2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
7 * Released under the GPL v2. (and only v2, not any later version)
11 #include <linux/bitops.h>
12 #include <api/fs/debugfs.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <sys/resource.h>
22 #include "thread_map.h"
24 #include "perf_regs.h"
26 #include "trace-event.h"
32 } perf_missing_features;
34 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
36 int __perf_evsel__sample_size(u64 sample_type)
38 u64 mask = sample_type & PERF_SAMPLE_MASK;
42 for (i = 0; i < 64; i++) {
43 if (mask & (1ULL << i))
53 * __perf_evsel__calc_id_pos - calculate id_pos.
54 * @sample_type: sample type
56 * This function returns the position of the event id (PERF_SAMPLE_ID or
57 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
60 static int __perf_evsel__calc_id_pos(u64 sample_type)
64 if (sample_type & PERF_SAMPLE_IDENTIFIER)
67 if (!(sample_type & PERF_SAMPLE_ID))
70 if (sample_type & PERF_SAMPLE_IP)
73 if (sample_type & PERF_SAMPLE_TID)
76 if (sample_type & PERF_SAMPLE_TIME)
79 if (sample_type & PERF_SAMPLE_ADDR)
86 * __perf_evsel__calc_is_pos - calculate is_pos.
87 * @sample_type: sample type
89 * This function returns the position (counting backwards) of the event id
90 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
91 * sample_id_all is used there is an id sample appended to non-sample events.
93 static int __perf_evsel__calc_is_pos(u64 sample_type)
97 if (sample_type & PERF_SAMPLE_IDENTIFIER)
100 if (!(sample_type & PERF_SAMPLE_ID))
103 if (sample_type & PERF_SAMPLE_CPU)
106 if (sample_type & PERF_SAMPLE_STREAM_ID)
112 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
114 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
115 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
118 void hists__init(struct hists *hists)
120 memset(hists, 0, sizeof(*hists));
121 hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
122 hists->entries_in = &hists->entries_in_array[0];
123 hists->entries_collapsed = RB_ROOT;
124 hists->entries = RB_ROOT;
125 pthread_mutex_init(&hists->lock, NULL);
128 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
129 enum perf_event_sample_format bit)
131 if (!(evsel->attr.sample_type & bit)) {
132 evsel->attr.sample_type |= bit;
133 evsel->sample_size += sizeof(u64);
134 perf_evsel__calc_id_pos(evsel);
138 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
139 enum perf_event_sample_format bit)
141 if (evsel->attr.sample_type & bit) {
142 evsel->attr.sample_type &= ~bit;
143 evsel->sample_size -= sizeof(u64);
144 perf_evsel__calc_id_pos(evsel);
148 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
149 bool can_sample_identifier)
151 if (can_sample_identifier) {
152 perf_evsel__reset_sample_bit(evsel, ID);
153 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
155 perf_evsel__set_sample_bit(evsel, ID);
157 evsel->attr.read_format |= PERF_FORMAT_ID;
160 void perf_evsel__init(struct perf_evsel *evsel,
161 struct perf_event_attr *attr, int idx)
165 evsel->leader = evsel;
168 INIT_LIST_HEAD(&evsel->node);
169 hists__init(&evsel->hists);
170 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
171 perf_evsel__calc_id_pos(evsel);
174 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
176 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
179 perf_evsel__init(evsel, attr, idx);
184 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
186 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
189 struct perf_event_attr attr = {
190 .type = PERF_TYPE_TRACEPOINT,
191 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
192 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
195 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
198 evsel->tp_format = trace_event__tp_format(sys, name);
199 if (evsel->tp_format == NULL)
202 event_attr_init(&attr);
203 attr.config = evsel->tp_format->id;
204 attr.sample_period = 1;
205 perf_evsel__init(evsel, &attr, idx);
216 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
224 "stalled-cycles-frontend",
225 "stalled-cycles-backend",
229 static const char *__perf_evsel__hw_name(u64 config)
231 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
232 return perf_evsel__hw_names[config];
234 return "unknown-hardware";
237 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
239 int colon = 0, r = 0;
240 struct perf_event_attr *attr = &evsel->attr;
241 bool exclude_guest_default = false;
243 #define MOD_PRINT(context, mod) do { \
244 if (!attr->exclude_##context) { \
245 if (!colon) colon = ++r; \
246 r += scnprintf(bf + r, size - r, "%c", mod); \
249 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
250 MOD_PRINT(kernel, 'k');
251 MOD_PRINT(user, 'u');
253 exclude_guest_default = true;
256 if (attr->precise_ip) {
259 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
260 exclude_guest_default = true;
263 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
264 MOD_PRINT(host, 'H');
265 MOD_PRINT(guest, 'G');
273 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
275 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
276 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
279 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
292 static const char *__perf_evsel__sw_name(u64 config)
294 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
295 return perf_evsel__sw_names[config];
296 return "unknown-software";
299 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
301 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
302 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
305 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
309 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
311 if (type & HW_BREAKPOINT_R)
312 r += scnprintf(bf + r, size - r, "r");
314 if (type & HW_BREAKPOINT_W)
315 r += scnprintf(bf + r, size - r, "w");
317 if (type & HW_BREAKPOINT_X)
318 r += scnprintf(bf + r, size - r, "x");
323 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
325 struct perf_event_attr *attr = &evsel->attr;
326 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
327 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
330 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
331 [PERF_EVSEL__MAX_ALIASES] = {
332 { "L1-dcache", "l1-d", "l1d", "L1-data", },
333 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
335 { "dTLB", "d-tlb", "Data-TLB", },
336 { "iTLB", "i-tlb", "Instruction-TLB", },
337 { "branch", "branches", "bpu", "btb", "bpc", },
341 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
342 [PERF_EVSEL__MAX_ALIASES] = {
343 { "load", "loads", "read", },
344 { "store", "stores", "write", },
345 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
348 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
349 [PERF_EVSEL__MAX_ALIASES] = {
350 { "refs", "Reference", "ops", "access", },
351 { "misses", "miss", },
354 #define C(x) PERF_COUNT_HW_CACHE_##x
355 #define CACHE_READ (1 << C(OP_READ))
356 #define CACHE_WRITE (1 << C(OP_WRITE))
357 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
358 #define COP(x) (1 << x)
361 * cache operartion stat
362 * L1I : Read and prefetch only
363 * ITLB and BPU : Read-only
365 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
366 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
367 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
368 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
369 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
370 [C(ITLB)] = (CACHE_READ),
371 [C(BPU)] = (CACHE_READ),
372 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
375 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
377 if (perf_evsel__hw_cache_stat[type] & COP(op))
378 return true; /* valid */
380 return false; /* invalid */
383 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
384 char *bf, size_t size)
387 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
388 perf_evsel__hw_cache_op[op][0],
389 perf_evsel__hw_cache_result[result][0]);
392 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
393 perf_evsel__hw_cache_op[op][1]);
396 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
398 u8 op, result, type = (config >> 0) & 0xff;
399 const char *err = "unknown-ext-hardware-cache-type";
401 if (type > PERF_COUNT_HW_CACHE_MAX)
404 op = (config >> 8) & 0xff;
405 err = "unknown-ext-hardware-cache-op";
406 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
409 result = (config >> 16) & 0xff;
410 err = "unknown-ext-hardware-cache-result";
411 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
414 err = "invalid-cache";
415 if (!perf_evsel__is_cache_op_valid(type, op))
418 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
420 return scnprintf(bf, size, "%s", err);
423 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
425 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
426 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
429 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
431 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
432 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
435 const char *perf_evsel__name(struct perf_evsel *evsel)
442 switch (evsel->attr.type) {
444 perf_evsel__raw_name(evsel, bf, sizeof(bf));
447 case PERF_TYPE_HARDWARE:
448 perf_evsel__hw_name(evsel, bf, sizeof(bf));
451 case PERF_TYPE_HW_CACHE:
452 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
455 case PERF_TYPE_SOFTWARE:
456 perf_evsel__sw_name(evsel, bf, sizeof(bf));
459 case PERF_TYPE_TRACEPOINT:
460 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
463 case PERF_TYPE_BREAKPOINT:
464 perf_evsel__bp_name(evsel, bf, sizeof(bf));
468 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
473 evsel->name = strdup(bf);
475 return evsel->name ?: "unknown";
478 const char *perf_evsel__group_name(struct perf_evsel *evsel)
480 return evsel->group_name ?: "anon group";
483 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
486 struct perf_evsel *pos;
487 const char *group_name = perf_evsel__group_name(evsel);
489 ret = scnprintf(buf, size, "%s", group_name);
491 ret += scnprintf(buf + ret, size - ret, " { %s",
492 perf_evsel__name(evsel));
494 for_each_group_member(pos, evsel)
495 ret += scnprintf(buf + ret, size - ret, ", %s",
496 perf_evsel__name(pos));
498 ret += scnprintf(buf + ret, size - ret, " }");
504 perf_evsel__config_callgraph(struct perf_evsel *evsel,
505 struct record_opts *opts)
507 bool function = perf_evsel__is_function_event(evsel);
508 struct perf_event_attr *attr = &evsel->attr;
510 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
512 if (opts->call_graph == CALLCHAIN_DWARF) {
514 perf_evsel__set_sample_bit(evsel, REGS_USER);
515 perf_evsel__set_sample_bit(evsel, STACK_USER);
516 attr->sample_regs_user = PERF_REGS_MASK;
517 attr->sample_stack_user = opts->stack_dump_size;
518 attr->exclude_callchain_user = 1;
520 pr_info("Cannot use DWARF unwind for function trace event,"
521 " falling back to framepointers.\n");
526 pr_info("Disabling user space callchains for function trace event.\n");
527 attr->exclude_callchain_user = 1;
532 * The enable_on_exec/disabled value strategy:
534 * 1) For any type of traced program:
535 * - all independent events and group leaders are disabled
536 * - all group members are enabled
538 * Group members are ruled by group leaders. They need to
539 * be enabled, because the group scheduling relies on that.
541 * 2) For traced programs executed by perf:
542 * - all independent events and group leaders have
544 * - we don't specifically enable or disable any event during
547 * Independent events and group leaders are initially disabled
548 * and get enabled by exec. Group members are ruled by group
549 * leaders as stated in 1).
551 * 3) For traced programs attached by perf (pid/tid):
552 * - we specifically enable or disable all events during
555 * When attaching events to already running traced we
556 * enable/disable events specifically, as there's no
557 * initial traced exec call.
559 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
561 struct perf_evsel *leader = evsel->leader;
562 struct perf_event_attr *attr = &evsel->attr;
563 int track = !evsel->idx; /* only the first counter needs these */
564 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
566 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
567 attr->inherit = !opts->no_inherit;
569 perf_evsel__set_sample_bit(evsel, IP);
570 perf_evsel__set_sample_bit(evsel, TID);
572 if (evsel->sample_read) {
573 perf_evsel__set_sample_bit(evsel, READ);
576 * We need ID even in case of single event, because
577 * PERF_SAMPLE_READ process ID specific data.
579 perf_evsel__set_sample_id(evsel, false);
582 * Apply group format only if we belong to group
583 * with more than one members.
585 if (leader->nr_members > 1) {
586 attr->read_format |= PERF_FORMAT_GROUP;
592 * We default some events to a 1 default interval. But keep
593 * it a weak assumption overridable by the user.
595 if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
596 opts->user_interval != ULLONG_MAX)) {
598 perf_evsel__set_sample_bit(evsel, PERIOD);
600 attr->sample_freq = opts->freq;
602 attr->sample_period = opts->default_interval;
607 * Disable sampling for all group members other
608 * than leader in case leader 'leads' the sampling.
610 if ((leader != evsel) && leader->sample_read) {
611 attr->sample_freq = 0;
612 attr->sample_period = 0;
615 if (opts->no_samples)
616 attr->sample_freq = 0;
618 if (opts->inherit_stat)
619 attr->inherit_stat = 1;
621 if (opts->sample_address) {
622 perf_evsel__set_sample_bit(evsel, ADDR);
623 attr->mmap_data = track;
626 if (opts->call_graph_enabled)
627 perf_evsel__config_callgraph(evsel, opts);
629 if (target__has_cpu(&opts->target))
630 perf_evsel__set_sample_bit(evsel, CPU);
633 perf_evsel__set_sample_bit(evsel, PERIOD);
635 if (!perf_missing_features.sample_id_all &&
636 (opts->sample_time || !opts->no_inherit ||
637 target__has_cpu(&opts->target) || per_cpu))
638 perf_evsel__set_sample_bit(evsel, TIME);
640 if (opts->raw_samples) {
641 perf_evsel__set_sample_bit(evsel, TIME);
642 perf_evsel__set_sample_bit(evsel, RAW);
643 perf_evsel__set_sample_bit(evsel, CPU);
646 if (opts->sample_address)
647 perf_evsel__set_sample_bit(evsel, DATA_SRC);
649 if (opts->no_buffering) {
651 attr->wakeup_events = 1;
653 if (opts->branch_stack) {
654 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
655 attr->branch_sample_type = opts->branch_stack;
658 if (opts->sample_weight)
659 perf_evsel__set_sample_bit(evsel, WEIGHT);
664 if (opts->sample_transaction)
665 perf_evsel__set_sample_bit(evsel, TRANSACTION);
668 * XXX see the function comment above
670 * Disabling only independent events or group leaders,
671 * keeping group members enabled.
673 if (perf_evsel__is_group_leader(evsel))
677 * Setting enable_on_exec for independent events and
678 * group leaders for traced executed by perf.
680 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
681 !opts->initial_delay)
682 attr->enable_on_exec = 1;
685 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
688 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
691 for (cpu = 0; cpu < ncpus; cpu++) {
692 for (thread = 0; thread < nthreads; thread++) {
693 FD(evsel, cpu, thread) = -1;
698 return evsel->fd != NULL ? 0 : -ENOMEM;
701 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
706 for (cpu = 0; cpu < ncpus; cpu++) {
707 for (thread = 0; thread < nthreads; thread++) {
708 int fd = FD(evsel, cpu, thread),
709 err = ioctl(fd, ioc, arg);
719 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
722 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
723 PERF_EVENT_IOC_SET_FILTER,
727 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
729 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
730 PERF_EVENT_IOC_ENABLE,
734 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
736 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
737 if (evsel->sample_id == NULL)
740 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
741 if (evsel->id == NULL) {
742 xyarray__delete(evsel->sample_id);
743 evsel->sample_id = NULL;
750 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
752 memset(evsel->counts, 0, (sizeof(*evsel->counts) +
753 (ncpus * sizeof(struct perf_counts_values))));
756 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
758 evsel->counts = zalloc((sizeof(*evsel->counts) +
759 (ncpus * sizeof(struct perf_counts_values))));
760 return evsel->counts != NULL ? 0 : -ENOMEM;
763 void perf_evsel__free_fd(struct perf_evsel *evsel)
765 xyarray__delete(evsel->fd);
769 void perf_evsel__free_id(struct perf_evsel *evsel)
771 xyarray__delete(evsel->sample_id);
772 evsel->sample_id = NULL;
776 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
780 for (cpu = 0; cpu < ncpus; cpu++)
781 for (thread = 0; thread < nthreads; ++thread) {
782 close(FD(evsel, cpu, thread));
783 FD(evsel, cpu, thread) = -1;
787 void perf_evsel__free_counts(struct perf_evsel *evsel)
789 zfree(&evsel->counts);
792 void perf_evsel__exit(struct perf_evsel *evsel)
794 assert(list_empty(&evsel->node));
795 perf_evsel__free_fd(evsel);
796 perf_evsel__free_id(evsel);
799 void perf_evsel__delete(struct perf_evsel *evsel)
801 perf_evsel__exit(evsel);
802 close_cgroup(evsel->cgrp);
803 zfree(&evsel->group_name);
804 if (evsel->tp_format)
805 pevent_free_format(evsel->tp_format);
810 static inline void compute_deltas(struct perf_evsel *evsel,
812 struct perf_counts_values *count)
814 struct perf_counts_values tmp;
816 if (!evsel->prev_raw_counts)
820 tmp = evsel->prev_raw_counts->aggr;
821 evsel->prev_raw_counts->aggr = *count;
823 tmp = evsel->prev_raw_counts->cpu[cpu];
824 evsel->prev_raw_counts->cpu[cpu] = *count;
827 count->val = count->val - tmp.val;
828 count->ena = count->ena - tmp.ena;
829 count->run = count->run - tmp.run;
832 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
833 int cpu, int thread, bool scale)
835 struct perf_counts_values count;
836 size_t nv = scale ? 3 : 1;
838 if (FD(evsel, cpu, thread) < 0)
841 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
844 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
847 compute_deltas(evsel, cpu, &count);
852 else if (count.run < count.ena)
853 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
855 count.ena = count.run = 0;
857 evsel->counts->cpu[cpu] = count;
861 int __perf_evsel__read(struct perf_evsel *evsel,
862 int ncpus, int nthreads, bool scale)
864 size_t nv = scale ? 3 : 1;
866 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
868 aggr->val = aggr->ena = aggr->run = 0;
870 for (cpu = 0; cpu < ncpus; cpu++) {
871 for (thread = 0; thread < nthreads; thread++) {
872 if (FD(evsel, cpu, thread) < 0)
875 if (readn(FD(evsel, cpu, thread),
876 &count, nv * sizeof(u64)) < 0)
879 aggr->val += count.val;
881 aggr->ena += count.ena;
882 aggr->run += count.run;
887 compute_deltas(evsel, -1, aggr);
889 evsel->counts->scaled = 0;
891 if (aggr->run == 0) {
892 evsel->counts->scaled = -1;
897 if (aggr->run < aggr->ena) {
898 evsel->counts->scaled = 1;
899 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
902 aggr->ena = aggr->run = 0;
907 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
909 struct perf_evsel *leader = evsel->leader;
912 if (perf_evsel__is_group_leader(evsel))
916 * Leader must be already processed/open,
921 fd = FD(leader, cpu, thread);
927 #define __PRINT_ATTR(fmt, cast, field) \
928 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
930 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
931 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
932 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
933 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
935 #define PRINT_ATTR2N(name1, field1, name2, field2) \
936 fprintf(fp, " %-19s %u %-19s %u\n", \
937 name1, attr->field1, name2, attr->field2)
939 #define PRINT_ATTR2(field1, field2) \
940 PRINT_ATTR2N(#field1, field1, #field2, field2)
942 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
946 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
947 ret += fprintf(fp, "perf_event_attr:\n");
949 ret += PRINT_ATTR_U32(type);
950 ret += PRINT_ATTR_U32(size);
951 ret += PRINT_ATTR_X64(config);
952 ret += PRINT_ATTR_U64(sample_period);
953 ret += PRINT_ATTR_U64(sample_freq);
954 ret += PRINT_ATTR_X64(sample_type);
955 ret += PRINT_ATTR_X64(read_format);
957 ret += PRINT_ATTR2(disabled, inherit);
958 ret += PRINT_ATTR2(pinned, exclusive);
959 ret += PRINT_ATTR2(exclude_user, exclude_kernel);
960 ret += PRINT_ATTR2(exclude_hv, exclude_idle);
961 ret += PRINT_ATTR2(mmap, comm);
962 ret += PRINT_ATTR2(freq, inherit_stat);
963 ret += PRINT_ATTR2(enable_on_exec, task);
964 ret += PRINT_ATTR2(watermark, precise_ip);
965 ret += PRINT_ATTR2(mmap_data, sample_id_all);
966 ret += PRINT_ATTR2(exclude_host, exclude_guest);
967 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
968 "excl.callchain_user", exclude_callchain_user);
969 ret += PRINT_ATTR_U32(mmap2);
971 ret += PRINT_ATTR_U32(wakeup_events);
972 ret += PRINT_ATTR_U32(wakeup_watermark);
973 ret += PRINT_ATTR_X32(bp_type);
974 ret += PRINT_ATTR_X64(bp_addr);
975 ret += PRINT_ATTR_X64(config1);
976 ret += PRINT_ATTR_U64(bp_len);
977 ret += PRINT_ATTR_X64(config2);
978 ret += PRINT_ATTR_X64(branch_sample_type);
979 ret += PRINT_ATTR_X64(sample_regs_user);
980 ret += PRINT_ATTR_U32(sample_stack_user);
982 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
987 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
988 struct thread_map *threads)
991 unsigned long flags = 0;
993 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
995 if (evsel->fd == NULL &&
996 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
1000 flags = PERF_FLAG_PID_CGROUP;
1001 pid = evsel->cgrp->fd;
1004 fallback_missing_features:
1005 if (perf_missing_features.mmap2)
1006 evsel->attr.mmap2 = 0;
1007 if (perf_missing_features.exclude_guest)
1008 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1010 if (perf_missing_features.sample_id_all)
1011 evsel->attr.sample_id_all = 0;
1014 perf_event_attr__fprintf(&evsel->attr, stderr);
1016 for (cpu = 0; cpu < cpus->nr; cpu++) {
1018 for (thread = 0; thread < threads->nr; thread++) {
1022 pid = threads->map[thread];
1024 group_fd = get_group_fd(evsel, cpu, thread);
1026 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1027 pid, cpus->map[cpu], group_fd, flags);
1029 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1033 if (FD(evsel, cpu, thread) < 0) {
1035 pr_debug2("sys_perf_event_open failed, error %d\n",
1039 set_rlimit = NO_CHANGE;
1047 * perf stat needs between 5 and 22 fds per CPU. When we run out
1048 * of them try to increase the limits.
1050 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1052 int old_errno = errno;
1054 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1055 if (set_rlimit == NO_CHANGE)
1056 l.rlim_cur = l.rlim_max;
1058 l.rlim_cur = l.rlim_max + 1000;
1059 l.rlim_max = l.rlim_cur;
1061 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1070 if (err != -EINVAL || cpu > 0 || thread > 0)
1073 if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1074 perf_missing_features.mmap2 = true;
1075 goto fallback_missing_features;
1076 } else if (!perf_missing_features.exclude_guest &&
1077 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1078 perf_missing_features.exclude_guest = true;
1079 goto fallback_missing_features;
1080 } else if (!perf_missing_features.sample_id_all) {
1081 perf_missing_features.sample_id_all = true;
1082 goto retry_sample_id;
1087 while (--thread >= 0) {
1088 close(FD(evsel, cpu, thread));
1089 FD(evsel, cpu, thread) = -1;
1091 thread = threads->nr;
1092 } while (--cpu >= 0);
1096 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1098 if (evsel->fd == NULL)
1101 perf_evsel__close_fd(evsel, ncpus, nthreads);
1102 perf_evsel__free_fd(evsel);
1114 struct thread_map map;
1116 } empty_thread_map = {
1121 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1122 struct thread_map *threads)
1125 /* Work around old compiler warnings about strict aliasing */
1126 cpus = &empty_cpu_map.map;
1129 if (threads == NULL)
1130 threads = &empty_thread_map.map;
1132 return __perf_evsel__open(evsel, cpus, threads);
1135 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1136 struct cpu_map *cpus)
1138 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1141 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1142 struct thread_map *threads)
1144 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1147 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1148 const union perf_event *event,
1149 struct perf_sample *sample)
1151 u64 type = evsel->attr.sample_type;
1152 const u64 *array = event->sample.array;
1153 bool swapped = evsel->needs_swap;
1156 array += ((event->header.size -
1157 sizeof(event->header)) / sizeof(u64)) - 1;
1159 if (type & PERF_SAMPLE_IDENTIFIER) {
1160 sample->id = *array;
1164 if (type & PERF_SAMPLE_CPU) {
1167 /* undo swap of u64, then swap on individual u32s */
1168 u.val64 = bswap_64(u.val64);
1169 u.val32[0] = bswap_32(u.val32[0]);
1172 sample->cpu = u.val32[0];
1176 if (type & PERF_SAMPLE_STREAM_ID) {
1177 sample->stream_id = *array;
1181 if (type & PERF_SAMPLE_ID) {
1182 sample->id = *array;
1186 if (type & PERF_SAMPLE_TIME) {
1187 sample->time = *array;
1191 if (type & PERF_SAMPLE_TID) {
1194 /* undo swap of u64, then swap on individual u32s */
1195 u.val64 = bswap_64(u.val64);
1196 u.val32[0] = bswap_32(u.val32[0]);
1197 u.val32[1] = bswap_32(u.val32[1]);
1200 sample->pid = u.val32[0];
1201 sample->tid = u.val32[1];
1208 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1211 return size > max_size || offset + size > endp;
1214 #define OVERFLOW_CHECK(offset, size, max_size) \
1216 if (overflow(endp, (max_size), (offset), (size))) \
1220 #define OVERFLOW_CHECK_u64(offset) \
1221 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1223 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1224 struct perf_sample *data)
1226 u64 type = evsel->attr.sample_type;
1227 bool swapped = evsel->needs_swap;
1229 u16 max_size = event->header.size;
1230 const void *endp = (void *)event + max_size;
1234 * used for cross-endian analysis. See git commit 65014ab3
1235 * for why this goofiness is needed.
1239 memset(data, 0, sizeof(*data));
1240 data->cpu = data->pid = data->tid = -1;
1241 data->stream_id = data->id = data->time = -1ULL;
1242 data->period = evsel->attr.sample_period;
1245 if (event->header.type != PERF_RECORD_SAMPLE) {
1246 if (!evsel->attr.sample_id_all)
1248 return perf_evsel__parse_id_sample(evsel, event, data);
1251 array = event->sample.array;
1254 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1255 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1256 * check the format does not go past the end of the event.
1258 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1262 if (type & PERF_SAMPLE_IDENTIFIER) {
1267 if (type & PERF_SAMPLE_IP) {
1272 if (type & PERF_SAMPLE_TID) {
1275 /* undo swap of u64, then swap on individual u32s */
1276 u.val64 = bswap_64(u.val64);
1277 u.val32[0] = bswap_32(u.val32[0]);
1278 u.val32[1] = bswap_32(u.val32[1]);
1281 data->pid = u.val32[0];
1282 data->tid = u.val32[1];
1286 if (type & PERF_SAMPLE_TIME) {
1287 data->time = *array;
1292 if (type & PERF_SAMPLE_ADDR) {
1293 data->addr = *array;
1297 if (type & PERF_SAMPLE_ID) {
1302 if (type & PERF_SAMPLE_STREAM_ID) {
1303 data->stream_id = *array;
1307 if (type & PERF_SAMPLE_CPU) {
1311 /* undo swap of u64, then swap on individual u32s */
1312 u.val64 = bswap_64(u.val64);
1313 u.val32[0] = bswap_32(u.val32[0]);
1316 data->cpu = u.val32[0];
1320 if (type & PERF_SAMPLE_PERIOD) {
1321 data->period = *array;
1325 if (type & PERF_SAMPLE_READ) {
1326 u64 read_format = evsel->attr.read_format;
1328 OVERFLOW_CHECK_u64(array);
1329 if (read_format & PERF_FORMAT_GROUP)
1330 data->read.group.nr = *array;
1332 data->read.one.value = *array;
1336 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1337 OVERFLOW_CHECK_u64(array);
1338 data->read.time_enabled = *array;
1342 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1343 OVERFLOW_CHECK_u64(array);
1344 data->read.time_running = *array;
1348 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1349 if (read_format & PERF_FORMAT_GROUP) {
1350 const u64 max_group_nr = UINT64_MAX /
1351 sizeof(struct sample_read_value);
1353 if (data->read.group.nr > max_group_nr)
1355 sz = data->read.group.nr *
1356 sizeof(struct sample_read_value);
1357 OVERFLOW_CHECK(array, sz, max_size);
1358 data->read.group.values =
1359 (struct sample_read_value *)array;
1360 array = (void *)array + sz;
1362 OVERFLOW_CHECK_u64(array);
1363 data->read.one.id = *array;
1368 if (type & PERF_SAMPLE_CALLCHAIN) {
1369 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1371 OVERFLOW_CHECK_u64(array);
1372 data->callchain = (struct ip_callchain *)array++;
1373 if (data->callchain->nr > max_callchain_nr)
1375 sz = data->callchain->nr * sizeof(u64);
1376 OVERFLOW_CHECK(array, sz, max_size);
1377 array = (void *)array + sz;
1380 if (type & PERF_SAMPLE_RAW) {
1381 OVERFLOW_CHECK_u64(array);
1383 if (WARN_ONCE(swapped,
1384 "Endianness of raw data not corrected!\n")) {
1385 /* undo swap of u64, then swap on individual u32s */
1386 u.val64 = bswap_64(u.val64);
1387 u.val32[0] = bswap_32(u.val32[0]);
1388 u.val32[1] = bswap_32(u.val32[1]);
1390 data->raw_size = u.val32[0];
1391 array = (void *)array + sizeof(u32);
1393 OVERFLOW_CHECK(array, data->raw_size, max_size);
1394 data->raw_data = (void *)array;
1395 array = (void *)array + data->raw_size;
1398 if (type & PERF_SAMPLE_BRANCH_STACK) {
1399 const u64 max_branch_nr = UINT64_MAX /
1400 sizeof(struct branch_entry);
1402 OVERFLOW_CHECK_u64(array);
1403 data->branch_stack = (struct branch_stack *)array++;
1405 if (data->branch_stack->nr > max_branch_nr)
1407 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1408 OVERFLOW_CHECK(array, sz, max_size);
1409 array = (void *)array + sz;
1412 if (type & PERF_SAMPLE_REGS_USER) {
1413 OVERFLOW_CHECK_u64(array);
1414 data->user_regs.abi = *array;
1417 if (data->user_regs.abi) {
1418 u64 mask = evsel->attr.sample_regs_user;
1420 sz = hweight_long(mask) * sizeof(u64);
1421 OVERFLOW_CHECK(array, sz, max_size);
1422 data->user_regs.mask = mask;
1423 data->user_regs.regs = (u64 *)array;
1424 array = (void *)array + sz;
1428 if (type & PERF_SAMPLE_STACK_USER) {
1429 OVERFLOW_CHECK_u64(array);
1432 data->user_stack.offset = ((char *)(array - 1)
1436 data->user_stack.size = 0;
1438 OVERFLOW_CHECK(array, sz, max_size);
1439 data->user_stack.data = (char *)array;
1440 array = (void *)array + sz;
1441 OVERFLOW_CHECK_u64(array);
1442 data->user_stack.size = *array++;
1443 if (WARN_ONCE(data->user_stack.size > sz,
1444 "user stack dump failure\n"))
1450 if (type & PERF_SAMPLE_WEIGHT) {
1451 OVERFLOW_CHECK_u64(array);
1452 data->weight = *array;
1456 data->data_src = PERF_MEM_DATA_SRC_NONE;
1457 if (type & PERF_SAMPLE_DATA_SRC) {
1458 OVERFLOW_CHECK_u64(array);
1459 data->data_src = *array;
1463 data->transaction = 0;
1464 if (type & PERF_SAMPLE_TRANSACTION) {
1465 OVERFLOW_CHECK_u64(array);
1466 data->transaction = *array;
1473 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1476 size_t sz, result = sizeof(struct sample_event);
1478 if (type & PERF_SAMPLE_IDENTIFIER)
1479 result += sizeof(u64);
1481 if (type & PERF_SAMPLE_IP)
1482 result += sizeof(u64);
1484 if (type & PERF_SAMPLE_TID)
1485 result += sizeof(u64);
1487 if (type & PERF_SAMPLE_TIME)
1488 result += sizeof(u64);
1490 if (type & PERF_SAMPLE_ADDR)
1491 result += sizeof(u64);
1493 if (type & PERF_SAMPLE_ID)
1494 result += sizeof(u64);
1496 if (type & PERF_SAMPLE_STREAM_ID)
1497 result += sizeof(u64);
1499 if (type & PERF_SAMPLE_CPU)
1500 result += sizeof(u64);
1502 if (type & PERF_SAMPLE_PERIOD)
1503 result += sizeof(u64);
1505 if (type & PERF_SAMPLE_READ) {
1506 result += sizeof(u64);
1507 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1508 result += sizeof(u64);
1509 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1510 result += sizeof(u64);
1511 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1512 if (read_format & PERF_FORMAT_GROUP) {
1513 sz = sample->read.group.nr *
1514 sizeof(struct sample_read_value);
1517 result += sizeof(u64);
1521 if (type & PERF_SAMPLE_CALLCHAIN) {
1522 sz = (sample->callchain->nr + 1) * sizeof(u64);
1526 if (type & PERF_SAMPLE_RAW) {
1527 result += sizeof(u32);
1528 result += sample->raw_size;
1531 if (type & PERF_SAMPLE_BRANCH_STACK) {
1532 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1537 if (type & PERF_SAMPLE_REGS_USER) {
1538 if (sample->user_regs.abi) {
1539 result += sizeof(u64);
1540 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1543 result += sizeof(u64);
1547 if (type & PERF_SAMPLE_STACK_USER) {
1548 sz = sample->user_stack.size;
1549 result += sizeof(u64);
1552 result += sizeof(u64);
1556 if (type & PERF_SAMPLE_WEIGHT)
1557 result += sizeof(u64);
1559 if (type & PERF_SAMPLE_DATA_SRC)
1560 result += sizeof(u64);
1562 if (type & PERF_SAMPLE_TRANSACTION)
1563 result += sizeof(u64);
1568 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1570 const struct perf_sample *sample,
1576 * used for cross-endian analysis. See git commit 65014ab3
1577 * for why this goofiness is needed.
1581 array = event->sample.array;
1583 if (type & PERF_SAMPLE_IDENTIFIER) {
1584 *array = sample->id;
1588 if (type & PERF_SAMPLE_IP) {
1589 *array = sample->ip;
1593 if (type & PERF_SAMPLE_TID) {
1594 u.val32[0] = sample->pid;
1595 u.val32[1] = sample->tid;
1598 * Inverse of what is done in perf_evsel__parse_sample
1600 u.val32[0] = bswap_32(u.val32[0]);
1601 u.val32[1] = bswap_32(u.val32[1]);
1602 u.val64 = bswap_64(u.val64);
1609 if (type & PERF_SAMPLE_TIME) {
1610 *array = sample->time;
1614 if (type & PERF_SAMPLE_ADDR) {
1615 *array = sample->addr;
1619 if (type & PERF_SAMPLE_ID) {
1620 *array = sample->id;
1624 if (type & PERF_SAMPLE_STREAM_ID) {
1625 *array = sample->stream_id;
1629 if (type & PERF_SAMPLE_CPU) {
1630 u.val32[0] = sample->cpu;
1633 * Inverse of what is done in perf_evsel__parse_sample
1635 u.val32[0] = bswap_32(u.val32[0]);
1636 u.val64 = bswap_64(u.val64);
1642 if (type & PERF_SAMPLE_PERIOD) {
1643 *array = sample->period;
1647 if (type & PERF_SAMPLE_READ) {
1648 if (read_format & PERF_FORMAT_GROUP)
1649 *array = sample->read.group.nr;
1651 *array = sample->read.one.value;
1654 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1655 *array = sample->read.time_enabled;
1659 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1660 *array = sample->read.time_running;
1664 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1665 if (read_format & PERF_FORMAT_GROUP) {
1666 sz = sample->read.group.nr *
1667 sizeof(struct sample_read_value);
1668 memcpy(array, sample->read.group.values, sz);
1669 array = (void *)array + sz;
1671 *array = sample->read.one.id;
1676 if (type & PERF_SAMPLE_CALLCHAIN) {
1677 sz = (sample->callchain->nr + 1) * sizeof(u64);
1678 memcpy(array, sample->callchain, sz);
1679 array = (void *)array + sz;
1682 if (type & PERF_SAMPLE_RAW) {
1683 u.val32[0] = sample->raw_size;
1684 if (WARN_ONCE(swapped,
1685 "Endianness of raw data not corrected!\n")) {
1687 * Inverse of what is done in perf_evsel__parse_sample
1689 u.val32[0] = bswap_32(u.val32[0]);
1690 u.val32[1] = bswap_32(u.val32[1]);
1691 u.val64 = bswap_64(u.val64);
1694 array = (void *)array + sizeof(u32);
1696 memcpy(array, sample->raw_data, sample->raw_size);
1697 array = (void *)array + sample->raw_size;
1700 if (type & PERF_SAMPLE_BRANCH_STACK) {
1701 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1703 memcpy(array, sample->branch_stack, sz);
1704 array = (void *)array + sz;
1707 if (type & PERF_SAMPLE_REGS_USER) {
1708 if (sample->user_regs.abi) {
1709 *array++ = sample->user_regs.abi;
1710 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1711 memcpy(array, sample->user_regs.regs, sz);
1712 array = (void *)array + sz;
1718 if (type & PERF_SAMPLE_STACK_USER) {
1719 sz = sample->user_stack.size;
1722 memcpy(array, sample->user_stack.data, sz);
1723 array = (void *)array + sz;
1728 if (type & PERF_SAMPLE_WEIGHT) {
1729 *array = sample->weight;
1733 if (type & PERF_SAMPLE_DATA_SRC) {
1734 *array = sample->data_src;
1738 if (type & PERF_SAMPLE_TRANSACTION) {
1739 *array = sample->transaction;
1746 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1748 return pevent_find_field(evsel->tp_format, name);
1751 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1754 struct format_field *field = perf_evsel__field(evsel, name);
1760 offset = field->offset;
1762 if (field->flags & FIELD_IS_DYNAMIC) {
1763 offset = *(int *)(sample->raw_data + field->offset);
1767 return sample->raw_data + offset;
1770 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1773 struct format_field *field = perf_evsel__field(evsel, name);
1780 ptr = sample->raw_data + field->offset;
1782 switch (field->size) {
1786 value = *(u16 *)ptr;
1789 value = *(u32 *)ptr;
1792 value = *(u64 *)ptr;
1798 if (!evsel->needs_swap)
1801 switch (field->size) {
1803 return bswap_16(value);
1805 return bswap_32(value);
1807 return bswap_64(value);
1815 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1821 ret += fprintf(fp, ",");
1823 ret += fprintf(fp, ":");
1827 va_start(args, fmt);
1828 ret += vfprintf(fp, fmt, args);
1833 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1838 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1841 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1848 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1849 struct bit_names *bits, bool *first)
1851 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1852 bool first_bit = true;
1855 if (value & bits[i].bit) {
1856 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1859 } while (bits[++i].name != NULL);
1864 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1866 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1867 struct bit_names bits[] = {
1868 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1869 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1870 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1871 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1872 bit_name(IDENTIFIER),
1876 return bits__fprintf(fp, "sample_type", value, bits, first);
1879 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1881 #define bit_name(n) { PERF_FORMAT_##n, #n }
1882 struct bit_names bits[] = {
1883 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1884 bit_name(ID), bit_name(GROUP),
1888 return bits__fprintf(fp, "read_format", value, bits, first);
1891 int perf_evsel__fprintf(struct perf_evsel *evsel,
1892 struct perf_attr_details *details, FILE *fp)
1897 if (details->event_group) {
1898 struct perf_evsel *pos;
1900 if (!perf_evsel__is_group_leader(evsel))
1903 if (evsel->nr_members > 1)
1904 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1906 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1907 for_each_group_member(pos, evsel)
1908 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1910 if (evsel->nr_members > 1)
1911 printed += fprintf(fp, "}");
1915 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1917 if (details->verbose || details->freq) {
1918 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1919 (u64)evsel->attr.sample_freq);
1922 if (details->verbose) {
1928 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1929 if (evsel->attr.read_format)
1930 printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1934 if_print(exclusive);
1935 if_print(exclude_user);
1936 if_print(exclude_kernel);
1937 if_print(exclude_hv);
1938 if_print(exclude_idle);
1943 if_print(inherit_stat);
1944 if_print(enable_on_exec);
1946 if_print(watermark);
1947 if_print(precise_ip);
1948 if_print(mmap_data);
1949 if_print(sample_id_all);
1950 if_print(exclude_host);
1951 if_print(exclude_guest);
1952 if_print(__reserved_1);
1953 if_print(wakeup_events);
1955 if_print(branch_sample_type);
1962 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1963 char *msg, size_t msgsize)
1965 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
1966 evsel->attr.type == PERF_TYPE_HARDWARE &&
1967 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1969 * If it's cycles then fall back to hrtimer based
1970 * cpu-clock-tick sw counter, which is always available even if
1973 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1976 scnprintf(msg, msgsize, "%s",
1977 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1979 evsel->attr.type = PERF_TYPE_SOFTWARE;
1980 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
1982 zfree(&evsel->name);
1989 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
1990 int err, char *msg, size_t size)
1995 return scnprintf(msg, size,
1996 "You may not have permission to collect %sstats.\n"
1997 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
1998 " -1 - Not paranoid at all\n"
1999 " 0 - Disallow raw tracepoint access for unpriv\n"
2000 " 1 - Disallow cpu events for unpriv\n"
2001 " 2 - Disallow kernel profiling for unpriv",
2002 target->system_wide ? "system-wide " : "");
2004 return scnprintf(msg, size, "The %s event is not supported.",
2005 perf_evsel__name(evsel));
2007 return scnprintf(msg, size, "%s",
2008 "Too many events are opened.\n"
2009 "Try again after reducing the number of events.");
2011 if (target->cpu_list)
2012 return scnprintf(msg, size, "%s",
2013 "No such device - did you specify an out-of-range profile CPU?\n");
2016 if (evsel->attr.precise_ip)
2017 return scnprintf(msg, size, "%s",
2018 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2019 #if defined(__i386__) || defined(__x86_64__)
2020 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2021 return scnprintf(msg, size, "%s",
2022 "No hardware sampling interrupt available.\n"
2023 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2030 return scnprintf(msg, size,
2031 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2032 "/bin/dmesg may provide additional information.\n"
2033 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2034 err, strerror(err), perf_evsel__name(evsel));