xref: /linux/tools/perf/util/evsel.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 #include "util/intel-tpebs.h"
63 
64 #include <linux/ctype.h>
65 
66 #ifdef HAVE_LIBTRACEEVENT
67 #include <traceevent/event-parse.h>
68 #endif
69 
70 struct perf_missing_features perf_missing_features;
71 
72 static clockid_t clockid;
73 
74 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
75 	NULL,
76 	"duration_time",
77 	"user_time",
78 	"system_time",
79 };
80 
81 const char *perf_tool_event__to_str(enum perf_tool_event ev)
82 {
83 	if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
84 		return perf_tool_event__tool_names[ev];
85 
86 	return NULL;
87 }
88 
89 enum perf_tool_event perf_tool_event__from_str(const char *str)
90 {
91 	int i;
92 
93 	perf_tool_event__for_each_event(i) {
94 		if (!strcmp(str, perf_tool_event__tool_names[i]))
95 			return i;
96 	}
97 	return PERF_TOOL_NONE;
98 }
99 
100 
101 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
102 {
103 	return 0;
104 }
105 
106 void __weak test_attr__ready(void) { }
107 
108 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
109 {
110 }
111 
112 static struct {
113 	size_t	size;
114 	int	(*init)(struct evsel *evsel);
115 	void	(*fini)(struct evsel *evsel);
116 } perf_evsel__object = {
117 	.size = sizeof(struct evsel),
118 	.init = evsel__no_extra_init,
119 	.fini = evsel__no_extra_fini,
120 };
121 
122 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
123 			 void (*fini)(struct evsel *evsel))
124 {
125 
126 	if (object_size == 0)
127 		goto set_methods;
128 
129 	if (perf_evsel__object.size > object_size)
130 		return -EINVAL;
131 
132 	perf_evsel__object.size = object_size;
133 
134 set_methods:
135 	if (init != NULL)
136 		perf_evsel__object.init = init;
137 
138 	if (fini != NULL)
139 		perf_evsel__object.fini = fini;
140 
141 	return 0;
142 }
143 
144 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
145 
146 int __evsel__sample_size(u64 sample_type)
147 {
148 	u64 mask = sample_type & PERF_SAMPLE_MASK;
149 	int size = 0;
150 	int i;
151 
152 	for (i = 0; i < 64; i++) {
153 		if (mask & (1ULL << i))
154 			size++;
155 	}
156 
157 	size *= sizeof(u64);
158 
159 	return size;
160 }
161 
162 /**
163  * __perf_evsel__calc_id_pos - calculate id_pos.
164  * @sample_type: sample type
165  *
166  * This function returns the position of the event id (PERF_SAMPLE_ID or
167  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
168  * perf_record_sample.
169  */
170 static int __perf_evsel__calc_id_pos(u64 sample_type)
171 {
172 	int idx = 0;
173 
174 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
175 		return 0;
176 
177 	if (!(sample_type & PERF_SAMPLE_ID))
178 		return -1;
179 
180 	if (sample_type & PERF_SAMPLE_IP)
181 		idx += 1;
182 
183 	if (sample_type & PERF_SAMPLE_TID)
184 		idx += 1;
185 
186 	if (sample_type & PERF_SAMPLE_TIME)
187 		idx += 1;
188 
189 	if (sample_type & PERF_SAMPLE_ADDR)
190 		idx += 1;
191 
192 	return idx;
193 }
194 
195 /**
196  * __perf_evsel__calc_is_pos - calculate is_pos.
197  * @sample_type: sample type
198  *
199  * This function returns the position (counting backwards) of the event id
200  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
201  * sample_id_all is used there is an id sample appended to non-sample events.
202  */
203 static int __perf_evsel__calc_is_pos(u64 sample_type)
204 {
205 	int idx = 1;
206 
207 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
208 		return 1;
209 
210 	if (!(sample_type & PERF_SAMPLE_ID))
211 		return -1;
212 
213 	if (sample_type & PERF_SAMPLE_CPU)
214 		idx += 1;
215 
216 	if (sample_type & PERF_SAMPLE_STREAM_ID)
217 		idx += 1;
218 
219 	return idx;
220 }
221 
222 void evsel__calc_id_pos(struct evsel *evsel)
223 {
224 	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
225 	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
226 }
227 
228 void __evsel__set_sample_bit(struct evsel *evsel,
229 				  enum perf_event_sample_format bit)
230 {
231 	if (!(evsel->core.attr.sample_type & bit)) {
232 		evsel->core.attr.sample_type |= bit;
233 		evsel->sample_size += sizeof(u64);
234 		evsel__calc_id_pos(evsel);
235 	}
236 }
237 
238 void __evsel__reset_sample_bit(struct evsel *evsel,
239 				    enum perf_event_sample_format bit)
240 {
241 	if (evsel->core.attr.sample_type & bit) {
242 		evsel->core.attr.sample_type &= ~bit;
243 		evsel->sample_size -= sizeof(u64);
244 		evsel__calc_id_pos(evsel);
245 	}
246 }
247 
248 void evsel__set_sample_id(struct evsel *evsel,
249 			       bool can_sample_identifier)
250 {
251 	if (can_sample_identifier) {
252 		evsel__reset_sample_bit(evsel, ID);
253 		evsel__set_sample_bit(evsel, IDENTIFIER);
254 	} else {
255 		evsel__set_sample_bit(evsel, ID);
256 	}
257 	evsel->core.attr.read_format |= PERF_FORMAT_ID;
258 }
259 
260 /**
261  * evsel__is_function_event - Return whether given evsel is a function
262  * trace event
263  *
264  * @evsel - evsel selector to be tested
265  *
266  * Return %true if event is function trace event
267  */
268 bool evsel__is_function_event(struct evsel *evsel)
269 {
270 #define FUNCTION_EVENT "ftrace:function"
271 
272 	return evsel->name &&
273 	       !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
274 
275 #undef FUNCTION_EVENT
276 }
277 
278 void evsel__init(struct evsel *evsel,
279 		 struct perf_event_attr *attr, int idx)
280 {
281 	perf_evsel__init(&evsel->core, attr, idx);
282 	evsel->tracking	   = !idx;
283 	evsel->unit	   = strdup("");
284 	evsel->scale	   = 1.0;
285 	evsel->max_events  = ULONG_MAX;
286 	evsel->evlist	   = NULL;
287 	evsel->bpf_obj	   = NULL;
288 	evsel->bpf_fd	   = -1;
289 	INIT_LIST_HEAD(&evsel->config_terms);
290 	INIT_LIST_HEAD(&evsel->bpf_counter_list);
291 	INIT_LIST_HEAD(&evsel->bpf_filters);
292 	perf_evsel__object.init(evsel);
293 	evsel->sample_size = __evsel__sample_size(attr->sample_type);
294 	evsel__calc_id_pos(evsel);
295 	evsel->cmdline_group_boundary = false;
296 	evsel->metric_events = NULL;
297 	evsel->per_pkg_mask  = NULL;
298 	evsel->collect_stat  = false;
299 	evsel->pmu_name      = NULL;
300 	evsel->group_pmu_name = NULL;
301 	evsel->skippable     = false;
302 }
303 
304 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
305 {
306 	struct evsel *evsel = zalloc(perf_evsel__object.size);
307 
308 	if (!evsel)
309 		return NULL;
310 	evsel__init(evsel, attr, idx);
311 
312 	if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
313 		evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
314 					    PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
315 		evsel->core.attr.sample_period = 1;
316 	}
317 
318 	if (evsel__is_clock(evsel)) {
319 		free((char *)evsel->unit);
320 		evsel->unit = strdup("msec");
321 		evsel->scale = 1e-6;
322 	}
323 
324 	return evsel;
325 }
326 
327 int copy_config_terms(struct list_head *dst, struct list_head *src)
328 {
329 	struct evsel_config_term *pos, *tmp;
330 
331 	list_for_each_entry(pos, src, list) {
332 		tmp = malloc(sizeof(*tmp));
333 		if (tmp == NULL)
334 			return -ENOMEM;
335 
336 		*tmp = *pos;
337 		if (tmp->free_str) {
338 			tmp->val.str = strdup(pos->val.str);
339 			if (tmp->val.str == NULL) {
340 				free(tmp);
341 				return -ENOMEM;
342 			}
343 		}
344 		list_add_tail(&tmp->list, dst);
345 	}
346 	return 0;
347 }
348 
349 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
350 {
351 	return copy_config_terms(&dst->config_terms, &src->config_terms);
352 }
353 
354 /**
355  * evsel__clone - create a new evsel copied from @orig
356  * @orig: original evsel
357  *
358  * The assumption is that @orig is not configured nor opened yet.
359  * So we only care about the attributes that can be set while it's parsed.
360  */
361 struct evsel *evsel__clone(struct evsel *orig)
362 {
363 	struct evsel *evsel;
364 
365 	BUG_ON(orig->core.fd);
366 	BUG_ON(orig->counts);
367 	BUG_ON(orig->priv);
368 	BUG_ON(orig->per_pkg_mask);
369 
370 	/* cannot handle BPF objects for now */
371 	if (orig->bpf_obj)
372 		return NULL;
373 
374 	evsel = evsel__new(&orig->core.attr);
375 	if (evsel == NULL)
376 		return NULL;
377 
378 	evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
379 	evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
380 	evsel->core.threads = perf_thread_map__get(orig->core.threads);
381 	evsel->core.nr_members = orig->core.nr_members;
382 	evsel->core.system_wide = orig->core.system_wide;
383 	evsel->core.requires_cpu = orig->core.requires_cpu;
384 	evsel->core.is_pmu_core = orig->core.is_pmu_core;
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->group_pmu_name) {
402 		evsel->group_pmu_name = strdup(orig->group_pmu_name);
403 		if (evsel->group_pmu_name == NULL)
404 			goto out_err;
405 	}
406 	if (orig->filter) {
407 		evsel->filter = strdup(orig->filter);
408 		if (evsel->filter == NULL)
409 			goto out_err;
410 	}
411 	if (orig->metric_id) {
412 		evsel->metric_id = strdup(orig->metric_id);
413 		if (evsel->metric_id == NULL)
414 			goto out_err;
415 	}
416 	evsel->cgrp = cgroup__get(orig->cgrp);
417 #ifdef HAVE_LIBTRACEEVENT
418 	evsel->tp_format = orig->tp_format;
419 #endif
420 	evsel->handler = orig->handler;
421 	evsel->core.leader = orig->core.leader;
422 
423 	evsel->max_events = orig->max_events;
424 	evsel->tool_event = orig->tool_event;
425 	free((char *)evsel->unit);
426 	evsel->unit = strdup(orig->unit);
427 	if (evsel->unit == NULL)
428 		goto out_err;
429 
430 	evsel->scale = orig->scale;
431 	evsel->snapshot = orig->snapshot;
432 	evsel->per_pkg = orig->per_pkg;
433 	evsel->percore = orig->percore;
434 	evsel->precise_max = orig->precise_max;
435 	evsel->is_libpfm_event = orig->is_libpfm_event;
436 
437 	evsel->exclude_GH = orig->exclude_GH;
438 	evsel->sample_read = orig->sample_read;
439 	evsel->auto_merge_stats = orig->auto_merge_stats;
440 	evsel->collect_stat = orig->collect_stat;
441 	evsel->weak_group = orig->weak_group;
442 	evsel->use_config_name = orig->use_config_name;
443 	evsel->pmu = orig->pmu;
444 
445 	if (evsel__copy_config_terms(evsel, orig) < 0)
446 		goto out_err;
447 
448 	return evsel;
449 
450 out_err:
451 	evsel__delete(evsel);
452 	return NULL;
453 }
454 
455 /*
456  * Returns pointer with encoded error via <linux/err.h> interface.
457  */
458 #ifdef HAVE_LIBTRACEEVENT
459 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx, bool format)
460 {
461 	struct evsel *evsel = zalloc(perf_evsel__object.size);
462 	int err = -ENOMEM;
463 
464 	if (evsel == NULL) {
465 		goto out_err;
466 	} else {
467 		struct perf_event_attr attr = {
468 			.type	       = PERF_TYPE_TRACEPOINT,
469 			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
470 					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
471 		};
472 
473 		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
474 			goto out_free;
475 
476 		event_attr_init(&attr);
477 
478 		if (format) {
479 			evsel->tp_format = trace_event__tp_format(sys, name);
480 			if (IS_ERR(evsel->tp_format)) {
481 				err = PTR_ERR(evsel->tp_format);
482 				goto out_free;
483 			}
484 			attr.config = evsel->tp_format->id;
485 		} else {
486 			attr.config = (__u64) -1;
487 		}
488 
489 
490 		attr.sample_period = 1;
491 		evsel__init(evsel, &attr, idx);
492 	}
493 
494 	return evsel;
495 
496 out_free:
497 	zfree(&evsel->name);
498 	free(evsel);
499 out_err:
500 	return ERR_PTR(err);
501 }
502 #endif
503 
504 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
505 	"cycles",
506 	"instructions",
507 	"cache-references",
508 	"cache-misses",
509 	"branches",
510 	"branch-misses",
511 	"bus-cycles",
512 	"stalled-cycles-frontend",
513 	"stalled-cycles-backend",
514 	"ref-cycles",
515 };
516 
517 char *evsel__bpf_counter_events;
518 
519 bool evsel__match_bpf_counter_events(const char *name)
520 {
521 	int name_len;
522 	bool match;
523 	char *ptr;
524 
525 	if (!evsel__bpf_counter_events)
526 		return false;
527 
528 	ptr = strstr(evsel__bpf_counter_events, name);
529 	name_len = strlen(name);
530 
531 	/* check name matches a full token in evsel__bpf_counter_events */
532 	match = (ptr != NULL) &&
533 		((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
534 		((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
535 
536 	return match;
537 }
538 
539 static const char *__evsel__hw_name(u64 config)
540 {
541 	if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
542 		return evsel__hw_names[config];
543 
544 	return "unknown-hardware";
545 }
546 
547 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
548 {
549 	int colon = 0, r = 0;
550 	struct perf_event_attr *attr = &evsel->core.attr;
551 	bool exclude_guest_default = false;
552 
553 #define MOD_PRINT(context, mod)	do {					\
554 		if (!attr->exclude_##context) {				\
555 			if (!colon) colon = ++r;			\
556 			r += scnprintf(bf + r, size - r, "%c", mod);	\
557 		} } while(0)
558 
559 	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
560 		MOD_PRINT(kernel, 'k');
561 		MOD_PRINT(user, 'u');
562 		MOD_PRINT(hv, 'h');
563 		exclude_guest_default = true;
564 	}
565 
566 	if (attr->precise_ip) {
567 		if (!colon)
568 			colon = ++r;
569 		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
570 		exclude_guest_default = true;
571 	}
572 
573 	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
574 		MOD_PRINT(host, 'H');
575 		MOD_PRINT(guest, 'G');
576 	}
577 #undef MOD_PRINT
578 	if (colon)
579 		bf[colon - 1] = ':';
580 	return r;
581 }
582 
583 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
584 {
585 	return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
586 }
587 
588 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
589 {
590 	int r = arch_evsel__hw_name(evsel, bf, size);
591 	return r + evsel__add_modifiers(evsel, bf + r, size - r);
592 }
593 
594 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
595 	"cpu-clock",
596 	"task-clock",
597 	"page-faults",
598 	"context-switches",
599 	"cpu-migrations",
600 	"minor-faults",
601 	"major-faults",
602 	"alignment-faults",
603 	"emulation-faults",
604 	"dummy",
605 };
606 
607 static const char *__evsel__sw_name(u64 config)
608 {
609 	if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
610 		return evsel__sw_names[config];
611 	return "unknown-software";
612 }
613 
614 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
615 {
616 	int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
617 	return r + evsel__add_modifiers(evsel, bf + r, size - r);
618 }
619 
620 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
621 {
622 	return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
623 }
624 
625 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
626 {
627 	int r;
628 
629 	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
630 
631 	if (type & HW_BREAKPOINT_R)
632 		r += scnprintf(bf + r, size - r, "r");
633 
634 	if (type & HW_BREAKPOINT_W)
635 		r += scnprintf(bf + r, size - r, "w");
636 
637 	if (type & HW_BREAKPOINT_X)
638 		r += scnprintf(bf + r, size - r, "x");
639 
640 	return r;
641 }
642 
643 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
644 {
645 	struct perf_event_attr *attr = &evsel->core.attr;
646 	int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
647 	return r + evsel__add_modifiers(evsel, bf + r, size - r);
648 }
649 
650 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
651  { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
652  { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
653  { "LLC",	"L2",							},
654  { "dTLB",	"d-tlb",	"Data-TLB",				},
655  { "iTLB",	"i-tlb",	"Instruction-TLB",			},
656  { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
657  { "node",								},
658 };
659 
660 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
661  { "load",	"loads",	"read",					},
662  { "store",	"stores",	"write",				},
663  { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
664 };
665 
666 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
667  { "refs",	"Reference",	"ops",		"access",		},
668  { "misses",	"miss",							},
669 };
670 
671 #define C(x)		PERF_COUNT_HW_CACHE_##x
672 #define CACHE_READ	(1 << C(OP_READ))
673 #define CACHE_WRITE	(1 << C(OP_WRITE))
674 #define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
675 #define COP(x)		(1 << x)
676 
677 /*
678  * cache operation stat
679  * L1I : Read and prefetch only
680  * ITLB and BPU : Read-only
681  */
682 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
683  [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
684  [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
685  [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
686  [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
687  [C(ITLB)]	= (CACHE_READ),
688  [C(BPU)]	= (CACHE_READ),
689  [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
690 };
691 
692 bool evsel__is_cache_op_valid(u8 type, u8 op)
693 {
694 	if (evsel__hw_cache_stat[type] & COP(op))
695 		return true;	/* valid */
696 	else
697 		return false;	/* invalid */
698 }
699 
700 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
701 {
702 	if (result) {
703 		return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
704 				 evsel__hw_cache_op[op][0],
705 				 evsel__hw_cache_result[result][0]);
706 	}
707 
708 	return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
709 			 evsel__hw_cache_op[op][1]);
710 }
711 
712 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
713 {
714 	u8 op, result, type = (config >>  0) & 0xff;
715 	const char *err = "unknown-ext-hardware-cache-type";
716 
717 	if (type >= PERF_COUNT_HW_CACHE_MAX)
718 		goto out_err;
719 
720 	op = (config >>  8) & 0xff;
721 	err = "unknown-ext-hardware-cache-op";
722 	if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
723 		goto out_err;
724 
725 	result = (config >> 16) & 0xff;
726 	err = "unknown-ext-hardware-cache-result";
727 	if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
728 		goto out_err;
729 
730 	err = "invalid-cache";
731 	if (!evsel__is_cache_op_valid(type, op))
732 		goto out_err;
733 
734 	return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
735 out_err:
736 	return scnprintf(bf, size, "%s", err);
737 }
738 
739 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
740 {
741 	int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
742 	return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
743 }
744 
745 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
746 {
747 	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
748 	return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
749 }
750 
751 const char *evsel__name(struct evsel *evsel)
752 {
753 	char bf[128];
754 
755 	if (!evsel)
756 		goto out_unknown;
757 
758 	if (evsel->name)
759 		return evsel->name;
760 
761 	switch (evsel->core.attr.type) {
762 	case PERF_TYPE_RAW:
763 		evsel__raw_name(evsel, bf, sizeof(bf));
764 		break;
765 
766 	case PERF_TYPE_HARDWARE:
767 		evsel__hw_name(evsel, bf, sizeof(bf));
768 		break;
769 
770 	case PERF_TYPE_HW_CACHE:
771 		evsel__hw_cache_name(evsel, bf, sizeof(bf));
772 		break;
773 
774 	case PERF_TYPE_SOFTWARE:
775 		if (evsel__is_tool(evsel))
776 			evsel__tool_name(evsel__tool_event(evsel), bf, sizeof(bf));
777 		else
778 			evsel__sw_name(evsel, bf, sizeof(bf));
779 		break;
780 
781 	case PERF_TYPE_TRACEPOINT:
782 		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
783 		break;
784 
785 	case PERF_TYPE_BREAKPOINT:
786 		evsel__bp_name(evsel, bf, sizeof(bf));
787 		break;
788 
789 	default:
790 		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
791 			  evsel->core.attr.type);
792 		break;
793 	}
794 
795 	evsel->name = strdup(bf);
796 
797 	if (evsel->name)
798 		return evsel->name;
799 out_unknown:
800 	return "unknown";
801 }
802 
803 bool evsel__name_is(struct evsel *evsel, const char *name)
804 {
805 	return !strcmp(evsel__name(evsel), name);
806 }
807 
808 const char *evsel__metric_id(const struct evsel *evsel)
809 {
810 	if (evsel->metric_id)
811 		return evsel->metric_id;
812 
813 	if (evsel__is_tool(evsel))
814 		return perf_tool_event__to_str(evsel__tool_event(evsel));
815 
816 	return "unknown";
817 }
818 
819 const char *evsel__group_name(struct evsel *evsel)
820 {
821 	return evsel->group_name ?: "anon group";
822 }
823 
824 /*
825  * Returns the group details for the specified leader,
826  * with following rules.
827  *
828  *  For record -e '{cycles,instructions}'
829  *    'anon group { cycles:u, instructions:u }'
830  *
831  *  For record -e 'cycles,instructions' and report --group
832  *    'cycles:u, instructions:u'
833  */
834 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
835 {
836 	int ret = 0;
837 	bool first = true;
838 	struct evsel *pos;
839 	const char *group_name = evsel__group_name(evsel);
840 
841 	if (!evsel->forced_leader)
842 		ret = scnprintf(buf, size, "%s { ", group_name);
843 
844 	for_each_group_evsel(pos, evsel) {
845 		if (symbol_conf.skip_empty &&
846 		    evsel__hists(pos)->stats.nr_samples == 0)
847 			continue;
848 
849 		ret += scnprintf(buf + ret, size - ret, "%s%s",
850 				 first ? "" : ", ", evsel__name(pos));
851 		first = false;
852 	}
853 
854 	if (!evsel->forced_leader)
855 		ret += scnprintf(buf + ret, size - ret, " }");
856 
857 	return ret;
858 }
859 
860 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
861 				      struct callchain_param *param)
862 {
863 	bool function = evsel__is_function_event(evsel);
864 	struct perf_event_attr *attr = &evsel->core.attr;
865 	const char *arch = perf_env__arch(evsel__env(evsel));
866 
867 	evsel__set_sample_bit(evsel, CALLCHAIN);
868 
869 	attr->sample_max_stack = param->max_stack;
870 
871 	if (opts->kernel_callchains)
872 		attr->exclude_callchain_user = 1;
873 	if (opts->user_callchains)
874 		attr->exclude_callchain_kernel = 1;
875 	if (param->record_mode == CALLCHAIN_LBR) {
876 		if (!opts->branch_stack) {
877 			if (attr->exclude_user) {
878 				pr_warning("LBR callstack option is only available "
879 					   "to get user callchain information. "
880 					   "Falling back to framepointers.\n");
881 			} else {
882 				evsel__set_sample_bit(evsel, BRANCH_STACK);
883 				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
884 							PERF_SAMPLE_BRANCH_CALL_STACK |
885 							PERF_SAMPLE_BRANCH_NO_CYCLES |
886 							PERF_SAMPLE_BRANCH_NO_FLAGS |
887 							PERF_SAMPLE_BRANCH_HW_INDEX;
888 			}
889 		} else
890 			 pr_warning("Cannot use LBR callstack with branch stack. "
891 				    "Falling back to framepointers.\n");
892 	}
893 
894 	if (param->record_mode == CALLCHAIN_DWARF) {
895 		if (!function) {
896 			evsel__set_sample_bit(evsel, REGS_USER);
897 			evsel__set_sample_bit(evsel, STACK_USER);
898 			if (opts->sample_user_regs &&
899 			    DWARF_MINIMAL_REGS(arch) != arch__user_reg_mask()) {
900 				attr->sample_regs_user |= DWARF_MINIMAL_REGS(arch);
901 				pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
902 					   "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
903 					   "so the minimal registers set (IP, SP) is explicitly forced.\n");
904 			} else {
905 				attr->sample_regs_user |= arch__user_reg_mask();
906 			}
907 			attr->sample_stack_user = param->dump_size;
908 			attr->exclude_callchain_user = 1;
909 		} else {
910 			pr_info("Cannot use DWARF unwind for function trace event,"
911 				" falling back to framepointers.\n");
912 		}
913 	}
914 
915 	if (function) {
916 		pr_info("Disabling user space callchains for function trace event.\n");
917 		attr->exclude_callchain_user = 1;
918 	}
919 }
920 
921 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
922 			     struct callchain_param *param)
923 {
924 	if (param->enabled)
925 		return __evsel__config_callchain(evsel, opts, param);
926 }
927 
928 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
929 {
930 	struct perf_event_attr *attr = &evsel->core.attr;
931 
932 	evsel__reset_sample_bit(evsel, CALLCHAIN);
933 	if (param->record_mode == CALLCHAIN_LBR) {
934 		evsel__reset_sample_bit(evsel, BRANCH_STACK);
935 		attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
936 					      PERF_SAMPLE_BRANCH_CALL_STACK |
937 					      PERF_SAMPLE_BRANCH_HW_INDEX);
938 	}
939 	if (param->record_mode == CALLCHAIN_DWARF) {
940 		evsel__reset_sample_bit(evsel, REGS_USER);
941 		evsel__reset_sample_bit(evsel, STACK_USER);
942 	}
943 }
944 
945 static void evsel__apply_config_terms(struct evsel *evsel,
946 				      struct record_opts *opts, bool track)
947 {
948 	struct evsel_config_term *term;
949 	struct list_head *config_terms = &evsel->config_terms;
950 	struct perf_event_attr *attr = &evsel->core.attr;
951 	/* callgraph default */
952 	struct callchain_param param = {
953 		.record_mode = callchain_param.record_mode,
954 	};
955 	u32 dump_size = 0;
956 	int max_stack = 0;
957 	const char *callgraph_buf = NULL;
958 
959 	list_for_each_entry(term, config_terms, list) {
960 		switch (term->type) {
961 		case EVSEL__CONFIG_TERM_PERIOD:
962 			if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
963 				attr->sample_period = term->val.period;
964 				attr->freq = 0;
965 				evsel__reset_sample_bit(evsel, PERIOD);
966 			}
967 			break;
968 		case EVSEL__CONFIG_TERM_FREQ:
969 			if (!(term->weak && opts->user_freq != UINT_MAX)) {
970 				attr->sample_freq = term->val.freq;
971 				attr->freq = 1;
972 				evsel__set_sample_bit(evsel, PERIOD);
973 			}
974 			break;
975 		case EVSEL__CONFIG_TERM_TIME:
976 			if (term->val.time)
977 				evsel__set_sample_bit(evsel, TIME);
978 			else
979 				evsel__reset_sample_bit(evsel, TIME);
980 			break;
981 		case EVSEL__CONFIG_TERM_CALLGRAPH:
982 			callgraph_buf = term->val.str;
983 			break;
984 		case EVSEL__CONFIG_TERM_BRANCH:
985 			if (term->val.str && strcmp(term->val.str, "no")) {
986 				evsel__set_sample_bit(evsel, BRANCH_STACK);
987 				parse_branch_str(term->val.str,
988 						 &attr->branch_sample_type);
989 			} else
990 				evsel__reset_sample_bit(evsel, BRANCH_STACK);
991 			break;
992 		case EVSEL__CONFIG_TERM_STACK_USER:
993 			dump_size = term->val.stack_user;
994 			break;
995 		case EVSEL__CONFIG_TERM_MAX_STACK:
996 			max_stack = term->val.max_stack;
997 			break;
998 		case EVSEL__CONFIG_TERM_MAX_EVENTS:
999 			evsel->max_events = term->val.max_events;
1000 			break;
1001 		case EVSEL__CONFIG_TERM_INHERIT:
1002 			/*
1003 			 * attr->inherit should has already been set by
1004 			 * evsel__config. If user explicitly set
1005 			 * inherit using config terms, override global
1006 			 * opt->no_inherit setting.
1007 			 */
1008 			attr->inherit = term->val.inherit ? 1 : 0;
1009 			break;
1010 		case EVSEL__CONFIG_TERM_OVERWRITE:
1011 			attr->write_backward = term->val.overwrite ? 1 : 0;
1012 			break;
1013 		case EVSEL__CONFIG_TERM_DRV_CFG:
1014 			break;
1015 		case EVSEL__CONFIG_TERM_PERCORE:
1016 			break;
1017 		case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1018 			attr->aux_output = term->val.aux_output ? 1 : 0;
1019 			break;
1020 		case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1021 			/* Already applied by auxtrace */
1022 			break;
1023 		case EVSEL__CONFIG_TERM_CFG_CHG:
1024 			break;
1025 		default:
1026 			break;
1027 		}
1028 	}
1029 
1030 	/* User explicitly set per-event callgraph, clear the old setting and reset. */
1031 	if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1032 		bool sample_address = false;
1033 
1034 		if (max_stack) {
1035 			param.max_stack = max_stack;
1036 			if (callgraph_buf == NULL)
1037 				callgraph_buf = "fp";
1038 		}
1039 
1040 		/* parse callgraph parameters */
1041 		if (callgraph_buf != NULL) {
1042 			if (!strcmp(callgraph_buf, "no")) {
1043 				param.enabled = false;
1044 				param.record_mode = CALLCHAIN_NONE;
1045 			} else {
1046 				param.enabled = true;
1047 				if (parse_callchain_record(callgraph_buf, &param)) {
1048 					pr_err("per-event callgraph setting for %s failed. "
1049 					       "Apply callgraph global setting for it\n",
1050 					       evsel->name);
1051 					return;
1052 				}
1053 				if (param.record_mode == CALLCHAIN_DWARF)
1054 					sample_address = true;
1055 			}
1056 		}
1057 		if (dump_size > 0) {
1058 			dump_size = round_up(dump_size, sizeof(u64));
1059 			param.dump_size = dump_size;
1060 		}
1061 
1062 		/* If global callgraph set, clear it */
1063 		if (callchain_param.enabled)
1064 			evsel__reset_callgraph(evsel, &callchain_param);
1065 
1066 		/* set perf-event callgraph */
1067 		if (param.enabled) {
1068 			if (sample_address) {
1069 				evsel__set_sample_bit(evsel, ADDR);
1070 				evsel__set_sample_bit(evsel, DATA_SRC);
1071 				evsel->core.attr.mmap_data = track;
1072 			}
1073 			evsel__config_callchain(evsel, opts, &param);
1074 		}
1075 	}
1076 }
1077 
1078 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1079 {
1080 	struct evsel_config_term *term, *found_term = NULL;
1081 
1082 	list_for_each_entry(term, &evsel->config_terms, list) {
1083 		if (term->type == type)
1084 			found_term = term;
1085 	}
1086 
1087 	return found_term;
1088 }
1089 
1090 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1091 {
1092 	evsel__set_sample_bit(evsel, WEIGHT);
1093 }
1094 
1095 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1096 				    struct perf_event_attr *attr __maybe_unused)
1097 {
1098 }
1099 
1100 static void evsel__set_default_freq_period(struct record_opts *opts,
1101 					   struct perf_event_attr *attr)
1102 {
1103 	if (opts->freq) {
1104 		attr->freq = 1;
1105 		attr->sample_freq = opts->freq;
1106 	} else {
1107 		attr->sample_period = opts->default_interval;
1108 	}
1109 }
1110 
1111 static bool evsel__is_offcpu_event(struct evsel *evsel)
1112 {
1113 	return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT);
1114 }
1115 
1116 /*
1117  * The enable_on_exec/disabled value strategy:
1118  *
1119  *  1) For any type of traced program:
1120  *    - all independent events and group leaders are disabled
1121  *    - all group members are enabled
1122  *
1123  *     Group members are ruled by group leaders. They need to
1124  *     be enabled, because the group scheduling relies on that.
1125  *
1126  *  2) For traced programs executed by perf:
1127  *     - all independent events and group leaders have
1128  *       enable_on_exec set
1129  *     - we don't specifically enable or disable any event during
1130  *       the record command
1131  *
1132  *     Independent events and group leaders are initially disabled
1133  *     and get enabled by exec. Group members are ruled by group
1134  *     leaders as stated in 1).
1135  *
1136  *  3) For traced programs attached by perf (pid/tid):
1137  *     - we specifically enable or disable all events during
1138  *       the record command
1139  *
1140  *     When attaching events to already running traced we
1141  *     enable/disable events specifically, as there's no
1142  *     initial traced exec call.
1143  */
1144 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1145 		   struct callchain_param *callchain)
1146 {
1147 	struct evsel *leader = evsel__leader(evsel);
1148 	struct perf_event_attr *attr = &evsel->core.attr;
1149 	int track = evsel->tracking;
1150 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1151 
1152 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1153 	attr->inherit	    = !opts->no_inherit;
1154 	attr->write_backward = opts->overwrite ? 1 : 0;
1155 	attr->read_format   = PERF_FORMAT_LOST;
1156 
1157 	evsel__set_sample_bit(evsel, IP);
1158 	evsel__set_sample_bit(evsel, TID);
1159 
1160 	if (evsel->sample_read) {
1161 		evsel__set_sample_bit(evsel, READ);
1162 
1163 		/*
1164 		 * We need ID even in case of single event, because
1165 		 * PERF_SAMPLE_READ process ID specific data.
1166 		 */
1167 		evsel__set_sample_id(evsel, false);
1168 
1169 		/*
1170 		 * Apply group format only if we belong to group
1171 		 * with more than one members.
1172 		 */
1173 		if (leader->core.nr_members > 1) {
1174 			attr->read_format |= PERF_FORMAT_GROUP;
1175 			attr->inherit = 0;
1176 		}
1177 	}
1178 
1179 	/*
1180 	 * We default some events to have a default interval. But keep
1181 	 * it a weak assumption overridable by the user.
1182 	 */
1183 	if ((evsel->is_libpfm_event && !attr->sample_period) ||
1184 	    (!evsel->is_libpfm_event && (!attr->sample_period ||
1185 					 opts->user_freq != UINT_MAX ||
1186 					 opts->user_interval != ULLONG_MAX)))
1187 		evsel__set_default_freq_period(opts, attr);
1188 
1189 	/*
1190 	 * If attr->freq was set (here or earlier), ask for period
1191 	 * to be sampled.
1192 	 */
1193 	if (attr->freq)
1194 		evsel__set_sample_bit(evsel, PERIOD);
1195 
1196 	if (opts->no_samples)
1197 		attr->sample_freq = 0;
1198 
1199 	if (opts->inherit_stat) {
1200 		evsel->core.attr.read_format |=
1201 			PERF_FORMAT_TOTAL_TIME_ENABLED |
1202 			PERF_FORMAT_TOTAL_TIME_RUNNING |
1203 			PERF_FORMAT_ID;
1204 		attr->inherit_stat = 1;
1205 	}
1206 
1207 	if (opts->sample_address) {
1208 		evsel__set_sample_bit(evsel, ADDR);
1209 		attr->mmap_data = track;
1210 	}
1211 
1212 	/*
1213 	 * We don't allow user space callchains for  function trace
1214 	 * event, due to issues with page faults while tracing page
1215 	 * fault handler and its overall trickiness nature.
1216 	 */
1217 	if (evsel__is_function_event(evsel))
1218 		evsel->core.attr.exclude_callchain_user = 1;
1219 
1220 	if (callchain && callchain->enabled && !evsel->no_aux_samples)
1221 		evsel__config_callchain(evsel, opts, callchain);
1222 
1223 	if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1224 	    !evsel__is_dummy_event(evsel)) {
1225 		attr->sample_regs_intr = opts->sample_intr_regs;
1226 		evsel__set_sample_bit(evsel, REGS_INTR);
1227 	}
1228 
1229 	if (opts->sample_user_regs && !evsel->no_aux_samples &&
1230 	    !evsel__is_dummy_event(evsel)) {
1231 		attr->sample_regs_user |= opts->sample_user_regs;
1232 		evsel__set_sample_bit(evsel, REGS_USER);
1233 	}
1234 
1235 	if (target__has_cpu(&opts->target) || opts->sample_cpu)
1236 		evsel__set_sample_bit(evsel, CPU);
1237 
1238 	/*
1239 	 * When the user explicitly disabled time don't force it here.
1240 	 */
1241 	if (opts->sample_time &&
1242 	    (!perf_missing_features.sample_id_all &&
1243 	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1244 	     opts->sample_time_set)))
1245 		evsel__set_sample_bit(evsel, TIME);
1246 
1247 	if (opts->raw_samples && !evsel->no_aux_samples) {
1248 		evsel__set_sample_bit(evsel, TIME);
1249 		evsel__set_sample_bit(evsel, RAW);
1250 		evsel__set_sample_bit(evsel, CPU);
1251 	}
1252 
1253 	if (opts->sample_address)
1254 		evsel__set_sample_bit(evsel, DATA_SRC);
1255 
1256 	if (opts->sample_phys_addr)
1257 		evsel__set_sample_bit(evsel, PHYS_ADDR);
1258 
1259 	if (opts->no_buffering) {
1260 		attr->watermark = 0;
1261 		attr->wakeup_events = 1;
1262 	}
1263 	if (opts->branch_stack && !evsel->no_aux_samples) {
1264 		evsel__set_sample_bit(evsel, BRANCH_STACK);
1265 		attr->branch_sample_type = opts->branch_stack;
1266 	}
1267 
1268 	if (opts->sample_weight)
1269 		arch_evsel__set_sample_weight(evsel);
1270 
1271 	attr->task     = track;
1272 	attr->mmap     = track;
1273 	attr->mmap2    = track && !perf_missing_features.mmap2;
1274 	attr->comm     = track;
1275 	attr->build_id = track && opts->build_id;
1276 
1277 	/*
1278 	 * ksymbol is tracked separately with text poke because it needs to be
1279 	 * system wide and enabled immediately.
1280 	 */
1281 	if (!opts->text_poke)
1282 		attr->ksymbol = track && !perf_missing_features.ksymbol;
1283 	attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1284 
1285 	if (opts->record_namespaces)
1286 		attr->namespaces  = track;
1287 
1288 	if (opts->record_cgroup) {
1289 		attr->cgroup = track && !perf_missing_features.cgroup;
1290 		evsel__set_sample_bit(evsel, CGROUP);
1291 	}
1292 
1293 	if (opts->sample_data_page_size)
1294 		evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1295 
1296 	if (opts->sample_code_page_size)
1297 		evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1298 
1299 	if (opts->record_switch_events)
1300 		attr->context_switch = track;
1301 
1302 	if (opts->sample_transaction)
1303 		evsel__set_sample_bit(evsel, TRANSACTION);
1304 
1305 	if (opts->running_time) {
1306 		evsel->core.attr.read_format |=
1307 			PERF_FORMAT_TOTAL_TIME_ENABLED |
1308 			PERF_FORMAT_TOTAL_TIME_RUNNING;
1309 	}
1310 
1311 	/*
1312 	 * XXX see the function comment above
1313 	 *
1314 	 * Disabling only independent events or group leaders,
1315 	 * keeping group members enabled.
1316 	 */
1317 	if (evsel__is_group_leader(evsel))
1318 		attr->disabled = 1;
1319 
1320 	/*
1321 	 * Setting enable_on_exec for independent events and
1322 	 * group leaders for traced executed by perf.
1323 	 */
1324 	if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1325 	    !opts->target.initial_delay)
1326 		attr->enable_on_exec = 1;
1327 
1328 	if (evsel->immediate) {
1329 		attr->disabled = 0;
1330 		attr->enable_on_exec = 0;
1331 	}
1332 
1333 	clockid = opts->clockid;
1334 	if (opts->use_clockid) {
1335 		attr->use_clockid = 1;
1336 		attr->clockid = opts->clockid;
1337 	}
1338 
1339 	if (evsel->precise_max)
1340 		attr->precise_ip = 3;
1341 
1342 	if (opts->all_user) {
1343 		attr->exclude_kernel = 1;
1344 		attr->exclude_user   = 0;
1345 	}
1346 
1347 	if (opts->all_kernel) {
1348 		attr->exclude_kernel = 0;
1349 		attr->exclude_user   = 1;
1350 	}
1351 
1352 	if (evsel->core.own_cpus || evsel->unit)
1353 		evsel->core.attr.read_format |= PERF_FORMAT_ID;
1354 
1355 	/*
1356 	 * Apply event specific term settings,
1357 	 * it overloads any global configuration.
1358 	 */
1359 	evsel__apply_config_terms(evsel, opts, track);
1360 
1361 	evsel->ignore_missing_thread = opts->ignore_missing_thread;
1362 
1363 	/* The --period option takes the precedence. */
1364 	if (opts->period_set) {
1365 		if (opts->period)
1366 			evsel__set_sample_bit(evsel, PERIOD);
1367 		else
1368 			evsel__reset_sample_bit(evsel, PERIOD);
1369 	}
1370 
1371 	/*
1372 	 * A dummy event never triggers any actual counter and therefore
1373 	 * cannot be used with branch_stack.
1374 	 *
1375 	 * For initial_delay, a dummy event is added implicitly.
1376 	 * The software event will trigger -EOPNOTSUPP error out,
1377 	 * if BRANCH_STACK bit is set.
1378 	 */
1379 	if (evsel__is_dummy_event(evsel))
1380 		evsel__reset_sample_bit(evsel, BRANCH_STACK);
1381 
1382 	if (evsel__is_offcpu_event(evsel))
1383 		evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1384 
1385 	arch__post_evsel_config(evsel, attr);
1386 }
1387 
1388 int evsel__set_filter(struct evsel *evsel, const char *filter)
1389 {
1390 	char *new_filter = strdup(filter);
1391 
1392 	if (new_filter != NULL) {
1393 		free(evsel->filter);
1394 		evsel->filter = new_filter;
1395 		return 0;
1396 	}
1397 
1398 	return -1;
1399 }
1400 
1401 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1402 {
1403 	char *new_filter;
1404 
1405 	if (evsel->filter == NULL)
1406 		return evsel__set_filter(evsel, filter);
1407 
1408 	if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1409 		free(evsel->filter);
1410 		evsel->filter = new_filter;
1411 		return 0;
1412 	}
1413 
1414 	return -1;
1415 }
1416 
1417 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1418 {
1419 	return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1420 }
1421 
1422 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1423 {
1424 	return evsel__append_filter(evsel, "%s,%s", filter);
1425 }
1426 
1427 /* Caller has to clear disabled after going through all CPUs. */
1428 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1429 {
1430 	return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1431 }
1432 
1433 int evsel__enable(struct evsel *evsel)
1434 {
1435 	int err = perf_evsel__enable(&evsel->core);
1436 
1437 	if (!err)
1438 		evsel->disabled = false;
1439 	return err;
1440 }
1441 
1442 /* Caller has to set disabled after going through all CPUs. */
1443 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1444 {
1445 	return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1446 }
1447 
1448 int evsel__disable(struct evsel *evsel)
1449 {
1450 	int err = perf_evsel__disable(&evsel->core);
1451 	/*
1452 	 * We mark it disabled here so that tools that disable a event can
1453 	 * ignore events after they disable it. I.e. the ring buffer may have
1454 	 * already a few more events queued up before the kernel got the stop
1455 	 * request.
1456 	 */
1457 	if (!err)
1458 		evsel->disabled = true;
1459 
1460 	return err;
1461 }
1462 
1463 void free_config_terms(struct list_head *config_terms)
1464 {
1465 	struct evsel_config_term *term, *h;
1466 
1467 	list_for_each_entry_safe(term, h, config_terms, list) {
1468 		list_del_init(&term->list);
1469 		if (term->free_str)
1470 			zfree(&term->val.str);
1471 		free(term);
1472 	}
1473 }
1474 
1475 static void evsel__free_config_terms(struct evsel *evsel)
1476 {
1477 	free_config_terms(&evsel->config_terms);
1478 }
1479 
1480 void evsel__exit(struct evsel *evsel)
1481 {
1482 	assert(list_empty(&evsel->core.node));
1483 	assert(evsel->evlist == NULL);
1484 	bpf_counter__destroy(evsel);
1485 	perf_bpf_filter__destroy(evsel);
1486 	evsel__free_counts(evsel);
1487 	perf_evsel__free_fd(&evsel->core);
1488 	perf_evsel__free_id(&evsel->core);
1489 	evsel__free_config_terms(evsel);
1490 	cgroup__put(evsel->cgrp);
1491 	perf_cpu_map__put(evsel->core.cpus);
1492 	perf_cpu_map__put(evsel->core.own_cpus);
1493 	perf_thread_map__put(evsel->core.threads);
1494 	zfree(&evsel->group_name);
1495 	zfree(&evsel->name);
1496 	zfree(&evsel->filter);
1497 	zfree(&evsel->pmu_name);
1498 	zfree(&evsel->group_pmu_name);
1499 	zfree(&evsel->unit);
1500 	zfree(&evsel->metric_id);
1501 	evsel__zero_per_pkg(evsel);
1502 	hashmap__free(evsel->per_pkg_mask);
1503 	evsel->per_pkg_mask = NULL;
1504 	zfree(&evsel->metric_events);
1505 	perf_evsel__object.fini(evsel);
1506 	if (evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME ||
1507 	    evsel__tool_event(evsel) == PERF_TOOL_USER_TIME)
1508 		xyarray__delete(evsel->start_times);
1509 }
1510 
1511 void evsel__delete(struct evsel *evsel)
1512 {
1513 	if (!evsel)
1514 		return;
1515 
1516 	evsel__exit(evsel);
1517 	free(evsel);
1518 }
1519 
1520 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1521 			   struct perf_counts_values *count)
1522 {
1523 	struct perf_counts_values tmp;
1524 
1525 	if (!evsel->prev_raw_counts)
1526 		return;
1527 
1528 	tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1529 	*perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1530 
1531 	count->val = count->val - tmp.val;
1532 	count->ena = count->ena - tmp.ena;
1533 	count->run = count->run - tmp.run;
1534 }
1535 
1536 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1537 {
1538 	struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1539 
1540 	return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1541 }
1542 
1543 static int evsel__read_retire_lat(struct evsel *evsel, int cpu_map_idx, int thread)
1544 {
1545 	return tpebs_set_evsel(evsel, cpu_map_idx, thread);
1546 }
1547 
1548 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1549 			     u64 val, u64 ena, u64 run, u64 lost)
1550 {
1551 	struct perf_counts_values *count;
1552 
1553 	count = perf_counts(counter->counts, cpu_map_idx, thread);
1554 
1555 	if (counter->retire_lat) {
1556 		evsel__read_retire_lat(counter, cpu_map_idx, thread);
1557 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1558 		return;
1559 	}
1560 
1561 	count->val    = val;
1562 	count->ena    = ena;
1563 	count->run    = run;
1564 	count->lost   = lost;
1565 
1566 	perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1567 }
1568 
1569 static bool evsel__group_has_tpebs(struct evsel *leader)
1570 {
1571 	struct evsel *evsel;
1572 
1573 	for_each_group_evsel(evsel, leader) {
1574 		if (evsel__is_retire_lat(evsel))
1575 			return true;
1576 	}
1577 	return false;
1578 }
1579 
1580 static u64 evsel__group_read_nr_members(struct evsel *leader)
1581 {
1582 	u64 nr = leader->core.nr_members;
1583 	struct evsel *evsel;
1584 
1585 	for_each_group_evsel(evsel, leader) {
1586 		if (evsel__is_retire_lat(evsel))
1587 			nr--;
1588 	}
1589 	return nr;
1590 }
1591 
1592 static u64 evsel__group_read_size(struct evsel *leader)
1593 {
1594 	u64 read_format = leader->core.attr.read_format;
1595 	int entry = sizeof(u64); /* value */
1596 	int size = 0;
1597 	int nr = 1;
1598 
1599 	if (!evsel__group_has_tpebs(leader))
1600 		return perf_evsel__read_size(&leader->core);
1601 
1602 	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1603 		size += sizeof(u64);
1604 
1605 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1606 		size += sizeof(u64);
1607 
1608 	if (read_format & PERF_FORMAT_ID)
1609 		entry += sizeof(u64);
1610 
1611 	if (read_format & PERF_FORMAT_LOST)
1612 		entry += sizeof(u64);
1613 
1614 	if (read_format & PERF_FORMAT_GROUP) {
1615 		nr = evsel__group_read_nr_members(leader);
1616 		size += sizeof(u64);
1617 	}
1618 
1619 	size += entry * nr;
1620 	return size;
1621 }
1622 
1623 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1624 {
1625 	u64 read_format = leader->core.attr.read_format;
1626 	struct sample_read_value *v;
1627 	u64 nr, ena = 0, run = 0, lost = 0;
1628 
1629 	nr = *data++;
1630 
1631 	if (nr != evsel__group_read_nr_members(leader))
1632 		return -EINVAL;
1633 
1634 	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1635 		ena = *data++;
1636 
1637 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1638 		run = *data++;
1639 
1640 	v = (void *)data;
1641 	sample_read_group__for_each(v, nr, read_format) {
1642 		struct evsel *counter;
1643 
1644 		counter = evlist__id2evsel(leader->evlist, v->id);
1645 		if (!counter)
1646 			return -EINVAL;
1647 
1648 		if (read_format & PERF_FORMAT_LOST)
1649 			lost = v->lost;
1650 
1651 		evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1652 	}
1653 
1654 	return 0;
1655 }
1656 
1657 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1658 {
1659 	struct perf_stat_evsel *ps = leader->stats;
1660 	u64 read_format = leader->core.attr.read_format;
1661 	int size = evsel__group_read_size(leader);
1662 	u64 *data = ps->group_data;
1663 
1664 	if (!(read_format & PERF_FORMAT_ID))
1665 		return -EINVAL;
1666 
1667 	if (!evsel__is_group_leader(leader))
1668 		return -EINVAL;
1669 
1670 	if (!data) {
1671 		data = zalloc(size);
1672 		if (!data)
1673 			return -ENOMEM;
1674 
1675 		ps->group_data = data;
1676 	}
1677 
1678 	if (FD(leader, cpu_map_idx, thread) < 0)
1679 		return -EINVAL;
1680 
1681 	if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1682 		return -errno;
1683 
1684 	return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1685 }
1686 
1687 static bool read_until_char(struct io *io, char e)
1688 {
1689 	int c;
1690 
1691 	do {
1692 		c = io__get_char(io);
1693 		if (c == -1)
1694 			return false;
1695 	} while (c != e);
1696 	return true;
1697 }
1698 
1699 static int read_stat_field(int fd, struct perf_cpu cpu, int field, __u64 *val)
1700 {
1701 	char buf[256];
1702 	struct io io;
1703 	int i;
1704 
1705 	io__init(&io, fd, buf, sizeof(buf));
1706 
1707 	/* Skip lines to relevant CPU. */
1708 	for (i = -1; i < cpu.cpu; i++) {
1709 		if (!read_until_char(&io, '\n'))
1710 			return -EINVAL;
1711 	}
1712 	/* Skip to "cpu". */
1713 	if (io__get_char(&io) != 'c') return -EINVAL;
1714 	if (io__get_char(&io) != 'p') return -EINVAL;
1715 	if (io__get_char(&io) != 'u') return -EINVAL;
1716 
1717 	/* Skip N of cpuN. */
1718 	if (!read_until_char(&io, ' '))
1719 		return -EINVAL;
1720 
1721 	i = 1;
1722 	while (true) {
1723 		if (io__get_dec(&io, val) != ' ')
1724 			break;
1725 		if (field == i)
1726 			return 0;
1727 		i++;
1728 	}
1729 	return -EINVAL;
1730 }
1731 
1732 static int read_pid_stat_field(int fd, int field, __u64 *val)
1733 {
1734 	char buf[256];
1735 	struct io io;
1736 	int c, i;
1737 
1738 	io__init(&io, fd, buf, sizeof(buf));
1739 	if (io__get_dec(&io, val) != ' ')
1740 		return -EINVAL;
1741 	if (field == 1)
1742 		return 0;
1743 
1744 	/* Skip comm. */
1745 	if (io__get_char(&io) != '(' || !read_until_char(&io, ')'))
1746 		return -EINVAL;
1747 	if (field == 2)
1748 		return -EINVAL; /* String can't be returned. */
1749 
1750 	/* Skip state */
1751 	if (io__get_char(&io) != ' ' || io__get_char(&io) == -1)
1752 		return -EINVAL;
1753 	if (field == 3)
1754 		return -EINVAL; /* String can't be returned. */
1755 
1756 	/* Loop over numeric fields*/
1757 	if (io__get_char(&io) != ' ')
1758 		return -EINVAL;
1759 
1760 	i = 4;
1761 	while (true) {
1762 		c = io__get_dec(&io, val);
1763 		if (c == -1)
1764 			return -EINVAL;
1765 		if (c == -2) {
1766 			/* Assume a -ve was read */
1767 			c = io__get_dec(&io, val);
1768 			*val *= -1;
1769 		}
1770 		if (c != ' ')
1771 			return -EINVAL;
1772 		if (field == i)
1773 			return 0;
1774 		i++;
1775 	}
1776 	return -EINVAL;
1777 }
1778 
1779 static int evsel__read_tool(struct evsel *evsel, int cpu_map_idx, int thread)
1780 {
1781 	__u64 *start_time, cur_time, delta_start;
1782 	int fd, err = 0;
1783 	struct perf_counts_values *count;
1784 	bool adjust = false;
1785 
1786 	count = perf_counts(evsel->counts, cpu_map_idx, thread);
1787 
1788 	switch (evsel__tool_event(evsel)) {
1789 	case PERF_TOOL_DURATION_TIME:
1790 		/*
1791 		 * Pretend duration_time is only on the first CPU and thread, or
1792 		 * else aggregation will scale duration_time by the number of
1793 		 * CPUs/threads.
1794 		 */
1795 		start_time = &evsel->start_time;
1796 		if (cpu_map_idx == 0 && thread == 0)
1797 			cur_time = rdclock();
1798 		else
1799 			cur_time = *start_time;
1800 		break;
1801 	case PERF_TOOL_USER_TIME:
1802 	case PERF_TOOL_SYSTEM_TIME: {
1803 		bool system = evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME;
1804 
1805 		start_time = xyarray__entry(evsel->start_times, cpu_map_idx, thread);
1806 		fd = FD(evsel, cpu_map_idx, thread);
1807 		lseek(fd, SEEK_SET, 0);
1808 		if (evsel->pid_stat) {
1809 			/* The event exists solely on 1 CPU. */
1810 			if (cpu_map_idx == 0)
1811 				err = read_pid_stat_field(fd, system ? 15 : 14, &cur_time);
1812 			else
1813 				cur_time = 0;
1814 		} else {
1815 			/* The event is for all threads. */
1816 			if (thread == 0) {
1817 				struct perf_cpu cpu = perf_cpu_map__cpu(evsel->core.cpus,
1818 									cpu_map_idx);
1819 
1820 				err = read_stat_field(fd, cpu, system ? 3 : 1, &cur_time);
1821 			} else {
1822 				cur_time = 0;
1823 			}
1824 		}
1825 		adjust = true;
1826 		break;
1827 	}
1828 	case PERF_TOOL_NONE:
1829 	case PERF_TOOL_MAX:
1830 	default:
1831 		err = -EINVAL;
1832 	}
1833 	if (err)
1834 		return err;
1835 
1836 	delta_start = cur_time - *start_time;
1837 	if (adjust) {
1838 		__u64 ticks_per_sec = sysconf(_SC_CLK_TCK);
1839 
1840 		delta_start *= 1000000000 / ticks_per_sec;
1841 	}
1842 	count->val    = delta_start;
1843 	count->ena    = count->run = delta_start;
1844 	count->lost   = 0;
1845 	return 0;
1846 }
1847 
1848 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1849 {
1850 	if (evsel__is_tool(evsel))
1851 		return evsel__read_tool(evsel, cpu_map_idx, thread);
1852 
1853 	if (evsel__is_retire_lat(evsel))
1854 		return evsel__read_retire_lat(evsel, cpu_map_idx, thread);
1855 
1856 	if (evsel->core.attr.read_format & PERF_FORMAT_GROUP)
1857 		return evsel__read_group(evsel, cpu_map_idx, thread);
1858 
1859 	return evsel__read_one(evsel, cpu_map_idx, thread);
1860 }
1861 
1862 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1863 {
1864 	struct perf_counts_values count;
1865 	size_t nv = scale ? 3 : 1;
1866 
1867 	if (FD(evsel, cpu_map_idx, thread) < 0)
1868 		return -EINVAL;
1869 
1870 	if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1871 		return -ENOMEM;
1872 
1873 	if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1874 		return -errno;
1875 
1876 	evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1877 	perf_counts_values__scale(&count, scale, NULL);
1878 	*perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1879 	return 0;
1880 }
1881 
1882 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1883 				  int cpu_map_idx)
1884 {
1885 	struct perf_cpu cpu;
1886 
1887 	cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1888 	return perf_cpu_map__idx(other->core.cpus, cpu);
1889 }
1890 
1891 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1892 {
1893 	struct evsel *leader = evsel__leader(evsel);
1894 
1895 	if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1896 	    (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1897 		return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1898 	}
1899 
1900 	return cpu_map_idx;
1901 }
1902 
1903 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1904 {
1905 	struct evsel *leader = evsel__leader(evsel);
1906 	int fd;
1907 
1908 	if (evsel__is_group_leader(evsel))
1909 		return -1;
1910 
1911 	/*
1912 	 * Leader must be already processed/open,
1913 	 * if not it's a bug.
1914 	 */
1915 	BUG_ON(!leader->core.fd);
1916 
1917 	cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1918 	if (cpu_map_idx == -1)
1919 		return -1;
1920 
1921 	fd = FD(leader, cpu_map_idx, thread);
1922 	BUG_ON(fd == -1 && !leader->skippable);
1923 
1924 	/*
1925 	 * When the leader has been skipped, return -2 to distinguish from no
1926 	 * group leader case.
1927 	 */
1928 	return fd == -1 ? -2 : fd;
1929 }
1930 
1931 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1932 {
1933 	for (int cpu = 0; cpu < nr_cpus; cpu++)
1934 		for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1935 			FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1936 }
1937 
1938 static int update_fds(struct evsel *evsel,
1939 		      int nr_cpus, int cpu_map_idx,
1940 		      int nr_threads, int thread_idx)
1941 {
1942 	struct evsel *pos;
1943 
1944 	if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1945 		return -EINVAL;
1946 
1947 	evlist__for_each_entry(evsel->evlist, pos) {
1948 		nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1949 
1950 		evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1951 
1952 		/*
1953 		 * Since fds for next evsel has not been created,
1954 		 * there is no need to iterate whole event list.
1955 		 */
1956 		if (pos == evsel)
1957 			break;
1958 	}
1959 	return 0;
1960 }
1961 
1962 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1963 					 int nr_cpus, int cpu_map_idx,
1964 					 struct perf_thread_map *threads,
1965 					 int thread, int err)
1966 {
1967 	pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1968 
1969 	if (!evsel->ignore_missing_thread)
1970 		return false;
1971 
1972 	/* The system wide setup does not work with threads. */
1973 	if (evsel->core.system_wide)
1974 		return false;
1975 
1976 	/* The -ESRCH is perf event syscall errno for pid's not found. */
1977 	if (err != -ESRCH)
1978 		return false;
1979 
1980 	/* If there's only one thread, let it fail. */
1981 	if (threads->nr == 1)
1982 		return false;
1983 
1984 	/*
1985 	 * We should remove fd for missing_thread first
1986 	 * because thread_map__remove() will decrease threads->nr.
1987 	 */
1988 	if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1989 		return false;
1990 
1991 	if (thread_map__remove(threads, thread))
1992 		return false;
1993 
1994 	pr_warning("WARNING: Ignored open failure for pid %d\n",
1995 		   ignore_pid);
1996 	return true;
1997 }
1998 
1999 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
2000 				void *priv __maybe_unused)
2001 {
2002 	return fprintf(fp, "  %-32s %s\n", name, val);
2003 }
2004 
2005 static void display_attr(struct perf_event_attr *attr)
2006 {
2007 	if (verbose >= 2 || debug_peo_args) {
2008 		fprintf(stderr, "%.60s\n", graph_dotted_line);
2009 		fprintf(stderr, "perf_event_attr:\n");
2010 		perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
2011 		fprintf(stderr, "%.60s\n", graph_dotted_line);
2012 	}
2013 }
2014 
2015 bool evsel__precise_ip_fallback(struct evsel *evsel)
2016 {
2017 	/* Do not try less precise if not requested. */
2018 	if (!evsel->precise_max)
2019 		return false;
2020 
2021 	/*
2022 	 * We tried all the precise_ip values, and it's
2023 	 * still failing, so leave it to standard fallback.
2024 	 */
2025 	if (!evsel->core.attr.precise_ip) {
2026 		evsel->core.attr.precise_ip = evsel->precise_ip_original;
2027 		return false;
2028 	}
2029 
2030 	if (!evsel->precise_ip_original)
2031 		evsel->precise_ip_original = evsel->core.attr.precise_ip;
2032 
2033 	evsel->core.attr.precise_ip--;
2034 	pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
2035 	display_attr(&evsel->core.attr);
2036 	return true;
2037 }
2038 
2039 static struct perf_cpu_map *empty_cpu_map;
2040 static struct perf_thread_map *empty_thread_map;
2041 
2042 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
2043 		struct perf_thread_map *threads)
2044 {
2045 	int nthreads = perf_thread_map__nr(threads);
2046 
2047 	if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
2048 	    (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
2049 		return -EINVAL;
2050 
2051 	if (cpus == NULL) {
2052 		if (empty_cpu_map == NULL) {
2053 			empty_cpu_map = perf_cpu_map__new_any_cpu();
2054 			if (empty_cpu_map == NULL)
2055 				return -ENOMEM;
2056 		}
2057 
2058 		cpus = empty_cpu_map;
2059 	}
2060 
2061 	if (threads == NULL) {
2062 		if (empty_thread_map == NULL) {
2063 			empty_thread_map = thread_map__new_by_tid(-1);
2064 			if (empty_thread_map == NULL)
2065 				return -ENOMEM;
2066 		}
2067 
2068 		threads = empty_thread_map;
2069 	}
2070 
2071 	if (evsel->core.fd == NULL &&
2072 	    perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
2073 		return -ENOMEM;
2074 
2075 	if ((evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME ||
2076 	     evsel__tool_event(evsel) == PERF_TOOL_USER_TIME) &&
2077 	    !evsel->start_times) {
2078 		evsel->start_times = xyarray__new(perf_cpu_map__nr(cpus), nthreads, sizeof(__u64));
2079 		if (!evsel->start_times)
2080 			return -ENOMEM;
2081 	}
2082 
2083 	evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
2084 	if (evsel->cgrp)
2085 		evsel->open_flags |= PERF_FLAG_PID_CGROUP;
2086 
2087 	return 0;
2088 }
2089 
2090 static void evsel__disable_missing_features(struct evsel *evsel)
2091 {
2092 	if (perf_missing_features.branch_counters)
2093 		evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_COUNTERS;
2094 	if (perf_missing_features.read_lost)
2095 		evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
2096 	if (perf_missing_features.weight_struct) {
2097 		evsel__set_sample_bit(evsel, WEIGHT);
2098 		evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
2099 	}
2100 	if (perf_missing_features.clockid_wrong)
2101 		evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
2102 	if (perf_missing_features.clockid) {
2103 		evsel->core.attr.use_clockid = 0;
2104 		evsel->core.attr.clockid = 0;
2105 	}
2106 	if (perf_missing_features.cloexec)
2107 		evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
2108 	if (perf_missing_features.mmap2)
2109 		evsel->core.attr.mmap2 = 0;
2110 	if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
2111 		evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
2112 	if (perf_missing_features.lbr_flags)
2113 		evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
2114 				     PERF_SAMPLE_BRANCH_NO_CYCLES);
2115 	if (perf_missing_features.group_read && evsel->core.attr.inherit)
2116 		evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
2117 	if (perf_missing_features.ksymbol)
2118 		evsel->core.attr.ksymbol = 0;
2119 	if (perf_missing_features.bpf)
2120 		evsel->core.attr.bpf_event = 0;
2121 	if (perf_missing_features.branch_hw_idx)
2122 		evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
2123 	if (perf_missing_features.sample_id_all)
2124 		evsel->core.attr.sample_id_all = 0;
2125 }
2126 
2127 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
2128 			struct perf_thread_map *threads)
2129 {
2130 	int err;
2131 
2132 	err = __evsel__prepare_open(evsel, cpus, threads);
2133 	if (err)
2134 		return err;
2135 
2136 	evsel__disable_missing_features(evsel);
2137 
2138 	return err;
2139 }
2140 
2141 bool evsel__detect_missing_features(struct evsel *evsel)
2142 {
2143 	/*
2144 	 * Must probe features in the order they were added to the
2145 	 * perf_event_attr interface.
2146 	 */
2147 	if (!perf_missing_features.branch_counters &&
2148 	    (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS)) {
2149 		perf_missing_features.branch_counters = true;
2150 		pr_debug2("switching off branch counters support\n");
2151 		return true;
2152 	} else if (!perf_missing_features.read_lost &&
2153 	    (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
2154 		perf_missing_features.read_lost = true;
2155 		pr_debug2("switching off PERF_FORMAT_LOST support\n");
2156 		return true;
2157 	} else if (!perf_missing_features.weight_struct &&
2158 	    (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
2159 		perf_missing_features.weight_struct = true;
2160 		pr_debug2("switching off weight struct support\n");
2161 		return true;
2162 	} else if (!perf_missing_features.code_page_size &&
2163 	    (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
2164 		perf_missing_features.code_page_size = true;
2165 		pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
2166 		return false;
2167 	} else if (!perf_missing_features.data_page_size &&
2168 	    (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
2169 		perf_missing_features.data_page_size = true;
2170 		pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
2171 		return false;
2172 	} else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
2173 		perf_missing_features.cgroup = true;
2174 		pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
2175 		return false;
2176 	} else if (!perf_missing_features.branch_hw_idx &&
2177 	    (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
2178 		perf_missing_features.branch_hw_idx = true;
2179 		pr_debug2("switching off branch HW index support\n");
2180 		return true;
2181 	} else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
2182 		perf_missing_features.aux_output = true;
2183 		pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
2184 		return false;
2185 	} else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
2186 		perf_missing_features.bpf = true;
2187 		pr_debug2_peo("switching off bpf_event\n");
2188 		return true;
2189 	} else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
2190 		perf_missing_features.ksymbol = true;
2191 		pr_debug2_peo("switching off ksymbol\n");
2192 		return true;
2193 	} else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
2194 		perf_missing_features.write_backward = true;
2195 		pr_debug2_peo("switching off write_backward\n");
2196 		return false;
2197 	} else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
2198 		perf_missing_features.clockid_wrong = true;
2199 		pr_debug2_peo("switching off clockid\n");
2200 		return true;
2201 	} else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
2202 		perf_missing_features.clockid = true;
2203 		pr_debug2_peo("switching off use_clockid\n");
2204 		return true;
2205 	} else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
2206 		perf_missing_features.cloexec = true;
2207 		pr_debug2_peo("switching off cloexec flag\n");
2208 		return true;
2209 	} else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
2210 		perf_missing_features.mmap2 = true;
2211 		pr_debug2_peo("switching off mmap2\n");
2212 		return true;
2213 	} else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
2214 		if (evsel->pmu == NULL)
2215 			evsel->pmu = evsel__find_pmu(evsel);
2216 
2217 		if (evsel->pmu)
2218 			evsel->pmu->missing_features.exclude_guest = true;
2219 		else {
2220 			/* we cannot find PMU, disable attrs now */
2221 			evsel->core.attr.exclude_host = false;
2222 			evsel->core.attr.exclude_guest = false;
2223 		}
2224 
2225 		if (evsel->exclude_GH) {
2226 			pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
2227 			return false;
2228 		}
2229 		if (!perf_missing_features.exclude_guest) {
2230 			perf_missing_features.exclude_guest = true;
2231 			pr_debug2_peo("switching off exclude_guest, exclude_host\n");
2232 		}
2233 		return true;
2234 	} else if (!perf_missing_features.sample_id_all) {
2235 		perf_missing_features.sample_id_all = true;
2236 		pr_debug2_peo("switching off sample_id_all\n");
2237 		return true;
2238 	} else if (!perf_missing_features.lbr_flags &&
2239 			(evsel->core.attr.branch_sample_type &
2240 			 (PERF_SAMPLE_BRANCH_NO_CYCLES |
2241 			  PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2242 		perf_missing_features.lbr_flags = true;
2243 		pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2244 		return true;
2245 	} else if (!perf_missing_features.group_read &&
2246 		    evsel->core.attr.inherit &&
2247 		   (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2248 		   evsel__is_group_leader(evsel)) {
2249 		perf_missing_features.group_read = true;
2250 		pr_debug2_peo("switching off group read\n");
2251 		return true;
2252 	} else {
2253 		return false;
2254 	}
2255 }
2256 
2257 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2258 		struct perf_thread_map *threads,
2259 		int start_cpu_map_idx, int end_cpu_map_idx)
2260 {
2261 	int idx, thread, nthreads;
2262 	int pid = -1, err, old_errno;
2263 	enum rlimit_action set_rlimit = NO_CHANGE;
2264 
2265 	if (evsel__tool_event(evsel) == PERF_TOOL_DURATION_TIME) {
2266 		if (evsel->core.attr.sample_period) /* no sampling */
2267 			return -EINVAL;
2268 		evsel->start_time = rdclock();
2269 		return 0;
2270 	}
2271 
2272 	if (evsel__is_retire_lat(evsel))
2273 		return tpebs_start(evsel->evlist);
2274 
2275 	err = __evsel__prepare_open(evsel, cpus, threads);
2276 	if (err)
2277 		return err;
2278 
2279 	if (cpus == NULL)
2280 		cpus = empty_cpu_map;
2281 
2282 	if (threads == NULL)
2283 		threads = empty_thread_map;
2284 
2285 	nthreads = perf_thread_map__nr(threads);
2286 
2287 	if (evsel->cgrp)
2288 		pid = evsel->cgrp->fd;
2289 
2290 fallback_missing_features:
2291 	evsel__disable_missing_features(evsel);
2292 
2293 	pr_debug3("Opening: %s\n", evsel__name(evsel));
2294 	display_attr(&evsel->core.attr);
2295 
2296 	for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2297 
2298 		for (thread = 0; thread < nthreads; thread++) {
2299 			int fd, group_fd;
2300 retry_open:
2301 			if (thread >= nthreads)
2302 				break;
2303 
2304 			if (!evsel->cgrp && !evsel->core.system_wide)
2305 				pid = perf_thread_map__pid(threads, thread);
2306 
2307 			if (evsel__tool_event(evsel) == PERF_TOOL_USER_TIME ||
2308 			    evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME) {
2309 				bool system = evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME;
2310 				__u64 *start_time = NULL;
2311 
2312 				if (evsel->core.attr.sample_period) {
2313 					/* no sampling */
2314 					err = -EINVAL;
2315 					goto out_close;
2316 				}
2317 				if (pid > -1) {
2318 					char buf[64];
2319 
2320 					snprintf(buf, sizeof(buf), "/proc/%d/stat", pid);
2321 					fd = open(buf, O_RDONLY);
2322 					evsel->pid_stat = true;
2323 				} else {
2324 					fd = open("/proc/stat", O_RDONLY);
2325 				}
2326 				FD(evsel, idx, thread) = fd;
2327 				if (fd < 0) {
2328 					err = -errno;
2329 					goto out_close;
2330 				}
2331 				start_time = xyarray__entry(evsel->start_times, idx, thread);
2332 				if (pid > -1) {
2333 					err = read_pid_stat_field(fd, system ? 15 : 14,
2334 								  start_time);
2335 				} else {
2336 					struct perf_cpu cpu;
2337 
2338 					cpu = perf_cpu_map__cpu(evsel->core.cpus, idx);
2339 					err = read_stat_field(fd, cpu, system ? 3 : 1,
2340 							      start_time);
2341 				}
2342 				if (err)
2343 					goto out_close;
2344 				continue;
2345 			}
2346 
2347 			group_fd = get_group_fd(evsel, idx, thread);
2348 
2349 			if (group_fd == -2) {
2350 				pr_debug("broken group leader for %s\n", evsel->name);
2351 				err = -EINVAL;
2352 				goto out_close;
2353 			}
2354 
2355 			test_attr__ready();
2356 
2357 			/* Debug message used by test scripts */
2358 			pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2359 				pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2360 
2361 			fd = sys_perf_event_open(&evsel->core.attr, pid,
2362 						perf_cpu_map__cpu(cpus, idx).cpu,
2363 						group_fd, evsel->open_flags);
2364 
2365 			FD(evsel, idx, thread) = fd;
2366 
2367 			if (fd < 0) {
2368 				err = -errno;
2369 
2370 				pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2371 					  err);
2372 				goto try_fallback;
2373 			}
2374 
2375 			bpf_counter__install_pe(evsel, idx, fd);
2376 
2377 			if (unlikely(test_attr__enabled)) {
2378 				test_attr__open(&evsel->core.attr, pid,
2379 						perf_cpu_map__cpu(cpus, idx),
2380 						fd, group_fd, evsel->open_flags);
2381 			}
2382 
2383 			/* Debug message used by test scripts */
2384 			pr_debug2_peo(" = %d\n", fd);
2385 
2386 			if (evsel->bpf_fd >= 0) {
2387 				int evt_fd = fd;
2388 				int bpf_fd = evsel->bpf_fd;
2389 
2390 				err = ioctl(evt_fd,
2391 					    PERF_EVENT_IOC_SET_BPF,
2392 					    bpf_fd);
2393 				if (err && errno != EEXIST) {
2394 					pr_err("failed to attach bpf fd %d: %s\n",
2395 					       bpf_fd, strerror(errno));
2396 					err = -EINVAL;
2397 					goto out_close;
2398 				}
2399 			}
2400 
2401 			set_rlimit = NO_CHANGE;
2402 
2403 			/*
2404 			 * If we succeeded but had to kill clockid, fail and
2405 			 * have evsel__open_strerror() print us a nice error.
2406 			 */
2407 			if (perf_missing_features.clockid ||
2408 			    perf_missing_features.clockid_wrong) {
2409 				err = -EINVAL;
2410 				goto out_close;
2411 			}
2412 		}
2413 	}
2414 
2415 	return 0;
2416 
2417 try_fallback:
2418 	if (evsel__precise_ip_fallback(evsel))
2419 		goto retry_open;
2420 
2421 	if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2422 					 idx, threads, thread, err)) {
2423 		/* We just removed 1 thread, so lower the upper nthreads limit. */
2424 		nthreads--;
2425 
2426 		/* ... and pretend like nothing have happened. */
2427 		err = 0;
2428 		goto retry_open;
2429 	}
2430 	/*
2431 	 * perf stat needs between 5 and 22 fds per CPU. When we run out
2432 	 * of them try to increase the limits.
2433 	 */
2434 	if (err == -EMFILE && rlimit__increase_nofile(&set_rlimit))
2435 		goto retry_open;
2436 
2437 	if (err != -EINVAL || idx > 0 || thread > 0)
2438 		goto out_close;
2439 
2440 	if (evsel__detect_missing_features(evsel))
2441 		goto fallback_missing_features;
2442 out_close:
2443 	if (err)
2444 		threads->err_thread = thread;
2445 
2446 	old_errno = errno;
2447 	do {
2448 		while (--thread >= 0) {
2449 			if (FD(evsel, idx, thread) >= 0)
2450 				close(FD(evsel, idx, thread));
2451 			FD(evsel, idx, thread) = -1;
2452 		}
2453 		thread = nthreads;
2454 	} while (--idx >= 0);
2455 	errno = old_errno;
2456 	return err;
2457 }
2458 
2459 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2460 		struct perf_thread_map *threads)
2461 {
2462 	return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2463 }
2464 
2465 void evsel__close(struct evsel *evsel)
2466 {
2467 	if (evsel__is_retire_lat(evsel))
2468 		tpebs_delete();
2469 	perf_evsel__close(&evsel->core);
2470 	perf_evsel__free_id(&evsel->core);
2471 }
2472 
2473 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2474 {
2475 	if (cpu_map_idx == -1)
2476 		return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2477 
2478 	return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2479 }
2480 
2481 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2482 {
2483 	return evsel__open(evsel, NULL, threads);
2484 }
2485 
2486 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2487 				       const union perf_event *event,
2488 				       struct perf_sample *sample)
2489 {
2490 	u64 type = evsel->core.attr.sample_type;
2491 	const __u64 *array = event->sample.array;
2492 	bool swapped = evsel->needs_swap;
2493 	union u64_swap u;
2494 
2495 	array += ((event->header.size -
2496 		   sizeof(event->header)) / sizeof(u64)) - 1;
2497 
2498 	if (type & PERF_SAMPLE_IDENTIFIER) {
2499 		sample->id = *array;
2500 		array--;
2501 	}
2502 
2503 	if (type & PERF_SAMPLE_CPU) {
2504 		u.val64 = *array;
2505 		if (swapped) {
2506 			/* undo swap of u64, then swap on individual u32s */
2507 			u.val64 = bswap_64(u.val64);
2508 			u.val32[0] = bswap_32(u.val32[0]);
2509 		}
2510 
2511 		sample->cpu = u.val32[0];
2512 		array--;
2513 	}
2514 
2515 	if (type & PERF_SAMPLE_STREAM_ID) {
2516 		sample->stream_id = *array;
2517 		array--;
2518 	}
2519 
2520 	if (type & PERF_SAMPLE_ID) {
2521 		sample->id = *array;
2522 		array--;
2523 	}
2524 
2525 	if (type & PERF_SAMPLE_TIME) {
2526 		sample->time = *array;
2527 		array--;
2528 	}
2529 
2530 	if (type & PERF_SAMPLE_TID) {
2531 		u.val64 = *array;
2532 		if (swapped) {
2533 			/* undo swap of u64, then swap on individual u32s */
2534 			u.val64 = bswap_64(u.val64);
2535 			u.val32[0] = bswap_32(u.val32[0]);
2536 			u.val32[1] = bswap_32(u.val32[1]);
2537 		}
2538 
2539 		sample->pid = u.val32[0];
2540 		sample->tid = u.val32[1];
2541 		array--;
2542 	}
2543 
2544 	return 0;
2545 }
2546 
2547 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2548 			    u64 size)
2549 {
2550 	return size > max_size || offset + size > endp;
2551 }
2552 
2553 #define OVERFLOW_CHECK(offset, size, max_size)				\
2554 	do {								\
2555 		if (overflow(endp, (max_size), (offset), (size)))	\
2556 			return -EFAULT;					\
2557 	} while (0)
2558 
2559 #define OVERFLOW_CHECK_u64(offset) \
2560 	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2561 
2562 static int
2563 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2564 {
2565 	/*
2566 	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2567 	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2568 	 * check the format does not go past the end of the event.
2569 	 */
2570 	if (sample_size + sizeof(event->header) > event->header.size)
2571 		return -EFAULT;
2572 
2573 	return 0;
2574 }
2575 
2576 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2577 					  const __u64 *array,
2578 					  u64 type __maybe_unused)
2579 {
2580 	data->weight = *array;
2581 }
2582 
2583 u64 evsel__bitfield_swap_branch_flags(u64 value)
2584 {
2585 	u64 new_val = 0;
2586 
2587 	/*
2588 	 * branch_flags
2589 	 * union {
2590 	 * 	u64 values;
2591 	 * 	struct {
2592 	 * 		mispred:1	//target mispredicted
2593 	 * 		predicted:1	//target predicted
2594 	 * 		in_tx:1		//in transaction
2595 	 * 		abort:1		//transaction abort
2596 	 * 		cycles:16	//cycle count to last branch
2597 	 * 		type:4		//branch type
2598 	 * 		spec:2		//branch speculation info
2599 	 * 		new_type:4	//additional branch type
2600 	 * 		priv:3		//privilege level
2601 	 * 		reserved:31
2602 	 * 	}
2603 	 * }
2604 	 *
2605 	 * Avoid bswap64() the entire branch_flag.value,
2606 	 * as it has variable bit-field sizes. Instead the
2607 	 * macro takes the bit-field position/size,
2608 	 * swaps it based on the host endianness.
2609 	 */
2610 	if (host_is_bigendian()) {
2611 		new_val = bitfield_swap(value, 0, 1);
2612 		new_val |= bitfield_swap(value, 1, 1);
2613 		new_val |= bitfield_swap(value, 2, 1);
2614 		new_val |= bitfield_swap(value, 3, 1);
2615 		new_val |= bitfield_swap(value, 4, 16);
2616 		new_val |= bitfield_swap(value, 20, 4);
2617 		new_val |= bitfield_swap(value, 24, 2);
2618 		new_val |= bitfield_swap(value, 26, 4);
2619 		new_val |= bitfield_swap(value, 30, 3);
2620 		new_val |= bitfield_swap(value, 33, 31);
2621 	} else {
2622 		new_val = bitfield_swap(value, 63, 1);
2623 		new_val |= bitfield_swap(value, 62, 1);
2624 		new_val |= bitfield_swap(value, 61, 1);
2625 		new_val |= bitfield_swap(value, 60, 1);
2626 		new_val |= bitfield_swap(value, 44, 16);
2627 		new_val |= bitfield_swap(value, 40, 4);
2628 		new_val |= bitfield_swap(value, 38, 2);
2629 		new_val |= bitfield_swap(value, 34, 4);
2630 		new_val |= bitfield_swap(value, 31, 3);
2631 		new_val |= bitfield_swap(value, 0, 31);
2632 	}
2633 
2634 	return new_val;
2635 }
2636 
2637 static inline bool evsel__has_branch_counters(const struct evsel *evsel)
2638 {
2639 	struct evsel *leader = evsel__leader(evsel);
2640 
2641 	/* The branch counters feature only supports group */
2642 	if (!leader || !evsel->evlist)
2643 		return false;
2644 
2645 	if (evsel->evlist->nr_br_cntr < 0)
2646 		evlist__update_br_cntr(evsel->evlist);
2647 
2648 	if (leader->br_cntr_nr > 0)
2649 		return true;
2650 
2651 	return false;
2652 }
2653 
2654 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2655 			struct perf_sample *data)
2656 {
2657 	u64 type = evsel->core.attr.sample_type;
2658 	bool swapped = evsel->needs_swap;
2659 	const __u64 *array;
2660 	u16 max_size = event->header.size;
2661 	const void *endp = (void *)event + max_size;
2662 	u64 sz;
2663 
2664 	/*
2665 	 * used for cross-endian analysis. See git commit 65014ab3
2666 	 * for why this goofiness is needed.
2667 	 */
2668 	union u64_swap u;
2669 
2670 	memset(data, 0, sizeof(*data));
2671 	data->cpu = data->pid = data->tid = -1;
2672 	data->stream_id = data->id = data->time = -1ULL;
2673 	data->period = evsel->core.attr.sample_period;
2674 	data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2675 	data->misc    = event->header.misc;
2676 	data->data_src = PERF_MEM_DATA_SRC_NONE;
2677 	data->vcpu = -1;
2678 
2679 	if (event->header.type != PERF_RECORD_SAMPLE) {
2680 		if (!evsel->core.attr.sample_id_all)
2681 			return 0;
2682 		return perf_evsel__parse_id_sample(evsel, event, data);
2683 	}
2684 
2685 	array = event->sample.array;
2686 
2687 	if (perf_event__check_size(event, evsel->sample_size))
2688 		return -EFAULT;
2689 
2690 	if (type & PERF_SAMPLE_IDENTIFIER) {
2691 		data->id = *array;
2692 		array++;
2693 	}
2694 
2695 	if (type & PERF_SAMPLE_IP) {
2696 		data->ip = *array;
2697 		array++;
2698 	}
2699 
2700 	if (type & PERF_SAMPLE_TID) {
2701 		u.val64 = *array;
2702 		if (swapped) {
2703 			/* undo swap of u64, then swap on individual u32s */
2704 			u.val64 = bswap_64(u.val64);
2705 			u.val32[0] = bswap_32(u.val32[0]);
2706 			u.val32[1] = bswap_32(u.val32[1]);
2707 		}
2708 
2709 		data->pid = u.val32[0];
2710 		data->tid = u.val32[1];
2711 		array++;
2712 	}
2713 
2714 	if (type & PERF_SAMPLE_TIME) {
2715 		data->time = *array;
2716 		array++;
2717 	}
2718 
2719 	if (type & PERF_SAMPLE_ADDR) {
2720 		data->addr = *array;
2721 		array++;
2722 	}
2723 
2724 	if (type & PERF_SAMPLE_ID) {
2725 		data->id = *array;
2726 		array++;
2727 	}
2728 
2729 	if (type & PERF_SAMPLE_STREAM_ID) {
2730 		data->stream_id = *array;
2731 		array++;
2732 	}
2733 
2734 	if (type & PERF_SAMPLE_CPU) {
2735 
2736 		u.val64 = *array;
2737 		if (swapped) {
2738 			/* undo swap of u64, then swap on individual u32s */
2739 			u.val64 = bswap_64(u.val64);
2740 			u.val32[0] = bswap_32(u.val32[0]);
2741 		}
2742 
2743 		data->cpu = u.val32[0];
2744 		array++;
2745 	}
2746 
2747 	if (type & PERF_SAMPLE_PERIOD) {
2748 		data->period = *array;
2749 		array++;
2750 	}
2751 
2752 	if (type & PERF_SAMPLE_READ) {
2753 		u64 read_format = evsel->core.attr.read_format;
2754 
2755 		OVERFLOW_CHECK_u64(array);
2756 		if (read_format & PERF_FORMAT_GROUP)
2757 			data->read.group.nr = *array;
2758 		else
2759 			data->read.one.value = *array;
2760 
2761 		array++;
2762 
2763 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2764 			OVERFLOW_CHECK_u64(array);
2765 			data->read.time_enabled = *array;
2766 			array++;
2767 		}
2768 
2769 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2770 			OVERFLOW_CHECK_u64(array);
2771 			data->read.time_running = *array;
2772 			array++;
2773 		}
2774 
2775 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2776 		if (read_format & PERF_FORMAT_GROUP) {
2777 			const u64 max_group_nr = UINT64_MAX /
2778 					sizeof(struct sample_read_value);
2779 
2780 			if (data->read.group.nr > max_group_nr)
2781 				return -EFAULT;
2782 
2783 			sz = data->read.group.nr * sample_read_value_size(read_format);
2784 			OVERFLOW_CHECK(array, sz, max_size);
2785 			data->read.group.values =
2786 					(struct sample_read_value *)array;
2787 			array = (void *)array + sz;
2788 		} else {
2789 			OVERFLOW_CHECK_u64(array);
2790 			data->read.one.id = *array;
2791 			array++;
2792 
2793 			if (read_format & PERF_FORMAT_LOST) {
2794 				OVERFLOW_CHECK_u64(array);
2795 				data->read.one.lost = *array;
2796 				array++;
2797 			}
2798 		}
2799 	}
2800 
2801 	if (type & PERF_SAMPLE_CALLCHAIN) {
2802 		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2803 
2804 		OVERFLOW_CHECK_u64(array);
2805 		data->callchain = (struct ip_callchain *)array++;
2806 		if (data->callchain->nr > max_callchain_nr)
2807 			return -EFAULT;
2808 		sz = data->callchain->nr * sizeof(u64);
2809 		OVERFLOW_CHECK(array, sz, max_size);
2810 		array = (void *)array + sz;
2811 	}
2812 
2813 	if (type & PERF_SAMPLE_RAW) {
2814 		OVERFLOW_CHECK_u64(array);
2815 		u.val64 = *array;
2816 
2817 		/*
2818 		 * Undo swap of u64, then swap on individual u32s,
2819 		 * get the size of the raw area and undo all of the
2820 		 * swap. The pevent interface handles endianness by
2821 		 * itself.
2822 		 */
2823 		if (swapped) {
2824 			u.val64 = bswap_64(u.val64);
2825 			u.val32[0] = bswap_32(u.val32[0]);
2826 			u.val32[1] = bswap_32(u.val32[1]);
2827 		}
2828 		data->raw_size = u.val32[0];
2829 
2830 		/*
2831 		 * The raw data is aligned on 64bits including the
2832 		 * u32 size, so it's safe to use mem_bswap_64.
2833 		 */
2834 		if (swapped)
2835 			mem_bswap_64((void *) array, data->raw_size);
2836 
2837 		array = (void *)array + sizeof(u32);
2838 
2839 		OVERFLOW_CHECK(array, data->raw_size, max_size);
2840 		data->raw_data = (void *)array;
2841 		array = (void *)array + data->raw_size;
2842 	}
2843 
2844 	if (type & PERF_SAMPLE_BRANCH_STACK) {
2845 		const u64 max_branch_nr = UINT64_MAX /
2846 					  sizeof(struct branch_entry);
2847 		struct branch_entry *e;
2848 		unsigned int i;
2849 
2850 		OVERFLOW_CHECK_u64(array);
2851 		data->branch_stack = (struct branch_stack *)array++;
2852 
2853 		if (data->branch_stack->nr > max_branch_nr)
2854 			return -EFAULT;
2855 
2856 		sz = data->branch_stack->nr * sizeof(struct branch_entry);
2857 		if (evsel__has_branch_hw_idx(evsel)) {
2858 			sz += sizeof(u64);
2859 			e = &data->branch_stack->entries[0];
2860 		} else {
2861 			data->no_hw_idx = true;
2862 			/*
2863 			 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2864 			 * only nr and entries[] will be output by kernel.
2865 			 */
2866 			e = (struct branch_entry *)&data->branch_stack->hw_idx;
2867 		}
2868 
2869 		if (swapped) {
2870 			/*
2871 			 * struct branch_flag does not have endian
2872 			 * specific bit field definition. And bswap
2873 			 * will not resolve the issue, since these
2874 			 * are bit fields.
2875 			 *
2876 			 * evsel__bitfield_swap_branch_flags() uses a
2877 			 * bitfield_swap macro to swap the bit position
2878 			 * based on the host endians.
2879 			 */
2880 			for (i = 0; i < data->branch_stack->nr; i++, e++)
2881 				e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2882 		}
2883 
2884 		OVERFLOW_CHECK(array, sz, max_size);
2885 		array = (void *)array + sz;
2886 
2887 		if (evsel__has_branch_counters(evsel)) {
2888 			data->branch_stack_cntr = (u64 *)array;
2889 			sz = data->branch_stack->nr * sizeof(u64);
2890 
2891 			OVERFLOW_CHECK(array, sz, max_size);
2892 			array = (void *)array + sz;
2893 		}
2894 	}
2895 
2896 	if (type & PERF_SAMPLE_REGS_USER) {
2897 		OVERFLOW_CHECK_u64(array);
2898 		data->user_regs.abi = *array;
2899 		array++;
2900 
2901 		if (data->user_regs.abi) {
2902 			u64 mask = evsel->core.attr.sample_regs_user;
2903 
2904 			sz = hweight64(mask) * sizeof(u64);
2905 			OVERFLOW_CHECK(array, sz, max_size);
2906 			data->user_regs.mask = mask;
2907 			data->user_regs.regs = (u64 *)array;
2908 			array = (void *)array + sz;
2909 		}
2910 	}
2911 
2912 	if (type & PERF_SAMPLE_STACK_USER) {
2913 		OVERFLOW_CHECK_u64(array);
2914 		sz = *array++;
2915 
2916 		data->user_stack.offset = ((char *)(array - 1)
2917 					  - (char *) event);
2918 
2919 		if (!sz) {
2920 			data->user_stack.size = 0;
2921 		} else {
2922 			OVERFLOW_CHECK(array, sz, max_size);
2923 			data->user_stack.data = (char *)array;
2924 			array = (void *)array + sz;
2925 			OVERFLOW_CHECK_u64(array);
2926 			data->user_stack.size = *array++;
2927 			if (WARN_ONCE(data->user_stack.size > sz,
2928 				      "user stack dump failure\n"))
2929 				return -EFAULT;
2930 		}
2931 	}
2932 
2933 	if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2934 		OVERFLOW_CHECK_u64(array);
2935 		arch_perf_parse_sample_weight(data, array, type);
2936 		array++;
2937 	}
2938 
2939 	if (type & PERF_SAMPLE_DATA_SRC) {
2940 		OVERFLOW_CHECK_u64(array);
2941 		data->data_src = *array;
2942 		array++;
2943 	}
2944 
2945 	if (type & PERF_SAMPLE_TRANSACTION) {
2946 		OVERFLOW_CHECK_u64(array);
2947 		data->transaction = *array;
2948 		array++;
2949 	}
2950 
2951 	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2952 	if (type & PERF_SAMPLE_REGS_INTR) {
2953 		OVERFLOW_CHECK_u64(array);
2954 		data->intr_regs.abi = *array;
2955 		array++;
2956 
2957 		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2958 			u64 mask = evsel->core.attr.sample_regs_intr;
2959 
2960 			sz = hweight64(mask) * sizeof(u64);
2961 			OVERFLOW_CHECK(array, sz, max_size);
2962 			data->intr_regs.mask = mask;
2963 			data->intr_regs.regs = (u64 *)array;
2964 			array = (void *)array + sz;
2965 		}
2966 	}
2967 
2968 	data->phys_addr = 0;
2969 	if (type & PERF_SAMPLE_PHYS_ADDR) {
2970 		data->phys_addr = *array;
2971 		array++;
2972 	}
2973 
2974 	data->cgroup = 0;
2975 	if (type & PERF_SAMPLE_CGROUP) {
2976 		data->cgroup = *array;
2977 		array++;
2978 	}
2979 
2980 	data->data_page_size = 0;
2981 	if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2982 		data->data_page_size = *array;
2983 		array++;
2984 	}
2985 
2986 	data->code_page_size = 0;
2987 	if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2988 		data->code_page_size = *array;
2989 		array++;
2990 	}
2991 
2992 	if (type & PERF_SAMPLE_AUX) {
2993 		OVERFLOW_CHECK_u64(array);
2994 		sz = *array++;
2995 
2996 		OVERFLOW_CHECK(array, sz, max_size);
2997 		/* Undo swap of data */
2998 		if (swapped)
2999 			mem_bswap_64((char *)array, sz);
3000 		data->aux_sample.size = sz;
3001 		data->aux_sample.data = (char *)array;
3002 		array = (void *)array + sz;
3003 	}
3004 
3005 	return 0;
3006 }
3007 
3008 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
3009 				  u64 *timestamp)
3010 {
3011 	u64 type = evsel->core.attr.sample_type;
3012 	const __u64 *array;
3013 
3014 	if (!(type & PERF_SAMPLE_TIME))
3015 		return -1;
3016 
3017 	if (event->header.type != PERF_RECORD_SAMPLE) {
3018 		struct perf_sample data = {
3019 			.time = -1ULL,
3020 		};
3021 
3022 		if (!evsel->core.attr.sample_id_all)
3023 			return -1;
3024 		if (perf_evsel__parse_id_sample(evsel, event, &data))
3025 			return -1;
3026 
3027 		*timestamp = data.time;
3028 		return 0;
3029 	}
3030 
3031 	array = event->sample.array;
3032 
3033 	if (perf_event__check_size(event, evsel->sample_size))
3034 		return -EFAULT;
3035 
3036 	if (type & PERF_SAMPLE_IDENTIFIER)
3037 		array++;
3038 
3039 	if (type & PERF_SAMPLE_IP)
3040 		array++;
3041 
3042 	if (type & PERF_SAMPLE_TID)
3043 		array++;
3044 
3045 	if (type & PERF_SAMPLE_TIME)
3046 		*timestamp = *array;
3047 
3048 	return 0;
3049 }
3050 
3051 u16 evsel__id_hdr_size(const struct evsel *evsel)
3052 {
3053 	u64 sample_type = evsel->core.attr.sample_type;
3054 	u16 size = 0;
3055 
3056 	if (sample_type & PERF_SAMPLE_TID)
3057 		size += sizeof(u64);
3058 
3059 	if (sample_type & PERF_SAMPLE_TIME)
3060 		size += sizeof(u64);
3061 
3062 	if (sample_type & PERF_SAMPLE_ID)
3063 		size += sizeof(u64);
3064 
3065 	if (sample_type & PERF_SAMPLE_STREAM_ID)
3066 		size += sizeof(u64);
3067 
3068 	if (sample_type & PERF_SAMPLE_CPU)
3069 		size += sizeof(u64);
3070 
3071 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
3072 		size += sizeof(u64);
3073 
3074 	return size;
3075 }
3076 
3077 #ifdef HAVE_LIBTRACEEVENT
3078 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
3079 {
3080 	return tep_find_field(evsel->tp_format, name);
3081 }
3082 
3083 struct tep_format_field *evsel__common_field(struct evsel *evsel, const char *name)
3084 {
3085 	return tep_find_common_field(evsel->tp_format, name);
3086 }
3087 
3088 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
3089 {
3090 	struct tep_format_field *field = evsel__field(evsel, name);
3091 	int offset;
3092 
3093 	if (!field)
3094 		return NULL;
3095 
3096 	offset = field->offset;
3097 
3098 	if (field->flags & TEP_FIELD_IS_DYNAMIC) {
3099 		offset = *(int *)(sample->raw_data + field->offset);
3100 		offset &= 0xffff;
3101 		if (tep_field_is_relative(field->flags))
3102 			offset += field->offset + field->size;
3103 	}
3104 
3105 	return sample->raw_data + offset;
3106 }
3107 
3108 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
3109 			 bool needs_swap)
3110 {
3111 	u64 value;
3112 	void *ptr = sample->raw_data + field->offset;
3113 
3114 	switch (field->size) {
3115 	case 1:
3116 		return *(u8 *)ptr;
3117 	case 2:
3118 		value = *(u16 *)ptr;
3119 		break;
3120 	case 4:
3121 		value = *(u32 *)ptr;
3122 		break;
3123 	case 8:
3124 		memcpy(&value, ptr, sizeof(u64));
3125 		break;
3126 	default:
3127 		return 0;
3128 	}
3129 
3130 	if (!needs_swap)
3131 		return value;
3132 
3133 	switch (field->size) {
3134 	case 2:
3135 		return bswap_16(value);
3136 	case 4:
3137 		return bswap_32(value);
3138 	case 8:
3139 		return bswap_64(value);
3140 	default:
3141 		return 0;
3142 	}
3143 
3144 	return 0;
3145 }
3146 
3147 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
3148 {
3149 	struct tep_format_field *field = evsel__field(evsel, name);
3150 
3151 	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
3152 }
3153 
3154 u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const char *name)
3155 {
3156 	struct tep_format_field *field = evsel__common_field(evsel, name);
3157 
3158 	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
3159 }
3160 
3161 char evsel__taskstate(struct evsel *evsel, struct perf_sample *sample, const char *name)
3162 {
3163 	static struct tep_format_field *prev_state_field;
3164 	static const char *states;
3165 	struct tep_format_field *field;
3166 	unsigned long long val;
3167 	unsigned int bit;
3168 	char state = '?'; /* '?' denotes unknown task state */
3169 
3170 	field = evsel__field(evsel, name);
3171 
3172 	if (!field)
3173 		return state;
3174 
3175 	if (!states || field != prev_state_field) {
3176 		states = parse_task_states(field);
3177 		if (!states)
3178 			return state;
3179 		prev_state_field = field;
3180 	}
3181 
3182 	/*
3183 	 * Note since the kernel exposes TASK_REPORT_MAX to userspace
3184 	 * to denote the 'preempted' state, we might as welll report
3185 	 * 'R' for this case, which make senses to users as well.
3186 	 *
3187 	 * We can change this if we have a good reason in the future.
3188 	 */
3189 	val = evsel__intval(evsel, sample, name);
3190 	bit = val ? ffs(val) : 0;
3191 	state = (!bit || bit > strlen(states)) ? 'R' : states[bit-1];
3192 	return state;
3193 }
3194 #endif
3195 
3196 bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
3197 		     char *msg, size_t msgsize)
3198 {
3199 	int paranoid;
3200 
3201 	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
3202 	    evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
3203 	    evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
3204 		/*
3205 		 * If it's cycles then fall back to hrtimer based cpu-clock sw
3206 		 * counter, which is always available even if no PMU support.
3207 		 *
3208 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
3209 		 * b0a873e).
3210 		 */
3211 		evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
3212 		evsel->core.attr.config = target__has_cpu(target)
3213 			? PERF_COUNT_SW_CPU_CLOCK
3214 			: PERF_COUNT_SW_TASK_CLOCK;
3215 		scnprintf(msg, msgsize,
3216 			"The cycles event is not supported, trying to fall back to %s",
3217 			target__has_cpu(target) ? "cpu-clock" : "task-clock");
3218 
3219 		zfree(&evsel->name);
3220 		return true;
3221 	} else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
3222 		   (paranoid = perf_event_paranoid()) > 1) {
3223 		const char *name = evsel__name(evsel);
3224 		char *new_name;
3225 		const char *sep = ":";
3226 
3227 		/* If event has exclude user then don't exclude kernel. */
3228 		if (evsel->core.attr.exclude_user)
3229 			return false;
3230 
3231 		/* Is there already the separator in the name. */
3232 		if (strchr(name, '/') ||
3233 		    (strchr(name, ':') && !evsel->is_libpfm_event))
3234 			sep = "";
3235 
3236 		if (asprintf(&new_name, "%s%su", name, sep) < 0)
3237 			return false;
3238 
3239 		free(evsel->name);
3240 		evsel->name = new_name;
3241 		scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
3242 			  "to fall back to excluding kernel and hypervisor "
3243 			  " samples", paranoid);
3244 		evsel->core.attr.exclude_kernel = 1;
3245 		evsel->core.attr.exclude_hv     = 1;
3246 
3247 		return true;
3248 	}
3249 
3250 	return false;
3251 }
3252 
3253 static bool find_process(const char *name)
3254 {
3255 	size_t len = strlen(name);
3256 	DIR *dir;
3257 	struct dirent *d;
3258 	int ret = -1;
3259 
3260 	dir = opendir(procfs__mountpoint());
3261 	if (!dir)
3262 		return false;
3263 
3264 	/* Walk through the directory. */
3265 	while (ret && (d = readdir(dir)) != NULL) {
3266 		char path[PATH_MAX];
3267 		char *data;
3268 		size_t size;
3269 
3270 		if ((d->d_type != DT_DIR) ||
3271 		     !strcmp(".", d->d_name) ||
3272 		     !strcmp("..", d->d_name))
3273 			continue;
3274 
3275 		scnprintf(path, sizeof(path), "%s/%s/comm",
3276 			  procfs__mountpoint(), d->d_name);
3277 
3278 		if (filename__read_str(path, &data, &size))
3279 			continue;
3280 
3281 		ret = strncmp(name, data, len);
3282 		free(data);
3283 	}
3284 
3285 	closedir(dir);
3286 	return ret ? false : true;
3287 }
3288 
3289 int __weak arch_evsel__open_strerror(struct evsel *evsel __maybe_unused,
3290 				     char *msg __maybe_unused,
3291 				     size_t size __maybe_unused)
3292 {
3293 	return 0;
3294 }
3295 
3296 int evsel__open_strerror(struct evsel *evsel, struct target *target,
3297 			 int err, char *msg, size_t size)
3298 {
3299 	char sbuf[STRERR_BUFSIZE];
3300 	int printed = 0, enforced = 0;
3301 	int ret;
3302 
3303 	switch (err) {
3304 	case EPERM:
3305 	case EACCES:
3306 		printed += scnprintf(msg + printed, size - printed,
3307 			"Access to performance monitoring and observability operations is limited.\n");
3308 
3309 		if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
3310 			if (enforced) {
3311 				printed += scnprintf(msg + printed, size - printed,
3312 					"Enforced MAC policy settings (SELinux) can limit access to performance\n"
3313 					"monitoring and observability operations. Inspect system audit records for\n"
3314 					"more perf_event access control information and adjusting the policy.\n");
3315 			}
3316 		}
3317 
3318 		if (err == EPERM)
3319 			printed += scnprintf(msg, size,
3320 				"No permission to enable %s event.\n\n", evsel__name(evsel));
3321 
3322 		return scnprintf(msg + printed, size - printed,
3323 		 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
3324 		 "access to performance monitoring and observability operations for processes\n"
3325 		 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
3326 		 "More information can be found at 'Perf events and tool security' document:\n"
3327 		 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
3328 		 "perf_event_paranoid setting is %d:\n"
3329 		 "  -1: Allow use of (almost) all events by all users\n"
3330 		 "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
3331 		 ">= 0: Disallow raw and ftrace function tracepoint access\n"
3332 		 ">= 1: Disallow CPU event access\n"
3333 		 ">= 2: Disallow kernel profiling\n"
3334 		 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
3335 		 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
3336 		 perf_event_paranoid());
3337 	case ENOENT:
3338 		return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
3339 	case EMFILE:
3340 		return scnprintf(msg, size, "%s",
3341 			 "Too many events are opened.\n"
3342 			 "Probably the maximum number of open file descriptors has been reached.\n"
3343 			 "Hint: Try again after reducing the number of events.\n"
3344 			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3345 	case ENOMEM:
3346 		if (evsel__has_callchain(evsel) &&
3347 		    access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3348 			return scnprintf(msg, size,
3349 					 "Not enough memory to setup event with callchain.\n"
3350 					 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3351 					 "Hint: Current value: %d", sysctl__max_stack());
3352 		break;
3353 	case ENODEV:
3354 		if (target->cpu_list)
3355 			return scnprintf(msg, size, "%s",
3356 	 "No such device - did you specify an out-of-range profile CPU?");
3357 		break;
3358 	case EOPNOTSUPP:
3359 		if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3360 			return scnprintf(msg, size,
3361 	"%s: PMU Hardware or event type doesn't support branch stack sampling.",
3362 					 evsel__name(evsel));
3363 		if (evsel->core.attr.aux_output)
3364 			return scnprintf(msg, size,
3365 	"%s: PMU Hardware doesn't support 'aux_output' feature",
3366 					 evsel__name(evsel));
3367 		if (evsel->core.attr.sample_period != 0)
3368 			return scnprintf(msg, size,
3369 	"%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3370 					 evsel__name(evsel));
3371 		if (evsel->core.attr.precise_ip)
3372 			return scnprintf(msg, size, "%s",
3373 	"\'precise\' request may not be supported. Try removing 'p' modifier.");
3374 #if defined(__i386__) || defined(__x86_64__)
3375 		if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3376 			return scnprintf(msg, size, "%s",
3377 	"No hardware sampling interrupt available.\n");
3378 #endif
3379 		break;
3380 	case EBUSY:
3381 		if (find_process("oprofiled"))
3382 			return scnprintf(msg, size,
3383 	"The PMU counters are busy/taken by another profiler.\n"
3384 	"We found oprofile daemon running, please stop it and try again.");
3385 		break;
3386 	case EINVAL:
3387 		if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3388 			return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3389 		if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3390 			return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3391 		if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3392 			return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3393 		if (perf_missing_features.clockid)
3394 			return scnprintf(msg, size, "clockid feature not supported.");
3395 		if (perf_missing_features.clockid_wrong)
3396 			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3397 		if (perf_missing_features.aux_output)
3398 			return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3399 		if (!target__has_cpu(target))
3400 			return scnprintf(msg, size,
3401 	"Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3402 					evsel__name(evsel));
3403 
3404 		break;
3405 	case ENODATA:
3406 		return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3407 				 "Please add an auxiliary event in front of the load latency event.");
3408 	default:
3409 		break;
3410 	}
3411 
3412 	ret = arch_evsel__open_strerror(evsel, msg, size);
3413 	if (ret)
3414 		return ret;
3415 
3416 	return scnprintf(msg, size,
3417 	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3418 	"/bin/dmesg | grep -i perf may provide additional information.\n",
3419 			 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3420 }
3421 
3422 struct perf_env *evsel__env(struct evsel *evsel)
3423 {
3424 	if (evsel && evsel->evlist && evsel->evlist->env)
3425 		return evsel->evlist->env;
3426 	return &perf_env;
3427 }
3428 
3429 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3430 {
3431 	int cpu_map_idx, thread;
3432 
3433 	if (evsel__is_retire_lat(evsel))
3434 		return 0;
3435 
3436 	for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3437 		for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3438 		     thread++) {
3439 			int fd = FD(evsel, cpu_map_idx, thread);
3440 
3441 			if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3442 						   cpu_map_idx, thread, fd) < 0)
3443 				return -1;
3444 		}
3445 	}
3446 
3447 	return 0;
3448 }
3449 
3450 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3451 {
3452 	struct perf_cpu_map *cpus = evsel->core.cpus;
3453 	struct perf_thread_map *threads = evsel->core.threads;
3454 
3455 	if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3456 		return -ENOMEM;
3457 
3458 	return store_evsel_ids(evsel, evlist);
3459 }
3460 
3461 void evsel__zero_per_pkg(struct evsel *evsel)
3462 {
3463 	struct hashmap_entry *cur;
3464 	size_t bkt;
3465 
3466 	if (evsel->per_pkg_mask) {
3467 		hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3468 			zfree(&cur->pkey);
3469 
3470 		hashmap__clear(evsel->per_pkg_mask);
3471 	}
3472 }
3473 
3474 /**
3475  * evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this
3476  *                    will be false on hybrid systems for hardware and legacy
3477  *                    cache events.
3478  */
3479 bool evsel__is_hybrid(const struct evsel *evsel)
3480 {
3481 	if (perf_pmus__num_core_pmus() == 1)
3482 		return false;
3483 
3484 	return evsel->core.is_pmu_core;
3485 }
3486 
3487 struct evsel *evsel__leader(const struct evsel *evsel)
3488 {
3489 	return container_of(evsel->core.leader, struct evsel, core);
3490 }
3491 
3492 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3493 {
3494 	return evsel->core.leader == &leader->core;
3495 }
3496 
3497 bool evsel__is_leader(struct evsel *evsel)
3498 {
3499 	return evsel__has_leader(evsel, evsel);
3500 }
3501 
3502 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3503 {
3504 	evsel->core.leader = &leader->core;
3505 }
3506 
3507 int evsel__source_count(const struct evsel *evsel)
3508 {
3509 	struct evsel *pos;
3510 	int count = 0;
3511 
3512 	evlist__for_each_entry(evsel->evlist, pos) {
3513 		if (pos->metric_leader == evsel)
3514 			count++;
3515 	}
3516 	return count;
3517 }
3518 
3519 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3520 {
3521 	return false;
3522 }
3523 
3524 /*
3525  * Remove an event from a given group (leader).
3526  * Some events, e.g., perf metrics Topdown events,
3527  * must always be grouped. Ignore the events.
3528  */
3529 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3530 {
3531 	if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3532 		evsel__set_leader(evsel, evsel);
3533 		evsel->core.nr_members = 0;
3534 		leader->core.nr_members--;
3535 	}
3536 }
3537