xref: /openbmc/linux/tools/perf/util/evsel.c (revision 6aa7de05)
1 /*
2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3  *
4  * Parts came from builtin-{top,stat,record}.c, see those files for further
5  * copyright notes.
6  *
7  * Released under the GPL v2. (and only v2, not any later version)
8  */
9 
10 #include <byteswap.h>
11 #include <errno.h>
12 #include <inttypes.h>
13 #include <linux/bitops.h>
14 #include <api/fs/fs.h>
15 #include <api/fs/tracing_path.h>
16 #include <traceevent/event-parse.h>
17 #include <linux/hw_breakpoint.h>
18 #include <linux/perf_event.h>
19 #include <linux/compiler.h>
20 #include <linux/err.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
24 #include <dirent.h>
25 #include "asm/bug.h"
26 #include "callchain.h"
27 #include "cgroup.h"
28 #include "event.h"
29 #include "evsel.h"
30 #include "evlist.h"
31 #include "util.h"
32 #include "cpumap.h"
33 #include "thread_map.h"
34 #include "target.h"
35 #include "perf_regs.h"
36 #include "debug.h"
37 #include "trace-event.h"
38 #include "stat.h"
39 #include "util/parse-branch-options.h"
40 
41 #include "sane_ctype.h"
42 
43 static struct {
44 	bool sample_id_all;
45 	bool exclude_guest;
46 	bool mmap2;
47 	bool cloexec;
48 	bool clockid;
49 	bool clockid_wrong;
50 	bool lbr_flags;
51 	bool write_backward;
52 	bool group_read;
53 } perf_missing_features;
54 
55 static clockid_t clockid;
56 
57 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
58 {
59 	return 0;
60 }
61 
62 void __weak test_attr__ready(void) { }
63 
64 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
65 {
66 }
67 
68 static struct {
69 	size_t	size;
70 	int	(*init)(struct perf_evsel *evsel);
71 	void	(*fini)(struct perf_evsel *evsel);
72 } perf_evsel__object = {
73 	.size = sizeof(struct perf_evsel),
74 	.init = perf_evsel__no_extra_init,
75 	.fini = perf_evsel__no_extra_fini,
76 };
77 
78 int perf_evsel__object_config(size_t object_size,
79 			      int (*init)(struct perf_evsel *evsel),
80 			      void (*fini)(struct perf_evsel *evsel))
81 {
82 
83 	if (object_size == 0)
84 		goto set_methods;
85 
86 	if (perf_evsel__object.size > object_size)
87 		return -EINVAL;
88 
89 	perf_evsel__object.size = object_size;
90 
91 set_methods:
92 	if (init != NULL)
93 		perf_evsel__object.init = init;
94 
95 	if (fini != NULL)
96 		perf_evsel__object.fini = fini;
97 
98 	return 0;
99 }
100 
101 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
102 
103 int __perf_evsel__sample_size(u64 sample_type)
104 {
105 	u64 mask = sample_type & PERF_SAMPLE_MASK;
106 	int size = 0;
107 	int i;
108 
109 	for (i = 0; i < 64; i++) {
110 		if (mask & (1ULL << i))
111 			size++;
112 	}
113 
114 	size *= sizeof(u64);
115 
116 	return size;
117 }
118 
119 /**
120  * __perf_evsel__calc_id_pos - calculate id_pos.
121  * @sample_type: sample type
122  *
123  * This function returns the position of the event id (PERF_SAMPLE_ID or
124  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
125  * sample_event.
126  */
127 static int __perf_evsel__calc_id_pos(u64 sample_type)
128 {
129 	int idx = 0;
130 
131 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
132 		return 0;
133 
134 	if (!(sample_type & PERF_SAMPLE_ID))
135 		return -1;
136 
137 	if (sample_type & PERF_SAMPLE_IP)
138 		idx += 1;
139 
140 	if (sample_type & PERF_SAMPLE_TID)
141 		idx += 1;
142 
143 	if (sample_type & PERF_SAMPLE_TIME)
144 		idx += 1;
145 
146 	if (sample_type & PERF_SAMPLE_ADDR)
147 		idx += 1;
148 
149 	return idx;
150 }
151 
152 /**
153  * __perf_evsel__calc_is_pos - calculate is_pos.
154  * @sample_type: sample type
155  *
156  * This function returns the position (counting backwards) of the event id
157  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
158  * sample_id_all is used there is an id sample appended to non-sample events.
159  */
160 static int __perf_evsel__calc_is_pos(u64 sample_type)
161 {
162 	int idx = 1;
163 
164 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
165 		return 1;
166 
167 	if (!(sample_type & PERF_SAMPLE_ID))
168 		return -1;
169 
170 	if (sample_type & PERF_SAMPLE_CPU)
171 		idx += 1;
172 
173 	if (sample_type & PERF_SAMPLE_STREAM_ID)
174 		idx += 1;
175 
176 	return idx;
177 }
178 
179 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
180 {
181 	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
182 	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
183 }
184 
185 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
186 				  enum perf_event_sample_format bit)
187 {
188 	if (!(evsel->attr.sample_type & bit)) {
189 		evsel->attr.sample_type |= bit;
190 		evsel->sample_size += sizeof(u64);
191 		perf_evsel__calc_id_pos(evsel);
192 	}
193 }
194 
195 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
196 				    enum perf_event_sample_format bit)
197 {
198 	if (evsel->attr.sample_type & bit) {
199 		evsel->attr.sample_type &= ~bit;
200 		evsel->sample_size -= sizeof(u64);
201 		perf_evsel__calc_id_pos(evsel);
202 	}
203 }
204 
205 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
206 			       bool can_sample_identifier)
207 {
208 	if (can_sample_identifier) {
209 		perf_evsel__reset_sample_bit(evsel, ID);
210 		perf_evsel__set_sample_bit(evsel, IDENTIFIER);
211 	} else {
212 		perf_evsel__set_sample_bit(evsel, ID);
213 	}
214 	evsel->attr.read_format |= PERF_FORMAT_ID;
215 }
216 
217 /**
218  * perf_evsel__is_function_event - Return whether given evsel is a function
219  * trace event
220  *
221  * @evsel - evsel selector to be tested
222  *
223  * Return %true if event is function trace event
224  */
225 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
226 {
227 #define FUNCTION_EVENT "ftrace:function"
228 
229 	return evsel->name &&
230 	       !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
231 
232 #undef FUNCTION_EVENT
233 }
234 
235 void perf_evsel__init(struct perf_evsel *evsel,
236 		      struct perf_event_attr *attr, int idx)
237 {
238 	evsel->idx	   = idx;
239 	evsel->tracking	   = !idx;
240 	evsel->attr	   = *attr;
241 	evsel->leader	   = evsel;
242 	evsel->unit	   = "";
243 	evsel->scale	   = 1.0;
244 	evsel->evlist	   = NULL;
245 	evsel->bpf_fd	   = -1;
246 	INIT_LIST_HEAD(&evsel->node);
247 	INIT_LIST_HEAD(&evsel->config_terms);
248 	perf_evsel__object.init(evsel);
249 	evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
250 	perf_evsel__calc_id_pos(evsel);
251 	evsel->cmdline_group_boundary = false;
252 	evsel->metric_expr   = NULL;
253 	evsel->metric_name   = NULL;
254 	evsel->metric_events = NULL;
255 	evsel->collect_stat  = false;
256 }
257 
258 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
259 {
260 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
261 
262 	if (evsel != NULL)
263 		perf_evsel__init(evsel, attr, idx);
264 
265 	if (perf_evsel__is_bpf_output(evsel)) {
266 		evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
267 					    PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
268 		evsel->attr.sample_period = 1;
269 	}
270 
271 	return evsel;
272 }
273 
274 static bool perf_event_can_profile_kernel(void)
275 {
276 	return geteuid() == 0 || perf_event_paranoid() == -1;
277 }
278 
279 struct perf_evsel *perf_evsel__new_cycles(bool precise)
280 {
281 	struct perf_event_attr attr = {
282 		.type	= PERF_TYPE_HARDWARE,
283 		.config	= PERF_COUNT_HW_CPU_CYCLES,
284 		.exclude_kernel	= !perf_event_can_profile_kernel(),
285 	};
286 	struct perf_evsel *evsel;
287 
288 	event_attr_init(&attr);
289 
290 	if (!precise)
291 		goto new_event;
292 	/*
293 	 * Unnamed union member, not supported as struct member named
294 	 * initializer in older compilers such as gcc 4.4.7
295 	 *
296 	 * Just for probing the precise_ip:
297 	 */
298 	attr.sample_period = 1;
299 
300 	perf_event_attr__set_max_precise_ip(&attr);
301 	/*
302 	 * Now let the usual logic to set up the perf_event_attr defaults
303 	 * to kick in when we return and before perf_evsel__open() is called.
304 	 */
305 	attr.sample_period = 0;
306 new_event:
307 	evsel = perf_evsel__new(&attr);
308 	if (evsel == NULL)
309 		goto out;
310 
311 	/* use asprintf() because free(evsel) assumes name is allocated */
312 	if (asprintf(&evsel->name, "cycles%s%s%.*s",
313 		     (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
314 		     attr.exclude_kernel ? "u" : "",
315 		     attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
316 		goto error_free;
317 out:
318 	return evsel;
319 error_free:
320 	perf_evsel__delete(evsel);
321 	evsel = NULL;
322 	goto out;
323 }
324 
325 /*
326  * Returns pointer with encoded error via <linux/err.h> interface.
327  */
328 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
329 {
330 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
331 	int err = -ENOMEM;
332 
333 	if (evsel == NULL) {
334 		goto out_err;
335 	} else {
336 		struct perf_event_attr attr = {
337 			.type	       = PERF_TYPE_TRACEPOINT,
338 			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
339 					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
340 		};
341 
342 		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
343 			goto out_free;
344 
345 		evsel->tp_format = trace_event__tp_format(sys, name);
346 		if (IS_ERR(evsel->tp_format)) {
347 			err = PTR_ERR(evsel->tp_format);
348 			goto out_free;
349 		}
350 
351 		event_attr_init(&attr);
352 		attr.config = evsel->tp_format->id;
353 		attr.sample_period = 1;
354 		perf_evsel__init(evsel, &attr, idx);
355 	}
356 
357 	return evsel;
358 
359 out_free:
360 	zfree(&evsel->name);
361 	free(evsel);
362 out_err:
363 	return ERR_PTR(err);
364 }
365 
366 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
367 	"cycles",
368 	"instructions",
369 	"cache-references",
370 	"cache-misses",
371 	"branches",
372 	"branch-misses",
373 	"bus-cycles",
374 	"stalled-cycles-frontend",
375 	"stalled-cycles-backend",
376 	"ref-cycles",
377 };
378 
379 static const char *__perf_evsel__hw_name(u64 config)
380 {
381 	if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
382 		return perf_evsel__hw_names[config];
383 
384 	return "unknown-hardware";
385 }
386 
387 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
388 {
389 	int colon = 0, r = 0;
390 	struct perf_event_attr *attr = &evsel->attr;
391 	bool exclude_guest_default = false;
392 
393 #define MOD_PRINT(context, mod)	do {					\
394 		if (!attr->exclude_##context) {				\
395 			if (!colon) colon = ++r;			\
396 			r += scnprintf(bf + r, size - r, "%c", mod);	\
397 		} } while(0)
398 
399 	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
400 		MOD_PRINT(kernel, 'k');
401 		MOD_PRINT(user, 'u');
402 		MOD_PRINT(hv, 'h');
403 		exclude_guest_default = true;
404 	}
405 
406 	if (attr->precise_ip) {
407 		if (!colon)
408 			colon = ++r;
409 		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
410 		exclude_guest_default = true;
411 	}
412 
413 	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
414 		MOD_PRINT(host, 'H');
415 		MOD_PRINT(guest, 'G');
416 	}
417 #undef MOD_PRINT
418 	if (colon)
419 		bf[colon - 1] = ':';
420 	return r;
421 }
422 
423 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
424 {
425 	int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
426 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
427 }
428 
429 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
430 	"cpu-clock",
431 	"task-clock",
432 	"page-faults",
433 	"context-switches",
434 	"cpu-migrations",
435 	"minor-faults",
436 	"major-faults",
437 	"alignment-faults",
438 	"emulation-faults",
439 	"dummy",
440 };
441 
442 static const char *__perf_evsel__sw_name(u64 config)
443 {
444 	if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
445 		return perf_evsel__sw_names[config];
446 	return "unknown-software";
447 }
448 
449 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
450 {
451 	int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
452 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
453 }
454 
455 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
456 {
457 	int r;
458 
459 	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
460 
461 	if (type & HW_BREAKPOINT_R)
462 		r += scnprintf(bf + r, size - r, "r");
463 
464 	if (type & HW_BREAKPOINT_W)
465 		r += scnprintf(bf + r, size - r, "w");
466 
467 	if (type & HW_BREAKPOINT_X)
468 		r += scnprintf(bf + r, size - r, "x");
469 
470 	return r;
471 }
472 
473 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
474 {
475 	struct perf_event_attr *attr = &evsel->attr;
476 	int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
477 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
478 }
479 
480 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
481 				[PERF_EVSEL__MAX_ALIASES] = {
482  { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
483  { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
484  { "LLC",	"L2",							},
485  { "dTLB",	"d-tlb",	"Data-TLB",				},
486  { "iTLB",	"i-tlb",	"Instruction-TLB",			},
487  { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
488  { "node",								},
489 };
490 
491 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
492 				   [PERF_EVSEL__MAX_ALIASES] = {
493  { "load",	"loads",	"read",					},
494  { "store",	"stores",	"write",				},
495  { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
496 };
497 
498 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
499 				       [PERF_EVSEL__MAX_ALIASES] = {
500  { "refs",	"Reference",	"ops",		"access",		},
501  { "misses",	"miss",							},
502 };
503 
504 #define C(x)		PERF_COUNT_HW_CACHE_##x
505 #define CACHE_READ	(1 << C(OP_READ))
506 #define CACHE_WRITE	(1 << C(OP_WRITE))
507 #define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
508 #define COP(x)		(1 << x)
509 
510 /*
511  * cache operartion stat
512  * L1I : Read and prefetch only
513  * ITLB and BPU : Read-only
514  */
515 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
516  [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
517  [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
518  [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
519  [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
520  [C(ITLB)]	= (CACHE_READ),
521  [C(BPU)]	= (CACHE_READ),
522  [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
523 };
524 
525 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
526 {
527 	if (perf_evsel__hw_cache_stat[type] & COP(op))
528 		return true;	/* valid */
529 	else
530 		return false;	/* invalid */
531 }
532 
533 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
534 					    char *bf, size_t size)
535 {
536 	if (result) {
537 		return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
538 				 perf_evsel__hw_cache_op[op][0],
539 				 perf_evsel__hw_cache_result[result][0]);
540 	}
541 
542 	return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
543 			 perf_evsel__hw_cache_op[op][1]);
544 }
545 
546 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
547 {
548 	u8 op, result, type = (config >>  0) & 0xff;
549 	const char *err = "unknown-ext-hardware-cache-type";
550 
551 	if (type >= PERF_COUNT_HW_CACHE_MAX)
552 		goto out_err;
553 
554 	op = (config >>  8) & 0xff;
555 	err = "unknown-ext-hardware-cache-op";
556 	if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
557 		goto out_err;
558 
559 	result = (config >> 16) & 0xff;
560 	err = "unknown-ext-hardware-cache-result";
561 	if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
562 		goto out_err;
563 
564 	err = "invalid-cache";
565 	if (!perf_evsel__is_cache_op_valid(type, op))
566 		goto out_err;
567 
568 	return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
569 out_err:
570 	return scnprintf(bf, size, "%s", err);
571 }
572 
573 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
574 {
575 	int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
576 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
577 }
578 
579 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
580 {
581 	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
582 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
583 }
584 
585 const char *perf_evsel__name(struct perf_evsel *evsel)
586 {
587 	char bf[128];
588 
589 	if (evsel->name)
590 		return evsel->name;
591 
592 	switch (evsel->attr.type) {
593 	case PERF_TYPE_RAW:
594 		perf_evsel__raw_name(evsel, bf, sizeof(bf));
595 		break;
596 
597 	case PERF_TYPE_HARDWARE:
598 		perf_evsel__hw_name(evsel, bf, sizeof(bf));
599 		break;
600 
601 	case PERF_TYPE_HW_CACHE:
602 		perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
603 		break;
604 
605 	case PERF_TYPE_SOFTWARE:
606 		perf_evsel__sw_name(evsel, bf, sizeof(bf));
607 		break;
608 
609 	case PERF_TYPE_TRACEPOINT:
610 		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
611 		break;
612 
613 	case PERF_TYPE_BREAKPOINT:
614 		perf_evsel__bp_name(evsel, bf, sizeof(bf));
615 		break;
616 
617 	default:
618 		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
619 			  evsel->attr.type);
620 		break;
621 	}
622 
623 	evsel->name = strdup(bf);
624 
625 	return evsel->name ?: "unknown";
626 }
627 
628 const char *perf_evsel__group_name(struct perf_evsel *evsel)
629 {
630 	return evsel->group_name ?: "anon group";
631 }
632 
633 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
634 {
635 	int ret;
636 	struct perf_evsel *pos;
637 	const char *group_name = perf_evsel__group_name(evsel);
638 
639 	ret = scnprintf(buf, size, "%s", group_name);
640 
641 	ret += scnprintf(buf + ret, size - ret, " { %s",
642 			 perf_evsel__name(evsel));
643 
644 	for_each_group_member(pos, evsel)
645 		ret += scnprintf(buf + ret, size - ret, ", %s",
646 				 perf_evsel__name(pos));
647 
648 	ret += scnprintf(buf + ret, size - ret, " }");
649 
650 	return ret;
651 }
652 
653 void perf_evsel__config_callchain(struct perf_evsel *evsel,
654 				  struct record_opts *opts,
655 				  struct callchain_param *param)
656 {
657 	bool function = perf_evsel__is_function_event(evsel);
658 	struct perf_event_attr *attr = &evsel->attr;
659 
660 	perf_evsel__set_sample_bit(evsel, CALLCHAIN);
661 
662 	attr->sample_max_stack = param->max_stack;
663 
664 	if (param->record_mode == CALLCHAIN_LBR) {
665 		if (!opts->branch_stack) {
666 			if (attr->exclude_user) {
667 				pr_warning("LBR callstack option is only available "
668 					   "to get user callchain information. "
669 					   "Falling back to framepointers.\n");
670 			} else {
671 				perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
672 				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
673 							PERF_SAMPLE_BRANCH_CALL_STACK |
674 							PERF_SAMPLE_BRANCH_NO_CYCLES |
675 							PERF_SAMPLE_BRANCH_NO_FLAGS;
676 			}
677 		} else
678 			 pr_warning("Cannot use LBR callstack with branch stack. "
679 				    "Falling back to framepointers.\n");
680 	}
681 
682 	if (param->record_mode == CALLCHAIN_DWARF) {
683 		if (!function) {
684 			perf_evsel__set_sample_bit(evsel, REGS_USER);
685 			perf_evsel__set_sample_bit(evsel, STACK_USER);
686 			attr->sample_regs_user = PERF_REGS_MASK;
687 			attr->sample_stack_user = param->dump_size;
688 			attr->exclude_callchain_user = 1;
689 		} else {
690 			pr_info("Cannot use DWARF unwind for function trace event,"
691 				" falling back to framepointers.\n");
692 		}
693 	}
694 
695 	if (function) {
696 		pr_info("Disabling user space callchains for function trace event.\n");
697 		attr->exclude_callchain_user = 1;
698 	}
699 }
700 
701 static void
702 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
703 			    struct callchain_param *param)
704 {
705 	struct perf_event_attr *attr = &evsel->attr;
706 
707 	perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
708 	if (param->record_mode == CALLCHAIN_LBR) {
709 		perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
710 		attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
711 					      PERF_SAMPLE_BRANCH_CALL_STACK);
712 	}
713 	if (param->record_mode == CALLCHAIN_DWARF) {
714 		perf_evsel__reset_sample_bit(evsel, REGS_USER);
715 		perf_evsel__reset_sample_bit(evsel, STACK_USER);
716 	}
717 }
718 
719 static void apply_config_terms(struct perf_evsel *evsel,
720 			       struct record_opts *opts)
721 {
722 	struct perf_evsel_config_term *term;
723 	struct list_head *config_terms = &evsel->config_terms;
724 	struct perf_event_attr *attr = &evsel->attr;
725 	struct callchain_param param;
726 	u32 dump_size = 0;
727 	int max_stack = 0;
728 	const char *callgraph_buf = NULL;
729 
730 	/* callgraph default */
731 	param.record_mode = callchain_param.record_mode;
732 
733 	list_for_each_entry(term, config_terms, list) {
734 		switch (term->type) {
735 		case PERF_EVSEL__CONFIG_TERM_PERIOD:
736 			attr->sample_period = term->val.period;
737 			attr->freq = 0;
738 			break;
739 		case PERF_EVSEL__CONFIG_TERM_FREQ:
740 			attr->sample_freq = term->val.freq;
741 			attr->freq = 1;
742 			break;
743 		case PERF_EVSEL__CONFIG_TERM_TIME:
744 			if (term->val.time)
745 				perf_evsel__set_sample_bit(evsel, TIME);
746 			else
747 				perf_evsel__reset_sample_bit(evsel, TIME);
748 			break;
749 		case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
750 			callgraph_buf = term->val.callgraph;
751 			break;
752 		case PERF_EVSEL__CONFIG_TERM_BRANCH:
753 			if (term->val.branch && strcmp(term->val.branch, "no")) {
754 				perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
755 				parse_branch_str(term->val.branch,
756 						 &attr->branch_sample_type);
757 			} else
758 				perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
759 			break;
760 		case PERF_EVSEL__CONFIG_TERM_STACK_USER:
761 			dump_size = term->val.stack_user;
762 			break;
763 		case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
764 			max_stack = term->val.max_stack;
765 			break;
766 		case PERF_EVSEL__CONFIG_TERM_INHERIT:
767 			/*
768 			 * attr->inherit should has already been set by
769 			 * perf_evsel__config. If user explicitly set
770 			 * inherit using config terms, override global
771 			 * opt->no_inherit setting.
772 			 */
773 			attr->inherit = term->val.inherit ? 1 : 0;
774 			break;
775 		case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
776 			attr->write_backward = term->val.overwrite ? 1 : 0;
777 			break;
778 		default:
779 			break;
780 		}
781 	}
782 
783 	/* User explicitly set per-event callgraph, clear the old setting and reset. */
784 	if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
785 		if (max_stack) {
786 			param.max_stack = max_stack;
787 			if (callgraph_buf == NULL)
788 				callgraph_buf = "fp";
789 		}
790 
791 		/* parse callgraph parameters */
792 		if (callgraph_buf != NULL) {
793 			if (!strcmp(callgraph_buf, "no")) {
794 				param.enabled = false;
795 				param.record_mode = CALLCHAIN_NONE;
796 			} else {
797 				param.enabled = true;
798 				if (parse_callchain_record(callgraph_buf, &param)) {
799 					pr_err("per-event callgraph setting for %s failed. "
800 					       "Apply callgraph global setting for it\n",
801 					       evsel->name);
802 					return;
803 				}
804 			}
805 		}
806 		if (dump_size > 0) {
807 			dump_size = round_up(dump_size, sizeof(u64));
808 			param.dump_size = dump_size;
809 		}
810 
811 		/* If global callgraph set, clear it */
812 		if (callchain_param.enabled)
813 			perf_evsel__reset_callgraph(evsel, &callchain_param);
814 
815 		/* set perf-event callgraph */
816 		if (param.enabled)
817 			perf_evsel__config_callchain(evsel, opts, &param);
818 	}
819 }
820 
821 /*
822  * The enable_on_exec/disabled value strategy:
823  *
824  *  1) For any type of traced program:
825  *    - all independent events and group leaders are disabled
826  *    - all group members are enabled
827  *
828  *     Group members are ruled by group leaders. They need to
829  *     be enabled, because the group scheduling relies on that.
830  *
831  *  2) For traced programs executed by perf:
832  *     - all independent events and group leaders have
833  *       enable_on_exec set
834  *     - we don't specifically enable or disable any event during
835  *       the record command
836  *
837  *     Independent events and group leaders are initially disabled
838  *     and get enabled by exec. Group members are ruled by group
839  *     leaders as stated in 1).
840  *
841  *  3) For traced programs attached by perf (pid/tid):
842  *     - we specifically enable or disable all events during
843  *       the record command
844  *
845  *     When attaching events to already running traced we
846  *     enable/disable events specifically, as there's no
847  *     initial traced exec call.
848  */
849 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
850 			struct callchain_param *callchain)
851 {
852 	struct perf_evsel *leader = evsel->leader;
853 	struct perf_event_attr *attr = &evsel->attr;
854 	int track = evsel->tracking;
855 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
856 
857 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
858 	attr->inherit	    = !opts->no_inherit;
859 	attr->write_backward = opts->overwrite ? 1 : 0;
860 
861 	perf_evsel__set_sample_bit(evsel, IP);
862 	perf_evsel__set_sample_bit(evsel, TID);
863 
864 	if (evsel->sample_read) {
865 		perf_evsel__set_sample_bit(evsel, READ);
866 
867 		/*
868 		 * We need ID even in case of single event, because
869 		 * PERF_SAMPLE_READ process ID specific data.
870 		 */
871 		perf_evsel__set_sample_id(evsel, false);
872 
873 		/*
874 		 * Apply group format only if we belong to group
875 		 * with more than one members.
876 		 */
877 		if (leader->nr_members > 1) {
878 			attr->read_format |= PERF_FORMAT_GROUP;
879 			attr->inherit = 0;
880 		}
881 	}
882 
883 	/*
884 	 * We default some events to have a default interval. But keep
885 	 * it a weak assumption overridable by the user.
886 	 */
887 	if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
888 				     opts->user_interval != ULLONG_MAX)) {
889 		if (opts->freq) {
890 			perf_evsel__set_sample_bit(evsel, PERIOD);
891 			attr->freq		= 1;
892 			attr->sample_freq	= opts->freq;
893 		} else {
894 			attr->sample_period = opts->default_interval;
895 		}
896 	}
897 
898 	/*
899 	 * Disable sampling for all group members other
900 	 * than leader in case leader 'leads' the sampling.
901 	 */
902 	if ((leader != evsel) && leader->sample_read) {
903 		attr->sample_freq   = 0;
904 		attr->sample_period = 0;
905 	}
906 
907 	if (opts->no_samples)
908 		attr->sample_freq = 0;
909 
910 	if (opts->inherit_stat) {
911 		evsel->attr.read_format |=
912 			PERF_FORMAT_TOTAL_TIME_ENABLED |
913 			PERF_FORMAT_TOTAL_TIME_RUNNING |
914 			PERF_FORMAT_ID;
915 		attr->inherit_stat = 1;
916 	}
917 
918 	if (opts->sample_address) {
919 		perf_evsel__set_sample_bit(evsel, ADDR);
920 		attr->mmap_data = track;
921 	}
922 
923 	/*
924 	 * We don't allow user space callchains for  function trace
925 	 * event, due to issues with page faults while tracing page
926 	 * fault handler and its overall trickiness nature.
927 	 */
928 	if (perf_evsel__is_function_event(evsel))
929 		evsel->attr.exclude_callchain_user = 1;
930 
931 	if (callchain && callchain->enabled && !evsel->no_aux_samples)
932 		perf_evsel__config_callchain(evsel, opts, callchain);
933 
934 	if (opts->sample_intr_regs) {
935 		attr->sample_regs_intr = opts->sample_intr_regs;
936 		perf_evsel__set_sample_bit(evsel, REGS_INTR);
937 	}
938 
939 	if (target__has_cpu(&opts->target) || opts->sample_cpu)
940 		perf_evsel__set_sample_bit(evsel, CPU);
941 
942 	if (opts->period)
943 		perf_evsel__set_sample_bit(evsel, PERIOD);
944 
945 	/*
946 	 * When the user explicitly disabled time don't force it here.
947 	 */
948 	if (opts->sample_time &&
949 	    (!perf_missing_features.sample_id_all &&
950 	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
951 	     opts->sample_time_set)))
952 		perf_evsel__set_sample_bit(evsel, TIME);
953 
954 	if (opts->raw_samples && !evsel->no_aux_samples) {
955 		perf_evsel__set_sample_bit(evsel, TIME);
956 		perf_evsel__set_sample_bit(evsel, RAW);
957 		perf_evsel__set_sample_bit(evsel, CPU);
958 	}
959 
960 	if (opts->sample_address)
961 		perf_evsel__set_sample_bit(evsel, DATA_SRC);
962 
963 	if (opts->sample_phys_addr)
964 		perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
965 
966 	if (opts->no_buffering) {
967 		attr->watermark = 0;
968 		attr->wakeup_events = 1;
969 	}
970 	if (opts->branch_stack && !evsel->no_aux_samples) {
971 		perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
972 		attr->branch_sample_type = opts->branch_stack;
973 	}
974 
975 	if (opts->sample_weight)
976 		perf_evsel__set_sample_bit(evsel, WEIGHT);
977 
978 	attr->task  = track;
979 	attr->mmap  = track;
980 	attr->mmap2 = track && !perf_missing_features.mmap2;
981 	attr->comm  = track;
982 
983 	if (opts->record_namespaces)
984 		attr->namespaces  = track;
985 
986 	if (opts->record_switch_events)
987 		attr->context_switch = track;
988 
989 	if (opts->sample_transaction)
990 		perf_evsel__set_sample_bit(evsel, TRANSACTION);
991 
992 	if (opts->running_time) {
993 		evsel->attr.read_format |=
994 			PERF_FORMAT_TOTAL_TIME_ENABLED |
995 			PERF_FORMAT_TOTAL_TIME_RUNNING;
996 	}
997 
998 	/*
999 	 * XXX see the function comment above
1000 	 *
1001 	 * Disabling only independent events or group leaders,
1002 	 * keeping group members enabled.
1003 	 */
1004 	if (perf_evsel__is_group_leader(evsel))
1005 		attr->disabled = 1;
1006 
1007 	/*
1008 	 * Setting enable_on_exec for independent events and
1009 	 * group leaders for traced executed by perf.
1010 	 */
1011 	if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1012 		!opts->initial_delay)
1013 		attr->enable_on_exec = 1;
1014 
1015 	if (evsel->immediate) {
1016 		attr->disabled = 0;
1017 		attr->enable_on_exec = 0;
1018 	}
1019 
1020 	clockid = opts->clockid;
1021 	if (opts->use_clockid) {
1022 		attr->use_clockid = 1;
1023 		attr->clockid = opts->clockid;
1024 	}
1025 
1026 	if (evsel->precise_max)
1027 		perf_event_attr__set_max_precise_ip(attr);
1028 
1029 	if (opts->all_user) {
1030 		attr->exclude_kernel = 1;
1031 		attr->exclude_user   = 0;
1032 	}
1033 
1034 	if (opts->all_kernel) {
1035 		attr->exclude_kernel = 0;
1036 		attr->exclude_user   = 1;
1037 	}
1038 
1039 	/*
1040 	 * Apply event specific term settings,
1041 	 * it overloads any global configuration.
1042 	 */
1043 	apply_config_terms(evsel, opts);
1044 
1045 	evsel->ignore_missing_thread = opts->ignore_missing_thread;
1046 }
1047 
1048 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1049 {
1050 	if (evsel->system_wide)
1051 		nthreads = 1;
1052 
1053 	evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1054 
1055 	if (evsel->fd) {
1056 		int cpu, thread;
1057 		for (cpu = 0; cpu < ncpus; cpu++) {
1058 			for (thread = 0; thread < nthreads; thread++) {
1059 				FD(evsel, cpu, thread) = -1;
1060 			}
1061 		}
1062 	}
1063 
1064 	return evsel->fd != NULL ? 0 : -ENOMEM;
1065 }
1066 
1067 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1068 			  int ioc,  void *arg)
1069 {
1070 	int cpu, thread;
1071 
1072 	for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1073 		for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1074 			int fd = FD(evsel, cpu, thread),
1075 			    err = ioctl(fd, ioc, arg);
1076 
1077 			if (err)
1078 				return err;
1079 		}
1080 	}
1081 
1082 	return 0;
1083 }
1084 
1085 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1086 {
1087 	return perf_evsel__run_ioctl(evsel,
1088 				     PERF_EVENT_IOC_SET_FILTER,
1089 				     (void *)filter);
1090 }
1091 
1092 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1093 {
1094 	char *new_filter = strdup(filter);
1095 
1096 	if (new_filter != NULL) {
1097 		free(evsel->filter);
1098 		evsel->filter = new_filter;
1099 		return 0;
1100 	}
1101 
1102 	return -1;
1103 }
1104 
1105 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1106 				     const char *fmt, const char *filter)
1107 {
1108 	char *new_filter;
1109 
1110 	if (evsel->filter == NULL)
1111 		return perf_evsel__set_filter(evsel, filter);
1112 
1113 	if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1114 		free(evsel->filter);
1115 		evsel->filter = new_filter;
1116 		return 0;
1117 	}
1118 
1119 	return -1;
1120 }
1121 
1122 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1123 {
1124 	return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1125 }
1126 
1127 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1128 {
1129 	return perf_evsel__append_filter(evsel, "%s,%s", filter);
1130 }
1131 
1132 int perf_evsel__enable(struct perf_evsel *evsel)
1133 {
1134 	return perf_evsel__run_ioctl(evsel,
1135 				     PERF_EVENT_IOC_ENABLE,
1136 				     0);
1137 }
1138 
1139 int perf_evsel__disable(struct perf_evsel *evsel)
1140 {
1141 	return perf_evsel__run_ioctl(evsel,
1142 				     PERF_EVENT_IOC_DISABLE,
1143 				     0);
1144 }
1145 
1146 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1147 {
1148 	if (ncpus == 0 || nthreads == 0)
1149 		return 0;
1150 
1151 	if (evsel->system_wide)
1152 		nthreads = 1;
1153 
1154 	evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1155 	if (evsel->sample_id == NULL)
1156 		return -ENOMEM;
1157 
1158 	evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1159 	if (evsel->id == NULL) {
1160 		xyarray__delete(evsel->sample_id);
1161 		evsel->sample_id = NULL;
1162 		return -ENOMEM;
1163 	}
1164 
1165 	return 0;
1166 }
1167 
1168 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1169 {
1170 	xyarray__delete(evsel->fd);
1171 	evsel->fd = NULL;
1172 }
1173 
1174 static void perf_evsel__free_id(struct perf_evsel *evsel)
1175 {
1176 	xyarray__delete(evsel->sample_id);
1177 	evsel->sample_id = NULL;
1178 	zfree(&evsel->id);
1179 }
1180 
1181 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1182 {
1183 	struct perf_evsel_config_term *term, *h;
1184 
1185 	list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1186 		list_del(&term->list);
1187 		free(term);
1188 	}
1189 }
1190 
1191 void perf_evsel__close_fd(struct perf_evsel *evsel)
1192 {
1193 	int cpu, thread;
1194 
1195 	for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1196 		for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1197 			close(FD(evsel, cpu, thread));
1198 			FD(evsel, cpu, thread) = -1;
1199 		}
1200 }
1201 
1202 void perf_evsel__exit(struct perf_evsel *evsel)
1203 {
1204 	assert(list_empty(&evsel->node));
1205 	assert(evsel->evlist == NULL);
1206 	perf_evsel__free_fd(evsel);
1207 	perf_evsel__free_id(evsel);
1208 	perf_evsel__free_config_terms(evsel);
1209 	close_cgroup(evsel->cgrp);
1210 	cpu_map__put(evsel->cpus);
1211 	cpu_map__put(evsel->own_cpus);
1212 	thread_map__put(evsel->threads);
1213 	zfree(&evsel->group_name);
1214 	zfree(&evsel->name);
1215 	perf_evsel__object.fini(evsel);
1216 }
1217 
1218 void perf_evsel__delete(struct perf_evsel *evsel)
1219 {
1220 	perf_evsel__exit(evsel);
1221 	free(evsel);
1222 }
1223 
1224 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1225 				struct perf_counts_values *count)
1226 {
1227 	struct perf_counts_values tmp;
1228 
1229 	if (!evsel->prev_raw_counts)
1230 		return;
1231 
1232 	if (cpu == -1) {
1233 		tmp = evsel->prev_raw_counts->aggr;
1234 		evsel->prev_raw_counts->aggr = *count;
1235 	} else {
1236 		tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1237 		*perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1238 	}
1239 
1240 	count->val = count->val - tmp.val;
1241 	count->ena = count->ena - tmp.ena;
1242 	count->run = count->run - tmp.run;
1243 }
1244 
1245 void perf_counts_values__scale(struct perf_counts_values *count,
1246 			       bool scale, s8 *pscaled)
1247 {
1248 	s8 scaled = 0;
1249 
1250 	if (scale) {
1251 		if (count->run == 0) {
1252 			scaled = -1;
1253 			count->val = 0;
1254 		} else if (count->run < count->ena) {
1255 			scaled = 1;
1256 			count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1257 		}
1258 	} else
1259 		count->ena = count->run = 0;
1260 
1261 	if (pscaled)
1262 		*pscaled = scaled;
1263 }
1264 
1265 static int perf_evsel__read_size(struct perf_evsel *evsel)
1266 {
1267 	u64 read_format = evsel->attr.read_format;
1268 	int entry = sizeof(u64); /* value */
1269 	int size = 0;
1270 	int nr = 1;
1271 
1272 	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1273 		size += sizeof(u64);
1274 
1275 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1276 		size += sizeof(u64);
1277 
1278 	if (read_format & PERF_FORMAT_ID)
1279 		entry += sizeof(u64);
1280 
1281 	if (read_format & PERF_FORMAT_GROUP) {
1282 		nr = evsel->nr_members;
1283 		size += sizeof(u64);
1284 	}
1285 
1286 	size += entry * nr;
1287 	return size;
1288 }
1289 
1290 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1291 		     struct perf_counts_values *count)
1292 {
1293 	size_t size = perf_evsel__read_size(evsel);
1294 
1295 	memset(count, 0, sizeof(*count));
1296 
1297 	if (FD(evsel, cpu, thread) < 0)
1298 		return -EINVAL;
1299 
1300 	if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1301 		return -errno;
1302 
1303 	return 0;
1304 }
1305 
1306 static int
1307 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1308 {
1309 	struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1310 
1311 	return perf_evsel__read(evsel, cpu, thread, count);
1312 }
1313 
1314 static void
1315 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1316 		      u64 val, u64 ena, u64 run)
1317 {
1318 	struct perf_counts_values *count;
1319 
1320 	count = perf_counts(counter->counts, cpu, thread);
1321 
1322 	count->val    = val;
1323 	count->ena    = ena;
1324 	count->run    = run;
1325 	count->loaded = true;
1326 }
1327 
1328 static int
1329 perf_evsel__process_group_data(struct perf_evsel *leader,
1330 			       int cpu, int thread, u64 *data)
1331 {
1332 	u64 read_format = leader->attr.read_format;
1333 	struct sample_read_value *v;
1334 	u64 nr, ena = 0, run = 0, i;
1335 
1336 	nr = *data++;
1337 
1338 	if (nr != (u64) leader->nr_members)
1339 		return -EINVAL;
1340 
1341 	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1342 		ena = *data++;
1343 
1344 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1345 		run = *data++;
1346 
1347 	v = (struct sample_read_value *) data;
1348 
1349 	perf_evsel__set_count(leader, cpu, thread,
1350 			      v[0].value, ena, run);
1351 
1352 	for (i = 1; i < nr; i++) {
1353 		struct perf_evsel *counter;
1354 
1355 		counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1356 		if (!counter)
1357 			return -EINVAL;
1358 
1359 		perf_evsel__set_count(counter, cpu, thread,
1360 				      v[i].value, ena, run);
1361 	}
1362 
1363 	return 0;
1364 }
1365 
1366 static int
1367 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1368 {
1369 	struct perf_stat_evsel *ps = leader->priv;
1370 	u64 read_format = leader->attr.read_format;
1371 	int size = perf_evsel__read_size(leader);
1372 	u64 *data = ps->group_data;
1373 
1374 	if (!(read_format & PERF_FORMAT_ID))
1375 		return -EINVAL;
1376 
1377 	if (!perf_evsel__is_group_leader(leader))
1378 		return -EINVAL;
1379 
1380 	if (!data) {
1381 		data = zalloc(size);
1382 		if (!data)
1383 			return -ENOMEM;
1384 
1385 		ps->group_data = data;
1386 	}
1387 
1388 	if (FD(leader, cpu, thread) < 0)
1389 		return -EINVAL;
1390 
1391 	if (readn(FD(leader, cpu, thread), data, size) <= 0)
1392 		return -errno;
1393 
1394 	return perf_evsel__process_group_data(leader, cpu, thread, data);
1395 }
1396 
1397 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1398 {
1399 	u64 read_format = evsel->attr.read_format;
1400 
1401 	if (read_format & PERF_FORMAT_GROUP)
1402 		return perf_evsel__read_group(evsel, cpu, thread);
1403 	else
1404 		return perf_evsel__read_one(evsel, cpu, thread);
1405 }
1406 
1407 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1408 			      int cpu, int thread, bool scale)
1409 {
1410 	struct perf_counts_values count;
1411 	size_t nv = scale ? 3 : 1;
1412 
1413 	if (FD(evsel, cpu, thread) < 0)
1414 		return -EINVAL;
1415 
1416 	if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1417 		return -ENOMEM;
1418 
1419 	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1420 		return -errno;
1421 
1422 	perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1423 	perf_counts_values__scale(&count, scale, NULL);
1424 	*perf_counts(evsel->counts, cpu, thread) = count;
1425 	return 0;
1426 }
1427 
1428 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1429 {
1430 	struct perf_evsel *leader = evsel->leader;
1431 	int fd;
1432 
1433 	if (perf_evsel__is_group_leader(evsel))
1434 		return -1;
1435 
1436 	/*
1437 	 * Leader must be already processed/open,
1438 	 * if not it's a bug.
1439 	 */
1440 	BUG_ON(!leader->fd);
1441 
1442 	fd = FD(leader, cpu, thread);
1443 	BUG_ON(fd == -1);
1444 
1445 	return fd;
1446 }
1447 
1448 struct bit_names {
1449 	int bit;
1450 	const char *name;
1451 };
1452 
1453 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1454 {
1455 	bool first_bit = true;
1456 	int i = 0;
1457 
1458 	do {
1459 		if (value & bits[i].bit) {
1460 			buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1461 			first_bit = false;
1462 		}
1463 	} while (bits[++i].name != NULL);
1464 }
1465 
1466 static void __p_sample_type(char *buf, size_t size, u64 value)
1467 {
1468 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1469 	struct bit_names bits[] = {
1470 		bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1471 		bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1472 		bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1473 		bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1474 		bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1475 		bit_name(WEIGHT), bit_name(PHYS_ADDR),
1476 		{ .name = NULL, }
1477 	};
1478 #undef bit_name
1479 	__p_bits(buf, size, value, bits);
1480 }
1481 
1482 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1483 {
1484 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1485 	struct bit_names bits[] = {
1486 		bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1487 		bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1488 		bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1489 		bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1490 		bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1491 		{ .name = NULL, }
1492 	};
1493 #undef bit_name
1494 	__p_bits(buf, size, value, bits);
1495 }
1496 
1497 static void __p_read_format(char *buf, size_t size, u64 value)
1498 {
1499 #define bit_name(n) { PERF_FORMAT_##n, #n }
1500 	struct bit_names bits[] = {
1501 		bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1502 		bit_name(ID), bit_name(GROUP),
1503 		{ .name = NULL, }
1504 	};
1505 #undef bit_name
1506 	__p_bits(buf, size, value, bits);
1507 }
1508 
1509 #define BUF_SIZE		1024
1510 
1511 #define p_hex(val)		snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1512 #define p_unsigned(val)		snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1513 #define p_signed(val)		snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1514 #define p_sample_type(val)	__p_sample_type(buf, BUF_SIZE, val)
1515 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1516 #define p_read_format(val)	__p_read_format(buf, BUF_SIZE, val)
1517 
1518 #define PRINT_ATTRn(_n, _f, _p)				\
1519 do {							\
1520 	if (attr->_f) {					\
1521 		_p(attr->_f);				\
1522 		ret += attr__fprintf(fp, _n, buf, priv);\
1523 	}						\
1524 } while (0)
1525 
1526 #define PRINT_ATTRf(_f, _p)	PRINT_ATTRn(#_f, _f, _p)
1527 
1528 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1529 			     attr__fprintf_f attr__fprintf, void *priv)
1530 {
1531 	char buf[BUF_SIZE];
1532 	int ret = 0;
1533 
1534 	PRINT_ATTRf(type, p_unsigned);
1535 	PRINT_ATTRf(size, p_unsigned);
1536 	PRINT_ATTRf(config, p_hex);
1537 	PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1538 	PRINT_ATTRf(sample_type, p_sample_type);
1539 	PRINT_ATTRf(read_format, p_read_format);
1540 
1541 	PRINT_ATTRf(disabled, p_unsigned);
1542 	PRINT_ATTRf(inherit, p_unsigned);
1543 	PRINT_ATTRf(pinned, p_unsigned);
1544 	PRINT_ATTRf(exclusive, p_unsigned);
1545 	PRINT_ATTRf(exclude_user, p_unsigned);
1546 	PRINT_ATTRf(exclude_kernel, p_unsigned);
1547 	PRINT_ATTRf(exclude_hv, p_unsigned);
1548 	PRINT_ATTRf(exclude_idle, p_unsigned);
1549 	PRINT_ATTRf(mmap, p_unsigned);
1550 	PRINT_ATTRf(comm, p_unsigned);
1551 	PRINT_ATTRf(freq, p_unsigned);
1552 	PRINT_ATTRf(inherit_stat, p_unsigned);
1553 	PRINT_ATTRf(enable_on_exec, p_unsigned);
1554 	PRINT_ATTRf(task, p_unsigned);
1555 	PRINT_ATTRf(watermark, p_unsigned);
1556 	PRINT_ATTRf(precise_ip, p_unsigned);
1557 	PRINT_ATTRf(mmap_data, p_unsigned);
1558 	PRINT_ATTRf(sample_id_all, p_unsigned);
1559 	PRINT_ATTRf(exclude_host, p_unsigned);
1560 	PRINT_ATTRf(exclude_guest, p_unsigned);
1561 	PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1562 	PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1563 	PRINT_ATTRf(mmap2, p_unsigned);
1564 	PRINT_ATTRf(comm_exec, p_unsigned);
1565 	PRINT_ATTRf(use_clockid, p_unsigned);
1566 	PRINT_ATTRf(context_switch, p_unsigned);
1567 	PRINT_ATTRf(write_backward, p_unsigned);
1568 
1569 	PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1570 	PRINT_ATTRf(bp_type, p_unsigned);
1571 	PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1572 	PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1573 	PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1574 	PRINT_ATTRf(sample_regs_user, p_hex);
1575 	PRINT_ATTRf(sample_stack_user, p_unsigned);
1576 	PRINT_ATTRf(clockid, p_signed);
1577 	PRINT_ATTRf(sample_regs_intr, p_hex);
1578 	PRINT_ATTRf(aux_watermark, p_unsigned);
1579 	PRINT_ATTRf(sample_max_stack, p_unsigned);
1580 
1581 	return ret;
1582 }
1583 
1584 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1585 				void *priv __maybe_unused)
1586 {
1587 	return fprintf(fp, "  %-32s %s\n", name, val);
1588 }
1589 
1590 static bool ignore_missing_thread(struct perf_evsel *evsel,
1591 				  struct thread_map *threads,
1592 				  int thread, int err)
1593 {
1594 	if (!evsel->ignore_missing_thread)
1595 		return false;
1596 
1597 	/* The system wide setup does not work with threads. */
1598 	if (evsel->system_wide)
1599 		return false;
1600 
1601 	/* The -ESRCH is perf event syscall errno for pid's not found. */
1602 	if (err != -ESRCH)
1603 		return false;
1604 
1605 	/* If there's only one thread, let it fail. */
1606 	if (threads->nr == 1)
1607 		return false;
1608 
1609 	if (thread_map__remove(threads, thread))
1610 		return false;
1611 
1612 	pr_warning("WARNING: Ignored open failure for pid %d\n",
1613 		   thread_map__pid(threads, thread));
1614 	return true;
1615 }
1616 
1617 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1618 		     struct thread_map *threads)
1619 {
1620 	int cpu, thread, nthreads;
1621 	unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1622 	int pid = -1, err;
1623 	enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1624 
1625 	if (perf_missing_features.write_backward && evsel->attr.write_backward)
1626 		return -EINVAL;
1627 
1628 	if (cpus == NULL) {
1629 		static struct cpu_map *empty_cpu_map;
1630 
1631 		if (empty_cpu_map == NULL) {
1632 			empty_cpu_map = cpu_map__dummy_new();
1633 			if (empty_cpu_map == NULL)
1634 				return -ENOMEM;
1635 		}
1636 
1637 		cpus = empty_cpu_map;
1638 	}
1639 
1640 	if (threads == NULL) {
1641 		static struct thread_map *empty_thread_map;
1642 
1643 		if (empty_thread_map == NULL) {
1644 			empty_thread_map = thread_map__new_by_tid(-1);
1645 			if (empty_thread_map == NULL)
1646 				return -ENOMEM;
1647 		}
1648 
1649 		threads = empty_thread_map;
1650 	}
1651 
1652 	if (evsel->system_wide)
1653 		nthreads = 1;
1654 	else
1655 		nthreads = threads->nr;
1656 
1657 	if (evsel->fd == NULL &&
1658 	    perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1659 		return -ENOMEM;
1660 
1661 	if (evsel->cgrp) {
1662 		flags |= PERF_FLAG_PID_CGROUP;
1663 		pid = evsel->cgrp->fd;
1664 	}
1665 
1666 fallback_missing_features:
1667 	if (perf_missing_features.clockid_wrong)
1668 		evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1669 	if (perf_missing_features.clockid) {
1670 		evsel->attr.use_clockid = 0;
1671 		evsel->attr.clockid = 0;
1672 	}
1673 	if (perf_missing_features.cloexec)
1674 		flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1675 	if (perf_missing_features.mmap2)
1676 		evsel->attr.mmap2 = 0;
1677 	if (perf_missing_features.exclude_guest)
1678 		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1679 	if (perf_missing_features.lbr_flags)
1680 		evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1681 				     PERF_SAMPLE_BRANCH_NO_CYCLES);
1682 	if (perf_missing_features.group_read && evsel->attr.inherit)
1683 		evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1684 retry_sample_id:
1685 	if (perf_missing_features.sample_id_all)
1686 		evsel->attr.sample_id_all = 0;
1687 
1688 	if (verbose >= 2) {
1689 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1690 		fprintf(stderr, "perf_event_attr:\n");
1691 		perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1692 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1693 	}
1694 
1695 	for (cpu = 0; cpu < cpus->nr; cpu++) {
1696 
1697 		for (thread = 0; thread < nthreads; thread++) {
1698 			int fd, group_fd;
1699 
1700 			if (!evsel->cgrp && !evsel->system_wide)
1701 				pid = thread_map__pid(threads, thread);
1702 
1703 			group_fd = get_group_fd(evsel, cpu, thread);
1704 retry_open:
1705 			pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1706 				  pid, cpus->map[cpu], group_fd, flags);
1707 
1708 			test_attr__ready();
1709 
1710 			fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1711 						 group_fd, flags);
1712 
1713 			FD(evsel, cpu, thread) = fd;
1714 
1715 			if (fd < 0) {
1716 				err = -errno;
1717 
1718 				if (ignore_missing_thread(evsel, threads, thread, err)) {
1719 					/*
1720 					 * We just removed 1 thread, so take a step
1721 					 * back on thread index and lower the upper
1722 					 * nthreads limit.
1723 					 */
1724 					nthreads--;
1725 					thread--;
1726 
1727 					/* ... and pretend like nothing have happened. */
1728 					err = 0;
1729 					continue;
1730 				}
1731 
1732 				pr_debug2("\nsys_perf_event_open failed, error %d\n",
1733 					  err);
1734 				goto try_fallback;
1735 			}
1736 
1737 			pr_debug2(" = %d\n", fd);
1738 
1739 			if (evsel->bpf_fd >= 0) {
1740 				int evt_fd = fd;
1741 				int bpf_fd = evsel->bpf_fd;
1742 
1743 				err = ioctl(evt_fd,
1744 					    PERF_EVENT_IOC_SET_BPF,
1745 					    bpf_fd);
1746 				if (err && errno != EEXIST) {
1747 					pr_err("failed to attach bpf fd %d: %s\n",
1748 					       bpf_fd, strerror(errno));
1749 					err = -EINVAL;
1750 					goto out_close;
1751 				}
1752 			}
1753 
1754 			set_rlimit = NO_CHANGE;
1755 
1756 			/*
1757 			 * If we succeeded but had to kill clockid, fail and
1758 			 * have perf_evsel__open_strerror() print us a nice
1759 			 * error.
1760 			 */
1761 			if (perf_missing_features.clockid ||
1762 			    perf_missing_features.clockid_wrong) {
1763 				err = -EINVAL;
1764 				goto out_close;
1765 			}
1766 		}
1767 	}
1768 
1769 	return 0;
1770 
1771 try_fallback:
1772 	/*
1773 	 * perf stat needs between 5 and 22 fds per CPU. When we run out
1774 	 * of them try to increase the limits.
1775 	 */
1776 	if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1777 		struct rlimit l;
1778 		int old_errno = errno;
1779 
1780 		if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1781 			if (set_rlimit == NO_CHANGE)
1782 				l.rlim_cur = l.rlim_max;
1783 			else {
1784 				l.rlim_cur = l.rlim_max + 1000;
1785 				l.rlim_max = l.rlim_cur;
1786 			}
1787 			if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1788 				set_rlimit++;
1789 				errno = old_errno;
1790 				goto retry_open;
1791 			}
1792 		}
1793 		errno = old_errno;
1794 	}
1795 
1796 	if (err != -EINVAL || cpu > 0 || thread > 0)
1797 		goto out_close;
1798 
1799 	/*
1800 	 * Must probe features in the order they were added to the
1801 	 * perf_event_attr interface.
1802 	 */
1803 	if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1804 		perf_missing_features.write_backward = true;
1805 		pr_debug2("switching off write_backward\n");
1806 		goto out_close;
1807 	} else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1808 		perf_missing_features.clockid_wrong = true;
1809 		pr_debug2("switching off clockid\n");
1810 		goto fallback_missing_features;
1811 	} else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1812 		perf_missing_features.clockid = true;
1813 		pr_debug2("switching off use_clockid\n");
1814 		goto fallback_missing_features;
1815 	} else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1816 		perf_missing_features.cloexec = true;
1817 		pr_debug2("switching off cloexec flag\n");
1818 		goto fallback_missing_features;
1819 	} else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1820 		perf_missing_features.mmap2 = true;
1821 		pr_debug2("switching off mmap2\n");
1822 		goto fallback_missing_features;
1823 	} else if (!perf_missing_features.exclude_guest &&
1824 		   (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1825 		perf_missing_features.exclude_guest = true;
1826 		pr_debug2("switching off exclude_guest, exclude_host\n");
1827 		goto fallback_missing_features;
1828 	} else if (!perf_missing_features.sample_id_all) {
1829 		perf_missing_features.sample_id_all = true;
1830 		pr_debug2("switching off sample_id_all\n");
1831 		goto retry_sample_id;
1832 	} else if (!perf_missing_features.lbr_flags &&
1833 			(evsel->attr.branch_sample_type &
1834 			 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1835 			  PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1836 		perf_missing_features.lbr_flags = true;
1837 		pr_debug2("switching off branch sample type no (cycles/flags)\n");
1838 		goto fallback_missing_features;
1839 	} else if (!perf_missing_features.group_read &&
1840 		    evsel->attr.inherit &&
1841 		   (evsel->attr.read_format & PERF_FORMAT_GROUP)) {
1842 		perf_missing_features.group_read = true;
1843 		pr_debug2("switching off group read\n");
1844 		goto fallback_missing_features;
1845 	}
1846 out_close:
1847 	do {
1848 		while (--thread >= 0) {
1849 			close(FD(evsel, cpu, thread));
1850 			FD(evsel, cpu, thread) = -1;
1851 		}
1852 		thread = nthreads;
1853 	} while (--cpu >= 0);
1854 	return err;
1855 }
1856 
1857 void perf_evsel__close(struct perf_evsel *evsel)
1858 {
1859 	if (evsel->fd == NULL)
1860 		return;
1861 
1862 	perf_evsel__close_fd(evsel);
1863 	perf_evsel__free_fd(evsel);
1864 }
1865 
1866 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1867 			     struct cpu_map *cpus)
1868 {
1869 	return perf_evsel__open(evsel, cpus, NULL);
1870 }
1871 
1872 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1873 				struct thread_map *threads)
1874 {
1875 	return perf_evsel__open(evsel, NULL, threads);
1876 }
1877 
1878 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1879 				       const union perf_event *event,
1880 				       struct perf_sample *sample)
1881 {
1882 	u64 type = evsel->attr.sample_type;
1883 	const u64 *array = event->sample.array;
1884 	bool swapped = evsel->needs_swap;
1885 	union u64_swap u;
1886 
1887 	array += ((event->header.size -
1888 		   sizeof(event->header)) / sizeof(u64)) - 1;
1889 
1890 	if (type & PERF_SAMPLE_IDENTIFIER) {
1891 		sample->id = *array;
1892 		array--;
1893 	}
1894 
1895 	if (type & PERF_SAMPLE_CPU) {
1896 		u.val64 = *array;
1897 		if (swapped) {
1898 			/* undo swap of u64, then swap on individual u32s */
1899 			u.val64 = bswap_64(u.val64);
1900 			u.val32[0] = bswap_32(u.val32[0]);
1901 		}
1902 
1903 		sample->cpu = u.val32[0];
1904 		array--;
1905 	}
1906 
1907 	if (type & PERF_SAMPLE_STREAM_ID) {
1908 		sample->stream_id = *array;
1909 		array--;
1910 	}
1911 
1912 	if (type & PERF_SAMPLE_ID) {
1913 		sample->id = *array;
1914 		array--;
1915 	}
1916 
1917 	if (type & PERF_SAMPLE_TIME) {
1918 		sample->time = *array;
1919 		array--;
1920 	}
1921 
1922 	if (type & PERF_SAMPLE_TID) {
1923 		u.val64 = *array;
1924 		if (swapped) {
1925 			/* undo swap of u64, then swap on individual u32s */
1926 			u.val64 = bswap_64(u.val64);
1927 			u.val32[0] = bswap_32(u.val32[0]);
1928 			u.val32[1] = bswap_32(u.val32[1]);
1929 		}
1930 
1931 		sample->pid = u.val32[0];
1932 		sample->tid = u.val32[1];
1933 		array--;
1934 	}
1935 
1936 	return 0;
1937 }
1938 
1939 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1940 			    u64 size)
1941 {
1942 	return size > max_size || offset + size > endp;
1943 }
1944 
1945 #define OVERFLOW_CHECK(offset, size, max_size)				\
1946 	do {								\
1947 		if (overflow(endp, (max_size), (offset), (size)))	\
1948 			return -EFAULT;					\
1949 	} while (0)
1950 
1951 #define OVERFLOW_CHECK_u64(offset) \
1952 	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1953 
1954 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1955 			     struct perf_sample *data)
1956 {
1957 	u64 type = evsel->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->attr.sample_period;
1974 	data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1975 
1976 	if (event->header.type != PERF_RECORD_SAMPLE) {
1977 		if (!evsel->attr.sample_id_all)
1978 			return 0;
1979 		return perf_evsel__parse_id_sample(evsel, event, data);
1980 	}
1981 
1982 	array = event->sample.array;
1983 
1984 	/*
1985 	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1986 	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1987 	 * check the format does not go past the end of the event.
1988 	 */
1989 	if (evsel->sample_size + sizeof(event->header) > event->header.size)
1990 		return -EFAULT;
1991 
1992 	data->id = -1ULL;
1993 	if (type & PERF_SAMPLE_IDENTIFIER) {
1994 		data->id = *array;
1995 		array++;
1996 	}
1997 
1998 	if (type & PERF_SAMPLE_IP) {
1999 		data->ip = *array;
2000 		array++;
2001 	}
2002 
2003 	if (type & PERF_SAMPLE_TID) {
2004 		u.val64 = *array;
2005 		if (swapped) {
2006 			/* undo swap of u64, then swap on individual u32s */
2007 			u.val64 = bswap_64(u.val64);
2008 			u.val32[0] = bswap_32(u.val32[0]);
2009 			u.val32[1] = bswap_32(u.val32[1]);
2010 		}
2011 
2012 		data->pid = u.val32[0];
2013 		data->tid = u.val32[1];
2014 		array++;
2015 	}
2016 
2017 	if (type & PERF_SAMPLE_TIME) {
2018 		data->time = *array;
2019 		array++;
2020 	}
2021 
2022 	data->addr = 0;
2023 	if (type & PERF_SAMPLE_ADDR) {
2024 		data->addr = *array;
2025 		array++;
2026 	}
2027 
2028 	if (type & PERF_SAMPLE_ID) {
2029 		data->id = *array;
2030 		array++;
2031 	}
2032 
2033 	if (type & PERF_SAMPLE_STREAM_ID) {
2034 		data->stream_id = *array;
2035 		array++;
2036 	}
2037 
2038 	if (type & PERF_SAMPLE_CPU) {
2039 
2040 		u.val64 = *array;
2041 		if (swapped) {
2042 			/* undo swap of u64, then swap on individual u32s */
2043 			u.val64 = bswap_64(u.val64);
2044 			u.val32[0] = bswap_32(u.val32[0]);
2045 		}
2046 
2047 		data->cpu = u.val32[0];
2048 		array++;
2049 	}
2050 
2051 	if (type & PERF_SAMPLE_PERIOD) {
2052 		data->period = *array;
2053 		array++;
2054 	}
2055 
2056 	if (type & PERF_SAMPLE_READ) {
2057 		u64 read_format = evsel->attr.read_format;
2058 
2059 		OVERFLOW_CHECK_u64(array);
2060 		if (read_format & PERF_FORMAT_GROUP)
2061 			data->read.group.nr = *array;
2062 		else
2063 			data->read.one.value = *array;
2064 
2065 		array++;
2066 
2067 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2068 			OVERFLOW_CHECK_u64(array);
2069 			data->read.time_enabled = *array;
2070 			array++;
2071 		}
2072 
2073 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2074 			OVERFLOW_CHECK_u64(array);
2075 			data->read.time_running = *array;
2076 			array++;
2077 		}
2078 
2079 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2080 		if (read_format & PERF_FORMAT_GROUP) {
2081 			const u64 max_group_nr = UINT64_MAX /
2082 					sizeof(struct sample_read_value);
2083 
2084 			if (data->read.group.nr > max_group_nr)
2085 				return -EFAULT;
2086 			sz = data->read.group.nr *
2087 			     sizeof(struct sample_read_value);
2088 			OVERFLOW_CHECK(array, sz, max_size);
2089 			data->read.group.values =
2090 					(struct sample_read_value *)array;
2091 			array = (void *)array + sz;
2092 		} else {
2093 			OVERFLOW_CHECK_u64(array);
2094 			data->read.one.id = *array;
2095 			array++;
2096 		}
2097 	}
2098 
2099 	if (type & PERF_SAMPLE_CALLCHAIN) {
2100 		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2101 
2102 		OVERFLOW_CHECK_u64(array);
2103 		data->callchain = (struct ip_callchain *)array++;
2104 		if (data->callchain->nr > max_callchain_nr)
2105 			return -EFAULT;
2106 		sz = data->callchain->nr * sizeof(u64);
2107 		OVERFLOW_CHECK(array, sz, max_size);
2108 		array = (void *)array + sz;
2109 	}
2110 
2111 	if (type & PERF_SAMPLE_RAW) {
2112 		OVERFLOW_CHECK_u64(array);
2113 		u.val64 = *array;
2114 		if (WARN_ONCE(swapped,
2115 			      "Endianness of raw data not corrected!\n")) {
2116 			/* undo swap of u64, then swap on individual u32s */
2117 			u.val64 = bswap_64(u.val64);
2118 			u.val32[0] = bswap_32(u.val32[0]);
2119 			u.val32[1] = bswap_32(u.val32[1]);
2120 		}
2121 		data->raw_size = u.val32[0];
2122 		array = (void *)array + sizeof(u32);
2123 
2124 		OVERFLOW_CHECK(array, data->raw_size, max_size);
2125 		data->raw_data = (void *)array;
2126 		array = (void *)array + data->raw_size;
2127 	}
2128 
2129 	if (type & PERF_SAMPLE_BRANCH_STACK) {
2130 		const u64 max_branch_nr = UINT64_MAX /
2131 					  sizeof(struct branch_entry);
2132 
2133 		OVERFLOW_CHECK_u64(array);
2134 		data->branch_stack = (struct branch_stack *)array++;
2135 
2136 		if (data->branch_stack->nr > max_branch_nr)
2137 			return -EFAULT;
2138 		sz = data->branch_stack->nr * sizeof(struct branch_entry);
2139 		OVERFLOW_CHECK(array, sz, max_size);
2140 		array = (void *)array + sz;
2141 	}
2142 
2143 	if (type & PERF_SAMPLE_REGS_USER) {
2144 		OVERFLOW_CHECK_u64(array);
2145 		data->user_regs.abi = *array;
2146 		array++;
2147 
2148 		if (data->user_regs.abi) {
2149 			u64 mask = evsel->attr.sample_regs_user;
2150 
2151 			sz = hweight_long(mask) * sizeof(u64);
2152 			OVERFLOW_CHECK(array, sz, max_size);
2153 			data->user_regs.mask = mask;
2154 			data->user_regs.regs = (u64 *)array;
2155 			array = (void *)array + sz;
2156 		}
2157 	}
2158 
2159 	if (type & PERF_SAMPLE_STACK_USER) {
2160 		OVERFLOW_CHECK_u64(array);
2161 		sz = *array++;
2162 
2163 		data->user_stack.offset = ((char *)(array - 1)
2164 					  - (char *) event);
2165 
2166 		if (!sz) {
2167 			data->user_stack.size = 0;
2168 		} else {
2169 			OVERFLOW_CHECK(array, sz, max_size);
2170 			data->user_stack.data = (char *)array;
2171 			array = (void *)array + sz;
2172 			OVERFLOW_CHECK_u64(array);
2173 			data->user_stack.size = *array++;
2174 			if (WARN_ONCE(data->user_stack.size > sz,
2175 				      "user stack dump failure\n"))
2176 				return -EFAULT;
2177 		}
2178 	}
2179 
2180 	if (type & PERF_SAMPLE_WEIGHT) {
2181 		OVERFLOW_CHECK_u64(array);
2182 		data->weight = *array;
2183 		array++;
2184 	}
2185 
2186 	data->data_src = PERF_MEM_DATA_SRC_NONE;
2187 	if (type & PERF_SAMPLE_DATA_SRC) {
2188 		OVERFLOW_CHECK_u64(array);
2189 		data->data_src = *array;
2190 		array++;
2191 	}
2192 
2193 	data->transaction = 0;
2194 	if (type & PERF_SAMPLE_TRANSACTION) {
2195 		OVERFLOW_CHECK_u64(array);
2196 		data->transaction = *array;
2197 		array++;
2198 	}
2199 
2200 	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2201 	if (type & PERF_SAMPLE_REGS_INTR) {
2202 		OVERFLOW_CHECK_u64(array);
2203 		data->intr_regs.abi = *array;
2204 		array++;
2205 
2206 		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2207 			u64 mask = evsel->attr.sample_regs_intr;
2208 
2209 			sz = hweight_long(mask) * sizeof(u64);
2210 			OVERFLOW_CHECK(array, sz, max_size);
2211 			data->intr_regs.mask = mask;
2212 			data->intr_regs.regs = (u64 *)array;
2213 			array = (void *)array + sz;
2214 		}
2215 	}
2216 
2217 	data->phys_addr = 0;
2218 	if (type & PERF_SAMPLE_PHYS_ADDR) {
2219 		data->phys_addr = *array;
2220 		array++;
2221 	}
2222 
2223 	return 0;
2224 }
2225 
2226 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2227 				     u64 read_format)
2228 {
2229 	size_t sz, result = sizeof(struct sample_event);
2230 
2231 	if (type & PERF_SAMPLE_IDENTIFIER)
2232 		result += sizeof(u64);
2233 
2234 	if (type & PERF_SAMPLE_IP)
2235 		result += sizeof(u64);
2236 
2237 	if (type & PERF_SAMPLE_TID)
2238 		result += sizeof(u64);
2239 
2240 	if (type & PERF_SAMPLE_TIME)
2241 		result += sizeof(u64);
2242 
2243 	if (type & PERF_SAMPLE_ADDR)
2244 		result += sizeof(u64);
2245 
2246 	if (type & PERF_SAMPLE_ID)
2247 		result += sizeof(u64);
2248 
2249 	if (type & PERF_SAMPLE_STREAM_ID)
2250 		result += sizeof(u64);
2251 
2252 	if (type & PERF_SAMPLE_CPU)
2253 		result += sizeof(u64);
2254 
2255 	if (type & PERF_SAMPLE_PERIOD)
2256 		result += sizeof(u64);
2257 
2258 	if (type & PERF_SAMPLE_READ) {
2259 		result += sizeof(u64);
2260 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2261 			result += sizeof(u64);
2262 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2263 			result += sizeof(u64);
2264 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2265 		if (read_format & PERF_FORMAT_GROUP) {
2266 			sz = sample->read.group.nr *
2267 			     sizeof(struct sample_read_value);
2268 			result += sz;
2269 		} else {
2270 			result += sizeof(u64);
2271 		}
2272 	}
2273 
2274 	if (type & PERF_SAMPLE_CALLCHAIN) {
2275 		sz = (sample->callchain->nr + 1) * sizeof(u64);
2276 		result += sz;
2277 	}
2278 
2279 	if (type & PERF_SAMPLE_RAW) {
2280 		result += sizeof(u32);
2281 		result += sample->raw_size;
2282 	}
2283 
2284 	if (type & PERF_SAMPLE_BRANCH_STACK) {
2285 		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2286 		sz += sizeof(u64);
2287 		result += sz;
2288 	}
2289 
2290 	if (type & PERF_SAMPLE_REGS_USER) {
2291 		if (sample->user_regs.abi) {
2292 			result += sizeof(u64);
2293 			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2294 			result += sz;
2295 		} else {
2296 			result += sizeof(u64);
2297 		}
2298 	}
2299 
2300 	if (type & PERF_SAMPLE_STACK_USER) {
2301 		sz = sample->user_stack.size;
2302 		result += sizeof(u64);
2303 		if (sz) {
2304 			result += sz;
2305 			result += sizeof(u64);
2306 		}
2307 	}
2308 
2309 	if (type & PERF_SAMPLE_WEIGHT)
2310 		result += sizeof(u64);
2311 
2312 	if (type & PERF_SAMPLE_DATA_SRC)
2313 		result += sizeof(u64);
2314 
2315 	if (type & PERF_SAMPLE_TRANSACTION)
2316 		result += sizeof(u64);
2317 
2318 	if (type & PERF_SAMPLE_REGS_INTR) {
2319 		if (sample->intr_regs.abi) {
2320 			result += sizeof(u64);
2321 			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2322 			result += sz;
2323 		} else {
2324 			result += sizeof(u64);
2325 		}
2326 	}
2327 
2328 	if (type & PERF_SAMPLE_PHYS_ADDR)
2329 		result += sizeof(u64);
2330 
2331 	return result;
2332 }
2333 
2334 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2335 				  u64 read_format,
2336 				  const struct perf_sample *sample,
2337 				  bool swapped)
2338 {
2339 	u64 *array;
2340 	size_t sz;
2341 	/*
2342 	 * used for cross-endian analysis. See git commit 65014ab3
2343 	 * for why this goofiness is needed.
2344 	 */
2345 	union u64_swap u;
2346 
2347 	array = event->sample.array;
2348 
2349 	if (type & PERF_SAMPLE_IDENTIFIER) {
2350 		*array = sample->id;
2351 		array++;
2352 	}
2353 
2354 	if (type & PERF_SAMPLE_IP) {
2355 		*array = sample->ip;
2356 		array++;
2357 	}
2358 
2359 	if (type & PERF_SAMPLE_TID) {
2360 		u.val32[0] = sample->pid;
2361 		u.val32[1] = sample->tid;
2362 		if (swapped) {
2363 			/*
2364 			 * Inverse of what is done in perf_evsel__parse_sample
2365 			 */
2366 			u.val32[0] = bswap_32(u.val32[0]);
2367 			u.val32[1] = bswap_32(u.val32[1]);
2368 			u.val64 = bswap_64(u.val64);
2369 		}
2370 
2371 		*array = u.val64;
2372 		array++;
2373 	}
2374 
2375 	if (type & PERF_SAMPLE_TIME) {
2376 		*array = sample->time;
2377 		array++;
2378 	}
2379 
2380 	if (type & PERF_SAMPLE_ADDR) {
2381 		*array = sample->addr;
2382 		array++;
2383 	}
2384 
2385 	if (type & PERF_SAMPLE_ID) {
2386 		*array = sample->id;
2387 		array++;
2388 	}
2389 
2390 	if (type & PERF_SAMPLE_STREAM_ID) {
2391 		*array = sample->stream_id;
2392 		array++;
2393 	}
2394 
2395 	if (type & PERF_SAMPLE_CPU) {
2396 		u.val32[0] = sample->cpu;
2397 		if (swapped) {
2398 			/*
2399 			 * Inverse of what is done in perf_evsel__parse_sample
2400 			 */
2401 			u.val32[0] = bswap_32(u.val32[0]);
2402 			u.val64 = bswap_64(u.val64);
2403 		}
2404 		*array = u.val64;
2405 		array++;
2406 	}
2407 
2408 	if (type & PERF_SAMPLE_PERIOD) {
2409 		*array = sample->period;
2410 		array++;
2411 	}
2412 
2413 	if (type & PERF_SAMPLE_READ) {
2414 		if (read_format & PERF_FORMAT_GROUP)
2415 			*array = sample->read.group.nr;
2416 		else
2417 			*array = sample->read.one.value;
2418 		array++;
2419 
2420 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2421 			*array = sample->read.time_enabled;
2422 			array++;
2423 		}
2424 
2425 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2426 			*array = sample->read.time_running;
2427 			array++;
2428 		}
2429 
2430 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2431 		if (read_format & PERF_FORMAT_GROUP) {
2432 			sz = sample->read.group.nr *
2433 			     sizeof(struct sample_read_value);
2434 			memcpy(array, sample->read.group.values, sz);
2435 			array = (void *)array + sz;
2436 		} else {
2437 			*array = sample->read.one.id;
2438 			array++;
2439 		}
2440 	}
2441 
2442 	if (type & PERF_SAMPLE_CALLCHAIN) {
2443 		sz = (sample->callchain->nr + 1) * sizeof(u64);
2444 		memcpy(array, sample->callchain, sz);
2445 		array = (void *)array + sz;
2446 	}
2447 
2448 	if (type & PERF_SAMPLE_RAW) {
2449 		u.val32[0] = sample->raw_size;
2450 		if (WARN_ONCE(swapped,
2451 			      "Endianness of raw data not corrected!\n")) {
2452 			/*
2453 			 * Inverse of what is done in perf_evsel__parse_sample
2454 			 */
2455 			u.val32[0] = bswap_32(u.val32[0]);
2456 			u.val32[1] = bswap_32(u.val32[1]);
2457 			u.val64 = bswap_64(u.val64);
2458 		}
2459 		*array = u.val64;
2460 		array = (void *)array + sizeof(u32);
2461 
2462 		memcpy(array, sample->raw_data, sample->raw_size);
2463 		array = (void *)array + sample->raw_size;
2464 	}
2465 
2466 	if (type & PERF_SAMPLE_BRANCH_STACK) {
2467 		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2468 		sz += sizeof(u64);
2469 		memcpy(array, sample->branch_stack, sz);
2470 		array = (void *)array + sz;
2471 	}
2472 
2473 	if (type & PERF_SAMPLE_REGS_USER) {
2474 		if (sample->user_regs.abi) {
2475 			*array++ = sample->user_regs.abi;
2476 			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2477 			memcpy(array, sample->user_regs.regs, sz);
2478 			array = (void *)array + sz;
2479 		} else {
2480 			*array++ = 0;
2481 		}
2482 	}
2483 
2484 	if (type & PERF_SAMPLE_STACK_USER) {
2485 		sz = sample->user_stack.size;
2486 		*array++ = sz;
2487 		if (sz) {
2488 			memcpy(array, sample->user_stack.data, sz);
2489 			array = (void *)array + sz;
2490 			*array++ = sz;
2491 		}
2492 	}
2493 
2494 	if (type & PERF_SAMPLE_WEIGHT) {
2495 		*array = sample->weight;
2496 		array++;
2497 	}
2498 
2499 	if (type & PERF_SAMPLE_DATA_SRC) {
2500 		*array = sample->data_src;
2501 		array++;
2502 	}
2503 
2504 	if (type & PERF_SAMPLE_TRANSACTION) {
2505 		*array = sample->transaction;
2506 		array++;
2507 	}
2508 
2509 	if (type & PERF_SAMPLE_REGS_INTR) {
2510 		if (sample->intr_regs.abi) {
2511 			*array++ = sample->intr_regs.abi;
2512 			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2513 			memcpy(array, sample->intr_regs.regs, sz);
2514 			array = (void *)array + sz;
2515 		} else {
2516 			*array++ = 0;
2517 		}
2518 	}
2519 
2520 	if (type & PERF_SAMPLE_PHYS_ADDR) {
2521 		*array = sample->phys_addr;
2522 		array++;
2523 	}
2524 
2525 	return 0;
2526 }
2527 
2528 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2529 {
2530 	return pevent_find_field(evsel->tp_format, name);
2531 }
2532 
2533 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2534 			 const char *name)
2535 {
2536 	struct format_field *field = perf_evsel__field(evsel, name);
2537 	int offset;
2538 
2539 	if (!field)
2540 		return NULL;
2541 
2542 	offset = field->offset;
2543 
2544 	if (field->flags & FIELD_IS_DYNAMIC) {
2545 		offset = *(int *)(sample->raw_data + field->offset);
2546 		offset &= 0xffff;
2547 	}
2548 
2549 	return sample->raw_data + offset;
2550 }
2551 
2552 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2553 			 bool needs_swap)
2554 {
2555 	u64 value;
2556 	void *ptr = sample->raw_data + field->offset;
2557 
2558 	switch (field->size) {
2559 	case 1:
2560 		return *(u8 *)ptr;
2561 	case 2:
2562 		value = *(u16 *)ptr;
2563 		break;
2564 	case 4:
2565 		value = *(u32 *)ptr;
2566 		break;
2567 	case 8:
2568 		memcpy(&value, ptr, sizeof(u64));
2569 		break;
2570 	default:
2571 		return 0;
2572 	}
2573 
2574 	if (!needs_swap)
2575 		return value;
2576 
2577 	switch (field->size) {
2578 	case 2:
2579 		return bswap_16(value);
2580 	case 4:
2581 		return bswap_32(value);
2582 	case 8:
2583 		return bswap_64(value);
2584 	default:
2585 		return 0;
2586 	}
2587 
2588 	return 0;
2589 }
2590 
2591 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2592 		       const char *name)
2593 {
2594 	struct format_field *field = perf_evsel__field(evsel, name);
2595 
2596 	if (!field)
2597 		return 0;
2598 
2599 	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2600 }
2601 
2602 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2603 			  char *msg, size_t msgsize)
2604 {
2605 	int paranoid;
2606 
2607 	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2608 	    evsel->attr.type   == PERF_TYPE_HARDWARE &&
2609 	    evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2610 		/*
2611 		 * If it's cycles then fall back to hrtimer based
2612 		 * cpu-clock-tick sw counter, which is always available even if
2613 		 * no PMU support.
2614 		 *
2615 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2616 		 * b0a873e).
2617 		 */
2618 		scnprintf(msg, msgsize, "%s",
2619 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2620 
2621 		evsel->attr.type   = PERF_TYPE_SOFTWARE;
2622 		evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2623 
2624 		zfree(&evsel->name);
2625 		return true;
2626 	} else if (err == EACCES && !evsel->attr.exclude_kernel &&
2627 		   (paranoid = perf_event_paranoid()) > 1) {
2628 		const char *name = perf_evsel__name(evsel);
2629 		char *new_name;
2630 
2631 		if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2632 			return false;
2633 
2634 		if (evsel->name)
2635 			free(evsel->name);
2636 		evsel->name = new_name;
2637 		scnprintf(msg, msgsize,
2638 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2639 		evsel->attr.exclude_kernel = 1;
2640 
2641 		return true;
2642 	}
2643 
2644 	return false;
2645 }
2646 
2647 static bool find_process(const char *name)
2648 {
2649 	size_t len = strlen(name);
2650 	DIR *dir;
2651 	struct dirent *d;
2652 	int ret = -1;
2653 
2654 	dir = opendir(procfs__mountpoint());
2655 	if (!dir)
2656 		return false;
2657 
2658 	/* Walk through the directory. */
2659 	while (ret && (d = readdir(dir)) != NULL) {
2660 		char path[PATH_MAX];
2661 		char *data;
2662 		size_t size;
2663 
2664 		if ((d->d_type != DT_DIR) ||
2665 		     !strcmp(".", d->d_name) ||
2666 		     !strcmp("..", d->d_name))
2667 			continue;
2668 
2669 		scnprintf(path, sizeof(path), "%s/%s/comm",
2670 			  procfs__mountpoint(), d->d_name);
2671 
2672 		if (filename__read_str(path, &data, &size))
2673 			continue;
2674 
2675 		ret = strncmp(name, data, len);
2676 		free(data);
2677 	}
2678 
2679 	closedir(dir);
2680 	return ret ? false : true;
2681 }
2682 
2683 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2684 			      int err, char *msg, size_t size)
2685 {
2686 	char sbuf[STRERR_BUFSIZE];
2687 	int printed = 0;
2688 
2689 	switch (err) {
2690 	case EPERM:
2691 	case EACCES:
2692 		if (err == EPERM)
2693 			printed = scnprintf(msg, size,
2694 				"No permission to enable %s event.\n\n",
2695 				perf_evsel__name(evsel));
2696 
2697 		return scnprintf(msg + printed, size - printed,
2698 		 "You may not have permission to collect %sstats.\n\n"
2699 		 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2700 		 "which controls use of the performance events system by\n"
2701 		 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2702 		 "The current value is %d:\n\n"
2703 		 "  -1: Allow use of (almost) all events by all users\n"
2704 		 "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2705 		 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2706 		 "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2707 		 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2708 		 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2709 		 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2710 		 "	kernel.perf_event_paranoid = -1\n" ,
2711 				 target->system_wide ? "system-wide " : "",
2712 				 perf_event_paranoid());
2713 	case ENOENT:
2714 		return scnprintf(msg, size, "The %s event is not supported.",
2715 				 perf_evsel__name(evsel));
2716 	case EMFILE:
2717 		return scnprintf(msg, size, "%s",
2718 			 "Too many events are opened.\n"
2719 			 "Probably the maximum number of open file descriptors has been reached.\n"
2720 			 "Hint: Try again after reducing the number of events.\n"
2721 			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2722 	case ENOMEM:
2723 		if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2724 		    access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2725 			return scnprintf(msg, size,
2726 					 "Not enough memory to setup event with callchain.\n"
2727 					 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2728 					 "Hint: Current value: %d", sysctl_perf_event_max_stack);
2729 		break;
2730 	case ENODEV:
2731 		if (target->cpu_list)
2732 			return scnprintf(msg, size, "%s",
2733 	 "No such device - did you specify an out-of-range profile CPU?");
2734 		break;
2735 	case EOPNOTSUPP:
2736 		if (evsel->attr.sample_period != 0)
2737 			return scnprintf(msg, size, "%s",
2738 	"PMU Hardware doesn't support sampling/overflow-interrupts.");
2739 		if (evsel->attr.precise_ip)
2740 			return scnprintf(msg, size, "%s",
2741 	"\'precise\' request may not be supported. Try removing 'p' modifier.");
2742 #if defined(__i386__) || defined(__x86_64__)
2743 		if (evsel->attr.type == PERF_TYPE_HARDWARE)
2744 			return scnprintf(msg, size, "%s",
2745 	"No hardware sampling interrupt available.\n"
2746 	"No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2747 #endif
2748 		break;
2749 	case EBUSY:
2750 		if (find_process("oprofiled"))
2751 			return scnprintf(msg, size,
2752 	"The PMU counters are busy/taken by another profiler.\n"
2753 	"We found oprofile daemon running, please stop it and try again.");
2754 		break;
2755 	case EINVAL:
2756 		if (evsel->attr.write_backward && perf_missing_features.write_backward)
2757 			return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2758 		if (perf_missing_features.clockid)
2759 			return scnprintf(msg, size, "clockid feature not supported.");
2760 		if (perf_missing_features.clockid_wrong)
2761 			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2762 		break;
2763 	default:
2764 		break;
2765 	}
2766 
2767 	return scnprintf(msg, size,
2768 	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2769 	"/bin/dmesg may provide additional information.\n"
2770 	"No CONFIG_PERF_EVENTS=y kernel support configured?",
2771 			 err, str_error_r(err, sbuf, sizeof(sbuf)),
2772 			 perf_evsel__name(evsel));
2773 }
2774 
2775 char *perf_evsel__env_arch(struct perf_evsel *evsel)
2776 {
2777 	if (evsel && evsel->evlist && evsel->evlist->env)
2778 		return evsel->evlist->env->arch;
2779 	return NULL;
2780 }
2781 
2782 char *perf_evsel__env_cpuid(struct perf_evsel *evsel)
2783 {
2784 	if (evsel && evsel->evlist && evsel->evlist->env)
2785 		return evsel->evlist->env->cpuid;
2786 	return NULL;
2787 }
2788