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