1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
5 * Parts came from builtin-{top,stat,record}.c, see those files for further
12 #include <linux/bitops.h>
13 #include <api/fs/fs.h>
14 #include <api/fs/tracing_path.h>
15 #include <traceevent/event-parse.h>
16 #include <linux/hw_breakpoint.h>
17 #include <linux/perf_event.h>
18 #include <linux/compiler.h>
19 #include <linux/err.h>
20 #include <linux/zalloc.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
26 #include <perf/evsel.h>
28 #include "bpf_counter.h"
29 #include "callchain.h"
35 #include "util/evsel_config.h"
36 #include "util/evsel_fprintf.h"
38 #include <perf/cpumap.h>
39 #include "thread_map.h"
41 #include "perf_regs.h"
44 #include "trace-event.h"
50 #include "pmu-hybrid.h"
51 #include "../perf-sys.h"
52 #include "util/parse-branch-options.h"
53 #include <internal/xyarray.h>
54 #include <internal/lib.h>
56 #include <linux/ctype.h>
58 struct perf_missing_features perf_missing_features;
60 static clockid_t clockid;
62 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
67 void __weak test_attr__ready(void) { }
69 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
75 int (*init)(struct evsel *evsel);
76 void (*fini)(struct evsel *evsel);
77 } perf_evsel__object = {
78 .size = sizeof(struct evsel),
79 .init = evsel__no_extra_init,
80 .fini = evsel__no_extra_fini,
83 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
84 void (*fini)(struct evsel *evsel))
90 if (perf_evsel__object.size > object_size)
93 perf_evsel__object.size = object_size;
97 perf_evsel__object.init = init;
100 perf_evsel__object.fini = fini;
105 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
107 int __evsel__sample_size(u64 sample_type)
109 u64 mask = sample_type & PERF_SAMPLE_MASK;
113 for (i = 0; i < 64; i++) {
114 if (mask & (1ULL << i))
124 * __perf_evsel__calc_id_pos - calculate id_pos.
125 * @sample_type: sample type
127 * This function returns the position of the event id (PERF_SAMPLE_ID or
128 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
129 * perf_record_sample.
131 static int __perf_evsel__calc_id_pos(u64 sample_type)
135 if (sample_type & PERF_SAMPLE_IDENTIFIER)
138 if (!(sample_type & PERF_SAMPLE_ID))
141 if (sample_type & PERF_SAMPLE_IP)
144 if (sample_type & PERF_SAMPLE_TID)
147 if (sample_type & PERF_SAMPLE_TIME)
150 if (sample_type & PERF_SAMPLE_ADDR)
157 * __perf_evsel__calc_is_pos - calculate is_pos.
158 * @sample_type: sample type
160 * This function returns the position (counting backwards) of the event id
161 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
162 * sample_id_all is used there is an id sample appended to non-sample events.
164 static int __perf_evsel__calc_is_pos(u64 sample_type)
168 if (sample_type & PERF_SAMPLE_IDENTIFIER)
171 if (!(sample_type & PERF_SAMPLE_ID))
174 if (sample_type & PERF_SAMPLE_CPU)
177 if (sample_type & PERF_SAMPLE_STREAM_ID)
183 void evsel__calc_id_pos(struct evsel *evsel)
185 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
186 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
189 void __evsel__set_sample_bit(struct evsel *evsel,
190 enum perf_event_sample_format bit)
192 if (!(evsel->core.attr.sample_type & bit)) {
193 evsel->core.attr.sample_type |= bit;
194 evsel->sample_size += sizeof(u64);
195 evsel__calc_id_pos(evsel);
199 void __evsel__reset_sample_bit(struct evsel *evsel,
200 enum perf_event_sample_format bit)
202 if (evsel->core.attr.sample_type & bit) {
203 evsel->core.attr.sample_type &= ~bit;
204 evsel->sample_size -= sizeof(u64);
205 evsel__calc_id_pos(evsel);
209 void evsel__set_sample_id(struct evsel *evsel,
210 bool can_sample_identifier)
212 if (can_sample_identifier) {
213 evsel__reset_sample_bit(evsel, ID);
214 evsel__set_sample_bit(evsel, IDENTIFIER);
216 evsel__set_sample_bit(evsel, ID);
218 evsel->core.attr.read_format |= PERF_FORMAT_ID;
222 * evsel__is_function_event - Return whether given evsel is a function
225 * @evsel - evsel selector to be tested
227 * Return %true if event is function trace event
229 bool evsel__is_function_event(struct evsel *evsel)
231 #define FUNCTION_EVENT "ftrace:function"
233 return evsel->name &&
234 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
236 #undef FUNCTION_EVENT
239 void evsel__init(struct evsel *evsel,
240 struct perf_event_attr *attr, int idx)
242 perf_evsel__init(&evsel->core, attr);
244 evsel->tracking = !idx;
245 evsel->leader = evsel;
248 evsel->max_events = ULONG_MAX;
249 evsel->evlist = NULL;
250 evsel->bpf_obj = NULL;
252 INIT_LIST_HEAD(&evsel->config_terms);
253 INIT_LIST_HEAD(&evsel->bpf_counter_list);
254 perf_evsel__object.init(evsel);
255 evsel->sample_size = __evsel__sample_size(attr->sample_type);
256 evsel__calc_id_pos(evsel);
257 evsel->cmdline_group_boundary = false;
258 evsel->metric_expr = NULL;
259 evsel->metric_name = NULL;
260 evsel->metric_events = NULL;
261 evsel->per_pkg_mask = NULL;
262 evsel->collect_stat = false;
263 evsel->pmu_name = NULL;
266 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
268 struct evsel *evsel = zalloc(perf_evsel__object.size);
272 evsel__init(evsel, attr, idx);
274 if (evsel__is_bpf_output(evsel)) {
275 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
276 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
277 evsel->core.attr.sample_period = 1;
280 if (evsel__is_clock(evsel)) {
282 * The evsel->unit points to static alias->unit
283 * so it's ok to use static string in here.
285 static const char *unit = "msec";
294 static bool perf_event_can_profile_kernel(void)
296 return perf_event_paranoid_check(1);
299 struct evsel *evsel__new_cycles(bool precise, __u32 type, __u64 config)
301 struct perf_event_attr attr = {
304 .exclude_kernel = !perf_event_can_profile_kernel(),
308 event_attr_init(&attr);
314 * Now let the usual logic to set up the perf_event_attr defaults
315 * to kick in when we return and before perf_evsel__open() is called.
318 evsel = evsel__new(&attr);
322 evsel->precise_max = true;
324 /* use asprintf() because free(evsel) assumes name is allocated */
325 if (asprintf(&evsel->name, "cycles%s%s%.*s",
326 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
327 attr.exclude_kernel ? "u" : "",
328 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
333 evsel__delete(evsel);
338 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
340 struct evsel_config_term *pos, *tmp;
342 list_for_each_entry(pos, &src->config_terms, list) {
343 tmp = malloc(sizeof(*tmp));
349 tmp->val.str = strdup(pos->val.str);
350 if (tmp->val.str == NULL) {
355 list_add_tail(&tmp->list, &dst->config_terms);
361 * evsel__clone - create a new evsel copied from @orig
362 * @orig: original evsel
364 * The assumption is that @orig is not configured nor opened yet.
365 * So we only care about the attributes that can be set while it's parsed.
367 struct evsel *evsel__clone(struct evsel *orig)
371 BUG_ON(orig->core.fd);
372 BUG_ON(orig->counts);
374 BUG_ON(orig->per_pkg_mask);
376 /* cannot handle BPF objects for now */
380 evsel = evsel__new(&orig->core.attr);
384 evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
385 evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
386 evsel->core.threads = perf_thread_map__get(orig->core.threads);
387 evsel->core.nr_members = orig->core.nr_members;
388 evsel->core.system_wide = orig->core.system_wide;
391 evsel->name = strdup(orig->name);
392 if (evsel->name == NULL)
395 if (orig->group_name) {
396 evsel->group_name = strdup(orig->group_name);
397 if (evsel->group_name == NULL)
400 if (orig->pmu_name) {
401 evsel->pmu_name = strdup(orig->pmu_name);
402 if (evsel->pmu_name == NULL)
406 evsel->filter = strdup(orig->filter);
407 if (evsel->filter == NULL)
410 evsel->cgrp = cgroup__get(orig->cgrp);
411 evsel->tp_format = orig->tp_format;
412 evsel->handler = orig->handler;
413 evsel->leader = orig->leader;
415 evsel->max_events = orig->max_events;
416 evsel->tool_event = orig->tool_event;
417 evsel->unit = orig->unit;
418 evsel->scale = orig->scale;
419 evsel->snapshot = orig->snapshot;
420 evsel->per_pkg = orig->per_pkg;
421 evsel->percore = orig->percore;
422 evsel->precise_max = orig->precise_max;
423 evsel->use_uncore_alias = orig->use_uncore_alias;
424 evsel->is_libpfm_event = orig->is_libpfm_event;
426 evsel->exclude_GH = orig->exclude_GH;
427 evsel->sample_read = orig->sample_read;
428 evsel->auto_merge_stats = orig->auto_merge_stats;
429 evsel->collect_stat = orig->collect_stat;
430 evsel->weak_group = orig->weak_group;
432 if (evsel__copy_config_terms(evsel, orig) < 0)
438 evsel__delete(evsel);
443 * Returns pointer with encoded error via <linux/err.h> interface.
445 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
447 struct evsel *evsel = zalloc(perf_evsel__object.size);
453 struct perf_event_attr attr = {
454 .type = PERF_TYPE_TRACEPOINT,
455 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
456 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
459 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
462 evsel->tp_format = trace_event__tp_format(sys, name);
463 if (IS_ERR(evsel->tp_format)) {
464 err = PTR_ERR(evsel->tp_format);
468 event_attr_init(&attr);
469 attr.config = evsel->tp_format->id;
470 attr.sample_period = 1;
471 evsel__init(evsel, &attr, idx);
483 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
491 "stalled-cycles-frontend",
492 "stalled-cycles-backend",
496 char *evsel__bpf_counter_events;
498 bool evsel__match_bpf_counter_events(const char *name)
504 if (!evsel__bpf_counter_events)
507 ptr = strstr(evsel__bpf_counter_events, name);
508 name_len = strlen(name);
510 /* check name matches a full token in evsel__bpf_counter_events */
511 match = (ptr != NULL) &&
512 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
513 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
518 static const char *__evsel__hw_name(u64 config)
520 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
521 return evsel__hw_names[config];
523 return "unknown-hardware";
526 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
528 int colon = 0, r = 0;
529 struct perf_event_attr *attr = &evsel->core.attr;
530 bool exclude_guest_default = false;
532 #define MOD_PRINT(context, mod) do { \
533 if (!attr->exclude_##context) { \
534 if (!colon) colon = ++r; \
535 r += scnprintf(bf + r, size - r, "%c", mod); \
538 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
539 MOD_PRINT(kernel, 'k');
540 MOD_PRINT(user, 'u');
542 exclude_guest_default = true;
545 if (attr->precise_ip) {
548 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
549 exclude_guest_default = true;
552 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
553 MOD_PRINT(host, 'H');
554 MOD_PRINT(guest, 'G');
562 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
564 int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
565 return r + evsel__add_modifiers(evsel, bf + r, size - r);
568 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
581 static const char *__evsel__sw_name(u64 config)
583 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
584 return evsel__sw_names[config];
585 return "unknown-software";
588 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
590 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
591 return r + evsel__add_modifiers(evsel, bf + r, size - r);
594 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
598 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
600 if (type & HW_BREAKPOINT_R)
601 r += scnprintf(bf + r, size - r, "r");
603 if (type & HW_BREAKPOINT_W)
604 r += scnprintf(bf + r, size - r, "w");
606 if (type & HW_BREAKPOINT_X)
607 r += scnprintf(bf + r, size - r, "x");
612 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
614 struct perf_event_attr *attr = &evsel->core.attr;
615 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
616 return r + evsel__add_modifiers(evsel, bf + r, size - r);
619 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
620 { "L1-dcache", "l1-d", "l1d", "L1-data", },
621 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
623 { "dTLB", "d-tlb", "Data-TLB", },
624 { "iTLB", "i-tlb", "Instruction-TLB", },
625 { "branch", "branches", "bpu", "btb", "bpc", },
629 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
630 { "load", "loads", "read", },
631 { "store", "stores", "write", },
632 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
635 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
636 { "refs", "Reference", "ops", "access", },
637 { "misses", "miss", },
640 #define C(x) PERF_COUNT_HW_CACHE_##x
641 #define CACHE_READ (1 << C(OP_READ))
642 #define CACHE_WRITE (1 << C(OP_WRITE))
643 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
644 #define COP(x) (1 << x)
647 * cache operation stat
648 * L1I : Read and prefetch only
649 * ITLB and BPU : Read-only
651 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
652 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
653 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
654 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
655 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
656 [C(ITLB)] = (CACHE_READ),
657 [C(BPU)] = (CACHE_READ),
658 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
661 bool evsel__is_cache_op_valid(u8 type, u8 op)
663 if (evsel__hw_cache_stat[type] & COP(op))
664 return true; /* valid */
666 return false; /* invalid */
669 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
672 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
673 evsel__hw_cache_op[op][0],
674 evsel__hw_cache_result[result][0]);
677 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
678 evsel__hw_cache_op[op][1]);
681 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
683 u8 op, result, type = (config >> 0) & 0xff;
684 const char *err = "unknown-ext-hardware-cache-type";
686 if (type >= PERF_COUNT_HW_CACHE_MAX)
689 op = (config >> 8) & 0xff;
690 err = "unknown-ext-hardware-cache-op";
691 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
694 result = (config >> 16) & 0xff;
695 err = "unknown-ext-hardware-cache-result";
696 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
699 err = "invalid-cache";
700 if (!evsel__is_cache_op_valid(type, op))
703 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
705 return scnprintf(bf, size, "%s", err);
708 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
710 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
711 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
714 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
716 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
717 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
720 static int evsel__tool_name(char *bf, size_t size)
722 int ret = scnprintf(bf, size, "duration_time");
726 const char *evsel__name(struct evsel *evsel)
736 switch (evsel->core.attr.type) {
738 evsel__raw_name(evsel, bf, sizeof(bf));
741 case PERF_TYPE_HARDWARE:
742 evsel__hw_name(evsel, bf, sizeof(bf));
745 case PERF_TYPE_HW_CACHE:
746 evsel__hw_cache_name(evsel, bf, sizeof(bf));
749 case PERF_TYPE_SOFTWARE:
750 if (evsel->tool_event)
751 evsel__tool_name(bf, sizeof(bf));
753 evsel__sw_name(evsel, bf, sizeof(bf));
756 case PERF_TYPE_TRACEPOINT:
757 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
760 case PERF_TYPE_BREAKPOINT:
761 evsel__bp_name(evsel, bf, sizeof(bf));
765 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
766 evsel->core.attr.type);
770 evsel->name = strdup(bf);
778 const char *evsel__group_name(struct evsel *evsel)
780 return evsel->group_name ?: "anon group";
784 * Returns the group details for the specified leader,
785 * with following rules.
787 * For record -e '{cycles,instructions}'
788 * 'anon group { cycles:u, instructions:u }'
790 * For record -e 'cycles,instructions' and report --group
791 * 'cycles:u, instructions:u'
793 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
797 const char *group_name = evsel__group_name(evsel);
799 if (!evsel->forced_leader)
800 ret = scnprintf(buf, size, "%s { ", group_name);
802 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
804 for_each_group_member(pos, evsel)
805 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
807 if (!evsel->forced_leader)
808 ret += scnprintf(buf + ret, size - ret, " }");
813 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
814 struct callchain_param *param)
816 bool function = evsel__is_function_event(evsel);
817 struct perf_event_attr *attr = &evsel->core.attr;
819 evsel__set_sample_bit(evsel, CALLCHAIN);
821 attr->sample_max_stack = param->max_stack;
823 if (opts->kernel_callchains)
824 attr->exclude_callchain_user = 1;
825 if (opts->user_callchains)
826 attr->exclude_callchain_kernel = 1;
827 if (param->record_mode == CALLCHAIN_LBR) {
828 if (!opts->branch_stack) {
829 if (attr->exclude_user) {
830 pr_warning("LBR callstack option is only available "
831 "to get user callchain information. "
832 "Falling back to framepointers.\n");
834 evsel__set_sample_bit(evsel, BRANCH_STACK);
835 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
836 PERF_SAMPLE_BRANCH_CALL_STACK |
837 PERF_SAMPLE_BRANCH_NO_CYCLES |
838 PERF_SAMPLE_BRANCH_NO_FLAGS |
839 PERF_SAMPLE_BRANCH_HW_INDEX;
842 pr_warning("Cannot use LBR callstack with branch stack. "
843 "Falling back to framepointers.\n");
846 if (param->record_mode == CALLCHAIN_DWARF) {
848 evsel__set_sample_bit(evsel, REGS_USER);
849 evsel__set_sample_bit(evsel, STACK_USER);
850 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
851 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
852 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
853 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
854 "so the minimal registers set (IP, SP) is explicitly forced.\n");
856 attr->sample_regs_user |= PERF_REGS_MASK;
858 attr->sample_stack_user = param->dump_size;
859 attr->exclude_callchain_user = 1;
861 pr_info("Cannot use DWARF unwind for function trace event,"
862 " falling back to framepointers.\n");
867 pr_info("Disabling user space callchains for function trace event.\n");
868 attr->exclude_callchain_user = 1;
872 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
873 struct callchain_param *param)
876 return __evsel__config_callchain(evsel, opts, param);
879 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
881 struct perf_event_attr *attr = &evsel->core.attr;
883 evsel__reset_sample_bit(evsel, CALLCHAIN);
884 if (param->record_mode == CALLCHAIN_LBR) {
885 evsel__reset_sample_bit(evsel, BRANCH_STACK);
886 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
887 PERF_SAMPLE_BRANCH_CALL_STACK |
888 PERF_SAMPLE_BRANCH_HW_INDEX);
890 if (param->record_mode == CALLCHAIN_DWARF) {
891 evsel__reset_sample_bit(evsel, REGS_USER);
892 evsel__reset_sample_bit(evsel, STACK_USER);
896 static void evsel__apply_config_terms(struct evsel *evsel,
897 struct record_opts *opts, bool track)
899 struct evsel_config_term *term;
900 struct list_head *config_terms = &evsel->config_terms;
901 struct perf_event_attr *attr = &evsel->core.attr;
902 /* callgraph default */
903 struct callchain_param param = {
904 .record_mode = callchain_param.record_mode,
908 const char *callgraph_buf = NULL;
910 list_for_each_entry(term, config_terms, list) {
911 switch (term->type) {
912 case EVSEL__CONFIG_TERM_PERIOD:
913 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
914 attr->sample_period = term->val.period;
916 evsel__reset_sample_bit(evsel, PERIOD);
919 case EVSEL__CONFIG_TERM_FREQ:
920 if (!(term->weak && opts->user_freq != UINT_MAX)) {
921 attr->sample_freq = term->val.freq;
923 evsel__set_sample_bit(evsel, PERIOD);
926 case EVSEL__CONFIG_TERM_TIME:
928 evsel__set_sample_bit(evsel, TIME);
930 evsel__reset_sample_bit(evsel, TIME);
932 case EVSEL__CONFIG_TERM_CALLGRAPH:
933 callgraph_buf = term->val.str;
935 case EVSEL__CONFIG_TERM_BRANCH:
936 if (term->val.str && strcmp(term->val.str, "no")) {
937 evsel__set_sample_bit(evsel, BRANCH_STACK);
938 parse_branch_str(term->val.str,
939 &attr->branch_sample_type);
941 evsel__reset_sample_bit(evsel, BRANCH_STACK);
943 case EVSEL__CONFIG_TERM_STACK_USER:
944 dump_size = term->val.stack_user;
946 case EVSEL__CONFIG_TERM_MAX_STACK:
947 max_stack = term->val.max_stack;
949 case EVSEL__CONFIG_TERM_MAX_EVENTS:
950 evsel->max_events = term->val.max_events;
952 case EVSEL__CONFIG_TERM_INHERIT:
954 * attr->inherit should has already been set by
955 * evsel__config. If user explicitly set
956 * inherit using config terms, override global
957 * opt->no_inherit setting.
959 attr->inherit = term->val.inherit ? 1 : 0;
961 case EVSEL__CONFIG_TERM_OVERWRITE:
962 attr->write_backward = term->val.overwrite ? 1 : 0;
964 case EVSEL__CONFIG_TERM_DRV_CFG:
966 case EVSEL__CONFIG_TERM_PERCORE:
968 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
969 attr->aux_output = term->val.aux_output ? 1 : 0;
971 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
972 /* Already applied by auxtrace */
974 case EVSEL__CONFIG_TERM_CFG_CHG:
981 /* User explicitly set per-event callgraph, clear the old setting and reset. */
982 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
983 bool sample_address = false;
986 param.max_stack = max_stack;
987 if (callgraph_buf == NULL)
988 callgraph_buf = "fp";
991 /* parse callgraph parameters */
992 if (callgraph_buf != NULL) {
993 if (!strcmp(callgraph_buf, "no")) {
994 param.enabled = false;
995 param.record_mode = CALLCHAIN_NONE;
997 param.enabled = true;
998 if (parse_callchain_record(callgraph_buf, ¶m)) {
999 pr_err("per-event callgraph setting for %s failed. "
1000 "Apply callgraph global setting for it\n",
1004 if (param.record_mode == CALLCHAIN_DWARF)
1005 sample_address = true;
1008 if (dump_size > 0) {
1009 dump_size = round_up(dump_size, sizeof(u64));
1010 param.dump_size = dump_size;
1013 /* If global callgraph set, clear it */
1014 if (callchain_param.enabled)
1015 evsel__reset_callgraph(evsel, &callchain_param);
1017 /* set perf-event callgraph */
1018 if (param.enabled) {
1019 if (sample_address) {
1020 evsel__set_sample_bit(evsel, ADDR);
1021 evsel__set_sample_bit(evsel, DATA_SRC);
1022 evsel->core.attr.mmap_data = track;
1024 evsel__config_callchain(evsel, opts, ¶m);
1029 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1031 struct evsel_config_term *term, *found_term = NULL;
1033 list_for_each_entry(term, &evsel->config_terms, list) {
1034 if (term->type == type)
1041 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1043 evsel__set_sample_bit(evsel, WEIGHT);
1047 * The enable_on_exec/disabled value strategy:
1049 * 1) For any type of traced program:
1050 * - all independent events and group leaders are disabled
1051 * - all group members are enabled
1053 * Group members are ruled by group leaders. They need to
1054 * be enabled, because the group scheduling relies on that.
1056 * 2) For traced programs executed by perf:
1057 * - all independent events and group leaders have
1058 * enable_on_exec set
1059 * - we don't specifically enable or disable any event during
1060 * the record command
1062 * Independent events and group leaders are initially disabled
1063 * and get enabled by exec. Group members are ruled by group
1064 * leaders as stated in 1).
1066 * 3) For traced programs attached by perf (pid/tid):
1067 * - we specifically enable or disable all events during
1068 * the record command
1070 * When attaching events to already running traced we
1071 * enable/disable events specifically, as there's no
1072 * initial traced exec call.
1074 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1075 struct callchain_param *callchain)
1077 struct evsel *leader = evsel->leader;
1078 struct perf_event_attr *attr = &evsel->core.attr;
1079 int track = evsel->tracking;
1080 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1082 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1083 attr->inherit = !opts->no_inherit;
1084 attr->write_backward = opts->overwrite ? 1 : 0;
1086 evsel__set_sample_bit(evsel, IP);
1087 evsel__set_sample_bit(evsel, TID);
1089 if (evsel->sample_read) {
1090 evsel__set_sample_bit(evsel, READ);
1093 * We need ID even in case of single event, because
1094 * PERF_SAMPLE_READ process ID specific data.
1096 evsel__set_sample_id(evsel, false);
1099 * Apply group format only if we belong to group
1100 * with more than one members.
1102 if (leader->core.nr_members > 1) {
1103 attr->read_format |= PERF_FORMAT_GROUP;
1109 * We default some events to have a default interval. But keep
1110 * it a weak assumption overridable by the user.
1112 if (!attr->sample_period) {
1115 attr->sample_freq = opts->freq;
1117 attr->sample_period = opts->default_interval;
1121 * If attr->freq was set (here or earlier), ask for period
1125 evsel__set_sample_bit(evsel, PERIOD);
1127 if (opts->no_samples)
1128 attr->sample_freq = 0;
1130 if (opts->inherit_stat) {
1131 evsel->core.attr.read_format |=
1132 PERF_FORMAT_TOTAL_TIME_ENABLED |
1133 PERF_FORMAT_TOTAL_TIME_RUNNING |
1135 attr->inherit_stat = 1;
1138 if (opts->sample_address) {
1139 evsel__set_sample_bit(evsel, ADDR);
1140 attr->mmap_data = track;
1144 * We don't allow user space callchains for function trace
1145 * event, due to issues with page faults while tracing page
1146 * fault handler and its overall trickiness nature.
1148 if (evsel__is_function_event(evsel))
1149 evsel->core.attr.exclude_callchain_user = 1;
1151 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1152 evsel__config_callchain(evsel, opts, callchain);
1154 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1155 !evsel__is_dummy_event(evsel)) {
1156 attr->sample_regs_intr = opts->sample_intr_regs;
1157 evsel__set_sample_bit(evsel, REGS_INTR);
1160 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1161 !evsel__is_dummy_event(evsel)) {
1162 attr->sample_regs_user |= opts->sample_user_regs;
1163 evsel__set_sample_bit(evsel, REGS_USER);
1166 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1167 evsel__set_sample_bit(evsel, CPU);
1170 * When the user explicitly disabled time don't force it here.
1172 if (opts->sample_time &&
1173 (!perf_missing_features.sample_id_all &&
1174 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1175 opts->sample_time_set)))
1176 evsel__set_sample_bit(evsel, TIME);
1178 if (opts->raw_samples && !evsel->no_aux_samples) {
1179 evsel__set_sample_bit(evsel, TIME);
1180 evsel__set_sample_bit(evsel, RAW);
1181 evsel__set_sample_bit(evsel, CPU);
1184 if (opts->sample_address)
1185 evsel__set_sample_bit(evsel, DATA_SRC);
1187 if (opts->sample_phys_addr)
1188 evsel__set_sample_bit(evsel, PHYS_ADDR);
1190 if (opts->no_buffering) {
1191 attr->watermark = 0;
1192 attr->wakeup_events = 1;
1194 if (opts->branch_stack && !evsel->no_aux_samples) {
1195 evsel__set_sample_bit(evsel, BRANCH_STACK);
1196 attr->branch_sample_type = opts->branch_stack;
1199 if (opts->sample_weight)
1200 arch_evsel__set_sample_weight(evsel);
1204 attr->mmap2 = track && !perf_missing_features.mmap2;
1206 attr->build_id = track && opts->build_id;
1209 * ksymbol is tracked separately with text poke because it needs to be
1210 * system wide and enabled immediately.
1212 if (!opts->text_poke)
1213 attr->ksymbol = track && !perf_missing_features.ksymbol;
1214 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1216 if (opts->record_namespaces)
1217 attr->namespaces = track;
1219 if (opts->record_cgroup) {
1220 attr->cgroup = track && !perf_missing_features.cgroup;
1221 evsel__set_sample_bit(evsel, CGROUP);
1224 if (opts->sample_data_page_size)
1225 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1227 if (opts->sample_code_page_size)
1228 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1230 if (opts->record_switch_events)
1231 attr->context_switch = track;
1233 if (opts->sample_transaction)
1234 evsel__set_sample_bit(evsel, TRANSACTION);
1236 if (opts->running_time) {
1237 evsel->core.attr.read_format |=
1238 PERF_FORMAT_TOTAL_TIME_ENABLED |
1239 PERF_FORMAT_TOTAL_TIME_RUNNING;
1243 * XXX see the function comment above
1245 * Disabling only independent events or group leaders,
1246 * keeping group members enabled.
1248 if (evsel__is_group_leader(evsel))
1252 * Setting enable_on_exec for independent events and
1253 * group leaders for traced executed by perf.
1255 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1256 !opts->initial_delay)
1257 attr->enable_on_exec = 1;
1259 if (evsel->immediate) {
1261 attr->enable_on_exec = 0;
1264 clockid = opts->clockid;
1265 if (opts->use_clockid) {
1266 attr->use_clockid = 1;
1267 attr->clockid = opts->clockid;
1270 if (evsel->precise_max)
1271 attr->precise_ip = 3;
1273 if (opts->all_user) {
1274 attr->exclude_kernel = 1;
1275 attr->exclude_user = 0;
1278 if (opts->all_kernel) {
1279 attr->exclude_kernel = 0;
1280 attr->exclude_user = 1;
1283 if (evsel->core.own_cpus || evsel->unit)
1284 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1287 * Apply event specific term settings,
1288 * it overloads any global configuration.
1290 evsel__apply_config_terms(evsel, opts, track);
1292 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1294 /* The --period option takes the precedence. */
1295 if (opts->period_set) {
1297 evsel__set_sample_bit(evsel, PERIOD);
1299 evsel__reset_sample_bit(evsel, PERIOD);
1303 * A dummy event never triggers any actual counter and therefore
1304 * cannot be used with branch_stack.
1306 * For initial_delay, a dummy event is added implicitly.
1307 * The software event will trigger -EOPNOTSUPP error out,
1308 * if BRANCH_STACK bit is set.
1310 if (evsel__is_dummy_event(evsel))
1311 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1314 int evsel__set_filter(struct evsel *evsel, const char *filter)
1316 char *new_filter = strdup(filter);
1318 if (new_filter != NULL) {
1319 free(evsel->filter);
1320 evsel->filter = new_filter;
1327 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1331 if (evsel->filter == NULL)
1332 return evsel__set_filter(evsel, filter);
1334 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1335 free(evsel->filter);
1336 evsel->filter = new_filter;
1343 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1345 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1348 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1350 return evsel__append_filter(evsel, "%s,%s", filter);
1353 /* Caller has to clear disabled after going through all CPUs. */
1354 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1356 return perf_evsel__enable_cpu(&evsel->core, cpu);
1359 int evsel__enable(struct evsel *evsel)
1361 int err = perf_evsel__enable(&evsel->core);
1364 evsel->disabled = false;
1368 /* Caller has to set disabled after going through all CPUs. */
1369 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1371 return perf_evsel__disable_cpu(&evsel->core, cpu);
1374 int evsel__disable(struct evsel *evsel)
1376 int err = perf_evsel__disable(&evsel->core);
1378 * We mark it disabled here so that tools that disable a event can
1379 * ignore events after they disable it. I.e. the ring buffer may have
1380 * already a few more events queued up before the kernel got the stop
1384 evsel->disabled = true;
1389 static void evsel__free_config_terms(struct evsel *evsel)
1391 struct evsel_config_term *term, *h;
1393 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1394 list_del_init(&term->list);
1396 zfree(&term->val.str);
1401 void evsel__exit(struct evsel *evsel)
1403 assert(list_empty(&evsel->core.node));
1404 assert(evsel->evlist == NULL);
1405 bpf_counter__destroy(evsel);
1406 evsel__free_counts(evsel);
1407 perf_evsel__free_fd(&evsel->core);
1408 perf_evsel__free_id(&evsel->core);
1409 evsel__free_config_terms(evsel);
1410 cgroup__put(evsel->cgrp);
1411 perf_cpu_map__put(evsel->core.cpus);
1412 perf_cpu_map__put(evsel->core.own_cpus);
1413 perf_thread_map__put(evsel->core.threads);
1414 zfree(&evsel->group_name);
1415 zfree(&evsel->name);
1416 zfree(&evsel->pmu_name);
1417 evsel__zero_per_pkg(evsel);
1418 hashmap__free(evsel->per_pkg_mask);
1419 evsel->per_pkg_mask = NULL;
1420 zfree(&evsel->metric_events);
1421 perf_evsel__object.fini(evsel);
1424 void evsel__delete(struct evsel *evsel)
1430 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1431 struct perf_counts_values *count)
1433 struct perf_counts_values tmp;
1435 if (!evsel->prev_raw_counts)
1439 tmp = evsel->prev_raw_counts->aggr;
1440 evsel->prev_raw_counts->aggr = *count;
1442 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1443 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1446 count->val = count->val - tmp.val;
1447 count->ena = count->ena - tmp.ena;
1448 count->run = count->run - tmp.run;
1451 void perf_counts_values__scale(struct perf_counts_values *count,
1452 bool scale, s8 *pscaled)
1457 if (count->run == 0) {
1460 } else if (count->run < count->ena) {
1462 count->val = (u64)((double) count->val * count->ena / count->run);
1470 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1472 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1474 return perf_evsel__read(&evsel->core, cpu, thread, count);
1477 static void evsel__set_count(struct evsel *counter, int cpu, int thread, u64 val, u64 ena, u64 run)
1479 struct perf_counts_values *count;
1481 count = perf_counts(counter->counts, cpu, thread);
1487 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1490 static int evsel__process_group_data(struct evsel *leader, int cpu, int thread, u64 *data)
1492 u64 read_format = leader->core.attr.read_format;
1493 struct sample_read_value *v;
1494 u64 nr, ena = 0, run = 0, i;
1498 if (nr != (u64) leader->core.nr_members)
1501 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1504 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1507 v = (struct sample_read_value *) data;
1509 evsel__set_count(leader, cpu, thread, v[0].value, ena, run);
1511 for (i = 1; i < nr; i++) {
1512 struct evsel *counter;
1514 counter = evlist__id2evsel(leader->evlist, v[i].id);
1518 evsel__set_count(counter, cpu, thread, v[i].value, ena, run);
1524 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1526 struct perf_stat_evsel *ps = leader->stats;
1527 u64 read_format = leader->core.attr.read_format;
1528 int size = perf_evsel__read_size(&leader->core);
1529 u64 *data = ps->group_data;
1531 if (!(read_format & PERF_FORMAT_ID))
1534 if (!evsel__is_group_leader(leader))
1538 data = zalloc(size);
1542 ps->group_data = data;
1545 if (FD(leader, cpu, thread) < 0)
1548 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1551 return evsel__process_group_data(leader, cpu, thread, data);
1554 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1556 u64 read_format = evsel->core.attr.read_format;
1558 if (read_format & PERF_FORMAT_GROUP)
1559 return evsel__read_group(evsel, cpu, thread);
1561 return evsel__read_one(evsel, cpu, thread);
1564 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1566 struct perf_counts_values count;
1567 size_t nv = scale ? 3 : 1;
1569 if (FD(evsel, cpu, thread) < 0)
1572 if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1575 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1578 evsel__compute_deltas(evsel, cpu, thread, &count);
1579 perf_counts_values__scale(&count, scale, NULL);
1580 *perf_counts(evsel->counts, cpu, thread) = count;
1584 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1586 struct evsel *leader = evsel->leader;
1589 if (evsel__is_group_leader(evsel))
1593 * Leader must be already processed/open,
1594 * if not it's a bug.
1596 BUG_ON(!leader->core.fd);
1598 fd = FD(leader, cpu, thread);
1604 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1606 for (int cpu = 0; cpu < nr_cpus; cpu++)
1607 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1608 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1611 static int update_fds(struct evsel *evsel,
1612 int nr_cpus, int cpu_idx,
1613 int nr_threads, int thread_idx)
1617 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1620 evlist__for_each_entry(evsel->evlist, pos) {
1621 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1623 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1626 * Since fds for next evsel has not been created,
1627 * there is no need to iterate whole event list.
1635 static bool ignore_missing_thread(struct evsel *evsel,
1636 int nr_cpus, int cpu,
1637 struct perf_thread_map *threads,
1638 int thread, int err)
1640 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1642 if (!evsel->ignore_missing_thread)
1645 /* The system wide setup does not work with threads. */
1646 if (evsel->core.system_wide)
1649 /* The -ESRCH is perf event syscall errno for pid's not found. */
1653 /* If there's only one thread, let it fail. */
1654 if (threads->nr == 1)
1658 * We should remove fd for missing_thread first
1659 * because thread_map__remove() will decrease threads->nr.
1661 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1664 if (thread_map__remove(threads, thread))
1667 pr_warning("WARNING: Ignored open failure for pid %d\n",
1672 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1673 void *priv __maybe_unused)
1675 return fprintf(fp, " %-32s %s\n", name, val);
1678 static void display_attr(struct perf_event_attr *attr)
1680 if (verbose >= 2 || debug_peo_args) {
1681 fprintf(stderr, "%.60s\n", graph_dotted_line);
1682 fprintf(stderr, "perf_event_attr:\n");
1683 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1684 fprintf(stderr, "%.60s\n", graph_dotted_line);
1688 static int perf_event_open(struct evsel *evsel,
1689 pid_t pid, int cpu, int group_fd,
1690 unsigned long flags)
1692 int precise_ip = evsel->core.attr.precise_ip;
1696 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1697 pid, cpu, group_fd, flags);
1699 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1703 /* Do not try less precise if not requested. */
1704 if (!evsel->precise_max)
1708 * We tried all the precise_ip values, and it's
1709 * still failing, so leave it to standard fallback.
1711 if (!evsel->core.attr.precise_ip) {
1712 evsel->core.attr.precise_ip = precise_ip;
1716 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1717 evsel->core.attr.precise_ip--;
1718 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1719 display_attr(&evsel->core.attr);
1725 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1726 struct perf_thread_map *threads,
1727 int start_cpu, int end_cpu)
1729 int cpu, thread, nthreads;
1730 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1731 int pid = -1, err, old_errno;
1732 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1734 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1735 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1739 static struct perf_cpu_map *empty_cpu_map;
1741 if (empty_cpu_map == NULL) {
1742 empty_cpu_map = perf_cpu_map__dummy_new();
1743 if (empty_cpu_map == NULL)
1747 cpus = empty_cpu_map;
1750 if (threads == NULL) {
1751 static struct perf_thread_map *empty_thread_map;
1753 if (empty_thread_map == NULL) {
1754 empty_thread_map = thread_map__new_by_tid(-1);
1755 if (empty_thread_map == NULL)
1759 threads = empty_thread_map;
1762 if (evsel->core.system_wide)
1765 nthreads = threads->nr;
1767 if (evsel->core.fd == NULL &&
1768 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1772 flags |= PERF_FLAG_PID_CGROUP;
1773 pid = evsel->cgrp->fd;
1776 fallback_missing_features:
1777 if (perf_missing_features.weight_struct) {
1778 evsel__set_sample_bit(evsel, WEIGHT);
1779 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1781 if (perf_missing_features.clockid_wrong)
1782 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1783 if (perf_missing_features.clockid) {
1784 evsel->core.attr.use_clockid = 0;
1785 evsel->core.attr.clockid = 0;
1787 if (perf_missing_features.cloexec)
1788 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1789 if (perf_missing_features.mmap2)
1790 evsel->core.attr.mmap2 = 0;
1791 if (perf_missing_features.exclude_guest)
1792 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1793 if (perf_missing_features.lbr_flags)
1794 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1795 PERF_SAMPLE_BRANCH_NO_CYCLES);
1796 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1797 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1798 if (perf_missing_features.ksymbol)
1799 evsel->core.attr.ksymbol = 0;
1800 if (perf_missing_features.bpf)
1801 evsel->core.attr.bpf_event = 0;
1802 if (perf_missing_features.branch_hw_idx)
1803 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1805 if (perf_missing_features.sample_id_all)
1806 evsel->core.attr.sample_id_all = 0;
1808 display_attr(&evsel->core.attr);
1810 for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1812 for (thread = 0; thread < nthreads; thread++) {
1815 if (!evsel->cgrp && !evsel->core.system_wide)
1816 pid = perf_thread_map__pid(threads, thread);
1818 group_fd = get_group_fd(evsel, cpu, thread);
1822 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1825 FD(evsel, cpu, thread) = fd;
1827 bpf_counter__install_pe(evsel, cpu, fd);
1829 if (unlikely(test_attr__enabled)) {
1830 test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
1831 fd, group_fd, flags);
1837 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1839 * We just removed 1 thread, so take a step
1840 * back on thread index and lower the upper
1846 /* ... and pretend like nothing have happened. */
1851 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1856 pr_debug2_peo(" = %d\n", fd);
1858 if (evsel->bpf_fd >= 0) {
1860 int bpf_fd = evsel->bpf_fd;
1863 PERF_EVENT_IOC_SET_BPF,
1865 if (err && errno != EEXIST) {
1866 pr_err("failed to attach bpf fd %d: %s\n",
1867 bpf_fd, strerror(errno));
1873 set_rlimit = NO_CHANGE;
1876 * If we succeeded but had to kill clockid, fail and
1877 * have evsel__open_strerror() print us a nice error.
1879 if (perf_missing_features.clockid ||
1880 perf_missing_features.clockid_wrong) {
1891 * perf stat needs between 5 and 22 fds per CPU. When we run out
1892 * of them try to increase the limits.
1894 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1898 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1899 if (set_rlimit == NO_CHANGE)
1900 l.rlim_cur = l.rlim_max;
1902 l.rlim_cur = l.rlim_max + 1000;
1903 l.rlim_max = l.rlim_cur;
1905 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1914 if (err != -EINVAL || cpu > 0 || thread > 0)
1918 * Must probe features in the order they were added to the
1919 * perf_event_attr interface.
1921 if (!perf_missing_features.weight_struct &&
1922 (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1923 perf_missing_features.weight_struct = true;
1924 pr_debug2("switching off weight struct support\n");
1925 goto fallback_missing_features;
1926 } else if (!perf_missing_features.code_page_size &&
1927 (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1928 perf_missing_features.code_page_size = true;
1929 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1931 } else if (!perf_missing_features.data_page_size &&
1932 (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1933 perf_missing_features.data_page_size = true;
1934 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1936 } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1937 perf_missing_features.cgroup = true;
1938 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1940 } else if (!perf_missing_features.branch_hw_idx &&
1941 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1942 perf_missing_features.branch_hw_idx = true;
1943 pr_debug2("switching off branch HW index support\n");
1944 goto fallback_missing_features;
1945 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1946 perf_missing_features.aux_output = true;
1947 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1949 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1950 perf_missing_features.bpf = true;
1951 pr_debug2_peo("switching off bpf_event\n");
1952 goto fallback_missing_features;
1953 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1954 perf_missing_features.ksymbol = true;
1955 pr_debug2_peo("switching off ksymbol\n");
1956 goto fallback_missing_features;
1957 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1958 perf_missing_features.write_backward = true;
1959 pr_debug2_peo("switching off write_backward\n");
1961 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1962 perf_missing_features.clockid_wrong = true;
1963 pr_debug2_peo("switching off clockid\n");
1964 goto fallback_missing_features;
1965 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1966 perf_missing_features.clockid = true;
1967 pr_debug2_peo("switching off use_clockid\n");
1968 goto fallback_missing_features;
1969 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1970 perf_missing_features.cloexec = true;
1971 pr_debug2_peo("switching off cloexec flag\n");
1972 goto fallback_missing_features;
1973 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1974 perf_missing_features.mmap2 = true;
1975 pr_debug2_peo("switching off mmap2\n");
1976 goto fallback_missing_features;
1977 } else if (!perf_missing_features.exclude_guest &&
1978 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1979 perf_missing_features.exclude_guest = true;
1980 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1981 goto fallback_missing_features;
1982 } else if (!perf_missing_features.sample_id_all) {
1983 perf_missing_features.sample_id_all = true;
1984 pr_debug2_peo("switching off sample_id_all\n");
1985 goto retry_sample_id;
1986 } else if (!perf_missing_features.lbr_flags &&
1987 (evsel->core.attr.branch_sample_type &
1988 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1989 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1990 perf_missing_features.lbr_flags = true;
1991 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1992 goto fallback_missing_features;
1993 } else if (!perf_missing_features.group_read &&
1994 evsel->core.attr.inherit &&
1995 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1996 evsel__is_group_leader(evsel)) {
1997 perf_missing_features.group_read = true;
1998 pr_debug2_peo("switching off group read\n");
1999 goto fallback_missing_features;
2003 threads->err_thread = thread;
2007 while (--thread >= 0) {
2008 if (FD(evsel, cpu, thread) >= 0)
2009 close(FD(evsel, cpu, thread));
2010 FD(evsel, cpu, thread) = -1;
2013 } while (--cpu >= 0);
2018 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2019 struct perf_thread_map *threads)
2021 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
2024 void evsel__close(struct evsel *evsel)
2026 perf_evsel__close(&evsel->core);
2027 perf_evsel__free_id(&evsel->core);
2030 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
2033 return evsel__open_cpu(evsel, cpus, NULL, 0,
2034 cpus ? cpus->nr : 1);
2036 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
2039 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2041 return evsel__open(evsel, NULL, threads);
2044 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2045 const union perf_event *event,
2046 struct perf_sample *sample)
2048 u64 type = evsel->core.attr.sample_type;
2049 const __u64 *array = event->sample.array;
2050 bool swapped = evsel->needs_swap;
2053 array += ((event->header.size -
2054 sizeof(event->header)) / sizeof(u64)) - 1;
2056 if (type & PERF_SAMPLE_IDENTIFIER) {
2057 sample->id = *array;
2061 if (type & PERF_SAMPLE_CPU) {
2064 /* undo swap of u64, then swap on individual u32s */
2065 u.val64 = bswap_64(u.val64);
2066 u.val32[0] = bswap_32(u.val32[0]);
2069 sample->cpu = u.val32[0];
2073 if (type & PERF_SAMPLE_STREAM_ID) {
2074 sample->stream_id = *array;
2078 if (type & PERF_SAMPLE_ID) {
2079 sample->id = *array;
2083 if (type & PERF_SAMPLE_TIME) {
2084 sample->time = *array;
2088 if (type & PERF_SAMPLE_TID) {
2091 /* undo swap of u64, then swap on individual u32s */
2092 u.val64 = bswap_64(u.val64);
2093 u.val32[0] = bswap_32(u.val32[0]);
2094 u.val32[1] = bswap_32(u.val32[1]);
2097 sample->pid = u.val32[0];
2098 sample->tid = u.val32[1];
2105 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2108 return size > max_size || offset + size > endp;
2111 #define OVERFLOW_CHECK(offset, size, max_size) \
2113 if (overflow(endp, (max_size), (offset), (size))) \
2117 #define OVERFLOW_CHECK_u64(offset) \
2118 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2121 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2124 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2125 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2126 * check the format does not go past the end of the event.
2128 if (sample_size + sizeof(event->header) > event->header.size)
2134 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2136 u64 type __maybe_unused)
2138 data->weight = *array;
2141 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2142 struct perf_sample *data)
2144 u64 type = evsel->core.attr.sample_type;
2145 bool swapped = evsel->needs_swap;
2147 u16 max_size = event->header.size;
2148 const void *endp = (void *)event + max_size;
2152 * used for cross-endian analysis. See git commit 65014ab3
2153 * for why this goofiness is needed.
2157 memset(data, 0, sizeof(*data));
2158 data->cpu = data->pid = data->tid = -1;
2159 data->stream_id = data->id = data->time = -1ULL;
2160 data->period = evsel->core.attr.sample_period;
2161 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2162 data->misc = event->header.misc;
2164 data->data_src = PERF_MEM_DATA_SRC_NONE;
2166 if (event->header.type != PERF_RECORD_SAMPLE) {
2167 if (!evsel->core.attr.sample_id_all)
2169 return perf_evsel__parse_id_sample(evsel, event, data);
2172 array = event->sample.array;
2174 if (perf_event__check_size(event, evsel->sample_size))
2177 if (type & PERF_SAMPLE_IDENTIFIER) {
2182 if (type & PERF_SAMPLE_IP) {
2187 if (type & PERF_SAMPLE_TID) {
2190 /* undo swap of u64, then swap on individual u32s */
2191 u.val64 = bswap_64(u.val64);
2192 u.val32[0] = bswap_32(u.val32[0]);
2193 u.val32[1] = bswap_32(u.val32[1]);
2196 data->pid = u.val32[0];
2197 data->tid = u.val32[1];
2201 if (type & PERF_SAMPLE_TIME) {
2202 data->time = *array;
2206 if (type & PERF_SAMPLE_ADDR) {
2207 data->addr = *array;
2211 if (type & PERF_SAMPLE_ID) {
2216 if (type & PERF_SAMPLE_STREAM_ID) {
2217 data->stream_id = *array;
2221 if (type & PERF_SAMPLE_CPU) {
2225 /* undo swap of u64, then swap on individual u32s */
2226 u.val64 = bswap_64(u.val64);
2227 u.val32[0] = bswap_32(u.val32[0]);
2230 data->cpu = u.val32[0];
2234 if (type & PERF_SAMPLE_PERIOD) {
2235 data->period = *array;
2239 if (type & PERF_SAMPLE_READ) {
2240 u64 read_format = evsel->core.attr.read_format;
2242 OVERFLOW_CHECK_u64(array);
2243 if (read_format & PERF_FORMAT_GROUP)
2244 data->read.group.nr = *array;
2246 data->read.one.value = *array;
2250 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2251 OVERFLOW_CHECK_u64(array);
2252 data->read.time_enabled = *array;
2256 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2257 OVERFLOW_CHECK_u64(array);
2258 data->read.time_running = *array;
2262 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2263 if (read_format & PERF_FORMAT_GROUP) {
2264 const u64 max_group_nr = UINT64_MAX /
2265 sizeof(struct sample_read_value);
2267 if (data->read.group.nr > max_group_nr)
2269 sz = data->read.group.nr *
2270 sizeof(struct sample_read_value);
2271 OVERFLOW_CHECK(array, sz, max_size);
2272 data->read.group.values =
2273 (struct sample_read_value *)array;
2274 array = (void *)array + sz;
2276 OVERFLOW_CHECK_u64(array);
2277 data->read.one.id = *array;
2282 if (type & PERF_SAMPLE_CALLCHAIN) {
2283 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2285 OVERFLOW_CHECK_u64(array);
2286 data->callchain = (struct ip_callchain *)array++;
2287 if (data->callchain->nr > max_callchain_nr)
2289 sz = data->callchain->nr * sizeof(u64);
2290 OVERFLOW_CHECK(array, sz, max_size);
2291 array = (void *)array + sz;
2294 if (type & PERF_SAMPLE_RAW) {
2295 OVERFLOW_CHECK_u64(array);
2299 * Undo swap of u64, then swap on individual u32s,
2300 * get the size of the raw area and undo all of the
2301 * swap. The pevent interface handles endianness by
2305 u.val64 = bswap_64(u.val64);
2306 u.val32[0] = bswap_32(u.val32[0]);
2307 u.val32[1] = bswap_32(u.val32[1]);
2309 data->raw_size = u.val32[0];
2312 * The raw data is aligned on 64bits including the
2313 * u32 size, so it's safe to use mem_bswap_64.
2316 mem_bswap_64((void *) array, data->raw_size);
2318 array = (void *)array + sizeof(u32);
2320 OVERFLOW_CHECK(array, data->raw_size, max_size);
2321 data->raw_data = (void *)array;
2322 array = (void *)array + data->raw_size;
2325 if (type & PERF_SAMPLE_BRANCH_STACK) {
2326 const u64 max_branch_nr = UINT64_MAX /
2327 sizeof(struct branch_entry);
2329 OVERFLOW_CHECK_u64(array);
2330 data->branch_stack = (struct branch_stack *)array++;
2332 if (data->branch_stack->nr > max_branch_nr)
2335 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2336 if (evsel__has_branch_hw_idx(evsel))
2339 data->no_hw_idx = true;
2340 OVERFLOW_CHECK(array, sz, max_size);
2341 array = (void *)array + sz;
2344 if (type & PERF_SAMPLE_REGS_USER) {
2345 OVERFLOW_CHECK_u64(array);
2346 data->user_regs.abi = *array;
2349 if (data->user_regs.abi) {
2350 u64 mask = evsel->core.attr.sample_regs_user;
2352 sz = hweight64(mask) * sizeof(u64);
2353 OVERFLOW_CHECK(array, sz, max_size);
2354 data->user_regs.mask = mask;
2355 data->user_regs.regs = (u64 *)array;
2356 array = (void *)array + sz;
2360 if (type & PERF_SAMPLE_STACK_USER) {
2361 OVERFLOW_CHECK_u64(array);
2364 data->user_stack.offset = ((char *)(array - 1)
2368 data->user_stack.size = 0;
2370 OVERFLOW_CHECK(array, sz, max_size);
2371 data->user_stack.data = (char *)array;
2372 array = (void *)array + sz;
2373 OVERFLOW_CHECK_u64(array);
2374 data->user_stack.size = *array++;
2375 if (WARN_ONCE(data->user_stack.size > sz,
2376 "user stack dump failure\n"))
2381 if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2382 OVERFLOW_CHECK_u64(array);
2383 arch_perf_parse_sample_weight(data, array, type);
2387 if (type & PERF_SAMPLE_DATA_SRC) {
2388 OVERFLOW_CHECK_u64(array);
2389 data->data_src = *array;
2393 if (type & PERF_SAMPLE_TRANSACTION) {
2394 OVERFLOW_CHECK_u64(array);
2395 data->transaction = *array;
2399 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2400 if (type & PERF_SAMPLE_REGS_INTR) {
2401 OVERFLOW_CHECK_u64(array);
2402 data->intr_regs.abi = *array;
2405 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2406 u64 mask = evsel->core.attr.sample_regs_intr;
2408 sz = hweight64(mask) * sizeof(u64);
2409 OVERFLOW_CHECK(array, sz, max_size);
2410 data->intr_regs.mask = mask;
2411 data->intr_regs.regs = (u64 *)array;
2412 array = (void *)array + sz;
2416 data->phys_addr = 0;
2417 if (type & PERF_SAMPLE_PHYS_ADDR) {
2418 data->phys_addr = *array;
2423 if (type & PERF_SAMPLE_CGROUP) {
2424 data->cgroup = *array;
2428 data->data_page_size = 0;
2429 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2430 data->data_page_size = *array;
2434 data->code_page_size = 0;
2435 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2436 data->code_page_size = *array;
2440 if (type & PERF_SAMPLE_AUX) {
2441 OVERFLOW_CHECK_u64(array);
2444 OVERFLOW_CHECK(array, sz, max_size);
2445 /* Undo swap of data */
2447 mem_bswap_64((char *)array, sz);
2448 data->aux_sample.size = sz;
2449 data->aux_sample.data = (char *)array;
2450 array = (void *)array + sz;
2456 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2459 u64 type = evsel->core.attr.sample_type;
2462 if (!(type & PERF_SAMPLE_TIME))
2465 if (event->header.type != PERF_RECORD_SAMPLE) {
2466 struct perf_sample data = {
2470 if (!evsel->core.attr.sample_id_all)
2472 if (perf_evsel__parse_id_sample(evsel, event, &data))
2475 *timestamp = data.time;
2479 array = event->sample.array;
2481 if (perf_event__check_size(event, evsel->sample_size))
2484 if (type & PERF_SAMPLE_IDENTIFIER)
2487 if (type & PERF_SAMPLE_IP)
2490 if (type & PERF_SAMPLE_TID)
2493 if (type & PERF_SAMPLE_TIME)
2494 *timestamp = *array;
2499 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2501 return tep_find_field(evsel->tp_format, name);
2504 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2506 struct tep_format_field *field = evsel__field(evsel, name);
2512 offset = field->offset;
2514 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2515 offset = *(int *)(sample->raw_data + field->offset);
2519 return sample->raw_data + offset;
2522 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2526 void *ptr = sample->raw_data + field->offset;
2528 switch (field->size) {
2532 value = *(u16 *)ptr;
2535 value = *(u32 *)ptr;
2538 memcpy(&value, ptr, sizeof(u64));
2547 switch (field->size) {
2549 return bswap_16(value);
2551 return bswap_32(value);
2553 return bswap_64(value);
2561 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2563 struct tep_format_field *field = evsel__field(evsel, name);
2568 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2571 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2575 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2576 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2577 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2579 * If it's cycles then fall back to hrtimer based
2580 * cpu-clock-tick sw counter, which is always available even if
2583 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2586 scnprintf(msg, msgsize, "%s",
2587 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2589 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2590 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2592 zfree(&evsel->name);
2594 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2595 (paranoid = perf_event_paranoid()) > 1) {
2596 const char *name = evsel__name(evsel);
2598 const char *sep = ":";
2600 /* If event has exclude user then don't exclude kernel. */
2601 if (evsel->core.attr.exclude_user)
2604 /* Is there already the separator in the name. */
2605 if (strchr(name, '/') ||
2606 (strchr(name, ':') && !evsel->is_libpfm_event))
2609 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2614 evsel->name = new_name;
2615 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2616 "to fall back to excluding kernel and hypervisor "
2617 " samples", paranoid);
2618 evsel->core.attr.exclude_kernel = 1;
2619 evsel->core.attr.exclude_hv = 1;
2627 static bool find_process(const char *name)
2629 size_t len = strlen(name);
2634 dir = opendir(procfs__mountpoint());
2638 /* Walk through the directory. */
2639 while (ret && (d = readdir(dir)) != NULL) {
2640 char path[PATH_MAX];
2644 if ((d->d_type != DT_DIR) ||
2645 !strcmp(".", d->d_name) ||
2646 !strcmp("..", d->d_name))
2649 scnprintf(path, sizeof(path), "%s/%s/comm",
2650 procfs__mountpoint(), d->d_name);
2652 if (filename__read_str(path, &data, &size))
2655 ret = strncmp(name, data, len);
2660 return ret ? false : true;
2663 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2664 int err, char *msg, size_t size)
2666 char sbuf[STRERR_BUFSIZE];
2667 int printed = 0, enforced = 0;
2672 printed += scnprintf(msg + printed, size - printed,
2673 "Access to performance monitoring and observability operations is limited.\n");
2675 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2677 printed += scnprintf(msg + printed, size - printed,
2678 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2679 "monitoring and observability operations. Inspect system audit records for\n"
2680 "more perf_event access control information and adjusting the policy.\n");
2685 printed += scnprintf(msg, size,
2686 "No permission to enable %s event.\n\n", evsel__name(evsel));
2688 return scnprintf(msg + printed, size - printed,
2689 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2690 "access to performance monitoring and observability operations for processes\n"
2691 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2692 "More information can be found at 'Perf events and tool security' document:\n"
2693 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2694 "perf_event_paranoid setting is %d:\n"
2695 " -1: Allow use of (almost) all events by all users\n"
2696 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2697 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2698 ">= 1: Disallow CPU event access\n"
2699 ">= 2: Disallow kernel profiling\n"
2700 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2701 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2702 perf_event_paranoid());
2704 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2706 return scnprintf(msg, size, "%s",
2707 "Too many events are opened.\n"
2708 "Probably the maximum number of open file descriptors has been reached.\n"
2709 "Hint: Try again after reducing the number of events.\n"
2710 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2712 if (evsel__has_callchain(evsel) &&
2713 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2714 return scnprintf(msg, size,
2715 "Not enough memory to setup event with callchain.\n"
2716 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2717 "Hint: Current value: %d", sysctl__max_stack());
2720 if (target->cpu_list)
2721 return scnprintf(msg, size, "%s",
2722 "No such device - did you specify an out-of-range profile CPU?");
2725 if (evsel->core.attr.aux_output)
2726 return scnprintf(msg, size,
2727 "%s: PMU Hardware doesn't support 'aux_output' feature",
2728 evsel__name(evsel));
2729 if (evsel->core.attr.sample_period != 0)
2730 return scnprintf(msg, size,
2731 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2732 evsel__name(evsel));
2733 if (evsel->core.attr.precise_ip)
2734 return scnprintf(msg, size, "%s",
2735 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2736 #if defined(__i386__) || defined(__x86_64__)
2737 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2738 return scnprintf(msg, size, "%s",
2739 "No hardware sampling interrupt available.\n");
2743 if (find_process("oprofiled"))
2744 return scnprintf(msg, size,
2745 "The PMU counters are busy/taken by another profiler.\n"
2746 "We found oprofile daemon running, please stop it and try again.");
2749 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
2750 return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
2751 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
2752 return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
2753 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2754 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2755 if (perf_missing_features.clockid)
2756 return scnprintf(msg, size, "clockid feature not supported.");
2757 if (perf_missing_features.clockid_wrong)
2758 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2759 if (perf_missing_features.aux_output)
2760 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2763 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
2764 "Please add an auxiliary event in front of the load latency event.");
2769 return scnprintf(msg, size,
2770 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2771 "/bin/dmesg | grep -i perf may provide additional information.\n",
2772 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2775 struct perf_env *evsel__env(struct evsel *evsel)
2777 if (evsel && evsel->evlist)
2778 return evsel->evlist->env;
2782 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2786 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2787 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2789 int fd = FD(evsel, cpu, thread);
2791 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2792 cpu, thread, fd) < 0)
2800 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2802 struct perf_cpu_map *cpus = evsel->core.cpus;
2803 struct perf_thread_map *threads = evsel->core.threads;
2805 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2808 return store_evsel_ids(evsel, evlist);
2811 void evsel__zero_per_pkg(struct evsel *evsel)
2813 struct hashmap_entry *cur;
2816 if (evsel->per_pkg_mask) {
2817 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
2818 free((char *)cur->key);
2820 hashmap__clear(evsel->per_pkg_mask);
2824 bool evsel__is_hybrid(struct evsel *evsel)
2826 return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);