]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blob - tools/perf/util/evsel.c
projects / mirror_ubuntu-zesty-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
perf evsel: Add missing decrement in id sample parsing
[mirror_ubuntu-zesty-kernel.git] / tools / perf / util / evsel.c
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 <lk/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 "evsel.h"
19 #include "evlist.h"
20 #include "util.h"
21 #include "cpumap.h"
22 #include "thread_map.h"
23 #include "target.h"
24 #include "perf_regs.h"
25 #include "debug.h"
26
27 static struct {
28 bool sample_id_all;
29 bool exclude_guest;
30 bool mmap2;
31 } perf_missing_features;
32
33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
34
35 int __perf_evsel__sample_size(u64 sample_type)
36 {
37 u64 mask = sample_type & PERF_SAMPLE_MASK;
38 int size = 0;
39 int i;
40
41 for (i = 0; i < 64; i++) {
42 if (mask & (1ULL << i))
43 size++;
44 }
45
46 size *= sizeof(u64);
47
48 return size;
49 }
50
51 /**
52 * __perf_evsel__calc_id_pos - calculate id_pos.
53 * @sample_type: sample type
54 *
55 * This function returns the position of the event id (PERF_SAMPLE_ID or
56 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
57 * sample_event.
58 */
59 static int __perf_evsel__calc_id_pos(u64 sample_type)
60 {
61 int idx = 0;
62
63 if (sample_type & PERF_SAMPLE_IDENTIFIER)
64 return 0;
65
66 if (!(sample_type & PERF_SAMPLE_ID))
67 return -1;
68
69 if (sample_type & PERF_SAMPLE_IP)
70 idx += 1;
71
72 if (sample_type & PERF_SAMPLE_TID)
73 idx += 1;
74
75 if (sample_type & PERF_SAMPLE_TIME)
76 idx += 1;
77
78 if (sample_type & PERF_SAMPLE_ADDR)
79 idx += 1;
80
81 return idx;
82 }
83
84 /**
85 * __perf_evsel__calc_is_pos - calculate is_pos.
86 * @sample_type: sample type
87 *
88 * This function returns the position (counting backwards) of the event id
89 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
90 * sample_id_all is used there is an id sample appended to non-sample events.
91 */
92 static int __perf_evsel__calc_is_pos(u64 sample_type)
93 {
94 int idx = 1;
95
96 if (sample_type & PERF_SAMPLE_IDENTIFIER)
97 return 1;
98
99 if (!(sample_type & PERF_SAMPLE_ID))
100 return -1;
101
102 if (sample_type & PERF_SAMPLE_CPU)
103 idx += 1;
104
105 if (sample_type & PERF_SAMPLE_STREAM_ID)
106 idx += 1;
107
108 return idx;
109 }
110
111 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
112 {
113 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
114 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
115 }
116
117 void hists__init(struct hists *hists)
118 {
119 memset(hists, 0, sizeof(*hists));
120 hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
121 hists->entries_in = &hists->entries_in_array[0];
122 hists->entries_collapsed = RB_ROOT;
123 hists->entries = RB_ROOT;
124 pthread_mutex_init(&hists->lock, NULL);
125 }
126
127 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
128 enum perf_event_sample_format bit)
129 {
130 if (!(evsel->attr.sample_type & bit)) {
131 evsel->attr.sample_type |= bit;
132 evsel->sample_size += sizeof(u64);
133 perf_evsel__calc_id_pos(evsel);
134 }
135 }
136
137 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
138 enum perf_event_sample_format bit)
139 {
140 if (evsel->attr.sample_type & bit) {
141 evsel->attr.sample_type &= ~bit;
142 evsel->sample_size -= sizeof(u64);
143 perf_evsel__calc_id_pos(evsel);
144 }
145 }
146
147 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
148 bool can_sample_identifier)
149 {
150 if (can_sample_identifier) {
151 perf_evsel__reset_sample_bit(evsel, ID);
152 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
153 } else {
154 perf_evsel__set_sample_bit(evsel, ID);
155 }
156 evsel->attr.read_format |= PERF_FORMAT_ID;
157 }
158
159 void perf_evsel__init(struct perf_evsel *evsel,
160 struct perf_event_attr *attr, int idx)
161 {
162 evsel->idx = idx;
163 evsel->attr = *attr;
164 evsel->leader = evsel;
165 INIT_LIST_HEAD(&evsel->node);
166 hists__init(&evsel->hists);
167 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
168 perf_evsel__calc_id_pos(evsel);
169 }
170
171 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
172 {
173 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
174
175 if (evsel != NULL)
176 perf_evsel__init(evsel, attr, idx);
177
178 return evsel;
179 }
180
181 struct event_format *event_format__new(const char *sys, const char *name)
182 {
183 int fd, n;
184 char *filename;
185 void *bf = NULL, *nbf;
186 size_t size = 0, alloc_size = 0;
187 struct event_format *format = NULL;
188
189 if (asprintf(&filename, "%s/%s/%s/format", tracing_events_path, sys, name) < 0)
190 goto out;
191
192 fd = open(filename, O_RDONLY);
193 if (fd < 0)
194 goto out_free_filename;
195
196 do {
197 if (size == alloc_size) {
198 alloc_size += BUFSIZ;
199 nbf = realloc(bf, alloc_size);
200 if (nbf == NULL)
201 goto out_free_bf;
202 bf = nbf;
203 }
204
205 n = read(fd, bf + size, alloc_size - size);
206 if (n < 0)
207 goto out_free_bf;
208 size += n;
209 } while (n > 0);
210
211 pevent_parse_format(&format, bf, size, sys);
212
213 out_free_bf:
214 free(bf);
215 close(fd);
216 out_free_filename:
217 free(filename);
218 out:
219 return format;
220 }
221
222 struct perf_evsel *perf_evsel__newtp(const char *sys, const char *name, int idx)
223 {
224 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
225
226 if (evsel != NULL) {
227 struct perf_event_attr attr = {
228 .type = PERF_TYPE_TRACEPOINT,
229 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
230 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
231 };
232
233 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
234 goto out_free;
235
236 evsel->tp_format = event_format__new(sys, name);
237 if (evsel->tp_format == NULL)
238 goto out_free;
239
240 event_attr_init(&attr);
241 attr.config = evsel->tp_format->id;
242 attr.sample_period = 1;
243 perf_evsel__init(evsel, &attr, idx);
244 }
245
246 return evsel;
247
248 out_free:
249 free(evsel->name);
250 free(evsel);
251 return NULL;
252 }
253
254 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
255 "cycles",
256 "instructions",
257 "cache-references",
258 "cache-misses",
259 "branches",
260 "branch-misses",
261 "bus-cycles",
262 "stalled-cycles-frontend",
263 "stalled-cycles-backend",
264 "ref-cycles",
265 };
266
267 static const char *__perf_evsel__hw_name(u64 config)
268 {
269 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
270 return perf_evsel__hw_names[config];
271
272 return "unknown-hardware";
273 }
274
275 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
276 {
277 int colon = 0, r = 0;
278 struct perf_event_attr *attr = &evsel->attr;
279 bool exclude_guest_default = false;
280
281 #define MOD_PRINT(context, mod) do { \
282 if (!attr->exclude_##context) { \
283 if (!colon) colon = ++r; \
284 r += scnprintf(bf + r, size - r, "%c", mod); \
285 } } while(0)
286
287 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
288 MOD_PRINT(kernel, 'k');
289 MOD_PRINT(user, 'u');
290 MOD_PRINT(hv, 'h');
291 exclude_guest_default = true;
292 }
293
294 if (attr->precise_ip) {
295 if (!colon)
296 colon = ++r;
297 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
298 exclude_guest_default = true;
299 }
300
301 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
302 MOD_PRINT(host, 'H');
303 MOD_PRINT(guest, 'G');
304 }
305 #undef MOD_PRINT
306 if (colon)
307 bf[colon - 1] = ':';
308 return r;
309 }
310
311 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
312 {
313 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
314 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
315 }
316
317 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
318 "cpu-clock",
319 "task-clock",
320 "page-faults",
321 "context-switches",
322 "cpu-migrations",
323 "minor-faults",
324 "major-faults",
325 "alignment-faults",
326 "emulation-faults",
327 "dummy",
328 };
329
330 static const char *__perf_evsel__sw_name(u64 config)
331 {
332 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
333 return perf_evsel__sw_names[config];
334 return "unknown-software";
335 }
336
337 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
338 {
339 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
340 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
341 }
342
343 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
344 {
345 int r;
346
347 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
348
349 if (type & HW_BREAKPOINT_R)
350 r += scnprintf(bf + r, size - r, "r");
351
352 if (type & HW_BREAKPOINT_W)
353 r += scnprintf(bf + r, size - r, "w");
354
355 if (type & HW_BREAKPOINT_X)
356 r += scnprintf(bf + r, size - r, "x");
357
358 return r;
359 }
360
361 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
362 {
363 struct perf_event_attr *attr = &evsel->attr;
364 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
365 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
366 }
367
368 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
369 [PERF_EVSEL__MAX_ALIASES] = {
370 { "L1-dcache", "l1-d", "l1d", "L1-data", },
371 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
372 { "LLC", "L2", },
373 { "dTLB", "d-tlb", "Data-TLB", },
374 { "iTLB", "i-tlb", "Instruction-TLB", },
375 { "branch", "branches", "bpu", "btb", "bpc", },
376 { "node", },
377 };
378
379 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
380 [PERF_EVSEL__MAX_ALIASES] = {
381 { "load", "loads", "read", },
382 { "store", "stores", "write", },
383 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
384 };
385
386 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
387 [PERF_EVSEL__MAX_ALIASES] = {
388 { "refs", "Reference", "ops", "access", },
389 { "misses", "miss", },
390 };
391
392 #define C(x) PERF_COUNT_HW_CACHE_##x
393 #define CACHE_READ (1 << C(OP_READ))
394 #define CACHE_WRITE (1 << C(OP_WRITE))
395 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
396 #define COP(x) (1 << x)
397
398 /*
399 * cache operartion stat
400 * L1I : Read and prefetch only
401 * ITLB and BPU : Read-only
402 */
403 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
404 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
405 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
406 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
407 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
408 [C(ITLB)] = (CACHE_READ),
409 [C(BPU)] = (CACHE_READ),
410 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
411 };
412
413 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
414 {
415 if (perf_evsel__hw_cache_stat[type] & COP(op))
416 return true; /* valid */
417 else
418 return false; /* invalid */
419 }
420
421 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
422 char *bf, size_t size)
423 {
424 if (result) {
425 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
426 perf_evsel__hw_cache_op[op][0],
427 perf_evsel__hw_cache_result[result][0]);
428 }
429
430 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
431 perf_evsel__hw_cache_op[op][1]);
432 }
433
434 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
435 {
436 u8 op, result, type = (config >> 0) & 0xff;
437 const char *err = "unknown-ext-hardware-cache-type";
438
439 if (type > PERF_COUNT_HW_CACHE_MAX)
440 goto out_err;
441
442 op = (config >> 8) & 0xff;
443 err = "unknown-ext-hardware-cache-op";
444 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
445 goto out_err;
446
447 result = (config >> 16) & 0xff;
448 err = "unknown-ext-hardware-cache-result";
449 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
450 goto out_err;
451
452 err = "invalid-cache";
453 if (!perf_evsel__is_cache_op_valid(type, op))
454 goto out_err;
455
456 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
457 out_err:
458 return scnprintf(bf, size, "%s", err);
459 }
460
461 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
462 {
463 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
464 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
465 }
466
467 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
468 {
469 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
470 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
471 }
472
473 const char *perf_evsel__name(struct perf_evsel *evsel)
474 {
475 char bf[128];
476
477 if (evsel->name)
478 return evsel->name;
479
480 switch (evsel->attr.type) {
481 case PERF_TYPE_RAW:
482 perf_evsel__raw_name(evsel, bf, sizeof(bf));
483 break;
484
485 case PERF_TYPE_HARDWARE:
486 perf_evsel__hw_name(evsel, bf, sizeof(bf));
487 break;
488
489 case PERF_TYPE_HW_CACHE:
490 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
491 break;
492
493 case PERF_TYPE_SOFTWARE:
494 perf_evsel__sw_name(evsel, bf, sizeof(bf));
495 break;
496
497 case PERF_TYPE_TRACEPOINT:
498 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
499 break;
500
501 case PERF_TYPE_BREAKPOINT:
502 perf_evsel__bp_name(evsel, bf, sizeof(bf));
503 break;
504
505 default:
506 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
507 evsel->attr.type);
508 break;
509 }
510
511 evsel->name = strdup(bf);
512
513 return evsel->name ?: "unknown";
514 }
515
516 const char *perf_evsel__group_name(struct perf_evsel *evsel)
517 {
518 return evsel->group_name ?: "anon group";
519 }
520
521 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
522 {
523 int ret;
524 struct perf_evsel *pos;
525 const char *group_name = perf_evsel__group_name(evsel);
526
527 ret = scnprintf(buf, size, "%s", group_name);
528
529 ret += scnprintf(buf + ret, size - ret, " { %s",
530 perf_evsel__name(evsel));
531
532 for_each_group_member(pos, evsel)
533 ret += scnprintf(buf + ret, size - ret, ", %s",
534 perf_evsel__name(pos));
535
536 ret += scnprintf(buf + ret, size - ret, " }");
537
538 return ret;
539 }
540
541 /*
542 * The enable_on_exec/disabled value strategy:
543 *
544 * 1) For any type of traced program:
545 * - all independent events and group leaders are disabled
546 * - all group members are enabled
547 *
548 * Group members are ruled by group leaders. They need to
549 * be enabled, because the group scheduling relies on that.
550 *
551 * 2) For traced programs executed by perf:
552 * - all independent events and group leaders have
553 * enable_on_exec set
554 * - we don't specifically enable or disable any event during
555 * the record command
556 *
557 * Independent events and group leaders are initially disabled
558 * and get enabled by exec. Group members are ruled by group
559 * leaders as stated in 1).
560 *
561 * 3) For traced programs attached by perf (pid/tid):
562 * - we specifically enable or disable all events during
563 * the record command
564 *
565 * When attaching events to already running traced we
566 * enable/disable events specifically, as there's no
567 * initial traced exec call.
568 */
569 void perf_evsel__config(struct perf_evsel *evsel,
570 struct perf_record_opts *opts)
571 {
572 struct perf_evsel *leader = evsel->leader;
573 struct perf_event_attr *attr = &evsel->attr;
574 int track = !evsel->idx; /* only the first counter needs these */
575
576 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
577 attr->inherit = !opts->no_inherit;
578
579 perf_evsel__set_sample_bit(evsel, IP);
580 perf_evsel__set_sample_bit(evsel, TID);
581
582 if (evsel->sample_read) {
583 perf_evsel__set_sample_bit(evsel, READ);
584
585 /*
586 * We need ID even in case of single event, because
587 * PERF_SAMPLE_READ process ID specific data.
588 */
589 perf_evsel__set_sample_id(evsel, false);
590
591 /*
592 * Apply group format only if we belong to group
593 * with more than one members.
594 */
595 if (leader->nr_members > 1) {
596 attr->read_format |= PERF_FORMAT_GROUP;
597 attr->inherit = 0;
598 }
599 }
600
601 /*
602 * We default some events to a 1 default interval. But keep
603 * it a weak assumption overridable by the user.
604 */
605 if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
606 opts->user_interval != ULLONG_MAX)) {
607 if (opts->freq) {
608 perf_evsel__set_sample_bit(evsel, PERIOD);
609 attr->freq = 1;
610 attr->sample_freq = opts->freq;
611 } else {
612 attr->sample_period = opts->default_interval;
613 }
614 }
615
616 /*
617 * Disable sampling for all group members other
618 * than leader in case leader 'leads' the sampling.
619 */
620 if ((leader != evsel) && leader->sample_read) {
621 attr->sample_freq = 0;
622 attr->sample_period = 0;
623 }
624
625 if (opts->no_samples)
626 attr->sample_freq = 0;
627
628 if (opts->inherit_stat)
629 attr->inherit_stat = 1;
630
631 if (opts->sample_address) {
632 perf_evsel__set_sample_bit(evsel, ADDR);
633 attr->mmap_data = track;
634 }
635
636 if (opts->call_graph) {
637 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
638
639 if (opts->call_graph == CALLCHAIN_DWARF) {
640 perf_evsel__set_sample_bit(evsel, REGS_USER);
641 perf_evsel__set_sample_bit(evsel, STACK_USER);
642 attr->sample_regs_user = PERF_REGS_MASK;
643 attr->sample_stack_user = opts->stack_dump_size;
644 attr->exclude_callchain_user = 1;
645 }
646 }
647
648 if (perf_target__has_cpu(&opts->target))
649 perf_evsel__set_sample_bit(evsel, CPU);
650
651 if (opts->period)
652 perf_evsel__set_sample_bit(evsel, PERIOD);
653
654 if (!perf_missing_features.sample_id_all &&
655 (opts->sample_time || !opts->no_inherit ||
656 perf_target__has_cpu(&opts->target)))
657 perf_evsel__set_sample_bit(evsel, TIME);
658
659 if (opts->raw_samples) {
660 perf_evsel__set_sample_bit(evsel, TIME);
661 perf_evsel__set_sample_bit(evsel, RAW);
662 perf_evsel__set_sample_bit(evsel, CPU);
663 }
664
665 if (opts->sample_address)
666 attr->sample_type |= PERF_SAMPLE_DATA_SRC;
667
668 if (opts->no_delay) {
669 attr->watermark = 0;
670 attr->wakeup_events = 1;
671 }
672 if (opts->branch_stack) {
673 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
674 attr->branch_sample_type = opts->branch_stack;
675 }
676
677 if (opts->sample_weight)
678 attr->sample_type |= PERF_SAMPLE_WEIGHT;
679
680 attr->mmap = track;
681 attr->mmap2 = track && !perf_missing_features.mmap2;
682 attr->comm = track;
683
684 if (opts->sample_transaction)
685 attr->sample_type |= PERF_SAMPLE_TRANSACTION;
686
687 /*
688 * XXX see the function comment above
689 *
690 * Disabling only independent events or group leaders,
691 * keeping group members enabled.
692 */
693 if (perf_evsel__is_group_leader(evsel))
694 attr->disabled = 1;
695
696 /*
697 * Setting enable_on_exec for independent events and
698 * group leaders for traced executed by perf.
699 */
700 if (perf_target__none(&opts->target) && perf_evsel__is_group_leader(evsel))
701 attr->enable_on_exec = 1;
702 }
703
704 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
705 {
706 int cpu, thread;
707 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
708
709 if (evsel->fd) {
710 for (cpu = 0; cpu < ncpus; cpu++) {
711 for (thread = 0; thread < nthreads; thread++) {
712 FD(evsel, cpu, thread) = -1;
713 }
714 }
715 }
716
717 return evsel->fd != NULL ? 0 : -ENOMEM;
718 }
719
720 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
721 int ioc, void *arg)
722 {
723 int cpu, thread;
724
725 for (cpu = 0; cpu < ncpus; cpu++) {
726 for (thread = 0; thread < nthreads; thread++) {
727 int fd = FD(evsel, cpu, thread),
728 err = ioctl(fd, ioc, arg);
729
730 if (err)
731 return err;
732 }
733 }
734
735 return 0;
736 }
737
738 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
739 const char *filter)
740 {
741 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
742 PERF_EVENT_IOC_SET_FILTER,
743 (void *)filter);
744 }
745
746 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
747 {
748 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
749 PERF_EVENT_IOC_ENABLE,
750 0);
751 }
752
753 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
754 {
755 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
756 if (evsel->sample_id == NULL)
757 return -ENOMEM;
758
759 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
760 if (evsel->id == NULL) {
761 xyarray__delete(evsel->sample_id);
762 evsel->sample_id = NULL;
763 return -ENOMEM;
764 }
765
766 return 0;
767 }
768
769 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
770 {
771 memset(evsel->counts, 0, (sizeof(*evsel->counts) +
772 (ncpus * sizeof(struct perf_counts_values))));
773 }
774
775 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
776 {
777 evsel->counts = zalloc((sizeof(*evsel->counts) +
778 (ncpus * sizeof(struct perf_counts_values))));
779 return evsel->counts != NULL ? 0 : -ENOMEM;
780 }
781
782 void perf_evsel__free_fd(struct perf_evsel *evsel)
783 {
784 xyarray__delete(evsel->fd);
785 evsel->fd = NULL;
786 }
787
788 void perf_evsel__free_id(struct perf_evsel *evsel)
789 {
790 xyarray__delete(evsel->sample_id);
791 evsel->sample_id = NULL;
792 free(evsel->id);
793 evsel->id = NULL;
794 }
795
796 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
797 {
798 int cpu, thread;
799
800 for (cpu = 0; cpu < ncpus; cpu++)
801 for (thread = 0; thread < nthreads; ++thread) {
802 close(FD(evsel, cpu, thread));
803 FD(evsel, cpu, thread) = -1;
804 }
805 }
806
807 void perf_evsel__free_counts(struct perf_evsel *evsel)
808 {
809 free(evsel->counts);
810 }
811
812 void perf_evsel__exit(struct perf_evsel *evsel)
813 {
814 assert(list_empty(&evsel->node));
815 perf_evsel__free_fd(evsel);
816 perf_evsel__free_id(evsel);
817 }
818
819 void perf_evsel__delete(struct perf_evsel *evsel)
820 {
821 perf_evsel__exit(evsel);
822 close_cgroup(evsel->cgrp);
823 free(evsel->group_name);
824 if (evsel->tp_format)
825 pevent_free_format(evsel->tp_format);
826 free(evsel->name);
827 free(evsel);
828 }
829
830 static inline void compute_deltas(struct perf_evsel *evsel,
831 int cpu,
832 struct perf_counts_values *count)
833 {
834 struct perf_counts_values tmp;
835
836 if (!evsel->prev_raw_counts)
837 return;
838
839 if (cpu == -1) {
840 tmp = evsel->prev_raw_counts->aggr;
841 evsel->prev_raw_counts->aggr = *count;
842 } else {
843 tmp = evsel->prev_raw_counts->cpu[cpu];
844 evsel->prev_raw_counts->cpu[cpu] = *count;
845 }
846
847 count->val = count->val - tmp.val;
848 count->ena = count->ena - tmp.ena;
849 count->run = count->run - tmp.run;
850 }
851
852 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
853 int cpu, int thread, bool scale)
854 {
855 struct perf_counts_values count;
856 size_t nv = scale ? 3 : 1;
857
858 if (FD(evsel, cpu, thread) < 0)
859 return -EINVAL;
860
861 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
862 return -ENOMEM;
863
864 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
865 return -errno;
866
867 compute_deltas(evsel, cpu, &count);
868
869 if (scale) {
870 if (count.run == 0)
871 count.val = 0;
872 else if (count.run < count.ena)
873 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
874 } else
875 count.ena = count.run = 0;
876
877 evsel->counts->cpu[cpu] = count;
878 return 0;
879 }
880
881 int __perf_evsel__read(struct perf_evsel *evsel,
882 int ncpus, int nthreads, bool scale)
883 {
884 size_t nv = scale ? 3 : 1;
885 int cpu, thread;
886 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
887
888 aggr->val = aggr->ena = aggr->run = 0;
889
890 for (cpu = 0; cpu < ncpus; cpu++) {
891 for (thread = 0; thread < nthreads; thread++) {
892 if (FD(evsel, cpu, thread) < 0)
893 continue;
894
895 if (readn(FD(evsel, cpu, thread),
896 &count, nv * sizeof(u64)) < 0)
897 return -errno;
898
899 aggr->val += count.val;
900 if (scale) {
901 aggr->ena += count.ena;
902 aggr->run += count.run;
903 }
904 }
905 }
906
907 compute_deltas(evsel, -1, aggr);
908
909 evsel->counts->scaled = 0;
910 if (scale) {
911 if (aggr->run == 0) {
912 evsel->counts->scaled = -1;
913 aggr->val = 0;
914 return 0;
915 }
916
917 if (aggr->run < aggr->ena) {
918 evsel->counts->scaled = 1;
919 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
920 }
921 } else
922 aggr->ena = aggr->run = 0;
923
924 return 0;
925 }
926
927 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
928 {
929 struct perf_evsel *leader = evsel->leader;
930 int fd;
931
932 if (perf_evsel__is_group_leader(evsel))
933 return -1;
934
935 /*
936 * Leader must be already processed/open,
937 * if not it's a bug.
938 */
939 BUG_ON(!leader->fd);
940
941 fd = FD(leader, cpu, thread);
942 BUG_ON(fd == -1);
943
944 return fd;
945 }
946
947 #define __PRINT_ATTR(fmt, cast, field) \
948 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
949
950 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
951 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
952 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
953 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
954
955 #define PRINT_ATTR2N(name1, field1, name2, field2) \
956 fprintf(fp, " %-19s %u %-19s %u\n", \
957 name1, attr->field1, name2, attr->field2)
958
959 #define PRINT_ATTR2(field1, field2) \
960 PRINT_ATTR2N(#field1, field1, #field2, field2)
961
962 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
963 {
964 size_t ret = 0;
965
966 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
967 ret += fprintf(fp, "perf_event_attr:\n");
968
969 ret += PRINT_ATTR_U32(type);
970 ret += PRINT_ATTR_U32(size);
971 ret += PRINT_ATTR_X64(config);
972 ret += PRINT_ATTR_U64(sample_period);
973 ret += PRINT_ATTR_U64(sample_freq);
974 ret += PRINT_ATTR_X64(sample_type);
975 ret += PRINT_ATTR_X64(read_format);
976
977 ret += PRINT_ATTR2(disabled, inherit);
978 ret += PRINT_ATTR2(pinned, exclusive);
979 ret += PRINT_ATTR2(exclude_user, exclude_kernel);
980 ret += PRINT_ATTR2(exclude_hv, exclude_idle);
981 ret += PRINT_ATTR2(mmap, comm);
982 ret += PRINT_ATTR2(freq, inherit_stat);
983 ret += PRINT_ATTR2(enable_on_exec, task);
984 ret += PRINT_ATTR2(watermark, precise_ip);
985 ret += PRINT_ATTR2(mmap_data, sample_id_all);
986 ret += PRINT_ATTR2(exclude_host, exclude_guest);
987 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
988 "excl.callchain_user", exclude_callchain_user);
989 ret += PRINT_ATTR_U32(mmap2);
990
991 ret += PRINT_ATTR_U32(wakeup_events);
992 ret += PRINT_ATTR_U32(wakeup_watermark);
993 ret += PRINT_ATTR_X32(bp_type);
994 ret += PRINT_ATTR_X64(bp_addr);
995 ret += PRINT_ATTR_X64(config1);
996 ret += PRINT_ATTR_U64(bp_len);
997 ret += PRINT_ATTR_X64(config2);
998 ret += PRINT_ATTR_X64(branch_sample_type);
999 ret += PRINT_ATTR_X64(sample_regs_user);
1000 ret += PRINT_ATTR_U32(sample_stack_user);
1001
1002 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1003
1004 return ret;
1005 }
1006
1007 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1008 struct thread_map *threads)
1009 {
1010 int cpu, thread;
1011 unsigned long flags = 0;
1012 int pid = -1, err;
1013 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1014
1015 if (evsel->fd == NULL &&
1016 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
1017 return -ENOMEM;
1018
1019 if (evsel->cgrp) {
1020 flags = PERF_FLAG_PID_CGROUP;
1021 pid = evsel->cgrp->fd;
1022 }
1023
1024 fallback_missing_features:
1025 if (perf_missing_features.mmap2)
1026 evsel->attr.mmap2 = 0;
1027 if (perf_missing_features.exclude_guest)
1028 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1029 retry_sample_id:
1030 if (perf_missing_features.sample_id_all)
1031 evsel->attr.sample_id_all = 0;
1032
1033 if (verbose >= 2)
1034 perf_event_attr__fprintf(&evsel->attr, stderr);
1035
1036 for (cpu = 0; cpu < cpus->nr; cpu++) {
1037
1038 for (thread = 0; thread < threads->nr; thread++) {
1039 int group_fd;
1040
1041 if (!evsel->cgrp)
1042 pid = threads->map[thread];
1043
1044 group_fd = get_group_fd(evsel, cpu, thread);
1045 retry_open:
1046 pr_debug2("perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1047 pid, cpus->map[cpu], group_fd, flags);
1048
1049 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1050 pid,
1051 cpus->map[cpu],
1052 group_fd, flags);
1053 if (FD(evsel, cpu, thread) < 0) {
1054 err = -errno;
1055 goto try_fallback;
1056 }
1057 set_rlimit = NO_CHANGE;
1058 }
1059 }
1060
1061 return 0;
1062
1063 try_fallback:
1064 /*
1065 * perf stat needs between 5 and 22 fds per CPU. When we run out
1066 * of them try to increase the limits.
1067 */
1068 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1069 struct rlimit l;
1070 int old_errno = errno;
1071
1072 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1073 if (set_rlimit == NO_CHANGE)
1074 l.rlim_cur = l.rlim_max;
1075 else {
1076 l.rlim_cur = l.rlim_max + 1000;
1077 l.rlim_max = l.rlim_cur;
1078 }
1079 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1080 set_rlimit++;
1081 errno = old_errno;
1082 goto retry_open;
1083 }
1084 }
1085 errno = old_errno;
1086 }
1087
1088 if (err != -EINVAL || cpu > 0 || thread > 0)
1089 goto out_close;
1090
1091 if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1092 perf_missing_features.mmap2 = true;
1093 goto fallback_missing_features;
1094 } else if (!perf_missing_features.exclude_guest &&
1095 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1096 perf_missing_features.exclude_guest = true;
1097 goto fallback_missing_features;
1098 } else if (!perf_missing_features.sample_id_all) {
1099 perf_missing_features.sample_id_all = true;
1100 goto retry_sample_id;
1101 }
1102
1103 out_close:
1104 do {
1105 while (--thread >= 0) {
1106 close(FD(evsel, cpu, thread));
1107 FD(evsel, cpu, thread) = -1;
1108 }
1109 thread = threads->nr;
1110 } while (--cpu >= 0);
1111 return err;
1112 }
1113
1114 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1115 {
1116 if (evsel->fd == NULL)
1117 return;
1118
1119 perf_evsel__close_fd(evsel, ncpus, nthreads);
1120 perf_evsel__free_fd(evsel);
1121 evsel->fd = NULL;
1122 }
1123
1124 static struct {
1125 struct cpu_map map;
1126 int cpus[1];
1127 } empty_cpu_map = {
1128 .map.nr = 1,
1129 .cpus = { -1, },
1130 };
1131
1132 static struct {
1133 struct thread_map map;
1134 int threads[1];
1135 } empty_thread_map = {
1136 .map.nr = 1,
1137 .threads = { -1, },
1138 };
1139
1140 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1141 struct thread_map *threads)
1142 {
1143 if (cpus == NULL) {
1144 /* Work around old compiler warnings about strict aliasing */
1145 cpus = &empty_cpu_map.map;
1146 }
1147
1148 if (threads == NULL)
1149 threads = &empty_thread_map.map;
1150
1151 return __perf_evsel__open(evsel, cpus, threads);
1152 }
1153
1154 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1155 struct cpu_map *cpus)
1156 {
1157 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1158 }
1159
1160 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1161 struct thread_map *threads)
1162 {
1163 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1164 }
1165
1166 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1167 const union perf_event *event,
1168 struct perf_sample *sample)
1169 {
1170 u64 type = evsel->attr.sample_type;
1171 const u64 *array = event->sample.array;
1172 bool swapped = evsel->needs_swap;
1173 union u64_swap u;
1174
1175 array += ((event->header.size -
1176 sizeof(event->header)) / sizeof(u64)) - 1;
1177
1178 if (type & PERF_SAMPLE_IDENTIFIER) {
1179 sample->id = *array;
1180 array--;
1181 }
1182
1183 if (type & PERF_SAMPLE_CPU) {
1184 u.val64 = *array;
1185 if (swapped) {
1186 /* undo swap of u64, then swap on individual u32s */
1187 u.val64 = bswap_64(u.val64);
1188 u.val32[0] = bswap_32(u.val32[0]);
1189 }
1190
1191 sample->cpu = u.val32[0];
1192 array--;
1193 }
1194
1195 if (type & PERF_SAMPLE_STREAM_ID) {
1196 sample->stream_id = *array;
1197 array--;
1198 }
1199
1200 if (type & PERF_SAMPLE_ID) {
1201 sample->id = *array;
1202 array--;
1203 }
1204
1205 if (type & PERF_SAMPLE_TIME) {
1206 sample->time = *array;
1207 array--;
1208 }
1209
1210 if (type & PERF_SAMPLE_TID) {
1211 u.val64 = *array;
1212 if (swapped) {
1213 /* undo swap of u64, then swap on individual u32s */
1214 u.val64 = bswap_64(u.val64);
1215 u.val32[0] = bswap_32(u.val32[0]);
1216 u.val32[1] = bswap_32(u.val32[1]);
1217 }
1218
1219 sample->pid = u.val32[0];
1220 sample->tid = u.val32[1];
1221 array--;
1222 }
1223
1224 return 0;
1225 }
1226
1227 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1228 u64 size)
1229 {
1230 return size > max_size || offset + size > endp;
1231 }
1232
1233 #define OVERFLOW_CHECK(offset, size, max_size) \
1234 do { \
1235 if (overflow(endp, (max_size), (offset), (size))) \
1236 return -EFAULT; \
1237 } while (0)
1238
1239 #define OVERFLOW_CHECK_u64(offset) \
1240 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1241
1242 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1243 struct perf_sample *data)
1244 {
1245 u64 type = evsel->attr.sample_type;
1246 bool swapped = evsel->needs_swap;
1247 const u64 *array;
1248 u16 max_size = event->header.size;
1249 const void *endp = (void *)event + max_size;
1250 u64 sz;
1251
1252 /*
1253 * used for cross-endian analysis. See git commit 65014ab3
1254 * for why this goofiness is needed.
1255 */
1256 union u64_swap u;
1257
1258 memset(data, 0, sizeof(*data));
1259 data->cpu = data->pid = data->tid = -1;
1260 data->stream_id = data->id = data->time = -1ULL;
1261 data->period = 1;
1262 data->weight = 0;
1263
1264 if (event->header.type != PERF_RECORD_SAMPLE) {
1265 if (!evsel->attr.sample_id_all)
1266 return 0;
1267 return perf_evsel__parse_id_sample(evsel, event, data);
1268 }
1269
1270 array = event->sample.array;
1271
1272 /*
1273 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1274 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1275 * check the format does not go past the end of the event.
1276 */
1277 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1278 return -EFAULT;
1279
1280 data->id = -1ULL;
1281 if (type & PERF_SAMPLE_IDENTIFIER) {
1282 data->id = *array;
1283 array++;
1284 }
1285
1286 if (type & PERF_SAMPLE_IP) {
1287 data->ip = *array;
1288 array++;
1289 }
1290
1291 if (type & PERF_SAMPLE_TID) {
1292 u.val64 = *array;
1293 if (swapped) {
1294 /* undo swap of u64, then swap on individual u32s */
1295 u.val64 = bswap_64(u.val64);
1296 u.val32[0] = bswap_32(u.val32[0]);
1297 u.val32[1] = bswap_32(u.val32[1]);
1298 }
1299
1300 data->pid = u.val32[0];
1301 data->tid = u.val32[1];
1302 array++;
1303 }
1304
1305 if (type & PERF_SAMPLE_TIME) {
1306 data->time = *array;
1307 array++;
1308 }
1309
1310 data->addr = 0;
1311 if (type & PERF_SAMPLE_ADDR) {
1312 data->addr = *array;
1313 array++;
1314 }
1315
1316 if (type & PERF_SAMPLE_ID) {
1317 data->id = *array;
1318 array++;
1319 }
1320
1321 if (type & PERF_SAMPLE_STREAM_ID) {
1322 data->stream_id = *array;
1323 array++;
1324 }
1325
1326 if (type & PERF_SAMPLE_CPU) {
1327
1328 u.val64 = *array;
1329 if (swapped) {
1330 /* undo swap of u64, then swap on individual u32s */
1331 u.val64 = bswap_64(u.val64);
1332 u.val32[0] = bswap_32(u.val32[0]);
1333 }
1334
1335 data->cpu = u.val32[0];
1336 array++;
1337 }
1338
1339 if (type & PERF_SAMPLE_PERIOD) {
1340 data->period = *array;
1341 array++;
1342 }
1343
1344 if (type & PERF_SAMPLE_READ) {
1345 u64 read_format = evsel->attr.read_format;
1346
1347 OVERFLOW_CHECK_u64(array);
1348 if (read_format & PERF_FORMAT_GROUP)
1349 data->read.group.nr = *array;
1350 else
1351 data->read.one.value = *array;
1352
1353 array++;
1354
1355 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1356 OVERFLOW_CHECK_u64(array);
1357 data->read.time_enabled = *array;
1358 array++;
1359 }
1360
1361 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1362 OVERFLOW_CHECK_u64(array);
1363 data->read.time_running = *array;
1364 array++;
1365 }
1366
1367 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1368 if (read_format & PERF_FORMAT_GROUP) {
1369 const u64 max_group_nr = UINT64_MAX /
1370 sizeof(struct sample_read_value);
1371
1372 if (data->read.group.nr > max_group_nr)
1373 return -EFAULT;
1374 sz = data->read.group.nr *
1375 sizeof(struct sample_read_value);
1376 OVERFLOW_CHECK(array, sz, max_size);
1377 data->read.group.values =
1378 (struct sample_read_value *)array;
1379 array = (void *)array + sz;
1380 } else {
1381 OVERFLOW_CHECK_u64(array);
1382 data->read.one.id = *array;
1383 array++;
1384 }
1385 }
1386
1387 if (type & PERF_SAMPLE_CALLCHAIN) {
1388 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1389
1390 OVERFLOW_CHECK_u64(array);
1391 data->callchain = (struct ip_callchain *)array++;
1392 if (data->callchain->nr > max_callchain_nr)
1393 return -EFAULT;
1394 sz = data->callchain->nr * sizeof(u64);
1395 OVERFLOW_CHECK(array, sz, max_size);
1396 array = (void *)array + sz;
1397 }
1398
1399 if (type & PERF_SAMPLE_RAW) {
1400 OVERFLOW_CHECK_u64(array);
1401 u.val64 = *array;
1402 if (WARN_ONCE(swapped,
1403 "Endianness of raw data not corrected!\n")) {
1404 /* undo swap of u64, then swap on individual u32s */
1405 u.val64 = bswap_64(u.val64);
1406 u.val32[0] = bswap_32(u.val32[0]);
1407 u.val32[1] = bswap_32(u.val32[1]);
1408 }
1409 data->raw_size = u.val32[0];
1410 array = (void *)array + sizeof(u32);
1411
1412 OVERFLOW_CHECK(array, data->raw_size, max_size);
1413 data->raw_data = (void *)array;
1414 array = (void *)array + data->raw_size;
1415 }
1416
1417 if (type & PERF_SAMPLE_BRANCH_STACK) {
1418 const u64 max_branch_nr = UINT64_MAX /
1419 sizeof(struct branch_entry);
1420
1421 OVERFLOW_CHECK_u64(array);
1422 data->branch_stack = (struct branch_stack *)array++;
1423
1424 if (data->branch_stack->nr > max_branch_nr)
1425 return -EFAULT;
1426 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1427 OVERFLOW_CHECK(array, sz, max_size);
1428 array = (void *)array + sz;
1429 }
1430
1431 if (type & PERF_SAMPLE_REGS_USER) {
1432 OVERFLOW_CHECK_u64(array);
1433 data->user_regs.abi = *array;
1434 array++;
1435
1436 if (data->user_regs.abi) {
1437 u64 regs_user = evsel->attr.sample_regs_user;
1438
1439 sz = hweight_long(regs_user) * sizeof(u64);
1440 OVERFLOW_CHECK(array, sz, max_size);
1441 data->user_regs.regs = (u64 *)array;
1442 array = (void *)array + sz;
1443 }
1444 }
1445
1446 if (type & PERF_SAMPLE_STACK_USER) {
1447 OVERFLOW_CHECK_u64(array);
1448 sz = *array++;
1449
1450 data->user_stack.offset = ((char *)(array - 1)
1451 - (char *) event);
1452
1453 if (!sz) {
1454 data->user_stack.size = 0;
1455 } else {
1456 OVERFLOW_CHECK(array, sz, max_size);
1457 data->user_stack.data = (char *)array;
1458 array = (void *)array + sz;
1459 OVERFLOW_CHECK_u64(array);
1460 data->user_stack.size = *array++;
1461 if (WARN_ONCE(data->user_stack.size > sz,
1462 "user stack dump failure\n"))
1463 return -EFAULT;
1464 }
1465 }
1466
1467 data->weight = 0;
1468 if (type & PERF_SAMPLE_WEIGHT) {
1469 OVERFLOW_CHECK_u64(array);
1470 data->weight = *array;
1471 array++;
1472 }
1473
1474 data->data_src = PERF_MEM_DATA_SRC_NONE;
1475 if (type & PERF_SAMPLE_DATA_SRC) {
1476 OVERFLOW_CHECK_u64(array);
1477 data->data_src = *array;
1478 array++;
1479 }
1480
1481 data->transaction = 0;
1482 if (type & PERF_SAMPLE_TRANSACTION) {
1483 data->transaction = *array;
1484 array++;
1485 }
1486
1487 return 0;
1488 }
1489
1490 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1491 u64 sample_regs_user, u64 read_format)
1492 {
1493 size_t sz, result = sizeof(struct sample_event);
1494
1495 if (type & PERF_SAMPLE_IDENTIFIER)
1496 result += sizeof(u64);
1497
1498 if (type & PERF_SAMPLE_IP)
1499 result += sizeof(u64);
1500
1501 if (type & PERF_SAMPLE_TID)
1502 result += sizeof(u64);
1503
1504 if (type & PERF_SAMPLE_TIME)
1505 result += sizeof(u64);
1506
1507 if (type & PERF_SAMPLE_ADDR)
1508 result += sizeof(u64);
1509
1510 if (type & PERF_SAMPLE_ID)
1511 result += sizeof(u64);
1512
1513 if (type & PERF_SAMPLE_STREAM_ID)
1514 result += sizeof(u64);
1515
1516 if (type & PERF_SAMPLE_CPU)
1517 result += sizeof(u64);
1518
1519 if (type & PERF_SAMPLE_PERIOD)
1520 result += sizeof(u64);
1521
1522 if (type & PERF_SAMPLE_READ) {
1523 result += sizeof(u64);
1524 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1525 result += sizeof(u64);
1526 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1527 result += sizeof(u64);
1528 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1529 if (read_format & PERF_FORMAT_GROUP) {
1530 sz = sample->read.group.nr *
1531 sizeof(struct sample_read_value);
1532 result += sz;
1533 } else {
1534 result += sizeof(u64);
1535 }
1536 }
1537
1538 if (type & PERF_SAMPLE_CALLCHAIN) {
1539 sz = (sample->callchain->nr + 1) * sizeof(u64);
1540 result += sz;
1541 }
1542
1543 if (type & PERF_SAMPLE_RAW) {
1544 result += sizeof(u32);
1545 result += sample->raw_size;
1546 }
1547
1548 if (type & PERF_SAMPLE_BRANCH_STACK) {
1549 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1550 sz += sizeof(u64);
1551 result += sz;
1552 }
1553
1554 if (type & PERF_SAMPLE_REGS_USER) {
1555 if (sample->user_regs.abi) {
1556 result += sizeof(u64);
1557 sz = hweight_long(sample_regs_user) * sizeof(u64);
1558 result += sz;
1559 } else {
1560 result += sizeof(u64);
1561 }
1562 }
1563
1564 if (type & PERF_SAMPLE_STACK_USER) {
1565 sz = sample->user_stack.size;
1566 result += sizeof(u64);
1567 if (sz) {
1568 result += sz;
1569 result += sizeof(u64);
1570 }
1571 }
1572
1573 if (type & PERF_SAMPLE_WEIGHT)
1574 result += sizeof(u64);
1575
1576 if (type & PERF_SAMPLE_DATA_SRC)
1577 result += sizeof(u64);
1578
1579 return result;
1580 }
1581
1582 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1583 u64 sample_regs_user, u64 read_format,
1584 const struct perf_sample *sample,
1585 bool swapped)
1586 {
1587 u64 *array;
1588 size_t sz;
1589 /*
1590 * used for cross-endian analysis. See git commit 65014ab3
1591 * for why this goofiness is needed.
1592 */
1593 union u64_swap u;
1594
1595 array = event->sample.array;
1596
1597 if (type & PERF_SAMPLE_IDENTIFIER) {
1598 *array = sample->id;
1599 array++;
1600 }
1601
1602 if (type & PERF_SAMPLE_IP) {
1603 *array = sample->ip;
1604 array++;
1605 }
1606
1607 if (type & PERF_SAMPLE_TID) {
1608 u.val32[0] = sample->pid;
1609 u.val32[1] = sample->tid;
1610 if (swapped) {
1611 /*
1612 * Inverse of what is done in perf_evsel__parse_sample
1613 */
1614 u.val32[0] = bswap_32(u.val32[0]);
1615 u.val32[1] = bswap_32(u.val32[1]);
1616 u.val64 = bswap_64(u.val64);
1617 }
1618
1619 *array = u.val64;
1620 array++;
1621 }
1622
1623 if (type & PERF_SAMPLE_TIME) {
1624 *array = sample->time;
1625 array++;
1626 }
1627
1628 if (type & PERF_SAMPLE_ADDR) {
1629 *array = sample->addr;
1630 array++;
1631 }
1632
1633 if (type & PERF_SAMPLE_ID) {
1634 *array = sample->id;
1635 array++;
1636 }
1637
1638 if (type & PERF_SAMPLE_STREAM_ID) {
1639 *array = sample->stream_id;
1640 array++;
1641 }
1642
1643 if (type & PERF_SAMPLE_CPU) {
1644 u.val32[0] = sample->cpu;
1645 if (swapped) {
1646 /*
1647 * Inverse of what is done in perf_evsel__parse_sample
1648 */
1649 u.val32[0] = bswap_32(u.val32[0]);
1650 u.val64 = bswap_64(u.val64);
1651 }
1652 *array = u.val64;
1653 array++;
1654 }
1655
1656 if (type & PERF_SAMPLE_PERIOD) {
1657 *array = sample->period;
1658 array++;
1659 }
1660
1661 if (type & PERF_SAMPLE_READ) {
1662 if (read_format & PERF_FORMAT_GROUP)
1663 *array = sample->read.group.nr;
1664 else
1665 *array = sample->read.one.value;
1666 array++;
1667
1668 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1669 *array = sample->read.time_enabled;
1670 array++;
1671 }
1672
1673 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1674 *array = sample->read.time_running;
1675 array++;
1676 }
1677
1678 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1679 if (read_format & PERF_FORMAT_GROUP) {
1680 sz = sample->read.group.nr *
1681 sizeof(struct sample_read_value);
1682 memcpy(array, sample->read.group.values, sz);
1683 array = (void *)array + sz;
1684 } else {
1685 *array = sample->read.one.id;
1686 array++;
1687 }
1688 }
1689
1690 if (type & PERF_SAMPLE_CALLCHAIN) {
1691 sz = (sample->callchain->nr + 1) * sizeof(u64);
1692 memcpy(array, sample->callchain, sz);
1693 array = (void *)array + sz;
1694 }
1695
1696 if (type & PERF_SAMPLE_RAW) {
1697 u.val32[0] = sample->raw_size;
1698 if (WARN_ONCE(swapped,
1699 "Endianness of raw data not corrected!\n")) {
1700 /*
1701 * Inverse of what is done in perf_evsel__parse_sample
1702 */
1703 u.val32[0] = bswap_32(u.val32[0]);
1704 u.val32[1] = bswap_32(u.val32[1]);
1705 u.val64 = bswap_64(u.val64);
1706 }
1707 *array = u.val64;
1708 array = (void *)array + sizeof(u32);
1709
1710 memcpy(array, sample->raw_data, sample->raw_size);
1711 array = (void *)array + sample->raw_size;
1712 }
1713
1714 if (type & PERF_SAMPLE_BRANCH_STACK) {
1715 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1716 sz += sizeof(u64);
1717 memcpy(array, sample->branch_stack, sz);
1718 array = (void *)array + sz;
1719 }
1720
1721 if (type & PERF_SAMPLE_REGS_USER) {
1722 if (sample->user_regs.abi) {
1723 *array++ = sample->user_regs.abi;
1724 sz = hweight_long(sample_regs_user) * sizeof(u64);
1725 memcpy(array, sample->user_regs.regs, sz);
1726 array = (void *)array + sz;
1727 } else {
1728 *array++ = 0;
1729 }
1730 }
1731
1732 if (type & PERF_SAMPLE_STACK_USER) {
1733 sz = sample->user_stack.size;
1734 *array++ = sz;
1735 if (sz) {
1736 memcpy(array, sample->user_stack.data, sz);
1737 array = (void *)array + sz;
1738 *array++ = sz;
1739 }
1740 }
1741
1742 if (type & PERF_SAMPLE_WEIGHT) {
1743 *array = sample->weight;
1744 array++;
1745 }
1746
1747 if (type & PERF_SAMPLE_DATA_SRC) {
1748 *array = sample->data_src;
1749 array++;
1750 }
1751
1752 return 0;
1753 }
1754
1755 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1756 {
1757 return pevent_find_field(evsel->tp_format, name);
1758 }
1759
1760 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1761 const char *name)
1762 {
1763 struct format_field *field = perf_evsel__field(evsel, name);
1764 int offset;
1765
1766 if (!field)
1767 return NULL;
1768
1769 offset = field->offset;
1770
1771 if (field->flags & FIELD_IS_DYNAMIC) {
1772 offset = *(int *)(sample->raw_data + field->offset);
1773 offset &= 0xffff;
1774 }
1775
1776 return sample->raw_data + offset;
1777 }
1778
1779 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1780 const char *name)
1781 {
1782 struct format_field *field = perf_evsel__field(evsel, name);
1783 void *ptr;
1784 u64 value;
1785
1786 if (!field)
1787 return 0;
1788
1789 ptr = sample->raw_data + field->offset;
1790
1791 switch (field->size) {
1792 case 1:
1793 return *(u8 *)ptr;
1794 case 2:
1795 value = *(u16 *)ptr;
1796 break;
1797 case 4:
1798 value = *(u32 *)ptr;
1799 break;
1800 case 8:
1801 value = *(u64 *)ptr;
1802 break;
1803 default:
1804 return 0;
1805 }
1806
1807 if (!evsel->needs_swap)
1808 return value;
1809
1810 switch (field->size) {
1811 case 2:
1812 return bswap_16(value);
1813 case 4:
1814 return bswap_32(value);
1815 case 8:
1816 return bswap_64(value);
1817 default:
1818 return 0;
1819 }
1820
1821 return 0;
1822 }
1823
1824 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1825 {
1826 va_list args;
1827 int ret = 0;
1828
1829 if (!*first) {
1830 ret += fprintf(fp, ",");
1831 } else {
1832 ret += fprintf(fp, ":");
1833 *first = false;
1834 }
1835
1836 va_start(args, fmt);
1837 ret += vfprintf(fp, fmt, args);
1838 va_end(args);
1839 return ret;
1840 }
1841
1842 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1843 {
1844 if (value == 0)
1845 return 0;
1846
1847 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1848 }
1849
1850 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1851
1852 struct bit_names {
1853 int bit;
1854 const char *name;
1855 };
1856
1857 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1858 struct bit_names *bits, bool *first)
1859 {
1860 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1861 bool first_bit = true;
1862
1863 do {
1864 if (value & bits[i].bit) {
1865 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1866 first_bit = false;
1867 }
1868 } while (bits[++i].name != NULL);
1869
1870 return printed;
1871 }
1872
1873 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1874 {
1875 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1876 struct bit_names bits[] = {
1877 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1878 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1879 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1880 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1881 bit_name(IDENTIFIER),
1882 { .name = NULL, }
1883 };
1884 #undef bit_name
1885 return bits__fprintf(fp, "sample_type", value, bits, first);
1886 }
1887
1888 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1889 {
1890 #define bit_name(n) { PERF_FORMAT_##n, #n }
1891 struct bit_names bits[] = {
1892 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1893 bit_name(ID), bit_name(GROUP),
1894 { .name = NULL, }
1895 };
1896 #undef bit_name
1897 return bits__fprintf(fp, "read_format", value, bits, first);
1898 }
1899
1900 int perf_evsel__fprintf(struct perf_evsel *evsel,
1901 struct perf_attr_details *details, FILE *fp)
1902 {
1903 bool first = true;
1904 int printed = 0;
1905
1906 if (details->event_group) {
1907 struct perf_evsel *pos;
1908
1909 if (!perf_evsel__is_group_leader(evsel))
1910 return 0;
1911
1912 if (evsel->nr_members > 1)
1913 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1914
1915 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1916 for_each_group_member(pos, evsel)
1917 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1918
1919 if (evsel->nr_members > 1)
1920 printed += fprintf(fp, "}");
1921 goto out;
1922 }
1923
1924 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1925
1926 if (details->verbose || details->freq) {
1927 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1928 (u64)evsel->attr.sample_freq);
1929 }
1930
1931 if (details->verbose) {
1932 if_print(type);
1933 if_print(config);
1934 if_print(config1);
1935 if_print(config2);
1936 if_print(size);
1937 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1938 if (evsel->attr.read_format)
1939 printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1940 if_print(disabled);
1941 if_print(inherit);
1942 if_print(pinned);
1943 if_print(exclusive);
1944 if_print(exclude_user);
1945 if_print(exclude_kernel);
1946 if_print(exclude_hv);
1947 if_print(exclude_idle);
1948 if_print(mmap);
1949 if_print(mmap2);
1950 if_print(comm);
1951 if_print(freq);
1952 if_print(inherit_stat);
1953 if_print(enable_on_exec);
1954 if_print(task);
1955 if_print(watermark);
1956 if_print(precise_ip);
1957 if_print(mmap_data);
1958 if_print(sample_id_all);
1959 if_print(exclude_host);
1960 if_print(exclude_guest);
1961 if_print(__reserved_1);
1962 if_print(wakeup_events);
1963 if_print(bp_type);
1964 if_print(branch_sample_type);
1965 }
1966 out:
1967 fputc('\n', fp);
1968 return ++printed;
1969 }
1970
1971 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1972 char *msg, size_t msgsize)
1973 {
1974 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
1975 evsel->attr.type == PERF_TYPE_HARDWARE &&
1976 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1977 /*
1978 * If it's cycles then fall back to hrtimer based
1979 * cpu-clock-tick sw counter, which is always available even if
1980 * no PMU support.
1981 *
1982 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1983 * b0a873e).
1984 */
1985 scnprintf(msg, msgsize, "%s",
1986 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1987
1988 evsel->attr.type = PERF_TYPE_SOFTWARE;
1989 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
1990
1991 free(evsel->name);
1992 evsel->name = NULL;
1993 return true;
1994 }
1995
1996 return false;
1997 }
1998
1999 int perf_evsel__open_strerror(struct perf_evsel *evsel,
2000 struct perf_target *target,
2001 int err, char *msg, size_t size)
2002 {
2003 switch (err) {
2004 case EPERM:
2005 case EACCES:
2006 return scnprintf(msg, size,
2007 "You may not have permission to collect %sstats.\n"
2008 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2009 " -1 - Not paranoid at all\n"
2010 " 0 - Disallow raw tracepoint access for unpriv\n"
2011 " 1 - Disallow cpu events for unpriv\n"
2012 " 2 - Disallow kernel profiling for unpriv",
2013 target->system_wide ? "system-wide " : "");
2014 case ENOENT:
2015 return scnprintf(msg, size, "The %s event is not supported.",
2016 perf_evsel__name(evsel));
2017 case EMFILE:
2018 return scnprintf(msg, size, "%s",
2019 "Too many events are opened.\n"
2020 "Try again after reducing the number of events.");
2021 case ENODEV:
2022 if (target->cpu_list)
2023 return scnprintf(msg, size, "%s",
2024 "No such device - did you specify an out-of-range profile CPU?\n");
2025 break;
2026 case EOPNOTSUPP:
2027 if (evsel->attr.precise_ip)
2028 return scnprintf(msg, size, "%s",
2029 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2030 #if defined(__i386__) || defined(__x86_64__)
2031 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2032 return scnprintf(msg, size, "%s",
2033 "No hardware sampling interrupt available.\n"
2034 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2035 #endif
2036 break;
2037 default:
2038 break;
2039 }
2040
2041 return scnprintf(msg, size,
2042 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2043 "/bin/dmesg may provide additional information.\n"
2044 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2045 err, strerror(err), perf_evsel__name(evsel));
2046 }
ubuntu-zesty-kernel mirror