xref: /linux/tools/perf/util/evsel.c (revision 8a79db5e83a5d52c74e6f3c40d6f312cf899213e)
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 perf_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 perf_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 = perf_evsel__no_extra_init,
77 	.fini = perf_evsel__no_extra_fini,
78 };
79 
80 int perf_evsel__object_config(size_t object_size,
81 			      int (*init)(struct evsel *evsel),
82 			      void (*fini)(struct evsel *evsel))
83 {
84 
85 	if (object_size == 0)
86 		goto set_methods;
87 
88 	if (perf_evsel__object.size > object_size)
89 		return -EINVAL;
90 
91 	perf_evsel__object.size = object_size;
92 
93 set_methods:
94 	if (init != NULL)
95 		perf_evsel__object.init = init;
96 
97 	if (fini != NULL)
98 		perf_evsel__object.fini = fini;
99 
100 	return 0;
101 }
102 
103 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
104 
105 int __perf_evsel__sample_size(u64 sample_type)
106 {
107 	u64 mask = sample_type & PERF_SAMPLE_MASK;
108 	int size = 0;
109 	int i;
110 
111 	for (i = 0; i < 64; i++) {
112 		if (mask & (1ULL << i))
113 			size++;
114 	}
115 
116 	size *= sizeof(u64);
117 
118 	return size;
119 }
120 
121 /**
122  * __perf_evsel__calc_id_pos - calculate id_pos.
123  * @sample_type: sample type
124  *
125  * This function returns the position of the event id (PERF_SAMPLE_ID or
126  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
127  * perf_record_sample.
128  */
129 static int __perf_evsel__calc_id_pos(u64 sample_type)
130 {
131 	int idx = 0;
132 
133 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
134 		return 0;
135 
136 	if (!(sample_type & PERF_SAMPLE_ID))
137 		return -1;
138 
139 	if (sample_type & PERF_SAMPLE_IP)
140 		idx += 1;
141 
142 	if (sample_type & PERF_SAMPLE_TID)
143 		idx += 1;
144 
145 	if (sample_type & PERF_SAMPLE_TIME)
146 		idx += 1;
147 
148 	if (sample_type & PERF_SAMPLE_ADDR)
149 		idx += 1;
150 
151 	return idx;
152 }
153 
154 /**
155  * __perf_evsel__calc_is_pos - calculate is_pos.
156  * @sample_type: sample type
157  *
158  * This function returns the position (counting backwards) of the event id
159  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
160  * sample_id_all is used there is an id sample appended to non-sample events.
161  */
162 static int __perf_evsel__calc_is_pos(u64 sample_type)
163 {
164 	int idx = 1;
165 
166 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
167 		return 1;
168 
169 	if (!(sample_type & PERF_SAMPLE_ID))
170 		return -1;
171 
172 	if (sample_type & PERF_SAMPLE_CPU)
173 		idx += 1;
174 
175 	if (sample_type & PERF_SAMPLE_STREAM_ID)
176 		idx += 1;
177 
178 	return idx;
179 }
180 
181 void perf_evsel__calc_id_pos(struct evsel *evsel)
182 {
183 	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
184 	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
185 }
186 
187 void __perf_evsel__set_sample_bit(struct evsel *evsel,
188 				  enum perf_event_sample_format bit)
189 {
190 	if (!(evsel->core.attr.sample_type & bit)) {
191 		evsel->core.attr.sample_type |= bit;
192 		evsel->sample_size += sizeof(u64);
193 		perf_evsel__calc_id_pos(evsel);
194 	}
195 }
196 
197 void __perf_evsel__reset_sample_bit(struct evsel *evsel,
198 				    enum perf_event_sample_format bit)
199 {
200 	if (evsel->core.attr.sample_type & bit) {
201 		evsel->core.attr.sample_type &= ~bit;
202 		evsel->sample_size -= sizeof(u64);
203 		perf_evsel__calc_id_pos(evsel);
204 	}
205 }
206 
207 void perf_evsel__set_sample_id(struct evsel *evsel,
208 			       bool can_sample_identifier)
209 {
210 	if (can_sample_identifier) {
211 		perf_evsel__reset_sample_bit(evsel, ID);
212 		perf_evsel__set_sample_bit(evsel, IDENTIFIER);
213 	} else {
214 		perf_evsel__set_sample_bit(evsel, ID);
215 	}
216 	evsel->core.attr.read_format |= PERF_FORMAT_ID;
217 }
218 
219 /**
220  * perf_evsel__is_function_event - Return whether given evsel is a function
221  * trace event
222  *
223  * @evsel - evsel selector to be tested
224  *
225  * Return %true if event is function trace event
226  */
227 bool perf_evsel__is_function_event(struct evsel *evsel)
228 {
229 #define FUNCTION_EVENT "ftrace:function"
230 
231 	return evsel->name &&
232 	       !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
233 
234 #undef FUNCTION_EVENT
235 }
236 
237 void evsel__init(struct evsel *evsel,
238 		 struct perf_event_attr *attr, int idx)
239 {
240 	perf_evsel__init(&evsel->core, attr);
241 	evsel->idx	   = idx;
242 	evsel->tracking	   = !idx;
243 	evsel->leader	   = evsel;
244 	evsel->unit	   = "";
245 	evsel->scale	   = 1.0;
246 	evsel->max_events  = ULONG_MAX;
247 	evsel->evlist	   = NULL;
248 	evsel->bpf_obj	   = NULL;
249 	evsel->bpf_fd	   = -1;
250 	INIT_LIST_HEAD(&evsel->config_terms);
251 	perf_evsel__object.init(evsel);
252 	evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
253 	perf_evsel__calc_id_pos(evsel);
254 	evsel->cmdline_group_boundary = false;
255 	evsel->metric_expr   = NULL;
256 	evsel->metric_name   = NULL;
257 	evsel->metric_events = NULL;
258 	evsel->collect_stat  = false;
259 	evsel->pmu_name      = NULL;
260 }
261 
262 struct evsel *perf_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 (perf_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 (perf_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 *perf_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 *perf_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 *perf_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 *__perf_evsel__hw_name(u64 config)
389 {
390 	if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
391 		return perf_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 perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
433 {
434 	int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config));
435 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
436 }
437 
438 const char *perf_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 *__perf_evsel__sw_name(u64 config)
452 {
453 	if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
454 		return perf_evsel__sw_names[config];
455 	return "unknown-software";
456 }
457 
458 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
459 {
460 	int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config));
461 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
462 }
463 
464 static int __perf_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 perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
483 {
484 	struct perf_event_attr *attr = &evsel->core.attr;
485 	int r = __perf_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 *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
490 				[PERF_EVSEL__MAX_ALIASES] = {
491  { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
492  { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
493  { "LLC",	"L2",							},
494  { "dTLB",	"d-tlb",	"Data-TLB",				},
495  { "iTLB",	"i-tlb",	"Instruction-TLB",			},
496  { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
497  { "node",								},
498 };
499 
500 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
501 				   [PERF_EVSEL__MAX_ALIASES] = {
502  { "load",	"loads",	"read",					},
503  { "store",	"stores",	"write",				},
504  { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
505 };
506 
507 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
508 				       [PERF_EVSEL__MAX_ALIASES] = {
509  { "refs",	"Reference",	"ops",		"access",		},
510  { "misses",	"miss",							},
511 };
512 
513 #define C(x)		PERF_COUNT_HW_CACHE_##x
514 #define CACHE_READ	(1 << C(OP_READ))
515 #define CACHE_WRITE	(1 << C(OP_WRITE))
516 #define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
517 #define COP(x)		(1 << x)
518 
519 /*
520  * cache operartion stat
521  * L1I : Read and prefetch only
522  * ITLB and BPU : Read-only
523  */
524 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
525  [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
526  [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
527  [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
528  [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
529  [C(ITLB)]	= (CACHE_READ),
530  [C(BPU)]	= (CACHE_READ),
531  [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
532 };
533 
534 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
535 {
536 	if (perf_evsel__hw_cache_stat[type] & COP(op))
537 		return true;	/* valid */
538 	else
539 		return false;	/* invalid */
540 }
541 
542 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
543 					    char *bf, size_t size)
544 {
545 	if (result) {
546 		return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
547 				 perf_evsel__hw_cache_op[op][0],
548 				 perf_evsel__hw_cache_result[result][0]);
549 	}
550 
551 	return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
552 			 perf_evsel__hw_cache_op[op][1]);
553 }
554 
555 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
556 {
557 	u8 op, result, type = (config >>  0) & 0xff;
558 	const char *err = "unknown-ext-hardware-cache-type";
559 
560 	if (type >= PERF_COUNT_HW_CACHE_MAX)
561 		goto out_err;
562 
563 	op = (config >>  8) & 0xff;
564 	err = "unknown-ext-hardware-cache-op";
565 	if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
566 		goto out_err;
567 
568 	result = (config >> 16) & 0xff;
569 	err = "unknown-ext-hardware-cache-result";
570 	if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
571 		goto out_err;
572 
573 	err = "invalid-cache";
574 	if (!perf_evsel__is_cache_op_valid(type, op))
575 		goto out_err;
576 
577 	return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
578 out_err:
579 	return scnprintf(bf, size, "%s", err);
580 }
581 
582 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
583 {
584 	int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size);
585 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
586 }
587 
588 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
589 {
590 	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
591 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
592 }
593 
594 static int perf_evsel__tool_name(char *bf, size_t size)
595 {
596 	int ret = scnprintf(bf, size, "duration_time");
597 	return ret;
598 }
599 
600 const char *perf_evsel__name(struct evsel *evsel)
601 {
602 	char bf[128];
603 
604 	if (!evsel)
605 		goto out_unknown;
606 
607 	if (evsel->name)
608 		return evsel->name;
609 
610 	switch (evsel->core.attr.type) {
611 	case PERF_TYPE_RAW:
612 		perf_evsel__raw_name(evsel, bf, sizeof(bf));
613 		break;
614 
615 	case PERF_TYPE_HARDWARE:
616 		perf_evsel__hw_name(evsel, bf, sizeof(bf));
617 		break;
618 
619 	case PERF_TYPE_HW_CACHE:
620 		perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
621 		break;
622 
623 	case PERF_TYPE_SOFTWARE:
624 		if (evsel->tool_event)
625 			perf_evsel__tool_name(bf, sizeof(bf));
626 		else
627 			perf_evsel__sw_name(evsel, bf, sizeof(bf));
628 		break;
629 
630 	case PERF_TYPE_TRACEPOINT:
631 		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
632 		break;
633 
634 	case PERF_TYPE_BREAKPOINT:
635 		perf_evsel__bp_name(evsel, bf, sizeof(bf));
636 		break;
637 
638 	default:
639 		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
640 			  evsel->core.attr.type);
641 		break;
642 	}
643 
644 	evsel->name = strdup(bf);
645 
646 	if (evsel->name)
647 		return evsel->name;
648 out_unknown:
649 	return "unknown";
650 }
651 
652 const char *perf_evsel__group_name(struct evsel *evsel)
653 {
654 	return evsel->group_name ?: "anon group";
655 }
656 
657 /*
658  * Returns the group details for the specified leader,
659  * with following rules.
660  *
661  *  For record -e '{cycles,instructions}'
662  *    'anon group { cycles:u, instructions:u }'
663  *
664  *  For record -e 'cycles,instructions' and report --group
665  *    'cycles:u, instructions:u'
666  */
667 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
668 {
669 	int ret = 0;
670 	struct evsel *pos;
671 	const char *group_name = perf_evsel__group_name(evsel);
672 
673 	if (!evsel->forced_leader)
674 		ret = scnprintf(buf, size, "%s { ", group_name);
675 
676 	ret += scnprintf(buf + ret, size - ret, "%s",
677 			 perf_evsel__name(evsel));
678 
679 	for_each_group_member(pos, evsel)
680 		ret += scnprintf(buf + ret, size - ret, ", %s",
681 				 perf_evsel__name(pos));
682 
683 	if (!evsel->forced_leader)
684 		ret += scnprintf(buf + ret, size - ret, " }");
685 
686 	return ret;
687 }
688 
689 static void __perf_evsel__config_callchain(struct evsel *evsel,
690 					   struct record_opts *opts,
691 					   struct callchain_param *param)
692 {
693 	bool function = perf_evsel__is_function_event(evsel);
694 	struct perf_event_attr *attr = &evsel->core.attr;
695 
696 	perf_evsel__set_sample_bit(evsel, CALLCHAIN);
697 
698 	attr->sample_max_stack = param->max_stack;
699 
700 	if (opts->kernel_callchains)
701 		attr->exclude_callchain_user = 1;
702 	if (opts->user_callchains)
703 		attr->exclude_callchain_kernel = 1;
704 	if (param->record_mode == CALLCHAIN_LBR) {
705 		if (!opts->branch_stack) {
706 			if (attr->exclude_user) {
707 				pr_warning("LBR callstack option is only available "
708 					   "to get user callchain information. "
709 					   "Falling back to framepointers.\n");
710 			} else {
711 				perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
712 				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
713 							PERF_SAMPLE_BRANCH_CALL_STACK |
714 							PERF_SAMPLE_BRANCH_NO_CYCLES |
715 							PERF_SAMPLE_BRANCH_NO_FLAGS;
716 			}
717 		} else
718 			 pr_warning("Cannot use LBR callstack with branch stack. "
719 				    "Falling back to framepointers.\n");
720 	}
721 
722 	if (param->record_mode == CALLCHAIN_DWARF) {
723 		if (!function) {
724 			perf_evsel__set_sample_bit(evsel, REGS_USER);
725 			perf_evsel__set_sample_bit(evsel, STACK_USER);
726 			if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
727 				attr->sample_regs_user |= DWARF_MINIMAL_REGS;
728 				pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
729 					   "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
730 					   "so the minimal registers set (IP, SP) is explicitly forced.\n");
731 			} else {
732 				attr->sample_regs_user |= PERF_REGS_MASK;
733 			}
734 			attr->sample_stack_user = param->dump_size;
735 			attr->exclude_callchain_user = 1;
736 		} else {
737 			pr_info("Cannot use DWARF unwind for function trace event,"
738 				" falling back to framepointers.\n");
739 		}
740 	}
741 
742 	if (function) {
743 		pr_info("Disabling user space callchains for function trace event.\n");
744 		attr->exclude_callchain_user = 1;
745 	}
746 }
747 
748 void perf_evsel__config_callchain(struct evsel *evsel,
749 				  struct record_opts *opts,
750 				  struct callchain_param *param)
751 {
752 	if (param->enabled)
753 		return __perf_evsel__config_callchain(evsel, opts, param);
754 }
755 
756 static void
757 perf_evsel__reset_callgraph(struct evsel *evsel,
758 			    struct callchain_param *param)
759 {
760 	struct perf_event_attr *attr = &evsel->core.attr;
761 
762 	perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
763 	if (param->record_mode == CALLCHAIN_LBR) {
764 		perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
765 		attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
766 					      PERF_SAMPLE_BRANCH_CALL_STACK);
767 	}
768 	if (param->record_mode == CALLCHAIN_DWARF) {
769 		perf_evsel__reset_sample_bit(evsel, REGS_USER);
770 		perf_evsel__reset_sample_bit(evsel, STACK_USER);
771 	}
772 }
773 
774 static void apply_config_terms(struct evsel *evsel,
775 			       struct record_opts *opts, bool track)
776 {
777 	struct perf_evsel_config_term *term;
778 	struct list_head *config_terms = &evsel->config_terms;
779 	struct perf_event_attr *attr = &evsel->core.attr;
780 	/* callgraph default */
781 	struct callchain_param param = {
782 		.record_mode = callchain_param.record_mode,
783 	};
784 	u32 dump_size = 0;
785 	int max_stack = 0;
786 	const char *callgraph_buf = NULL;
787 
788 	list_for_each_entry(term, config_terms, list) {
789 		switch (term->type) {
790 		case PERF_EVSEL__CONFIG_TERM_PERIOD:
791 			if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
792 				attr->sample_period = term->val.period;
793 				attr->freq = 0;
794 				perf_evsel__reset_sample_bit(evsel, PERIOD);
795 			}
796 			break;
797 		case PERF_EVSEL__CONFIG_TERM_FREQ:
798 			if (!(term->weak && opts->user_freq != UINT_MAX)) {
799 				attr->sample_freq = term->val.freq;
800 				attr->freq = 1;
801 				perf_evsel__set_sample_bit(evsel, PERIOD);
802 			}
803 			break;
804 		case PERF_EVSEL__CONFIG_TERM_TIME:
805 			if (term->val.time)
806 				perf_evsel__set_sample_bit(evsel, TIME);
807 			else
808 				perf_evsel__reset_sample_bit(evsel, TIME);
809 			break;
810 		case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
811 			callgraph_buf = term->val.callgraph;
812 			break;
813 		case PERF_EVSEL__CONFIG_TERM_BRANCH:
814 			if (term->val.branch && strcmp(term->val.branch, "no")) {
815 				perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
816 				parse_branch_str(term->val.branch,
817 						 &attr->branch_sample_type);
818 			} else
819 				perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
820 			break;
821 		case PERF_EVSEL__CONFIG_TERM_STACK_USER:
822 			dump_size = term->val.stack_user;
823 			break;
824 		case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
825 			max_stack = term->val.max_stack;
826 			break;
827 		case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
828 			evsel->max_events = term->val.max_events;
829 			break;
830 		case PERF_EVSEL__CONFIG_TERM_INHERIT:
831 			/*
832 			 * attr->inherit should has already been set by
833 			 * perf_evsel__config. If user explicitly set
834 			 * inherit using config terms, override global
835 			 * opt->no_inherit setting.
836 			 */
837 			attr->inherit = term->val.inherit ? 1 : 0;
838 			break;
839 		case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
840 			attr->write_backward = term->val.overwrite ? 1 : 0;
841 			break;
842 		case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
843 			break;
844 		case PERF_EVSEL__CONFIG_TERM_PERCORE:
845 			break;
846 		case PERF_EVSEL__CONFIG_TERM_AUX_OUTPUT:
847 			attr->aux_output = term->val.aux_output ? 1 : 0;
848 			break;
849 		case PERF_EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
850 			/* Already applied by auxtrace */
851 			break;
852 		case PERF_EVSEL__CONFIG_TERM_CFG_CHG:
853 			break;
854 		default:
855 			break;
856 		}
857 	}
858 
859 	/* User explicitly set per-event callgraph, clear the old setting and reset. */
860 	if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
861 		bool sample_address = false;
862 
863 		if (max_stack) {
864 			param.max_stack = max_stack;
865 			if (callgraph_buf == NULL)
866 				callgraph_buf = "fp";
867 		}
868 
869 		/* parse callgraph parameters */
870 		if (callgraph_buf != NULL) {
871 			if (!strcmp(callgraph_buf, "no")) {
872 				param.enabled = false;
873 				param.record_mode = CALLCHAIN_NONE;
874 			} else {
875 				param.enabled = true;
876 				if (parse_callchain_record(callgraph_buf, &param)) {
877 					pr_err("per-event callgraph setting for %s failed. "
878 					       "Apply callgraph global setting for it\n",
879 					       evsel->name);
880 					return;
881 				}
882 				if (param.record_mode == CALLCHAIN_DWARF)
883 					sample_address = true;
884 			}
885 		}
886 		if (dump_size > 0) {
887 			dump_size = round_up(dump_size, sizeof(u64));
888 			param.dump_size = dump_size;
889 		}
890 
891 		/* If global callgraph set, clear it */
892 		if (callchain_param.enabled)
893 			perf_evsel__reset_callgraph(evsel, &callchain_param);
894 
895 		/* set perf-event callgraph */
896 		if (param.enabled) {
897 			if (sample_address) {
898 				perf_evsel__set_sample_bit(evsel, ADDR);
899 				perf_evsel__set_sample_bit(evsel, DATA_SRC);
900 				evsel->core.attr.mmap_data = track;
901 			}
902 			perf_evsel__config_callchain(evsel, opts, &param);
903 		}
904 	}
905 }
906 
907 static bool is_dummy_event(struct evsel *evsel)
908 {
909 	return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
910 	       (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
911 }
912 
913 struct perf_evsel_config_term *__perf_evsel__get_config_term(struct evsel *evsel,
914 							     enum evsel_term_type type)
915 {
916 	struct perf_evsel_config_term *term, *found_term = NULL;
917 
918 	list_for_each_entry(term, &evsel->config_terms, list) {
919 		if (term->type == type)
920 			found_term = term;
921 	}
922 
923 	return found_term;
924 }
925 
926 /*
927  * The enable_on_exec/disabled value strategy:
928  *
929  *  1) For any type of traced program:
930  *    - all independent events and group leaders are disabled
931  *    - all group members are enabled
932  *
933  *     Group members are ruled by group leaders. They need to
934  *     be enabled, because the group scheduling relies on that.
935  *
936  *  2) For traced programs executed by perf:
937  *     - all independent events and group leaders have
938  *       enable_on_exec set
939  *     - we don't specifically enable or disable any event during
940  *       the record command
941  *
942  *     Independent events and group leaders are initially disabled
943  *     and get enabled by exec. Group members are ruled by group
944  *     leaders as stated in 1).
945  *
946  *  3) For traced programs attached by perf (pid/tid):
947  *     - we specifically enable or disable all events during
948  *       the record command
949  *
950  *     When attaching events to already running traced we
951  *     enable/disable events specifically, as there's no
952  *     initial traced exec call.
953  */
954 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
955 			struct callchain_param *callchain)
956 {
957 	struct evsel *leader = evsel->leader;
958 	struct perf_event_attr *attr = &evsel->core.attr;
959 	int track = evsel->tracking;
960 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
961 
962 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
963 	attr->inherit	    = !opts->no_inherit;
964 	attr->write_backward = opts->overwrite ? 1 : 0;
965 
966 	perf_evsel__set_sample_bit(evsel, IP);
967 	perf_evsel__set_sample_bit(evsel, TID);
968 
969 	if (evsel->sample_read) {
970 		perf_evsel__set_sample_bit(evsel, READ);
971 
972 		/*
973 		 * We need ID even in case of single event, because
974 		 * PERF_SAMPLE_READ process ID specific data.
975 		 */
976 		perf_evsel__set_sample_id(evsel, false);
977 
978 		/*
979 		 * Apply group format only if we belong to group
980 		 * with more than one members.
981 		 */
982 		if (leader->core.nr_members > 1) {
983 			attr->read_format |= PERF_FORMAT_GROUP;
984 			attr->inherit = 0;
985 		}
986 	}
987 
988 	/*
989 	 * We default some events to have a default interval. But keep
990 	 * it a weak assumption overridable by the user.
991 	 */
992 	if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
993 				     opts->user_interval != ULLONG_MAX)) {
994 		if (opts->freq) {
995 			perf_evsel__set_sample_bit(evsel, PERIOD);
996 			attr->freq		= 1;
997 			attr->sample_freq	= opts->freq;
998 		} else {
999 			attr->sample_period = opts->default_interval;
1000 		}
1001 	}
1002 
1003 	/*
1004 	 * Disable sampling for all group members other
1005 	 * than leader in case leader 'leads' the sampling.
1006 	 */
1007 	if ((leader != evsel) && leader->sample_read) {
1008 		attr->freq           = 0;
1009 		attr->sample_freq    = 0;
1010 		attr->sample_period  = 0;
1011 		attr->write_backward = 0;
1012 
1013 		/*
1014 		 * We don't get sample for slave events, we make them
1015 		 * when delivering group leader sample. Set the slave
1016 		 * event to follow the master sample_type to ease up
1017 		 * report.
1018 		 */
1019 		attr->sample_type = leader->core.attr.sample_type;
1020 	}
1021 
1022 	if (opts->no_samples)
1023 		attr->sample_freq = 0;
1024 
1025 	if (opts->inherit_stat) {
1026 		evsel->core.attr.read_format |=
1027 			PERF_FORMAT_TOTAL_TIME_ENABLED |
1028 			PERF_FORMAT_TOTAL_TIME_RUNNING |
1029 			PERF_FORMAT_ID;
1030 		attr->inherit_stat = 1;
1031 	}
1032 
1033 	if (opts->sample_address) {
1034 		perf_evsel__set_sample_bit(evsel, ADDR);
1035 		attr->mmap_data = track;
1036 	}
1037 
1038 	/*
1039 	 * We don't allow user space callchains for  function trace
1040 	 * event, due to issues with page faults while tracing page
1041 	 * fault handler and its overall trickiness nature.
1042 	 */
1043 	if (perf_evsel__is_function_event(evsel))
1044 		evsel->core.attr.exclude_callchain_user = 1;
1045 
1046 	if (callchain && callchain->enabled && !evsel->no_aux_samples)
1047 		perf_evsel__config_callchain(evsel, opts, callchain);
1048 
1049 	if (opts->sample_intr_regs) {
1050 		attr->sample_regs_intr = opts->sample_intr_regs;
1051 		perf_evsel__set_sample_bit(evsel, REGS_INTR);
1052 	}
1053 
1054 	if (opts->sample_user_regs) {
1055 		attr->sample_regs_user |= opts->sample_user_regs;
1056 		perf_evsel__set_sample_bit(evsel, REGS_USER);
1057 	}
1058 
1059 	if (target__has_cpu(&opts->target) || opts->sample_cpu)
1060 		perf_evsel__set_sample_bit(evsel, CPU);
1061 
1062 	/*
1063 	 * When the user explicitly disabled time don't force it here.
1064 	 */
1065 	if (opts->sample_time &&
1066 	    (!perf_missing_features.sample_id_all &&
1067 	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1068 	     opts->sample_time_set)))
1069 		perf_evsel__set_sample_bit(evsel, TIME);
1070 
1071 	if (opts->raw_samples && !evsel->no_aux_samples) {
1072 		perf_evsel__set_sample_bit(evsel, TIME);
1073 		perf_evsel__set_sample_bit(evsel, RAW);
1074 		perf_evsel__set_sample_bit(evsel, CPU);
1075 	}
1076 
1077 	if (opts->sample_address)
1078 		perf_evsel__set_sample_bit(evsel, DATA_SRC);
1079 
1080 	if (opts->sample_phys_addr)
1081 		perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1082 
1083 	if (opts->no_buffering) {
1084 		attr->watermark = 0;
1085 		attr->wakeup_events = 1;
1086 	}
1087 	if (opts->branch_stack && !evsel->no_aux_samples) {
1088 		perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1089 		attr->branch_sample_type = opts->branch_stack;
1090 	}
1091 
1092 	if (opts->sample_weight)
1093 		perf_evsel__set_sample_bit(evsel, WEIGHT);
1094 
1095 	attr->task  = track;
1096 	attr->mmap  = track;
1097 	attr->mmap2 = track && !perf_missing_features.mmap2;
1098 	attr->comm  = track;
1099 	attr->ksymbol = track && !perf_missing_features.ksymbol;
1100 	attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1101 
1102 	if (opts->record_namespaces)
1103 		attr->namespaces  = track;
1104 
1105 	if (opts->record_switch_events)
1106 		attr->context_switch = track;
1107 
1108 	if (opts->sample_transaction)
1109 		perf_evsel__set_sample_bit(evsel, TRANSACTION);
1110 
1111 	if (opts->running_time) {
1112 		evsel->core.attr.read_format |=
1113 			PERF_FORMAT_TOTAL_TIME_ENABLED |
1114 			PERF_FORMAT_TOTAL_TIME_RUNNING;
1115 	}
1116 
1117 	/*
1118 	 * XXX see the function comment above
1119 	 *
1120 	 * Disabling only independent events or group leaders,
1121 	 * keeping group members enabled.
1122 	 */
1123 	if (perf_evsel__is_group_leader(evsel))
1124 		attr->disabled = 1;
1125 
1126 	/*
1127 	 * Setting enable_on_exec for independent events and
1128 	 * group leaders for traced executed by perf.
1129 	 */
1130 	if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1131 		!opts->initial_delay)
1132 		attr->enable_on_exec = 1;
1133 
1134 	if (evsel->immediate) {
1135 		attr->disabled = 0;
1136 		attr->enable_on_exec = 0;
1137 	}
1138 
1139 	clockid = opts->clockid;
1140 	if (opts->use_clockid) {
1141 		attr->use_clockid = 1;
1142 		attr->clockid = opts->clockid;
1143 	}
1144 
1145 	if (evsel->precise_max)
1146 		attr->precise_ip = 3;
1147 
1148 	if (opts->all_user) {
1149 		attr->exclude_kernel = 1;
1150 		attr->exclude_user   = 0;
1151 	}
1152 
1153 	if (opts->all_kernel) {
1154 		attr->exclude_kernel = 0;
1155 		attr->exclude_user   = 1;
1156 	}
1157 
1158 	if (evsel->core.own_cpus || evsel->unit)
1159 		evsel->core.attr.read_format |= PERF_FORMAT_ID;
1160 
1161 	/*
1162 	 * Apply event specific term settings,
1163 	 * it overloads any global configuration.
1164 	 */
1165 	apply_config_terms(evsel, opts, track);
1166 
1167 	evsel->ignore_missing_thread = opts->ignore_missing_thread;
1168 
1169 	/* The --period option takes the precedence. */
1170 	if (opts->period_set) {
1171 		if (opts->period)
1172 			perf_evsel__set_sample_bit(evsel, PERIOD);
1173 		else
1174 			perf_evsel__reset_sample_bit(evsel, PERIOD);
1175 	}
1176 
1177 	/*
1178 	 * For initial_delay, a dummy event is added implicitly.
1179 	 * The software event will trigger -EOPNOTSUPP error out,
1180 	 * if BRANCH_STACK bit is set.
1181 	 */
1182 	if (opts->initial_delay && is_dummy_event(evsel))
1183 		perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1184 }
1185 
1186 int perf_evsel__set_filter(struct evsel *evsel, const char *filter)
1187 {
1188 	char *new_filter = strdup(filter);
1189 
1190 	if (new_filter != NULL) {
1191 		free(evsel->filter);
1192 		evsel->filter = new_filter;
1193 		return 0;
1194 	}
1195 
1196 	return -1;
1197 }
1198 
1199 static int perf_evsel__append_filter(struct evsel *evsel,
1200 				     const char *fmt, const char *filter)
1201 {
1202 	char *new_filter;
1203 
1204 	if (evsel->filter == NULL)
1205 		return perf_evsel__set_filter(evsel, filter);
1206 
1207 	if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1208 		free(evsel->filter);
1209 		evsel->filter = new_filter;
1210 		return 0;
1211 	}
1212 
1213 	return -1;
1214 }
1215 
1216 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1217 {
1218 	return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1219 }
1220 
1221 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1222 {
1223 	return perf_evsel__append_filter(evsel, "%s,%s", filter);
1224 }
1225 
1226 int evsel__enable(struct evsel *evsel)
1227 {
1228 	int err = perf_evsel__enable(&evsel->core);
1229 
1230 	if (!err)
1231 		evsel->disabled = false;
1232 
1233 	return err;
1234 }
1235 
1236 int evsel__disable(struct evsel *evsel)
1237 {
1238 	int err = perf_evsel__disable(&evsel->core);
1239 	/*
1240 	 * We mark it disabled here so that tools that disable a event can
1241 	 * ignore events after they disable it. I.e. the ring buffer may have
1242 	 * already a few more events queued up before the kernel got the stop
1243 	 * request.
1244 	 */
1245 	if (!err)
1246 		evsel->disabled = true;
1247 
1248 	return err;
1249 }
1250 
1251 static void perf_evsel__free_config_terms(struct evsel *evsel)
1252 {
1253 	struct perf_evsel_config_term *term, *h;
1254 
1255 	list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1256 		list_del_init(&term->list);
1257 		free(term);
1258 	}
1259 }
1260 
1261 void perf_evsel__exit(struct evsel *evsel)
1262 {
1263 	assert(list_empty(&evsel->core.node));
1264 	assert(evsel->evlist == NULL);
1265 	perf_evsel__free_counts(evsel);
1266 	perf_evsel__free_fd(&evsel->core);
1267 	perf_evsel__free_id(&evsel->core);
1268 	perf_evsel__free_config_terms(evsel);
1269 	cgroup__put(evsel->cgrp);
1270 	perf_cpu_map__put(evsel->core.cpus);
1271 	perf_cpu_map__put(evsel->core.own_cpus);
1272 	perf_thread_map__put(evsel->core.threads);
1273 	zfree(&evsel->group_name);
1274 	zfree(&evsel->name);
1275 	perf_evsel__object.fini(evsel);
1276 }
1277 
1278 void evsel__delete(struct evsel *evsel)
1279 {
1280 	perf_evsel__exit(evsel);
1281 	free(evsel);
1282 }
1283 
1284 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1285 				struct perf_counts_values *count)
1286 {
1287 	struct perf_counts_values tmp;
1288 
1289 	if (!evsel->prev_raw_counts)
1290 		return;
1291 
1292 	if (cpu == -1) {
1293 		tmp = evsel->prev_raw_counts->aggr;
1294 		evsel->prev_raw_counts->aggr = *count;
1295 	} else {
1296 		tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1297 		*perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1298 	}
1299 
1300 	count->val = count->val - tmp.val;
1301 	count->ena = count->ena - tmp.ena;
1302 	count->run = count->run - tmp.run;
1303 }
1304 
1305 void perf_counts_values__scale(struct perf_counts_values *count,
1306 			       bool scale, s8 *pscaled)
1307 {
1308 	s8 scaled = 0;
1309 
1310 	if (scale) {
1311 		if (count->run == 0) {
1312 			scaled = -1;
1313 			count->val = 0;
1314 		} else if (count->run < count->ena) {
1315 			scaled = 1;
1316 			count->val = (u64)((double) count->val * count->ena / count->run);
1317 		}
1318 	}
1319 
1320 	if (pscaled)
1321 		*pscaled = scaled;
1322 }
1323 
1324 static int
1325 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread)
1326 {
1327 	struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1328 
1329 	return perf_evsel__read(&evsel->core, cpu, thread, count);
1330 }
1331 
1332 static void
1333 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1334 		      u64 val, u64 ena, u64 run)
1335 {
1336 	struct perf_counts_values *count;
1337 
1338 	count = perf_counts(counter->counts, cpu, thread);
1339 
1340 	count->val    = val;
1341 	count->ena    = ena;
1342 	count->run    = run;
1343 
1344 	perf_counts__set_loaded(counter->counts, cpu, thread, true);
1345 }
1346 
1347 static int
1348 perf_evsel__process_group_data(struct evsel *leader,
1349 			       int cpu, int thread, u64 *data)
1350 {
1351 	u64 read_format = leader->core.attr.read_format;
1352 	struct sample_read_value *v;
1353 	u64 nr, ena = 0, run = 0, i;
1354 
1355 	nr = *data++;
1356 
1357 	if (nr != (u64) leader->core.nr_members)
1358 		return -EINVAL;
1359 
1360 	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1361 		ena = *data++;
1362 
1363 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1364 		run = *data++;
1365 
1366 	v = (struct sample_read_value *) data;
1367 
1368 	perf_evsel__set_count(leader, cpu, thread,
1369 			      v[0].value, ena, run);
1370 
1371 	for (i = 1; i < nr; i++) {
1372 		struct evsel *counter;
1373 
1374 		counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1375 		if (!counter)
1376 			return -EINVAL;
1377 
1378 		perf_evsel__set_count(counter, cpu, thread,
1379 				      v[i].value, ena, run);
1380 	}
1381 
1382 	return 0;
1383 }
1384 
1385 static int
1386 perf_evsel__read_group(struct evsel *leader, int cpu, int thread)
1387 {
1388 	struct perf_stat_evsel *ps = leader->stats;
1389 	u64 read_format = leader->core.attr.read_format;
1390 	int size = perf_evsel__read_size(&leader->core);
1391 	u64 *data = ps->group_data;
1392 
1393 	if (!(read_format & PERF_FORMAT_ID))
1394 		return -EINVAL;
1395 
1396 	if (!perf_evsel__is_group_leader(leader))
1397 		return -EINVAL;
1398 
1399 	if (!data) {
1400 		data = zalloc(size);
1401 		if (!data)
1402 			return -ENOMEM;
1403 
1404 		ps->group_data = data;
1405 	}
1406 
1407 	if (FD(leader, cpu, thread) < 0)
1408 		return -EINVAL;
1409 
1410 	if (readn(FD(leader, cpu, thread), data, size) <= 0)
1411 		return -errno;
1412 
1413 	return perf_evsel__process_group_data(leader, cpu, thread, data);
1414 }
1415 
1416 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1417 {
1418 	u64 read_format = evsel->core.attr.read_format;
1419 
1420 	if (read_format & PERF_FORMAT_GROUP)
1421 		return perf_evsel__read_group(evsel, cpu, thread);
1422 	else
1423 		return perf_evsel__read_one(evsel, cpu, thread);
1424 }
1425 
1426 int __perf_evsel__read_on_cpu(struct evsel *evsel,
1427 			      int cpu, int thread, bool scale)
1428 {
1429 	struct perf_counts_values count;
1430 	size_t nv = scale ? 3 : 1;
1431 
1432 	if (FD(evsel, cpu, thread) < 0)
1433 		return -EINVAL;
1434 
1435 	if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1436 		return -ENOMEM;
1437 
1438 	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1439 		return -errno;
1440 
1441 	perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1442 	perf_counts_values__scale(&count, scale, NULL);
1443 	*perf_counts(evsel->counts, cpu, thread) = count;
1444 	return 0;
1445 }
1446 
1447 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1448 {
1449 	struct evsel *leader = evsel->leader;
1450 	int fd;
1451 
1452 	if (perf_evsel__is_group_leader(evsel))
1453 		return -1;
1454 
1455 	/*
1456 	 * Leader must be already processed/open,
1457 	 * if not it's a bug.
1458 	 */
1459 	BUG_ON(!leader->core.fd);
1460 
1461 	fd = FD(leader, cpu, thread);
1462 	BUG_ON(fd == -1);
1463 
1464 	return fd;
1465 }
1466 
1467 static void perf_evsel__remove_fd(struct evsel *pos,
1468 				  int nr_cpus, int nr_threads,
1469 				  int thread_idx)
1470 {
1471 	for (int cpu = 0; cpu < nr_cpus; cpu++)
1472 		for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1473 			FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1474 }
1475 
1476 static int update_fds(struct evsel *evsel,
1477 		      int nr_cpus, int cpu_idx,
1478 		      int nr_threads, int thread_idx)
1479 {
1480 	struct evsel *pos;
1481 
1482 	if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1483 		return -EINVAL;
1484 
1485 	evlist__for_each_entry(evsel->evlist, pos) {
1486 		nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1487 
1488 		perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1489 
1490 		/*
1491 		 * Since fds for next evsel has not been created,
1492 		 * there is no need to iterate whole event list.
1493 		 */
1494 		if (pos == evsel)
1495 			break;
1496 	}
1497 	return 0;
1498 }
1499 
1500 static bool ignore_missing_thread(struct evsel *evsel,
1501 				  int nr_cpus, int cpu,
1502 				  struct perf_thread_map *threads,
1503 				  int thread, int err)
1504 {
1505 	pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1506 
1507 	if (!evsel->ignore_missing_thread)
1508 		return false;
1509 
1510 	/* The system wide setup does not work with threads. */
1511 	if (evsel->core.system_wide)
1512 		return false;
1513 
1514 	/* The -ESRCH is perf event syscall errno for pid's not found. */
1515 	if (err != -ESRCH)
1516 		return false;
1517 
1518 	/* If there's only one thread, let it fail. */
1519 	if (threads->nr == 1)
1520 		return false;
1521 
1522 	/*
1523 	 * We should remove fd for missing_thread first
1524 	 * because thread_map__remove() will decrease threads->nr.
1525 	 */
1526 	if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1527 		return false;
1528 
1529 	if (thread_map__remove(threads, thread))
1530 		return false;
1531 
1532 	pr_warning("WARNING: Ignored open failure for pid %d\n",
1533 		   ignore_pid);
1534 	return true;
1535 }
1536 
1537 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1538 				void *priv __maybe_unused)
1539 {
1540 	return fprintf(fp, "  %-32s %s\n", name, val);
1541 }
1542 
1543 static void display_attr(struct perf_event_attr *attr)
1544 {
1545 	if (verbose >= 2 || debug_peo_args) {
1546 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1547 		fprintf(stderr, "perf_event_attr:\n");
1548 		perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1549 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1550 	}
1551 }
1552 
1553 static int perf_event_open(struct evsel *evsel,
1554 			   pid_t pid, int cpu, int group_fd,
1555 			   unsigned long flags)
1556 {
1557 	int precise_ip = evsel->core.attr.precise_ip;
1558 	int fd;
1559 
1560 	while (1) {
1561 		pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1562 			  pid, cpu, group_fd, flags);
1563 
1564 		fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1565 		if (fd >= 0)
1566 			break;
1567 
1568 		/* Do not try less precise if not requested. */
1569 		if (!evsel->precise_max)
1570 			break;
1571 
1572 		/*
1573 		 * We tried all the precise_ip values, and it's
1574 		 * still failing, so leave it to standard fallback.
1575 		 */
1576 		if (!evsel->core.attr.precise_ip) {
1577 			evsel->core.attr.precise_ip = precise_ip;
1578 			break;
1579 		}
1580 
1581 		pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1582 		evsel->core.attr.precise_ip--;
1583 		pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1584 		display_attr(&evsel->core.attr);
1585 	}
1586 
1587 	return fd;
1588 }
1589 
1590 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1591 		struct perf_thread_map *threads)
1592 {
1593 	int cpu, thread, nthreads;
1594 	unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1595 	int pid = -1, err, old_errno;
1596 	enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1597 
1598 	if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1599 	    (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1600 		return -EINVAL;
1601 
1602 	if (cpus == NULL) {
1603 		static struct perf_cpu_map *empty_cpu_map;
1604 
1605 		if (empty_cpu_map == NULL) {
1606 			empty_cpu_map = perf_cpu_map__dummy_new();
1607 			if (empty_cpu_map == NULL)
1608 				return -ENOMEM;
1609 		}
1610 
1611 		cpus = empty_cpu_map;
1612 	}
1613 
1614 	if (threads == NULL) {
1615 		static struct perf_thread_map *empty_thread_map;
1616 
1617 		if (empty_thread_map == NULL) {
1618 			empty_thread_map = thread_map__new_by_tid(-1);
1619 			if (empty_thread_map == NULL)
1620 				return -ENOMEM;
1621 		}
1622 
1623 		threads = empty_thread_map;
1624 	}
1625 
1626 	if (evsel->core.system_wide)
1627 		nthreads = 1;
1628 	else
1629 		nthreads = threads->nr;
1630 
1631 	if (evsel->core.fd == NULL &&
1632 	    perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1633 		return -ENOMEM;
1634 
1635 	if (evsel->cgrp) {
1636 		flags |= PERF_FLAG_PID_CGROUP;
1637 		pid = evsel->cgrp->fd;
1638 	}
1639 
1640 fallback_missing_features:
1641 	if (perf_missing_features.clockid_wrong)
1642 		evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1643 	if (perf_missing_features.clockid) {
1644 		evsel->core.attr.use_clockid = 0;
1645 		evsel->core.attr.clockid = 0;
1646 	}
1647 	if (perf_missing_features.cloexec)
1648 		flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1649 	if (perf_missing_features.mmap2)
1650 		evsel->core.attr.mmap2 = 0;
1651 	if (perf_missing_features.exclude_guest)
1652 		evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1653 	if (perf_missing_features.lbr_flags)
1654 		evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1655 				     PERF_SAMPLE_BRANCH_NO_CYCLES);
1656 	if (perf_missing_features.group_read && evsel->core.attr.inherit)
1657 		evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1658 	if (perf_missing_features.ksymbol)
1659 		evsel->core.attr.ksymbol = 0;
1660 	if (perf_missing_features.bpf)
1661 		evsel->core.attr.bpf_event = 0;
1662 retry_sample_id:
1663 	if (perf_missing_features.sample_id_all)
1664 		evsel->core.attr.sample_id_all = 0;
1665 
1666 	display_attr(&evsel->core.attr);
1667 
1668 	for (cpu = 0; cpu < cpus->nr; cpu++) {
1669 
1670 		for (thread = 0; thread < nthreads; thread++) {
1671 			int fd, group_fd;
1672 
1673 			if (!evsel->cgrp && !evsel->core.system_wide)
1674 				pid = perf_thread_map__pid(threads, thread);
1675 
1676 			group_fd = get_group_fd(evsel, cpu, thread);
1677 retry_open:
1678 			test_attr__ready();
1679 
1680 			fd = perf_event_open(evsel, pid, cpus->map[cpu],
1681 					     group_fd, flags);
1682 
1683 			FD(evsel, cpu, thread) = fd;
1684 
1685 			if (fd < 0) {
1686 				err = -errno;
1687 
1688 				if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1689 					/*
1690 					 * We just removed 1 thread, so take a step
1691 					 * back on thread index and lower the upper
1692 					 * nthreads limit.
1693 					 */
1694 					nthreads--;
1695 					thread--;
1696 
1697 					/* ... and pretend like nothing have happened. */
1698 					err = 0;
1699 					continue;
1700 				}
1701 
1702 				pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1703 					  err);
1704 				goto try_fallback;
1705 			}
1706 
1707 			pr_debug2_peo(" = %d\n", fd);
1708 
1709 			if (evsel->bpf_fd >= 0) {
1710 				int evt_fd = fd;
1711 				int bpf_fd = evsel->bpf_fd;
1712 
1713 				err = ioctl(evt_fd,
1714 					    PERF_EVENT_IOC_SET_BPF,
1715 					    bpf_fd);
1716 				if (err && errno != EEXIST) {
1717 					pr_err("failed to attach bpf fd %d: %s\n",
1718 					       bpf_fd, strerror(errno));
1719 					err = -EINVAL;
1720 					goto out_close;
1721 				}
1722 			}
1723 
1724 			set_rlimit = NO_CHANGE;
1725 
1726 			/*
1727 			 * If we succeeded but had to kill clockid, fail and
1728 			 * have perf_evsel__open_strerror() print us a nice
1729 			 * error.
1730 			 */
1731 			if (perf_missing_features.clockid ||
1732 			    perf_missing_features.clockid_wrong) {
1733 				err = -EINVAL;
1734 				goto out_close;
1735 			}
1736 		}
1737 	}
1738 
1739 	return 0;
1740 
1741 try_fallback:
1742 	/*
1743 	 * perf stat needs between 5 and 22 fds per CPU. When we run out
1744 	 * of them try to increase the limits.
1745 	 */
1746 	if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1747 		struct rlimit l;
1748 
1749 		old_errno = errno;
1750 		if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1751 			if (set_rlimit == NO_CHANGE)
1752 				l.rlim_cur = l.rlim_max;
1753 			else {
1754 				l.rlim_cur = l.rlim_max + 1000;
1755 				l.rlim_max = l.rlim_cur;
1756 			}
1757 			if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1758 				set_rlimit++;
1759 				errno = old_errno;
1760 				goto retry_open;
1761 			}
1762 		}
1763 		errno = old_errno;
1764 	}
1765 
1766 	if (err != -EINVAL || cpu > 0 || thread > 0)
1767 		goto out_close;
1768 
1769 	/*
1770 	 * Must probe features in the order they were added to the
1771 	 * perf_event_attr interface.
1772 	 */
1773 	if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1774 		perf_missing_features.aux_output = true;
1775 		pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1776 		goto out_close;
1777 	} else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1778 		perf_missing_features.bpf = true;
1779 		pr_debug2_peo("switching off bpf_event\n");
1780 		goto fallback_missing_features;
1781 	} else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1782 		perf_missing_features.ksymbol = true;
1783 		pr_debug2_peo("switching off ksymbol\n");
1784 		goto fallback_missing_features;
1785 	} else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1786 		perf_missing_features.write_backward = true;
1787 		pr_debug2_peo("switching off write_backward\n");
1788 		goto out_close;
1789 	} else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1790 		perf_missing_features.clockid_wrong = true;
1791 		pr_debug2_peo("switching off clockid\n");
1792 		goto fallback_missing_features;
1793 	} else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1794 		perf_missing_features.clockid = true;
1795 		pr_debug2_peo("switching off use_clockid\n");
1796 		goto fallback_missing_features;
1797 	} else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1798 		perf_missing_features.cloexec = true;
1799 		pr_debug2_peo("switching off cloexec flag\n");
1800 		goto fallback_missing_features;
1801 	} else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1802 		perf_missing_features.mmap2 = true;
1803 		pr_debug2_peo("switching off mmap2\n");
1804 		goto fallback_missing_features;
1805 	} else if (!perf_missing_features.exclude_guest &&
1806 		   (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1807 		perf_missing_features.exclude_guest = true;
1808 		pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1809 		goto fallback_missing_features;
1810 	} else if (!perf_missing_features.sample_id_all) {
1811 		perf_missing_features.sample_id_all = true;
1812 		pr_debug2_peo("switching off sample_id_all\n");
1813 		goto retry_sample_id;
1814 	} else if (!perf_missing_features.lbr_flags &&
1815 			(evsel->core.attr.branch_sample_type &
1816 			 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1817 			  PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1818 		perf_missing_features.lbr_flags = true;
1819 		pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1820 		goto fallback_missing_features;
1821 	} else if (!perf_missing_features.group_read &&
1822 		    evsel->core.attr.inherit &&
1823 		   (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1824 		   perf_evsel__is_group_leader(evsel)) {
1825 		perf_missing_features.group_read = true;
1826 		pr_debug2_peo("switching off group read\n");
1827 		goto fallback_missing_features;
1828 	}
1829 out_close:
1830 	if (err)
1831 		threads->err_thread = thread;
1832 
1833 	old_errno = errno;
1834 	do {
1835 		while (--thread >= 0) {
1836 			if (FD(evsel, cpu, thread) >= 0)
1837 				close(FD(evsel, cpu, thread));
1838 			FD(evsel, cpu, thread) = -1;
1839 		}
1840 		thread = nthreads;
1841 	} while (--cpu >= 0);
1842 	errno = old_errno;
1843 	return err;
1844 }
1845 
1846 void evsel__close(struct evsel *evsel)
1847 {
1848 	perf_evsel__close(&evsel->core);
1849 	perf_evsel__free_id(&evsel->core);
1850 }
1851 
1852 int perf_evsel__open_per_cpu(struct evsel *evsel,
1853 			     struct perf_cpu_map *cpus)
1854 {
1855 	return evsel__open(evsel, cpus, NULL);
1856 }
1857 
1858 int perf_evsel__open_per_thread(struct evsel *evsel,
1859 				struct perf_thread_map *threads)
1860 {
1861 	return evsel__open(evsel, NULL, threads);
1862 }
1863 
1864 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1865 				       const union perf_event *event,
1866 				       struct perf_sample *sample)
1867 {
1868 	u64 type = evsel->core.attr.sample_type;
1869 	const __u64 *array = event->sample.array;
1870 	bool swapped = evsel->needs_swap;
1871 	union u64_swap u;
1872 
1873 	array += ((event->header.size -
1874 		   sizeof(event->header)) / sizeof(u64)) - 1;
1875 
1876 	if (type & PERF_SAMPLE_IDENTIFIER) {
1877 		sample->id = *array;
1878 		array--;
1879 	}
1880 
1881 	if (type & PERF_SAMPLE_CPU) {
1882 		u.val64 = *array;
1883 		if (swapped) {
1884 			/* undo swap of u64, then swap on individual u32s */
1885 			u.val64 = bswap_64(u.val64);
1886 			u.val32[0] = bswap_32(u.val32[0]);
1887 		}
1888 
1889 		sample->cpu = u.val32[0];
1890 		array--;
1891 	}
1892 
1893 	if (type & PERF_SAMPLE_STREAM_ID) {
1894 		sample->stream_id = *array;
1895 		array--;
1896 	}
1897 
1898 	if (type & PERF_SAMPLE_ID) {
1899 		sample->id = *array;
1900 		array--;
1901 	}
1902 
1903 	if (type & PERF_SAMPLE_TIME) {
1904 		sample->time = *array;
1905 		array--;
1906 	}
1907 
1908 	if (type & PERF_SAMPLE_TID) {
1909 		u.val64 = *array;
1910 		if (swapped) {
1911 			/* undo swap of u64, then swap on individual u32s */
1912 			u.val64 = bswap_64(u.val64);
1913 			u.val32[0] = bswap_32(u.val32[0]);
1914 			u.val32[1] = bswap_32(u.val32[1]);
1915 		}
1916 
1917 		sample->pid = u.val32[0];
1918 		sample->tid = u.val32[1];
1919 		array--;
1920 	}
1921 
1922 	return 0;
1923 }
1924 
1925 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1926 			    u64 size)
1927 {
1928 	return size > max_size || offset + size > endp;
1929 }
1930 
1931 #define OVERFLOW_CHECK(offset, size, max_size)				\
1932 	do {								\
1933 		if (overflow(endp, (max_size), (offset), (size)))	\
1934 			return -EFAULT;					\
1935 	} while (0)
1936 
1937 #define OVERFLOW_CHECK_u64(offset) \
1938 	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1939 
1940 static int
1941 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1942 {
1943 	/*
1944 	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1945 	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1946 	 * check the format does not go past the end of the event.
1947 	 */
1948 	if (sample_size + sizeof(event->header) > event->header.size)
1949 		return -EFAULT;
1950 
1951 	return 0;
1952 }
1953 
1954 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1955 			     struct perf_sample *data)
1956 {
1957 	u64 type = evsel->core.attr.sample_type;
1958 	bool swapped = evsel->needs_swap;
1959 	const __u64 *array;
1960 	u16 max_size = event->header.size;
1961 	const void *endp = (void *)event + max_size;
1962 	u64 sz;
1963 
1964 	/*
1965 	 * used for cross-endian analysis. See git commit 65014ab3
1966 	 * for why this goofiness is needed.
1967 	 */
1968 	union u64_swap u;
1969 
1970 	memset(data, 0, sizeof(*data));
1971 	data->cpu = data->pid = data->tid = -1;
1972 	data->stream_id = data->id = data->time = -1ULL;
1973 	data->period = evsel->core.attr.sample_period;
1974 	data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1975 	data->misc    = event->header.misc;
1976 	data->id = -1ULL;
1977 	data->data_src = PERF_MEM_DATA_SRC_NONE;
1978 
1979 	if (event->header.type != PERF_RECORD_SAMPLE) {
1980 		if (!evsel->core.attr.sample_id_all)
1981 			return 0;
1982 		return perf_evsel__parse_id_sample(evsel, event, data);
1983 	}
1984 
1985 	array = event->sample.array;
1986 
1987 	if (perf_event__check_size(event, evsel->sample_size))
1988 		return -EFAULT;
1989 
1990 	if (type & PERF_SAMPLE_IDENTIFIER) {
1991 		data->id = *array;
1992 		array++;
1993 	}
1994 
1995 	if (type & PERF_SAMPLE_IP) {
1996 		data->ip = *array;
1997 		array++;
1998 	}
1999 
2000 	if (type & PERF_SAMPLE_TID) {
2001 		u.val64 = *array;
2002 		if (swapped) {
2003 			/* undo swap of u64, then swap on individual u32s */
2004 			u.val64 = bswap_64(u.val64);
2005 			u.val32[0] = bswap_32(u.val32[0]);
2006 			u.val32[1] = bswap_32(u.val32[1]);
2007 		}
2008 
2009 		data->pid = u.val32[0];
2010 		data->tid = u.val32[1];
2011 		array++;
2012 	}
2013 
2014 	if (type & PERF_SAMPLE_TIME) {
2015 		data->time = *array;
2016 		array++;
2017 	}
2018 
2019 	if (type & PERF_SAMPLE_ADDR) {
2020 		data->addr = *array;
2021 		array++;
2022 	}
2023 
2024 	if (type & PERF_SAMPLE_ID) {
2025 		data->id = *array;
2026 		array++;
2027 	}
2028 
2029 	if (type & PERF_SAMPLE_STREAM_ID) {
2030 		data->stream_id = *array;
2031 		array++;
2032 	}
2033 
2034 	if (type & PERF_SAMPLE_CPU) {
2035 
2036 		u.val64 = *array;
2037 		if (swapped) {
2038 			/* undo swap of u64, then swap on individual u32s */
2039 			u.val64 = bswap_64(u.val64);
2040 			u.val32[0] = bswap_32(u.val32[0]);
2041 		}
2042 
2043 		data->cpu = u.val32[0];
2044 		array++;
2045 	}
2046 
2047 	if (type & PERF_SAMPLE_PERIOD) {
2048 		data->period = *array;
2049 		array++;
2050 	}
2051 
2052 	if (type & PERF_SAMPLE_READ) {
2053 		u64 read_format = evsel->core.attr.read_format;
2054 
2055 		OVERFLOW_CHECK_u64(array);
2056 		if (read_format & PERF_FORMAT_GROUP)
2057 			data->read.group.nr = *array;
2058 		else
2059 			data->read.one.value = *array;
2060 
2061 		array++;
2062 
2063 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2064 			OVERFLOW_CHECK_u64(array);
2065 			data->read.time_enabled = *array;
2066 			array++;
2067 		}
2068 
2069 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2070 			OVERFLOW_CHECK_u64(array);
2071 			data->read.time_running = *array;
2072 			array++;
2073 		}
2074 
2075 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2076 		if (read_format & PERF_FORMAT_GROUP) {
2077 			const u64 max_group_nr = UINT64_MAX /
2078 					sizeof(struct sample_read_value);
2079 
2080 			if (data->read.group.nr > max_group_nr)
2081 				return -EFAULT;
2082 			sz = data->read.group.nr *
2083 			     sizeof(struct sample_read_value);
2084 			OVERFLOW_CHECK(array, sz, max_size);
2085 			data->read.group.values =
2086 					(struct sample_read_value *)array;
2087 			array = (void *)array + sz;
2088 		} else {
2089 			OVERFLOW_CHECK_u64(array);
2090 			data->read.one.id = *array;
2091 			array++;
2092 		}
2093 	}
2094 
2095 	if (evsel__has_callchain(evsel)) {
2096 		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2097 
2098 		OVERFLOW_CHECK_u64(array);
2099 		data->callchain = (struct ip_callchain *)array++;
2100 		if (data->callchain->nr > max_callchain_nr)
2101 			return -EFAULT;
2102 		sz = data->callchain->nr * sizeof(u64);
2103 		OVERFLOW_CHECK(array, sz, max_size);
2104 		array = (void *)array + sz;
2105 	}
2106 
2107 	if (type & PERF_SAMPLE_RAW) {
2108 		OVERFLOW_CHECK_u64(array);
2109 		u.val64 = *array;
2110 
2111 		/*
2112 		 * Undo swap of u64, then swap on individual u32s,
2113 		 * get the size of the raw area and undo all of the
2114 		 * swap. The pevent interface handles endianity by
2115 		 * itself.
2116 		 */
2117 		if (swapped) {
2118 			u.val64 = bswap_64(u.val64);
2119 			u.val32[0] = bswap_32(u.val32[0]);
2120 			u.val32[1] = bswap_32(u.val32[1]);
2121 		}
2122 		data->raw_size = u.val32[0];
2123 
2124 		/*
2125 		 * The raw data is aligned on 64bits including the
2126 		 * u32 size, so it's safe to use mem_bswap_64.
2127 		 */
2128 		if (swapped)
2129 			mem_bswap_64((void *) array, data->raw_size);
2130 
2131 		array = (void *)array + sizeof(u32);
2132 
2133 		OVERFLOW_CHECK(array, data->raw_size, max_size);
2134 		data->raw_data = (void *)array;
2135 		array = (void *)array + data->raw_size;
2136 	}
2137 
2138 	if (type & PERF_SAMPLE_BRANCH_STACK) {
2139 		const u64 max_branch_nr = UINT64_MAX /
2140 					  sizeof(struct branch_entry);
2141 
2142 		OVERFLOW_CHECK_u64(array);
2143 		data->branch_stack = (struct branch_stack *)array++;
2144 
2145 		if (data->branch_stack->nr > max_branch_nr)
2146 			return -EFAULT;
2147 		sz = data->branch_stack->nr * sizeof(struct branch_entry);
2148 		OVERFLOW_CHECK(array, sz, max_size);
2149 		array = (void *)array + sz;
2150 	}
2151 
2152 	if (type & PERF_SAMPLE_REGS_USER) {
2153 		OVERFLOW_CHECK_u64(array);
2154 		data->user_regs.abi = *array;
2155 		array++;
2156 
2157 		if (data->user_regs.abi) {
2158 			u64 mask = evsel->core.attr.sample_regs_user;
2159 
2160 			sz = hweight64(mask) * sizeof(u64);
2161 			OVERFLOW_CHECK(array, sz, max_size);
2162 			data->user_regs.mask = mask;
2163 			data->user_regs.regs = (u64 *)array;
2164 			array = (void *)array + sz;
2165 		}
2166 	}
2167 
2168 	if (type & PERF_SAMPLE_STACK_USER) {
2169 		OVERFLOW_CHECK_u64(array);
2170 		sz = *array++;
2171 
2172 		data->user_stack.offset = ((char *)(array - 1)
2173 					  - (char *) event);
2174 
2175 		if (!sz) {
2176 			data->user_stack.size = 0;
2177 		} else {
2178 			OVERFLOW_CHECK(array, sz, max_size);
2179 			data->user_stack.data = (char *)array;
2180 			array = (void *)array + sz;
2181 			OVERFLOW_CHECK_u64(array);
2182 			data->user_stack.size = *array++;
2183 			if (WARN_ONCE(data->user_stack.size > sz,
2184 				      "user stack dump failure\n"))
2185 				return -EFAULT;
2186 		}
2187 	}
2188 
2189 	if (type & PERF_SAMPLE_WEIGHT) {
2190 		OVERFLOW_CHECK_u64(array);
2191 		data->weight = *array;
2192 		array++;
2193 	}
2194 
2195 	if (type & PERF_SAMPLE_DATA_SRC) {
2196 		OVERFLOW_CHECK_u64(array);
2197 		data->data_src = *array;
2198 		array++;
2199 	}
2200 
2201 	if (type & PERF_SAMPLE_TRANSACTION) {
2202 		OVERFLOW_CHECK_u64(array);
2203 		data->transaction = *array;
2204 		array++;
2205 	}
2206 
2207 	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2208 	if (type & PERF_SAMPLE_REGS_INTR) {
2209 		OVERFLOW_CHECK_u64(array);
2210 		data->intr_regs.abi = *array;
2211 		array++;
2212 
2213 		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2214 			u64 mask = evsel->core.attr.sample_regs_intr;
2215 
2216 			sz = hweight64(mask) * sizeof(u64);
2217 			OVERFLOW_CHECK(array, sz, max_size);
2218 			data->intr_regs.mask = mask;
2219 			data->intr_regs.regs = (u64 *)array;
2220 			array = (void *)array + sz;
2221 		}
2222 	}
2223 
2224 	data->phys_addr = 0;
2225 	if (type & PERF_SAMPLE_PHYS_ADDR) {
2226 		data->phys_addr = *array;
2227 		array++;
2228 	}
2229 
2230 	if (type & PERF_SAMPLE_AUX) {
2231 		OVERFLOW_CHECK_u64(array);
2232 		sz = *array++;
2233 
2234 		OVERFLOW_CHECK(array, sz, max_size);
2235 		/* Undo swap of data */
2236 		if (swapped)
2237 			mem_bswap_64((char *)array, sz);
2238 		data->aux_sample.size = sz;
2239 		data->aux_sample.data = (char *)array;
2240 		array = (void *)array + sz;
2241 	}
2242 
2243 	return 0;
2244 }
2245 
2246 int perf_evsel__parse_sample_timestamp(struct evsel *evsel,
2247 				       union perf_event *event,
2248 				       u64 *timestamp)
2249 {
2250 	u64 type = evsel->core.attr.sample_type;
2251 	const __u64 *array;
2252 
2253 	if (!(type & PERF_SAMPLE_TIME))
2254 		return -1;
2255 
2256 	if (event->header.type != PERF_RECORD_SAMPLE) {
2257 		struct perf_sample data = {
2258 			.time = -1ULL,
2259 		};
2260 
2261 		if (!evsel->core.attr.sample_id_all)
2262 			return -1;
2263 		if (perf_evsel__parse_id_sample(evsel, event, &data))
2264 			return -1;
2265 
2266 		*timestamp = data.time;
2267 		return 0;
2268 	}
2269 
2270 	array = event->sample.array;
2271 
2272 	if (perf_event__check_size(event, evsel->sample_size))
2273 		return -EFAULT;
2274 
2275 	if (type & PERF_SAMPLE_IDENTIFIER)
2276 		array++;
2277 
2278 	if (type & PERF_SAMPLE_IP)
2279 		array++;
2280 
2281 	if (type & PERF_SAMPLE_TID)
2282 		array++;
2283 
2284 	if (type & PERF_SAMPLE_TIME)
2285 		*timestamp = *array;
2286 
2287 	return 0;
2288 }
2289 
2290 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name)
2291 {
2292 	return tep_find_field(evsel->tp_format, name);
2293 }
2294 
2295 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample,
2296 			 const char *name)
2297 {
2298 	struct tep_format_field *field = perf_evsel__field(evsel, name);
2299 	int offset;
2300 
2301 	if (!field)
2302 		return NULL;
2303 
2304 	offset = field->offset;
2305 
2306 	if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2307 		offset = *(int *)(sample->raw_data + field->offset);
2308 		offset &= 0xffff;
2309 	}
2310 
2311 	return sample->raw_data + offset;
2312 }
2313 
2314 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2315 			 bool needs_swap)
2316 {
2317 	u64 value;
2318 	void *ptr = sample->raw_data + field->offset;
2319 
2320 	switch (field->size) {
2321 	case 1:
2322 		return *(u8 *)ptr;
2323 	case 2:
2324 		value = *(u16 *)ptr;
2325 		break;
2326 	case 4:
2327 		value = *(u32 *)ptr;
2328 		break;
2329 	case 8:
2330 		memcpy(&value, ptr, sizeof(u64));
2331 		break;
2332 	default:
2333 		return 0;
2334 	}
2335 
2336 	if (!needs_swap)
2337 		return value;
2338 
2339 	switch (field->size) {
2340 	case 2:
2341 		return bswap_16(value);
2342 	case 4:
2343 		return bswap_32(value);
2344 	case 8:
2345 		return bswap_64(value);
2346 	default:
2347 		return 0;
2348 	}
2349 
2350 	return 0;
2351 }
2352 
2353 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample,
2354 		       const char *name)
2355 {
2356 	struct tep_format_field *field = perf_evsel__field(evsel, name);
2357 
2358 	if (!field)
2359 		return 0;
2360 
2361 	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2362 }
2363 
2364 bool perf_evsel__fallback(struct evsel *evsel, int err,
2365 			  char *msg, size_t msgsize)
2366 {
2367 	int paranoid;
2368 
2369 	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2370 	    evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2371 	    evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2372 		/*
2373 		 * If it's cycles then fall back to hrtimer based
2374 		 * cpu-clock-tick sw counter, which is always available even if
2375 		 * no PMU support.
2376 		 *
2377 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2378 		 * b0a873e).
2379 		 */
2380 		scnprintf(msg, msgsize, "%s",
2381 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2382 
2383 		evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2384 		evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2385 
2386 		zfree(&evsel->name);
2387 		return true;
2388 	} else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2389 		   (paranoid = perf_event_paranoid()) > 1) {
2390 		const char *name = perf_evsel__name(evsel);
2391 		char *new_name;
2392 		const char *sep = ":";
2393 
2394 		/* Is there already the separator in the name. */
2395 		if (strchr(name, '/') ||
2396 		    strchr(name, ':'))
2397 			sep = "";
2398 
2399 		if (asprintf(&new_name, "%s%su", name, sep) < 0)
2400 			return false;
2401 
2402 		if (evsel->name)
2403 			free(evsel->name);
2404 		evsel->name = new_name;
2405 		scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2406 			  "to fall back to excluding kernel and hypervisor "
2407 			  " samples", paranoid);
2408 		evsel->core.attr.exclude_kernel = 1;
2409 		evsel->core.attr.exclude_hv     = 1;
2410 
2411 		return true;
2412 	}
2413 
2414 	return false;
2415 }
2416 
2417 static bool find_process(const char *name)
2418 {
2419 	size_t len = strlen(name);
2420 	DIR *dir;
2421 	struct dirent *d;
2422 	int ret = -1;
2423 
2424 	dir = opendir(procfs__mountpoint());
2425 	if (!dir)
2426 		return false;
2427 
2428 	/* Walk through the directory. */
2429 	while (ret && (d = readdir(dir)) != NULL) {
2430 		char path[PATH_MAX];
2431 		char *data;
2432 		size_t size;
2433 
2434 		if ((d->d_type != DT_DIR) ||
2435 		     !strcmp(".", d->d_name) ||
2436 		     !strcmp("..", d->d_name))
2437 			continue;
2438 
2439 		scnprintf(path, sizeof(path), "%s/%s/comm",
2440 			  procfs__mountpoint(), d->d_name);
2441 
2442 		if (filename__read_str(path, &data, &size))
2443 			continue;
2444 
2445 		ret = strncmp(name, data, len);
2446 		free(data);
2447 	}
2448 
2449 	closedir(dir);
2450 	return ret ? false : true;
2451 }
2452 
2453 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target,
2454 			      int err, char *msg, size_t size)
2455 {
2456 	char sbuf[STRERR_BUFSIZE];
2457 	int printed = 0;
2458 
2459 	switch (err) {
2460 	case EPERM:
2461 	case EACCES:
2462 		if (err == EPERM)
2463 			printed = scnprintf(msg, size,
2464 				"No permission to enable %s event.\n\n",
2465 				perf_evsel__name(evsel));
2466 
2467 		return scnprintf(msg + printed, size - printed,
2468 		 "You may not have permission to collect %sstats.\n\n"
2469 		 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2470 		 "which controls use of the performance events system by\n"
2471 		 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2472 		 "The current value is %d:\n\n"
2473 		 "  -1: Allow use of (almost) all events by all users\n"
2474 		 "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2475 		 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2476 		 "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2477 		 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2478 		 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2479 		 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2480 		 "	kernel.perf_event_paranoid = -1\n" ,
2481 				 target->system_wide ? "system-wide " : "",
2482 				 perf_event_paranoid());
2483 	case ENOENT:
2484 		return scnprintf(msg, size, "The %s event is not supported.",
2485 				 perf_evsel__name(evsel));
2486 	case EMFILE:
2487 		return scnprintf(msg, size, "%s",
2488 			 "Too many events are opened.\n"
2489 			 "Probably the maximum number of open file descriptors has been reached.\n"
2490 			 "Hint: Try again after reducing the number of events.\n"
2491 			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2492 	case ENOMEM:
2493 		if (evsel__has_callchain(evsel) &&
2494 		    access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2495 			return scnprintf(msg, size,
2496 					 "Not enough memory to setup event with callchain.\n"
2497 					 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2498 					 "Hint: Current value: %d", sysctl__max_stack());
2499 		break;
2500 	case ENODEV:
2501 		if (target->cpu_list)
2502 			return scnprintf(msg, size, "%s",
2503 	 "No such device - did you specify an out-of-range profile CPU?");
2504 		break;
2505 	case EOPNOTSUPP:
2506 		if (evsel->core.attr.sample_period != 0)
2507 			return scnprintf(msg, size,
2508 	"%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2509 					 perf_evsel__name(evsel));
2510 		if (evsel->core.attr.precise_ip)
2511 			return scnprintf(msg, size, "%s",
2512 	"\'precise\' request may not be supported. Try removing 'p' modifier.");
2513 #if defined(__i386__) || defined(__x86_64__)
2514 		if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2515 			return scnprintf(msg, size, "%s",
2516 	"No hardware sampling interrupt available.\n");
2517 #endif
2518 		break;
2519 	case EBUSY:
2520 		if (find_process("oprofiled"))
2521 			return scnprintf(msg, size,
2522 	"The PMU counters are busy/taken by another profiler.\n"
2523 	"We found oprofile daemon running, please stop it and try again.");
2524 		break;
2525 	case EINVAL:
2526 		if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2527 			return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2528 		if (perf_missing_features.clockid)
2529 			return scnprintf(msg, size, "clockid feature not supported.");
2530 		if (perf_missing_features.clockid_wrong)
2531 			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2532 		if (perf_missing_features.aux_output)
2533 			return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2534 		break;
2535 	default:
2536 		break;
2537 	}
2538 
2539 	return scnprintf(msg, size,
2540 	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2541 	"/bin/dmesg | grep -i perf may provide additional information.\n",
2542 			 err, str_error_r(err, sbuf, sizeof(sbuf)),
2543 			 perf_evsel__name(evsel));
2544 }
2545 
2546 struct perf_env *perf_evsel__env(struct evsel *evsel)
2547 {
2548 	if (evsel && evsel->evlist)
2549 		return evsel->evlist->env;
2550 	return &perf_env;
2551 }
2552 
2553 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2554 {
2555 	int cpu, thread;
2556 
2557 	for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2558 		for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2559 		     thread++) {
2560 			int fd = FD(evsel, cpu, thread);
2561 
2562 			if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2563 						   cpu, thread, fd) < 0)
2564 				return -1;
2565 		}
2566 	}
2567 
2568 	return 0;
2569 }
2570 
2571 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2572 {
2573 	struct perf_cpu_map *cpus = evsel->core.cpus;
2574 	struct perf_thread_map *threads = evsel->core.threads;
2575 
2576 	if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2577 		return -ENOMEM;
2578 
2579 	return store_evsel_ids(evsel, evlist);
2580 }
2581