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->collect_stat = false;
258 evsel->pmu_name = NULL;
261 struct evsel *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 (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 (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 *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 *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 *evsel__hw_names[PERF_COUNT_HW_MAX] = {
382 "stalled-cycles-frontend",
383 "stalled-cycles-backend",
387 static const char *__evsel__hw_name(u64 config)
389 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
390 return 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 evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
433 int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
434 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
437 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
450 static const char *__evsel__sw_name(u64 config)
452 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
453 return evsel__sw_names[config];
454 return "unknown-software";
457 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
459 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
460 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
463 static int __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 evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
483 struct perf_event_attr *attr = &evsel->core.attr;
484 int r = __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 *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
489 { "L1-dcache", "l1-d", "l1d", "L1-data", },
490 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
492 { "dTLB", "d-tlb", "Data-TLB", },
493 { "iTLB", "i-tlb", "Instruction-TLB", },
494 { "branch", "branches", "bpu", "btb", "bpc", },
498 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
499 { "load", "loads", "read", },
500 { "store", "stores", "write", },
501 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
504 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
505 { "refs", "Reference", "ops", "access", },
506 { "misses", "miss", },
509 #define C(x) PERF_COUNT_HW_CACHE_##x
510 #define CACHE_READ (1 << C(OP_READ))
511 #define CACHE_WRITE (1 << C(OP_WRITE))
512 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
513 #define COP(x) (1 << x)
516 * cache operartion stat
517 * L1I : Read and prefetch only
518 * ITLB and BPU : Read-only
520 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
521 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
522 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
523 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
524 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
525 [C(ITLB)] = (CACHE_READ),
526 [C(BPU)] = (CACHE_READ),
527 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
530 bool evsel__is_cache_op_valid(u8 type, u8 op)
532 if (evsel__hw_cache_stat[type] & COP(op))
533 return true; /* valid */
535 return false; /* invalid */
538 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
541 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
542 evsel__hw_cache_op[op][0],
543 evsel__hw_cache_result[result][0]);
546 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
547 evsel__hw_cache_op[op][1]);
550 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
552 u8 op, result, type = (config >> 0) & 0xff;
553 const char *err = "unknown-ext-hardware-cache-type";
555 if (type >= PERF_COUNT_HW_CACHE_MAX)
558 op = (config >> 8) & 0xff;
559 err = "unknown-ext-hardware-cache-op";
560 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
563 result = (config >> 16) & 0xff;
564 err = "unknown-ext-hardware-cache-result";
565 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
568 err = "invalid-cache";
569 if (!evsel__is_cache_op_valid(type, op))
572 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
574 return scnprintf(bf, size, "%s", err);
577 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
579 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
580 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
583 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
585 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
586 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
589 static int evsel__tool_name(char *bf, size_t size)
591 int ret = scnprintf(bf, size, "duration_time");
595 const char *evsel__name(struct evsel *evsel)
605 switch (evsel->core.attr.type) {
607 evsel__raw_name(evsel, bf, sizeof(bf));
610 case PERF_TYPE_HARDWARE:
611 evsel__hw_name(evsel, bf, sizeof(bf));
614 case PERF_TYPE_HW_CACHE:
615 evsel__hw_cache_name(evsel, bf, sizeof(bf));
618 case PERF_TYPE_SOFTWARE:
619 if (evsel->tool_event)
620 evsel__tool_name(bf, sizeof(bf));
622 evsel__sw_name(evsel, bf, sizeof(bf));
625 case PERF_TYPE_TRACEPOINT:
626 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
629 case PERF_TYPE_BREAKPOINT:
630 evsel__bp_name(evsel, bf, sizeof(bf));
634 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
635 evsel->core.attr.type);
639 evsel->name = strdup(bf);
647 const char *evsel__group_name(struct evsel *evsel)
649 return evsel->group_name ?: "anon group";
653 * Returns the group details for the specified leader,
654 * with following rules.
656 * For record -e '{cycles,instructions}'
657 * 'anon group { cycles:u, instructions:u }'
659 * For record -e 'cycles,instructions' and report --group
660 * 'cycles:u, instructions:u'
662 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
666 const char *group_name = evsel__group_name(evsel);
668 if (!evsel->forced_leader)
669 ret = scnprintf(buf, size, "%s { ", group_name);
671 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
673 for_each_group_member(pos, evsel)
674 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
676 if (!evsel->forced_leader)
677 ret += scnprintf(buf + ret, size - ret, " }");
682 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
683 struct callchain_param *param)
685 bool function = evsel__is_function_event(evsel);
686 struct perf_event_attr *attr = &evsel->core.attr;
688 evsel__set_sample_bit(evsel, CALLCHAIN);
690 attr->sample_max_stack = param->max_stack;
692 if (opts->kernel_callchains)
693 attr->exclude_callchain_user = 1;
694 if (opts->user_callchains)
695 attr->exclude_callchain_kernel = 1;
696 if (param->record_mode == CALLCHAIN_LBR) {
697 if (!opts->branch_stack) {
698 if (attr->exclude_user) {
699 pr_warning("LBR callstack option is only available "
700 "to get user callchain information. "
701 "Falling back to framepointers.\n");
703 evsel__set_sample_bit(evsel, BRANCH_STACK);
704 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
705 PERF_SAMPLE_BRANCH_CALL_STACK |
706 PERF_SAMPLE_BRANCH_NO_CYCLES |
707 PERF_SAMPLE_BRANCH_NO_FLAGS |
708 PERF_SAMPLE_BRANCH_HW_INDEX;
711 pr_warning("Cannot use LBR callstack with branch stack. "
712 "Falling back to framepointers.\n");
715 if (param->record_mode == CALLCHAIN_DWARF) {
717 evsel__set_sample_bit(evsel, REGS_USER);
718 evsel__set_sample_bit(evsel, STACK_USER);
719 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
720 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
721 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
722 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
723 "so the minimal registers set (IP, SP) is explicitly forced.\n");
725 attr->sample_regs_user |= PERF_REGS_MASK;
727 attr->sample_stack_user = param->dump_size;
728 attr->exclude_callchain_user = 1;
730 pr_info("Cannot use DWARF unwind for function trace event,"
731 " falling back to framepointers.\n");
736 pr_info("Disabling user space callchains for function trace event.\n");
737 attr->exclude_callchain_user = 1;
741 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
742 struct callchain_param *param)
745 return __evsel__config_callchain(evsel, opts, param);
749 perf_evsel__reset_callgraph(struct evsel *evsel,
750 struct callchain_param *param)
752 struct perf_event_attr *attr = &evsel->core.attr;
754 evsel__reset_sample_bit(evsel, CALLCHAIN);
755 if (param->record_mode == CALLCHAIN_LBR) {
756 evsel__reset_sample_bit(evsel, BRANCH_STACK);
757 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
758 PERF_SAMPLE_BRANCH_CALL_STACK |
759 PERF_SAMPLE_BRANCH_HW_INDEX);
761 if (param->record_mode == CALLCHAIN_DWARF) {
762 evsel__reset_sample_bit(evsel, REGS_USER);
763 evsel__reset_sample_bit(evsel, STACK_USER);
767 static void evsel__apply_config_terms(struct evsel *evsel,
768 struct record_opts *opts, bool track)
770 struct evsel_config_term *term;
771 struct list_head *config_terms = &evsel->config_terms;
772 struct perf_event_attr *attr = &evsel->core.attr;
773 /* callgraph default */
774 struct callchain_param param = {
775 .record_mode = callchain_param.record_mode,
779 const char *callgraph_buf = NULL;
781 list_for_each_entry(term, config_terms, list) {
782 switch (term->type) {
783 case EVSEL__CONFIG_TERM_PERIOD:
784 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
785 attr->sample_period = term->val.period;
787 evsel__reset_sample_bit(evsel, PERIOD);
790 case EVSEL__CONFIG_TERM_FREQ:
791 if (!(term->weak && opts->user_freq != UINT_MAX)) {
792 attr->sample_freq = term->val.freq;
794 evsel__set_sample_bit(evsel, PERIOD);
797 case EVSEL__CONFIG_TERM_TIME:
799 evsel__set_sample_bit(evsel, TIME);
801 evsel__reset_sample_bit(evsel, TIME);
803 case EVSEL__CONFIG_TERM_CALLGRAPH:
804 callgraph_buf = term->val.str;
806 case EVSEL__CONFIG_TERM_BRANCH:
807 if (term->val.str && strcmp(term->val.str, "no")) {
808 evsel__set_sample_bit(evsel, BRANCH_STACK);
809 parse_branch_str(term->val.str,
810 &attr->branch_sample_type);
812 evsel__reset_sample_bit(evsel, BRANCH_STACK);
814 case EVSEL__CONFIG_TERM_STACK_USER:
815 dump_size = term->val.stack_user;
817 case EVSEL__CONFIG_TERM_MAX_STACK:
818 max_stack = term->val.max_stack;
820 case EVSEL__CONFIG_TERM_MAX_EVENTS:
821 evsel->max_events = term->val.max_events;
823 case EVSEL__CONFIG_TERM_INHERIT:
825 * attr->inherit should has already been set by
826 * evsel__config. If user explicitly set
827 * inherit using config terms, override global
828 * opt->no_inherit setting.
830 attr->inherit = term->val.inherit ? 1 : 0;
832 case EVSEL__CONFIG_TERM_OVERWRITE:
833 attr->write_backward = term->val.overwrite ? 1 : 0;
835 case EVSEL__CONFIG_TERM_DRV_CFG:
837 case EVSEL__CONFIG_TERM_PERCORE:
839 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
840 attr->aux_output = term->val.aux_output ? 1 : 0;
842 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
843 /* Already applied by auxtrace */
845 case EVSEL__CONFIG_TERM_CFG_CHG:
852 /* User explicitly set per-event callgraph, clear the old setting and reset. */
853 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
854 bool sample_address = false;
857 param.max_stack = max_stack;
858 if (callgraph_buf == NULL)
859 callgraph_buf = "fp";
862 /* parse callgraph parameters */
863 if (callgraph_buf != NULL) {
864 if (!strcmp(callgraph_buf, "no")) {
865 param.enabled = false;
866 param.record_mode = CALLCHAIN_NONE;
868 param.enabled = true;
869 if (parse_callchain_record(callgraph_buf, ¶m)) {
870 pr_err("per-event callgraph setting for %s failed. "
871 "Apply callgraph global setting for it\n",
875 if (param.record_mode == CALLCHAIN_DWARF)
876 sample_address = true;
880 dump_size = round_up(dump_size, sizeof(u64));
881 param.dump_size = dump_size;
884 /* If global callgraph set, clear it */
885 if (callchain_param.enabled)
886 perf_evsel__reset_callgraph(evsel, &callchain_param);
888 /* set perf-event callgraph */
890 if (sample_address) {
891 evsel__set_sample_bit(evsel, ADDR);
892 evsel__set_sample_bit(evsel, DATA_SRC);
893 evsel->core.attr.mmap_data = track;
895 evsel__config_callchain(evsel, opts, ¶m);
900 static bool is_dummy_event(struct evsel *evsel)
902 return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
903 (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
906 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
908 struct evsel_config_term *term, *found_term = NULL;
910 list_for_each_entry(term, &evsel->config_terms, list) {
911 if (term->type == type)
919 * The enable_on_exec/disabled value strategy:
921 * 1) For any type of traced program:
922 * - all independent events and group leaders are disabled
923 * - all group members are enabled
925 * Group members are ruled by group leaders. They need to
926 * be enabled, because the group scheduling relies on that.
928 * 2) For traced programs executed by perf:
929 * - all independent events and group leaders have
931 * - we don't specifically enable or disable any event during
934 * Independent events and group leaders are initially disabled
935 * and get enabled by exec. Group members are ruled by group
936 * leaders as stated in 1).
938 * 3) For traced programs attached by perf (pid/tid):
939 * - we specifically enable or disable all events during
942 * When attaching events to already running traced we
943 * enable/disable events specifically, as there's no
944 * initial traced exec call.
946 void evsel__config(struct evsel *evsel, struct record_opts *opts,
947 struct callchain_param *callchain)
949 struct evsel *leader = evsel->leader;
950 struct perf_event_attr *attr = &evsel->core.attr;
951 int track = evsel->tracking;
952 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
954 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
955 attr->inherit = !opts->no_inherit;
956 attr->write_backward = opts->overwrite ? 1 : 0;
958 evsel__set_sample_bit(evsel, IP);
959 evsel__set_sample_bit(evsel, TID);
961 if (evsel->sample_read) {
962 evsel__set_sample_bit(evsel, READ);
965 * We need ID even in case of single event, because
966 * PERF_SAMPLE_READ process ID specific data.
968 evsel__set_sample_id(evsel, false);
971 * Apply group format only if we belong to group
972 * with more than one members.
974 if (leader->core.nr_members > 1) {
975 attr->read_format |= PERF_FORMAT_GROUP;
981 * We default some events to have a default interval. But keep
982 * it a weak assumption overridable by the user.
984 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
985 opts->user_interval != ULLONG_MAX)) {
987 evsel__set_sample_bit(evsel, PERIOD);
989 attr->sample_freq = opts->freq;
991 attr->sample_period = opts->default_interval;
995 if (opts->no_samples)
996 attr->sample_freq = 0;
998 if (opts->inherit_stat) {
999 evsel->core.attr.read_format |=
1000 PERF_FORMAT_TOTAL_TIME_ENABLED |
1001 PERF_FORMAT_TOTAL_TIME_RUNNING |
1003 attr->inherit_stat = 1;
1006 if (opts->sample_address) {
1007 evsel__set_sample_bit(evsel, ADDR);
1008 attr->mmap_data = track;
1012 * We don't allow user space callchains for function trace
1013 * event, due to issues with page faults while tracing page
1014 * fault handler and its overall trickiness nature.
1016 if (evsel__is_function_event(evsel))
1017 evsel->core.attr.exclude_callchain_user = 1;
1019 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1020 evsel__config_callchain(evsel, opts, callchain);
1022 if (opts->sample_intr_regs) {
1023 attr->sample_regs_intr = opts->sample_intr_regs;
1024 evsel__set_sample_bit(evsel, REGS_INTR);
1027 if (opts->sample_user_regs) {
1028 attr->sample_regs_user |= opts->sample_user_regs;
1029 evsel__set_sample_bit(evsel, REGS_USER);
1032 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1033 evsel__set_sample_bit(evsel, CPU);
1036 * When the user explicitly disabled time don't force it here.
1038 if (opts->sample_time &&
1039 (!perf_missing_features.sample_id_all &&
1040 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1041 opts->sample_time_set)))
1042 evsel__set_sample_bit(evsel, TIME);
1044 if (opts->raw_samples && !evsel->no_aux_samples) {
1045 evsel__set_sample_bit(evsel, TIME);
1046 evsel__set_sample_bit(evsel, RAW);
1047 evsel__set_sample_bit(evsel, CPU);
1050 if (opts->sample_address)
1051 evsel__set_sample_bit(evsel, DATA_SRC);
1053 if (opts->sample_phys_addr)
1054 evsel__set_sample_bit(evsel, PHYS_ADDR);
1056 if (opts->no_buffering) {
1057 attr->watermark = 0;
1058 attr->wakeup_events = 1;
1060 if (opts->branch_stack && !evsel->no_aux_samples) {
1061 evsel__set_sample_bit(evsel, BRANCH_STACK);
1062 attr->branch_sample_type = opts->branch_stack;
1065 if (opts->sample_weight)
1066 evsel__set_sample_bit(evsel, WEIGHT);
1070 attr->mmap2 = track && !perf_missing_features.mmap2;
1072 attr->ksymbol = track && !perf_missing_features.ksymbol;
1073 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1075 if (opts->record_namespaces)
1076 attr->namespaces = track;
1078 if (opts->record_cgroup) {
1079 attr->cgroup = track && !perf_missing_features.cgroup;
1080 evsel__set_sample_bit(evsel, CGROUP);
1083 if (opts->record_switch_events)
1084 attr->context_switch = track;
1086 if (opts->sample_transaction)
1087 evsel__set_sample_bit(evsel, TRANSACTION);
1089 if (opts->running_time) {
1090 evsel->core.attr.read_format |=
1091 PERF_FORMAT_TOTAL_TIME_ENABLED |
1092 PERF_FORMAT_TOTAL_TIME_RUNNING;
1096 * XXX see the function comment above
1098 * Disabling only independent events or group leaders,
1099 * keeping group members enabled.
1101 if (evsel__is_group_leader(evsel))
1105 * Setting enable_on_exec for independent events and
1106 * group leaders for traced executed by perf.
1108 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1109 !opts->initial_delay)
1110 attr->enable_on_exec = 1;
1112 if (evsel->immediate) {
1114 attr->enable_on_exec = 0;
1117 clockid = opts->clockid;
1118 if (opts->use_clockid) {
1119 attr->use_clockid = 1;
1120 attr->clockid = opts->clockid;
1123 if (evsel->precise_max)
1124 attr->precise_ip = 3;
1126 if (opts->all_user) {
1127 attr->exclude_kernel = 1;
1128 attr->exclude_user = 0;
1131 if (opts->all_kernel) {
1132 attr->exclude_kernel = 0;
1133 attr->exclude_user = 1;
1136 if (evsel->core.own_cpus || evsel->unit)
1137 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1140 * Apply event specific term settings,
1141 * it overloads any global configuration.
1143 evsel__apply_config_terms(evsel, opts, track);
1145 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1147 /* The --period option takes the precedence. */
1148 if (opts->period_set) {
1150 evsel__set_sample_bit(evsel, PERIOD);
1152 evsel__reset_sample_bit(evsel, PERIOD);
1156 * For initial_delay, a dummy event is added implicitly.
1157 * The software event will trigger -EOPNOTSUPP error out,
1158 * if BRANCH_STACK bit is set.
1160 if (opts->initial_delay && is_dummy_event(evsel))
1161 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1164 int evsel__set_filter(struct evsel *evsel, const char *filter)
1166 char *new_filter = strdup(filter);
1168 if (new_filter != NULL) {
1169 free(evsel->filter);
1170 evsel->filter = new_filter;
1177 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1181 if (evsel->filter == NULL)
1182 return evsel__set_filter(evsel, filter);
1184 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1185 free(evsel->filter);
1186 evsel->filter = new_filter;
1193 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1195 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1198 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1200 return evsel__append_filter(evsel, "%s,%s", filter);
1203 /* Caller has to clear disabled after going through all CPUs. */
1204 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1206 return perf_evsel__enable_cpu(&evsel->core, cpu);
1209 int evsel__enable(struct evsel *evsel)
1211 int err = perf_evsel__enable(&evsel->core);
1214 evsel->disabled = false;
1218 /* Caller has to set disabled after going through all CPUs. */
1219 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1221 return perf_evsel__disable_cpu(&evsel->core, cpu);
1224 int evsel__disable(struct evsel *evsel)
1226 int err = perf_evsel__disable(&evsel->core);
1228 * We mark it disabled here so that tools that disable a event can
1229 * ignore events after they disable it. I.e. the ring buffer may have
1230 * already a few more events queued up before the kernel got the stop
1234 evsel->disabled = true;
1239 static void evsel__free_config_terms(struct evsel *evsel)
1241 struct evsel_config_term *term, *h;
1243 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1244 list_del_init(&term->list);
1246 zfree(&term->val.str);
1251 void evsel__exit(struct evsel *evsel)
1253 assert(list_empty(&evsel->core.node));
1254 assert(evsel->evlist == NULL);
1255 evsel__free_counts(evsel);
1256 perf_evsel__free_fd(&evsel->core);
1257 perf_evsel__free_id(&evsel->core);
1258 evsel__free_config_terms(evsel);
1259 cgroup__put(evsel->cgrp);
1260 perf_cpu_map__put(evsel->core.cpus);
1261 perf_cpu_map__put(evsel->core.own_cpus);
1262 perf_thread_map__put(evsel->core.threads);
1263 zfree(&evsel->group_name);
1264 zfree(&evsel->name);
1265 zfree(&evsel->pmu_name);
1266 perf_evsel__object.fini(evsel);
1269 void evsel__delete(struct evsel *evsel)
1275 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1276 struct perf_counts_values *count)
1278 struct perf_counts_values tmp;
1280 if (!evsel->prev_raw_counts)
1284 tmp = evsel->prev_raw_counts->aggr;
1285 evsel->prev_raw_counts->aggr = *count;
1287 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1288 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1291 count->val = count->val - tmp.val;
1292 count->ena = count->ena - tmp.ena;
1293 count->run = count->run - tmp.run;
1296 void perf_counts_values__scale(struct perf_counts_values *count,
1297 bool scale, s8 *pscaled)
1302 if (count->run == 0) {
1305 } else if (count->run < count->ena) {
1307 count->val = (u64)((double) count->val * count->ena / count->run);
1315 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1317 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1319 return perf_evsel__read(&evsel->core, cpu, thread, count);
1323 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1324 u64 val, u64 ena, u64 run)
1326 struct perf_counts_values *count;
1328 count = perf_counts(counter->counts, cpu, thread);
1334 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1338 perf_evsel__process_group_data(struct evsel *leader,
1339 int cpu, int thread, u64 *data)
1341 u64 read_format = leader->core.attr.read_format;
1342 struct sample_read_value *v;
1343 u64 nr, ena = 0, run = 0, i;
1347 if (nr != (u64) leader->core.nr_members)
1350 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1353 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1356 v = (struct sample_read_value *) data;
1358 perf_evsel__set_count(leader, cpu, thread,
1359 v[0].value, ena, run);
1361 for (i = 1; i < nr; i++) {
1362 struct evsel *counter;
1364 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1368 perf_evsel__set_count(counter, cpu, thread,
1369 v[i].value, ena, run);
1375 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1377 struct perf_stat_evsel *ps = leader->stats;
1378 u64 read_format = leader->core.attr.read_format;
1379 int size = perf_evsel__read_size(&leader->core);
1380 u64 *data = ps->group_data;
1382 if (!(read_format & PERF_FORMAT_ID))
1385 if (!evsel__is_group_leader(leader))
1389 data = zalloc(size);
1393 ps->group_data = data;
1396 if (FD(leader, cpu, thread) < 0)
1399 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1402 return perf_evsel__process_group_data(leader, cpu, thread, data);
1405 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1407 u64 read_format = evsel->core.attr.read_format;
1409 if (read_format & PERF_FORMAT_GROUP)
1410 return evsel__read_group(evsel, cpu, thread);
1412 return evsel__read_one(evsel, cpu, thread);
1415 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1417 struct perf_counts_values count;
1418 size_t nv = scale ? 3 : 1;
1420 if (FD(evsel, cpu, thread) < 0)
1423 if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1426 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1429 evsel__compute_deltas(evsel, cpu, thread, &count);
1430 perf_counts_values__scale(&count, scale, NULL);
1431 *perf_counts(evsel->counts, cpu, thread) = count;
1435 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1437 struct evsel *leader = evsel->leader;
1440 if (evsel__is_group_leader(evsel))
1444 * Leader must be already processed/open,
1445 * if not it's a bug.
1447 BUG_ON(!leader->core.fd);
1449 fd = FD(leader, cpu, thread);
1455 static void perf_evsel__remove_fd(struct evsel *pos,
1456 int nr_cpus, int nr_threads,
1459 for (int cpu = 0; cpu < nr_cpus; cpu++)
1460 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1461 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1464 static int update_fds(struct evsel *evsel,
1465 int nr_cpus, int cpu_idx,
1466 int nr_threads, int thread_idx)
1470 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1473 evlist__for_each_entry(evsel->evlist, pos) {
1474 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1476 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1479 * Since fds for next evsel has not been created,
1480 * there is no need to iterate whole event list.
1488 static bool ignore_missing_thread(struct evsel *evsel,
1489 int nr_cpus, int cpu,
1490 struct perf_thread_map *threads,
1491 int thread, int err)
1493 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1495 if (!evsel->ignore_missing_thread)
1498 /* The system wide setup does not work with threads. */
1499 if (evsel->core.system_wide)
1502 /* The -ESRCH is perf event syscall errno for pid's not found. */
1506 /* If there's only one thread, let it fail. */
1507 if (threads->nr == 1)
1511 * We should remove fd for missing_thread first
1512 * because thread_map__remove() will decrease threads->nr.
1514 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1517 if (thread_map__remove(threads, thread))
1520 pr_warning("WARNING: Ignored open failure for pid %d\n",
1525 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1526 void *priv __maybe_unused)
1528 return fprintf(fp, " %-32s %s\n", name, val);
1531 static void display_attr(struct perf_event_attr *attr)
1533 if (verbose >= 2 || debug_peo_args) {
1534 fprintf(stderr, "%.60s\n", graph_dotted_line);
1535 fprintf(stderr, "perf_event_attr:\n");
1536 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1537 fprintf(stderr, "%.60s\n", graph_dotted_line);
1541 static int perf_event_open(struct evsel *evsel,
1542 pid_t pid, int cpu, int group_fd,
1543 unsigned long flags)
1545 int precise_ip = evsel->core.attr.precise_ip;
1549 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1550 pid, cpu, group_fd, flags);
1552 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1556 /* Do not try less precise if not requested. */
1557 if (!evsel->precise_max)
1561 * We tried all the precise_ip values, and it's
1562 * still failing, so leave it to standard fallback.
1564 if (!evsel->core.attr.precise_ip) {
1565 evsel->core.attr.precise_ip = precise_ip;
1569 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1570 evsel->core.attr.precise_ip--;
1571 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1572 display_attr(&evsel->core.attr);
1578 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1579 struct perf_thread_map *threads,
1580 int start_cpu, int end_cpu)
1582 int cpu, thread, nthreads;
1583 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1584 int pid = -1, err, old_errno;
1585 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1587 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1588 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1592 static struct perf_cpu_map *empty_cpu_map;
1594 if (empty_cpu_map == NULL) {
1595 empty_cpu_map = perf_cpu_map__dummy_new();
1596 if (empty_cpu_map == NULL)
1600 cpus = empty_cpu_map;
1603 if (threads == NULL) {
1604 static struct perf_thread_map *empty_thread_map;
1606 if (empty_thread_map == NULL) {
1607 empty_thread_map = thread_map__new_by_tid(-1);
1608 if (empty_thread_map == NULL)
1612 threads = empty_thread_map;
1615 if (evsel->core.system_wide)
1618 nthreads = threads->nr;
1620 if (evsel->core.fd == NULL &&
1621 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1625 flags |= PERF_FLAG_PID_CGROUP;
1626 pid = evsel->cgrp->fd;
1629 fallback_missing_features:
1630 if (perf_missing_features.clockid_wrong)
1631 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1632 if (perf_missing_features.clockid) {
1633 evsel->core.attr.use_clockid = 0;
1634 evsel->core.attr.clockid = 0;
1636 if (perf_missing_features.cloexec)
1637 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1638 if (perf_missing_features.mmap2)
1639 evsel->core.attr.mmap2 = 0;
1640 if (perf_missing_features.exclude_guest)
1641 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1642 if (perf_missing_features.lbr_flags)
1643 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1644 PERF_SAMPLE_BRANCH_NO_CYCLES);
1645 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1646 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1647 if (perf_missing_features.ksymbol)
1648 evsel->core.attr.ksymbol = 0;
1649 if (perf_missing_features.bpf)
1650 evsel->core.attr.bpf_event = 0;
1651 if (perf_missing_features.branch_hw_idx)
1652 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1654 if (perf_missing_features.sample_id_all)
1655 evsel->core.attr.sample_id_all = 0;
1657 display_attr(&evsel->core.attr);
1659 for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1661 for (thread = 0; thread < nthreads; thread++) {
1664 if (!evsel->cgrp && !evsel->core.system_wide)
1665 pid = perf_thread_map__pid(threads, thread);
1667 group_fd = get_group_fd(evsel, cpu, thread);
1671 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1674 FD(evsel, cpu, thread) = fd;
1679 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1681 * We just removed 1 thread, so take a step
1682 * back on thread index and lower the upper
1688 /* ... and pretend like nothing have happened. */
1693 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1698 pr_debug2_peo(" = %d\n", fd);
1700 if (evsel->bpf_fd >= 0) {
1702 int bpf_fd = evsel->bpf_fd;
1705 PERF_EVENT_IOC_SET_BPF,
1707 if (err && errno != EEXIST) {
1708 pr_err("failed to attach bpf fd %d: %s\n",
1709 bpf_fd, strerror(errno));
1715 set_rlimit = NO_CHANGE;
1718 * If we succeeded but had to kill clockid, fail and
1719 * have evsel__open_strerror() print us a nice error.
1721 if (perf_missing_features.clockid ||
1722 perf_missing_features.clockid_wrong) {
1733 * perf stat needs between 5 and 22 fds per CPU. When we run out
1734 * of them try to increase the limits.
1736 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1740 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1741 if (set_rlimit == NO_CHANGE)
1742 l.rlim_cur = l.rlim_max;
1744 l.rlim_cur = l.rlim_max + 1000;
1745 l.rlim_max = l.rlim_cur;
1747 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1756 if (err != -EINVAL || cpu > 0 || thread > 0)
1760 * Must probe features in the order they were added to the
1761 * perf_event_attr interface.
1763 if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1764 perf_missing_features.cgroup = true;
1765 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1767 } else if (!perf_missing_features.branch_hw_idx &&
1768 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1769 perf_missing_features.branch_hw_idx = true;
1770 pr_debug2("switching off branch HW index support\n");
1771 goto fallback_missing_features;
1772 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1773 perf_missing_features.aux_output = true;
1774 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1776 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1777 perf_missing_features.bpf = true;
1778 pr_debug2_peo("switching off bpf_event\n");
1779 goto fallback_missing_features;
1780 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1781 perf_missing_features.ksymbol = true;
1782 pr_debug2_peo("switching off ksymbol\n");
1783 goto fallback_missing_features;
1784 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1785 perf_missing_features.write_backward = true;
1786 pr_debug2_peo("switching off write_backward\n");
1788 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1789 perf_missing_features.clockid_wrong = true;
1790 pr_debug2_peo("switching off clockid\n");
1791 goto fallback_missing_features;
1792 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1793 perf_missing_features.clockid = true;
1794 pr_debug2_peo("switching off use_clockid\n");
1795 goto fallback_missing_features;
1796 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1797 perf_missing_features.cloexec = true;
1798 pr_debug2_peo("switching off cloexec flag\n");
1799 goto fallback_missing_features;
1800 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1801 perf_missing_features.mmap2 = true;
1802 pr_debug2_peo("switching off mmap2\n");
1803 goto fallback_missing_features;
1804 } else if (!perf_missing_features.exclude_guest &&
1805 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1806 perf_missing_features.exclude_guest = true;
1807 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1808 goto fallback_missing_features;
1809 } else if (!perf_missing_features.sample_id_all) {
1810 perf_missing_features.sample_id_all = true;
1811 pr_debug2_peo("switching off sample_id_all\n");
1812 goto retry_sample_id;
1813 } else if (!perf_missing_features.lbr_flags &&
1814 (evsel->core.attr.branch_sample_type &
1815 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1816 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1817 perf_missing_features.lbr_flags = true;
1818 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1819 goto fallback_missing_features;
1820 } else if (!perf_missing_features.group_read &&
1821 evsel->core.attr.inherit &&
1822 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1823 evsel__is_group_leader(evsel)) {
1824 perf_missing_features.group_read = true;
1825 pr_debug2_peo("switching off group read\n");
1826 goto fallback_missing_features;
1830 threads->err_thread = thread;
1834 while (--thread >= 0) {
1835 if (FD(evsel, cpu, thread) >= 0)
1836 close(FD(evsel, cpu, thread));
1837 FD(evsel, cpu, thread) = -1;
1840 } while (--cpu >= 0);
1845 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1846 struct perf_thread_map *threads)
1848 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1851 void evsel__close(struct evsel *evsel)
1853 perf_evsel__close(&evsel->core);
1854 perf_evsel__free_id(&evsel->core);
1857 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
1860 return evsel__open_cpu(evsel, cpus, NULL, 0,
1861 cpus ? cpus->nr : 1);
1863 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1866 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
1868 return evsel__open(evsel, NULL, threads);
1871 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1872 const union perf_event *event,
1873 struct perf_sample *sample)
1875 u64 type = evsel->core.attr.sample_type;
1876 const __u64 *array = event->sample.array;
1877 bool swapped = evsel->needs_swap;
1880 array += ((event->header.size -
1881 sizeof(event->header)) / sizeof(u64)) - 1;
1883 if (type & PERF_SAMPLE_IDENTIFIER) {
1884 sample->id = *array;
1888 if (type & PERF_SAMPLE_CPU) {
1891 /* undo swap of u64, then swap on individual u32s */
1892 u.val64 = bswap_64(u.val64);
1893 u.val32[0] = bswap_32(u.val32[0]);
1896 sample->cpu = u.val32[0];
1900 if (type & PERF_SAMPLE_STREAM_ID) {
1901 sample->stream_id = *array;
1905 if (type & PERF_SAMPLE_ID) {
1906 sample->id = *array;
1910 if (type & PERF_SAMPLE_TIME) {
1911 sample->time = *array;
1915 if (type & PERF_SAMPLE_TID) {
1918 /* undo swap of u64, then swap on individual u32s */
1919 u.val64 = bswap_64(u.val64);
1920 u.val32[0] = bswap_32(u.val32[0]);
1921 u.val32[1] = bswap_32(u.val32[1]);
1924 sample->pid = u.val32[0];
1925 sample->tid = u.val32[1];
1932 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1935 return size > max_size || offset + size > endp;
1938 #define OVERFLOW_CHECK(offset, size, max_size) \
1940 if (overflow(endp, (max_size), (offset), (size))) \
1944 #define OVERFLOW_CHECK_u64(offset) \
1945 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1948 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1951 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1952 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1953 * check the format does not go past the end of the event.
1955 if (sample_size + sizeof(event->header) > event->header.size)
1961 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1962 struct perf_sample *data)
1964 u64 type = evsel->core.attr.sample_type;
1965 bool swapped = evsel->needs_swap;
1967 u16 max_size = event->header.size;
1968 const void *endp = (void *)event + max_size;
1972 * used for cross-endian analysis. See git commit 65014ab3
1973 * for why this goofiness is needed.
1977 memset(data, 0, sizeof(*data));
1978 data->cpu = data->pid = data->tid = -1;
1979 data->stream_id = data->id = data->time = -1ULL;
1980 data->period = evsel->core.attr.sample_period;
1981 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1982 data->misc = event->header.misc;
1984 data->data_src = PERF_MEM_DATA_SRC_NONE;
1986 if (event->header.type != PERF_RECORD_SAMPLE) {
1987 if (!evsel->core.attr.sample_id_all)
1989 return perf_evsel__parse_id_sample(evsel, event, data);
1992 array = event->sample.array;
1994 if (perf_event__check_size(event, evsel->sample_size))
1997 if (type & PERF_SAMPLE_IDENTIFIER) {
2002 if (type & PERF_SAMPLE_IP) {
2007 if (type & PERF_SAMPLE_TID) {
2010 /* undo swap of u64, then swap on individual u32s */
2011 u.val64 = bswap_64(u.val64);
2012 u.val32[0] = bswap_32(u.val32[0]);
2013 u.val32[1] = bswap_32(u.val32[1]);
2016 data->pid = u.val32[0];
2017 data->tid = u.val32[1];
2021 if (type & PERF_SAMPLE_TIME) {
2022 data->time = *array;
2026 if (type & PERF_SAMPLE_ADDR) {
2027 data->addr = *array;
2031 if (type & PERF_SAMPLE_ID) {
2036 if (type & PERF_SAMPLE_STREAM_ID) {
2037 data->stream_id = *array;
2041 if (type & PERF_SAMPLE_CPU) {
2045 /* undo swap of u64, then swap on individual u32s */
2046 u.val64 = bswap_64(u.val64);
2047 u.val32[0] = bswap_32(u.val32[0]);
2050 data->cpu = u.val32[0];
2054 if (type & PERF_SAMPLE_PERIOD) {
2055 data->period = *array;
2059 if (type & PERF_SAMPLE_READ) {
2060 u64 read_format = evsel->core.attr.read_format;
2062 OVERFLOW_CHECK_u64(array);
2063 if (read_format & PERF_FORMAT_GROUP)
2064 data->read.group.nr = *array;
2066 data->read.one.value = *array;
2070 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2071 OVERFLOW_CHECK_u64(array);
2072 data->read.time_enabled = *array;
2076 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2077 OVERFLOW_CHECK_u64(array);
2078 data->read.time_running = *array;
2082 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2083 if (read_format & PERF_FORMAT_GROUP) {
2084 const u64 max_group_nr = UINT64_MAX /
2085 sizeof(struct sample_read_value);
2087 if (data->read.group.nr > max_group_nr)
2089 sz = data->read.group.nr *
2090 sizeof(struct sample_read_value);
2091 OVERFLOW_CHECK(array, sz, max_size);
2092 data->read.group.values =
2093 (struct sample_read_value *)array;
2094 array = (void *)array + sz;
2096 OVERFLOW_CHECK_u64(array);
2097 data->read.one.id = *array;
2102 if (type & PERF_SAMPLE_CALLCHAIN) {
2103 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2105 OVERFLOW_CHECK_u64(array);
2106 data->callchain = (struct ip_callchain *)array++;
2107 if (data->callchain->nr > max_callchain_nr)
2109 sz = data->callchain->nr * sizeof(u64);
2110 OVERFLOW_CHECK(array, sz, max_size);
2111 array = (void *)array + sz;
2114 if (type & PERF_SAMPLE_RAW) {
2115 OVERFLOW_CHECK_u64(array);
2119 * Undo swap of u64, then swap on individual u32s,
2120 * get the size of the raw area and undo all of the
2121 * swap. The pevent interface handles endianity by
2125 u.val64 = bswap_64(u.val64);
2126 u.val32[0] = bswap_32(u.val32[0]);
2127 u.val32[1] = bswap_32(u.val32[1]);
2129 data->raw_size = u.val32[0];
2132 * The raw data is aligned on 64bits including the
2133 * u32 size, so it's safe to use mem_bswap_64.
2136 mem_bswap_64((void *) array, data->raw_size);
2138 array = (void *)array + sizeof(u32);
2140 OVERFLOW_CHECK(array, data->raw_size, max_size);
2141 data->raw_data = (void *)array;
2142 array = (void *)array + data->raw_size;
2145 if (type & PERF_SAMPLE_BRANCH_STACK) {
2146 const u64 max_branch_nr = UINT64_MAX /
2147 sizeof(struct branch_entry);
2149 OVERFLOW_CHECK_u64(array);
2150 data->branch_stack = (struct branch_stack *)array++;
2152 if (data->branch_stack->nr > max_branch_nr)
2155 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2156 if (evsel__has_branch_hw_idx(evsel))
2159 data->no_hw_idx = true;
2160 OVERFLOW_CHECK(array, sz, max_size);
2161 array = (void *)array + sz;
2164 if (type & PERF_SAMPLE_REGS_USER) {
2165 OVERFLOW_CHECK_u64(array);
2166 data->user_regs.abi = *array;
2169 if (data->user_regs.abi) {
2170 u64 mask = evsel->core.attr.sample_regs_user;
2172 sz = hweight64(mask) * sizeof(u64);
2173 OVERFLOW_CHECK(array, sz, max_size);
2174 data->user_regs.mask = mask;
2175 data->user_regs.regs = (u64 *)array;
2176 array = (void *)array + sz;
2180 if (type & PERF_SAMPLE_STACK_USER) {
2181 OVERFLOW_CHECK_u64(array);
2184 data->user_stack.offset = ((char *)(array - 1)
2188 data->user_stack.size = 0;
2190 OVERFLOW_CHECK(array, sz, max_size);
2191 data->user_stack.data = (char *)array;
2192 array = (void *)array + sz;
2193 OVERFLOW_CHECK_u64(array);
2194 data->user_stack.size = *array++;
2195 if (WARN_ONCE(data->user_stack.size > sz,
2196 "user stack dump failure\n"))
2201 if (type & PERF_SAMPLE_WEIGHT) {
2202 OVERFLOW_CHECK_u64(array);
2203 data->weight = *array;
2207 if (type & PERF_SAMPLE_DATA_SRC) {
2208 OVERFLOW_CHECK_u64(array);
2209 data->data_src = *array;
2213 if (type & PERF_SAMPLE_TRANSACTION) {
2214 OVERFLOW_CHECK_u64(array);
2215 data->transaction = *array;
2219 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2220 if (type & PERF_SAMPLE_REGS_INTR) {
2221 OVERFLOW_CHECK_u64(array);
2222 data->intr_regs.abi = *array;
2225 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2226 u64 mask = evsel->core.attr.sample_regs_intr;
2228 sz = hweight64(mask) * sizeof(u64);
2229 OVERFLOW_CHECK(array, sz, max_size);
2230 data->intr_regs.mask = mask;
2231 data->intr_regs.regs = (u64 *)array;
2232 array = (void *)array + sz;
2236 data->phys_addr = 0;
2237 if (type & PERF_SAMPLE_PHYS_ADDR) {
2238 data->phys_addr = *array;
2243 if (type & PERF_SAMPLE_CGROUP) {
2244 data->cgroup = *array;
2248 if (type & PERF_SAMPLE_AUX) {
2249 OVERFLOW_CHECK_u64(array);
2252 OVERFLOW_CHECK(array, sz, max_size);
2253 /* Undo swap of data */
2255 mem_bswap_64((char *)array, sz);
2256 data->aux_sample.size = sz;
2257 data->aux_sample.data = (char *)array;
2258 array = (void *)array + sz;
2264 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2267 u64 type = evsel->core.attr.sample_type;
2270 if (!(type & PERF_SAMPLE_TIME))
2273 if (event->header.type != PERF_RECORD_SAMPLE) {
2274 struct perf_sample data = {
2278 if (!evsel->core.attr.sample_id_all)
2280 if (perf_evsel__parse_id_sample(evsel, event, &data))
2283 *timestamp = data.time;
2287 array = event->sample.array;
2289 if (perf_event__check_size(event, evsel->sample_size))
2292 if (type & PERF_SAMPLE_IDENTIFIER)
2295 if (type & PERF_SAMPLE_IP)
2298 if (type & PERF_SAMPLE_TID)
2301 if (type & PERF_SAMPLE_TIME)
2302 *timestamp = *array;
2307 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2309 return tep_find_field(evsel->tp_format, name);
2312 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2314 struct tep_format_field *field = evsel__field(evsel, name);
2320 offset = field->offset;
2322 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2323 offset = *(int *)(sample->raw_data + field->offset);
2327 return sample->raw_data + offset;
2330 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2334 void *ptr = sample->raw_data + field->offset;
2336 switch (field->size) {
2340 value = *(u16 *)ptr;
2343 value = *(u32 *)ptr;
2346 memcpy(&value, ptr, sizeof(u64));
2355 switch (field->size) {
2357 return bswap_16(value);
2359 return bswap_32(value);
2361 return bswap_64(value);
2369 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2371 struct tep_format_field *field = evsel__field(evsel, name);
2376 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2379 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2383 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2384 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2385 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2387 * If it's cycles then fall back to hrtimer based
2388 * cpu-clock-tick sw counter, which is always available even if
2391 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2394 scnprintf(msg, msgsize, "%s",
2395 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2397 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2398 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2400 zfree(&evsel->name);
2402 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2403 (paranoid = perf_event_paranoid()) > 1) {
2404 const char *name = evsel__name(evsel);
2406 const char *sep = ":";
2408 /* If event has exclude user then don't exclude kernel. */
2409 if (evsel->core.attr.exclude_user)
2412 /* Is there already the separator in the name. */
2413 if (strchr(name, '/') ||
2417 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2422 evsel->name = new_name;
2423 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2424 "to fall back to excluding kernel and hypervisor "
2425 " samples", paranoid);
2426 evsel->core.attr.exclude_kernel = 1;
2427 evsel->core.attr.exclude_hv = 1;
2435 static bool find_process(const char *name)
2437 size_t len = strlen(name);
2442 dir = opendir(procfs__mountpoint());
2446 /* Walk through the directory. */
2447 while (ret && (d = readdir(dir)) != NULL) {
2448 char path[PATH_MAX];
2452 if ((d->d_type != DT_DIR) ||
2453 !strcmp(".", d->d_name) ||
2454 !strcmp("..", d->d_name))
2457 scnprintf(path, sizeof(path), "%s/%s/comm",
2458 procfs__mountpoint(), d->d_name);
2460 if (filename__read_str(path, &data, &size))
2463 ret = strncmp(name, data, len);
2468 return ret ? false : true;
2471 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2472 int err, char *msg, size_t size)
2474 char sbuf[STRERR_BUFSIZE];
2481 printed = scnprintf(msg, size,
2482 "No permission to enable %s event.\n\n", evsel__name(evsel));
2484 return scnprintf(msg + printed, size - printed,
2485 "You may not have permission to collect %sstats.\n\n"
2486 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2487 "which controls use of the performance events system by\n"
2488 "unprivileged users (without CAP_PERFMON or CAP_SYS_ADMIN).\n\n"
2489 "The current value is %d:\n\n"
2490 " -1: Allow use of (almost) all events by all users\n"
2491 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2492 ">= 0: Disallow ftrace function tracepoint by users without CAP_PERFMON or CAP_SYS_ADMIN\n"
2493 " Disallow raw tracepoint access by users without CAP_SYS_PERFMON or CAP_SYS_ADMIN\n"
2494 ">= 1: Disallow CPU event access by users without CAP_PERFMON or CAP_SYS_ADMIN\n"
2495 ">= 2: Disallow kernel profiling by users without CAP_PERFMON or CAP_SYS_ADMIN\n\n"
2496 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2497 " kernel.perf_event_paranoid = -1\n" ,
2498 target->system_wide ? "system-wide " : "",
2499 perf_event_paranoid());
2501 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2503 return scnprintf(msg, size, "%s",
2504 "Too many events are opened.\n"
2505 "Probably the maximum number of open file descriptors has been reached.\n"
2506 "Hint: Try again after reducing the number of events.\n"
2507 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2509 if (evsel__has_callchain(evsel) &&
2510 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2511 return scnprintf(msg, size,
2512 "Not enough memory to setup event with callchain.\n"
2513 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2514 "Hint: Current value: %d", sysctl__max_stack());
2517 if (target->cpu_list)
2518 return scnprintf(msg, size, "%s",
2519 "No such device - did you specify an out-of-range profile CPU?");
2522 if (evsel->core.attr.sample_period != 0)
2523 return scnprintf(msg, size,
2524 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2525 evsel__name(evsel));
2526 if (evsel->core.attr.precise_ip)
2527 return scnprintf(msg, size, "%s",
2528 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2529 #if defined(__i386__) || defined(__x86_64__)
2530 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2531 return scnprintf(msg, size, "%s",
2532 "No hardware sampling interrupt available.\n");
2536 if (find_process("oprofiled"))
2537 return scnprintf(msg, size,
2538 "The PMU counters are busy/taken by another profiler.\n"
2539 "We found oprofile daemon running, please stop it and try again.");
2542 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2543 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2544 if (perf_missing_features.clockid)
2545 return scnprintf(msg, size, "clockid feature not supported.");
2546 if (perf_missing_features.clockid_wrong)
2547 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2548 if (perf_missing_features.aux_output)
2549 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2555 return scnprintf(msg, size,
2556 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2557 "/bin/dmesg | grep -i perf may provide additional information.\n",
2558 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2561 struct perf_env *evsel__env(struct evsel *evsel)
2563 if (evsel && evsel->evlist)
2564 return evsel->evlist->env;
2568 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2572 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2573 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2575 int fd = FD(evsel, cpu, thread);
2577 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2578 cpu, thread, fd) < 0)
2586 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2588 struct perf_cpu_map *cpus = evsel->core.cpus;
2589 struct perf_thread_map *threads = evsel->core.threads;
2591 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2594 return store_evsel_ids(evsel, evlist);