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 "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 "../perf-sys.h"
49 #include "util/parse-branch-options.h"
50 #include <internal/xyarray.h>
51 #include <internal/lib.h>
53 #include <linux/ctype.h>
55 struct perf_missing_features perf_missing_features;
57 static clockid_t clockid;
59 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
64 void __weak test_attr__ready(void) { }
66 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
72 int (*init)(struct evsel *evsel);
73 void (*fini)(struct evsel *evsel);
74 } perf_evsel__object = {
75 .size = sizeof(struct evsel),
76 .init = evsel__no_extra_init,
77 .fini = evsel__no_extra_fini,
80 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
81 void (*fini)(struct evsel *evsel))
87 if (perf_evsel__object.size > object_size)
90 perf_evsel__object.size = object_size;
94 perf_evsel__object.init = init;
97 perf_evsel__object.fini = fini;
102 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
104 int __evsel__sample_size(u64 sample_type)
106 u64 mask = sample_type & PERF_SAMPLE_MASK;
110 for (i = 0; i < 64; i++) {
111 if (mask & (1ULL << i))
121 * __perf_evsel__calc_id_pos - calculate id_pos.
122 * @sample_type: sample type
124 * This function returns the position of the event id (PERF_SAMPLE_ID or
125 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
126 * perf_record_sample.
128 static int __perf_evsel__calc_id_pos(u64 sample_type)
132 if (sample_type & PERF_SAMPLE_IDENTIFIER)
135 if (!(sample_type & PERF_SAMPLE_ID))
138 if (sample_type & PERF_SAMPLE_IP)
141 if (sample_type & PERF_SAMPLE_TID)
144 if (sample_type & PERF_SAMPLE_TIME)
147 if (sample_type & PERF_SAMPLE_ADDR)
154 * __perf_evsel__calc_is_pos - calculate is_pos.
155 * @sample_type: sample type
157 * This function returns the position (counting backwards) of the event id
158 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
159 * sample_id_all is used there is an id sample appended to non-sample events.
161 static int __perf_evsel__calc_is_pos(u64 sample_type)
165 if (sample_type & PERF_SAMPLE_IDENTIFIER)
168 if (!(sample_type & PERF_SAMPLE_ID))
171 if (sample_type & PERF_SAMPLE_CPU)
174 if (sample_type & PERF_SAMPLE_STREAM_ID)
180 void evsel__calc_id_pos(struct evsel *evsel)
182 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
183 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
186 void __evsel__set_sample_bit(struct evsel *evsel,
187 enum perf_event_sample_format bit)
189 if (!(evsel->core.attr.sample_type & bit)) {
190 evsel->core.attr.sample_type |= bit;
191 evsel->sample_size += sizeof(u64);
192 evsel__calc_id_pos(evsel);
196 void __evsel__reset_sample_bit(struct evsel *evsel,
197 enum perf_event_sample_format bit)
199 if (evsel->core.attr.sample_type & bit) {
200 evsel->core.attr.sample_type &= ~bit;
201 evsel->sample_size -= sizeof(u64);
202 evsel__calc_id_pos(evsel);
206 void evsel__set_sample_id(struct evsel *evsel,
207 bool can_sample_identifier)
209 if (can_sample_identifier) {
210 evsel__reset_sample_bit(evsel, ID);
211 evsel__set_sample_bit(evsel, IDENTIFIER);
213 evsel__set_sample_bit(evsel, ID);
215 evsel->core.attr.read_format |= PERF_FORMAT_ID;
219 * evsel__is_function_event - Return whether given evsel is a function
222 * @evsel - evsel selector to be tested
224 * Return %true if event is function trace event
226 bool evsel__is_function_event(struct evsel *evsel)
228 #define FUNCTION_EVENT "ftrace:function"
230 return evsel->name &&
231 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
233 #undef FUNCTION_EVENT
236 void evsel__init(struct evsel *evsel,
237 struct perf_event_attr *attr, int idx)
239 perf_evsel__init(&evsel->core, attr);
241 evsel->tracking = !idx;
242 evsel->leader = evsel;
245 evsel->max_events = ULONG_MAX;
246 evsel->evlist = NULL;
247 evsel->bpf_obj = NULL;
249 INIT_LIST_HEAD(&evsel->config_terms);
250 perf_evsel__object.init(evsel);
251 evsel->sample_size = __evsel__sample_size(attr->sample_type);
252 evsel__calc_id_pos(evsel);
253 evsel->cmdline_group_boundary = false;
254 evsel->metric_expr = NULL;
255 evsel->metric_name = NULL;
256 evsel->metric_events = NULL;
257 evsel->per_pkg_mask = NULL;
258 evsel->collect_stat = false;
259 evsel->pmu_name = NULL;
262 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
264 struct evsel *evsel = zalloc(perf_evsel__object.size);
268 evsel__init(evsel, attr, idx);
270 if (evsel__is_bpf_output(evsel)) {
271 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
272 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
273 evsel->core.attr.sample_period = 1;
276 if (evsel__is_clock(evsel)) {
278 * The evsel->unit points to static alias->unit
279 * so it's ok to use static string in here.
281 static const char *unit = "msec";
290 static bool perf_event_can_profile_kernel(void)
292 return perf_event_paranoid_check(1);
295 struct evsel *evsel__new_cycles(bool precise)
297 struct perf_event_attr attr = {
298 .type = PERF_TYPE_HARDWARE,
299 .config = PERF_COUNT_HW_CPU_CYCLES,
300 .exclude_kernel = !perf_event_can_profile_kernel(),
304 event_attr_init(&attr);
310 * Now let the usual logic to set up the perf_event_attr defaults
311 * to kick in when we return and before perf_evsel__open() is called.
314 evsel = evsel__new(&attr);
318 evsel->precise_max = true;
320 /* use asprintf() because free(evsel) assumes name is allocated */
321 if (asprintf(&evsel->name, "cycles%s%s%.*s",
322 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
323 attr.exclude_kernel ? "u" : "",
324 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
329 evsel__delete(evsel);
334 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
336 struct evsel_config_term *pos, *tmp;
338 list_for_each_entry(pos, &src->config_terms, list) {
339 tmp = malloc(sizeof(*tmp));
345 tmp->val.str = strdup(pos->val.str);
346 if (tmp->val.str == NULL) {
351 list_add_tail(&tmp->list, &dst->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 evsel->cgrp = cgroup__get(orig->cgrp);
407 evsel->tp_format = orig->tp_format;
408 evsel->handler = orig->handler;
409 evsel->leader = orig->leader;
411 evsel->max_events = orig->max_events;
412 evsel->tool_event = orig->tool_event;
413 evsel->unit = orig->unit;
414 evsel->scale = orig->scale;
415 evsel->snapshot = orig->snapshot;
416 evsel->per_pkg = orig->per_pkg;
417 evsel->percore = orig->percore;
418 evsel->precise_max = orig->precise_max;
419 evsel->use_uncore_alias = orig->use_uncore_alias;
420 evsel->is_libpfm_event = orig->is_libpfm_event;
422 evsel->exclude_GH = orig->exclude_GH;
423 evsel->sample_read = orig->sample_read;
424 evsel->auto_merge_stats = orig->auto_merge_stats;
425 evsel->collect_stat = orig->collect_stat;
426 evsel->weak_group = orig->weak_group;
428 if (evsel__copy_config_terms(evsel, orig) < 0)
434 evsel__delete(evsel);
439 * Returns pointer with encoded error via <linux/err.h> interface.
441 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
443 struct evsel *evsel = zalloc(perf_evsel__object.size);
449 struct perf_event_attr attr = {
450 .type = PERF_TYPE_TRACEPOINT,
451 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
452 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
455 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
458 evsel->tp_format = trace_event__tp_format(sys, name);
459 if (IS_ERR(evsel->tp_format)) {
460 err = PTR_ERR(evsel->tp_format);
464 event_attr_init(&attr);
465 attr.config = evsel->tp_format->id;
466 attr.sample_period = 1;
467 evsel__init(evsel, &attr, idx);
479 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
487 "stalled-cycles-frontend",
488 "stalled-cycles-backend",
492 static const char *__evsel__hw_name(u64 config)
494 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
495 return evsel__hw_names[config];
497 return "unknown-hardware";
500 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
502 int colon = 0, r = 0;
503 struct perf_event_attr *attr = &evsel->core.attr;
504 bool exclude_guest_default = false;
506 #define MOD_PRINT(context, mod) do { \
507 if (!attr->exclude_##context) { \
508 if (!colon) colon = ++r; \
509 r += scnprintf(bf + r, size - r, "%c", mod); \
512 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
513 MOD_PRINT(kernel, 'k');
514 MOD_PRINT(user, 'u');
516 exclude_guest_default = true;
519 if (attr->precise_ip) {
522 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
523 exclude_guest_default = true;
526 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
527 MOD_PRINT(host, 'H');
528 MOD_PRINT(guest, 'G');
536 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
538 int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
539 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
542 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
555 static const char *__evsel__sw_name(u64 config)
557 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
558 return evsel__sw_names[config];
559 return "unknown-software";
562 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
564 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
565 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
568 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
572 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
574 if (type & HW_BREAKPOINT_R)
575 r += scnprintf(bf + r, size - r, "r");
577 if (type & HW_BREAKPOINT_W)
578 r += scnprintf(bf + r, size - r, "w");
580 if (type & HW_BREAKPOINT_X)
581 r += scnprintf(bf + r, size - r, "x");
586 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
588 struct perf_event_attr *attr = &evsel->core.attr;
589 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
590 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
593 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
594 { "L1-dcache", "l1-d", "l1d", "L1-data", },
595 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
597 { "dTLB", "d-tlb", "Data-TLB", },
598 { "iTLB", "i-tlb", "Instruction-TLB", },
599 { "branch", "branches", "bpu", "btb", "bpc", },
603 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
604 { "load", "loads", "read", },
605 { "store", "stores", "write", },
606 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
609 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
610 { "refs", "Reference", "ops", "access", },
611 { "misses", "miss", },
614 #define C(x) PERF_COUNT_HW_CACHE_##x
615 #define CACHE_READ (1 << C(OP_READ))
616 #define CACHE_WRITE (1 << C(OP_WRITE))
617 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
618 #define COP(x) (1 << x)
621 * cache operartion stat
622 * L1I : Read and prefetch only
623 * ITLB and BPU : Read-only
625 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
626 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
627 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
628 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
629 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
630 [C(ITLB)] = (CACHE_READ),
631 [C(BPU)] = (CACHE_READ),
632 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
635 bool evsel__is_cache_op_valid(u8 type, u8 op)
637 if (evsel__hw_cache_stat[type] & COP(op))
638 return true; /* valid */
640 return false; /* invalid */
643 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
646 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
647 evsel__hw_cache_op[op][0],
648 evsel__hw_cache_result[result][0]);
651 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
652 evsel__hw_cache_op[op][1]);
655 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
657 u8 op, result, type = (config >> 0) & 0xff;
658 const char *err = "unknown-ext-hardware-cache-type";
660 if (type >= PERF_COUNT_HW_CACHE_MAX)
663 op = (config >> 8) & 0xff;
664 err = "unknown-ext-hardware-cache-op";
665 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
668 result = (config >> 16) & 0xff;
669 err = "unknown-ext-hardware-cache-result";
670 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
673 err = "invalid-cache";
674 if (!evsel__is_cache_op_valid(type, op))
677 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
679 return scnprintf(bf, size, "%s", err);
682 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
684 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
685 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
688 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
690 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
691 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
694 static int evsel__tool_name(char *bf, size_t size)
696 int ret = scnprintf(bf, size, "duration_time");
700 const char *evsel__name(struct evsel *evsel)
710 switch (evsel->core.attr.type) {
712 evsel__raw_name(evsel, bf, sizeof(bf));
715 case PERF_TYPE_HARDWARE:
716 evsel__hw_name(evsel, bf, sizeof(bf));
719 case PERF_TYPE_HW_CACHE:
720 evsel__hw_cache_name(evsel, bf, sizeof(bf));
723 case PERF_TYPE_SOFTWARE:
724 if (evsel->tool_event)
725 evsel__tool_name(bf, sizeof(bf));
727 evsel__sw_name(evsel, bf, sizeof(bf));
730 case PERF_TYPE_TRACEPOINT:
731 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
734 case PERF_TYPE_BREAKPOINT:
735 evsel__bp_name(evsel, bf, sizeof(bf));
739 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
740 evsel->core.attr.type);
744 evsel->name = strdup(bf);
752 const char *evsel__group_name(struct evsel *evsel)
754 return evsel->group_name ?: "anon group";
758 * Returns the group details for the specified leader,
759 * with following rules.
761 * For record -e '{cycles,instructions}'
762 * 'anon group { cycles:u, instructions:u }'
764 * For record -e 'cycles,instructions' and report --group
765 * 'cycles:u, instructions:u'
767 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
771 const char *group_name = evsel__group_name(evsel);
773 if (!evsel->forced_leader)
774 ret = scnprintf(buf, size, "%s { ", group_name);
776 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
778 for_each_group_member(pos, evsel)
779 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
781 if (!evsel->forced_leader)
782 ret += scnprintf(buf + ret, size - ret, " }");
787 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
788 struct callchain_param *param)
790 bool function = evsel__is_function_event(evsel);
791 struct perf_event_attr *attr = &evsel->core.attr;
793 evsel__set_sample_bit(evsel, CALLCHAIN);
795 attr->sample_max_stack = param->max_stack;
797 if (opts->kernel_callchains)
798 attr->exclude_callchain_user = 1;
799 if (opts->user_callchains)
800 attr->exclude_callchain_kernel = 1;
801 if (param->record_mode == CALLCHAIN_LBR) {
802 if (!opts->branch_stack) {
803 if (attr->exclude_user) {
804 pr_warning("LBR callstack option is only available "
805 "to get user callchain information. "
806 "Falling back to framepointers.\n");
808 evsel__set_sample_bit(evsel, BRANCH_STACK);
809 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
810 PERF_SAMPLE_BRANCH_CALL_STACK |
811 PERF_SAMPLE_BRANCH_NO_CYCLES |
812 PERF_SAMPLE_BRANCH_NO_FLAGS |
813 PERF_SAMPLE_BRANCH_HW_INDEX;
816 pr_warning("Cannot use LBR callstack with branch stack. "
817 "Falling back to framepointers.\n");
820 if (param->record_mode == CALLCHAIN_DWARF) {
822 evsel__set_sample_bit(evsel, REGS_USER);
823 evsel__set_sample_bit(evsel, STACK_USER);
824 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
825 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
826 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
827 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
828 "so the minimal registers set (IP, SP) is explicitly forced.\n");
830 attr->sample_regs_user |= PERF_REGS_MASK;
832 attr->sample_stack_user = param->dump_size;
833 attr->exclude_callchain_user = 1;
835 pr_info("Cannot use DWARF unwind for function trace event,"
836 " falling back to framepointers.\n");
841 pr_info("Disabling user space callchains for function trace event.\n");
842 attr->exclude_callchain_user = 1;
846 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
847 struct callchain_param *param)
850 return __evsel__config_callchain(evsel, opts, param);
854 perf_evsel__reset_callgraph(struct evsel *evsel,
855 struct callchain_param *param)
857 struct perf_event_attr *attr = &evsel->core.attr;
859 evsel__reset_sample_bit(evsel, CALLCHAIN);
860 if (param->record_mode == CALLCHAIN_LBR) {
861 evsel__reset_sample_bit(evsel, BRANCH_STACK);
862 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
863 PERF_SAMPLE_BRANCH_CALL_STACK |
864 PERF_SAMPLE_BRANCH_HW_INDEX);
866 if (param->record_mode == CALLCHAIN_DWARF) {
867 evsel__reset_sample_bit(evsel, REGS_USER);
868 evsel__reset_sample_bit(evsel, STACK_USER);
872 static void evsel__apply_config_terms(struct evsel *evsel,
873 struct record_opts *opts, bool track)
875 struct evsel_config_term *term;
876 struct list_head *config_terms = &evsel->config_terms;
877 struct perf_event_attr *attr = &evsel->core.attr;
878 /* callgraph default */
879 struct callchain_param param = {
880 .record_mode = callchain_param.record_mode,
884 const char *callgraph_buf = NULL;
886 list_for_each_entry(term, config_terms, list) {
887 switch (term->type) {
888 case EVSEL__CONFIG_TERM_PERIOD:
889 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
890 attr->sample_period = term->val.period;
892 evsel__reset_sample_bit(evsel, PERIOD);
895 case EVSEL__CONFIG_TERM_FREQ:
896 if (!(term->weak && opts->user_freq != UINT_MAX)) {
897 attr->sample_freq = term->val.freq;
899 evsel__set_sample_bit(evsel, PERIOD);
902 case EVSEL__CONFIG_TERM_TIME:
904 evsel__set_sample_bit(evsel, TIME);
906 evsel__reset_sample_bit(evsel, TIME);
908 case EVSEL__CONFIG_TERM_CALLGRAPH:
909 callgraph_buf = term->val.str;
911 case EVSEL__CONFIG_TERM_BRANCH:
912 if (term->val.str && strcmp(term->val.str, "no")) {
913 evsel__set_sample_bit(evsel, BRANCH_STACK);
914 parse_branch_str(term->val.str,
915 &attr->branch_sample_type);
917 evsel__reset_sample_bit(evsel, BRANCH_STACK);
919 case EVSEL__CONFIG_TERM_STACK_USER:
920 dump_size = term->val.stack_user;
922 case EVSEL__CONFIG_TERM_MAX_STACK:
923 max_stack = term->val.max_stack;
925 case EVSEL__CONFIG_TERM_MAX_EVENTS:
926 evsel->max_events = term->val.max_events;
928 case EVSEL__CONFIG_TERM_INHERIT:
930 * attr->inherit should has already been set by
931 * evsel__config. If user explicitly set
932 * inherit using config terms, override global
933 * opt->no_inherit setting.
935 attr->inherit = term->val.inherit ? 1 : 0;
937 case EVSEL__CONFIG_TERM_OVERWRITE:
938 attr->write_backward = term->val.overwrite ? 1 : 0;
940 case EVSEL__CONFIG_TERM_DRV_CFG:
942 case EVSEL__CONFIG_TERM_PERCORE:
944 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
945 attr->aux_output = term->val.aux_output ? 1 : 0;
947 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
948 /* Already applied by auxtrace */
950 case EVSEL__CONFIG_TERM_CFG_CHG:
957 /* User explicitly set per-event callgraph, clear the old setting and reset. */
958 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
959 bool sample_address = false;
962 param.max_stack = max_stack;
963 if (callgraph_buf == NULL)
964 callgraph_buf = "fp";
967 /* parse callgraph parameters */
968 if (callgraph_buf != NULL) {
969 if (!strcmp(callgraph_buf, "no")) {
970 param.enabled = false;
971 param.record_mode = CALLCHAIN_NONE;
973 param.enabled = true;
974 if (parse_callchain_record(callgraph_buf, ¶m)) {
975 pr_err("per-event callgraph setting for %s failed. "
976 "Apply callgraph global setting for it\n",
980 if (param.record_mode == CALLCHAIN_DWARF)
981 sample_address = true;
985 dump_size = round_up(dump_size, sizeof(u64));
986 param.dump_size = dump_size;
989 /* If global callgraph set, clear it */
990 if (callchain_param.enabled)
991 perf_evsel__reset_callgraph(evsel, &callchain_param);
993 /* set perf-event callgraph */
995 if (sample_address) {
996 evsel__set_sample_bit(evsel, ADDR);
997 evsel__set_sample_bit(evsel, DATA_SRC);
998 evsel->core.attr.mmap_data = track;
1000 evsel__config_callchain(evsel, opts, ¶m);
1005 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1007 struct evsel_config_term *term, *found_term = NULL;
1009 list_for_each_entry(term, &evsel->config_terms, list) {
1010 if (term->type == type)
1017 static void evsel__set_default_freq_period(struct record_opts *opts,
1018 struct perf_event_attr *attr)
1022 attr->sample_freq = opts->freq;
1024 attr->sample_period = opts->default_interval;
1029 * The enable_on_exec/disabled value strategy:
1031 * 1) For any type of traced program:
1032 * - all independent events and group leaders are disabled
1033 * - all group members are enabled
1035 * Group members are ruled by group leaders. They need to
1036 * be enabled, because the group scheduling relies on that.
1038 * 2) For traced programs executed by perf:
1039 * - all independent events and group leaders have
1040 * enable_on_exec set
1041 * - we don't specifically enable or disable any event during
1042 * the record command
1044 * Independent events and group leaders are initially disabled
1045 * and get enabled by exec. Group members are ruled by group
1046 * leaders as stated in 1).
1048 * 3) For traced programs attached by perf (pid/tid):
1049 * - we specifically enable or disable all events during
1050 * the record command
1052 * When attaching events to already running traced we
1053 * enable/disable events specifically, as there's no
1054 * initial traced exec call.
1056 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1057 struct callchain_param *callchain)
1059 struct evsel *leader = evsel->leader;
1060 struct perf_event_attr *attr = &evsel->core.attr;
1061 int track = evsel->tracking;
1062 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1064 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1065 attr->inherit = !opts->no_inherit;
1066 attr->write_backward = opts->overwrite ? 1 : 0;
1068 evsel__set_sample_bit(evsel, IP);
1069 evsel__set_sample_bit(evsel, TID);
1071 if (evsel->sample_read) {
1072 evsel__set_sample_bit(evsel, READ);
1075 * We need ID even in case of single event, because
1076 * PERF_SAMPLE_READ process ID specific data.
1078 evsel__set_sample_id(evsel, false);
1081 * Apply group format only if we belong to group
1082 * with more than one members.
1084 if (leader->core.nr_members > 1) {
1085 attr->read_format |= PERF_FORMAT_GROUP;
1091 * We default some events to have a default interval. But keep
1092 * it a weak assumption overridable by the user.
1094 if ((evsel->is_libpfm_event && !attr->sample_period) ||
1095 (!evsel->is_libpfm_event && (!attr->sample_period ||
1096 opts->user_freq != UINT_MAX ||
1097 opts->user_interval != ULLONG_MAX)))
1098 evsel__set_default_freq_period(opts, attr);
1101 * If attr->freq was set (here or earlier), ask for period
1105 evsel__set_sample_bit(evsel, PERIOD);
1107 if (opts->no_samples)
1108 attr->sample_freq = 0;
1110 if (opts->inherit_stat) {
1111 evsel->core.attr.read_format |=
1112 PERF_FORMAT_TOTAL_TIME_ENABLED |
1113 PERF_FORMAT_TOTAL_TIME_RUNNING |
1115 attr->inherit_stat = 1;
1118 if (opts->sample_address) {
1119 evsel__set_sample_bit(evsel, ADDR);
1120 attr->mmap_data = track;
1124 * We don't allow user space callchains for function trace
1125 * event, due to issues with page faults while tracing page
1126 * fault handler and its overall trickiness nature.
1128 if (evsel__is_function_event(evsel))
1129 evsel->core.attr.exclude_callchain_user = 1;
1131 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1132 evsel__config_callchain(evsel, opts, callchain);
1134 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1135 !evsel__is_dummy_event(evsel)) {
1136 attr->sample_regs_intr = opts->sample_intr_regs;
1137 evsel__set_sample_bit(evsel, REGS_INTR);
1140 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1141 !evsel__is_dummy_event(evsel)) {
1142 attr->sample_regs_user |= opts->sample_user_regs;
1143 evsel__set_sample_bit(evsel, REGS_USER);
1146 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1147 evsel__set_sample_bit(evsel, CPU);
1150 * When the user explicitly disabled time don't force it here.
1152 if (opts->sample_time &&
1153 (!perf_missing_features.sample_id_all &&
1154 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1155 opts->sample_time_set)))
1156 evsel__set_sample_bit(evsel, TIME);
1158 if (opts->raw_samples && !evsel->no_aux_samples) {
1159 evsel__set_sample_bit(evsel, TIME);
1160 evsel__set_sample_bit(evsel, RAW);
1161 evsel__set_sample_bit(evsel, CPU);
1164 if (opts->sample_address)
1165 evsel__set_sample_bit(evsel, DATA_SRC);
1167 if (opts->sample_phys_addr)
1168 evsel__set_sample_bit(evsel, PHYS_ADDR);
1170 if (opts->no_buffering) {
1171 attr->watermark = 0;
1172 attr->wakeup_events = 1;
1174 if (opts->branch_stack && !evsel->no_aux_samples) {
1175 evsel__set_sample_bit(evsel, BRANCH_STACK);
1176 attr->branch_sample_type = opts->branch_stack;
1179 if (opts->sample_weight)
1180 evsel__set_sample_bit(evsel, WEIGHT);
1184 attr->mmap2 = track && !perf_missing_features.mmap2;
1187 * ksymbol is tracked separately with text poke because it needs to be
1188 * system wide and enabled immediately.
1190 if (!opts->text_poke)
1191 attr->ksymbol = track && !perf_missing_features.ksymbol;
1192 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1194 if (opts->record_namespaces)
1195 attr->namespaces = track;
1197 if (opts->record_cgroup) {
1198 attr->cgroup = track && !perf_missing_features.cgroup;
1199 evsel__set_sample_bit(evsel, CGROUP);
1202 if (opts->record_switch_events)
1203 attr->context_switch = track;
1205 if (opts->sample_transaction)
1206 evsel__set_sample_bit(evsel, TRANSACTION);
1208 if (opts->running_time) {
1209 evsel->core.attr.read_format |=
1210 PERF_FORMAT_TOTAL_TIME_ENABLED |
1211 PERF_FORMAT_TOTAL_TIME_RUNNING;
1215 * XXX see the function comment above
1217 * Disabling only independent events or group leaders,
1218 * keeping group members enabled.
1220 if (evsel__is_group_leader(evsel))
1224 * Setting enable_on_exec for independent events and
1225 * group leaders for traced executed by perf.
1227 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1228 !opts->initial_delay)
1229 attr->enable_on_exec = 1;
1231 if (evsel->immediate) {
1233 attr->enable_on_exec = 0;
1236 clockid = opts->clockid;
1237 if (opts->use_clockid) {
1238 attr->use_clockid = 1;
1239 attr->clockid = opts->clockid;
1242 if (evsel->precise_max)
1243 attr->precise_ip = 3;
1245 if (opts->all_user) {
1246 attr->exclude_kernel = 1;
1247 attr->exclude_user = 0;
1250 if (opts->all_kernel) {
1251 attr->exclude_kernel = 0;
1252 attr->exclude_user = 1;
1255 if (evsel->core.own_cpus || evsel->unit)
1256 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1259 * Apply event specific term settings,
1260 * it overloads any global configuration.
1262 evsel__apply_config_terms(evsel, opts, track);
1264 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1266 /* The --period option takes the precedence. */
1267 if (opts->period_set) {
1269 evsel__set_sample_bit(evsel, PERIOD);
1271 evsel__reset_sample_bit(evsel, PERIOD);
1275 * A dummy event never triggers any actual counter and therefore
1276 * cannot be used with branch_stack.
1278 * For initial_delay, a dummy event is added implicitly.
1279 * The software event will trigger -EOPNOTSUPP error out,
1280 * if BRANCH_STACK bit is set.
1282 if (evsel__is_dummy_event(evsel))
1283 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1286 int evsel__set_filter(struct evsel *evsel, const char *filter)
1288 char *new_filter = strdup(filter);
1290 if (new_filter != NULL) {
1291 free(evsel->filter);
1292 evsel->filter = new_filter;
1299 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1303 if (evsel->filter == NULL)
1304 return evsel__set_filter(evsel, filter);
1306 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1307 free(evsel->filter);
1308 evsel->filter = new_filter;
1315 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1317 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1320 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1322 return evsel__append_filter(evsel, "%s,%s", filter);
1325 /* Caller has to clear disabled after going through all CPUs. */
1326 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1328 return perf_evsel__enable_cpu(&evsel->core, cpu);
1331 int evsel__enable(struct evsel *evsel)
1333 int err = perf_evsel__enable(&evsel->core);
1336 evsel->disabled = false;
1340 /* Caller has to set disabled after going through all CPUs. */
1341 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1343 return perf_evsel__disable_cpu(&evsel->core, cpu);
1346 int evsel__disable(struct evsel *evsel)
1348 int err = perf_evsel__disable(&evsel->core);
1350 * We mark it disabled here so that tools that disable a event can
1351 * ignore events after they disable it. I.e. the ring buffer may have
1352 * already a few more events queued up before the kernel got the stop
1356 evsel->disabled = true;
1361 static void evsel__free_config_terms(struct evsel *evsel)
1363 struct evsel_config_term *term, *h;
1365 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1366 list_del_init(&term->list);
1368 zfree(&term->val.str);
1373 void evsel__exit(struct evsel *evsel)
1375 assert(list_empty(&evsel->core.node));
1376 assert(evsel->evlist == NULL);
1377 evsel__free_counts(evsel);
1378 perf_evsel__free_fd(&evsel->core);
1379 perf_evsel__free_id(&evsel->core);
1380 evsel__free_config_terms(evsel);
1381 cgroup__put(evsel->cgrp);
1382 perf_cpu_map__put(evsel->core.cpus);
1383 perf_cpu_map__put(evsel->core.own_cpus);
1384 perf_thread_map__put(evsel->core.threads);
1385 zfree(&evsel->group_name);
1386 zfree(&evsel->name);
1387 zfree(&evsel->pmu_name);
1388 zfree(&evsel->per_pkg_mask);
1389 zfree(&evsel->metric_events);
1390 perf_evsel__object.fini(evsel);
1393 void evsel__delete(struct evsel *evsel)
1399 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1400 struct perf_counts_values *count)
1402 struct perf_counts_values tmp;
1404 if (!evsel->prev_raw_counts)
1408 tmp = evsel->prev_raw_counts->aggr;
1409 evsel->prev_raw_counts->aggr = *count;
1411 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1412 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1415 count->val = count->val - tmp.val;
1416 count->ena = count->ena - tmp.ena;
1417 count->run = count->run - tmp.run;
1420 void perf_counts_values__scale(struct perf_counts_values *count,
1421 bool scale, s8 *pscaled)
1426 if (count->run == 0) {
1429 } else if (count->run < count->ena) {
1431 count->val = (u64)((double) count->val * count->ena / count->run);
1439 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1441 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1443 return perf_evsel__read(&evsel->core, cpu, thread, count);
1447 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1448 u64 val, u64 ena, u64 run)
1450 struct perf_counts_values *count;
1452 count = perf_counts(counter->counts, cpu, thread);
1458 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1462 perf_evsel__process_group_data(struct evsel *leader,
1463 int cpu, int thread, u64 *data)
1465 u64 read_format = leader->core.attr.read_format;
1466 struct sample_read_value *v;
1467 u64 nr, ena = 0, run = 0, i;
1471 if (nr != (u64) leader->core.nr_members)
1474 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1477 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1480 v = (struct sample_read_value *) data;
1482 perf_evsel__set_count(leader, cpu, thread,
1483 v[0].value, ena, run);
1485 for (i = 1; i < nr; i++) {
1486 struct evsel *counter;
1488 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1492 perf_evsel__set_count(counter, cpu, thread,
1493 v[i].value, ena, run);
1499 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1501 struct perf_stat_evsel *ps = leader->stats;
1502 u64 read_format = leader->core.attr.read_format;
1503 int size = perf_evsel__read_size(&leader->core);
1504 u64 *data = ps->group_data;
1506 if (!(read_format & PERF_FORMAT_ID))
1509 if (!evsel__is_group_leader(leader))
1513 data = zalloc(size);
1517 ps->group_data = data;
1520 if (FD(leader, cpu, thread) < 0)
1523 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1526 return perf_evsel__process_group_data(leader, cpu, thread, data);
1529 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1531 u64 read_format = evsel->core.attr.read_format;
1533 if (read_format & PERF_FORMAT_GROUP)
1534 return evsel__read_group(evsel, cpu, thread);
1536 return evsel__read_one(evsel, cpu, thread);
1539 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1541 struct perf_counts_values count;
1542 size_t nv = scale ? 3 : 1;
1544 if (FD(evsel, cpu, thread) < 0)
1547 if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1550 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1553 evsel__compute_deltas(evsel, cpu, thread, &count);
1554 perf_counts_values__scale(&count, scale, NULL);
1555 *perf_counts(evsel->counts, cpu, thread) = count;
1559 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1561 struct evsel *leader = evsel->leader;
1564 if (evsel__is_group_leader(evsel))
1568 * Leader must be already processed/open,
1569 * if not it's a bug.
1571 BUG_ON(!leader->core.fd);
1573 fd = FD(leader, cpu, thread);
1579 static void perf_evsel__remove_fd(struct evsel *pos,
1580 int nr_cpus, int nr_threads,
1583 for (int cpu = 0; cpu < nr_cpus; cpu++)
1584 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1585 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1588 static int update_fds(struct evsel *evsel,
1589 int nr_cpus, int cpu_idx,
1590 int nr_threads, int thread_idx)
1594 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1597 evlist__for_each_entry(evsel->evlist, pos) {
1598 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1600 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1603 * Since fds for next evsel has not been created,
1604 * there is no need to iterate whole event list.
1612 static bool ignore_missing_thread(struct evsel *evsel,
1613 int nr_cpus, int cpu,
1614 struct perf_thread_map *threads,
1615 int thread, int err)
1617 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1619 if (!evsel->ignore_missing_thread)
1622 /* The system wide setup does not work with threads. */
1623 if (evsel->core.system_wide)
1626 /* The -ESRCH is perf event syscall errno for pid's not found. */
1630 /* If there's only one thread, let it fail. */
1631 if (threads->nr == 1)
1635 * We should remove fd for missing_thread first
1636 * because thread_map__remove() will decrease threads->nr.
1638 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1641 if (thread_map__remove(threads, thread))
1644 pr_warning("WARNING: Ignored open failure for pid %d\n",
1649 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1650 void *priv __maybe_unused)
1652 return fprintf(fp, " %-32s %s\n", name, val);
1655 static void display_attr(struct perf_event_attr *attr)
1657 if (verbose >= 2 || debug_peo_args) {
1658 fprintf(stderr, "%.60s\n", graph_dotted_line);
1659 fprintf(stderr, "perf_event_attr:\n");
1660 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1661 fprintf(stderr, "%.60s\n", graph_dotted_line);
1665 static int perf_event_open(struct evsel *evsel,
1666 pid_t pid, int cpu, int group_fd,
1667 unsigned long flags)
1669 int precise_ip = evsel->core.attr.precise_ip;
1673 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1674 pid, cpu, group_fd, flags);
1676 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1680 /* Do not try less precise if not requested. */
1681 if (!evsel->precise_max)
1685 * We tried all the precise_ip values, and it's
1686 * still failing, so leave it to standard fallback.
1688 if (!evsel->core.attr.precise_ip) {
1689 evsel->core.attr.precise_ip = precise_ip;
1693 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1694 evsel->core.attr.precise_ip--;
1695 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1696 display_attr(&evsel->core.attr);
1702 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1703 struct perf_thread_map *threads,
1704 int start_cpu, int end_cpu)
1706 int cpu, thread, nthreads;
1707 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1708 int pid = -1, err, old_errno;
1709 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1711 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1712 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1716 static struct perf_cpu_map *empty_cpu_map;
1718 if (empty_cpu_map == NULL) {
1719 empty_cpu_map = perf_cpu_map__dummy_new();
1720 if (empty_cpu_map == NULL)
1724 cpus = empty_cpu_map;
1727 if (threads == NULL) {
1728 static struct perf_thread_map *empty_thread_map;
1730 if (empty_thread_map == NULL) {
1731 empty_thread_map = thread_map__new_by_tid(-1);
1732 if (empty_thread_map == NULL)
1736 threads = empty_thread_map;
1739 if (evsel->core.system_wide)
1742 nthreads = threads->nr;
1744 if (evsel->core.fd == NULL &&
1745 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1749 flags |= PERF_FLAG_PID_CGROUP;
1750 pid = evsel->cgrp->fd;
1753 fallback_missing_features:
1754 if (perf_missing_features.clockid_wrong)
1755 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1756 if (perf_missing_features.clockid) {
1757 evsel->core.attr.use_clockid = 0;
1758 evsel->core.attr.clockid = 0;
1760 if (perf_missing_features.cloexec)
1761 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1762 if (perf_missing_features.mmap2)
1763 evsel->core.attr.mmap2 = 0;
1764 if (perf_missing_features.exclude_guest)
1765 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1766 if (perf_missing_features.lbr_flags)
1767 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1768 PERF_SAMPLE_BRANCH_NO_CYCLES);
1769 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1770 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1771 if (perf_missing_features.ksymbol)
1772 evsel->core.attr.ksymbol = 0;
1773 if (perf_missing_features.bpf)
1774 evsel->core.attr.bpf_event = 0;
1775 if (perf_missing_features.branch_hw_idx)
1776 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1778 if (perf_missing_features.sample_id_all)
1779 evsel->core.attr.sample_id_all = 0;
1781 display_attr(&evsel->core.attr);
1783 for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1785 for (thread = 0; thread < nthreads; thread++) {
1788 if (!evsel->cgrp && !evsel->core.system_wide)
1789 pid = perf_thread_map__pid(threads, thread);
1791 group_fd = get_group_fd(evsel, cpu, thread);
1795 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1798 FD(evsel, cpu, thread) = fd;
1800 if (unlikely(test_attr__enabled)) {
1801 test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
1802 fd, group_fd, flags);
1808 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1810 * We just removed 1 thread, so take a step
1811 * back on thread index and lower the upper
1817 /* ... and pretend like nothing have happened. */
1822 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1827 pr_debug2_peo(" = %d\n", fd);
1829 if (evsel->bpf_fd >= 0) {
1831 int bpf_fd = evsel->bpf_fd;
1834 PERF_EVENT_IOC_SET_BPF,
1836 if (err && errno != EEXIST) {
1837 pr_err("failed to attach bpf fd %d: %s\n",
1838 bpf_fd, strerror(errno));
1844 set_rlimit = NO_CHANGE;
1847 * If we succeeded but had to kill clockid, fail and
1848 * have evsel__open_strerror() print us a nice error.
1850 if (perf_missing_features.clockid ||
1851 perf_missing_features.clockid_wrong) {
1862 * perf stat needs between 5 and 22 fds per CPU. When we run out
1863 * of them try to increase the limits.
1865 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1869 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1870 if (set_rlimit == NO_CHANGE)
1871 l.rlim_cur = l.rlim_max;
1873 l.rlim_cur = l.rlim_max + 1000;
1874 l.rlim_max = l.rlim_cur;
1876 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1885 if (err != -EINVAL || cpu > 0 || thread > 0)
1889 * Must probe features in the order they were added to the
1890 * perf_event_attr interface.
1892 if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1893 perf_missing_features.cgroup = true;
1894 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1896 } else if (!perf_missing_features.branch_hw_idx &&
1897 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1898 perf_missing_features.branch_hw_idx = true;
1899 pr_debug2("switching off branch HW index support\n");
1900 goto fallback_missing_features;
1901 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1902 perf_missing_features.aux_output = true;
1903 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1905 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1906 perf_missing_features.bpf = true;
1907 pr_debug2_peo("switching off bpf_event\n");
1908 goto fallback_missing_features;
1909 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1910 perf_missing_features.ksymbol = true;
1911 pr_debug2_peo("switching off ksymbol\n");
1912 goto fallback_missing_features;
1913 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1914 perf_missing_features.write_backward = true;
1915 pr_debug2_peo("switching off write_backward\n");
1917 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1918 perf_missing_features.clockid_wrong = true;
1919 pr_debug2_peo("switching off clockid\n");
1920 goto fallback_missing_features;
1921 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1922 perf_missing_features.clockid = true;
1923 pr_debug2_peo("switching off use_clockid\n");
1924 goto fallback_missing_features;
1925 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1926 perf_missing_features.cloexec = true;
1927 pr_debug2_peo("switching off cloexec flag\n");
1928 goto fallback_missing_features;
1929 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1930 perf_missing_features.mmap2 = true;
1931 pr_debug2_peo("switching off mmap2\n");
1932 goto fallback_missing_features;
1933 } else if (!perf_missing_features.exclude_guest &&
1934 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1935 perf_missing_features.exclude_guest = true;
1936 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1937 goto fallback_missing_features;
1938 } else if (!perf_missing_features.sample_id_all) {
1939 perf_missing_features.sample_id_all = true;
1940 pr_debug2_peo("switching off sample_id_all\n");
1941 goto retry_sample_id;
1942 } else if (!perf_missing_features.lbr_flags &&
1943 (evsel->core.attr.branch_sample_type &
1944 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1945 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1946 perf_missing_features.lbr_flags = true;
1947 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1948 goto fallback_missing_features;
1949 } else if (!perf_missing_features.group_read &&
1950 evsel->core.attr.inherit &&
1951 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1952 evsel__is_group_leader(evsel)) {
1953 perf_missing_features.group_read = true;
1954 pr_debug2_peo("switching off group read\n");
1955 goto fallback_missing_features;
1959 threads->err_thread = thread;
1963 while (--thread >= 0) {
1964 if (FD(evsel, cpu, thread) >= 0)
1965 close(FD(evsel, cpu, thread));
1966 FD(evsel, cpu, thread) = -1;
1969 } while (--cpu >= 0);
1974 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1975 struct perf_thread_map *threads)
1977 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1980 void evsel__close(struct evsel *evsel)
1982 perf_evsel__close(&evsel->core);
1983 perf_evsel__free_id(&evsel->core);
1986 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
1989 return evsel__open_cpu(evsel, cpus, NULL, 0,
1990 cpus ? cpus->nr : 1);
1992 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1995 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
1997 return evsel__open(evsel, NULL, threads);
2000 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2001 const union perf_event *event,
2002 struct perf_sample *sample)
2004 u64 type = evsel->core.attr.sample_type;
2005 const __u64 *array = event->sample.array;
2006 bool swapped = evsel->needs_swap;
2009 array += ((event->header.size -
2010 sizeof(event->header)) / sizeof(u64)) - 1;
2012 if (type & PERF_SAMPLE_IDENTIFIER) {
2013 sample->id = *array;
2017 if (type & PERF_SAMPLE_CPU) {
2020 /* undo swap of u64, then swap on individual u32s */
2021 u.val64 = bswap_64(u.val64);
2022 u.val32[0] = bswap_32(u.val32[0]);
2025 sample->cpu = u.val32[0];
2029 if (type & PERF_SAMPLE_STREAM_ID) {
2030 sample->stream_id = *array;
2034 if (type & PERF_SAMPLE_ID) {
2035 sample->id = *array;
2039 if (type & PERF_SAMPLE_TIME) {
2040 sample->time = *array;
2044 if (type & PERF_SAMPLE_TID) {
2047 /* undo swap of u64, then swap on individual u32s */
2048 u.val64 = bswap_64(u.val64);
2049 u.val32[0] = bswap_32(u.val32[0]);
2050 u.val32[1] = bswap_32(u.val32[1]);
2053 sample->pid = u.val32[0];
2054 sample->tid = u.val32[1];
2061 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2064 return size > max_size || offset + size > endp;
2067 #define OVERFLOW_CHECK(offset, size, max_size) \
2069 if (overflow(endp, (max_size), (offset), (size))) \
2073 #define OVERFLOW_CHECK_u64(offset) \
2074 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2077 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2080 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2081 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2082 * check the format does not go past the end of the event.
2084 if (sample_size + sizeof(event->header) > event->header.size)
2090 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2091 struct perf_sample *data)
2093 u64 type = evsel->core.attr.sample_type;
2094 bool swapped = evsel->needs_swap;
2096 u16 max_size = event->header.size;
2097 const void *endp = (void *)event + max_size;
2101 * used for cross-endian analysis. See git commit 65014ab3
2102 * for why this goofiness is needed.
2106 memset(data, 0, sizeof(*data));
2107 data->cpu = data->pid = data->tid = -1;
2108 data->stream_id = data->id = data->time = -1ULL;
2109 data->period = evsel->core.attr.sample_period;
2110 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2111 data->misc = event->header.misc;
2112 data->data_src = PERF_MEM_DATA_SRC_NONE;
2114 if (event->header.type != PERF_RECORD_SAMPLE) {
2115 if (!evsel->core.attr.sample_id_all)
2117 return perf_evsel__parse_id_sample(evsel, event, data);
2120 array = event->sample.array;
2122 if (perf_event__check_size(event, evsel->sample_size))
2125 if (type & PERF_SAMPLE_IDENTIFIER) {
2130 if (type & PERF_SAMPLE_IP) {
2135 if (type & PERF_SAMPLE_TID) {
2138 /* undo swap of u64, then swap on individual u32s */
2139 u.val64 = bswap_64(u.val64);
2140 u.val32[0] = bswap_32(u.val32[0]);
2141 u.val32[1] = bswap_32(u.val32[1]);
2144 data->pid = u.val32[0];
2145 data->tid = u.val32[1];
2149 if (type & PERF_SAMPLE_TIME) {
2150 data->time = *array;
2154 if (type & PERF_SAMPLE_ADDR) {
2155 data->addr = *array;
2159 if (type & PERF_SAMPLE_ID) {
2164 if (type & PERF_SAMPLE_STREAM_ID) {
2165 data->stream_id = *array;
2169 if (type & PERF_SAMPLE_CPU) {
2173 /* undo swap of u64, then swap on individual u32s */
2174 u.val64 = bswap_64(u.val64);
2175 u.val32[0] = bswap_32(u.val32[0]);
2178 data->cpu = u.val32[0];
2182 if (type & PERF_SAMPLE_PERIOD) {
2183 data->period = *array;
2187 if (type & PERF_SAMPLE_READ) {
2188 u64 read_format = evsel->core.attr.read_format;
2190 OVERFLOW_CHECK_u64(array);
2191 if (read_format & PERF_FORMAT_GROUP)
2192 data->read.group.nr = *array;
2194 data->read.one.value = *array;
2198 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2199 OVERFLOW_CHECK_u64(array);
2200 data->read.time_enabled = *array;
2204 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2205 OVERFLOW_CHECK_u64(array);
2206 data->read.time_running = *array;
2210 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2211 if (read_format & PERF_FORMAT_GROUP) {
2212 const u64 max_group_nr = UINT64_MAX /
2213 sizeof(struct sample_read_value);
2215 if (data->read.group.nr > max_group_nr)
2217 sz = data->read.group.nr *
2218 sizeof(struct sample_read_value);
2219 OVERFLOW_CHECK(array, sz, max_size);
2220 data->read.group.values =
2221 (struct sample_read_value *)array;
2222 array = (void *)array + sz;
2224 OVERFLOW_CHECK_u64(array);
2225 data->read.one.id = *array;
2230 if (type & PERF_SAMPLE_CALLCHAIN) {
2231 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2233 OVERFLOW_CHECK_u64(array);
2234 data->callchain = (struct ip_callchain *)array++;
2235 if (data->callchain->nr > max_callchain_nr)
2237 sz = data->callchain->nr * sizeof(u64);
2238 OVERFLOW_CHECK(array, sz, max_size);
2239 array = (void *)array + sz;
2242 if (type & PERF_SAMPLE_RAW) {
2243 OVERFLOW_CHECK_u64(array);
2247 * Undo swap of u64, then swap on individual u32s,
2248 * get the size of the raw area and undo all of the
2249 * swap. The pevent interface handles endianity by
2253 u.val64 = bswap_64(u.val64);
2254 u.val32[0] = bswap_32(u.val32[0]);
2255 u.val32[1] = bswap_32(u.val32[1]);
2257 data->raw_size = u.val32[0];
2260 * The raw data is aligned on 64bits including the
2261 * u32 size, so it's safe to use mem_bswap_64.
2264 mem_bswap_64((void *) array, data->raw_size);
2266 array = (void *)array + sizeof(u32);
2268 OVERFLOW_CHECK(array, data->raw_size, max_size);
2269 data->raw_data = (void *)array;
2270 array = (void *)array + data->raw_size;
2273 if (type & PERF_SAMPLE_BRANCH_STACK) {
2274 const u64 max_branch_nr = UINT64_MAX /
2275 sizeof(struct branch_entry);
2277 OVERFLOW_CHECK_u64(array);
2278 data->branch_stack = (struct branch_stack *)array++;
2280 if (data->branch_stack->nr > max_branch_nr)
2283 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2284 if (evsel__has_branch_hw_idx(evsel))
2287 data->no_hw_idx = true;
2288 OVERFLOW_CHECK(array, sz, max_size);
2289 array = (void *)array + sz;
2292 if (type & PERF_SAMPLE_REGS_USER) {
2293 OVERFLOW_CHECK_u64(array);
2294 data->user_regs.abi = *array;
2297 if (data->user_regs.abi) {
2298 u64 mask = evsel->core.attr.sample_regs_user;
2300 sz = hweight64(mask) * sizeof(u64);
2301 OVERFLOW_CHECK(array, sz, max_size);
2302 data->user_regs.mask = mask;
2303 data->user_regs.regs = (u64 *)array;
2304 array = (void *)array + sz;
2308 if (type & PERF_SAMPLE_STACK_USER) {
2309 OVERFLOW_CHECK_u64(array);
2312 data->user_stack.offset = ((char *)(array - 1)
2316 data->user_stack.size = 0;
2318 OVERFLOW_CHECK(array, sz, max_size);
2319 data->user_stack.data = (char *)array;
2320 array = (void *)array + sz;
2321 OVERFLOW_CHECK_u64(array);
2322 data->user_stack.size = *array++;
2323 if (WARN_ONCE(data->user_stack.size > sz,
2324 "user stack dump failure\n"))
2329 if (type & PERF_SAMPLE_WEIGHT) {
2330 OVERFLOW_CHECK_u64(array);
2331 data->weight = *array;
2335 if (type & PERF_SAMPLE_DATA_SRC) {
2336 OVERFLOW_CHECK_u64(array);
2337 data->data_src = *array;
2341 if (type & PERF_SAMPLE_TRANSACTION) {
2342 OVERFLOW_CHECK_u64(array);
2343 data->transaction = *array;
2347 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2348 if (type & PERF_SAMPLE_REGS_INTR) {
2349 OVERFLOW_CHECK_u64(array);
2350 data->intr_regs.abi = *array;
2353 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2354 u64 mask = evsel->core.attr.sample_regs_intr;
2356 sz = hweight64(mask) * sizeof(u64);
2357 OVERFLOW_CHECK(array, sz, max_size);
2358 data->intr_regs.mask = mask;
2359 data->intr_regs.regs = (u64 *)array;
2360 array = (void *)array + sz;
2364 data->phys_addr = 0;
2365 if (type & PERF_SAMPLE_PHYS_ADDR) {
2366 data->phys_addr = *array;
2371 if (type & PERF_SAMPLE_CGROUP) {
2372 data->cgroup = *array;
2376 if (type & PERF_SAMPLE_AUX) {
2377 OVERFLOW_CHECK_u64(array);
2380 OVERFLOW_CHECK(array, sz, max_size);
2381 /* Undo swap of data */
2383 mem_bswap_64((char *)array, sz);
2384 data->aux_sample.size = sz;
2385 data->aux_sample.data = (char *)array;
2386 array = (void *)array + sz;
2392 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2395 u64 type = evsel->core.attr.sample_type;
2398 if (!(type & PERF_SAMPLE_TIME))
2401 if (event->header.type != PERF_RECORD_SAMPLE) {
2402 struct perf_sample data = {
2406 if (!evsel->core.attr.sample_id_all)
2408 if (perf_evsel__parse_id_sample(evsel, event, &data))
2411 *timestamp = data.time;
2415 array = event->sample.array;
2417 if (perf_event__check_size(event, evsel->sample_size))
2420 if (type & PERF_SAMPLE_IDENTIFIER)
2423 if (type & PERF_SAMPLE_IP)
2426 if (type & PERF_SAMPLE_TID)
2429 if (type & PERF_SAMPLE_TIME)
2430 *timestamp = *array;
2435 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2437 return tep_find_field(evsel->tp_format, name);
2440 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2442 struct tep_format_field *field = evsel__field(evsel, name);
2448 offset = field->offset;
2450 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2451 offset = *(int *)(sample->raw_data + field->offset);
2455 return sample->raw_data + offset;
2458 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2462 void *ptr = sample->raw_data + field->offset;
2464 switch (field->size) {
2468 value = *(u16 *)ptr;
2471 value = *(u32 *)ptr;
2474 memcpy(&value, ptr, sizeof(u64));
2483 switch (field->size) {
2485 return bswap_16(value);
2487 return bswap_32(value);
2489 return bswap_64(value);
2497 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2499 struct tep_format_field *field = evsel__field(evsel, name);
2504 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2507 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2511 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2512 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2513 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2515 * If it's cycles then fall back to hrtimer based
2516 * cpu-clock-tick sw counter, which is always available even if
2519 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2522 scnprintf(msg, msgsize, "%s",
2523 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2525 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2526 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2528 zfree(&evsel->name);
2530 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2531 (paranoid = perf_event_paranoid()) > 1) {
2532 const char *name = evsel__name(evsel);
2534 const char *sep = ":";
2536 /* If event has exclude user then don't exclude kernel. */
2537 if (evsel->core.attr.exclude_user)
2540 /* Is there already the separator in the name. */
2541 if (strchr(name, '/') ||
2542 (strchr(name, ':') && !evsel->is_libpfm_event))
2545 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2550 evsel->name = new_name;
2551 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2552 "to fall back to excluding kernel and hypervisor "
2553 " samples", paranoid);
2554 evsel->core.attr.exclude_kernel = 1;
2555 evsel->core.attr.exclude_hv = 1;
2563 static bool find_process(const char *name)
2565 size_t len = strlen(name);
2570 dir = opendir(procfs__mountpoint());
2574 /* Walk through the directory. */
2575 while (ret && (d = readdir(dir)) != NULL) {
2576 char path[PATH_MAX];
2580 if ((d->d_type != DT_DIR) ||
2581 !strcmp(".", d->d_name) ||
2582 !strcmp("..", d->d_name))
2585 scnprintf(path, sizeof(path), "%s/%s/comm",
2586 procfs__mountpoint(), d->d_name);
2588 if (filename__read_str(path, &data, &size))
2591 ret = strncmp(name, data, len);
2596 return ret ? false : true;
2599 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2600 int err, char *msg, size_t size)
2602 char sbuf[STRERR_BUFSIZE];
2603 int printed = 0, enforced = 0;
2608 printed += scnprintf(msg + printed, size - printed,
2609 "Access to performance monitoring and observability operations is limited.\n");
2611 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2613 printed += scnprintf(msg + printed, size - printed,
2614 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2615 "monitoring and observability operations. Inspect system audit records for\n"
2616 "more perf_event access control information and adjusting the policy.\n");
2621 printed += scnprintf(msg, size,
2622 "No permission to enable %s event.\n\n", evsel__name(evsel));
2624 return scnprintf(msg + printed, size - printed,
2625 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2626 "access to performance monitoring and observability operations for processes\n"
2627 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2628 "More information can be found at 'Perf events and tool security' document:\n"
2629 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2630 "perf_event_paranoid setting is %d:\n"
2631 " -1: Allow use of (almost) all events by all users\n"
2632 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2633 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2634 ">= 1: Disallow CPU event access\n"
2635 ">= 2: Disallow kernel profiling\n"
2636 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2637 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2638 perf_event_paranoid());
2640 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2642 return scnprintf(msg, size, "%s",
2643 "Too many events are opened.\n"
2644 "Probably the maximum number of open file descriptors has been reached.\n"
2645 "Hint: Try again after reducing the number of events.\n"
2646 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2648 if (evsel__has_callchain(evsel) &&
2649 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2650 return scnprintf(msg, size,
2651 "Not enough memory to setup event with callchain.\n"
2652 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2653 "Hint: Current value: %d", sysctl__max_stack());
2656 if (target->cpu_list)
2657 return scnprintf(msg, size, "%s",
2658 "No such device - did you specify an out-of-range profile CPU?");
2661 if (evsel->core.attr.aux_output)
2662 return scnprintf(msg, size,
2663 "%s: PMU Hardware doesn't support 'aux_output' feature",
2664 evsel__name(evsel));
2665 if (evsel->core.attr.sample_period != 0)
2666 return scnprintf(msg, size,
2667 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2668 evsel__name(evsel));
2669 if (evsel->core.attr.precise_ip)
2670 return scnprintf(msg, size, "%s",
2671 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2672 #if defined(__i386__) || defined(__x86_64__)
2673 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2674 return scnprintf(msg, size, "%s",
2675 "No hardware sampling interrupt available.\n");
2679 if (find_process("oprofiled"))
2680 return scnprintf(msg, size,
2681 "The PMU counters are busy/taken by another profiler.\n"
2682 "We found oprofile daemon running, please stop it and try again.");
2685 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2686 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2687 if (perf_missing_features.clockid)
2688 return scnprintf(msg, size, "clockid feature not supported.");
2689 if (perf_missing_features.clockid_wrong)
2690 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2691 if (perf_missing_features.aux_output)
2692 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2698 return scnprintf(msg, size,
2699 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2700 "/bin/dmesg | grep -i perf may provide additional information.\n",
2701 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2704 struct perf_env *evsel__env(struct evsel *evsel)
2706 if (evsel && evsel->evlist)
2707 return evsel->evlist->env;
2711 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2715 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2716 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2718 int fd = FD(evsel, cpu, thread);
2720 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2721 cpu, thread, fd) < 0)
2729 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2731 struct perf_cpu_map *cpus = evsel->core.cpus;
2732 struct perf_thread_map *threads = evsel->core.threads;
2734 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2737 return store_evsel_ids(evsel, evlist);