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 <linux/hw_breakpoint.h>
16 #include <linux/perf_event.h>
17 #include <linux/compiler.h>
18 #include <linux/err.h>
19 #include <linux/zalloc.h>
20 #include <sys/ioctl.h>
21 #include <sys/resource.h>
22 #include <sys/types.h>
25 #include <perf/evsel.h>
27 #include "bpf_counter.h"
28 #include "callchain.h"
34 #include "util/evsel_config.h"
35 #include "util/evsel_fprintf.h"
37 #include <perf/cpumap.h>
38 #include "thread_map.h"
40 #include "perf_regs.h"
43 #include "trace-event.h"
48 #include "util/hashmap.h"
49 #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>
55 #include <internal/threadmap.h>
57 #include <linux/ctype.h>
59 #ifdef HAVE_LIBTRACEEVENT
60 #include <traceevent/event-parse.h>
63 struct perf_missing_features perf_missing_features;
65 static clockid_t clockid;
67 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
74 const char *perf_tool_event__to_str(enum perf_tool_event ev)
76 if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
77 return perf_tool_event__tool_names[ev];
82 enum perf_tool_event perf_tool_event__from_str(const char *str)
86 perf_tool_event__for_each_event(i) {
87 if (!strcmp(str, perf_tool_event__tool_names[i]))
90 return PERF_TOOL_NONE;
94 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
99 void __weak test_attr__ready(void) { }
101 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
107 int (*init)(struct evsel *evsel);
108 void (*fini)(struct evsel *evsel);
109 } perf_evsel__object = {
110 .size = sizeof(struct evsel),
111 .init = evsel__no_extra_init,
112 .fini = evsel__no_extra_fini,
115 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
116 void (*fini)(struct evsel *evsel))
119 if (object_size == 0)
122 if (perf_evsel__object.size > object_size)
125 perf_evsel__object.size = object_size;
129 perf_evsel__object.init = init;
132 perf_evsel__object.fini = fini;
137 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
139 int __evsel__sample_size(u64 sample_type)
141 u64 mask = sample_type & PERF_SAMPLE_MASK;
145 for (i = 0; i < 64; i++) {
146 if (mask & (1ULL << i))
156 * __perf_evsel__calc_id_pos - calculate id_pos.
157 * @sample_type: sample type
159 * This function returns the position of the event id (PERF_SAMPLE_ID or
160 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
161 * perf_record_sample.
163 static int __perf_evsel__calc_id_pos(u64 sample_type)
167 if (sample_type & PERF_SAMPLE_IDENTIFIER)
170 if (!(sample_type & PERF_SAMPLE_ID))
173 if (sample_type & PERF_SAMPLE_IP)
176 if (sample_type & PERF_SAMPLE_TID)
179 if (sample_type & PERF_SAMPLE_TIME)
182 if (sample_type & PERF_SAMPLE_ADDR)
189 * __perf_evsel__calc_is_pos - calculate is_pos.
190 * @sample_type: sample type
192 * This function returns the position (counting backwards) of the event id
193 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
194 * sample_id_all is used there is an id sample appended to non-sample events.
196 static int __perf_evsel__calc_is_pos(u64 sample_type)
200 if (sample_type & PERF_SAMPLE_IDENTIFIER)
203 if (!(sample_type & PERF_SAMPLE_ID))
206 if (sample_type & PERF_SAMPLE_CPU)
209 if (sample_type & PERF_SAMPLE_STREAM_ID)
215 void evsel__calc_id_pos(struct evsel *evsel)
217 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
218 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
221 void __evsel__set_sample_bit(struct evsel *evsel,
222 enum perf_event_sample_format bit)
224 if (!(evsel->core.attr.sample_type & bit)) {
225 evsel->core.attr.sample_type |= bit;
226 evsel->sample_size += sizeof(u64);
227 evsel__calc_id_pos(evsel);
231 void __evsel__reset_sample_bit(struct evsel *evsel,
232 enum perf_event_sample_format bit)
234 if (evsel->core.attr.sample_type & bit) {
235 evsel->core.attr.sample_type &= ~bit;
236 evsel->sample_size -= sizeof(u64);
237 evsel__calc_id_pos(evsel);
241 void evsel__set_sample_id(struct evsel *evsel,
242 bool can_sample_identifier)
244 if (can_sample_identifier) {
245 evsel__reset_sample_bit(evsel, ID);
246 evsel__set_sample_bit(evsel, IDENTIFIER);
248 evsel__set_sample_bit(evsel, ID);
250 evsel->core.attr.read_format |= PERF_FORMAT_ID;
254 * evsel__is_function_event - Return whether given evsel is a function
257 * @evsel - evsel selector to be tested
259 * Return %true if event is function trace event
261 bool evsel__is_function_event(struct evsel *evsel)
263 #define FUNCTION_EVENT "ftrace:function"
265 return evsel->name &&
266 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
268 #undef FUNCTION_EVENT
271 void evsel__init(struct evsel *evsel,
272 struct perf_event_attr *attr, int idx)
274 perf_evsel__init(&evsel->core, attr, idx);
275 evsel->tracking = !idx;
276 evsel->unit = strdup("");
278 evsel->max_events = ULONG_MAX;
279 evsel->evlist = NULL;
280 evsel->bpf_obj = NULL;
282 INIT_LIST_HEAD(&evsel->config_terms);
283 INIT_LIST_HEAD(&evsel->bpf_counter_list);
284 perf_evsel__object.init(evsel);
285 evsel->sample_size = __evsel__sample_size(attr->sample_type);
286 evsel__calc_id_pos(evsel);
287 evsel->cmdline_group_boundary = false;
288 evsel->metric_expr = NULL;
289 evsel->metric_name = NULL;
290 evsel->metric_events = NULL;
291 evsel->per_pkg_mask = NULL;
292 evsel->collect_stat = false;
293 evsel->pmu_name = NULL;
296 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
298 struct evsel *evsel = zalloc(perf_evsel__object.size);
302 evsel__init(evsel, attr, idx);
304 if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
305 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
306 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
307 evsel->core.attr.sample_period = 1;
310 if (evsel__is_clock(evsel)) {
311 free((char *)evsel->unit);
312 evsel->unit = strdup("msec");
319 static bool perf_event_can_profile_kernel(void)
321 return perf_event_paranoid_check(1);
324 struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
326 struct perf_event_attr attr = {
329 .exclude_kernel = !perf_event_can_profile_kernel(),
333 event_attr_init(&attr);
336 * Now let the usual logic to set up the perf_event_attr defaults
337 * to kick in when we return and before perf_evsel__open() is called.
339 evsel = evsel__new(&attr);
343 arch_evsel__fixup_new_cycles(&evsel->core.attr);
345 evsel->precise_max = true;
347 /* use asprintf() because free(evsel) assumes name is allocated */
348 if (asprintf(&evsel->name, "cycles%s%s%.*s",
349 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
350 attr.exclude_kernel ? "u" : "",
351 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
356 evsel__delete(evsel);
361 int copy_config_terms(struct list_head *dst, struct list_head *src)
363 struct evsel_config_term *pos, *tmp;
365 list_for_each_entry(pos, src, list) {
366 tmp = malloc(sizeof(*tmp));
372 tmp->val.str = strdup(pos->val.str);
373 if (tmp->val.str == NULL) {
378 list_add_tail(&tmp->list, dst);
383 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
385 return copy_config_terms(&dst->config_terms, &src->config_terms);
389 * evsel__clone - create a new evsel copied from @orig
390 * @orig: original evsel
392 * The assumption is that @orig is not configured nor opened yet.
393 * So we only care about the attributes that can be set while it's parsed.
395 struct evsel *evsel__clone(struct evsel *orig)
399 BUG_ON(orig->core.fd);
400 BUG_ON(orig->counts);
402 BUG_ON(orig->per_pkg_mask);
404 /* cannot handle BPF objects for now */
408 evsel = evsel__new(&orig->core.attr);
412 evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
413 evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
414 evsel->core.threads = perf_thread_map__get(orig->core.threads);
415 evsel->core.nr_members = orig->core.nr_members;
416 evsel->core.system_wide = orig->core.system_wide;
417 evsel->core.requires_cpu = orig->core.requires_cpu;
420 evsel->name = strdup(orig->name);
421 if (evsel->name == NULL)
424 if (orig->group_name) {
425 evsel->group_name = strdup(orig->group_name);
426 if (evsel->group_name == NULL)
429 if (orig->pmu_name) {
430 evsel->pmu_name = strdup(orig->pmu_name);
431 if (evsel->pmu_name == NULL)
435 evsel->filter = strdup(orig->filter);
436 if (evsel->filter == NULL)
439 if (orig->metric_id) {
440 evsel->metric_id = strdup(orig->metric_id);
441 if (evsel->metric_id == NULL)
444 evsel->cgrp = cgroup__get(orig->cgrp);
445 #ifdef HAVE_LIBTRACEEVENT
446 evsel->tp_format = orig->tp_format;
448 evsel->handler = orig->handler;
449 evsel->core.leader = orig->core.leader;
451 evsel->max_events = orig->max_events;
452 evsel->tool_event = orig->tool_event;
453 free((char *)evsel->unit);
454 evsel->unit = strdup(orig->unit);
455 if (evsel->unit == NULL)
458 evsel->scale = orig->scale;
459 evsel->snapshot = orig->snapshot;
460 evsel->per_pkg = orig->per_pkg;
461 evsel->percore = orig->percore;
462 evsel->precise_max = orig->precise_max;
463 evsel->use_uncore_alias = orig->use_uncore_alias;
464 evsel->is_libpfm_event = orig->is_libpfm_event;
466 evsel->exclude_GH = orig->exclude_GH;
467 evsel->sample_read = orig->sample_read;
468 evsel->auto_merge_stats = orig->auto_merge_stats;
469 evsel->collect_stat = orig->collect_stat;
470 evsel->weak_group = orig->weak_group;
471 evsel->use_config_name = orig->use_config_name;
472 evsel->pmu = orig->pmu;
474 if (evsel__copy_config_terms(evsel, orig) < 0)
480 evsel__delete(evsel);
485 * Returns pointer with encoded error via <linux/err.h> interface.
487 #ifdef HAVE_LIBTRACEEVENT
488 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
490 struct evsel *evsel = zalloc(perf_evsel__object.size);
496 struct perf_event_attr attr = {
497 .type = PERF_TYPE_TRACEPOINT,
498 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
499 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
502 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
505 evsel->tp_format = trace_event__tp_format(sys, name);
506 if (IS_ERR(evsel->tp_format)) {
507 err = PTR_ERR(evsel->tp_format);
511 event_attr_init(&attr);
512 attr.config = evsel->tp_format->id;
513 attr.sample_period = 1;
514 evsel__init(evsel, &attr, idx);
527 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
535 "stalled-cycles-frontend",
536 "stalled-cycles-backend",
540 char *evsel__bpf_counter_events;
542 bool evsel__match_bpf_counter_events(const char *name)
548 if (!evsel__bpf_counter_events)
551 ptr = strstr(evsel__bpf_counter_events, name);
552 name_len = strlen(name);
554 /* check name matches a full token in evsel__bpf_counter_events */
555 match = (ptr != NULL) &&
556 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
557 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
562 static const char *__evsel__hw_name(u64 config)
564 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
565 return evsel__hw_names[config];
567 return "unknown-hardware";
570 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
572 int colon = 0, r = 0;
573 struct perf_event_attr *attr = &evsel->core.attr;
574 bool exclude_guest_default = false;
576 #define MOD_PRINT(context, mod) do { \
577 if (!attr->exclude_##context) { \
578 if (!colon) colon = ++r; \
579 r += scnprintf(bf + r, size - r, "%c", mod); \
582 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
583 MOD_PRINT(kernel, 'k');
584 MOD_PRINT(user, 'u');
586 exclude_guest_default = true;
589 if (attr->precise_ip) {
592 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
593 exclude_guest_default = true;
596 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
597 MOD_PRINT(host, 'H');
598 MOD_PRINT(guest, 'G');
606 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
608 return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
611 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
613 int r = arch_evsel__hw_name(evsel, bf, size);
614 return r + evsel__add_modifiers(evsel, bf + r, size - r);
617 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
630 static const char *__evsel__sw_name(u64 config)
632 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
633 return evsel__sw_names[config];
634 return "unknown-software";
637 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
639 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
640 return r + evsel__add_modifiers(evsel, bf + r, size - r);
643 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
645 return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
648 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
652 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
654 if (type & HW_BREAKPOINT_R)
655 r += scnprintf(bf + r, size - r, "r");
657 if (type & HW_BREAKPOINT_W)
658 r += scnprintf(bf + r, size - r, "w");
660 if (type & HW_BREAKPOINT_X)
661 r += scnprintf(bf + r, size - r, "x");
666 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
668 struct perf_event_attr *attr = &evsel->core.attr;
669 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
670 return r + evsel__add_modifiers(evsel, bf + r, size - r);
673 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
674 { "L1-dcache", "l1-d", "l1d", "L1-data", },
675 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
677 { "dTLB", "d-tlb", "Data-TLB", },
678 { "iTLB", "i-tlb", "Instruction-TLB", },
679 { "branch", "branches", "bpu", "btb", "bpc", },
683 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
684 { "load", "loads", "read", },
685 { "store", "stores", "write", },
686 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
689 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
690 { "refs", "Reference", "ops", "access", },
691 { "misses", "miss", },
694 #define C(x) PERF_COUNT_HW_CACHE_##x
695 #define CACHE_READ (1 << C(OP_READ))
696 #define CACHE_WRITE (1 << C(OP_WRITE))
697 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
698 #define COP(x) (1 << x)
701 * cache operation stat
702 * L1I : Read and prefetch only
703 * ITLB and BPU : Read-only
705 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
706 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
707 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
708 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
709 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
710 [C(ITLB)] = (CACHE_READ),
711 [C(BPU)] = (CACHE_READ),
712 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
715 bool evsel__is_cache_op_valid(u8 type, u8 op)
717 if (evsel__hw_cache_stat[type] & COP(op))
718 return true; /* valid */
720 return false; /* invalid */
723 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
726 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
727 evsel__hw_cache_op[op][0],
728 evsel__hw_cache_result[result][0]);
731 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
732 evsel__hw_cache_op[op][1]);
735 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
737 u8 op, result, type = (config >> 0) & 0xff;
738 const char *err = "unknown-ext-hardware-cache-type";
740 if (type >= PERF_COUNT_HW_CACHE_MAX)
743 op = (config >> 8) & 0xff;
744 err = "unknown-ext-hardware-cache-op";
745 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
748 result = (config >> 16) & 0xff;
749 err = "unknown-ext-hardware-cache-result";
750 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
753 err = "invalid-cache";
754 if (!evsel__is_cache_op_valid(type, op))
757 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
759 return scnprintf(bf, size, "%s", err);
762 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
764 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
765 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
768 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
770 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
771 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
774 const char *evsel__name(struct evsel *evsel)
784 switch (evsel->core.attr.type) {
786 evsel__raw_name(evsel, bf, sizeof(bf));
789 case PERF_TYPE_HARDWARE:
790 evsel__hw_name(evsel, bf, sizeof(bf));
793 case PERF_TYPE_HW_CACHE:
794 evsel__hw_cache_name(evsel, bf, sizeof(bf));
797 case PERF_TYPE_SOFTWARE:
798 if (evsel__is_tool(evsel))
799 evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
801 evsel__sw_name(evsel, bf, sizeof(bf));
804 case PERF_TYPE_TRACEPOINT:
805 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
808 case PERF_TYPE_BREAKPOINT:
809 evsel__bp_name(evsel, bf, sizeof(bf));
813 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
814 evsel->core.attr.type);
818 evsel->name = strdup(bf);
826 const char *evsel__metric_id(const struct evsel *evsel)
828 if (evsel->metric_id)
829 return evsel->metric_id;
831 if (evsel__is_tool(evsel))
832 return perf_tool_event__to_str(evsel->tool_event);
837 const char *evsel__group_name(struct evsel *evsel)
839 return evsel->group_name ?: "anon group";
843 * Returns the group details for the specified leader,
844 * with following rules.
846 * For record -e '{cycles,instructions}'
847 * 'anon group { cycles:u, instructions:u }'
849 * For record -e 'cycles,instructions' and report --group
850 * 'cycles:u, instructions:u'
852 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
856 const char *group_name = evsel__group_name(evsel);
858 if (!evsel->forced_leader)
859 ret = scnprintf(buf, size, "%s { ", group_name);
861 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
863 for_each_group_member(pos, evsel)
864 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
866 if (!evsel->forced_leader)
867 ret += scnprintf(buf + ret, size - ret, " }");
872 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
873 struct callchain_param *param)
875 bool function = evsel__is_function_event(evsel);
876 struct perf_event_attr *attr = &evsel->core.attr;
878 evsel__set_sample_bit(evsel, CALLCHAIN);
880 attr->sample_max_stack = param->max_stack;
882 if (opts->kernel_callchains)
883 attr->exclude_callchain_user = 1;
884 if (opts->user_callchains)
885 attr->exclude_callchain_kernel = 1;
886 if (param->record_mode == CALLCHAIN_LBR) {
887 if (!opts->branch_stack) {
888 if (attr->exclude_user) {
889 pr_warning("LBR callstack option is only available "
890 "to get user callchain information. "
891 "Falling back to framepointers.\n");
893 evsel__set_sample_bit(evsel, BRANCH_STACK);
894 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
895 PERF_SAMPLE_BRANCH_CALL_STACK |
896 PERF_SAMPLE_BRANCH_NO_CYCLES |
897 PERF_SAMPLE_BRANCH_NO_FLAGS |
898 PERF_SAMPLE_BRANCH_HW_INDEX;
901 pr_warning("Cannot use LBR callstack with branch stack. "
902 "Falling back to framepointers.\n");
905 if (param->record_mode == CALLCHAIN_DWARF) {
907 evsel__set_sample_bit(evsel, REGS_USER);
908 evsel__set_sample_bit(evsel, STACK_USER);
909 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
910 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
911 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
912 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
913 "so the minimal registers set (IP, SP) is explicitly forced.\n");
915 attr->sample_regs_user |= arch__user_reg_mask();
917 attr->sample_stack_user = param->dump_size;
918 attr->exclude_callchain_user = 1;
920 pr_info("Cannot use DWARF unwind for function trace event,"
921 " falling back to framepointers.\n");
926 pr_info("Disabling user space callchains for function trace event.\n");
927 attr->exclude_callchain_user = 1;
931 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
932 struct callchain_param *param)
935 return __evsel__config_callchain(evsel, opts, param);
938 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
940 struct perf_event_attr *attr = &evsel->core.attr;
942 evsel__reset_sample_bit(evsel, CALLCHAIN);
943 if (param->record_mode == CALLCHAIN_LBR) {
944 evsel__reset_sample_bit(evsel, BRANCH_STACK);
945 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
946 PERF_SAMPLE_BRANCH_CALL_STACK |
947 PERF_SAMPLE_BRANCH_HW_INDEX);
949 if (param->record_mode == CALLCHAIN_DWARF) {
950 evsel__reset_sample_bit(evsel, REGS_USER);
951 evsel__reset_sample_bit(evsel, STACK_USER);
955 static void evsel__apply_config_terms(struct evsel *evsel,
956 struct record_opts *opts, bool track)
958 struct evsel_config_term *term;
959 struct list_head *config_terms = &evsel->config_terms;
960 struct perf_event_attr *attr = &evsel->core.attr;
961 /* callgraph default */
962 struct callchain_param param = {
963 .record_mode = callchain_param.record_mode,
967 const char *callgraph_buf = NULL;
969 list_for_each_entry(term, config_terms, list) {
970 switch (term->type) {
971 case EVSEL__CONFIG_TERM_PERIOD:
972 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
973 attr->sample_period = term->val.period;
975 evsel__reset_sample_bit(evsel, PERIOD);
978 case EVSEL__CONFIG_TERM_FREQ:
979 if (!(term->weak && opts->user_freq != UINT_MAX)) {
980 attr->sample_freq = term->val.freq;
982 evsel__set_sample_bit(evsel, PERIOD);
985 case EVSEL__CONFIG_TERM_TIME:
987 evsel__set_sample_bit(evsel, TIME);
989 evsel__reset_sample_bit(evsel, TIME);
991 case EVSEL__CONFIG_TERM_CALLGRAPH:
992 callgraph_buf = term->val.str;
994 case EVSEL__CONFIG_TERM_BRANCH:
995 if (term->val.str && strcmp(term->val.str, "no")) {
996 evsel__set_sample_bit(evsel, BRANCH_STACK);
997 parse_branch_str(term->val.str,
998 &attr->branch_sample_type);
1000 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1002 case EVSEL__CONFIG_TERM_STACK_USER:
1003 dump_size = term->val.stack_user;
1005 case EVSEL__CONFIG_TERM_MAX_STACK:
1006 max_stack = term->val.max_stack;
1008 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1009 evsel->max_events = term->val.max_events;
1011 case EVSEL__CONFIG_TERM_INHERIT:
1013 * attr->inherit should has already been set by
1014 * evsel__config. If user explicitly set
1015 * inherit using config terms, override global
1016 * opt->no_inherit setting.
1018 attr->inherit = term->val.inherit ? 1 : 0;
1020 case EVSEL__CONFIG_TERM_OVERWRITE:
1021 attr->write_backward = term->val.overwrite ? 1 : 0;
1023 case EVSEL__CONFIG_TERM_DRV_CFG:
1025 case EVSEL__CONFIG_TERM_PERCORE:
1027 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1028 attr->aux_output = term->val.aux_output ? 1 : 0;
1030 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1031 /* Already applied by auxtrace */
1033 case EVSEL__CONFIG_TERM_CFG_CHG:
1040 /* User explicitly set per-event callgraph, clear the old setting and reset. */
1041 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1042 bool sample_address = false;
1045 param.max_stack = max_stack;
1046 if (callgraph_buf == NULL)
1047 callgraph_buf = "fp";
1050 /* parse callgraph parameters */
1051 if (callgraph_buf != NULL) {
1052 if (!strcmp(callgraph_buf, "no")) {
1053 param.enabled = false;
1054 param.record_mode = CALLCHAIN_NONE;
1056 param.enabled = true;
1057 if (parse_callchain_record(callgraph_buf, ¶m)) {
1058 pr_err("per-event callgraph setting for %s failed. "
1059 "Apply callgraph global setting for it\n",
1063 if (param.record_mode == CALLCHAIN_DWARF)
1064 sample_address = true;
1067 if (dump_size > 0) {
1068 dump_size = round_up(dump_size, sizeof(u64));
1069 param.dump_size = dump_size;
1072 /* If global callgraph set, clear it */
1073 if (callchain_param.enabled)
1074 evsel__reset_callgraph(evsel, &callchain_param);
1076 /* set perf-event callgraph */
1077 if (param.enabled) {
1078 if (sample_address) {
1079 evsel__set_sample_bit(evsel, ADDR);
1080 evsel__set_sample_bit(evsel, DATA_SRC);
1081 evsel->core.attr.mmap_data = track;
1083 evsel__config_callchain(evsel, opts, ¶m);
1088 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1090 struct evsel_config_term *term, *found_term = NULL;
1092 list_for_each_entry(term, &evsel->config_terms, list) {
1093 if (term->type == type)
1100 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1102 evsel__set_sample_bit(evsel, WEIGHT);
1105 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1109 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1110 struct perf_event_attr *attr __maybe_unused)
1114 static void evsel__set_default_freq_period(struct record_opts *opts,
1115 struct perf_event_attr *attr)
1119 attr->sample_freq = opts->freq;
1121 attr->sample_period = opts->default_interval;
1125 static bool evsel__is_offcpu_event(struct evsel *evsel)
1127 return evsel__is_bpf_output(evsel) && !strcmp(evsel->name, OFFCPU_EVENT);
1131 * The enable_on_exec/disabled value strategy:
1133 * 1) For any type of traced program:
1134 * - all independent events and group leaders are disabled
1135 * - all group members are enabled
1137 * Group members are ruled by group leaders. They need to
1138 * be enabled, because the group scheduling relies on that.
1140 * 2) For traced programs executed by perf:
1141 * - all independent events and group leaders have
1142 * enable_on_exec set
1143 * - we don't specifically enable or disable any event during
1144 * the record command
1146 * Independent events and group leaders are initially disabled
1147 * and get enabled by exec. Group members are ruled by group
1148 * leaders as stated in 1).
1150 * 3) For traced programs attached by perf (pid/tid):
1151 * - we specifically enable or disable all events during
1152 * the record command
1154 * When attaching events to already running traced we
1155 * enable/disable events specifically, as there's no
1156 * initial traced exec call.
1158 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1159 struct callchain_param *callchain)
1161 struct evsel *leader = evsel__leader(evsel);
1162 struct perf_event_attr *attr = &evsel->core.attr;
1163 int track = evsel->tracking;
1164 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1166 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1167 attr->inherit = !opts->no_inherit;
1168 attr->write_backward = opts->overwrite ? 1 : 0;
1169 attr->read_format = PERF_FORMAT_LOST;
1171 evsel__set_sample_bit(evsel, IP);
1172 evsel__set_sample_bit(evsel, TID);
1174 if (evsel->sample_read) {
1175 evsel__set_sample_bit(evsel, READ);
1178 * We need ID even in case of single event, because
1179 * PERF_SAMPLE_READ process ID specific data.
1181 evsel__set_sample_id(evsel, false);
1184 * Apply group format only if we belong to group
1185 * with more than one members.
1187 if (leader->core.nr_members > 1) {
1188 attr->read_format |= PERF_FORMAT_GROUP;
1194 * We default some events to have a default interval. But keep
1195 * it a weak assumption overridable by the user.
1197 if ((evsel->is_libpfm_event && !attr->sample_period) ||
1198 (!evsel->is_libpfm_event && (!attr->sample_period ||
1199 opts->user_freq != UINT_MAX ||
1200 opts->user_interval != ULLONG_MAX)))
1201 evsel__set_default_freq_period(opts, attr);
1204 * If attr->freq was set (here or earlier), ask for period
1208 evsel__set_sample_bit(evsel, PERIOD);
1210 if (opts->no_samples)
1211 attr->sample_freq = 0;
1213 if (opts->inherit_stat) {
1214 evsel->core.attr.read_format |=
1215 PERF_FORMAT_TOTAL_TIME_ENABLED |
1216 PERF_FORMAT_TOTAL_TIME_RUNNING |
1218 attr->inherit_stat = 1;
1221 if (opts->sample_address) {
1222 evsel__set_sample_bit(evsel, ADDR);
1223 attr->mmap_data = track;
1227 * We don't allow user space callchains for function trace
1228 * event, due to issues with page faults while tracing page
1229 * fault handler and its overall trickiness nature.
1231 if (evsel__is_function_event(evsel))
1232 evsel->core.attr.exclude_callchain_user = 1;
1234 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1235 evsel__config_callchain(evsel, opts, callchain);
1237 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1238 !evsel__is_dummy_event(evsel)) {
1239 attr->sample_regs_intr = opts->sample_intr_regs;
1240 evsel__set_sample_bit(evsel, REGS_INTR);
1243 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1244 !evsel__is_dummy_event(evsel)) {
1245 attr->sample_regs_user |= opts->sample_user_regs;
1246 evsel__set_sample_bit(evsel, REGS_USER);
1249 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1250 evsel__set_sample_bit(evsel, CPU);
1253 * When the user explicitly disabled time don't force it here.
1255 if (opts->sample_time &&
1256 (!perf_missing_features.sample_id_all &&
1257 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1258 opts->sample_time_set)))
1259 evsel__set_sample_bit(evsel, TIME);
1261 if (opts->raw_samples && !evsel->no_aux_samples) {
1262 evsel__set_sample_bit(evsel, TIME);
1263 evsel__set_sample_bit(evsel, RAW);
1264 evsel__set_sample_bit(evsel, CPU);
1267 if (opts->sample_address)
1268 evsel__set_sample_bit(evsel, DATA_SRC);
1270 if (opts->sample_phys_addr)
1271 evsel__set_sample_bit(evsel, PHYS_ADDR);
1273 if (opts->no_buffering) {
1274 attr->watermark = 0;
1275 attr->wakeup_events = 1;
1277 if (opts->branch_stack && !evsel->no_aux_samples) {
1278 evsel__set_sample_bit(evsel, BRANCH_STACK);
1279 attr->branch_sample_type = opts->branch_stack;
1282 if (opts->sample_weight)
1283 arch_evsel__set_sample_weight(evsel);
1287 attr->mmap2 = track && !perf_missing_features.mmap2;
1289 attr->build_id = track && opts->build_id;
1292 * ksymbol is tracked separately with text poke because it needs to be
1293 * system wide and enabled immediately.
1295 if (!opts->text_poke)
1296 attr->ksymbol = track && !perf_missing_features.ksymbol;
1297 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1299 if (opts->record_namespaces)
1300 attr->namespaces = track;
1302 if (opts->record_cgroup) {
1303 attr->cgroup = track && !perf_missing_features.cgroup;
1304 evsel__set_sample_bit(evsel, CGROUP);
1307 if (opts->sample_data_page_size)
1308 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1310 if (opts->sample_code_page_size)
1311 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1313 if (opts->record_switch_events)
1314 attr->context_switch = track;
1316 if (opts->sample_transaction)
1317 evsel__set_sample_bit(evsel, TRANSACTION);
1319 if (opts->running_time) {
1320 evsel->core.attr.read_format |=
1321 PERF_FORMAT_TOTAL_TIME_ENABLED |
1322 PERF_FORMAT_TOTAL_TIME_RUNNING;
1326 * XXX see the function comment above
1328 * Disabling only independent events or group leaders,
1329 * keeping group members enabled.
1331 if (evsel__is_group_leader(evsel))
1335 * Setting enable_on_exec for independent events and
1336 * group leaders for traced executed by perf.
1338 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1339 !opts->initial_delay)
1340 attr->enable_on_exec = 1;
1342 if (evsel->immediate) {
1344 attr->enable_on_exec = 0;
1347 clockid = opts->clockid;
1348 if (opts->use_clockid) {
1349 attr->use_clockid = 1;
1350 attr->clockid = opts->clockid;
1353 if (evsel->precise_max)
1354 attr->precise_ip = 3;
1356 if (opts->all_user) {
1357 attr->exclude_kernel = 1;
1358 attr->exclude_user = 0;
1361 if (opts->all_kernel) {
1362 attr->exclude_kernel = 0;
1363 attr->exclude_user = 1;
1366 if (evsel->core.own_cpus || evsel->unit)
1367 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1370 * Apply event specific term settings,
1371 * it overloads any global configuration.
1373 evsel__apply_config_terms(evsel, opts, track);
1375 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1377 /* The --period option takes the precedence. */
1378 if (opts->period_set) {
1380 evsel__set_sample_bit(evsel, PERIOD);
1382 evsel__reset_sample_bit(evsel, PERIOD);
1386 * A dummy event never triggers any actual counter and therefore
1387 * cannot be used with branch_stack.
1389 * For initial_delay, a dummy event is added implicitly.
1390 * The software event will trigger -EOPNOTSUPP error out,
1391 * if BRANCH_STACK bit is set.
1393 if (evsel__is_dummy_event(evsel))
1394 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1396 if (evsel__is_offcpu_event(evsel))
1397 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1399 arch__post_evsel_config(evsel, attr);
1402 int evsel__set_filter(struct evsel *evsel, const char *filter)
1404 char *new_filter = strdup(filter);
1406 if (new_filter != NULL) {
1407 free(evsel->filter);
1408 evsel->filter = new_filter;
1415 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1419 if (evsel->filter == NULL)
1420 return evsel__set_filter(evsel, filter);
1422 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1423 free(evsel->filter);
1424 evsel->filter = new_filter;
1431 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1433 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1436 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1438 return evsel__append_filter(evsel, "%s,%s", filter);
1441 /* Caller has to clear disabled after going through all CPUs. */
1442 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1444 return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1447 int evsel__enable(struct evsel *evsel)
1449 int err = perf_evsel__enable(&evsel->core);
1452 evsel->disabled = false;
1456 /* Caller has to set disabled after going through all CPUs. */
1457 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1459 return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1462 int evsel__disable(struct evsel *evsel)
1464 int err = perf_evsel__disable(&evsel->core);
1466 * We mark it disabled here so that tools that disable a event can
1467 * ignore events after they disable it. I.e. the ring buffer may have
1468 * already a few more events queued up before the kernel got the stop
1472 evsel->disabled = true;
1477 void free_config_terms(struct list_head *config_terms)
1479 struct evsel_config_term *term, *h;
1481 list_for_each_entry_safe(term, h, config_terms, list) {
1482 list_del_init(&term->list);
1484 zfree(&term->val.str);
1489 static void evsel__free_config_terms(struct evsel *evsel)
1491 free_config_terms(&evsel->config_terms);
1494 void evsel__exit(struct evsel *evsel)
1496 assert(list_empty(&evsel->core.node));
1497 assert(evsel->evlist == NULL);
1498 bpf_counter__destroy(evsel);
1499 evsel__free_counts(evsel);
1500 perf_evsel__free_fd(&evsel->core);
1501 perf_evsel__free_id(&evsel->core);
1502 evsel__free_config_terms(evsel);
1503 cgroup__put(evsel->cgrp);
1504 perf_cpu_map__put(evsel->core.cpus);
1505 perf_cpu_map__put(evsel->core.own_cpus);
1506 perf_thread_map__put(evsel->core.threads);
1507 zfree(&evsel->group_name);
1508 zfree(&evsel->name);
1509 zfree(&evsel->pmu_name);
1510 zfree(&evsel->unit);
1511 zfree(&evsel->metric_id);
1512 evsel__zero_per_pkg(evsel);
1513 hashmap__free(evsel->per_pkg_mask);
1514 evsel->per_pkg_mask = NULL;
1515 zfree(&evsel->metric_events);
1516 perf_evsel__object.fini(evsel);
1519 void evsel__delete(struct evsel *evsel)
1525 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1526 struct perf_counts_values *count)
1528 struct perf_counts_values tmp;
1530 if (!evsel->prev_raw_counts)
1533 tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1534 *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1536 count->val = count->val - tmp.val;
1537 count->ena = count->ena - tmp.ena;
1538 count->run = count->run - tmp.run;
1541 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1543 struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1545 return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1548 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1549 u64 val, u64 ena, u64 run, u64 lost)
1551 struct perf_counts_values *count;
1553 count = perf_counts(counter->counts, cpu_map_idx, thread);
1560 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1563 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1565 u64 read_format = leader->core.attr.read_format;
1566 struct sample_read_value *v;
1567 u64 nr, ena = 0, run = 0, lost = 0;
1571 if (nr != (u64) leader->core.nr_members)
1574 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1577 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1581 sample_read_group__for_each(v, nr, read_format) {
1582 struct evsel *counter;
1584 counter = evlist__id2evsel(leader->evlist, v->id);
1588 if (read_format & PERF_FORMAT_LOST)
1591 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1597 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1599 struct perf_stat_evsel *ps = leader->stats;
1600 u64 read_format = leader->core.attr.read_format;
1601 int size = perf_evsel__read_size(&leader->core);
1602 u64 *data = ps->group_data;
1604 if (!(read_format & PERF_FORMAT_ID))
1607 if (!evsel__is_group_leader(leader))
1611 data = zalloc(size);
1615 ps->group_data = data;
1618 if (FD(leader, cpu_map_idx, thread) < 0)
1621 if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1624 return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1627 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1629 u64 read_format = evsel->core.attr.read_format;
1631 if (read_format & PERF_FORMAT_GROUP)
1632 return evsel__read_group(evsel, cpu_map_idx, thread);
1634 return evsel__read_one(evsel, cpu_map_idx, thread);
1637 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1639 struct perf_counts_values count;
1640 size_t nv = scale ? 3 : 1;
1642 if (FD(evsel, cpu_map_idx, thread) < 0)
1645 if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1648 if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1651 evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1652 perf_counts_values__scale(&count, scale, NULL);
1653 *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1657 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1660 struct perf_cpu cpu;
1662 cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1663 return perf_cpu_map__idx(other->core.cpus, cpu);
1666 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1668 struct evsel *leader = evsel__leader(evsel);
1670 if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1671 (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1672 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1678 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1680 struct evsel *leader = evsel__leader(evsel);
1683 if (evsel__is_group_leader(evsel))
1687 * Leader must be already processed/open,
1688 * if not it's a bug.
1690 BUG_ON(!leader->core.fd);
1692 cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1693 if (cpu_map_idx == -1)
1696 fd = FD(leader, cpu_map_idx, thread);
1702 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1704 for (int cpu = 0; cpu < nr_cpus; cpu++)
1705 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1706 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1709 static int update_fds(struct evsel *evsel,
1710 int nr_cpus, int cpu_map_idx,
1711 int nr_threads, int thread_idx)
1715 if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1718 evlist__for_each_entry(evsel->evlist, pos) {
1719 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1721 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1724 * Since fds for next evsel has not been created,
1725 * there is no need to iterate whole event list.
1733 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1734 int nr_cpus, int cpu_map_idx,
1735 struct perf_thread_map *threads,
1736 int thread, int err)
1738 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1740 if (!evsel->ignore_missing_thread)
1743 /* The system wide setup does not work with threads. */
1744 if (evsel->core.system_wide)
1747 /* The -ESRCH is perf event syscall errno for pid's not found. */
1751 /* If there's only one thread, let it fail. */
1752 if (threads->nr == 1)
1756 * We should remove fd for missing_thread first
1757 * because thread_map__remove() will decrease threads->nr.
1759 if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1762 if (thread_map__remove(threads, thread))
1765 pr_warning("WARNING: Ignored open failure for pid %d\n",
1770 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1771 void *priv __maybe_unused)
1773 return fprintf(fp, " %-32s %s\n", name, val);
1776 static void display_attr(struct perf_event_attr *attr)
1778 if (verbose >= 2 || debug_peo_args) {
1779 fprintf(stderr, "%.60s\n", graph_dotted_line);
1780 fprintf(stderr, "perf_event_attr:\n");
1781 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1782 fprintf(stderr, "%.60s\n", graph_dotted_line);
1786 bool evsel__precise_ip_fallback(struct evsel *evsel)
1788 /* Do not try less precise if not requested. */
1789 if (!evsel->precise_max)
1793 * We tried all the precise_ip values, and it's
1794 * still failing, so leave it to standard fallback.
1796 if (!evsel->core.attr.precise_ip) {
1797 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1801 if (!evsel->precise_ip_original)
1802 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1804 evsel->core.attr.precise_ip--;
1805 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1806 display_attr(&evsel->core.attr);
1810 static struct perf_cpu_map *empty_cpu_map;
1811 static struct perf_thread_map *empty_thread_map;
1813 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1814 struct perf_thread_map *threads)
1816 int nthreads = perf_thread_map__nr(threads);
1818 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1819 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1823 if (empty_cpu_map == NULL) {
1824 empty_cpu_map = perf_cpu_map__dummy_new();
1825 if (empty_cpu_map == NULL)
1829 cpus = empty_cpu_map;
1832 if (threads == NULL) {
1833 if (empty_thread_map == NULL) {
1834 empty_thread_map = thread_map__new_by_tid(-1);
1835 if (empty_thread_map == NULL)
1839 threads = empty_thread_map;
1842 if (evsel->core.fd == NULL &&
1843 perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1846 evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1848 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1853 static void evsel__disable_missing_features(struct evsel *evsel)
1855 if (perf_missing_features.read_lost)
1856 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1857 if (perf_missing_features.weight_struct) {
1858 evsel__set_sample_bit(evsel, WEIGHT);
1859 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1861 if (perf_missing_features.clockid_wrong)
1862 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1863 if (perf_missing_features.clockid) {
1864 evsel->core.attr.use_clockid = 0;
1865 evsel->core.attr.clockid = 0;
1867 if (perf_missing_features.cloexec)
1868 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1869 if (perf_missing_features.mmap2)
1870 evsel->core.attr.mmap2 = 0;
1871 if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1872 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1873 if (perf_missing_features.lbr_flags)
1874 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1875 PERF_SAMPLE_BRANCH_NO_CYCLES);
1876 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1877 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1878 if (perf_missing_features.ksymbol)
1879 evsel->core.attr.ksymbol = 0;
1880 if (perf_missing_features.bpf)
1881 evsel->core.attr.bpf_event = 0;
1882 if (perf_missing_features.branch_hw_idx)
1883 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1884 if (perf_missing_features.sample_id_all)
1885 evsel->core.attr.sample_id_all = 0;
1888 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1889 struct perf_thread_map *threads)
1893 err = __evsel__prepare_open(evsel, cpus, threads);
1897 evsel__disable_missing_features(evsel);
1902 bool evsel__detect_missing_features(struct evsel *evsel)
1905 * Must probe features in the order they were added to the
1906 * perf_event_attr interface.
1908 if (!perf_missing_features.read_lost &&
1909 (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1910 perf_missing_features.read_lost = true;
1911 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1913 } else if (!perf_missing_features.weight_struct &&
1914 (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1915 perf_missing_features.weight_struct = true;
1916 pr_debug2("switching off weight struct support\n");
1918 } else if (!perf_missing_features.code_page_size &&
1919 (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1920 perf_missing_features.code_page_size = true;
1921 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1923 } else if (!perf_missing_features.data_page_size &&
1924 (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1925 perf_missing_features.data_page_size = true;
1926 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1928 } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1929 perf_missing_features.cgroup = true;
1930 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1932 } else if (!perf_missing_features.branch_hw_idx &&
1933 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1934 perf_missing_features.branch_hw_idx = true;
1935 pr_debug2("switching off branch HW index support\n");
1937 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1938 perf_missing_features.aux_output = true;
1939 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1941 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1942 perf_missing_features.bpf = true;
1943 pr_debug2_peo("switching off bpf_event\n");
1945 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1946 perf_missing_features.ksymbol = true;
1947 pr_debug2_peo("switching off ksymbol\n");
1949 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1950 perf_missing_features.write_backward = true;
1951 pr_debug2_peo("switching off write_backward\n");
1953 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1954 perf_missing_features.clockid_wrong = true;
1955 pr_debug2_peo("switching off clockid\n");
1957 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1958 perf_missing_features.clockid = true;
1959 pr_debug2_peo("switching off use_clockid\n");
1961 } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1962 perf_missing_features.cloexec = true;
1963 pr_debug2_peo("switching off cloexec flag\n");
1965 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1966 perf_missing_features.mmap2 = true;
1967 pr_debug2_peo("switching off mmap2\n");
1969 } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1970 if (evsel->pmu == NULL)
1971 evsel->pmu = evsel__find_pmu(evsel);
1974 evsel->pmu->missing_features.exclude_guest = true;
1976 /* we cannot find PMU, disable attrs now */
1977 evsel->core.attr.exclude_host = false;
1978 evsel->core.attr.exclude_guest = false;
1981 if (evsel->exclude_GH) {
1982 pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1985 if (!perf_missing_features.exclude_guest) {
1986 perf_missing_features.exclude_guest = true;
1987 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1990 } else if (!perf_missing_features.sample_id_all) {
1991 perf_missing_features.sample_id_all = true;
1992 pr_debug2_peo("switching off sample_id_all\n");
1994 } else if (!perf_missing_features.lbr_flags &&
1995 (evsel->core.attr.branch_sample_type &
1996 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1997 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1998 perf_missing_features.lbr_flags = true;
1999 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2001 } else if (!perf_missing_features.group_read &&
2002 evsel->core.attr.inherit &&
2003 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2004 evsel__is_group_leader(evsel)) {
2005 perf_missing_features.group_read = true;
2006 pr_debug2_peo("switching off group read\n");
2013 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2018 if (*set_rlimit < INCREASED_MAX) {
2021 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2022 if (*set_rlimit == NO_CHANGE) {
2023 l.rlim_cur = l.rlim_max;
2025 l.rlim_cur = l.rlim_max + 1000;
2026 l.rlim_max = l.rlim_cur;
2028 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2040 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2041 struct perf_thread_map *threads,
2042 int start_cpu_map_idx, int end_cpu_map_idx)
2044 int idx, thread, nthreads;
2045 int pid = -1, err, old_errno;
2046 enum rlimit_action set_rlimit = NO_CHANGE;
2048 err = __evsel__prepare_open(evsel, cpus, threads);
2053 cpus = empty_cpu_map;
2055 if (threads == NULL)
2056 threads = empty_thread_map;
2058 nthreads = perf_thread_map__nr(threads);
2061 pid = evsel->cgrp->fd;
2063 fallback_missing_features:
2064 evsel__disable_missing_features(evsel);
2066 display_attr(&evsel->core.attr);
2068 for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2070 for (thread = 0; thread < nthreads; thread++) {
2073 if (thread >= nthreads)
2076 if (!evsel->cgrp && !evsel->core.system_wide)
2077 pid = perf_thread_map__pid(threads, thread);
2079 group_fd = get_group_fd(evsel, idx, thread);
2083 /* Debug message used by test scripts */
2084 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
2085 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2087 fd = sys_perf_event_open(&evsel->core.attr, pid,
2088 perf_cpu_map__cpu(cpus, idx).cpu,
2089 group_fd, evsel->open_flags);
2091 FD(evsel, idx, thread) = fd;
2096 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2101 bpf_counter__install_pe(evsel, idx, fd);
2103 if (unlikely(test_attr__enabled)) {
2104 test_attr__open(&evsel->core.attr, pid,
2105 perf_cpu_map__cpu(cpus, idx),
2106 fd, group_fd, evsel->open_flags);
2109 /* Debug message used by test scripts */
2110 pr_debug2_peo(" = %d\n", fd);
2112 if (evsel->bpf_fd >= 0) {
2114 int bpf_fd = evsel->bpf_fd;
2117 PERF_EVENT_IOC_SET_BPF,
2119 if (err && errno != EEXIST) {
2120 pr_err("failed to attach bpf fd %d: %s\n",
2121 bpf_fd, strerror(errno));
2127 set_rlimit = NO_CHANGE;
2130 * If we succeeded but had to kill clockid, fail and
2131 * have evsel__open_strerror() print us a nice error.
2133 if (perf_missing_features.clockid ||
2134 perf_missing_features.clockid_wrong) {
2144 if (evsel__precise_ip_fallback(evsel))
2147 if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2148 idx, threads, thread, err)) {
2149 /* We just removed 1 thread, so lower the upper nthreads limit. */
2152 /* ... and pretend like nothing have happened. */
2157 * perf stat needs between 5 and 22 fds per CPU. When we run out
2158 * of them try to increase the limits.
2160 if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2163 if (err != -EINVAL || idx > 0 || thread > 0)
2166 if (evsel__detect_missing_features(evsel))
2167 goto fallback_missing_features;
2170 threads->err_thread = thread;
2174 while (--thread >= 0) {
2175 if (FD(evsel, idx, thread) >= 0)
2176 close(FD(evsel, idx, thread));
2177 FD(evsel, idx, thread) = -1;
2180 } while (--idx >= 0);
2185 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2186 struct perf_thread_map *threads)
2188 return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2191 void evsel__close(struct evsel *evsel)
2193 perf_evsel__close(&evsel->core);
2194 perf_evsel__free_id(&evsel->core);
2197 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2199 if (cpu_map_idx == -1)
2200 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2202 return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2205 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2207 return evsel__open(evsel, NULL, threads);
2210 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2211 const union perf_event *event,
2212 struct perf_sample *sample)
2214 u64 type = evsel->core.attr.sample_type;
2215 const __u64 *array = event->sample.array;
2216 bool swapped = evsel->needs_swap;
2219 array += ((event->header.size -
2220 sizeof(event->header)) / sizeof(u64)) - 1;
2222 if (type & PERF_SAMPLE_IDENTIFIER) {
2223 sample->id = *array;
2227 if (type & PERF_SAMPLE_CPU) {
2230 /* undo swap of u64, then swap on individual u32s */
2231 u.val64 = bswap_64(u.val64);
2232 u.val32[0] = bswap_32(u.val32[0]);
2235 sample->cpu = u.val32[0];
2239 if (type & PERF_SAMPLE_STREAM_ID) {
2240 sample->stream_id = *array;
2244 if (type & PERF_SAMPLE_ID) {
2245 sample->id = *array;
2249 if (type & PERF_SAMPLE_TIME) {
2250 sample->time = *array;
2254 if (type & PERF_SAMPLE_TID) {
2257 /* undo swap of u64, then swap on individual u32s */
2258 u.val64 = bswap_64(u.val64);
2259 u.val32[0] = bswap_32(u.val32[0]);
2260 u.val32[1] = bswap_32(u.val32[1]);
2263 sample->pid = u.val32[0];
2264 sample->tid = u.val32[1];
2271 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2274 return size > max_size || offset + size > endp;
2277 #define OVERFLOW_CHECK(offset, size, max_size) \
2279 if (overflow(endp, (max_size), (offset), (size))) \
2283 #define OVERFLOW_CHECK_u64(offset) \
2284 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2287 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2290 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2291 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2292 * check the format does not go past the end of the event.
2294 if (sample_size + sizeof(event->header) > event->header.size)
2300 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2302 u64 type __maybe_unused)
2304 data->weight = *array;
2307 u64 evsel__bitfield_swap_branch_flags(u64 value)
2316 * mispred:1 //target mispredicted
2317 * predicted:1 //target predicted
2318 * in_tx:1 //in transaction
2319 * abort:1 //transaction abort
2320 * cycles:16 //cycle count to last branch
2321 * type:4 //branch type
2326 * Avoid bswap64() the entire branch_flag.value,
2327 * as it has variable bit-field sizes. Instead the
2328 * macro takes the bit-field position/size,
2329 * swaps it based on the host endianness.
2331 if (host_is_bigendian()) {
2332 new_val = bitfield_swap(value, 0, 1);
2333 new_val |= bitfield_swap(value, 1, 1);
2334 new_val |= bitfield_swap(value, 2, 1);
2335 new_val |= bitfield_swap(value, 3, 1);
2336 new_val |= bitfield_swap(value, 4, 16);
2337 new_val |= bitfield_swap(value, 20, 4);
2338 new_val |= bitfield_swap(value, 24, 40);
2340 new_val = bitfield_swap(value, 63, 1);
2341 new_val |= bitfield_swap(value, 62, 1);
2342 new_val |= bitfield_swap(value, 61, 1);
2343 new_val |= bitfield_swap(value, 60, 1);
2344 new_val |= bitfield_swap(value, 44, 16);
2345 new_val |= bitfield_swap(value, 40, 4);
2346 new_val |= bitfield_swap(value, 0, 40);
2352 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2353 struct perf_sample *data)
2355 u64 type = evsel->core.attr.sample_type;
2356 bool swapped = evsel->needs_swap;
2358 u16 max_size = event->header.size;
2359 const void *endp = (void *)event + max_size;
2363 * used for cross-endian analysis. See git commit 65014ab3
2364 * for why this goofiness is needed.
2368 memset(data, 0, sizeof(*data));
2369 data->cpu = data->pid = data->tid = -1;
2370 data->stream_id = data->id = data->time = -1ULL;
2371 data->period = evsel->core.attr.sample_period;
2372 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2373 data->misc = event->header.misc;
2375 data->data_src = PERF_MEM_DATA_SRC_NONE;
2378 if (event->header.type != PERF_RECORD_SAMPLE) {
2379 if (!evsel->core.attr.sample_id_all)
2381 return perf_evsel__parse_id_sample(evsel, event, data);
2384 array = event->sample.array;
2386 if (perf_event__check_size(event, evsel->sample_size))
2389 if (type & PERF_SAMPLE_IDENTIFIER) {
2394 if (type & PERF_SAMPLE_IP) {
2399 if (type & PERF_SAMPLE_TID) {
2402 /* undo swap of u64, then swap on individual u32s */
2403 u.val64 = bswap_64(u.val64);
2404 u.val32[0] = bswap_32(u.val32[0]);
2405 u.val32[1] = bswap_32(u.val32[1]);
2408 data->pid = u.val32[0];
2409 data->tid = u.val32[1];
2413 if (type & PERF_SAMPLE_TIME) {
2414 data->time = *array;
2418 if (type & PERF_SAMPLE_ADDR) {
2419 data->addr = *array;
2423 if (type & PERF_SAMPLE_ID) {
2428 if (type & PERF_SAMPLE_STREAM_ID) {
2429 data->stream_id = *array;
2433 if (type & PERF_SAMPLE_CPU) {
2437 /* undo swap of u64, then swap on individual u32s */
2438 u.val64 = bswap_64(u.val64);
2439 u.val32[0] = bswap_32(u.val32[0]);
2442 data->cpu = u.val32[0];
2446 if (type & PERF_SAMPLE_PERIOD) {
2447 data->period = *array;
2451 if (type & PERF_SAMPLE_READ) {
2452 u64 read_format = evsel->core.attr.read_format;
2454 OVERFLOW_CHECK_u64(array);
2455 if (read_format & PERF_FORMAT_GROUP)
2456 data->read.group.nr = *array;
2458 data->read.one.value = *array;
2462 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2463 OVERFLOW_CHECK_u64(array);
2464 data->read.time_enabled = *array;
2468 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2469 OVERFLOW_CHECK_u64(array);
2470 data->read.time_running = *array;
2474 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2475 if (read_format & PERF_FORMAT_GROUP) {
2476 const u64 max_group_nr = UINT64_MAX /
2477 sizeof(struct sample_read_value);
2479 if (data->read.group.nr > max_group_nr)
2482 sz = data->read.group.nr * sample_read_value_size(read_format);
2483 OVERFLOW_CHECK(array, sz, max_size);
2484 data->read.group.values =
2485 (struct sample_read_value *)array;
2486 array = (void *)array + sz;
2488 OVERFLOW_CHECK_u64(array);
2489 data->read.one.id = *array;
2492 if (read_format & PERF_FORMAT_LOST) {
2493 OVERFLOW_CHECK_u64(array);
2494 data->read.one.lost = *array;
2500 if (type & PERF_SAMPLE_CALLCHAIN) {
2501 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2503 OVERFLOW_CHECK_u64(array);
2504 data->callchain = (struct ip_callchain *)array++;
2505 if (data->callchain->nr > max_callchain_nr)
2507 sz = data->callchain->nr * sizeof(u64);
2508 OVERFLOW_CHECK(array, sz, max_size);
2509 array = (void *)array + sz;
2512 if (type & PERF_SAMPLE_RAW) {
2513 OVERFLOW_CHECK_u64(array);
2517 * Undo swap of u64, then swap on individual u32s,
2518 * get the size of the raw area and undo all of the
2519 * swap. The pevent interface handles endianness by
2523 u.val64 = bswap_64(u.val64);
2524 u.val32[0] = bswap_32(u.val32[0]);
2525 u.val32[1] = bswap_32(u.val32[1]);
2527 data->raw_size = u.val32[0];
2530 * The raw data is aligned on 64bits including the
2531 * u32 size, so it's safe to use mem_bswap_64.
2534 mem_bswap_64((void *) array, data->raw_size);
2536 array = (void *)array + sizeof(u32);
2538 OVERFLOW_CHECK(array, data->raw_size, max_size);
2539 data->raw_data = (void *)array;
2540 array = (void *)array + data->raw_size;
2543 if (type & PERF_SAMPLE_BRANCH_STACK) {
2544 const u64 max_branch_nr = UINT64_MAX /
2545 sizeof(struct branch_entry);
2546 struct branch_entry *e;
2549 OVERFLOW_CHECK_u64(array);
2550 data->branch_stack = (struct branch_stack *)array++;
2552 if (data->branch_stack->nr > max_branch_nr)
2555 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2556 if (evsel__has_branch_hw_idx(evsel)) {
2558 e = &data->branch_stack->entries[0];
2560 data->no_hw_idx = true;
2562 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2563 * only nr and entries[] will be output by kernel.
2565 e = (struct branch_entry *)&data->branch_stack->hw_idx;
2570 * struct branch_flag does not have endian
2571 * specific bit field definition. And bswap
2572 * will not resolve the issue, since these
2575 * evsel__bitfield_swap_branch_flags() uses a
2576 * bitfield_swap macro to swap the bit position
2577 * based on the host endians.
2579 for (i = 0; i < data->branch_stack->nr; i++, e++)
2580 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2583 OVERFLOW_CHECK(array, sz, max_size);
2584 array = (void *)array + sz;
2587 if (type & PERF_SAMPLE_REGS_USER) {
2588 OVERFLOW_CHECK_u64(array);
2589 data->user_regs.abi = *array;
2592 if (data->user_regs.abi) {
2593 u64 mask = evsel->core.attr.sample_regs_user;
2595 sz = hweight64(mask) * sizeof(u64);
2596 OVERFLOW_CHECK(array, sz, max_size);
2597 data->user_regs.mask = mask;
2598 data->user_regs.regs = (u64 *)array;
2599 array = (void *)array + sz;
2603 if (type & PERF_SAMPLE_STACK_USER) {
2604 OVERFLOW_CHECK_u64(array);
2607 data->user_stack.offset = ((char *)(array - 1)
2611 data->user_stack.size = 0;
2613 OVERFLOW_CHECK(array, sz, max_size);
2614 data->user_stack.data = (char *)array;
2615 array = (void *)array + sz;
2616 OVERFLOW_CHECK_u64(array);
2617 data->user_stack.size = *array++;
2618 if (WARN_ONCE(data->user_stack.size > sz,
2619 "user stack dump failure\n"))
2624 if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2625 OVERFLOW_CHECK_u64(array);
2626 arch_perf_parse_sample_weight(data, array, type);
2630 if (type & PERF_SAMPLE_DATA_SRC) {
2631 OVERFLOW_CHECK_u64(array);
2632 data->data_src = *array;
2636 if (type & PERF_SAMPLE_TRANSACTION) {
2637 OVERFLOW_CHECK_u64(array);
2638 data->transaction = *array;
2642 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2643 if (type & PERF_SAMPLE_REGS_INTR) {
2644 OVERFLOW_CHECK_u64(array);
2645 data->intr_regs.abi = *array;
2648 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2649 u64 mask = evsel->core.attr.sample_regs_intr;
2651 sz = hweight64(mask) * sizeof(u64);
2652 OVERFLOW_CHECK(array, sz, max_size);
2653 data->intr_regs.mask = mask;
2654 data->intr_regs.regs = (u64 *)array;
2655 array = (void *)array + sz;
2659 data->phys_addr = 0;
2660 if (type & PERF_SAMPLE_PHYS_ADDR) {
2661 data->phys_addr = *array;
2666 if (type & PERF_SAMPLE_CGROUP) {
2667 data->cgroup = *array;
2671 data->data_page_size = 0;
2672 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2673 data->data_page_size = *array;
2677 data->code_page_size = 0;
2678 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2679 data->code_page_size = *array;
2683 if (type & PERF_SAMPLE_AUX) {
2684 OVERFLOW_CHECK_u64(array);
2687 OVERFLOW_CHECK(array, sz, max_size);
2688 /* Undo swap of data */
2690 mem_bswap_64((char *)array, sz);
2691 data->aux_sample.size = sz;
2692 data->aux_sample.data = (char *)array;
2693 array = (void *)array + sz;
2699 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2702 u64 type = evsel->core.attr.sample_type;
2705 if (!(type & PERF_SAMPLE_TIME))
2708 if (event->header.type != PERF_RECORD_SAMPLE) {
2709 struct perf_sample data = {
2713 if (!evsel->core.attr.sample_id_all)
2715 if (perf_evsel__parse_id_sample(evsel, event, &data))
2718 *timestamp = data.time;
2722 array = event->sample.array;
2724 if (perf_event__check_size(event, evsel->sample_size))
2727 if (type & PERF_SAMPLE_IDENTIFIER)
2730 if (type & PERF_SAMPLE_IP)
2733 if (type & PERF_SAMPLE_TID)
2736 if (type & PERF_SAMPLE_TIME)
2737 *timestamp = *array;
2742 u16 evsel__id_hdr_size(struct evsel *evsel)
2744 u64 sample_type = evsel->core.attr.sample_type;
2747 if (sample_type & PERF_SAMPLE_TID)
2748 size += sizeof(u64);
2750 if (sample_type & PERF_SAMPLE_TIME)
2751 size += sizeof(u64);
2753 if (sample_type & PERF_SAMPLE_ID)
2754 size += sizeof(u64);
2756 if (sample_type & PERF_SAMPLE_STREAM_ID)
2757 size += sizeof(u64);
2759 if (sample_type & PERF_SAMPLE_CPU)
2760 size += sizeof(u64);
2762 if (sample_type & PERF_SAMPLE_IDENTIFIER)
2763 size += sizeof(u64);
2768 #ifdef HAVE_LIBTRACEEVENT
2769 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2771 return tep_find_field(evsel->tp_format, name);
2774 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2776 struct tep_format_field *field = evsel__field(evsel, name);
2782 offset = field->offset;
2784 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2785 offset = *(int *)(sample->raw_data + field->offset);
2787 #ifdef HAVE_LIBTRACEEVENT_TEP_FIELD_IS_RELATIVE
2788 if (field->flags & TEP_FIELD_IS_RELATIVE)
2789 offset += field->offset + field->size;
2793 return sample->raw_data + offset;
2796 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2800 void *ptr = sample->raw_data + field->offset;
2802 switch (field->size) {
2806 value = *(u16 *)ptr;
2809 value = *(u32 *)ptr;
2812 memcpy(&value, ptr, sizeof(u64));
2821 switch (field->size) {
2823 return bswap_16(value);
2825 return bswap_32(value);
2827 return bswap_64(value);
2835 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2837 struct tep_format_field *field = evsel__field(evsel, name);
2842 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2846 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2850 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2851 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2852 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2854 * If it's cycles then fall back to hrtimer based
2855 * cpu-clock-tick sw counter, which is always available even if
2858 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2861 scnprintf(msg, msgsize, "%s",
2862 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2864 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2865 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2867 zfree(&evsel->name);
2869 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2870 (paranoid = perf_event_paranoid()) > 1) {
2871 const char *name = evsel__name(evsel);
2873 const char *sep = ":";
2875 /* If event has exclude user then don't exclude kernel. */
2876 if (evsel->core.attr.exclude_user)
2879 /* Is there already the separator in the name. */
2880 if (strchr(name, '/') ||
2881 (strchr(name, ':') && !evsel->is_libpfm_event))
2884 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2889 evsel->name = new_name;
2890 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2891 "to fall back to excluding kernel and hypervisor "
2892 " samples", paranoid);
2893 evsel->core.attr.exclude_kernel = 1;
2894 evsel->core.attr.exclude_hv = 1;
2902 static bool find_process(const char *name)
2904 size_t len = strlen(name);
2909 dir = opendir(procfs__mountpoint());
2913 /* Walk through the directory. */
2914 while (ret && (d = readdir(dir)) != NULL) {
2915 char path[PATH_MAX];
2919 if ((d->d_type != DT_DIR) ||
2920 !strcmp(".", d->d_name) ||
2921 !strcmp("..", d->d_name))
2924 scnprintf(path, sizeof(path), "%s/%s/comm",
2925 procfs__mountpoint(), d->d_name);
2927 if (filename__read_str(path, &data, &size))
2930 ret = strncmp(name, data, len);
2935 return ret ? false : true;
2938 static bool is_amd(const char *arch, const char *cpuid)
2940 return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2943 static bool is_amd_ibs(struct evsel *evsel)
2945 return evsel->core.attr.precise_ip
2946 || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2949 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2950 int err, char *msg, size_t size)
2952 struct perf_env *env = evsel__env(evsel);
2953 const char *arch = perf_env__arch(env);
2954 const char *cpuid = perf_env__cpuid(env);
2955 char sbuf[STRERR_BUFSIZE];
2956 int printed = 0, enforced = 0;
2961 printed += scnprintf(msg + printed, size - printed,
2962 "Access to performance monitoring and observability operations is limited.\n");
2964 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2966 printed += scnprintf(msg + printed, size - printed,
2967 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2968 "monitoring and observability operations. Inspect system audit records for\n"
2969 "more perf_event access control information and adjusting the policy.\n");
2974 printed += scnprintf(msg, size,
2975 "No permission to enable %s event.\n\n", evsel__name(evsel));
2977 return scnprintf(msg + printed, size - printed,
2978 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2979 "access to performance monitoring and observability operations for processes\n"
2980 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2981 "More information can be found at 'Perf events and tool security' document:\n"
2982 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2983 "perf_event_paranoid setting is %d:\n"
2984 " -1: Allow use of (almost) all events by all users\n"
2985 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2986 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2987 ">= 1: Disallow CPU event access\n"
2988 ">= 2: Disallow kernel profiling\n"
2989 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2990 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2991 perf_event_paranoid());
2993 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2995 return scnprintf(msg, size, "%s",
2996 "Too many events are opened.\n"
2997 "Probably the maximum number of open file descriptors has been reached.\n"
2998 "Hint: Try again after reducing the number of events.\n"
2999 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3001 if (evsel__has_callchain(evsel) &&
3002 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3003 return scnprintf(msg, size,
3004 "Not enough memory to setup event with callchain.\n"
3005 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3006 "Hint: Current value: %d", sysctl__max_stack());
3009 if (target->cpu_list)
3010 return scnprintf(msg, size, "%s",
3011 "No such device - did you specify an out-of-range profile CPU?");
3014 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3015 return scnprintf(msg, size,
3016 "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3017 evsel__name(evsel));
3018 if (evsel->core.attr.aux_output)
3019 return scnprintf(msg, size,
3020 "%s: PMU Hardware doesn't support 'aux_output' feature",
3021 evsel__name(evsel));
3022 if (evsel->core.attr.sample_period != 0)
3023 return scnprintf(msg, size,
3024 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3025 evsel__name(evsel));
3026 if (evsel->core.attr.precise_ip)
3027 return scnprintf(msg, size, "%s",
3028 "\'precise\' request may not be supported. Try removing 'p' modifier.");
3029 #if defined(__i386__) || defined(__x86_64__)
3030 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3031 return scnprintf(msg, size, "%s",
3032 "No hardware sampling interrupt available.\n");
3036 if (find_process("oprofiled"))
3037 return scnprintf(msg, size,
3038 "The PMU counters are busy/taken by another profiler.\n"
3039 "We found oprofile daemon running, please stop it and try again.");
3042 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3043 return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3044 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3045 return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3046 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3047 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3048 if (perf_missing_features.clockid)
3049 return scnprintf(msg, size, "clockid feature not supported.");
3050 if (perf_missing_features.clockid_wrong)
3051 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3052 if (perf_missing_features.aux_output)
3053 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3054 if (!target__has_cpu(target))
3055 return scnprintf(msg, size,
3056 "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3057 evsel__name(evsel));
3058 if (is_amd(arch, cpuid)) {
3059 if (is_amd_ibs(evsel)) {
3060 if (evsel->core.attr.exclude_kernel)
3061 return scnprintf(msg, size,
3062 "AMD IBS can't exclude kernel events. Try running at a higher privilege level.");
3063 if (!evsel->core.system_wide)
3064 return scnprintf(msg, size,
3065 "AMD IBS may only be available in system-wide/per-cpu mode. Try using -a, or -C and workload affinity");
3071 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3072 "Please add an auxiliary event in front of the load latency event.");
3077 return scnprintf(msg, size,
3078 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3079 "/bin/dmesg | grep -i perf may provide additional information.\n",
3080 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3083 struct perf_env *evsel__env(struct evsel *evsel)
3085 if (evsel && evsel->evlist && evsel->evlist->env)
3086 return evsel->evlist->env;
3090 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3092 int cpu_map_idx, thread;
3094 for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3095 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3097 int fd = FD(evsel, cpu_map_idx, thread);
3099 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3100 cpu_map_idx, thread, fd) < 0)
3108 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3110 struct perf_cpu_map *cpus = evsel->core.cpus;
3111 struct perf_thread_map *threads = evsel->core.threads;
3113 if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3116 return store_evsel_ids(evsel, evlist);
3119 void evsel__zero_per_pkg(struct evsel *evsel)
3121 struct hashmap_entry *cur;
3124 if (evsel->per_pkg_mask) {
3125 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3126 free((void *)cur->pkey);
3128 hashmap__clear(evsel->per_pkg_mask);
3132 bool evsel__is_hybrid(const struct evsel *evsel)
3134 return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3137 struct evsel *evsel__leader(struct evsel *evsel)
3139 return container_of(evsel->core.leader, struct evsel, core);
3142 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3144 return evsel->core.leader == &leader->core;
3147 bool evsel__is_leader(struct evsel *evsel)
3149 return evsel__has_leader(evsel, evsel);
3152 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3154 evsel->core.leader = &leader->core;
3157 int evsel__source_count(const struct evsel *evsel)
3162 evlist__for_each_entry(evsel->evlist, pos) {
3163 if (pos->metric_leader == evsel)
3169 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3175 * Remove an event from a given group (leader).
3176 * Some events, e.g., perf metrics Topdown events,
3177 * must always be grouped. Ignore the events.
3179 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3181 if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3182 evsel__set_leader(evsel, evsel);
3183 evsel->core.nr_members = 0;
3184 leader->core.nr_members--;