xref: /openbmc/linux/tools/perf/util/evsel.c (revision 8684014d)
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 
30 static struct {
31 	bool sample_id_all;
32 	bool exclude_guest;
33 	bool mmap2;
34 	bool cloexec;
35 } perf_missing_features;
36 
37 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
38 {
39 	return 0;
40 }
41 
42 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
43 {
44 }
45 
46 static struct {
47 	size_t	size;
48 	int	(*init)(struct perf_evsel *evsel);
49 	void	(*fini)(struct perf_evsel *evsel);
50 } perf_evsel__object = {
51 	.size = sizeof(struct perf_evsel),
52 	.init = perf_evsel__no_extra_init,
53 	.fini = perf_evsel__no_extra_fini,
54 };
55 
56 int perf_evsel__object_config(size_t object_size,
57 			      int (*init)(struct perf_evsel *evsel),
58 			      void (*fini)(struct perf_evsel *evsel))
59 {
60 
61 	if (object_size == 0)
62 		goto set_methods;
63 
64 	if (perf_evsel__object.size > object_size)
65 		return -EINVAL;
66 
67 	perf_evsel__object.size = object_size;
68 
69 set_methods:
70 	if (init != NULL)
71 		perf_evsel__object.init = init;
72 
73 	if (fini != NULL)
74 		perf_evsel__object.fini = fini;
75 
76 	return 0;
77 }
78 
79 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
80 
81 int __perf_evsel__sample_size(u64 sample_type)
82 {
83 	u64 mask = sample_type & PERF_SAMPLE_MASK;
84 	int size = 0;
85 	int i;
86 
87 	for (i = 0; i < 64; i++) {
88 		if (mask & (1ULL << i))
89 			size++;
90 	}
91 
92 	size *= sizeof(u64);
93 
94 	return size;
95 }
96 
97 /**
98  * __perf_evsel__calc_id_pos - calculate id_pos.
99  * @sample_type: sample type
100  *
101  * This function returns the position of the event id (PERF_SAMPLE_ID or
102  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
103  * sample_event.
104  */
105 static int __perf_evsel__calc_id_pos(u64 sample_type)
106 {
107 	int idx = 0;
108 
109 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
110 		return 0;
111 
112 	if (!(sample_type & PERF_SAMPLE_ID))
113 		return -1;
114 
115 	if (sample_type & PERF_SAMPLE_IP)
116 		idx += 1;
117 
118 	if (sample_type & PERF_SAMPLE_TID)
119 		idx += 1;
120 
121 	if (sample_type & PERF_SAMPLE_TIME)
122 		idx += 1;
123 
124 	if (sample_type & PERF_SAMPLE_ADDR)
125 		idx += 1;
126 
127 	return idx;
128 }
129 
130 /**
131  * __perf_evsel__calc_is_pos - calculate is_pos.
132  * @sample_type: sample type
133  *
134  * This function returns the position (counting backwards) of the event id
135  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
136  * sample_id_all is used there is an id sample appended to non-sample events.
137  */
138 static int __perf_evsel__calc_is_pos(u64 sample_type)
139 {
140 	int idx = 1;
141 
142 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
143 		return 1;
144 
145 	if (!(sample_type & PERF_SAMPLE_ID))
146 		return -1;
147 
148 	if (sample_type & PERF_SAMPLE_CPU)
149 		idx += 1;
150 
151 	if (sample_type & PERF_SAMPLE_STREAM_ID)
152 		idx += 1;
153 
154 	return idx;
155 }
156 
157 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
158 {
159 	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
160 	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
161 }
162 
163 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
164 				  enum perf_event_sample_format bit)
165 {
166 	if (!(evsel->attr.sample_type & bit)) {
167 		evsel->attr.sample_type |= bit;
168 		evsel->sample_size += sizeof(u64);
169 		perf_evsel__calc_id_pos(evsel);
170 	}
171 }
172 
173 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
174 				    enum perf_event_sample_format bit)
175 {
176 	if (evsel->attr.sample_type & bit) {
177 		evsel->attr.sample_type &= ~bit;
178 		evsel->sample_size -= sizeof(u64);
179 		perf_evsel__calc_id_pos(evsel);
180 	}
181 }
182 
183 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
184 			       bool can_sample_identifier)
185 {
186 	if (can_sample_identifier) {
187 		perf_evsel__reset_sample_bit(evsel, ID);
188 		perf_evsel__set_sample_bit(evsel, IDENTIFIER);
189 	} else {
190 		perf_evsel__set_sample_bit(evsel, ID);
191 	}
192 	evsel->attr.read_format |= PERF_FORMAT_ID;
193 }
194 
195 void perf_evsel__init(struct perf_evsel *evsel,
196 		      struct perf_event_attr *attr, int idx)
197 {
198 	evsel->idx	   = idx;
199 	evsel->tracking	   = !idx;
200 	evsel->attr	   = *attr;
201 	evsel->leader	   = evsel;
202 	evsel->unit	   = "";
203 	evsel->scale	   = 1.0;
204 	INIT_LIST_HEAD(&evsel->node);
205 	perf_evsel__object.init(evsel);
206 	evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
207 	perf_evsel__calc_id_pos(evsel);
208 }
209 
210 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
211 {
212 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
213 
214 	if (evsel != NULL)
215 		perf_evsel__init(evsel, attr, idx);
216 
217 	return evsel;
218 }
219 
220 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
221 {
222 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
223 
224 	if (evsel != NULL) {
225 		struct perf_event_attr attr = {
226 			.type	       = PERF_TYPE_TRACEPOINT,
227 			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
228 					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
229 		};
230 
231 		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
232 			goto out_free;
233 
234 		evsel->tp_format = trace_event__tp_format(sys, name);
235 		if (evsel->tp_format == NULL)
236 			goto out_free;
237 
238 		event_attr_init(&attr);
239 		attr.config = evsel->tp_format->id;
240 		attr.sample_period = 1;
241 		perf_evsel__init(evsel, &attr, idx);
242 	}
243 
244 	return evsel;
245 
246 out_free:
247 	zfree(&evsel->name);
248 	free(evsel);
249 	return NULL;
250 }
251 
252 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
253 	"cycles",
254 	"instructions",
255 	"cache-references",
256 	"cache-misses",
257 	"branches",
258 	"branch-misses",
259 	"bus-cycles",
260 	"stalled-cycles-frontend",
261 	"stalled-cycles-backend",
262 	"ref-cycles",
263 };
264 
265 static const char *__perf_evsel__hw_name(u64 config)
266 {
267 	if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
268 		return perf_evsel__hw_names[config];
269 
270 	return "unknown-hardware";
271 }
272 
273 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
274 {
275 	int colon = 0, r = 0;
276 	struct perf_event_attr *attr = &evsel->attr;
277 	bool exclude_guest_default = false;
278 
279 #define MOD_PRINT(context, mod)	do {					\
280 		if (!attr->exclude_##context) {				\
281 			if (!colon) colon = ++r;			\
282 			r += scnprintf(bf + r, size - r, "%c", mod);	\
283 		} } while(0)
284 
285 	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
286 		MOD_PRINT(kernel, 'k');
287 		MOD_PRINT(user, 'u');
288 		MOD_PRINT(hv, 'h');
289 		exclude_guest_default = true;
290 	}
291 
292 	if (attr->precise_ip) {
293 		if (!colon)
294 			colon = ++r;
295 		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
296 		exclude_guest_default = true;
297 	}
298 
299 	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
300 		MOD_PRINT(host, 'H');
301 		MOD_PRINT(guest, 'G');
302 	}
303 #undef MOD_PRINT
304 	if (colon)
305 		bf[colon - 1] = ':';
306 	return r;
307 }
308 
309 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
310 {
311 	int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
312 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
313 }
314 
315 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
316 	"cpu-clock",
317 	"task-clock",
318 	"page-faults",
319 	"context-switches",
320 	"cpu-migrations",
321 	"minor-faults",
322 	"major-faults",
323 	"alignment-faults",
324 	"emulation-faults",
325 	"dummy",
326 };
327 
328 static const char *__perf_evsel__sw_name(u64 config)
329 {
330 	if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
331 		return perf_evsel__sw_names[config];
332 	return "unknown-software";
333 }
334 
335 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
336 {
337 	int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
338 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
339 }
340 
341 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
342 {
343 	int r;
344 
345 	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
346 
347 	if (type & HW_BREAKPOINT_R)
348 		r += scnprintf(bf + r, size - r, "r");
349 
350 	if (type & HW_BREAKPOINT_W)
351 		r += scnprintf(bf + r, size - r, "w");
352 
353 	if (type & HW_BREAKPOINT_X)
354 		r += scnprintf(bf + r, size - r, "x");
355 
356 	return r;
357 }
358 
359 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
360 {
361 	struct perf_event_attr *attr = &evsel->attr;
362 	int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
363 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
364 }
365 
366 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
367 				[PERF_EVSEL__MAX_ALIASES] = {
368  { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
369  { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
370  { "LLC",	"L2",							},
371  { "dTLB",	"d-tlb",	"Data-TLB",				},
372  { "iTLB",	"i-tlb",	"Instruction-TLB",			},
373  { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
374  { "node",								},
375 };
376 
377 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
378 				   [PERF_EVSEL__MAX_ALIASES] = {
379  { "load",	"loads",	"read",					},
380  { "store",	"stores",	"write",				},
381  { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
382 };
383 
384 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
385 				       [PERF_EVSEL__MAX_ALIASES] = {
386  { "refs",	"Reference",	"ops",		"access",		},
387  { "misses",	"miss",							},
388 };
389 
390 #define C(x)		PERF_COUNT_HW_CACHE_##x
391 #define CACHE_READ	(1 << C(OP_READ))
392 #define CACHE_WRITE	(1 << C(OP_WRITE))
393 #define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
394 #define COP(x)		(1 << x)
395 
396 /*
397  * cache operartion stat
398  * L1I : Read and prefetch only
399  * ITLB and BPU : Read-only
400  */
401 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
402  [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
403  [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
404  [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
405  [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
406  [C(ITLB)]	= (CACHE_READ),
407  [C(BPU)]	= (CACHE_READ),
408  [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
409 };
410 
411 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
412 {
413 	if (perf_evsel__hw_cache_stat[type] & COP(op))
414 		return true;	/* valid */
415 	else
416 		return false;	/* invalid */
417 }
418 
419 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
420 					    char *bf, size_t size)
421 {
422 	if (result) {
423 		return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
424 				 perf_evsel__hw_cache_op[op][0],
425 				 perf_evsel__hw_cache_result[result][0]);
426 	}
427 
428 	return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
429 			 perf_evsel__hw_cache_op[op][1]);
430 }
431 
432 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
433 {
434 	u8 op, result, type = (config >>  0) & 0xff;
435 	const char *err = "unknown-ext-hardware-cache-type";
436 
437 	if (type > PERF_COUNT_HW_CACHE_MAX)
438 		goto out_err;
439 
440 	op = (config >>  8) & 0xff;
441 	err = "unknown-ext-hardware-cache-op";
442 	if (op > PERF_COUNT_HW_CACHE_OP_MAX)
443 		goto out_err;
444 
445 	result = (config >> 16) & 0xff;
446 	err = "unknown-ext-hardware-cache-result";
447 	if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
448 		goto out_err;
449 
450 	err = "invalid-cache";
451 	if (!perf_evsel__is_cache_op_valid(type, op))
452 		goto out_err;
453 
454 	return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
455 out_err:
456 	return scnprintf(bf, size, "%s", err);
457 }
458 
459 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
460 {
461 	int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
462 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
463 }
464 
465 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
466 {
467 	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
468 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
469 }
470 
471 const char *perf_evsel__name(struct perf_evsel *evsel)
472 {
473 	char bf[128];
474 
475 	if (evsel->name)
476 		return evsel->name;
477 
478 	switch (evsel->attr.type) {
479 	case PERF_TYPE_RAW:
480 		perf_evsel__raw_name(evsel, bf, sizeof(bf));
481 		break;
482 
483 	case PERF_TYPE_HARDWARE:
484 		perf_evsel__hw_name(evsel, bf, sizeof(bf));
485 		break;
486 
487 	case PERF_TYPE_HW_CACHE:
488 		perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
489 		break;
490 
491 	case PERF_TYPE_SOFTWARE:
492 		perf_evsel__sw_name(evsel, bf, sizeof(bf));
493 		break;
494 
495 	case PERF_TYPE_TRACEPOINT:
496 		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
497 		break;
498 
499 	case PERF_TYPE_BREAKPOINT:
500 		perf_evsel__bp_name(evsel, bf, sizeof(bf));
501 		break;
502 
503 	default:
504 		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
505 			  evsel->attr.type);
506 		break;
507 	}
508 
509 	evsel->name = strdup(bf);
510 
511 	return evsel->name ?: "unknown";
512 }
513 
514 const char *perf_evsel__group_name(struct perf_evsel *evsel)
515 {
516 	return evsel->group_name ?: "anon group";
517 }
518 
519 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
520 {
521 	int ret;
522 	struct perf_evsel *pos;
523 	const char *group_name = perf_evsel__group_name(evsel);
524 
525 	ret = scnprintf(buf, size, "%s", group_name);
526 
527 	ret += scnprintf(buf + ret, size - ret, " { %s",
528 			 perf_evsel__name(evsel));
529 
530 	for_each_group_member(pos, evsel)
531 		ret += scnprintf(buf + ret, size - ret, ", %s",
532 				 perf_evsel__name(pos));
533 
534 	ret += scnprintf(buf + ret, size - ret, " }");
535 
536 	return ret;
537 }
538 
539 static void
540 perf_evsel__config_callgraph(struct perf_evsel *evsel)
541 {
542 	bool function = perf_evsel__is_function_event(evsel);
543 	struct perf_event_attr *attr = &evsel->attr;
544 
545 	perf_evsel__set_sample_bit(evsel, CALLCHAIN);
546 
547 	if (callchain_param.record_mode == CALLCHAIN_DWARF) {
548 		if (!function) {
549 			perf_evsel__set_sample_bit(evsel, REGS_USER);
550 			perf_evsel__set_sample_bit(evsel, STACK_USER);
551 			attr->sample_regs_user = PERF_REGS_MASK;
552 			attr->sample_stack_user = callchain_param.dump_size;
553 			attr->exclude_callchain_user = 1;
554 		} else {
555 			pr_info("Cannot use DWARF unwind for function trace event,"
556 				" falling back to framepointers.\n");
557 		}
558 	}
559 
560 	if (function) {
561 		pr_info("Disabling user space callchains for function trace event.\n");
562 		attr->exclude_callchain_user = 1;
563 	}
564 }
565 
566 /*
567  * The enable_on_exec/disabled value strategy:
568  *
569  *  1) For any type of traced program:
570  *    - all independent events and group leaders are disabled
571  *    - all group members are enabled
572  *
573  *     Group members are ruled by group leaders. They need to
574  *     be enabled, because the group scheduling relies on that.
575  *
576  *  2) For traced programs executed by perf:
577  *     - all independent events and group leaders have
578  *       enable_on_exec set
579  *     - we don't specifically enable or disable any event during
580  *       the record command
581  *
582  *     Independent events and group leaders are initially disabled
583  *     and get enabled by exec. Group members are ruled by group
584  *     leaders as stated in 1).
585  *
586  *  3) For traced programs attached by perf (pid/tid):
587  *     - we specifically enable or disable all events during
588  *       the record command
589  *
590  *     When attaching events to already running traced we
591  *     enable/disable events specifically, as there's no
592  *     initial traced exec call.
593  */
594 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
595 {
596 	struct perf_evsel *leader = evsel->leader;
597 	struct perf_event_attr *attr = &evsel->attr;
598 	int track = evsel->tracking;
599 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
600 
601 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
602 	attr->inherit	    = !opts->no_inherit;
603 
604 	perf_evsel__set_sample_bit(evsel, IP);
605 	perf_evsel__set_sample_bit(evsel, TID);
606 
607 	if (evsel->sample_read) {
608 		perf_evsel__set_sample_bit(evsel, READ);
609 
610 		/*
611 		 * We need ID even in case of single event, because
612 		 * PERF_SAMPLE_READ process ID specific data.
613 		 */
614 		perf_evsel__set_sample_id(evsel, false);
615 
616 		/*
617 		 * Apply group format only if we belong to group
618 		 * with more than one members.
619 		 */
620 		if (leader->nr_members > 1) {
621 			attr->read_format |= PERF_FORMAT_GROUP;
622 			attr->inherit = 0;
623 		}
624 	}
625 
626 	/*
627 	 * We default some events to have a default interval. But keep
628 	 * it a weak assumption overridable by the user.
629 	 */
630 	if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
631 				     opts->user_interval != ULLONG_MAX)) {
632 		if (opts->freq) {
633 			perf_evsel__set_sample_bit(evsel, PERIOD);
634 			attr->freq		= 1;
635 			attr->sample_freq	= opts->freq;
636 		} else {
637 			attr->sample_period = opts->default_interval;
638 		}
639 	}
640 
641 	/*
642 	 * Disable sampling for all group members other
643 	 * than leader in case leader 'leads' the sampling.
644 	 */
645 	if ((leader != evsel) && leader->sample_read) {
646 		attr->sample_freq   = 0;
647 		attr->sample_period = 0;
648 	}
649 
650 	if (opts->no_samples)
651 		attr->sample_freq = 0;
652 
653 	if (opts->inherit_stat)
654 		attr->inherit_stat = 1;
655 
656 	if (opts->sample_address) {
657 		perf_evsel__set_sample_bit(evsel, ADDR);
658 		attr->mmap_data = track;
659 	}
660 
661 	/*
662 	 * We don't allow user space callchains for  function trace
663 	 * event, due to issues with page faults while tracing page
664 	 * fault handler and its overall trickiness nature.
665 	 */
666 	if (perf_evsel__is_function_event(evsel))
667 		evsel->attr.exclude_callchain_user = 1;
668 
669 	if (callchain_param.enabled && !evsel->no_aux_samples)
670 		perf_evsel__config_callgraph(evsel);
671 
672 	if (opts->sample_intr_regs) {
673 		attr->sample_regs_intr = PERF_REGS_MASK;
674 		perf_evsel__set_sample_bit(evsel, REGS_INTR);
675 	}
676 
677 	if (target__has_cpu(&opts->target))
678 		perf_evsel__set_sample_bit(evsel, CPU);
679 
680 	if (opts->period)
681 		perf_evsel__set_sample_bit(evsel, PERIOD);
682 
683 	/*
684 	 * When the user explicitely disabled time don't force it here.
685 	 */
686 	if (opts->sample_time &&
687 	    (!perf_missing_features.sample_id_all &&
688 	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu)))
689 		perf_evsel__set_sample_bit(evsel, TIME);
690 
691 	if (opts->raw_samples && !evsel->no_aux_samples) {
692 		perf_evsel__set_sample_bit(evsel, TIME);
693 		perf_evsel__set_sample_bit(evsel, RAW);
694 		perf_evsel__set_sample_bit(evsel, CPU);
695 	}
696 
697 	if (opts->sample_address)
698 		perf_evsel__set_sample_bit(evsel, DATA_SRC);
699 
700 	if (opts->no_buffering) {
701 		attr->watermark = 0;
702 		attr->wakeup_events = 1;
703 	}
704 	if (opts->branch_stack && !evsel->no_aux_samples) {
705 		perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
706 		attr->branch_sample_type = opts->branch_stack;
707 	}
708 
709 	if (opts->sample_weight)
710 		perf_evsel__set_sample_bit(evsel, WEIGHT);
711 
712 	attr->mmap  = track;
713 	attr->mmap2 = track && !perf_missing_features.mmap2;
714 	attr->comm  = track;
715 
716 	if (opts->sample_transaction)
717 		perf_evsel__set_sample_bit(evsel, TRANSACTION);
718 
719 	/*
720 	 * XXX see the function comment above
721 	 *
722 	 * Disabling only independent events or group leaders,
723 	 * keeping group members enabled.
724 	 */
725 	if (perf_evsel__is_group_leader(evsel))
726 		attr->disabled = 1;
727 
728 	/*
729 	 * Setting enable_on_exec for independent events and
730 	 * group leaders for traced executed by perf.
731 	 */
732 	if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
733 		!opts->initial_delay)
734 		attr->enable_on_exec = 1;
735 
736 	if (evsel->immediate) {
737 		attr->disabled = 0;
738 		attr->enable_on_exec = 0;
739 	}
740 }
741 
742 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
743 {
744 	int cpu, thread;
745 
746 	if (evsel->system_wide)
747 		nthreads = 1;
748 
749 	evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
750 
751 	if (evsel->fd) {
752 		for (cpu = 0; cpu < ncpus; cpu++) {
753 			for (thread = 0; thread < nthreads; thread++) {
754 				FD(evsel, cpu, thread) = -1;
755 			}
756 		}
757 	}
758 
759 	return evsel->fd != NULL ? 0 : -ENOMEM;
760 }
761 
762 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
763 			  int ioc,  void *arg)
764 {
765 	int cpu, thread;
766 
767 	if (evsel->system_wide)
768 		nthreads = 1;
769 
770 	for (cpu = 0; cpu < ncpus; cpu++) {
771 		for (thread = 0; thread < nthreads; thread++) {
772 			int fd = FD(evsel, cpu, thread),
773 			    err = ioctl(fd, ioc, arg);
774 
775 			if (err)
776 				return err;
777 		}
778 	}
779 
780 	return 0;
781 }
782 
783 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
784 			   const char *filter)
785 {
786 	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
787 				     PERF_EVENT_IOC_SET_FILTER,
788 				     (void *)filter);
789 }
790 
791 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
792 {
793 	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
794 				     PERF_EVENT_IOC_ENABLE,
795 				     0);
796 }
797 
798 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
799 {
800 	if (evsel->system_wide)
801 		nthreads = 1;
802 
803 	evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
804 	if (evsel->sample_id == NULL)
805 		return -ENOMEM;
806 
807 	evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
808 	if (evsel->id == NULL) {
809 		xyarray__delete(evsel->sample_id);
810 		evsel->sample_id = NULL;
811 		return -ENOMEM;
812 	}
813 
814 	return 0;
815 }
816 
817 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
818 {
819 	memset(evsel->counts, 0, (sizeof(*evsel->counts) +
820 				 (ncpus * sizeof(struct perf_counts_values))));
821 }
822 
823 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
824 {
825 	evsel->counts = zalloc((sizeof(*evsel->counts) +
826 				(ncpus * sizeof(struct perf_counts_values))));
827 	return evsel->counts != NULL ? 0 : -ENOMEM;
828 }
829 
830 static void perf_evsel__free_fd(struct perf_evsel *evsel)
831 {
832 	xyarray__delete(evsel->fd);
833 	evsel->fd = NULL;
834 }
835 
836 static void perf_evsel__free_id(struct perf_evsel *evsel)
837 {
838 	xyarray__delete(evsel->sample_id);
839 	evsel->sample_id = NULL;
840 	zfree(&evsel->id);
841 }
842 
843 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
844 {
845 	int cpu, thread;
846 
847 	if (evsel->system_wide)
848 		nthreads = 1;
849 
850 	for (cpu = 0; cpu < ncpus; cpu++)
851 		for (thread = 0; thread < nthreads; ++thread) {
852 			close(FD(evsel, cpu, thread));
853 			FD(evsel, cpu, thread) = -1;
854 		}
855 }
856 
857 void perf_evsel__free_counts(struct perf_evsel *evsel)
858 {
859 	zfree(&evsel->counts);
860 }
861 
862 void perf_evsel__exit(struct perf_evsel *evsel)
863 {
864 	assert(list_empty(&evsel->node));
865 	perf_evsel__free_fd(evsel);
866 	perf_evsel__free_id(evsel);
867 	close_cgroup(evsel->cgrp);
868 	zfree(&evsel->group_name);
869 	zfree(&evsel->name);
870 	perf_evsel__object.fini(evsel);
871 }
872 
873 void perf_evsel__delete(struct perf_evsel *evsel)
874 {
875 	perf_evsel__exit(evsel);
876 	free(evsel);
877 }
878 
879 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu,
880 				struct perf_counts_values *count)
881 {
882 	struct perf_counts_values tmp;
883 
884 	if (!evsel->prev_raw_counts)
885 		return;
886 
887 	if (cpu == -1) {
888 		tmp = evsel->prev_raw_counts->aggr;
889 		evsel->prev_raw_counts->aggr = *count;
890 	} else {
891 		tmp = evsel->prev_raw_counts->cpu[cpu];
892 		evsel->prev_raw_counts->cpu[cpu] = *count;
893 	}
894 
895 	count->val = count->val - tmp.val;
896 	count->ena = count->ena - tmp.ena;
897 	count->run = count->run - tmp.run;
898 }
899 
900 void perf_counts_values__scale(struct perf_counts_values *count,
901 			       bool scale, s8 *pscaled)
902 {
903 	s8 scaled = 0;
904 
905 	if (scale) {
906 		if (count->run == 0) {
907 			scaled = -1;
908 			count->val = 0;
909 		} else if (count->run < count->ena) {
910 			scaled = 1;
911 			count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
912 		}
913 	} else
914 		count->ena = count->run = 0;
915 
916 	if (pscaled)
917 		*pscaled = scaled;
918 }
919 
920 int perf_evsel__read_cb(struct perf_evsel *evsel, int cpu, int thread,
921 			perf_evsel__read_cb_t cb)
922 {
923 	struct perf_counts_values count;
924 
925 	memset(&count, 0, sizeof(count));
926 
927 	if (FD(evsel, cpu, thread) < 0)
928 		return -EINVAL;
929 
930 	if (readn(FD(evsel, cpu, thread), &count, sizeof(count)) < 0)
931 		return -errno;
932 
933 	return cb(evsel, cpu, thread, &count);
934 }
935 
936 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
937 			      int cpu, int thread, bool scale)
938 {
939 	struct perf_counts_values count;
940 	size_t nv = scale ? 3 : 1;
941 
942 	if (FD(evsel, cpu, thread) < 0)
943 		return -EINVAL;
944 
945 	if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
946 		return -ENOMEM;
947 
948 	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
949 		return -errno;
950 
951 	perf_evsel__compute_deltas(evsel, cpu, &count);
952 	perf_counts_values__scale(&count, scale, NULL);
953 	evsel->counts->cpu[cpu] = count;
954 	return 0;
955 }
956 
957 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
958 {
959 	struct perf_evsel *leader = evsel->leader;
960 	int fd;
961 
962 	if (perf_evsel__is_group_leader(evsel))
963 		return -1;
964 
965 	/*
966 	 * Leader must be already processed/open,
967 	 * if not it's a bug.
968 	 */
969 	BUG_ON(!leader->fd);
970 
971 	fd = FD(leader, cpu, thread);
972 	BUG_ON(fd == -1);
973 
974 	return fd;
975 }
976 
977 #define __PRINT_ATTR(fmt, cast, field)  \
978 	fprintf(fp, "  %-19s "fmt"\n", #field, cast attr->field)
979 
980 #define PRINT_ATTR_U32(field)  __PRINT_ATTR("%u" , , field)
981 #define PRINT_ATTR_X32(field)  __PRINT_ATTR("%#x", , field)
982 #define PRINT_ATTR_U64(field)  __PRINT_ATTR("%" PRIu64, (uint64_t), field)
983 #define PRINT_ATTR_X64(field)  __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
984 
985 #define PRINT_ATTR2N(name1, field1, name2, field2)	\
986 	fprintf(fp, "  %-19s %u    %-19s %u\n",		\
987 	name1, attr->field1, name2, attr->field2)
988 
989 #define PRINT_ATTR2(field1, field2) \
990 	PRINT_ATTR2N(#field1, field1, #field2, field2)
991 
992 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
993 {
994 	size_t ret = 0;
995 
996 	ret += fprintf(fp, "%.60s\n", graph_dotted_line);
997 	ret += fprintf(fp, "perf_event_attr:\n");
998 
999 	ret += PRINT_ATTR_U32(type);
1000 	ret += PRINT_ATTR_U32(size);
1001 	ret += PRINT_ATTR_X64(config);
1002 	ret += PRINT_ATTR_U64(sample_period);
1003 	ret += PRINT_ATTR_U64(sample_freq);
1004 	ret += PRINT_ATTR_X64(sample_type);
1005 	ret += PRINT_ATTR_X64(read_format);
1006 
1007 	ret += PRINT_ATTR2(disabled, inherit);
1008 	ret += PRINT_ATTR2(pinned, exclusive);
1009 	ret += PRINT_ATTR2(exclude_user, exclude_kernel);
1010 	ret += PRINT_ATTR2(exclude_hv, exclude_idle);
1011 	ret += PRINT_ATTR2(mmap, comm);
1012 	ret += PRINT_ATTR2(mmap2, comm_exec);
1013 	ret += PRINT_ATTR2(freq, inherit_stat);
1014 	ret += PRINT_ATTR2(enable_on_exec, task);
1015 	ret += PRINT_ATTR2(watermark, precise_ip);
1016 	ret += PRINT_ATTR2(mmap_data, sample_id_all);
1017 	ret += PRINT_ATTR2(exclude_host, exclude_guest);
1018 	ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
1019 			    "excl.callchain_user", exclude_callchain_user);
1020 
1021 	ret += PRINT_ATTR_U32(wakeup_events);
1022 	ret += PRINT_ATTR_U32(wakeup_watermark);
1023 	ret += PRINT_ATTR_X32(bp_type);
1024 	ret += PRINT_ATTR_X64(bp_addr);
1025 	ret += PRINT_ATTR_X64(config1);
1026 	ret += PRINT_ATTR_U64(bp_len);
1027 	ret += PRINT_ATTR_X64(config2);
1028 	ret += PRINT_ATTR_X64(branch_sample_type);
1029 	ret += PRINT_ATTR_X64(sample_regs_user);
1030 	ret += PRINT_ATTR_U32(sample_stack_user);
1031 	ret += PRINT_ATTR_X64(sample_regs_intr);
1032 
1033 	ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1034 
1035 	return ret;
1036 }
1037 
1038 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1039 			      struct thread_map *threads)
1040 {
1041 	int cpu, thread, nthreads;
1042 	unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1043 	int pid = -1, err;
1044 	enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1045 
1046 	if (evsel->system_wide)
1047 		nthreads = 1;
1048 	else
1049 		nthreads = threads->nr;
1050 
1051 	if (evsel->fd == NULL &&
1052 	    perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1053 		return -ENOMEM;
1054 
1055 	if (evsel->cgrp) {
1056 		flags |= PERF_FLAG_PID_CGROUP;
1057 		pid = evsel->cgrp->fd;
1058 	}
1059 
1060 fallback_missing_features:
1061 	if (perf_missing_features.cloexec)
1062 		flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1063 	if (perf_missing_features.mmap2)
1064 		evsel->attr.mmap2 = 0;
1065 	if (perf_missing_features.exclude_guest)
1066 		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1067 retry_sample_id:
1068 	if (perf_missing_features.sample_id_all)
1069 		evsel->attr.sample_id_all = 0;
1070 
1071 	if (verbose >= 2)
1072 		perf_event_attr__fprintf(&evsel->attr, stderr);
1073 
1074 	for (cpu = 0; cpu < cpus->nr; cpu++) {
1075 
1076 		for (thread = 0; thread < nthreads; thread++) {
1077 			int group_fd;
1078 
1079 			if (!evsel->cgrp && !evsel->system_wide)
1080 				pid = threads->map[thread];
1081 
1082 			group_fd = get_group_fd(evsel, cpu, thread);
1083 retry_open:
1084 			pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1085 				  pid, cpus->map[cpu], group_fd, flags);
1086 
1087 			FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1088 								     pid,
1089 								     cpus->map[cpu],
1090 								     group_fd, flags);
1091 			if (FD(evsel, cpu, thread) < 0) {
1092 				err = -errno;
1093 				pr_debug2("sys_perf_event_open failed, error %d\n",
1094 					  err);
1095 				goto try_fallback;
1096 			}
1097 			set_rlimit = NO_CHANGE;
1098 		}
1099 	}
1100 
1101 	return 0;
1102 
1103 try_fallback:
1104 	/*
1105 	 * perf stat needs between 5 and 22 fds per CPU. When we run out
1106 	 * of them try to increase the limits.
1107 	 */
1108 	if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1109 		struct rlimit l;
1110 		int old_errno = errno;
1111 
1112 		if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1113 			if (set_rlimit == NO_CHANGE)
1114 				l.rlim_cur = l.rlim_max;
1115 			else {
1116 				l.rlim_cur = l.rlim_max + 1000;
1117 				l.rlim_max = l.rlim_cur;
1118 			}
1119 			if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1120 				set_rlimit++;
1121 				errno = old_errno;
1122 				goto retry_open;
1123 			}
1124 		}
1125 		errno = old_errno;
1126 	}
1127 
1128 	if (err != -EINVAL || cpu > 0 || thread > 0)
1129 		goto out_close;
1130 
1131 	if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1132 		perf_missing_features.cloexec = true;
1133 		goto fallback_missing_features;
1134 	} else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1135 		perf_missing_features.mmap2 = true;
1136 		goto fallback_missing_features;
1137 	} else if (!perf_missing_features.exclude_guest &&
1138 		   (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1139 		perf_missing_features.exclude_guest = true;
1140 		goto fallback_missing_features;
1141 	} else if (!perf_missing_features.sample_id_all) {
1142 		perf_missing_features.sample_id_all = true;
1143 		goto retry_sample_id;
1144 	}
1145 
1146 out_close:
1147 	do {
1148 		while (--thread >= 0) {
1149 			close(FD(evsel, cpu, thread));
1150 			FD(evsel, cpu, thread) = -1;
1151 		}
1152 		thread = nthreads;
1153 	} while (--cpu >= 0);
1154 	return err;
1155 }
1156 
1157 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1158 {
1159 	if (evsel->fd == NULL)
1160 		return;
1161 
1162 	perf_evsel__close_fd(evsel, ncpus, nthreads);
1163 	perf_evsel__free_fd(evsel);
1164 }
1165 
1166 static struct {
1167 	struct cpu_map map;
1168 	int cpus[1];
1169 } empty_cpu_map = {
1170 	.map.nr	= 1,
1171 	.cpus	= { -1, },
1172 };
1173 
1174 static struct {
1175 	struct thread_map map;
1176 	int threads[1];
1177 } empty_thread_map = {
1178 	.map.nr	 = 1,
1179 	.threads = { -1, },
1180 };
1181 
1182 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1183 		     struct thread_map *threads)
1184 {
1185 	if (cpus == NULL) {
1186 		/* Work around old compiler warnings about strict aliasing */
1187 		cpus = &empty_cpu_map.map;
1188 	}
1189 
1190 	if (threads == NULL)
1191 		threads = &empty_thread_map.map;
1192 
1193 	return __perf_evsel__open(evsel, cpus, threads);
1194 }
1195 
1196 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1197 			     struct cpu_map *cpus)
1198 {
1199 	return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1200 }
1201 
1202 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1203 				struct thread_map *threads)
1204 {
1205 	return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1206 }
1207 
1208 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1209 				       const union perf_event *event,
1210 				       struct perf_sample *sample)
1211 {
1212 	u64 type = evsel->attr.sample_type;
1213 	const u64 *array = event->sample.array;
1214 	bool swapped = evsel->needs_swap;
1215 	union u64_swap u;
1216 
1217 	array += ((event->header.size -
1218 		   sizeof(event->header)) / sizeof(u64)) - 1;
1219 
1220 	if (type & PERF_SAMPLE_IDENTIFIER) {
1221 		sample->id = *array;
1222 		array--;
1223 	}
1224 
1225 	if (type & PERF_SAMPLE_CPU) {
1226 		u.val64 = *array;
1227 		if (swapped) {
1228 			/* undo swap of u64, then swap on individual u32s */
1229 			u.val64 = bswap_64(u.val64);
1230 			u.val32[0] = bswap_32(u.val32[0]);
1231 		}
1232 
1233 		sample->cpu = u.val32[0];
1234 		array--;
1235 	}
1236 
1237 	if (type & PERF_SAMPLE_STREAM_ID) {
1238 		sample->stream_id = *array;
1239 		array--;
1240 	}
1241 
1242 	if (type & PERF_SAMPLE_ID) {
1243 		sample->id = *array;
1244 		array--;
1245 	}
1246 
1247 	if (type & PERF_SAMPLE_TIME) {
1248 		sample->time = *array;
1249 		array--;
1250 	}
1251 
1252 	if (type & PERF_SAMPLE_TID) {
1253 		u.val64 = *array;
1254 		if (swapped) {
1255 			/* undo swap of u64, then swap on individual u32s */
1256 			u.val64 = bswap_64(u.val64);
1257 			u.val32[0] = bswap_32(u.val32[0]);
1258 			u.val32[1] = bswap_32(u.val32[1]);
1259 		}
1260 
1261 		sample->pid = u.val32[0];
1262 		sample->tid = u.val32[1];
1263 		array--;
1264 	}
1265 
1266 	return 0;
1267 }
1268 
1269 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1270 			    u64 size)
1271 {
1272 	return size > max_size || offset + size > endp;
1273 }
1274 
1275 #define OVERFLOW_CHECK(offset, size, max_size)				\
1276 	do {								\
1277 		if (overflow(endp, (max_size), (offset), (size)))	\
1278 			return -EFAULT;					\
1279 	} while (0)
1280 
1281 #define OVERFLOW_CHECK_u64(offset) \
1282 	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1283 
1284 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1285 			     struct perf_sample *data)
1286 {
1287 	u64 type = evsel->attr.sample_type;
1288 	bool swapped = evsel->needs_swap;
1289 	const u64 *array;
1290 	u16 max_size = event->header.size;
1291 	const void *endp = (void *)event + max_size;
1292 	u64 sz;
1293 
1294 	/*
1295 	 * used for cross-endian analysis. See git commit 65014ab3
1296 	 * for why this goofiness is needed.
1297 	 */
1298 	union u64_swap u;
1299 
1300 	memset(data, 0, sizeof(*data));
1301 	data->cpu = data->pid = data->tid = -1;
1302 	data->stream_id = data->id = data->time = -1ULL;
1303 	data->period = evsel->attr.sample_period;
1304 	data->weight = 0;
1305 
1306 	if (event->header.type != PERF_RECORD_SAMPLE) {
1307 		if (!evsel->attr.sample_id_all)
1308 			return 0;
1309 		return perf_evsel__parse_id_sample(evsel, event, data);
1310 	}
1311 
1312 	array = event->sample.array;
1313 
1314 	/*
1315 	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1316 	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1317 	 * check the format does not go past the end of the event.
1318 	 */
1319 	if (evsel->sample_size + sizeof(event->header) > event->header.size)
1320 		return -EFAULT;
1321 
1322 	data->id = -1ULL;
1323 	if (type & PERF_SAMPLE_IDENTIFIER) {
1324 		data->id = *array;
1325 		array++;
1326 	}
1327 
1328 	if (type & PERF_SAMPLE_IP) {
1329 		data->ip = *array;
1330 		array++;
1331 	}
1332 
1333 	if (type & PERF_SAMPLE_TID) {
1334 		u.val64 = *array;
1335 		if (swapped) {
1336 			/* undo swap of u64, then swap on individual u32s */
1337 			u.val64 = bswap_64(u.val64);
1338 			u.val32[0] = bswap_32(u.val32[0]);
1339 			u.val32[1] = bswap_32(u.val32[1]);
1340 		}
1341 
1342 		data->pid = u.val32[0];
1343 		data->tid = u.val32[1];
1344 		array++;
1345 	}
1346 
1347 	if (type & PERF_SAMPLE_TIME) {
1348 		data->time = *array;
1349 		array++;
1350 	}
1351 
1352 	data->addr = 0;
1353 	if (type & PERF_SAMPLE_ADDR) {
1354 		data->addr = *array;
1355 		array++;
1356 	}
1357 
1358 	if (type & PERF_SAMPLE_ID) {
1359 		data->id = *array;
1360 		array++;
1361 	}
1362 
1363 	if (type & PERF_SAMPLE_STREAM_ID) {
1364 		data->stream_id = *array;
1365 		array++;
1366 	}
1367 
1368 	if (type & PERF_SAMPLE_CPU) {
1369 
1370 		u.val64 = *array;
1371 		if (swapped) {
1372 			/* undo swap of u64, then swap on individual u32s */
1373 			u.val64 = bswap_64(u.val64);
1374 			u.val32[0] = bswap_32(u.val32[0]);
1375 		}
1376 
1377 		data->cpu = u.val32[0];
1378 		array++;
1379 	}
1380 
1381 	if (type & PERF_SAMPLE_PERIOD) {
1382 		data->period = *array;
1383 		array++;
1384 	}
1385 
1386 	if (type & PERF_SAMPLE_READ) {
1387 		u64 read_format = evsel->attr.read_format;
1388 
1389 		OVERFLOW_CHECK_u64(array);
1390 		if (read_format & PERF_FORMAT_GROUP)
1391 			data->read.group.nr = *array;
1392 		else
1393 			data->read.one.value = *array;
1394 
1395 		array++;
1396 
1397 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1398 			OVERFLOW_CHECK_u64(array);
1399 			data->read.time_enabled = *array;
1400 			array++;
1401 		}
1402 
1403 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1404 			OVERFLOW_CHECK_u64(array);
1405 			data->read.time_running = *array;
1406 			array++;
1407 		}
1408 
1409 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1410 		if (read_format & PERF_FORMAT_GROUP) {
1411 			const u64 max_group_nr = UINT64_MAX /
1412 					sizeof(struct sample_read_value);
1413 
1414 			if (data->read.group.nr > max_group_nr)
1415 				return -EFAULT;
1416 			sz = data->read.group.nr *
1417 			     sizeof(struct sample_read_value);
1418 			OVERFLOW_CHECK(array, sz, max_size);
1419 			data->read.group.values =
1420 					(struct sample_read_value *)array;
1421 			array = (void *)array + sz;
1422 		} else {
1423 			OVERFLOW_CHECK_u64(array);
1424 			data->read.one.id = *array;
1425 			array++;
1426 		}
1427 	}
1428 
1429 	if (type & PERF_SAMPLE_CALLCHAIN) {
1430 		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1431 
1432 		OVERFLOW_CHECK_u64(array);
1433 		data->callchain = (struct ip_callchain *)array++;
1434 		if (data->callchain->nr > max_callchain_nr)
1435 			return -EFAULT;
1436 		sz = data->callchain->nr * sizeof(u64);
1437 		OVERFLOW_CHECK(array, sz, max_size);
1438 		array = (void *)array + sz;
1439 	}
1440 
1441 	if (type & PERF_SAMPLE_RAW) {
1442 		OVERFLOW_CHECK_u64(array);
1443 		u.val64 = *array;
1444 		if (WARN_ONCE(swapped,
1445 			      "Endianness of raw data not corrected!\n")) {
1446 			/* undo swap of u64, then swap on individual u32s */
1447 			u.val64 = bswap_64(u.val64);
1448 			u.val32[0] = bswap_32(u.val32[0]);
1449 			u.val32[1] = bswap_32(u.val32[1]);
1450 		}
1451 		data->raw_size = u.val32[0];
1452 		array = (void *)array + sizeof(u32);
1453 
1454 		OVERFLOW_CHECK(array, data->raw_size, max_size);
1455 		data->raw_data = (void *)array;
1456 		array = (void *)array + data->raw_size;
1457 	}
1458 
1459 	if (type & PERF_SAMPLE_BRANCH_STACK) {
1460 		const u64 max_branch_nr = UINT64_MAX /
1461 					  sizeof(struct branch_entry);
1462 
1463 		OVERFLOW_CHECK_u64(array);
1464 		data->branch_stack = (struct branch_stack *)array++;
1465 
1466 		if (data->branch_stack->nr > max_branch_nr)
1467 			return -EFAULT;
1468 		sz = data->branch_stack->nr * sizeof(struct branch_entry);
1469 		OVERFLOW_CHECK(array, sz, max_size);
1470 		array = (void *)array + sz;
1471 	}
1472 
1473 	if (type & PERF_SAMPLE_REGS_USER) {
1474 		OVERFLOW_CHECK_u64(array);
1475 		data->user_regs.abi = *array;
1476 		array++;
1477 
1478 		if (data->user_regs.abi) {
1479 			u64 mask = evsel->attr.sample_regs_user;
1480 
1481 			sz = hweight_long(mask) * sizeof(u64);
1482 			OVERFLOW_CHECK(array, sz, max_size);
1483 			data->user_regs.mask = mask;
1484 			data->user_regs.regs = (u64 *)array;
1485 			array = (void *)array + sz;
1486 		}
1487 	}
1488 
1489 	if (type & PERF_SAMPLE_STACK_USER) {
1490 		OVERFLOW_CHECK_u64(array);
1491 		sz = *array++;
1492 
1493 		data->user_stack.offset = ((char *)(array - 1)
1494 					  - (char *) event);
1495 
1496 		if (!sz) {
1497 			data->user_stack.size = 0;
1498 		} else {
1499 			OVERFLOW_CHECK(array, sz, max_size);
1500 			data->user_stack.data = (char *)array;
1501 			array = (void *)array + sz;
1502 			OVERFLOW_CHECK_u64(array);
1503 			data->user_stack.size = *array++;
1504 			if (WARN_ONCE(data->user_stack.size > sz,
1505 				      "user stack dump failure\n"))
1506 				return -EFAULT;
1507 		}
1508 	}
1509 
1510 	data->weight = 0;
1511 	if (type & PERF_SAMPLE_WEIGHT) {
1512 		OVERFLOW_CHECK_u64(array);
1513 		data->weight = *array;
1514 		array++;
1515 	}
1516 
1517 	data->data_src = PERF_MEM_DATA_SRC_NONE;
1518 	if (type & PERF_SAMPLE_DATA_SRC) {
1519 		OVERFLOW_CHECK_u64(array);
1520 		data->data_src = *array;
1521 		array++;
1522 	}
1523 
1524 	data->transaction = 0;
1525 	if (type & PERF_SAMPLE_TRANSACTION) {
1526 		OVERFLOW_CHECK_u64(array);
1527 		data->transaction = *array;
1528 		array++;
1529 	}
1530 
1531 	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1532 	if (type & PERF_SAMPLE_REGS_INTR) {
1533 		OVERFLOW_CHECK_u64(array);
1534 		data->intr_regs.abi = *array;
1535 		array++;
1536 
1537 		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1538 			u64 mask = evsel->attr.sample_regs_intr;
1539 
1540 			sz = hweight_long(mask) * sizeof(u64);
1541 			OVERFLOW_CHECK(array, sz, max_size);
1542 			data->intr_regs.mask = mask;
1543 			data->intr_regs.regs = (u64 *)array;
1544 			array = (void *)array + sz;
1545 		}
1546 	}
1547 
1548 	return 0;
1549 }
1550 
1551 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1552 				     u64 read_format)
1553 {
1554 	size_t sz, result = sizeof(struct sample_event);
1555 
1556 	if (type & PERF_SAMPLE_IDENTIFIER)
1557 		result += sizeof(u64);
1558 
1559 	if (type & PERF_SAMPLE_IP)
1560 		result += sizeof(u64);
1561 
1562 	if (type & PERF_SAMPLE_TID)
1563 		result += sizeof(u64);
1564 
1565 	if (type & PERF_SAMPLE_TIME)
1566 		result += sizeof(u64);
1567 
1568 	if (type & PERF_SAMPLE_ADDR)
1569 		result += sizeof(u64);
1570 
1571 	if (type & PERF_SAMPLE_ID)
1572 		result += sizeof(u64);
1573 
1574 	if (type & PERF_SAMPLE_STREAM_ID)
1575 		result += sizeof(u64);
1576 
1577 	if (type & PERF_SAMPLE_CPU)
1578 		result += sizeof(u64);
1579 
1580 	if (type & PERF_SAMPLE_PERIOD)
1581 		result += sizeof(u64);
1582 
1583 	if (type & PERF_SAMPLE_READ) {
1584 		result += sizeof(u64);
1585 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1586 			result += sizeof(u64);
1587 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1588 			result += sizeof(u64);
1589 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1590 		if (read_format & PERF_FORMAT_GROUP) {
1591 			sz = sample->read.group.nr *
1592 			     sizeof(struct sample_read_value);
1593 			result += sz;
1594 		} else {
1595 			result += sizeof(u64);
1596 		}
1597 	}
1598 
1599 	if (type & PERF_SAMPLE_CALLCHAIN) {
1600 		sz = (sample->callchain->nr + 1) * sizeof(u64);
1601 		result += sz;
1602 	}
1603 
1604 	if (type & PERF_SAMPLE_RAW) {
1605 		result += sizeof(u32);
1606 		result += sample->raw_size;
1607 	}
1608 
1609 	if (type & PERF_SAMPLE_BRANCH_STACK) {
1610 		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1611 		sz += sizeof(u64);
1612 		result += sz;
1613 	}
1614 
1615 	if (type & PERF_SAMPLE_REGS_USER) {
1616 		if (sample->user_regs.abi) {
1617 			result += sizeof(u64);
1618 			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1619 			result += sz;
1620 		} else {
1621 			result += sizeof(u64);
1622 		}
1623 	}
1624 
1625 	if (type & PERF_SAMPLE_STACK_USER) {
1626 		sz = sample->user_stack.size;
1627 		result += sizeof(u64);
1628 		if (sz) {
1629 			result += sz;
1630 			result += sizeof(u64);
1631 		}
1632 	}
1633 
1634 	if (type & PERF_SAMPLE_WEIGHT)
1635 		result += sizeof(u64);
1636 
1637 	if (type & PERF_SAMPLE_DATA_SRC)
1638 		result += sizeof(u64);
1639 
1640 	if (type & PERF_SAMPLE_TRANSACTION)
1641 		result += sizeof(u64);
1642 
1643 	if (type & PERF_SAMPLE_REGS_INTR) {
1644 		if (sample->intr_regs.abi) {
1645 			result += sizeof(u64);
1646 			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1647 			result += sz;
1648 		} else {
1649 			result += sizeof(u64);
1650 		}
1651 	}
1652 
1653 	return result;
1654 }
1655 
1656 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1657 				  u64 read_format,
1658 				  const struct perf_sample *sample,
1659 				  bool swapped)
1660 {
1661 	u64 *array;
1662 	size_t sz;
1663 	/*
1664 	 * used for cross-endian analysis. See git commit 65014ab3
1665 	 * for why this goofiness is needed.
1666 	 */
1667 	union u64_swap u;
1668 
1669 	array = event->sample.array;
1670 
1671 	if (type & PERF_SAMPLE_IDENTIFIER) {
1672 		*array = sample->id;
1673 		array++;
1674 	}
1675 
1676 	if (type & PERF_SAMPLE_IP) {
1677 		*array = sample->ip;
1678 		array++;
1679 	}
1680 
1681 	if (type & PERF_SAMPLE_TID) {
1682 		u.val32[0] = sample->pid;
1683 		u.val32[1] = sample->tid;
1684 		if (swapped) {
1685 			/*
1686 			 * Inverse of what is done in perf_evsel__parse_sample
1687 			 */
1688 			u.val32[0] = bswap_32(u.val32[0]);
1689 			u.val32[1] = bswap_32(u.val32[1]);
1690 			u.val64 = bswap_64(u.val64);
1691 		}
1692 
1693 		*array = u.val64;
1694 		array++;
1695 	}
1696 
1697 	if (type & PERF_SAMPLE_TIME) {
1698 		*array = sample->time;
1699 		array++;
1700 	}
1701 
1702 	if (type & PERF_SAMPLE_ADDR) {
1703 		*array = sample->addr;
1704 		array++;
1705 	}
1706 
1707 	if (type & PERF_SAMPLE_ID) {
1708 		*array = sample->id;
1709 		array++;
1710 	}
1711 
1712 	if (type & PERF_SAMPLE_STREAM_ID) {
1713 		*array = sample->stream_id;
1714 		array++;
1715 	}
1716 
1717 	if (type & PERF_SAMPLE_CPU) {
1718 		u.val32[0] = sample->cpu;
1719 		if (swapped) {
1720 			/*
1721 			 * Inverse of what is done in perf_evsel__parse_sample
1722 			 */
1723 			u.val32[0] = bswap_32(u.val32[0]);
1724 			u.val64 = bswap_64(u.val64);
1725 		}
1726 		*array = u.val64;
1727 		array++;
1728 	}
1729 
1730 	if (type & PERF_SAMPLE_PERIOD) {
1731 		*array = sample->period;
1732 		array++;
1733 	}
1734 
1735 	if (type & PERF_SAMPLE_READ) {
1736 		if (read_format & PERF_FORMAT_GROUP)
1737 			*array = sample->read.group.nr;
1738 		else
1739 			*array = sample->read.one.value;
1740 		array++;
1741 
1742 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1743 			*array = sample->read.time_enabled;
1744 			array++;
1745 		}
1746 
1747 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1748 			*array = sample->read.time_running;
1749 			array++;
1750 		}
1751 
1752 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1753 		if (read_format & PERF_FORMAT_GROUP) {
1754 			sz = sample->read.group.nr *
1755 			     sizeof(struct sample_read_value);
1756 			memcpy(array, sample->read.group.values, sz);
1757 			array = (void *)array + sz;
1758 		} else {
1759 			*array = sample->read.one.id;
1760 			array++;
1761 		}
1762 	}
1763 
1764 	if (type & PERF_SAMPLE_CALLCHAIN) {
1765 		sz = (sample->callchain->nr + 1) * sizeof(u64);
1766 		memcpy(array, sample->callchain, sz);
1767 		array = (void *)array + sz;
1768 	}
1769 
1770 	if (type & PERF_SAMPLE_RAW) {
1771 		u.val32[0] = sample->raw_size;
1772 		if (WARN_ONCE(swapped,
1773 			      "Endianness of raw data not corrected!\n")) {
1774 			/*
1775 			 * Inverse of what is done in perf_evsel__parse_sample
1776 			 */
1777 			u.val32[0] = bswap_32(u.val32[0]);
1778 			u.val32[1] = bswap_32(u.val32[1]);
1779 			u.val64 = bswap_64(u.val64);
1780 		}
1781 		*array = u.val64;
1782 		array = (void *)array + sizeof(u32);
1783 
1784 		memcpy(array, sample->raw_data, sample->raw_size);
1785 		array = (void *)array + sample->raw_size;
1786 	}
1787 
1788 	if (type & PERF_SAMPLE_BRANCH_STACK) {
1789 		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1790 		sz += sizeof(u64);
1791 		memcpy(array, sample->branch_stack, sz);
1792 		array = (void *)array + sz;
1793 	}
1794 
1795 	if (type & PERF_SAMPLE_REGS_USER) {
1796 		if (sample->user_regs.abi) {
1797 			*array++ = sample->user_regs.abi;
1798 			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1799 			memcpy(array, sample->user_regs.regs, sz);
1800 			array = (void *)array + sz;
1801 		} else {
1802 			*array++ = 0;
1803 		}
1804 	}
1805 
1806 	if (type & PERF_SAMPLE_STACK_USER) {
1807 		sz = sample->user_stack.size;
1808 		*array++ = sz;
1809 		if (sz) {
1810 			memcpy(array, sample->user_stack.data, sz);
1811 			array = (void *)array + sz;
1812 			*array++ = sz;
1813 		}
1814 	}
1815 
1816 	if (type & PERF_SAMPLE_WEIGHT) {
1817 		*array = sample->weight;
1818 		array++;
1819 	}
1820 
1821 	if (type & PERF_SAMPLE_DATA_SRC) {
1822 		*array = sample->data_src;
1823 		array++;
1824 	}
1825 
1826 	if (type & PERF_SAMPLE_TRANSACTION) {
1827 		*array = sample->transaction;
1828 		array++;
1829 	}
1830 
1831 	if (type & PERF_SAMPLE_REGS_INTR) {
1832 		if (sample->intr_regs.abi) {
1833 			*array++ = sample->intr_regs.abi;
1834 			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1835 			memcpy(array, sample->intr_regs.regs, sz);
1836 			array = (void *)array + sz;
1837 		} else {
1838 			*array++ = 0;
1839 		}
1840 	}
1841 
1842 	return 0;
1843 }
1844 
1845 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1846 {
1847 	return pevent_find_field(evsel->tp_format, name);
1848 }
1849 
1850 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1851 			 const char *name)
1852 {
1853 	struct format_field *field = perf_evsel__field(evsel, name);
1854 	int offset;
1855 
1856 	if (!field)
1857 		return NULL;
1858 
1859 	offset = field->offset;
1860 
1861 	if (field->flags & FIELD_IS_DYNAMIC) {
1862 		offset = *(int *)(sample->raw_data + field->offset);
1863 		offset &= 0xffff;
1864 	}
1865 
1866 	return sample->raw_data + offset;
1867 }
1868 
1869 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1870 		       const char *name)
1871 {
1872 	struct format_field *field = perf_evsel__field(evsel, name);
1873 	void *ptr;
1874 	u64 value;
1875 
1876 	if (!field)
1877 		return 0;
1878 
1879 	ptr = sample->raw_data + field->offset;
1880 
1881 	switch (field->size) {
1882 	case 1:
1883 		return *(u8 *)ptr;
1884 	case 2:
1885 		value = *(u16 *)ptr;
1886 		break;
1887 	case 4:
1888 		value = *(u32 *)ptr;
1889 		break;
1890 	case 8:
1891 		value = *(u64 *)ptr;
1892 		break;
1893 	default:
1894 		return 0;
1895 	}
1896 
1897 	if (!evsel->needs_swap)
1898 		return value;
1899 
1900 	switch (field->size) {
1901 	case 2:
1902 		return bswap_16(value);
1903 	case 4:
1904 		return bswap_32(value);
1905 	case 8:
1906 		return bswap_64(value);
1907 	default:
1908 		return 0;
1909 	}
1910 
1911 	return 0;
1912 }
1913 
1914 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1915 {
1916 	va_list args;
1917 	int ret = 0;
1918 
1919 	if (!*first) {
1920 		ret += fprintf(fp, ",");
1921 	} else {
1922 		ret += fprintf(fp, ":");
1923 		*first = false;
1924 	}
1925 
1926 	va_start(args, fmt);
1927 	ret += vfprintf(fp, fmt, args);
1928 	va_end(args);
1929 	return ret;
1930 }
1931 
1932 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1933 {
1934 	if (value == 0)
1935 		return 0;
1936 
1937 	return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1938 }
1939 
1940 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1941 
1942 struct bit_names {
1943 	int bit;
1944 	const char *name;
1945 };
1946 
1947 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1948 			 struct bit_names *bits, bool *first)
1949 {
1950 	int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1951 	bool first_bit = true;
1952 
1953 	do {
1954 		if (value & bits[i].bit) {
1955 			printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1956 			first_bit = false;
1957 		}
1958 	} while (bits[++i].name != NULL);
1959 
1960 	return printed;
1961 }
1962 
1963 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1964 {
1965 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1966 	struct bit_names bits[] = {
1967 		bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1968 		bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1969 		bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1970 		bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1971 		bit_name(IDENTIFIER), bit_name(REGS_INTR),
1972 		{ .name = NULL, }
1973 	};
1974 #undef bit_name
1975 	return bits__fprintf(fp, "sample_type", value, bits, first);
1976 }
1977 
1978 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1979 {
1980 #define bit_name(n) { PERF_FORMAT_##n, #n }
1981 	struct bit_names bits[] = {
1982 		bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1983 		bit_name(ID), bit_name(GROUP),
1984 		{ .name = NULL, }
1985 	};
1986 #undef bit_name
1987 	return bits__fprintf(fp, "read_format", value, bits, first);
1988 }
1989 
1990 int perf_evsel__fprintf(struct perf_evsel *evsel,
1991 			struct perf_attr_details *details, FILE *fp)
1992 {
1993 	bool first = true;
1994 	int printed = 0;
1995 
1996 	if (details->event_group) {
1997 		struct perf_evsel *pos;
1998 
1999 		if (!perf_evsel__is_group_leader(evsel))
2000 			return 0;
2001 
2002 		if (evsel->nr_members > 1)
2003 			printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2004 
2005 		printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2006 		for_each_group_member(pos, evsel)
2007 			printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2008 
2009 		if (evsel->nr_members > 1)
2010 			printed += fprintf(fp, "}");
2011 		goto out;
2012 	}
2013 
2014 	printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2015 
2016 	if (details->verbose || details->freq) {
2017 		printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2018 					 (u64)evsel->attr.sample_freq);
2019 	}
2020 
2021 	if (details->verbose) {
2022 		if_print(type);
2023 		if_print(config);
2024 		if_print(config1);
2025 		if_print(config2);
2026 		if_print(size);
2027 		printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
2028 		if (evsel->attr.read_format)
2029 			printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
2030 		if_print(disabled);
2031 		if_print(inherit);
2032 		if_print(pinned);
2033 		if_print(exclusive);
2034 		if_print(exclude_user);
2035 		if_print(exclude_kernel);
2036 		if_print(exclude_hv);
2037 		if_print(exclude_idle);
2038 		if_print(mmap);
2039 		if_print(mmap2);
2040 		if_print(comm);
2041 		if_print(comm_exec);
2042 		if_print(freq);
2043 		if_print(inherit_stat);
2044 		if_print(enable_on_exec);
2045 		if_print(task);
2046 		if_print(watermark);
2047 		if_print(precise_ip);
2048 		if_print(mmap_data);
2049 		if_print(sample_id_all);
2050 		if_print(exclude_host);
2051 		if_print(exclude_guest);
2052 		if_print(__reserved_1);
2053 		if_print(wakeup_events);
2054 		if_print(bp_type);
2055 		if_print(branch_sample_type);
2056 	}
2057 out:
2058 	fputc('\n', fp);
2059 	return ++printed;
2060 }
2061 
2062 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2063 			  char *msg, size_t msgsize)
2064 {
2065 	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2066 	    evsel->attr.type   == PERF_TYPE_HARDWARE &&
2067 	    evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2068 		/*
2069 		 * If it's cycles then fall back to hrtimer based
2070 		 * cpu-clock-tick sw counter, which is always available even if
2071 		 * no PMU support.
2072 		 *
2073 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2074 		 * b0a873e).
2075 		 */
2076 		scnprintf(msg, msgsize, "%s",
2077 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2078 
2079 		evsel->attr.type   = PERF_TYPE_SOFTWARE;
2080 		evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2081 
2082 		zfree(&evsel->name);
2083 		return true;
2084 	}
2085 
2086 	return false;
2087 }
2088 
2089 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2090 			      int err, char *msg, size_t size)
2091 {
2092 	char sbuf[STRERR_BUFSIZE];
2093 
2094 	switch (err) {
2095 	case EPERM:
2096 	case EACCES:
2097 		return scnprintf(msg, size,
2098 		 "You may not have permission to collect %sstats.\n"
2099 		 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2100 		 " -1 - Not paranoid at all\n"
2101 		 "  0 - Disallow raw tracepoint access for unpriv\n"
2102 		 "  1 - Disallow cpu events for unpriv\n"
2103 		 "  2 - Disallow kernel profiling for unpriv",
2104 				 target->system_wide ? "system-wide " : "");
2105 	case ENOENT:
2106 		return scnprintf(msg, size, "The %s event is not supported.",
2107 				 perf_evsel__name(evsel));
2108 	case EMFILE:
2109 		return scnprintf(msg, size, "%s",
2110 			 "Too many events are opened.\n"
2111 			 "Try again after reducing the number of events.");
2112 	case ENODEV:
2113 		if (target->cpu_list)
2114 			return scnprintf(msg, size, "%s",
2115 	 "No such device - did you specify an out-of-range profile CPU?\n");
2116 		break;
2117 	case EOPNOTSUPP:
2118 		if (evsel->attr.precise_ip)
2119 			return scnprintf(msg, size, "%s",
2120 	"\'precise\' request may not be supported. Try removing 'p' modifier.");
2121 #if defined(__i386__) || defined(__x86_64__)
2122 		if (evsel->attr.type == PERF_TYPE_HARDWARE)
2123 			return scnprintf(msg, size, "%s",
2124 	"No hardware sampling interrupt available.\n"
2125 	"No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2126 #endif
2127 		break;
2128 	case EBUSY:
2129 		if (find_process("oprofiled"))
2130 			return scnprintf(msg, size,
2131 	"The PMU counters are busy/taken by another profiler.\n"
2132 	"We found oprofile daemon running, please stop it and try again.");
2133 		break;
2134 	default:
2135 		break;
2136 	}
2137 
2138 	return scnprintf(msg, size,
2139 	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2140 	"/bin/dmesg may provide additional information.\n"
2141 	"No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2142 			 err, strerror_r(err, sbuf, sizeof(sbuf)),
2143 			 perf_evsel__name(evsel));
2144 }
2145