xref: /openbmc/linux/tools/perf/util/evsel.c (revision a16be368)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4  *
5  * Parts came from builtin-{top,stat,record}.c, see those files for further
6  * copyright notes.
7  */
8 
9 #include <byteswap.h>
10 #include <errno.h>
11 #include <inttypes.h>
12 #include <linux/bitops.h>
13 #include <api/fs/fs.h>
14 #include <api/fs/tracing_path.h>
15 #include <traceevent/event-parse.h>
16 #include <linux/hw_breakpoint.h>
17 #include <linux/perf_event.h>
18 #include <linux/compiler.h>
19 #include <linux/err.h>
20 #include <linux/zalloc.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
24 #include <dirent.h>
25 #include <stdlib.h>
26 #include <perf/evsel.h>
27 #include "asm/bug.h"
28 #include "callchain.h"
29 #include "cgroup.h"
30 #include "counts.h"
31 #include "event.h"
32 #include "evsel.h"
33 #include "util/env.h"
34 #include "util/evsel_config.h"
35 #include "util/evsel_fprintf.h"
36 #include "evlist.h"
37 #include <perf/cpumap.h>
38 #include "thread_map.h"
39 #include "target.h"
40 #include "perf_regs.h"
41 #include "record.h"
42 #include "debug.h"
43 #include "trace-event.h"
44 #include "stat.h"
45 #include "string2.h"
46 #include "memswap.h"
47 #include "util.h"
48 #include "../perf-sys.h"
49 #include "util/parse-branch-options.h"
50 #include <internal/xyarray.h>
51 #include <internal/lib.h>
52 
53 #include <linux/ctype.h>
54 
55 struct perf_missing_features perf_missing_features;
56 
57 static clockid_t clockid;
58 
59 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
60 {
61 	return 0;
62 }
63 
64 void __weak test_attr__ready(void) { }
65 
66 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
67 {
68 }
69 
70 static struct {
71 	size_t	size;
72 	int	(*init)(struct evsel *evsel);
73 	void	(*fini)(struct evsel *evsel);
74 } perf_evsel__object = {
75 	.size = sizeof(struct evsel),
76 	.init = evsel__no_extra_init,
77 	.fini = evsel__no_extra_fini,
78 };
79 
80 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
81 			 void (*fini)(struct evsel *evsel))
82 {
83 
84 	if (object_size == 0)
85 		goto set_methods;
86 
87 	if (perf_evsel__object.size > object_size)
88 		return -EINVAL;
89 
90 	perf_evsel__object.size = object_size;
91 
92 set_methods:
93 	if (init != NULL)
94 		perf_evsel__object.init = init;
95 
96 	if (fini != NULL)
97 		perf_evsel__object.fini = fini;
98 
99 	return 0;
100 }
101 
102 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
103 
104 int __evsel__sample_size(u64 sample_type)
105 {
106 	u64 mask = sample_type & PERF_SAMPLE_MASK;
107 	int size = 0;
108 	int i;
109 
110 	for (i = 0; i < 64; i++) {
111 		if (mask & (1ULL << i))
112 			size++;
113 	}
114 
115 	size *= sizeof(u64);
116 
117 	return size;
118 }
119 
120 /**
121  * __perf_evsel__calc_id_pos - calculate id_pos.
122  * @sample_type: sample type
123  *
124  * This function returns the position of the event id (PERF_SAMPLE_ID or
125  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
126  * perf_record_sample.
127  */
128 static int __perf_evsel__calc_id_pos(u64 sample_type)
129 {
130 	int idx = 0;
131 
132 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
133 		return 0;
134 
135 	if (!(sample_type & PERF_SAMPLE_ID))
136 		return -1;
137 
138 	if (sample_type & PERF_SAMPLE_IP)
139 		idx += 1;
140 
141 	if (sample_type & PERF_SAMPLE_TID)
142 		idx += 1;
143 
144 	if (sample_type & PERF_SAMPLE_TIME)
145 		idx += 1;
146 
147 	if (sample_type & PERF_SAMPLE_ADDR)
148 		idx += 1;
149 
150 	return idx;
151 }
152 
153 /**
154  * __perf_evsel__calc_is_pos - calculate is_pos.
155  * @sample_type: sample type
156  *
157  * This function returns the position (counting backwards) of the event id
158  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
159  * sample_id_all is used there is an id sample appended to non-sample events.
160  */
161 static int __perf_evsel__calc_is_pos(u64 sample_type)
162 {
163 	int idx = 1;
164 
165 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
166 		return 1;
167 
168 	if (!(sample_type & PERF_SAMPLE_ID))
169 		return -1;
170 
171 	if (sample_type & PERF_SAMPLE_CPU)
172 		idx += 1;
173 
174 	if (sample_type & PERF_SAMPLE_STREAM_ID)
175 		idx += 1;
176 
177 	return idx;
178 }
179 
180 void evsel__calc_id_pos(struct evsel *evsel)
181 {
182 	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
183 	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
184 }
185 
186 void __evsel__set_sample_bit(struct evsel *evsel,
187 				  enum perf_event_sample_format bit)
188 {
189 	if (!(evsel->core.attr.sample_type & bit)) {
190 		evsel->core.attr.sample_type |= bit;
191 		evsel->sample_size += sizeof(u64);
192 		evsel__calc_id_pos(evsel);
193 	}
194 }
195 
196 void __evsel__reset_sample_bit(struct evsel *evsel,
197 				    enum perf_event_sample_format bit)
198 {
199 	if (evsel->core.attr.sample_type & bit) {
200 		evsel->core.attr.sample_type &= ~bit;
201 		evsel->sample_size -= sizeof(u64);
202 		evsel__calc_id_pos(evsel);
203 	}
204 }
205 
206 void evsel__set_sample_id(struct evsel *evsel,
207 			       bool can_sample_identifier)
208 {
209 	if (can_sample_identifier) {
210 		evsel__reset_sample_bit(evsel, ID);
211 		evsel__set_sample_bit(evsel, IDENTIFIER);
212 	} else {
213 		evsel__set_sample_bit(evsel, ID);
214 	}
215 	evsel->core.attr.read_format |= PERF_FORMAT_ID;
216 }
217 
218 /**
219  * evsel__is_function_event - Return whether given evsel is a function
220  * trace event
221  *
222  * @evsel - evsel selector to be tested
223  *
224  * Return %true if event is function trace event
225  */
226 bool evsel__is_function_event(struct evsel *evsel)
227 {
228 #define FUNCTION_EVENT "ftrace:function"
229 
230 	return evsel->name &&
231 	       !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
232 
233 #undef FUNCTION_EVENT
234 }
235 
236 void evsel__init(struct evsel *evsel,
237 		 struct perf_event_attr *attr, int idx)
238 {
239 	perf_evsel__init(&evsel->core, attr);
240 	evsel->idx	   = idx;
241 	evsel->tracking	   = !idx;
242 	evsel->leader	   = evsel;
243 	evsel->unit	   = "";
244 	evsel->scale	   = 1.0;
245 	evsel->max_events  = ULONG_MAX;
246 	evsel->evlist	   = NULL;
247 	evsel->bpf_obj	   = NULL;
248 	evsel->bpf_fd	   = -1;
249 	INIT_LIST_HEAD(&evsel->config_terms);
250 	perf_evsel__object.init(evsel);
251 	evsel->sample_size = __evsel__sample_size(attr->sample_type);
252 	evsel__calc_id_pos(evsel);
253 	evsel->cmdline_group_boundary = false;
254 	evsel->metric_expr   = NULL;
255 	evsel->metric_name   = NULL;
256 	evsel->metric_events = NULL;
257 	evsel->per_pkg_mask  = NULL;
258 	evsel->collect_stat  = false;
259 	evsel->pmu_name      = NULL;
260 }
261 
262 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
263 {
264 	struct evsel *evsel = zalloc(perf_evsel__object.size);
265 
266 	if (!evsel)
267 		return NULL;
268 	evsel__init(evsel, attr, idx);
269 
270 	if (evsel__is_bpf_output(evsel)) {
271 		evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
272 					    PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
273 		evsel->core.attr.sample_period = 1;
274 	}
275 
276 	if (evsel__is_clock(evsel)) {
277 		/*
278 		 * The evsel->unit points to static alias->unit
279 		 * so it's ok to use static string in here.
280 		 */
281 		static const char *unit = "msec";
282 
283 		evsel->unit = unit;
284 		evsel->scale = 1e-6;
285 	}
286 
287 	return evsel;
288 }
289 
290 static bool perf_event_can_profile_kernel(void)
291 {
292 	return perf_event_paranoid_check(1);
293 }
294 
295 struct evsel *evsel__new_cycles(bool precise)
296 {
297 	struct perf_event_attr attr = {
298 		.type	= PERF_TYPE_HARDWARE,
299 		.config	= PERF_COUNT_HW_CPU_CYCLES,
300 		.exclude_kernel	= !perf_event_can_profile_kernel(),
301 	};
302 	struct evsel *evsel;
303 
304 	event_attr_init(&attr);
305 
306 	if (!precise)
307 		goto new_event;
308 
309 	/*
310 	 * Now let the usual logic to set up the perf_event_attr defaults
311 	 * to kick in when we return and before perf_evsel__open() is called.
312 	 */
313 new_event:
314 	evsel = evsel__new(&attr);
315 	if (evsel == NULL)
316 		goto out;
317 
318 	evsel->precise_max = true;
319 
320 	/* use asprintf() because free(evsel) assumes name is allocated */
321 	if (asprintf(&evsel->name, "cycles%s%s%.*s",
322 		     (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
323 		     attr.exclude_kernel ? "u" : "",
324 		     attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
325 		goto error_free;
326 out:
327 	return evsel;
328 error_free:
329 	evsel__delete(evsel);
330 	evsel = NULL;
331 	goto out;
332 }
333 
334 /*
335  * Returns pointer with encoded error via <linux/err.h> interface.
336  */
337 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
338 {
339 	struct evsel *evsel = zalloc(perf_evsel__object.size);
340 	int err = -ENOMEM;
341 
342 	if (evsel == NULL) {
343 		goto out_err;
344 	} else {
345 		struct perf_event_attr attr = {
346 			.type	       = PERF_TYPE_TRACEPOINT,
347 			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
348 					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
349 		};
350 
351 		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
352 			goto out_free;
353 
354 		evsel->tp_format = trace_event__tp_format(sys, name);
355 		if (IS_ERR(evsel->tp_format)) {
356 			err = PTR_ERR(evsel->tp_format);
357 			goto out_free;
358 		}
359 
360 		event_attr_init(&attr);
361 		attr.config = evsel->tp_format->id;
362 		attr.sample_period = 1;
363 		evsel__init(evsel, &attr, idx);
364 	}
365 
366 	return evsel;
367 
368 out_free:
369 	zfree(&evsel->name);
370 	free(evsel);
371 out_err:
372 	return ERR_PTR(err);
373 }
374 
375 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
376 	"cycles",
377 	"instructions",
378 	"cache-references",
379 	"cache-misses",
380 	"branches",
381 	"branch-misses",
382 	"bus-cycles",
383 	"stalled-cycles-frontend",
384 	"stalled-cycles-backend",
385 	"ref-cycles",
386 };
387 
388 static const char *__evsel__hw_name(u64 config)
389 {
390 	if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
391 		return evsel__hw_names[config];
392 
393 	return "unknown-hardware";
394 }
395 
396 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
397 {
398 	int colon = 0, r = 0;
399 	struct perf_event_attr *attr = &evsel->core.attr;
400 	bool exclude_guest_default = false;
401 
402 #define MOD_PRINT(context, mod)	do {					\
403 		if (!attr->exclude_##context) {				\
404 			if (!colon) colon = ++r;			\
405 			r += scnprintf(bf + r, size - r, "%c", mod);	\
406 		} } while(0)
407 
408 	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
409 		MOD_PRINT(kernel, 'k');
410 		MOD_PRINT(user, 'u');
411 		MOD_PRINT(hv, 'h');
412 		exclude_guest_default = true;
413 	}
414 
415 	if (attr->precise_ip) {
416 		if (!colon)
417 			colon = ++r;
418 		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
419 		exclude_guest_default = true;
420 	}
421 
422 	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
423 		MOD_PRINT(host, 'H');
424 		MOD_PRINT(guest, 'G');
425 	}
426 #undef MOD_PRINT
427 	if (colon)
428 		bf[colon - 1] = ':';
429 	return r;
430 }
431 
432 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
433 {
434 	int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
435 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
436 }
437 
438 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
439 	"cpu-clock",
440 	"task-clock",
441 	"page-faults",
442 	"context-switches",
443 	"cpu-migrations",
444 	"minor-faults",
445 	"major-faults",
446 	"alignment-faults",
447 	"emulation-faults",
448 	"dummy",
449 };
450 
451 static const char *__evsel__sw_name(u64 config)
452 {
453 	if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
454 		return evsel__sw_names[config];
455 	return "unknown-software";
456 }
457 
458 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
459 {
460 	int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
461 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
462 }
463 
464 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
465 {
466 	int r;
467 
468 	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
469 
470 	if (type & HW_BREAKPOINT_R)
471 		r += scnprintf(bf + r, size - r, "r");
472 
473 	if (type & HW_BREAKPOINT_W)
474 		r += scnprintf(bf + r, size - r, "w");
475 
476 	if (type & HW_BREAKPOINT_X)
477 		r += scnprintf(bf + r, size - r, "x");
478 
479 	return r;
480 }
481 
482 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
483 {
484 	struct perf_event_attr *attr = &evsel->core.attr;
485 	int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
486 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
487 }
488 
489 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
490  { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
491  { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
492  { "LLC",	"L2",							},
493  { "dTLB",	"d-tlb",	"Data-TLB",				},
494  { "iTLB",	"i-tlb",	"Instruction-TLB",			},
495  { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
496  { "node",								},
497 };
498 
499 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
500  { "load",	"loads",	"read",					},
501  { "store",	"stores",	"write",				},
502  { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
503 };
504 
505 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
506  { "refs",	"Reference",	"ops",		"access",		},
507  { "misses",	"miss",							},
508 };
509 
510 #define C(x)		PERF_COUNT_HW_CACHE_##x
511 #define CACHE_READ	(1 << C(OP_READ))
512 #define CACHE_WRITE	(1 << C(OP_WRITE))
513 #define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
514 #define COP(x)		(1 << x)
515 
516 /*
517  * cache operartion stat
518  * L1I : Read and prefetch only
519  * ITLB and BPU : Read-only
520  */
521 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
522  [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
523  [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
524  [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
525  [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
526  [C(ITLB)]	= (CACHE_READ),
527  [C(BPU)]	= (CACHE_READ),
528  [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
529 };
530 
531 bool evsel__is_cache_op_valid(u8 type, u8 op)
532 {
533 	if (evsel__hw_cache_stat[type] & COP(op))
534 		return true;	/* valid */
535 	else
536 		return false;	/* invalid */
537 }
538 
539 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
540 {
541 	if (result) {
542 		return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
543 				 evsel__hw_cache_op[op][0],
544 				 evsel__hw_cache_result[result][0]);
545 	}
546 
547 	return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
548 			 evsel__hw_cache_op[op][1]);
549 }
550 
551 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
552 {
553 	u8 op, result, type = (config >>  0) & 0xff;
554 	const char *err = "unknown-ext-hardware-cache-type";
555 
556 	if (type >= PERF_COUNT_HW_CACHE_MAX)
557 		goto out_err;
558 
559 	op = (config >>  8) & 0xff;
560 	err = "unknown-ext-hardware-cache-op";
561 	if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
562 		goto out_err;
563 
564 	result = (config >> 16) & 0xff;
565 	err = "unknown-ext-hardware-cache-result";
566 	if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
567 		goto out_err;
568 
569 	err = "invalid-cache";
570 	if (!evsel__is_cache_op_valid(type, op))
571 		goto out_err;
572 
573 	return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
574 out_err:
575 	return scnprintf(bf, size, "%s", err);
576 }
577 
578 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
579 {
580 	int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
581 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
582 }
583 
584 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
585 {
586 	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
587 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
588 }
589 
590 static int evsel__tool_name(char *bf, size_t size)
591 {
592 	int ret = scnprintf(bf, size, "duration_time");
593 	return ret;
594 }
595 
596 const char *evsel__name(struct evsel *evsel)
597 {
598 	char bf[128];
599 
600 	if (!evsel)
601 		goto out_unknown;
602 
603 	if (evsel->name)
604 		return evsel->name;
605 
606 	switch (evsel->core.attr.type) {
607 	case PERF_TYPE_RAW:
608 		evsel__raw_name(evsel, bf, sizeof(bf));
609 		break;
610 
611 	case PERF_TYPE_HARDWARE:
612 		evsel__hw_name(evsel, bf, sizeof(bf));
613 		break;
614 
615 	case PERF_TYPE_HW_CACHE:
616 		evsel__hw_cache_name(evsel, bf, sizeof(bf));
617 		break;
618 
619 	case PERF_TYPE_SOFTWARE:
620 		if (evsel->tool_event)
621 			evsel__tool_name(bf, sizeof(bf));
622 		else
623 			evsel__sw_name(evsel, bf, sizeof(bf));
624 		break;
625 
626 	case PERF_TYPE_TRACEPOINT:
627 		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
628 		break;
629 
630 	case PERF_TYPE_BREAKPOINT:
631 		evsel__bp_name(evsel, bf, sizeof(bf));
632 		break;
633 
634 	default:
635 		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
636 			  evsel->core.attr.type);
637 		break;
638 	}
639 
640 	evsel->name = strdup(bf);
641 
642 	if (evsel->name)
643 		return evsel->name;
644 out_unknown:
645 	return "unknown";
646 }
647 
648 const char *evsel__group_name(struct evsel *evsel)
649 {
650 	return evsel->group_name ?: "anon group";
651 }
652 
653 /*
654  * Returns the group details for the specified leader,
655  * with following rules.
656  *
657  *  For record -e '{cycles,instructions}'
658  *    'anon group { cycles:u, instructions:u }'
659  *
660  *  For record -e 'cycles,instructions' and report --group
661  *    'cycles:u, instructions:u'
662  */
663 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
664 {
665 	int ret = 0;
666 	struct evsel *pos;
667 	const char *group_name = evsel__group_name(evsel);
668 
669 	if (!evsel->forced_leader)
670 		ret = scnprintf(buf, size, "%s { ", group_name);
671 
672 	ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
673 
674 	for_each_group_member(pos, evsel)
675 		ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
676 
677 	if (!evsel->forced_leader)
678 		ret += scnprintf(buf + ret, size - ret, " }");
679 
680 	return ret;
681 }
682 
683 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
684 				      struct callchain_param *param)
685 {
686 	bool function = evsel__is_function_event(evsel);
687 	struct perf_event_attr *attr = &evsel->core.attr;
688 
689 	evsel__set_sample_bit(evsel, CALLCHAIN);
690 
691 	attr->sample_max_stack = param->max_stack;
692 
693 	if (opts->kernel_callchains)
694 		attr->exclude_callchain_user = 1;
695 	if (opts->user_callchains)
696 		attr->exclude_callchain_kernel = 1;
697 	if (param->record_mode == CALLCHAIN_LBR) {
698 		if (!opts->branch_stack) {
699 			if (attr->exclude_user) {
700 				pr_warning("LBR callstack option is only available "
701 					   "to get user callchain information. "
702 					   "Falling back to framepointers.\n");
703 			} else {
704 				evsel__set_sample_bit(evsel, BRANCH_STACK);
705 				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
706 							PERF_SAMPLE_BRANCH_CALL_STACK |
707 							PERF_SAMPLE_BRANCH_NO_CYCLES |
708 							PERF_SAMPLE_BRANCH_NO_FLAGS |
709 							PERF_SAMPLE_BRANCH_HW_INDEX;
710 			}
711 		} else
712 			 pr_warning("Cannot use LBR callstack with branch stack. "
713 				    "Falling back to framepointers.\n");
714 	}
715 
716 	if (param->record_mode == CALLCHAIN_DWARF) {
717 		if (!function) {
718 			evsel__set_sample_bit(evsel, REGS_USER);
719 			evsel__set_sample_bit(evsel, STACK_USER);
720 			if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
721 				attr->sample_regs_user |= DWARF_MINIMAL_REGS;
722 				pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
723 					   "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
724 					   "so the minimal registers set (IP, SP) is explicitly forced.\n");
725 			} else {
726 				attr->sample_regs_user |= PERF_REGS_MASK;
727 			}
728 			attr->sample_stack_user = param->dump_size;
729 			attr->exclude_callchain_user = 1;
730 		} else {
731 			pr_info("Cannot use DWARF unwind for function trace event,"
732 				" falling back to framepointers.\n");
733 		}
734 	}
735 
736 	if (function) {
737 		pr_info("Disabling user space callchains for function trace event.\n");
738 		attr->exclude_callchain_user = 1;
739 	}
740 }
741 
742 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
743 			     struct callchain_param *param)
744 {
745 	if (param->enabled)
746 		return __evsel__config_callchain(evsel, opts, param);
747 }
748 
749 static void
750 perf_evsel__reset_callgraph(struct evsel *evsel,
751 			    struct callchain_param *param)
752 {
753 	struct perf_event_attr *attr = &evsel->core.attr;
754 
755 	evsel__reset_sample_bit(evsel, CALLCHAIN);
756 	if (param->record_mode == CALLCHAIN_LBR) {
757 		evsel__reset_sample_bit(evsel, BRANCH_STACK);
758 		attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
759 					      PERF_SAMPLE_BRANCH_CALL_STACK |
760 					      PERF_SAMPLE_BRANCH_HW_INDEX);
761 	}
762 	if (param->record_mode == CALLCHAIN_DWARF) {
763 		evsel__reset_sample_bit(evsel, REGS_USER);
764 		evsel__reset_sample_bit(evsel, STACK_USER);
765 	}
766 }
767 
768 static void evsel__apply_config_terms(struct evsel *evsel,
769 				      struct record_opts *opts, bool track)
770 {
771 	struct evsel_config_term *term;
772 	struct list_head *config_terms = &evsel->config_terms;
773 	struct perf_event_attr *attr = &evsel->core.attr;
774 	/* callgraph default */
775 	struct callchain_param param = {
776 		.record_mode = callchain_param.record_mode,
777 	};
778 	u32 dump_size = 0;
779 	int max_stack = 0;
780 	const char *callgraph_buf = NULL;
781 
782 	list_for_each_entry(term, config_terms, list) {
783 		switch (term->type) {
784 		case EVSEL__CONFIG_TERM_PERIOD:
785 			if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
786 				attr->sample_period = term->val.period;
787 				attr->freq = 0;
788 				evsel__reset_sample_bit(evsel, PERIOD);
789 			}
790 			break;
791 		case EVSEL__CONFIG_TERM_FREQ:
792 			if (!(term->weak && opts->user_freq != UINT_MAX)) {
793 				attr->sample_freq = term->val.freq;
794 				attr->freq = 1;
795 				evsel__set_sample_bit(evsel, PERIOD);
796 			}
797 			break;
798 		case EVSEL__CONFIG_TERM_TIME:
799 			if (term->val.time)
800 				evsel__set_sample_bit(evsel, TIME);
801 			else
802 				evsel__reset_sample_bit(evsel, TIME);
803 			break;
804 		case EVSEL__CONFIG_TERM_CALLGRAPH:
805 			callgraph_buf = term->val.str;
806 			break;
807 		case EVSEL__CONFIG_TERM_BRANCH:
808 			if (term->val.str && strcmp(term->val.str, "no")) {
809 				evsel__set_sample_bit(evsel, BRANCH_STACK);
810 				parse_branch_str(term->val.str,
811 						 &attr->branch_sample_type);
812 			} else
813 				evsel__reset_sample_bit(evsel, BRANCH_STACK);
814 			break;
815 		case EVSEL__CONFIG_TERM_STACK_USER:
816 			dump_size = term->val.stack_user;
817 			break;
818 		case EVSEL__CONFIG_TERM_MAX_STACK:
819 			max_stack = term->val.max_stack;
820 			break;
821 		case EVSEL__CONFIG_TERM_MAX_EVENTS:
822 			evsel->max_events = term->val.max_events;
823 			break;
824 		case EVSEL__CONFIG_TERM_INHERIT:
825 			/*
826 			 * attr->inherit should has already been set by
827 			 * evsel__config. If user explicitly set
828 			 * inherit using config terms, override global
829 			 * opt->no_inherit setting.
830 			 */
831 			attr->inherit = term->val.inherit ? 1 : 0;
832 			break;
833 		case EVSEL__CONFIG_TERM_OVERWRITE:
834 			attr->write_backward = term->val.overwrite ? 1 : 0;
835 			break;
836 		case EVSEL__CONFIG_TERM_DRV_CFG:
837 			break;
838 		case EVSEL__CONFIG_TERM_PERCORE:
839 			break;
840 		case EVSEL__CONFIG_TERM_AUX_OUTPUT:
841 			attr->aux_output = term->val.aux_output ? 1 : 0;
842 			break;
843 		case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
844 			/* Already applied by auxtrace */
845 			break;
846 		case EVSEL__CONFIG_TERM_CFG_CHG:
847 			break;
848 		default:
849 			break;
850 		}
851 	}
852 
853 	/* User explicitly set per-event callgraph, clear the old setting and reset. */
854 	if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
855 		bool sample_address = false;
856 
857 		if (max_stack) {
858 			param.max_stack = max_stack;
859 			if (callgraph_buf == NULL)
860 				callgraph_buf = "fp";
861 		}
862 
863 		/* parse callgraph parameters */
864 		if (callgraph_buf != NULL) {
865 			if (!strcmp(callgraph_buf, "no")) {
866 				param.enabled = false;
867 				param.record_mode = CALLCHAIN_NONE;
868 			} else {
869 				param.enabled = true;
870 				if (parse_callchain_record(callgraph_buf, &param)) {
871 					pr_err("per-event callgraph setting for %s failed. "
872 					       "Apply callgraph global setting for it\n",
873 					       evsel->name);
874 					return;
875 				}
876 				if (param.record_mode == CALLCHAIN_DWARF)
877 					sample_address = true;
878 			}
879 		}
880 		if (dump_size > 0) {
881 			dump_size = round_up(dump_size, sizeof(u64));
882 			param.dump_size = dump_size;
883 		}
884 
885 		/* If global callgraph set, clear it */
886 		if (callchain_param.enabled)
887 			perf_evsel__reset_callgraph(evsel, &callchain_param);
888 
889 		/* set perf-event callgraph */
890 		if (param.enabled) {
891 			if (sample_address) {
892 				evsel__set_sample_bit(evsel, ADDR);
893 				evsel__set_sample_bit(evsel, DATA_SRC);
894 				evsel->core.attr.mmap_data = track;
895 			}
896 			evsel__config_callchain(evsel, opts, &param);
897 		}
898 	}
899 }
900 
901 static bool is_dummy_event(struct evsel *evsel)
902 {
903 	return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
904 	       (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
905 }
906 
907 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
908 {
909 	struct evsel_config_term *term, *found_term = NULL;
910 
911 	list_for_each_entry(term, &evsel->config_terms, list) {
912 		if (term->type == type)
913 			found_term = term;
914 	}
915 
916 	return found_term;
917 }
918 
919 /*
920  * The enable_on_exec/disabled value strategy:
921  *
922  *  1) For any type of traced program:
923  *    - all independent events and group leaders are disabled
924  *    - all group members are enabled
925  *
926  *     Group members are ruled by group leaders. They need to
927  *     be enabled, because the group scheduling relies on that.
928  *
929  *  2) For traced programs executed by perf:
930  *     - all independent events and group leaders have
931  *       enable_on_exec set
932  *     - we don't specifically enable or disable any event during
933  *       the record command
934  *
935  *     Independent events and group leaders are initially disabled
936  *     and get enabled by exec. Group members are ruled by group
937  *     leaders as stated in 1).
938  *
939  *  3) For traced programs attached by perf (pid/tid):
940  *     - we specifically enable or disable all events during
941  *       the record command
942  *
943  *     When attaching events to already running traced we
944  *     enable/disable events specifically, as there's no
945  *     initial traced exec call.
946  */
947 void evsel__config(struct evsel *evsel, struct record_opts *opts,
948 		   struct callchain_param *callchain)
949 {
950 	struct evsel *leader = evsel->leader;
951 	struct perf_event_attr *attr = &evsel->core.attr;
952 	int track = evsel->tracking;
953 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
954 
955 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
956 	attr->inherit	    = !opts->no_inherit;
957 	attr->write_backward = opts->overwrite ? 1 : 0;
958 
959 	evsel__set_sample_bit(evsel, IP);
960 	evsel__set_sample_bit(evsel, TID);
961 
962 	if (evsel->sample_read) {
963 		evsel__set_sample_bit(evsel, READ);
964 
965 		/*
966 		 * We need ID even in case of single event, because
967 		 * PERF_SAMPLE_READ process ID specific data.
968 		 */
969 		evsel__set_sample_id(evsel, false);
970 
971 		/*
972 		 * Apply group format only if we belong to group
973 		 * with more than one members.
974 		 */
975 		if (leader->core.nr_members > 1) {
976 			attr->read_format |= PERF_FORMAT_GROUP;
977 			attr->inherit = 0;
978 		}
979 	}
980 
981 	/*
982 	 * We default some events to have a default interval. But keep
983 	 * it a weak assumption overridable by the user.
984 	 */
985 	if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
986 				     opts->user_interval != ULLONG_MAX)) {
987 		if (opts->freq) {
988 			evsel__set_sample_bit(evsel, PERIOD);
989 			attr->freq		= 1;
990 			attr->sample_freq	= opts->freq;
991 		} else {
992 			attr->sample_period = opts->default_interval;
993 		}
994 	}
995 
996 	if (opts->no_samples)
997 		attr->sample_freq = 0;
998 
999 	if (opts->inherit_stat) {
1000 		evsel->core.attr.read_format |=
1001 			PERF_FORMAT_TOTAL_TIME_ENABLED |
1002 			PERF_FORMAT_TOTAL_TIME_RUNNING |
1003 			PERF_FORMAT_ID;
1004 		attr->inherit_stat = 1;
1005 	}
1006 
1007 	if (opts->sample_address) {
1008 		evsel__set_sample_bit(evsel, ADDR);
1009 		attr->mmap_data = track;
1010 	}
1011 
1012 	/*
1013 	 * We don't allow user space callchains for  function trace
1014 	 * event, due to issues with page faults while tracing page
1015 	 * fault handler and its overall trickiness nature.
1016 	 */
1017 	if (evsel__is_function_event(evsel))
1018 		evsel->core.attr.exclude_callchain_user = 1;
1019 
1020 	if (callchain && callchain->enabled && !evsel->no_aux_samples)
1021 		evsel__config_callchain(evsel, opts, callchain);
1022 
1023 	if (opts->sample_intr_regs) {
1024 		attr->sample_regs_intr = opts->sample_intr_regs;
1025 		evsel__set_sample_bit(evsel, REGS_INTR);
1026 	}
1027 
1028 	if (opts->sample_user_regs) {
1029 		attr->sample_regs_user |= opts->sample_user_regs;
1030 		evsel__set_sample_bit(evsel, REGS_USER);
1031 	}
1032 
1033 	if (target__has_cpu(&opts->target) || opts->sample_cpu)
1034 		evsel__set_sample_bit(evsel, CPU);
1035 
1036 	/*
1037 	 * When the user explicitly disabled time don't force it here.
1038 	 */
1039 	if (opts->sample_time &&
1040 	    (!perf_missing_features.sample_id_all &&
1041 	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1042 	     opts->sample_time_set)))
1043 		evsel__set_sample_bit(evsel, TIME);
1044 
1045 	if (opts->raw_samples && !evsel->no_aux_samples) {
1046 		evsel__set_sample_bit(evsel, TIME);
1047 		evsel__set_sample_bit(evsel, RAW);
1048 		evsel__set_sample_bit(evsel, CPU);
1049 	}
1050 
1051 	if (opts->sample_address)
1052 		evsel__set_sample_bit(evsel, DATA_SRC);
1053 
1054 	if (opts->sample_phys_addr)
1055 		evsel__set_sample_bit(evsel, PHYS_ADDR);
1056 
1057 	if (opts->no_buffering) {
1058 		attr->watermark = 0;
1059 		attr->wakeup_events = 1;
1060 	}
1061 	if (opts->branch_stack && !evsel->no_aux_samples) {
1062 		evsel__set_sample_bit(evsel, BRANCH_STACK);
1063 		attr->branch_sample_type = opts->branch_stack;
1064 	}
1065 
1066 	if (opts->sample_weight)
1067 		evsel__set_sample_bit(evsel, WEIGHT);
1068 
1069 	attr->task  = track;
1070 	attr->mmap  = track;
1071 	attr->mmap2 = track && !perf_missing_features.mmap2;
1072 	attr->comm  = track;
1073 	attr->ksymbol = track && !perf_missing_features.ksymbol;
1074 	attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1075 
1076 	if (opts->record_namespaces)
1077 		attr->namespaces  = track;
1078 
1079 	if (opts->record_cgroup) {
1080 		attr->cgroup = track && !perf_missing_features.cgroup;
1081 		evsel__set_sample_bit(evsel, CGROUP);
1082 	}
1083 
1084 	if (opts->record_switch_events)
1085 		attr->context_switch = track;
1086 
1087 	if (opts->sample_transaction)
1088 		evsel__set_sample_bit(evsel, TRANSACTION);
1089 
1090 	if (opts->running_time) {
1091 		evsel->core.attr.read_format |=
1092 			PERF_FORMAT_TOTAL_TIME_ENABLED |
1093 			PERF_FORMAT_TOTAL_TIME_RUNNING;
1094 	}
1095 
1096 	/*
1097 	 * XXX see the function comment above
1098 	 *
1099 	 * Disabling only independent events or group leaders,
1100 	 * keeping group members enabled.
1101 	 */
1102 	if (evsel__is_group_leader(evsel))
1103 		attr->disabled = 1;
1104 
1105 	/*
1106 	 * Setting enable_on_exec for independent events and
1107 	 * group leaders for traced executed by perf.
1108 	 */
1109 	if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1110 	    !opts->initial_delay)
1111 		attr->enable_on_exec = 1;
1112 
1113 	if (evsel->immediate) {
1114 		attr->disabled = 0;
1115 		attr->enable_on_exec = 0;
1116 	}
1117 
1118 	clockid = opts->clockid;
1119 	if (opts->use_clockid) {
1120 		attr->use_clockid = 1;
1121 		attr->clockid = opts->clockid;
1122 	}
1123 
1124 	if (evsel->precise_max)
1125 		attr->precise_ip = 3;
1126 
1127 	if (opts->all_user) {
1128 		attr->exclude_kernel = 1;
1129 		attr->exclude_user   = 0;
1130 	}
1131 
1132 	if (opts->all_kernel) {
1133 		attr->exclude_kernel = 0;
1134 		attr->exclude_user   = 1;
1135 	}
1136 
1137 	if (evsel->core.own_cpus || evsel->unit)
1138 		evsel->core.attr.read_format |= PERF_FORMAT_ID;
1139 
1140 	/*
1141 	 * Apply event specific term settings,
1142 	 * it overloads any global configuration.
1143 	 */
1144 	evsel__apply_config_terms(evsel, opts, track);
1145 
1146 	evsel->ignore_missing_thread = opts->ignore_missing_thread;
1147 
1148 	/* The --period option takes the precedence. */
1149 	if (opts->period_set) {
1150 		if (opts->period)
1151 			evsel__set_sample_bit(evsel, PERIOD);
1152 		else
1153 			evsel__reset_sample_bit(evsel, PERIOD);
1154 	}
1155 
1156 	/*
1157 	 * A dummy event never triggers any actual counter and therefore
1158 	 * cannot be used with branch_stack.
1159 	 *
1160 	 * For initial_delay, a dummy event is added implicitly.
1161 	 * The software event will trigger -EOPNOTSUPP error out,
1162 	 * if BRANCH_STACK bit is set.
1163 	 */
1164 	if (is_dummy_event(evsel))
1165 		evsel__reset_sample_bit(evsel, BRANCH_STACK);
1166 }
1167 
1168 int evsel__set_filter(struct evsel *evsel, const char *filter)
1169 {
1170 	char *new_filter = strdup(filter);
1171 
1172 	if (new_filter != NULL) {
1173 		free(evsel->filter);
1174 		evsel->filter = new_filter;
1175 		return 0;
1176 	}
1177 
1178 	return -1;
1179 }
1180 
1181 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1182 {
1183 	char *new_filter;
1184 
1185 	if (evsel->filter == NULL)
1186 		return evsel__set_filter(evsel, filter);
1187 
1188 	if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1189 		free(evsel->filter);
1190 		evsel->filter = new_filter;
1191 		return 0;
1192 	}
1193 
1194 	return -1;
1195 }
1196 
1197 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1198 {
1199 	return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1200 }
1201 
1202 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1203 {
1204 	return evsel__append_filter(evsel, "%s,%s", filter);
1205 }
1206 
1207 /* Caller has to clear disabled after going through all CPUs. */
1208 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1209 {
1210 	return perf_evsel__enable_cpu(&evsel->core, cpu);
1211 }
1212 
1213 int evsel__enable(struct evsel *evsel)
1214 {
1215 	int err = perf_evsel__enable(&evsel->core);
1216 
1217 	if (!err)
1218 		evsel->disabled = false;
1219 	return err;
1220 }
1221 
1222 /* Caller has to set disabled after going through all CPUs. */
1223 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1224 {
1225 	return perf_evsel__disable_cpu(&evsel->core, cpu);
1226 }
1227 
1228 int evsel__disable(struct evsel *evsel)
1229 {
1230 	int err = perf_evsel__disable(&evsel->core);
1231 	/*
1232 	 * We mark it disabled here so that tools that disable a event can
1233 	 * ignore events after they disable it. I.e. the ring buffer may have
1234 	 * already a few more events queued up before the kernel got the stop
1235 	 * request.
1236 	 */
1237 	if (!err)
1238 		evsel->disabled = true;
1239 
1240 	return err;
1241 }
1242 
1243 static void evsel__free_config_terms(struct evsel *evsel)
1244 {
1245 	struct evsel_config_term *term, *h;
1246 
1247 	list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1248 		list_del_init(&term->list);
1249 		if (term->free_str)
1250 			zfree(&term->val.str);
1251 		free(term);
1252 	}
1253 }
1254 
1255 void evsel__exit(struct evsel *evsel)
1256 {
1257 	assert(list_empty(&evsel->core.node));
1258 	assert(evsel->evlist == NULL);
1259 	evsel__free_counts(evsel);
1260 	perf_evsel__free_fd(&evsel->core);
1261 	perf_evsel__free_id(&evsel->core);
1262 	evsel__free_config_terms(evsel);
1263 	cgroup__put(evsel->cgrp);
1264 	perf_cpu_map__put(evsel->core.cpus);
1265 	perf_cpu_map__put(evsel->core.own_cpus);
1266 	perf_thread_map__put(evsel->core.threads);
1267 	zfree(&evsel->group_name);
1268 	zfree(&evsel->name);
1269 	zfree(&evsel->pmu_name);
1270 	zfree(&evsel->per_pkg_mask);
1271 	zfree(&evsel->metric_events);
1272 	perf_evsel__object.fini(evsel);
1273 }
1274 
1275 void evsel__delete(struct evsel *evsel)
1276 {
1277 	evsel__exit(evsel);
1278 	free(evsel);
1279 }
1280 
1281 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1282 			   struct perf_counts_values *count)
1283 {
1284 	struct perf_counts_values tmp;
1285 
1286 	if (!evsel->prev_raw_counts)
1287 		return;
1288 
1289 	if (cpu == -1) {
1290 		tmp = evsel->prev_raw_counts->aggr;
1291 		evsel->prev_raw_counts->aggr = *count;
1292 	} else {
1293 		tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1294 		*perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1295 	}
1296 
1297 	count->val = count->val - tmp.val;
1298 	count->ena = count->ena - tmp.ena;
1299 	count->run = count->run - tmp.run;
1300 }
1301 
1302 void perf_counts_values__scale(struct perf_counts_values *count,
1303 			       bool scale, s8 *pscaled)
1304 {
1305 	s8 scaled = 0;
1306 
1307 	if (scale) {
1308 		if (count->run == 0) {
1309 			scaled = -1;
1310 			count->val = 0;
1311 		} else if (count->run < count->ena) {
1312 			scaled = 1;
1313 			count->val = (u64)((double) count->val * count->ena / count->run);
1314 		}
1315 	}
1316 
1317 	if (pscaled)
1318 		*pscaled = scaled;
1319 }
1320 
1321 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1322 {
1323 	struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1324 
1325 	return perf_evsel__read(&evsel->core, cpu, thread, count);
1326 }
1327 
1328 static void
1329 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1330 		      u64 val, u64 ena, u64 run)
1331 {
1332 	struct perf_counts_values *count;
1333 
1334 	count = perf_counts(counter->counts, cpu, thread);
1335 
1336 	count->val    = val;
1337 	count->ena    = ena;
1338 	count->run    = run;
1339 
1340 	perf_counts__set_loaded(counter->counts, cpu, thread, true);
1341 }
1342 
1343 static int
1344 perf_evsel__process_group_data(struct evsel *leader,
1345 			       int cpu, int thread, u64 *data)
1346 {
1347 	u64 read_format = leader->core.attr.read_format;
1348 	struct sample_read_value *v;
1349 	u64 nr, ena = 0, run = 0, i;
1350 
1351 	nr = *data++;
1352 
1353 	if (nr != (u64) leader->core.nr_members)
1354 		return -EINVAL;
1355 
1356 	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1357 		ena = *data++;
1358 
1359 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1360 		run = *data++;
1361 
1362 	v = (struct sample_read_value *) data;
1363 
1364 	perf_evsel__set_count(leader, cpu, thread,
1365 			      v[0].value, ena, run);
1366 
1367 	for (i = 1; i < nr; i++) {
1368 		struct evsel *counter;
1369 
1370 		counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1371 		if (!counter)
1372 			return -EINVAL;
1373 
1374 		perf_evsel__set_count(counter, cpu, thread,
1375 				      v[i].value, ena, run);
1376 	}
1377 
1378 	return 0;
1379 }
1380 
1381 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1382 {
1383 	struct perf_stat_evsel *ps = leader->stats;
1384 	u64 read_format = leader->core.attr.read_format;
1385 	int size = perf_evsel__read_size(&leader->core);
1386 	u64 *data = ps->group_data;
1387 
1388 	if (!(read_format & PERF_FORMAT_ID))
1389 		return -EINVAL;
1390 
1391 	if (!evsel__is_group_leader(leader))
1392 		return -EINVAL;
1393 
1394 	if (!data) {
1395 		data = zalloc(size);
1396 		if (!data)
1397 			return -ENOMEM;
1398 
1399 		ps->group_data = data;
1400 	}
1401 
1402 	if (FD(leader, cpu, thread) < 0)
1403 		return -EINVAL;
1404 
1405 	if (readn(FD(leader, cpu, thread), data, size) <= 0)
1406 		return -errno;
1407 
1408 	return perf_evsel__process_group_data(leader, cpu, thread, data);
1409 }
1410 
1411 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1412 {
1413 	u64 read_format = evsel->core.attr.read_format;
1414 
1415 	if (read_format & PERF_FORMAT_GROUP)
1416 		return evsel__read_group(evsel, cpu, thread);
1417 
1418 	return evsel__read_one(evsel, cpu, thread);
1419 }
1420 
1421 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1422 {
1423 	struct perf_counts_values count;
1424 	size_t nv = scale ? 3 : 1;
1425 
1426 	if (FD(evsel, cpu, thread) < 0)
1427 		return -EINVAL;
1428 
1429 	if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1430 		return -ENOMEM;
1431 
1432 	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1433 		return -errno;
1434 
1435 	evsel__compute_deltas(evsel, cpu, thread, &count);
1436 	perf_counts_values__scale(&count, scale, NULL);
1437 	*perf_counts(evsel->counts, cpu, thread) = count;
1438 	return 0;
1439 }
1440 
1441 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1442 {
1443 	struct evsel *leader = evsel->leader;
1444 	int fd;
1445 
1446 	if (evsel__is_group_leader(evsel))
1447 		return -1;
1448 
1449 	/*
1450 	 * Leader must be already processed/open,
1451 	 * if not it's a bug.
1452 	 */
1453 	BUG_ON(!leader->core.fd);
1454 
1455 	fd = FD(leader, cpu, thread);
1456 	BUG_ON(fd == -1);
1457 
1458 	return fd;
1459 }
1460 
1461 static void perf_evsel__remove_fd(struct evsel *pos,
1462 				  int nr_cpus, int nr_threads,
1463 				  int thread_idx)
1464 {
1465 	for (int cpu = 0; cpu < nr_cpus; cpu++)
1466 		for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1467 			FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1468 }
1469 
1470 static int update_fds(struct evsel *evsel,
1471 		      int nr_cpus, int cpu_idx,
1472 		      int nr_threads, int thread_idx)
1473 {
1474 	struct evsel *pos;
1475 
1476 	if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1477 		return -EINVAL;
1478 
1479 	evlist__for_each_entry(evsel->evlist, pos) {
1480 		nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1481 
1482 		perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1483 
1484 		/*
1485 		 * Since fds for next evsel has not been created,
1486 		 * there is no need to iterate whole event list.
1487 		 */
1488 		if (pos == evsel)
1489 			break;
1490 	}
1491 	return 0;
1492 }
1493 
1494 static bool ignore_missing_thread(struct evsel *evsel,
1495 				  int nr_cpus, int cpu,
1496 				  struct perf_thread_map *threads,
1497 				  int thread, int err)
1498 {
1499 	pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1500 
1501 	if (!evsel->ignore_missing_thread)
1502 		return false;
1503 
1504 	/* The system wide setup does not work with threads. */
1505 	if (evsel->core.system_wide)
1506 		return false;
1507 
1508 	/* The -ESRCH is perf event syscall errno for pid's not found. */
1509 	if (err != -ESRCH)
1510 		return false;
1511 
1512 	/* If there's only one thread, let it fail. */
1513 	if (threads->nr == 1)
1514 		return false;
1515 
1516 	/*
1517 	 * We should remove fd for missing_thread first
1518 	 * because thread_map__remove() will decrease threads->nr.
1519 	 */
1520 	if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1521 		return false;
1522 
1523 	if (thread_map__remove(threads, thread))
1524 		return false;
1525 
1526 	pr_warning("WARNING: Ignored open failure for pid %d\n",
1527 		   ignore_pid);
1528 	return true;
1529 }
1530 
1531 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1532 				void *priv __maybe_unused)
1533 {
1534 	return fprintf(fp, "  %-32s %s\n", name, val);
1535 }
1536 
1537 static void display_attr(struct perf_event_attr *attr)
1538 {
1539 	if (verbose >= 2 || debug_peo_args) {
1540 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1541 		fprintf(stderr, "perf_event_attr:\n");
1542 		perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1543 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1544 	}
1545 }
1546 
1547 static int perf_event_open(struct evsel *evsel,
1548 			   pid_t pid, int cpu, int group_fd,
1549 			   unsigned long flags)
1550 {
1551 	int precise_ip = evsel->core.attr.precise_ip;
1552 	int fd;
1553 
1554 	while (1) {
1555 		pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1556 			  pid, cpu, group_fd, flags);
1557 
1558 		fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1559 		if (fd >= 0)
1560 			break;
1561 
1562 		/* Do not try less precise if not requested. */
1563 		if (!evsel->precise_max)
1564 			break;
1565 
1566 		/*
1567 		 * We tried all the precise_ip values, and it's
1568 		 * still failing, so leave it to standard fallback.
1569 		 */
1570 		if (!evsel->core.attr.precise_ip) {
1571 			evsel->core.attr.precise_ip = precise_ip;
1572 			break;
1573 		}
1574 
1575 		pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1576 		evsel->core.attr.precise_ip--;
1577 		pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1578 		display_attr(&evsel->core.attr);
1579 	}
1580 
1581 	return fd;
1582 }
1583 
1584 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1585 		struct perf_thread_map *threads,
1586 		int start_cpu, int end_cpu)
1587 {
1588 	int cpu, thread, nthreads;
1589 	unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1590 	int pid = -1, err, old_errno;
1591 	enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1592 
1593 	if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1594 	    (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1595 		return -EINVAL;
1596 
1597 	if (cpus == NULL) {
1598 		static struct perf_cpu_map *empty_cpu_map;
1599 
1600 		if (empty_cpu_map == NULL) {
1601 			empty_cpu_map = perf_cpu_map__dummy_new();
1602 			if (empty_cpu_map == NULL)
1603 				return -ENOMEM;
1604 		}
1605 
1606 		cpus = empty_cpu_map;
1607 	}
1608 
1609 	if (threads == NULL) {
1610 		static struct perf_thread_map *empty_thread_map;
1611 
1612 		if (empty_thread_map == NULL) {
1613 			empty_thread_map = thread_map__new_by_tid(-1);
1614 			if (empty_thread_map == NULL)
1615 				return -ENOMEM;
1616 		}
1617 
1618 		threads = empty_thread_map;
1619 	}
1620 
1621 	if (evsel->core.system_wide)
1622 		nthreads = 1;
1623 	else
1624 		nthreads = threads->nr;
1625 
1626 	if (evsel->core.fd == NULL &&
1627 	    perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1628 		return -ENOMEM;
1629 
1630 	if (evsel->cgrp) {
1631 		flags |= PERF_FLAG_PID_CGROUP;
1632 		pid = evsel->cgrp->fd;
1633 	}
1634 
1635 fallback_missing_features:
1636 	if (perf_missing_features.clockid_wrong)
1637 		evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1638 	if (perf_missing_features.clockid) {
1639 		evsel->core.attr.use_clockid = 0;
1640 		evsel->core.attr.clockid = 0;
1641 	}
1642 	if (perf_missing_features.cloexec)
1643 		flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1644 	if (perf_missing_features.mmap2)
1645 		evsel->core.attr.mmap2 = 0;
1646 	if (perf_missing_features.exclude_guest)
1647 		evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1648 	if (perf_missing_features.lbr_flags)
1649 		evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1650 				     PERF_SAMPLE_BRANCH_NO_CYCLES);
1651 	if (perf_missing_features.group_read && evsel->core.attr.inherit)
1652 		evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1653 	if (perf_missing_features.ksymbol)
1654 		evsel->core.attr.ksymbol = 0;
1655 	if (perf_missing_features.bpf)
1656 		evsel->core.attr.bpf_event = 0;
1657 	if (perf_missing_features.branch_hw_idx)
1658 		evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1659 retry_sample_id:
1660 	if (perf_missing_features.sample_id_all)
1661 		evsel->core.attr.sample_id_all = 0;
1662 
1663 	display_attr(&evsel->core.attr);
1664 
1665 	for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1666 
1667 		for (thread = 0; thread < nthreads; thread++) {
1668 			int fd, group_fd;
1669 
1670 			if (!evsel->cgrp && !evsel->core.system_wide)
1671 				pid = perf_thread_map__pid(threads, thread);
1672 
1673 			group_fd = get_group_fd(evsel, cpu, thread);
1674 retry_open:
1675 			test_attr__ready();
1676 
1677 			fd = perf_event_open(evsel, pid, cpus->map[cpu],
1678 					     group_fd, flags);
1679 
1680 			FD(evsel, cpu, thread) = fd;
1681 
1682 			if (fd < 0) {
1683 				err = -errno;
1684 
1685 				if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1686 					/*
1687 					 * We just removed 1 thread, so take a step
1688 					 * back on thread index and lower the upper
1689 					 * nthreads limit.
1690 					 */
1691 					nthreads--;
1692 					thread--;
1693 
1694 					/* ... and pretend like nothing have happened. */
1695 					err = 0;
1696 					continue;
1697 				}
1698 
1699 				pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1700 					  err);
1701 				goto try_fallback;
1702 			}
1703 
1704 			pr_debug2_peo(" = %d\n", fd);
1705 
1706 			if (evsel->bpf_fd >= 0) {
1707 				int evt_fd = fd;
1708 				int bpf_fd = evsel->bpf_fd;
1709 
1710 				err = ioctl(evt_fd,
1711 					    PERF_EVENT_IOC_SET_BPF,
1712 					    bpf_fd);
1713 				if (err && errno != EEXIST) {
1714 					pr_err("failed to attach bpf fd %d: %s\n",
1715 					       bpf_fd, strerror(errno));
1716 					err = -EINVAL;
1717 					goto out_close;
1718 				}
1719 			}
1720 
1721 			set_rlimit = NO_CHANGE;
1722 
1723 			/*
1724 			 * If we succeeded but had to kill clockid, fail and
1725 			 * have evsel__open_strerror() print us a nice error.
1726 			 */
1727 			if (perf_missing_features.clockid ||
1728 			    perf_missing_features.clockid_wrong) {
1729 				err = -EINVAL;
1730 				goto out_close;
1731 			}
1732 		}
1733 	}
1734 
1735 	return 0;
1736 
1737 try_fallback:
1738 	/*
1739 	 * perf stat needs between 5 and 22 fds per CPU. When we run out
1740 	 * of them try to increase the limits.
1741 	 */
1742 	if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1743 		struct rlimit l;
1744 
1745 		old_errno = errno;
1746 		if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1747 			if (set_rlimit == NO_CHANGE)
1748 				l.rlim_cur = l.rlim_max;
1749 			else {
1750 				l.rlim_cur = l.rlim_max + 1000;
1751 				l.rlim_max = l.rlim_cur;
1752 			}
1753 			if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1754 				set_rlimit++;
1755 				errno = old_errno;
1756 				goto retry_open;
1757 			}
1758 		}
1759 		errno = old_errno;
1760 	}
1761 
1762 	if (err != -EINVAL || cpu > 0 || thread > 0)
1763 		goto out_close;
1764 
1765 	/*
1766 	 * Must probe features in the order they were added to the
1767 	 * perf_event_attr interface.
1768 	 */
1769         if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1770 		perf_missing_features.cgroup = true;
1771 		pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1772 		goto out_close;
1773         } else if (!perf_missing_features.branch_hw_idx &&
1774 	    (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1775 		perf_missing_features.branch_hw_idx = true;
1776 		pr_debug2("switching off branch HW index support\n");
1777 		goto fallback_missing_features;
1778 	} else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1779 		perf_missing_features.aux_output = true;
1780 		pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1781 		goto out_close;
1782 	} else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1783 		perf_missing_features.bpf = true;
1784 		pr_debug2_peo("switching off bpf_event\n");
1785 		goto fallback_missing_features;
1786 	} else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1787 		perf_missing_features.ksymbol = true;
1788 		pr_debug2_peo("switching off ksymbol\n");
1789 		goto fallback_missing_features;
1790 	} else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1791 		perf_missing_features.write_backward = true;
1792 		pr_debug2_peo("switching off write_backward\n");
1793 		goto out_close;
1794 	} else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1795 		perf_missing_features.clockid_wrong = true;
1796 		pr_debug2_peo("switching off clockid\n");
1797 		goto fallback_missing_features;
1798 	} else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1799 		perf_missing_features.clockid = true;
1800 		pr_debug2_peo("switching off use_clockid\n");
1801 		goto fallback_missing_features;
1802 	} else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1803 		perf_missing_features.cloexec = true;
1804 		pr_debug2_peo("switching off cloexec flag\n");
1805 		goto fallback_missing_features;
1806 	} else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1807 		perf_missing_features.mmap2 = true;
1808 		pr_debug2_peo("switching off mmap2\n");
1809 		goto fallback_missing_features;
1810 	} else if (!perf_missing_features.exclude_guest &&
1811 		   (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1812 		perf_missing_features.exclude_guest = true;
1813 		pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1814 		goto fallback_missing_features;
1815 	} else if (!perf_missing_features.sample_id_all) {
1816 		perf_missing_features.sample_id_all = true;
1817 		pr_debug2_peo("switching off sample_id_all\n");
1818 		goto retry_sample_id;
1819 	} else if (!perf_missing_features.lbr_flags &&
1820 			(evsel->core.attr.branch_sample_type &
1821 			 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1822 			  PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1823 		perf_missing_features.lbr_flags = true;
1824 		pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1825 		goto fallback_missing_features;
1826 	} else if (!perf_missing_features.group_read &&
1827 		    evsel->core.attr.inherit &&
1828 		   (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1829 		   evsel__is_group_leader(evsel)) {
1830 		perf_missing_features.group_read = true;
1831 		pr_debug2_peo("switching off group read\n");
1832 		goto fallback_missing_features;
1833 	}
1834 out_close:
1835 	if (err)
1836 		threads->err_thread = thread;
1837 
1838 	old_errno = errno;
1839 	do {
1840 		while (--thread >= 0) {
1841 			if (FD(evsel, cpu, thread) >= 0)
1842 				close(FD(evsel, cpu, thread));
1843 			FD(evsel, cpu, thread) = -1;
1844 		}
1845 		thread = nthreads;
1846 	} while (--cpu >= 0);
1847 	errno = old_errno;
1848 	return err;
1849 }
1850 
1851 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1852 		struct perf_thread_map *threads)
1853 {
1854 	return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1855 }
1856 
1857 void evsel__close(struct evsel *evsel)
1858 {
1859 	perf_evsel__close(&evsel->core);
1860 	perf_evsel__free_id(&evsel->core);
1861 }
1862 
1863 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
1864 {
1865 	if (cpu == -1)
1866 		return evsel__open_cpu(evsel, cpus, NULL, 0,
1867 					cpus ? cpus->nr : 1);
1868 
1869 	return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1870 }
1871 
1872 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
1873 {
1874 	return evsel__open(evsel, NULL, threads);
1875 }
1876 
1877 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1878 				       const union perf_event *event,
1879 				       struct perf_sample *sample)
1880 {
1881 	u64 type = evsel->core.attr.sample_type;
1882 	const __u64 *array = event->sample.array;
1883 	bool swapped = evsel->needs_swap;
1884 	union u64_swap u;
1885 
1886 	array += ((event->header.size -
1887 		   sizeof(event->header)) / sizeof(u64)) - 1;
1888 
1889 	if (type & PERF_SAMPLE_IDENTIFIER) {
1890 		sample->id = *array;
1891 		array--;
1892 	}
1893 
1894 	if (type & PERF_SAMPLE_CPU) {
1895 		u.val64 = *array;
1896 		if (swapped) {
1897 			/* undo swap of u64, then swap on individual u32s */
1898 			u.val64 = bswap_64(u.val64);
1899 			u.val32[0] = bswap_32(u.val32[0]);
1900 		}
1901 
1902 		sample->cpu = u.val32[0];
1903 		array--;
1904 	}
1905 
1906 	if (type & PERF_SAMPLE_STREAM_ID) {
1907 		sample->stream_id = *array;
1908 		array--;
1909 	}
1910 
1911 	if (type & PERF_SAMPLE_ID) {
1912 		sample->id = *array;
1913 		array--;
1914 	}
1915 
1916 	if (type & PERF_SAMPLE_TIME) {
1917 		sample->time = *array;
1918 		array--;
1919 	}
1920 
1921 	if (type & PERF_SAMPLE_TID) {
1922 		u.val64 = *array;
1923 		if (swapped) {
1924 			/* undo swap of u64, then swap on individual u32s */
1925 			u.val64 = bswap_64(u.val64);
1926 			u.val32[0] = bswap_32(u.val32[0]);
1927 			u.val32[1] = bswap_32(u.val32[1]);
1928 		}
1929 
1930 		sample->pid = u.val32[0];
1931 		sample->tid = u.val32[1];
1932 		array--;
1933 	}
1934 
1935 	return 0;
1936 }
1937 
1938 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1939 			    u64 size)
1940 {
1941 	return size > max_size || offset + size > endp;
1942 }
1943 
1944 #define OVERFLOW_CHECK(offset, size, max_size)				\
1945 	do {								\
1946 		if (overflow(endp, (max_size), (offset), (size)))	\
1947 			return -EFAULT;					\
1948 	} while (0)
1949 
1950 #define OVERFLOW_CHECK_u64(offset) \
1951 	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1952 
1953 static int
1954 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1955 {
1956 	/*
1957 	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1958 	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1959 	 * check the format does not go past the end of the event.
1960 	 */
1961 	if (sample_size + sizeof(event->header) > event->header.size)
1962 		return -EFAULT;
1963 
1964 	return 0;
1965 }
1966 
1967 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1968 			struct perf_sample *data)
1969 {
1970 	u64 type = evsel->core.attr.sample_type;
1971 	bool swapped = evsel->needs_swap;
1972 	const __u64 *array;
1973 	u16 max_size = event->header.size;
1974 	const void *endp = (void *)event + max_size;
1975 	u64 sz;
1976 
1977 	/*
1978 	 * used for cross-endian analysis. See git commit 65014ab3
1979 	 * for why this goofiness is needed.
1980 	 */
1981 	union u64_swap u;
1982 
1983 	memset(data, 0, sizeof(*data));
1984 	data->cpu = data->pid = data->tid = -1;
1985 	data->stream_id = data->id = data->time = -1ULL;
1986 	data->period = evsel->core.attr.sample_period;
1987 	data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1988 	data->misc    = event->header.misc;
1989 	data->id = -1ULL;
1990 	data->data_src = PERF_MEM_DATA_SRC_NONE;
1991 
1992 	if (event->header.type != PERF_RECORD_SAMPLE) {
1993 		if (!evsel->core.attr.sample_id_all)
1994 			return 0;
1995 		return perf_evsel__parse_id_sample(evsel, event, data);
1996 	}
1997 
1998 	array = event->sample.array;
1999 
2000 	if (perf_event__check_size(event, evsel->sample_size))
2001 		return -EFAULT;
2002 
2003 	if (type & PERF_SAMPLE_IDENTIFIER) {
2004 		data->id = *array;
2005 		array++;
2006 	}
2007 
2008 	if (type & PERF_SAMPLE_IP) {
2009 		data->ip = *array;
2010 		array++;
2011 	}
2012 
2013 	if (type & PERF_SAMPLE_TID) {
2014 		u.val64 = *array;
2015 		if (swapped) {
2016 			/* undo swap of u64, then swap on individual u32s */
2017 			u.val64 = bswap_64(u.val64);
2018 			u.val32[0] = bswap_32(u.val32[0]);
2019 			u.val32[1] = bswap_32(u.val32[1]);
2020 		}
2021 
2022 		data->pid = u.val32[0];
2023 		data->tid = u.val32[1];
2024 		array++;
2025 	}
2026 
2027 	if (type & PERF_SAMPLE_TIME) {
2028 		data->time = *array;
2029 		array++;
2030 	}
2031 
2032 	if (type & PERF_SAMPLE_ADDR) {
2033 		data->addr = *array;
2034 		array++;
2035 	}
2036 
2037 	if (type & PERF_SAMPLE_ID) {
2038 		data->id = *array;
2039 		array++;
2040 	}
2041 
2042 	if (type & PERF_SAMPLE_STREAM_ID) {
2043 		data->stream_id = *array;
2044 		array++;
2045 	}
2046 
2047 	if (type & PERF_SAMPLE_CPU) {
2048 
2049 		u.val64 = *array;
2050 		if (swapped) {
2051 			/* undo swap of u64, then swap on individual u32s */
2052 			u.val64 = bswap_64(u.val64);
2053 			u.val32[0] = bswap_32(u.val32[0]);
2054 		}
2055 
2056 		data->cpu = u.val32[0];
2057 		array++;
2058 	}
2059 
2060 	if (type & PERF_SAMPLE_PERIOD) {
2061 		data->period = *array;
2062 		array++;
2063 	}
2064 
2065 	if (type & PERF_SAMPLE_READ) {
2066 		u64 read_format = evsel->core.attr.read_format;
2067 
2068 		OVERFLOW_CHECK_u64(array);
2069 		if (read_format & PERF_FORMAT_GROUP)
2070 			data->read.group.nr = *array;
2071 		else
2072 			data->read.one.value = *array;
2073 
2074 		array++;
2075 
2076 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2077 			OVERFLOW_CHECK_u64(array);
2078 			data->read.time_enabled = *array;
2079 			array++;
2080 		}
2081 
2082 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2083 			OVERFLOW_CHECK_u64(array);
2084 			data->read.time_running = *array;
2085 			array++;
2086 		}
2087 
2088 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2089 		if (read_format & PERF_FORMAT_GROUP) {
2090 			const u64 max_group_nr = UINT64_MAX /
2091 					sizeof(struct sample_read_value);
2092 
2093 			if (data->read.group.nr > max_group_nr)
2094 				return -EFAULT;
2095 			sz = data->read.group.nr *
2096 			     sizeof(struct sample_read_value);
2097 			OVERFLOW_CHECK(array, sz, max_size);
2098 			data->read.group.values =
2099 					(struct sample_read_value *)array;
2100 			array = (void *)array + sz;
2101 		} else {
2102 			OVERFLOW_CHECK_u64(array);
2103 			data->read.one.id = *array;
2104 			array++;
2105 		}
2106 	}
2107 
2108 	if (type & PERF_SAMPLE_CALLCHAIN) {
2109 		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2110 
2111 		OVERFLOW_CHECK_u64(array);
2112 		data->callchain = (struct ip_callchain *)array++;
2113 		if (data->callchain->nr > max_callchain_nr)
2114 			return -EFAULT;
2115 		sz = data->callchain->nr * sizeof(u64);
2116 		OVERFLOW_CHECK(array, sz, max_size);
2117 		array = (void *)array + sz;
2118 	}
2119 
2120 	if (type & PERF_SAMPLE_RAW) {
2121 		OVERFLOW_CHECK_u64(array);
2122 		u.val64 = *array;
2123 
2124 		/*
2125 		 * Undo swap of u64, then swap on individual u32s,
2126 		 * get the size of the raw area and undo all of the
2127 		 * swap. The pevent interface handles endianity by
2128 		 * itself.
2129 		 */
2130 		if (swapped) {
2131 			u.val64 = bswap_64(u.val64);
2132 			u.val32[0] = bswap_32(u.val32[0]);
2133 			u.val32[1] = bswap_32(u.val32[1]);
2134 		}
2135 		data->raw_size = u.val32[0];
2136 
2137 		/*
2138 		 * The raw data is aligned on 64bits including the
2139 		 * u32 size, so it's safe to use mem_bswap_64.
2140 		 */
2141 		if (swapped)
2142 			mem_bswap_64((void *) array, data->raw_size);
2143 
2144 		array = (void *)array + sizeof(u32);
2145 
2146 		OVERFLOW_CHECK(array, data->raw_size, max_size);
2147 		data->raw_data = (void *)array;
2148 		array = (void *)array + data->raw_size;
2149 	}
2150 
2151 	if (type & PERF_SAMPLE_BRANCH_STACK) {
2152 		const u64 max_branch_nr = UINT64_MAX /
2153 					  sizeof(struct branch_entry);
2154 
2155 		OVERFLOW_CHECK_u64(array);
2156 		data->branch_stack = (struct branch_stack *)array++;
2157 
2158 		if (data->branch_stack->nr > max_branch_nr)
2159 			return -EFAULT;
2160 
2161 		sz = data->branch_stack->nr * sizeof(struct branch_entry);
2162 		if (evsel__has_branch_hw_idx(evsel))
2163 			sz += sizeof(u64);
2164 		else
2165 			data->no_hw_idx = true;
2166 		OVERFLOW_CHECK(array, sz, max_size);
2167 		array = (void *)array + sz;
2168 	}
2169 
2170 	if (type & PERF_SAMPLE_REGS_USER) {
2171 		OVERFLOW_CHECK_u64(array);
2172 		data->user_regs.abi = *array;
2173 		array++;
2174 
2175 		if (data->user_regs.abi) {
2176 			u64 mask = evsel->core.attr.sample_regs_user;
2177 
2178 			sz = hweight64(mask) * sizeof(u64);
2179 			OVERFLOW_CHECK(array, sz, max_size);
2180 			data->user_regs.mask = mask;
2181 			data->user_regs.regs = (u64 *)array;
2182 			array = (void *)array + sz;
2183 		}
2184 	}
2185 
2186 	if (type & PERF_SAMPLE_STACK_USER) {
2187 		OVERFLOW_CHECK_u64(array);
2188 		sz = *array++;
2189 
2190 		data->user_stack.offset = ((char *)(array - 1)
2191 					  - (char *) event);
2192 
2193 		if (!sz) {
2194 			data->user_stack.size = 0;
2195 		} else {
2196 			OVERFLOW_CHECK(array, sz, max_size);
2197 			data->user_stack.data = (char *)array;
2198 			array = (void *)array + sz;
2199 			OVERFLOW_CHECK_u64(array);
2200 			data->user_stack.size = *array++;
2201 			if (WARN_ONCE(data->user_stack.size > sz,
2202 				      "user stack dump failure\n"))
2203 				return -EFAULT;
2204 		}
2205 	}
2206 
2207 	if (type & PERF_SAMPLE_WEIGHT) {
2208 		OVERFLOW_CHECK_u64(array);
2209 		data->weight = *array;
2210 		array++;
2211 	}
2212 
2213 	if (type & PERF_SAMPLE_DATA_SRC) {
2214 		OVERFLOW_CHECK_u64(array);
2215 		data->data_src = *array;
2216 		array++;
2217 	}
2218 
2219 	if (type & PERF_SAMPLE_TRANSACTION) {
2220 		OVERFLOW_CHECK_u64(array);
2221 		data->transaction = *array;
2222 		array++;
2223 	}
2224 
2225 	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2226 	if (type & PERF_SAMPLE_REGS_INTR) {
2227 		OVERFLOW_CHECK_u64(array);
2228 		data->intr_regs.abi = *array;
2229 		array++;
2230 
2231 		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2232 			u64 mask = evsel->core.attr.sample_regs_intr;
2233 
2234 			sz = hweight64(mask) * sizeof(u64);
2235 			OVERFLOW_CHECK(array, sz, max_size);
2236 			data->intr_regs.mask = mask;
2237 			data->intr_regs.regs = (u64 *)array;
2238 			array = (void *)array + sz;
2239 		}
2240 	}
2241 
2242 	data->phys_addr = 0;
2243 	if (type & PERF_SAMPLE_PHYS_ADDR) {
2244 		data->phys_addr = *array;
2245 		array++;
2246 	}
2247 
2248 	data->cgroup = 0;
2249 	if (type & PERF_SAMPLE_CGROUP) {
2250 		data->cgroup = *array;
2251 		array++;
2252 	}
2253 
2254 	if (type & PERF_SAMPLE_AUX) {
2255 		OVERFLOW_CHECK_u64(array);
2256 		sz = *array++;
2257 
2258 		OVERFLOW_CHECK(array, sz, max_size);
2259 		/* Undo swap of data */
2260 		if (swapped)
2261 			mem_bswap_64((char *)array, sz);
2262 		data->aux_sample.size = sz;
2263 		data->aux_sample.data = (char *)array;
2264 		array = (void *)array + sz;
2265 	}
2266 
2267 	return 0;
2268 }
2269 
2270 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2271 				  u64 *timestamp)
2272 {
2273 	u64 type = evsel->core.attr.sample_type;
2274 	const __u64 *array;
2275 
2276 	if (!(type & PERF_SAMPLE_TIME))
2277 		return -1;
2278 
2279 	if (event->header.type != PERF_RECORD_SAMPLE) {
2280 		struct perf_sample data = {
2281 			.time = -1ULL,
2282 		};
2283 
2284 		if (!evsel->core.attr.sample_id_all)
2285 			return -1;
2286 		if (perf_evsel__parse_id_sample(evsel, event, &data))
2287 			return -1;
2288 
2289 		*timestamp = data.time;
2290 		return 0;
2291 	}
2292 
2293 	array = event->sample.array;
2294 
2295 	if (perf_event__check_size(event, evsel->sample_size))
2296 		return -EFAULT;
2297 
2298 	if (type & PERF_SAMPLE_IDENTIFIER)
2299 		array++;
2300 
2301 	if (type & PERF_SAMPLE_IP)
2302 		array++;
2303 
2304 	if (type & PERF_SAMPLE_TID)
2305 		array++;
2306 
2307 	if (type & PERF_SAMPLE_TIME)
2308 		*timestamp = *array;
2309 
2310 	return 0;
2311 }
2312 
2313 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2314 {
2315 	return tep_find_field(evsel->tp_format, name);
2316 }
2317 
2318 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2319 {
2320 	struct tep_format_field *field = evsel__field(evsel, name);
2321 	int offset;
2322 
2323 	if (!field)
2324 		return NULL;
2325 
2326 	offset = field->offset;
2327 
2328 	if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2329 		offset = *(int *)(sample->raw_data + field->offset);
2330 		offset &= 0xffff;
2331 	}
2332 
2333 	return sample->raw_data + offset;
2334 }
2335 
2336 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2337 			 bool needs_swap)
2338 {
2339 	u64 value;
2340 	void *ptr = sample->raw_data + field->offset;
2341 
2342 	switch (field->size) {
2343 	case 1:
2344 		return *(u8 *)ptr;
2345 	case 2:
2346 		value = *(u16 *)ptr;
2347 		break;
2348 	case 4:
2349 		value = *(u32 *)ptr;
2350 		break;
2351 	case 8:
2352 		memcpy(&value, ptr, sizeof(u64));
2353 		break;
2354 	default:
2355 		return 0;
2356 	}
2357 
2358 	if (!needs_swap)
2359 		return value;
2360 
2361 	switch (field->size) {
2362 	case 2:
2363 		return bswap_16(value);
2364 	case 4:
2365 		return bswap_32(value);
2366 	case 8:
2367 		return bswap_64(value);
2368 	default:
2369 		return 0;
2370 	}
2371 
2372 	return 0;
2373 }
2374 
2375 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2376 {
2377 	struct tep_format_field *field = evsel__field(evsel, name);
2378 
2379 	if (!field)
2380 		return 0;
2381 
2382 	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2383 }
2384 
2385 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2386 {
2387 	int paranoid;
2388 
2389 	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2390 	    evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2391 	    evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2392 		/*
2393 		 * If it's cycles then fall back to hrtimer based
2394 		 * cpu-clock-tick sw counter, which is always available even if
2395 		 * no PMU support.
2396 		 *
2397 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2398 		 * b0a873e).
2399 		 */
2400 		scnprintf(msg, msgsize, "%s",
2401 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2402 
2403 		evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2404 		evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2405 
2406 		zfree(&evsel->name);
2407 		return true;
2408 	} else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2409 		   (paranoid = perf_event_paranoid()) > 1) {
2410 		const char *name = evsel__name(evsel);
2411 		char *new_name;
2412 		const char *sep = ":";
2413 
2414 		/* If event has exclude user then don't exclude kernel. */
2415 		if (evsel->core.attr.exclude_user)
2416 			return false;
2417 
2418 		/* Is there already the separator in the name. */
2419 		if (strchr(name, '/') ||
2420 		    (strchr(name, ':') && !evsel->is_libpfm_event))
2421 			sep = "";
2422 
2423 		if (asprintf(&new_name, "%s%su", name, sep) < 0)
2424 			return false;
2425 
2426 		if (evsel->name)
2427 			free(evsel->name);
2428 		evsel->name = new_name;
2429 		scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2430 			  "to fall back to excluding kernel and hypervisor "
2431 			  " samples", paranoid);
2432 		evsel->core.attr.exclude_kernel = 1;
2433 		evsel->core.attr.exclude_hv     = 1;
2434 
2435 		return true;
2436 	}
2437 
2438 	return false;
2439 }
2440 
2441 static bool find_process(const char *name)
2442 {
2443 	size_t len = strlen(name);
2444 	DIR *dir;
2445 	struct dirent *d;
2446 	int ret = -1;
2447 
2448 	dir = opendir(procfs__mountpoint());
2449 	if (!dir)
2450 		return false;
2451 
2452 	/* Walk through the directory. */
2453 	while (ret && (d = readdir(dir)) != NULL) {
2454 		char path[PATH_MAX];
2455 		char *data;
2456 		size_t size;
2457 
2458 		if ((d->d_type != DT_DIR) ||
2459 		     !strcmp(".", d->d_name) ||
2460 		     !strcmp("..", d->d_name))
2461 			continue;
2462 
2463 		scnprintf(path, sizeof(path), "%s/%s/comm",
2464 			  procfs__mountpoint(), d->d_name);
2465 
2466 		if (filename__read_str(path, &data, &size))
2467 			continue;
2468 
2469 		ret = strncmp(name, data, len);
2470 		free(data);
2471 	}
2472 
2473 	closedir(dir);
2474 	return ret ? false : true;
2475 }
2476 
2477 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2478 			 int err, char *msg, size_t size)
2479 {
2480 	char sbuf[STRERR_BUFSIZE];
2481 	int printed = 0, enforced = 0;
2482 
2483 	switch (err) {
2484 	case EPERM:
2485 	case EACCES:
2486 		printed += scnprintf(msg + printed, size - printed,
2487 			"Access to performance monitoring and observability operations is limited.\n");
2488 
2489 		if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2490 			if (enforced) {
2491 				printed += scnprintf(msg + printed, size - printed,
2492 					"Enforced MAC policy settings (SELinux) can limit access to performance\n"
2493 					"monitoring and observability operations. Inspect system audit records for\n"
2494 					"more perf_event access control information and adjusting the policy.\n");
2495 			}
2496 		}
2497 
2498 		if (err == EPERM)
2499 			printed += scnprintf(msg, size,
2500 				"No permission to enable %s event.\n\n", evsel__name(evsel));
2501 
2502 		return scnprintf(msg + printed, size - printed,
2503 		 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2504 		 "access to performance monitoring and observability operations for users\n"
2505 		 "without CAP_PERFMON or CAP_SYS_ADMIN Linux capability.\n"
2506 		 "perf_event_paranoid setting is %d:\n"
2507 		 "  -1: Allow use of (almost) all events by all users\n"
2508 		 "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2509 		 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2510 		 ">= 1: Disallow CPU event access\n"
2511 		 ">= 2: Disallow kernel profiling\n"
2512 		 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2513 		 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2514 		 perf_event_paranoid());
2515 	case ENOENT:
2516 		return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2517 	case EMFILE:
2518 		return scnprintf(msg, size, "%s",
2519 			 "Too many events are opened.\n"
2520 			 "Probably the maximum number of open file descriptors has been reached.\n"
2521 			 "Hint: Try again after reducing the number of events.\n"
2522 			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2523 	case ENOMEM:
2524 		if (evsel__has_callchain(evsel) &&
2525 		    access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2526 			return scnprintf(msg, size,
2527 					 "Not enough memory to setup event with callchain.\n"
2528 					 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2529 					 "Hint: Current value: %d", sysctl__max_stack());
2530 		break;
2531 	case ENODEV:
2532 		if (target->cpu_list)
2533 			return scnprintf(msg, size, "%s",
2534 	 "No such device - did you specify an out-of-range profile CPU?");
2535 		break;
2536 	case EOPNOTSUPP:
2537 		if (evsel->core.attr.sample_period != 0)
2538 			return scnprintf(msg, size,
2539 	"%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2540 					 evsel__name(evsel));
2541 		if (evsel->core.attr.precise_ip)
2542 			return scnprintf(msg, size, "%s",
2543 	"\'precise\' request may not be supported. Try removing 'p' modifier.");
2544 #if defined(__i386__) || defined(__x86_64__)
2545 		if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2546 			return scnprintf(msg, size, "%s",
2547 	"No hardware sampling interrupt available.\n");
2548 #endif
2549 		break;
2550 	case EBUSY:
2551 		if (find_process("oprofiled"))
2552 			return scnprintf(msg, size,
2553 	"The PMU counters are busy/taken by another profiler.\n"
2554 	"We found oprofile daemon running, please stop it and try again.");
2555 		break;
2556 	case EINVAL:
2557 		if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2558 			return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2559 		if (perf_missing_features.clockid)
2560 			return scnprintf(msg, size, "clockid feature not supported.");
2561 		if (perf_missing_features.clockid_wrong)
2562 			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2563 		if (perf_missing_features.aux_output)
2564 			return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2565 		break;
2566 	default:
2567 		break;
2568 	}
2569 
2570 	return scnprintf(msg, size,
2571 	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2572 	"/bin/dmesg | grep -i perf may provide additional information.\n",
2573 			 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2574 }
2575 
2576 struct perf_env *evsel__env(struct evsel *evsel)
2577 {
2578 	if (evsel && evsel->evlist)
2579 		return evsel->evlist->env;
2580 	return &perf_env;
2581 }
2582 
2583 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2584 {
2585 	int cpu, thread;
2586 
2587 	for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2588 		for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2589 		     thread++) {
2590 			int fd = FD(evsel, cpu, thread);
2591 
2592 			if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2593 						   cpu, thread, fd) < 0)
2594 				return -1;
2595 		}
2596 	}
2597 
2598 	return 0;
2599 }
2600 
2601 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2602 {
2603 	struct perf_cpu_map *cpus = evsel->core.cpus;
2604 	struct perf_thread_map *threads = evsel->core.threads;
2605 
2606 	if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2607 		return -ENOMEM;
2608 
2609 	return store_evsel_ids(evsel, evlist);
2610 }
2611