xref: /linux/tools/perf/util/evsel.c (revision b43ab901d671e3e3cad425ea5e9a3c74e266dcdd)
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 "asm/bug.h"
12 #include "evsel.h"
13 #include "evlist.h"
14 #include "util.h"
15 #include "cpumap.h"
16 #include "thread_map.h"
17 
18 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
19 #define GROUP_FD(group_fd, cpu) (*(int *)xyarray__entry(group_fd, cpu, 0))
20 
21 int __perf_evsel__sample_size(u64 sample_type)
22 {
23 	u64 mask = sample_type & PERF_SAMPLE_MASK;
24 	int size = 0;
25 	int i;
26 
27 	for (i = 0; i < 64; i++) {
28 		if (mask & (1ULL << i))
29 			size++;
30 	}
31 
32 	size *= sizeof(u64);
33 
34 	return size;
35 }
36 
37 static void hists__init(struct hists *hists)
38 {
39 	memset(hists, 0, sizeof(*hists));
40 	hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
41 	hists->entries_in = &hists->entries_in_array[0];
42 	hists->entries_collapsed = RB_ROOT;
43 	hists->entries = RB_ROOT;
44 	pthread_mutex_init(&hists->lock, NULL);
45 }
46 
47 void perf_evsel__init(struct perf_evsel *evsel,
48 		      struct perf_event_attr *attr, int idx)
49 {
50 	evsel->idx	   = idx;
51 	evsel->attr	   = *attr;
52 	INIT_LIST_HEAD(&evsel->node);
53 	hists__init(&evsel->hists);
54 }
55 
56 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
57 {
58 	struct perf_evsel *evsel = zalloc(sizeof(*evsel));
59 
60 	if (evsel != NULL)
61 		perf_evsel__init(evsel, attr, idx);
62 
63 	return evsel;
64 }
65 
66 void perf_evsel__config(struct perf_evsel *evsel, struct perf_record_opts *opts)
67 {
68 	struct perf_event_attr *attr = &evsel->attr;
69 	int track = !evsel->idx; /* only the first counter needs these */
70 
71 	attr->sample_id_all = opts->sample_id_all_avail ? 1 : 0;
72 	attr->inherit	    = !opts->no_inherit;
73 	attr->read_format   = PERF_FORMAT_TOTAL_TIME_ENABLED |
74 			      PERF_FORMAT_TOTAL_TIME_RUNNING |
75 			      PERF_FORMAT_ID;
76 
77 	attr->sample_type  |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
78 
79 	/*
80 	 * We default some events to a 1 default interval. But keep
81 	 * it a weak assumption overridable by the user.
82 	 */
83 	if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
84 				     opts->user_interval != ULLONG_MAX)) {
85 		if (opts->freq) {
86 			attr->sample_type	|= PERF_SAMPLE_PERIOD;
87 			attr->freq		= 1;
88 			attr->sample_freq	= opts->freq;
89 		} else {
90 			attr->sample_period = opts->default_interval;
91 		}
92 	}
93 
94 	if (opts->no_samples)
95 		attr->sample_freq = 0;
96 
97 	if (opts->inherit_stat)
98 		attr->inherit_stat = 1;
99 
100 	if (opts->sample_address) {
101 		attr->sample_type	|= PERF_SAMPLE_ADDR;
102 		attr->mmap_data = track;
103 	}
104 
105 	if (opts->call_graph)
106 		attr->sample_type	|= PERF_SAMPLE_CALLCHAIN;
107 
108 	if (opts->system_wide)
109 		attr->sample_type	|= PERF_SAMPLE_CPU;
110 
111 	if (opts->period)
112 		attr->sample_type	|= PERF_SAMPLE_PERIOD;
113 
114 	if (opts->sample_id_all_avail &&
115 	    (opts->sample_time || opts->system_wide ||
116 	     !opts->no_inherit || opts->cpu_list))
117 		attr->sample_type	|= PERF_SAMPLE_TIME;
118 
119 	if (opts->raw_samples) {
120 		attr->sample_type	|= PERF_SAMPLE_TIME;
121 		attr->sample_type	|= PERF_SAMPLE_RAW;
122 		attr->sample_type	|= PERF_SAMPLE_CPU;
123 	}
124 
125 	if (opts->no_delay) {
126 		attr->watermark = 0;
127 		attr->wakeup_events = 1;
128 	}
129 
130 	attr->mmap = track;
131 	attr->comm = track;
132 
133 	if (opts->target_pid == -1 && opts->target_tid == -1 && !opts->system_wide) {
134 		attr->disabled = 1;
135 		attr->enable_on_exec = 1;
136 	}
137 }
138 
139 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
140 {
141 	int cpu, thread;
142 	evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
143 
144 	if (evsel->fd) {
145 		for (cpu = 0; cpu < ncpus; cpu++) {
146 			for (thread = 0; thread < nthreads; thread++) {
147 				FD(evsel, cpu, thread) = -1;
148 			}
149 		}
150 	}
151 
152 	return evsel->fd != NULL ? 0 : -ENOMEM;
153 }
154 
155 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
156 {
157 	evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
158 	if (evsel->sample_id == NULL)
159 		return -ENOMEM;
160 
161 	evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
162 	if (evsel->id == NULL) {
163 		xyarray__delete(evsel->sample_id);
164 		evsel->sample_id = NULL;
165 		return -ENOMEM;
166 	}
167 
168 	return 0;
169 }
170 
171 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
172 {
173 	evsel->counts = zalloc((sizeof(*evsel->counts) +
174 				(ncpus * sizeof(struct perf_counts_values))));
175 	return evsel->counts != NULL ? 0 : -ENOMEM;
176 }
177 
178 void perf_evsel__free_fd(struct perf_evsel *evsel)
179 {
180 	xyarray__delete(evsel->fd);
181 	evsel->fd = NULL;
182 }
183 
184 void perf_evsel__free_id(struct perf_evsel *evsel)
185 {
186 	xyarray__delete(evsel->sample_id);
187 	evsel->sample_id = NULL;
188 	free(evsel->id);
189 	evsel->id = NULL;
190 }
191 
192 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
193 {
194 	int cpu, thread;
195 
196 	for (cpu = 0; cpu < ncpus; cpu++)
197 		for (thread = 0; thread < nthreads; ++thread) {
198 			close(FD(evsel, cpu, thread));
199 			FD(evsel, cpu, thread) = -1;
200 		}
201 }
202 
203 void perf_evsel__exit(struct perf_evsel *evsel)
204 {
205 	assert(list_empty(&evsel->node));
206 	xyarray__delete(evsel->fd);
207 	xyarray__delete(evsel->sample_id);
208 	free(evsel->id);
209 }
210 
211 void perf_evsel__delete(struct perf_evsel *evsel)
212 {
213 	perf_evsel__exit(evsel);
214 	close_cgroup(evsel->cgrp);
215 	free(evsel->name);
216 	free(evsel);
217 }
218 
219 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
220 			      int cpu, int thread, bool scale)
221 {
222 	struct perf_counts_values count;
223 	size_t nv = scale ? 3 : 1;
224 
225 	if (FD(evsel, cpu, thread) < 0)
226 		return -EINVAL;
227 
228 	if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
229 		return -ENOMEM;
230 
231 	if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
232 		return -errno;
233 
234 	if (scale) {
235 		if (count.run == 0)
236 			count.val = 0;
237 		else if (count.run < count.ena)
238 			count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
239 	} else
240 		count.ena = count.run = 0;
241 
242 	evsel->counts->cpu[cpu] = count;
243 	return 0;
244 }
245 
246 int __perf_evsel__read(struct perf_evsel *evsel,
247 		       int ncpus, int nthreads, bool scale)
248 {
249 	size_t nv = scale ? 3 : 1;
250 	int cpu, thread;
251 	struct perf_counts_values *aggr = &evsel->counts->aggr, count;
252 
253 	aggr->val = aggr->ena = aggr->run = 0;
254 
255 	for (cpu = 0; cpu < ncpus; cpu++) {
256 		for (thread = 0; thread < nthreads; thread++) {
257 			if (FD(evsel, cpu, thread) < 0)
258 				continue;
259 
260 			if (readn(FD(evsel, cpu, thread),
261 				  &count, nv * sizeof(u64)) < 0)
262 				return -errno;
263 
264 			aggr->val += count.val;
265 			if (scale) {
266 				aggr->ena += count.ena;
267 				aggr->run += count.run;
268 			}
269 		}
270 	}
271 
272 	evsel->counts->scaled = 0;
273 	if (scale) {
274 		if (aggr->run == 0) {
275 			evsel->counts->scaled = -1;
276 			aggr->val = 0;
277 			return 0;
278 		}
279 
280 		if (aggr->run < aggr->ena) {
281 			evsel->counts->scaled = 1;
282 			aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
283 		}
284 	} else
285 		aggr->ena = aggr->run = 0;
286 
287 	return 0;
288 }
289 
290 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
291 			      struct thread_map *threads, bool group,
292 			      struct xyarray *group_fds)
293 {
294 	int cpu, thread;
295 	unsigned long flags = 0;
296 	int pid = -1, err;
297 
298 	if (evsel->fd == NULL &&
299 	    perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
300 		return -ENOMEM;
301 
302 	if (evsel->cgrp) {
303 		flags = PERF_FLAG_PID_CGROUP;
304 		pid = evsel->cgrp->fd;
305 	}
306 
307 	for (cpu = 0; cpu < cpus->nr; cpu++) {
308 		int group_fd = group_fds ? GROUP_FD(group_fds, cpu) : -1;
309 
310 		for (thread = 0; thread < threads->nr; thread++) {
311 
312 			if (!evsel->cgrp)
313 				pid = threads->map[thread];
314 
315 			FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
316 								     pid,
317 								     cpus->map[cpu],
318 								     group_fd, flags);
319 			if (FD(evsel, cpu, thread) < 0) {
320 				err = -errno;
321 				goto out_close;
322 			}
323 
324 			if (group && group_fd == -1)
325 				group_fd = FD(evsel, cpu, thread);
326 		}
327 	}
328 
329 	return 0;
330 
331 out_close:
332 	do {
333 		while (--thread >= 0) {
334 			close(FD(evsel, cpu, thread));
335 			FD(evsel, cpu, thread) = -1;
336 		}
337 		thread = threads->nr;
338 	} while (--cpu >= 0);
339 	return err;
340 }
341 
342 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
343 {
344 	if (evsel->fd == NULL)
345 		return;
346 
347 	perf_evsel__close_fd(evsel, ncpus, nthreads);
348 	perf_evsel__free_fd(evsel);
349 	evsel->fd = NULL;
350 }
351 
352 static struct {
353 	struct cpu_map map;
354 	int cpus[1];
355 } empty_cpu_map = {
356 	.map.nr	= 1,
357 	.cpus	= { -1, },
358 };
359 
360 static struct {
361 	struct thread_map map;
362 	int threads[1];
363 } empty_thread_map = {
364 	.map.nr	 = 1,
365 	.threads = { -1, },
366 };
367 
368 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
369 		     struct thread_map *threads, bool group,
370 		     struct xyarray *group_fd)
371 {
372 	if (cpus == NULL) {
373 		/* Work around old compiler warnings about strict aliasing */
374 		cpus = &empty_cpu_map.map;
375 	}
376 
377 	if (threads == NULL)
378 		threads = &empty_thread_map.map;
379 
380 	return __perf_evsel__open(evsel, cpus, threads, group, group_fd);
381 }
382 
383 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
384 			     struct cpu_map *cpus, bool group,
385 			     struct xyarray *group_fd)
386 {
387 	return __perf_evsel__open(evsel, cpus, &empty_thread_map.map, group,
388 				  group_fd);
389 }
390 
391 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
392 				struct thread_map *threads, bool group,
393 				struct xyarray *group_fd)
394 {
395 	return __perf_evsel__open(evsel, &empty_cpu_map.map, threads, group,
396 				  group_fd);
397 }
398 
399 static int perf_event__parse_id_sample(const union perf_event *event, u64 type,
400 				       struct perf_sample *sample)
401 {
402 	const u64 *array = event->sample.array;
403 
404 	array += ((event->header.size -
405 		   sizeof(event->header)) / sizeof(u64)) - 1;
406 
407 	if (type & PERF_SAMPLE_CPU) {
408 		u32 *p = (u32 *)array;
409 		sample->cpu = *p;
410 		array--;
411 	}
412 
413 	if (type & PERF_SAMPLE_STREAM_ID) {
414 		sample->stream_id = *array;
415 		array--;
416 	}
417 
418 	if (type & PERF_SAMPLE_ID) {
419 		sample->id = *array;
420 		array--;
421 	}
422 
423 	if (type & PERF_SAMPLE_TIME) {
424 		sample->time = *array;
425 		array--;
426 	}
427 
428 	if (type & PERF_SAMPLE_TID) {
429 		u32 *p = (u32 *)array;
430 		sample->pid = p[0];
431 		sample->tid = p[1];
432 	}
433 
434 	return 0;
435 }
436 
437 static bool sample_overlap(const union perf_event *event,
438 			   const void *offset, u64 size)
439 {
440 	const void *base = event;
441 
442 	if (offset + size > base + event->header.size)
443 		return true;
444 
445 	return false;
446 }
447 
448 int perf_event__parse_sample(const union perf_event *event, u64 type,
449 			     int sample_size, bool sample_id_all,
450 			     struct perf_sample *data, bool swapped)
451 {
452 	const u64 *array;
453 
454 	/*
455 	 * used for cross-endian analysis. See git commit 65014ab3
456 	 * for why this goofiness is needed.
457 	 */
458 	union {
459 		u64 val64;
460 		u32 val32[2];
461 	} u;
462 
463 	memset(data, 0, sizeof(*data));
464 	data->cpu = data->pid = data->tid = -1;
465 	data->stream_id = data->id = data->time = -1ULL;
466 
467 	if (event->header.type != PERF_RECORD_SAMPLE) {
468 		if (!sample_id_all)
469 			return 0;
470 		return perf_event__parse_id_sample(event, type, data);
471 	}
472 
473 	array = event->sample.array;
474 
475 	if (sample_size + sizeof(event->header) > event->header.size)
476 		return -EFAULT;
477 
478 	if (type & PERF_SAMPLE_IP) {
479 		data->ip = event->ip.ip;
480 		array++;
481 	}
482 
483 	if (type & PERF_SAMPLE_TID) {
484 		u.val64 = *array;
485 		if (swapped) {
486 			/* undo swap of u64, then swap on individual u32s */
487 			u.val64 = bswap_64(u.val64);
488 			u.val32[0] = bswap_32(u.val32[0]);
489 			u.val32[1] = bswap_32(u.val32[1]);
490 		}
491 
492 		data->pid = u.val32[0];
493 		data->tid = u.val32[1];
494 		array++;
495 	}
496 
497 	if (type & PERF_SAMPLE_TIME) {
498 		data->time = *array;
499 		array++;
500 	}
501 
502 	data->addr = 0;
503 	if (type & PERF_SAMPLE_ADDR) {
504 		data->addr = *array;
505 		array++;
506 	}
507 
508 	data->id = -1ULL;
509 	if (type & PERF_SAMPLE_ID) {
510 		data->id = *array;
511 		array++;
512 	}
513 
514 	if (type & PERF_SAMPLE_STREAM_ID) {
515 		data->stream_id = *array;
516 		array++;
517 	}
518 
519 	if (type & PERF_SAMPLE_CPU) {
520 
521 		u.val64 = *array;
522 		if (swapped) {
523 			/* undo swap of u64, then swap on individual u32s */
524 			u.val64 = bswap_64(u.val64);
525 			u.val32[0] = bswap_32(u.val32[0]);
526 		}
527 
528 		data->cpu = u.val32[0];
529 		array++;
530 	}
531 
532 	if (type & PERF_SAMPLE_PERIOD) {
533 		data->period = *array;
534 		array++;
535 	}
536 
537 	if (type & PERF_SAMPLE_READ) {
538 		fprintf(stderr, "PERF_SAMPLE_READ is unsuported for now\n");
539 		return -1;
540 	}
541 
542 	if (type & PERF_SAMPLE_CALLCHAIN) {
543 		if (sample_overlap(event, array, sizeof(data->callchain->nr)))
544 			return -EFAULT;
545 
546 		data->callchain = (struct ip_callchain *)array;
547 
548 		if (sample_overlap(event, array, data->callchain->nr))
549 			return -EFAULT;
550 
551 		array += 1 + data->callchain->nr;
552 	}
553 
554 	if (type & PERF_SAMPLE_RAW) {
555 		const u64 *pdata;
556 
557 		u.val64 = *array;
558 		if (WARN_ONCE(swapped,
559 			      "Endianness of raw data not corrected!\n")) {
560 			/* undo swap of u64, then swap on individual u32s */
561 			u.val64 = bswap_64(u.val64);
562 			u.val32[0] = bswap_32(u.val32[0]);
563 			u.val32[1] = bswap_32(u.val32[1]);
564 		}
565 
566 		if (sample_overlap(event, array, sizeof(u32)))
567 			return -EFAULT;
568 
569 		data->raw_size = u.val32[0];
570 		pdata = (void *) array + sizeof(u32);
571 
572 		if (sample_overlap(event, pdata, data->raw_size))
573 			return -EFAULT;
574 
575 		data->raw_data = (void *) pdata;
576 	}
577 
578 	return 0;
579 }
580 
581 int perf_event__synthesize_sample(union perf_event *event, u64 type,
582 				  const struct perf_sample *sample,
583 				  bool swapped)
584 {
585 	u64 *array;
586 
587 	/*
588 	 * used for cross-endian analysis. See git commit 65014ab3
589 	 * for why this goofiness is needed.
590 	 */
591 	union {
592 		u64 val64;
593 		u32 val32[2];
594 	} u;
595 
596 	array = event->sample.array;
597 
598 	if (type & PERF_SAMPLE_IP) {
599 		event->ip.ip = sample->ip;
600 		array++;
601 	}
602 
603 	if (type & PERF_SAMPLE_TID) {
604 		u.val32[0] = sample->pid;
605 		u.val32[1] = sample->tid;
606 		if (swapped) {
607 			/*
608 			 * Inverse of what is done in perf_event__parse_sample
609 			 */
610 			u.val32[0] = bswap_32(u.val32[0]);
611 			u.val32[1] = bswap_32(u.val32[1]);
612 			u.val64 = bswap_64(u.val64);
613 		}
614 
615 		*array = u.val64;
616 		array++;
617 	}
618 
619 	if (type & PERF_SAMPLE_TIME) {
620 		*array = sample->time;
621 		array++;
622 	}
623 
624 	if (type & PERF_SAMPLE_ADDR) {
625 		*array = sample->addr;
626 		array++;
627 	}
628 
629 	if (type & PERF_SAMPLE_ID) {
630 		*array = sample->id;
631 		array++;
632 	}
633 
634 	if (type & PERF_SAMPLE_STREAM_ID) {
635 		*array = sample->stream_id;
636 		array++;
637 	}
638 
639 	if (type & PERF_SAMPLE_CPU) {
640 		u.val32[0] = sample->cpu;
641 		if (swapped) {
642 			/*
643 			 * Inverse of what is done in perf_event__parse_sample
644 			 */
645 			u.val32[0] = bswap_32(u.val32[0]);
646 			u.val64 = bswap_64(u.val64);
647 		}
648 		*array = u.val64;
649 		array++;
650 	}
651 
652 	if (type & PERF_SAMPLE_PERIOD) {
653 		*array = sample->period;
654 		array++;
655 	}
656 
657 	return 0;
658 }
659