xref: /linux/tools/perf/util/evsel.c (revision 200323768787a0ee02e01c35c1aff13dc9d77dde)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4  *
5  * Parts came from builtin-{top,stat,record}.c, see those files for further
6  * copyright notes.
7  */
8 
9 #include <byteswap.h>
10 #include <errno.h>
11 #include <inttypes.h>
12 #include <linux/bitops.h>
13 #include <api/fs/fs.h>
14 #include <api/fs/tracing_path.h>
15 #include <linux/hw_breakpoint.h>
16 #include <linux/perf_event.h>
17 #include <linux/compiler.h>
18 #include <linux/err.h>
19 #include <linux/zalloc.h>
20 #include <sys/ioctl.h>
21 #include <sys/resource.h>
22 #include <sys/types.h>
23 #include <dirent.h>
24 #include <stdlib.h>
25 #include <perf/evsel.h>
26 #include "asm/bug.h"
27 #include "bpf_counter.h"
28 #include "callchain.h"
29 #include "cgroup.h"
30 #include "counts.h"
31 #include "event.h"
32 #include "evsel.h"
33 #include "util/env.h"
34 #include "util/evsel_config.h"
35 #include "util/evsel_fprintf.h"
36 #include "evlist.h"
37 #include <perf/cpumap.h>
38 #include "thread_map.h"
39 #include "target.h"
40 #include "perf_regs.h"
41 #include "record.h"
42 #include "debug.h"
43 #include "trace-event.h"
44 #include "stat.h"
45 #include "string2.h"
46 #include "memswap.h"
47 #include "util.h"
48 #include "util/hashmap.h"
49 #include "off_cpu.h"
50 #include "pmu.h"
51 #include "pmus.h"
52 #include "../perf-sys.h"
53 #include "util/parse-branch-options.h"
54 #include "util/bpf-filter.h"
55 #include <internal/xyarray.h>
56 #include <internal/lib.h>
57 #include <internal/threadmap.h>
58 
59 #include <linux/ctype.h>
60 
61 #ifdef HAVE_LIBTRACEEVENT
62 #include <traceevent/event-parse.h>
63 #endif
64 
65 struct perf_missing_features perf_missing_features;
66 
67 static clockid_t clockid;
68 
69 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
70 	NULL,
71 	"duration_time",
72 	"user_time",
73 	"system_time",
74 };
75 
76 const char *perf_tool_event__to_str(enum perf_tool_event ev)
77 {
78 	if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
79 		return perf_tool_event__tool_names[ev];
80 
81 	return NULL;
82 }
83 
84 enum perf_tool_event perf_tool_event__from_str(const char *str)
85 {
86 	int i;
87 
88 	perf_tool_event__for_each_event(i) {
89 		if (!strcmp(str, perf_tool_event__tool_names[i]))
90 			return i;
91 	}
92 	return PERF_TOOL_NONE;
93 }
94 
95 
96 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
97 {
98 	return 0;
99 }
100 
101 void __weak test_attr__ready(void) { }
102 
103 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
104 {
105 }
106 
107 static struct {
108 	size_t	size;
109 	int	(*init)(struct evsel *evsel);
110 	void	(*fini)(struct evsel *evsel);
111 } perf_evsel__object = {
112 	.size = sizeof(struct evsel),
113 	.init = evsel__no_extra_init,
114 	.fini = evsel__no_extra_fini,
115 };
116 
117 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
118 			 void (*fini)(struct evsel *evsel))
119 {
120 
121 	if (object_size == 0)
122 		goto set_methods;
123 
124 	if (perf_evsel__object.size > object_size)
125 		return -EINVAL;
126 
127 	perf_evsel__object.size = object_size;
128 
129 set_methods:
130 	if (init != NULL)
131 		perf_evsel__object.init = init;
132 
133 	if (fini != NULL)
134 		perf_evsel__object.fini = fini;
135 
136 	return 0;
137 }
138 
139 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
140 
141 int __evsel__sample_size(u64 sample_type)
142 {
143 	u64 mask = sample_type & PERF_SAMPLE_MASK;
144 	int size = 0;
145 	int i;
146 
147 	for (i = 0; i < 64; i++) {
148 		if (mask & (1ULL << i))
149 			size++;
150 	}
151 
152 	size *= sizeof(u64);
153 
154 	return size;
155 }
156 
157 /**
158  * __perf_evsel__calc_id_pos - calculate id_pos.
159  * @sample_type: sample type
160  *
161  * This function returns the position of the event id (PERF_SAMPLE_ID or
162  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
163  * perf_record_sample.
164  */
165 static int __perf_evsel__calc_id_pos(u64 sample_type)
166 {
167 	int idx = 0;
168 
169 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
170 		return 0;
171 
172 	if (!(sample_type & PERF_SAMPLE_ID))
173 		return -1;
174 
175 	if (sample_type & PERF_SAMPLE_IP)
176 		idx += 1;
177 
178 	if (sample_type & PERF_SAMPLE_TID)
179 		idx += 1;
180 
181 	if (sample_type & PERF_SAMPLE_TIME)
182 		idx += 1;
183 
184 	if (sample_type & PERF_SAMPLE_ADDR)
185 		idx += 1;
186 
187 	return idx;
188 }
189 
190 /**
191  * __perf_evsel__calc_is_pos - calculate is_pos.
192  * @sample_type: sample type
193  *
194  * This function returns the position (counting backwards) of the event id
195  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
196  * sample_id_all is used there is an id sample appended to non-sample events.
197  */
198 static int __perf_evsel__calc_is_pos(u64 sample_type)
199 {
200 	int idx = 1;
201 
202 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
203 		return 1;
204 
205 	if (!(sample_type & PERF_SAMPLE_ID))
206 		return -1;
207 
208 	if (sample_type & PERF_SAMPLE_CPU)
209 		idx += 1;
210 
211 	if (sample_type & PERF_SAMPLE_STREAM_ID)
212 		idx += 1;
213 
214 	return idx;
215 }
216 
217 void evsel__calc_id_pos(struct evsel *evsel)
218 {
219 	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
220 	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
221 }
222 
223 void __evsel__set_sample_bit(struct evsel *evsel,
224 				  enum perf_event_sample_format bit)
225 {
226 	if (!(evsel->core.attr.sample_type & bit)) {
227 		evsel->core.attr.sample_type |= bit;
228 		evsel->sample_size += sizeof(u64);
229 		evsel__calc_id_pos(evsel);
230 	}
231 }
232 
233 void __evsel__reset_sample_bit(struct evsel *evsel,
234 				    enum perf_event_sample_format bit)
235 {
236 	if (evsel->core.attr.sample_type & bit) {
237 		evsel->core.attr.sample_type &= ~bit;
238 		evsel->sample_size -= sizeof(u64);
239 		evsel__calc_id_pos(evsel);
240 	}
241 }
242 
243 void evsel__set_sample_id(struct evsel *evsel,
244 			       bool can_sample_identifier)
245 {
246 	if (can_sample_identifier) {
247 		evsel__reset_sample_bit(evsel, ID);
248 		evsel__set_sample_bit(evsel, IDENTIFIER);
249 	} else {
250 		evsel__set_sample_bit(evsel, ID);
251 	}
252 	evsel->core.attr.read_format |= PERF_FORMAT_ID;
253 }
254 
255 /**
256  * evsel__is_function_event - Return whether given evsel is a function
257  * trace event
258  *
259  * @evsel - evsel selector to be tested
260  *
261  * Return %true if event is function trace event
262  */
263 bool evsel__is_function_event(struct evsel *evsel)
264 {
265 #define FUNCTION_EVENT "ftrace:function"
266 
267 	return evsel->name &&
268 	       !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
269 
270 #undef FUNCTION_EVENT
271 }
272 
273 void evsel__init(struct evsel *evsel,
274 		 struct perf_event_attr *attr, int idx)
275 {
276 	perf_evsel__init(&evsel->core, attr, idx);
277 	evsel->tracking	   = !idx;
278 	evsel->unit	   = strdup("");
279 	evsel->scale	   = 1.0;
280 	evsel->max_events  = ULONG_MAX;
281 	evsel->evlist	   = NULL;
282 	evsel->bpf_obj	   = NULL;
283 	evsel->bpf_fd	   = -1;
284 	INIT_LIST_HEAD(&evsel->config_terms);
285 	INIT_LIST_HEAD(&evsel->bpf_counter_list);
286 	perf_evsel__object.init(evsel);
287 	evsel->sample_size = __evsel__sample_size(attr->sample_type);
288 	evsel__calc_id_pos(evsel);
289 	evsel->cmdline_group_boundary = false;
290 	evsel->metric_events = NULL;
291 	evsel->per_pkg_mask  = NULL;
292 	evsel->collect_stat  = false;
293 	evsel->pmu_name      = NULL;
294 	evsel->group_pmu_name = NULL;
295 	evsel->skippable     = false;
296 }
297 
298 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
299 {
300 	struct evsel *evsel = zalloc(perf_evsel__object.size);
301 
302 	if (!evsel)
303 		return NULL;
304 	evsel__init(evsel, attr, idx);
305 
306 	if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
307 		evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
308 					    PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
309 		evsel->core.attr.sample_period = 1;
310 	}
311 
312 	if (evsel__is_clock(evsel)) {
313 		free((char *)evsel->unit);
314 		evsel->unit = strdup("msec");
315 		evsel->scale = 1e-6;
316 	}
317 
318 	return evsel;
319 }
320 
321 int copy_config_terms(struct list_head *dst, struct list_head *src)
322 {
323 	struct evsel_config_term *pos, *tmp;
324 
325 	list_for_each_entry(pos, src, list) {
326 		tmp = malloc(sizeof(*tmp));
327 		if (tmp == NULL)
328 			return -ENOMEM;
329 
330 		*tmp = *pos;
331 		if (tmp->free_str) {
332 			tmp->val.str = strdup(pos->val.str);
333 			if (tmp->val.str == NULL) {
334 				free(tmp);
335 				return -ENOMEM;
336 			}
337 		}
338 		list_add_tail(&tmp->list, dst);
339 	}
340 	return 0;
341 }
342 
343 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
344 {
345 	return copy_config_terms(&dst->config_terms, &src->config_terms);
346 }
347 
348 /**
349  * evsel__clone - create a new evsel copied from @orig
350  * @orig: original evsel
351  *
352  * The assumption is that @orig is not configured nor opened yet.
353  * So we only care about the attributes that can be set while it's parsed.
354  */
355 struct evsel *evsel__clone(struct evsel *orig)
356 {
357 	struct evsel *evsel;
358 
359 	BUG_ON(orig->core.fd);
360 	BUG_ON(orig->counts);
361 	BUG_ON(orig->priv);
362 	BUG_ON(orig->per_pkg_mask);
363 
364 	/* cannot handle BPF objects for now */
365 	if (orig->bpf_obj)
366 		return NULL;
367 
368 	evsel = evsel__new(&orig->core.attr);
369 	if (evsel == NULL)
370 		return NULL;
371 
372 	evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
373 	evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
374 	evsel->core.threads = perf_thread_map__get(orig->core.threads);
375 	evsel->core.nr_members = orig->core.nr_members;
376 	evsel->core.system_wide = orig->core.system_wide;
377 	evsel->core.requires_cpu = orig->core.requires_cpu;
378 	evsel->core.is_pmu_core = orig->core.is_pmu_core;
379 
380 	if (orig->name) {
381 		evsel->name = strdup(orig->name);
382 		if (evsel->name == NULL)
383 			goto out_err;
384 	}
385 	if (orig->group_name) {
386 		evsel->group_name = strdup(orig->group_name);
387 		if (evsel->group_name == NULL)
388 			goto out_err;
389 	}
390 	if (orig->pmu_name) {
391 		evsel->pmu_name = strdup(orig->pmu_name);
392 		if (evsel->pmu_name == NULL)
393 			goto out_err;
394 	}
395 	if (orig->group_pmu_name) {
396 		evsel->group_pmu_name = strdup(orig->group_pmu_name);
397 		if (evsel->group_pmu_name == NULL)
398 			goto out_err;
399 	}
400 	if (orig->filter) {
401 		evsel->filter = strdup(orig->filter);
402 		if (evsel->filter == NULL)
403 			goto out_err;
404 	}
405 	if (orig->metric_id) {
406 		evsel->metric_id = strdup(orig->metric_id);
407 		if (evsel->metric_id == NULL)
408 			goto out_err;
409 	}
410 	evsel->cgrp = cgroup__get(orig->cgrp);
411 #ifdef HAVE_LIBTRACEEVENT
412 	evsel->tp_format = orig->tp_format;
413 #endif
414 	evsel->handler = orig->handler;
415 	evsel->core.leader = orig->core.leader;
416 
417 	evsel->max_events = orig->max_events;
418 	evsel->tool_event = orig->tool_event;
419 	free((char *)evsel->unit);
420 	evsel->unit = strdup(orig->unit);
421 	if (evsel->unit == NULL)
422 		goto out_err;
423 
424 	evsel->scale = orig->scale;
425 	evsel->snapshot = orig->snapshot;
426 	evsel->per_pkg = orig->per_pkg;
427 	evsel->percore = orig->percore;
428 	evsel->precise_max = orig->precise_max;
429 	evsel->is_libpfm_event = orig->is_libpfm_event;
430 
431 	evsel->exclude_GH = orig->exclude_GH;
432 	evsel->sample_read = orig->sample_read;
433 	evsel->auto_merge_stats = orig->auto_merge_stats;
434 	evsel->collect_stat = orig->collect_stat;
435 	evsel->weak_group = orig->weak_group;
436 	evsel->use_config_name = orig->use_config_name;
437 	evsel->pmu = orig->pmu;
438 
439 	if (evsel__copy_config_terms(evsel, orig) < 0)
440 		goto out_err;
441 
442 	return evsel;
443 
444 out_err:
445 	evsel__delete(evsel);
446 	return NULL;
447 }
448 
449 /*
450  * Returns pointer with encoded error via <linux/err.h> interface.
451  */
452 #ifdef HAVE_LIBTRACEEVENT
453 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
454 {
455 	struct evsel *evsel = zalloc(perf_evsel__object.size);
456 	int err = -ENOMEM;
457 
458 	if (evsel == NULL) {
459 		goto out_err;
460 	} else {
461 		struct perf_event_attr attr = {
462 			.type	       = PERF_TYPE_TRACEPOINT,
463 			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
464 					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
465 		};
466 
467 		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
468 			goto out_free;
469 
470 		evsel->tp_format = trace_event__tp_format(sys, name);
471 		if (IS_ERR(evsel->tp_format)) {
472 			err = PTR_ERR(evsel->tp_format);
473 			goto out_free;
474 		}
475 
476 		event_attr_init(&attr);
477 		attr.config = evsel->tp_format->id;
478 		attr.sample_period = 1;
479 		evsel__init(evsel, &attr, idx);
480 	}
481 
482 	return evsel;
483 
484 out_free:
485 	zfree(&evsel->name);
486 	free(evsel);
487 out_err:
488 	return ERR_PTR(err);
489 }
490 #endif
491 
492 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
493 	"cycles",
494 	"instructions",
495 	"cache-references",
496 	"cache-misses",
497 	"branches",
498 	"branch-misses",
499 	"bus-cycles",
500 	"stalled-cycles-frontend",
501 	"stalled-cycles-backend",
502 	"ref-cycles",
503 };
504 
505 char *evsel__bpf_counter_events;
506 
507 bool evsel__match_bpf_counter_events(const char *name)
508 {
509 	int name_len;
510 	bool match;
511 	char *ptr;
512 
513 	if (!evsel__bpf_counter_events)
514 		return false;
515 
516 	ptr = strstr(evsel__bpf_counter_events, name);
517 	name_len = strlen(name);
518 
519 	/* check name matches a full token in evsel__bpf_counter_events */
520 	match = (ptr != NULL) &&
521 		((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
522 		((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
523 
524 	return match;
525 }
526 
527 static const char *__evsel__hw_name(u64 config)
528 {
529 	if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
530 		return evsel__hw_names[config];
531 
532 	return "unknown-hardware";
533 }
534 
535 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
536 {
537 	int colon = 0, r = 0;
538 	struct perf_event_attr *attr = &evsel->core.attr;
539 	bool exclude_guest_default = false;
540 
541 #define MOD_PRINT(context, mod)	do {					\
542 		if (!attr->exclude_##context) {				\
543 			if (!colon) colon = ++r;			\
544 			r += scnprintf(bf + r, size - r, "%c", mod);	\
545 		} } while(0)
546 
547 	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
548 		MOD_PRINT(kernel, 'k');
549 		MOD_PRINT(user, 'u');
550 		MOD_PRINT(hv, 'h');
551 		exclude_guest_default = true;
552 	}
553 
554 	if (attr->precise_ip) {
555 		if (!colon)
556 			colon = ++r;
557 		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
558 		exclude_guest_default = true;
559 	}
560 
561 	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
562 		MOD_PRINT(host, 'H');
563 		MOD_PRINT(guest, 'G');
564 	}
565 #undef MOD_PRINT
566 	if (colon)
567 		bf[colon - 1] = ':';
568 	return r;
569 }
570 
571 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
572 {
573 	return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
574 }
575 
576 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
577 {
578 	int r = arch_evsel__hw_name(evsel, bf, size);
579 	return r + evsel__add_modifiers(evsel, bf + r, size - r);
580 }
581 
582 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
583 	"cpu-clock",
584 	"task-clock",
585 	"page-faults",
586 	"context-switches",
587 	"cpu-migrations",
588 	"minor-faults",
589 	"major-faults",
590 	"alignment-faults",
591 	"emulation-faults",
592 	"dummy",
593 };
594 
595 static const char *__evsel__sw_name(u64 config)
596 {
597 	if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
598 		return evsel__sw_names[config];
599 	return "unknown-software";
600 }
601 
602 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
603 {
604 	int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
605 	return r + evsel__add_modifiers(evsel, bf + r, size - r);
606 }
607 
608 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
609 {
610 	return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
611 }
612 
613 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
614 {
615 	int r;
616 
617 	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
618 
619 	if (type & HW_BREAKPOINT_R)
620 		r += scnprintf(bf + r, size - r, "r");
621 
622 	if (type & HW_BREAKPOINT_W)
623 		r += scnprintf(bf + r, size - r, "w");
624 
625 	if (type & HW_BREAKPOINT_X)
626 		r += scnprintf(bf + r, size - r, "x");
627 
628 	return r;
629 }
630 
631 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
632 {
633 	struct perf_event_attr *attr = &evsel->core.attr;
634 	int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
635 	return r + evsel__add_modifiers(evsel, bf + r, size - r);
636 }
637 
638 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
639  { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
640  { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
641  { "LLC",	"L2",							},
642  { "dTLB",	"d-tlb",	"Data-TLB",				},
643  { "iTLB",	"i-tlb",	"Instruction-TLB",			},
644  { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
645  { "node",								},
646 };
647 
648 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
649  { "load",	"loads",	"read",					},
650  { "store",	"stores",	"write",				},
651  { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
652 };
653 
654 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
655  { "refs",	"Reference",	"ops",		"access",		},
656  { "misses",	"miss",							},
657 };
658 
659 #define C(x)		PERF_COUNT_HW_CACHE_##x
660 #define CACHE_READ	(1 << C(OP_READ))
661 #define CACHE_WRITE	(1 << C(OP_WRITE))
662 #define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
663 #define COP(x)		(1 << x)
664 
665 /*
666  * cache operation stat
667  * L1I : Read and prefetch only
668  * ITLB and BPU : Read-only
669  */
670 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
671  [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
672  [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
673  [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
674  [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
675  [C(ITLB)]	= (CACHE_READ),
676  [C(BPU)]	= (CACHE_READ),
677  [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
678 };
679 
680 bool evsel__is_cache_op_valid(u8 type, u8 op)
681 {
682 	if (evsel__hw_cache_stat[type] & COP(op))
683 		return true;	/* valid */
684 	else
685 		return false;	/* invalid */
686 }
687 
688 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
689 {
690 	if (result) {
691 		return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
692 				 evsel__hw_cache_op[op][0],
693 				 evsel__hw_cache_result[result][0]);
694 	}
695 
696 	return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
697 			 evsel__hw_cache_op[op][1]);
698 }
699 
700 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
701 {
702 	u8 op, result, type = (config >>  0) & 0xff;
703 	const char *err = "unknown-ext-hardware-cache-type";
704 
705 	if (type >= PERF_COUNT_HW_CACHE_MAX)
706 		goto out_err;
707 
708 	op = (config >>  8) & 0xff;
709 	err = "unknown-ext-hardware-cache-op";
710 	if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
711 		goto out_err;
712 
713 	result = (config >> 16) & 0xff;
714 	err = "unknown-ext-hardware-cache-result";
715 	if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
716 		goto out_err;
717 
718 	err = "invalid-cache";
719 	if (!evsel__is_cache_op_valid(type, op))
720 		goto out_err;
721 
722 	return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
723 out_err:
724 	return scnprintf(bf, size, "%s", err);
725 }
726 
727 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
728 {
729 	int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
730 	return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
731 }
732 
733 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
734 {
735 	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
736 	return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
737 }
738 
739 const char *evsel__name(struct evsel *evsel)
740 {
741 	char bf[128];
742 
743 	if (!evsel)
744 		goto out_unknown;
745 
746 	if (evsel->name)
747 		return evsel->name;
748 
749 	switch (evsel->core.attr.type) {
750 	case PERF_TYPE_RAW:
751 		evsel__raw_name(evsel, bf, sizeof(bf));
752 		break;
753 
754 	case PERF_TYPE_HARDWARE:
755 		evsel__hw_name(evsel, bf, sizeof(bf));
756 		break;
757 
758 	case PERF_TYPE_HW_CACHE:
759 		evsel__hw_cache_name(evsel, bf, sizeof(bf));
760 		break;
761 
762 	case PERF_TYPE_SOFTWARE:
763 		if (evsel__is_tool(evsel))
764 			evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
765 		else
766 			evsel__sw_name(evsel, bf, sizeof(bf));
767 		break;
768 
769 	case PERF_TYPE_TRACEPOINT:
770 		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
771 		break;
772 
773 	case PERF_TYPE_BREAKPOINT:
774 		evsel__bp_name(evsel, bf, sizeof(bf));
775 		break;
776 
777 	default:
778 		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
779 			  evsel->core.attr.type);
780 		break;
781 	}
782 
783 	evsel->name = strdup(bf);
784 
785 	if (evsel->name)
786 		return evsel->name;
787 out_unknown:
788 	return "unknown";
789 }
790 
791 bool evsel__name_is(struct evsel *evsel, const char *name)
792 {
793 	return !strcmp(evsel__name(evsel), name);
794 }
795 
796 const char *evsel__metric_id(const struct evsel *evsel)
797 {
798 	if (evsel->metric_id)
799 		return evsel->metric_id;
800 
801 	if (evsel__is_tool(evsel))
802 		return perf_tool_event__to_str(evsel->tool_event);
803 
804 	return "unknown";
805 }
806 
807 const char *evsel__group_name(struct evsel *evsel)
808 {
809 	return evsel->group_name ?: "anon group";
810 }
811 
812 /*
813  * Returns the group details for the specified leader,
814  * with following rules.
815  *
816  *  For record -e '{cycles,instructions}'
817  *    'anon group { cycles:u, instructions:u }'
818  *
819  *  For record -e 'cycles,instructions' and report --group
820  *    'cycles:u, instructions:u'
821  */
822 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
823 {
824 	int ret = 0;
825 	struct evsel *pos;
826 	const char *group_name = evsel__group_name(evsel);
827 
828 	if (!evsel->forced_leader)
829 		ret = scnprintf(buf, size, "%s { ", group_name);
830 
831 	ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
832 
833 	for_each_group_member(pos, evsel)
834 		ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
835 
836 	if (!evsel->forced_leader)
837 		ret += scnprintf(buf + ret, size - ret, " }");
838 
839 	return ret;
840 }
841 
842 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
843 				      struct callchain_param *param)
844 {
845 	bool function = evsel__is_function_event(evsel);
846 	struct perf_event_attr *attr = &evsel->core.attr;
847 
848 	evsel__set_sample_bit(evsel, CALLCHAIN);
849 
850 	attr->sample_max_stack = param->max_stack;
851 
852 	if (opts->kernel_callchains)
853 		attr->exclude_callchain_user = 1;
854 	if (opts->user_callchains)
855 		attr->exclude_callchain_kernel = 1;
856 	if (param->record_mode == CALLCHAIN_LBR) {
857 		if (!opts->branch_stack) {
858 			if (attr->exclude_user) {
859 				pr_warning("LBR callstack option is only available "
860 					   "to get user callchain information. "
861 					   "Falling back to framepointers.\n");
862 			} else {
863 				evsel__set_sample_bit(evsel, BRANCH_STACK);
864 				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
865 							PERF_SAMPLE_BRANCH_CALL_STACK |
866 							PERF_SAMPLE_BRANCH_NO_CYCLES |
867 							PERF_SAMPLE_BRANCH_NO_FLAGS |
868 							PERF_SAMPLE_BRANCH_HW_INDEX;
869 			}
870 		} else
871 			 pr_warning("Cannot use LBR callstack with branch stack. "
872 				    "Falling back to framepointers.\n");
873 	}
874 
875 	if (param->record_mode == CALLCHAIN_DWARF) {
876 		if (!function) {
877 			evsel__set_sample_bit(evsel, REGS_USER);
878 			evsel__set_sample_bit(evsel, STACK_USER);
879 			if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
880 				attr->sample_regs_user |= DWARF_MINIMAL_REGS;
881 				pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
882 					   "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
883 					   "so the minimal registers set (IP, SP) is explicitly forced.\n");
884 			} else {
885 				attr->sample_regs_user |= arch__user_reg_mask();
886 			}
887 			attr->sample_stack_user = param->dump_size;
888 			attr->exclude_callchain_user = 1;
889 		} else {
890 			pr_info("Cannot use DWARF unwind for function trace event,"
891 				" falling back to framepointers.\n");
892 		}
893 	}
894 
895 	if (function) {
896 		pr_info("Disabling user space callchains for function trace event.\n");
897 		attr->exclude_callchain_user = 1;
898 	}
899 }
900 
901 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
902 			     struct callchain_param *param)
903 {
904 	if (param->enabled)
905 		return __evsel__config_callchain(evsel, opts, param);
906 }
907 
908 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
909 {
910 	struct perf_event_attr *attr = &evsel->core.attr;
911 
912 	evsel__reset_sample_bit(evsel, CALLCHAIN);
913 	if (param->record_mode == CALLCHAIN_LBR) {
914 		evsel__reset_sample_bit(evsel, BRANCH_STACK);
915 		attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
916 					      PERF_SAMPLE_BRANCH_CALL_STACK |
917 					      PERF_SAMPLE_BRANCH_HW_INDEX);
918 	}
919 	if (param->record_mode == CALLCHAIN_DWARF) {
920 		evsel__reset_sample_bit(evsel, REGS_USER);
921 		evsel__reset_sample_bit(evsel, STACK_USER);
922 	}
923 }
924 
925 static void evsel__apply_config_terms(struct evsel *evsel,
926 				      struct record_opts *opts, bool track)
927 {
928 	struct evsel_config_term *term;
929 	struct list_head *config_terms = &evsel->config_terms;
930 	struct perf_event_attr *attr = &evsel->core.attr;
931 	/* callgraph default */
932 	struct callchain_param param = {
933 		.record_mode = callchain_param.record_mode,
934 	};
935 	u32 dump_size = 0;
936 	int max_stack = 0;
937 	const char *callgraph_buf = NULL;
938 
939 	list_for_each_entry(term, config_terms, list) {
940 		switch (term->type) {
941 		case EVSEL__CONFIG_TERM_PERIOD:
942 			if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
943 				attr->sample_period = term->val.period;
944 				attr->freq = 0;
945 				evsel__reset_sample_bit(evsel, PERIOD);
946 			}
947 			break;
948 		case EVSEL__CONFIG_TERM_FREQ:
949 			if (!(term->weak && opts->user_freq != UINT_MAX)) {
950 				attr->sample_freq = term->val.freq;
951 				attr->freq = 1;
952 				evsel__set_sample_bit(evsel, PERIOD);
953 			}
954 			break;
955 		case EVSEL__CONFIG_TERM_TIME:
956 			if (term->val.time)
957 				evsel__set_sample_bit(evsel, TIME);
958 			else
959 				evsel__reset_sample_bit(evsel, TIME);
960 			break;
961 		case EVSEL__CONFIG_TERM_CALLGRAPH:
962 			callgraph_buf = term->val.str;
963 			break;
964 		case EVSEL__CONFIG_TERM_BRANCH:
965 			if (term->val.str && strcmp(term->val.str, "no")) {
966 				evsel__set_sample_bit(evsel, BRANCH_STACK);
967 				parse_branch_str(term->val.str,
968 						 &attr->branch_sample_type);
969 			} else
970 				evsel__reset_sample_bit(evsel, BRANCH_STACK);
971 			break;
972 		case EVSEL__CONFIG_TERM_STACK_USER:
973 			dump_size = term->val.stack_user;
974 			break;
975 		case EVSEL__CONFIG_TERM_MAX_STACK:
976 			max_stack = term->val.max_stack;
977 			break;
978 		case EVSEL__CONFIG_TERM_MAX_EVENTS:
979 			evsel->max_events = term->val.max_events;
980 			break;
981 		case EVSEL__CONFIG_TERM_INHERIT:
982 			/*
983 			 * attr->inherit should has already been set by
984 			 * evsel__config. If user explicitly set
985 			 * inherit using config terms, override global
986 			 * opt->no_inherit setting.
987 			 */
988 			attr->inherit = term->val.inherit ? 1 : 0;
989 			break;
990 		case EVSEL__CONFIG_TERM_OVERWRITE:
991 			attr->write_backward = term->val.overwrite ? 1 : 0;
992 			break;
993 		case EVSEL__CONFIG_TERM_DRV_CFG:
994 			break;
995 		case EVSEL__CONFIG_TERM_PERCORE:
996 			break;
997 		case EVSEL__CONFIG_TERM_AUX_OUTPUT:
998 			attr->aux_output = term->val.aux_output ? 1 : 0;
999 			break;
1000 		case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1001 			/* Already applied by auxtrace */
1002 			break;
1003 		case EVSEL__CONFIG_TERM_CFG_CHG:
1004 			break;
1005 		default:
1006 			break;
1007 		}
1008 	}
1009 
1010 	/* User explicitly set per-event callgraph, clear the old setting and reset. */
1011 	if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1012 		bool sample_address = false;
1013 
1014 		if (max_stack) {
1015 			param.max_stack = max_stack;
1016 			if (callgraph_buf == NULL)
1017 				callgraph_buf = "fp";
1018 		}
1019 
1020 		/* parse callgraph parameters */
1021 		if (callgraph_buf != NULL) {
1022 			if (!strcmp(callgraph_buf, "no")) {
1023 				param.enabled = false;
1024 				param.record_mode = CALLCHAIN_NONE;
1025 			} else {
1026 				param.enabled = true;
1027 				if (parse_callchain_record(callgraph_buf, &param)) {
1028 					pr_err("per-event callgraph setting for %s failed. "
1029 					       "Apply callgraph global setting for it\n",
1030 					       evsel->name);
1031 					return;
1032 				}
1033 				if (param.record_mode == CALLCHAIN_DWARF)
1034 					sample_address = true;
1035 			}
1036 		}
1037 		if (dump_size > 0) {
1038 			dump_size = round_up(dump_size, sizeof(u64));
1039 			param.dump_size = dump_size;
1040 		}
1041 
1042 		/* If global callgraph set, clear it */
1043 		if (callchain_param.enabled)
1044 			evsel__reset_callgraph(evsel, &callchain_param);
1045 
1046 		/* set perf-event callgraph */
1047 		if (param.enabled) {
1048 			if (sample_address) {
1049 				evsel__set_sample_bit(evsel, ADDR);
1050 				evsel__set_sample_bit(evsel, DATA_SRC);
1051 				evsel->core.attr.mmap_data = track;
1052 			}
1053 			evsel__config_callchain(evsel, opts, &param);
1054 		}
1055 	}
1056 }
1057 
1058 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1059 {
1060 	struct evsel_config_term *term, *found_term = NULL;
1061 
1062 	list_for_each_entry(term, &evsel->config_terms, list) {
1063 		if (term->type == type)
1064 			found_term = term;
1065 	}
1066 
1067 	return found_term;
1068 }
1069 
1070 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1071 {
1072 	evsel__set_sample_bit(evsel, WEIGHT);
1073 }
1074 
1075 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1076 				    struct perf_event_attr *attr __maybe_unused)
1077 {
1078 }
1079 
1080 static void evsel__set_default_freq_period(struct record_opts *opts,
1081 					   struct perf_event_attr *attr)
1082 {
1083 	if (opts->freq) {
1084 		attr->freq = 1;
1085 		attr->sample_freq = opts->freq;
1086 	} else {
1087 		attr->sample_period = opts->default_interval;
1088 	}
1089 }
1090 
1091 static bool evsel__is_offcpu_event(struct evsel *evsel)
1092 {
1093 	return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT);
1094 }
1095 
1096 /*
1097  * The enable_on_exec/disabled value strategy:
1098  *
1099  *  1) For any type of traced program:
1100  *    - all independent events and group leaders are disabled
1101  *    - all group members are enabled
1102  *
1103  *     Group members are ruled by group leaders. They need to
1104  *     be enabled, because the group scheduling relies on that.
1105  *
1106  *  2) For traced programs executed by perf:
1107  *     - all independent events and group leaders have
1108  *       enable_on_exec set
1109  *     - we don't specifically enable or disable any event during
1110  *       the record command
1111  *
1112  *     Independent events and group leaders are initially disabled
1113  *     and get enabled by exec. Group members are ruled by group
1114  *     leaders as stated in 1).
1115  *
1116  *  3) For traced programs attached by perf (pid/tid):
1117  *     - we specifically enable or disable all events during
1118  *       the record command
1119  *
1120  *     When attaching events to already running traced we
1121  *     enable/disable events specifically, as there's no
1122  *     initial traced exec call.
1123  */
1124 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1125 		   struct callchain_param *callchain)
1126 {
1127 	struct evsel *leader = evsel__leader(evsel);
1128 	struct perf_event_attr *attr = &evsel->core.attr;
1129 	int track = evsel->tracking;
1130 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1131 
1132 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1133 	attr->inherit	    = !opts->no_inherit;
1134 	attr->write_backward = opts->overwrite ? 1 : 0;
1135 	attr->read_format   = PERF_FORMAT_LOST;
1136 
1137 	evsel__set_sample_bit(evsel, IP);
1138 	evsel__set_sample_bit(evsel, TID);
1139 
1140 	if (evsel->sample_read) {
1141 		evsel__set_sample_bit(evsel, READ);
1142 
1143 		/*
1144 		 * We need ID even in case of single event, because
1145 		 * PERF_SAMPLE_READ process ID specific data.
1146 		 */
1147 		evsel__set_sample_id(evsel, false);
1148 
1149 		/*
1150 		 * Apply group format only if we belong to group
1151 		 * with more than one members.
1152 		 */
1153 		if (leader->core.nr_members > 1) {
1154 			attr->read_format |= PERF_FORMAT_GROUP;
1155 			attr->inherit = 0;
1156 		}
1157 	}
1158 
1159 	/*
1160 	 * We default some events to have a default interval. But keep
1161 	 * it a weak assumption overridable by the user.
1162 	 */
1163 	if ((evsel->is_libpfm_event && !attr->sample_period) ||
1164 	    (!evsel->is_libpfm_event && (!attr->sample_period ||
1165 					 opts->user_freq != UINT_MAX ||
1166 					 opts->user_interval != ULLONG_MAX)))
1167 		evsel__set_default_freq_period(opts, attr);
1168 
1169 	/*
1170 	 * If attr->freq was set (here or earlier), ask for period
1171 	 * to be sampled.
1172 	 */
1173 	if (attr->freq)
1174 		evsel__set_sample_bit(evsel, PERIOD);
1175 
1176 	if (opts->no_samples)
1177 		attr->sample_freq = 0;
1178 
1179 	if (opts->inherit_stat) {
1180 		evsel->core.attr.read_format |=
1181 			PERF_FORMAT_TOTAL_TIME_ENABLED |
1182 			PERF_FORMAT_TOTAL_TIME_RUNNING |
1183 			PERF_FORMAT_ID;
1184 		attr->inherit_stat = 1;
1185 	}
1186 
1187 	if (opts->sample_address) {
1188 		evsel__set_sample_bit(evsel, ADDR);
1189 		attr->mmap_data = track;
1190 	}
1191 
1192 	/*
1193 	 * We don't allow user space callchains for  function trace
1194 	 * event, due to issues with page faults while tracing page
1195 	 * fault handler and its overall trickiness nature.
1196 	 */
1197 	if (evsel__is_function_event(evsel))
1198 		evsel->core.attr.exclude_callchain_user = 1;
1199 
1200 	if (callchain && callchain->enabled && !evsel->no_aux_samples)
1201 		evsel__config_callchain(evsel, opts, callchain);
1202 
1203 	if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1204 	    !evsel__is_dummy_event(evsel)) {
1205 		attr->sample_regs_intr = opts->sample_intr_regs;
1206 		evsel__set_sample_bit(evsel, REGS_INTR);
1207 	}
1208 
1209 	if (opts->sample_user_regs && !evsel->no_aux_samples &&
1210 	    !evsel__is_dummy_event(evsel)) {
1211 		attr->sample_regs_user |= opts->sample_user_regs;
1212 		evsel__set_sample_bit(evsel, REGS_USER);
1213 	}
1214 
1215 	if (target__has_cpu(&opts->target) || opts->sample_cpu)
1216 		evsel__set_sample_bit(evsel, CPU);
1217 
1218 	/*
1219 	 * When the user explicitly disabled time don't force it here.
1220 	 */
1221 	if (opts->sample_time &&
1222 	    (!perf_missing_features.sample_id_all &&
1223 	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1224 	     opts->sample_time_set)))
1225 		evsel__set_sample_bit(evsel, TIME);
1226 
1227 	if (opts->raw_samples && !evsel->no_aux_samples) {
1228 		evsel__set_sample_bit(evsel, TIME);
1229 		evsel__set_sample_bit(evsel, RAW);
1230 		evsel__set_sample_bit(evsel, CPU);
1231 	}
1232 
1233 	if (opts->sample_address)
1234 		evsel__set_sample_bit(evsel, DATA_SRC);
1235 
1236 	if (opts->sample_phys_addr)
1237 		evsel__set_sample_bit(evsel, PHYS_ADDR);
1238 
1239 	if (opts->no_buffering) {
1240 		attr->watermark = 0;
1241 		attr->wakeup_events = 1;
1242 	}
1243 	if (opts->branch_stack && !evsel->no_aux_samples) {
1244 		evsel__set_sample_bit(evsel, BRANCH_STACK);
1245 		attr->branch_sample_type = opts->branch_stack;
1246 	}
1247 
1248 	if (opts->sample_weight)
1249 		arch_evsel__set_sample_weight(evsel);
1250 
1251 	attr->task     = track;
1252 	attr->mmap     = track;
1253 	attr->mmap2    = track && !perf_missing_features.mmap2;
1254 	attr->comm     = track;
1255 	attr->build_id = track && opts->build_id;
1256 
1257 	/*
1258 	 * ksymbol is tracked separately with text poke because it needs to be
1259 	 * system wide and enabled immediately.
1260 	 */
1261 	if (!opts->text_poke)
1262 		attr->ksymbol = track && !perf_missing_features.ksymbol;
1263 	attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1264 
1265 	if (opts->record_namespaces)
1266 		attr->namespaces  = track;
1267 
1268 	if (opts->record_cgroup) {
1269 		attr->cgroup = track && !perf_missing_features.cgroup;
1270 		evsel__set_sample_bit(evsel, CGROUP);
1271 	}
1272 
1273 	if (opts->sample_data_page_size)
1274 		evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1275 
1276 	if (opts->sample_code_page_size)
1277 		evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1278 
1279 	if (opts->record_switch_events)
1280 		attr->context_switch = track;
1281 
1282 	if (opts->sample_transaction)
1283 		evsel__set_sample_bit(evsel, TRANSACTION);
1284 
1285 	if (opts->running_time) {
1286 		evsel->core.attr.read_format |=
1287 			PERF_FORMAT_TOTAL_TIME_ENABLED |
1288 			PERF_FORMAT_TOTAL_TIME_RUNNING;
1289 	}
1290 
1291 	/*
1292 	 * XXX see the function comment above
1293 	 *
1294 	 * Disabling only independent events or group leaders,
1295 	 * keeping group members enabled.
1296 	 */
1297 	if (evsel__is_group_leader(evsel))
1298 		attr->disabled = 1;
1299 
1300 	/*
1301 	 * Setting enable_on_exec for independent events and
1302 	 * group leaders for traced executed by perf.
1303 	 */
1304 	if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1305 	    !opts->target.initial_delay)
1306 		attr->enable_on_exec = 1;
1307 
1308 	if (evsel->immediate) {
1309 		attr->disabled = 0;
1310 		attr->enable_on_exec = 0;
1311 	}
1312 
1313 	clockid = opts->clockid;
1314 	if (opts->use_clockid) {
1315 		attr->use_clockid = 1;
1316 		attr->clockid = opts->clockid;
1317 	}
1318 
1319 	if (evsel->precise_max)
1320 		attr->precise_ip = 3;
1321 
1322 	if (opts->all_user) {
1323 		attr->exclude_kernel = 1;
1324 		attr->exclude_user   = 0;
1325 	}
1326 
1327 	if (opts->all_kernel) {
1328 		attr->exclude_kernel = 0;
1329 		attr->exclude_user   = 1;
1330 	}
1331 
1332 	if (evsel->core.own_cpus || evsel->unit)
1333 		evsel->core.attr.read_format |= PERF_FORMAT_ID;
1334 
1335 	/*
1336 	 * Apply event specific term settings,
1337 	 * it overloads any global configuration.
1338 	 */
1339 	evsel__apply_config_terms(evsel, opts, track);
1340 
1341 	evsel->ignore_missing_thread = opts->ignore_missing_thread;
1342 
1343 	/* The --period option takes the precedence. */
1344 	if (opts->period_set) {
1345 		if (opts->period)
1346 			evsel__set_sample_bit(evsel, PERIOD);
1347 		else
1348 			evsel__reset_sample_bit(evsel, PERIOD);
1349 	}
1350 
1351 	/*
1352 	 * A dummy event never triggers any actual counter and therefore
1353 	 * cannot be used with branch_stack.
1354 	 *
1355 	 * For initial_delay, a dummy event is added implicitly.
1356 	 * The software event will trigger -EOPNOTSUPP error out,
1357 	 * if BRANCH_STACK bit is set.
1358 	 */
1359 	if (evsel__is_dummy_event(evsel))
1360 		evsel__reset_sample_bit(evsel, BRANCH_STACK);
1361 
1362 	if (evsel__is_offcpu_event(evsel))
1363 		evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1364 
1365 	arch__post_evsel_config(evsel, attr);
1366 }
1367 
1368 int evsel__set_filter(struct evsel *evsel, const char *filter)
1369 {
1370 	char *new_filter = strdup(filter);
1371 
1372 	if (new_filter != NULL) {
1373 		free(evsel->filter);
1374 		evsel->filter = new_filter;
1375 		return 0;
1376 	}
1377 
1378 	return -1;
1379 }
1380 
1381 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1382 {
1383 	char *new_filter;
1384 
1385 	if (evsel->filter == NULL)
1386 		return evsel__set_filter(evsel, filter);
1387 
1388 	if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1389 		free(evsel->filter);
1390 		evsel->filter = new_filter;
1391 		return 0;
1392 	}
1393 
1394 	return -1;
1395 }
1396 
1397 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1398 {
1399 	return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1400 }
1401 
1402 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1403 {
1404 	return evsel__append_filter(evsel, "%s,%s", filter);
1405 }
1406 
1407 /* Caller has to clear disabled after going through all CPUs. */
1408 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1409 {
1410 	return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1411 }
1412 
1413 int evsel__enable(struct evsel *evsel)
1414 {
1415 	int err = perf_evsel__enable(&evsel->core);
1416 
1417 	if (!err)
1418 		evsel->disabled = false;
1419 	return err;
1420 }
1421 
1422 /* Caller has to set disabled after going through all CPUs. */
1423 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1424 {
1425 	return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1426 }
1427 
1428 int evsel__disable(struct evsel *evsel)
1429 {
1430 	int err = perf_evsel__disable(&evsel->core);
1431 	/*
1432 	 * We mark it disabled here so that tools that disable a event can
1433 	 * ignore events after they disable it. I.e. the ring buffer may have
1434 	 * already a few more events queued up before the kernel got the stop
1435 	 * request.
1436 	 */
1437 	if (!err)
1438 		evsel->disabled = true;
1439 
1440 	return err;
1441 }
1442 
1443 void free_config_terms(struct list_head *config_terms)
1444 {
1445 	struct evsel_config_term *term, *h;
1446 
1447 	list_for_each_entry_safe(term, h, config_terms, list) {
1448 		list_del_init(&term->list);
1449 		if (term->free_str)
1450 			zfree(&term->val.str);
1451 		free(term);
1452 	}
1453 }
1454 
1455 static void evsel__free_config_terms(struct evsel *evsel)
1456 {
1457 	free_config_terms(&evsel->config_terms);
1458 }
1459 
1460 void evsel__exit(struct evsel *evsel)
1461 {
1462 	assert(list_empty(&evsel->core.node));
1463 	assert(evsel->evlist == NULL);
1464 	bpf_counter__destroy(evsel);
1465 	perf_bpf_filter__destroy(evsel);
1466 	evsel__free_counts(evsel);
1467 	perf_evsel__free_fd(&evsel->core);
1468 	perf_evsel__free_id(&evsel->core);
1469 	evsel__free_config_terms(evsel);
1470 	cgroup__put(evsel->cgrp);
1471 	perf_cpu_map__put(evsel->core.cpus);
1472 	perf_cpu_map__put(evsel->core.own_cpus);
1473 	perf_thread_map__put(evsel->core.threads);
1474 	zfree(&evsel->group_name);
1475 	zfree(&evsel->name);
1476 	zfree(&evsel->pmu_name);
1477 	zfree(&evsel->group_pmu_name);
1478 	zfree(&evsel->unit);
1479 	zfree(&evsel->metric_id);
1480 	evsel__zero_per_pkg(evsel);
1481 	hashmap__free(evsel->per_pkg_mask);
1482 	evsel->per_pkg_mask = NULL;
1483 	zfree(&evsel->metric_events);
1484 	perf_evsel__object.fini(evsel);
1485 }
1486 
1487 void evsel__delete(struct evsel *evsel)
1488 {
1489 	if (!evsel)
1490 		return;
1491 
1492 	evsel__exit(evsel);
1493 	free(evsel);
1494 }
1495 
1496 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1497 			   struct perf_counts_values *count)
1498 {
1499 	struct perf_counts_values tmp;
1500 
1501 	if (!evsel->prev_raw_counts)
1502 		return;
1503 
1504 	tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1505 	*perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1506 
1507 	count->val = count->val - tmp.val;
1508 	count->ena = count->ena - tmp.ena;
1509 	count->run = count->run - tmp.run;
1510 }
1511 
1512 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1513 {
1514 	struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1515 
1516 	return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1517 }
1518 
1519 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1520 			     u64 val, u64 ena, u64 run, u64 lost)
1521 {
1522 	struct perf_counts_values *count;
1523 
1524 	count = perf_counts(counter->counts, cpu_map_idx, thread);
1525 
1526 	count->val    = val;
1527 	count->ena    = ena;
1528 	count->run    = run;
1529 	count->lost   = lost;
1530 
1531 	perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1532 }
1533 
1534 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1535 {
1536 	u64 read_format = leader->core.attr.read_format;
1537 	struct sample_read_value *v;
1538 	u64 nr, ena = 0, run = 0, lost = 0;
1539 
1540 	nr = *data++;
1541 
1542 	if (nr != (u64) leader->core.nr_members)
1543 		return -EINVAL;
1544 
1545 	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1546 		ena = *data++;
1547 
1548 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1549 		run = *data++;
1550 
1551 	v = (void *)data;
1552 	sample_read_group__for_each(v, nr, read_format) {
1553 		struct evsel *counter;
1554 
1555 		counter = evlist__id2evsel(leader->evlist, v->id);
1556 		if (!counter)
1557 			return -EINVAL;
1558 
1559 		if (read_format & PERF_FORMAT_LOST)
1560 			lost = v->lost;
1561 
1562 		evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1563 	}
1564 
1565 	return 0;
1566 }
1567 
1568 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1569 {
1570 	struct perf_stat_evsel *ps = leader->stats;
1571 	u64 read_format = leader->core.attr.read_format;
1572 	int size = perf_evsel__read_size(&leader->core);
1573 	u64 *data = ps->group_data;
1574 
1575 	if (!(read_format & PERF_FORMAT_ID))
1576 		return -EINVAL;
1577 
1578 	if (!evsel__is_group_leader(leader))
1579 		return -EINVAL;
1580 
1581 	if (!data) {
1582 		data = zalloc(size);
1583 		if (!data)
1584 			return -ENOMEM;
1585 
1586 		ps->group_data = data;
1587 	}
1588 
1589 	if (FD(leader, cpu_map_idx, thread) < 0)
1590 		return -EINVAL;
1591 
1592 	if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1593 		return -errno;
1594 
1595 	return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1596 }
1597 
1598 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1599 {
1600 	u64 read_format = evsel->core.attr.read_format;
1601 
1602 	if (read_format & PERF_FORMAT_GROUP)
1603 		return evsel__read_group(evsel, cpu_map_idx, thread);
1604 
1605 	return evsel__read_one(evsel, cpu_map_idx, thread);
1606 }
1607 
1608 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1609 {
1610 	struct perf_counts_values count;
1611 	size_t nv = scale ? 3 : 1;
1612 
1613 	if (FD(evsel, cpu_map_idx, thread) < 0)
1614 		return -EINVAL;
1615 
1616 	if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1617 		return -ENOMEM;
1618 
1619 	if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1620 		return -errno;
1621 
1622 	evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1623 	perf_counts_values__scale(&count, scale, NULL);
1624 	*perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1625 	return 0;
1626 }
1627 
1628 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1629 				  int cpu_map_idx)
1630 {
1631 	struct perf_cpu cpu;
1632 
1633 	cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1634 	return perf_cpu_map__idx(other->core.cpus, cpu);
1635 }
1636 
1637 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1638 {
1639 	struct evsel *leader = evsel__leader(evsel);
1640 
1641 	if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1642 	    (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1643 		return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1644 	}
1645 
1646 	return cpu_map_idx;
1647 }
1648 
1649 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1650 {
1651 	struct evsel *leader = evsel__leader(evsel);
1652 	int fd;
1653 
1654 	if (evsel__is_group_leader(evsel))
1655 		return -1;
1656 
1657 	/*
1658 	 * Leader must be already processed/open,
1659 	 * if not it's a bug.
1660 	 */
1661 	BUG_ON(!leader->core.fd);
1662 
1663 	cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1664 	if (cpu_map_idx == -1)
1665 		return -1;
1666 
1667 	fd = FD(leader, cpu_map_idx, thread);
1668 	BUG_ON(fd == -1 && !leader->skippable);
1669 
1670 	/*
1671 	 * When the leader has been skipped, return -2 to distinguish from no
1672 	 * group leader case.
1673 	 */
1674 	return fd == -1 ? -2 : fd;
1675 }
1676 
1677 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1678 {
1679 	for (int cpu = 0; cpu < nr_cpus; cpu++)
1680 		for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1681 			FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1682 }
1683 
1684 static int update_fds(struct evsel *evsel,
1685 		      int nr_cpus, int cpu_map_idx,
1686 		      int nr_threads, int thread_idx)
1687 {
1688 	struct evsel *pos;
1689 
1690 	if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1691 		return -EINVAL;
1692 
1693 	evlist__for_each_entry(evsel->evlist, pos) {
1694 		nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1695 
1696 		evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1697 
1698 		/*
1699 		 * Since fds for next evsel has not been created,
1700 		 * there is no need to iterate whole event list.
1701 		 */
1702 		if (pos == evsel)
1703 			break;
1704 	}
1705 	return 0;
1706 }
1707 
1708 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1709 					 int nr_cpus, int cpu_map_idx,
1710 					 struct perf_thread_map *threads,
1711 					 int thread, int err)
1712 {
1713 	pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1714 
1715 	if (!evsel->ignore_missing_thread)
1716 		return false;
1717 
1718 	/* The system wide setup does not work with threads. */
1719 	if (evsel->core.system_wide)
1720 		return false;
1721 
1722 	/* The -ESRCH is perf event syscall errno for pid's not found. */
1723 	if (err != -ESRCH)
1724 		return false;
1725 
1726 	/* If there's only one thread, let it fail. */
1727 	if (threads->nr == 1)
1728 		return false;
1729 
1730 	/*
1731 	 * We should remove fd for missing_thread first
1732 	 * because thread_map__remove() will decrease threads->nr.
1733 	 */
1734 	if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1735 		return false;
1736 
1737 	if (thread_map__remove(threads, thread))
1738 		return false;
1739 
1740 	pr_warning("WARNING: Ignored open failure for pid %d\n",
1741 		   ignore_pid);
1742 	return true;
1743 }
1744 
1745 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1746 				void *priv __maybe_unused)
1747 {
1748 	return fprintf(fp, "  %-32s %s\n", name, val);
1749 }
1750 
1751 static void display_attr(struct perf_event_attr *attr)
1752 {
1753 	if (verbose >= 2 || debug_peo_args) {
1754 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1755 		fprintf(stderr, "perf_event_attr:\n");
1756 		perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1757 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1758 	}
1759 }
1760 
1761 bool evsel__precise_ip_fallback(struct evsel *evsel)
1762 {
1763 	/* Do not try less precise if not requested. */
1764 	if (!evsel->precise_max)
1765 		return false;
1766 
1767 	/*
1768 	 * We tried all the precise_ip values, and it's
1769 	 * still failing, so leave it to standard fallback.
1770 	 */
1771 	if (!evsel->core.attr.precise_ip) {
1772 		evsel->core.attr.precise_ip = evsel->precise_ip_original;
1773 		return false;
1774 	}
1775 
1776 	if (!evsel->precise_ip_original)
1777 		evsel->precise_ip_original = evsel->core.attr.precise_ip;
1778 
1779 	evsel->core.attr.precise_ip--;
1780 	pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1781 	display_attr(&evsel->core.attr);
1782 	return true;
1783 }
1784 
1785 static struct perf_cpu_map *empty_cpu_map;
1786 static struct perf_thread_map *empty_thread_map;
1787 
1788 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1789 		struct perf_thread_map *threads)
1790 {
1791 	int nthreads = perf_thread_map__nr(threads);
1792 
1793 	if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1794 	    (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1795 		return -EINVAL;
1796 
1797 	if (cpus == NULL) {
1798 		if (empty_cpu_map == NULL) {
1799 			empty_cpu_map = perf_cpu_map__dummy_new();
1800 			if (empty_cpu_map == NULL)
1801 				return -ENOMEM;
1802 		}
1803 
1804 		cpus = empty_cpu_map;
1805 	}
1806 
1807 	if (threads == NULL) {
1808 		if (empty_thread_map == NULL) {
1809 			empty_thread_map = thread_map__new_by_tid(-1);
1810 			if (empty_thread_map == NULL)
1811 				return -ENOMEM;
1812 		}
1813 
1814 		threads = empty_thread_map;
1815 	}
1816 
1817 	if (evsel->core.fd == NULL &&
1818 	    perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1819 		return -ENOMEM;
1820 
1821 	evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1822 	if (evsel->cgrp)
1823 		evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1824 
1825 	return 0;
1826 }
1827 
1828 static void evsel__disable_missing_features(struct evsel *evsel)
1829 {
1830 	if (perf_missing_features.read_lost)
1831 		evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1832 	if (perf_missing_features.weight_struct) {
1833 		evsel__set_sample_bit(evsel, WEIGHT);
1834 		evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1835 	}
1836 	if (perf_missing_features.clockid_wrong)
1837 		evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1838 	if (perf_missing_features.clockid) {
1839 		evsel->core.attr.use_clockid = 0;
1840 		evsel->core.attr.clockid = 0;
1841 	}
1842 	if (perf_missing_features.cloexec)
1843 		evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1844 	if (perf_missing_features.mmap2)
1845 		evsel->core.attr.mmap2 = 0;
1846 	if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1847 		evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1848 	if (perf_missing_features.lbr_flags)
1849 		evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1850 				     PERF_SAMPLE_BRANCH_NO_CYCLES);
1851 	if (perf_missing_features.group_read && evsel->core.attr.inherit)
1852 		evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1853 	if (perf_missing_features.ksymbol)
1854 		evsel->core.attr.ksymbol = 0;
1855 	if (perf_missing_features.bpf)
1856 		evsel->core.attr.bpf_event = 0;
1857 	if (perf_missing_features.branch_hw_idx)
1858 		evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1859 	if (perf_missing_features.sample_id_all)
1860 		evsel->core.attr.sample_id_all = 0;
1861 }
1862 
1863 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1864 			struct perf_thread_map *threads)
1865 {
1866 	int err;
1867 
1868 	err = __evsel__prepare_open(evsel, cpus, threads);
1869 	if (err)
1870 		return err;
1871 
1872 	evsel__disable_missing_features(evsel);
1873 
1874 	return err;
1875 }
1876 
1877 bool evsel__detect_missing_features(struct evsel *evsel)
1878 {
1879 	/*
1880 	 * Must probe features in the order they were added to the
1881 	 * perf_event_attr interface.
1882 	 */
1883 	if (!perf_missing_features.read_lost &&
1884 	    (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1885 		perf_missing_features.read_lost = true;
1886 		pr_debug2("switching off PERF_FORMAT_LOST support\n");
1887 		return true;
1888 	} else if (!perf_missing_features.weight_struct &&
1889 	    (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1890 		perf_missing_features.weight_struct = true;
1891 		pr_debug2("switching off weight struct support\n");
1892 		return true;
1893 	} else if (!perf_missing_features.code_page_size &&
1894 	    (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1895 		perf_missing_features.code_page_size = true;
1896 		pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1897 		return false;
1898 	} else if (!perf_missing_features.data_page_size &&
1899 	    (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1900 		perf_missing_features.data_page_size = true;
1901 		pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1902 		return false;
1903 	} else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1904 		perf_missing_features.cgroup = true;
1905 		pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1906 		return false;
1907 	} else if (!perf_missing_features.branch_hw_idx &&
1908 	    (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1909 		perf_missing_features.branch_hw_idx = true;
1910 		pr_debug2("switching off branch HW index support\n");
1911 		return true;
1912 	} else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1913 		perf_missing_features.aux_output = true;
1914 		pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1915 		return false;
1916 	} else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1917 		perf_missing_features.bpf = true;
1918 		pr_debug2_peo("switching off bpf_event\n");
1919 		return true;
1920 	} else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1921 		perf_missing_features.ksymbol = true;
1922 		pr_debug2_peo("switching off ksymbol\n");
1923 		return true;
1924 	} else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1925 		perf_missing_features.write_backward = true;
1926 		pr_debug2_peo("switching off write_backward\n");
1927 		return false;
1928 	} else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1929 		perf_missing_features.clockid_wrong = true;
1930 		pr_debug2_peo("switching off clockid\n");
1931 		return true;
1932 	} else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1933 		perf_missing_features.clockid = true;
1934 		pr_debug2_peo("switching off use_clockid\n");
1935 		return true;
1936 	} else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1937 		perf_missing_features.cloexec = true;
1938 		pr_debug2_peo("switching off cloexec flag\n");
1939 		return true;
1940 	} else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1941 		perf_missing_features.mmap2 = true;
1942 		pr_debug2_peo("switching off mmap2\n");
1943 		return true;
1944 	} else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1945 		if (evsel->pmu == NULL)
1946 			evsel->pmu = evsel__find_pmu(evsel);
1947 
1948 		if (evsel->pmu)
1949 			evsel->pmu->missing_features.exclude_guest = true;
1950 		else {
1951 			/* we cannot find PMU, disable attrs now */
1952 			evsel->core.attr.exclude_host = false;
1953 			evsel->core.attr.exclude_guest = false;
1954 		}
1955 
1956 		if (evsel->exclude_GH) {
1957 			pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1958 			return false;
1959 		}
1960 		if (!perf_missing_features.exclude_guest) {
1961 			perf_missing_features.exclude_guest = true;
1962 			pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1963 		}
1964 		return true;
1965 	} else if (!perf_missing_features.sample_id_all) {
1966 		perf_missing_features.sample_id_all = true;
1967 		pr_debug2_peo("switching off sample_id_all\n");
1968 		return true;
1969 	} else if (!perf_missing_features.lbr_flags &&
1970 			(evsel->core.attr.branch_sample_type &
1971 			 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1972 			  PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1973 		perf_missing_features.lbr_flags = true;
1974 		pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1975 		return true;
1976 	} else if (!perf_missing_features.group_read &&
1977 		    evsel->core.attr.inherit &&
1978 		   (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1979 		   evsel__is_group_leader(evsel)) {
1980 		perf_missing_features.group_read = true;
1981 		pr_debug2_peo("switching off group read\n");
1982 		return true;
1983 	} else {
1984 		return false;
1985 	}
1986 }
1987 
1988 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
1989 {
1990 	int old_errno;
1991 	struct rlimit l;
1992 
1993 	if (*set_rlimit < INCREASED_MAX) {
1994 		old_errno = errno;
1995 
1996 		if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1997 			if (*set_rlimit == NO_CHANGE) {
1998 				l.rlim_cur = l.rlim_max;
1999 			} else {
2000 				l.rlim_cur = l.rlim_max + 1000;
2001 				l.rlim_max = l.rlim_cur;
2002 			}
2003 			if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2004 				(*set_rlimit) += 1;
2005 				errno = old_errno;
2006 				return true;
2007 			}
2008 		}
2009 		errno = old_errno;
2010 	}
2011 
2012 	return false;
2013 }
2014 
2015 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2016 		struct perf_thread_map *threads,
2017 		int start_cpu_map_idx, int end_cpu_map_idx)
2018 {
2019 	int idx, thread, nthreads;
2020 	int pid = -1, err, old_errno;
2021 	enum rlimit_action set_rlimit = NO_CHANGE;
2022 
2023 	err = __evsel__prepare_open(evsel, cpus, threads);
2024 	if (err)
2025 		return err;
2026 
2027 	if (cpus == NULL)
2028 		cpus = empty_cpu_map;
2029 
2030 	if (threads == NULL)
2031 		threads = empty_thread_map;
2032 
2033 	nthreads = perf_thread_map__nr(threads);
2034 
2035 	if (evsel->cgrp)
2036 		pid = evsel->cgrp->fd;
2037 
2038 fallback_missing_features:
2039 	evsel__disable_missing_features(evsel);
2040 
2041 	display_attr(&evsel->core.attr);
2042 
2043 	for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2044 
2045 		for (thread = 0; thread < nthreads; thread++) {
2046 			int fd, group_fd;
2047 retry_open:
2048 			if (thread >= nthreads)
2049 				break;
2050 
2051 			if (!evsel->cgrp && !evsel->core.system_wide)
2052 				pid = perf_thread_map__pid(threads, thread);
2053 
2054 			group_fd = get_group_fd(evsel, idx, thread);
2055 
2056 			if (group_fd == -2) {
2057 				pr_debug("broken group leader for %s\n", evsel->name);
2058 				err = -EINVAL;
2059 				goto out_close;
2060 			}
2061 
2062 			test_attr__ready();
2063 
2064 			/* Debug message used by test scripts */
2065 			pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2066 				pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2067 
2068 			fd = sys_perf_event_open(&evsel->core.attr, pid,
2069 						perf_cpu_map__cpu(cpus, idx).cpu,
2070 						group_fd, evsel->open_flags);
2071 
2072 			FD(evsel, idx, thread) = fd;
2073 
2074 			if (fd < 0) {
2075 				err = -errno;
2076 
2077 				pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2078 					  err);
2079 				goto try_fallback;
2080 			}
2081 
2082 			bpf_counter__install_pe(evsel, idx, fd);
2083 
2084 			if (unlikely(test_attr__enabled)) {
2085 				test_attr__open(&evsel->core.attr, pid,
2086 						perf_cpu_map__cpu(cpus, idx),
2087 						fd, group_fd, evsel->open_flags);
2088 			}
2089 
2090 			/* Debug message used by test scripts */
2091 			pr_debug2_peo(" = %d\n", fd);
2092 
2093 			if (evsel->bpf_fd >= 0) {
2094 				int evt_fd = fd;
2095 				int bpf_fd = evsel->bpf_fd;
2096 
2097 				err = ioctl(evt_fd,
2098 					    PERF_EVENT_IOC_SET_BPF,
2099 					    bpf_fd);
2100 				if (err && errno != EEXIST) {
2101 					pr_err("failed to attach bpf fd %d: %s\n",
2102 					       bpf_fd, strerror(errno));
2103 					err = -EINVAL;
2104 					goto out_close;
2105 				}
2106 			}
2107 
2108 			set_rlimit = NO_CHANGE;
2109 
2110 			/*
2111 			 * If we succeeded but had to kill clockid, fail and
2112 			 * have evsel__open_strerror() print us a nice error.
2113 			 */
2114 			if (perf_missing_features.clockid ||
2115 			    perf_missing_features.clockid_wrong) {
2116 				err = -EINVAL;
2117 				goto out_close;
2118 			}
2119 		}
2120 	}
2121 
2122 	return 0;
2123 
2124 try_fallback:
2125 	if (evsel__precise_ip_fallback(evsel))
2126 		goto retry_open;
2127 
2128 	if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2129 					 idx, threads, thread, err)) {
2130 		/* We just removed 1 thread, so lower the upper nthreads limit. */
2131 		nthreads--;
2132 
2133 		/* ... and pretend like nothing have happened. */
2134 		err = 0;
2135 		goto retry_open;
2136 	}
2137 	/*
2138 	 * perf stat needs between 5 and 22 fds per CPU. When we run out
2139 	 * of them try to increase the limits.
2140 	 */
2141 	if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2142 		goto retry_open;
2143 
2144 	if (err != -EINVAL || idx > 0 || thread > 0)
2145 		goto out_close;
2146 
2147 	if (evsel__detect_missing_features(evsel))
2148 		goto fallback_missing_features;
2149 out_close:
2150 	if (err)
2151 		threads->err_thread = thread;
2152 
2153 	old_errno = errno;
2154 	do {
2155 		while (--thread >= 0) {
2156 			if (FD(evsel, idx, thread) >= 0)
2157 				close(FD(evsel, idx, thread));
2158 			FD(evsel, idx, thread) = -1;
2159 		}
2160 		thread = nthreads;
2161 	} while (--idx >= 0);
2162 	errno = old_errno;
2163 	return err;
2164 }
2165 
2166 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2167 		struct perf_thread_map *threads)
2168 {
2169 	return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2170 }
2171 
2172 void evsel__close(struct evsel *evsel)
2173 {
2174 	perf_evsel__close(&evsel->core);
2175 	perf_evsel__free_id(&evsel->core);
2176 }
2177 
2178 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2179 {
2180 	if (cpu_map_idx == -1)
2181 		return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2182 
2183 	return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2184 }
2185 
2186 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2187 {
2188 	return evsel__open(evsel, NULL, threads);
2189 }
2190 
2191 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2192 				       const union perf_event *event,
2193 				       struct perf_sample *sample)
2194 {
2195 	u64 type = evsel->core.attr.sample_type;
2196 	const __u64 *array = event->sample.array;
2197 	bool swapped = evsel->needs_swap;
2198 	union u64_swap u;
2199 
2200 	array += ((event->header.size -
2201 		   sizeof(event->header)) / sizeof(u64)) - 1;
2202 
2203 	if (type & PERF_SAMPLE_IDENTIFIER) {
2204 		sample->id = *array;
2205 		array--;
2206 	}
2207 
2208 	if (type & PERF_SAMPLE_CPU) {
2209 		u.val64 = *array;
2210 		if (swapped) {
2211 			/* undo swap of u64, then swap on individual u32s */
2212 			u.val64 = bswap_64(u.val64);
2213 			u.val32[0] = bswap_32(u.val32[0]);
2214 		}
2215 
2216 		sample->cpu = u.val32[0];
2217 		array--;
2218 	}
2219 
2220 	if (type & PERF_SAMPLE_STREAM_ID) {
2221 		sample->stream_id = *array;
2222 		array--;
2223 	}
2224 
2225 	if (type & PERF_SAMPLE_ID) {
2226 		sample->id = *array;
2227 		array--;
2228 	}
2229 
2230 	if (type & PERF_SAMPLE_TIME) {
2231 		sample->time = *array;
2232 		array--;
2233 	}
2234 
2235 	if (type & PERF_SAMPLE_TID) {
2236 		u.val64 = *array;
2237 		if (swapped) {
2238 			/* undo swap of u64, then swap on individual u32s */
2239 			u.val64 = bswap_64(u.val64);
2240 			u.val32[0] = bswap_32(u.val32[0]);
2241 			u.val32[1] = bswap_32(u.val32[1]);
2242 		}
2243 
2244 		sample->pid = u.val32[0];
2245 		sample->tid = u.val32[1];
2246 		array--;
2247 	}
2248 
2249 	return 0;
2250 }
2251 
2252 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2253 			    u64 size)
2254 {
2255 	return size > max_size || offset + size > endp;
2256 }
2257 
2258 #define OVERFLOW_CHECK(offset, size, max_size)				\
2259 	do {								\
2260 		if (overflow(endp, (max_size), (offset), (size)))	\
2261 			return -EFAULT;					\
2262 	} while (0)
2263 
2264 #define OVERFLOW_CHECK_u64(offset) \
2265 	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2266 
2267 static int
2268 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2269 {
2270 	/*
2271 	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2272 	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2273 	 * check the format does not go past the end of the event.
2274 	 */
2275 	if (sample_size + sizeof(event->header) > event->header.size)
2276 		return -EFAULT;
2277 
2278 	return 0;
2279 }
2280 
2281 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2282 					  const __u64 *array,
2283 					  u64 type __maybe_unused)
2284 {
2285 	data->weight = *array;
2286 }
2287 
2288 u64 evsel__bitfield_swap_branch_flags(u64 value)
2289 {
2290 	u64 new_val = 0;
2291 
2292 	/*
2293 	 * branch_flags
2294 	 * union {
2295 	 * 	u64 values;
2296 	 * 	struct {
2297 	 * 		mispred:1	//target mispredicted
2298 	 * 		predicted:1	//target predicted
2299 	 * 		in_tx:1		//in transaction
2300 	 * 		abort:1		//transaction abort
2301 	 * 		cycles:16	//cycle count to last branch
2302 	 * 		type:4		//branch type
2303 	 * 		spec:2		//branch speculation info
2304 	 * 		new_type:4	//additional branch type
2305 	 * 		priv:3		//privilege level
2306 	 * 		reserved:31
2307 	 * 	}
2308 	 * }
2309 	 *
2310 	 * Avoid bswap64() the entire branch_flag.value,
2311 	 * as it has variable bit-field sizes. Instead the
2312 	 * macro takes the bit-field position/size,
2313 	 * swaps it based on the host endianness.
2314 	 */
2315 	if (host_is_bigendian()) {
2316 		new_val = bitfield_swap(value, 0, 1);
2317 		new_val |= bitfield_swap(value, 1, 1);
2318 		new_val |= bitfield_swap(value, 2, 1);
2319 		new_val |= bitfield_swap(value, 3, 1);
2320 		new_val |= bitfield_swap(value, 4, 16);
2321 		new_val |= bitfield_swap(value, 20, 4);
2322 		new_val |= bitfield_swap(value, 24, 2);
2323 		new_val |= bitfield_swap(value, 26, 4);
2324 		new_val |= bitfield_swap(value, 30, 3);
2325 		new_val |= bitfield_swap(value, 33, 31);
2326 	} else {
2327 		new_val = bitfield_swap(value, 63, 1);
2328 		new_val |= bitfield_swap(value, 62, 1);
2329 		new_val |= bitfield_swap(value, 61, 1);
2330 		new_val |= bitfield_swap(value, 60, 1);
2331 		new_val |= bitfield_swap(value, 44, 16);
2332 		new_val |= bitfield_swap(value, 40, 4);
2333 		new_val |= bitfield_swap(value, 38, 2);
2334 		new_val |= bitfield_swap(value, 34, 4);
2335 		new_val |= bitfield_swap(value, 31, 3);
2336 		new_val |= bitfield_swap(value, 0, 31);
2337 	}
2338 
2339 	return new_val;
2340 }
2341 
2342 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2343 			struct perf_sample *data)
2344 {
2345 	u64 type = evsel->core.attr.sample_type;
2346 	bool swapped = evsel->needs_swap;
2347 	const __u64 *array;
2348 	u16 max_size = event->header.size;
2349 	const void *endp = (void *)event + max_size;
2350 	u64 sz;
2351 
2352 	/*
2353 	 * used for cross-endian analysis. See git commit 65014ab3
2354 	 * for why this goofiness is needed.
2355 	 */
2356 	union u64_swap u;
2357 
2358 	memset(data, 0, sizeof(*data));
2359 	data->cpu = data->pid = data->tid = -1;
2360 	data->stream_id = data->id = data->time = -1ULL;
2361 	data->period = evsel->core.attr.sample_period;
2362 	data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2363 	data->misc    = event->header.misc;
2364 	data->id = -1ULL;
2365 	data->data_src = PERF_MEM_DATA_SRC_NONE;
2366 	data->vcpu = -1;
2367 
2368 	if (event->header.type != PERF_RECORD_SAMPLE) {
2369 		if (!evsel->core.attr.sample_id_all)
2370 			return 0;
2371 		return perf_evsel__parse_id_sample(evsel, event, data);
2372 	}
2373 
2374 	array = event->sample.array;
2375 
2376 	if (perf_event__check_size(event, evsel->sample_size))
2377 		return -EFAULT;
2378 
2379 	if (type & PERF_SAMPLE_IDENTIFIER) {
2380 		data->id = *array;
2381 		array++;
2382 	}
2383 
2384 	if (type & PERF_SAMPLE_IP) {
2385 		data->ip = *array;
2386 		array++;
2387 	}
2388 
2389 	if (type & PERF_SAMPLE_TID) {
2390 		u.val64 = *array;
2391 		if (swapped) {
2392 			/* undo swap of u64, then swap on individual u32s */
2393 			u.val64 = bswap_64(u.val64);
2394 			u.val32[0] = bswap_32(u.val32[0]);
2395 			u.val32[1] = bswap_32(u.val32[1]);
2396 		}
2397 
2398 		data->pid = u.val32[0];
2399 		data->tid = u.val32[1];
2400 		array++;
2401 	}
2402 
2403 	if (type & PERF_SAMPLE_TIME) {
2404 		data->time = *array;
2405 		array++;
2406 	}
2407 
2408 	if (type & PERF_SAMPLE_ADDR) {
2409 		data->addr = *array;
2410 		array++;
2411 	}
2412 
2413 	if (type & PERF_SAMPLE_ID) {
2414 		data->id = *array;
2415 		array++;
2416 	}
2417 
2418 	if (type & PERF_SAMPLE_STREAM_ID) {
2419 		data->stream_id = *array;
2420 		array++;
2421 	}
2422 
2423 	if (type & PERF_SAMPLE_CPU) {
2424 
2425 		u.val64 = *array;
2426 		if (swapped) {
2427 			/* undo swap of u64, then swap on individual u32s */
2428 			u.val64 = bswap_64(u.val64);
2429 			u.val32[0] = bswap_32(u.val32[0]);
2430 		}
2431 
2432 		data->cpu = u.val32[0];
2433 		array++;
2434 	}
2435 
2436 	if (type & PERF_SAMPLE_PERIOD) {
2437 		data->period = *array;
2438 		array++;
2439 	}
2440 
2441 	if (type & PERF_SAMPLE_READ) {
2442 		u64 read_format = evsel->core.attr.read_format;
2443 
2444 		OVERFLOW_CHECK_u64(array);
2445 		if (read_format & PERF_FORMAT_GROUP)
2446 			data->read.group.nr = *array;
2447 		else
2448 			data->read.one.value = *array;
2449 
2450 		array++;
2451 
2452 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2453 			OVERFLOW_CHECK_u64(array);
2454 			data->read.time_enabled = *array;
2455 			array++;
2456 		}
2457 
2458 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2459 			OVERFLOW_CHECK_u64(array);
2460 			data->read.time_running = *array;
2461 			array++;
2462 		}
2463 
2464 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2465 		if (read_format & PERF_FORMAT_GROUP) {
2466 			const u64 max_group_nr = UINT64_MAX /
2467 					sizeof(struct sample_read_value);
2468 
2469 			if (data->read.group.nr > max_group_nr)
2470 				return -EFAULT;
2471 
2472 			sz = data->read.group.nr * sample_read_value_size(read_format);
2473 			OVERFLOW_CHECK(array, sz, max_size);
2474 			data->read.group.values =
2475 					(struct sample_read_value *)array;
2476 			array = (void *)array + sz;
2477 		} else {
2478 			OVERFLOW_CHECK_u64(array);
2479 			data->read.one.id = *array;
2480 			array++;
2481 
2482 			if (read_format & PERF_FORMAT_LOST) {
2483 				OVERFLOW_CHECK_u64(array);
2484 				data->read.one.lost = *array;
2485 				array++;
2486 			}
2487 		}
2488 	}
2489 
2490 	if (type & PERF_SAMPLE_CALLCHAIN) {
2491 		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2492 
2493 		OVERFLOW_CHECK_u64(array);
2494 		data->callchain = (struct ip_callchain *)array++;
2495 		if (data->callchain->nr > max_callchain_nr)
2496 			return -EFAULT;
2497 		sz = data->callchain->nr * sizeof(u64);
2498 		OVERFLOW_CHECK(array, sz, max_size);
2499 		array = (void *)array + sz;
2500 	}
2501 
2502 	if (type & PERF_SAMPLE_RAW) {
2503 		OVERFLOW_CHECK_u64(array);
2504 		u.val64 = *array;
2505 
2506 		/*
2507 		 * Undo swap of u64, then swap on individual u32s,
2508 		 * get the size of the raw area and undo all of the
2509 		 * swap. The pevent interface handles endianness by
2510 		 * itself.
2511 		 */
2512 		if (swapped) {
2513 			u.val64 = bswap_64(u.val64);
2514 			u.val32[0] = bswap_32(u.val32[0]);
2515 			u.val32[1] = bswap_32(u.val32[1]);
2516 		}
2517 		data->raw_size = u.val32[0];
2518 
2519 		/*
2520 		 * The raw data is aligned on 64bits including the
2521 		 * u32 size, so it's safe to use mem_bswap_64.
2522 		 */
2523 		if (swapped)
2524 			mem_bswap_64((void *) array, data->raw_size);
2525 
2526 		array = (void *)array + sizeof(u32);
2527 
2528 		OVERFLOW_CHECK(array, data->raw_size, max_size);
2529 		data->raw_data = (void *)array;
2530 		array = (void *)array + data->raw_size;
2531 	}
2532 
2533 	if (type & PERF_SAMPLE_BRANCH_STACK) {
2534 		const u64 max_branch_nr = UINT64_MAX /
2535 					  sizeof(struct branch_entry);
2536 		struct branch_entry *e;
2537 		unsigned int i;
2538 
2539 		OVERFLOW_CHECK_u64(array);
2540 		data->branch_stack = (struct branch_stack *)array++;
2541 
2542 		if (data->branch_stack->nr > max_branch_nr)
2543 			return -EFAULT;
2544 
2545 		sz = data->branch_stack->nr * sizeof(struct branch_entry);
2546 		if (evsel__has_branch_hw_idx(evsel)) {
2547 			sz += sizeof(u64);
2548 			e = &data->branch_stack->entries[0];
2549 		} else {
2550 			data->no_hw_idx = true;
2551 			/*
2552 			 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2553 			 * only nr and entries[] will be output by kernel.
2554 			 */
2555 			e = (struct branch_entry *)&data->branch_stack->hw_idx;
2556 		}
2557 
2558 		if (swapped) {
2559 			/*
2560 			 * struct branch_flag does not have endian
2561 			 * specific bit field definition. And bswap
2562 			 * will not resolve the issue, since these
2563 			 * are bit fields.
2564 			 *
2565 			 * evsel__bitfield_swap_branch_flags() uses a
2566 			 * bitfield_swap macro to swap the bit position
2567 			 * based on the host endians.
2568 			 */
2569 			for (i = 0; i < data->branch_stack->nr; i++, e++)
2570 				e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2571 		}
2572 
2573 		OVERFLOW_CHECK(array, sz, max_size);
2574 		array = (void *)array + sz;
2575 	}
2576 
2577 	if (type & PERF_SAMPLE_REGS_USER) {
2578 		OVERFLOW_CHECK_u64(array);
2579 		data->user_regs.abi = *array;
2580 		array++;
2581 
2582 		if (data->user_regs.abi) {
2583 			u64 mask = evsel->core.attr.sample_regs_user;
2584 
2585 			sz = hweight64(mask) * sizeof(u64);
2586 			OVERFLOW_CHECK(array, sz, max_size);
2587 			data->user_regs.mask = mask;
2588 			data->user_regs.regs = (u64 *)array;
2589 			array = (void *)array + sz;
2590 		}
2591 	}
2592 
2593 	if (type & PERF_SAMPLE_STACK_USER) {
2594 		OVERFLOW_CHECK_u64(array);
2595 		sz = *array++;
2596 
2597 		data->user_stack.offset = ((char *)(array - 1)
2598 					  - (char *) event);
2599 
2600 		if (!sz) {
2601 			data->user_stack.size = 0;
2602 		} else {
2603 			OVERFLOW_CHECK(array, sz, max_size);
2604 			data->user_stack.data = (char *)array;
2605 			array = (void *)array + sz;
2606 			OVERFLOW_CHECK_u64(array);
2607 			data->user_stack.size = *array++;
2608 			if (WARN_ONCE(data->user_stack.size > sz,
2609 				      "user stack dump failure\n"))
2610 				return -EFAULT;
2611 		}
2612 	}
2613 
2614 	if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2615 		OVERFLOW_CHECK_u64(array);
2616 		arch_perf_parse_sample_weight(data, array, type);
2617 		array++;
2618 	}
2619 
2620 	if (type & PERF_SAMPLE_DATA_SRC) {
2621 		OVERFLOW_CHECK_u64(array);
2622 		data->data_src = *array;
2623 		array++;
2624 	}
2625 
2626 	if (type & PERF_SAMPLE_TRANSACTION) {
2627 		OVERFLOW_CHECK_u64(array);
2628 		data->transaction = *array;
2629 		array++;
2630 	}
2631 
2632 	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2633 	if (type & PERF_SAMPLE_REGS_INTR) {
2634 		OVERFLOW_CHECK_u64(array);
2635 		data->intr_regs.abi = *array;
2636 		array++;
2637 
2638 		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2639 			u64 mask = evsel->core.attr.sample_regs_intr;
2640 
2641 			sz = hweight64(mask) * sizeof(u64);
2642 			OVERFLOW_CHECK(array, sz, max_size);
2643 			data->intr_regs.mask = mask;
2644 			data->intr_regs.regs = (u64 *)array;
2645 			array = (void *)array + sz;
2646 		}
2647 	}
2648 
2649 	data->phys_addr = 0;
2650 	if (type & PERF_SAMPLE_PHYS_ADDR) {
2651 		data->phys_addr = *array;
2652 		array++;
2653 	}
2654 
2655 	data->cgroup = 0;
2656 	if (type & PERF_SAMPLE_CGROUP) {
2657 		data->cgroup = *array;
2658 		array++;
2659 	}
2660 
2661 	data->data_page_size = 0;
2662 	if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2663 		data->data_page_size = *array;
2664 		array++;
2665 	}
2666 
2667 	data->code_page_size = 0;
2668 	if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2669 		data->code_page_size = *array;
2670 		array++;
2671 	}
2672 
2673 	if (type & PERF_SAMPLE_AUX) {
2674 		OVERFLOW_CHECK_u64(array);
2675 		sz = *array++;
2676 
2677 		OVERFLOW_CHECK(array, sz, max_size);
2678 		/* Undo swap of data */
2679 		if (swapped)
2680 			mem_bswap_64((char *)array, sz);
2681 		data->aux_sample.size = sz;
2682 		data->aux_sample.data = (char *)array;
2683 		array = (void *)array + sz;
2684 	}
2685 
2686 	return 0;
2687 }
2688 
2689 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2690 				  u64 *timestamp)
2691 {
2692 	u64 type = evsel->core.attr.sample_type;
2693 	const __u64 *array;
2694 
2695 	if (!(type & PERF_SAMPLE_TIME))
2696 		return -1;
2697 
2698 	if (event->header.type != PERF_RECORD_SAMPLE) {
2699 		struct perf_sample data = {
2700 			.time = -1ULL,
2701 		};
2702 
2703 		if (!evsel->core.attr.sample_id_all)
2704 			return -1;
2705 		if (perf_evsel__parse_id_sample(evsel, event, &data))
2706 			return -1;
2707 
2708 		*timestamp = data.time;
2709 		return 0;
2710 	}
2711 
2712 	array = event->sample.array;
2713 
2714 	if (perf_event__check_size(event, evsel->sample_size))
2715 		return -EFAULT;
2716 
2717 	if (type & PERF_SAMPLE_IDENTIFIER)
2718 		array++;
2719 
2720 	if (type & PERF_SAMPLE_IP)
2721 		array++;
2722 
2723 	if (type & PERF_SAMPLE_TID)
2724 		array++;
2725 
2726 	if (type & PERF_SAMPLE_TIME)
2727 		*timestamp = *array;
2728 
2729 	return 0;
2730 }
2731 
2732 u16 evsel__id_hdr_size(struct evsel *evsel)
2733 {
2734 	u64 sample_type = evsel->core.attr.sample_type;
2735 	u16 size = 0;
2736 
2737 	if (sample_type & PERF_SAMPLE_TID)
2738 		size += sizeof(u64);
2739 
2740 	if (sample_type & PERF_SAMPLE_TIME)
2741 		size += sizeof(u64);
2742 
2743 	if (sample_type & PERF_SAMPLE_ID)
2744 		size += sizeof(u64);
2745 
2746 	if (sample_type & PERF_SAMPLE_STREAM_ID)
2747 		size += sizeof(u64);
2748 
2749 	if (sample_type & PERF_SAMPLE_CPU)
2750 		size += sizeof(u64);
2751 
2752 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
2753 		size += sizeof(u64);
2754 
2755 	return size;
2756 }
2757 
2758 #ifdef HAVE_LIBTRACEEVENT
2759 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2760 {
2761 	return tep_find_field(evsel->tp_format, name);
2762 }
2763 
2764 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2765 {
2766 	struct tep_format_field *field = evsel__field(evsel, name);
2767 	int offset;
2768 
2769 	if (!field)
2770 		return NULL;
2771 
2772 	offset = field->offset;
2773 
2774 	if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2775 		offset = *(int *)(sample->raw_data + field->offset);
2776 		offset &= 0xffff;
2777 		if (tep_field_is_relative(field->flags))
2778 			offset += field->offset + field->size;
2779 	}
2780 
2781 	return sample->raw_data + offset;
2782 }
2783 
2784 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2785 			 bool needs_swap)
2786 {
2787 	u64 value;
2788 	void *ptr = sample->raw_data + field->offset;
2789 
2790 	switch (field->size) {
2791 	case 1:
2792 		return *(u8 *)ptr;
2793 	case 2:
2794 		value = *(u16 *)ptr;
2795 		break;
2796 	case 4:
2797 		value = *(u32 *)ptr;
2798 		break;
2799 	case 8:
2800 		memcpy(&value, ptr, sizeof(u64));
2801 		break;
2802 	default:
2803 		return 0;
2804 	}
2805 
2806 	if (!needs_swap)
2807 		return value;
2808 
2809 	switch (field->size) {
2810 	case 2:
2811 		return bswap_16(value);
2812 	case 4:
2813 		return bswap_32(value);
2814 	case 8:
2815 		return bswap_64(value);
2816 	default:
2817 		return 0;
2818 	}
2819 
2820 	return 0;
2821 }
2822 
2823 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2824 {
2825 	struct tep_format_field *field = evsel__field(evsel, name);
2826 
2827 	if (!field)
2828 		return 0;
2829 
2830 	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2831 }
2832 #endif
2833 
2834 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2835 {
2836 	int paranoid;
2837 
2838 	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2839 	    evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2840 	    evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2841 		/*
2842 		 * If it's cycles then fall back to hrtimer based
2843 		 * cpu-clock-tick sw counter, which is always available even if
2844 		 * no PMU support.
2845 		 *
2846 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2847 		 * b0a873e).
2848 		 */
2849 		scnprintf(msg, msgsize, "%s",
2850 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2851 
2852 		evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2853 		evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2854 
2855 		zfree(&evsel->name);
2856 		return true;
2857 	} else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2858 		   (paranoid = perf_event_paranoid()) > 1) {
2859 		const char *name = evsel__name(evsel);
2860 		char *new_name;
2861 		const char *sep = ":";
2862 
2863 		/* If event has exclude user then don't exclude kernel. */
2864 		if (evsel->core.attr.exclude_user)
2865 			return false;
2866 
2867 		/* Is there already the separator in the name. */
2868 		if (strchr(name, '/') ||
2869 		    (strchr(name, ':') && !evsel->is_libpfm_event))
2870 			sep = "";
2871 
2872 		if (asprintf(&new_name, "%s%su", name, sep) < 0)
2873 			return false;
2874 
2875 		free(evsel->name);
2876 		evsel->name = new_name;
2877 		scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2878 			  "to fall back to excluding kernel and hypervisor "
2879 			  " samples", paranoid);
2880 		evsel->core.attr.exclude_kernel = 1;
2881 		evsel->core.attr.exclude_hv     = 1;
2882 
2883 		return true;
2884 	}
2885 
2886 	return false;
2887 }
2888 
2889 static bool find_process(const char *name)
2890 {
2891 	size_t len = strlen(name);
2892 	DIR *dir;
2893 	struct dirent *d;
2894 	int ret = -1;
2895 
2896 	dir = opendir(procfs__mountpoint());
2897 	if (!dir)
2898 		return false;
2899 
2900 	/* Walk through the directory. */
2901 	while (ret && (d = readdir(dir)) != NULL) {
2902 		char path[PATH_MAX];
2903 		char *data;
2904 		size_t size;
2905 
2906 		if ((d->d_type != DT_DIR) ||
2907 		     !strcmp(".", d->d_name) ||
2908 		     !strcmp("..", d->d_name))
2909 			continue;
2910 
2911 		scnprintf(path, sizeof(path), "%s/%s/comm",
2912 			  procfs__mountpoint(), d->d_name);
2913 
2914 		if (filename__read_str(path, &data, &size))
2915 			continue;
2916 
2917 		ret = strncmp(name, data, len);
2918 		free(data);
2919 	}
2920 
2921 	closedir(dir);
2922 	return ret ? false : true;
2923 }
2924 
2925 static bool is_amd(const char *arch, const char *cpuid)
2926 {
2927 	return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2928 }
2929 
2930 static bool is_amd_ibs(struct evsel *evsel)
2931 {
2932 	return evsel->core.attr.precise_ip
2933 	    || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2934 }
2935 
2936 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2937 			 int err, char *msg, size_t size)
2938 {
2939 	struct perf_env *env = evsel__env(evsel);
2940 	const char *arch = perf_env__arch(env);
2941 	const char *cpuid = perf_env__cpuid(env);
2942 	char sbuf[STRERR_BUFSIZE];
2943 	int printed = 0, enforced = 0;
2944 
2945 	switch (err) {
2946 	case EPERM:
2947 	case EACCES:
2948 		printed += scnprintf(msg + printed, size - printed,
2949 			"Access to performance monitoring and observability operations is limited.\n");
2950 
2951 		if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2952 			if (enforced) {
2953 				printed += scnprintf(msg + printed, size - printed,
2954 					"Enforced MAC policy settings (SELinux) can limit access to performance\n"
2955 					"monitoring and observability operations. Inspect system audit records for\n"
2956 					"more perf_event access control information and adjusting the policy.\n");
2957 			}
2958 		}
2959 
2960 		if (err == EPERM)
2961 			printed += scnprintf(msg, size,
2962 				"No permission to enable %s event.\n\n", evsel__name(evsel));
2963 
2964 		return scnprintf(msg + printed, size - printed,
2965 		 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2966 		 "access to performance monitoring and observability operations for processes\n"
2967 		 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2968 		 "More information can be found at 'Perf events and tool security' document:\n"
2969 		 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2970 		 "perf_event_paranoid setting is %d:\n"
2971 		 "  -1: Allow use of (almost) all events by all users\n"
2972 		 "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2973 		 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2974 		 ">= 1: Disallow CPU event access\n"
2975 		 ">= 2: Disallow kernel profiling\n"
2976 		 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2977 		 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2978 		 perf_event_paranoid());
2979 	case ENOENT:
2980 		return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2981 	case EMFILE:
2982 		return scnprintf(msg, size, "%s",
2983 			 "Too many events are opened.\n"
2984 			 "Probably the maximum number of open file descriptors has been reached.\n"
2985 			 "Hint: Try again after reducing the number of events.\n"
2986 			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2987 	case ENOMEM:
2988 		if (evsel__has_callchain(evsel) &&
2989 		    access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2990 			return scnprintf(msg, size,
2991 					 "Not enough memory to setup event with callchain.\n"
2992 					 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2993 					 "Hint: Current value: %d", sysctl__max_stack());
2994 		break;
2995 	case ENODEV:
2996 		if (target->cpu_list)
2997 			return scnprintf(msg, size, "%s",
2998 	 "No such device - did you specify an out-of-range profile CPU?");
2999 		break;
3000 	case EOPNOTSUPP:
3001 		if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3002 			return scnprintf(msg, size,
3003 	"%s: PMU Hardware or event type doesn't support branch stack sampling.",
3004 					 evsel__name(evsel));
3005 		if (evsel->core.attr.aux_output)
3006 			return scnprintf(msg, size,
3007 	"%s: PMU Hardware doesn't support 'aux_output' feature",
3008 					 evsel__name(evsel));
3009 		if (evsel->core.attr.sample_period != 0)
3010 			return scnprintf(msg, size,
3011 	"%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3012 					 evsel__name(evsel));
3013 		if (evsel->core.attr.precise_ip)
3014 			return scnprintf(msg, size, "%s",
3015 	"\'precise\' request may not be supported. Try removing 'p' modifier.");
3016 #if defined(__i386__) || defined(__x86_64__)
3017 		if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3018 			return scnprintf(msg, size, "%s",
3019 	"No hardware sampling interrupt available.\n");
3020 #endif
3021 		break;
3022 	case EBUSY:
3023 		if (find_process("oprofiled"))
3024 			return scnprintf(msg, size,
3025 	"The PMU counters are busy/taken by another profiler.\n"
3026 	"We found oprofile daemon running, please stop it and try again.");
3027 		break;
3028 	case EINVAL:
3029 		if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3030 			return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3031 		if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3032 			return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3033 		if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3034 			return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3035 		if (perf_missing_features.clockid)
3036 			return scnprintf(msg, size, "clockid feature not supported.");
3037 		if (perf_missing_features.clockid_wrong)
3038 			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3039 		if (perf_missing_features.aux_output)
3040 			return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3041 		if (!target__has_cpu(target))
3042 			return scnprintf(msg, size,
3043 	"Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3044 					evsel__name(evsel));
3045 		if (is_amd(arch, cpuid)) {
3046 			if (is_amd_ibs(evsel)) {
3047 				if (evsel->core.attr.exclude_kernel)
3048 					return scnprintf(msg, size,
3049 	"AMD IBS can't exclude kernel events.  Try running at a higher privilege level.");
3050 				if (!evsel->core.system_wide)
3051 					return scnprintf(msg, size,
3052 	"AMD IBS may only be available in system-wide/per-cpu mode.  Try using -a, or -C and workload affinity");
3053 			}
3054 		}
3055 
3056 		break;
3057 	case ENODATA:
3058 		return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3059 				 "Please add an auxiliary event in front of the load latency event.");
3060 	default:
3061 		break;
3062 	}
3063 
3064 	return scnprintf(msg, size,
3065 	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3066 	"/bin/dmesg | grep -i perf may provide additional information.\n",
3067 			 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3068 }
3069 
3070 struct perf_env *evsel__env(struct evsel *evsel)
3071 {
3072 	if (evsel && evsel->evlist && evsel->evlist->env)
3073 		return evsel->evlist->env;
3074 	return &perf_env;
3075 }
3076 
3077 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3078 {
3079 	int cpu_map_idx, thread;
3080 
3081 	for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3082 		for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3083 		     thread++) {
3084 			int fd = FD(evsel, cpu_map_idx, thread);
3085 
3086 			if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3087 						   cpu_map_idx, thread, fd) < 0)
3088 				return -1;
3089 		}
3090 	}
3091 
3092 	return 0;
3093 }
3094 
3095 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3096 {
3097 	struct perf_cpu_map *cpus = evsel->core.cpus;
3098 	struct perf_thread_map *threads = evsel->core.threads;
3099 
3100 	if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3101 		return -ENOMEM;
3102 
3103 	return store_evsel_ids(evsel, evlist);
3104 }
3105 
3106 void evsel__zero_per_pkg(struct evsel *evsel)
3107 {
3108 	struct hashmap_entry *cur;
3109 	size_t bkt;
3110 
3111 	if (evsel->per_pkg_mask) {
3112 		hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3113 			zfree(&cur->pkey);
3114 
3115 		hashmap__clear(evsel->per_pkg_mask);
3116 	}
3117 }
3118 
3119 /**
3120  * evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this
3121  *                    will be false on hybrid systems for hardware and legacy
3122  *                    cache events.
3123  */
3124 bool evsel__is_hybrid(const struct evsel *evsel)
3125 {
3126 	if (perf_pmus__num_core_pmus() == 1)
3127 		return false;
3128 
3129 	return evsel->core.is_pmu_core;
3130 }
3131 
3132 struct evsel *evsel__leader(const struct evsel *evsel)
3133 {
3134 	return container_of(evsel->core.leader, struct evsel, core);
3135 }
3136 
3137 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3138 {
3139 	return evsel->core.leader == &leader->core;
3140 }
3141 
3142 bool evsel__is_leader(struct evsel *evsel)
3143 {
3144 	return evsel__has_leader(evsel, evsel);
3145 }
3146 
3147 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3148 {
3149 	evsel->core.leader = &leader->core;
3150 }
3151 
3152 int evsel__source_count(const struct evsel *evsel)
3153 {
3154 	struct evsel *pos;
3155 	int count = 0;
3156 
3157 	evlist__for_each_entry(evsel->evlist, pos) {
3158 		if (pos->metric_leader == evsel)
3159 			count++;
3160 	}
3161 	return count;
3162 }
3163 
3164 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3165 {
3166 	return false;
3167 }
3168 
3169 /*
3170  * Remove an event from a given group (leader).
3171  * Some events, e.g., perf metrics Topdown events,
3172  * must always be grouped. Ignore the events.
3173  */
3174 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3175 {
3176 	if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3177 		evsel__set_leader(evsel, evsel);
3178 		evsel->core.nr_members = 0;
3179 		leader->core.nr_members--;
3180 	}
3181 }
3182