPLE_STACK_USER) {
2253 sz = sample->user_stack.size;
2256 memcpy(array, sample->user_stack.data, sz);
2257 array = (void *)array + sz;
2262 if (type & PERF_SAMPLE_WEIGHT) {
2263 *array = sample->weight;
2267 if (type & PERF_SAMPLE_DATA_SRC) {
2268 *array = sample->data_src;
2272 if (type & PERF_SAMPLE_TRANSACTION) {
2273 *array = sample->transaction;
2277 if (type & PERF_SAMPLE_REGS_INTR) {
2278 if (sample->intr_regs.abi) {
2279 *array++ = sample->intr_regs.abi;
2280 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2281 memcpy(array, sample->intr_regs.regs, sz);
2282 array = (void *)array + sz;
2291 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2293 return pevent_find_field(evsel->tp_format, name);
2296 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2299 struct format_field *field = perf_evsel__field(evsel, name);
2305 offset = field->offset;
2307 if (field->flags & FIELD_IS_DYNAMIC) {
2308 offset = *(int *)(sample->raw_data + field->offset);
2312 return sample->raw_data + offset;
2315 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2319 void *ptr = sample->raw_data + field->offset;
2321 switch (field->size) {
2325 value = *(u16 *)ptr;
2328 value = *(u32 *)ptr;
2331 memcpy(&value, ptr, sizeof(u64));
2340 switch (field->size) {
2342 return bswap_16(value);
2344 return bswap_32(value);
2346 return bswap_64(value);
2354 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2357 struct format_field *field = perf_evsel__field(evsel, name);
2362 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2365 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2366 char *msg, size_t msgsize)
2370 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2371 evsel->attr.type == PERF_TYPE_HARDWARE &&
2372 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2374 * If it's cycles then fall back to hrtimer based
2375 * cpu-clock-tick sw counter, which is always available even if
2378 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2381 scnprintf(msg, msgsize, "%s",
2382 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2384 evsel->attr.type = PERF_TYPE_SOFTWARE;
2385 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2387 zfree(&evsel->name);
2389 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2390 (paranoid = perf_event_paranoid()) > 1) {
2391 const char *name = perf_evsel__name(evsel);
2394 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2399 evsel->name = new_name;
2400 scnprintf(msg, msgsize,
2401 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2402 evsel->attr.exclude_kernel = 1;
2410 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2411 int err, char *msg, size_t size)
2413 char sbuf[STRERR_BUFSIZE];
2418 return scnprintf(msg, size,
2419 "You may not have permission to collect %sstats.\n\n"
2420 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2421 "which controls use of the performance events system by\n"
2422 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2423 "The current value is %d:\n\n"
2424 " -1: Allow use of (almost) all events by all users\n"
2425 ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2426 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2427 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2428 target->system_wide ? "system-wide " : "",
2429 perf_event_paranoid());
2431 return scnprintf(msg, size, "The %s event is not supported.",
2432 perf_evsel__name(evsel));
2434 return scnprintf(msg, size, "%s",
2435 "Too many events are opened.\n"
2436 "Probably the maximum number of open file descriptors has been reached.\n"
2437 "Hint: Try again after reducing the number of events.\n"
2438 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2440 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2441 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2442 return scnprintf(msg, size,
2443 "Not enough memory to setup event with callchain.\n"
2444 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2445 "Hint: Current value: %d", sysctl_perf_event_max_stack);
2448 if (target->cpu_list)
2449 return scnprintf(msg, size, "%s",
2450 "No such device - did you specify an out-of-range profile CPU?");
2453 if (evsel->attr.sample_period != 0)
2454 return scnprintf(msg, size, "%s",
2455 "PMU Hardware doesn't support sampling/overflow-interrupts.");
2456 if (evsel->attr.precise_ip)
2457 return scnprintf(msg, size, "%s",
2458 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2459 #if defined(__i386__) || defined(__x86_64__)
2460 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2461 return scnprintf(msg, size, "%s",
2462 "No hardware sampling interrupt available.\n"
2463 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2467 if (find_process("oprofiled"))
2468 return scnprintf(msg, size,
2469 "The PMU counters are busy/taken by another profiler.\n"
2470 "We found oprofile daemon running, please stop it and try again.");
2473 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2474 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2475 if (perf_missing_features.clockid)
2476 return scnprintf(msg, size, "clockid feature not supported.");
2477 if (perf_missing_features.clockid_wrong)
2478 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2484 return scnprintf(msg, size,
2485 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2486 "/bin/dmesg may provide additional information.\n"
2487 "No CONFIG_PERF_EVENTS=y kernel support configured?",
2488 err, str_error_r(err, sbuf, sizeof(sbuf)),
2489 perf_evsel__name(evsel));
2492 char *perf_evsel__env_arch(struct perf_evsel *evsel)
2494 if (evsel && evsel->evlist && evsel->evlist->env)
2495 return evsel->evlist->env->arch;