xref: /linux/tools/perf/util/evsel.c (revision 005438a8eef063495ac059d128eea71b58de50e5)
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 <linux/bitops.h>
12 #include <api/fs/debugfs.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <sys/resource.h>
17 #include "asm/bug.h"
18 #include "callchain.h"
19 #include "cgroup.h"
20 #include "evsel.h"
21 #include "evlist.h"
22 #include "util.h"
23 #include "cpumap.h"
24 #include "thread_map.h"
25 #include "target.h"
26 #include "perf_regs.h"
27 #include "debug.h"
28 #include "trace-event.h"
29 #include "stat.h"
30 
31 static struct {
32 	bool sample_id_all;
33 	bool exclude_guest;
34 	bool mmap2;
35 	bool cloexec;
36 	bool clockid;
37 	bool clockid_wrong;
38 } perf_missing_features;
39 
40 static clockid_t clockid;
41 
42 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
43 {
44 	return 0;
45 }
46 
47 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
48 {
49 }
50 
51 static struct {
52 	size_t	size;
53 	int	(*init)(struct perf_evsel *evsel);
54 	void	(*fini)(struct perf_evsel *evsel);
55 } perf_evsel__object = {
56 	.size = sizeof(struct perf_evsel),
57 	.init = perf_evsel__no_extra_init,
58 	.fini = perf_evsel__no_extra_fini,
59 };
60 
61 int perf_evsel__object_config(size_t object_size,
62 			      int (*init)(struct perf_evsel *evsel),
63 			      void (*fini)(struct perf_evsel *evsel))
64 {
65 
66 	if (object_size == 0)
67 		goto set_methods;
68 
69 	if (perf_evsel__object.size > object_size)
70 		return -EINVAL;
71 
72 	perf_evsel__object.size = object_size;
73 
74 set_methods:
75 	if (init != NULL)
76 		perf_evsel__object.init = init;
77 
78 	if (fini != NULL)
79 		perf_evsel__object.fini = fini;
80 
81 	return 0;
82 }
83 
84 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
85 
86 int __perf_evsel__sample_size(u64 sample_type)
87 {
88 	u64 mask = sample_type & PERF_SAMPLE_MASK;
89 	int size = 0;
90 	int i;
91 
92 	for (i = 0; i < 64; i++) {
93 		if (mask & (1ULL << i))
94 			size++;
95 	}
96 
97 	size *= sizeof(u64);
98 
99 	return size;
100 }
101 
102 /**
103  * __perf_evsel__calc_id_pos - calculate id_pos.
104  * @sample_type: sample type
105  *
106  * This function returns the position of the event id (PERF_SAMPLE_ID or
107  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
108  * sample_event.
109  */
110 static int __perf_evsel__calc_id_pos(u64 sample_type)
111 {
112 	int idx = 0;
113 
114 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
115 		return 0;
116 
117 	if (!(sample_type & PERF_SAMPLE_ID))
118 		return -1;
119 
120 	if (sample_type & PERF_SAMPLE_IP)
121 		idx += 1;
122 
123 	if (sample_type & PERF_SAMPLE_TID)
124 		idx += 1;
125 
126 	if (sample_type & PERF_SAMPLE_TIME)
127 		idx += 1;
128 
129 	if (sample_type & PERF_SAMPLE_ADDR)
130 		idx += 1;
131 
132 	return idx;
133 }
134 
135 /**
136  * __perf_evsel__calc_is_pos - calculate is_pos.
137  * @sample_type: sample type
138  *
139  * This function returns the position (counting backwards) of the event id
140  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
141  * sample_id_all is used there is an id sample appended to non-sample events.
142  */
143 static int __perf_evsel__calc_is_pos(u64 sample_type)
144 {
145 	int idx = 1;
146 
147 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
148 		return 1;
149 
150 	if (!(sample_type & PERF_SAMPLE_ID))
151 		return -1;
152 
153 	if (sample_type & PERF_SAMPLE_CPU)
154 		idx += 1;
155 
156 	if (sample_type & PERF_SAMPLE_STREAM_ID)
157 		idx += 1;
158 
159 	return idx;
160 }
161 
162 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
163 {
164 	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
165 	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
166 }
167 
168 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
169 				  enum perf_event_sample_format bit)
170 {
171 	if (!(evsel->attr.sample_type & bit)) {
172 		evsel->attr.sample_type |= bit;
173 		evsel->sample_size += sizeof(u64);
174 		perf_evsel__calc_id_pos(evsel);
175 	}
176 }
177 
178 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
179 				    enum perf_event_sample_format bit)
180 {
181 	if (evsel->attr.sample_type & bit) {
182 		evsel->attr.sample_type &= ~bit;
183 		evsel->sample_size -= sizeof(u64);
184 		perf_evsel__calc_id_pos(evsel);
185 	}
186 }
187 
188 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
189 			       bool can_sample_identifier)
190 {
191 	if (can_sample_identifier) {
192 		perf_evsel__reset_sample_bit(evsel, ID);
193 		perf_evsel__set_sample_bit(evsel, IDENTIFIER);
194 	} else {
195 		perf_evsel__set_sample_bit(evsel, ID);
196 	}
197 	evsel->attr.read_format |= PERF_FORMAT_ID;
198 }
199 
200 void perf_evsel__init(struct perf_evsel *evsel,
201 		      struct perf_event_attr *attr, int idx)
202 {
203 	evsel->idx	   = idx;
204 	evsel->tracking	   = !idx;
205 	evsel->attr	   = *attr;
206 	evsel->leader	   = evsel;
207 	evsel->unit	   = "";
208 	evsel->scale	   = 1.0;
209 	INIT_LIST_HEAD(&evsel->node);
210 	perf_evsel__object.init(evsel);
211 	evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
212 	perf_evsel__calc_id_pos(evsel);
213 }
214 
215 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
216 {
217 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
218 
219 	if (evsel != NULL)
220 		perf_evsel__init(evsel, attr, idx);
221 
222 	return evsel;
223 }
224 
225 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
226 {
227 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
228 
229 	if (evsel != NULL) {
230 		struct perf_event_attr attr = {
231 			.type	       = PERF_TYPE_TRACEPOINT,
232 			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
233 					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
234 		};
235 
236 		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
237 			goto out_free;
238 
239 		evsel->tp_format = trace_event__tp_format(sys, name);
240 		if (evsel->tp_format == NULL)
241 			goto out_free;
242 
243 		event_attr_init(&attr);
244 		attr.config = evsel->tp_format->id;
245 		attr.sample_period = 1;
246 		perf_evsel__init(evsel, &attr, idx);
247 	}
248 
249 	return evsel;
250 
251 out_free:
252 	zfree(&evsel->name);
253 	free(evsel);
254 	return NULL;
255 }
256 
257 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
258 	"cycles",
259 	"instructions",
260 	"cache-references",
261 	"cache-misses",
262 	"branches",
263 	"branch-misses",
264 	"bus-cycles",
265 	"stalled-cycles-frontend",
266 	"stalled-cycles-backend",
267 	"ref-cycles",
268 };
269 
270 static const char *__perf_evsel__hw_name(u64 config)
271 {
272 	if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
273 		return perf_evsel__hw_names[config];
274 
275 	return "unknown-hardware";
276 }
277 
278 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
279 {
280 	int colon = 0, r = 0;
281 	struct perf_event_attr *attr = &evsel->attr;
282 	bool exclude_guest_default = false;
283 
284 #define MOD_PRINT(context, mod)	do {					\
285 		if (!attr->exclude_##context) {				\
286 			if (!colon) colon = ++r;			\
287 			r += scnprintf(bf + r, size - r, "%c", mod);	\
288 		} } while(0)
289 
290 	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
291 		MOD_PRINT(kernel, 'k');
292 		MOD_PRINT(user, 'u');
293 		MOD_PRINT(hv, 'h');
294 		exclude_guest_default = true;
295 	}
296 
297 	if (attr->precise_ip) {
298 		if (!colon)
299 			colon = ++r;
300 		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
301 		exclude_guest_default = true;
302 	}
303 
304 	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
305 		MOD_PRINT(host, 'H');
306 		MOD_PRINT(guest, 'G');
307 	}
308 #undef MOD_PRINT
309 	if (colon)
310 		bf[colon - 1] = ':';
311 	return r;
312 }
313 
314 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
315 {
316 	int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
317 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
318 }
319 
320 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
321 	"cpu-clock",
322 	"task-clock",
323 	"page-faults",
324 	"context-switches",
325 	"cpu-migrations",
326 	"minor-faults",
327 	"major-faults",
328 	"alignment-faults",
329 	"emulation-faults",
330 	"dummy",
331 };
332 
333 static const char *__perf_evsel__sw_name(u64 config)
334 {
335 	if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
336 		return perf_evsel__sw_names[config];
337 	return "unknown-software";
338 }
339 
340 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
341 {
342 	int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
343 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
344 }
345 
346 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
347 {
348 	int r;
349 
350 	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
351 
352 	if (type & HW_BREAKPOINT_R)
353 		r += scnprintf(bf + r, size - r, "r");
354 
355 	if (type & HW_BREAKPOINT_W)
356 		r += scnprintf(bf + r, size - r, "w");
357 
358 	if (type & HW_BREAKPOINT_X)
359 		r += scnprintf(bf + r, size - r, "x");
360 
361 	return r;
362 }
363 
364 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
365 {
366 	struct perf_event_attr *attr = &evsel->attr;
367 	int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
368 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
369 }
370 
371 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
372 				[PERF_EVSEL__MAX_ALIASES] = {
373  { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
374  { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
375  { "LLC",	"L2",							},
376  { "dTLB",	"d-tlb",	"Data-TLB",				},
377  { "iTLB",	"i-tlb",	"Instruction-TLB",			},
378  { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
379  { "node",								},
380 };
381 
382 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
383 				   [PERF_EVSEL__MAX_ALIASES] = {
384  { "load",	"loads",	"read",					},
385  { "store",	"stores",	"write",				},
386  { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
387 };
388 
389 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
390 				       [PERF_EVSEL__MAX_ALIASES] = {
391  { "refs",	"Reference",	"ops",		"access",		},
392  { "misses",	"miss",							},
393 };
394 
395 #define C(x)		PERF_COUNT_HW_CACHE_##x
396 #define CACHE_READ	(1 << C(OP_READ))
397 #define CACHE_WRITE	(1 << C(OP_WRITE))
398 #define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
399 #define COP(x)		(1 << x)
400 
401 /*
402  * cache operartion stat
403  * L1I : Read and prefetch only
404  * ITLB and BPU : Read-only
405  */
406 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
407  [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
408  [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
409  [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
410  [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
411  [C(ITLB)]	= (CACHE_READ),
412  [C(BPU)]	= (CACHE_READ),
413  [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
414 };
415 
416 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
417 {
418 	if (perf_evsel__hw_cache_stat[type] & COP(op))
419 		return true;	/* valid */
420 	else
421 		return false;	/* invalid */
422 }
423 
424 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
425 					    char *bf, size_t size)
426 {
427 	if (result) {
428 		return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
429 				 perf_evsel__hw_cache_op[op][0],
430 				 perf_evsel__hw_cache_result[result][0]);
431 	}
432 
433 	return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
434 			 perf_evsel__hw_cache_op[op][1]);
435 }
436 
437 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
438 {
439 	u8 op, result, type = (config >>  0) & 0xff;
440 	const char *err = "unknown-ext-hardware-cache-type";
441 
442 	if (type > PERF_COUNT_HW_CACHE_MAX)
443 		goto out_err;
444 
445 	op = (config >>  8) & 0xff;
446 	err = "unknown-ext-hardware-cache-op";
447 	if (op > PERF_COUNT_HW_CACHE_OP_MAX)
448 		goto out_err;
449 
450 	result = (config >> 16) & 0xff;
451 	err = "unknown-ext-hardware-cache-result";
452 	if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
453 		goto out_err;
454 
455 	err = "invalid-cache";
456 	if (!perf_evsel__is_cache_op_valid(type, op))
457 		goto out_err;
458 
459 	return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
460 out_err:
461 	return scnprintf(bf, size, "%s", err);
462 }
463 
464 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
465 {
466 	int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
467 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
468 }
469 
470 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
471 {
472 	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
473 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
474 }
475 
476 const char *perf_evsel__name(struct perf_evsel *evsel)
477 {
478 	char bf[128];
479 
480 	if (evsel->name)
481 		return evsel->name;
482 
483 	switch (evsel->attr.type) {
484 	case PERF_TYPE_RAW:
485 		perf_evsel__raw_name(evsel, bf, sizeof(bf));
486 		break;
487 
488 	case PERF_TYPE_HARDWARE:
489 		perf_evsel__hw_name(evsel, bf, sizeof(bf));
490 		break;
491 
492 	case PERF_TYPE_HW_CACHE:
493 		perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
494 		break;
495 
496 	case PERF_TYPE_SOFTWARE:
497 		perf_evsel__sw_name(evsel, bf, sizeof(bf));
498 		break;
499 
500 	case PERF_TYPE_TRACEPOINT:
501 		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
502 		break;
503 
504 	case PERF_TYPE_BREAKPOINT:
505 		perf_evsel__bp_name(evsel, bf, sizeof(bf));
506 		break;
507 
508 	default:
509 		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
510 			  evsel->attr.type);
511 		break;
512 	}
513 
514 	evsel->name = strdup(bf);
515 
516 	return evsel->name ?: "unknown";
517 }
518 
519 const char *perf_evsel__group_name(struct perf_evsel *evsel)
520 {
521 	return evsel->group_name ?: "anon group";
522 }
523 
524 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
525 {
526 	int ret;
527 	struct perf_evsel *pos;
528 	const char *group_name = perf_evsel__group_name(evsel);
529 
530 	ret = scnprintf(buf, size, "%s", group_name);
531 
532 	ret += scnprintf(buf + ret, size - ret, " { %s",
533 			 perf_evsel__name(evsel));
534 
535 	for_each_group_member(pos, evsel)
536 		ret += scnprintf(buf + ret, size - ret, ", %s",
537 				 perf_evsel__name(pos));
538 
539 	ret += scnprintf(buf + ret, size - ret, " }");
540 
541 	return ret;
542 }
543 
544 static void
545 perf_evsel__config_callgraph(struct perf_evsel *evsel,
546 			     struct record_opts *opts)
547 {
548 	bool function = perf_evsel__is_function_event(evsel);
549 	struct perf_event_attr *attr = &evsel->attr;
550 
551 	perf_evsel__set_sample_bit(evsel, CALLCHAIN);
552 
553 	if (callchain_param.record_mode == CALLCHAIN_LBR) {
554 		if (!opts->branch_stack) {
555 			if (attr->exclude_user) {
556 				pr_warning("LBR callstack option is only available "
557 					   "to get user callchain information. "
558 					   "Falling back to framepointers.\n");
559 			} else {
560 				perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
561 				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
562 							PERF_SAMPLE_BRANCH_CALL_STACK;
563 			}
564 		} else
565 			 pr_warning("Cannot use LBR callstack with branch stack. "
566 				    "Falling back to framepointers.\n");
567 	}
568 
569 	if (callchain_param.record_mode == CALLCHAIN_DWARF) {
570 		if (!function) {
571 			perf_evsel__set_sample_bit(evsel, REGS_USER);
572 			perf_evsel__set_sample_bit(evsel, STACK_USER);
573 			attr->sample_regs_user = PERF_REGS_MASK;
574 			attr->sample_stack_user = callchain_param.dump_size;
575 			attr->exclude_callchain_user = 1;
576 		} else {
577 			pr_info("Cannot use DWARF unwind for function trace event,"
578 				" falling back to framepointers.\n");
579 		}
580 	}
581 
582 	if (function) {
583 		pr_info("Disabling user space callchains for function trace event.\n");
584 		attr->exclude_callchain_user = 1;
585 	}
586 }
587 
588 /*
589  * The enable_on_exec/disabled value strategy:
590  *
591  *  1) For any type of traced program:
592  *    - all independent events and group leaders are disabled
593  *    - all group members are enabled
594  *
595  *     Group members are ruled by group leaders. They need to
596  *     be enabled, because the group scheduling relies on that.
597  *
598  *  2) For traced programs executed by perf:
599  *     - all independent events and group leaders have
600  *       enable_on_exec set
601  *     - we don't specifically enable or disable any event during
602  *       the record command
603  *
604  *     Independent events and group leaders are initially disabled
605  *     and get enabled by exec. Group members are ruled by group
606  *     leaders as stated in 1).
607  *
608  *  3) For traced programs attached by perf (pid/tid):
609  *     - we specifically enable or disable all events during
610  *       the record command
611  *
612  *     When attaching events to already running traced we
613  *     enable/disable events specifically, as there's no
614  *     initial traced exec call.
615  */
616 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
617 {
618 	struct perf_evsel *leader = evsel->leader;
619 	struct perf_event_attr *attr = &evsel->attr;
620 	int track = evsel->tracking;
621 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
622 
623 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
624 	attr->inherit	    = !opts->no_inherit;
625 
626 	perf_evsel__set_sample_bit(evsel, IP);
627 	perf_evsel__set_sample_bit(evsel, TID);
628 
629 	if (evsel->sample_read) {
630 		perf_evsel__set_sample_bit(evsel, READ);
631 
632 		/*
633 		 * We need ID even in case of single event, because
634 		 * PERF_SAMPLE_READ process ID specific data.
635 		 */
636 		perf_evsel__set_sample_id(evsel, false);
637 
638 		/*
639 		 * Apply group format only if we belong to group
640 		 * with more than one members.
641 		 */
642 		if (leader->nr_members > 1) {
643 			attr->read_format |= PERF_FORMAT_GROUP;
644 			attr->inherit = 0;
645 		}
646 	}
647 
648 	/*
649 	 * We default some events to have a default interval. But keep
650 	 * it a weak assumption overridable by the user.
651 	 */
652 	if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
653 				     opts->user_interval != ULLONG_MAX)) {
654 		if (opts->freq) {
655 			perf_evsel__set_sample_bit(evsel, PERIOD);
656 			attr->freq		= 1;
657 			attr->sample_freq	= opts->freq;
658 		} else {
659 			attr->sample_period = opts->default_interval;
660 		}
661 	}
662 
663 	/*
664 	 * Disable sampling for all group members other
665 	 * than leader in case leader 'leads' the sampling.
666 	 */
667 	if ((leader != evsel) && leader->sample_read) {
668 		attr->sample_freq   = 0;
669 		attr->sample_period = 0;
670 	}
671 
672 	if (opts->no_samples)
673 		attr->sample_freq = 0;
674 
675 	if (opts->inherit_stat)
676 		attr->inherit_stat = 1;
677 
678 	if (opts->sample_address) {
679 		perf_evsel__set_sample_bit(evsel, ADDR);
680 		attr->mmap_data = track;
681 	}
682 
683 	/*
684 	 * We don't allow user space callchains for  function trace
685 	 * event, due to issues with page faults while tracing page
686 	 * fault handler and its overall trickiness nature.
687 	 */
688 	if (perf_evsel__is_function_event(evsel))
689 		evsel->attr.exclude_callchain_user = 1;
690 
691 	if (callchain_param.enabled && !evsel->no_aux_samples)
692 		perf_evsel__config_callgraph(evsel, opts);
693 
694 	if (opts->sample_intr_regs) {
695 		attr->sample_regs_intr = PERF_REGS_MASK;
696 		perf_evsel__set_sample_bit(evsel, REGS_INTR);
697 	}
698 
699 	if (target__has_cpu(&opts->target))
700 		perf_evsel__set_sample_bit(evsel, CPU);
701 
702 	if (opts->period)
703 		perf_evsel__set_sample_bit(evsel, PERIOD);
704 
705 	/*
706 	 * When the user explicitely disabled time don't force it here.
707 	 */
708 	if (opts->sample_time &&
709 	    (!perf_missing_features.sample_id_all &&
710 	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
711 	     opts->sample_time_set)))
712 		perf_evsel__set_sample_bit(evsel, TIME);
713 
714 	if (opts->raw_samples && !evsel->no_aux_samples) {
715 		perf_evsel__set_sample_bit(evsel, TIME);
716 		perf_evsel__set_sample_bit(evsel, RAW);
717 		perf_evsel__set_sample_bit(evsel, CPU);
718 	}
719 
720 	if (opts->sample_address)
721 		perf_evsel__set_sample_bit(evsel, DATA_SRC);
722 
723 	if (opts->no_buffering) {
724 		attr->watermark = 0;
725 		attr->wakeup_events = 1;
726 	}
727 	if (opts->branch_stack && !evsel->no_aux_samples) {
728 		perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
729 		attr->branch_sample_type = opts->branch_stack;
730 	}
731 
732 	if (opts->sample_weight)
733 		perf_evsel__set_sample_bit(evsel, WEIGHT);
734 
735 	attr->task  = track;
736 	attr->mmap  = track;
737 	attr->mmap2 = track && !perf_missing_features.mmap2;
738 	attr->comm  = track;
739 
740 	if (opts->sample_transaction)
741 		perf_evsel__set_sample_bit(evsel, TRANSACTION);
742 
743 	if (opts->running_time) {
744 		evsel->attr.read_format |=
745 			PERF_FORMAT_TOTAL_TIME_ENABLED |
746 			PERF_FORMAT_TOTAL_TIME_RUNNING;
747 	}
748 
749 	/*
750 	 * XXX see the function comment above
751 	 *
752 	 * Disabling only independent events or group leaders,
753 	 * keeping group members enabled.
754 	 */
755 	if (perf_evsel__is_group_leader(evsel))
756 		attr->disabled = 1;
757 
758 	/*
759 	 * Setting enable_on_exec for independent events and
760 	 * group leaders for traced executed by perf.
761 	 */
762 	if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
763 		!opts->initial_delay)
764 		attr->enable_on_exec = 1;
765 
766 	if (evsel->immediate) {
767 		attr->disabled = 0;
768 		attr->enable_on_exec = 0;
769 	}
770 
771 	clockid = opts->clockid;
772 	if (opts->use_clockid) {
773 		attr->use_clockid = 1;
774 		attr->clockid = opts->clockid;
775 	}
776 }
777 
778 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
779 {
780 	int cpu, thread;
781 
782 	if (evsel->system_wide)
783 		nthreads = 1;
784 
785 	evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
786 
787 	if (evsel->fd) {
788 		for (cpu = 0; cpu < ncpus; cpu++) {
789 			for (thread = 0; thread < nthreads; thread++) {
790 				FD(evsel, cpu, thread) = -1;
791 			}
792 		}
793 	}
794 
795 	return evsel->fd != NULL ? 0 : -ENOMEM;
796 }
797 
798 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
799 			  int ioc,  void *arg)
800 {
801 	int cpu, thread;
802 
803 	if (evsel->system_wide)
804 		nthreads = 1;
805 
806 	for (cpu = 0; cpu < ncpus; cpu++) {
807 		for (thread = 0; thread < nthreads; thread++) {
808 			int fd = FD(evsel, cpu, thread),
809 			    err = ioctl(fd, ioc, arg);
810 
811 			if (err)
812 				return err;
813 		}
814 	}
815 
816 	return 0;
817 }
818 
819 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
820 			     const char *filter)
821 {
822 	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
823 				     PERF_EVENT_IOC_SET_FILTER,
824 				     (void *)filter);
825 }
826 
827 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
828 {
829 	char *new_filter = strdup(filter);
830 
831 	if (new_filter != NULL) {
832 		free(evsel->filter);
833 		evsel->filter = new_filter;
834 		return 0;
835 	}
836 
837 	return -1;
838 }
839 
840 int perf_evsel__append_filter(struct perf_evsel *evsel,
841 			      const char *op, const char *filter)
842 {
843 	char *new_filter;
844 
845 	if (evsel->filter == NULL)
846 		return perf_evsel__set_filter(evsel, filter);
847 
848 	if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
849 		free(evsel->filter);
850 		evsel->filter = new_filter;
851 		return 0;
852 	}
853 
854 	return -1;
855 }
856 
857 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
858 {
859 	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
860 				     PERF_EVENT_IOC_ENABLE,
861 				     0);
862 }
863 
864 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
865 {
866 	if (ncpus == 0 || nthreads == 0)
867 		return 0;
868 
869 	if (evsel->system_wide)
870 		nthreads = 1;
871 
872 	evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
873 	if (evsel->sample_id == NULL)
874 		return -ENOMEM;
875 
876 	evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
877 	if (evsel->id == NULL) {
878 		xyarray__delete(evsel->sample_id);
879 		evsel->sample_id = NULL;
880 		return -ENOMEM;
881 	}
882 
883 	return 0;
884 }
885 
886 static void perf_evsel__free_fd(struct perf_evsel *evsel)
887 {
888 	xyarray__delete(evsel->fd);
889 	evsel->fd = NULL;
890 }
891 
892 static void perf_evsel__free_id(struct perf_evsel *evsel)
893 {
894 	xyarray__delete(evsel->sample_id);
895 	evsel->sample_id = NULL;
896 	zfree(&evsel->id);
897 }
898 
899 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
900 {
901 	int cpu, thread;
902 
903 	if (evsel->system_wide)
904 		nthreads = 1;
905 
906 	for (cpu = 0; cpu < ncpus; cpu++)
907 		for (thread = 0; thread < nthreads; ++thread) {
908 			close(FD(evsel, cpu, thread));
909 			FD(evsel, cpu, thread) = -1;
910 		}
911 }
912 
913 void perf_evsel__exit(struct perf_evsel *evsel)
914 {
915 	assert(list_empty(&evsel->node));
916 	perf_evsel__free_fd(evsel);
917 	perf_evsel__free_id(evsel);
918 	close_cgroup(evsel->cgrp);
919 	cpu_map__put(evsel->cpus);
920 	thread_map__put(evsel->threads);
921 	zfree(&evsel->group_name);
922 	zfree(&evsel->name);
923 	perf_evsel__object.fini(evsel);
924 }
925 
926 void perf_evsel__delete(struct perf_evsel *evsel)
927 {
928 	perf_evsel__exit(evsel);
929 	free(evsel);
930 }
931 
932 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
933 				struct perf_counts_values *count)
934 {
935 	struct perf_counts_values tmp;
936 
937 	if (!evsel->prev_raw_counts)
938 		return;
939 
940 	if (cpu == -1) {
941 		tmp = evsel->prev_raw_counts->aggr;
942 		evsel->prev_raw_counts->aggr = *count;
943 	} else {
944 		tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
945 		*perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
946 	}
947 
948 	count->val = count->val - tmp.val;
949 	count->ena = count->ena - tmp.ena;
950 	count->run = count->run - tmp.run;
951 }
952 
953 void perf_counts_values__scale(struct perf_counts_values *count,
954 			       bool scale, s8 *pscaled)
955 {
956 	s8 scaled = 0;
957 
958 	if (scale) {
959 		if (count->run == 0) {
960 			scaled = -1;
961 			count->val = 0;
962 		} else if (count->run < count->ena) {
963 			scaled = 1;
964 			count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
965 		}
966 	} else
967 		count->ena = count->run = 0;
968 
969 	if (pscaled)
970 		*pscaled = scaled;
971 }
972 
973 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
974 		     struct perf_counts_values *count)
975 {
976 	memset(count, 0, sizeof(*count));
977 
978 	if (FD(evsel, cpu, thread) < 0)
979 		return -EINVAL;
980 
981 	if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
982 		return -errno;
983 
984 	return 0;
985 }
986 
987 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
988 			      int cpu, int thread, bool scale)
989 {
990 	struct perf_counts_values count;
991 	size_t nv = scale ? 3 : 1;
992 
993 	if (FD(evsel, cpu, thread) < 0)
994 		return -EINVAL;
995 
996 	if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
997 		return -ENOMEM;
998 
999 	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1000 		return -errno;
1001 
1002 	perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1003 	perf_counts_values__scale(&count, scale, NULL);
1004 	*perf_counts(evsel->counts, cpu, thread) = count;
1005 	return 0;
1006 }
1007 
1008 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1009 {
1010 	struct perf_evsel *leader = evsel->leader;
1011 	int fd;
1012 
1013 	if (perf_evsel__is_group_leader(evsel))
1014 		return -1;
1015 
1016 	/*
1017 	 * Leader must be already processed/open,
1018 	 * if not it's a bug.
1019 	 */
1020 	BUG_ON(!leader->fd);
1021 
1022 	fd = FD(leader, cpu, thread);
1023 	BUG_ON(fd == -1);
1024 
1025 	return fd;
1026 }
1027 
1028 struct bit_names {
1029 	int bit;
1030 	const char *name;
1031 };
1032 
1033 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1034 {
1035 	bool first_bit = true;
1036 	int i = 0;
1037 
1038 	do {
1039 		if (value & bits[i].bit) {
1040 			buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1041 			first_bit = false;
1042 		}
1043 	} while (bits[++i].name != NULL);
1044 }
1045 
1046 static void __p_sample_type(char *buf, size_t size, u64 value)
1047 {
1048 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1049 	struct bit_names bits[] = {
1050 		bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1051 		bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1052 		bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1053 		bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1054 		bit_name(IDENTIFIER), bit_name(REGS_INTR),
1055 		{ .name = NULL, }
1056 	};
1057 #undef bit_name
1058 	__p_bits(buf, size, value, bits);
1059 }
1060 
1061 static void __p_read_format(char *buf, size_t size, u64 value)
1062 {
1063 #define bit_name(n) { PERF_FORMAT_##n, #n }
1064 	struct bit_names bits[] = {
1065 		bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1066 		bit_name(ID), bit_name(GROUP),
1067 		{ .name = NULL, }
1068 	};
1069 #undef bit_name
1070 	__p_bits(buf, size, value, bits);
1071 }
1072 
1073 #define BUF_SIZE		1024
1074 
1075 #define p_hex(val)		snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1076 #define p_unsigned(val)		snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1077 #define p_signed(val)		snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1078 #define p_sample_type(val)	__p_sample_type(buf, BUF_SIZE, val)
1079 #define p_read_format(val)	__p_read_format(buf, BUF_SIZE, val)
1080 
1081 #define PRINT_ATTRn(_n, _f, _p)				\
1082 do {							\
1083 	if (attr->_f) {					\
1084 		_p(attr->_f);				\
1085 		ret += attr__fprintf(fp, _n, buf, priv);\
1086 	}						\
1087 } while (0)
1088 
1089 #define PRINT_ATTRf(_f, _p)	PRINT_ATTRn(#_f, _f, _p)
1090 
1091 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1092 			     attr__fprintf_f attr__fprintf, void *priv)
1093 {
1094 	char buf[BUF_SIZE];
1095 	int ret = 0;
1096 
1097 	PRINT_ATTRf(type, p_unsigned);
1098 	PRINT_ATTRf(size, p_unsigned);
1099 	PRINT_ATTRf(config, p_hex);
1100 	PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1101 	PRINT_ATTRf(sample_type, p_sample_type);
1102 	PRINT_ATTRf(read_format, p_read_format);
1103 
1104 	PRINT_ATTRf(disabled, p_unsigned);
1105 	PRINT_ATTRf(inherit, p_unsigned);
1106 	PRINT_ATTRf(pinned, p_unsigned);
1107 	PRINT_ATTRf(exclusive, p_unsigned);
1108 	PRINT_ATTRf(exclude_user, p_unsigned);
1109 	PRINT_ATTRf(exclude_kernel, p_unsigned);
1110 	PRINT_ATTRf(exclude_hv, p_unsigned);
1111 	PRINT_ATTRf(exclude_idle, p_unsigned);
1112 	PRINT_ATTRf(mmap, p_unsigned);
1113 	PRINT_ATTRf(comm, p_unsigned);
1114 	PRINT_ATTRf(freq, p_unsigned);
1115 	PRINT_ATTRf(inherit_stat, p_unsigned);
1116 	PRINT_ATTRf(enable_on_exec, p_unsigned);
1117 	PRINT_ATTRf(task, p_unsigned);
1118 	PRINT_ATTRf(watermark, p_unsigned);
1119 	PRINT_ATTRf(precise_ip, p_unsigned);
1120 	PRINT_ATTRf(mmap_data, p_unsigned);
1121 	PRINT_ATTRf(sample_id_all, p_unsigned);
1122 	PRINT_ATTRf(exclude_host, p_unsigned);
1123 	PRINT_ATTRf(exclude_guest, p_unsigned);
1124 	PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1125 	PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1126 	PRINT_ATTRf(mmap2, p_unsigned);
1127 	PRINT_ATTRf(comm_exec, p_unsigned);
1128 	PRINT_ATTRf(use_clockid, p_unsigned);
1129 
1130 	PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1131 	PRINT_ATTRf(bp_type, p_unsigned);
1132 	PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1133 	PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1134 	PRINT_ATTRf(sample_regs_user, p_hex);
1135 	PRINT_ATTRf(sample_stack_user, p_unsigned);
1136 	PRINT_ATTRf(clockid, p_signed);
1137 	PRINT_ATTRf(sample_regs_intr, p_hex);
1138 	PRINT_ATTRf(aux_watermark, p_unsigned);
1139 
1140 	return ret;
1141 }
1142 
1143 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1144 				void *priv __attribute__((unused)))
1145 {
1146 	return fprintf(fp, "  %-32s %s\n", name, val);
1147 }
1148 
1149 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1150 			      struct thread_map *threads)
1151 {
1152 	int cpu, thread, nthreads;
1153 	unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1154 	int pid = -1, err;
1155 	enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1156 
1157 	if (evsel->system_wide)
1158 		nthreads = 1;
1159 	else
1160 		nthreads = threads->nr;
1161 
1162 	if (evsel->fd == NULL &&
1163 	    perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1164 		return -ENOMEM;
1165 
1166 	if (evsel->cgrp) {
1167 		flags |= PERF_FLAG_PID_CGROUP;
1168 		pid = evsel->cgrp->fd;
1169 	}
1170 
1171 fallback_missing_features:
1172 	if (perf_missing_features.clockid_wrong)
1173 		evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1174 	if (perf_missing_features.clockid) {
1175 		evsel->attr.use_clockid = 0;
1176 		evsel->attr.clockid = 0;
1177 	}
1178 	if (perf_missing_features.cloexec)
1179 		flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1180 	if (perf_missing_features.mmap2)
1181 		evsel->attr.mmap2 = 0;
1182 	if (perf_missing_features.exclude_guest)
1183 		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1184 retry_sample_id:
1185 	if (perf_missing_features.sample_id_all)
1186 		evsel->attr.sample_id_all = 0;
1187 
1188 	if (verbose >= 2) {
1189 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1190 		fprintf(stderr, "perf_event_attr:\n");
1191 		perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1192 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1193 	}
1194 
1195 	for (cpu = 0; cpu < cpus->nr; cpu++) {
1196 
1197 		for (thread = 0; thread < nthreads; thread++) {
1198 			int group_fd;
1199 
1200 			if (!evsel->cgrp && !evsel->system_wide)
1201 				pid = thread_map__pid(threads, thread);
1202 
1203 			group_fd = get_group_fd(evsel, cpu, thread);
1204 retry_open:
1205 			pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1206 				  pid, cpus->map[cpu], group_fd, flags);
1207 
1208 			FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1209 								     pid,
1210 								     cpus->map[cpu],
1211 								     group_fd, flags);
1212 			if (FD(evsel, cpu, thread) < 0) {
1213 				err = -errno;
1214 				pr_debug2("sys_perf_event_open failed, error %d\n",
1215 					  err);
1216 				goto try_fallback;
1217 			}
1218 			set_rlimit = NO_CHANGE;
1219 
1220 			/*
1221 			 * If we succeeded but had to kill clockid, fail and
1222 			 * have perf_evsel__open_strerror() print us a nice
1223 			 * error.
1224 			 */
1225 			if (perf_missing_features.clockid ||
1226 			    perf_missing_features.clockid_wrong) {
1227 				err = -EINVAL;
1228 				goto out_close;
1229 			}
1230 		}
1231 	}
1232 
1233 	return 0;
1234 
1235 try_fallback:
1236 	/*
1237 	 * perf stat needs between 5 and 22 fds per CPU. When we run out
1238 	 * of them try to increase the limits.
1239 	 */
1240 	if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1241 		struct rlimit l;
1242 		int old_errno = errno;
1243 
1244 		if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1245 			if (set_rlimit == NO_CHANGE)
1246 				l.rlim_cur = l.rlim_max;
1247 			else {
1248 				l.rlim_cur = l.rlim_max + 1000;
1249 				l.rlim_max = l.rlim_cur;
1250 			}
1251 			if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1252 				set_rlimit++;
1253 				errno = old_errno;
1254 				goto retry_open;
1255 			}
1256 		}
1257 		errno = old_errno;
1258 	}
1259 
1260 	if (err != -EINVAL || cpu > 0 || thread > 0)
1261 		goto out_close;
1262 
1263 	/*
1264 	 * Must probe features in the order they were added to the
1265 	 * perf_event_attr interface.
1266 	 */
1267 	if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1268 		perf_missing_features.clockid_wrong = true;
1269 		goto fallback_missing_features;
1270 	} else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1271 		perf_missing_features.clockid = true;
1272 		goto fallback_missing_features;
1273 	} else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1274 		perf_missing_features.cloexec = true;
1275 		goto fallback_missing_features;
1276 	} else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1277 		perf_missing_features.mmap2 = true;
1278 		goto fallback_missing_features;
1279 	} else if (!perf_missing_features.exclude_guest &&
1280 		   (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1281 		perf_missing_features.exclude_guest = true;
1282 		goto fallback_missing_features;
1283 	} else if (!perf_missing_features.sample_id_all) {
1284 		perf_missing_features.sample_id_all = true;
1285 		goto retry_sample_id;
1286 	}
1287 
1288 out_close:
1289 	do {
1290 		while (--thread >= 0) {
1291 			close(FD(evsel, cpu, thread));
1292 			FD(evsel, cpu, thread) = -1;
1293 		}
1294 		thread = nthreads;
1295 	} while (--cpu >= 0);
1296 	return err;
1297 }
1298 
1299 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1300 {
1301 	if (evsel->fd == NULL)
1302 		return;
1303 
1304 	perf_evsel__close_fd(evsel, ncpus, nthreads);
1305 	perf_evsel__free_fd(evsel);
1306 }
1307 
1308 static struct {
1309 	struct cpu_map map;
1310 	int cpus[1];
1311 } empty_cpu_map = {
1312 	.map.nr	= 1,
1313 	.cpus	= { -1, },
1314 };
1315 
1316 static struct {
1317 	struct thread_map map;
1318 	int threads[1];
1319 } empty_thread_map = {
1320 	.map.nr	 = 1,
1321 	.threads = { -1, },
1322 };
1323 
1324 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1325 		     struct thread_map *threads)
1326 {
1327 	if (cpus == NULL) {
1328 		/* Work around old compiler warnings about strict aliasing */
1329 		cpus = &empty_cpu_map.map;
1330 	}
1331 
1332 	if (threads == NULL)
1333 		threads = &empty_thread_map.map;
1334 
1335 	return __perf_evsel__open(evsel, cpus, threads);
1336 }
1337 
1338 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1339 			     struct cpu_map *cpus)
1340 {
1341 	return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1342 }
1343 
1344 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1345 				struct thread_map *threads)
1346 {
1347 	return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1348 }
1349 
1350 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1351 				       const union perf_event *event,
1352 				       struct perf_sample *sample)
1353 {
1354 	u64 type = evsel->attr.sample_type;
1355 	const u64 *array = event->sample.array;
1356 	bool swapped = evsel->needs_swap;
1357 	union u64_swap u;
1358 
1359 	array += ((event->header.size -
1360 		   sizeof(event->header)) / sizeof(u64)) - 1;
1361 
1362 	if (type & PERF_SAMPLE_IDENTIFIER) {
1363 		sample->id = *array;
1364 		array--;
1365 	}
1366 
1367 	if (type & PERF_SAMPLE_CPU) {
1368 		u.val64 = *array;
1369 		if (swapped) {
1370 			/* undo swap of u64, then swap on individual u32s */
1371 			u.val64 = bswap_64(u.val64);
1372 			u.val32[0] = bswap_32(u.val32[0]);
1373 		}
1374 
1375 		sample->cpu = u.val32[0];
1376 		array--;
1377 	}
1378 
1379 	if (type & PERF_SAMPLE_STREAM_ID) {
1380 		sample->stream_id = *array;
1381 		array--;
1382 	}
1383 
1384 	if (type & PERF_SAMPLE_ID) {
1385 		sample->id = *array;
1386 		array--;
1387 	}
1388 
1389 	if (type & PERF_SAMPLE_TIME) {
1390 		sample->time = *array;
1391 		array--;
1392 	}
1393 
1394 	if (type & PERF_SAMPLE_TID) {
1395 		u.val64 = *array;
1396 		if (swapped) {
1397 			/* undo swap of u64, then swap on individual u32s */
1398 			u.val64 = bswap_64(u.val64);
1399 			u.val32[0] = bswap_32(u.val32[0]);
1400 			u.val32[1] = bswap_32(u.val32[1]);
1401 		}
1402 
1403 		sample->pid = u.val32[0];
1404 		sample->tid = u.val32[1];
1405 		array--;
1406 	}
1407 
1408 	return 0;
1409 }
1410 
1411 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1412 			    u64 size)
1413 {
1414 	return size > max_size || offset + size > endp;
1415 }
1416 
1417 #define OVERFLOW_CHECK(offset, size, max_size)				\
1418 	do {								\
1419 		if (overflow(endp, (max_size), (offset), (size)))	\
1420 			return -EFAULT;					\
1421 	} while (0)
1422 
1423 #define OVERFLOW_CHECK_u64(offset) \
1424 	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1425 
1426 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1427 			     struct perf_sample *data)
1428 {
1429 	u64 type = evsel->attr.sample_type;
1430 	bool swapped = evsel->needs_swap;
1431 	const u64 *array;
1432 	u16 max_size = event->header.size;
1433 	const void *endp = (void *)event + max_size;
1434 	u64 sz;
1435 
1436 	/*
1437 	 * used for cross-endian analysis. See git commit 65014ab3
1438 	 * for why this goofiness is needed.
1439 	 */
1440 	union u64_swap u;
1441 
1442 	memset(data, 0, sizeof(*data));
1443 	data->cpu = data->pid = data->tid = -1;
1444 	data->stream_id = data->id = data->time = -1ULL;
1445 	data->period = evsel->attr.sample_period;
1446 	data->weight = 0;
1447 
1448 	if (event->header.type != PERF_RECORD_SAMPLE) {
1449 		if (!evsel->attr.sample_id_all)
1450 			return 0;
1451 		return perf_evsel__parse_id_sample(evsel, event, data);
1452 	}
1453 
1454 	array = event->sample.array;
1455 
1456 	/*
1457 	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1458 	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1459 	 * check the format does not go past the end of the event.
1460 	 */
1461 	if (evsel->sample_size + sizeof(event->header) > event->header.size)
1462 		return -EFAULT;
1463 
1464 	data->id = -1ULL;
1465 	if (type & PERF_SAMPLE_IDENTIFIER) {
1466 		data->id = *array;
1467 		array++;
1468 	}
1469 
1470 	if (type & PERF_SAMPLE_IP) {
1471 		data->ip = *array;
1472 		array++;
1473 	}
1474 
1475 	if (type & PERF_SAMPLE_TID) {
1476 		u.val64 = *array;
1477 		if (swapped) {
1478 			/* undo swap of u64, then swap on individual u32s */
1479 			u.val64 = bswap_64(u.val64);
1480 			u.val32[0] = bswap_32(u.val32[0]);
1481 			u.val32[1] = bswap_32(u.val32[1]);
1482 		}
1483 
1484 		data->pid = u.val32[0];
1485 		data->tid = u.val32[1];
1486 		array++;
1487 	}
1488 
1489 	if (type & PERF_SAMPLE_TIME) {
1490 		data->time = *array;
1491 		array++;
1492 	}
1493 
1494 	data->addr = 0;
1495 	if (type & PERF_SAMPLE_ADDR) {
1496 		data->addr = *array;
1497 		array++;
1498 	}
1499 
1500 	if (type & PERF_SAMPLE_ID) {
1501 		data->id = *array;
1502 		array++;
1503 	}
1504 
1505 	if (type & PERF_SAMPLE_STREAM_ID) {
1506 		data->stream_id = *array;
1507 		array++;
1508 	}
1509 
1510 	if (type & PERF_SAMPLE_CPU) {
1511 
1512 		u.val64 = *array;
1513 		if (swapped) {
1514 			/* undo swap of u64, then swap on individual u32s */
1515 			u.val64 = bswap_64(u.val64);
1516 			u.val32[0] = bswap_32(u.val32[0]);
1517 		}
1518 
1519 		data->cpu = u.val32[0];
1520 		array++;
1521 	}
1522 
1523 	if (type & PERF_SAMPLE_PERIOD) {
1524 		data->period = *array;
1525 		array++;
1526 	}
1527 
1528 	if (type & PERF_SAMPLE_READ) {
1529 		u64 read_format = evsel->attr.read_format;
1530 
1531 		OVERFLOW_CHECK_u64(array);
1532 		if (read_format & PERF_FORMAT_GROUP)
1533 			data->read.group.nr = *array;
1534 		else
1535 			data->read.one.value = *array;
1536 
1537 		array++;
1538 
1539 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1540 			OVERFLOW_CHECK_u64(array);
1541 			data->read.time_enabled = *array;
1542 			array++;
1543 		}
1544 
1545 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1546 			OVERFLOW_CHECK_u64(array);
1547 			data->read.time_running = *array;
1548 			array++;
1549 		}
1550 
1551 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1552 		if (read_format & PERF_FORMAT_GROUP) {
1553 			const u64 max_group_nr = UINT64_MAX /
1554 					sizeof(struct sample_read_value);
1555 
1556 			if (data->read.group.nr > max_group_nr)
1557 				return -EFAULT;
1558 			sz = data->read.group.nr *
1559 			     sizeof(struct sample_read_value);
1560 			OVERFLOW_CHECK(array, sz, max_size);
1561 			data->read.group.values =
1562 					(struct sample_read_value *)array;
1563 			array = (void *)array + sz;
1564 		} else {
1565 			OVERFLOW_CHECK_u64(array);
1566 			data->read.one.id = *array;
1567 			array++;
1568 		}
1569 	}
1570 
1571 	if (type & PERF_SAMPLE_CALLCHAIN) {
1572 		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1573 
1574 		OVERFLOW_CHECK_u64(array);
1575 		data->callchain = (struct ip_callchain *)array++;
1576 		if (data->callchain->nr > max_callchain_nr)
1577 			return -EFAULT;
1578 		sz = data->callchain->nr * sizeof(u64);
1579 		OVERFLOW_CHECK(array, sz, max_size);
1580 		array = (void *)array + sz;
1581 	}
1582 
1583 	if (type & PERF_SAMPLE_RAW) {
1584 		OVERFLOW_CHECK_u64(array);
1585 		u.val64 = *array;
1586 		if (WARN_ONCE(swapped,
1587 			      "Endianness of raw data not corrected!\n")) {
1588 			/* undo swap of u64, then swap on individual u32s */
1589 			u.val64 = bswap_64(u.val64);
1590 			u.val32[0] = bswap_32(u.val32[0]);
1591 			u.val32[1] = bswap_32(u.val32[1]);
1592 		}
1593 		data->raw_size = u.val32[0];
1594 		array = (void *)array + sizeof(u32);
1595 
1596 		OVERFLOW_CHECK(array, data->raw_size, max_size);
1597 		data->raw_data = (void *)array;
1598 		array = (void *)array + data->raw_size;
1599 	}
1600 
1601 	if (type & PERF_SAMPLE_BRANCH_STACK) {
1602 		const u64 max_branch_nr = UINT64_MAX /
1603 					  sizeof(struct branch_entry);
1604 
1605 		OVERFLOW_CHECK_u64(array);
1606 		data->branch_stack = (struct branch_stack *)array++;
1607 
1608 		if (data->branch_stack->nr > max_branch_nr)
1609 			return -EFAULT;
1610 		sz = data->branch_stack->nr * sizeof(struct branch_entry);
1611 		OVERFLOW_CHECK(array, sz, max_size);
1612 		array = (void *)array + sz;
1613 	}
1614 
1615 	if (type & PERF_SAMPLE_REGS_USER) {
1616 		OVERFLOW_CHECK_u64(array);
1617 		data->user_regs.abi = *array;
1618 		array++;
1619 
1620 		if (data->user_regs.abi) {
1621 			u64 mask = evsel->attr.sample_regs_user;
1622 
1623 			sz = hweight_long(mask) * sizeof(u64);
1624 			OVERFLOW_CHECK(array, sz, max_size);
1625 			data->user_regs.mask = mask;
1626 			data->user_regs.regs = (u64 *)array;
1627 			array = (void *)array + sz;
1628 		}
1629 	}
1630 
1631 	if (type & PERF_SAMPLE_STACK_USER) {
1632 		OVERFLOW_CHECK_u64(array);
1633 		sz = *array++;
1634 
1635 		data->user_stack.offset = ((char *)(array - 1)
1636 					  - (char *) event);
1637 
1638 		if (!sz) {
1639 			data->user_stack.size = 0;
1640 		} else {
1641 			OVERFLOW_CHECK(array, sz, max_size);
1642 			data->user_stack.data = (char *)array;
1643 			array = (void *)array + sz;
1644 			OVERFLOW_CHECK_u64(array);
1645 			data->user_stack.size = *array++;
1646 			if (WARN_ONCE(data->user_stack.size > sz,
1647 				      "user stack dump failure\n"))
1648 				return -EFAULT;
1649 		}
1650 	}
1651 
1652 	data->weight = 0;
1653 	if (type & PERF_SAMPLE_WEIGHT) {
1654 		OVERFLOW_CHECK_u64(array);
1655 		data->weight = *array;
1656 		array++;
1657 	}
1658 
1659 	data->data_src = PERF_MEM_DATA_SRC_NONE;
1660 	if (type & PERF_SAMPLE_DATA_SRC) {
1661 		OVERFLOW_CHECK_u64(array);
1662 		data->data_src = *array;
1663 		array++;
1664 	}
1665 
1666 	data->transaction = 0;
1667 	if (type & PERF_SAMPLE_TRANSACTION) {
1668 		OVERFLOW_CHECK_u64(array);
1669 		data->transaction = *array;
1670 		array++;
1671 	}
1672 
1673 	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1674 	if (type & PERF_SAMPLE_REGS_INTR) {
1675 		OVERFLOW_CHECK_u64(array);
1676 		data->intr_regs.abi = *array;
1677 		array++;
1678 
1679 		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1680 			u64 mask = evsel->attr.sample_regs_intr;
1681 
1682 			sz = hweight_long(mask) * sizeof(u64);
1683 			OVERFLOW_CHECK(array, sz, max_size);
1684 			data->intr_regs.mask = mask;
1685 			data->intr_regs.regs = (u64 *)array;
1686 			array = (void *)array + sz;
1687 		}
1688 	}
1689 
1690 	return 0;
1691 }
1692 
1693 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1694 				     u64 read_format)
1695 {
1696 	size_t sz, result = sizeof(struct sample_event);
1697 
1698 	if (type & PERF_SAMPLE_IDENTIFIER)
1699 		result += sizeof(u64);
1700 
1701 	if (type & PERF_SAMPLE_IP)
1702 		result += sizeof(u64);
1703 
1704 	if (type & PERF_SAMPLE_TID)
1705 		result += sizeof(u64);
1706 
1707 	if (type & PERF_SAMPLE_TIME)
1708 		result += sizeof(u64);
1709 
1710 	if (type & PERF_SAMPLE_ADDR)
1711 		result += sizeof(u64);
1712 
1713 	if (type & PERF_SAMPLE_ID)
1714 		result += sizeof(u64);
1715 
1716 	if (type & PERF_SAMPLE_STREAM_ID)
1717 		result += sizeof(u64);
1718 
1719 	if (type & PERF_SAMPLE_CPU)
1720 		result += sizeof(u64);
1721 
1722 	if (type & PERF_SAMPLE_PERIOD)
1723 		result += sizeof(u64);
1724 
1725 	if (type & PERF_SAMPLE_READ) {
1726 		result += sizeof(u64);
1727 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1728 			result += sizeof(u64);
1729 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1730 			result += sizeof(u64);
1731 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1732 		if (read_format & PERF_FORMAT_GROUP) {
1733 			sz = sample->read.group.nr *
1734 			     sizeof(struct sample_read_value);
1735 			result += sz;
1736 		} else {
1737 			result += sizeof(u64);
1738 		}
1739 	}
1740 
1741 	if (type & PERF_SAMPLE_CALLCHAIN) {
1742 		sz = (sample->callchain->nr + 1) * sizeof(u64);
1743 		result += sz;
1744 	}
1745 
1746 	if (type & PERF_SAMPLE_RAW) {
1747 		result += sizeof(u32);
1748 		result += sample->raw_size;
1749 	}
1750 
1751 	if (type & PERF_SAMPLE_BRANCH_STACK) {
1752 		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1753 		sz += sizeof(u64);
1754 		result += sz;
1755 	}
1756 
1757 	if (type & PERF_SAMPLE_REGS_USER) {
1758 		if (sample->user_regs.abi) {
1759 			result += sizeof(u64);
1760 			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1761 			result += sz;
1762 		} else {
1763 			result += sizeof(u64);
1764 		}
1765 	}
1766 
1767 	if (type & PERF_SAMPLE_STACK_USER) {
1768 		sz = sample->user_stack.size;
1769 		result += sizeof(u64);
1770 		if (sz) {
1771 			result += sz;
1772 			result += sizeof(u64);
1773 		}
1774 	}
1775 
1776 	if (type & PERF_SAMPLE_WEIGHT)
1777 		result += sizeof(u64);
1778 
1779 	if (type & PERF_SAMPLE_DATA_SRC)
1780 		result += sizeof(u64);
1781 
1782 	if (type & PERF_SAMPLE_TRANSACTION)
1783 		result += sizeof(u64);
1784 
1785 	if (type & PERF_SAMPLE_REGS_INTR) {
1786 		if (sample->intr_regs.abi) {
1787 			result += sizeof(u64);
1788 			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1789 			result += sz;
1790 		} else {
1791 			result += sizeof(u64);
1792 		}
1793 	}
1794 
1795 	return result;
1796 }
1797 
1798 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1799 				  u64 read_format,
1800 				  const struct perf_sample *sample,
1801 				  bool swapped)
1802 {
1803 	u64 *array;
1804 	size_t sz;
1805 	/*
1806 	 * used for cross-endian analysis. See git commit 65014ab3
1807 	 * for why this goofiness is needed.
1808 	 */
1809 	union u64_swap u;
1810 
1811 	array = event->sample.array;
1812 
1813 	if (type & PERF_SAMPLE_IDENTIFIER) {
1814 		*array = sample->id;
1815 		array++;
1816 	}
1817 
1818 	if (type & PERF_SAMPLE_IP) {
1819 		*array = sample->ip;
1820 		array++;
1821 	}
1822 
1823 	if (type & PERF_SAMPLE_TID) {
1824 		u.val32[0] = sample->pid;
1825 		u.val32[1] = sample->tid;
1826 		if (swapped) {
1827 			/*
1828 			 * Inverse of what is done in perf_evsel__parse_sample
1829 			 */
1830 			u.val32[0] = bswap_32(u.val32[0]);
1831 			u.val32[1] = bswap_32(u.val32[1]);
1832 			u.val64 = bswap_64(u.val64);
1833 		}
1834 
1835 		*array = u.val64;
1836 		array++;
1837 	}
1838 
1839 	if (type & PERF_SAMPLE_TIME) {
1840 		*array = sample->time;
1841 		array++;
1842 	}
1843 
1844 	if (type & PERF_SAMPLE_ADDR) {
1845 		*array = sample->addr;
1846 		array++;
1847 	}
1848 
1849 	if (type & PERF_SAMPLE_ID) {
1850 		*array = sample->id;
1851 		array++;
1852 	}
1853 
1854 	if (type & PERF_SAMPLE_STREAM_ID) {
1855 		*array = sample->stream_id;
1856 		array++;
1857 	}
1858 
1859 	if (type & PERF_SAMPLE_CPU) {
1860 		u.val32[0] = sample->cpu;
1861 		if (swapped) {
1862 			/*
1863 			 * Inverse of what is done in perf_evsel__parse_sample
1864 			 */
1865 			u.val32[0] = bswap_32(u.val32[0]);
1866 			u.val64 = bswap_64(u.val64);
1867 		}
1868 		*array = u.val64;
1869 		array++;
1870 	}
1871 
1872 	if (type & PERF_SAMPLE_PERIOD) {
1873 		*array = sample->period;
1874 		array++;
1875 	}
1876 
1877 	if (type & PERF_SAMPLE_READ) {
1878 		if (read_format & PERF_FORMAT_GROUP)
1879 			*array = sample->read.group.nr;
1880 		else
1881 			*array = sample->read.one.value;
1882 		array++;
1883 
1884 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1885 			*array = sample->read.time_enabled;
1886 			array++;
1887 		}
1888 
1889 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1890 			*array = sample->read.time_running;
1891 			array++;
1892 		}
1893 
1894 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1895 		if (read_format & PERF_FORMAT_GROUP) {
1896 			sz = sample->read.group.nr *
1897 			     sizeof(struct sample_read_value);
1898 			memcpy(array, sample->read.group.values, sz);
1899 			array = (void *)array + sz;
1900 		} else {
1901 			*array = sample->read.one.id;
1902 			array++;
1903 		}
1904 	}
1905 
1906 	if (type & PERF_SAMPLE_CALLCHAIN) {
1907 		sz = (sample->callchain->nr + 1) * sizeof(u64);
1908 		memcpy(array, sample->callchain, sz);
1909 		array = (void *)array + sz;
1910 	}
1911 
1912 	if (type & PERF_SAMPLE_RAW) {
1913 		u.val32[0] = sample->raw_size;
1914 		if (WARN_ONCE(swapped,
1915 			      "Endianness of raw data not corrected!\n")) {
1916 			/*
1917 			 * Inverse of what is done in perf_evsel__parse_sample
1918 			 */
1919 			u.val32[0] = bswap_32(u.val32[0]);
1920 			u.val32[1] = bswap_32(u.val32[1]);
1921 			u.val64 = bswap_64(u.val64);
1922 		}
1923 		*array = u.val64;
1924 		array = (void *)array + sizeof(u32);
1925 
1926 		memcpy(array, sample->raw_data, sample->raw_size);
1927 		array = (void *)array + sample->raw_size;
1928 	}
1929 
1930 	if (type & PERF_SAMPLE_BRANCH_STACK) {
1931 		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1932 		sz += sizeof(u64);
1933 		memcpy(array, sample->branch_stack, sz);
1934 		array = (void *)array + sz;
1935 	}
1936 
1937 	if (type & PERF_SAMPLE_REGS_USER) {
1938 		if (sample->user_regs.abi) {
1939 			*array++ = sample->user_regs.abi;
1940 			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1941 			memcpy(array, sample->user_regs.regs, sz);
1942 			array = (void *)array + sz;
1943 		} else {
1944 			*array++ = 0;
1945 		}
1946 	}
1947 
1948 	if (type & PERF_SAMPLE_STACK_USER) {
1949 		sz = sample->user_stack.size;
1950 		*array++ = sz;
1951 		if (sz) {
1952 			memcpy(array, sample->user_stack.data, sz);
1953 			array = (void *)array + sz;
1954 			*array++ = sz;
1955 		}
1956 	}
1957 
1958 	if (type & PERF_SAMPLE_WEIGHT) {
1959 		*array = sample->weight;
1960 		array++;
1961 	}
1962 
1963 	if (type & PERF_SAMPLE_DATA_SRC) {
1964 		*array = sample->data_src;
1965 		array++;
1966 	}
1967 
1968 	if (type & PERF_SAMPLE_TRANSACTION) {
1969 		*array = sample->transaction;
1970 		array++;
1971 	}
1972 
1973 	if (type & PERF_SAMPLE_REGS_INTR) {
1974 		if (sample->intr_regs.abi) {
1975 			*array++ = sample->intr_regs.abi;
1976 			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1977 			memcpy(array, sample->intr_regs.regs, sz);
1978 			array = (void *)array + sz;
1979 		} else {
1980 			*array++ = 0;
1981 		}
1982 	}
1983 
1984 	return 0;
1985 }
1986 
1987 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1988 {
1989 	return pevent_find_field(evsel->tp_format, name);
1990 }
1991 
1992 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1993 			 const char *name)
1994 {
1995 	struct format_field *field = perf_evsel__field(evsel, name);
1996 	int offset;
1997 
1998 	if (!field)
1999 		return NULL;
2000 
2001 	offset = field->offset;
2002 
2003 	if (field->flags & FIELD_IS_DYNAMIC) {
2004 		offset = *(int *)(sample->raw_data + field->offset);
2005 		offset &= 0xffff;
2006 	}
2007 
2008 	return sample->raw_data + offset;
2009 }
2010 
2011 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2012 		       const char *name)
2013 {
2014 	struct format_field *field = perf_evsel__field(evsel, name);
2015 	void *ptr;
2016 	u64 value;
2017 
2018 	if (!field)
2019 		return 0;
2020 
2021 	ptr = sample->raw_data + field->offset;
2022 
2023 	switch (field->size) {
2024 	case 1:
2025 		return *(u8 *)ptr;
2026 	case 2:
2027 		value = *(u16 *)ptr;
2028 		break;
2029 	case 4:
2030 		value = *(u32 *)ptr;
2031 		break;
2032 	case 8:
2033 		memcpy(&value, ptr, sizeof(u64));
2034 		break;
2035 	default:
2036 		return 0;
2037 	}
2038 
2039 	if (!evsel->needs_swap)
2040 		return value;
2041 
2042 	switch (field->size) {
2043 	case 2:
2044 		return bswap_16(value);
2045 	case 4:
2046 		return bswap_32(value);
2047 	case 8:
2048 		return bswap_64(value);
2049 	default:
2050 		return 0;
2051 	}
2052 
2053 	return 0;
2054 }
2055 
2056 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2057 {
2058 	va_list args;
2059 	int ret = 0;
2060 
2061 	if (!*first) {
2062 		ret += fprintf(fp, ",");
2063 	} else {
2064 		ret += fprintf(fp, ":");
2065 		*first = false;
2066 	}
2067 
2068 	va_start(args, fmt);
2069 	ret += vfprintf(fp, fmt, args);
2070 	va_end(args);
2071 	return ret;
2072 }
2073 
2074 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2075 {
2076 	return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2077 }
2078 
2079 int perf_evsel__fprintf(struct perf_evsel *evsel,
2080 			struct perf_attr_details *details, FILE *fp)
2081 {
2082 	bool first = true;
2083 	int printed = 0;
2084 
2085 	if (details->event_group) {
2086 		struct perf_evsel *pos;
2087 
2088 		if (!perf_evsel__is_group_leader(evsel))
2089 			return 0;
2090 
2091 		if (evsel->nr_members > 1)
2092 			printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2093 
2094 		printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2095 		for_each_group_member(pos, evsel)
2096 			printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2097 
2098 		if (evsel->nr_members > 1)
2099 			printed += fprintf(fp, "}");
2100 		goto out;
2101 	}
2102 
2103 	printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2104 
2105 	if (details->verbose) {
2106 		printed += perf_event_attr__fprintf(fp, &evsel->attr,
2107 						    __print_attr__fprintf, &first);
2108 	} else if (details->freq) {
2109 		printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2110 					 (u64)evsel->attr.sample_freq);
2111 	}
2112 out:
2113 	fputc('\n', fp);
2114 	return ++printed;
2115 }
2116 
2117 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2118 			  char *msg, size_t msgsize)
2119 {
2120 	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2121 	    evsel->attr.type   == PERF_TYPE_HARDWARE &&
2122 	    evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2123 		/*
2124 		 * If it's cycles then fall back to hrtimer based
2125 		 * cpu-clock-tick sw counter, which is always available even if
2126 		 * no PMU support.
2127 		 *
2128 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2129 		 * b0a873e).
2130 		 */
2131 		scnprintf(msg, msgsize, "%s",
2132 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2133 
2134 		evsel->attr.type   = PERF_TYPE_SOFTWARE;
2135 		evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2136 
2137 		zfree(&evsel->name);
2138 		return true;
2139 	}
2140 
2141 	return false;
2142 }
2143 
2144 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2145 			      int err, char *msg, size_t size)
2146 {
2147 	char sbuf[STRERR_BUFSIZE];
2148 
2149 	switch (err) {
2150 	case EPERM:
2151 	case EACCES:
2152 		return scnprintf(msg, size,
2153 		 "You may not have permission to collect %sstats.\n"
2154 		 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2155 		 " -1 - Not paranoid at all\n"
2156 		 "  0 - Disallow raw tracepoint access for unpriv\n"
2157 		 "  1 - Disallow cpu events for unpriv\n"
2158 		 "  2 - Disallow kernel profiling for unpriv",
2159 				 target->system_wide ? "system-wide " : "");
2160 	case ENOENT:
2161 		return scnprintf(msg, size, "The %s event is not supported.",
2162 				 perf_evsel__name(evsel));
2163 	case EMFILE:
2164 		return scnprintf(msg, size, "%s",
2165 			 "Too many events are opened.\n"
2166 			 "Probably the maximum number of open file descriptors has been reached.\n"
2167 			 "Hint: Try again after reducing the number of events.\n"
2168 			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2169 	case ENODEV:
2170 		if (target->cpu_list)
2171 			return scnprintf(msg, size, "%s",
2172 	 "No such device - did you specify an out-of-range profile CPU?\n");
2173 		break;
2174 	case EOPNOTSUPP:
2175 		if (evsel->attr.precise_ip)
2176 			return scnprintf(msg, size, "%s",
2177 	"\'precise\' request may not be supported. Try removing 'p' modifier.");
2178 #if defined(__i386__) || defined(__x86_64__)
2179 		if (evsel->attr.type == PERF_TYPE_HARDWARE)
2180 			return scnprintf(msg, size, "%s",
2181 	"No hardware sampling interrupt available.\n"
2182 	"No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2183 #endif
2184 		break;
2185 	case EBUSY:
2186 		if (find_process("oprofiled"))
2187 			return scnprintf(msg, size,
2188 	"The PMU counters are busy/taken by another profiler.\n"
2189 	"We found oprofile daemon running, please stop it and try again.");
2190 		break;
2191 	case EINVAL:
2192 		if (perf_missing_features.clockid)
2193 			return scnprintf(msg, size, "clockid feature not supported.");
2194 		if (perf_missing_features.clockid_wrong)
2195 			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2196 		break;
2197 	default:
2198 		break;
2199 	}
2200 
2201 	return scnprintf(msg, size,
2202 	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2203 	"/bin/dmesg may provide additional information.\n"
2204 	"No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2205 			 err, strerror_r(err, sbuf, sizeof(sbuf)),
2206 			 perf_evsel__name(evsel));
2207 }
2208