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, idx);
243 evsel->tracking = !idx;
244 evsel->unit = strdup("");
246 evsel->max_events = ULONG_MAX;
247 evsel->evlist = NULL;
248 evsel->bpf_obj = NULL;
250 INIT_LIST_HEAD(&evsel->config_terms);
251 INIT_LIST_HEAD(&evsel->bpf_counter_list);
252 perf_evsel__object.init(evsel);
253 evsel->sample_size = __evsel__sample_size(attr->sample_type);
254 evsel__calc_id_pos(evsel);
255 evsel->cmdline_group_boundary = false;
256 evsel->metric_expr = NULL;
257 evsel->metric_name = NULL;
258 evsel->metric_events = NULL;
259 evsel->per_pkg_mask = NULL;
260 evsel->collect_stat = false;
261 evsel->pmu_name = NULL;
264 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
266 struct evsel *evsel = zalloc(perf_evsel__object.size);
270 evsel__init(evsel, attr, idx);
272 if (evsel__is_bpf_output(evsel)) {
273 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
274 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
275 evsel->core.attr.sample_period = 1;
278 if (evsel__is_clock(evsel)) {
279 free((char *)evsel->unit);
280 evsel->unit = strdup("msec");
287 static bool perf_event_can_profile_kernel(void)
289 return perf_event_paranoid_check(1);
292 struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
294 struct perf_event_attr attr = {
297 .exclude_kernel = !perf_event_can_profile_kernel(),
301 event_attr_init(&attr);
304 * Now let the usual logic to set up the perf_event_attr defaults
305 * to kick in when we return and before perf_evsel__open() is called.
307 evsel = evsel__new(&attr);
311 arch_evsel__fixup_new_cycles(&evsel->core.attr);
313 evsel->precise_max = true;
315 /* use asprintf() because free(evsel) assumes name is allocated */
316 if (asprintf(&evsel->name, "cycles%s%s%.*s",
317 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
318 attr.exclude_kernel ? "u" : "",
319 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
324 evsel__delete(evsel);
329 int copy_config_terms(struct list_head *dst, struct list_head *src)
331 struct evsel_config_term *pos, *tmp;
333 list_for_each_entry(pos, src, list) {
334 tmp = malloc(sizeof(*tmp));
340 tmp->val.str = strdup(pos->val.str);
341 if (tmp->val.str == NULL) {
346 list_add_tail(&tmp->list, dst);
351 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
353 return copy_config_terms(&dst->config_terms, &src->config_terms);
357 * evsel__clone - create a new evsel copied from @orig
358 * @orig: original evsel
360 * The assumption is that @orig is not configured nor opened yet.
361 * So we only care about the attributes that can be set while it's parsed.
363 struct evsel *evsel__clone(struct evsel *orig)
367 BUG_ON(orig->core.fd);
368 BUG_ON(orig->counts);
370 BUG_ON(orig->per_pkg_mask);
372 /* cannot handle BPF objects for now */
376 evsel = evsel__new(&orig->core.attr);
380 evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
381 evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
382 evsel->core.threads = perf_thread_map__get(orig->core.threads);
383 evsel->core.nr_members = orig->core.nr_members;
384 evsel->core.system_wide = orig->core.system_wide;
387 evsel->name = strdup(orig->name);
388 if (evsel->name == NULL)
391 if (orig->group_name) {
392 evsel->group_name = strdup(orig->group_name);
393 if (evsel->group_name == NULL)
396 if (orig->pmu_name) {
397 evsel->pmu_name = strdup(orig->pmu_name);
398 if (evsel->pmu_name == NULL)
402 evsel->filter = strdup(orig->filter);
403 if (evsel->filter == NULL)
406 if (orig->metric_id) {
407 evsel->metric_id = strdup(orig->metric_id);
408 if (evsel->metric_id == NULL)
411 evsel->cgrp = cgroup__get(orig->cgrp);
412 evsel->tp_format = orig->tp_format;
413 evsel->handler = orig->handler;
414 evsel->core.leader = orig->core.leader;
416 evsel->max_events = orig->max_events;
417 evsel->tool_event = orig->tool_event;
418 free((char *)evsel->unit);
419 evsel->unit = strdup(orig->unit);
420 if (evsel->unit == NULL)
423 evsel->scale = orig->scale;
424 evsel->snapshot = orig->snapshot;
425 evsel->per_pkg = orig->per_pkg;
426 evsel->percore = orig->percore;
427 evsel->precise_max = orig->precise_max;
428 evsel->use_uncore_alias = orig->use_uncore_alias;
429 evsel->is_libpfm_event = orig->is_libpfm_event;
431 evsel->exclude_GH = orig->exclude_GH;
432 evsel->sample_read = orig->sample_read;
433 evsel->auto_merge_stats = orig->auto_merge_stats;
434 evsel->collect_stat = orig->collect_stat;
435 evsel->weak_group = orig->weak_group;
436 evsel->use_config_name = orig->use_config_name;
438 if (evsel__copy_config_terms(evsel, orig) < 0)
444 evsel__delete(evsel);
449 * Returns pointer with encoded error via <linux/err.h> interface.
451 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
453 struct evsel *evsel = zalloc(perf_evsel__object.size);
459 struct perf_event_attr attr = {
460 .type = PERF_TYPE_TRACEPOINT,
461 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
462 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
465 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
468 evsel->tp_format = trace_event__tp_format(sys, name);
469 if (IS_ERR(evsel->tp_format)) {
470 err = PTR_ERR(evsel->tp_format);
474 event_attr_init(&attr);
475 attr.config = evsel->tp_format->id;
476 attr.sample_period = 1;
477 evsel__init(evsel, &attr, idx);
489 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
497 "stalled-cycles-frontend",
498 "stalled-cycles-backend",
502 char *evsel__bpf_counter_events;
504 bool evsel__match_bpf_counter_events(const char *name)
510 if (!evsel__bpf_counter_events)
513 ptr = strstr(evsel__bpf_counter_events, name);
514 name_len = strlen(name);
516 /* check name matches a full token in evsel__bpf_counter_events */
517 match = (ptr != NULL) &&
518 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
519 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
524 static const char *__evsel__hw_name(u64 config)
526 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
527 return evsel__hw_names[config];
529 return "unknown-hardware";
532 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
534 int colon = 0, r = 0;
535 struct perf_event_attr *attr = &evsel->core.attr;
536 bool exclude_guest_default = false;
538 #define MOD_PRINT(context, mod) do { \
539 if (!attr->exclude_##context) { \
540 if (!colon) colon = ++r; \
541 r += scnprintf(bf + r, size - r, "%c", mod); \
544 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
545 MOD_PRINT(kernel, 'k');
546 MOD_PRINT(user, 'u');
548 exclude_guest_default = true;
551 if (attr->precise_ip) {
554 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
555 exclude_guest_default = true;
558 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
559 MOD_PRINT(host, 'H');
560 MOD_PRINT(guest, 'G');
568 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
570 int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
571 return r + evsel__add_modifiers(evsel, bf + r, size - r);
574 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
587 static const char *__evsel__sw_name(u64 config)
589 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
590 return evsel__sw_names[config];
591 return "unknown-software";
594 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
596 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
597 return r + evsel__add_modifiers(evsel, bf + r, size - r);
600 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
604 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
606 if (type & HW_BREAKPOINT_R)
607 r += scnprintf(bf + r, size - r, "r");
609 if (type & HW_BREAKPOINT_W)
610 r += scnprintf(bf + r, size - r, "w");
612 if (type & HW_BREAKPOINT_X)
613 r += scnprintf(bf + r, size - r, "x");
618 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
620 struct perf_event_attr *attr = &evsel->core.attr;
621 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
622 return r + evsel__add_modifiers(evsel, bf + r, size - r);
625 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
626 { "L1-dcache", "l1-d", "l1d", "L1-data", },
627 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
629 { "dTLB", "d-tlb", "Data-TLB", },
630 { "iTLB", "i-tlb", "Instruction-TLB", },
631 { "branch", "branches", "bpu", "btb", "bpc", },
635 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
636 { "load", "loads", "read", },
637 { "store", "stores", "write", },
638 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
641 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
642 { "refs", "Reference", "ops", "access", },
643 { "misses", "miss", },
646 #define C(x) PERF_COUNT_HW_CACHE_##x
647 #define CACHE_READ (1 << C(OP_READ))
648 #define CACHE_WRITE (1 << C(OP_WRITE))
649 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
650 #define COP(x) (1 << x)
653 * cache operation stat
654 * L1I : Read and prefetch only
655 * ITLB and BPU : Read-only
657 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
658 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
659 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
660 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
661 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
662 [C(ITLB)] = (CACHE_READ),
663 [C(BPU)] = (CACHE_READ),
664 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
667 bool evsel__is_cache_op_valid(u8 type, u8 op)
669 if (evsel__hw_cache_stat[type] & COP(op))
670 return true; /* valid */
672 return false; /* invalid */
675 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
678 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
679 evsel__hw_cache_op[op][0],
680 evsel__hw_cache_result[result][0]);
683 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
684 evsel__hw_cache_op[op][1]);
687 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
689 u8 op, result, type = (config >> 0) & 0xff;
690 const char *err = "unknown-ext-hardware-cache-type";
692 if (type >= PERF_COUNT_HW_CACHE_MAX)
695 op = (config >> 8) & 0xff;
696 err = "unknown-ext-hardware-cache-op";
697 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
700 result = (config >> 16) & 0xff;
701 err = "unknown-ext-hardware-cache-result";
702 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
705 err = "invalid-cache";
706 if (!evsel__is_cache_op_valid(type, op))
709 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
711 return scnprintf(bf, size, "%s", err);
714 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
716 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
717 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
720 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
722 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
723 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
726 static int evsel__tool_name(char *bf, size_t size)
728 int ret = scnprintf(bf, size, "duration_time");
732 const char *evsel__name(struct evsel *evsel)
742 switch (evsel->core.attr.type) {
744 evsel__raw_name(evsel, bf, sizeof(bf));
747 case PERF_TYPE_HARDWARE:
748 evsel__hw_name(evsel, bf, sizeof(bf));
751 case PERF_TYPE_HW_CACHE:
752 evsel__hw_cache_name(evsel, bf, sizeof(bf));
755 case PERF_TYPE_SOFTWARE:
756 if (evsel->tool_event)
757 evsel__tool_name(bf, sizeof(bf));
759 evsel__sw_name(evsel, bf, sizeof(bf));
762 case PERF_TYPE_TRACEPOINT:
763 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
766 case PERF_TYPE_BREAKPOINT:
767 evsel__bp_name(evsel, bf, sizeof(bf));
771 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
772 evsel->core.attr.type);
776 evsel->name = strdup(bf);
784 const char *evsel__metric_id(const struct evsel *evsel)
786 if (evsel->metric_id)
787 return evsel->metric_id;
789 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE && evsel->tool_event)
790 return "duration_time";
795 const char *evsel__group_name(struct evsel *evsel)
797 return evsel->group_name ?: "anon group";
801 * Returns the group details for the specified leader,
802 * with following rules.
804 * For record -e '{cycles,instructions}'
805 * 'anon group { cycles:u, instructions:u }'
807 * For record -e 'cycles,instructions' and report --group
808 * 'cycles:u, instructions:u'
810 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
814 const char *group_name = evsel__group_name(evsel);
816 if (!evsel->forced_leader)
817 ret = scnprintf(buf, size, "%s { ", group_name);
819 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
821 for_each_group_member(pos, evsel)
822 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
824 if (!evsel->forced_leader)
825 ret += scnprintf(buf + ret, size - ret, " }");
830 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
831 struct callchain_param *param)
833 bool function = evsel__is_function_event(evsel);
834 struct perf_event_attr *attr = &evsel->core.attr;
836 evsel__set_sample_bit(evsel, CALLCHAIN);
838 attr->sample_max_stack = param->max_stack;
840 if (opts->kernel_callchains)
841 attr->exclude_callchain_user = 1;
842 if (opts->user_callchains)
843 attr->exclude_callchain_kernel = 1;
844 if (param->record_mode == CALLCHAIN_LBR) {
845 if (!opts->branch_stack) {
846 if (attr->exclude_user) {
847 pr_warning("LBR callstack option is only available "
848 "to get user callchain information. "
849 "Falling back to framepointers.\n");
851 evsel__set_sample_bit(evsel, BRANCH_STACK);
852 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
853 PERF_SAMPLE_BRANCH_CALL_STACK |
854 PERF_SAMPLE_BRANCH_NO_CYCLES |
855 PERF_SAMPLE_BRANCH_NO_FLAGS |
856 PERF_SAMPLE_BRANCH_HW_INDEX;
859 pr_warning("Cannot use LBR callstack with branch stack. "
860 "Falling back to framepointers.\n");
863 if (param->record_mode == CALLCHAIN_DWARF) {
865 evsel__set_sample_bit(evsel, REGS_USER);
866 evsel__set_sample_bit(evsel, STACK_USER);
867 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
868 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
869 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
870 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
871 "so the minimal registers set (IP, SP) is explicitly forced.\n");
873 attr->sample_regs_user |= PERF_REGS_MASK;
875 attr->sample_stack_user = param->dump_size;
876 attr->exclude_callchain_user = 1;
878 pr_info("Cannot use DWARF unwind for function trace event,"
879 " falling back to framepointers.\n");
884 pr_info("Disabling user space callchains for function trace event.\n");
885 attr->exclude_callchain_user = 1;
889 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
890 struct callchain_param *param)
893 return __evsel__config_callchain(evsel, opts, param);
896 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
898 struct perf_event_attr *attr = &evsel->core.attr;
900 evsel__reset_sample_bit(evsel, CALLCHAIN);
901 if (param->record_mode == CALLCHAIN_LBR) {
902 evsel__reset_sample_bit(evsel, BRANCH_STACK);
903 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
904 PERF_SAMPLE_BRANCH_CALL_STACK |
905 PERF_SAMPLE_BRANCH_HW_INDEX);
907 if (param->record_mode == CALLCHAIN_DWARF) {
908 evsel__reset_sample_bit(evsel, REGS_USER);
909 evsel__reset_sample_bit(evsel, STACK_USER);
913 static void evsel__apply_config_terms(struct evsel *evsel,
914 struct record_opts *opts, bool track)
916 struct evsel_config_term *term;
917 struct list_head *config_terms = &evsel->config_terms;
918 struct perf_event_attr *attr = &evsel->core.attr;
919 /* callgraph default */
920 struct callchain_param param = {
921 .record_mode = callchain_param.record_mode,
925 const char *callgraph_buf = NULL;
927 list_for_each_entry(term, config_terms, list) {
928 switch (term->type) {
929 case EVSEL__CONFIG_TERM_PERIOD:
930 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
931 attr->sample_period = term->val.period;
933 evsel__reset_sample_bit(evsel, PERIOD);
936 case EVSEL__CONFIG_TERM_FREQ:
937 if (!(term->weak && opts->user_freq != UINT_MAX)) {
938 attr->sample_freq = term->val.freq;
940 evsel__set_sample_bit(evsel, PERIOD);
943 case EVSEL__CONFIG_TERM_TIME:
945 evsel__set_sample_bit(evsel, TIME);
947 evsel__reset_sample_bit(evsel, TIME);
949 case EVSEL__CONFIG_TERM_CALLGRAPH:
950 callgraph_buf = term->val.str;
952 case EVSEL__CONFIG_TERM_BRANCH:
953 if (term->val.str && strcmp(term->val.str, "no")) {
954 evsel__set_sample_bit(evsel, BRANCH_STACK);
955 parse_branch_str(term->val.str,
956 &attr->branch_sample_type);
958 evsel__reset_sample_bit(evsel, BRANCH_STACK);
960 case EVSEL__CONFIG_TERM_STACK_USER:
961 dump_size = term->val.stack_user;
963 case EVSEL__CONFIG_TERM_MAX_STACK:
964 max_stack = term->val.max_stack;
966 case EVSEL__CONFIG_TERM_MAX_EVENTS:
967 evsel->max_events = term->val.max_events;
969 case EVSEL__CONFIG_TERM_INHERIT:
971 * attr->inherit should has already been set by
972 * evsel__config. If user explicitly set
973 * inherit using config terms, override global
974 * opt->no_inherit setting.
976 attr->inherit = term->val.inherit ? 1 : 0;
978 case EVSEL__CONFIG_TERM_OVERWRITE:
979 attr->write_backward = term->val.overwrite ? 1 : 0;
981 case EVSEL__CONFIG_TERM_DRV_CFG:
983 case EVSEL__CONFIG_TERM_PERCORE:
985 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
986 attr->aux_output = term->val.aux_output ? 1 : 0;
988 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
989 /* Already applied by auxtrace */
991 case EVSEL__CONFIG_TERM_CFG_CHG:
998 /* User explicitly set per-event callgraph, clear the old setting and reset. */
999 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1000 bool sample_address = false;
1003 param.max_stack = max_stack;
1004 if (callgraph_buf == NULL)
1005 callgraph_buf = "fp";
1008 /* parse callgraph parameters */
1009 if (callgraph_buf != NULL) {
1010 if (!strcmp(callgraph_buf, "no")) {
1011 param.enabled = false;
1012 param.record_mode = CALLCHAIN_NONE;
1014 param.enabled = true;
1015 if (parse_callchain_record(callgraph_buf, ¶m)) {
1016 pr_err("per-event callgraph setting for %s failed. "
1017 "Apply callgraph global setting for it\n",
1021 if (param.record_mode == CALLCHAIN_DWARF)
1022 sample_address = true;
1025 if (dump_size > 0) {
1026 dump_size = round_up(dump_size, sizeof(u64));
1027 param.dump_size = dump_size;
1030 /* If global callgraph set, clear it */
1031 if (callchain_param.enabled)
1032 evsel__reset_callgraph(evsel, &callchain_param);
1034 /* set perf-event callgraph */
1035 if (param.enabled) {
1036 if (sample_address) {
1037 evsel__set_sample_bit(evsel, ADDR);
1038 evsel__set_sample_bit(evsel, DATA_SRC);
1039 evsel->core.attr.mmap_data = track;
1041 evsel__config_callchain(evsel, opts, ¶m);
1046 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1048 struct evsel_config_term *term, *found_term = NULL;
1050 list_for_each_entry(term, &evsel->config_terms, list) {
1051 if (term->type == type)
1058 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1060 evsel__set_sample_bit(evsel, WEIGHT);
1063 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1068 * The enable_on_exec/disabled value strategy:
1070 * 1) For any type of traced program:
1071 * - all independent events and group leaders are disabled
1072 * - all group members are enabled
1074 * Group members are ruled by group leaders. They need to
1075 * be enabled, because the group scheduling relies on that.
1077 * 2) For traced programs executed by perf:
1078 * - all independent events and group leaders have
1079 * enable_on_exec set
1080 * - we don't specifically enable or disable any event during
1081 * the record command
1083 * Independent events and group leaders are initially disabled
1084 * and get enabled by exec. Group members are ruled by group
1085 * leaders as stated in 1).
1087 * 3) For traced programs attached by perf (pid/tid):
1088 * - we specifically enable or disable all events during
1089 * the record command
1091 * When attaching events to already running traced we
1092 * enable/disable events specifically, as there's no
1093 * initial traced exec call.
1095 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1096 struct callchain_param *callchain)
1098 struct evsel *leader = evsel__leader(evsel);
1099 struct perf_event_attr *attr = &evsel->core.attr;
1100 int track = evsel->tracking;
1101 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1103 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1104 attr->inherit = !opts->no_inherit;
1105 attr->write_backward = opts->overwrite ? 1 : 0;
1107 evsel__set_sample_bit(evsel, IP);
1108 evsel__set_sample_bit(evsel, TID);
1110 if (evsel->sample_read) {
1111 evsel__set_sample_bit(evsel, READ);
1114 * We need ID even in case of single event, because
1115 * PERF_SAMPLE_READ process ID specific data.
1117 evsel__set_sample_id(evsel, false);
1120 * Apply group format only if we belong to group
1121 * with more than one members.
1123 if (leader->core.nr_members > 1) {
1124 attr->read_format |= PERF_FORMAT_GROUP;
1130 * We default some events to have a default interval. But keep
1131 * it a weak assumption overridable by the user.
1133 if (!attr->sample_period) {
1136 attr->sample_freq = opts->freq;
1138 attr->sample_period = opts->default_interval;
1142 * If attr->freq was set (here or earlier), ask for period
1146 evsel__set_sample_bit(evsel, PERIOD);
1148 if (opts->no_samples)
1149 attr->sample_freq = 0;
1151 if (opts->inherit_stat) {
1152 evsel->core.attr.read_format |=
1153 PERF_FORMAT_TOTAL_TIME_ENABLED |
1154 PERF_FORMAT_TOTAL_TIME_RUNNING |
1156 attr->inherit_stat = 1;
1159 if (opts->sample_address) {
1160 evsel__set_sample_bit(evsel, ADDR);
1161 attr->mmap_data = track;
1165 * We don't allow user space callchains for function trace
1166 * event, due to issues with page faults while tracing page
1167 * fault handler and its overall trickiness nature.
1169 if (evsel__is_function_event(evsel))
1170 evsel->core.attr.exclude_callchain_user = 1;
1172 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1173 evsel__config_callchain(evsel, opts, callchain);
1175 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1176 !evsel__is_dummy_event(evsel)) {
1177 attr->sample_regs_intr = opts->sample_intr_regs;
1178 evsel__set_sample_bit(evsel, REGS_INTR);
1181 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1182 !evsel__is_dummy_event(evsel)) {
1183 attr->sample_regs_user |= opts->sample_user_regs;
1184 evsel__set_sample_bit(evsel, REGS_USER);
1187 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1188 evsel__set_sample_bit(evsel, CPU);
1191 * When the user explicitly disabled time don't force it here.
1193 if (opts->sample_time &&
1194 (!perf_missing_features.sample_id_all &&
1195 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1196 opts->sample_time_set)))
1197 evsel__set_sample_bit(evsel, TIME);
1199 if (opts->raw_samples && !evsel->no_aux_samples) {
1200 evsel__set_sample_bit(evsel, TIME);
1201 evsel__set_sample_bit(evsel, RAW);
1202 evsel__set_sample_bit(evsel, CPU);
1205 if (opts->sample_address)
1206 evsel__set_sample_bit(evsel, DATA_SRC);
1208 if (opts->sample_phys_addr)
1209 evsel__set_sample_bit(evsel, PHYS_ADDR);
1211 if (opts->no_buffering) {
1212 attr->watermark = 0;
1213 attr->wakeup_events = 1;
1215 if (opts->branch_stack && !evsel->no_aux_samples) {
1216 evsel__set_sample_bit(evsel, BRANCH_STACK);
1217 attr->branch_sample_type = opts->branch_stack;
1220 if (opts->sample_weight)
1221 arch_evsel__set_sample_weight(evsel);
1225 attr->mmap2 = track && !perf_missing_features.mmap2;
1227 attr->build_id = track && opts->build_id;
1230 * ksymbol is tracked separately with text poke because it needs to be
1231 * system wide and enabled immediately.
1233 if (!opts->text_poke)
1234 attr->ksymbol = track && !perf_missing_features.ksymbol;
1235 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1237 if (opts->record_namespaces)
1238 attr->namespaces = track;
1240 if (opts->record_cgroup) {
1241 attr->cgroup = track && !perf_missing_features.cgroup;
1242 evsel__set_sample_bit(evsel, CGROUP);
1245 if (opts->sample_data_page_size)
1246 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1248 if (opts->sample_code_page_size)
1249 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1251 if (opts->record_switch_events)
1252 attr->context_switch = track;
1254 if (opts->sample_transaction)
1255 evsel__set_sample_bit(evsel, TRANSACTION);
1257 if (opts->running_time) {
1258 evsel->core.attr.read_format |=
1259 PERF_FORMAT_TOTAL_TIME_ENABLED |
1260 PERF_FORMAT_TOTAL_TIME_RUNNING;
1264 * XXX see the function comment above
1266 * Disabling only independent events or group leaders,
1267 * keeping group members enabled.
1269 if (evsel__is_group_leader(evsel))
1273 * Setting enable_on_exec for independent events and
1274 * group leaders for traced executed by perf.
1276 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1277 !opts->initial_delay)
1278 attr->enable_on_exec = 1;
1280 if (evsel->immediate) {
1282 attr->enable_on_exec = 0;
1285 clockid = opts->clockid;
1286 if (opts->use_clockid) {
1287 attr->use_clockid = 1;
1288 attr->clockid = opts->clockid;
1291 if (evsel->precise_max)
1292 attr->precise_ip = 3;
1294 if (opts->all_user) {
1295 attr->exclude_kernel = 1;
1296 attr->exclude_user = 0;
1299 if (opts->all_kernel) {
1300 attr->exclude_kernel = 0;
1301 attr->exclude_user = 1;
1304 if (evsel->core.own_cpus || evsel->unit)
1305 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1308 * Apply event specific term settings,
1309 * it overloads any global configuration.
1311 evsel__apply_config_terms(evsel, opts, track);
1313 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1315 /* The --period option takes the precedence. */
1316 if (opts->period_set) {
1318 evsel__set_sample_bit(evsel, PERIOD);
1320 evsel__reset_sample_bit(evsel, PERIOD);
1324 * A dummy event never triggers any actual counter and therefore
1325 * cannot be used with branch_stack.
1327 * For initial_delay, a dummy event is added implicitly.
1328 * The software event will trigger -EOPNOTSUPP error out,
1329 * if BRANCH_STACK bit is set.
1331 if (evsel__is_dummy_event(evsel))
1332 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1335 int evsel__set_filter(struct evsel *evsel, const char *filter)
1337 char *new_filter = strdup(filter);
1339 if (new_filter != NULL) {
1340 free(evsel->filter);
1341 evsel->filter = new_filter;
1348 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1352 if (evsel->filter == NULL)
1353 return evsel__set_filter(evsel, filter);
1355 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1356 free(evsel->filter);
1357 evsel->filter = new_filter;
1364 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1366 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1369 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1371 return evsel__append_filter(evsel, "%s,%s", filter);
1374 /* Caller has to clear disabled after going through all CPUs. */
1375 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1377 return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1380 int evsel__enable(struct evsel *evsel)
1382 int err = perf_evsel__enable(&evsel->core);
1385 evsel->disabled = false;
1389 /* Caller has to set disabled after going through all CPUs. */
1390 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1392 return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1395 int evsel__disable(struct evsel *evsel)
1397 int err = perf_evsel__disable(&evsel->core);
1399 * We mark it disabled here so that tools that disable a event can
1400 * ignore events after they disable it. I.e. the ring buffer may have
1401 * already a few more events queued up before the kernel got the stop
1405 evsel->disabled = true;
1410 void free_config_terms(struct list_head *config_terms)
1412 struct evsel_config_term *term, *h;
1414 list_for_each_entry_safe(term, h, config_terms, list) {
1415 list_del_init(&term->list);
1417 zfree(&term->val.str);
1422 static void evsel__free_config_terms(struct evsel *evsel)
1424 free_config_terms(&evsel->config_terms);
1427 void evsel__exit(struct evsel *evsel)
1429 assert(list_empty(&evsel->core.node));
1430 assert(evsel->evlist == NULL);
1431 bpf_counter__destroy(evsel);
1432 evsel__free_counts(evsel);
1433 perf_evsel__free_fd(&evsel->core);
1434 perf_evsel__free_id(&evsel->core);
1435 evsel__free_config_terms(evsel);
1436 cgroup__put(evsel->cgrp);
1437 perf_cpu_map__put(evsel->core.cpus);
1438 perf_cpu_map__put(evsel->core.own_cpus);
1439 perf_thread_map__put(evsel->core.threads);
1440 zfree(&evsel->group_name);
1441 zfree(&evsel->name);
1442 zfree(&evsel->pmu_name);
1443 zfree(&evsel->unit);
1444 zfree(&evsel->metric_id);
1445 evsel__zero_per_pkg(evsel);
1446 hashmap__free(evsel->per_pkg_mask);
1447 evsel->per_pkg_mask = NULL;
1448 zfree(&evsel->metric_events);
1449 perf_evsel__object.fini(evsel);
1452 void evsel__delete(struct evsel *evsel)
1458 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1459 struct perf_counts_values *count)
1461 struct perf_counts_values tmp;
1463 if (!evsel->prev_raw_counts)
1466 if (cpu_map_idx == -1) {
1467 tmp = evsel->prev_raw_counts->aggr;
1468 evsel->prev_raw_counts->aggr = *count;
1470 tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1471 *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1474 count->val = count->val - tmp.val;
1475 count->ena = count->ena - tmp.ena;
1476 count->run = count->run - tmp.run;
1479 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1481 struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1483 return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1486 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1487 u64 val, u64 ena, u64 run)
1489 struct perf_counts_values *count;
1491 count = perf_counts(counter->counts, cpu_map_idx, thread);
1497 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1500 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1502 u64 read_format = leader->core.attr.read_format;
1503 struct sample_read_value *v;
1504 u64 nr, ena = 0, run = 0, i;
1508 if (nr != (u64) leader->core.nr_members)
1511 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1514 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1517 v = (struct sample_read_value *) data;
1519 evsel__set_count(leader, cpu_map_idx, thread, v[0].value, ena, run);
1521 for (i = 1; i < nr; i++) {
1522 struct evsel *counter;
1524 counter = evlist__id2evsel(leader->evlist, v[i].id);
1528 evsel__set_count(counter, cpu_map_idx, thread, v[i].value, ena, run);
1534 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1536 struct perf_stat_evsel *ps = leader->stats;
1537 u64 read_format = leader->core.attr.read_format;
1538 int size = perf_evsel__read_size(&leader->core);
1539 u64 *data = ps->group_data;
1541 if (!(read_format & PERF_FORMAT_ID))
1544 if (!evsel__is_group_leader(leader))
1548 data = zalloc(size);
1552 ps->group_data = data;
1555 if (FD(leader, cpu_map_idx, thread) < 0)
1558 if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1561 return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1564 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1566 u64 read_format = evsel->core.attr.read_format;
1568 if (read_format & PERF_FORMAT_GROUP)
1569 return evsel__read_group(evsel, cpu_map_idx, thread);
1571 return evsel__read_one(evsel, cpu_map_idx, thread);
1574 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1576 struct perf_counts_values count;
1577 size_t nv = scale ? 3 : 1;
1579 if (FD(evsel, cpu_map_idx, thread) < 0)
1582 if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1585 if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1588 evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1589 perf_counts_values__scale(&count, scale, NULL);
1590 *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1594 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1597 struct perf_cpu cpu;
1599 cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1600 return perf_cpu_map__idx(other->core.cpus, cpu);
1603 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1605 struct evsel *leader = evsel__leader(evsel);
1607 if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1608 (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1609 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1615 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1617 struct evsel *leader = evsel__leader(evsel);
1620 if (evsel__is_group_leader(evsel))
1624 * Leader must be already processed/open,
1625 * if not it's a bug.
1627 BUG_ON(!leader->core.fd);
1629 cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1630 if (cpu_map_idx == -1)
1633 fd = FD(leader, cpu_map_idx, thread);
1639 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1641 for (int cpu = 0; cpu < nr_cpus; cpu++)
1642 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1643 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1646 static int update_fds(struct evsel *evsel,
1647 int nr_cpus, int cpu_map_idx,
1648 int nr_threads, int thread_idx)
1652 if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1655 evlist__for_each_entry(evsel->evlist, pos) {
1656 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1658 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1661 * Since fds for next evsel has not been created,
1662 * there is no need to iterate whole event list.
1670 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1671 int nr_cpus, int cpu_map_idx,
1672 struct perf_thread_map *threads,
1673 int thread, int err)
1675 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1677 if (!evsel->ignore_missing_thread)
1680 /* The system wide setup does not work with threads. */
1681 if (evsel->core.system_wide)
1684 /* The -ESRCH is perf event syscall errno for pid's not found. */
1688 /* If there's only one thread, let it fail. */
1689 if (threads->nr == 1)
1693 * We should remove fd for missing_thread first
1694 * because thread_map__remove() will decrease threads->nr.
1696 if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1699 if (thread_map__remove(threads, thread))
1702 pr_warning("WARNING: Ignored open failure for pid %d\n",
1707 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1708 void *priv __maybe_unused)
1710 return fprintf(fp, " %-32s %s\n", name, val);
1713 static void display_attr(struct perf_event_attr *attr)
1715 if (verbose >= 2 || debug_peo_args) {
1716 fprintf(stderr, "%.60s\n", graph_dotted_line);
1717 fprintf(stderr, "perf_event_attr:\n");
1718 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1719 fprintf(stderr, "%.60s\n", graph_dotted_line);
1723 bool evsel__precise_ip_fallback(struct evsel *evsel)
1725 /* Do not try less precise if not requested. */
1726 if (!evsel->precise_max)
1730 * We tried all the precise_ip values, and it's
1731 * still failing, so leave it to standard fallback.
1733 if (!evsel->core.attr.precise_ip) {
1734 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1738 if (!evsel->precise_ip_original)
1739 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1741 evsel->core.attr.precise_ip--;
1742 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1743 display_attr(&evsel->core.attr);
1747 static struct perf_cpu_map *empty_cpu_map;
1748 static struct perf_thread_map *empty_thread_map;
1750 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1751 struct perf_thread_map *threads)
1755 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1756 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1760 if (empty_cpu_map == NULL) {
1761 empty_cpu_map = perf_cpu_map__dummy_new();
1762 if (empty_cpu_map == NULL)
1766 cpus = empty_cpu_map;
1769 if (threads == NULL) {
1770 if (empty_thread_map == NULL) {
1771 empty_thread_map = thread_map__new_by_tid(-1);
1772 if (empty_thread_map == NULL)
1776 threads = empty_thread_map;
1779 if (evsel->core.system_wide)
1782 nthreads = threads->nr;
1784 if (evsel->core.fd == NULL &&
1785 perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1788 evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1790 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1795 static void evsel__disable_missing_features(struct evsel *evsel)
1797 if (perf_missing_features.weight_struct) {
1798 evsel__set_sample_bit(evsel, WEIGHT);
1799 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1801 if (perf_missing_features.clockid_wrong)
1802 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1803 if (perf_missing_features.clockid) {
1804 evsel->core.attr.use_clockid = 0;
1805 evsel->core.attr.clockid = 0;
1807 if (perf_missing_features.cloexec)
1808 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1809 if (perf_missing_features.mmap2)
1810 evsel->core.attr.mmap2 = 0;
1811 if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1812 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1813 if (perf_missing_features.lbr_flags)
1814 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1815 PERF_SAMPLE_BRANCH_NO_CYCLES);
1816 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1817 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1818 if (perf_missing_features.ksymbol)
1819 evsel->core.attr.ksymbol = 0;
1820 if (perf_missing_features.bpf)
1821 evsel->core.attr.bpf_event = 0;
1822 if (perf_missing_features.branch_hw_idx)
1823 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1824 if (perf_missing_features.sample_id_all)
1825 evsel->core.attr.sample_id_all = 0;
1828 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1829 struct perf_thread_map *threads)
1833 err = __evsel__prepare_open(evsel, cpus, threads);
1837 evsel__disable_missing_features(evsel);
1842 bool evsel__detect_missing_features(struct evsel *evsel)
1845 * Must probe features in the order they were added to the
1846 * perf_event_attr interface.
1848 if (!perf_missing_features.weight_struct &&
1849 (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1850 perf_missing_features.weight_struct = true;
1851 pr_debug2("switching off weight struct support\n");
1853 } else if (!perf_missing_features.code_page_size &&
1854 (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1855 perf_missing_features.code_page_size = true;
1856 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1858 } else if (!perf_missing_features.data_page_size &&
1859 (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1860 perf_missing_features.data_page_size = true;
1861 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1863 } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1864 perf_missing_features.cgroup = true;
1865 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1867 } else if (!perf_missing_features.branch_hw_idx &&
1868 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1869 perf_missing_features.branch_hw_idx = true;
1870 pr_debug2("switching off branch HW index support\n");
1872 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1873 perf_missing_features.aux_output = true;
1874 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1876 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1877 perf_missing_features.bpf = true;
1878 pr_debug2_peo("switching off bpf_event\n");
1880 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1881 perf_missing_features.ksymbol = true;
1882 pr_debug2_peo("switching off ksymbol\n");
1884 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1885 perf_missing_features.write_backward = true;
1886 pr_debug2_peo("switching off write_backward\n");
1888 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1889 perf_missing_features.clockid_wrong = true;
1890 pr_debug2_peo("switching off clockid\n");
1892 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1893 perf_missing_features.clockid = true;
1894 pr_debug2_peo("switching off use_clockid\n");
1896 } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1897 perf_missing_features.cloexec = true;
1898 pr_debug2_peo("switching off cloexec flag\n");
1900 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1901 perf_missing_features.mmap2 = true;
1902 pr_debug2_peo("switching off mmap2\n");
1904 } else if ((evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) &&
1905 (evsel->pmu == NULL || evsel->pmu->missing_features.exclude_guest)) {
1906 if (evsel->pmu == NULL) {
1907 evsel->pmu = evsel__find_pmu(evsel);
1909 evsel->pmu->missing_features.exclude_guest = true;
1911 /* we cannot find PMU, disable attrs now */
1912 evsel->core.attr.exclude_host = false;
1913 evsel->core.attr.exclude_guest = false;
1917 if (evsel->exclude_GH) {
1918 pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1921 if (!perf_missing_features.exclude_guest) {
1922 perf_missing_features.exclude_guest = true;
1923 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1926 } else if (!perf_missing_features.sample_id_all) {
1927 perf_missing_features.sample_id_all = true;
1928 pr_debug2_peo("switching off sample_id_all\n");
1930 } else if (!perf_missing_features.lbr_flags &&
1931 (evsel->core.attr.branch_sample_type &
1932 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1933 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1934 perf_missing_features.lbr_flags = true;
1935 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1937 } else if (!perf_missing_features.group_read &&
1938 evsel->core.attr.inherit &&
1939 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1940 evsel__is_group_leader(evsel)) {
1941 perf_missing_features.group_read = true;
1942 pr_debug2_peo("switching off group read\n");
1949 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
1954 if (*set_rlimit < INCREASED_MAX) {
1957 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1958 if (*set_rlimit == NO_CHANGE) {
1959 l.rlim_cur = l.rlim_max;
1961 l.rlim_cur = l.rlim_max + 1000;
1962 l.rlim_max = l.rlim_cur;
1964 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1976 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1977 struct perf_thread_map *threads,
1978 int start_cpu_map_idx, int end_cpu_map_idx)
1980 int idx, thread, nthreads;
1981 int pid = -1, err, old_errno;
1982 enum rlimit_action set_rlimit = NO_CHANGE;
1984 err = __evsel__prepare_open(evsel, cpus, threads);
1989 cpus = empty_cpu_map;
1991 if (threads == NULL)
1992 threads = empty_thread_map;
1994 if (evsel->core.system_wide)
1997 nthreads = threads->nr;
2000 pid = evsel->cgrp->fd;
2002 fallback_missing_features:
2003 evsel__disable_missing_features(evsel);
2005 display_attr(&evsel->core.attr);
2007 for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2009 for (thread = 0; thread < nthreads; thread++) {
2012 if (thread >= nthreads)
2015 if (!evsel->cgrp && !evsel->core.system_wide)
2016 pid = perf_thread_map__pid(threads, thread);
2018 group_fd = get_group_fd(evsel, idx, thread);
2022 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
2023 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2025 fd = sys_perf_event_open(&evsel->core.attr, pid,
2026 perf_cpu_map__cpu(cpus, idx).cpu,
2027 group_fd, evsel->open_flags);
2029 FD(evsel, idx, thread) = fd;
2034 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2039 bpf_counter__install_pe(evsel, idx, fd);
2041 if (unlikely(test_attr__enabled)) {
2042 test_attr__open(&evsel->core.attr, pid,
2043 perf_cpu_map__cpu(cpus, idx),
2044 fd, group_fd, evsel->open_flags);
2047 pr_debug2_peo(" = %d\n", fd);
2049 if (evsel->bpf_fd >= 0) {
2051 int bpf_fd = evsel->bpf_fd;
2054 PERF_EVENT_IOC_SET_BPF,
2056 if (err && errno != EEXIST) {
2057 pr_err("failed to attach bpf fd %d: %s\n",
2058 bpf_fd, strerror(errno));
2064 set_rlimit = NO_CHANGE;
2067 * If we succeeded but had to kill clockid, fail and
2068 * have evsel__open_strerror() print us a nice error.
2070 if (perf_missing_features.clockid ||
2071 perf_missing_features.clockid_wrong) {
2081 if (evsel__precise_ip_fallback(evsel))
2084 if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2085 idx, threads, thread, err)) {
2086 /* We just removed 1 thread, so lower the upper nthreads limit. */
2089 /* ... and pretend like nothing have happened. */
2094 * perf stat needs between 5 and 22 fds per CPU. When we run out
2095 * of them try to increase the limits.
2097 if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2100 if (err != -EINVAL || idx > 0 || thread > 0)
2103 if (evsel__detect_missing_features(evsel))
2104 goto fallback_missing_features;
2107 threads->err_thread = thread;
2111 while (--thread >= 0) {
2112 if (FD(evsel, idx, thread) >= 0)
2113 close(FD(evsel, idx, thread));
2114 FD(evsel, idx, thread) = -1;
2117 } while (--idx >= 0);
2122 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2123 struct perf_thread_map *threads)
2125 return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2128 void evsel__close(struct evsel *evsel)
2130 perf_evsel__close(&evsel->core);
2131 perf_evsel__free_id(&evsel->core);
2134 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2136 if (cpu_map_idx == -1)
2137 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2139 return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2142 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2144 return evsel__open(evsel, NULL, threads);
2147 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2148 const union perf_event *event,
2149 struct perf_sample *sample)
2151 u64 type = evsel->core.attr.sample_type;
2152 const __u64 *array = event->sample.array;
2153 bool swapped = evsel->needs_swap;
2156 array += ((event->header.size -
2157 sizeof(event->header)) / sizeof(u64)) - 1;
2159 if (type & PERF_SAMPLE_IDENTIFIER) {
2160 sample->id = *array;
2164 if (type & PERF_SAMPLE_CPU) {
2167 /* undo swap of u64, then swap on individual u32s */
2168 u.val64 = bswap_64(u.val64);
2169 u.val32[0] = bswap_32(u.val32[0]);
2172 sample->cpu = u.val32[0];
2176 if (type & PERF_SAMPLE_STREAM_ID) {
2177 sample->stream_id = *array;
2181 if (type & PERF_SAMPLE_ID) {
2182 sample->id = *array;
2186 if (type & PERF_SAMPLE_TIME) {
2187 sample->time = *array;
2191 if (type & PERF_SAMPLE_TID) {
2194 /* undo swap of u64, then swap on individual u32s */
2195 u.val64 = bswap_64(u.val64);
2196 u.val32[0] = bswap_32(u.val32[0]);
2197 u.val32[1] = bswap_32(u.val32[1]);
2200 sample->pid = u.val32[0];
2201 sample->tid = u.val32[1];
2208 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2211 return size > max_size || offset + size > endp;
2214 #define OVERFLOW_CHECK(offset, size, max_size) \
2216 if (overflow(endp, (max_size), (offset), (size))) \
2220 #define OVERFLOW_CHECK_u64(offset) \
2221 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2224 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2227 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2228 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2229 * check the format does not go past the end of the event.
2231 if (sample_size + sizeof(event->header) > event->header.size)
2237 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2239 u64 type __maybe_unused)
2241 data->weight = *array;
2244 u64 evsel__bitfield_swap_branch_flags(u64 value)
2253 * mispred:1 //target mispredicted
2254 * predicted:1 //target predicted
2255 * in_tx:1 //in transaction
2256 * abort:1 //transaction abort
2257 * cycles:16 //cycle count to last branch
2258 * type:4 //branch type
2263 * Avoid bswap64() the entire branch_flag.value,
2264 * as it has variable bit-field sizes. Instead the
2265 * macro takes the bit-field position/size,
2266 * swaps it based on the host endianness.
2268 * tep_is_bigendian() is used here instead of
2269 * bigendian() to avoid python test fails.
2271 if (tep_is_bigendian()) {
2272 new_val = bitfield_swap(value, 0, 1);
2273 new_val |= bitfield_swap(value, 1, 1);
2274 new_val |= bitfield_swap(value, 2, 1);
2275 new_val |= bitfield_swap(value, 3, 1);
2276 new_val |= bitfield_swap(value, 4, 16);
2277 new_val |= bitfield_swap(value, 20, 4);
2278 new_val |= bitfield_swap(value, 24, 40);
2280 new_val = bitfield_swap(value, 63, 1);
2281 new_val |= bitfield_swap(value, 62, 1);
2282 new_val |= bitfield_swap(value, 61, 1);
2283 new_val |= bitfield_swap(value, 60, 1);
2284 new_val |= bitfield_swap(value, 44, 16);
2285 new_val |= bitfield_swap(value, 40, 4);
2286 new_val |= bitfield_swap(value, 0, 40);
2292 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2293 struct perf_sample *data)
2295 u64 type = evsel->core.attr.sample_type;
2296 bool swapped = evsel->needs_swap;
2298 u16 max_size = event->header.size;
2299 const void *endp = (void *)event + max_size;
2303 * used for cross-endian analysis. See git commit 65014ab3
2304 * for why this goofiness is needed.
2308 memset(data, 0, sizeof(*data));
2309 data->cpu = data->pid = data->tid = -1;
2310 data->stream_id = data->id = data->time = -1ULL;
2311 data->period = evsel->core.attr.sample_period;
2312 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2313 data->misc = event->header.misc;
2315 data->data_src = PERF_MEM_DATA_SRC_NONE;
2317 if (event->header.type != PERF_RECORD_SAMPLE) {
2318 if (!evsel->core.attr.sample_id_all)
2320 return perf_evsel__parse_id_sample(evsel, event, data);
2323 array = event->sample.array;
2325 if (perf_event__check_size(event, evsel->sample_size))
2328 if (type & PERF_SAMPLE_IDENTIFIER) {
2333 if (type & PERF_SAMPLE_IP) {
2338 if (type & PERF_SAMPLE_TID) {
2341 /* undo swap of u64, then swap on individual u32s */
2342 u.val64 = bswap_64(u.val64);
2343 u.val32[0] = bswap_32(u.val32[0]);
2344 u.val32[1] = bswap_32(u.val32[1]);
2347 data->pid = u.val32[0];
2348 data->tid = u.val32[1];
2352 if (type & PERF_SAMPLE_TIME) {
2353 data->time = *array;
2357 if (type & PERF_SAMPLE_ADDR) {
2358 data->addr = *array;
2362 if (type & PERF_SAMPLE_ID) {
2367 if (type & PERF_SAMPLE_STREAM_ID) {
2368 data->stream_id = *array;
2372 if (type & PERF_SAMPLE_CPU) {
2376 /* undo swap of u64, then swap on individual u32s */
2377 u.val64 = bswap_64(u.val64);
2378 u.val32[0] = bswap_32(u.val32[0]);
2381 data->cpu = u.val32[0];
2385 if (type & PERF_SAMPLE_PERIOD) {
2386 data->period = *array;
2390 if (type & PERF_SAMPLE_READ) {
2391 u64 read_format = evsel->core.attr.read_format;
2393 OVERFLOW_CHECK_u64(array);
2394 if (read_format & PERF_FORMAT_GROUP)
2395 data->read.group.nr = *array;
2397 data->read.one.value = *array;
2401 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2402 OVERFLOW_CHECK_u64(array);
2403 data->read.time_enabled = *array;
2407 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2408 OVERFLOW_CHECK_u64(array);
2409 data->read.time_running = *array;
2413 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2414 if (read_format & PERF_FORMAT_GROUP) {
2415 const u64 max_group_nr = UINT64_MAX /
2416 sizeof(struct sample_read_value);
2418 if (data->read.group.nr > max_group_nr)
2420 sz = data->read.group.nr *
2421 sizeof(struct sample_read_value);
2422 OVERFLOW_CHECK(array, sz, max_size);
2423 data->read.group.values =
2424 (struct sample_read_value *)array;
2425 array = (void *)array + sz;
2427 OVERFLOW_CHECK_u64(array);
2428 data->read.one.id = *array;
2433 if (type & PERF_SAMPLE_CALLCHAIN) {
2434 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2436 OVERFLOW_CHECK_u64(array);
2437 data->callchain = (struct ip_callchain *)array++;
2438 if (data->callchain->nr > max_callchain_nr)
2440 sz = data->callchain->nr * sizeof(u64);
2441 OVERFLOW_CHECK(array, sz, max_size);
2442 array = (void *)array + sz;
2445 if (type & PERF_SAMPLE_RAW) {
2446 OVERFLOW_CHECK_u64(array);
2450 * Undo swap of u64, then swap on individual u32s,
2451 * get the size of the raw area and undo all of the
2452 * swap. The pevent interface handles endianness by
2456 u.val64 = bswap_64(u.val64);
2457 u.val32[0] = bswap_32(u.val32[0]);
2458 u.val32[1] = bswap_32(u.val32[1]);
2460 data->raw_size = u.val32[0];
2463 * The raw data is aligned on 64bits including the
2464 * u32 size, so it's safe to use mem_bswap_64.
2467 mem_bswap_64((void *) array, data->raw_size);
2469 array = (void *)array + sizeof(u32);
2471 OVERFLOW_CHECK(array, data->raw_size, max_size);
2472 data->raw_data = (void *)array;
2473 array = (void *)array + data->raw_size;
2476 if (type & PERF_SAMPLE_BRANCH_STACK) {
2477 const u64 max_branch_nr = UINT64_MAX /
2478 sizeof(struct branch_entry);
2479 struct branch_entry *e;
2482 OVERFLOW_CHECK_u64(array);
2483 data->branch_stack = (struct branch_stack *)array++;
2485 if (data->branch_stack->nr > max_branch_nr)
2488 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2489 if (evsel__has_branch_hw_idx(evsel)) {
2491 e = &data->branch_stack->entries[0];
2493 data->no_hw_idx = true;
2495 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2496 * only nr and entries[] will be output by kernel.
2498 e = (struct branch_entry *)&data->branch_stack->hw_idx;
2503 * struct branch_flag does not have endian
2504 * specific bit field definition. And bswap
2505 * will not resolve the issue, since these
2508 * evsel__bitfield_swap_branch_flags() uses a
2509 * bitfield_swap macro to swap the bit position
2510 * based on the host endians.
2512 for (i = 0; i < data->branch_stack->nr; i++, e++)
2513 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2516 OVERFLOW_CHECK(array, sz, max_size);
2517 array = (void *)array + sz;
2520 if (type & PERF_SAMPLE_REGS_USER) {
2521 OVERFLOW_CHECK_u64(array);
2522 data->user_regs.abi = *array;
2525 if (data->user_regs.abi) {
2526 u64 mask = evsel->core.attr.sample_regs_user;
2528 sz = hweight64(mask) * sizeof(u64);
2529 OVERFLOW_CHECK(array, sz, max_size);
2530 data->user_regs.mask = mask;
2531 data->user_regs.regs = (u64 *)array;
2532 array = (void *)array + sz;
2536 if (type & PERF_SAMPLE_STACK_USER) {
2537 OVERFLOW_CHECK_u64(array);
2540 data->user_stack.offset = ((char *)(array - 1)
2544 data->user_stack.size = 0;
2546 OVERFLOW_CHECK(array, sz, max_size);
2547 data->user_stack.data = (char *)array;
2548 array = (void *)array + sz;
2549 OVERFLOW_CHECK_u64(array);
2550 data->user_stack.size = *array++;
2551 if (WARN_ONCE(data->user_stack.size > sz,
2552 "user stack dump failure\n"))
2557 if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2558 OVERFLOW_CHECK_u64(array);
2559 arch_perf_parse_sample_weight(data, array, type);
2563 if (type & PERF_SAMPLE_DATA_SRC) {
2564 OVERFLOW_CHECK_u64(array);
2565 data->data_src = *array;
2569 if (type & PERF_SAMPLE_TRANSACTION) {
2570 OVERFLOW_CHECK_u64(array);
2571 data->transaction = *array;
2575 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2576 if (type & PERF_SAMPLE_REGS_INTR) {
2577 OVERFLOW_CHECK_u64(array);
2578 data->intr_regs.abi = *array;
2581 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2582 u64 mask = evsel->core.attr.sample_regs_intr;
2584 sz = hweight64(mask) * sizeof(u64);
2585 OVERFLOW_CHECK(array, sz, max_size);
2586 data->intr_regs.mask = mask;
2587 data->intr_regs.regs = (u64 *)array;
2588 array = (void *)array + sz;
2592 data->phys_addr = 0;
2593 if (type & PERF_SAMPLE_PHYS_ADDR) {
2594 data->phys_addr = *array;
2599 if (type & PERF_SAMPLE_CGROUP) {
2600 data->cgroup = *array;
2604 data->data_page_size = 0;
2605 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2606 data->data_page_size = *array;
2610 data->code_page_size = 0;
2611 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2612 data->code_page_size = *array;
2616 if (type & PERF_SAMPLE_AUX) {
2617 OVERFLOW_CHECK_u64(array);
2620 OVERFLOW_CHECK(array, sz, max_size);
2621 /* Undo swap of data */
2623 mem_bswap_64((char *)array, sz);
2624 data->aux_sample.size = sz;
2625 data->aux_sample.data = (char *)array;
2626 array = (void *)array + sz;
2632 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2635 u64 type = evsel->core.attr.sample_type;
2638 if (!(type & PERF_SAMPLE_TIME))
2641 if (event->header.type != PERF_RECORD_SAMPLE) {
2642 struct perf_sample data = {
2646 if (!evsel->core.attr.sample_id_all)
2648 if (perf_evsel__parse_id_sample(evsel, event, &data))
2651 *timestamp = data.time;
2655 array = event->sample.array;
2657 if (perf_event__check_size(event, evsel->sample_size))
2660 if (type & PERF_SAMPLE_IDENTIFIER)
2663 if (type & PERF_SAMPLE_IP)
2666 if (type & PERF_SAMPLE_TID)
2669 if (type & PERF_SAMPLE_TIME)
2670 *timestamp = *array;
2675 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2677 return tep_find_field(evsel->tp_format, name);
2680 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2682 struct tep_format_field *field = evsel__field(evsel, name);
2688 offset = field->offset;
2690 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2691 offset = *(int *)(sample->raw_data + field->offset);
2693 if (field->flags & TEP_FIELD_IS_RELATIVE)
2694 offset += field->offset + field->size;
2697 return sample->raw_data + offset;
2700 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2704 void *ptr = sample->raw_data + field->offset;
2706 switch (field->size) {
2710 value = *(u16 *)ptr;
2713 value = *(u32 *)ptr;
2716 memcpy(&value, ptr, sizeof(u64));
2725 switch (field->size) {
2727 return bswap_16(value);
2729 return bswap_32(value);
2731 return bswap_64(value);
2739 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2741 struct tep_format_field *field = evsel__field(evsel, name);
2746 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2749 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2753 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2754 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2755 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2757 * If it's cycles then fall back to hrtimer based
2758 * cpu-clock-tick sw counter, which is always available even if
2761 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2764 scnprintf(msg, msgsize, "%s",
2765 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2767 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2768 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2770 zfree(&evsel->name);
2772 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2773 (paranoid = perf_event_paranoid()) > 1) {
2774 const char *name = evsel__name(evsel);
2776 const char *sep = ":";
2778 /* If event has exclude user then don't exclude kernel. */
2779 if (evsel->core.attr.exclude_user)
2782 /* Is there already the separator in the name. */
2783 if (strchr(name, '/') ||
2784 (strchr(name, ':') && !evsel->is_libpfm_event))
2787 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2792 evsel->name = new_name;
2793 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2794 "to fall back to excluding kernel and hypervisor "
2795 " samples", paranoid);
2796 evsel->core.attr.exclude_kernel = 1;
2797 evsel->core.attr.exclude_hv = 1;
2805 static bool find_process(const char *name)
2807 size_t len = strlen(name);
2812 dir = opendir(procfs__mountpoint());
2816 /* Walk through the directory. */
2817 while (ret && (d = readdir(dir)) != NULL) {
2818 char path[PATH_MAX];
2822 if ((d->d_type != DT_DIR) ||
2823 !strcmp(".", d->d_name) ||
2824 !strcmp("..", d->d_name))
2827 scnprintf(path, sizeof(path), "%s/%s/comm",
2828 procfs__mountpoint(), d->d_name);
2830 if (filename__read_str(path, &data, &size))
2833 ret = strncmp(name, data, len);
2838 return ret ? false : true;
2841 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2842 int err, char *msg, size_t size)
2844 char sbuf[STRERR_BUFSIZE];
2845 int printed = 0, enforced = 0;
2850 printed += scnprintf(msg + printed, size - printed,
2851 "Access to performance monitoring and observability operations is limited.\n");
2853 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2855 printed += scnprintf(msg + printed, size - printed,
2856 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2857 "monitoring and observability operations. Inspect system audit records for\n"
2858 "more perf_event access control information and adjusting the policy.\n");
2863 printed += scnprintf(msg, size,
2864 "No permission to enable %s event.\n\n", evsel__name(evsel));
2866 return scnprintf(msg + printed, size - printed,
2867 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2868 "access to performance monitoring and observability operations for processes\n"
2869 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2870 "More information can be found at 'Perf events and tool security' document:\n"
2871 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2872 "perf_event_paranoid setting is %d:\n"
2873 " -1: Allow use of (almost) all events by all users\n"
2874 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2875 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2876 ">= 1: Disallow CPU event access\n"
2877 ">= 2: Disallow kernel profiling\n"
2878 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2879 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2880 perf_event_paranoid());
2882 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2884 return scnprintf(msg, size, "%s",
2885 "Too many events are opened.\n"
2886 "Probably the maximum number of open file descriptors has been reached.\n"
2887 "Hint: Try again after reducing the number of events.\n"
2888 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2890 if (evsel__has_callchain(evsel) &&
2891 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2892 return scnprintf(msg, size,
2893 "Not enough memory to setup event with callchain.\n"
2894 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2895 "Hint: Current value: %d", sysctl__max_stack());
2898 if (target->cpu_list)
2899 return scnprintf(msg, size, "%s",
2900 "No such device - did you specify an out-of-range profile CPU?");
2903 if (evsel->core.attr.aux_output)
2904 return scnprintf(msg, size,
2905 "%s: PMU Hardware doesn't support 'aux_output' feature",
2906 evsel__name(evsel));
2907 if (evsel->core.attr.sample_period != 0)
2908 return scnprintf(msg, size,
2909 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2910 evsel__name(evsel));
2911 if (evsel->core.attr.precise_ip)
2912 return scnprintf(msg, size, "%s",
2913 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2914 #if defined(__i386__) || defined(__x86_64__)
2915 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2916 return scnprintf(msg, size, "%s",
2917 "No hardware sampling interrupt available.\n");
2921 if (find_process("oprofiled"))
2922 return scnprintf(msg, size,
2923 "The PMU counters are busy/taken by another profiler.\n"
2924 "We found oprofile daemon running, please stop it and try again.");
2927 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
2928 return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
2929 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
2930 return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
2931 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2932 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2933 if (perf_missing_features.clockid)
2934 return scnprintf(msg, size, "clockid feature not supported.");
2935 if (perf_missing_features.clockid_wrong)
2936 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2937 if (perf_missing_features.aux_output)
2938 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2939 if (!target__has_cpu(target))
2940 return scnprintf(msg, size,
2941 "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
2942 evsel__name(evsel));
2945 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
2946 "Please add an auxiliary event in front of the load latency event.");
2951 return scnprintf(msg, size,
2952 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2953 "/bin/dmesg | grep -i perf may provide additional information.\n",
2954 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2957 struct perf_env *evsel__env(struct evsel *evsel)
2959 if (evsel && evsel->evlist)
2960 return evsel->evlist->env;
2964 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2966 int cpu_map_idx, thread;
2968 for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
2969 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2971 int fd = FD(evsel, cpu_map_idx, thread);
2973 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2974 cpu_map_idx, thread, fd) < 0)
2982 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2984 struct perf_cpu_map *cpus = evsel->core.cpus;
2985 struct perf_thread_map *threads = evsel->core.threads;
2987 if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
2990 return store_evsel_ids(evsel, evlist);
2993 void evsel__zero_per_pkg(struct evsel *evsel)
2995 struct hashmap_entry *cur;
2998 if (evsel->per_pkg_mask) {
2999 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3000 free((char *)cur->key);
3002 hashmap__clear(evsel->per_pkg_mask);
3006 bool evsel__is_hybrid(struct evsel *evsel)
3008 return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3011 struct evsel *evsel__leader(struct evsel *evsel)
3013 return container_of(evsel->core.leader, struct evsel, core);
3016 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3018 return evsel->core.leader == &leader->core;
3021 bool evsel__is_leader(struct evsel *evsel)
3023 return evsel__has_leader(evsel, evsel);
3026 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3028 evsel->core.leader = &leader->core;
3031 int evsel__source_count(const struct evsel *evsel)
3036 evlist__for_each_entry(evsel->evlist, pos) {
3037 if (pos->metric_leader == evsel)