xref: /linux/tools/perf/util/evsel.c (revision 5dafea097ac65bd01cc86801c399ae41dce79756)
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/tracing_path.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <linux/err.h>
17 #include <sys/resource.h>
18 #include "asm/bug.h"
19 #include "callchain.h"
20 #include "cgroup.h"
21 #include "evsel.h"
22 #include "evlist.h"
23 #include "util.h"
24 #include "cpumap.h"
25 #include "thread_map.h"
26 #include "target.h"
27 #include "perf_regs.h"
28 #include "debug.h"
29 #include "trace-event.h"
30 #include "stat.h"
31 
32 static struct {
33 	bool sample_id_all;
34 	bool exclude_guest;
35 	bool mmap2;
36 	bool cloexec;
37 	bool clockid;
38 	bool clockid_wrong;
39 	bool lbr_flags;
40 	bool write_backward;
41 } perf_missing_features;
42 
43 static clockid_t clockid;
44 
45 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
46 {
47 	return 0;
48 }
49 
50 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
51 {
52 }
53 
54 static struct {
55 	size_t	size;
56 	int	(*init)(struct perf_evsel *evsel);
57 	void	(*fini)(struct perf_evsel *evsel);
58 } perf_evsel__object = {
59 	.size = sizeof(struct perf_evsel),
60 	.init = perf_evsel__no_extra_init,
61 	.fini = perf_evsel__no_extra_fini,
62 };
63 
64 int perf_evsel__object_config(size_t object_size,
65 			      int (*init)(struct perf_evsel *evsel),
66 			      void (*fini)(struct perf_evsel *evsel))
67 {
68 
69 	if (object_size == 0)
70 		goto set_methods;
71 
72 	if (perf_evsel__object.size > object_size)
73 		return -EINVAL;
74 
75 	perf_evsel__object.size = object_size;
76 
77 set_methods:
78 	if (init != NULL)
79 		perf_evsel__object.init = init;
80 
81 	if (fini != NULL)
82 		perf_evsel__object.fini = fini;
83 
84 	return 0;
85 }
86 
87 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
88 
89 int __perf_evsel__sample_size(u64 sample_type)
90 {
91 	u64 mask = sample_type & PERF_SAMPLE_MASK;
92 	int size = 0;
93 	int i;
94 
95 	for (i = 0; i < 64; i++) {
96 		if (mask & (1ULL << i))
97 			size++;
98 	}
99 
100 	size *= sizeof(u64);
101 
102 	return size;
103 }
104 
105 /**
106  * __perf_evsel__calc_id_pos - calculate id_pos.
107  * @sample_type: sample type
108  *
109  * This function returns the position of the event id (PERF_SAMPLE_ID or
110  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
111  * sample_event.
112  */
113 static int __perf_evsel__calc_id_pos(u64 sample_type)
114 {
115 	int idx = 0;
116 
117 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
118 		return 0;
119 
120 	if (!(sample_type & PERF_SAMPLE_ID))
121 		return -1;
122 
123 	if (sample_type & PERF_SAMPLE_IP)
124 		idx += 1;
125 
126 	if (sample_type & PERF_SAMPLE_TID)
127 		idx += 1;
128 
129 	if (sample_type & PERF_SAMPLE_TIME)
130 		idx += 1;
131 
132 	if (sample_type & PERF_SAMPLE_ADDR)
133 		idx += 1;
134 
135 	return idx;
136 }
137 
138 /**
139  * __perf_evsel__calc_is_pos - calculate is_pos.
140  * @sample_type: sample type
141  *
142  * This function returns the position (counting backwards) of the event id
143  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
144  * sample_id_all is used there is an id sample appended to non-sample events.
145  */
146 static int __perf_evsel__calc_is_pos(u64 sample_type)
147 {
148 	int idx = 1;
149 
150 	if (sample_type & PERF_SAMPLE_IDENTIFIER)
151 		return 1;
152 
153 	if (!(sample_type & PERF_SAMPLE_ID))
154 		return -1;
155 
156 	if (sample_type & PERF_SAMPLE_CPU)
157 		idx += 1;
158 
159 	if (sample_type & PERF_SAMPLE_STREAM_ID)
160 		idx += 1;
161 
162 	return idx;
163 }
164 
165 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
166 {
167 	evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
168 	evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
169 }
170 
171 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
172 				  enum perf_event_sample_format bit)
173 {
174 	if (!(evsel->attr.sample_type & bit)) {
175 		evsel->attr.sample_type |= bit;
176 		evsel->sample_size += sizeof(u64);
177 		perf_evsel__calc_id_pos(evsel);
178 	}
179 }
180 
181 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
182 				    enum perf_event_sample_format bit)
183 {
184 	if (evsel->attr.sample_type & bit) {
185 		evsel->attr.sample_type &= ~bit;
186 		evsel->sample_size -= sizeof(u64);
187 		perf_evsel__calc_id_pos(evsel);
188 	}
189 }
190 
191 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
192 			       bool can_sample_identifier)
193 {
194 	if (can_sample_identifier) {
195 		perf_evsel__reset_sample_bit(evsel, ID);
196 		perf_evsel__set_sample_bit(evsel, IDENTIFIER);
197 	} else {
198 		perf_evsel__set_sample_bit(evsel, ID);
199 	}
200 	evsel->attr.read_format |= PERF_FORMAT_ID;
201 }
202 
203 void perf_evsel__init(struct perf_evsel *evsel,
204 		      struct perf_event_attr *attr, int idx)
205 {
206 	evsel->idx	   = idx;
207 	evsel->tracking	   = !idx;
208 	evsel->attr	   = *attr;
209 	evsel->leader	   = evsel;
210 	evsel->unit	   = "";
211 	evsel->scale	   = 1.0;
212 	evsel->evlist	   = NULL;
213 	evsel->bpf_fd	   = -1;
214 	INIT_LIST_HEAD(&evsel->node);
215 	INIT_LIST_HEAD(&evsel->config_terms);
216 	perf_evsel__object.init(evsel);
217 	evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
218 	perf_evsel__calc_id_pos(evsel);
219 	evsel->cmdline_group_boundary = false;
220 }
221 
222 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
223 {
224 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
225 
226 	if (evsel != NULL)
227 		perf_evsel__init(evsel, attr, idx);
228 
229 	if (perf_evsel__is_bpf_output(evsel)) {
230 		evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
231 					    PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
232 		evsel->attr.sample_period = 1;
233 	}
234 
235 	return evsel;
236 }
237 
238 /*
239  * Returns pointer with encoded error via <linux/err.h> interface.
240  */
241 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
242 {
243 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
244 	int err = -ENOMEM;
245 
246 	if (evsel == NULL) {
247 		goto out_err;
248 	} else {
249 		struct perf_event_attr attr = {
250 			.type	       = PERF_TYPE_TRACEPOINT,
251 			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
252 					  PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
253 		};
254 
255 		if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
256 			goto out_free;
257 
258 		evsel->tp_format = trace_event__tp_format(sys, name);
259 		if (IS_ERR(evsel->tp_format)) {
260 			err = PTR_ERR(evsel->tp_format);
261 			goto out_free;
262 		}
263 
264 		event_attr_init(&attr);
265 		attr.config = evsel->tp_format->id;
266 		attr.sample_period = 1;
267 		perf_evsel__init(evsel, &attr, idx);
268 	}
269 
270 	return evsel;
271 
272 out_free:
273 	zfree(&evsel->name);
274 	free(evsel);
275 out_err:
276 	return ERR_PTR(err);
277 }
278 
279 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
280 	"cycles",
281 	"instructions",
282 	"cache-references",
283 	"cache-misses",
284 	"branches",
285 	"branch-misses",
286 	"bus-cycles",
287 	"stalled-cycles-frontend",
288 	"stalled-cycles-backend",
289 	"ref-cycles",
290 };
291 
292 static const char *__perf_evsel__hw_name(u64 config)
293 {
294 	if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
295 		return perf_evsel__hw_names[config];
296 
297 	return "unknown-hardware";
298 }
299 
300 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
301 {
302 	int colon = 0, r = 0;
303 	struct perf_event_attr *attr = &evsel->attr;
304 	bool exclude_guest_default = false;
305 
306 #define MOD_PRINT(context, mod)	do {					\
307 		if (!attr->exclude_##context) {				\
308 			if (!colon) colon = ++r;			\
309 			r += scnprintf(bf + r, size - r, "%c", mod);	\
310 		} } while(0)
311 
312 	if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
313 		MOD_PRINT(kernel, 'k');
314 		MOD_PRINT(user, 'u');
315 		MOD_PRINT(hv, 'h');
316 		exclude_guest_default = true;
317 	}
318 
319 	if (attr->precise_ip) {
320 		if (!colon)
321 			colon = ++r;
322 		r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
323 		exclude_guest_default = true;
324 	}
325 
326 	if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
327 		MOD_PRINT(host, 'H');
328 		MOD_PRINT(guest, 'G');
329 	}
330 #undef MOD_PRINT
331 	if (colon)
332 		bf[colon - 1] = ':';
333 	return r;
334 }
335 
336 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
337 {
338 	int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
339 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
340 }
341 
342 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
343 	"cpu-clock",
344 	"task-clock",
345 	"page-faults",
346 	"context-switches",
347 	"cpu-migrations",
348 	"minor-faults",
349 	"major-faults",
350 	"alignment-faults",
351 	"emulation-faults",
352 	"dummy",
353 };
354 
355 static const char *__perf_evsel__sw_name(u64 config)
356 {
357 	if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
358 		return perf_evsel__sw_names[config];
359 	return "unknown-software";
360 }
361 
362 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
363 {
364 	int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
365 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
366 }
367 
368 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
369 {
370 	int r;
371 
372 	r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
373 
374 	if (type & HW_BREAKPOINT_R)
375 		r += scnprintf(bf + r, size - r, "r");
376 
377 	if (type & HW_BREAKPOINT_W)
378 		r += scnprintf(bf + r, size - r, "w");
379 
380 	if (type & HW_BREAKPOINT_X)
381 		r += scnprintf(bf + r, size - r, "x");
382 
383 	return r;
384 }
385 
386 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
387 {
388 	struct perf_event_attr *attr = &evsel->attr;
389 	int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
390 	return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
391 }
392 
393 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
394 				[PERF_EVSEL__MAX_ALIASES] = {
395  { "L1-dcache",	"l1-d",		"l1d",		"L1-data",		},
396  { "L1-icache",	"l1-i",		"l1i",		"L1-instruction",	},
397  { "LLC",	"L2",							},
398  { "dTLB",	"d-tlb",	"Data-TLB",				},
399  { "iTLB",	"i-tlb",	"Instruction-TLB",			},
400  { "branch",	"branches",	"bpu",		"btb",		"bpc",	},
401  { "node",								},
402 };
403 
404 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
405 				   [PERF_EVSEL__MAX_ALIASES] = {
406  { "load",	"loads",	"read",					},
407  { "store",	"stores",	"write",				},
408  { "prefetch",	"prefetches",	"speculative-read", "speculative-load",	},
409 };
410 
411 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
412 				       [PERF_EVSEL__MAX_ALIASES] = {
413  { "refs",	"Reference",	"ops",		"access",		},
414  { "misses",	"miss",							},
415 };
416 
417 #define C(x)		PERF_COUNT_HW_CACHE_##x
418 #define CACHE_READ	(1 << C(OP_READ))
419 #define CACHE_WRITE	(1 << C(OP_WRITE))
420 #define CACHE_PREFETCH	(1 << C(OP_PREFETCH))
421 #define COP(x)		(1 << x)
422 
423 /*
424  * cache operartion stat
425  * L1I : Read and prefetch only
426  * ITLB and BPU : Read-only
427  */
428 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
429  [C(L1D)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
430  [C(L1I)]	= (CACHE_READ | CACHE_PREFETCH),
431  [C(LL)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
432  [C(DTLB)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
433  [C(ITLB)]	= (CACHE_READ),
434  [C(BPU)]	= (CACHE_READ),
435  [C(NODE)]	= (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
436 };
437 
438 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
439 {
440 	if (perf_evsel__hw_cache_stat[type] & COP(op))
441 		return true;	/* valid */
442 	else
443 		return false;	/* invalid */
444 }
445 
446 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
447 					    char *bf, size_t size)
448 {
449 	if (result) {
450 		return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
451 				 perf_evsel__hw_cache_op[op][0],
452 				 perf_evsel__hw_cache_result[result][0]);
453 	}
454 
455 	return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
456 			 perf_evsel__hw_cache_op[op][1]);
457 }
458 
459 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
460 {
461 	u8 op, result, type = (config >>  0) & 0xff;
462 	const char *err = "unknown-ext-hardware-cache-type";
463 
464 	if (type > PERF_COUNT_HW_CACHE_MAX)
465 		goto out_err;
466 
467 	op = (config >>  8) & 0xff;
468 	err = "unknown-ext-hardware-cache-op";
469 	if (op > PERF_COUNT_HW_CACHE_OP_MAX)
470 		goto out_err;
471 
472 	result = (config >> 16) & 0xff;
473 	err = "unknown-ext-hardware-cache-result";
474 	if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
475 		goto out_err;
476 
477 	err = "invalid-cache";
478 	if (!perf_evsel__is_cache_op_valid(type, op))
479 		goto out_err;
480 
481 	return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
482 out_err:
483 	return scnprintf(bf, size, "%s", err);
484 }
485 
486 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
487 {
488 	int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
489 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
490 }
491 
492 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
493 {
494 	int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
495 	return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
496 }
497 
498 const char *perf_evsel__name(struct perf_evsel *evsel)
499 {
500 	char bf[128];
501 
502 	if (evsel->name)
503 		return evsel->name;
504 
505 	switch (evsel->attr.type) {
506 	case PERF_TYPE_RAW:
507 		perf_evsel__raw_name(evsel, bf, sizeof(bf));
508 		break;
509 
510 	case PERF_TYPE_HARDWARE:
511 		perf_evsel__hw_name(evsel, bf, sizeof(bf));
512 		break;
513 
514 	case PERF_TYPE_HW_CACHE:
515 		perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
516 		break;
517 
518 	case PERF_TYPE_SOFTWARE:
519 		perf_evsel__sw_name(evsel, bf, sizeof(bf));
520 		break;
521 
522 	case PERF_TYPE_TRACEPOINT:
523 		scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
524 		break;
525 
526 	case PERF_TYPE_BREAKPOINT:
527 		perf_evsel__bp_name(evsel, bf, sizeof(bf));
528 		break;
529 
530 	default:
531 		scnprintf(bf, sizeof(bf), "unknown attr type: %d",
532 			  evsel->attr.type);
533 		break;
534 	}
535 
536 	evsel->name = strdup(bf);
537 
538 	return evsel->name ?: "unknown";
539 }
540 
541 const char *perf_evsel__group_name(struct perf_evsel *evsel)
542 {
543 	return evsel->group_name ?: "anon group";
544 }
545 
546 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
547 {
548 	int ret;
549 	struct perf_evsel *pos;
550 	const char *group_name = perf_evsel__group_name(evsel);
551 
552 	ret = scnprintf(buf, size, "%s", group_name);
553 
554 	ret += scnprintf(buf + ret, size - ret, " { %s",
555 			 perf_evsel__name(evsel));
556 
557 	for_each_group_member(pos, evsel)
558 		ret += scnprintf(buf + ret, size - ret, ", %s",
559 				 perf_evsel__name(pos));
560 
561 	ret += scnprintf(buf + ret, size - ret, " }");
562 
563 	return ret;
564 }
565 
566 void perf_evsel__config_callchain(struct perf_evsel *evsel,
567 				  struct record_opts *opts,
568 				  struct callchain_param *param)
569 {
570 	bool function = perf_evsel__is_function_event(evsel);
571 	struct perf_event_attr *attr = &evsel->attr;
572 
573 	perf_evsel__set_sample_bit(evsel, CALLCHAIN);
574 
575 	attr->sample_max_stack = param->max_stack;
576 
577 	if (param->record_mode == CALLCHAIN_LBR) {
578 		if (!opts->branch_stack) {
579 			if (attr->exclude_user) {
580 				pr_warning("LBR callstack option is only available "
581 					   "to get user callchain information. "
582 					   "Falling back to framepointers.\n");
583 			} else {
584 				perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
585 				attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
586 							PERF_SAMPLE_BRANCH_CALL_STACK |
587 							PERF_SAMPLE_BRANCH_NO_CYCLES |
588 							PERF_SAMPLE_BRANCH_NO_FLAGS;
589 			}
590 		} else
591 			 pr_warning("Cannot use LBR callstack with branch stack. "
592 				    "Falling back to framepointers.\n");
593 	}
594 
595 	if (param->record_mode == CALLCHAIN_DWARF) {
596 		if (!function) {
597 			perf_evsel__set_sample_bit(evsel, REGS_USER);
598 			perf_evsel__set_sample_bit(evsel, STACK_USER);
599 			attr->sample_regs_user = PERF_REGS_MASK;
600 			attr->sample_stack_user = param->dump_size;
601 			attr->exclude_callchain_user = 1;
602 		} else {
603 			pr_info("Cannot use DWARF unwind for function trace event,"
604 				" falling back to framepointers.\n");
605 		}
606 	}
607 
608 	if (function) {
609 		pr_info("Disabling user space callchains for function trace event.\n");
610 		attr->exclude_callchain_user = 1;
611 	}
612 }
613 
614 static void
615 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
616 			    struct callchain_param *param)
617 {
618 	struct perf_event_attr *attr = &evsel->attr;
619 
620 	perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
621 	if (param->record_mode == CALLCHAIN_LBR) {
622 		perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
623 		attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
624 					      PERF_SAMPLE_BRANCH_CALL_STACK);
625 	}
626 	if (param->record_mode == CALLCHAIN_DWARF) {
627 		perf_evsel__reset_sample_bit(evsel, REGS_USER);
628 		perf_evsel__reset_sample_bit(evsel, STACK_USER);
629 	}
630 }
631 
632 static void apply_config_terms(struct perf_evsel *evsel,
633 			       struct record_opts *opts)
634 {
635 	struct perf_evsel_config_term *term;
636 	struct list_head *config_terms = &evsel->config_terms;
637 	struct perf_event_attr *attr = &evsel->attr;
638 	struct callchain_param param;
639 	u32 dump_size = 0;
640 	int max_stack = 0;
641 	const char *callgraph_buf = NULL;
642 
643 	/* callgraph default */
644 	param.record_mode = callchain_param.record_mode;
645 
646 	list_for_each_entry(term, config_terms, list) {
647 		switch (term->type) {
648 		case PERF_EVSEL__CONFIG_TERM_PERIOD:
649 			attr->sample_period = term->val.period;
650 			attr->freq = 0;
651 			break;
652 		case PERF_EVSEL__CONFIG_TERM_FREQ:
653 			attr->sample_freq = term->val.freq;
654 			attr->freq = 1;
655 			break;
656 		case PERF_EVSEL__CONFIG_TERM_TIME:
657 			if (term->val.time)
658 				perf_evsel__set_sample_bit(evsel, TIME);
659 			else
660 				perf_evsel__reset_sample_bit(evsel, TIME);
661 			break;
662 		case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
663 			callgraph_buf = term->val.callgraph;
664 			break;
665 		case PERF_EVSEL__CONFIG_TERM_STACK_USER:
666 			dump_size = term->val.stack_user;
667 			break;
668 		case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
669 			max_stack = term->val.max_stack;
670 			break;
671 		case PERF_EVSEL__CONFIG_TERM_INHERIT:
672 			/*
673 			 * attr->inherit should has already been set by
674 			 * perf_evsel__config. If user explicitly set
675 			 * inherit using config terms, override global
676 			 * opt->no_inherit setting.
677 			 */
678 			attr->inherit = term->val.inherit ? 1 : 0;
679 			break;
680 		default:
681 			break;
682 		}
683 	}
684 
685 	/* User explicitly set per-event callgraph, clear the old setting and reset. */
686 	if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
687 		if (max_stack) {
688 			param.max_stack = max_stack;
689 			if (callgraph_buf == NULL)
690 				callgraph_buf = "fp";
691 		}
692 
693 		/* parse callgraph parameters */
694 		if (callgraph_buf != NULL) {
695 			if (!strcmp(callgraph_buf, "no")) {
696 				param.enabled = false;
697 				param.record_mode = CALLCHAIN_NONE;
698 			} else {
699 				param.enabled = true;
700 				if (parse_callchain_record(callgraph_buf, &param)) {
701 					pr_err("per-event callgraph setting for %s failed. "
702 					       "Apply callgraph global setting for it\n",
703 					       evsel->name);
704 					return;
705 				}
706 			}
707 		}
708 		if (dump_size > 0) {
709 			dump_size = round_up(dump_size, sizeof(u64));
710 			param.dump_size = dump_size;
711 		}
712 
713 		/* If global callgraph set, clear it */
714 		if (callchain_param.enabled)
715 			perf_evsel__reset_callgraph(evsel, &callchain_param);
716 
717 		/* set perf-event callgraph */
718 		if (param.enabled)
719 			perf_evsel__config_callchain(evsel, opts, &param);
720 	}
721 }
722 
723 /*
724  * The enable_on_exec/disabled value strategy:
725  *
726  *  1) For any type of traced program:
727  *    - all independent events and group leaders are disabled
728  *    - all group members are enabled
729  *
730  *     Group members are ruled by group leaders. They need to
731  *     be enabled, because the group scheduling relies on that.
732  *
733  *  2) For traced programs executed by perf:
734  *     - all independent events and group leaders have
735  *       enable_on_exec set
736  *     - we don't specifically enable or disable any event during
737  *       the record command
738  *
739  *     Independent events and group leaders are initially disabled
740  *     and get enabled by exec. Group members are ruled by group
741  *     leaders as stated in 1).
742  *
743  *  3) For traced programs attached by perf (pid/tid):
744  *     - we specifically enable or disable all events during
745  *       the record command
746  *
747  *     When attaching events to already running traced we
748  *     enable/disable events specifically, as there's no
749  *     initial traced exec call.
750  */
751 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
752 			struct callchain_param *callchain)
753 {
754 	struct perf_evsel *leader = evsel->leader;
755 	struct perf_event_attr *attr = &evsel->attr;
756 	int track = evsel->tracking;
757 	bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
758 
759 	attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
760 	attr->inherit	    = !opts->no_inherit;
761 
762 	perf_evsel__set_sample_bit(evsel, IP);
763 	perf_evsel__set_sample_bit(evsel, TID);
764 
765 	if (evsel->sample_read) {
766 		perf_evsel__set_sample_bit(evsel, READ);
767 
768 		/*
769 		 * We need ID even in case of single event, because
770 		 * PERF_SAMPLE_READ process ID specific data.
771 		 */
772 		perf_evsel__set_sample_id(evsel, false);
773 
774 		/*
775 		 * Apply group format only if we belong to group
776 		 * with more than one members.
777 		 */
778 		if (leader->nr_members > 1) {
779 			attr->read_format |= PERF_FORMAT_GROUP;
780 			attr->inherit = 0;
781 		}
782 	}
783 
784 	/*
785 	 * We default some events to have a default interval. But keep
786 	 * it a weak assumption overridable by the user.
787 	 */
788 	if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
789 				     opts->user_interval != ULLONG_MAX)) {
790 		if (opts->freq) {
791 			perf_evsel__set_sample_bit(evsel, PERIOD);
792 			attr->freq		= 1;
793 			attr->sample_freq	= opts->freq;
794 		} else {
795 			attr->sample_period = opts->default_interval;
796 		}
797 	}
798 
799 	/*
800 	 * Disable sampling for all group members other
801 	 * than leader in case leader 'leads' the sampling.
802 	 */
803 	if ((leader != evsel) && leader->sample_read) {
804 		attr->sample_freq   = 0;
805 		attr->sample_period = 0;
806 	}
807 
808 	if (opts->no_samples)
809 		attr->sample_freq = 0;
810 
811 	if (opts->inherit_stat)
812 		attr->inherit_stat = 1;
813 
814 	if (opts->sample_address) {
815 		perf_evsel__set_sample_bit(evsel, ADDR);
816 		attr->mmap_data = track;
817 	}
818 
819 	/*
820 	 * We don't allow user space callchains for  function trace
821 	 * event, due to issues with page faults while tracing page
822 	 * fault handler and its overall trickiness nature.
823 	 */
824 	if (perf_evsel__is_function_event(evsel))
825 		evsel->attr.exclude_callchain_user = 1;
826 
827 	if (callchain && callchain->enabled && !evsel->no_aux_samples)
828 		perf_evsel__config_callchain(evsel, opts, callchain);
829 
830 	if (opts->sample_intr_regs) {
831 		attr->sample_regs_intr = opts->sample_intr_regs;
832 		perf_evsel__set_sample_bit(evsel, REGS_INTR);
833 	}
834 
835 	if (target__has_cpu(&opts->target))
836 		perf_evsel__set_sample_bit(evsel, CPU);
837 
838 	if (opts->period)
839 		perf_evsel__set_sample_bit(evsel, PERIOD);
840 
841 	/*
842 	 * When the user explicitly disabled time don't force it here.
843 	 */
844 	if (opts->sample_time &&
845 	    (!perf_missing_features.sample_id_all &&
846 	    (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
847 	     opts->sample_time_set)))
848 		perf_evsel__set_sample_bit(evsel, TIME);
849 
850 	if (opts->raw_samples && !evsel->no_aux_samples) {
851 		perf_evsel__set_sample_bit(evsel, TIME);
852 		perf_evsel__set_sample_bit(evsel, RAW);
853 		perf_evsel__set_sample_bit(evsel, CPU);
854 	}
855 
856 	if (opts->sample_address)
857 		perf_evsel__set_sample_bit(evsel, DATA_SRC);
858 
859 	if (opts->no_buffering) {
860 		attr->watermark = 0;
861 		attr->wakeup_events = 1;
862 	}
863 	if (opts->branch_stack && !evsel->no_aux_samples) {
864 		perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
865 		attr->branch_sample_type = opts->branch_stack;
866 	}
867 
868 	if (opts->sample_weight)
869 		perf_evsel__set_sample_bit(evsel, WEIGHT);
870 
871 	attr->task  = track;
872 	attr->mmap  = track;
873 	attr->mmap2 = track && !perf_missing_features.mmap2;
874 	attr->comm  = track;
875 
876 	if (opts->record_switch_events)
877 		attr->context_switch = track;
878 
879 	if (opts->sample_transaction)
880 		perf_evsel__set_sample_bit(evsel, TRANSACTION);
881 
882 	if (opts->running_time) {
883 		evsel->attr.read_format |=
884 			PERF_FORMAT_TOTAL_TIME_ENABLED |
885 			PERF_FORMAT_TOTAL_TIME_RUNNING;
886 	}
887 
888 	/*
889 	 * XXX see the function comment above
890 	 *
891 	 * Disabling only independent events or group leaders,
892 	 * keeping group members enabled.
893 	 */
894 	if (perf_evsel__is_group_leader(evsel))
895 		attr->disabled = 1;
896 
897 	/*
898 	 * Setting enable_on_exec for independent events and
899 	 * group leaders for traced executed by perf.
900 	 */
901 	if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
902 		!opts->initial_delay)
903 		attr->enable_on_exec = 1;
904 
905 	if (evsel->immediate) {
906 		attr->disabled = 0;
907 		attr->enable_on_exec = 0;
908 	}
909 
910 	clockid = opts->clockid;
911 	if (opts->use_clockid) {
912 		attr->use_clockid = 1;
913 		attr->clockid = opts->clockid;
914 	}
915 
916 	if (evsel->precise_max)
917 		perf_event_attr__set_max_precise_ip(attr);
918 
919 	if (opts->all_user) {
920 		attr->exclude_kernel = 1;
921 		attr->exclude_user   = 0;
922 	}
923 
924 	if (opts->all_kernel) {
925 		attr->exclude_kernel = 0;
926 		attr->exclude_user   = 1;
927 	}
928 
929 	/*
930 	 * Apply event specific term settings,
931 	 * it overloads any global configuration.
932 	 */
933 	apply_config_terms(evsel, opts);
934 }
935 
936 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
937 {
938 	int cpu, thread;
939 
940 	if (evsel->system_wide)
941 		nthreads = 1;
942 
943 	evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
944 
945 	if (evsel->fd) {
946 		for (cpu = 0; cpu < ncpus; cpu++) {
947 			for (thread = 0; thread < nthreads; thread++) {
948 				FD(evsel, cpu, thread) = -1;
949 			}
950 		}
951 	}
952 
953 	return evsel->fd != NULL ? 0 : -ENOMEM;
954 }
955 
956 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
957 			  int ioc,  void *arg)
958 {
959 	int cpu, thread;
960 
961 	if (evsel->system_wide)
962 		nthreads = 1;
963 
964 	for (cpu = 0; cpu < ncpus; cpu++) {
965 		for (thread = 0; thread < nthreads; thread++) {
966 			int fd = FD(evsel, cpu, thread),
967 			    err = ioctl(fd, ioc, arg);
968 
969 			if (err)
970 				return err;
971 		}
972 	}
973 
974 	return 0;
975 }
976 
977 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
978 			     const char *filter)
979 {
980 	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
981 				     PERF_EVENT_IOC_SET_FILTER,
982 				     (void *)filter);
983 }
984 
985 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
986 {
987 	char *new_filter = strdup(filter);
988 
989 	if (new_filter != NULL) {
990 		free(evsel->filter);
991 		evsel->filter = new_filter;
992 		return 0;
993 	}
994 
995 	return -1;
996 }
997 
998 int perf_evsel__append_filter(struct perf_evsel *evsel,
999 			      const char *op, const char *filter)
1000 {
1001 	char *new_filter;
1002 
1003 	if (evsel->filter == NULL)
1004 		return perf_evsel__set_filter(evsel, filter);
1005 
1006 	if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
1007 		free(evsel->filter);
1008 		evsel->filter = new_filter;
1009 		return 0;
1010 	}
1011 
1012 	return -1;
1013 }
1014 
1015 int perf_evsel__enable(struct perf_evsel *evsel)
1016 {
1017 	int nthreads = thread_map__nr(evsel->threads);
1018 	int ncpus = cpu_map__nr(evsel->cpus);
1019 
1020 	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1021 				     PERF_EVENT_IOC_ENABLE,
1022 				     0);
1023 }
1024 
1025 int perf_evsel__disable(struct perf_evsel *evsel)
1026 {
1027 	int nthreads = thread_map__nr(evsel->threads);
1028 	int ncpus = cpu_map__nr(evsel->cpus);
1029 
1030 	return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1031 				     PERF_EVENT_IOC_DISABLE,
1032 				     0);
1033 }
1034 
1035 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1036 {
1037 	if (ncpus == 0 || nthreads == 0)
1038 		return 0;
1039 
1040 	if (evsel->system_wide)
1041 		nthreads = 1;
1042 
1043 	evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1044 	if (evsel->sample_id == NULL)
1045 		return -ENOMEM;
1046 
1047 	evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1048 	if (evsel->id == NULL) {
1049 		xyarray__delete(evsel->sample_id);
1050 		evsel->sample_id = NULL;
1051 		return -ENOMEM;
1052 	}
1053 
1054 	return 0;
1055 }
1056 
1057 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1058 {
1059 	xyarray__delete(evsel->fd);
1060 	evsel->fd = NULL;
1061 }
1062 
1063 static void perf_evsel__free_id(struct perf_evsel *evsel)
1064 {
1065 	xyarray__delete(evsel->sample_id);
1066 	evsel->sample_id = NULL;
1067 	zfree(&evsel->id);
1068 }
1069 
1070 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1071 {
1072 	struct perf_evsel_config_term *term, *h;
1073 
1074 	list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1075 		list_del(&term->list);
1076 		free(term);
1077 	}
1078 }
1079 
1080 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1081 {
1082 	int cpu, thread;
1083 
1084 	if (evsel->system_wide)
1085 		nthreads = 1;
1086 
1087 	for (cpu = 0; cpu < ncpus; cpu++)
1088 		for (thread = 0; thread < nthreads; ++thread) {
1089 			close(FD(evsel, cpu, thread));
1090 			FD(evsel, cpu, thread) = -1;
1091 		}
1092 }
1093 
1094 void perf_evsel__exit(struct perf_evsel *evsel)
1095 {
1096 	assert(list_empty(&evsel->node));
1097 	assert(evsel->evlist == NULL);
1098 	perf_evsel__free_fd(evsel);
1099 	perf_evsel__free_id(evsel);
1100 	perf_evsel__free_config_terms(evsel);
1101 	close_cgroup(evsel->cgrp);
1102 	cpu_map__put(evsel->cpus);
1103 	cpu_map__put(evsel->own_cpus);
1104 	thread_map__put(evsel->threads);
1105 	zfree(&evsel->group_name);
1106 	zfree(&evsel->name);
1107 	perf_evsel__object.fini(evsel);
1108 }
1109 
1110 void perf_evsel__delete(struct perf_evsel *evsel)
1111 {
1112 	perf_evsel__exit(evsel);
1113 	free(evsel);
1114 }
1115 
1116 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1117 				struct perf_counts_values *count)
1118 {
1119 	struct perf_counts_values tmp;
1120 
1121 	if (!evsel->prev_raw_counts)
1122 		return;
1123 
1124 	if (cpu == -1) {
1125 		tmp = evsel->prev_raw_counts->aggr;
1126 		evsel->prev_raw_counts->aggr = *count;
1127 	} else {
1128 		tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1129 		*perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1130 	}
1131 
1132 	count->val = count->val - tmp.val;
1133 	count->ena = count->ena - tmp.ena;
1134 	count->run = count->run - tmp.run;
1135 }
1136 
1137 void perf_counts_values__scale(struct perf_counts_values *count,
1138 			       bool scale, s8 *pscaled)
1139 {
1140 	s8 scaled = 0;
1141 
1142 	if (scale) {
1143 		if (count->run == 0) {
1144 			scaled = -1;
1145 			count->val = 0;
1146 		} else if (count->run < count->ena) {
1147 			scaled = 1;
1148 			count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1149 		}
1150 	} else
1151 		count->ena = count->run = 0;
1152 
1153 	if (pscaled)
1154 		*pscaled = scaled;
1155 }
1156 
1157 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1158 		     struct perf_counts_values *count)
1159 {
1160 	memset(count, 0, sizeof(*count));
1161 
1162 	if (FD(evsel, cpu, thread) < 0)
1163 		return -EINVAL;
1164 
1165 	if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1166 		return -errno;
1167 
1168 	return 0;
1169 }
1170 
1171 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1172 			      int cpu, int thread, bool scale)
1173 {
1174 	struct perf_counts_values count;
1175 	size_t nv = scale ? 3 : 1;
1176 
1177 	if (FD(evsel, cpu, thread) < 0)
1178 		return -EINVAL;
1179 
1180 	if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1181 		return -ENOMEM;
1182 
1183 	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1184 		return -errno;
1185 
1186 	perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1187 	perf_counts_values__scale(&count, scale, NULL);
1188 	*perf_counts(evsel->counts, cpu, thread) = count;
1189 	return 0;
1190 }
1191 
1192 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1193 {
1194 	struct perf_evsel *leader = evsel->leader;
1195 	int fd;
1196 
1197 	if (perf_evsel__is_group_leader(evsel))
1198 		return -1;
1199 
1200 	/*
1201 	 * Leader must be already processed/open,
1202 	 * if not it's a bug.
1203 	 */
1204 	BUG_ON(!leader->fd);
1205 
1206 	fd = FD(leader, cpu, thread);
1207 	BUG_ON(fd == -1);
1208 
1209 	return fd;
1210 }
1211 
1212 struct bit_names {
1213 	int bit;
1214 	const char *name;
1215 };
1216 
1217 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1218 {
1219 	bool first_bit = true;
1220 	int i = 0;
1221 
1222 	do {
1223 		if (value & bits[i].bit) {
1224 			buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1225 			first_bit = false;
1226 		}
1227 	} while (bits[++i].name != NULL);
1228 }
1229 
1230 static void __p_sample_type(char *buf, size_t size, u64 value)
1231 {
1232 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1233 	struct bit_names bits[] = {
1234 		bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1235 		bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1236 		bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1237 		bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1238 		bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1239 		bit_name(WEIGHT),
1240 		{ .name = NULL, }
1241 	};
1242 #undef bit_name
1243 	__p_bits(buf, size, value, bits);
1244 }
1245 
1246 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1247 {
1248 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1249 	struct bit_names bits[] = {
1250 		bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1251 		bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1252 		bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1253 		bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1254 		bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1255 		{ .name = NULL, }
1256 	};
1257 #undef bit_name
1258 	__p_bits(buf, size, value, bits);
1259 }
1260 
1261 static void __p_read_format(char *buf, size_t size, u64 value)
1262 {
1263 #define bit_name(n) { PERF_FORMAT_##n, #n }
1264 	struct bit_names bits[] = {
1265 		bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1266 		bit_name(ID), bit_name(GROUP),
1267 		{ .name = NULL, }
1268 	};
1269 #undef bit_name
1270 	__p_bits(buf, size, value, bits);
1271 }
1272 
1273 #define BUF_SIZE		1024
1274 
1275 #define p_hex(val)		snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1276 #define p_unsigned(val)		snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1277 #define p_signed(val)		snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1278 #define p_sample_type(val)	__p_sample_type(buf, BUF_SIZE, val)
1279 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1280 #define p_read_format(val)	__p_read_format(buf, BUF_SIZE, val)
1281 
1282 #define PRINT_ATTRn(_n, _f, _p)				\
1283 do {							\
1284 	if (attr->_f) {					\
1285 		_p(attr->_f);				\
1286 		ret += attr__fprintf(fp, _n, buf, priv);\
1287 	}						\
1288 } while (0)
1289 
1290 #define PRINT_ATTRf(_f, _p)	PRINT_ATTRn(#_f, _f, _p)
1291 
1292 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1293 			     attr__fprintf_f attr__fprintf, void *priv)
1294 {
1295 	char buf[BUF_SIZE];
1296 	int ret = 0;
1297 
1298 	PRINT_ATTRf(type, p_unsigned);
1299 	PRINT_ATTRf(size, p_unsigned);
1300 	PRINT_ATTRf(config, p_hex);
1301 	PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1302 	PRINT_ATTRf(sample_type, p_sample_type);
1303 	PRINT_ATTRf(read_format, p_read_format);
1304 
1305 	PRINT_ATTRf(disabled, p_unsigned);
1306 	PRINT_ATTRf(inherit, p_unsigned);
1307 	PRINT_ATTRf(pinned, p_unsigned);
1308 	PRINT_ATTRf(exclusive, p_unsigned);
1309 	PRINT_ATTRf(exclude_user, p_unsigned);
1310 	PRINT_ATTRf(exclude_kernel, p_unsigned);
1311 	PRINT_ATTRf(exclude_hv, p_unsigned);
1312 	PRINT_ATTRf(exclude_idle, p_unsigned);
1313 	PRINT_ATTRf(mmap, p_unsigned);
1314 	PRINT_ATTRf(comm, p_unsigned);
1315 	PRINT_ATTRf(freq, p_unsigned);
1316 	PRINT_ATTRf(inherit_stat, p_unsigned);
1317 	PRINT_ATTRf(enable_on_exec, p_unsigned);
1318 	PRINT_ATTRf(task, p_unsigned);
1319 	PRINT_ATTRf(watermark, p_unsigned);
1320 	PRINT_ATTRf(precise_ip, p_unsigned);
1321 	PRINT_ATTRf(mmap_data, p_unsigned);
1322 	PRINT_ATTRf(sample_id_all, p_unsigned);
1323 	PRINT_ATTRf(exclude_host, p_unsigned);
1324 	PRINT_ATTRf(exclude_guest, p_unsigned);
1325 	PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1326 	PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1327 	PRINT_ATTRf(mmap2, p_unsigned);
1328 	PRINT_ATTRf(comm_exec, p_unsigned);
1329 	PRINT_ATTRf(use_clockid, p_unsigned);
1330 	PRINT_ATTRf(context_switch, p_unsigned);
1331 	PRINT_ATTRf(write_backward, p_unsigned);
1332 
1333 	PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1334 	PRINT_ATTRf(bp_type, p_unsigned);
1335 	PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1336 	PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1337 	PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1338 	PRINT_ATTRf(sample_regs_user, p_hex);
1339 	PRINT_ATTRf(sample_stack_user, p_unsigned);
1340 	PRINT_ATTRf(clockid, p_signed);
1341 	PRINT_ATTRf(sample_regs_intr, p_hex);
1342 	PRINT_ATTRf(aux_watermark, p_unsigned);
1343 	PRINT_ATTRf(sample_max_stack, p_unsigned);
1344 
1345 	return ret;
1346 }
1347 
1348 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1349 				void *priv __attribute__((unused)))
1350 {
1351 	return fprintf(fp, "  %-32s %s\n", name, val);
1352 }
1353 
1354 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1355 			      struct thread_map *threads)
1356 {
1357 	int cpu, thread, nthreads;
1358 	unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1359 	int pid = -1, err;
1360 	enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1361 
1362 	if (evsel->system_wide)
1363 		nthreads = 1;
1364 	else
1365 		nthreads = threads->nr;
1366 
1367 	if (evsel->fd == NULL &&
1368 	    perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1369 		return -ENOMEM;
1370 
1371 	if (evsel->cgrp) {
1372 		flags |= PERF_FLAG_PID_CGROUP;
1373 		pid = evsel->cgrp->fd;
1374 	}
1375 
1376 fallback_missing_features:
1377 	if (perf_missing_features.clockid_wrong)
1378 		evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1379 	if (perf_missing_features.clockid) {
1380 		evsel->attr.use_clockid = 0;
1381 		evsel->attr.clockid = 0;
1382 	}
1383 	if (perf_missing_features.cloexec)
1384 		flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1385 	if (perf_missing_features.mmap2)
1386 		evsel->attr.mmap2 = 0;
1387 	if (perf_missing_features.exclude_guest)
1388 		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1389 	if (perf_missing_features.lbr_flags)
1390 		evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1391 				     PERF_SAMPLE_BRANCH_NO_CYCLES);
1392 	if (perf_missing_features.write_backward) {
1393 		if (evsel->overwrite)
1394 			return -EINVAL;
1395 		evsel->attr.write_backward = false;
1396 	}
1397 retry_sample_id:
1398 	if (perf_missing_features.sample_id_all)
1399 		evsel->attr.sample_id_all = 0;
1400 
1401 	if (verbose >= 2) {
1402 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1403 		fprintf(stderr, "perf_event_attr:\n");
1404 		perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1405 		fprintf(stderr, "%.60s\n", graph_dotted_line);
1406 	}
1407 
1408 	for (cpu = 0; cpu < cpus->nr; cpu++) {
1409 
1410 		for (thread = 0; thread < nthreads; thread++) {
1411 			int group_fd;
1412 
1413 			if (!evsel->cgrp && !evsel->system_wide)
1414 				pid = thread_map__pid(threads, thread);
1415 
1416 			group_fd = get_group_fd(evsel, cpu, thread);
1417 retry_open:
1418 			pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1419 				  pid, cpus->map[cpu], group_fd, flags);
1420 
1421 			FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1422 								     pid,
1423 								     cpus->map[cpu],
1424 								     group_fd, flags);
1425 			if (FD(evsel, cpu, thread) < 0) {
1426 				err = -errno;
1427 				pr_debug2("sys_perf_event_open failed, error %d\n",
1428 					  err);
1429 				goto try_fallback;
1430 			}
1431 
1432 			if (evsel->bpf_fd >= 0) {
1433 				int evt_fd = FD(evsel, cpu, thread);
1434 				int bpf_fd = evsel->bpf_fd;
1435 
1436 				err = ioctl(evt_fd,
1437 					    PERF_EVENT_IOC_SET_BPF,
1438 					    bpf_fd);
1439 				if (err && errno != EEXIST) {
1440 					pr_err("failed to attach bpf fd %d: %s\n",
1441 					       bpf_fd, strerror(errno));
1442 					err = -EINVAL;
1443 					goto out_close;
1444 				}
1445 			}
1446 
1447 			set_rlimit = NO_CHANGE;
1448 
1449 			/*
1450 			 * If we succeeded but had to kill clockid, fail and
1451 			 * have perf_evsel__open_strerror() print us a nice
1452 			 * error.
1453 			 */
1454 			if (perf_missing_features.clockid ||
1455 			    perf_missing_features.clockid_wrong) {
1456 				err = -EINVAL;
1457 				goto out_close;
1458 			}
1459 		}
1460 	}
1461 
1462 	return 0;
1463 
1464 try_fallback:
1465 	/*
1466 	 * perf stat needs between 5 and 22 fds per CPU. When we run out
1467 	 * of them try to increase the limits.
1468 	 */
1469 	if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1470 		struct rlimit l;
1471 		int old_errno = errno;
1472 
1473 		if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1474 			if (set_rlimit == NO_CHANGE)
1475 				l.rlim_cur = l.rlim_max;
1476 			else {
1477 				l.rlim_cur = l.rlim_max + 1000;
1478 				l.rlim_max = l.rlim_cur;
1479 			}
1480 			if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1481 				set_rlimit++;
1482 				errno = old_errno;
1483 				goto retry_open;
1484 			}
1485 		}
1486 		errno = old_errno;
1487 	}
1488 
1489 	if (err != -EINVAL || cpu > 0 || thread > 0)
1490 		goto out_close;
1491 
1492 	/*
1493 	 * Must probe features in the order they were added to the
1494 	 * perf_event_attr interface.
1495 	 */
1496 	if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1497 		perf_missing_features.write_backward = true;
1498 		goto fallback_missing_features;
1499 	} else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1500 		perf_missing_features.clockid_wrong = true;
1501 		goto fallback_missing_features;
1502 	} else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1503 		perf_missing_features.clockid = true;
1504 		goto fallback_missing_features;
1505 	} else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1506 		perf_missing_features.cloexec = true;
1507 		goto fallback_missing_features;
1508 	} else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1509 		perf_missing_features.mmap2 = true;
1510 		goto fallback_missing_features;
1511 	} else if (!perf_missing_features.exclude_guest &&
1512 		   (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1513 		perf_missing_features.exclude_guest = true;
1514 		goto fallback_missing_features;
1515 	} else if (!perf_missing_features.sample_id_all) {
1516 		perf_missing_features.sample_id_all = true;
1517 		goto retry_sample_id;
1518 	} else if (!perf_missing_features.lbr_flags &&
1519 			(evsel->attr.branch_sample_type &
1520 			 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1521 			  PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1522 		perf_missing_features.lbr_flags = true;
1523 		goto fallback_missing_features;
1524 	}
1525 out_close:
1526 	do {
1527 		while (--thread >= 0) {
1528 			close(FD(evsel, cpu, thread));
1529 			FD(evsel, cpu, thread) = -1;
1530 		}
1531 		thread = nthreads;
1532 	} while (--cpu >= 0);
1533 	return err;
1534 }
1535 
1536 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1537 {
1538 	if (evsel->fd == NULL)
1539 		return;
1540 
1541 	perf_evsel__close_fd(evsel, ncpus, nthreads);
1542 	perf_evsel__free_fd(evsel);
1543 }
1544 
1545 static struct {
1546 	struct cpu_map map;
1547 	int cpus[1];
1548 } empty_cpu_map = {
1549 	.map.nr	= 1,
1550 	.cpus	= { -1, },
1551 };
1552 
1553 static struct {
1554 	struct thread_map map;
1555 	int threads[1];
1556 } empty_thread_map = {
1557 	.map.nr	 = 1,
1558 	.threads = { -1, },
1559 };
1560 
1561 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1562 		     struct thread_map *threads)
1563 {
1564 	if (cpus == NULL) {
1565 		/* Work around old compiler warnings about strict aliasing */
1566 		cpus = &empty_cpu_map.map;
1567 	}
1568 
1569 	if (threads == NULL)
1570 		threads = &empty_thread_map.map;
1571 
1572 	return __perf_evsel__open(evsel, cpus, threads);
1573 }
1574 
1575 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1576 			     struct cpu_map *cpus)
1577 {
1578 	return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1579 }
1580 
1581 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1582 				struct thread_map *threads)
1583 {
1584 	return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1585 }
1586 
1587 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1588 				       const union perf_event *event,
1589 				       struct perf_sample *sample)
1590 {
1591 	u64 type = evsel->attr.sample_type;
1592 	const u64 *array = event->sample.array;
1593 	bool swapped = evsel->needs_swap;
1594 	union u64_swap u;
1595 
1596 	array += ((event->header.size -
1597 		   sizeof(event->header)) / sizeof(u64)) - 1;
1598 
1599 	if (type & PERF_SAMPLE_IDENTIFIER) {
1600 		sample->id = *array;
1601 		array--;
1602 	}
1603 
1604 	if (type & PERF_SAMPLE_CPU) {
1605 		u.val64 = *array;
1606 		if (swapped) {
1607 			/* undo swap of u64, then swap on individual u32s */
1608 			u.val64 = bswap_64(u.val64);
1609 			u.val32[0] = bswap_32(u.val32[0]);
1610 		}
1611 
1612 		sample->cpu = u.val32[0];
1613 		array--;
1614 	}
1615 
1616 	if (type & PERF_SAMPLE_STREAM_ID) {
1617 		sample->stream_id = *array;
1618 		array--;
1619 	}
1620 
1621 	if (type & PERF_SAMPLE_ID) {
1622 		sample->id = *array;
1623 		array--;
1624 	}
1625 
1626 	if (type & PERF_SAMPLE_TIME) {
1627 		sample->time = *array;
1628 		array--;
1629 	}
1630 
1631 	if (type & PERF_SAMPLE_TID) {
1632 		u.val64 = *array;
1633 		if (swapped) {
1634 			/* undo swap of u64, then swap on individual u32s */
1635 			u.val64 = bswap_64(u.val64);
1636 			u.val32[0] = bswap_32(u.val32[0]);
1637 			u.val32[1] = bswap_32(u.val32[1]);
1638 		}
1639 
1640 		sample->pid = u.val32[0];
1641 		sample->tid = u.val32[1];
1642 		array--;
1643 	}
1644 
1645 	return 0;
1646 }
1647 
1648 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1649 			    u64 size)
1650 {
1651 	return size > max_size || offset + size > endp;
1652 }
1653 
1654 #define OVERFLOW_CHECK(offset, size, max_size)				\
1655 	do {								\
1656 		if (overflow(endp, (max_size), (offset), (size)))	\
1657 			return -EFAULT;					\
1658 	} while (0)
1659 
1660 #define OVERFLOW_CHECK_u64(offset) \
1661 	OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1662 
1663 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1664 			     struct perf_sample *data)
1665 {
1666 	u64 type = evsel->attr.sample_type;
1667 	bool swapped = evsel->needs_swap;
1668 	const u64 *array;
1669 	u16 max_size = event->header.size;
1670 	const void *endp = (void *)event + max_size;
1671 	u64 sz;
1672 
1673 	/*
1674 	 * used for cross-endian analysis. See git commit 65014ab3
1675 	 * for why this goofiness is needed.
1676 	 */
1677 	union u64_swap u;
1678 
1679 	memset(data, 0, sizeof(*data));
1680 	data->cpu = data->pid = data->tid = -1;
1681 	data->stream_id = data->id = data->time = -1ULL;
1682 	data->period = evsel->attr.sample_period;
1683 	data->weight = 0;
1684 	data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1685 
1686 	if (event->header.type != PERF_RECORD_SAMPLE) {
1687 		if (!evsel->attr.sample_id_all)
1688 			return 0;
1689 		return perf_evsel__parse_id_sample(evsel, event, data);
1690 	}
1691 
1692 	array = event->sample.array;
1693 
1694 	/*
1695 	 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1696 	 * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1697 	 * check the format does not go past the end of the event.
1698 	 */
1699 	if (evsel->sample_size + sizeof(event->header) > event->header.size)
1700 		return -EFAULT;
1701 
1702 	data->id = -1ULL;
1703 	if (type & PERF_SAMPLE_IDENTIFIER) {
1704 		data->id = *array;
1705 		array++;
1706 	}
1707 
1708 	if (type & PERF_SAMPLE_IP) {
1709 		data->ip = *array;
1710 		array++;
1711 	}
1712 
1713 	if (type & PERF_SAMPLE_TID) {
1714 		u.val64 = *array;
1715 		if (swapped) {
1716 			/* undo swap of u64, then swap on individual u32s */
1717 			u.val64 = bswap_64(u.val64);
1718 			u.val32[0] = bswap_32(u.val32[0]);
1719 			u.val32[1] = bswap_32(u.val32[1]);
1720 		}
1721 
1722 		data->pid = u.val32[0];
1723 		data->tid = u.val32[1];
1724 		array++;
1725 	}
1726 
1727 	if (type & PERF_SAMPLE_TIME) {
1728 		data->time = *array;
1729 		array++;
1730 	}
1731 
1732 	data->addr = 0;
1733 	if (type & PERF_SAMPLE_ADDR) {
1734 		data->addr = *array;
1735 		array++;
1736 	}
1737 
1738 	if (type & PERF_SAMPLE_ID) {
1739 		data->id = *array;
1740 		array++;
1741 	}
1742 
1743 	if (type & PERF_SAMPLE_STREAM_ID) {
1744 		data->stream_id = *array;
1745 		array++;
1746 	}
1747 
1748 	if (type & PERF_SAMPLE_CPU) {
1749 
1750 		u.val64 = *array;
1751 		if (swapped) {
1752 			/* undo swap of u64, then swap on individual u32s */
1753 			u.val64 = bswap_64(u.val64);
1754 			u.val32[0] = bswap_32(u.val32[0]);
1755 		}
1756 
1757 		data->cpu = u.val32[0];
1758 		array++;
1759 	}
1760 
1761 	if (type & PERF_SAMPLE_PERIOD) {
1762 		data->period = *array;
1763 		array++;
1764 	}
1765 
1766 	if (type & PERF_SAMPLE_READ) {
1767 		u64 read_format = evsel->attr.read_format;
1768 
1769 		OVERFLOW_CHECK_u64(array);
1770 		if (read_format & PERF_FORMAT_GROUP)
1771 			data->read.group.nr = *array;
1772 		else
1773 			data->read.one.value = *array;
1774 
1775 		array++;
1776 
1777 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1778 			OVERFLOW_CHECK_u64(array);
1779 			data->read.time_enabled = *array;
1780 			array++;
1781 		}
1782 
1783 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1784 			OVERFLOW_CHECK_u64(array);
1785 			data->read.time_running = *array;
1786 			array++;
1787 		}
1788 
1789 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1790 		if (read_format & PERF_FORMAT_GROUP) {
1791 			const u64 max_group_nr = UINT64_MAX /
1792 					sizeof(struct sample_read_value);
1793 
1794 			if (data->read.group.nr > max_group_nr)
1795 				return -EFAULT;
1796 			sz = data->read.group.nr *
1797 			     sizeof(struct sample_read_value);
1798 			OVERFLOW_CHECK(array, sz, max_size);
1799 			data->read.group.values =
1800 					(struct sample_read_value *)array;
1801 			array = (void *)array + sz;
1802 		} else {
1803 			OVERFLOW_CHECK_u64(array);
1804 			data->read.one.id = *array;
1805 			array++;
1806 		}
1807 	}
1808 
1809 	if (type & PERF_SAMPLE_CALLCHAIN) {
1810 		const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1811 
1812 		OVERFLOW_CHECK_u64(array);
1813 		data->callchain = (struct ip_callchain *)array++;
1814 		if (data->callchain->nr > max_callchain_nr)
1815 			return -EFAULT;
1816 		sz = data->callchain->nr * sizeof(u64);
1817 		OVERFLOW_CHECK(array, sz, max_size);
1818 		array = (void *)array + sz;
1819 	}
1820 
1821 	if (type & PERF_SAMPLE_RAW) {
1822 		OVERFLOW_CHECK_u64(array);
1823 		u.val64 = *array;
1824 		if (WARN_ONCE(swapped,
1825 			      "Endianness of raw data not corrected!\n")) {
1826 			/* undo swap of u64, then swap on individual u32s */
1827 			u.val64 = bswap_64(u.val64);
1828 			u.val32[0] = bswap_32(u.val32[0]);
1829 			u.val32[1] = bswap_32(u.val32[1]);
1830 		}
1831 		data->raw_size = u.val32[0];
1832 		array = (void *)array + sizeof(u32);
1833 
1834 		OVERFLOW_CHECK(array, data->raw_size, max_size);
1835 		data->raw_data = (void *)array;
1836 		array = (void *)array + data->raw_size;
1837 	}
1838 
1839 	if (type & PERF_SAMPLE_BRANCH_STACK) {
1840 		const u64 max_branch_nr = UINT64_MAX /
1841 					  sizeof(struct branch_entry);
1842 
1843 		OVERFLOW_CHECK_u64(array);
1844 		data->branch_stack = (struct branch_stack *)array++;
1845 
1846 		if (data->branch_stack->nr > max_branch_nr)
1847 			return -EFAULT;
1848 		sz = data->branch_stack->nr * sizeof(struct branch_entry);
1849 		OVERFLOW_CHECK(array, sz, max_size);
1850 		array = (void *)array + sz;
1851 	}
1852 
1853 	if (type & PERF_SAMPLE_REGS_USER) {
1854 		OVERFLOW_CHECK_u64(array);
1855 		data->user_regs.abi = *array;
1856 		array++;
1857 
1858 		if (data->user_regs.abi) {
1859 			u64 mask = evsel->attr.sample_regs_user;
1860 
1861 			sz = hweight_long(mask) * sizeof(u64);
1862 			OVERFLOW_CHECK(array, sz, max_size);
1863 			data->user_regs.mask = mask;
1864 			data->user_regs.regs = (u64 *)array;
1865 			array = (void *)array + sz;
1866 		}
1867 	}
1868 
1869 	if (type & PERF_SAMPLE_STACK_USER) {
1870 		OVERFLOW_CHECK_u64(array);
1871 		sz = *array++;
1872 
1873 		data->user_stack.offset = ((char *)(array - 1)
1874 					  - (char *) event);
1875 
1876 		if (!sz) {
1877 			data->user_stack.size = 0;
1878 		} else {
1879 			OVERFLOW_CHECK(array, sz, max_size);
1880 			data->user_stack.data = (char *)array;
1881 			array = (void *)array + sz;
1882 			OVERFLOW_CHECK_u64(array);
1883 			data->user_stack.size = *array++;
1884 			if (WARN_ONCE(data->user_stack.size > sz,
1885 				      "user stack dump failure\n"))
1886 				return -EFAULT;
1887 		}
1888 	}
1889 
1890 	data->weight = 0;
1891 	if (type & PERF_SAMPLE_WEIGHT) {
1892 		OVERFLOW_CHECK_u64(array);
1893 		data->weight = *array;
1894 		array++;
1895 	}
1896 
1897 	data->data_src = PERF_MEM_DATA_SRC_NONE;
1898 	if (type & PERF_SAMPLE_DATA_SRC) {
1899 		OVERFLOW_CHECK_u64(array);
1900 		data->data_src = *array;
1901 		array++;
1902 	}
1903 
1904 	data->transaction = 0;
1905 	if (type & PERF_SAMPLE_TRANSACTION) {
1906 		OVERFLOW_CHECK_u64(array);
1907 		data->transaction = *array;
1908 		array++;
1909 	}
1910 
1911 	data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1912 	if (type & PERF_SAMPLE_REGS_INTR) {
1913 		OVERFLOW_CHECK_u64(array);
1914 		data->intr_regs.abi = *array;
1915 		array++;
1916 
1917 		if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1918 			u64 mask = evsel->attr.sample_regs_intr;
1919 
1920 			sz = hweight_long(mask) * sizeof(u64);
1921 			OVERFLOW_CHECK(array, sz, max_size);
1922 			data->intr_regs.mask = mask;
1923 			data->intr_regs.regs = (u64 *)array;
1924 			array = (void *)array + sz;
1925 		}
1926 	}
1927 
1928 	return 0;
1929 }
1930 
1931 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1932 				     u64 read_format)
1933 {
1934 	size_t sz, result = sizeof(struct sample_event);
1935 
1936 	if (type & PERF_SAMPLE_IDENTIFIER)
1937 		result += sizeof(u64);
1938 
1939 	if (type & PERF_SAMPLE_IP)
1940 		result += sizeof(u64);
1941 
1942 	if (type & PERF_SAMPLE_TID)
1943 		result += sizeof(u64);
1944 
1945 	if (type & PERF_SAMPLE_TIME)
1946 		result += sizeof(u64);
1947 
1948 	if (type & PERF_SAMPLE_ADDR)
1949 		result += sizeof(u64);
1950 
1951 	if (type & PERF_SAMPLE_ID)
1952 		result += sizeof(u64);
1953 
1954 	if (type & PERF_SAMPLE_STREAM_ID)
1955 		result += sizeof(u64);
1956 
1957 	if (type & PERF_SAMPLE_CPU)
1958 		result += sizeof(u64);
1959 
1960 	if (type & PERF_SAMPLE_PERIOD)
1961 		result += sizeof(u64);
1962 
1963 	if (type & PERF_SAMPLE_READ) {
1964 		result += sizeof(u64);
1965 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1966 			result += sizeof(u64);
1967 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1968 			result += sizeof(u64);
1969 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1970 		if (read_format & PERF_FORMAT_GROUP) {
1971 			sz = sample->read.group.nr *
1972 			     sizeof(struct sample_read_value);
1973 			result += sz;
1974 		} else {
1975 			result += sizeof(u64);
1976 		}
1977 	}
1978 
1979 	if (type & PERF_SAMPLE_CALLCHAIN) {
1980 		sz = (sample->callchain->nr + 1) * sizeof(u64);
1981 		result += sz;
1982 	}
1983 
1984 	if (type & PERF_SAMPLE_RAW) {
1985 		result += sizeof(u32);
1986 		result += sample->raw_size;
1987 	}
1988 
1989 	if (type & PERF_SAMPLE_BRANCH_STACK) {
1990 		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1991 		sz += sizeof(u64);
1992 		result += sz;
1993 	}
1994 
1995 	if (type & PERF_SAMPLE_REGS_USER) {
1996 		if (sample->user_regs.abi) {
1997 			result += sizeof(u64);
1998 			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1999 			result += sz;
2000 		} else {
2001 			result += sizeof(u64);
2002 		}
2003 	}
2004 
2005 	if (type & PERF_SAMPLE_STACK_USER) {
2006 		sz = sample->user_stack.size;
2007 		result += sizeof(u64);
2008 		if (sz) {
2009 			result += sz;
2010 			result += sizeof(u64);
2011 		}
2012 	}
2013 
2014 	if (type & PERF_SAMPLE_WEIGHT)
2015 		result += sizeof(u64);
2016 
2017 	if (type & PERF_SAMPLE_DATA_SRC)
2018 		result += sizeof(u64);
2019 
2020 	if (type & PERF_SAMPLE_TRANSACTION)
2021 		result += sizeof(u64);
2022 
2023 	if (type & PERF_SAMPLE_REGS_INTR) {
2024 		if (sample->intr_regs.abi) {
2025 			result += sizeof(u64);
2026 			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2027 			result += sz;
2028 		} else {
2029 			result += sizeof(u64);
2030 		}
2031 	}
2032 
2033 	return result;
2034 }
2035 
2036 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2037 				  u64 read_format,
2038 				  const struct perf_sample *sample,
2039 				  bool swapped)
2040 {
2041 	u64 *array;
2042 	size_t sz;
2043 	/*
2044 	 * used for cross-endian analysis. See git commit 65014ab3
2045 	 * for why this goofiness is needed.
2046 	 */
2047 	union u64_swap u;
2048 
2049 	array = event->sample.array;
2050 
2051 	if (type & PERF_SAMPLE_IDENTIFIER) {
2052 		*array = sample->id;
2053 		array++;
2054 	}
2055 
2056 	if (type & PERF_SAMPLE_IP) {
2057 		*array = sample->ip;
2058 		array++;
2059 	}
2060 
2061 	if (type & PERF_SAMPLE_TID) {
2062 		u.val32[0] = sample->pid;
2063 		u.val32[1] = sample->tid;
2064 		if (swapped) {
2065 			/*
2066 			 * Inverse of what is done in perf_evsel__parse_sample
2067 			 */
2068 			u.val32[0] = bswap_32(u.val32[0]);
2069 			u.val32[1] = bswap_32(u.val32[1]);
2070 			u.val64 = bswap_64(u.val64);
2071 		}
2072 
2073 		*array = u.val64;
2074 		array++;
2075 	}
2076 
2077 	if (type & PERF_SAMPLE_TIME) {
2078 		*array = sample->time;
2079 		array++;
2080 	}
2081 
2082 	if (type & PERF_SAMPLE_ADDR) {
2083 		*array = sample->addr;
2084 		array++;
2085 	}
2086 
2087 	if (type & PERF_SAMPLE_ID) {
2088 		*array = sample->id;
2089 		array++;
2090 	}
2091 
2092 	if (type & PERF_SAMPLE_STREAM_ID) {
2093 		*array = sample->stream_id;
2094 		array++;
2095 	}
2096 
2097 	if (type & PERF_SAMPLE_CPU) {
2098 		u.val32[0] = sample->cpu;
2099 		if (swapped) {
2100 			/*
2101 			 * Inverse of what is done in perf_evsel__parse_sample
2102 			 */
2103 			u.val32[0] = bswap_32(u.val32[0]);
2104 			u.val64 = bswap_64(u.val64);
2105 		}
2106 		*array = u.val64;
2107 		array++;
2108 	}
2109 
2110 	if (type & PERF_SAMPLE_PERIOD) {
2111 		*array = sample->period;
2112 		array++;
2113 	}
2114 
2115 	if (type & PERF_SAMPLE_READ) {
2116 		if (read_format & PERF_FORMAT_GROUP)
2117 			*array = sample->read.group.nr;
2118 		else
2119 			*array = sample->read.one.value;
2120 		array++;
2121 
2122 		if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2123 			*array = sample->read.time_enabled;
2124 			array++;
2125 		}
2126 
2127 		if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2128 			*array = sample->read.time_running;
2129 			array++;
2130 		}
2131 
2132 		/* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2133 		if (read_format & PERF_FORMAT_GROUP) {
2134 			sz = sample->read.group.nr *
2135 			     sizeof(struct sample_read_value);
2136 			memcpy(array, sample->read.group.values, sz);
2137 			array = (void *)array + sz;
2138 		} else {
2139 			*array = sample->read.one.id;
2140 			array++;
2141 		}
2142 	}
2143 
2144 	if (type & PERF_SAMPLE_CALLCHAIN) {
2145 		sz = (sample->callchain->nr + 1) * sizeof(u64);
2146 		memcpy(array, sample->callchain, sz);
2147 		array = (void *)array + sz;
2148 	}
2149 
2150 	if (type & PERF_SAMPLE_RAW) {
2151 		u.val32[0] = sample->raw_size;
2152 		if (WARN_ONCE(swapped,
2153 			      "Endianness of raw data not corrected!\n")) {
2154 			/*
2155 			 * Inverse of what is done in perf_evsel__parse_sample
2156 			 */
2157 			u.val32[0] = bswap_32(u.val32[0]);
2158 			u.val32[1] = bswap_32(u.val32[1]);
2159 			u.val64 = bswap_64(u.val64);
2160 		}
2161 		*array = u.val64;
2162 		array = (void *)array + sizeof(u32);
2163 
2164 		memcpy(array, sample->raw_data, sample->raw_size);
2165 		array = (void *)array + sample->raw_size;
2166 	}
2167 
2168 	if (type & PERF_SAMPLE_BRANCH_STACK) {
2169 		sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2170 		sz += sizeof(u64);
2171 		memcpy(array, sample->branch_stack, sz);
2172 		array = (void *)array + sz;
2173 	}
2174 
2175 	if (type & PERF_SAMPLE_REGS_USER) {
2176 		if (sample->user_regs.abi) {
2177 			*array++ = sample->user_regs.abi;
2178 			sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2179 			memcpy(array, sample->user_regs.regs, sz);
2180 			array = (void *)array + sz;
2181 		} else {
2182 			*array++ = 0;
2183 		}
2184 	}
2185 
2186 	if (type & PERF_SAMPLE_STACK_USER) {
2187 		sz = sample->user_stack.size;
2188 		*array++ = sz;
2189 		if (sz) {
2190 			memcpy(array, sample->user_stack.data, sz);
2191 			array = (void *)array + sz;
2192 			*array++ = sz;
2193 		}
2194 	}
2195 
2196 	if (type & PERF_SAMPLE_WEIGHT) {
2197 		*array = sample->weight;
2198 		array++;
2199 	}
2200 
2201 	if (type & PERF_SAMPLE_DATA_SRC) {
2202 		*array = sample->data_src;
2203 		array++;
2204 	}
2205 
2206 	if (type & PERF_SAMPLE_TRANSACTION) {
2207 		*array = sample->transaction;
2208 		array++;
2209 	}
2210 
2211 	if (type & PERF_SAMPLE_REGS_INTR) {
2212 		if (sample->intr_regs.abi) {
2213 			*array++ = sample->intr_regs.abi;
2214 			sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2215 			memcpy(array, sample->intr_regs.regs, sz);
2216 			array = (void *)array + sz;
2217 		} else {
2218 			*array++ = 0;
2219 		}
2220 	}
2221 
2222 	return 0;
2223 }
2224 
2225 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2226 {
2227 	return pevent_find_field(evsel->tp_format, name);
2228 }
2229 
2230 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2231 			 const char *name)
2232 {
2233 	struct format_field *field = perf_evsel__field(evsel, name);
2234 	int offset;
2235 
2236 	if (!field)
2237 		return NULL;
2238 
2239 	offset = field->offset;
2240 
2241 	if (field->flags & FIELD_IS_DYNAMIC) {
2242 		offset = *(int *)(sample->raw_data + field->offset);
2243 		offset &= 0xffff;
2244 	}
2245 
2246 	return sample->raw_data + offset;
2247 }
2248 
2249 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2250 			 bool needs_swap)
2251 {
2252 	u64 value;
2253 	void *ptr = sample->raw_data + field->offset;
2254 
2255 	switch (field->size) {
2256 	case 1:
2257 		return *(u8 *)ptr;
2258 	case 2:
2259 		value = *(u16 *)ptr;
2260 		break;
2261 	case 4:
2262 		value = *(u32 *)ptr;
2263 		break;
2264 	case 8:
2265 		memcpy(&value, ptr, sizeof(u64));
2266 		break;
2267 	default:
2268 		return 0;
2269 	}
2270 
2271 	if (!needs_swap)
2272 		return value;
2273 
2274 	switch (field->size) {
2275 	case 2:
2276 		return bswap_16(value);
2277 	case 4:
2278 		return bswap_32(value);
2279 	case 8:
2280 		return bswap_64(value);
2281 	default:
2282 		return 0;
2283 	}
2284 
2285 	return 0;
2286 }
2287 
2288 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2289 		       const char *name)
2290 {
2291 	struct format_field *field = perf_evsel__field(evsel, name);
2292 
2293 	if (!field)
2294 		return 0;
2295 
2296 	return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2297 }
2298 
2299 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2300 			  char *msg, size_t msgsize)
2301 {
2302 	int paranoid;
2303 
2304 	if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2305 	    evsel->attr.type   == PERF_TYPE_HARDWARE &&
2306 	    evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2307 		/*
2308 		 * If it's cycles then fall back to hrtimer based
2309 		 * cpu-clock-tick sw counter, which is always available even if
2310 		 * no PMU support.
2311 		 *
2312 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2313 		 * b0a873e).
2314 		 */
2315 		scnprintf(msg, msgsize, "%s",
2316 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2317 
2318 		evsel->attr.type   = PERF_TYPE_SOFTWARE;
2319 		evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2320 
2321 		zfree(&evsel->name);
2322 		return true;
2323 	} else if (err == EACCES && !evsel->attr.exclude_kernel &&
2324 		   (paranoid = perf_event_paranoid()) > 1) {
2325 		const char *name = perf_evsel__name(evsel);
2326 		char *new_name;
2327 
2328 		if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2329 			return false;
2330 
2331 		if (evsel->name)
2332 			free(evsel->name);
2333 		evsel->name = new_name;
2334 		scnprintf(msg, msgsize,
2335 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2336 		evsel->attr.exclude_kernel = 1;
2337 
2338 		return true;
2339 	}
2340 
2341 	return false;
2342 }
2343 
2344 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2345 			      int err, char *msg, size_t size)
2346 {
2347 	char sbuf[STRERR_BUFSIZE];
2348 
2349 	switch (err) {
2350 	case EPERM:
2351 	case EACCES:
2352 		return scnprintf(msg, size,
2353 		 "You may not have permission to collect %sstats.\n\n"
2354 		 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2355 		 "which controls use of the performance events system by\n"
2356 		 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2357 		 "The current value is %d:\n\n"
2358 		 "  -1: Allow use of (almost) all events by all users\n"
2359 		 ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2360 		 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2361 		 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2362 				 target->system_wide ? "system-wide " : "",
2363 				 perf_event_paranoid());
2364 	case ENOENT:
2365 		return scnprintf(msg, size, "The %s event is not supported.",
2366 				 perf_evsel__name(evsel));
2367 	case EMFILE:
2368 		return scnprintf(msg, size, "%s",
2369 			 "Too many events are opened.\n"
2370 			 "Probably the maximum number of open file descriptors has been reached.\n"
2371 			 "Hint: Try again after reducing the number of events.\n"
2372 			 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2373 	case ENOMEM:
2374 		if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2375 		    access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2376 			return scnprintf(msg, size,
2377 					 "Not enough memory to setup event with callchain.\n"
2378 					 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2379 					 "Hint: Current value: %d", sysctl_perf_event_max_stack);
2380 		break;
2381 	case ENODEV:
2382 		if (target->cpu_list)
2383 			return scnprintf(msg, size, "%s",
2384 	 "No such device - did you specify an out-of-range profile CPU?");
2385 		break;
2386 	case EOPNOTSUPP:
2387 		if (evsel->attr.sample_period != 0)
2388 			return scnprintf(msg, size, "%s",
2389 	"PMU Hardware doesn't support sampling/overflow-interrupts.");
2390 		if (evsel->attr.precise_ip)
2391 			return scnprintf(msg, size, "%s",
2392 	"\'precise\' request may not be supported. Try removing 'p' modifier.");
2393 #if defined(__i386__) || defined(__x86_64__)
2394 		if (evsel->attr.type == PERF_TYPE_HARDWARE)
2395 			return scnprintf(msg, size, "%s",
2396 	"No hardware sampling interrupt available.\n"
2397 	"No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2398 #endif
2399 		break;
2400 	case EBUSY:
2401 		if (find_process("oprofiled"))
2402 			return scnprintf(msg, size,
2403 	"The PMU counters are busy/taken by another profiler.\n"
2404 	"We found oprofile daemon running, please stop it and try again.");
2405 		break;
2406 	case EINVAL:
2407 		if (evsel->overwrite && perf_missing_features.write_backward)
2408 			return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2409 		if (perf_missing_features.clockid)
2410 			return scnprintf(msg, size, "clockid feature not supported.");
2411 		if (perf_missing_features.clockid_wrong)
2412 			return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2413 		break;
2414 	default:
2415 		break;
2416 	}
2417 
2418 	return scnprintf(msg, size,
2419 	"The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2420 	"/bin/dmesg may provide additional information.\n"
2421 	"No CONFIG_PERF_EVENTS=y kernel support configured?",
2422 			 err, strerror_r(err, sbuf, sizeof(sbuf)),
2423 			 perf_evsel__name(evsel));
2424 }
2425