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"
33 #include "util/evsel_config.h"
34 #include "util/evsel_fprintf.h"
36 #include <perf/cpumap.h>
37 #include "thread_map.h"
39 #include "perf_regs.h"
42 #include "trace-event.h"
47 #include "../perf-sys.h"
48 #include "util/parse-branch-options.h"
49 #include <internal/xyarray.h>
50 #include <internal/lib.h>
52 #include <linux/ctype.h>
54 struct perf_missing_features perf_missing_features;
56 static clockid_t clockid;
58 static int perf_evsel__no_extra_init(struct evsel *evsel __maybe_unused)
63 void __weak test_attr__ready(void) { }
65 static void perf_evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
71 int (*init)(struct evsel *evsel);
72 void (*fini)(struct evsel *evsel);
73 } perf_evsel__object = {
74 .size = sizeof(struct evsel),
75 .init = perf_evsel__no_extra_init,
76 .fini = perf_evsel__no_extra_fini,
79 int perf_evsel__object_config(size_t object_size,
80 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 __perf_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 perf_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 __perf_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 perf_evsel__calc_id_pos(evsel);
196 void __perf_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 perf_evsel__calc_id_pos(evsel);
206 void perf_evsel__set_sample_id(struct evsel *evsel,
207 bool can_sample_identifier)
209 if (can_sample_identifier) {
210 perf_evsel__reset_sample_bit(evsel, ID);
211 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
213 perf_evsel__set_sample_bit(evsel, ID);
215 evsel->core.attr.read_format |= PERF_FORMAT_ID;
219 * perf_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 perf_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 = __perf_evsel__sample_size(attr->sample_type);
252 perf_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->collect_stat = false;
258 evsel->pmu_name = NULL;
261 struct evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
263 struct evsel *evsel = zalloc(perf_evsel__object.size);
267 evsel__init(evsel, attr, idx);
269 if (perf_evsel__is_bpf_output(evsel)) {
270 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
271 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
272 evsel->core.attr.sample_period = 1;
275 if (perf_evsel__is_clock(evsel)) {
277 * The evsel->unit points to static alias->unit
278 * so it's ok to use static string in here.
280 static const char *unit = "msec";
289 static bool perf_event_can_profile_kernel(void)
291 return perf_event_paranoid_check(1);
294 struct evsel *perf_evsel__new_cycles(bool precise)
296 struct perf_event_attr attr = {
297 .type = PERF_TYPE_HARDWARE,
298 .config = PERF_COUNT_HW_CPU_CYCLES,
299 .exclude_kernel = !perf_event_can_profile_kernel(),
303 event_attr_init(&attr);
309 * Now let the usual logic to set up the perf_event_attr defaults
310 * to kick in when we return and before perf_evsel__open() is called.
313 evsel = evsel__new(&attr);
317 evsel->precise_max = true;
319 /* use asprintf() because free(evsel) assumes name is allocated */
320 if (asprintf(&evsel->name, "cycles%s%s%.*s",
321 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
322 attr.exclude_kernel ? "u" : "",
323 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
328 evsel__delete(evsel);
334 * Returns pointer with encoded error via <linux/err.h> interface.
336 struct evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
338 struct evsel *evsel = zalloc(perf_evsel__object.size);
344 struct perf_event_attr attr = {
345 .type = PERF_TYPE_TRACEPOINT,
346 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
347 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
350 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
353 evsel->tp_format = trace_event__tp_format(sys, name);
354 if (IS_ERR(evsel->tp_format)) {
355 err = PTR_ERR(evsel->tp_format);
359 event_attr_init(&attr);
360 attr.config = evsel->tp_format->id;
361 attr.sample_period = 1;
362 evsel__init(evsel, &attr, idx);
374 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
382 "stalled-cycles-frontend",
383 "stalled-cycles-backend",
387 static const char *__perf_evsel__hw_name(u64 config)
389 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
390 return perf_evsel__hw_names[config];
392 return "unknown-hardware";
395 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
397 int colon = 0, r = 0;
398 struct perf_event_attr *attr = &evsel->core.attr;
399 bool exclude_guest_default = false;
401 #define MOD_PRINT(context, mod) do { \
402 if (!attr->exclude_##context) { \
403 if (!colon) colon = ++r; \
404 r += scnprintf(bf + r, size - r, "%c", mod); \
407 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
408 MOD_PRINT(kernel, 'k');
409 MOD_PRINT(user, 'u');
411 exclude_guest_default = true;
414 if (attr->precise_ip) {
417 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
418 exclude_guest_default = true;
421 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
422 MOD_PRINT(host, 'H');
423 MOD_PRINT(guest, 'G');
431 static int perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
433 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config));
434 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
437 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
450 static const char *__perf_evsel__sw_name(u64 config)
452 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
453 return perf_evsel__sw_names[config];
454 return "unknown-software";
457 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
459 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config));
460 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
463 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
467 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
469 if (type & HW_BREAKPOINT_R)
470 r += scnprintf(bf + r, size - r, "r");
472 if (type & HW_BREAKPOINT_W)
473 r += scnprintf(bf + r, size - r, "w");
475 if (type & HW_BREAKPOINT_X)
476 r += scnprintf(bf + r, size - r, "x");
481 static int perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
483 struct perf_event_attr *attr = &evsel->core.attr;
484 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
485 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
488 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
489 [PERF_EVSEL__MAX_ALIASES] = {
490 { "L1-dcache", "l1-d", "l1d", "L1-data", },
491 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
493 { "dTLB", "d-tlb", "Data-TLB", },
494 { "iTLB", "i-tlb", "Instruction-TLB", },
495 { "branch", "branches", "bpu", "btb", "bpc", },
499 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
500 [PERF_EVSEL__MAX_ALIASES] = {
501 { "load", "loads", "read", },
502 { "store", "stores", "write", },
503 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
506 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
507 [PERF_EVSEL__MAX_ALIASES] = {
508 { "refs", "Reference", "ops", "access", },
509 { "misses", "miss", },
512 #define C(x) PERF_COUNT_HW_CACHE_##x
513 #define CACHE_READ (1 << C(OP_READ))
514 #define CACHE_WRITE (1 << C(OP_WRITE))
515 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
516 #define COP(x) (1 << x)
519 * cache operartion stat
520 * L1I : Read and prefetch only
521 * ITLB and BPU : Read-only
523 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
524 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
525 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
526 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
527 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
528 [C(ITLB)] = (CACHE_READ),
529 [C(BPU)] = (CACHE_READ),
530 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
533 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
535 if (perf_evsel__hw_cache_stat[type] & COP(op))
536 return true; /* valid */
538 return false; /* invalid */
541 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
542 char *bf, size_t size)
545 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
546 perf_evsel__hw_cache_op[op][0],
547 perf_evsel__hw_cache_result[result][0]);
550 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
551 perf_evsel__hw_cache_op[op][1]);
554 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
556 u8 op, result, type = (config >> 0) & 0xff;
557 const char *err = "unknown-ext-hardware-cache-type";
559 if (type >= PERF_COUNT_HW_CACHE_MAX)
562 op = (config >> 8) & 0xff;
563 err = "unknown-ext-hardware-cache-op";
564 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
567 result = (config >> 16) & 0xff;
568 err = "unknown-ext-hardware-cache-result";
569 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
572 err = "invalid-cache";
573 if (!perf_evsel__is_cache_op_valid(type, op))
576 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
578 return scnprintf(bf, size, "%s", err);
581 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
583 int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size);
584 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
587 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
589 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
590 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
593 static int perf_evsel__tool_name(char *bf, size_t size)
595 int ret = scnprintf(bf, size, "duration_time");
599 const char *perf_evsel__name(struct evsel *evsel)
609 switch (evsel->core.attr.type) {
611 perf_evsel__raw_name(evsel, bf, sizeof(bf));
614 case PERF_TYPE_HARDWARE:
615 perf_evsel__hw_name(evsel, bf, sizeof(bf));
618 case PERF_TYPE_HW_CACHE:
619 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
622 case PERF_TYPE_SOFTWARE:
623 if (evsel->tool_event)
624 perf_evsel__tool_name(bf, sizeof(bf));
626 perf_evsel__sw_name(evsel, bf, sizeof(bf));
629 case PERF_TYPE_TRACEPOINT:
630 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
633 case PERF_TYPE_BREAKPOINT:
634 perf_evsel__bp_name(evsel, bf, sizeof(bf));
638 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
639 evsel->core.attr.type);
643 evsel->name = strdup(bf);
651 const char *perf_evsel__group_name(struct evsel *evsel)
653 return evsel->group_name ?: "anon group";
657 * Returns the group details for the specified leader,
658 * with following rules.
660 * For record -e '{cycles,instructions}'
661 * 'anon group { cycles:u, instructions:u }'
663 * For record -e 'cycles,instructions' and report --group
664 * 'cycles:u, instructions:u'
666 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
670 const char *group_name = perf_evsel__group_name(evsel);
672 if (!evsel->forced_leader)
673 ret = scnprintf(buf, size, "%s { ", group_name);
675 ret += scnprintf(buf + ret, size - ret, "%s",
676 perf_evsel__name(evsel));
678 for_each_group_member(pos, evsel)
679 ret += scnprintf(buf + ret, size - ret, ", %s",
680 perf_evsel__name(pos));
682 if (!evsel->forced_leader)
683 ret += scnprintf(buf + ret, size - ret, " }");
688 static void __perf_evsel__config_callchain(struct evsel *evsel,
689 struct record_opts *opts,
690 struct callchain_param *param)
692 bool function = perf_evsel__is_function_event(evsel);
693 struct perf_event_attr *attr = &evsel->core.attr;
695 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
697 attr->sample_max_stack = param->max_stack;
699 if (opts->kernel_callchains)
700 attr->exclude_callchain_user = 1;
701 if (opts->user_callchains)
702 attr->exclude_callchain_kernel = 1;
703 if (param->record_mode == CALLCHAIN_LBR) {
704 if (!opts->branch_stack) {
705 if (attr->exclude_user) {
706 pr_warning("LBR callstack option is only available "
707 "to get user callchain information. "
708 "Falling back to framepointers.\n");
710 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
711 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
712 PERF_SAMPLE_BRANCH_CALL_STACK |
713 PERF_SAMPLE_BRANCH_NO_CYCLES |
714 PERF_SAMPLE_BRANCH_NO_FLAGS;
717 pr_warning("Cannot use LBR callstack with branch stack. "
718 "Falling back to framepointers.\n");
721 if (param->record_mode == CALLCHAIN_DWARF) {
723 perf_evsel__set_sample_bit(evsel, REGS_USER);
724 perf_evsel__set_sample_bit(evsel, STACK_USER);
725 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
726 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
727 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
728 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
729 "so the minimal registers set (IP, SP) is explicitly forced.\n");
731 attr->sample_regs_user |= PERF_REGS_MASK;
733 attr->sample_stack_user = param->dump_size;
734 attr->exclude_callchain_user = 1;
736 pr_info("Cannot use DWARF unwind for function trace event,"
737 " falling back to framepointers.\n");
742 pr_info("Disabling user space callchains for function trace event.\n");
743 attr->exclude_callchain_user = 1;
747 void perf_evsel__config_callchain(struct evsel *evsel,
748 struct record_opts *opts,
749 struct callchain_param *param)
752 return __perf_evsel__config_callchain(evsel, opts, param);
756 perf_evsel__reset_callgraph(struct evsel *evsel,
757 struct callchain_param *param)
759 struct perf_event_attr *attr = &evsel->core.attr;
761 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
762 if (param->record_mode == CALLCHAIN_LBR) {
763 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
764 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
765 PERF_SAMPLE_BRANCH_CALL_STACK);
767 if (param->record_mode == CALLCHAIN_DWARF) {
768 perf_evsel__reset_sample_bit(evsel, REGS_USER);
769 perf_evsel__reset_sample_bit(evsel, STACK_USER);
773 static void apply_config_terms(struct evsel *evsel,
774 struct record_opts *opts, bool track)
776 struct perf_evsel_config_term *term;
777 struct list_head *config_terms = &evsel->config_terms;
778 struct perf_event_attr *attr = &evsel->core.attr;
779 /* callgraph default */
780 struct callchain_param param = {
781 .record_mode = callchain_param.record_mode,
785 const char *callgraph_buf = NULL;
787 list_for_each_entry(term, config_terms, list) {
788 switch (term->type) {
789 case PERF_EVSEL__CONFIG_TERM_PERIOD:
790 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
791 attr->sample_period = term->val.period;
793 perf_evsel__reset_sample_bit(evsel, PERIOD);
796 case PERF_EVSEL__CONFIG_TERM_FREQ:
797 if (!(term->weak && opts->user_freq != UINT_MAX)) {
798 attr->sample_freq = term->val.freq;
800 perf_evsel__set_sample_bit(evsel, PERIOD);
803 case PERF_EVSEL__CONFIG_TERM_TIME:
805 perf_evsel__set_sample_bit(evsel, TIME);
807 perf_evsel__reset_sample_bit(evsel, TIME);
809 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
810 callgraph_buf = term->val.callgraph;
812 case PERF_EVSEL__CONFIG_TERM_BRANCH:
813 if (term->val.branch && strcmp(term->val.branch, "no")) {
814 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
815 parse_branch_str(term->val.branch,
816 &attr->branch_sample_type);
818 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
820 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
821 dump_size = term->val.stack_user;
823 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
824 max_stack = term->val.max_stack;
826 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
827 evsel->max_events = term->val.max_events;
829 case PERF_EVSEL__CONFIG_TERM_INHERIT:
831 * attr->inherit should has already been set by
832 * perf_evsel__config. If user explicitly set
833 * inherit using config terms, override global
834 * opt->no_inherit setting.
836 attr->inherit = term->val.inherit ? 1 : 0;
838 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
839 attr->write_backward = term->val.overwrite ? 1 : 0;
841 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
843 case PERF_EVSEL__CONFIG_TERM_PERCORE:
845 case PERF_EVSEL__CONFIG_TERM_AUX_OUTPUT:
846 attr->aux_output = term->val.aux_output ? 1 : 0;
853 /* User explicitly set per-event callgraph, clear the old setting and reset. */
854 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
855 bool sample_address = false;
858 param.max_stack = max_stack;
859 if (callgraph_buf == NULL)
860 callgraph_buf = "fp";
863 /* parse callgraph parameters */
864 if (callgraph_buf != NULL) {
865 if (!strcmp(callgraph_buf, "no")) {
866 param.enabled = false;
867 param.record_mode = CALLCHAIN_NONE;
869 param.enabled = true;
870 if (parse_callchain_record(callgraph_buf, ¶m)) {
871 pr_err("per-event callgraph setting for %s failed. "
872 "Apply callgraph global setting for it\n",
876 if (param.record_mode == CALLCHAIN_DWARF)
877 sample_address = true;
881 dump_size = round_up(dump_size, sizeof(u64));
882 param.dump_size = dump_size;
885 /* If global callgraph set, clear it */
886 if (callchain_param.enabled)
887 perf_evsel__reset_callgraph(evsel, &callchain_param);
889 /* set perf-event callgraph */
891 if (sample_address) {
892 perf_evsel__set_sample_bit(evsel, ADDR);
893 perf_evsel__set_sample_bit(evsel, DATA_SRC);
894 evsel->core.attr.mmap_data = track;
896 perf_evsel__config_callchain(evsel, opts, ¶m);
901 static bool is_dummy_event(struct evsel *evsel)
903 return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
904 (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
908 * The enable_on_exec/disabled value strategy:
910 * 1) For any type of traced program:
911 * - all independent events and group leaders are disabled
912 * - all group members are enabled
914 * Group members are ruled by group leaders. They need to
915 * be enabled, because the group scheduling relies on that.
917 * 2) For traced programs executed by perf:
918 * - all independent events and group leaders have
920 * - we don't specifically enable or disable any event during
923 * Independent events and group leaders are initially disabled
924 * and get enabled by exec. Group members are ruled by group
925 * leaders as stated in 1).
927 * 3) For traced programs attached by perf (pid/tid):
928 * - we specifically enable or disable all events during
931 * When attaching events to already running traced we
932 * enable/disable events specifically, as there's no
933 * initial traced exec call.
935 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
936 struct callchain_param *callchain)
938 struct evsel *leader = evsel->leader;
939 struct perf_event_attr *attr = &evsel->core.attr;
940 int track = evsel->tracking;
941 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
943 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
944 attr->inherit = !opts->no_inherit;
945 attr->write_backward = opts->overwrite ? 1 : 0;
947 perf_evsel__set_sample_bit(evsel, IP);
948 perf_evsel__set_sample_bit(evsel, TID);
950 if (evsel->sample_read) {
951 perf_evsel__set_sample_bit(evsel, READ);
954 * We need ID even in case of single event, because
955 * PERF_SAMPLE_READ process ID specific data.
957 perf_evsel__set_sample_id(evsel, false);
960 * Apply group format only if we belong to group
961 * with more than one members.
963 if (leader->core.nr_members > 1) {
964 attr->read_format |= PERF_FORMAT_GROUP;
970 * We default some events to have a default interval. But keep
971 * it a weak assumption overridable by the user.
973 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
974 opts->user_interval != ULLONG_MAX)) {
976 perf_evsel__set_sample_bit(evsel, PERIOD);
978 attr->sample_freq = opts->freq;
980 attr->sample_period = opts->default_interval;
985 * Disable sampling for all group members other
986 * than leader in case leader 'leads' the sampling.
988 if ((leader != evsel) && leader->sample_read) {
990 attr->sample_freq = 0;
991 attr->sample_period = 0;
992 attr->write_backward = 0;
995 * We don't get sample for slave events, we make them
996 * when delivering group leader sample. Set the slave
997 * event to follow the master sample_type to ease up
1000 attr->sample_type = leader->core.attr.sample_type;
1003 if (opts->no_samples)
1004 attr->sample_freq = 0;
1006 if (opts->inherit_stat) {
1007 evsel->core.attr.read_format |=
1008 PERF_FORMAT_TOTAL_TIME_ENABLED |
1009 PERF_FORMAT_TOTAL_TIME_RUNNING |
1011 attr->inherit_stat = 1;
1014 if (opts->sample_address) {
1015 perf_evsel__set_sample_bit(evsel, ADDR);
1016 attr->mmap_data = track;
1020 * We don't allow user space callchains for function trace
1021 * event, due to issues with page faults while tracing page
1022 * fault handler and its overall trickiness nature.
1024 if (perf_evsel__is_function_event(evsel))
1025 evsel->core.attr.exclude_callchain_user = 1;
1027 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1028 perf_evsel__config_callchain(evsel, opts, callchain);
1030 if (opts->sample_intr_regs) {
1031 attr->sample_regs_intr = opts->sample_intr_regs;
1032 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1035 if (opts->sample_user_regs) {
1036 attr->sample_regs_user |= opts->sample_user_regs;
1037 perf_evsel__set_sample_bit(evsel, REGS_USER);
1040 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1041 perf_evsel__set_sample_bit(evsel, CPU);
1044 * When the user explicitly disabled time don't force it here.
1046 if (opts->sample_time &&
1047 (!perf_missing_features.sample_id_all &&
1048 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1049 opts->sample_time_set)))
1050 perf_evsel__set_sample_bit(evsel, TIME);
1052 if (opts->raw_samples && !evsel->no_aux_samples) {
1053 perf_evsel__set_sample_bit(evsel, TIME);
1054 perf_evsel__set_sample_bit(evsel, RAW);
1055 perf_evsel__set_sample_bit(evsel, CPU);
1058 if (opts->sample_address)
1059 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1061 if (opts->sample_phys_addr)
1062 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1064 if (opts->no_buffering) {
1065 attr->watermark = 0;
1066 attr->wakeup_events = 1;
1068 if (opts->branch_stack && !evsel->no_aux_samples) {
1069 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1070 attr->branch_sample_type = opts->branch_stack;
1073 if (opts->sample_weight)
1074 perf_evsel__set_sample_bit(evsel, WEIGHT);
1078 attr->mmap2 = track && !perf_missing_features.mmap2;
1080 attr->ksymbol = track && !perf_missing_features.ksymbol;
1081 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1083 if (opts->record_namespaces)
1084 attr->namespaces = track;
1086 if (opts->record_switch_events)
1087 attr->context_switch = track;
1089 if (opts->sample_transaction)
1090 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1092 if (opts->running_time) {
1093 evsel->core.attr.read_format |=
1094 PERF_FORMAT_TOTAL_TIME_ENABLED |
1095 PERF_FORMAT_TOTAL_TIME_RUNNING;
1099 * XXX see the function comment above
1101 * Disabling only independent events or group leaders,
1102 * keeping group members enabled.
1104 if (perf_evsel__is_group_leader(evsel))
1108 * Setting enable_on_exec for independent events and
1109 * group leaders for traced executed by perf.
1111 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1112 !opts->initial_delay)
1113 attr->enable_on_exec = 1;
1115 if (evsel->immediate) {
1117 attr->enable_on_exec = 0;
1120 clockid = opts->clockid;
1121 if (opts->use_clockid) {
1122 attr->use_clockid = 1;
1123 attr->clockid = opts->clockid;
1126 if (evsel->precise_max)
1127 attr->precise_ip = 3;
1129 if (opts->all_user) {
1130 attr->exclude_kernel = 1;
1131 attr->exclude_user = 0;
1134 if (opts->all_kernel) {
1135 attr->exclude_kernel = 0;
1136 attr->exclude_user = 1;
1139 if (evsel->core.own_cpus || evsel->unit)
1140 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1143 * Apply event specific term settings,
1144 * it overloads any global configuration.
1146 apply_config_terms(evsel, opts, track);
1148 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1150 /* The --period option takes the precedence. */
1151 if (opts->period_set) {
1153 perf_evsel__set_sample_bit(evsel, PERIOD);
1155 perf_evsel__reset_sample_bit(evsel, PERIOD);
1159 * For initial_delay, a dummy event is added implicitly.
1160 * The software event will trigger -EOPNOTSUPP error out,
1161 * if BRANCH_STACK bit is set.
1163 if (opts->initial_delay && is_dummy_event(evsel))
1164 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1167 int perf_evsel__set_filter(struct evsel *evsel, const char *filter)
1169 char *new_filter = strdup(filter);
1171 if (new_filter != NULL) {
1172 free(evsel->filter);
1173 evsel->filter = new_filter;
1180 static int perf_evsel__append_filter(struct evsel *evsel,
1181 const char *fmt, const char *filter)
1185 if (evsel->filter == NULL)
1186 return perf_evsel__set_filter(evsel, filter);
1188 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1189 free(evsel->filter);
1190 evsel->filter = new_filter;
1197 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1199 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1202 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1204 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1207 int evsel__enable(struct evsel *evsel)
1209 int err = perf_evsel__enable(&evsel->core);
1212 evsel->disabled = false;
1217 int evsel__disable(struct evsel *evsel)
1219 int err = perf_evsel__disable(&evsel->core);
1221 * We mark it disabled here so that tools that disable a event can
1222 * ignore events after they disable it. I.e. the ring buffer may have
1223 * already a few more events queued up before the kernel got the stop
1227 evsel->disabled = true;
1232 static void perf_evsel__free_config_terms(struct evsel *evsel)
1234 struct perf_evsel_config_term *term, *h;
1236 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1237 list_del_init(&term->list);
1242 void perf_evsel__exit(struct evsel *evsel)
1244 assert(list_empty(&evsel->core.node));
1245 assert(evsel->evlist == NULL);
1246 perf_evsel__free_counts(evsel);
1247 perf_evsel__free_fd(&evsel->core);
1248 perf_evsel__free_id(&evsel->core);
1249 perf_evsel__free_config_terms(evsel);
1250 cgroup__put(evsel->cgrp);
1251 perf_cpu_map__put(evsel->core.cpus);
1252 perf_cpu_map__put(evsel->core.own_cpus);
1253 perf_thread_map__put(evsel->core.threads);
1254 zfree(&evsel->group_name);
1255 zfree(&evsel->name);
1256 perf_evsel__object.fini(evsel);
1259 void evsel__delete(struct evsel *evsel)
1261 perf_evsel__exit(evsel);
1265 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1266 struct perf_counts_values *count)
1268 struct perf_counts_values tmp;
1270 if (!evsel->prev_raw_counts)
1274 tmp = evsel->prev_raw_counts->aggr;
1275 evsel->prev_raw_counts->aggr = *count;
1277 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1278 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1281 count->val = count->val - tmp.val;
1282 count->ena = count->ena - tmp.ena;
1283 count->run = count->run - tmp.run;
1286 void perf_counts_values__scale(struct perf_counts_values *count,
1287 bool scale, s8 *pscaled)
1292 if (count->run == 0) {
1295 } else if (count->run < count->ena) {
1297 count->val = (u64)((double) count->val * count->ena / count->run);
1306 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread)
1308 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1310 return perf_evsel__read(&evsel->core, cpu, thread, count);
1314 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1315 u64 val, u64 ena, u64 run)
1317 struct perf_counts_values *count;
1319 count = perf_counts(counter->counts, cpu, thread);
1325 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1329 perf_evsel__process_group_data(struct evsel *leader,
1330 int cpu, int thread, u64 *data)
1332 u64 read_format = leader->core.attr.read_format;
1333 struct sample_read_value *v;
1334 u64 nr, ena = 0, run = 0, i;
1338 if (nr != (u64) leader->core.nr_members)
1341 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1344 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1347 v = (struct sample_read_value *) data;
1349 perf_evsel__set_count(leader, cpu, thread,
1350 v[0].value, ena, run);
1352 for (i = 1; i < nr; i++) {
1353 struct evsel *counter;
1355 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1359 perf_evsel__set_count(counter, cpu, thread,
1360 v[i].value, ena, run);
1367 perf_evsel__read_group(struct evsel *leader, int cpu, int thread)
1369 struct perf_stat_evsel *ps = leader->stats;
1370 u64 read_format = leader->core.attr.read_format;
1371 int size = perf_evsel__read_size(&leader->core);
1372 u64 *data = ps->group_data;
1374 if (!(read_format & PERF_FORMAT_ID))
1377 if (!perf_evsel__is_group_leader(leader))
1381 data = zalloc(size);
1385 ps->group_data = data;
1388 if (FD(leader, cpu, thread) < 0)
1391 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1394 return perf_evsel__process_group_data(leader, cpu, thread, data);
1397 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1399 u64 read_format = evsel->core.attr.read_format;
1401 if (read_format & PERF_FORMAT_GROUP)
1402 return perf_evsel__read_group(evsel, cpu, thread);
1404 return perf_evsel__read_one(evsel, cpu, thread);
1407 int __perf_evsel__read_on_cpu(struct evsel *evsel,
1408 int cpu, int thread, bool scale)
1410 struct perf_counts_values count;
1411 size_t nv = scale ? 3 : 1;
1413 if (FD(evsel, cpu, thread) < 0)
1416 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1419 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1422 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1423 perf_counts_values__scale(&count, scale, NULL);
1424 *perf_counts(evsel->counts, cpu, thread) = count;
1428 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1430 struct evsel *leader = evsel->leader;
1433 if (perf_evsel__is_group_leader(evsel))
1437 * Leader must be already processed/open,
1438 * if not it's a bug.
1440 BUG_ON(!leader->core.fd);
1442 fd = FD(leader, cpu, thread);
1448 static void perf_evsel__remove_fd(struct evsel *pos,
1449 int nr_cpus, int nr_threads,
1452 for (int cpu = 0; cpu < nr_cpus; cpu++)
1453 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1454 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1457 static int update_fds(struct evsel *evsel,
1458 int nr_cpus, int cpu_idx,
1459 int nr_threads, int thread_idx)
1463 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1466 evlist__for_each_entry(evsel->evlist, pos) {
1467 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1469 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1472 * Since fds for next evsel has not been created,
1473 * there is no need to iterate whole event list.
1481 static bool ignore_missing_thread(struct evsel *evsel,
1482 int nr_cpus, int cpu,
1483 struct perf_thread_map *threads,
1484 int thread, int err)
1486 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1488 if (!evsel->ignore_missing_thread)
1491 /* The system wide setup does not work with threads. */
1492 if (evsel->core.system_wide)
1495 /* The -ESRCH is perf event syscall errno for pid's not found. */
1499 /* If there's only one thread, let it fail. */
1500 if (threads->nr == 1)
1504 * We should remove fd for missing_thread first
1505 * because thread_map__remove() will decrease threads->nr.
1507 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1510 if (thread_map__remove(threads, thread))
1513 pr_warning("WARNING: Ignored open failure for pid %d\n",
1518 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1519 void *priv __maybe_unused)
1521 return fprintf(fp, " %-32s %s\n", name, val);
1524 static void display_attr(struct perf_event_attr *attr)
1527 fprintf(stderr, "%.60s\n", graph_dotted_line);
1528 fprintf(stderr, "perf_event_attr:\n");
1529 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1530 fprintf(stderr, "%.60s\n", graph_dotted_line);
1534 static int perf_event_open(struct evsel *evsel,
1535 pid_t pid, int cpu, int group_fd,
1536 unsigned long flags)
1538 int precise_ip = evsel->core.attr.precise_ip;
1542 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1543 pid, cpu, group_fd, flags);
1545 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1549 /* Do not try less precise if not requested. */
1550 if (!evsel->precise_max)
1554 * We tried all the precise_ip values, and it's
1555 * still failing, so leave it to standard fallback.
1557 if (!evsel->core.attr.precise_ip) {
1558 evsel->core.attr.precise_ip = precise_ip;
1562 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1563 evsel->core.attr.precise_ip--;
1564 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1565 display_attr(&evsel->core.attr);
1571 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1572 struct perf_thread_map *threads)
1574 int cpu, thread, nthreads;
1575 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1577 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1579 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1580 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1584 static struct perf_cpu_map *empty_cpu_map;
1586 if (empty_cpu_map == NULL) {
1587 empty_cpu_map = perf_cpu_map__dummy_new();
1588 if (empty_cpu_map == NULL)
1592 cpus = empty_cpu_map;
1595 if (threads == NULL) {
1596 static struct perf_thread_map *empty_thread_map;
1598 if (empty_thread_map == NULL) {
1599 empty_thread_map = thread_map__new_by_tid(-1);
1600 if (empty_thread_map == NULL)
1604 threads = empty_thread_map;
1607 if (evsel->core.system_wide)
1610 nthreads = threads->nr;
1612 if (evsel->core.fd == NULL &&
1613 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1617 flags |= PERF_FLAG_PID_CGROUP;
1618 pid = evsel->cgrp->fd;
1621 fallback_missing_features:
1622 if (perf_missing_features.clockid_wrong)
1623 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1624 if (perf_missing_features.clockid) {
1625 evsel->core.attr.use_clockid = 0;
1626 evsel->core.attr.clockid = 0;
1628 if (perf_missing_features.cloexec)
1629 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1630 if (perf_missing_features.mmap2)
1631 evsel->core.attr.mmap2 = 0;
1632 if (perf_missing_features.exclude_guest)
1633 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1634 if (perf_missing_features.lbr_flags)
1635 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1636 PERF_SAMPLE_BRANCH_NO_CYCLES);
1637 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1638 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1639 if (perf_missing_features.ksymbol)
1640 evsel->core.attr.ksymbol = 0;
1641 if (perf_missing_features.bpf)
1642 evsel->core.attr.bpf_event = 0;
1644 if (perf_missing_features.sample_id_all)
1645 evsel->core.attr.sample_id_all = 0;
1647 display_attr(&evsel->core.attr);
1649 for (cpu = 0; cpu < cpus->nr; cpu++) {
1651 for (thread = 0; thread < nthreads; thread++) {
1654 if (!evsel->cgrp && !evsel->core.system_wide)
1655 pid = perf_thread_map__pid(threads, thread);
1657 group_fd = get_group_fd(evsel, cpu, thread);
1661 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1664 FD(evsel, cpu, thread) = fd;
1669 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1671 * We just removed 1 thread, so take a step
1672 * back on thread index and lower the upper
1678 /* ... and pretend like nothing have happened. */
1683 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1688 pr_debug2(" = %d\n", fd);
1690 if (evsel->bpf_fd >= 0) {
1692 int bpf_fd = evsel->bpf_fd;
1695 PERF_EVENT_IOC_SET_BPF,
1697 if (err && errno != EEXIST) {
1698 pr_err("failed to attach bpf fd %d: %s\n",
1699 bpf_fd, strerror(errno));
1705 set_rlimit = NO_CHANGE;
1708 * If we succeeded but had to kill clockid, fail and
1709 * have perf_evsel__open_strerror() print us a nice
1712 if (perf_missing_features.clockid ||
1713 perf_missing_features.clockid_wrong) {
1724 * perf stat needs between 5 and 22 fds per CPU. When we run out
1725 * of them try to increase the limits.
1727 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1729 int old_errno = errno;
1731 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1732 if (set_rlimit == NO_CHANGE)
1733 l.rlim_cur = l.rlim_max;
1735 l.rlim_cur = l.rlim_max + 1000;
1736 l.rlim_max = l.rlim_cur;
1738 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1747 if (err != -EINVAL || cpu > 0 || thread > 0)
1751 * Must probe features in the order they were added to the
1752 * perf_event_attr interface.
1754 if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1755 perf_missing_features.aux_output = true;
1756 pr_debug2("Kernel has no attr.aux_output support, bailing out\n");
1758 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1759 perf_missing_features.bpf = true;
1760 pr_debug2("switching off bpf_event\n");
1761 goto fallback_missing_features;
1762 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1763 perf_missing_features.ksymbol = true;
1764 pr_debug2("switching off ksymbol\n");
1765 goto fallback_missing_features;
1766 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1767 perf_missing_features.write_backward = true;
1768 pr_debug2("switching off write_backward\n");
1770 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1771 perf_missing_features.clockid_wrong = true;
1772 pr_debug2("switching off clockid\n");
1773 goto fallback_missing_features;
1774 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1775 perf_missing_features.clockid = true;
1776 pr_debug2("switching off use_clockid\n");
1777 goto fallback_missing_features;
1778 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1779 perf_missing_features.cloexec = true;
1780 pr_debug2("switching off cloexec flag\n");
1781 goto fallback_missing_features;
1782 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1783 perf_missing_features.mmap2 = true;
1784 pr_debug2("switching off mmap2\n");
1785 goto fallback_missing_features;
1786 } else if (!perf_missing_features.exclude_guest &&
1787 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1788 perf_missing_features.exclude_guest = true;
1789 pr_debug2("switching off exclude_guest, exclude_host\n");
1790 goto fallback_missing_features;
1791 } else if (!perf_missing_features.sample_id_all) {
1792 perf_missing_features.sample_id_all = true;
1793 pr_debug2("switching off sample_id_all\n");
1794 goto retry_sample_id;
1795 } else if (!perf_missing_features.lbr_flags &&
1796 (evsel->core.attr.branch_sample_type &
1797 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1798 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1799 perf_missing_features.lbr_flags = true;
1800 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1801 goto fallback_missing_features;
1802 } else if (!perf_missing_features.group_read &&
1803 evsel->core.attr.inherit &&
1804 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1805 perf_evsel__is_group_leader(evsel)) {
1806 perf_missing_features.group_read = true;
1807 pr_debug2("switching off group read\n");
1808 goto fallback_missing_features;
1812 threads->err_thread = thread;
1815 while (--thread >= 0) {
1816 close(FD(evsel, cpu, thread));
1817 FD(evsel, cpu, thread) = -1;
1820 } while (--cpu >= 0);
1824 void evsel__close(struct evsel *evsel)
1826 perf_evsel__close(&evsel->core);
1827 perf_evsel__free_id(&evsel->core);
1830 int perf_evsel__open_per_cpu(struct evsel *evsel,
1831 struct perf_cpu_map *cpus)
1833 return evsel__open(evsel, cpus, NULL);
1836 int perf_evsel__open_per_thread(struct evsel *evsel,
1837 struct perf_thread_map *threads)
1839 return evsel__open(evsel, NULL, threads);
1842 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1843 const union perf_event *event,
1844 struct perf_sample *sample)
1846 u64 type = evsel->core.attr.sample_type;
1847 const __u64 *array = event->sample.array;
1848 bool swapped = evsel->needs_swap;
1851 array += ((event->header.size -
1852 sizeof(event->header)) / sizeof(u64)) - 1;
1854 if (type & PERF_SAMPLE_IDENTIFIER) {
1855 sample->id = *array;
1859 if (type & PERF_SAMPLE_CPU) {
1862 /* undo swap of u64, then swap on individual u32s */
1863 u.val64 = bswap_64(u.val64);
1864 u.val32[0] = bswap_32(u.val32[0]);
1867 sample->cpu = u.val32[0];
1871 if (type & PERF_SAMPLE_STREAM_ID) {
1872 sample->stream_id = *array;
1876 if (type & PERF_SAMPLE_ID) {
1877 sample->id = *array;
1881 if (type & PERF_SAMPLE_TIME) {
1882 sample->time = *array;
1886 if (type & PERF_SAMPLE_TID) {
1889 /* undo swap of u64, then swap on individual u32s */
1890 u.val64 = bswap_64(u.val64);
1891 u.val32[0] = bswap_32(u.val32[0]);
1892 u.val32[1] = bswap_32(u.val32[1]);
1895 sample->pid = u.val32[0];
1896 sample->tid = u.val32[1];
1903 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1906 return size > max_size || offset + size > endp;
1909 #define OVERFLOW_CHECK(offset, size, max_size) \
1911 if (overflow(endp, (max_size), (offset), (size))) \
1915 #define OVERFLOW_CHECK_u64(offset) \
1916 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1919 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1922 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1923 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1924 * check the format does not go past the end of the event.
1926 if (sample_size + sizeof(event->header) > event->header.size)
1932 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1933 struct perf_sample *data)
1935 u64 type = evsel->core.attr.sample_type;
1936 bool swapped = evsel->needs_swap;
1938 u16 max_size = event->header.size;
1939 const void *endp = (void *)event + max_size;
1943 * used for cross-endian analysis. See git commit 65014ab3
1944 * for why this goofiness is needed.
1948 memset(data, 0, sizeof(*data));
1949 data->cpu = data->pid = data->tid = -1;
1950 data->stream_id = data->id = data->time = -1ULL;
1951 data->period = evsel->core.attr.sample_period;
1952 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1953 data->misc = event->header.misc;
1955 data->data_src = PERF_MEM_DATA_SRC_NONE;
1957 if (event->header.type != PERF_RECORD_SAMPLE) {
1958 if (!evsel->core.attr.sample_id_all)
1960 return perf_evsel__parse_id_sample(evsel, event, data);
1963 array = event->sample.array;
1965 if (perf_event__check_size(event, evsel->sample_size))
1968 if (type & PERF_SAMPLE_IDENTIFIER) {
1973 if (type & PERF_SAMPLE_IP) {
1978 if (type & PERF_SAMPLE_TID) {
1981 /* undo swap of u64, then swap on individual u32s */
1982 u.val64 = bswap_64(u.val64);
1983 u.val32[0] = bswap_32(u.val32[0]);
1984 u.val32[1] = bswap_32(u.val32[1]);
1987 data->pid = u.val32[0];
1988 data->tid = u.val32[1];
1992 if (type & PERF_SAMPLE_TIME) {
1993 data->time = *array;
1997 if (type & PERF_SAMPLE_ADDR) {
1998 data->addr = *array;
2002 if (type & PERF_SAMPLE_ID) {
2007 if (type & PERF_SAMPLE_STREAM_ID) {
2008 data->stream_id = *array;
2012 if (type & PERF_SAMPLE_CPU) {
2016 /* undo swap of u64, then swap on individual u32s */
2017 u.val64 = bswap_64(u.val64);
2018 u.val32[0] = bswap_32(u.val32[0]);
2021 data->cpu = u.val32[0];
2025 if (type & PERF_SAMPLE_PERIOD) {
2026 data->period = *array;
2030 if (type & PERF_SAMPLE_READ) {
2031 u64 read_format = evsel->core.attr.read_format;
2033 OVERFLOW_CHECK_u64(array);
2034 if (read_format & PERF_FORMAT_GROUP)
2035 data->read.group.nr = *array;
2037 data->read.one.value = *array;
2041 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2042 OVERFLOW_CHECK_u64(array);
2043 data->read.time_enabled = *array;
2047 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2048 OVERFLOW_CHECK_u64(array);
2049 data->read.time_running = *array;
2053 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2054 if (read_format & PERF_FORMAT_GROUP) {
2055 const u64 max_group_nr = UINT64_MAX /
2056 sizeof(struct sample_read_value);
2058 if (data->read.group.nr > max_group_nr)
2060 sz = data->read.group.nr *
2061 sizeof(struct sample_read_value);
2062 OVERFLOW_CHECK(array, sz, max_size);
2063 data->read.group.values =
2064 (struct sample_read_value *)array;
2065 array = (void *)array + sz;
2067 OVERFLOW_CHECK_u64(array);
2068 data->read.one.id = *array;
2073 if (evsel__has_callchain(evsel)) {
2074 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2076 OVERFLOW_CHECK_u64(array);
2077 data->callchain = (struct ip_callchain *)array++;
2078 if (data->callchain->nr > max_callchain_nr)
2080 sz = data->callchain->nr * sizeof(u64);
2081 OVERFLOW_CHECK(array, sz, max_size);
2082 array = (void *)array + sz;
2085 if (type & PERF_SAMPLE_RAW) {
2086 OVERFLOW_CHECK_u64(array);
2090 * Undo swap of u64, then swap on individual u32s,
2091 * get the size of the raw area and undo all of the
2092 * swap. The pevent interface handles endianity by
2096 u.val64 = bswap_64(u.val64);
2097 u.val32[0] = bswap_32(u.val32[0]);
2098 u.val32[1] = bswap_32(u.val32[1]);
2100 data->raw_size = u.val32[0];
2103 * The raw data is aligned on 64bits including the
2104 * u32 size, so it's safe to use mem_bswap_64.
2107 mem_bswap_64((void *) array, data->raw_size);
2109 array = (void *)array + sizeof(u32);
2111 OVERFLOW_CHECK(array, data->raw_size, max_size);
2112 data->raw_data = (void *)array;
2113 array = (void *)array + data->raw_size;
2116 if (type & PERF_SAMPLE_BRANCH_STACK) {
2117 const u64 max_branch_nr = UINT64_MAX /
2118 sizeof(struct branch_entry);
2120 OVERFLOW_CHECK_u64(array);
2121 data->branch_stack = (struct branch_stack *)array++;
2123 if (data->branch_stack->nr > max_branch_nr)
2125 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2126 OVERFLOW_CHECK(array, sz, max_size);
2127 array = (void *)array + sz;
2130 if (type & PERF_SAMPLE_REGS_USER) {
2131 OVERFLOW_CHECK_u64(array);
2132 data->user_regs.abi = *array;
2135 if (data->user_regs.abi) {
2136 u64 mask = evsel->core.attr.sample_regs_user;
2138 sz = hweight64(mask) * sizeof(u64);
2139 OVERFLOW_CHECK(array, sz, max_size);
2140 data->user_regs.mask = mask;
2141 data->user_regs.regs = (u64 *)array;
2142 array = (void *)array + sz;
2146 if (type & PERF_SAMPLE_STACK_USER) {
2147 OVERFLOW_CHECK_u64(array);
2150 data->user_stack.offset = ((char *)(array - 1)
2154 data->user_stack.size = 0;
2156 OVERFLOW_CHECK(array, sz, max_size);
2157 data->user_stack.data = (char *)array;
2158 array = (void *)array + sz;
2159 OVERFLOW_CHECK_u64(array);
2160 data->user_stack.size = *array++;
2161 if (WARN_ONCE(data->user_stack.size > sz,
2162 "user stack dump failure\n"))
2167 if (type & PERF_SAMPLE_WEIGHT) {
2168 OVERFLOW_CHECK_u64(array);
2169 data->weight = *array;
2173 if (type & PERF_SAMPLE_DATA_SRC) {
2174 OVERFLOW_CHECK_u64(array);
2175 data->data_src = *array;
2179 if (type & PERF_SAMPLE_TRANSACTION) {
2180 OVERFLOW_CHECK_u64(array);
2181 data->transaction = *array;
2185 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2186 if (type & PERF_SAMPLE_REGS_INTR) {
2187 OVERFLOW_CHECK_u64(array);
2188 data->intr_regs.abi = *array;
2191 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2192 u64 mask = evsel->core.attr.sample_regs_intr;
2194 sz = hweight64(mask) * sizeof(u64);
2195 OVERFLOW_CHECK(array, sz, max_size);
2196 data->intr_regs.mask = mask;
2197 data->intr_regs.regs = (u64 *)array;
2198 array = (void *)array + sz;
2202 data->phys_addr = 0;
2203 if (type & PERF_SAMPLE_PHYS_ADDR) {
2204 data->phys_addr = *array;
2211 int perf_evsel__parse_sample_timestamp(struct evsel *evsel,
2212 union perf_event *event,
2215 u64 type = evsel->core.attr.sample_type;
2218 if (!(type & PERF_SAMPLE_TIME))
2221 if (event->header.type != PERF_RECORD_SAMPLE) {
2222 struct perf_sample data = {
2226 if (!evsel->core.attr.sample_id_all)
2228 if (perf_evsel__parse_id_sample(evsel, event, &data))
2231 *timestamp = data.time;
2235 array = event->sample.array;
2237 if (perf_event__check_size(event, evsel->sample_size))
2240 if (type & PERF_SAMPLE_IDENTIFIER)
2243 if (type & PERF_SAMPLE_IP)
2246 if (type & PERF_SAMPLE_TID)
2249 if (type & PERF_SAMPLE_TIME)
2250 *timestamp = *array;
2255 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name)
2257 return tep_find_field(evsel->tp_format, name);
2260 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample,
2263 struct tep_format_field *field = perf_evsel__field(evsel, name);
2269 offset = field->offset;
2271 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2272 offset = *(int *)(sample->raw_data + field->offset);
2276 return sample->raw_data + offset;
2279 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2283 void *ptr = sample->raw_data + field->offset;
2285 switch (field->size) {
2289 value = *(u16 *)ptr;
2292 value = *(u32 *)ptr;
2295 memcpy(&value, ptr, sizeof(u64));
2304 switch (field->size) {
2306 return bswap_16(value);
2308 return bswap_32(value);
2310 return bswap_64(value);
2318 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample,
2321 struct tep_format_field *field = perf_evsel__field(evsel, name);
2326 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2329 bool perf_evsel__fallback(struct evsel *evsel, int err,
2330 char *msg, size_t msgsize)
2334 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2335 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2336 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2338 * If it's cycles then fall back to hrtimer based
2339 * cpu-clock-tick sw counter, which is always available even if
2342 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2345 scnprintf(msg, msgsize, "%s",
2346 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2348 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2349 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2351 zfree(&evsel->name);
2353 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2354 (paranoid = perf_event_paranoid()) > 1) {
2355 const char *name = perf_evsel__name(evsel);
2357 const char *sep = ":";
2359 /* Is there already the separator in the name. */
2360 if (strchr(name, '/') ||
2364 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2369 evsel->name = new_name;
2370 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2371 "to fall back to excluding kernel and hypervisor "
2372 " samples", paranoid);
2373 evsel->core.attr.exclude_kernel = 1;
2374 evsel->core.attr.exclude_hv = 1;
2382 static bool find_process(const char *name)
2384 size_t len = strlen(name);
2389 dir = opendir(procfs__mountpoint());
2393 /* Walk through the directory. */
2394 while (ret && (d = readdir(dir)) != NULL) {
2395 char path[PATH_MAX];
2399 if ((d->d_type != DT_DIR) ||
2400 !strcmp(".", d->d_name) ||
2401 !strcmp("..", d->d_name))
2404 scnprintf(path, sizeof(path), "%s/%s/comm",
2405 procfs__mountpoint(), d->d_name);
2407 if (filename__read_str(path, &data, &size))
2410 ret = strncmp(name, data, len);
2415 return ret ? false : true;
2418 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target,
2419 int err, char *msg, size_t size)
2421 char sbuf[STRERR_BUFSIZE];
2428 printed = scnprintf(msg, size,
2429 "No permission to enable %s event.\n\n",
2430 perf_evsel__name(evsel));
2432 return scnprintf(msg + printed, size - printed,
2433 "You may not have permission to collect %sstats.\n\n"
2434 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2435 "which controls use of the performance events system by\n"
2436 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2437 "The current value is %d:\n\n"
2438 " -1: Allow use of (almost) all events by all users\n"
2439 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2440 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2441 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2442 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2443 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2444 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2445 " kernel.perf_event_paranoid = -1\n" ,
2446 target->system_wide ? "system-wide " : "",
2447 perf_event_paranoid());
2449 return scnprintf(msg, size, "The %s event is not supported.",
2450 perf_evsel__name(evsel));
2452 return scnprintf(msg, size, "%s",
2453 "Too many events are opened.\n"
2454 "Probably the maximum number of open file descriptors has been reached.\n"
2455 "Hint: Try again after reducing the number of events.\n"
2456 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2458 if (evsel__has_callchain(evsel) &&
2459 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2460 return scnprintf(msg, size,
2461 "Not enough memory to setup event with callchain.\n"
2462 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2463 "Hint: Current value: %d", sysctl__max_stack());
2466 if (target->cpu_list)
2467 return scnprintf(msg, size, "%s",
2468 "No such device - did you specify an out-of-range profile CPU?");
2471 if (evsel->core.attr.sample_period != 0)
2472 return scnprintf(msg, size,
2473 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2474 perf_evsel__name(evsel));
2475 if (evsel->core.attr.precise_ip)
2476 return scnprintf(msg, size, "%s",
2477 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2478 #if defined(__i386__) || defined(__x86_64__)
2479 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2480 return scnprintf(msg, size, "%s",
2481 "No hardware sampling interrupt available.\n");
2485 if (find_process("oprofiled"))
2486 return scnprintf(msg, size,
2487 "The PMU counters are busy/taken by another profiler.\n"
2488 "We found oprofile daemon running, please stop it and try again.");
2491 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2492 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2493 if (perf_missing_features.clockid)
2494 return scnprintf(msg, size, "clockid feature not supported.");
2495 if (perf_missing_features.clockid_wrong)
2496 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2497 if (perf_missing_features.aux_output)
2498 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2504 return scnprintf(msg, size,
2505 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2506 "/bin/dmesg | grep -i perf may provide additional information.\n",
2507 err, str_error_r(err, sbuf, sizeof(sbuf)),
2508 perf_evsel__name(evsel));
2511 struct perf_env *perf_evsel__env(struct evsel *evsel)
2513 if (evsel && evsel->evlist)
2514 return evsel->evlist->env;
2518 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2522 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2523 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2525 int fd = FD(evsel, cpu, thread);
2527 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2528 cpu, thread, fd) < 0)
2536 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2538 struct perf_cpu_map *cpus = evsel->core.cpus;
2539 struct perf_thread_map *threads = evsel->core.threads;
2541 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2544 return store_evsel_ids(evsel, evlist);