?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'block-5.11-2021-01-10' of git://git.kernel.dk/linux-block
[linux/fpc-iii.git] / tools / perf / util / evsel.c
blobc26ea82220bd8625a753104e18de740373201d5a
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/fs/fs.h>
14 #include <api/fs/tracing_path.h>
15 #include <traceevent/event-parse.h>
16 #include <linux/hw_breakpoint.h>
17 #include <linux/perf_event.h>
18 #include <linux/compiler.h>
19 #include <linux/err.h>
20 #include <linux/zalloc.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
24 #include <dirent.h>
25 #include <stdlib.h>
26 #include <perf/evsel.h>
27 #include "asm/bug.h"
28 #include "callchain.h"
29 #include "cgroup.h"
30 #include "counts.h"
31 #include "event.h"
32 #include "evsel.h"
33 #include "util/env.h"
34 #include "util/evsel_config.h"
35 #include "util/evsel_fprintf.h"
36 #include "evlist.h"
37 #include <perf/cpumap.h>
38 #include "thread_map.h"
39 #include "target.h"
40 #include "perf_regs.h"
41 #include "record.h"
42 #include "debug.h"
43 #include "trace-event.h"
44 #include "stat.h"
45 #include "string2.h"
46 #include "memswap.h"
47 #include "util.h"
48 #include "../perf-sys.h"
49 #include "util/parse-branch-options.h"
50 #include <internal/xyarray.h>
51 #include <internal/lib.h>
52
53 #include <linux/ctype.h>
54
55 struct perf_missing_features perf_missing_features;
56
57 static clockid_t clockid;
58
59 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
60 {
61 return 0;
62 }
63
64 void __weak test_attr__ready(void) { }
65
66 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
67 {
68 }
69
70 static struct {
71 size_t size;
72 int (*init)(struct evsel *evsel);
73 void (*fini)(struct evsel *evsel);
74 } perf_evsel__object = {
75 .size = sizeof(struct evsel),
76 .init = evsel__no_extra_init,
77 .fini = evsel__no_extra_fini,
78 };
79
80 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
81 void (*fini)(struct evsel *evsel))
82 {
83
84 if (object_size == 0)
85 goto set_methods;
86
87 if (perf_evsel__object.size > object_size)
88 return -EINVAL;
89
90 perf_evsel__object.size = object_size;
91
92 set_methods:
93 if (init != NULL)
94 perf_evsel__object.init = init;
95
96 if (fini != NULL)
97 perf_evsel__object.fini = fini;
98
99 return 0;
100 }
101
102 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
103
104 int __evsel__sample_size(u64 sample_type)
105 {
106 u64 mask = sample_type & PERF_SAMPLE_MASK;
107 int size = 0;
108 int i;
109
110 for (i = 0; i < 64; i++) {
111 if (mask & (1ULL << i))
112 size++;
113 }
114
115 size *= sizeof(u64);
116
117 return size;
118 }
119
120 /**
121 * __perf_evsel__calc_id_pos - calculate id_pos.
122 * @sample_type: sample type
123 *
124 * This function returns the position of the event id (PERF_SAMPLE_ID or
125 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
126 * perf_record_sample.
127 */
128 static int __perf_evsel__calc_id_pos(u64 sample_type)
129 {
130 int idx = 0;
131
132 if (sample_type & PERF_SAMPLE_IDENTIFIER)
133 return 0;
134
135 if (!(sample_type & PERF_SAMPLE_ID))
136 return -1;
137
138 if (sample_type & PERF_SAMPLE_IP)
139 idx += 1;
140
141 if (sample_type & PERF_SAMPLE_TID)
142 idx += 1;
143
144 if (sample_type & PERF_SAMPLE_TIME)
145 idx += 1;
146
147 if (sample_type & PERF_SAMPLE_ADDR)
148 idx += 1;
149
150 return idx;
151 }
152
153 /**
154 * __perf_evsel__calc_is_pos - calculate is_pos.
155 * @sample_type: sample type
156 *
157 * This function returns the position (counting backwards) of the event id
158 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
159 * sample_id_all is used there is an id sample appended to non-sample events.
160 */
161 static int __perf_evsel__calc_is_pos(u64 sample_type)
162 {
163 int idx = 1;
164
165 if (sample_type & PERF_SAMPLE_IDENTIFIER)
166 return 1;
167
168 if (!(sample_type & PERF_SAMPLE_ID))
169 return -1;
170
171 if (sample_type & PERF_SAMPLE_CPU)
172 idx += 1;
173
174 if (sample_type & PERF_SAMPLE_STREAM_ID)
175 idx += 1;
176
177 return idx;
178 }
179
180 void evsel__calc_id_pos(struct evsel *evsel)
181 {
182 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
183 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
184 }
185
186 void __evsel__set_sample_bit(struct evsel *evsel,
187 enum perf_event_sample_format bit)
188 {
189 if (!(evsel->core.attr.sample_type & bit)) {
190 evsel->core.attr.sample_type |= bit;
191 evsel->sample_size += sizeof(u64);
192 evsel__calc_id_pos(evsel);
193 }
194 }
195
196 void __evsel__reset_sample_bit(struct evsel *evsel,
197 enum perf_event_sample_format bit)
198 {
199 if (evsel->core.attr.sample_type & bit) {
200 evsel->core.attr.sample_type &= ~bit;
201 evsel->sample_size -= sizeof(u64);
202 evsel__calc_id_pos(evsel);
203 }
204 }
205
206 void evsel__set_sample_id(struct evsel *evsel,
207 bool can_sample_identifier)
208 {
209 if (can_sample_identifier) {
210 evsel__reset_sample_bit(evsel, ID);
211 evsel__set_sample_bit(evsel, IDENTIFIER);
212 } else {
213 evsel__set_sample_bit(evsel, ID);
214 }
215 evsel->core.attr.read_format |= PERF_FORMAT_ID;
216 }
217
218 /**
219 * evsel__is_function_event - Return whether given evsel is a function
220 * trace event
221 *
222 * @evsel - evsel selector to be tested
223 *
224 * Return %true if event is function trace event
225 */
226 bool evsel__is_function_event(struct evsel *evsel)
227 {
228 #define FUNCTION_EVENT "ftrace:function"
229
230 return evsel->name &&
231 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
232
233 #undef FUNCTION_EVENT
234 }
235
236 void evsel__init(struct evsel *evsel,
237 struct perf_event_attr *attr, int idx)
238 {
239 perf_evsel__init(&evsel->core, attr);
240 evsel->idx = idx;
241 evsel->tracking = !idx;
242 evsel->leader = evsel;
243 evsel->unit = "";
244 evsel->scale = 1.0;
245 evsel->max_events = ULONG_MAX;
246 evsel->evlist = NULL;
247 evsel->bpf_obj = NULL;
248 evsel->bpf_fd = -1;
249 INIT_LIST_HEAD(&evsel->config_terms);
250 perf_evsel__object.init(evsel);
251 evsel->sample_size = __evsel__sample_size(attr->sample_type);
252 evsel__calc_id_pos(evsel);
253 evsel->cmdline_group_boundary = false;
254 evsel->metric_expr = NULL;
255 evsel->metric_name = NULL;
256 evsel->metric_events = NULL;
257 evsel->per_pkg_mask = NULL;
258 evsel->collect_stat = false;
259 evsel->pmu_name = NULL;
260 }
261
262 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
263 {
264 struct evsel *evsel = zalloc(perf_evsel__object.size);
265
266 if (!evsel)
267 return NULL;
268 evsel__init(evsel, attr, idx);
269
270 if (evsel__is_bpf_output(evsel)) {
271 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
272 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
273 evsel->core.attr.sample_period = 1;
274 }
275
276 if (evsel__is_clock(evsel)) {
277 /*
278 * The evsel->unit points to static alias->unit
279 * so it's ok to use static string in here.
280 */
281 static const char *unit = "msec";
282
283 evsel->unit = unit;
284 evsel->scale = 1e-6;
285 }
286
287 return evsel;
288 }
289
290 static bool perf_event_can_profile_kernel(void)
291 {
292 return perf_event_paranoid_check(1);
293 }
294
295 struct evsel *evsel__new_cycles(bool precise)
296 {
297 struct perf_event_attr attr = {
298 .type = PERF_TYPE_HARDWARE,
299 .config = PERF_COUNT_HW_CPU_CYCLES,
300 .exclude_kernel = !perf_event_can_profile_kernel(),
301 };
302 struct evsel *evsel;
303
304 event_attr_init(&attr);
305
306 if (!precise)
307 goto new_event;
308
309 /*
310 * Now let the usual logic to set up the perf_event_attr defaults
311 * to kick in when we return and before perf_evsel__open() is called.
312 */
313 new_event:
314 evsel = evsel__new(&attr);
315 if (evsel == NULL)
316 goto out;
317
318 evsel->precise_max = true;
319
320 /* use asprintf() because free(evsel) assumes name is allocated */
321 if (asprintf(&evsel->name, "cycles%s%s%.*s",
322 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
323 attr.exclude_kernel ? "u" : "",
324 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
325 goto error_free;
326 out:
327 return evsel;
328 error_free:
329 evsel__delete(evsel);
330 evsel = NULL;
331 goto out;
332 }
333
334 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
335 {
336 struct evsel_config_term *pos, *tmp;
337
338 list_for_each_entry(pos, &src->config_terms, list) {
339 tmp = malloc(sizeof(*tmp));
340 if (tmp == NULL)
341 return -ENOMEM;
342
343 *tmp = *pos;
344 if (tmp->free_str) {
345 tmp->val.str = strdup(pos->val.str);
346 if (tmp->val.str == NULL) {
347 free(tmp);
348 return -ENOMEM;
349 }
350 }
351 list_add_tail(&tmp->list, &dst->config_terms);
352 }
353 return 0;
354 }
355
356 /**
357 * evsel__clone - create a new evsel copied from @orig
358 * @orig: original evsel
359 *
360 * The assumption is that @orig is not configured nor opened yet.
361 * So we only care about the attributes that can be set while it's parsed.
362 */
363 struct evsel *evsel__clone(struct evsel *orig)
364 {
365 struct evsel *evsel;
366
367 BUG_ON(orig->core.fd);
368 BUG_ON(orig->counts);
369 BUG_ON(orig->priv);
370 BUG_ON(orig->per_pkg_mask);
371
372 /* cannot handle BPF objects for now */
373 if (orig->bpf_obj)
374 return NULL;
375
376 evsel = evsel__new(&orig->core.attr);
377 if (evsel == NULL)
378 return NULL;
379
380 evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
381 evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
382 evsel->core.threads = perf_thread_map__get(orig->core.threads);
383 evsel->core.nr_members = orig->core.nr_members;
384 evsel->core.system_wide = orig->core.system_wide;
385
386 if (orig->name) {
387 evsel->name = strdup(orig->name);
388 if (evsel->name == NULL)
389 goto out_err;
390 }
391 if (orig->group_name) {
392 evsel->group_name = strdup(orig->group_name);
393 if (evsel->group_name == NULL)
394 goto out_err;
395 }
396 if (orig->pmu_name) {
397 evsel->pmu_name = strdup(orig->pmu_name);
398 if (evsel->pmu_name == NULL)
399 goto out_err;
400 }
401 if (orig->filter) {
402 evsel->filter = strdup(orig->filter);
403 if (evsel->filter == NULL)
404 goto out_err;
405 }
406 evsel->cgrp = cgroup__get(orig->cgrp);
407 evsel->tp_format = orig->tp_format;
408 evsel->handler = orig->handler;
409 evsel->leader = orig->leader;
410
411 evsel->max_events = orig->max_events;
412 evsel->tool_event = orig->tool_event;
413 evsel->unit = orig->unit;
414 evsel->scale = orig->scale;
415 evsel->snapshot = orig->snapshot;
416 evsel->per_pkg = orig->per_pkg;
417 evsel->percore = orig->percore;
418 evsel->precise_max = orig->precise_max;
419 evsel->use_uncore_alias = orig->use_uncore_alias;
420 evsel->is_libpfm_event = orig->is_libpfm_event;
421
422 evsel->exclude_GH = orig->exclude_GH;
423 evsel->sample_read = orig->sample_read;
424 evsel->auto_merge_stats = orig->auto_merge_stats;
425 evsel->collect_stat = orig->collect_stat;
426 evsel->weak_group = orig->weak_group;
427
428 if (evsel__copy_config_terms(evsel, orig) < 0)
429 goto out_err;
430
431 return evsel;
432
433 out_err:
434 evsel__delete(evsel);
435 return NULL;
436 }
437
438 /*
439 * Returns pointer with encoded error via <linux/err.h> interface.
440 */
441 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
442 {
443 struct evsel *evsel = zalloc(perf_evsel__object.size);
444 int err = -ENOMEM;
445
446 if (evsel == NULL) {
447 goto out_err;
448 } else {
449 struct perf_event_attr attr = {
450 .type = PERF_TYPE_TRACEPOINT,
451 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
452 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
453 };
454
455 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
456 goto out_free;
457
458 evsel->tp_format = trace_event__tp_format(sys, name);
459 if (IS_ERR(evsel->tp_format)) {
460 err = PTR_ERR(evsel->tp_format);
461 goto out_free;
462 }
463
464 event_attr_init(&attr);
465 attr.config = evsel->tp_format->id;
466 attr.sample_period = 1;
467 evsel__init(evsel, &attr, idx);
468 }
469
470 return evsel;
471
472 out_free:
473 zfree(&evsel->name);
474 free(evsel);
475 out_err:
476 return ERR_PTR(err);
477 }
478
479 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
480 "cycles",
481 "instructions",
482 "cache-references",
483 "cache-misses",
484 "branches",
485 "branch-misses",
486 "bus-cycles",
487 "stalled-cycles-frontend",
488 "stalled-cycles-backend",
489 "ref-cycles",
490 };
491
492 static const char *__evsel__hw_name(u64 config)
493 {
494 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
495 return evsel__hw_names[config];
496
497 return "unknown-hardware";
498 }
499
500 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
501 {
502 int colon = 0, r = 0;
503 struct perf_event_attr *attr = &evsel->core.attr;
504 bool exclude_guest_default = false;
505
506 #define MOD_PRINT(context, mod) do { \
507 if (!attr->exclude_##context) { \
508 if (!colon) colon = ++r; \
509 r += scnprintf(bf + r, size - r, "%c", mod); \
510 } } while(0)
511
512 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
513 MOD_PRINT(kernel, 'k');
514 MOD_PRINT(user, 'u');
515 MOD_PRINT(hv, 'h');
516 exclude_guest_default = true;
517 }
518
519 if (attr->precise_ip) {
520 if (!colon)
521 colon = ++r;
522 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
523 exclude_guest_default = true;
524 }
525
526 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
527 MOD_PRINT(host, 'H');
528 MOD_PRINT(guest, 'G');
529 }
530 #undef MOD_PRINT
531 if (colon)
532 bf[colon - 1] = ':';
533 return r;
534 }
535
536 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
537 {
538 int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
539 return r + evsel__add_modifiers(evsel, bf + r, size - r);
540 }
541
542 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
543 "cpu-clock",
544 "task-clock",
545 "page-faults",
546 "context-switches",
547 "cpu-migrations",
548 "minor-faults",
549 "major-faults",
550 "alignment-faults",
551 "emulation-faults",
552 "dummy",
553 };
554
555 static const char *__evsel__sw_name(u64 config)
556 {
557 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
558 return evsel__sw_names[config];
559 return "unknown-software";
560 }
561
562 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
563 {
564 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
565 return r + evsel__add_modifiers(evsel, bf + r, size - r);
566 }
567
568 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
569 {
570 int r;
571
572 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
573
574 if (type & HW_BREAKPOINT_R)
575 r += scnprintf(bf + r, size - r, "r");
576
577 if (type & HW_BREAKPOINT_W)
578 r += scnprintf(bf + r, size - r, "w");
579
580 if (type & HW_BREAKPOINT_X)
581 r += scnprintf(bf + r, size - r, "x");
582
583 return r;
584 }
585
586 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
587 {
588 struct perf_event_attr *attr = &evsel->core.attr;
589 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
590 return r + evsel__add_modifiers(evsel, bf + r, size - r);
591 }
592
593 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
594 { "L1-dcache", "l1-d", "l1d", "L1-data", },
595 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
596 { "LLC", "L2", },
597 { "dTLB", "d-tlb", "Data-TLB", },
598 { "iTLB", "i-tlb", "Instruction-TLB", },
599 { "branch", "branches", "bpu", "btb", "bpc", },
600 { "node", },
601 };
602
603 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
604 { "load", "loads", "read", },
605 { "store", "stores", "write", },
606 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
607 };
608
609 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
610 { "refs", "Reference", "ops", "access", },
611 { "misses", "miss", },
612 };
613
614 #define C(x) PERF_COUNT_HW_CACHE_##x
615 #define CACHE_READ (1 << C(OP_READ))
616 #define CACHE_WRITE (1 << C(OP_WRITE))
617 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
618 #define COP(x) (1 << x)
619
620 /*
621 * cache operartion stat
622 * L1I : Read and prefetch only
623 * ITLB and BPU : Read-only
624 */
625 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
626 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
627 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
628 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
629 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
630 [C(ITLB)] = (CACHE_READ),
631 [C(BPU)] = (CACHE_READ),
632 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
633 };
634
635 bool evsel__is_cache_op_valid(u8 type, u8 op)
636 {
637 if (evsel__hw_cache_stat[type] & COP(op))
638 return true; /* valid */
639 else
640 return false; /* invalid */
641 }
642
643 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
644 {
645 if (result) {
646 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
647 evsel__hw_cache_op[op][0],
648 evsel__hw_cache_result[result][0]);
649 }
650
651 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
652 evsel__hw_cache_op[op][1]);
653 }
654
655 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
656 {
657 u8 op, result, type = (config >> 0) & 0xff;
658 const char *err = "unknown-ext-hardware-cache-type";
659
660 if (type >= PERF_COUNT_HW_CACHE_MAX)
661 goto out_err;
662
663 op = (config >> 8) & 0xff;
664 err = "unknown-ext-hardware-cache-op";
665 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
666 goto out_err;
667
668 result = (config >> 16) & 0xff;
669 err = "unknown-ext-hardware-cache-result";
670 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
671 goto out_err;
672
673 err = "invalid-cache";
674 if (!evsel__is_cache_op_valid(type, op))
675 goto out_err;
676
677 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
678 out_err:
679 return scnprintf(bf, size, "%s", err);
680 }
681
682 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
683 {
684 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
685 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
686 }
687
688 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
689 {
690 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
691 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
692 }
693
694 static int evsel__tool_name(char *bf, size_t size)
695 {
696 int ret = scnprintf(bf, size, "duration_time");
697 return ret;
698 }
699
700 const char *evsel__name(struct evsel *evsel)
701 {
702 char bf[128];
703
704 if (!evsel)
705 goto out_unknown;
706
707 if (evsel->name)
708 return evsel->name;
709
710 switch (evsel->core.attr.type) {
711 case PERF_TYPE_RAW:
712 evsel__raw_name(evsel, bf, sizeof(bf));
713 break;
714
715 case PERF_TYPE_HARDWARE:
716 evsel__hw_name(evsel, bf, sizeof(bf));
717 break;
718
719 case PERF_TYPE_HW_CACHE:
720 evsel__hw_cache_name(evsel, bf, sizeof(bf));
721 break;
722
723 case PERF_TYPE_SOFTWARE:
724 if (evsel->tool_event)
725 evsel__tool_name(bf, sizeof(bf));
726 else
727 evsel__sw_name(evsel, bf, sizeof(bf));
728 break;
729
730 case PERF_TYPE_TRACEPOINT:
731 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
732 break;
733
734 case PERF_TYPE_BREAKPOINT:
735 evsel__bp_name(evsel, bf, sizeof(bf));
736 break;
737
738 default:
739 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
740 evsel->core.attr.type);
741 break;
742 }
743
744 evsel->name = strdup(bf);
745
746 if (evsel->name)
747 return evsel->name;
748 out_unknown:
749 return "unknown";
750 }
751
752 const char *evsel__group_name(struct evsel *evsel)
753 {
754 return evsel->group_name ?: "anon group";
755 }
756
757 /*
758 * Returns the group details for the specified leader,
759 * with following rules.
760 *
761 * For record -e '{cycles,instructions}'
762 * 'anon group { cycles:u, instructions:u }'
763 *
764 * For record -e 'cycles,instructions' and report --group
765 * 'cycles:u, instructions:u'
766 */
767 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
768 {
769 int ret = 0;
770 struct evsel *pos;
771 const char *group_name = evsel__group_name(evsel);
772
773 if (!evsel->forced_leader)
774 ret = scnprintf(buf, size, "%s { ", group_name);
775
776 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
777
778 for_each_group_member(pos, evsel)
779 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
780
781 if (!evsel->forced_leader)
782 ret += scnprintf(buf + ret, size - ret, " }");
783
784 return ret;
785 }
786
787 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
788 struct callchain_param *param)
789 {
790 bool function = evsel__is_function_event(evsel);
791 struct perf_event_attr *attr = &evsel->core.attr;
792
793 evsel__set_sample_bit(evsel, CALLCHAIN);
794
795 attr->sample_max_stack = param->max_stack;
796
797 if (opts->kernel_callchains)
798 attr->exclude_callchain_user = 1;
799 if (opts->user_callchains)
800 attr->exclude_callchain_kernel = 1;
801 if (param->record_mode == CALLCHAIN_LBR) {
802 if (!opts->branch_stack) {
803 if (attr->exclude_user) {
804 pr_warning("LBR callstack option is only available "
805 "to get user callchain information. "
806 "Falling back to framepointers.\n");
807 } else {
808 evsel__set_sample_bit(evsel, BRANCH_STACK);
809 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
810 PERF_SAMPLE_BRANCH_CALL_STACK |
811 PERF_SAMPLE_BRANCH_NO_CYCLES |
812 PERF_SAMPLE_BRANCH_NO_FLAGS |
813 PERF_SAMPLE_BRANCH_HW_INDEX;
814 }
815 } else
816 pr_warning("Cannot use LBR callstack with branch stack. "
817 "Falling back to framepointers.\n");
818 }
819
820 if (param->record_mode == CALLCHAIN_DWARF) {
821 if (!function) {
822 evsel__set_sample_bit(evsel, REGS_USER);
823 evsel__set_sample_bit(evsel, STACK_USER);
824 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
825 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
826 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
827 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
828 "so the minimal registers set (IP, SP) is explicitly forced.\n");
829 } else {
830 attr->sample_regs_user |= PERF_REGS_MASK;
831 }
832 attr->sample_stack_user = param->dump_size;
833 attr->exclude_callchain_user = 1;
834 } else {
835 pr_info("Cannot use DWARF unwind for function trace event,"
836 " falling back to framepointers.\n");
837 }
838 }
839
840 if (function) {
841 pr_info("Disabling user space callchains for function trace event.\n");
842 attr->exclude_callchain_user = 1;
843 }
844 }
845
846 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
847 struct callchain_param *param)
848 {
849 if (param->enabled)
850 return __evsel__config_callchain(evsel, opts, param);
851 }
852
853 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
854 {
855 struct perf_event_attr *attr = &evsel->core.attr;
856
857 evsel__reset_sample_bit(evsel, CALLCHAIN);
858 if (param->record_mode == CALLCHAIN_LBR) {
859 evsel__reset_sample_bit(evsel, BRANCH_STACK);
860 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
861 PERF_SAMPLE_BRANCH_CALL_STACK |
862 PERF_SAMPLE_BRANCH_HW_INDEX);
863 }
864 if (param->record_mode == CALLCHAIN_DWARF) {
865 evsel__reset_sample_bit(evsel, REGS_USER);
866 evsel__reset_sample_bit(evsel, STACK_USER);
867 }
868 }
869
870 static void evsel__apply_config_terms(struct evsel *evsel,
871 struct record_opts *opts, bool track)
872 {
873 struct evsel_config_term *term;
874 struct list_head *config_terms = &evsel->config_terms;
875 struct perf_event_attr *attr = &evsel->core.attr;
876 /* callgraph default */
877 struct callchain_param param = {
878 .record_mode = callchain_param.record_mode,
879 };
880 u32 dump_size = 0;
881 int max_stack = 0;
882 const char *callgraph_buf = NULL;
883
884 list_for_each_entry(term, config_terms, list) {
885 switch (term->type) {
886 case EVSEL__CONFIG_TERM_PERIOD:
887 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
888 attr->sample_period = term->val.period;
889 attr->freq = 0;
890 evsel__reset_sample_bit(evsel, PERIOD);
891 }
892 break;
893 case EVSEL__CONFIG_TERM_FREQ:
894 if (!(term->weak && opts->user_freq != UINT_MAX)) {
895 attr->sample_freq = term->val.freq;
896 attr->freq = 1;
897 evsel__set_sample_bit(evsel, PERIOD);
898 }
899 break;
900 case EVSEL__CONFIG_TERM_TIME:
901 if (term->val.time)
902 evsel__set_sample_bit(evsel, TIME);
903 else
904 evsel__reset_sample_bit(evsel, TIME);
905 break;
906 case EVSEL__CONFIG_TERM_CALLGRAPH:
907 callgraph_buf = term->val.str;
908 break;
909 case EVSEL__CONFIG_TERM_BRANCH:
910 if (term->val.str && strcmp(term->val.str, "no")) {
911 evsel__set_sample_bit(evsel, BRANCH_STACK);
912 parse_branch_str(term->val.str,
913 &attr->branch_sample_type);
914 } else
915 evsel__reset_sample_bit(evsel, BRANCH_STACK);
916 break;
917 case EVSEL__CONFIG_TERM_STACK_USER:
918 dump_size = term->val.stack_user;
919 break;
920 case EVSEL__CONFIG_TERM_MAX_STACK:
921 max_stack = term->val.max_stack;
922 break;
923 case EVSEL__CONFIG_TERM_MAX_EVENTS:
924 evsel->max_events = term->val.max_events;
925 break;
926 case EVSEL__CONFIG_TERM_INHERIT:
927 /*
928 * attr->inherit should has already been set by
929 * evsel__config. If user explicitly set
930 * inherit using config terms, override global
931 * opt->no_inherit setting.
932 */
933 attr->inherit = term->val.inherit ? 1 : 0;
934 break;
935 case EVSEL__CONFIG_TERM_OVERWRITE:
936 attr->write_backward = term->val.overwrite ? 1 : 0;
937 break;
938 case EVSEL__CONFIG_TERM_DRV_CFG:
939 break;
940 case EVSEL__CONFIG_TERM_PERCORE:
941 break;
942 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
943 attr->aux_output = term->val.aux_output ? 1 : 0;
944 break;
945 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
946 /* Already applied by auxtrace */
947 break;
948 case EVSEL__CONFIG_TERM_CFG_CHG:
949 break;
950 default:
951 break;
952 }
953 }
954
955 /* User explicitly set per-event callgraph, clear the old setting and reset. */
956 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
957 bool sample_address = false;
958
959 if (max_stack) {
960 param.max_stack = max_stack;
961 if (callgraph_buf == NULL)
962 callgraph_buf = "fp";
963 }
964
965 /* parse callgraph parameters */
966 if (callgraph_buf != NULL) {
967 if (!strcmp(callgraph_buf, "no")) {
968 param.enabled = false;
969 param.record_mode = CALLCHAIN_NONE;
970 } else {
971 param.enabled = true;
972 if (parse_callchain_record(callgraph_buf, &param)) {
973 pr_err("per-event callgraph setting for %s failed. "
974 "Apply callgraph global setting for it\n",
975 evsel->name);
976 return;
977 }
978 if (param.record_mode == CALLCHAIN_DWARF)
979 sample_address = true;
980 }
981 }
982 if (dump_size > 0) {
983 dump_size = round_up(dump_size, sizeof(u64));
984 param.dump_size = dump_size;
985 }
986
987 /* If global callgraph set, clear it */
988 if (callchain_param.enabled)
989 evsel__reset_callgraph(evsel, &callchain_param);
990
991 /* set perf-event callgraph */
992 if (param.enabled) {
993 if (sample_address) {
994 evsel__set_sample_bit(evsel, ADDR);
995 evsel__set_sample_bit(evsel, DATA_SRC);
996 evsel->core.attr.mmap_data = track;
997 }
998 evsel__config_callchain(evsel, opts, &param);
999 }
1000 }
1001 }
1002
1003 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1004 {
1005 struct evsel_config_term *term, *found_term = NULL;
1006
1007 list_for_each_entry(term, &evsel->config_terms, list) {
1008 if (term->type == type)
1009 found_term = term;
1010 }
1011
1012 return found_term;
1013 }
1014
1015 /*
1016 * The enable_on_exec/disabled value strategy:
1017 *
1018 * 1) For any type of traced program:
1019 * - all independent events and group leaders are disabled
1020 * - all group members are enabled
1021 *
1022 * Group members are ruled by group leaders. They need to
1023 * be enabled, because the group scheduling relies on that.
1024 *
1025 * 2) For traced programs executed by perf:
1026 * - all independent events and group leaders have
1027 * enable_on_exec set
1028 * - we don't specifically enable or disable any event during
1029 * the record command
1030 *
1031 * Independent events and group leaders are initially disabled
1032 * and get enabled by exec. Group members are ruled by group
1033 * leaders as stated in 1).
1034 *
1035 * 3) For traced programs attached by perf (pid/tid):
1036 * - we specifically enable or disable all events during
1037 * the record command
1038 *
1039 * When attaching events to already running traced we
1040 * enable/disable events specifically, as there's no
1041 * initial traced exec call.
1042 */
1043 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1044 struct callchain_param *callchain)
1045 {
1046 struct evsel *leader = evsel->leader;
1047 struct perf_event_attr *attr = &evsel->core.attr;
1048 int track = evsel->tracking;
1049 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1050
1051 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1052 attr->inherit = !opts->no_inherit;
1053 attr->write_backward = opts->overwrite ? 1 : 0;
1054
1055 evsel__set_sample_bit(evsel, IP);
1056 evsel__set_sample_bit(evsel, TID);
1057
1058 if (evsel->sample_read) {
1059 evsel__set_sample_bit(evsel, READ);
1060
1061 /*
1062 * We need ID even in case of single event, because
1063 * PERF_SAMPLE_READ process ID specific data.
1064 */
1065 evsel__set_sample_id(evsel, false);
1066
1067 /*
1068 * Apply group format only if we belong to group
1069 * with more than one members.
1070 */
1071 if (leader->core.nr_members > 1) {
1072 attr->read_format |= PERF_FORMAT_GROUP;
1073 attr->inherit = 0;
1074 }
1075 }
1076
1077 /*
1078 * We default some events to have a default interval. But keep
1079 * it a weak assumption overridable by the user.
1080 */
1081 if (!attr->sample_period) {
1082 if (opts->freq) {
1083 attr->freq = 1;
1084 attr->sample_freq = opts->freq;
1085 } else {
1086 attr->sample_period = opts->default_interval;
1087 }
1088 }
1089 /*
1090 * If attr->freq was set (here or earlier), ask for period
1091 * to be sampled.
1092 */
1093 if (attr->freq)
1094 evsel__set_sample_bit(evsel, PERIOD);
1095
1096 if (opts->no_samples)
1097 attr->sample_freq = 0;
1098
1099 if (opts->inherit_stat) {
1100 evsel->core.attr.read_format |=
1101 PERF_FORMAT_TOTAL_TIME_ENABLED |
1102 PERF_FORMAT_TOTAL_TIME_RUNNING |
1103 PERF_FORMAT_ID;
1104 attr->inherit_stat = 1;
1105 }
1106
1107 if (opts->sample_address) {
1108 evsel__set_sample_bit(evsel, ADDR);
1109 attr->mmap_data = track;
1110 }
1111
1112 /*
1113 * We don't allow user space callchains for function trace
1114 * event, due to issues with page faults while tracing page
1115 * fault handler and its overall trickiness nature.
1116 */
1117 if (evsel__is_function_event(evsel))
1118 evsel->core.attr.exclude_callchain_user = 1;
1119
1120 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1121 evsel__config_callchain(evsel, opts, callchain);
1122
1123 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1124 !evsel__is_dummy_event(evsel)) {
1125 attr->sample_regs_intr = opts->sample_intr_regs;
1126 evsel__set_sample_bit(evsel, REGS_INTR);
1127 }
1128
1129 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1130 !evsel__is_dummy_event(evsel)) {
1131 attr->sample_regs_user |= opts->sample_user_regs;
1132 evsel__set_sample_bit(evsel, REGS_USER);
1133 }
1134
1135 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1136 evsel__set_sample_bit(evsel, CPU);
1137
1138 /*
1139 * When the user explicitly disabled time don't force it here.
1140 */
1141 if (opts->sample_time &&
1142 (!perf_missing_features.sample_id_all &&
1143 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1144 opts->sample_time_set)))
1145 evsel__set_sample_bit(evsel, TIME);
1146
1147 if (opts->raw_samples && !evsel->no_aux_samples) {
1148 evsel__set_sample_bit(evsel, TIME);
1149 evsel__set_sample_bit(evsel, RAW);
1150 evsel__set_sample_bit(evsel, CPU);
1151 }
1152
1153 if (opts->sample_address)
1154 evsel__set_sample_bit(evsel, DATA_SRC);
1155
1156 if (opts->sample_phys_addr)
1157 evsel__set_sample_bit(evsel, PHYS_ADDR);
1158
1159 if (opts->no_buffering) {
1160 attr->watermark = 0;
1161 attr->wakeup_events = 1;
1162 }
1163 if (opts->branch_stack && !evsel->no_aux_samples) {
1164 evsel__set_sample_bit(evsel, BRANCH_STACK);
1165 attr->branch_sample_type = opts->branch_stack;
1166 }
1167
1168 if (opts->sample_weight)
1169 evsel__set_sample_bit(evsel, WEIGHT);
1170
1171 attr->task = track;
1172 attr->mmap = track;
1173 attr->mmap2 = track && !perf_missing_features.mmap2;
1174 attr->comm = track;
1175 /*
1176 * ksymbol is tracked separately with text poke because it needs to be
1177 * system wide and enabled immediately.
1178 */
1179 if (!opts->text_poke)
1180 attr->ksymbol = track && !perf_missing_features.ksymbol;
1181 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1182
1183 if (opts->record_namespaces)
1184 attr->namespaces = track;
1185
1186 if (opts->record_cgroup) {
1187 attr->cgroup = track && !perf_missing_features.cgroup;
1188 evsel__set_sample_bit(evsel, CGROUP);
1189 }
1190
1191 if (opts->sample_data_page_size)
1192 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1193
1194 if (opts->record_switch_events)
1195 attr->context_switch = track;
1196
1197 if (opts->sample_transaction)
1198 evsel__set_sample_bit(evsel, TRANSACTION);
1199
1200 if (opts->running_time) {
1201 evsel->core.attr.read_format |=
1202 PERF_FORMAT_TOTAL_TIME_ENABLED |
1203 PERF_FORMAT_TOTAL_TIME_RUNNING;
1204 }
1205
1206 /*
1207 * XXX see the function comment above
1208 *
1209 * Disabling only independent events or group leaders,
1210 * keeping group members enabled.
1211 */
1212 if (evsel__is_group_leader(evsel))
1213 attr->disabled = 1;
1214
1215 /*
1216 * Setting enable_on_exec for independent events and
1217 * group leaders for traced executed by perf.
1218 */
1219 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1220 !opts->initial_delay)
1221 attr->enable_on_exec = 1;
1222
1223 if (evsel->immediate) {
1224 attr->disabled = 0;
1225 attr->enable_on_exec = 0;
1226 }
1227
1228 clockid = opts->clockid;
1229 if (opts->use_clockid) {
1230 attr->use_clockid = 1;
1231 attr->clockid = opts->clockid;
1232 }
1233
1234 if (evsel->precise_max)
1235 attr->precise_ip = 3;
1236
1237 if (opts->all_user) {
1238 attr->exclude_kernel = 1;
1239 attr->exclude_user = 0;
1240 }
1241
1242 if (opts->all_kernel) {
1243 attr->exclude_kernel = 0;
1244 attr->exclude_user = 1;
1245 }
1246
1247 if (evsel->core.own_cpus || evsel->unit)
1248 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1249
1250 /*
1251 * Apply event specific term settings,
1252 * it overloads any global configuration.
1253 */
1254 evsel__apply_config_terms(evsel, opts, track);
1255
1256 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1257
1258 /* The --period option takes the precedence. */
1259 if (opts->period_set) {
1260 if (opts->period)
1261 evsel__set_sample_bit(evsel, PERIOD);
1262 else
1263 evsel__reset_sample_bit(evsel, PERIOD);
1264 }
1265
1266 /*
1267 * A dummy event never triggers any actual counter and therefore
1268 * cannot be used with branch_stack.
1269 *
1270 * For initial_delay, a dummy event is added implicitly.
1271 * The software event will trigger -EOPNOTSUPP error out,
1272 * if BRANCH_STACK bit is set.
1273 */
1274 if (evsel__is_dummy_event(evsel))
1275 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1276 }
1277
1278 int evsel__set_filter(struct evsel *evsel, const char *filter)
1279 {
1280 char *new_filter = strdup(filter);
1281
1282 if (new_filter != NULL) {
1283 free(evsel->filter);
1284 evsel->filter = new_filter;
1285 return 0;
1286 }
1287
1288 return -1;
1289 }
1290
1291 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1292 {
1293 char *new_filter;
1294
1295 if (evsel->filter == NULL)
1296 return evsel__set_filter(evsel, filter);
1297
1298 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1299 free(evsel->filter);
1300 evsel->filter = new_filter;
1301 return 0;
1302 }
1303
1304 return -1;
1305 }
1306
1307 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1308 {
1309 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1310 }
1311
1312 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1313 {
1314 return evsel__append_filter(evsel, "%s,%s", filter);
1315 }
1316
1317 /* Caller has to clear disabled after going through all CPUs. */
1318 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1319 {
1320 return perf_evsel__enable_cpu(&evsel->core, cpu);
1321 }
1322
1323 int evsel__enable(struct evsel *evsel)
1324 {
1325 int err = perf_evsel__enable(&evsel->core);
1326
1327 if (!err)
1328 evsel->disabled = false;
1329 return err;
1330 }
1331
1332 /* Caller has to set disabled after going through all CPUs. */
1333 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1334 {
1335 return perf_evsel__disable_cpu(&evsel->core, cpu);
1336 }
1337
1338 int evsel__disable(struct evsel *evsel)
1339 {
1340 int err = perf_evsel__disable(&evsel->core);
1341 /*
1342 * We mark it disabled here so that tools that disable a event can
1343 * ignore events after they disable it. I.e. the ring buffer may have
1344 * already a few more events queued up before the kernel got the stop
1345 * request.
1346 */
1347 if (!err)
1348 evsel->disabled = true;
1349
1350 return err;
1351 }
1352
1353 static void evsel__free_config_terms(struct evsel *evsel)
1354 {
1355 struct evsel_config_term *term, *h;
1356
1357 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1358 list_del_init(&term->list);
1359 if (term->free_str)
1360 zfree(&term->val.str);
1361 free(term);
1362 }
1363 }
1364
1365 void evsel__exit(struct evsel *evsel)
1366 {
1367 assert(list_empty(&evsel->core.node));
1368 assert(evsel->evlist == NULL);
1369 evsel__free_counts(evsel);
1370 perf_evsel__free_fd(&evsel->core);
1371 perf_evsel__free_id(&evsel->core);
1372 evsel__free_config_terms(evsel);
1373 cgroup__put(evsel->cgrp);
1374 perf_cpu_map__put(evsel->core.cpus);
1375 perf_cpu_map__put(evsel->core.own_cpus);
1376 perf_thread_map__put(evsel->core.threads);
1377 zfree(&evsel->group_name);
1378 zfree(&evsel->name);
1379 zfree(&evsel->pmu_name);
1380 zfree(&evsel->per_pkg_mask);
1381 zfree(&evsel->metric_events);
1382 perf_evsel__object.fini(evsel);
1383 }
1384
1385 void evsel__delete(struct evsel *evsel)
1386 {
1387 evsel__exit(evsel);
1388 free(evsel);
1389 }
1390
1391 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1392 struct perf_counts_values *count)
1393 {
1394 struct perf_counts_values tmp;
1395
1396 if (!evsel->prev_raw_counts)
1397 return;
1398
1399 if (cpu == -1) {
1400 tmp = evsel->prev_raw_counts->aggr;
1401 evsel->prev_raw_counts->aggr = *count;
1402 } else {
1403 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1404 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1405 }
1406
1407 count->val = count->val - tmp.val;
1408 count->ena = count->ena - tmp.ena;
1409 count->run = count->run - tmp.run;
1410 }
1411
1412 void perf_counts_values__scale(struct perf_counts_values *count,
1413 bool scale, s8 *pscaled)
1414 {
1415 s8 scaled = 0;
1416
1417 if (scale) {
1418 if (count->run == 0) {
1419 scaled = -1;
1420 count->val = 0;
1421 } else if (count->run < count->ena) {
1422 scaled = 1;
1423 count->val = (u64)((double) count->val * count->ena / count->run);
1424 }
1425 }
1426
1427 if (pscaled)
1428 *pscaled = scaled;
1429 }
1430
1431 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1432 {
1433 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1434
1435 return perf_evsel__read(&evsel->core, cpu, thread, count);
1436 }
1437
1438 static void evsel__set_count(struct evsel *counter, int cpu, int thread, u64 val, u64 ena, u64 run)
1439 {
1440 struct perf_counts_values *count;
1441
1442 count = perf_counts(counter->counts, cpu, thread);
1443
1444 count->val = val;
1445 count->ena = ena;
1446 count->run = run;
1447
1448 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1449 }
1450
1451 static int evsel__process_group_data(struct evsel *leader, int cpu, int thread, u64 *data)
1452 {
1453 u64 read_format = leader->core.attr.read_format;
1454 struct sample_read_value *v;
1455 u64 nr, ena = 0, run = 0, i;
1456
1457 nr = *data++;
1458
1459 if (nr != (u64) leader->core.nr_members)
1460 return -EINVAL;
1461
1462 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1463 ena = *data++;
1464
1465 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1466 run = *data++;
1467
1468 v = (struct sample_read_value *) data;
1469
1470 evsel__set_count(leader, cpu, thread, v[0].value, ena, run);
1471
1472 for (i = 1; i < nr; i++) {
1473 struct evsel *counter;
1474
1475 counter = evlist__id2evsel(leader->evlist, v[i].id);
1476 if (!counter)
1477 return -EINVAL;
1478
1479 evsel__set_count(counter, cpu, thread, v[i].value, ena, run);
1480 }
1481
1482 return 0;
1483 }
1484
1485 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1486 {
1487 struct perf_stat_evsel *ps = leader->stats;
1488 u64 read_format = leader->core.attr.read_format;
1489 int size = perf_evsel__read_size(&leader->core);
1490 u64 *data = ps->group_data;
1491
1492 if (!(read_format & PERF_FORMAT_ID))
1493 return -EINVAL;
1494
1495 if (!evsel__is_group_leader(leader))
1496 return -EINVAL;
1497
1498 if (!data) {
1499 data = zalloc(size);
1500 if (!data)
1501 return -ENOMEM;
1502
1503 ps->group_data = data;
1504 }
1505
1506 if (FD(leader, cpu, thread) < 0)
1507 return -EINVAL;
1508
1509 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1510 return -errno;
1511
1512 return evsel__process_group_data(leader, cpu, thread, data);
1513 }
1514
1515 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1516 {
1517 u64 read_format = evsel->core.attr.read_format;
1518
1519 if (read_format & PERF_FORMAT_GROUP)
1520 return evsel__read_group(evsel, cpu, thread);
1521
1522 return evsel__read_one(evsel, cpu, thread);
1523 }
1524
1525 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1526 {
1527 struct perf_counts_values count;
1528 size_t nv = scale ? 3 : 1;
1529
1530 if (FD(evsel, cpu, thread) < 0)
1531 return -EINVAL;
1532
1533 if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1534 return -ENOMEM;
1535
1536 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1537 return -errno;
1538
1539 evsel__compute_deltas(evsel, cpu, thread, &count);
1540 perf_counts_values__scale(&count, scale, NULL);
1541 *perf_counts(evsel->counts, cpu, thread) = count;
1542 return 0;
1543 }
1544
1545 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1546 {
1547 struct evsel *leader = evsel->leader;
1548 int fd;
1549
1550 if (evsel__is_group_leader(evsel))
1551 return -1;
1552
1553 /*
1554 * Leader must be already processed/open,
1555 * if not it's a bug.
1556 */
1557 BUG_ON(!leader->core.fd);
1558
1559 fd = FD(leader, cpu, thread);
1560 BUG_ON(fd == -1);
1561
1562 return fd;
1563 }
1564
1565 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1566 {
1567 for (int cpu = 0; cpu < nr_cpus; cpu++)
1568 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1569 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1570 }
1571
1572 static int update_fds(struct evsel *evsel,
1573 int nr_cpus, int cpu_idx,
1574 int nr_threads, int thread_idx)
1575 {
1576 struct evsel *pos;
1577
1578 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1579 return -EINVAL;
1580
1581 evlist__for_each_entry(evsel->evlist, pos) {
1582 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1583
1584 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1585
1586 /*
1587 * Since fds for next evsel has not been created,
1588 * there is no need to iterate whole event list.
1589 */
1590 if (pos == evsel)
1591 break;
1592 }
1593 return 0;
1594 }
1595
1596 static bool ignore_missing_thread(struct evsel *evsel,
1597 int nr_cpus, int cpu,
1598 struct perf_thread_map *threads,
1599 int thread, int err)
1600 {
1601 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1602
1603 if (!evsel->ignore_missing_thread)
1604 return false;
1605
1606 /* The system wide setup does not work with threads. */
1607 if (evsel->core.system_wide)
1608 return false;
1609
1610 /* The -ESRCH is perf event syscall errno for pid's not found. */
1611 if (err != -ESRCH)
1612 return false;
1613
1614 /* If there's only one thread, let it fail. */
1615 if (threads->nr == 1)
1616 return false;
1617
1618 /*
1619 * We should remove fd for missing_thread first
1620 * because thread_map__remove() will decrease threads->nr.
1621 */
1622 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1623 return false;
1624
1625 if (thread_map__remove(threads, thread))
1626 return false;
1627
1628 pr_warning("WARNING: Ignored open failure for pid %d\n",
1629 ignore_pid);
1630 return true;
1631 }
1632
1633 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1634 void *priv __maybe_unused)
1635 {
1636 return fprintf(fp, " %-32s %s\n", name, val);
1637 }
1638
1639 static void display_attr(struct perf_event_attr *attr)
1640 {
1641 if (verbose >= 2 || debug_peo_args) {
1642 fprintf(stderr, "%.60s\n", graph_dotted_line);
1643 fprintf(stderr, "perf_event_attr:\n");
1644 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1645 fprintf(stderr, "%.60s\n", graph_dotted_line);
1646 }
1647 }
1648
1649 static int perf_event_open(struct evsel *evsel,
1650 pid_t pid, int cpu, int group_fd,
1651 unsigned long flags)
1652 {
1653 int precise_ip = evsel->core.attr.precise_ip;
1654 int fd;
1655
1656 while (1) {
1657 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1658 pid, cpu, group_fd, flags);
1659
1660 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1661 if (fd >= 0)
1662 break;
1663
1664 /* Do not try less precise if not requested. */
1665 if (!evsel->precise_max)
1666 break;
1667
1668 /*
1669 * We tried all the precise_ip values, and it's
1670 * still failing, so leave it to standard fallback.
1671 */
1672 if (!evsel->core.attr.precise_ip) {
1673 evsel->core.attr.precise_ip = precise_ip;
1674 break;
1675 }
1676
1677 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1678 evsel->core.attr.precise_ip--;
1679 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1680 display_attr(&evsel->core.attr);
1681 }
1682
1683 return fd;
1684 }
1685
1686 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1687 struct perf_thread_map *threads,
1688 int start_cpu, int end_cpu)
1689 {
1690 int cpu, thread, nthreads;
1691 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1692 int pid = -1, err, old_errno;
1693 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1694
1695 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1696 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1697 return -EINVAL;
1698
1699 if (cpus == NULL) {
1700 static struct perf_cpu_map *empty_cpu_map;
1701
1702 if (empty_cpu_map == NULL) {
1703 empty_cpu_map = perf_cpu_map__dummy_new();
1704 if (empty_cpu_map == NULL)
1705 return -ENOMEM;
1706 }
1707
1708 cpus = empty_cpu_map;
1709 }
1710
1711 if (threads == NULL) {
1712 static struct perf_thread_map *empty_thread_map;
1713
1714 if (empty_thread_map == NULL) {
1715 empty_thread_map = thread_map__new_by_tid(-1);
1716 if (empty_thread_map == NULL)
1717 return -ENOMEM;
1718 }
1719
1720 threads = empty_thread_map;
1721 }
1722
1723 if (evsel->core.system_wide)
1724 nthreads = 1;
1725 else
1726 nthreads = threads->nr;
1727
1728 if (evsel->core.fd == NULL &&
1729 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1730 return -ENOMEM;
1731
1732 if (evsel->cgrp) {
1733 flags |= PERF_FLAG_PID_CGROUP;
1734 pid = evsel->cgrp->fd;
1735 }
1736
1737 fallback_missing_features:
1738 if (perf_missing_features.clockid_wrong)
1739 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1740 if (perf_missing_features.clockid) {
1741 evsel->core.attr.use_clockid = 0;
1742 evsel->core.attr.clockid = 0;
1743 }
1744 if (perf_missing_features.cloexec)
1745 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1746 if (perf_missing_features.mmap2)
1747 evsel->core.attr.mmap2 = 0;
1748 if (perf_missing_features.exclude_guest)
1749 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1750 if (perf_missing_features.lbr_flags)
1751 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1752 PERF_SAMPLE_BRANCH_NO_CYCLES);
1753 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1754 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1755 if (perf_missing_features.ksymbol)
1756 evsel->core.attr.ksymbol = 0;
1757 if (perf_missing_features.bpf)
1758 evsel->core.attr.bpf_event = 0;
1759 if (perf_missing_features.branch_hw_idx)
1760 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1761 retry_sample_id:
1762 if (perf_missing_features.sample_id_all)
1763 evsel->core.attr.sample_id_all = 0;
1764
1765 display_attr(&evsel->core.attr);
1766
1767 for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1768
1769 for (thread = 0; thread < nthreads; thread++) {
1770 int fd, group_fd;
1771
1772 if (!evsel->cgrp && !evsel->core.system_wide)
1773 pid = perf_thread_map__pid(threads, thread);
1774
1775 group_fd = get_group_fd(evsel, cpu, thread);
1776 retry_open:
1777 test_attr__ready();
1778
1779 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1780 group_fd, flags);
1781
1782 FD(evsel, cpu, thread) = fd;
1783
1784 if (unlikely(test_attr__enabled)) {
1785 test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
1786 fd, group_fd, flags);
1787 }
1788
1789 if (fd < 0) {
1790 err = -errno;
1791
1792 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1793 /*
1794 * We just removed 1 thread, so take a step
1795 * back on thread index and lower the upper
1796 * nthreads limit.
1797 */
1798 nthreads--;
1799 thread--;
1800
1801 /* ... and pretend like nothing have happened. */
1802 err = 0;
1803 continue;
1804 }
1805
1806 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1807 err);
1808 goto try_fallback;
1809 }
1810
1811 pr_debug2_peo(" = %d\n", fd);
1812
1813 if (evsel->bpf_fd >= 0) {
1814 int evt_fd = fd;
1815 int bpf_fd = evsel->bpf_fd;
1816
1817 err = ioctl(evt_fd,
1818 PERF_EVENT_IOC_SET_BPF,
1819 bpf_fd);
1820 if (err && errno != EEXIST) {
1821 pr_err("failed to attach bpf fd %d: %s\n",
1822 bpf_fd, strerror(errno));
1823 err = -EINVAL;
1824 goto out_close;
1825 }
1826 }
1827
1828 set_rlimit = NO_CHANGE;
1829
1830 /*
1831 * If we succeeded but had to kill clockid, fail and
1832 * have evsel__open_strerror() print us a nice error.
1833 */
1834 if (perf_missing_features.clockid ||
1835 perf_missing_features.clockid_wrong) {
1836 err = -EINVAL;
1837 goto out_close;
1838 }
1839 }
1840 }
1841
1842 return 0;
1843
1844 try_fallback:
1845 /*
1846 * perf stat needs between 5 and 22 fds per CPU. When we run out
1847 * of them try to increase the limits.
1848 */
1849 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1850 struct rlimit l;
1851
1852 old_errno = errno;
1853 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1854 if (set_rlimit == NO_CHANGE)
1855 l.rlim_cur = l.rlim_max;
1856 else {
1857 l.rlim_cur = l.rlim_max + 1000;
1858 l.rlim_max = l.rlim_cur;
1859 }
1860 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1861 set_rlimit++;
1862 errno = old_errno;
1863 goto retry_open;
1864 }
1865 }
1866 errno = old_errno;
1867 }
1868
1869 if (err != -EINVAL || cpu > 0 || thread > 0)
1870 goto out_close;
1871
1872 /*
1873 * Must probe features in the order they were added to the
1874 * perf_event_attr interface.
1875 */
1876 if (!perf_missing_features.data_page_size &&
1877 (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1878 perf_missing_features.data_page_size = true;
1879 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1880 goto out_close;
1881 } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1882 perf_missing_features.cgroup = true;
1883 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1884 goto out_close;
1885 } else if (!perf_missing_features.branch_hw_idx &&
1886 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1887 perf_missing_features.branch_hw_idx = true;
1888 pr_debug2("switching off branch HW index support\n");
1889 goto fallback_missing_features;
1890 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1891 perf_missing_features.aux_output = true;
1892 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1893 goto out_close;
1894 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1895 perf_missing_features.bpf = true;
1896 pr_debug2_peo("switching off bpf_event\n");
1897 goto fallback_missing_features;
1898 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1899 perf_missing_features.ksymbol = true;
1900 pr_debug2_peo("switching off ksymbol\n");
1901 goto fallback_missing_features;
1902 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1903 perf_missing_features.write_backward = true;
1904 pr_debug2_peo("switching off write_backward\n");
1905 goto out_close;
1906 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1907 perf_missing_features.clockid_wrong = true;
1908 pr_debug2_peo("switching off clockid\n");
1909 goto fallback_missing_features;
1910 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1911 perf_missing_features.clockid = true;
1912 pr_debug2_peo("switching off use_clockid\n");
1913 goto fallback_missing_features;
1914 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1915 perf_missing_features.cloexec = true;
1916 pr_debug2_peo("switching off cloexec flag\n");
1917 goto fallback_missing_features;
1918 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1919 perf_missing_features.mmap2 = true;
1920 pr_debug2_peo("switching off mmap2\n");
1921 goto fallback_missing_features;
1922 } else if (!perf_missing_features.exclude_guest &&
1923 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1924 perf_missing_features.exclude_guest = true;
1925 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1926 goto fallback_missing_features;
1927 } else if (!perf_missing_features.sample_id_all) {
1928 perf_missing_features.sample_id_all = true;
1929 pr_debug2_peo("switching off sample_id_all\n");
1930 goto retry_sample_id;
1931 } else if (!perf_missing_features.lbr_flags &&
1932 (evsel->core.attr.branch_sample_type &
1933 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1934 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1935 perf_missing_features.lbr_flags = true;
1936 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1937 goto fallback_missing_features;
1938 } else if (!perf_missing_features.group_read &&
1939 evsel->core.attr.inherit &&
1940 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1941 evsel__is_group_leader(evsel)) {
1942 perf_missing_features.group_read = true;
1943 pr_debug2_peo("switching off group read\n");
1944 goto fallback_missing_features;
1945 }
1946 out_close:
1947 if (err)
1948 threads->err_thread = thread;
1949
1950 old_errno = errno;
1951 do {
1952 while (--thread >= 0) {
1953 if (FD(evsel, cpu, thread) >= 0)
1954 close(FD(evsel, cpu, thread));
1955 FD(evsel, cpu, thread) = -1;
1956 }
1957 thread = nthreads;
1958 } while (--cpu >= 0);
1959 errno = old_errno;
1960 return err;
1961 }
1962
1963 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1964 struct perf_thread_map *threads)
1965 {
1966 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1967 }
1968
1969 void evsel__close(struct evsel *evsel)
1970 {
1971 perf_evsel__close(&evsel->core);
1972 perf_evsel__free_id(&evsel->core);
1973 }
1974
1975 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
1976 {
1977 if (cpu == -1)
1978 return evsel__open_cpu(evsel, cpus, NULL, 0,
1979 cpus ? cpus->nr : 1);
1980
1981 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1982 }
1983
1984 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
1985 {
1986 return evsel__open(evsel, NULL, threads);
1987 }
1988
1989 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1990 const union perf_event *event,
1991 struct perf_sample *sample)
1992 {
1993 u64 type = evsel->core.attr.sample_type;
1994 const __u64 *array = event->sample.array;
1995 bool swapped = evsel->needs_swap;
1996 union u64_swap u;
1997
1998 array += ((event->header.size -
1999 sizeof(event->header)) / sizeof(u64)) - 1;
2000
2001 if (type & PERF_SAMPLE_IDENTIFIER) {
2002 sample->id = *array;
2003 array--;
2004 }
2005
2006 if (type & PERF_SAMPLE_CPU) {
2007 u.val64 = *array;
2008 if (swapped) {
2009 /* undo swap of u64, then swap on individual u32s */
2010 u.val64 = bswap_64(u.val64);
2011 u.val32[0] = bswap_32(u.val32[0]);
2012 }
2013
2014 sample->cpu = u.val32[0];
2015 array--;
2016 }
2017
2018 if (type & PERF_SAMPLE_STREAM_ID) {
2019 sample->stream_id = *array;
2020 array--;
2021 }
2022
2023 if (type & PERF_SAMPLE_ID) {
2024 sample->id = *array;
2025 array--;
2026 }
2027
2028 if (type & PERF_SAMPLE_TIME) {
2029 sample->time = *array;
2030 array--;
2031 }
2032
2033 if (type & PERF_SAMPLE_TID) {
2034 u.val64 = *array;
2035 if (swapped) {
2036 /* undo swap of u64, then swap on individual u32s */
2037 u.val64 = bswap_64(u.val64);
2038 u.val32[0] = bswap_32(u.val32[0]);
2039 u.val32[1] = bswap_32(u.val32[1]);
2040 }
2041
2042 sample->pid = u.val32[0];
2043 sample->tid = u.val32[1];
2044 array--;
2045 }
2046
2047 return 0;
2048 }
2049
2050 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2051 u64 size)
2052 {
2053 return size > max_size || offset + size > endp;
2054 }
2055
2056 #define OVERFLOW_CHECK(offset, size, max_size) \
2057 do { \
2058 if (overflow(endp, (max_size), (offset), (size))) \
2059 return -EFAULT; \
2060 } while (0)
2061
2062 #define OVERFLOW_CHECK_u64(offset) \
2063 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2064
2065 static int
2066 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2067 {
2068 /*
2069 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2070 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2071 * check the format does not go past the end of the event.
2072 */
2073 if (sample_size + sizeof(event->header) > event->header.size)
2074 return -EFAULT;
2075
2076 return 0;
2077 }
2078
2079 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2080 struct perf_sample *data)
2081 {
2082 u64 type = evsel->core.attr.sample_type;
2083 bool swapped = evsel->needs_swap;
2084 const __u64 *array;
2085 u16 max_size = event->header.size;
2086 const void *endp = (void *)event + max_size;
2087 u64 sz;
2088
2089 /*
2090 * used for cross-endian analysis. See git commit 65014ab3
2091 * for why this goofiness is needed.
2092 */
2093 union u64_swap u;
2094
2095 memset(data, 0, sizeof(*data));
2096 data->cpu = data->pid = data->tid = -1;
2097 data->stream_id = data->id = data->time = -1ULL;
2098 data->period = evsel->core.attr.sample_period;
2099 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2100 data->misc = event->header.misc;
2101 data->id = -1ULL;
2102 data->data_src = PERF_MEM_DATA_SRC_NONE;
2103
2104 if (event->header.type != PERF_RECORD_SAMPLE) {
2105 if (!evsel->core.attr.sample_id_all)
2106 return 0;
2107 return perf_evsel__parse_id_sample(evsel, event, data);
2108 }
2109
2110 array = event->sample.array;
2111
2112 if (perf_event__check_size(event, evsel->sample_size))
2113 return -EFAULT;
2114
2115 if (type & PERF_SAMPLE_IDENTIFIER) {
2116 data->id = *array;
2117 array++;
2118 }
2119
2120 if (type & PERF_SAMPLE_IP) {
2121 data->ip = *array;
2122 array++;
2123 }
2124
2125 if (type & PERF_SAMPLE_TID) {
2126 u.val64 = *array;
2127 if (swapped) {
2128 /* undo swap of u64, then swap on individual u32s */
2129 u.val64 = bswap_64(u.val64);
2130 u.val32[0] = bswap_32(u.val32[0]);
2131 u.val32[1] = bswap_32(u.val32[1]);
2132 }
2133
2134 data->pid = u.val32[0];
2135 data->tid = u.val32[1];
2136 array++;
2137 }
2138
2139 if (type & PERF_SAMPLE_TIME) {
2140 data->time = *array;
2141 array++;
2142 }
2143
2144 if (type & PERF_SAMPLE_ADDR) {
2145 data->addr = *array;
2146 array++;
2147 }
2148
2149 if (type & PERF_SAMPLE_ID) {
2150 data->id = *array;
2151 array++;
2152 }
2153
2154 if (type & PERF_SAMPLE_STREAM_ID) {
2155 data->stream_id = *array;
2156 array++;
2157 }
2158
2159 if (type & PERF_SAMPLE_CPU) {
2160
2161 u.val64 = *array;
2162 if (swapped) {
2163 /* undo swap of u64, then swap on individual u32s */
2164 u.val64 = bswap_64(u.val64);
2165 u.val32[0] = bswap_32(u.val32[0]);
2166 }
2167
2168 data->cpu = u.val32[0];
2169 array++;
2170 }
2171
2172 if (type & PERF_SAMPLE_PERIOD) {
2173 data->period = *array;
2174 array++;
2175 }
2176
2177 if (type & PERF_SAMPLE_READ) {
2178 u64 read_format = evsel->core.attr.read_format;
2179
2180 OVERFLOW_CHECK_u64(array);
2181 if (read_format & PERF_FORMAT_GROUP)
2182 data->read.group.nr = *array;
2183 else
2184 data->read.one.value = *array;
2185
2186 array++;
2187
2188 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2189 OVERFLOW_CHECK_u64(array);
2190 data->read.time_enabled = *array;
2191 array++;
2192 }
2193
2194 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2195 OVERFLOW_CHECK_u64(array);
2196 data->read.time_running = *array;
2197 array++;
2198 }
2199
2200 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2201 if (read_format & PERF_FORMAT_GROUP) {
2202 const u64 max_group_nr = UINT64_MAX /
2203 sizeof(struct sample_read_value);
2204
2205 if (data->read.group.nr > max_group_nr)
2206 return -EFAULT;
2207 sz = data->read.group.nr *
2208 sizeof(struct sample_read_value);
2209 OVERFLOW_CHECK(array, sz, max_size);
2210 data->read.group.values =
2211 (struct sample_read_value *)array;
2212 array = (void *)array + sz;
2213 } else {
2214 OVERFLOW_CHECK_u64(array);
2215 data->read.one.id = *array;
2216 array++;
2217 }
2218 }
2219
2220 if (type & PERF_SAMPLE_CALLCHAIN) {
2221 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2222
2223 OVERFLOW_CHECK_u64(array);
2224 data->callchain = (struct ip_callchain *)array++;
2225 if (data->callchain->nr > max_callchain_nr)
2226 return -EFAULT;
2227 sz = data->callchain->nr * sizeof(u64);
2228 OVERFLOW_CHECK(array, sz, max_size);
2229 array = (void *)array + sz;
2230 }
2231
2232 if (type & PERF_SAMPLE_RAW) {
2233 OVERFLOW_CHECK_u64(array);
2234 u.val64 = *array;
2235
2236 /*
2237 * Undo swap of u64, then swap on individual u32s,
2238 * get the size of the raw area and undo all of the
2239 * swap. The pevent interface handles endianity by
2240 * itself.
2241 */
2242 if (swapped) {
2243 u.val64 = bswap_64(u.val64);
2244 u.val32[0] = bswap_32(u.val32[0]);
2245 u.val32[1] = bswap_32(u.val32[1]);
2246 }
2247 data->raw_size = u.val32[0];
2248
2249 /*
2250 * The raw data is aligned on 64bits including the
2251 * u32 size, so it's safe to use mem_bswap_64.
2252 */
2253 if (swapped)
2254 mem_bswap_64((void *) array, data->raw_size);
2255
2256 array = (void *)array + sizeof(u32);
2257
2258 OVERFLOW_CHECK(array, data->raw_size, max_size);
2259 data->raw_data = (void *)array;
2260 array = (void *)array + data->raw_size;
2261 }
2262
2263 if (type & PERF_SAMPLE_BRANCH_STACK) {
2264 const u64 max_branch_nr = UINT64_MAX /
2265 sizeof(struct branch_entry);
2266
2267 OVERFLOW_CHECK_u64(array);
2268 data->branch_stack = (struct branch_stack *)array++;
2269
2270 if (data->branch_stack->nr > max_branch_nr)
2271 return -EFAULT;
2272
2273 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2274 if (evsel__has_branch_hw_idx(evsel))
2275 sz += sizeof(u64);
2276 else
2277 data->no_hw_idx = true;
2278 OVERFLOW_CHECK(array, sz, max_size);
2279 array = (void *)array + sz;
2280 }
2281
2282 if (type & PERF_SAMPLE_REGS_USER) {
2283 OVERFLOW_CHECK_u64(array);
2284 data->user_regs.abi = *array;
2285 array++;
2286
2287 if (data->user_regs.abi) {
2288 u64 mask = evsel->core.attr.sample_regs_user;
2289
2290 sz = hweight64(mask) * sizeof(u64);
2291 OVERFLOW_CHECK(array, sz, max_size);
2292 data->user_regs.mask = mask;
2293 data->user_regs.regs = (u64 *)array;
2294 array = (void *)array + sz;
2295 }
2296 }
2297
2298 if (type & PERF_SAMPLE_STACK_USER) {
2299 OVERFLOW_CHECK_u64(array);
2300 sz = *array++;
2301
2302 data->user_stack.offset = ((char *)(array - 1)
2303 - (char *) event);
2304
2305 if (!sz) {
2306 data->user_stack.size = 0;
2307 } else {
2308 OVERFLOW_CHECK(array, sz, max_size);
2309 data->user_stack.data = (char *)array;
2310 array = (void *)array + sz;
2311 OVERFLOW_CHECK_u64(array);
2312 data->user_stack.size = *array++;
2313 if (WARN_ONCE(data->user_stack.size > sz,
2314 "user stack dump failure\n"))
2315 return -EFAULT;
2316 }
2317 }
2318
2319 if (type & PERF_SAMPLE_WEIGHT) {
2320 OVERFLOW_CHECK_u64(array);
2321 data->weight = *array;
2322 array++;
2323 }
2324
2325 if (type & PERF_SAMPLE_DATA_SRC) {
2326 OVERFLOW_CHECK_u64(array);
2327 data->data_src = *array;
2328 array++;
2329 }
2330
2331 if (type & PERF_SAMPLE_TRANSACTION) {
2332 OVERFLOW_CHECK_u64(array);
2333 data->transaction = *array;
2334 array++;
2335 }
2336
2337 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2338 if (type & PERF_SAMPLE_REGS_INTR) {
2339 OVERFLOW_CHECK_u64(array);
2340 data->intr_regs.abi = *array;
2341 array++;
2342
2343 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2344 u64 mask = evsel->core.attr.sample_regs_intr;
2345
2346 sz = hweight64(mask) * sizeof(u64);
2347 OVERFLOW_CHECK(array, sz, max_size);
2348 data->intr_regs.mask = mask;
2349 data->intr_regs.regs = (u64 *)array;
2350 array = (void *)array + sz;
2351 }
2352 }
2353
2354 data->phys_addr = 0;
2355 if (type & PERF_SAMPLE_PHYS_ADDR) {
2356 data->phys_addr = *array;
2357 array++;
2358 }
2359
2360 data->cgroup = 0;
2361 if (type & PERF_SAMPLE_CGROUP) {
2362 data->cgroup = *array;
2363 array++;
2364 }
2365
2366 data->data_page_size = 0;
2367 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2368 data->data_page_size = *array;
2369 array++;
2370 }
2371
2372 if (type & PERF_SAMPLE_AUX) {
2373 OVERFLOW_CHECK_u64(array);
2374 sz = *array++;
2375
2376 OVERFLOW_CHECK(array, sz, max_size);
2377 /* Undo swap of data */
2378 if (swapped)
2379 mem_bswap_64((char *)array, sz);
2380 data->aux_sample.size = sz;
2381 data->aux_sample.data = (char *)array;
2382 array = (void *)array + sz;
2383 }
2384
2385 return 0;
2386 }
2387
2388 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2389 u64 *timestamp)
2390 {
2391 u64 type = evsel->core.attr.sample_type;
2392 const __u64 *array;
2393
2394 if (!(type & PERF_SAMPLE_TIME))
2395 return -1;
2396
2397 if (event->header.type != PERF_RECORD_SAMPLE) {
2398 struct perf_sample data = {
2399 .time = -1ULL,
2400 };
2401
2402 if (!evsel->core.attr.sample_id_all)
2403 return -1;
2404 if (perf_evsel__parse_id_sample(evsel, event, &data))
2405 return -1;
2406
2407 *timestamp = data.time;
2408 return 0;
2409 }
2410
2411 array = event->sample.array;
2412
2413 if (perf_event__check_size(event, evsel->sample_size))
2414 return -EFAULT;
2415
2416 if (type & PERF_SAMPLE_IDENTIFIER)
2417 array++;
2418
2419 if (type & PERF_SAMPLE_IP)
2420 array++;
2421
2422 if (type & PERF_SAMPLE_TID)
2423 array++;
2424
2425 if (type & PERF_SAMPLE_TIME)
2426 *timestamp = *array;
2427
2428 return 0;
2429 }
2430
2431 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2432 {
2433 return tep_find_field(evsel->tp_format, name);
2434 }
2435
2436 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2437 {
2438 struct tep_format_field *field = evsel__field(evsel, name);
2439 int offset;
2440
2441 if (!field)
2442 return NULL;
2443
2444 offset = field->offset;
2445
2446 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2447 offset = *(int *)(sample->raw_data + field->offset);
2448 offset &= 0xffff;
2449 }
2450
2451 return sample->raw_data + offset;
2452 }
2453
2454 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2455 bool needs_swap)
2456 {
2457 u64 value;
2458 void *ptr = sample->raw_data + field->offset;
2459
2460 switch (field->size) {
2461 case 1:
2462 return *(u8 *)ptr;
2463 case 2:
2464 value = *(u16 *)ptr;
2465 break;
2466 case 4:
2467 value = *(u32 *)ptr;
2468 break;
2469 case 8:
2470 memcpy(&value, ptr, sizeof(u64));
2471 break;
2472 default:
2473 return 0;
2474 }
2475
2476 if (!needs_swap)
2477 return value;
2478
2479 switch (field->size) {
2480 case 2:
2481 return bswap_16(value);
2482 case 4:
2483 return bswap_32(value);
2484 case 8:
2485 return bswap_64(value);
2486 default:
2487 return 0;
2488 }
2489
2490 return 0;
2491 }
2492
2493 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2494 {
2495 struct tep_format_field *field = evsel__field(evsel, name);
2496
2497 if (!field)
2498 return 0;
2499
2500 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2501 }
2502
2503 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2504 {
2505 int paranoid;
2506
2507 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2508 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2509 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2510 /*
2511 * If it's cycles then fall back to hrtimer based
2512 * cpu-clock-tick sw counter, which is always available even if
2513 * no PMU support.
2514 *
2515 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2516 * b0a873e).
2517 */
2518 scnprintf(msg, msgsize, "%s",
2519 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2520
2521 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2522 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2523
2524 zfree(&evsel->name);
2525 return true;
2526 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2527 (paranoid = perf_event_paranoid()) > 1) {
2528 const char *name = evsel__name(evsel);
2529 char *new_name;
2530 const char *sep = ":";
2531
2532 /* If event has exclude user then don't exclude kernel. */
2533 if (evsel->core.attr.exclude_user)
2534 return false;
2535
2536 /* Is there already the separator in the name. */
2537 if (strchr(name, '/') ||
2538 (strchr(name, ':') && !evsel->is_libpfm_event))
2539 sep = "";
2540
2541 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2542 return false;
2543
2544 if (evsel->name)
2545 free(evsel->name);
2546 evsel->name = new_name;
2547 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2548 "to fall back to excluding kernel and hypervisor "
2549 " samples", paranoid);
2550 evsel->core.attr.exclude_kernel = 1;
2551 evsel->core.attr.exclude_hv = 1;
2552
2553 return true;
2554 }
2555
2556 return false;
2557 }
2558
2559 static bool find_process(const char *name)
2560 {
2561 size_t len = strlen(name);
2562 DIR *dir;
2563 struct dirent *d;
2564 int ret = -1;
2565
2566 dir = opendir(procfs__mountpoint());
2567 if (!dir)
2568 return false;
2569
2570 /* Walk through the directory. */
2571 while (ret && (d = readdir(dir)) != NULL) {
2572 char path[PATH_MAX];
2573 char *data;
2574 size_t size;
2575
2576 if ((d->d_type != DT_DIR) ||
2577 !strcmp(".", d->d_name) ||
2578 !strcmp("..", d->d_name))
2579 continue;
2580
2581 scnprintf(path, sizeof(path), "%s/%s/comm",
2582 procfs__mountpoint(), d->d_name);
2583
2584 if (filename__read_str(path, &data, &size))
2585 continue;
2586
2587 ret = strncmp(name, data, len);
2588 free(data);
2589 }
2590
2591 closedir(dir);
2592 return ret ? false : true;
2593 }
2594
2595 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2596 int err, char *msg, size_t size)
2597 {
2598 char sbuf[STRERR_BUFSIZE];
2599 int printed = 0, enforced = 0;
2600
2601 switch (err) {
2602 case EPERM:
2603 case EACCES:
2604 printed += scnprintf(msg + printed, size - printed,
2605 "Access to performance monitoring and observability operations is limited.\n");
2606
2607 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2608 if (enforced) {
2609 printed += scnprintf(msg + printed, size - printed,
2610 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2611 "monitoring and observability operations. Inspect system audit records for\n"
2612 "more perf_event access control information and adjusting the policy.\n");
2613 }
2614 }
2615
2616 if (err == EPERM)
2617 printed += scnprintf(msg, size,
2618 "No permission to enable %s event.\n\n", evsel__name(evsel));
2619
2620 return scnprintf(msg + printed, size - printed,
2621 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2622 "access to performance monitoring and observability operations for processes\n"
2623 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2624 "More information can be found at 'Perf events and tool security' document:\n"
2625 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2626 "perf_event_paranoid setting is %d:\n"
2627 " -1: Allow use of (almost) all events by all users\n"
2628 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2629 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2630 ">= 1: Disallow CPU event access\n"
2631 ">= 2: Disallow kernel profiling\n"
2632 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2633 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2634 perf_event_paranoid());
2635 case ENOENT:
2636 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2637 case EMFILE:
2638 return scnprintf(msg, size, "%s",
2639 "Too many events are opened.\n"
2640 "Probably the maximum number of open file descriptors has been reached.\n"
2641 "Hint: Try again after reducing the number of events.\n"
2642 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2643 case ENOMEM:
2644 if (evsel__has_callchain(evsel) &&
2645 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2646 return scnprintf(msg, size,
2647 "Not enough memory to setup event with callchain.\n"
2648 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2649 "Hint: Current value: %d", sysctl__max_stack());
2650 break;
2651 case ENODEV:
2652 if (target->cpu_list)
2653 return scnprintf(msg, size, "%s",
2654 "No such device - did you specify an out-of-range profile CPU?");
2655 break;
2656 case EOPNOTSUPP:
2657 if (evsel->core.attr.aux_output)
2658 return scnprintf(msg, size,
2659 "%s: PMU Hardware doesn't support 'aux_output' feature",
2660 evsel__name(evsel));
2661 if (evsel->core.attr.sample_period != 0)
2662 return scnprintf(msg, size,
2663 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2664 evsel__name(evsel));
2665 if (evsel->core.attr.precise_ip)
2666 return scnprintf(msg, size, "%s",
2667 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2668 #if defined(__i386__) || defined(__x86_64__)
2669 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2670 return scnprintf(msg, size, "%s",
2671 "No hardware sampling interrupt available.\n");
2672 #endif
2673 break;
2674 case EBUSY:
2675 if (find_process("oprofiled"))
2676 return scnprintf(msg, size,
2677 "The PMU counters are busy/taken by another profiler.\n"
2678 "We found oprofile daemon running, please stop it and try again.");
2679 break;
2680 case EINVAL:
2681 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
2682 return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
2683 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2684 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2685 if (perf_missing_features.clockid)
2686 return scnprintf(msg, size, "clockid feature not supported.");
2687 if (perf_missing_features.clockid_wrong)
2688 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2689 if (perf_missing_features.aux_output)
2690 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2691 break;
2692 default:
2693 break;
2694 }
2695
2696 return scnprintf(msg, size,
2697 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2698 "/bin/dmesg | grep -i perf may provide additional information.\n",
2699 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2700 }
2701
2702 struct perf_env *evsel__env(struct evsel *evsel)
2703 {
2704 if (evsel && evsel->evlist)
2705 return evsel->evlist->env;
2706 return &perf_env;
2707 }
2708
2709 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2710 {
2711 int cpu, thread;
2712
2713 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2714 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2715 thread++) {
2716 int fd = FD(evsel, cpu, thread);
2717
2718 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2719 cpu, thread, fd) < 0)
2720 return -1;
2721 }
2722 }
2723
2724 return 0;
2725 }
2726
2727 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2728 {
2729 struct perf_cpu_map *cpus = evsel->core.cpus;
2730 struct perf_thread_map *threads = evsel->core.threads;
2731
2732 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2733 return -ENOMEM;
2734
2735 return store_evsel_ids(evsel, evlist);
2736 }
Linux with added FPC-III support