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 perf_evsel__no_extra_init(struct evsel *evsel __maybe_unused)
64 void __weak test_attr__ready(void) { }
66 static void perf_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 = perf_evsel__no_extra_init,
77 .fini = perf_evsel__no_extra_fini,
80 int perf_evsel__object_config(size_t object_size,
81 int (*init)(struct evsel *evsel),
82 void (*fini)(struct evsel *evsel))
88 if (perf_evsel__object.size > object_size)
91 perf_evsel__object.size = object_size;
95 perf_evsel__object.init = init;
98 perf_evsel__object.fini = fini;
103 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
105 int __perf_evsel__sample_size(u64 sample_type)
107 u64 mask = sample_type & PERF_SAMPLE_MASK;
111 for (i = 0; i < 64; i++) {
112 if (mask & (1ULL << i))
122 * __perf_evsel__calc_id_pos - calculate id_pos.
123 * @sample_type: sample type
125 * This function returns the position of the event id (PERF_SAMPLE_ID or
126 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
127 * perf_record_sample.
129 static int __perf_evsel__calc_id_pos(u64 sample_type)
133 if (sample_type & PERF_SAMPLE_IDENTIFIER)
136 if (!(sample_type & PERF_SAMPLE_ID))
139 if (sample_type & PERF_SAMPLE_IP)
142 if (sample_type & PERF_SAMPLE_TID)
145 if (sample_type & PERF_SAMPLE_TIME)
148 if (sample_type & PERF_SAMPLE_ADDR)
155 * __perf_evsel__calc_is_pos - calculate is_pos.
156 * @sample_type: sample type
158 * This function returns the position (counting backwards) of the event id
159 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
160 * sample_id_all is used there is an id sample appended to non-sample events.
162 static int __perf_evsel__calc_is_pos(u64 sample_type)
166 if (sample_type & PERF_SAMPLE_IDENTIFIER)
169 if (!(sample_type & PERF_SAMPLE_ID))
172 if (sample_type & PERF_SAMPLE_CPU)
175 if (sample_type & PERF_SAMPLE_STREAM_ID)
181 void perf_evsel__calc_id_pos(struct evsel *evsel)
183 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
184 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
187 void __perf_evsel__set_sample_bit(struct evsel *evsel,
188 enum perf_event_sample_format bit)
190 if (!(evsel->core.attr.sample_type & bit)) {
191 evsel->core.attr.sample_type |= bit;
192 evsel->sample_size += sizeof(u64);
193 perf_evsel__calc_id_pos(evsel);
197 void __perf_evsel__reset_sample_bit(struct evsel *evsel,
198 enum perf_event_sample_format bit)
200 if (evsel->core.attr.sample_type & bit) {
201 evsel->core.attr.sample_type &= ~bit;
202 evsel->sample_size -= sizeof(u64);
203 perf_evsel__calc_id_pos(evsel);
207 void perf_evsel__set_sample_id(struct evsel *evsel,
208 bool can_sample_identifier)
210 if (can_sample_identifier) {
211 perf_evsel__reset_sample_bit(evsel, ID);
212 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
214 perf_evsel__set_sample_bit(evsel, ID);
216 evsel->core.attr.read_format |= PERF_FORMAT_ID;
220 * perf_evsel__is_function_event - Return whether given evsel is a function
223 * @evsel - evsel selector to be tested
225 * Return %true if event is function trace event
227 bool perf_evsel__is_function_event(struct evsel *evsel)
229 #define FUNCTION_EVENT "ftrace:function"
231 return evsel->name &&
232 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
234 #undef FUNCTION_EVENT
237 void evsel__init(struct evsel *evsel,
238 struct perf_event_attr *attr, int idx)
240 perf_evsel__init(&evsel->core, attr);
242 evsel->tracking = !idx;
243 evsel->leader = evsel;
246 evsel->max_events = ULONG_MAX;
247 evsel->evlist = NULL;
248 evsel->bpf_obj = NULL;
250 INIT_LIST_HEAD(&evsel->config_terms);
251 perf_evsel__object.init(evsel);
252 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
253 perf_evsel__calc_id_pos(evsel);
254 evsel->cmdline_group_boundary = false;
255 evsel->metric_expr = NULL;
256 evsel->metric_name = NULL;
257 evsel->metric_events = NULL;
258 evsel->collect_stat = false;
259 evsel->pmu_name = NULL;
262 struct evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
264 struct evsel *evsel = zalloc(perf_evsel__object.size);
268 evsel__init(evsel, attr, idx);
270 if (perf_evsel__is_bpf_output(evsel)) {
271 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
272 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
273 evsel->core.attr.sample_period = 1;
276 if (perf_evsel__is_clock(evsel)) {
278 * The evsel->unit points to static alias->unit
279 * so it's ok to use static string in here.
281 static const char *unit = "msec";
290 static bool perf_event_can_profile_kernel(void)
292 return perf_event_paranoid_check(1);
295 struct evsel *perf_evsel__new_cycles(bool precise)
297 struct perf_event_attr attr = {
298 .type = PERF_TYPE_HARDWARE,
299 .config = PERF_COUNT_HW_CPU_CYCLES,
300 .exclude_kernel = !perf_event_can_profile_kernel(),
304 event_attr_init(&attr);
310 * Now let the usual logic to set up the perf_event_attr defaults
311 * to kick in when we return and before perf_evsel__open() is called.
314 evsel = evsel__new(&attr);
318 evsel->precise_max = true;
320 /* use asprintf() because free(evsel) assumes name is allocated */
321 if (asprintf(&evsel->name, "cycles%s%s%.*s",
322 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
323 attr.exclude_kernel ? "u" : "",
324 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
329 evsel__delete(evsel);
335 * Returns pointer with encoded error via <linux/err.h> interface.
337 struct evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
339 struct evsel *evsel = zalloc(perf_evsel__object.size);
345 struct perf_event_attr attr = {
346 .type = PERF_TYPE_TRACEPOINT,
347 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
348 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
351 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
354 evsel->tp_format = trace_event__tp_format(sys, name);
355 if (IS_ERR(evsel->tp_format)) {
356 err = PTR_ERR(evsel->tp_format);
360 event_attr_init(&attr);
361 attr.config = evsel->tp_format->id;
362 attr.sample_period = 1;
363 evsel__init(evsel, &attr, idx);
375 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
383 "stalled-cycles-frontend",
384 "stalled-cycles-backend",
388 static const char *__perf_evsel__hw_name(u64 config)
390 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
391 return perf_evsel__hw_names[config];
393 return "unknown-hardware";
396 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
398 int colon = 0, r = 0;
399 struct perf_event_attr *attr = &evsel->core.attr;
400 bool exclude_guest_default = false;
402 #define MOD_PRINT(context, mod) do { \
403 if (!attr->exclude_##context) { \
404 if (!colon) colon = ++r; \
405 r += scnprintf(bf + r, size - r, "%c", mod); \
408 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
409 MOD_PRINT(kernel, 'k');
410 MOD_PRINT(user, 'u');
412 exclude_guest_default = true;
415 if (attr->precise_ip) {
418 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
419 exclude_guest_default = true;
422 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
423 MOD_PRINT(host, 'H');
424 MOD_PRINT(guest, 'G');
432 static int perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
434 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config));
435 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
438 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
451 static const char *__perf_evsel__sw_name(u64 config)
453 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
454 return perf_evsel__sw_names[config];
455 return "unknown-software";
458 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
460 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config));
461 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
464 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
468 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
470 if (type & HW_BREAKPOINT_R)
471 r += scnprintf(bf + r, size - r, "r");
473 if (type & HW_BREAKPOINT_W)
474 r += scnprintf(bf + r, size - r, "w");
476 if (type & HW_BREAKPOINT_X)
477 r += scnprintf(bf + r, size - r, "x");
482 static int perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
484 struct perf_event_attr *attr = &evsel->core.attr;
485 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
486 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
489 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
490 [PERF_EVSEL__MAX_ALIASES] = {
491 { "L1-dcache", "l1-d", "l1d", "L1-data", },
492 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
494 { "dTLB", "d-tlb", "Data-TLB", },
495 { "iTLB", "i-tlb", "Instruction-TLB", },
496 { "branch", "branches", "bpu", "btb", "bpc", },
500 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
501 [PERF_EVSEL__MAX_ALIASES] = {
502 { "load", "loads", "read", },
503 { "store", "stores", "write", },
504 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
507 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
508 [PERF_EVSEL__MAX_ALIASES] = {
509 { "refs", "Reference", "ops", "access", },
510 { "misses", "miss", },
513 #define C(x) PERF_COUNT_HW_CACHE_##x
514 #define CACHE_READ (1 << C(OP_READ))
515 #define CACHE_WRITE (1 << C(OP_WRITE))
516 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
517 #define COP(x) (1 << x)
520 * cache operartion stat
521 * L1I : Read and prefetch only
522 * ITLB and BPU : Read-only
524 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
525 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
526 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
527 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
528 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
529 [C(ITLB)] = (CACHE_READ),
530 [C(BPU)] = (CACHE_READ),
531 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
534 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
536 if (perf_evsel__hw_cache_stat[type] & COP(op))
537 return true; /* valid */
539 return false; /* invalid */
542 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
543 char *bf, size_t size)
546 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
547 perf_evsel__hw_cache_op[op][0],
548 perf_evsel__hw_cache_result[result][0]);
551 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
552 perf_evsel__hw_cache_op[op][1]);
555 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
557 u8 op, result, type = (config >> 0) & 0xff;
558 const char *err = "unknown-ext-hardware-cache-type";
560 if (type >= PERF_COUNT_HW_CACHE_MAX)
563 op = (config >> 8) & 0xff;
564 err = "unknown-ext-hardware-cache-op";
565 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
568 result = (config >> 16) & 0xff;
569 err = "unknown-ext-hardware-cache-result";
570 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
573 err = "invalid-cache";
574 if (!perf_evsel__is_cache_op_valid(type, op))
577 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
579 return scnprintf(bf, size, "%s", err);
582 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
584 int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size);
585 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
588 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
590 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
591 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
594 static int perf_evsel__tool_name(char *bf, size_t size)
596 int ret = scnprintf(bf, size, "duration_time");
600 const char *perf_evsel__name(struct evsel *evsel)
610 switch (evsel->core.attr.type) {
612 perf_evsel__raw_name(evsel, bf, sizeof(bf));
615 case PERF_TYPE_HARDWARE:
616 perf_evsel__hw_name(evsel, bf, sizeof(bf));
619 case PERF_TYPE_HW_CACHE:
620 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
623 case PERF_TYPE_SOFTWARE:
624 if (evsel->tool_event)
625 perf_evsel__tool_name(bf, sizeof(bf));
627 perf_evsel__sw_name(evsel, bf, sizeof(bf));
630 case PERF_TYPE_TRACEPOINT:
631 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
634 case PERF_TYPE_BREAKPOINT:
635 perf_evsel__bp_name(evsel, bf, sizeof(bf));
639 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
640 evsel->core.attr.type);
644 evsel->name = strdup(bf);
652 const char *perf_evsel__group_name(struct evsel *evsel)
654 return evsel->group_name ?: "anon group";
658 * Returns the group details for the specified leader,
659 * with following rules.
661 * For record -e '{cycles,instructions}'
662 * 'anon group { cycles:u, instructions:u }'
664 * For record -e 'cycles,instructions' and report --group
665 * 'cycles:u, instructions:u'
667 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
671 const char *group_name = perf_evsel__group_name(evsel);
673 if (!evsel->forced_leader)
674 ret = scnprintf(buf, size, "%s { ", group_name);
676 ret += scnprintf(buf + ret, size - ret, "%s",
677 perf_evsel__name(evsel));
679 for_each_group_member(pos, evsel)
680 ret += scnprintf(buf + ret, size - ret, ", %s",
681 perf_evsel__name(pos));
683 if (!evsel->forced_leader)
684 ret += scnprintf(buf + ret, size - ret, " }");
689 static void __perf_evsel__config_callchain(struct evsel *evsel,
690 struct record_opts *opts,
691 struct callchain_param *param)
693 bool function = perf_evsel__is_function_event(evsel);
694 struct perf_event_attr *attr = &evsel->core.attr;
696 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
698 attr->sample_max_stack = param->max_stack;
700 if (opts->kernel_callchains)
701 attr->exclude_callchain_user = 1;
702 if (opts->user_callchains)
703 attr->exclude_callchain_kernel = 1;
704 if (param->record_mode == CALLCHAIN_LBR) {
705 if (!opts->branch_stack) {
706 if (attr->exclude_user) {
707 pr_warning("LBR callstack option is only available "
708 "to get user callchain information. "
709 "Falling back to framepointers.\n");
711 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
712 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
713 PERF_SAMPLE_BRANCH_CALL_STACK |
714 PERF_SAMPLE_BRANCH_NO_CYCLES |
715 PERF_SAMPLE_BRANCH_NO_FLAGS;
718 pr_warning("Cannot use LBR callstack with branch stack. "
719 "Falling back to framepointers.\n");
722 if (param->record_mode == CALLCHAIN_DWARF) {
724 perf_evsel__set_sample_bit(evsel, REGS_USER);
725 perf_evsel__set_sample_bit(evsel, STACK_USER);
726 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
727 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
728 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
729 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
730 "so the minimal registers set (IP, SP) is explicitly forced.\n");
732 attr->sample_regs_user |= PERF_REGS_MASK;
734 attr->sample_stack_user = param->dump_size;
735 attr->exclude_callchain_user = 1;
737 pr_info("Cannot use DWARF unwind for function trace event,"
738 " falling back to framepointers.\n");
743 pr_info("Disabling user space callchains for function trace event.\n");
744 attr->exclude_callchain_user = 1;
748 void perf_evsel__config_callchain(struct evsel *evsel,
749 struct record_opts *opts,
750 struct callchain_param *param)
753 return __perf_evsel__config_callchain(evsel, opts, param);
757 perf_evsel__reset_callgraph(struct evsel *evsel,
758 struct callchain_param *param)
760 struct perf_event_attr *attr = &evsel->core.attr;
762 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
763 if (param->record_mode == CALLCHAIN_LBR) {
764 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
765 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
766 PERF_SAMPLE_BRANCH_CALL_STACK);
768 if (param->record_mode == CALLCHAIN_DWARF) {
769 perf_evsel__reset_sample_bit(evsel, REGS_USER);
770 perf_evsel__reset_sample_bit(evsel, STACK_USER);
774 static void apply_config_terms(struct evsel *evsel,
775 struct record_opts *opts, bool track)
777 struct perf_evsel_config_term *term;
778 struct list_head *config_terms = &evsel->config_terms;
779 struct perf_event_attr *attr = &evsel->core.attr;
780 /* callgraph default */
781 struct callchain_param param = {
782 .record_mode = callchain_param.record_mode,
786 const char *callgraph_buf = NULL;
788 list_for_each_entry(term, config_terms, list) {
789 switch (term->type) {
790 case PERF_EVSEL__CONFIG_TERM_PERIOD:
791 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
792 attr->sample_period = term->val.period;
794 perf_evsel__reset_sample_bit(evsel, PERIOD);
797 case PERF_EVSEL__CONFIG_TERM_FREQ:
798 if (!(term->weak && opts->user_freq != UINT_MAX)) {
799 attr->sample_freq = term->val.freq;
801 perf_evsel__set_sample_bit(evsel, PERIOD);
804 case PERF_EVSEL__CONFIG_TERM_TIME:
806 perf_evsel__set_sample_bit(evsel, TIME);
808 perf_evsel__reset_sample_bit(evsel, TIME);
810 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
811 callgraph_buf = term->val.callgraph;
813 case PERF_EVSEL__CONFIG_TERM_BRANCH:
814 if (term->val.branch && strcmp(term->val.branch, "no")) {
815 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
816 parse_branch_str(term->val.branch,
817 &attr->branch_sample_type);
819 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
821 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
822 dump_size = term->val.stack_user;
824 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
825 max_stack = term->val.max_stack;
827 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
828 evsel->max_events = term->val.max_events;
830 case PERF_EVSEL__CONFIG_TERM_INHERIT:
832 * attr->inherit should has already been set by
833 * perf_evsel__config. If user explicitly set
834 * inherit using config terms, override global
835 * opt->no_inherit setting.
837 attr->inherit = term->val.inherit ? 1 : 0;
839 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
840 attr->write_backward = term->val.overwrite ? 1 : 0;
842 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
844 case PERF_EVSEL__CONFIG_TERM_PERCORE:
846 case PERF_EVSEL__CONFIG_TERM_AUX_OUTPUT:
847 attr->aux_output = term->val.aux_output ? 1 : 0;
854 /* User explicitly set per-event callgraph, clear the old setting and reset. */
855 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
856 bool sample_address = false;
859 param.max_stack = max_stack;
860 if (callgraph_buf == NULL)
861 callgraph_buf = "fp";
864 /* parse callgraph parameters */
865 if (callgraph_buf != NULL) {
866 if (!strcmp(callgraph_buf, "no")) {
867 param.enabled = false;
868 param.record_mode = CALLCHAIN_NONE;
870 param.enabled = true;
871 if (parse_callchain_record(callgraph_buf, ¶m)) {
872 pr_err("per-event callgraph setting for %s failed. "
873 "Apply callgraph global setting for it\n",
877 if (param.record_mode == CALLCHAIN_DWARF)
878 sample_address = true;
882 dump_size = round_up(dump_size, sizeof(u64));
883 param.dump_size = dump_size;
886 /* If global callgraph set, clear it */
887 if (callchain_param.enabled)
888 perf_evsel__reset_callgraph(evsel, &callchain_param);
890 /* set perf-event callgraph */
892 if (sample_address) {
893 perf_evsel__set_sample_bit(evsel, ADDR);
894 perf_evsel__set_sample_bit(evsel, DATA_SRC);
895 evsel->core.attr.mmap_data = track;
897 perf_evsel__config_callchain(evsel, opts, ¶m);
902 static bool is_dummy_event(struct evsel *evsel)
904 return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
905 (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
909 * The enable_on_exec/disabled value strategy:
911 * 1) For any type of traced program:
912 * - all independent events and group leaders are disabled
913 * - all group members are enabled
915 * Group members are ruled by group leaders. They need to
916 * be enabled, because the group scheduling relies on that.
918 * 2) For traced programs executed by perf:
919 * - all independent events and group leaders have
921 * - we don't specifically enable or disable any event during
924 * Independent events and group leaders are initially disabled
925 * and get enabled by exec. Group members are ruled by group
926 * leaders as stated in 1).
928 * 3) For traced programs attached by perf (pid/tid):
929 * - we specifically enable or disable all events during
932 * When attaching events to already running traced we
933 * enable/disable events specifically, as there's no
934 * initial traced exec call.
936 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
937 struct callchain_param *callchain)
939 struct evsel *leader = evsel->leader;
940 struct perf_event_attr *attr = &evsel->core.attr;
941 int track = evsel->tracking;
942 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
944 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
945 attr->inherit = !opts->no_inherit;
946 attr->write_backward = opts->overwrite ? 1 : 0;
948 perf_evsel__set_sample_bit(evsel, IP);
949 perf_evsel__set_sample_bit(evsel, TID);
951 if (evsel->sample_read) {
952 perf_evsel__set_sample_bit(evsel, READ);
955 * We need ID even in case of single event, because
956 * PERF_SAMPLE_READ process ID specific data.
958 perf_evsel__set_sample_id(evsel, false);
961 * Apply group format only if we belong to group
962 * with more than one members.
964 if (leader->core.nr_members > 1) {
965 attr->read_format |= PERF_FORMAT_GROUP;
971 * We default some events to have a default interval. But keep
972 * it a weak assumption overridable by the user.
974 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
975 opts->user_interval != ULLONG_MAX)) {
977 perf_evsel__set_sample_bit(evsel, PERIOD);
979 attr->sample_freq = opts->freq;
981 attr->sample_period = opts->default_interval;
986 * Disable sampling for all group members other
987 * than leader in case leader 'leads' the sampling.
989 if ((leader != evsel) && leader->sample_read) {
991 attr->sample_freq = 0;
992 attr->sample_period = 0;
993 attr->write_backward = 0;
996 * We don't get sample for slave events, we make them
997 * when delivering group leader sample. Set the slave
998 * event to follow the master sample_type to ease up
1001 attr->sample_type = leader->core.attr.sample_type;
1004 if (opts->no_samples)
1005 attr->sample_freq = 0;
1007 if (opts->inherit_stat) {
1008 evsel->core.attr.read_format |=
1009 PERF_FORMAT_TOTAL_TIME_ENABLED |
1010 PERF_FORMAT_TOTAL_TIME_RUNNING |
1012 attr->inherit_stat = 1;
1015 if (opts->sample_address) {
1016 perf_evsel__set_sample_bit(evsel, ADDR);
1017 attr->mmap_data = track;
1021 * We don't allow user space callchains for function trace
1022 * event, due to issues with page faults while tracing page
1023 * fault handler and its overall trickiness nature.
1025 if (perf_evsel__is_function_event(evsel))
1026 evsel->core.attr.exclude_callchain_user = 1;
1028 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1029 perf_evsel__config_callchain(evsel, opts, callchain);
1031 if (opts->sample_intr_regs) {
1032 attr->sample_regs_intr = opts->sample_intr_regs;
1033 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1036 if (opts->sample_user_regs) {
1037 attr->sample_regs_user |= opts->sample_user_regs;
1038 perf_evsel__set_sample_bit(evsel, REGS_USER);
1041 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1042 perf_evsel__set_sample_bit(evsel, CPU);
1045 * When the user explicitly disabled time don't force it here.
1047 if (opts->sample_time &&
1048 (!perf_missing_features.sample_id_all &&
1049 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1050 opts->sample_time_set)))
1051 perf_evsel__set_sample_bit(evsel, TIME);
1053 if (opts->raw_samples && !evsel->no_aux_samples) {
1054 perf_evsel__set_sample_bit(evsel, TIME);
1055 perf_evsel__set_sample_bit(evsel, RAW);
1056 perf_evsel__set_sample_bit(evsel, CPU);
1059 if (opts->sample_address)
1060 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1062 if (opts->sample_phys_addr)
1063 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1065 if (opts->no_buffering) {
1066 attr->watermark = 0;
1067 attr->wakeup_events = 1;
1069 if (opts->branch_stack && !evsel->no_aux_samples) {
1070 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1071 attr->branch_sample_type = opts->branch_stack;
1074 if (opts->sample_weight)
1075 perf_evsel__set_sample_bit(evsel, WEIGHT);
1079 attr->mmap2 = track && !perf_missing_features.mmap2;
1081 attr->ksymbol = track && !perf_missing_features.ksymbol;
1082 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1084 if (opts->record_namespaces)
1085 attr->namespaces = track;
1087 if (opts->record_switch_events)
1088 attr->context_switch = track;
1090 if (opts->sample_transaction)
1091 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1093 if (opts->running_time) {
1094 evsel->core.attr.read_format |=
1095 PERF_FORMAT_TOTAL_TIME_ENABLED |
1096 PERF_FORMAT_TOTAL_TIME_RUNNING;
1100 * XXX see the function comment above
1102 * Disabling only independent events or group leaders,
1103 * keeping group members enabled.
1105 if (perf_evsel__is_group_leader(evsel))
1109 * Setting enable_on_exec for independent events and
1110 * group leaders for traced executed by perf.
1112 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1113 !opts->initial_delay)
1114 attr->enable_on_exec = 1;
1116 if (evsel->immediate) {
1118 attr->enable_on_exec = 0;
1121 clockid = opts->clockid;
1122 if (opts->use_clockid) {
1123 attr->use_clockid = 1;
1124 attr->clockid = opts->clockid;
1127 if (evsel->precise_max)
1128 attr->precise_ip = 3;
1130 if (opts->all_user) {
1131 attr->exclude_kernel = 1;
1132 attr->exclude_user = 0;
1135 if (opts->all_kernel) {
1136 attr->exclude_kernel = 0;
1137 attr->exclude_user = 1;
1140 if (evsel->core.own_cpus || evsel->unit)
1141 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1144 * Apply event specific term settings,
1145 * it overloads any global configuration.
1147 apply_config_terms(evsel, opts, track);
1149 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1151 /* The --period option takes the precedence. */
1152 if (opts->period_set) {
1154 perf_evsel__set_sample_bit(evsel, PERIOD);
1156 perf_evsel__reset_sample_bit(evsel, PERIOD);
1160 * For initial_delay, a dummy event is added implicitly.
1161 * The software event will trigger -EOPNOTSUPP error out,
1162 * if BRANCH_STACK bit is set.
1164 if (opts->initial_delay && is_dummy_event(evsel))
1165 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1168 int perf_evsel__set_filter(struct evsel *evsel, const char *filter)
1170 char *new_filter = strdup(filter);
1172 if (new_filter != NULL) {
1173 free(evsel->filter);
1174 evsel->filter = new_filter;
1181 static int perf_evsel__append_filter(struct evsel *evsel,
1182 const char *fmt, const char *filter)
1186 if (evsel->filter == NULL)
1187 return perf_evsel__set_filter(evsel, filter);
1189 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1190 free(evsel->filter);
1191 evsel->filter = new_filter;
1198 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1200 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1203 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1205 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1208 int evsel__enable(struct evsel *evsel)
1210 int err = perf_evsel__enable(&evsel->core);
1213 evsel->disabled = false;
1218 int evsel__disable(struct evsel *evsel)
1220 int err = perf_evsel__disable(&evsel->core);
1222 * We mark it disabled here so that tools that disable a event can
1223 * ignore events after they disable it. I.e. the ring buffer may have
1224 * already a few more events queued up before the kernel got the stop
1228 evsel->disabled = true;
1233 static void perf_evsel__free_config_terms(struct evsel *evsel)
1235 struct perf_evsel_config_term *term, *h;
1237 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1238 list_del_init(&term->list);
1243 void perf_evsel__exit(struct evsel *evsel)
1245 assert(list_empty(&evsel->core.node));
1246 assert(evsel->evlist == NULL);
1247 perf_evsel__free_counts(evsel);
1248 perf_evsel__free_fd(&evsel->core);
1249 perf_evsel__free_id(&evsel->core);
1250 perf_evsel__free_config_terms(evsel);
1251 cgroup__put(evsel->cgrp);
1252 perf_cpu_map__put(evsel->core.cpus);
1253 perf_cpu_map__put(evsel->core.own_cpus);
1254 perf_thread_map__put(evsel->core.threads);
1255 zfree(&evsel->group_name);
1256 zfree(&evsel->name);
1257 perf_evsel__object.fini(evsel);
1260 void evsel__delete(struct evsel *evsel)
1262 perf_evsel__exit(evsel);
1266 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1267 struct perf_counts_values *count)
1269 struct perf_counts_values tmp;
1271 if (!evsel->prev_raw_counts)
1275 tmp = evsel->prev_raw_counts->aggr;
1276 evsel->prev_raw_counts->aggr = *count;
1278 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1279 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1282 count->val = count->val - tmp.val;
1283 count->ena = count->ena - tmp.ena;
1284 count->run = count->run - tmp.run;
1287 void perf_counts_values__scale(struct perf_counts_values *count,
1288 bool scale, s8 *pscaled)
1293 if (count->run == 0) {
1296 } else if (count->run < count->ena) {
1298 count->val = (u64)((double) count->val * count->ena / count->run);
1307 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread)
1309 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1311 return perf_evsel__read(&evsel->core, cpu, thread, count);
1315 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1316 u64 val, u64 ena, u64 run)
1318 struct perf_counts_values *count;
1320 count = perf_counts(counter->counts, cpu, thread);
1326 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1330 perf_evsel__process_group_data(struct evsel *leader,
1331 int cpu, int thread, u64 *data)
1333 u64 read_format = leader->core.attr.read_format;
1334 struct sample_read_value *v;
1335 u64 nr, ena = 0, run = 0, i;
1339 if (nr != (u64) leader->core.nr_members)
1342 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1345 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1348 v = (struct sample_read_value *) data;
1350 perf_evsel__set_count(leader, cpu, thread,
1351 v[0].value, ena, run);
1353 for (i = 1; i < nr; i++) {
1354 struct evsel *counter;
1356 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1360 perf_evsel__set_count(counter, cpu, thread,
1361 v[i].value, ena, run);
1368 perf_evsel__read_group(struct evsel *leader, int cpu, int thread)
1370 struct perf_stat_evsel *ps = leader->stats;
1371 u64 read_format = leader->core.attr.read_format;
1372 int size = perf_evsel__read_size(&leader->core);
1373 u64 *data = ps->group_data;
1375 if (!(read_format & PERF_FORMAT_ID))
1378 if (!perf_evsel__is_group_leader(leader))
1382 data = zalloc(size);
1386 ps->group_data = data;
1389 if (FD(leader, cpu, thread) < 0)
1392 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1395 return perf_evsel__process_group_data(leader, cpu, thread, data);
1398 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1400 u64 read_format = evsel->core.attr.read_format;
1402 if (read_format & PERF_FORMAT_GROUP)
1403 return perf_evsel__read_group(evsel, cpu, thread);
1405 return perf_evsel__read_one(evsel, cpu, thread);
1408 int __perf_evsel__read_on_cpu(struct evsel *evsel,
1409 int cpu, int thread, bool scale)
1411 struct perf_counts_values count;
1412 size_t nv = scale ? 3 : 1;
1414 if (FD(evsel, cpu, thread) < 0)
1417 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1420 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1423 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1424 perf_counts_values__scale(&count, scale, NULL);
1425 *perf_counts(evsel->counts, cpu, thread) = count;
1429 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1431 struct evsel *leader = evsel->leader;
1434 if (perf_evsel__is_group_leader(evsel))
1438 * Leader must be already processed/open,
1439 * if not it's a bug.
1441 BUG_ON(!leader->core.fd);
1443 fd = FD(leader, cpu, thread);
1449 static void perf_evsel__remove_fd(struct evsel *pos,
1450 int nr_cpus, int nr_threads,
1453 for (int cpu = 0; cpu < nr_cpus; cpu++)
1454 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1455 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1458 static int update_fds(struct evsel *evsel,
1459 int nr_cpus, int cpu_idx,
1460 int nr_threads, int thread_idx)
1464 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1467 evlist__for_each_entry(evsel->evlist, pos) {
1468 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1470 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1473 * Since fds for next evsel has not been created,
1474 * there is no need to iterate whole event list.
1482 static bool ignore_missing_thread(struct evsel *evsel,
1483 int nr_cpus, int cpu,
1484 struct perf_thread_map *threads,
1485 int thread, int err)
1487 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1489 if (!evsel->ignore_missing_thread)
1492 /* The system wide setup does not work with threads. */
1493 if (evsel->core.system_wide)
1496 /* The -ESRCH is perf event syscall errno for pid's not found. */
1500 /* If there's only one thread, let it fail. */
1501 if (threads->nr == 1)
1505 * We should remove fd for missing_thread first
1506 * because thread_map__remove() will decrease threads->nr.
1508 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1511 if (thread_map__remove(threads, thread))
1514 pr_warning("WARNING: Ignored open failure for pid %d\n",
1519 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1520 void *priv __maybe_unused)
1522 return fprintf(fp, " %-32s %s\n", name, val);
1525 static void display_attr(struct perf_event_attr *attr)
1528 fprintf(stderr, "%.60s\n", graph_dotted_line);
1529 fprintf(stderr, "perf_event_attr:\n");
1530 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1531 fprintf(stderr, "%.60s\n", graph_dotted_line);
1535 static int perf_event_open(struct evsel *evsel,
1536 pid_t pid, int cpu, int group_fd,
1537 unsigned long flags)
1539 int precise_ip = evsel->core.attr.precise_ip;
1543 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1544 pid, cpu, group_fd, flags);
1546 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1550 /* Do not try less precise if not requested. */
1551 if (!evsel->precise_max)
1555 * We tried all the precise_ip values, and it's
1556 * still failing, so leave it to standard fallback.
1558 if (!evsel->core.attr.precise_ip) {
1559 evsel->core.attr.precise_ip = precise_ip;
1563 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1564 evsel->core.attr.precise_ip--;
1565 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1566 display_attr(&evsel->core.attr);
1572 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1573 struct perf_thread_map *threads)
1575 int cpu, thread, nthreads;
1576 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1578 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1580 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1581 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1585 static struct perf_cpu_map *empty_cpu_map;
1587 if (empty_cpu_map == NULL) {
1588 empty_cpu_map = perf_cpu_map__dummy_new();
1589 if (empty_cpu_map == NULL)
1593 cpus = empty_cpu_map;
1596 if (threads == NULL) {
1597 static struct perf_thread_map *empty_thread_map;
1599 if (empty_thread_map == NULL) {
1600 empty_thread_map = thread_map__new_by_tid(-1);
1601 if (empty_thread_map == NULL)
1605 threads = empty_thread_map;
1608 if (evsel->core.system_wide)
1611 nthreads = threads->nr;
1613 if (evsel->core.fd == NULL &&
1614 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1618 flags |= PERF_FLAG_PID_CGROUP;
1619 pid = evsel->cgrp->fd;
1622 fallback_missing_features:
1623 if (perf_missing_features.clockid_wrong)
1624 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1625 if (perf_missing_features.clockid) {
1626 evsel->core.attr.use_clockid = 0;
1627 evsel->core.attr.clockid = 0;
1629 if (perf_missing_features.cloexec)
1630 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1631 if (perf_missing_features.mmap2)
1632 evsel->core.attr.mmap2 = 0;
1633 if (perf_missing_features.exclude_guest)
1634 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1635 if (perf_missing_features.lbr_flags)
1636 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1637 PERF_SAMPLE_BRANCH_NO_CYCLES);
1638 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1639 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1640 if (perf_missing_features.ksymbol)
1641 evsel->core.attr.ksymbol = 0;
1642 if (perf_missing_features.bpf)
1643 evsel->core.attr.bpf_event = 0;
1645 if (perf_missing_features.sample_id_all)
1646 evsel->core.attr.sample_id_all = 0;
1648 display_attr(&evsel->core.attr);
1650 for (cpu = 0; cpu < cpus->nr; cpu++) {
1652 for (thread = 0; thread < nthreads; thread++) {
1655 if (!evsel->cgrp && !evsel->core.system_wide)
1656 pid = perf_thread_map__pid(threads, thread);
1658 group_fd = get_group_fd(evsel, cpu, thread);
1662 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1665 FD(evsel, cpu, thread) = fd;
1670 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1672 * We just removed 1 thread, so take a step
1673 * back on thread index and lower the upper
1679 /* ... and pretend like nothing have happened. */
1684 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1689 pr_debug2(" = %d\n", fd);
1691 if (evsel->bpf_fd >= 0) {
1693 int bpf_fd = evsel->bpf_fd;
1696 PERF_EVENT_IOC_SET_BPF,
1698 if (err && errno != EEXIST) {
1699 pr_err("failed to attach bpf fd %d: %s\n",
1700 bpf_fd, strerror(errno));
1706 set_rlimit = NO_CHANGE;
1709 * If we succeeded but had to kill clockid, fail and
1710 * have perf_evsel__open_strerror() print us a nice
1713 if (perf_missing_features.clockid ||
1714 perf_missing_features.clockid_wrong) {
1725 * perf stat needs between 5 and 22 fds per CPU. When we run out
1726 * of them try to increase the limits.
1728 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1730 int old_errno = errno;
1732 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1733 if (set_rlimit == NO_CHANGE)
1734 l.rlim_cur = l.rlim_max;
1736 l.rlim_cur = l.rlim_max + 1000;
1737 l.rlim_max = l.rlim_cur;
1739 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1748 if (err != -EINVAL || cpu > 0 || thread > 0)
1752 * Must probe features in the order they were added to the
1753 * perf_event_attr interface.
1755 if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1756 perf_missing_features.aux_output = true;
1757 pr_debug2("Kernel has no attr.aux_output support, bailing out\n");
1759 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1760 perf_missing_features.bpf = true;
1761 pr_debug2("switching off bpf_event\n");
1762 goto fallback_missing_features;
1763 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1764 perf_missing_features.ksymbol = true;
1765 pr_debug2("switching off ksymbol\n");
1766 goto fallback_missing_features;
1767 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1768 perf_missing_features.write_backward = true;
1769 pr_debug2("switching off write_backward\n");
1771 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1772 perf_missing_features.clockid_wrong = true;
1773 pr_debug2("switching off clockid\n");
1774 goto fallback_missing_features;
1775 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1776 perf_missing_features.clockid = true;
1777 pr_debug2("switching off use_clockid\n");
1778 goto fallback_missing_features;
1779 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1780 perf_missing_features.cloexec = true;
1781 pr_debug2("switching off cloexec flag\n");
1782 goto fallback_missing_features;
1783 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1784 perf_missing_features.mmap2 = true;
1785 pr_debug2("switching off mmap2\n");
1786 goto fallback_missing_features;
1787 } else if (!perf_missing_features.exclude_guest &&
1788 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1789 perf_missing_features.exclude_guest = true;
1790 pr_debug2("switching off exclude_guest, exclude_host\n");
1791 goto fallback_missing_features;
1792 } else if (!perf_missing_features.sample_id_all) {
1793 perf_missing_features.sample_id_all = true;
1794 pr_debug2("switching off sample_id_all\n");
1795 goto retry_sample_id;
1796 } else if (!perf_missing_features.lbr_flags &&
1797 (evsel->core.attr.branch_sample_type &
1798 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1799 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1800 perf_missing_features.lbr_flags = true;
1801 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1802 goto fallback_missing_features;
1803 } else if (!perf_missing_features.group_read &&
1804 evsel->core.attr.inherit &&
1805 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1806 perf_evsel__is_group_leader(evsel)) {
1807 perf_missing_features.group_read = true;
1808 pr_debug2("switching off group read\n");
1809 goto fallback_missing_features;
1813 threads->err_thread = thread;
1816 while (--thread >= 0) {
1817 close(FD(evsel, cpu, thread));
1818 FD(evsel, cpu, thread) = -1;
1821 } while (--cpu >= 0);
1825 void evsel__close(struct evsel *evsel)
1827 perf_evsel__close(&evsel->core);
1828 perf_evsel__free_id(&evsel->core);
1831 int perf_evsel__open_per_cpu(struct evsel *evsel,
1832 struct perf_cpu_map *cpus)
1834 return evsel__open(evsel, cpus, NULL);
1837 int perf_evsel__open_per_thread(struct evsel *evsel,
1838 struct perf_thread_map *threads)
1840 return evsel__open(evsel, NULL, threads);
1843 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1844 const union perf_event *event,
1845 struct perf_sample *sample)
1847 u64 type = evsel->core.attr.sample_type;
1848 const __u64 *array = event->sample.array;
1849 bool swapped = evsel->needs_swap;
1852 array += ((event->header.size -
1853 sizeof(event->header)) / sizeof(u64)) - 1;
1855 if (type & PERF_SAMPLE_IDENTIFIER) {
1856 sample->id = *array;
1860 if (type & PERF_SAMPLE_CPU) {
1863 /* undo swap of u64, then swap on individual u32s */
1864 u.val64 = bswap_64(u.val64);
1865 u.val32[0] = bswap_32(u.val32[0]);
1868 sample->cpu = u.val32[0];
1872 if (type & PERF_SAMPLE_STREAM_ID) {
1873 sample->stream_id = *array;
1877 if (type & PERF_SAMPLE_ID) {
1878 sample->id = *array;
1882 if (type & PERF_SAMPLE_TIME) {
1883 sample->time = *array;
1887 if (type & PERF_SAMPLE_TID) {
1890 /* undo swap of u64, then swap on individual u32s */
1891 u.val64 = bswap_64(u.val64);
1892 u.val32[0] = bswap_32(u.val32[0]);
1893 u.val32[1] = bswap_32(u.val32[1]);
1896 sample->pid = u.val32[0];
1897 sample->tid = u.val32[1];
1904 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1907 return size > max_size || offset + size > endp;
1910 #define OVERFLOW_CHECK(offset, size, max_size) \
1912 if (overflow(endp, (max_size), (offset), (size))) \
1916 #define OVERFLOW_CHECK_u64(offset) \
1917 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1920 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1923 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1924 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1925 * check the format does not go past the end of the event.
1927 if (sample_size + sizeof(event->header) > event->header.size)
1933 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1934 struct perf_sample *data)
1936 u64 type = evsel->core.attr.sample_type;
1937 bool swapped = evsel->needs_swap;
1939 u16 max_size = event->header.size;
1940 const void *endp = (void *)event + max_size;
1944 * used for cross-endian analysis. See git commit 65014ab3
1945 * for why this goofiness is needed.
1949 memset(data, 0, sizeof(*data));
1950 data->cpu = data->pid = data->tid = -1;
1951 data->stream_id = data->id = data->time = -1ULL;
1952 data->period = evsel->core.attr.sample_period;
1953 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1954 data->misc = event->header.misc;
1956 data->data_src = PERF_MEM_DATA_SRC_NONE;
1958 if (event->header.type != PERF_RECORD_SAMPLE) {
1959 if (!evsel->core.attr.sample_id_all)
1961 return perf_evsel__parse_id_sample(evsel, event, data);
1964 array = event->sample.array;
1966 if (perf_event__check_size(event, evsel->sample_size))
1969 if (type & PERF_SAMPLE_IDENTIFIER) {
1974 if (type & PERF_SAMPLE_IP) {
1979 if (type & PERF_SAMPLE_TID) {
1982 /* undo swap of u64, then swap on individual u32s */
1983 u.val64 = bswap_64(u.val64);
1984 u.val32[0] = bswap_32(u.val32[0]);
1985 u.val32[1] = bswap_32(u.val32[1]);
1988 data->pid = u.val32[0];
1989 data->tid = u.val32[1];
1993 if (type & PERF_SAMPLE_TIME) {
1994 data->time = *array;
1998 if (type & PERF_SAMPLE_ADDR) {
1999 data->addr = *array;
2003 if (type & PERF_SAMPLE_ID) {
2008 if (type & PERF_SAMPLE_STREAM_ID) {
2009 data->stream_id = *array;
2013 if (type & PERF_SAMPLE_CPU) {
2017 /* undo swap of u64, then swap on individual u32s */
2018 u.val64 = bswap_64(u.val64);
2019 u.val32[0] = bswap_32(u.val32[0]);
2022 data->cpu = u.val32[0];
2026 if (type & PERF_SAMPLE_PERIOD) {
2027 data->period = *array;
2031 if (type & PERF_SAMPLE_READ) {
2032 u64 read_format = evsel->core.attr.read_format;
2034 OVERFLOW_CHECK_u64(array);
2035 if (read_format & PERF_FORMAT_GROUP)
2036 data->read.group.nr = *array;
2038 data->read.one.value = *array;
2042 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2043 OVERFLOW_CHECK_u64(array);
2044 data->read.time_enabled = *array;
2048 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2049 OVERFLOW_CHECK_u64(array);
2050 data->read.time_running = *array;
2054 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2055 if (read_format & PERF_FORMAT_GROUP) {
2056 const u64 max_group_nr = UINT64_MAX /
2057 sizeof(struct sample_read_value);
2059 if (data->read.group.nr > max_group_nr)
2061 sz = data->read.group.nr *
2062 sizeof(struct sample_read_value);
2063 OVERFLOW_CHECK(array, sz, max_size);
2064 data->read.group.values =
2065 (struct sample_read_value *)array;
2066 array = (void *)array + sz;
2068 OVERFLOW_CHECK_u64(array);
2069 data->read.one.id = *array;
2074 if (evsel__has_callchain(evsel)) {
2075 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2077 OVERFLOW_CHECK_u64(array);
2078 data->callchain = (struct ip_callchain *)array++;
2079 if (data->callchain->nr > max_callchain_nr)
2081 sz = data->callchain->nr * sizeof(u64);
2082 OVERFLOW_CHECK(array, sz, max_size);
2083 array = (void *)array + sz;
2086 if (type & PERF_SAMPLE_RAW) {
2087 OVERFLOW_CHECK_u64(array);
2091 * Undo swap of u64, then swap on individual u32s,
2092 * get the size of the raw area and undo all of the
2093 * swap. The pevent interface handles endianity by
2097 u.val64 = bswap_64(u.val64);
2098 u.val32[0] = bswap_32(u.val32[0]);
2099 u.val32[1] = bswap_32(u.val32[1]);
2101 data->raw_size = u.val32[0];
2104 * The raw data is aligned on 64bits including the
2105 * u32 size, so it's safe to use mem_bswap_64.
2108 mem_bswap_64((void *) array, data->raw_size);
2110 array = (void *)array + sizeof(u32);
2112 OVERFLOW_CHECK(array, data->raw_size, max_size);
2113 data->raw_data = (void *)array;
2114 array = (void *)array + data->raw_size;
2117 if (type & PERF_SAMPLE_BRANCH_STACK) {
2118 const u64 max_branch_nr = UINT64_MAX /
2119 sizeof(struct branch_entry);
2121 OVERFLOW_CHECK_u64(array);
2122 data->branch_stack = (struct branch_stack *)array++;
2124 if (data->branch_stack->nr > max_branch_nr)
2126 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2127 OVERFLOW_CHECK(array, sz, max_size);
2128 array = (void *)array + sz;
2131 if (type & PERF_SAMPLE_REGS_USER) {
2132 OVERFLOW_CHECK_u64(array);
2133 data->user_regs.abi = *array;
2136 if (data->user_regs.abi) {
2137 u64 mask = evsel->core.attr.sample_regs_user;
2139 sz = hweight64(mask) * sizeof(u64);
2140 OVERFLOW_CHECK(array, sz, max_size);
2141 data->user_regs.mask = mask;
2142 data->user_regs.regs = (u64 *)array;
2143 array = (void *)array + sz;
2147 if (type & PERF_SAMPLE_STACK_USER) {
2148 OVERFLOW_CHECK_u64(array);
2151 data->user_stack.offset = ((char *)(array - 1)
2155 data->user_stack.size = 0;
2157 OVERFLOW_CHECK(array, sz, max_size);
2158 data->user_stack.data = (char *)array;
2159 array = (void *)array + sz;
2160 OVERFLOW_CHECK_u64(array);
2161 data->user_stack.size = *array++;
2162 if (WARN_ONCE(data->user_stack.size > sz,
2163 "user stack dump failure\n"))
2168 if (type & PERF_SAMPLE_WEIGHT) {
2169 OVERFLOW_CHECK_u64(array);
2170 data->weight = *array;
2174 if (type & PERF_SAMPLE_DATA_SRC) {
2175 OVERFLOW_CHECK_u64(array);
2176 data->data_src = *array;
2180 if (type & PERF_SAMPLE_TRANSACTION) {
2181 OVERFLOW_CHECK_u64(array);
2182 data->transaction = *array;
2186 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2187 if (type & PERF_SAMPLE_REGS_INTR) {
2188 OVERFLOW_CHECK_u64(array);
2189 data->intr_regs.abi = *array;
2192 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2193 u64 mask = evsel->core.attr.sample_regs_intr;
2195 sz = hweight64(mask) * sizeof(u64);
2196 OVERFLOW_CHECK(array, sz, max_size);
2197 data->intr_regs.mask = mask;
2198 data->intr_regs.regs = (u64 *)array;
2199 array = (void *)array + sz;
2203 data->phys_addr = 0;
2204 if (type & PERF_SAMPLE_PHYS_ADDR) {
2205 data->phys_addr = *array;
2212 int perf_evsel__parse_sample_timestamp(struct evsel *evsel,
2213 union perf_event *event,
2216 u64 type = evsel->core.attr.sample_type;
2219 if (!(type & PERF_SAMPLE_TIME))
2222 if (event->header.type != PERF_RECORD_SAMPLE) {
2223 struct perf_sample data = {
2227 if (!evsel->core.attr.sample_id_all)
2229 if (perf_evsel__parse_id_sample(evsel, event, &data))
2232 *timestamp = data.time;
2236 array = event->sample.array;
2238 if (perf_event__check_size(event, evsel->sample_size))
2241 if (type & PERF_SAMPLE_IDENTIFIER)
2244 if (type & PERF_SAMPLE_IP)
2247 if (type & PERF_SAMPLE_TID)
2250 if (type & PERF_SAMPLE_TIME)
2251 *timestamp = *array;
2256 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name)
2258 return tep_find_field(evsel->tp_format, name);
2261 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample,
2264 struct tep_format_field *field = perf_evsel__field(evsel, name);
2270 offset = field->offset;
2272 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2273 offset = *(int *)(sample->raw_data + field->offset);
2277 return sample->raw_data + offset;
2280 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2284 void *ptr = sample->raw_data + field->offset;
2286 switch (field->size) {
2290 value = *(u16 *)ptr;
2293 value = *(u32 *)ptr;
2296 memcpy(&value, ptr, sizeof(u64));
2305 switch (field->size) {
2307 return bswap_16(value);
2309 return bswap_32(value);
2311 return bswap_64(value);
2319 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample,
2322 struct tep_format_field *field = perf_evsel__field(evsel, name);
2327 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2330 bool perf_evsel__fallback(struct evsel *evsel, int err,
2331 char *msg, size_t msgsize)
2335 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2336 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2337 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2339 * If it's cycles then fall back to hrtimer based
2340 * cpu-clock-tick sw counter, which is always available even if
2343 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2346 scnprintf(msg, msgsize, "%s",
2347 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2349 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2350 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2352 zfree(&evsel->name);
2354 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2355 (paranoid = perf_event_paranoid()) > 1) {
2356 const char *name = perf_evsel__name(evsel);
2358 const char *sep = ":";
2360 /* Is there already the separator in the name. */
2361 if (strchr(name, '/') ||
2365 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2370 evsel->name = new_name;
2371 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2372 "to fall back to excluding kernel and hypervisor "
2373 " samples", paranoid);
2374 evsel->core.attr.exclude_kernel = 1;
2375 evsel->core.attr.exclude_hv = 1;
2383 static bool find_process(const char *name)
2385 size_t len = strlen(name);
2390 dir = opendir(procfs__mountpoint());
2394 /* Walk through the directory. */
2395 while (ret && (d = readdir(dir)) != NULL) {
2396 char path[PATH_MAX];
2400 if ((d->d_type != DT_DIR) ||
2401 !strcmp(".", d->d_name) ||
2402 !strcmp("..", d->d_name))
2405 scnprintf(path, sizeof(path), "%s/%s/comm",
2406 procfs__mountpoint(), d->d_name);
2408 if (filename__read_str(path, &data, &size))
2411 ret = strncmp(name, data, len);
2416 return ret ? false : true;
2419 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target,
2420 int err, char *msg, size_t size)
2422 char sbuf[STRERR_BUFSIZE];
2429 printed = scnprintf(msg, size,
2430 "No permission to enable %s event.\n\n",
2431 perf_evsel__name(evsel));
2433 return scnprintf(msg + printed, size - printed,
2434 "You may not have permission to collect %sstats.\n\n"
2435 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2436 "which controls use of the performance events system by\n"
2437 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2438 "The current value is %d:\n\n"
2439 " -1: Allow use of (almost) all events by all users\n"
2440 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2441 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2442 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2443 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2444 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2445 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2446 " kernel.perf_event_paranoid = -1\n" ,
2447 target->system_wide ? "system-wide " : "",
2448 perf_event_paranoid());
2450 return scnprintf(msg, size, "The %s event is not supported.",
2451 perf_evsel__name(evsel));
2453 return scnprintf(msg, size, "%s",
2454 "Too many events are opened.\n"
2455 "Probably the maximum number of open file descriptors has been reached.\n"
2456 "Hint: Try again after reducing the number of events.\n"
2457 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2459 if (evsel__has_callchain(evsel) &&
2460 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2461 return scnprintf(msg, size,
2462 "Not enough memory to setup event with callchain.\n"
2463 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2464 "Hint: Current value: %d", sysctl__max_stack());
2467 if (target->cpu_list)
2468 return scnprintf(msg, size, "%s",
2469 "No such device - did you specify an out-of-range profile CPU?");
2472 if (evsel->core.attr.sample_period != 0)
2473 return scnprintf(msg, size,
2474 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2475 perf_evsel__name(evsel));
2476 if (evsel->core.attr.precise_ip)
2477 return scnprintf(msg, size, "%s",
2478 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2479 #if defined(__i386__) || defined(__x86_64__)
2480 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2481 return scnprintf(msg, size, "%s",
2482 "No hardware sampling interrupt available.\n");
2486 if (find_process("oprofiled"))
2487 return scnprintf(msg, size,
2488 "The PMU counters are busy/taken by another profiler.\n"
2489 "We found oprofile daemon running, please stop it and try again.");
2492 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2493 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2494 if (perf_missing_features.clockid)
2495 return scnprintf(msg, size, "clockid feature not supported.");
2496 if (perf_missing_features.clockid_wrong)
2497 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2498 if (perf_missing_features.aux_output)
2499 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2505 return scnprintf(msg, size,
2506 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2507 "/bin/dmesg | grep -i perf may provide additional information.\n",
2508 err, str_error_r(err, sbuf, sizeof(sbuf)),
2509 perf_evsel__name(evsel));
2512 struct perf_env *perf_evsel__env(struct evsel *evsel)
2514 if (evsel && evsel->evlist)
2515 return evsel->evlist->env;
2519 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2523 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2524 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2526 int fd = FD(evsel, cpu, thread);
2528 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2529 cpu, thread, fd) < 0)
2537 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2539 struct perf_cpu_map *cpus = evsel->core.cpus;
2540 struct perf_thread_map *threads = evsel->core.threads;
2542 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2545 return store_evsel_ids(evsel, evlist);