1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 *
5 * Parts came from builtin-{top,stat,record}.c, see those files for further
6 * copyright notes.
7 */
8
9 #include <byteswap.h>
10 #include <errno.h>
11 #include <inttypes.h>
12 #include <linux/bitops.h>
13 #include <api/io.h>
14 #include <api/fs/fs.h>
15 #include <api/fs/tracing_path.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>
24 #include <dirent.h>
25 #include <stdlib.h>
26 #include <perf/evsel.h>
27 #include "asm/bug.h"
28 #include "bpf_counter.h"
29 #include "callchain.h"
30 #include "cgroup.h"
31 #include "counts.h"
32 #include "event.h"
33 #include "evsel.h"
34 #include "time-utils.h"
35 #include "util/env.h"
36 #include "util/evsel_config.h"
37 #include "util/evsel_fprintf.h"
38 #include "evlist.h"
39 #include <perf/cpumap.h>
40 #include "thread_map.h"
41 #include "target.h"
42 #include "perf_regs.h"
43 #include "record.h"
44 #include "debug.h"
45 #include "trace-event.h"
46 #include "stat.h"
47 #include "string2.h"
48 #include "memswap.h"
49 #include "util.h"
50 #include "util/hashmap.h"
51 #include "off_cpu.h"
52 #include "pmu.h"
53 #include "pmus.h"
54 #include "rlimit.h"
55 #include "../perf-sys.h"
56 #include "util/parse-branch-options.h"
57 #include "util/bpf-filter.h"
58 #include "util/hist.h"
59 #include <internal/xyarray.h>
60 #include <internal/lib.h>
61 #include <internal/threadmap.h>
62
63 #include <linux/ctype.h>
64
65 #ifdef HAVE_LIBTRACEEVENT
66 #include <traceevent/event-parse.h>
67 #endif
68
69 struct perf_missing_features perf_missing_features;
70
71 static clockid_t clockid;
72
73 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
74 NULL,
75 "duration_time",
76 "user_time",
77 "system_time",
78 };
79
perf_tool_event__to_str(enum perf_tool_event ev)80 const char *perf_tool_event__to_str(enum perf_tool_event ev)
81 {
82 if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
83 return perf_tool_event__tool_names[ev];
84
85 return NULL;
86 }
87
perf_tool_event__from_str(const char * str)88 enum perf_tool_event perf_tool_event__from_str(const char *str)
89 {
90 int i;
91
92 perf_tool_event__for_each_event(i) {
93 if (!strcmp(str, perf_tool_event__tool_names[i]))
94 return i;
95 }
96 return PERF_TOOL_NONE;
97 }
98
99
evsel__no_extra_init(struct evsel * evsel __maybe_unused)100 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
101 {
102 return 0;
103 }
104
test_attr__ready(void)105 void __weak test_attr__ready(void) { }
106
evsel__no_extra_fini(struct evsel * evsel __maybe_unused)107 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
108 {
109 }
110
111 static struct {
112 size_t size;
113 int (*init)(struct evsel *evsel);
114 void (*fini)(struct evsel *evsel);
115 } perf_evsel__object = {
116 .size = sizeof(struct evsel),
117 .init = evsel__no_extra_init,
118 .fini = evsel__no_extra_fini,
119 };
120
evsel__object_config(size_t object_size,int (* init)(struct evsel * evsel),void (* fini)(struct evsel * evsel))121 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
122 void (*fini)(struct evsel *evsel))
123 {
124
125 if (object_size == 0)
126 goto set_methods;
127
128 if (perf_evsel__object.size > object_size)
129 return -EINVAL;
130
131 perf_evsel__object.size = object_size;
132
133 set_methods:
134 if (init != NULL)
135 perf_evsel__object.init = init;
136
137 if (fini != NULL)
138 perf_evsel__object.fini = fini;
139
140 return 0;
141 }
142
143 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
144
__evsel__sample_size(u64 sample_type)145 int __evsel__sample_size(u64 sample_type)
146 {
147 u64 mask = sample_type & PERF_SAMPLE_MASK;
148 int size = 0;
149 int i;
150
151 for (i = 0; i < 64; i++) {
152 if (mask & (1ULL << i))
153 size++;
154 }
155
156 size *= sizeof(u64);
157
158 return size;
159 }
160
161 /**
162 * __perf_evsel__calc_id_pos - calculate id_pos.
163 * @sample_type: sample type
164 *
165 * This function returns the position of the event id (PERF_SAMPLE_ID or
166 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
167 * perf_record_sample.
168 */
__perf_evsel__calc_id_pos(u64 sample_type)169 static int __perf_evsel__calc_id_pos(u64 sample_type)
170 {
171 int idx = 0;
172
173 if (sample_type & PERF_SAMPLE_IDENTIFIER)
174 return 0;
175
176 if (!(sample_type & PERF_SAMPLE_ID))
177 return -1;
178
179 if (sample_type & PERF_SAMPLE_IP)
180 idx += 1;
181
182 if (sample_type & PERF_SAMPLE_TID)
183 idx += 1;
184
185 if (sample_type & PERF_SAMPLE_TIME)
186 idx += 1;
187
188 if (sample_type & PERF_SAMPLE_ADDR)
189 idx += 1;
190
191 return idx;
192 }
193
194 /**
195 * __perf_evsel__calc_is_pos - calculate is_pos.
196 * @sample_type: sample type
197 *
198 * This function returns the position (counting backwards) of the event id
199 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
200 * sample_id_all is used there is an id sample appended to non-sample events.
201 */
__perf_evsel__calc_is_pos(u64 sample_type)202 static int __perf_evsel__calc_is_pos(u64 sample_type)
203 {
204 int idx = 1;
205
206 if (sample_type & PERF_SAMPLE_IDENTIFIER)
207 return 1;
208
209 if (!(sample_type & PERF_SAMPLE_ID))
210 return -1;
211
212 if (sample_type & PERF_SAMPLE_CPU)
213 idx += 1;
214
215 if (sample_type & PERF_SAMPLE_STREAM_ID)
216 idx += 1;
217
218 return idx;
219 }
220
evsel__calc_id_pos(struct evsel * evsel)221 void evsel__calc_id_pos(struct evsel *evsel)
222 {
223 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
224 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
225 }
226
__evsel__set_sample_bit(struct evsel * evsel,enum perf_event_sample_format bit)227 void __evsel__set_sample_bit(struct evsel *evsel,
228 enum perf_event_sample_format bit)
229 {
230 if (!(evsel->core.attr.sample_type & bit)) {
231 evsel->core.attr.sample_type |= bit;
232 evsel->sample_size += sizeof(u64);
233 evsel__calc_id_pos(evsel);
234 }
235 }
236
__evsel__reset_sample_bit(struct evsel * evsel,enum perf_event_sample_format bit)237 void __evsel__reset_sample_bit(struct evsel *evsel,
238 enum perf_event_sample_format bit)
239 {
240 if (evsel->core.attr.sample_type & bit) {
241 evsel->core.attr.sample_type &= ~bit;
242 evsel->sample_size -= sizeof(u64);
243 evsel__calc_id_pos(evsel);
244 }
245 }
246
evsel__set_sample_id(struct evsel * evsel,bool can_sample_identifier)247 void evsel__set_sample_id(struct evsel *evsel,
248 bool can_sample_identifier)
249 {
250 if (can_sample_identifier) {
251 evsel__reset_sample_bit(evsel, ID);
252 evsel__set_sample_bit(evsel, IDENTIFIER);
253 } else {
254 evsel__set_sample_bit(evsel, ID);
255 }
256 evsel->core.attr.read_format |= PERF_FORMAT_ID;
257 }
258
259 /**
260 * evsel__is_function_event - Return whether given evsel is a function
261 * trace event
262 *
263 * @evsel - evsel selector to be tested
264 *
265 * Return %true if event is function trace event
266 */
evsel__is_function_event(struct evsel * evsel)267 bool evsel__is_function_event(struct evsel *evsel)
268 {
269 #define FUNCTION_EVENT "ftrace:function"
270
271 return evsel->name &&
272 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
273
274 #undef FUNCTION_EVENT
275 }
276
evsel__init(struct evsel * evsel,struct perf_event_attr * attr,int idx)277 void evsel__init(struct evsel *evsel,
278 struct perf_event_attr *attr, int idx)
279 {
280 perf_evsel__init(&evsel->core, attr, idx);
281 evsel->tracking = !idx;
282 evsel->unit = strdup("");
283 evsel->scale = 1.0;
284 evsel->max_events = ULONG_MAX;
285 evsel->evlist = NULL;
286 evsel->bpf_obj = NULL;
287 evsel->bpf_fd = -1;
288 INIT_LIST_HEAD(&evsel->config_terms);
289 INIT_LIST_HEAD(&evsel->bpf_counter_list);
290 INIT_LIST_HEAD(&evsel->bpf_filters);
291 perf_evsel__object.init(evsel);
292 evsel->sample_size = __evsel__sample_size(attr->sample_type);
293 evsel__calc_id_pos(evsel);
294 evsel->cmdline_group_boundary = false;
295 evsel->metric_events = NULL;
296 evsel->per_pkg_mask = NULL;
297 evsel->collect_stat = false;
298 evsel->pmu_name = NULL;
299 evsel->group_pmu_name = NULL;
300 evsel->skippable = false;
301 }
302
evsel__new_idx(struct perf_event_attr * attr,int idx)303 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
304 {
305 struct evsel *evsel = zalloc(perf_evsel__object.size);
306
307 if (!evsel)
308 return NULL;
309 evsel__init(evsel, attr, idx);
310
311 if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
312 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
313 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
314 evsel->core.attr.sample_period = 1;
315 }
316
317 if (evsel__is_clock(evsel)) {
318 free((char *)evsel->unit);
319 evsel->unit = strdup("msec");
320 evsel->scale = 1e-6;
321 }
322
323 return evsel;
324 }
325
copy_config_terms(struct list_head * dst,struct list_head * src)326 int copy_config_terms(struct list_head *dst, struct list_head *src)
327 {
328 struct evsel_config_term *pos, *tmp;
329
330 list_for_each_entry(pos, src, list) {
331 tmp = malloc(sizeof(*tmp));
332 if (tmp == NULL)
333 return -ENOMEM;
334
335 *tmp = *pos;
336 if (tmp->free_str) {
337 tmp->val.str = strdup(pos->val.str);
338 if (tmp->val.str == NULL) {
339 free(tmp);
340 return -ENOMEM;
341 }
342 }
343 list_add_tail(&tmp->list, dst);
344 }
345 return 0;
346 }
347
evsel__copy_config_terms(struct evsel * dst,struct evsel * src)348 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
349 {
350 return copy_config_terms(&dst->config_terms, &src->config_terms);
351 }
352
353 /**
354 * evsel__clone - create a new evsel copied from @orig
355 * @orig: original evsel
356 *
357 * The assumption is that @orig is not configured nor opened yet.
358 * So we only care about the attributes that can be set while it's parsed.
359 */
evsel__clone(struct evsel * orig)360 struct evsel *evsel__clone(struct evsel *orig)
361 {
362 struct evsel *evsel;
363
364 BUG_ON(orig->core.fd);
365 BUG_ON(orig->counts);
366 BUG_ON(orig->priv);
367 BUG_ON(orig->per_pkg_mask);
368
369 /* cannot handle BPF objects for now */
370 if (orig->bpf_obj)
371 return NULL;
372
373 evsel = evsel__new(&orig->core.attr);
374 if (evsel == NULL)
375 return NULL;
376
377 evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
378 evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
379 evsel->core.threads = perf_thread_map__get(orig->core.threads);
380 evsel->core.nr_members = orig->core.nr_members;
381 evsel->core.system_wide = orig->core.system_wide;
382 evsel->core.requires_cpu = orig->core.requires_cpu;
383 evsel->core.is_pmu_core = orig->core.is_pmu_core;
384
385 if (orig->name) {
386 evsel->name = strdup(orig->name);
387 if (evsel->name == NULL)
388 goto out_err;
389 }
390 if (orig->group_name) {
391 evsel->group_name = strdup(orig->group_name);
392 if (evsel->group_name == NULL)
393 goto out_err;
394 }
395 if (orig->pmu_name) {
396 evsel->pmu_name = strdup(orig->pmu_name);
397 if (evsel->pmu_name == NULL)
398 goto out_err;
399 }
400 if (orig->group_pmu_name) {
401 evsel->group_pmu_name = strdup(orig->group_pmu_name);
402 if (evsel->group_pmu_name == NULL)
403 goto out_err;
404 }
405 if (orig->filter) {
406 evsel->filter = strdup(orig->filter);
407 if (evsel->filter == NULL)
408 goto out_err;
409 }
410 if (orig->metric_id) {
411 evsel->metric_id = strdup(orig->metric_id);
412 if (evsel->metric_id == NULL)
413 goto out_err;
414 }
415 evsel->cgrp = cgroup__get(orig->cgrp);
416 #ifdef HAVE_LIBTRACEEVENT
417 evsel->tp_format = orig->tp_format;
418 #endif
419 evsel->handler = orig->handler;
420 evsel->core.leader = orig->core.leader;
421
422 evsel->max_events = orig->max_events;
423 evsel->tool_event = orig->tool_event;
424 free((char *)evsel->unit);
425 evsel->unit = strdup(orig->unit);
426 if (evsel->unit == NULL)
427 goto out_err;
428
429 evsel->scale = orig->scale;
430 evsel->snapshot = orig->snapshot;
431 evsel->per_pkg = orig->per_pkg;
432 evsel->percore = orig->percore;
433 evsel->precise_max = orig->precise_max;
434 evsel->is_libpfm_event = orig->is_libpfm_event;
435
436 evsel->exclude_GH = orig->exclude_GH;
437 evsel->sample_read = orig->sample_read;
438 evsel->auto_merge_stats = orig->auto_merge_stats;
439 evsel->collect_stat = orig->collect_stat;
440 evsel->weak_group = orig->weak_group;
441 evsel->use_config_name = orig->use_config_name;
442 evsel->pmu = orig->pmu;
443
444 if (evsel__copy_config_terms(evsel, orig) < 0)
445 goto out_err;
446
447 return evsel;
448
449 out_err:
450 evsel__delete(evsel);
451 return NULL;
452 }
453
454 /*
455 * Returns pointer with encoded error via <linux/err.h> interface.
456 */
457 #ifdef HAVE_LIBTRACEEVENT
evsel__newtp_idx(const char * sys,const char * name,int idx,bool format)458 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx, bool format)
459 {
460 struct evsel *evsel = zalloc(perf_evsel__object.size);
461 int err = -ENOMEM;
462
463 if (evsel == NULL) {
464 goto out_err;
465 } else {
466 struct perf_event_attr attr = {
467 .type = PERF_TYPE_TRACEPOINT,
468 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
469 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
470 };
471
472 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
473 goto out_free;
474
475 event_attr_init(&attr);
476
477 if (format) {
478 evsel->tp_format = trace_event__tp_format(sys, name);
479 if (IS_ERR(evsel->tp_format)) {
480 err = PTR_ERR(evsel->tp_format);
481 goto out_free;
482 }
483 attr.config = evsel->tp_format->id;
484 } else {
485 attr.config = (__u64) -1;
486 }
487
488
489 attr.sample_period = 1;
490 evsel__init(evsel, &attr, idx);
491 }
492
493 return evsel;
494
495 out_free:
496 zfree(&evsel->name);
497 free(evsel);
498 out_err:
499 return ERR_PTR(err);
500 }
501 #endif
502
503 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
504 "cycles",
505 "instructions",
506 "cache-references",
507 "cache-misses",
508 "branches",
509 "branch-misses",
510 "bus-cycles",
511 "stalled-cycles-frontend",
512 "stalled-cycles-backend",
513 "ref-cycles",
514 };
515
516 char *evsel__bpf_counter_events;
517
evsel__match_bpf_counter_events(const char * name)518 bool evsel__match_bpf_counter_events(const char *name)
519 {
520 int name_len;
521 bool match;
522 char *ptr;
523
524 if (!evsel__bpf_counter_events)
525 return false;
526
527 ptr = strstr(evsel__bpf_counter_events, name);
528 name_len = strlen(name);
529
530 /* check name matches a full token in evsel__bpf_counter_events */
531 match = (ptr != NULL) &&
532 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
533 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
534
535 return match;
536 }
537
__evsel__hw_name(u64 config)538 static const char *__evsel__hw_name(u64 config)
539 {
540 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
541 return evsel__hw_names[config];
542
543 return "unknown-hardware";
544 }
545
evsel__add_modifiers(struct evsel * evsel,char * bf,size_t size)546 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
547 {
548 int colon = 0, r = 0;
549 struct perf_event_attr *attr = &evsel->core.attr;
550 bool exclude_guest_default = false;
551
552 #define MOD_PRINT(context, mod) do { \
553 if (!attr->exclude_##context) { \
554 if (!colon) colon = ++r; \
555 r += scnprintf(bf + r, size - r, "%c", mod); \
556 } } while(0)
557
558 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
559 MOD_PRINT(kernel, 'k');
560 MOD_PRINT(user, 'u');
561 MOD_PRINT(hv, 'h');
562 exclude_guest_default = true;
563 }
564
565 if (attr->precise_ip) {
566 if (!colon)
567 colon = ++r;
568 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
569 exclude_guest_default = true;
570 }
571
572 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
573 MOD_PRINT(host, 'H');
574 MOD_PRINT(guest, 'G');
575 }
576 #undef MOD_PRINT
577 if (colon)
578 bf[colon - 1] = ':';
579 return r;
580 }
581
arch_evsel__hw_name(struct evsel * evsel,char * bf,size_t size)582 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
583 {
584 return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
585 }
586
evsel__hw_name(struct evsel * evsel,char * bf,size_t size)587 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
588 {
589 int r = arch_evsel__hw_name(evsel, bf, size);
590 return r + evsel__add_modifiers(evsel, bf + r, size - r);
591 }
592
593 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
594 "cpu-clock",
595 "task-clock",
596 "page-faults",
597 "context-switches",
598 "cpu-migrations",
599 "minor-faults",
600 "major-faults",
601 "alignment-faults",
602 "emulation-faults",
603 "dummy",
604 };
605
__evsel__sw_name(u64 config)606 static const char *__evsel__sw_name(u64 config)
607 {
608 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
609 return evsel__sw_names[config];
610 return "unknown-software";
611 }
612
evsel__sw_name(struct evsel * evsel,char * bf,size_t size)613 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
614 {
615 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
616 return r + evsel__add_modifiers(evsel, bf + r, size - r);
617 }
618
evsel__tool_name(enum perf_tool_event ev,char * bf,size_t size)619 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
620 {
621 return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
622 }
623
__evsel__bp_name(char * bf,size_t size,u64 addr,u64 type)624 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
625 {
626 int r;
627
628 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
629
630 if (type & HW_BREAKPOINT_R)
631 r += scnprintf(bf + r, size - r, "r");
632
633 if (type & HW_BREAKPOINT_W)
634 r += scnprintf(bf + r, size - r, "w");
635
636 if (type & HW_BREAKPOINT_X)
637 r += scnprintf(bf + r, size - r, "x");
638
639 return r;
640 }
641
evsel__bp_name(struct evsel * evsel,char * bf,size_t size)642 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
643 {
644 struct perf_event_attr *attr = &evsel->core.attr;
645 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
646 return r + evsel__add_modifiers(evsel, bf + r, size - r);
647 }
648
649 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
650 { "L1-dcache", "l1-d", "l1d", "L1-data", },
651 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
652 { "LLC", "L2", },
653 { "dTLB", "d-tlb", "Data-TLB", },
654 { "iTLB", "i-tlb", "Instruction-TLB", },
655 { "branch", "branches", "bpu", "btb", "bpc", },
656 { "node", },
657 };
658
659 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
660 { "load", "loads", "read", },
661 { "store", "stores", "write", },
662 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
663 };
664
665 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
666 { "refs", "Reference", "ops", "access", },
667 { "misses", "miss", },
668 };
669
670 #define C(x) PERF_COUNT_HW_CACHE_##x
671 #define CACHE_READ (1 << C(OP_READ))
672 #define CACHE_WRITE (1 << C(OP_WRITE))
673 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
674 #define COP(x) (1 << x)
675
676 /*
677 * cache operation stat
678 * L1I : Read and prefetch only
679 * ITLB and BPU : Read-only
680 */
681 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
682 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
683 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
684 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
685 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
686 [C(ITLB)] = (CACHE_READ),
687 [C(BPU)] = (CACHE_READ),
688 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
689 };
690
evsel__is_cache_op_valid(u8 type,u8 op)691 bool evsel__is_cache_op_valid(u8 type, u8 op)
692 {
693 if (evsel__hw_cache_stat[type] & COP(op))
694 return true; /* valid */
695 else
696 return false; /* invalid */
697 }
698
__evsel__hw_cache_type_op_res_name(u8 type,u8 op,u8 result,char * bf,size_t size)699 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
700 {
701 if (result) {
702 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
703 evsel__hw_cache_op[op][0],
704 evsel__hw_cache_result[result][0]);
705 }
706
707 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
708 evsel__hw_cache_op[op][1]);
709 }
710
__evsel__hw_cache_name(u64 config,char * bf,size_t size)711 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
712 {
713 u8 op, result, type = (config >> 0) & 0xff;
714 const char *err = "unknown-ext-hardware-cache-type";
715
716 if (type >= PERF_COUNT_HW_CACHE_MAX)
717 goto out_err;
718
719 op = (config >> 8) & 0xff;
720 err = "unknown-ext-hardware-cache-op";
721 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
722 goto out_err;
723
724 result = (config >> 16) & 0xff;
725 err = "unknown-ext-hardware-cache-result";
726 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
727 goto out_err;
728
729 err = "invalid-cache";
730 if (!evsel__is_cache_op_valid(type, op))
731 goto out_err;
732
733 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
734 out_err:
735 return scnprintf(bf, size, "%s", err);
736 }
737
evsel__hw_cache_name(struct evsel * evsel,char * bf,size_t size)738 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
739 {
740 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
741 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
742 }
743
evsel__raw_name(struct evsel * evsel,char * bf,size_t size)744 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
745 {
746 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
747 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
748 }
749
evsel__name(struct evsel * evsel)750 const char *evsel__name(struct evsel *evsel)
751 {
752 char bf[128];
753
754 if (!evsel)
755 goto out_unknown;
756
757 if (evsel->name)
758 return evsel->name;
759
760 switch (evsel->core.attr.type) {
761 case PERF_TYPE_RAW:
762 evsel__raw_name(evsel, bf, sizeof(bf));
763 break;
764
765 case PERF_TYPE_HARDWARE:
766 evsel__hw_name(evsel, bf, sizeof(bf));
767 break;
768
769 case PERF_TYPE_HW_CACHE:
770 evsel__hw_cache_name(evsel, bf, sizeof(bf));
771 break;
772
773 case PERF_TYPE_SOFTWARE:
774 if (evsel__is_tool(evsel))
775 evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
776 else
777 evsel__sw_name(evsel, bf, sizeof(bf));
778 break;
779
780 case PERF_TYPE_TRACEPOINT:
781 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
782 break;
783
784 case PERF_TYPE_BREAKPOINT:
785 evsel__bp_name(evsel, bf, sizeof(bf));
786 break;
787
788 default:
789 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
790 evsel->core.attr.type);
791 break;
792 }
793
794 evsel->name = strdup(bf);
795
796 if (evsel->name)
797 return evsel->name;
798 out_unknown:
799 return "unknown";
800 }
801
evsel__name_is(struct evsel * evsel,const char * name)802 bool evsel__name_is(struct evsel *evsel, const char *name)
803 {
804 return !strcmp(evsel__name(evsel), name);
805 }
806
evsel__metric_id(const struct evsel * evsel)807 const char *evsel__metric_id(const struct evsel *evsel)
808 {
809 if (evsel->metric_id)
810 return evsel->metric_id;
811
812 if (evsel__is_tool(evsel))
813 return perf_tool_event__to_str(evsel->tool_event);
814
815 return "unknown";
816 }
817
evsel__group_name(struct evsel * evsel)818 const char *evsel__group_name(struct evsel *evsel)
819 {
820 return evsel->group_name ?: "anon group";
821 }
822
823 /*
824 * Returns the group details for the specified leader,
825 * with following rules.
826 *
827 * For record -e '{cycles,instructions}'
828 * 'anon group { cycles:u, instructions:u }'
829 *
830 * For record -e 'cycles,instructions' and report --group
831 * 'cycles:u, instructions:u'
832 */
evsel__group_desc(struct evsel * evsel,char * buf,size_t size)833 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
834 {
835 int ret = 0;
836 bool first = true;
837 struct evsel *pos;
838 const char *group_name = evsel__group_name(evsel);
839
840 if (!evsel->forced_leader)
841 ret = scnprintf(buf, size, "%s { ", group_name);
842
843 for_each_group_evsel(pos, evsel) {
844 if (symbol_conf.skip_empty &&
845 evsel__hists(pos)->stats.nr_samples == 0)
846 continue;
847
848 ret += scnprintf(buf + ret, size - ret, "%s%s",
849 first ? "" : ", ", evsel__name(pos));
850 first = false;
851 }
852
853 if (!evsel->forced_leader)
854 ret += scnprintf(buf + ret, size - ret, " }");
855
856 return ret;
857 }
858
__evsel__config_callchain(struct evsel * evsel,struct record_opts * opts,struct callchain_param * param)859 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
860 struct callchain_param *param)
861 {
862 bool function = evsel__is_function_event(evsel);
863 struct perf_event_attr *attr = &evsel->core.attr;
864 const char *arch = perf_env__arch(evsel__env(evsel));
865
866 evsel__set_sample_bit(evsel, CALLCHAIN);
867
868 attr->sample_max_stack = param->max_stack;
869
870 if (opts->kernel_callchains)
871 attr->exclude_callchain_user = 1;
872 if (opts->user_callchains)
873 attr->exclude_callchain_kernel = 1;
874 if (param->record_mode == CALLCHAIN_LBR) {
875 if (!opts->branch_stack) {
876 if (attr->exclude_user) {
877 pr_warning("LBR callstack option is only available "
878 "to get user callchain information. "
879 "Falling back to framepointers.\n");
880 } else {
881 evsel__set_sample_bit(evsel, BRANCH_STACK);
882 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
883 PERF_SAMPLE_BRANCH_CALL_STACK |
884 PERF_SAMPLE_BRANCH_NO_CYCLES |
885 PERF_SAMPLE_BRANCH_NO_FLAGS |
886 PERF_SAMPLE_BRANCH_HW_INDEX;
887 }
888 } else
889 pr_warning("Cannot use LBR callstack with branch stack. "
890 "Falling back to framepointers.\n");
891 }
892
893 if (param->record_mode == CALLCHAIN_DWARF) {
894 if (!function) {
895 evsel__set_sample_bit(evsel, REGS_USER);
896 evsel__set_sample_bit(evsel, STACK_USER);
897 if (opts->sample_user_regs &&
898 DWARF_MINIMAL_REGS(arch) != arch__user_reg_mask()) {
899 attr->sample_regs_user |= DWARF_MINIMAL_REGS(arch);
900 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
901 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
902 "so the minimal registers set (IP, SP) is explicitly forced.\n");
903 } else {
904 attr->sample_regs_user |= arch__user_reg_mask();
905 }
906 attr->sample_stack_user = param->dump_size;
907 attr->exclude_callchain_user = 1;
908 } else {
909 pr_info("Cannot use DWARF unwind for function trace event,"
910 " falling back to framepointers.\n");
911 }
912 }
913
914 if (function) {
915 pr_info("Disabling user space callchains for function trace event.\n");
916 attr->exclude_callchain_user = 1;
917 }
918 }
919
evsel__config_callchain(struct evsel * evsel,struct record_opts * opts,struct callchain_param * param)920 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
921 struct callchain_param *param)
922 {
923 if (param->enabled)
924 return __evsel__config_callchain(evsel, opts, param);
925 }
926
evsel__reset_callgraph(struct evsel * evsel,struct callchain_param * param)927 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
928 {
929 struct perf_event_attr *attr = &evsel->core.attr;
930
931 evsel__reset_sample_bit(evsel, CALLCHAIN);
932 if (param->record_mode == CALLCHAIN_LBR) {
933 evsel__reset_sample_bit(evsel, BRANCH_STACK);
934 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
935 PERF_SAMPLE_BRANCH_CALL_STACK |
936 PERF_SAMPLE_BRANCH_HW_INDEX);
937 }
938 if (param->record_mode == CALLCHAIN_DWARF) {
939 evsel__reset_sample_bit(evsel, REGS_USER);
940 evsel__reset_sample_bit(evsel, STACK_USER);
941 }
942 }
943
evsel__apply_config_terms(struct evsel * evsel,struct record_opts * opts,bool track)944 static void evsel__apply_config_terms(struct evsel *evsel,
945 struct record_opts *opts, bool track)
946 {
947 struct evsel_config_term *term;
948 struct list_head *config_terms = &evsel->config_terms;
949 struct perf_event_attr *attr = &evsel->core.attr;
950 /* callgraph default */
951 struct callchain_param param = {
952 .record_mode = callchain_param.record_mode,
953 };
954 u32 dump_size = 0;
955 int max_stack = 0;
956 const char *callgraph_buf = NULL;
957
958 list_for_each_entry(term, config_terms, list) {
959 switch (term->type) {
960 case EVSEL__CONFIG_TERM_PERIOD:
961 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
962 attr->sample_period = term->val.period;
963 attr->freq = 0;
964 evsel__reset_sample_bit(evsel, PERIOD);
965 }
966 break;
967 case EVSEL__CONFIG_TERM_FREQ:
968 if (!(term->weak && opts->user_freq != UINT_MAX)) {
969 attr->sample_freq = term->val.freq;
970 attr->freq = 1;
971 evsel__set_sample_bit(evsel, PERIOD);
972 }
973 break;
974 case EVSEL__CONFIG_TERM_TIME:
975 if (term->val.time)
976 evsel__set_sample_bit(evsel, TIME);
977 else
978 evsel__reset_sample_bit(evsel, TIME);
979 break;
980 case EVSEL__CONFIG_TERM_CALLGRAPH:
981 callgraph_buf = term->val.str;
982 break;
983 case EVSEL__CONFIG_TERM_BRANCH:
984 if (term->val.str && strcmp(term->val.str, "no")) {
985 evsel__set_sample_bit(evsel, BRANCH_STACK);
986 parse_branch_str(term->val.str,
987 &attr->branch_sample_type);
988 } else
989 evsel__reset_sample_bit(evsel, BRANCH_STACK);
990 break;
991 case EVSEL__CONFIG_TERM_STACK_USER:
992 dump_size = term->val.stack_user;
993 break;
994 case EVSEL__CONFIG_TERM_MAX_STACK:
995 max_stack = term->val.max_stack;
996 break;
997 case EVSEL__CONFIG_TERM_MAX_EVENTS:
998 evsel->max_events = term->val.max_events;
999 break;
1000 case EVSEL__CONFIG_TERM_INHERIT:
1001 /*
1002 * attr->inherit should has already been set by
1003 * evsel__config. If user explicitly set
1004 * inherit using config terms, override global
1005 * opt->no_inherit setting.
1006 */
1007 attr->inherit = term->val.inherit ? 1 : 0;
1008 break;
1009 case EVSEL__CONFIG_TERM_OVERWRITE:
1010 attr->write_backward = term->val.overwrite ? 1 : 0;
1011 break;
1012 case EVSEL__CONFIG_TERM_DRV_CFG:
1013 break;
1014 case EVSEL__CONFIG_TERM_PERCORE:
1015 break;
1016 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1017 attr->aux_output = term->val.aux_output ? 1 : 0;
1018 break;
1019 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1020 /* Already applied by auxtrace */
1021 break;
1022 case EVSEL__CONFIG_TERM_CFG_CHG:
1023 break;
1024 default:
1025 break;
1026 }
1027 }
1028
1029 /* User explicitly set per-event callgraph, clear the old setting and reset. */
1030 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1031 bool sample_address = false;
1032
1033 if (max_stack) {
1034 param.max_stack = max_stack;
1035 if (callgraph_buf == NULL)
1036 callgraph_buf = "fp";
1037 }
1038
1039 /* parse callgraph parameters */
1040 if (callgraph_buf != NULL) {
1041 if (!strcmp(callgraph_buf, "no")) {
1042 param.enabled = false;
1043 param.record_mode = CALLCHAIN_NONE;
1044 } else {
1045 param.enabled = true;
1046 if (parse_callchain_record(callgraph_buf, ¶m)) {
1047 pr_err("per-event callgraph setting for %s failed. "
1048 "Apply callgraph global setting for it\n",
1049 evsel->name);
1050 return;
1051 }
1052 if (param.record_mode == CALLCHAIN_DWARF)
1053 sample_address = true;
1054 }
1055 }
1056 if (dump_size > 0) {
1057 dump_size = round_up(dump_size, sizeof(u64));
1058 param.dump_size = dump_size;
1059 }
1060
1061 /* If global callgraph set, clear it */
1062 if (callchain_param.enabled)
1063 evsel__reset_callgraph(evsel, &callchain_param);
1064
1065 /* set perf-event callgraph */
1066 if (param.enabled) {
1067 if (sample_address) {
1068 evsel__set_sample_bit(evsel, ADDR);
1069 evsel__set_sample_bit(evsel, DATA_SRC);
1070 evsel->core.attr.mmap_data = track;
1071 }
1072 evsel__config_callchain(evsel, opts, ¶m);
1073 }
1074 }
1075 }
1076
__evsel__get_config_term(struct evsel * evsel,enum evsel_term_type type)1077 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1078 {
1079 struct evsel_config_term *term, *found_term = NULL;
1080
1081 list_for_each_entry(term, &evsel->config_terms, list) {
1082 if (term->type == type)
1083 found_term = term;
1084 }
1085
1086 return found_term;
1087 }
1088
arch_evsel__set_sample_weight(struct evsel * evsel)1089 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1090 {
1091 evsel__set_sample_bit(evsel, WEIGHT);
1092 }
1093
arch__post_evsel_config(struct evsel * evsel __maybe_unused,struct perf_event_attr * attr __maybe_unused)1094 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1095 struct perf_event_attr *attr __maybe_unused)
1096 {
1097 }
1098
evsel__set_default_freq_period(struct record_opts * opts,struct perf_event_attr * attr)1099 static void evsel__set_default_freq_period(struct record_opts *opts,
1100 struct perf_event_attr *attr)
1101 {
1102 if (opts->freq) {
1103 attr->freq = 1;
1104 attr->sample_freq = opts->freq;
1105 } else {
1106 attr->sample_period = opts->default_interval;
1107 }
1108 }
1109
evsel__is_offcpu_event(struct evsel * evsel)1110 static bool evsel__is_offcpu_event(struct evsel *evsel)
1111 {
1112 return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT);
1113 }
1114
1115 /*
1116 * The enable_on_exec/disabled value strategy:
1117 *
1118 * 1) For any type of traced program:
1119 * - all independent events and group leaders are disabled
1120 * - all group members are enabled
1121 *
1122 * Group members are ruled by group leaders. They need to
1123 * be enabled, because the group scheduling relies on that.
1124 *
1125 * 2) For traced programs executed by perf:
1126 * - all independent events and group leaders have
1127 * enable_on_exec set
1128 * - we don't specifically enable or disable any event during
1129 * the record command
1130 *
1131 * Independent events and group leaders are initially disabled
1132 * and get enabled by exec. Group members are ruled by group
1133 * leaders as stated in 1).
1134 *
1135 * 3) For traced programs attached by perf (pid/tid):
1136 * - we specifically enable or disable all events during
1137 * the record command
1138 *
1139 * When attaching events to already running traced we
1140 * enable/disable events specifically, as there's no
1141 * initial traced exec call.
1142 */
evsel__config(struct evsel * evsel,struct record_opts * opts,struct callchain_param * callchain)1143 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1144 struct callchain_param *callchain)
1145 {
1146 struct evsel *leader = evsel__leader(evsel);
1147 struct perf_event_attr *attr = &evsel->core.attr;
1148 int track = evsel->tracking;
1149 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1150
1151 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1152 attr->inherit = !opts->no_inherit;
1153 attr->write_backward = opts->overwrite ? 1 : 0;
1154 attr->read_format = PERF_FORMAT_LOST;
1155
1156 evsel__set_sample_bit(evsel, IP);
1157 evsel__set_sample_bit(evsel, TID);
1158
1159 if (evsel->sample_read) {
1160 evsel__set_sample_bit(evsel, READ);
1161
1162 /*
1163 * We need ID even in case of single event, because
1164 * PERF_SAMPLE_READ process ID specific data.
1165 */
1166 evsel__set_sample_id(evsel, false);
1167
1168 /*
1169 * Apply group format only if we belong to group
1170 * with more than one members.
1171 */
1172 if (leader->core.nr_members > 1) {
1173 attr->read_format |= PERF_FORMAT_GROUP;
1174 attr->inherit = 0;
1175 }
1176 }
1177
1178 /*
1179 * We default some events to have a default interval. But keep
1180 * it a weak assumption overridable by the user.
1181 */
1182 if ((evsel->is_libpfm_event && !attr->sample_period) ||
1183 (!evsel->is_libpfm_event && (!attr->sample_period ||
1184 opts->user_freq != UINT_MAX ||
1185 opts->user_interval != ULLONG_MAX)))
1186 evsel__set_default_freq_period(opts, attr);
1187
1188 /*
1189 * If attr->freq was set (here or earlier), ask for period
1190 * to be sampled.
1191 */
1192 if (attr->freq)
1193 evsel__set_sample_bit(evsel, PERIOD);
1194
1195 if (opts->no_samples)
1196 attr->sample_freq = 0;
1197
1198 if (opts->inherit_stat) {
1199 evsel->core.attr.read_format |=
1200 PERF_FORMAT_TOTAL_TIME_ENABLED |
1201 PERF_FORMAT_TOTAL_TIME_RUNNING |
1202 PERF_FORMAT_ID;
1203 attr->inherit_stat = 1;
1204 }
1205
1206 if (opts->sample_address) {
1207 evsel__set_sample_bit(evsel, ADDR);
1208 attr->mmap_data = track;
1209 }
1210
1211 /*
1212 * We don't allow user space callchains for function trace
1213 * event, due to issues with page faults while tracing page
1214 * fault handler and its overall trickiness nature.
1215 */
1216 if (evsel__is_function_event(evsel))
1217 evsel->core.attr.exclude_callchain_user = 1;
1218
1219 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1220 evsel__config_callchain(evsel, opts, callchain);
1221
1222 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1223 !evsel__is_dummy_event(evsel)) {
1224 attr->sample_regs_intr = opts->sample_intr_regs;
1225 evsel__set_sample_bit(evsel, REGS_INTR);
1226 }
1227
1228 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1229 !evsel__is_dummy_event(evsel)) {
1230 attr->sample_regs_user |= opts->sample_user_regs;
1231 evsel__set_sample_bit(evsel, REGS_USER);
1232 }
1233
1234 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1235 evsel__set_sample_bit(evsel, CPU);
1236
1237 /*
1238 * When the user explicitly disabled time don't force it here.
1239 */
1240 if (opts->sample_time &&
1241 (!perf_missing_features.sample_id_all &&
1242 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1243 opts->sample_time_set)))
1244 evsel__set_sample_bit(evsel, TIME);
1245
1246 if (opts->raw_samples && !evsel->no_aux_samples) {
1247 evsel__set_sample_bit(evsel, TIME);
1248 evsel__set_sample_bit(evsel, RAW);
1249 evsel__set_sample_bit(evsel, CPU);
1250 }
1251
1252 if (opts->sample_address)
1253 evsel__set_sample_bit(evsel, DATA_SRC);
1254
1255 if (opts->sample_phys_addr)
1256 evsel__set_sample_bit(evsel, PHYS_ADDR);
1257
1258 if (opts->no_buffering) {
1259 attr->watermark = 0;
1260 attr->wakeup_events = 1;
1261 }
1262 if (opts->branch_stack && !evsel->no_aux_samples) {
1263 evsel__set_sample_bit(evsel, BRANCH_STACK);
1264 attr->branch_sample_type = opts->branch_stack;
1265 }
1266
1267 if (opts->sample_weight)
1268 arch_evsel__set_sample_weight(evsel);
1269
1270 attr->task = track;
1271 attr->mmap = track;
1272 attr->mmap2 = track && !perf_missing_features.mmap2;
1273 attr->comm = track;
1274 attr->build_id = track && opts->build_id;
1275
1276 /*
1277 * ksymbol is tracked separately with text poke because it needs to be
1278 * system wide and enabled immediately.
1279 */
1280 if (!opts->text_poke)
1281 attr->ksymbol = track && !perf_missing_features.ksymbol;
1282 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1283
1284 if (opts->record_namespaces)
1285 attr->namespaces = track;
1286
1287 if (opts->record_cgroup) {
1288 attr->cgroup = track && !perf_missing_features.cgroup;
1289 evsel__set_sample_bit(evsel, CGROUP);
1290 }
1291
1292 if (opts->sample_data_page_size)
1293 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1294
1295 if (opts->sample_code_page_size)
1296 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1297
1298 if (opts->record_switch_events)
1299 attr->context_switch = track;
1300
1301 if (opts->sample_transaction)
1302 evsel__set_sample_bit(evsel, TRANSACTION);
1303
1304 if (opts->running_time) {
1305 evsel->core.attr.read_format |=
1306 PERF_FORMAT_TOTAL_TIME_ENABLED |
1307 PERF_FORMAT_TOTAL_TIME_RUNNING;
1308 }
1309
1310 /*
1311 * XXX see the function comment above
1312 *
1313 * Disabling only independent events or group leaders,
1314 * keeping group members enabled.
1315 */
1316 if (evsel__is_group_leader(evsel))
1317 attr->disabled = 1;
1318
1319 /*
1320 * Setting enable_on_exec for independent events and
1321 * group leaders for traced executed by perf.
1322 */
1323 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1324 !opts->target.initial_delay)
1325 attr->enable_on_exec = 1;
1326
1327 if (evsel->immediate) {
1328 attr->disabled = 0;
1329 attr->enable_on_exec = 0;
1330 }
1331
1332 clockid = opts->clockid;
1333 if (opts->use_clockid) {
1334 attr->use_clockid = 1;
1335 attr->clockid = opts->clockid;
1336 }
1337
1338 if (evsel->precise_max)
1339 attr->precise_ip = 3;
1340
1341 if (opts->all_user) {
1342 attr->exclude_kernel = 1;
1343 attr->exclude_user = 0;
1344 }
1345
1346 if (opts->all_kernel) {
1347 attr->exclude_kernel = 0;
1348 attr->exclude_user = 1;
1349 }
1350
1351 if (evsel->core.own_cpus || evsel->unit)
1352 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1353
1354 /*
1355 * Apply event specific term settings,
1356 * it overloads any global configuration.
1357 */
1358 evsel__apply_config_terms(evsel, opts, track);
1359
1360 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1361
1362 /* The --period option takes the precedence. */
1363 if (opts->period_set) {
1364 if (opts->period)
1365 evsel__set_sample_bit(evsel, PERIOD);
1366 else
1367 evsel__reset_sample_bit(evsel, PERIOD);
1368 }
1369
1370 /*
1371 * A dummy event never triggers any actual counter and therefore
1372 * cannot be used with branch_stack.
1373 *
1374 * For initial_delay, a dummy event is added implicitly.
1375 * The software event will trigger -EOPNOTSUPP error out,
1376 * if BRANCH_STACK bit is set.
1377 */
1378 if (evsel__is_dummy_event(evsel))
1379 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1380
1381 if (evsel__is_offcpu_event(evsel))
1382 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1383
1384 arch__post_evsel_config(evsel, attr);
1385 }
1386
evsel__set_filter(struct evsel * evsel,const char * filter)1387 int evsel__set_filter(struct evsel *evsel, const char *filter)
1388 {
1389 char *new_filter = strdup(filter);
1390
1391 if (new_filter != NULL) {
1392 free(evsel->filter);
1393 evsel->filter = new_filter;
1394 return 0;
1395 }
1396
1397 return -1;
1398 }
1399
evsel__append_filter(struct evsel * evsel,const char * fmt,const char * filter)1400 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1401 {
1402 char *new_filter;
1403
1404 if (evsel->filter == NULL)
1405 return evsel__set_filter(evsel, filter);
1406
1407 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1408 free(evsel->filter);
1409 evsel->filter = new_filter;
1410 return 0;
1411 }
1412
1413 return -1;
1414 }
1415
evsel__append_tp_filter(struct evsel * evsel,const char * filter)1416 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1417 {
1418 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1419 }
1420
evsel__append_addr_filter(struct evsel * evsel,const char * filter)1421 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1422 {
1423 return evsel__append_filter(evsel, "%s,%s", filter);
1424 }
1425
1426 /* Caller has to clear disabled after going through all CPUs. */
evsel__enable_cpu(struct evsel * evsel,int cpu_map_idx)1427 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1428 {
1429 return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1430 }
1431
evsel__enable(struct evsel * evsel)1432 int evsel__enable(struct evsel *evsel)
1433 {
1434 int err = perf_evsel__enable(&evsel->core);
1435
1436 if (!err)
1437 evsel->disabled = false;
1438 return err;
1439 }
1440
1441 /* Caller has to set disabled after going through all CPUs. */
evsel__disable_cpu(struct evsel * evsel,int cpu_map_idx)1442 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1443 {
1444 return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1445 }
1446
evsel__disable(struct evsel * evsel)1447 int evsel__disable(struct evsel *evsel)
1448 {
1449 int err = perf_evsel__disable(&evsel->core);
1450 /*
1451 * We mark it disabled here so that tools that disable a event can
1452 * ignore events after they disable it. I.e. the ring buffer may have
1453 * already a few more events queued up before the kernel got the stop
1454 * request.
1455 */
1456 if (!err)
1457 evsel->disabled = true;
1458
1459 return err;
1460 }
1461
free_config_terms(struct list_head * config_terms)1462 void free_config_terms(struct list_head *config_terms)
1463 {
1464 struct evsel_config_term *term, *h;
1465
1466 list_for_each_entry_safe(term, h, config_terms, list) {
1467 list_del_init(&term->list);
1468 if (term->free_str)
1469 zfree(&term->val.str);
1470 free(term);
1471 }
1472 }
1473
evsel__free_config_terms(struct evsel * evsel)1474 static void evsel__free_config_terms(struct evsel *evsel)
1475 {
1476 free_config_terms(&evsel->config_terms);
1477 }
1478
evsel__exit(struct evsel * evsel)1479 void evsel__exit(struct evsel *evsel)
1480 {
1481 assert(list_empty(&evsel->core.node));
1482 assert(evsel->evlist == NULL);
1483 bpf_counter__destroy(evsel);
1484 perf_bpf_filter__destroy(evsel);
1485 evsel__free_counts(evsel);
1486 perf_evsel__free_fd(&evsel->core);
1487 perf_evsel__free_id(&evsel->core);
1488 evsel__free_config_terms(evsel);
1489 cgroup__put(evsel->cgrp);
1490 perf_cpu_map__put(evsel->core.cpus);
1491 perf_cpu_map__put(evsel->core.own_cpus);
1492 perf_thread_map__put(evsel->core.threads);
1493 zfree(&evsel->group_name);
1494 zfree(&evsel->name);
1495 zfree(&evsel->filter);
1496 zfree(&evsel->pmu_name);
1497 zfree(&evsel->group_pmu_name);
1498 zfree(&evsel->unit);
1499 zfree(&evsel->metric_id);
1500 evsel__zero_per_pkg(evsel);
1501 hashmap__free(evsel->per_pkg_mask);
1502 evsel->per_pkg_mask = NULL;
1503 zfree(&evsel->metric_events);
1504 perf_evsel__object.fini(evsel);
1505 if (evsel->tool_event == PERF_TOOL_SYSTEM_TIME ||
1506 evsel->tool_event == PERF_TOOL_USER_TIME)
1507 xyarray__delete(evsel->start_times);
1508 }
1509
evsel__delete(struct evsel * evsel)1510 void evsel__delete(struct evsel *evsel)
1511 {
1512 if (!evsel)
1513 return;
1514
1515 evsel__exit(evsel);
1516 free(evsel);
1517 }
1518
evsel__compute_deltas(struct evsel * evsel,int cpu_map_idx,int thread,struct perf_counts_values * count)1519 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1520 struct perf_counts_values *count)
1521 {
1522 struct perf_counts_values tmp;
1523
1524 if (!evsel->prev_raw_counts)
1525 return;
1526
1527 tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1528 *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1529
1530 count->val = count->val - tmp.val;
1531 count->ena = count->ena - tmp.ena;
1532 count->run = count->run - tmp.run;
1533 }
1534
evsel__read_one(struct evsel * evsel,int cpu_map_idx,int thread)1535 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1536 {
1537 struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1538
1539 return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1540 }
1541
evsel__set_count(struct evsel * counter,int cpu_map_idx,int thread,u64 val,u64 ena,u64 run,u64 lost)1542 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1543 u64 val, u64 ena, u64 run, u64 lost)
1544 {
1545 struct perf_counts_values *count;
1546
1547 count = perf_counts(counter->counts, cpu_map_idx, thread);
1548
1549 count->val = val;
1550 count->ena = ena;
1551 count->run = run;
1552 count->lost = lost;
1553
1554 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1555 }
1556
evsel__process_group_data(struct evsel * leader,int cpu_map_idx,int thread,u64 * data)1557 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1558 {
1559 u64 read_format = leader->core.attr.read_format;
1560 struct sample_read_value *v;
1561 u64 nr, ena = 0, run = 0, lost = 0;
1562
1563 nr = *data++;
1564
1565 if (nr != (u64) leader->core.nr_members)
1566 return -EINVAL;
1567
1568 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1569 ena = *data++;
1570
1571 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1572 run = *data++;
1573
1574 v = (void *)data;
1575 sample_read_group__for_each(v, nr, read_format) {
1576 struct evsel *counter;
1577
1578 counter = evlist__id2evsel(leader->evlist, v->id);
1579 if (!counter)
1580 return -EINVAL;
1581
1582 if (read_format & PERF_FORMAT_LOST)
1583 lost = v->lost;
1584
1585 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1586 }
1587
1588 return 0;
1589 }
1590
evsel__read_group(struct evsel * leader,int cpu_map_idx,int thread)1591 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1592 {
1593 struct perf_stat_evsel *ps = leader->stats;
1594 u64 read_format = leader->core.attr.read_format;
1595 int size = perf_evsel__read_size(&leader->core);
1596 u64 *data = ps->group_data;
1597
1598 if (!(read_format & PERF_FORMAT_ID))
1599 return -EINVAL;
1600
1601 if (!evsel__is_group_leader(leader))
1602 return -EINVAL;
1603
1604 if (!data) {
1605 data = zalloc(size);
1606 if (!data)
1607 return -ENOMEM;
1608
1609 ps->group_data = data;
1610 }
1611
1612 if (FD(leader, cpu_map_idx, thread) < 0)
1613 return -EINVAL;
1614
1615 if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1616 return -errno;
1617
1618 return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1619 }
1620
read_until_char(struct io * io,char e)1621 static bool read_until_char(struct io *io, char e)
1622 {
1623 int c;
1624
1625 do {
1626 c = io__get_char(io);
1627 if (c == -1)
1628 return false;
1629 } while (c != e);
1630 return true;
1631 }
1632
read_stat_field(int fd,struct perf_cpu cpu,int field,__u64 * val)1633 static int read_stat_field(int fd, struct perf_cpu cpu, int field, __u64 *val)
1634 {
1635 char buf[256];
1636 struct io io;
1637 int i;
1638
1639 io__init(&io, fd, buf, sizeof(buf));
1640
1641 /* Skip lines to relevant CPU. */
1642 for (i = -1; i < cpu.cpu; i++) {
1643 if (!read_until_char(&io, '\n'))
1644 return -EINVAL;
1645 }
1646 /* Skip to "cpu". */
1647 if (io__get_char(&io) != 'c') return -EINVAL;
1648 if (io__get_char(&io) != 'p') return -EINVAL;
1649 if (io__get_char(&io) != 'u') return -EINVAL;
1650
1651 /* Skip N of cpuN. */
1652 if (!read_until_char(&io, ' '))
1653 return -EINVAL;
1654
1655 i = 1;
1656 while (true) {
1657 if (io__get_dec(&io, val) != ' ')
1658 break;
1659 if (field == i)
1660 return 0;
1661 i++;
1662 }
1663 return -EINVAL;
1664 }
1665
read_pid_stat_field(int fd,int field,__u64 * val)1666 static int read_pid_stat_field(int fd, int field, __u64 *val)
1667 {
1668 char buf[256];
1669 struct io io;
1670 int c, i;
1671
1672 io__init(&io, fd, buf, sizeof(buf));
1673 if (io__get_dec(&io, val) != ' ')
1674 return -EINVAL;
1675 if (field == 1)
1676 return 0;
1677
1678 /* Skip comm. */
1679 if (io__get_char(&io) != '(' || !read_until_char(&io, ')'))
1680 return -EINVAL;
1681 if (field == 2)
1682 return -EINVAL; /* String can't be returned. */
1683
1684 /* Skip state */
1685 if (io__get_char(&io) != ' ' || io__get_char(&io) == -1)
1686 return -EINVAL;
1687 if (field == 3)
1688 return -EINVAL; /* String can't be returned. */
1689
1690 /* Loop over numeric fields*/
1691 if (io__get_char(&io) != ' ')
1692 return -EINVAL;
1693
1694 i = 4;
1695 while (true) {
1696 c = io__get_dec(&io, val);
1697 if (c == -1)
1698 return -EINVAL;
1699 if (c == -2) {
1700 /* Assume a -ve was read */
1701 c = io__get_dec(&io, val);
1702 *val *= -1;
1703 }
1704 if (c != ' ')
1705 return -EINVAL;
1706 if (field == i)
1707 return 0;
1708 i++;
1709 }
1710 return -EINVAL;
1711 }
1712
evsel__read_tool(struct evsel * evsel,int cpu_map_idx,int thread)1713 static int evsel__read_tool(struct evsel *evsel, int cpu_map_idx, int thread)
1714 {
1715 __u64 *start_time, cur_time, delta_start;
1716 int fd, err = 0;
1717 struct perf_counts_values *count;
1718 bool adjust = false;
1719
1720 count = perf_counts(evsel->counts, cpu_map_idx, thread);
1721
1722 switch (evsel->tool_event) {
1723 case PERF_TOOL_DURATION_TIME:
1724 /*
1725 * Pretend duration_time is only on the first CPU and thread, or
1726 * else aggregation will scale duration_time by the number of
1727 * CPUs/threads.
1728 */
1729 start_time = &evsel->start_time;
1730 if (cpu_map_idx == 0 && thread == 0)
1731 cur_time = rdclock();
1732 else
1733 cur_time = *start_time;
1734 break;
1735 case PERF_TOOL_USER_TIME:
1736 case PERF_TOOL_SYSTEM_TIME: {
1737 bool system = evsel->tool_event == PERF_TOOL_SYSTEM_TIME;
1738
1739 start_time = xyarray__entry(evsel->start_times, cpu_map_idx, thread);
1740 fd = FD(evsel, cpu_map_idx, thread);
1741 lseek(fd, SEEK_SET, 0);
1742 if (evsel->pid_stat) {
1743 /* The event exists solely on 1 CPU. */
1744 if (cpu_map_idx == 0)
1745 err = read_pid_stat_field(fd, system ? 15 : 14, &cur_time);
1746 else
1747 cur_time = 0;
1748 } else {
1749 /* The event is for all threads. */
1750 if (thread == 0) {
1751 struct perf_cpu cpu = perf_cpu_map__cpu(evsel->core.cpus,
1752 cpu_map_idx);
1753
1754 err = read_stat_field(fd, cpu, system ? 3 : 1, &cur_time);
1755 } else {
1756 cur_time = 0;
1757 }
1758 }
1759 adjust = true;
1760 break;
1761 }
1762 case PERF_TOOL_NONE:
1763 case PERF_TOOL_MAX:
1764 default:
1765 err = -EINVAL;
1766 }
1767 if (err)
1768 return err;
1769
1770 delta_start = cur_time - *start_time;
1771 if (adjust) {
1772 __u64 ticks_per_sec = sysconf(_SC_CLK_TCK);
1773
1774 delta_start *= 1000000000 / ticks_per_sec;
1775 }
1776 count->val = delta_start;
1777 count->ena = count->run = delta_start;
1778 count->lost = 0;
1779 return 0;
1780 }
1781
evsel__read_counter(struct evsel * evsel,int cpu_map_idx,int thread)1782 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1783 {
1784 if (evsel__is_tool(evsel))
1785 return evsel__read_tool(evsel, cpu_map_idx, thread);
1786
1787 if (evsel->core.attr.read_format & PERF_FORMAT_GROUP)
1788 return evsel__read_group(evsel, cpu_map_idx, thread);
1789
1790 return evsel__read_one(evsel, cpu_map_idx, thread);
1791 }
1792
__evsel__read_on_cpu(struct evsel * evsel,int cpu_map_idx,int thread,bool scale)1793 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1794 {
1795 struct perf_counts_values count;
1796 size_t nv = scale ? 3 : 1;
1797
1798 if (FD(evsel, cpu_map_idx, thread) < 0)
1799 return -EINVAL;
1800
1801 if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1802 return -ENOMEM;
1803
1804 if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1805 return -errno;
1806
1807 evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1808 perf_counts_values__scale(&count, scale, NULL);
1809 *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1810 return 0;
1811 }
1812
evsel__match_other_cpu(struct evsel * evsel,struct evsel * other,int cpu_map_idx)1813 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1814 int cpu_map_idx)
1815 {
1816 struct perf_cpu cpu;
1817
1818 cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1819 return perf_cpu_map__idx(other->core.cpus, cpu);
1820 }
1821
evsel__hybrid_group_cpu_map_idx(struct evsel * evsel,int cpu_map_idx)1822 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1823 {
1824 struct evsel *leader = evsel__leader(evsel);
1825
1826 if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1827 (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1828 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1829 }
1830
1831 return cpu_map_idx;
1832 }
1833
get_group_fd(struct evsel * evsel,int cpu_map_idx,int thread)1834 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1835 {
1836 struct evsel *leader = evsel__leader(evsel);
1837 int fd;
1838
1839 if (evsel__is_group_leader(evsel))
1840 return -1;
1841
1842 /*
1843 * Leader must be already processed/open,
1844 * if not it's a bug.
1845 */
1846 BUG_ON(!leader->core.fd);
1847
1848 cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1849 if (cpu_map_idx == -1)
1850 return -1;
1851
1852 fd = FD(leader, cpu_map_idx, thread);
1853 BUG_ON(fd == -1 && !leader->skippable);
1854
1855 /*
1856 * When the leader has been skipped, return -2 to distinguish from no
1857 * group leader case.
1858 */
1859 return fd == -1 ? -2 : fd;
1860 }
1861
evsel__remove_fd(struct evsel * pos,int nr_cpus,int nr_threads,int thread_idx)1862 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1863 {
1864 for (int cpu = 0; cpu < nr_cpus; cpu++)
1865 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1866 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1867 }
1868
update_fds(struct evsel * evsel,int nr_cpus,int cpu_map_idx,int nr_threads,int thread_idx)1869 static int update_fds(struct evsel *evsel,
1870 int nr_cpus, int cpu_map_idx,
1871 int nr_threads, int thread_idx)
1872 {
1873 struct evsel *pos;
1874
1875 if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1876 return -EINVAL;
1877
1878 evlist__for_each_entry(evsel->evlist, pos) {
1879 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1880
1881 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1882
1883 /*
1884 * Since fds for next evsel has not been created,
1885 * there is no need to iterate whole event list.
1886 */
1887 if (pos == evsel)
1888 break;
1889 }
1890 return 0;
1891 }
1892
evsel__ignore_missing_thread(struct evsel * evsel,int nr_cpus,int cpu_map_idx,struct perf_thread_map * threads,int thread,int err)1893 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1894 int nr_cpus, int cpu_map_idx,
1895 struct perf_thread_map *threads,
1896 int thread, int err)
1897 {
1898 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1899
1900 if (!evsel->ignore_missing_thread)
1901 return false;
1902
1903 /* The system wide setup does not work with threads. */
1904 if (evsel->core.system_wide)
1905 return false;
1906
1907 /* The -ESRCH is perf event syscall errno for pid's not found. */
1908 if (err != -ESRCH)
1909 return false;
1910
1911 /* If there's only one thread, let it fail. */
1912 if (threads->nr == 1)
1913 return false;
1914
1915 /*
1916 * We should remove fd for missing_thread first
1917 * because thread_map__remove() will decrease threads->nr.
1918 */
1919 if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1920 return false;
1921
1922 if (thread_map__remove(threads, thread))
1923 return false;
1924
1925 pr_warning("WARNING: Ignored open failure for pid %d\n",
1926 ignore_pid);
1927 return true;
1928 }
1929
__open_attr__fprintf(FILE * fp,const char * name,const char * val,void * priv __maybe_unused)1930 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1931 void *priv __maybe_unused)
1932 {
1933 return fprintf(fp, " %-32s %s\n", name, val);
1934 }
1935
display_attr(struct perf_event_attr * attr)1936 static void display_attr(struct perf_event_attr *attr)
1937 {
1938 if (verbose >= 2 || debug_peo_args) {
1939 fprintf(stderr, "%.60s\n", graph_dotted_line);
1940 fprintf(stderr, "perf_event_attr:\n");
1941 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1942 fprintf(stderr, "%.60s\n", graph_dotted_line);
1943 }
1944 }
1945
evsel__precise_ip_fallback(struct evsel * evsel)1946 bool evsel__precise_ip_fallback(struct evsel *evsel)
1947 {
1948 /* Do not try less precise if not requested. */
1949 if (!evsel->precise_max)
1950 return false;
1951
1952 /*
1953 * We tried all the precise_ip values, and it's
1954 * still failing, so leave it to standard fallback.
1955 */
1956 if (!evsel->core.attr.precise_ip) {
1957 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1958 return false;
1959 }
1960
1961 if (!evsel->precise_ip_original)
1962 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1963
1964 evsel->core.attr.precise_ip--;
1965 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1966 display_attr(&evsel->core.attr);
1967 return true;
1968 }
1969
1970 static struct perf_cpu_map *empty_cpu_map;
1971 static struct perf_thread_map *empty_thread_map;
1972
__evsel__prepare_open(struct evsel * evsel,struct perf_cpu_map * cpus,struct perf_thread_map * threads)1973 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1974 struct perf_thread_map *threads)
1975 {
1976 int nthreads = perf_thread_map__nr(threads);
1977
1978 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1979 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1980 return -EINVAL;
1981
1982 if (cpus == NULL) {
1983 if (empty_cpu_map == NULL) {
1984 empty_cpu_map = perf_cpu_map__new_any_cpu();
1985 if (empty_cpu_map == NULL)
1986 return -ENOMEM;
1987 }
1988
1989 cpus = empty_cpu_map;
1990 }
1991
1992 if (threads == NULL) {
1993 if (empty_thread_map == NULL) {
1994 empty_thread_map = thread_map__new_by_tid(-1);
1995 if (empty_thread_map == NULL)
1996 return -ENOMEM;
1997 }
1998
1999 threads = empty_thread_map;
2000 }
2001
2002 if (evsel->core.fd == NULL &&
2003 perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
2004 return -ENOMEM;
2005
2006 if ((evsel->tool_event == PERF_TOOL_SYSTEM_TIME ||
2007 evsel->tool_event == PERF_TOOL_USER_TIME) &&
2008 !evsel->start_times) {
2009 evsel->start_times = xyarray__new(perf_cpu_map__nr(cpus), nthreads, sizeof(__u64));
2010 if (!evsel->start_times)
2011 return -ENOMEM;
2012 }
2013
2014 evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
2015 if (evsel->cgrp)
2016 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
2017
2018 return 0;
2019 }
2020
evsel__disable_missing_features(struct evsel * evsel)2021 static void evsel__disable_missing_features(struct evsel *evsel)
2022 {
2023 if (perf_missing_features.branch_counters)
2024 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_COUNTERS;
2025 if (perf_missing_features.read_lost)
2026 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
2027 if (perf_missing_features.weight_struct) {
2028 evsel__set_sample_bit(evsel, WEIGHT);
2029 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
2030 }
2031 if (perf_missing_features.clockid_wrong)
2032 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
2033 if (perf_missing_features.clockid) {
2034 evsel->core.attr.use_clockid = 0;
2035 evsel->core.attr.clockid = 0;
2036 }
2037 if (perf_missing_features.cloexec)
2038 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
2039 if (perf_missing_features.mmap2)
2040 evsel->core.attr.mmap2 = 0;
2041 if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
2042 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
2043 if (perf_missing_features.lbr_flags)
2044 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
2045 PERF_SAMPLE_BRANCH_NO_CYCLES);
2046 if (perf_missing_features.group_read && evsel->core.attr.inherit)
2047 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
2048 if (perf_missing_features.ksymbol)
2049 evsel->core.attr.ksymbol = 0;
2050 if (perf_missing_features.bpf)
2051 evsel->core.attr.bpf_event = 0;
2052 if (perf_missing_features.branch_hw_idx)
2053 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
2054 if (perf_missing_features.sample_id_all)
2055 evsel->core.attr.sample_id_all = 0;
2056 }
2057
evsel__prepare_open(struct evsel * evsel,struct perf_cpu_map * cpus,struct perf_thread_map * threads)2058 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
2059 struct perf_thread_map *threads)
2060 {
2061 int err;
2062
2063 err = __evsel__prepare_open(evsel, cpus, threads);
2064 if (err)
2065 return err;
2066
2067 evsel__disable_missing_features(evsel);
2068
2069 return err;
2070 }
2071
evsel__detect_missing_features(struct evsel * evsel)2072 bool evsel__detect_missing_features(struct evsel *evsel)
2073 {
2074 /*
2075 * Must probe features in the order they were added to the
2076 * perf_event_attr interface.
2077 */
2078 if (!perf_missing_features.branch_counters &&
2079 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS)) {
2080 perf_missing_features.branch_counters = true;
2081 pr_debug2("switching off branch counters support\n");
2082 return true;
2083 } else if (!perf_missing_features.read_lost &&
2084 (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
2085 perf_missing_features.read_lost = true;
2086 pr_debug2("switching off PERF_FORMAT_LOST support\n");
2087 return true;
2088 } else if (!perf_missing_features.weight_struct &&
2089 (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
2090 perf_missing_features.weight_struct = true;
2091 pr_debug2("switching off weight struct support\n");
2092 return true;
2093 } else if (!perf_missing_features.code_page_size &&
2094 (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
2095 perf_missing_features.code_page_size = true;
2096 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
2097 return false;
2098 } else if (!perf_missing_features.data_page_size &&
2099 (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
2100 perf_missing_features.data_page_size = true;
2101 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
2102 return false;
2103 } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
2104 perf_missing_features.cgroup = true;
2105 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
2106 return false;
2107 } else if (!perf_missing_features.branch_hw_idx &&
2108 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
2109 perf_missing_features.branch_hw_idx = true;
2110 pr_debug2("switching off branch HW index support\n");
2111 return true;
2112 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
2113 perf_missing_features.aux_output = true;
2114 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
2115 return false;
2116 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
2117 perf_missing_features.bpf = true;
2118 pr_debug2_peo("switching off bpf_event\n");
2119 return true;
2120 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
2121 perf_missing_features.ksymbol = true;
2122 pr_debug2_peo("switching off ksymbol\n");
2123 return true;
2124 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
2125 perf_missing_features.write_backward = true;
2126 pr_debug2_peo("switching off write_backward\n");
2127 return false;
2128 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
2129 perf_missing_features.clockid_wrong = true;
2130 pr_debug2_peo("switching off clockid\n");
2131 return true;
2132 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
2133 perf_missing_features.clockid = true;
2134 pr_debug2_peo("switching off use_clockid\n");
2135 return true;
2136 } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
2137 perf_missing_features.cloexec = true;
2138 pr_debug2_peo("switching off cloexec flag\n");
2139 return true;
2140 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
2141 perf_missing_features.mmap2 = true;
2142 pr_debug2_peo("switching off mmap2\n");
2143 return true;
2144 } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
2145 if (evsel->pmu == NULL)
2146 evsel->pmu = evsel__find_pmu(evsel);
2147
2148 if (evsel->pmu)
2149 evsel->pmu->missing_features.exclude_guest = true;
2150 else {
2151 /* we cannot find PMU, disable attrs now */
2152 evsel->core.attr.exclude_host = false;
2153 evsel->core.attr.exclude_guest = false;
2154 }
2155
2156 if (evsel->exclude_GH) {
2157 pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
2158 return false;
2159 }
2160 if (!perf_missing_features.exclude_guest) {
2161 perf_missing_features.exclude_guest = true;
2162 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
2163 }
2164 return true;
2165 } else if (!perf_missing_features.sample_id_all) {
2166 perf_missing_features.sample_id_all = true;
2167 pr_debug2_peo("switching off sample_id_all\n");
2168 return true;
2169 } else if (!perf_missing_features.lbr_flags &&
2170 (evsel->core.attr.branch_sample_type &
2171 (PERF_SAMPLE_BRANCH_NO_CYCLES |
2172 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2173 perf_missing_features.lbr_flags = true;
2174 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2175 return true;
2176 } else if (!perf_missing_features.group_read &&
2177 evsel->core.attr.inherit &&
2178 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2179 evsel__is_group_leader(evsel)) {
2180 perf_missing_features.group_read = true;
2181 pr_debug2_peo("switching off group read\n");
2182 return true;
2183 } else {
2184 return false;
2185 }
2186 }
2187
evsel__open_cpu(struct evsel * evsel,struct perf_cpu_map * cpus,struct perf_thread_map * threads,int start_cpu_map_idx,int end_cpu_map_idx)2188 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2189 struct perf_thread_map *threads,
2190 int start_cpu_map_idx, int end_cpu_map_idx)
2191 {
2192 int idx, thread, nthreads;
2193 int pid = -1, err, old_errno;
2194 enum rlimit_action set_rlimit = NO_CHANGE;
2195
2196 if (evsel->tool_event == PERF_TOOL_DURATION_TIME) {
2197 if (evsel->core.attr.sample_period) /* no sampling */
2198 return -EINVAL;
2199 evsel->start_time = rdclock();
2200 return 0;
2201 }
2202
2203 err = __evsel__prepare_open(evsel, cpus, threads);
2204 if (err)
2205 return err;
2206
2207 if (cpus == NULL)
2208 cpus = empty_cpu_map;
2209
2210 if (threads == NULL)
2211 threads = empty_thread_map;
2212
2213 nthreads = perf_thread_map__nr(threads);
2214
2215 if (evsel->cgrp)
2216 pid = evsel->cgrp->fd;
2217
2218 fallback_missing_features:
2219 evsel__disable_missing_features(evsel);
2220
2221 pr_debug3("Opening: %s\n", evsel__name(evsel));
2222 display_attr(&evsel->core.attr);
2223
2224 for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2225
2226 for (thread = 0; thread < nthreads; thread++) {
2227 int fd, group_fd;
2228 retry_open:
2229 if (thread >= nthreads)
2230 break;
2231
2232 if (!evsel->cgrp && !evsel->core.system_wide)
2233 pid = perf_thread_map__pid(threads, thread);
2234
2235 if (evsel->tool_event == PERF_TOOL_USER_TIME ||
2236 evsel->tool_event == PERF_TOOL_SYSTEM_TIME) {
2237 bool system = evsel->tool_event == PERF_TOOL_SYSTEM_TIME;
2238 __u64 *start_time = NULL;
2239
2240 if (evsel->core.attr.sample_period) {
2241 /* no sampling */
2242 err = -EINVAL;
2243 goto out_close;
2244 }
2245 if (pid > -1) {
2246 char buf[64];
2247
2248 snprintf(buf, sizeof(buf), "/proc/%d/stat", pid);
2249 fd = open(buf, O_RDONLY);
2250 evsel->pid_stat = true;
2251 } else {
2252 fd = open("/proc/stat", O_RDONLY);
2253 }
2254 FD(evsel, idx, thread) = fd;
2255 if (fd < 0) {
2256 err = -errno;
2257 goto out_close;
2258 }
2259 start_time = xyarray__entry(evsel->start_times, idx, thread);
2260 if (pid > -1) {
2261 err = read_pid_stat_field(fd, system ? 15 : 14,
2262 start_time);
2263 } else {
2264 struct perf_cpu cpu;
2265
2266 cpu = perf_cpu_map__cpu(evsel->core.cpus, idx);
2267 err = read_stat_field(fd, cpu, system ? 3 : 1,
2268 start_time);
2269 }
2270 if (err)
2271 goto out_close;
2272 continue;
2273 }
2274
2275 group_fd = get_group_fd(evsel, idx, thread);
2276
2277 if (group_fd == -2) {
2278 pr_debug("broken group leader for %s\n", evsel->name);
2279 err = -EINVAL;
2280 goto out_close;
2281 }
2282
2283 test_attr__ready();
2284
2285 /* Debug message used by test scripts */
2286 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
2287 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2288
2289 fd = sys_perf_event_open(&evsel->core.attr, pid,
2290 perf_cpu_map__cpu(cpus, idx).cpu,
2291 group_fd, evsel->open_flags);
2292
2293 FD(evsel, idx, thread) = fd;
2294
2295 if (fd < 0) {
2296 err = -errno;
2297
2298 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2299 err);
2300 goto try_fallback;
2301 }
2302
2303 bpf_counter__install_pe(evsel, idx, fd);
2304
2305 if (unlikely(test_attr__enabled)) {
2306 test_attr__open(&evsel->core.attr, pid,
2307 perf_cpu_map__cpu(cpus, idx),
2308 fd, group_fd, evsel->open_flags);
2309 }
2310
2311 /* Debug message used by test scripts */
2312 pr_debug2_peo(" = %d\n", fd);
2313
2314 if (evsel->bpf_fd >= 0) {
2315 int evt_fd = fd;
2316 int bpf_fd = evsel->bpf_fd;
2317
2318 err = ioctl(evt_fd,
2319 PERF_EVENT_IOC_SET_BPF,
2320 bpf_fd);
2321 if (err && errno != EEXIST) {
2322 pr_err("failed to attach bpf fd %d: %s\n",
2323 bpf_fd, strerror(errno));
2324 err = -EINVAL;
2325 goto out_close;
2326 }
2327 }
2328
2329 set_rlimit = NO_CHANGE;
2330
2331 /*
2332 * If we succeeded but had to kill clockid, fail and
2333 * have evsel__open_strerror() print us a nice error.
2334 */
2335 if (perf_missing_features.clockid ||
2336 perf_missing_features.clockid_wrong) {
2337 err = -EINVAL;
2338 goto out_close;
2339 }
2340 }
2341 }
2342
2343 return 0;
2344
2345 try_fallback:
2346 if (evsel__precise_ip_fallback(evsel))
2347 goto retry_open;
2348
2349 if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2350 idx, threads, thread, err)) {
2351 /* We just removed 1 thread, so lower the upper nthreads limit. */
2352 nthreads--;
2353
2354 /* ... and pretend like nothing have happened. */
2355 err = 0;
2356 goto retry_open;
2357 }
2358 /*
2359 * perf stat needs between 5 and 22 fds per CPU. When we run out
2360 * of them try to increase the limits.
2361 */
2362 if (err == -EMFILE && rlimit__increase_nofile(&set_rlimit))
2363 goto retry_open;
2364
2365 if (err != -EINVAL || idx > 0 || thread > 0)
2366 goto out_close;
2367
2368 if (evsel__detect_missing_features(evsel))
2369 goto fallback_missing_features;
2370 out_close:
2371 if (err)
2372 threads->err_thread = thread;
2373
2374 old_errno = errno;
2375 do {
2376 while (--thread >= 0) {
2377 if (FD(evsel, idx, thread) >= 0)
2378 close(FD(evsel, idx, thread));
2379 FD(evsel, idx, thread) = -1;
2380 }
2381 thread = nthreads;
2382 } while (--idx >= 0);
2383 errno = old_errno;
2384 return err;
2385 }
2386
evsel__open(struct evsel * evsel,struct perf_cpu_map * cpus,struct perf_thread_map * threads)2387 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2388 struct perf_thread_map *threads)
2389 {
2390 return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2391 }
2392
evsel__close(struct evsel * evsel)2393 void evsel__close(struct evsel *evsel)
2394 {
2395 perf_evsel__close(&evsel->core);
2396 perf_evsel__free_id(&evsel->core);
2397 }
2398
evsel__open_per_cpu(struct evsel * evsel,struct perf_cpu_map * cpus,int cpu_map_idx)2399 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2400 {
2401 if (cpu_map_idx == -1)
2402 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2403
2404 return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2405 }
2406
evsel__open_per_thread(struct evsel * evsel,struct perf_thread_map * threads)2407 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2408 {
2409 return evsel__open(evsel, NULL, threads);
2410 }
2411
perf_evsel__parse_id_sample(const struct evsel * evsel,const union perf_event * event,struct perf_sample * sample)2412 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2413 const union perf_event *event,
2414 struct perf_sample *sample)
2415 {
2416 u64 type = evsel->core.attr.sample_type;
2417 const __u64 *array = event->sample.array;
2418 bool swapped = evsel->needs_swap;
2419 union u64_swap u;
2420
2421 array += ((event->header.size -
2422 sizeof(event->header)) / sizeof(u64)) - 1;
2423
2424 if (type & PERF_SAMPLE_IDENTIFIER) {
2425 sample->id = *array;
2426 array--;
2427 }
2428
2429 if (type & PERF_SAMPLE_CPU) {
2430 u.val64 = *array;
2431 if (swapped) {
2432 /* undo swap of u64, then swap on individual u32s */
2433 u.val64 = bswap_64(u.val64);
2434 u.val32[0] = bswap_32(u.val32[0]);
2435 }
2436
2437 sample->cpu = u.val32[0];
2438 array--;
2439 }
2440
2441 if (type & PERF_SAMPLE_STREAM_ID) {
2442 sample->stream_id = *array;
2443 array--;
2444 }
2445
2446 if (type & PERF_SAMPLE_ID) {
2447 sample->id = *array;
2448 array--;
2449 }
2450
2451 if (type & PERF_SAMPLE_TIME) {
2452 sample->time = *array;
2453 array--;
2454 }
2455
2456 if (type & PERF_SAMPLE_TID) {
2457 u.val64 = *array;
2458 if (swapped) {
2459 /* undo swap of u64, then swap on individual u32s */
2460 u.val64 = bswap_64(u.val64);
2461 u.val32[0] = bswap_32(u.val32[0]);
2462 u.val32[1] = bswap_32(u.val32[1]);
2463 }
2464
2465 sample->pid = u.val32[0];
2466 sample->tid = u.val32[1];
2467 array--;
2468 }
2469
2470 return 0;
2471 }
2472
overflow(const void * endp,u16 max_size,const void * offset,u64 size)2473 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2474 u64 size)
2475 {
2476 return size > max_size || offset + size > endp;
2477 }
2478
2479 #define OVERFLOW_CHECK(offset, size, max_size) \
2480 do { \
2481 if (overflow(endp, (max_size), (offset), (size))) \
2482 return -EFAULT; \
2483 } while (0)
2484
2485 #define OVERFLOW_CHECK_u64(offset) \
2486 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2487
2488 static int
perf_event__check_size(union perf_event * event,unsigned int sample_size)2489 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2490 {
2491 /*
2492 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2493 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2494 * check the format does not go past the end of the event.
2495 */
2496 if (sample_size + sizeof(event->header) > event->header.size)
2497 return -EFAULT;
2498
2499 return 0;
2500 }
2501
arch_perf_parse_sample_weight(struct perf_sample * data,const __u64 * array,u64 type __maybe_unused)2502 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2503 const __u64 *array,
2504 u64 type __maybe_unused)
2505 {
2506 data->weight = *array;
2507 }
2508
evsel__bitfield_swap_branch_flags(u64 value)2509 u64 evsel__bitfield_swap_branch_flags(u64 value)
2510 {
2511 u64 new_val = 0;
2512
2513 /*
2514 * branch_flags
2515 * union {
2516 * u64 values;
2517 * struct {
2518 * mispred:1 //target mispredicted
2519 * predicted:1 //target predicted
2520 * in_tx:1 //in transaction
2521 * abort:1 //transaction abort
2522 * cycles:16 //cycle count to last branch
2523 * type:4 //branch type
2524 * spec:2 //branch speculation info
2525 * new_type:4 //additional branch type
2526 * priv:3 //privilege level
2527 * reserved:31
2528 * }
2529 * }
2530 *
2531 * Avoid bswap64() the entire branch_flag.value,
2532 * as it has variable bit-field sizes. Instead the
2533 * macro takes the bit-field position/size,
2534 * swaps it based on the host endianness.
2535 */
2536 if (host_is_bigendian()) {
2537 new_val = bitfield_swap(value, 0, 1);
2538 new_val |= bitfield_swap(value, 1, 1);
2539 new_val |= bitfield_swap(value, 2, 1);
2540 new_val |= bitfield_swap(value, 3, 1);
2541 new_val |= bitfield_swap(value, 4, 16);
2542 new_val |= bitfield_swap(value, 20, 4);
2543 new_val |= bitfield_swap(value, 24, 2);
2544 new_val |= bitfield_swap(value, 26, 4);
2545 new_val |= bitfield_swap(value, 30, 3);
2546 new_val |= bitfield_swap(value, 33, 31);
2547 } else {
2548 new_val = bitfield_swap(value, 63, 1);
2549 new_val |= bitfield_swap(value, 62, 1);
2550 new_val |= bitfield_swap(value, 61, 1);
2551 new_val |= bitfield_swap(value, 60, 1);
2552 new_val |= bitfield_swap(value, 44, 16);
2553 new_val |= bitfield_swap(value, 40, 4);
2554 new_val |= bitfield_swap(value, 38, 2);
2555 new_val |= bitfield_swap(value, 34, 4);
2556 new_val |= bitfield_swap(value, 31, 3);
2557 new_val |= bitfield_swap(value, 0, 31);
2558 }
2559
2560 return new_val;
2561 }
2562
evsel__has_branch_counters(const struct evsel * evsel)2563 static inline bool evsel__has_branch_counters(const struct evsel *evsel)
2564 {
2565 struct evsel *cur, *leader = evsel__leader(evsel);
2566
2567 /* The branch counters feature only supports group */
2568 if (!leader || !evsel->evlist)
2569 return false;
2570
2571 evlist__for_each_entry(evsel->evlist, cur) {
2572 if ((leader == evsel__leader(cur)) &&
2573 (cur->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS))
2574 return true;
2575 }
2576 return false;
2577 }
2578
evsel__parse_sample(struct evsel * evsel,union perf_event * event,struct perf_sample * data)2579 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2580 struct perf_sample *data)
2581 {
2582 u64 type = evsel->core.attr.sample_type;
2583 bool swapped = evsel->needs_swap;
2584 const __u64 *array;
2585 u16 max_size = event->header.size;
2586 const void *endp = (void *)event + max_size;
2587 u64 sz;
2588
2589 /*
2590 * used for cross-endian analysis. See git commit 65014ab3
2591 * for why this goofiness is needed.
2592 */
2593 union u64_swap u;
2594
2595 memset(data, 0, sizeof(*data));
2596 data->cpu = data->pid = data->tid = -1;
2597 data->stream_id = data->id = data->time = -1ULL;
2598 data->period = evsel->core.attr.sample_period;
2599 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2600 data->misc = event->header.misc;
2601 data->data_src = PERF_MEM_DATA_SRC_NONE;
2602 data->vcpu = -1;
2603
2604 if (event->header.type != PERF_RECORD_SAMPLE) {
2605 if (!evsel->core.attr.sample_id_all)
2606 return 0;
2607 return perf_evsel__parse_id_sample(evsel, event, data);
2608 }
2609
2610 array = event->sample.array;
2611
2612 if (perf_event__check_size(event, evsel->sample_size))
2613 return -EFAULT;
2614
2615 if (type & PERF_SAMPLE_IDENTIFIER) {
2616 data->id = *array;
2617 array++;
2618 }
2619
2620 if (type & PERF_SAMPLE_IP) {
2621 data->ip = *array;
2622 array++;
2623 }
2624
2625 if (type & PERF_SAMPLE_TID) {
2626 u.val64 = *array;
2627 if (swapped) {
2628 /* undo swap of u64, then swap on individual u32s */
2629 u.val64 = bswap_64(u.val64);
2630 u.val32[0] = bswap_32(u.val32[0]);
2631 u.val32[1] = bswap_32(u.val32[1]);
2632 }
2633
2634 data->pid = u.val32[0];
2635 data->tid = u.val32[1];
2636 array++;
2637 }
2638
2639 if (type & PERF_SAMPLE_TIME) {
2640 data->time = *array;
2641 array++;
2642 }
2643
2644 if (type & PERF_SAMPLE_ADDR) {
2645 data->addr = *array;
2646 array++;
2647 }
2648
2649 if (type & PERF_SAMPLE_ID) {
2650 data->id = *array;
2651 array++;
2652 }
2653
2654 if (type & PERF_SAMPLE_STREAM_ID) {
2655 data->stream_id = *array;
2656 array++;
2657 }
2658
2659 if (type & PERF_SAMPLE_CPU) {
2660
2661 u.val64 = *array;
2662 if (swapped) {
2663 /* undo swap of u64, then swap on individual u32s */
2664 u.val64 = bswap_64(u.val64);
2665 u.val32[0] = bswap_32(u.val32[0]);
2666 }
2667
2668 data->cpu = u.val32[0];
2669 array++;
2670 }
2671
2672 if (type & PERF_SAMPLE_PERIOD) {
2673 data->period = *array;
2674 array++;
2675 }
2676
2677 if (type & PERF_SAMPLE_READ) {
2678 u64 read_format = evsel->core.attr.read_format;
2679
2680 OVERFLOW_CHECK_u64(array);
2681 if (read_format & PERF_FORMAT_GROUP)
2682 data->read.group.nr = *array;
2683 else
2684 data->read.one.value = *array;
2685
2686 array++;
2687
2688 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2689 OVERFLOW_CHECK_u64(array);
2690 data->read.time_enabled = *array;
2691 array++;
2692 }
2693
2694 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2695 OVERFLOW_CHECK_u64(array);
2696 data->read.time_running = *array;
2697 array++;
2698 }
2699
2700 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2701 if (read_format & PERF_FORMAT_GROUP) {
2702 const u64 max_group_nr = UINT64_MAX /
2703 sizeof(struct sample_read_value);
2704
2705 if (data->read.group.nr > max_group_nr)
2706 return -EFAULT;
2707
2708 sz = data->read.group.nr * sample_read_value_size(read_format);
2709 OVERFLOW_CHECK(array, sz, max_size);
2710 data->read.group.values =
2711 (struct sample_read_value *)array;
2712 array = (void *)array + sz;
2713 } else {
2714 OVERFLOW_CHECK_u64(array);
2715 data->read.one.id = *array;
2716 array++;
2717
2718 if (read_format & PERF_FORMAT_LOST) {
2719 OVERFLOW_CHECK_u64(array);
2720 data->read.one.lost = *array;
2721 array++;
2722 }
2723 }
2724 }
2725
2726 if (type & PERF_SAMPLE_CALLCHAIN) {
2727 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2728
2729 OVERFLOW_CHECK_u64(array);
2730 data->callchain = (struct ip_callchain *)array++;
2731 if (data->callchain->nr > max_callchain_nr)
2732 return -EFAULT;
2733 sz = data->callchain->nr * sizeof(u64);
2734 OVERFLOW_CHECK(array, sz, max_size);
2735 array = (void *)array + sz;
2736 }
2737
2738 if (type & PERF_SAMPLE_RAW) {
2739 OVERFLOW_CHECK_u64(array);
2740 u.val64 = *array;
2741
2742 /*
2743 * Undo swap of u64, then swap on individual u32s,
2744 * get the size of the raw area and undo all of the
2745 * swap. The pevent interface handles endianness by
2746 * itself.
2747 */
2748 if (swapped) {
2749 u.val64 = bswap_64(u.val64);
2750 u.val32[0] = bswap_32(u.val32[0]);
2751 u.val32[1] = bswap_32(u.val32[1]);
2752 }
2753 data->raw_size = u.val32[0];
2754
2755 /*
2756 * The raw data is aligned on 64bits including the
2757 * u32 size, so it's safe to use mem_bswap_64.
2758 */
2759 if (swapped)
2760 mem_bswap_64((void *) array, data->raw_size);
2761
2762 array = (void *)array + sizeof(u32);
2763
2764 OVERFLOW_CHECK(array, data->raw_size, max_size);
2765 data->raw_data = (void *)array;
2766 array = (void *)array + data->raw_size;
2767 }
2768
2769 if (type & PERF_SAMPLE_BRANCH_STACK) {
2770 const u64 max_branch_nr = UINT64_MAX /
2771 sizeof(struct branch_entry);
2772 struct branch_entry *e;
2773 unsigned int i;
2774
2775 OVERFLOW_CHECK_u64(array);
2776 data->branch_stack = (struct branch_stack *)array++;
2777
2778 if (data->branch_stack->nr > max_branch_nr)
2779 return -EFAULT;
2780
2781 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2782 if (evsel__has_branch_hw_idx(evsel)) {
2783 sz += sizeof(u64);
2784 e = &data->branch_stack->entries[0];
2785 } else {
2786 data->no_hw_idx = true;
2787 /*
2788 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2789 * only nr and entries[] will be output by kernel.
2790 */
2791 e = (struct branch_entry *)&data->branch_stack->hw_idx;
2792 }
2793
2794 if (swapped) {
2795 /*
2796 * struct branch_flag does not have endian
2797 * specific bit field definition. And bswap
2798 * will not resolve the issue, since these
2799 * are bit fields.
2800 *
2801 * evsel__bitfield_swap_branch_flags() uses a
2802 * bitfield_swap macro to swap the bit position
2803 * based on the host endians.
2804 */
2805 for (i = 0; i < data->branch_stack->nr; i++, e++)
2806 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2807 }
2808
2809 OVERFLOW_CHECK(array, sz, max_size);
2810 array = (void *)array + sz;
2811
2812 if (evsel__has_branch_counters(evsel)) {
2813 OVERFLOW_CHECK_u64(array);
2814
2815 data->branch_stack_cntr = (u64 *)array;
2816 sz = data->branch_stack->nr * sizeof(u64);
2817
2818 OVERFLOW_CHECK(array, sz, max_size);
2819 array = (void *)array + sz;
2820 }
2821 }
2822
2823 if (type & PERF_SAMPLE_REGS_USER) {
2824 OVERFLOW_CHECK_u64(array);
2825 data->user_regs.abi = *array;
2826 array++;
2827
2828 if (data->user_regs.abi) {
2829 u64 mask = evsel->core.attr.sample_regs_user;
2830
2831 sz = hweight64(mask) * sizeof(u64);
2832 OVERFLOW_CHECK(array, sz, max_size);
2833 data->user_regs.mask = mask;
2834 data->user_regs.regs = (u64 *)array;
2835 array = (void *)array + sz;
2836 }
2837 }
2838
2839 if (type & PERF_SAMPLE_STACK_USER) {
2840 OVERFLOW_CHECK_u64(array);
2841 sz = *array++;
2842
2843 data->user_stack.offset = ((char *)(array - 1)
2844 - (char *) event);
2845
2846 if (!sz) {
2847 data->user_stack.size = 0;
2848 } else {
2849 OVERFLOW_CHECK(array, sz, max_size);
2850 data->user_stack.data = (char *)array;
2851 array = (void *)array + sz;
2852 OVERFLOW_CHECK_u64(array);
2853 data->user_stack.size = *array++;
2854 if (WARN_ONCE(data->user_stack.size > sz,
2855 "user stack dump failure\n"))
2856 return -EFAULT;
2857 }
2858 }
2859
2860 if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2861 OVERFLOW_CHECK_u64(array);
2862 arch_perf_parse_sample_weight(data, array, type);
2863 array++;
2864 }
2865
2866 if (type & PERF_SAMPLE_DATA_SRC) {
2867 OVERFLOW_CHECK_u64(array);
2868 data->data_src = *array;
2869 array++;
2870 }
2871
2872 if (type & PERF_SAMPLE_TRANSACTION) {
2873 OVERFLOW_CHECK_u64(array);
2874 data->transaction = *array;
2875 array++;
2876 }
2877
2878 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2879 if (type & PERF_SAMPLE_REGS_INTR) {
2880 OVERFLOW_CHECK_u64(array);
2881 data->intr_regs.abi = *array;
2882 array++;
2883
2884 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2885 u64 mask = evsel->core.attr.sample_regs_intr;
2886
2887 sz = hweight64(mask) * sizeof(u64);
2888 OVERFLOW_CHECK(array, sz, max_size);
2889 data->intr_regs.mask = mask;
2890 data->intr_regs.regs = (u64 *)array;
2891 array = (void *)array + sz;
2892 }
2893 }
2894
2895 data->phys_addr = 0;
2896 if (type & PERF_SAMPLE_PHYS_ADDR) {
2897 data->phys_addr = *array;
2898 array++;
2899 }
2900
2901 data->cgroup = 0;
2902 if (type & PERF_SAMPLE_CGROUP) {
2903 data->cgroup = *array;
2904 array++;
2905 }
2906
2907 data->data_page_size = 0;
2908 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2909 data->data_page_size = *array;
2910 array++;
2911 }
2912
2913 data->code_page_size = 0;
2914 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2915 data->code_page_size = *array;
2916 array++;
2917 }
2918
2919 if (type & PERF_SAMPLE_AUX) {
2920 OVERFLOW_CHECK_u64(array);
2921 sz = *array++;
2922
2923 OVERFLOW_CHECK(array, sz, max_size);
2924 /* Undo swap of data */
2925 if (swapped)
2926 mem_bswap_64((char *)array, sz);
2927 data->aux_sample.size = sz;
2928 data->aux_sample.data = (char *)array;
2929 array = (void *)array + sz;
2930 }
2931
2932 return 0;
2933 }
2934
evsel__parse_sample_timestamp(struct evsel * evsel,union perf_event * event,u64 * timestamp)2935 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2936 u64 *timestamp)
2937 {
2938 u64 type = evsel->core.attr.sample_type;
2939 const __u64 *array;
2940
2941 if (!(type & PERF_SAMPLE_TIME))
2942 return -1;
2943
2944 if (event->header.type != PERF_RECORD_SAMPLE) {
2945 struct perf_sample data = {
2946 .time = -1ULL,
2947 };
2948
2949 if (!evsel->core.attr.sample_id_all)
2950 return -1;
2951 if (perf_evsel__parse_id_sample(evsel, event, &data))
2952 return -1;
2953
2954 *timestamp = data.time;
2955 return 0;
2956 }
2957
2958 array = event->sample.array;
2959
2960 if (perf_event__check_size(event, evsel->sample_size))
2961 return -EFAULT;
2962
2963 if (type & PERF_SAMPLE_IDENTIFIER)
2964 array++;
2965
2966 if (type & PERF_SAMPLE_IP)
2967 array++;
2968
2969 if (type & PERF_SAMPLE_TID)
2970 array++;
2971
2972 if (type & PERF_SAMPLE_TIME)
2973 *timestamp = *array;
2974
2975 return 0;
2976 }
2977
evsel__id_hdr_size(struct evsel * evsel)2978 u16 evsel__id_hdr_size(struct evsel *evsel)
2979 {
2980 u64 sample_type = evsel->core.attr.sample_type;
2981 u16 size = 0;
2982
2983 if (sample_type & PERF_SAMPLE_TID)
2984 size += sizeof(u64);
2985
2986 if (sample_type & PERF_SAMPLE_TIME)
2987 size += sizeof(u64);
2988
2989 if (sample_type & PERF_SAMPLE_ID)
2990 size += sizeof(u64);
2991
2992 if (sample_type & PERF_SAMPLE_STREAM_ID)
2993 size += sizeof(u64);
2994
2995 if (sample_type & PERF_SAMPLE_CPU)
2996 size += sizeof(u64);
2997
2998 if (sample_type & PERF_SAMPLE_IDENTIFIER)
2999 size += sizeof(u64);
3000
3001 return size;
3002 }
3003
3004 #ifdef HAVE_LIBTRACEEVENT
evsel__field(struct evsel * evsel,const char * name)3005 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
3006 {
3007 return tep_find_field(evsel->tp_format, name);
3008 }
3009
evsel__common_field(struct evsel * evsel,const char * name)3010 struct tep_format_field *evsel__common_field(struct evsel *evsel, const char *name)
3011 {
3012 return tep_find_common_field(evsel->tp_format, name);
3013 }
3014
evsel__rawptr(struct evsel * evsel,struct perf_sample * sample,const char * name)3015 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
3016 {
3017 struct tep_format_field *field = evsel__field(evsel, name);
3018 int offset;
3019
3020 if (!field)
3021 return NULL;
3022
3023 offset = field->offset;
3024
3025 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
3026 offset = *(int *)(sample->raw_data + field->offset);
3027 offset &= 0xffff;
3028 if (tep_field_is_relative(field->flags))
3029 offset += field->offset + field->size;
3030 }
3031
3032 return sample->raw_data + offset;
3033 }
3034
format_field__intval(struct tep_format_field * field,struct perf_sample * sample,bool needs_swap)3035 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
3036 bool needs_swap)
3037 {
3038 u64 value;
3039 void *ptr = sample->raw_data + field->offset;
3040
3041 switch (field->size) {
3042 case 1:
3043 return *(u8 *)ptr;
3044 case 2:
3045 value = *(u16 *)ptr;
3046 break;
3047 case 4:
3048 value = *(u32 *)ptr;
3049 break;
3050 case 8:
3051 memcpy(&value, ptr, sizeof(u64));
3052 break;
3053 default:
3054 return 0;
3055 }
3056
3057 if (!needs_swap)
3058 return value;
3059
3060 switch (field->size) {
3061 case 2:
3062 return bswap_16(value);
3063 case 4:
3064 return bswap_32(value);
3065 case 8:
3066 return bswap_64(value);
3067 default:
3068 return 0;
3069 }
3070
3071 return 0;
3072 }
3073
evsel__intval(struct evsel * evsel,struct perf_sample * sample,const char * name)3074 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
3075 {
3076 struct tep_format_field *field = evsel__field(evsel, name);
3077
3078 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
3079 }
3080
evsel__intval_common(struct evsel * evsel,struct perf_sample * sample,const char * name)3081 u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const char *name)
3082 {
3083 struct tep_format_field *field = evsel__common_field(evsel, name);
3084
3085 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
3086 }
3087
evsel__taskstate(struct evsel * evsel,struct perf_sample * sample,const char * name)3088 char evsel__taskstate(struct evsel *evsel, struct perf_sample *sample, const char *name)
3089 {
3090 static struct tep_format_field *prev_state_field;
3091 static const char *states;
3092 struct tep_format_field *field;
3093 unsigned long long val;
3094 unsigned int bit;
3095 char state = '?'; /* '?' denotes unknown task state */
3096
3097 field = evsel__field(evsel, name);
3098
3099 if (!field)
3100 return state;
3101
3102 if (!states || field != prev_state_field) {
3103 states = parse_task_states(field);
3104 if (!states)
3105 return state;
3106 prev_state_field = field;
3107 }
3108
3109 /*
3110 * Note since the kernel exposes TASK_REPORT_MAX to userspace
3111 * to denote the 'preempted' state, we might as welll report
3112 * 'R' for this case, which make senses to users as well.
3113 *
3114 * We can change this if we have a good reason in the future.
3115 */
3116 val = evsel__intval(evsel, sample, name);
3117 bit = val ? ffs(val) : 0;
3118 state = (!bit || bit > strlen(states)) ? 'R' : states[bit-1];
3119 return state;
3120 }
3121 #endif
3122
evsel__fallback(struct evsel * evsel,struct target * target,int err,char * msg,size_t msgsize)3123 bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
3124 char *msg, size_t msgsize)
3125 {
3126 int paranoid;
3127
3128 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
3129 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
3130 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
3131 /*
3132 * If it's cycles then fall back to hrtimer based cpu-clock sw
3133 * counter, which is always available even if no PMU support.
3134 *
3135 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
3136 * b0a873e).
3137 */
3138 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
3139 evsel->core.attr.config = target__has_cpu(target)
3140 ? PERF_COUNT_SW_CPU_CLOCK
3141 : PERF_COUNT_SW_TASK_CLOCK;
3142 scnprintf(msg, msgsize,
3143 "The cycles event is not supported, trying to fall back to %s",
3144 target__has_cpu(target) ? "cpu-clock" : "task-clock");
3145
3146 zfree(&evsel->name);
3147 return true;
3148 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
3149 (paranoid = perf_event_paranoid()) > 1) {
3150 const char *name = evsel__name(evsel);
3151 char *new_name;
3152 const char *sep = ":";
3153
3154 /* If event has exclude user then don't exclude kernel. */
3155 if (evsel->core.attr.exclude_user)
3156 return false;
3157
3158 /* Is there already the separator in the name. */
3159 if (strchr(name, '/') ||
3160 (strchr(name, ':') && !evsel->is_libpfm_event))
3161 sep = "";
3162
3163 if (asprintf(&new_name, "%s%su", name, sep) < 0)
3164 return false;
3165
3166 free(evsel->name);
3167 evsel->name = new_name;
3168 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
3169 "to fall back to excluding kernel and hypervisor "
3170 " samples", paranoid);
3171 evsel->core.attr.exclude_kernel = 1;
3172 evsel->core.attr.exclude_hv = 1;
3173
3174 return true;
3175 }
3176
3177 return false;
3178 }
3179
find_process(const char * name)3180 static bool find_process(const char *name)
3181 {
3182 size_t len = strlen(name);
3183 DIR *dir;
3184 struct dirent *d;
3185 int ret = -1;
3186
3187 dir = opendir(procfs__mountpoint());
3188 if (!dir)
3189 return false;
3190
3191 /* Walk through the directory. */
3192 while (ret && (d = readdir(dir)) != NULL) {
3193 char path[PATH_MAX];
3194 char *data;
3195 size_t size;
3196
3197 if ((d->d_type != DT_DIR) ||
3198 !strcmp(".", d->d_name) ||
3199 !strcmp("..", d->d_name))
3200 continue;
3201
3202 scnprintf(path, sizeof(path), "%s/%s/comm",
3203 procfs__mountpoint(), d->d_name);
3204
3205 if (filename__read_str(path, &data, &size))
3206 continue;
3207
3208 ret = strncmp(name, data, len);
3209 free(data);
3210 }
3211
3212 closedir(dir);
3213 return ret ? false : true;
3214 }
3215
arch_evsel__open_strerror(struct evsel * evsel __maybe_unused,char * msg __maybe_unused,size_t size __maybe_unused)3216 int __weak arch_evsel__open_strerror(struct evsel *evsel __maybe_unused,
3217 char *msg __maybe_unused,
3218 size_t size __maybe_unused)
3219 {
3220 return 0;
3221 }
3222
evsel__open_strerror(struct evsel * evsel,struct target * target,int err,char * msg,size_t size)3223 int evsel__open_strerror(struct evsel *evsel, struct target *target,
3224 int err, char *msg, size_t size)
3225 {
3226 char sbuf[STRERR_BUFSIZE];
3227 int printed = 0, enforced = 0;
3228 int ret;
3229
3230 switch (err) {
3231 case EPERM:
3232 case EACCES:
3233 printed += scnprintf(msg + printed, size - printed,
3234 "Access to performance monitoring and observability operations is limited.\n");
3235
3236 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
3237 if (enforced) {
3238 printed += scnprintf(msg + printed, size - printed,
3239 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
3240 "monitoring and observability operations. Inspect system audit records for\n"
3241 "more perf_event access control information and adjusting the policy.\n");
3242 }
3243 }
3244
3245 if (err == EPERM)
3246 printed += scnprintf(msg, size,
3247 "No permission to enable %s event.\n\n", evsel__name(evsel));
3248
3249 return scnprintf(msg + printed, size - printed,
3250 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
3251 "access to performance monitoring and observability operations for processes\n"
3252 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
3253 "More information can be found at 'Perf events and tool security' document:\n"
3254 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
3255 "perf_event_paranoid setting is %d:\n"
3256 " -1: Allow use of (almost) all events by all users\n"
3257 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
3258 ">= 0: Disallow raw and ftrace function tracepoint access\n"
3259 ">= 1: Disallow CPU event access\n"
3260 ">= 2: Disallow kernel profiling\n"
3261 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
3262 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
3263 perf_event_paranoid());
3264 case ENOENT:
3265 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
3266 case EMFILE:
3267 return scnprintf(msg, size, "%s",
3268 "Too many events are opened.\n"
3269 "Probably the maximum number of open file descriptors has been reached.\n"
3270 "Hint: Try again after reducing the number of events.\n"
3271 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3272 case ENOMEM:
3273 if (evsel__has_callchain(evsel) &&
3274 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3275 return scnprintf(msg, size,
3276 "Not enough memory to setup event with callchain.\n"
3277 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3278 "Hint: Current value: %d", sysctl__max_stack());
3279 break;
3280 case ENODEV:
3281 if (target->cpu_list)
3282 return scnprintf(msg, size, "%s",
3283 "No such device - did you specify an out-of-range profile CPU?");
3284 break;
3285 case EOPNOTSUPP:
3286 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3287 return scnprintf(msg, size,
3288 "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3289 evsel__name(evsel));
3290 if (evsel->core.attr.aux_output)
3291 return scnprintf(msg, size,
3292 "%s: PMU Hardware doesn't support 'aux_output' feature",
3293 evsel__name(evsel));
3294 if (evsel->core.attr.sample_period != 0)
3295 return scnprintf(msg, size,
3296 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3297 evsel__name(evsel));
3298 if (evsel->core.attr.precise_ip)
3299 return scnprintf(msg, size, "%s",
3300 "\'precise\' request may not be supported. Try removing 'p' modifier.");
3301 #if defined(__i386__) || defined(__x86_64__)
3302 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3303 return scnprintf(msg, size, "%s",
3304 "No hardware sampling interrupt available.\n");
3305 #endif
3306 break;
3307 case EBUSY:
3308 if (find_process("oprofiled"))
3309 return scnprintf(msg, size,
3310 "The PMU counters are busy/taken by another profiler.\n"
3311 "We found oprofile daemon running, please stop it and try again.");
3312 break;
3313 case EINVAL:
3314 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3315 return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3316 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3317 return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3318 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3319 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3320 if (perf_missing_features.clockid)
3321 return scnprintf(msg, size, "clockid feature not supported.");
3322 if (perf_missing_features.clockid_wrong)
3323 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3324 if (perf_missing_features.aux_output)
3325 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3326 if (!target__has_cpu(target))
3327 return scnprintf(msg, size,
3328 "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3329 evsel__name(evsel));
3330
3331 break;
3332 case ENODATA:
3333 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3334 "Please add an auxiliary event in front of the load latency event.");
3335 default:
3336 break;
3337 }
3338
3339 ret = arch_evsel__open_strerror(evsel, msg, size);
3340 if (ret)
3341 return ret;
3342
3343 return scnprintf(msg, size,
3344 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3345 "/bin/dmesg | grep -i perf may provide additional information.\n",
3346 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3347 }
3348
evsel__env(struct evsel * evsel)3349 struct perf_env *evsel__env(struct evsel *evsel)
3350 {
3351 if (evsel && evsel->evlist && evsel->evlist->env)
3352 return evsel->evlist->env;
3353 return &perf_env;
3354 }
3355
store_evsel_ids(struct evsel * evsel,struct evlist * evlist)3356 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3357 {
3358 int cpu_map_idx, thread;
3359
3360 for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3361 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3362 thread++) {
3363 int fd = FD(evsel, cpu_map_idx, thread);
3364
3365 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3366 cpu_map_idx, thread, fd) < 0)
3367 return -1;
3368 }
3369 }
3370
3371 return 0;
3372 }
3373
evsel__store_ids(struct evsel * evsel,struct evlist * evlist)3374 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3375 {
3376 struct perf_cpu_map *cpus = evsel->core.cpus;
3377 struct perf_thread_map *threads = evsel->core.threads;
3378
3379 if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3380 return -ENOMEM;
3381
3382 return store_evsel_ids(evsel, evlist);
3383 }
3384
evsel__zero_per_pkg(struct evsel * evsel)3385 void evsel__zero_per_pkg(struct evsel *evsel)
3386 {
3387 struct hashmap_entry *cur;
3388 size_t bkt;
3389
3390 if (evsel->per_pkg_mask) {
3391 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3392 zfree(&cur->pkey);
3393
3394 hashmap__clear(evsel->per_pkg_mask);
3395 }
3396 }
3397
3398 /**
3399 * evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this
3400 * will be false on hybrid systems for hardware and legacy
3401 * cache events.
3402 */
evsel__is_hybrid(const struct evsel * evsel)3403 bool evsel__is_hybrid(const struct evsel *evsel)
3404 {
3405 if (perf_pmus__num_core_pmus() == 1)
3406 return false;
3407
3408 return evsel->core.is_pmu_core;
3409 }
3410
evsel__leader(const struct evsel * evsel)3411 struct evsel *evsel__leader(const struct evsel *evsel)
3412 {
3413 return container_of(evsel->core.leader, struct evsel, core);
3414 }
3415
evsel__has_leader(struct evsel * evsel,struct evsel * leader)3416 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3417 {
3418 return evsel->core.leader == &leader->core;
3419 }
3420
evsel__is_leader(struct evsel * evsel)3421 bool evsel__is_leader(struct evsel *evsel)
3422 {
3423 return evsel__has_leader(evsel, evsel);
3424 }
3425
evsel__set_leader(struct evsel * evsel,struct evsel * leader)3426 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3427 {
3428 evsel->core.leader = &leader->core;
3429 }
3430
evsel__source_count(const struct evsel * evsel)3431 int evsel__source_count(const struct evsel *evsel)
3432 {
3433 struct evsel *pos;
3434 int count = 0;
3435
3436 evlist__for_each_entry(evsel->evlist, pos) {
3437 if (pos->metric_leader == evsel)
3438 count++;
3439 }
3440 return count;
3441 }
3442
arch_evsel__must_be_in_group(const struct evsel * evsel __maybe_unused)3443 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3444 {
3445 return false;
3446 }
3447
3448 /*
3449 * Remove an event from a given group (leader).
3450 * Some events, e.g., perf metrics Topdown events,
3451 * must always be grouped. Ignore the events.
3452 */
evsel__remove_from_group(struct evsel * evsel,struct evsel * leader)3453 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3454 {
3455 if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3456 evsel__set_leader(evsel, evsel);
3457 evsel->core.nr_members = 0;
3458 leader->core.nr_members--;
3459 }
3460 }
3461