]> git.proxmox.com Git - mirror_ubuntu-focal-kernel.git/blob - tools/perf/util/evsel.c
projects / mirror_ubuntu-focal-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
libperf: Add threads to struct perf_evsel
[mirror_ubuntu-focal-kernel.git] / tools / perf / util / evsel.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 *
5 * Parts came from builtin-{top,stat,record}.c, see those files for further
6 * copyright notes.
7 */
8
9 #include <byteswap.h>
10 #include <errno.h>
11 #include <inttypes.h>
12 #include <linux/bitops.h>
13 #include <api/fs/fs.h>
14 #include <api/fs/tracing_path.h>
15 #include <traceevent/event-parse.h>
16 #include <linux/hw_breakpoint.h>
17 #include <linux/perf_event.h>
18 #include <linux/compiler.h>
19 #include <linux/err.h>
20 #include <linux/zalloc.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
24 #include <dirent.h>
25 #include <perf/evsel.h>
26 #include "asm/bug.h"
27 #include "callchain.h"
28 #include "cgroup.h"
29 #include "event.h"
30 #include "evsel.h"
31 #include "evlist.h"
32 #include "cpumap.h"
33 #include "thread_map.h"
34 #include "target.h"
35 #include "perf_regs.h"
36 #include "debug.h"
37 #include "trace-event.h"
38 #include "stat.h"
39 #include "string2.h"
40 #include "memswap.h"
41 #include "util/parse-branch-options.h"
42
43 #include <linux/ctype.h>
44
45 struct perf_missing_features perf_missing_features;
46
47 static clockid_t clockid;
48
49 static int perf_evsel__no_extra_init(struct evsel *evsel __maybe_unused)
50 {
51 return 0;
52 }
53
54 void __weak test_attr__ready(void) { }
55
56 static void perf_evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
57 {
58 }
59
60 static struct {
61 size_t size;
62 int (*init)(struct evsel *evsel);
63 void (*fini)(struct evsel *evsel);
64 } perf_evsel__object = {
65 .size = sizeof(struct evsel),
66 .init = perf_evsel__no_extra_init,
67 .fini = perf_evsel__no_extra_fini,
68 };
69
70 int perf_evsel__object_config(size_t object_size,
71 int (*init)(struct evsel *evsel),
72 void (*fini)(struct evsel *evsel))
73 {
74
75 if (object_size == 0)
76 goto set_methods;
77
78 if (perf_evsel__object.size > object_size)
79 return -EINVAL;
80
81 perf_evsel__object.size = object_size;
82
83 set_methods:
84 if (init != NULL)
85 perf_evsel__object.init = init;
86
87 if (fini != NULL)
88 perf_evsel__object.fini = fini;
89
90 return 0;
91 }
92
93 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
94
95 int __perf_evsel__sample_size(u64 sample_type)
96 {
97 u64 mask = sample_type & PERF_SAMPLE_MASK;
98 int size = 0;
99 int i;
100
101 for (i = 0; i < 64; i++) {
102 if (mask & (1ULL << i))
103 size++;
104 }
105
106 size *= sizeof(u64);
107
108 return size;
109 }
110
111 /**
112 * __perf_evsel__calc_id_pos - calculate id_pos.
113 * @sample_type: sample type
114 *
115 * This function returns the position of the event id (PERF_SAMPLE_ID or
116 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
117 * sample_event.
118 */
119 static int __perf_evsel__calc_id_pos(u64 sample_type)
120 {
121 int idx = 0;
122
123 if (sample_type & PERF_SAMPLE_IDENTIFIER)
124 return 0;
125
126 if (!(sample_type & PERF_SAMPLE_ID))
127 return -1;
128
129 if (sample_type & PERF_SAMPLE_IP)
130 idx += 1;
131
132 if (sample_type & PERF_SAMPLE_TID)
133 idx += 1;
134
135 if (sample_type & PERF_SAMPLE_TIME)
136 idx += 1;
137
138 if (sample_type & PERF_SAMPLE_ADDR)
139 idx += 1;
140
141 return idx;
142 }
143
144 /**
145 * __perf_evsel__calc_is_pos - calculate is_pos.
146 * @sample_type: sample type
147 *
148 * This function returns the position (counting backwards) of the event id
149 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
150 * sample_id_all is used there is an id sample appended to non-sample events.
151 */
152 static int __perf_evsel__calc_is_pos(u64 sample_type)
153 {
154 int idx = 1;
155
156 if (sample_type & PERF_SAMPLE_IDENTIFIER)
157 return 1;
158
159 if (!(sample_type & PERF_SAMPLE_ID))
160 return -1;
161
162 if (sample_type & PERF_SAMPLE_CPU)
163 idx += 1;
164
165 if (sample_type & PERF_SAMPLE_STREAM_ID)
166 idx += 1;
167
168 return idx;
169 }
170
171 void perf_evsel__calc_id_pos(struct evsel *evsel)
172 {
173 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
174 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
175 }
176
177 void __perf_evsel__set_sample_bit(struct evsel *evsel,
178 enum perf_event_sample_format bit)
179 {
180 if (!(evsel->core.attr.sample_type & bit)) {
181 evsel->core.attr.sample_type |= bit;
182 evsel->sample_size += sizeof(u64);
183 perf_evsel__calc_id_pos(evsel);
184 }
185 }
186
187 void __perf_evsel__reset_sample_bit(struct evsel *evsel,
188 enum perf_event_sample_format bit)
189 {
190 if (evsel->core.attr.sample_type & bit) {
191 evsel->core.attr.sample_type &= ~bit;
192 evsel->sample_size -= sizeof(u64);
193 perf_evsel__calc_id_pos(evsel);
194 }
195 }
196
197 void perf_evsel__set_sample_id(struct evsel *evsel,
198 bool can_sample_identifier)
199 {
200 if (can_sample_identifier) {
201 perf_evsel__reset_sample_bit(evsel, ID);
202 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
203 } else {
204 perf_evsel__set_sample_bit(evsel, ID);
205 }
206 evsel->core.attr.read_format |= PERF_FORMAT_ID;
207 }
208
209 /**
210 * perf_evsel__is_function_event - Return whether given evsel is a function
211 * trace event
212 *
213 * @evsel - evsel selector to be tested
214 *
215 * Return %true if event is function trace event
216 */
217 bool perf_evsel__is_function_event(struct evsel *evsel)
218 {
219 #define FUNCTION_EVENT "ftrace:function"
220
221 return evsel->name &&
222 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
223
224 #undef FUNCTION_EVENT
225 }
226
227 void evsel__init(struct evsel *evsel,
228 struct perf_event_attr *attr, int idx)
229 {
230 perf_evsel__init(&evsel->core, attr);
231 evsel->idx = idx;
232 evsel->tracking = !idx;
233 evsel->leader = evsel;
234 evsel->unit = "";
235 evsel->scale = 1.0;
236 evsel->max_events = ULONG_MAX;
237 evsel->evlist = NULL;
238 evsel->bpf_obj = NULL;
239 evsel->bpf_fd = -1;
240 INIT_LIST_HEAD(&evsel->config_terms);
241 perf_evsel__object.init(evsel);
242 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
243 perf_evsel__calc_id_pos(evsel);
244 evsel->cmdline_group_boundary = false;
245 evsel->metric_expr = NULL;
246 evsel->metric_name = NULL;
247 evsel->metric_events = NULL;
248 evsel->collect_stat = false;
249 evsel->pmu_name = NULL;
250 }
251
252 struct evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
253 {
254 struct evsel *evsel = zalloc(perf_evsel__object.size);
255
256 if (!evsel)
257 return NULL;
258 evsel__init(evsel, attr, idx);
259
260 if (perf_evsel__is_bpf_output(evsel)) {
261 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
262 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
263 evsel->core.attr.sample_period = 1;
264 }
265
266 if (perf_evsel__is_clock(evsel)) {
267 /*
268 * The evsel->unit points to static alias->unit
269 * so it's ok to use static string in here.
270 */
271 static const char *unit = "msec";
272
273 evsel->unit = unit;
274 evsel->scale = 1e-6;
275 }
276
277 return evsel;
278 }
279
280 static bool perf_event_can_profile_kernel(void)
281 {
282 return geteuid() == 0 || perf_event_paranoid() == -1;
283 }
284
285 struct evsel *perf_evsel__new_cycles(bool precise)
286 {
287 struct perf_event_attr attr = {
288 .type = PERF_TYPE_HARDWARE,
289 .config = PERF_COUNT_HW_CPU_CYCLES,
290 .exclude_kernel = !perf_event_can_profile_kernel(),
291 };
292 struct evsel *evsel;
293
294 event_attr_init(&attr);
295
296 if (!precise)
297 goto new_event;
298
299 /*
300 * Now let the usual logic to set up the perf_event_attr defaults
301 * to kick in when we return and before perf_evsel__open() is called.
302 */
303 new_event:
304 evsel = evsel__new(&attr);
305 if (evsel == NULL)
306 goto out;
307
308 evsel->precise_max = true;
309
310 /* use asprintf() because free(evsel) assumes name is allocated */
311 if (asprintf(&evsel->name, "cycles%s%s%.*s",
312 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
313 attr.exclude_kernel ? "u" : "",
314 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
315 goto error_free;
316 out:
317 return evsel;
318 error_free:
319 evsel__delete(evsel);
320 evsel = NULL;
321 goto out;
322 }
323
324 /*
325 * Returns pointer with encoded error via <linux/err.h> interface.
326 */
327 struct evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
328 {
329 struct evsel *evsel = zalloc(perf_evsel__object.size);
330 int err = -ENOMEM;
331
332 if (evsel == NULL) {
333 goto out_err;
334 } else {
335 struct perf_event_attr attr = {
336 .type = PERF_TYPE_TRACEPOINT,
337 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
338 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
339 };
340
341 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
342 goto out_free;
343
344 evsel->tp_format = trace_event__tp_format(sys, name);
345 if (IS_ERR(evsel->tp_format)) {
346 err = PTR_ERR(evsel->tp_format);
347 goto out_free;
348 }
349
350 event_attr_init(&attr);
351 attr.config = evsel->tp_format->id;
352 attr.sample_period = 1;
353 evsel__init(evsel, &attr, idx);
354 }
355
356 return evsel;
357
358 out_free:
359 zfree(&evsel->name);
360 free(evsel);
361 out_err:
362 return ERR_PTR(err);
363 }
364
365 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
366 "cycles",
367 "instructions",
368 "cache-references",
369 "cache-misses",
370 "branches",
371 "branch-misses",
372 "bus-cycles",
373 "stalled-cycles-frontend",
374 "stalled-cycles-backend",
375 "ref-cycles",
376 };
377
378 static const char *__perf_evsel__hw_name(u64 config)
379 {
380 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
381 return perf_evsel__hw_names[config];
382
383 return "unknown-hardware";
384 }
385
386 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
387 {
388 int colon = 0, r = 0;
389 struct perf_event_attr *attr = &evsel->core.attr;
390 bool exclude_guest_default = false;
391
392 #define MOD_PRINT(context, mod) do { \
393 if (!attr->exclude_##context) { \
394 if (!colon) colon = ++r; \
395 r += scnprintf(bf + r, size - r, "%c", mod); \
396 } } while(0)
397
398 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
399 MOD_PRINT(kernel, 'k');
400 MOD_PRINT(user, 'u');
401 MOD_PRINT(hv, 'h');
402 exclude_guest_default = true;
403 }
404
405 if (attr->precise_ip) {
406 if (!colon)
407 colon = ++r;
408 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
409 exclude_guest_default = true;
410 }
411
412 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
413 MOD_PRINT(host, 'H');
414 MOD_PRINT(guest, 'G');
415 }
416 #undef MOD_PRINT
417 if (colon)
418 bf[colon - 1] = ':';
419 return r;
420 }
421
422 static int perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
423 {
424 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config));
425 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
426 }
427
428 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
429 "cpu-clock",
430 "task-clock",
431 "page-faults",
432 "context-switches",
433 "cpu-migrations",
434 "minor-faults",
435 "major-faults",
436 "alignment-faults",
437 "emulation-faults",
438 "dummy",
439 };
440
441 static const char *__perf_evsel__sw_name(u64 config)
442 {
443 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
444 return perf_evsel__sw_names[config];
445 return "unknown-software";
446 }
447
448 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
449 {
450 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config));
451 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
452 }
453
454 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
455 {
456 int r;
457
458 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
459
460 if (type & HW_BREAKPOINT_R)
461 r += scnprintf(bf + r, size - r, "r");
462
463 if (type & HW_BREAKPOINT_W)
464 r += scnprintf(bf + r, size - r, "w");
465
466 if (type & HW_BREAKPOINT_X)
467 r += scnprintf(bf + r, size - r, "x");
468
469 return r;
470 }
471
472 static int perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
473 {
474 struct perf_event_attr *attr = &evsel->core.attr;
475 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
476 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
477 }
478
479 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
480 [PERF_EVSEL__MAX_ALIASES] = {
481 { "L1-dcache", "l1-d", "l1d", "L1-data", },
482 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
483 { "LLC", "L2", },
484 { "dTLB", "d-tlb", "Data-TLB", },
485 { "iTLB", "i-tlb", "Instruction-TLB", },
486 { "branch", "branches", "bpu", "btb", "bpc", },
487 { "node", },
488 };
489
490 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
491 [PERF_EVSEL__MAX_ALIASES] = {
492 { "load", "loads", "read", },
493 { "store", "stores", "write", },
494 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
495 };
496
497 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
498 [PERF_EVSEL__MAX_ALIASES] = {
499 { "refs", "Reference", "ops", "access", },
500 { "misses", "miss", },
501 };
502
503 #define C(x) PERF_COUNT_HW_CACHE_##x
504 #define CACHE_READ (1 << C(OP_READ))
505 #define CACHE_WRITE (1 << C(OP_WRITE))
506 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
507 #define COP(x) (1 << x)
508
509 /*
510 * cache operartion stat
511 * L1I : Read and prefetch only
512 * ITLB and BPU : Read-only
513 */
514 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
515 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
516 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
517 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
518 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
519 [C(ITLB)] = (CACHE_READ),
520 [C(BPU)] = (CACHE_READ),
521 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
522 };
523
524 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
525 {
526 if (perf_evsel__hw_cache_stat[type] & COP(op))
527 return true; /* valid */
528 else
529 return false; /* invalid */
530 }
531
532 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
533 char *bf, size_t size)
534 {
535 if (result) {
536 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
537 perf_evsel__hw_cache_op[op][0],
538 perf_evsel__hw_cache_result[result][0]);
539 }
540
541 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
542 perf_evsel__hw_cache_op[op][1]);
543 }
544
545 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
546 {
547 u8 op, result, type = (config >> 0) & 0xff;
548 const char *err = "unknown-ext-hardware-cache-type";
549
550 if (type >= PERF_COUNT_HW_CACHE_MAX)
551 goto out_err;
552
553 op = (config >> 8) & 0xff;
554 err = "unknown-ext-hardware-cache-op";
555 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
556 goto out_err;
557
558 result = (config >> 16) & 0xff;
559 err = "unknown-ext-hardware-cache-result";
560 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
561 goto out_err;
562
563 err = "invalid-cache";
564 if (!perf_evsel__is_cache_op_valid(type, op))
565 goto out_err;
566
567 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
568 out_err:
569 return scnprintf(bf, size, "%s", err);
570 }
571
572 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
573 {
574 int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size);
575 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
576 }
577
578 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
579 {
580 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
581 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
582 }
583
584 static int perf_evsel__tool_name(char *bf, size_t size)
585 {
586 int ret = scnprintf(bf, size, "duration_time");
587 return ret;
588 }
589
590 const char *perf_evsel__name(struct evsel *evsel)
591 {
592 char bf[128];
593
594 if (!evsel)
595 goto out_unknown;
596
597 if (evsel->name)
598 return evsel->name;
599
600 switch (evsel->core.attr.type) {
601 case PERF_TYPE_RAW:
602 perf_evsel__raw_name(evsel, bf, sizeof(bf));
603 break;
604
605 case PERF_TYPE_HARDWARE:
606 perf_evsel__hw_name(evsel, bf, sizeof(bf));
607 break;
608
609 case PERF_TYPE_HW_CACHE:
610 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
611 break;
612
613 case PERF_TYPE_SOFTWARE:
614 if (evsel->tool_event)
615 perf_evsel__tool_name(bf, sizeof(bf));
616 else
617 perf_evsel__sw_name(evsel, bf, sizeof(bf));
618 break;
619
620 case PERF_TYPE_TRACEPOINT:
621 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
622 break;
623
624 case PERF_TYPE_BREAKPOINT:
625 perf_evsel__bp_name(evsel, bf, sizeof(bf));
626 break;
627
628 default:
629 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
630 evsel->core.attr.type);
631 break;
632 }
633
634 evsel->name = strdup(bf);
635
636 if (evsel->name)
637 return evsel->name;
638 out_unknown:
639 return "unknown";
640 }
641
642 const char *perf_evsel__group_name(struct evsel *evsel)
643 {
644 return evsel->group_name ?: "anon group";
645 }
646
647 /*
648 * Returns the group details for the specified leader,
649 * with following rules.
650 *
651 * For record -e '{cycles,instructions}'
652 * 'anon group { cycles:u, instructions:u }'
653 *
654 * For record -e 'cycles,instructions' and report --group
655 * 'cycles:u, instructions:u'
656 */
657 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
658 {
659 int ret = 0;
660 struct evsel *pos;
661 const char *group_name = perf_evsel__group_name(evsel);
662
663 if (!evsel->forced_leader)
664 ret = scnprintf(buf, size, "%s { ", group_name);
665
666 ret += scnprintf(buf + ret, size - ret, "%s",
667 perf_evsel__name(evsel));
668
669 for_each_group_member(pos, evsel)
670 ret += scnprintf(buf + ret, size - ret, ", %s",
671 perf_evsel__name(pos));
672
673 if (!evsel->forced_leader)
674 ret += scnprintf(buf + ret, size - ret, " }");
675
676 return ret;
677 }
678
679 static void __perf_evsel__config_callchain(struct evsel *evsel,
680 struct record_opts *opts,
681 struct callchain_param *param)
682 {
683 bool function = perf_evsel__is_function_event(evsel);
684 struct perf_event_attr *attr = &evsel->core.attr;
685
686 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
687
688 attr->sample_max_stack = param->max_stack;
689
690 if (opts->kernel_callchains)
691 attr->exclude_callchain_user = 1;
692 if (opts->user_callchains)
693 attr->exclude_callchain_kernel = 1;
694 if (param->record_mode == CALLCHAIN_LBR) {
695 if (!opts->branch_stack) {
696 if (attr->exclude_user) {
697 pr_warning("LBR callstack option is only available "
698 "to get user callchain information. "
699 "Falling back to framepointers.\n");
700 } else {
701 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
702 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
703 PERF_SAMPLE_BRANCH_CALL_STACK |
704 PERF_SAMPLE_BRANCH_NO_CYCLES |
705 PERF_SAMPLE_BRANCH_NO_FLAGS;
706 }
707 } else
708 pr_warning("Cannot use LBR callstack with branch stack. "
709 "Falling back to framepointers.\n");
710 }
711
712 if (param->record_mode == CALLCHAIN_DWARF) {
713 if (!function) {
714 perf_evsel__set_sample_bit(evsel, REGS_USER);
715 perf_evsel__set_sample_bit(evsel, STACK_USER);
716 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
717 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
718 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
719 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
720 "so the minimal registers set (IP, SP) is explicitly forced.\n");
721 } else {
722 attr->sample_regs_user |= PERF_REGS_MASK;
723 }
724 attr->sample_stack_user = param->dump_size;
725 attr->exclude_callchain_user = 1;
726 } else {
727 pr_info("Cannot use DWARF unwind for function trace event,"
728 " falling back to framepointers.\n");
729 }
730 }
731
732 if (function) {
733 pr_info("Disabling user space callchains for function trace event.\n");
734 attr->exclude_callchain_user = 1;
735 }
736 }
737
738 void perf_evsel__config_callchain(struct evsel *evsel,
739 struct record_opts *opts,
740 struct callchain_param *param)
741 {
742 if (param->enabled)
743 return __perf_evsel__config_callchain(evsel, opts, param);
744 }
745
746 static void
747 perf_evsel__reset_callgraph(struct evsel *evsel,
748 struct callchain_param *param)
749 {
750 struct perf_event_attr *attr = &evsel->core.attr;
751
752 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
753 if (param->record_mode == CALLCHAIN_LBR) {
754 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
755 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
756 PERF_SAMPLE_BRANCH_CALL_STACK);
757 }
758 if (param->record_mode == CALLCHAIN_DWARF) {
759 perf_evsel__reset_sample_bit(evsel, REGS_USER);
760 perf_evsel__reset_sample_bit(evsel, STACK_USER);
761 }
762 }
763
764 static void apply_config_terms(struct evsel *evsel,
765 struct record_opts *opts, bool track)
766 {
767 struct perf_evsel_config_term *term;
768 struct list_head *config_terms = &evsel->config_terms;
769 struct perf_event_attr *attr = &evsel->core.attr;
770 /* callgraph default */
771 struct callchain_param param = {
772 .record_mode = callchain_param.record_mode,
773 };
774 u32 dump_size = 0;
775 int max_stack = 0;
776 const char *callgraph_buf = NULL;
777
778 list_for_each_entry(term, config_terms, list) {
779 switch (term->type) {
780 case PERF_EVSEL__CONFIG_TERM_PERIOD:
781 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
782 attr->sample_period = term->val.period;
783 attr->freq = 0;
784 perf_evsel__reset_sample_bit(evsel, PERIOD);
785 }
786 break;
787 case PERF_EVSEL__CONFIG_TERM_FREQ:
788 if (!(term->weak && opts->user_freq != UINT_MAX)) {
789 attr->sample_freq = term->val.freq;
790 attr->freq = 1;
791 perf_evsel__set_sample_bit(evsel, PERIOD);
792 }
793 break;
794 case PERF_EVSEL__CONFIG_TERM_TIME:
795 if (term->val.time)
796 perf_evsel__set_sample_bit(evsel, TIME);
797 else
798 perf_evsel__reset_sample_bit(evsel, TIME);
799 break;
800 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
801 callgraph_buf = term->val.callgraph;
802 break;
803 case PERF_EVSEL__CONFIG_TERM_BRANCH:
804 if (term->val.branch && strcmp(term->val.branch, "no")) {
805 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
806 parse_branch_str(term->val.branch,
807 &attr->branch_sample_type);
808 } else
809 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
810 break;
811 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
812 dump_size = term->val.stack_user;
813 break;
814 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
815 max_stack = term->val.max_stack;
816 break;
817 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
818 evsel->max_events = term->val.max_events;
819 break;
820 case PERF_EVSEL__CONFIG_TERM_INHERIT:
821 /*
822 * attr->inherit should has already been set by
823 * perf_evsel__config. If user explicitly set
824 * inherit using config terms, override global
825 * opt->no_inherit setting.
826 */
827 attr->inherit = term->val.inherit ? 1 : 0;
828 break;
829 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
830 attr->write_backward = term->val.overwrite ? 1 : 0;
831 break;
832 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
833 break;
834 case PERF_EVSEL__CONFIG_TERM_PERCORE:
835 break;
836 default:
837 break;
838 }
839 }
840
841 /* User explicitly set per-event callgraph, clear the old setting and reset. */
842 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
843 bool sample_address = false;
844
845 if (max_stack) {
846 param.max_stack = max_stack;
847 if (callgraph_buf == NULL)
848 callgraph_buf = "fp";
849 }
850
851 /* parse callgraph parameters */
852 if (callgraph_buf != NULL) {
853 if (!strcmp(callgraph_buf, "no")) {
854 param.enabled = false;
855 param.record_mode = CALLCHAIN_NONE;
856 } else {
857 param.enabled = true;
858 if (parse_callchain_record(callgraph_buf, &param)) {
859 pr_err("per-event callgraph setting for %s failed. "
860 "Apply callgraph global setting for it\n",
861 evsel->name);
862 return;
863 }
864 if (param.record_mode == CALLCHAIN_DWARF)
865 sample_address = true;
866 }
867 }
868 if (dump_size > 0) {
869 dump_size = round_up(dump_size, sizeof(u64));
870 param.dump_size = dump_size;
871 }
872
873 /* If global callgraph set, clear it */
874 if (callchain_param.enabled)
875 perf_evsel__reset_callgraph(evsel, &callchain_param);
876
877 /* set perf-event callgraph */
878 if (param.enabled) {
879 if (sample_address) {
880 perf_evsel__set_sample_bit(evsel, ADDR);
881 perf_evsel__set_sample_bit(evsel, DATA_SRC);
882 evsel->core.attr.mmap_data = track;
883 }
884 perf_evsel__config_callchain(evsel, opts, &param);
885 }
886 }
887 }
888
889 static bool is_dummy_event(struct evsel *evsel)
890 {
891 return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
892 (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
893 }
894
895 /*
896 * The enable_on_exec/disabled value strategy:
897 *
898 * 1) For any type of traced program:
899 * - all independent events and group leaders are disabled
900 * - all group members are enabled
901 *
902 * Group members are ruled by group leaders. They need to
903 * be enabled, because the group scheduling relies on that.
904 *
905 * 2) For traced programs executed by perf:
906 * - all independent events and group leaders have
907 * enable_on_exec set
908 * - we don't specifically enable or disable any event during
909 * the record command
910 *
911 * Independent events and group leaders are initially disabled
912 * and get enabled by exec. Group members are ruled by group
913 * leaders as stated in 1).
914 *
915 * 3) For traced programs attached by perf (pid/tid):
916 * - we specifically enable or disable all events during
917 * the record command
918 *
919 * When attaching events to already running traced we
920 * enable/disable events specifically, as there's no
921 * initial traced exec call.
922 */
923 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
924 struct callchain_param *callchain)
925 {
926 struct evsel *leader = evsel->leader;
927 struct perf_event_attr *attr = &evsel->core.attr;
928 int track = evsel->tracking;
929 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
930
931 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
932 attr->inherit = !opts->no_inherit;
933 attr->write_backward = opts->overwrite ? 1 : 0;
934
935 perf_evsel__set_sample_bit(evsel, IP);
936 perf_evsel__set_sample_bit(evsel, TID);
937
938 if (evsel->sample_read) {
939 perf_evsel__set_sample_bit(evsel, READ);
940
941 /*
942 * We need ID even in case of single event, because
943 * PERF_SAMPLE_READ process ID specific data.
944 */
945 perf_evsel__set_sample_id(evsel, false);
946
947 /*
948 * Apply group format only if we belong to group
949 * with more than one members.
950 */
951 if (leader->nr_members > 1) {
952 attr->read_format |= PERF_FORMAT_GROUP;
953 attr->inherit = 0;
954 }
955 }
956
957 /*
958 * We default some events to have a default interval. But keep
959 * it a weak assumption overridable by the user.
960 */
961 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
962 opts->user_interval != ULLONG_MAX)) {
963 if (opts->freq) {
964 perf_evsel__set_sample_bit(evsel, PERIOD);
965 attr->freq = 1;
966 attr->sample_freq = opts->freq;
967 } else {
968 attr->sample_period = opts->default_interval;
969 }
970 }
971
972 /*
973 * Disable sampling for all group members other
974 * than leader in case leader 'leads' the sampling.
975 */
976 if ((leader != evsel) && leader->sample_read) {
977 attr->freq = 0;
978 attr->sample_freq = 0;
979 attr->sample_period = 0;
980 attr->write_backward = 0;
981
982 /*
983 * We don't get sample for slave events, we make them
984 * when delivering group leader sample. Set the slave
985 * event to follow the master sample_type to ease up
986 * report.
987 */
988 attr->sample_type = leader->core.attr.sample_type;
989 }
990
991 if (opts->no_samples)
992 attr->sample_freq = 0;
993
994 if (opts->inherit_stat) {
995 evsel->core.attr.read_format |=
996 PERF_FORMAT_TOTAL_TIME_ENABLED |
997 PERF_FORMAT_TOTAL_TIME_RUNNING |
998 PERF_FORMAT_ID;
999 attr->inherit_stat = 1;
1000 }
1001
1002 if (opts->sample_address) {
1003 perf_evsel__set_sample_bit(evsel, ADDR);
1004 attr->mmap_data = track;
1005 }
1006
1007 /*
1008 * We don't allow user space callchains for function trace
1009 * event, due to issues with page faults while tracing page
1010 * fault handler and its overall trickiness nature.
1011 */
1012 if (perf_evsel__is_function_event(evsel))
1013 evsel->core.attr.exclude_callchain_user = 1;
1014
1015 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1016 perf_evsel__config_callchain(evsel, opts, callchain);
1017
1018 if (opts->sample_intr_regs) {
1019 attr->sample_regs_intr = opts->sample_intr_regs;
1020 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1021 }
1022
1023 if (opts->sample_user_regs) {
1024 attr->sample_regs_user |= opts->sample_user_regs;
1025 perf_evsel__set_sample_bit(evsel, REGS_USER);
1026 }
1027
1028 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1029 perf_evsel__set_sample_bit(evsel, CPU);
1030
1031 /*
1032 * When the user explicitly disabled time don't force it here.
1033 */
1034 if (opts->sample_time &&
1035 (!perf_missing_features.sample_id_all &&
1036 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1037 opts->sample_time_set)))
1038 perf_evsel__set_sample_bit(evsel, TIME);
1039
1040 if (opts->raw_samples && !evsel->no_aux_samples) {
1041 perf_evsel__set_sample_bit(evsel, TIME);
1042 perf_evsel__set_sample_bit(evsel, RAW);
1043 perf_evsel__set_sample_bit(evsel, CPU);
1044 }
1045
1046 if (opts->sample_address)
1047 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1048
1049 if (opts->sample_phys_addr)
1050 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1051
1052 if (opts->no_buffering) {
1053 attr->watermark = 0;
1054 attr->wakeup_events = 1;
1055 }
1056 if (opts->branch_stack && !evsel->no_aux_samples) {
1057 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1058 attr->branch_sample_type = opts->branch_stack;
1059 }
1060
1061 if (opts->sample_weight)
1062 perf_evsel__set_sample_bit(evsel, WEIGHT);
1063
1064 attr->task = track;
1065 attr->mmap = track;
1066 attr->mmap2 = track && !perf_missing_features.mmap2;
1067 attr->comm = track;
1068 attr->ksymbol = track && !perf_missing_features.ksymbol;
1069 attr->bpf_event = track && !opts->no_bpf_event &&
1070 !perf_missing_features.bpf_event;
1071
1072 if (opts->record_namespaces)
1073 attr->namespaces = track;
1074
1075 if (opts->record_switch_events)
1076 attr->context_switch = track;
1077
1078 if (opts->sample_transaction)
1079 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1080
1081 if (opts->running_time) {
1082 evsel->core.attr.read_format |=
1083 PERF_FORMAT_TOTAL_TIME_ENABLED |
1084 PERF_FORMAT_TOTAL_TIME_RUNNING;
1085 }
1086
1087 /*
1088 * XXX see the function comment above
1089 *
1090 * Disabling only independent events or group leaders,
1091 * keeping group members enabled.
1092 */
1093 if (perf_evsel__is_group_leader(evsel))
1094 attr->disabled = 1;
1095
1096 /*
1097 * Setting enable_on_exec for independent events and
1098 * group leaders for traced executed by perf.
1099 */
1100 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1101 !opts->initial_delay)
1102 attr->enable_on_exec = 1;
1103
1104 if (evsel->immediate) {
1105 attr->disabled = 0;
1106 attr->enable_on_exec = 0;
1107 }
1108
1109 clockid = opts->clockid;
1110 if (opts->use_clockid) {
1111 attr->use_clockid = 1;
1112 attr->clockid = opts->clockid;
1113 }
1114
1115 if (evsel->precise_max)
1116 attr->precise_ip = 3;
1117
1118 if (opts->all_user) {
1119 attr->exclude_kernel = 1;
1120 attr->exclude_user = 0;
1121 }
1122
1123 if (opts->all_kernel) {
1124 attr->exclude_kernel = 0;
1125 attr->exclude_user = 1;
1126 }
1127
1128 if (evsel->core.own_cpus || evsel->unit)
1129 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1130
1131 /*
1132 * Apply event specific term settings,
1133 * it overloads any global configuration.
1134 */
1135 apply_config_terms(evsel, opts, track);
1136
1137 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1138
1139 /* The --period option takes the precedence. */
1140 if (opts->period_set) {
1141 if (opts->period)
1142 perf_evsel__set_sample_bit(evsel, PERIOD);
1143 else
1144 perf_evsel__reset_sample_bit(evsel, PERIOD);
1145 }
1146
1147 /*
1148 * For initial_delay, a dummy event is added implicitly.
1149 * The software event will trigger -EOPNOTSUPP error out,
1150 * if BRANCH_STACK bit is set.
1151 */
1152 if (opts->initial_delay && is_dummy_event(evsel))
1153 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1154 }
1155
1156 static int perf_evsel__alloc_fd(struct evsel *evsel, int ncpus, int nthreads)
1157 {
1158 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1159
1160 if (evsel->fd) {
1161 int cpu, thread;
1162 for (cpu = 0; cpu < ncpus; cpu++) {
1163 for (thread = 0; thread < nthreads; thread++) {
1164 FD(evsel, cpu, thread) = -1;
1165 }
1166 }
1167 }
1168
1169 return evsel->fd != NULL ? 0 : -ENOMEM;
1170 }
1171
1172 static int perf_evsel__run_ioctl(struct evsel *evsel,
1173 int ioc, void *arg)
1174 {
1175 int cpu, thread;
1176
1177 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1178 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1179 int fd = FD(evsel, cpu, thread),
1180 err = ioctl(fd, ioc, arg);
1181
1182 if (err)
1183 return err;
1184 }
1185 }
1186
1187 return 0;
1188 }
1189
1190 int evsel__apply_filter(struct evsel *evsel, const char *filter)
1191 {
1192 return perf_evsel__run_ioctl(evsel,
1193 PERF_EVENT_IOC_SET_FILTER,
1194 (void *)filter);
1195 }
1196
1197 int perf_evsel__set_filter(struct evsel *evsel, const char *filter)
1198 {
1199 char *new_filter = strdup(filter);
1200
1201 if (new_filter != NULL) {
1202 free(evsel->filter);
1203 evsel->filter = new_filter;
1204 return 0;
1205 }
1206
1207 return -1;
1208 }
1209
1210 static int perf_evsel__append_filter(struct evsel *evsel,
1211 const char *fmt, const char *filter)
1212 {
1213 char *new_filter;
1214
1215 if (evsel->filter == NULL)
1216 return perf_evsel__set_filter(evsel, filter);
1217
1218 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1219 free(evsel->filter);
1220 evsel->filter = new_filter;
1221 return 0;
1222 }
1223
1224 return -1;
1225 }
1226
1227 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1228 {
1229 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1230 }
1231
1232 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1233 {
1234 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1235 }
1236
1237 int evsel__enable(struct evsel *evsel)
1238 {
1239 int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1240
1241 if (!err)
1242 evsel->disabled = false;
1243
1244 return err;
1245 }
1246
1247 int evsel__disable(struct evsel *evsel)
1248 {
1249 int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1250 /*
1251 * We mark it disabled here so that tools that disable a event can
1252 * ignore events after they disable it. I.e. the ring buffer may have
1253 * already a few more events queued up before the kernel got the stop
1254 * request.
1255 */
1256 if (!err)
1257 evsel->disabled = true;
1258
1259 return err;
1260 }
1261
1262 int perf_evsel__alloc_id(struct evsel *evsel, int ncpus, int nthreads)
1263 {
1264 if (ncpus == 0 || nthreads == 0)
1265 return 0;
1266
1267 if (evsel->system_wide)
1268 nthreads = 1;
1269
1270 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1271 if (evsel->sample_id == NULL)
1272 return -ENOMEM;
1273
1274 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1275 if (evsel->id == NULL) {
1276 xyarray__delete(evsel->sample_id);
1277 evsel->sample_id = NULL;
1278 return -ENOMEM;
1279 }
1280
1281 return 0;
1282 }
1283
1284 static void perf_evsel__free_fd(struct evsel *evsel)
1285 {
1286 xyarray__delete(evsel->fd);
1287 evsel->fd = NULL;
1288 }
1289
1290 static void perf_evsel__free_id(struct evsel *evsel)
1291 {
1292 xyarray__delete(evsel->sample_id);
1293 evsel->sample_id = NULL;
1294 zfree(&evsel->id);
1295 evsel->ids = 0;
1296 }
1297
1298 static void perf_evsel__free_config_terms(struct evsel *evsel)
1299 {
1300 struct perf_evsel_config_term *term, *h;
1301
1302 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1303 list_del_init(&term->list);
1304 free(term);
1305 }
1306 }
1307
1308 void perf_evsel__close_fd(struct evsel *evsel)
1309 {
1310 int cpu, thread;
1311
1312 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1313 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1314 close(FD(evsel, cpu, thread));
1315 FD(evsel, cpu, thread) = -1;
1316 }
1317 }
1318
1319 void perf_evsel__exit(struct evsel *evsel)
1320 {
1321 assert(list_empty(&evsel->core.node));
1322 assert(evsel->evlist == NULL);
1323 perf_evsel__free_counts(evsel);
1324 perf_evsel__free_fd(evsel);
1325 perf_evsel__free_id(evsel);
1326 perf_evsel__free_config_terms(evsel);
1327 cgroup__put(evsel->cgrp);
1328 perf_cpu_map__put(evsel->core.cpus);
1329 perf_cpu_map__put(evsel->core.own_cpus);
1330 perf_thread_map__put(evsel->core.threads);
1331 zfree(&evsel->group_name);
1332 zfree(&evsel->name);
1333 perf_evsel__object.fini(evsel);
1334 }
1335
1336 void evsel__delete(struct evsel *evsel)
1337 {
1338 perf_evsel__exit(evsel);
1339 free(evsel);
1340 }
1341
1342 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1343 struct perf_counts_values *count)
1344 {
1345 struct perf_counts_values tmp;
1346
1347 if (!evsel->prev_raw_counts)
1348 return;
1349
1350 if (cpu == -1) {
1351 tmp = evsel->prev_raw_counts->aggr;
1352 evsel->prev_raw_counts->aggr = *count;
1353 } else {
1354 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1355 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1356 }
1357
1358 count->val = count->val - tmp.val;
1359 count->ena = count->ena - tmp.ena;
1360 count->run = count->run - tmp.run;
1361 }
1362
1363 void perf_counts_values__scale(struct perf_counts_values *count,
1364 bool scale, s8 *pscaled)
1365 {
1366 s8 scaled = 0;
1367
1368 if (scale) {
1369 if (count->run == 0) {
1370 scaled = -1;
1371 count->val = 0;
1372 } else if (count->run < count->ena) {
1373 scaled = 1;
1374 count->val = (u64)((double) count->val * count->ena / count->run);
1375 }
1376 }
1377
1378 if (pscaled)
1379 *pscaled = scaled;
1380 }
1381
1382 static int perf_evsel__read_size(struct evsel *evsel)
1383 {
1384 u64 read_format = evsel->core.attr.read_format;
1385 int entry = sizeof(u64); /* value */
1386 int size = 0;
1387 int nr = 1;
1388
1389 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1390 size += sizeof(u64);
1391
1392 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1393 size += sizeof(u64);
1394
1395 if (read_format & PERF_FORMAT_ID)
1396 entry += sizeof(u64);
1397
1398 if (read_format & PERF_FORMAT_GROUP) {
1399 nr = evsel->nr_members;
1400 size += sizeof(u64);
1401 }
1402
1403 size += entry * nr;
1404 return size;
1405 }
1406
1407 int perf_evsel__read(struct evsel *evsel, int cpu, int thread,
1408 struct perf_counts_values *count)
1409 {
1410 size_t size = perf_evsel__read_size(evsel);
1411
1412 memset(count, 0, sizeof(*count));
1413
1414 if (FD(evsel, cpu, thread) < 0)
1415 return -EINVAL;
1416
1417 if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1418 return -errno;
1419
1420 return 0;
1421 }
1422
1423 static int
1424 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread)
1425 {
1426 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1427
1428 return perf_evsel__read(evsel, cpu, thread, count);
1429 }
1430
1431 static void
1432 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1433 u64 val, u64 ena, u64 run)
1434 {
1435 struct perf_counts_values *count;
1436
1437 count = perf_counts(counter->counts, cpu, thread);
1438
1439 count->val = val;
1440 count->ena = ena;
1441 count->run = run;
1442
1443 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1444 }
1445
1446 static int
1447 perf_evsel__process_group_data(struct evsel *leader,
1448 int cpu, int thread, u64 *data)
1449 {
1450 u64 read_format = leader->core.attr.read_format;
1451 struct sample_read_value *v;
1452 u64 nr, ena = 0, run = 0, i;
1453
1454 nr = *data++;
1455
1456 if (nr != (u64) leader->nr_members)
1457 return -EINVAL;
1458
1459 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1460 ena = *data++;
1461
1462 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1463 run = *data++;
1464
1465 v = (struct sample_read_value *) data;
1466
1467 perf_evsel__set_count(leader, cpu, thread,
1468 v[0].value, ena, run);
1469
1470 for (i = 1; i < nr; i++) {
1471 struct evsel *counter;
1472
1473 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1474 if (!counter)
1475 return -EINVAL;
1476
1477 perf_evsel__set_count(counter, cpu, thread,
1478 v[i].value, ena, run);
1479 }
1480
1481 return 0;
1482 }
1483
1484 static int
1485 perf_evsel__read_group(struct evsel *leader, int cpu, int thread)
1486 {
1487 struct perf_stat_evsel *ps = leader->stats;
1488 u64 read_format = leader->core.attr.read_format;
1489 int size = perf_evsel__read_size(leader);
1490 u64 *data = ps->group_data;
1491
1492 if (!(read_format & PERF_FORMAT_ID))
1493 return -EINVAL;
1494
1495 if (!perf_evsel__is_group_leader(leader))
1496 return -EINVAL;
1497
1498 if (!data) {
1499 data = zalloc(size);
1500 if (!data)
1501 return -ENOMEM;
1502
1503 ps->group_data = data;
1504 }
1505
1506 if (FD(leader, cpu, thread) < 0)
1507 return -EINVAL;
1508
1509 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1510 return -errno;
1511
1512 return perf_evsel__process_group_data(leader, cpu, thread, data);
1513 }
1514
1515 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1516 {
1517 u64 read_format = evsel->core.attr.read_format;
1518
1519 if (read_format & PERF_FORMAT_GROUP)
1520 return perf_evsel__read_group(evsel, cpu, thread);
1521 else
1522 return perf_evsel__read_one(evsel, cpu, thread);
1523 }
1524
1525 int __perf_evsel__read_on_cpu(struct evsel *evsel,
1526 int cpu, int thread, bool scale)
1527 {
1528 struct perf_counts_values count;
1529 size_t nv = scale ? 3 : 1;
1530
1531 if (FD(evsel, cpu, thread) < 0)
1532 return -EINVAL;
1533
1534 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1535 return -ENOMEM;
1536
1537 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1538 return -errno;
1539
1540 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1541 perf_counts_values__scale(&count, scale, NULL);
1542 *perf_counts(evsel->counts, cpu, thread) = count;
1543 return 0;
1544 }
1545
1546 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1547 {
1548 struct evsel *leader = evsel->leader;
1549 int fd;
1550
1551 if (perf_evsel__is_group_leader(evsel))
1552 return -1;
1553
1554 /*
1555 * Leader must be already processed/open,
1556 * if not it's a bug.
1557 */
1558 BUG_ON(!leader->fd);
1559
1560 fd = FD(leader, cpu, thread);
1561 BUG_ON(fd == -1);
1562
1563 return fd;
1564 }
1565
1566 struct bit_names {
1567 int bit;
1568 const char *name;
1569 };
1570
1571 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1572 {
1573 bool first_bit = true;
1574 int i = 0;
1575
1576 do {
1577 if (value & bits[i].bit) {
1578 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1579 first_bit = false;
1580 }
1581 } while (bits[++i].name != NULL);
1582 }
1583
1584 static void __p_sample_type(char *buf, size_t size, u64 value)
1585 {
1586 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1587 struct bit_names bits[] = {
1588 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1589 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1590 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1591 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1592 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1593 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1594 { .name = NULL, }
1595 };
1596 #undef bit_name
1597 __p_bits(buf, size, value, bits);
1598 }
1599
1600 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1601 {
1602 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1603 struct bit_names bits[] = {
1604 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1605 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1606 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1607 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1608 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1609 { .name = NULL, }
1610 };
1611 #undef bit_name
1612 __p_bits(buf, size, value, bits);
1613 }
1614
1615 static void __p_read_format(char *buf, size_t size, u64 value)
1616 {
1617 #define bit_name(n) { PERF_FORMAT_##n, #n }
1618 struct bit_names bits[] = {
1619 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1620 bit_name(ID), bit_name(GROUP),
1621 { .name = NULL, }
1622 };
1623 #undef bit_name
1624 __p_bits(buf, size, value, bits);
1625 }
1626
1627 #define BUF_SIZE 1024
1628
1629 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1630 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1631 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1632 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1633 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1634 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1635
1636 #define PRINT_ATTRn(_n, _f, _p) \
1637 do { \
1638 if (attr->_f) { \
1639 _p(attr->_f); \
1640 ret += attr__fprintf(fp, _n, buf, priv);\
1641 } \
1642 } while (0)
1643
1644 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1645
1646 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1647 attr__fprintf_f attr__fprintf, void *priv)
1648 {
1649 char buf[BUF_SIZE];
1650 int ret = 0;
1651
1652 PRINT_ATTRf(type, p_unsigned);
1653 PRINT_ATTRf(size, p_unsigned);
1654 PRINT_ATTRf(config, p_hex);
1655 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1656 PRINT_ATTRf(sample_type, p_sample_type);
1657 PRINT_ATTRf(read_format, p_read_format);
1658
1659 PRINT_ATTRf(disabled, p_unsigned);
1660 PRINT_ATTRf(inherit, p_unsigned);
1661 PRINT_ATTRf(pinned, p_unsigned);
1662 PRINT_ATTRf(exclusive, p_unsigned);
1663 PRINT_ATTRf(exclude_user, p_unsigned);
1664 PRINT_ATTRf(exclude_kernel, p_unsigned);
1665 PRINT_ATTRf(exclude_hv, p_unsigned);
1666 PRINT_ATTRf(exclude_idle, p_unsigned);
1667 PRINT_ATTRf(mmap, p_unsigned);
1668 PRINT_ATTRf(comm, p_unsigned);
1669 PRINT_ATTRf(freq, p_unsigned);
1670 PRINT_ATTRf(inherit_stat, p_unsigned);
1671 PRINT_ATTRf(enable_on_exec, p_unsigned);
1672 PRINT_ATTRf(task, p_unsigned);
1673 PRINT_ATTRf(watermark, p_unsigned);
1674 PRINT_ATTRf(precise_ip, p_unsigned);
1675 PRINT_ATTRf(mmap_data, p_unsigned);
1676 PRINT_ATTRf(sample_id_all, p_unsigned);
1677 PRINT_ATTRf(exclude_host, p_unsigned);
1678 PRINT_ATTRf(exclude_guest, p_unsigned);
1679 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1680 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1681 PRINT_ATTRf(mmap2, p_unsigned);
1682 PRINT_ATTRf(comm_exec, p_unsigned);
1683 PRINT_ATTRf(use_clockid, p_unsigned);
1684 PRINT_ATTRf(context_switch, p_unsigned);
1685 PRINT_ATTRf(write_backward, p_unsigned);
1686 PRINT_ATTRf(namespaces, p_unsigned);
1687 PRINT_ATTRf(ksymbol, p_unsigned);
1688 PRINT_ATTRf(bpf_event, p_unsigned);
1689
1690 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1691 PRINT_ATTRf(bp_type, p_unsigned);
1692 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1693 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1694 PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1695 PRINT_ATTRf(sample_regs_user, p_hex);
1696 PRINT_ATTRf(sample_stack_user, p_unsigned);
1697 PRINT_ATTRf(clockid, p_signed);
1698 PRINT_ATTRf(sample_regs_intr, p_hex);
1699 PRINT_ATTRf(aux_watermark, p_unsigned);
1700 PRINT_ATTRf(sample_max_stack, p_unsigned);
1701
1702 return ret;
1703 }
1704
1705 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1706 void *priv __maybe_unused)
1707 {
1708 return fprintf(fp, " %-32s %s\n", name, val);
1709 }
1710
1711 static void perf_evsel__remove_fd(struct evsel *pos,
1712 int nr_cpus, int nr_threads,
1713 int thread_idx)
1714 {
1715 for (int cpu = 0; cpu < nr_cpus; cpu++)
1716 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1717 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1718 }
1719
1720 static int update_fds(struct evsel *evsel,
1721 int nr_cpus, int cpu_idx,
1722 int nr_threads, int thread_idx)
1723 {
1724 struct evsel *pos;
1725
1726 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1727 return -EINVAL;
1728
1729 evlist__for_each_entry(evsel->evlist, pos) {
1730 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1731
1732 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1733
1734 /*
1735 * Since fds for next evsel has not been created,
1736 * there is no need to iterate whole event list.
1737 */
1738 if (pos == evsel)
1739 break;
1740 }
1741 return 0;
1742 }
1743
1744 static bool ignore_missing_thread(struct evsel *evsel,
1745 int nr_cpus, int cpu,
1746 struct perf_thread_map *threads,
1747 int thread, int err)
1748 {
1749 pid_t ignore_pid = thread_map__pid(threads, thread);
1750
1751 if (!evsel->ignore_missing_thread)
1752 return false;
1753
1754 /* The system wide setup does not work with threads. */
1755 if (evsel->system_wide)
1756 return false;
1757
1758 /* The -ESRCH is perf event syscall errno for pid's not found. */
1759 if (err != -ESRCH)
1760 return false;
1761
1762 /* If there's only one thread, let it fail. */
1763 if (threads->nr == 1)
1764 return false;
1765
1766 /*
1767 * We should remove fd for missing_thread first
1768 * because thread_map__remove() will decrease threads->nr.
1769 */
1770 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1771 return false;
1772
1773 if (thread_map__remove(threads, thread))
1774 return false;
1775
1776 pr_warning("WARNING: Ignored open failure for pid %d\n",
1777 ignore_pid);
1778 return true;
1779 }
1780
1781 static void display_attr(struct perf_event_attr *attr)
1782 {
1783 if (verbose >= 2) {
1784 fprintf(stderr, "%.60s\n", graph_dotted_line);
1785 fprintf(stderr, "perf_event_attr:\n");
1786 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1787 fprintf(stderr, "%.60s\n", graph_dotted_line);
1788 }
1789 }
1790
1791 static int perf_event_open(struct evsel *evsel,
1792 pid_t pid, int cpu, int group_fd,
1793 unsigned long flags)
1794 {
1795 int precise_ip = evsel->core.attr.precise_ip;
1796 int fd;
1797
1798 while (1) {
1799 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1800 pid, cpu, group_fd, flags);
1801
1802 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1803 if (fd >= 0)
1804 break;
1805
1806 /* Do not try less precise if not requested. */
1807 if (!evsel->precise_max)
1808 break;
1809
1810 /*
1811 * We tried all the precise_ip values, and it's
1812 * still failing, so leave it to standard fallback.
1813 */
1814 if (!evsel->core.attr.precise_ip) {
1815 evsel->core.attr.precise_ip = precise_ip;
1816 break;
1817 }
1818
1819 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1820 evsel->core.attr.precise_ip--;
1821 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1822 display_attr(&evsel->core.attr);
1823 }
1824
1825 return fd;
1826 }
1827
1828 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1829 struct perf_thread_map *threads)
1830 {
1831 int cpu, thread, nthreads;
1832 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1833 int pid = -1, err;
1834 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1835
1836 if (perf_missing_features.write_backward && evsel->core.attr.write_backward)
1837 return -EINVAL;
1838
1839 if (cpus == NULL) {
1840 static struct perf_cpu_map *empty_cpu_map;
1841
1842 if (empty_cpu_map == NULL) {
1843 empty_cpu_map = perf_cpu_map__dummy_new();
1844 if (empty_cpu_map == NULL)
1845 return -ENOMEM;
1846 }
1847
1848 cpus = empty_cpu_map;
1849 }
1850
1851 if (threads == NULL) {
1852 static struct perf_thread_map *empty_thread_map;
1853
1854 if (empty_thread_map == NULL) {
1855 empty_thread_map = thread_map__new_by_tid(-1);
1856 if (empty_thread_map == NULL)
1857 return -ENOMEM;
1858 }
1859
1860 threads = empty_thread_map;
1861 }
1862
1863 if (evsel->system_wide)
1864 nthreads = 1;
1865 else
1866 nthreads = threads->nr;
1867
1868 if (evsel->fd == NULL &&
1869 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1870 return -ENOMEM;
1871
1872 if (evsel->cgrp) {
1873 flags |= PERF_FLAG_PID_CGROUP;
1874 pid = evsel->cgrp->fd;
1875 }
1876
1877 fallback_missing_features:
1878 if (perf_missing_features.clockid_wrong)
1879 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1880 if (perf_missing_features.clockid) {
1881 evsel->core.attr.use_clockid = 0;
1882 evsel->core.attr.clockid = 0;
1883 }
1884 if (perf_missing_features.cloexec)
1885 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1886 if (perf_missing_features.mmap2)
1887 evsel->core.attr.mmap2 = 0;
1888 if (perf_missing_features.exclude_guest)
1889 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1890 if (perf_missing_features.lbr_flags)
1891 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1892 PERF_SAMPLE_BRANCH_NO_CYCLES);
1893 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1894 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1895 if (perf_missing_features.ksymbol)
1896 evsel->core.attr.ksymbol = 0;
1897 if (perf_missing_features.bpf_event)
1898 evsel->core.attr.bpf_event = 0;
1899 retry_sample_id:
1900 if (perf_missing_features.sample_id_all)
1901 evsel->core.attr.sample_id_all = 0;
1902
1903 display_attr(&evsel->core.attr);
1904
1905 for (cpu = 0; cpu < cpus->nr; cpu++) {
1906
1907 for (thread = 0; thread < nthreads; thread++) {
1908 int fd, group_fd;
1909
1910 if (!evsel->cgrp && !evsel->system_wide)
1911 pid = thread_map__pid(threads, thread);
1912
1913 group_fd = get_group_fd(evsel, cpu, thread);
1914 retry_open:
1915 test_attr__ready();
1916
1917 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1918 group_fd, flags);
1919
1920 FD(evsel, cpu, thread) = fd;
1921
1922 if (fd < 0) {
1923 err = -errno;
1924
1925 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1926 /*
1927 * We just removed 1 thread, so take a step
1928 * back on thread index and lower the upper
1929 * nthreads limit.
1930 */
1931 nthreads--;
1932 thread--;
1933
1934 /* ... and pretend like nothing have happened. */
1935 err = 0;
1936 continue;
1937 }
1938
1939 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1940 err);
1941 goto try_fallback;
1942 }
1943
1944 pr_debug2(" = %d\n", fd);
1945
1946 if (evsel->bpf_fd >= 0) {
1947 int evt_fd = fd;
1948 int bpf_fd = evsel->bpf_fd;
1949
1950 err = ioctl(evt_fd,
1951 PERF_EVENT_IOC_SET_BPF,
1952 bpf_fd);
1953 if (err && errno != EEXIST) {
1954 pr_err("failed to attach bpf fd %d: %s\n",
1955 bpf_fd, strerror(errno));
1956 err = -EINVAL;
1957 goto out_close;
1958 }
1959 }
1960
1961 set_rlimit = NO_CHANGE;
1962
1963 /*
1964 * If we succeeded but had to kill clockid, fail and
1965 * have perf_evsel__open_strerror() print us a nice
1966 * error.
1967 */
1968 if (perf_missing_features.clockid ||
1969 perf_missing_features.clockid_wrong) {
1970 err = -EINVAL;
1971 goto out_close;
1972 }
1973 }
1974 }
1975
1976 return 0;
1977
1978 try_fallback:
1979 /*
1980 * perf stat needs between 5 and 22 fds per CPU. When we run out
1981 * of them try to increase the limits.
1982 */
1983 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1984 struct rlimit l;
1985 int old_errno = errno;
1986
1987 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1988 if (set_rlimit == NO_CHANGE)
1989 l.rlim_cur = l.rlim_max;
1990 else {
1991 l.rlim_cur = l.rlim_max + 1000;
1992 l.rlim_max = l.rlim_cur;
1993 }
1994 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1995 set_rlimit++;
1996 errno = old_errno;
1997 goto retry_open;
1998 }
1999 }
2000 errno = old_errno;
2001 }
2002
2003 if (err != -EINVAL || cpu > 0 || thread > 0)
2004 goto out_close;
2005
2006 /*
2007 * Must probe features in the order they were added to the
2008 * perf_event_attr interface.
2009 */
2010 if (!perf_missing_features.bpf_event && evsel->core.attr.bpf_event) {
2011 perf_missing_features.bpf_event = true;
2012 pr_debug2("switching off bpf_event\n");
2013 goto fallback_missing_features;
2014 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
2015 perf_missing_features.ksymbol = true;
2016 pr_debug2("switching off ksymbol\n");
2017 goto fallback_missing_features;
2018 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
2019 perf_missing_features.write_backward = true;
2020 pr_debug2("switching off write_backward\n");
2021 goto out_close;
2022 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
2023 perf_missing_features.clockid_wrong = true;
2024 pr_debug2("switching off clockid\n");
2025 goto fallback_missing_features;
2026 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
2027 perf_missing_features.clockid = true;
2028 pr_debug2("switching off use_clockid\n");
2029 goto fallback_missing_features;
2030 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
2031 perf_missing_features.cloexec = true;
2032 pr_debug2("switching off cloexec flag\n");
2033 goto fallback_missing_features;
2034 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
2035 perf_missing_features.mmap2 = true;
2036 pr_debug2("switching off mmap2\n");
2037 goto fallback_missing_features;
2038 } else if (!perf_missing_features.exclude_guest &&
2039 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
2040 perf_missing_features.exclude_guest = true;
2041 pr_debug2("switching off exclude_guest, exclude_host\n");
2042 goto fallback_missing_features;
2043 } else if (!perf_missing_features.sample_id_all) {
2044 perf_missing_features.sample_id_all = true;
2045 pr_debug2("switching off sample_id_all\n");
2046 goto retry_sample_id;
2047 } else if (!perf_missing_features.lbr_flags &&
2048 (evsel->core.attr.branch_sample_type &
2049 (PERF_SAMPLE_BRANCH_NO_CYCLES |
2050 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2051 perf_missing_features.lbr_flags = true;
2052 pr_debug2("switching off branch sample type no (cycles/flags)\n");
2053 goto fallback_missing_features;
2054 } else if (!perf_missing_features.group_read &&
2055 evsel->core.attr.inherit &&
2056 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2057 perf_evsel__is_group_leader(evsel)) {
2058 perf_missing_features.group_read = true;
2059 pr_debug2("switching off group read\n");
2060 goto fallback_missing_features;
2061 }
2062 out_close:
2063 if (err)
2064 threads->err_thread = thread;
2065
2066 do {
2067 while (--thread >= 0) {
2068 close(FD(evsel, cpu, thread));
2069 FD(evsel, cpu, thread) = -1;
2070 }
2071 thread = nthreads;
2072 } while (--cpu >= 0);
2073 return err;
2074 }
2075
2076 void perf_evsel__close(struct evsel *evsel)
2077 {
2078 if (evsel->fd == NULL)
2079 return;
2080
2081 perf_evsel__close_fd(evsel);
2082 perf_evsel__free_fd(evsel);
2083 perf_evsel__free_id(evsel);
2084 }
2085
2086 int perf_evsel__open_per_cpu(struct evsel *evsel,
2087 struct perf_cpu_map *cpus)
2088 {
2089 return evsel__open(evsel, cpus, NULL);
2090 }
2091
2092 int perf_evsel__open_per_thread(struct evsel *evsel,
2093 struct perf_thread_map *threads)
2094 {
2095 return evsel__open(evsel, NULL, threads);
2096 }
2097
2098 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2099 const union perf_event *event,
2100 struct perf_sample *sample)
2101 {
2102 u64 type = evsel->core.attr.sample_type;
2103 const u64 *array = event->sample.array;
2104 bool swapped = evsel->needs_swap;
2105 union u64_swap u;
2106
2107 array += ((event->header.size -
2108 sizeof(event->header)) / sizeof(u64)) - 1;
2109
2110 if (type & PERF_SAMPLE_IDENTIFIER) {
2111 sample->id = *array;
2112 array--;
2113 }
2114
2115 if (type & PERF_SAMPLE_CPU) {
2116 u.val64 = *array;
2117 if (swapped) {
2118 /* undo swap of u64, then swap on individual u32s */
2119 u.val64 = bswap_64(u.val64);
2120 u.val32[0] = bswap_32(u.val32[0]);
2121 }
2122
2123 sample->cpu = u.val32[0];
2124 array--;
2125 }
2126
2127 if (type & PERF_SAMPLE_STREAM_ID) {
2128 sample->stream_id = *array;
2129 array--;
2130 }
2131
2132 if (type & PERF_SAMPLE_ID) {
2133 sample->id = *array;
2134 array--;
2135 }
2136
2137 if (type & PERF_SAMPLE_TIME) {
2138 sample->time = *array;
2139 array--;
2140 }
2141
2142 if (type & PERF_SAMPLE_TID) {
2143 u.val64 = *array;
2144 if (swapped) {
2145 /* undo swap of u64, then swap on individual u32s */
2146 u.val64 = bswap_64(u.val64);
2147 u.val32[0] = bswap_32(u.val32[0]);
2148 u.val32[1] = bswap_32(u.val32[1]);
2149 }
2150
2151 sample->pid = u.val32[0];
2152 sample->tid = u.val32[1];
2153 array--;
2154 }
2155
2156 return 0;
2157 }
2158
2159 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2160 u64 size)
2161 {
2162 return size > max_size || offset + size > endp;
2163 }
2164
2165 #define OVERFLOW_CHECK(offset, size, max_size) \
2166 do { \
2167 if (overflow(endp, (max_size), (offset), (size))) \
2168 return -EFAULT; \
2169 } while (0)
2170
2171 #define OVERFLOW_CHECK_u64(offset) \
2172 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2173
2174 static int
2175 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2176 {
2177 /*
2178 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2179 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2180 * check the format does not go past the end of the event.
2181 */
2182 if (sample_size + sizeof(event->header) > event->header.size)
2183 return -EFAULT;
2184
2185 return 0;
2186 }
2187
2188 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2189 struct perf_sample *data)
2190 {
2191 u64 type = evsel->core.attr.sample_type;
2192 bool swapped = evsel->needs_swap;
2193 const u64 *array;
2194 u16 max_size = event->header.size;
2195 const void *endp = (void *)event + max_size;
2196 u64 sz;
2197
2198 /*
2199 * used for cross-endian analysis. See git commit 65014ab3
2200 * for why this goofiness is needed.
2201 */
2202 union u64_swap u;
2203
2204 memset(data, 0, sizeof(*data));
2205 data->cpu = data->pid = data->tid = -1;
2206 data->stream_id = data->id = data->time = -1ULL;
2207 data->period = evsel->core.attr.sample_period;
2208 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2209 data->misc = event->header.misc;
2210 data->id = -1ULL;
2211 data->data_src = PERF_MEM_DATA_SRC_NONE;
2212
2213 if (event->header.type != PERF_RECORD_SAMPLE) {
2214 if (!evsel->core.attr.sample_id_all)
2215 return 0;
2216 return perf_evsel__parse_id_sample(evsel, event, data);
2217 }
2218
2219 array = event->sample.array;
2220
2221 if (perf_event__check_size(event, evsel->sample_size))
2222 return -EFAULT;
2223
2224 if (type & PERF_SAMPLE_IDENTIFIER) {
2225 data->id = *array;
2226 array++;
2227 }
2228
2229 if (type & PERF_SAMPLE_IP) {
2230 data->ip = *array;
2231 array++;
2232 }
2233
2234 if (type & PERF_SAMPLE_TID) {
2235 u.val64 = *array;
2236 if (swapped) {
2237 /* undo swap of u64, then swap on individual u32s */
2238 u.val64 = bswap_64(u.val64);
2239 u.val32[0] = bswap_32(u.val32[0]);
2240 u.val32[1] = bswap_32(u.val32[1]);
2241 }
2242
2243 data->pid = u.val32[0];
2244 data->tid = u.val32[1];
2245 array++;
2246 }
2247
2248 if (type & PERF_SAMPLE_TIME) {
2249 data->time = *array;
2250 array++;
2251 }
2252
2253 if (type & PERF_SAMPLE_ADDR) {
2254 data->addr = *array;
2255 array++;
2256 }
2257
2258 if (type & PERF_SAMPLE_ID) {
2259 data->id = *array;
2260 array++;
2261 }
2262
2263 if (type & PERF_SAMPLE_STREAM_ID) {
2264 data->stream_id = *array;
2265 array++;
2266 }
2267
2268 if (type & PERF_SAMPLE_CPU) {
2269
2270 u.val64 = *array;
2271 if (swapped) {
2272 /* undo swap of u64, then swap on individual u32s */
2273 u.val64 = bswap_64(u.val64);
2274 u.val32[0] = bswap_32(u.val32[0]);
2275 }
2276
2277 data->cpu = u.val32[0];
2278 array++;
2279 }
2280
2281 if (type & PERF_SAMPLE_PERIOD) {
2282 data->period = *array;
2283 array++;
2284 }
2285
2286 if (type & PERF_SAMPLE_READ) {
2287 u64 read_format = evsel->core.attr.read_format;
2288
2289 OVERFLOW_CHECK_u64(array);
2290 if (read_format & PERF_FORMAT_GROUP)
2291 data->read.group.nr = *array;
2292 else
2293 data->read.one.value = *array;
2294
2295 array++;
2296
2297 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2298 OVERFLOW_CHECK_u64(array);
2299 data->read.time_enabled = *array;
2300 array++;
2301 }
2302
2303 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2304 OVERFLOW_CHECK_u64(array);
2305 data->read.time_running = *array;
2306 array++;
2307 }
2308
2309 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2310 if (read_format & PERF_FORMAT_GROUP) {
2311 const u64 max_group_nr = UINT64_MAX /
2312 sizeof(struct sample_read_value);
2313
2314 if (data->read.group.nr > max_group_nr)
2315 return -EFAULT;
2316 sz = data->read.group.nr *
2317 sizeof(struct sample_read_value);
2318 OVERFLOW_CHECK(array, sz, max_size);
2319 data->read.group.values =
2320 (struct sample_read_value *)array;
2321 array = (void *)array + sz;
2322 } else {
2323 OVERFLOW_CHECK_u64(array);
2324 data->read.one.id = *array;
2325 array++;
2326 }
2327 }
2328
2329 if (evsel__has_callchain(evsel)) {
2330 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2331
2332 OVERFLOW_CHECK_u64(array);
2333 data->callchain = (struct ip_callchain *)array++;
2334 if (data->callchain->nr > max_callchain_nr)
2335 return -EFAULT;
2336 sz = data->callchain->nr * sizeof(u64);
2337 OVERFLOW_CHECK(array, sz, max_size);
2338 array = (void *)array + sz;
2339 }
2340
2341 if (type & PERF_SAMPLE_RAW) {
2342 OVERFLOW_CHECK_u64(array);
2343 u.val64 = *array;
2344
2345 /*
2346 * Undo swap of u64, then swap on individual u32s,
2347 * get the size of the raw area and undo all of the
2348 * swap. The pevent interface handles endianity by
2349 * itself.
2350 */
2351 if (swapped) {
2352 u.val64 = bswap_64(u.val64);
2353 u.val32[0] = bswap_32(u.val32[0]);
2354 u.val32[1] = bswap_32(u.val32[1]);
2355 }
2356 data->raw_size = u.val32[0];
2357
2358 /*
2359 * The raw data is aligned on 64bits including the
2360 * u32 size, so it's safe to use mem_bswap_64.
2361 */
2362 if (swapped)
2363 mem_bswap_64((void *) array, data->raw_size);
2364
2365 array = (void *)array + sizeof(u32);
2366
2367 OVERFLOW_CHECK(array, data->raw_size, max_size);
2368 data->raw_data = (void *)array;
2369 array = (void *)array + data->raw_size;
2370 }
2371
2372 if (type & PERF_SAMPLE_BRANCH_STACK) {
2373 const u64 max_branch_nr = UINT64_MAX /
2374 sizeof(struct branch_entry);
2375
2376 OVERFLOW_CHECK_u64(array);
2377 data->branch_stack = (struct branch_stack *)array++;
2378
2379 if (data->branch_stack->nr > max_branch_nr)
2380 return -EFAULT;
2381 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2382 OVERFLOW_CHECK(array, sz, max_size);
2383 array = (void *)array + sz;
2384 }
2385
2386 if (type & PERF_SAMPLE_REGS_USER) {
2387 OVERFLOW_CHECK_u64(array);
2388 data->user_regs.abi = *array;
2389 array++;
2390
2391 if (data->user_regs.abi) {
2392 u64 mask = evsel->core.attr.sample_regs_user;
2393
2394 sz = hweight64(mask) * sizeof(u64);
2395 OVERFLOW_CHECK(array, sz, max_size);
2396 data->user_regs.mask = mask;
2397 data->user_regs.regs = (u64 *)array;
2398 array = (void *)array + sz;
2399 }
2400 }
2401
2402 if (type & PERF_SAMPLE_STACK_USER) {
2403 OVERFLOW_CHECK_u64(array);
2404 sz = *array++;
2405
2406 data->user_stack.offset = ((char *)(array - 1)
2407 - (char *) event);
2408
2409 if (!sz) {
2410 data->user_stack.size = 0;
2411 } else {
2412 OVERFLOW_CHECK(array, sz, max_size);
2413 data->user_stack.data = (char *)array;
2414 array = (void *)array + sz;
2415 OVERFLOW_CHECK_u64(array);
2416 data->user_stack.size = *array++;
2417 if (WARN_ONCE(data->user_stack.size > sz,
2418 "user stack dump failure\n"))
2419 return -EFAULT;
2420 }
2421 }
2422
2423 if (type & PERF_SAMPLE_WEIGHT) {
2424 OVERFLOW_CHECK_u64(array);
2425 data->weight = *array;
2426 array++;
2427 }
2428
2429 if (type & PERF_SAMPLE_DATA_SRC) {
2430 OVERFLOW_CHECK_u64(array);
2431 data->data_src = *array;
2432 array++;
2433 }
2434
2435 if (type & PERF_SAMPLE_TRANSACTION) {
2436 OVERFLOW_CHECK_u64(array);
2437 data->transaction = *array;
2438 array++;
2439 }
2440
2441 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2442 if (type & PERF_SAMPLE_REGS_INTR) {
2443 OVERFLOW_CHECK_u64(array);
2444 data->intr_regs.abi = *array;
2445 array++;
2446
2447 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2448 u64 mask = evsel->core.attr.sample_regs_intr;
2449
2450 sz = hweight64(mask) * sizeof(u64);
2451 OVERFLOW_CHECK(array, sz, max_size);
2452 data->intr_regs.mask = mask;
2453 data->intr_regs.regs = (u64 *)array;
2454 array = (void *)array + sz;
2455 }
2456 }
2457
2458 data->phys_addr = 0;
2459 if (type & PERF_SAMPLE_PHYS_ADDR) {
2460 data->phys_addr = *array;
2461 array++;
2462 }
2463
2464 return 0;
2465 }
2466
2467 int perf_evsel__parse_sample_timestamp(struct evsel *evsel,
2468 union perf_event *event,
2469 u64 *timestamp)
2470 {
2471 u64 type = evsel->core.attr.sample_type;
2472 const u64 *array;
2473
2474 if (!(type & PERF_SAMPLE_TIME))
2475 return -1;
2476
2477 if (event->header.type != PERF_RECORD_SAMPLE) {
2478 struct perf_sample data = {
2479 .time = -1ULL,
2480 };
2481
2482 if (!evsel->core.attr.sample_id_all)
2483 return -1;
2484 if (perf_evsel__parse_id_sample(evsel, event, &data))
2485 return -1;
2486
2487 *timestamp = data.time;
2488 return 0;
2489 }
2490
2491 array = event->sample.array;
2492
2493 if (perf_event__check_size(event, evsel->sample_size))
2494 return -EFAULT;
2495
2496 if (type & PERF_SAMPLE_IDENTIFIER)
2497 array++;
2498
2499 if (type & PERF_SAMPLE_IP)
2500 array++;
2501
2502 if (type & PERF_SAMPLE_TID)
2503 array++;
2504
2505 if (type & PERF_SAMPLE_TIME)
2506 *timestamp = *array;
2507
2508 return 0;
2509 }
2510
2511 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2512 u64 read_format)
2513 {
2514 size_t sz, result = sizeof(struct sample_event);
2515
2516 if (type & PERF_SAMPLE_IDENTIFIER)
2517 result += sizeof(u64);
2518
2519 if (type & PERF_SAMPLE_IP)
2520 result += sizeof(u64);
2521
2522 if (type & PERF_SAMPLE_TID)
2523 result += sizeof(u64);
2524
2525 if (type & PERF_SAMPLE_TIME)
2526 result += sizeof(u64);
2527
2528 if (type & PERF_SAMPLE_ADDR)
2529 result += sizeof(u64);
2530
2531 if (type & PERF_SAMPLE_ID)
2532 result += sizeof(u64);
2533
2534 if (type & PERF_SAMPLE_STREAM_ID)
2535 result += sizeof(u64);
2536
2537 if (type & PERF_SAMPLE_CPU)
2538 result += sizeof(u64);
2539
2540 if (type & PERF_SAMPLE_PERIOD)
2541 result += sizeof(u64);
2542
2543 if (type & PERF_SAMPLE_READ) {
2544 result += sizeof(u64);
2545 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2546 result += sizeof(u64);
2547 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2548 result += sizeof(u64);
2549 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2550 if (read_format & PERF_FORMAT_GROUP) {
2551 sz = sample->read.group.nr *
2552 sizeof(struct sample_read_value);
2553 result += sz;
2554 } else {
2555 result += sizeof(u64);
2556 }
2557 }
2558
2559 if (type & PERF_SAMPLE_CALLCHAIN) {
2560 sz = (sample->callchain->nr + 1) * sizeof(u64);
2561 result += sz;
2562 }
2563
2564 if (type & PERF_SAMPLE_RAW) {
2565 result += sizeof(u32);
2566 result += sample->raw_size;
2567 }
2568
2569 if (type & PERF_SAMPLE_BRANCH_STACK) {
2570 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2571 sz += sizeof(u64);
2572 result += sz;
2573 }
2574
2575 if (type & PERF_SAMPLE_REGS_USER) {
2576 if (sample->user_regs.abi) {
2577 result += sizeof(u64);
2578 sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2579 result += sz;
2580 } else {
2581 result += sizeof(u64);
2582 }
2583 }
2584
2585 if (type & PERF_SAMPLE_STACK_USER) {
2586 sz = sample->user_stack.size;
2587 result += sizeof(u64);
2588 if (sz) {
2589 result += sz;
2590 result += sizeof(u64);
2591 }
2592 }
2593
2594 if (type & PERF_SAMPLE_WEIGHT)
2595 result += sizeof(u64);
2596
2597 if (type & PERF_SAMPLE_DATA_SRC)
2598 result += sizeof(u64);
2599
2600 if (type & PERF_SAMPLE_TRANSACTION)
2601 result += sizeof(u64);
2602
2603 if (type & PERF_SAMPLE_REGS_INTR) {
2604 if (sample->intr_regs.abi) {
2605 result += sizeof(u64);
2606 sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2607 result += sz;
2608 } else {
2609 result += sizeof(u64);
2610 }
2611 }
2612
2613 if (type & PERF_SAMPLE_PHYS_ADDR)
2614 result += sizeof(u64);
2615
2616 return result;
2617 }
2618
2619 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2620 u64 read_format,
2621 const struct perf_sample *sample)
2622 {
2623 u64 *array;
2624 size_t sz;
2625 /*
2626 * used for cross-endian analysis. See git commit 65014ab3
2627 * for why this goofiness is needed.
2628 */
2629 union u64_swap u;
2630
2631 array = event->sample.array;
2632
2633 if (type & PERF_SAMPLE_IDENTIFIER) {
2634 *array = sample->id;
2635 array++;
2636 }
2637
2638 if (type & PERF_SAMPLE_IP) {
2639 *array = sample->ip;
2640 array++;
2641 }
2642
2643 if (type & PERF_SAMPLE_TID) {
2644 u.val32[0] = sample->pid;
2645 u.val32[1] = sample->tid;
2646 *array = u.val64;
2647 array++;
2648 }
2649
2650 if (type & PERF_SAMPLE_TIME) {
2651 *array = sample->time;
2652 array++;
2653 }
2654
2655 if (type & PERF_SAMPLE_ADDR) {
2656 *array = sample->addr;
2657 array++;
2658 }
2659
2660 if (type & PERF_SAMPLE_ID) {
2661 *array = sample->id;
2662 array++;
2663 }
2664
2665 if (type & PERF_SAMPLE_STREAM_ID) {
2666 *array = sample->stream_id;
2667 array++;
2668 }
2669
2670 if (type & PERF_SAMPLE_CPU) {
2671 u.val32[0] = sample->cpu;
2672 u.val32[1] = 0;
2673 *array = u.val64;
2674 array++;
2675 }
2676
2677 if (type & PERF_SAMPLE_PERIOD) {
2678 *array = sample->period;
2679 array++;
2680 }
2681
2682 if (type & PERF_SAMPLE_READ) {
2683 if (read_format & PERF_FORMAT_GROUP)
2684 *array = sample->read.group.nr;
2685 else
2686 *array = sample->read.one.value;
2687 array++;
2688
2689 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2690 *array = sample->read.time_enabled;
2691 array++;
2692 }
2693
2694 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2695 *array = sample->read.time_running;
2696 array++;
2697 }
2698
2699 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2700 if (read_format & PERF_FORMAT_GROUP) {
2701 sz = sample->read.group.nr *
2702 sizeof(struct sample_read_value);
2703 memcpy(array, sample->read.group.values, sz);
2704 array = (void *)array + sz;
2705 } else {
2706 *array = sample->read.one.id;
2707 array++;
2708 }
2709 }
2710
2711 if (type & PERF_SAMPLE_CALLCHAIN) {
2712 sz = (sample->callchain->nr + 1) * sizeof(u64);
2713 memcpy(array, sample->callchain, sz);
2714 array = (void *)array + sz;
2715 }
2716
2717 if (type & PERF_SAMPLE_RAW) {
2718 u.val32[0] = sample->raw_size;
2719 *array = u.val64;
2720 array = (void *)array + sizeof(u32);
2721
2722 memcpy(array, sample->raw_data, sample->raw_size);
2723 array = (void *)array + sample->raw_size;
2724 }
2725
2726 if (type & PERF_SAMPLE_BRANCH_STACK) {
2727 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2728 sz += sizeof(u64);
2729 memcpy(array, sample->branch_stack, sz);
2730 array = (void *)array + sz;
2731 }
2732
2733 if (type & PERF_SAMPLE_REGS_USER) {
2734 if (sample->user_regs.abi) {
2735 *array++ = sample->user_regs.abi;
2736 sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2737 memcpy(array, sample->user_regs.regs, sz);
2738 array = (void *)array + sz;
2739 } else {
2740 *array++ = 0;
2741 }
2742 }
2743
2744 if (type & PERF_SAMPLE_STACK_USER) {
2745 sz = sample->user_stack.size;
2746 *array++ = sz;
2747 if (sz) {
2748 memcpy(array, sample->user_stack.data, sz);
2749 array = (void *)array + sz;
2750 *array++ = sz;
2751 }
2752 }
2753
2754 if (type & PERF_SAMPLE_WEIGHT) {
2755 *array = sample->weight;
2756 array++;
2757 }
2758
2759 if (type & PERF_SAMPLE_DATA_SRC) {
2760 *array = sample->data_src;
2761 array++;
2762 }
2763
2764 if (type & PERF_SAMPLE_TRANSACTION) {
2765 *array = sample->transaction;
2766 array++;
2767 }
2768
2769 if (type & PERF_SAMPLE_REGS_INTR) {
2770 if (sample->intr_regs.abi) {
2771 *array++ = sample->intr_regs.abi;
2772 sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2773 memcpy(array, sample->intr_regs.regs, sz);
2774 array = (void *)array + sz;
2775 } else {
2776 *array++ = 0;
2777 }
2778 }
2779
2780 if (type & PERF_SAMPLE_PHYS_ADDR) {
2781 *array = sample->phys_addr;
2782 array++;
2783 }
2784
2785 return 0;
2786 }
2787
2788 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name)
2789 {
2790 return tep_find_field(evsel->tp_format, name);
2791 }
2792
2793 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample,
2794 const char *name)
2795 {
2796 struct tep_format_field *field = perf_evsel__field(evsel, name);
2797 int offset;
2798
2799 if (!field)
2800 return NULL;
2801
2802 offset = field->offset;
2803
2804 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2805 offset = *(int *)(sample->raw_data + field->offset);
2806 offset &= 0xffff;
2807 }
2808
2809 return sample->raw_data + offset;
2810 }
2811
2812 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2813 bool needs_swap)
2814 {
2815 u64 value;
2816 void *ptr = sample->raw_data + field->offset;
2817
2818 switch (field->size) {
2819 case 1:
2820 return *(u8 *)ptr;
2821 case 2:
2822 value = *(u16 *)ptr;
2823 break;
2824 case 4:
2825 value = *(u32 *)ptr;
2826 break;
2827 case 8:
2828 memcpy(&value, ptr, sizeof(u64));
2829 break;
2830 default:
2831 return 0;
2832 }
2833
2834 if (!needs_swap)
2835 return value;
2836
2837 switch (field->size) {
2838 case 2:
2839 return bswap_16(value);
2840 case 4:
2841 return bswap_32(value);
2842 case 8:
2843 return bswap_64(value);
2844 default:
2845 return 0;
2846 }
2847
2848 return 0;
2849 }
2850
2851 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample,
2852 const char *name)
2853 {
2854 struct tep_format_field *field = perf_evsel__field(evsel, name);
2855
2856 if (!field)
2857 return 0;
2858
2859 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2860 }
2861
2862 bool perf_evsel__fallback(struct evsel *evsel, int err,
2863 char *msg, size_t msgsize)
2864 {
2865 int paranoid;
2866
2867 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2868 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2869 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2870 /*
2871 * If it's cycles then fall back to hrtimer based
2872 * cpu-clock-tick sw counter, which is always available even if
2873 * no PMU support.
2874 *
2875 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2876 * b0a873e).
2877 */
2878 scnprintf(msg, msgsize, "%s",
2879 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2880
2881 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2882 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2883
2884 zfree(&evsel->name);
2885 return true;
2886 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2887 (paranoid = perf_event_paranoid()) > 1) {
2888 const char *name = perf_evsel__name(evsel);
2889 char *new_name;
2890 const char *sep = ":";
2891
2892 /* Is there already the separator in the name. */
2893 if (strchr(name, '/') ||
2894 strchr(name, ':'))
2895 sep = "";
2896
2897 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2898 return false;
2899
2900 if (evsel->name)
2901 free(evsel->name);
2902 evsel->name = new_name;
2903 scnprintf(msg, msgsize,
2904 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2905 evsel->core.attr.exclude_kernel = 1;
2906
2907 return true;
2908 }
2909
2910 return false;
2911 }
2912
2913 static bool find_process(const char *name)
2914 {
2915 size_t len = strlen(name);
2916 DIR *dir;
2917 struct dirent *d;
2918 int ret = -1;
2919
2920 dir = opendir(procfs__mountpoint());
2921 if (!dir)
2922 return false;
2923
2924 /* Walk through the directory. */
2925 while (ret && (d = readdir(dir)) != NULL) {
2926 char path[PATH_MAX];
2927 char *data;
2928 size_t size;
2929
2930 if ((d->d_type != DT_DIR) ||
2931 !strcmp(".", d->d_name) ||
2932 !strcmp("..", d->d_name))
2933 continue;
2934
2935 scnprintf(path, sizeof(path), "%s/%s/comm",
2936 procfs__mountpoint(), d->d_name);
2937
2938 if (filename__read_str(path, &data, &size))
2939 continue;
2940
2941 ret = strncmp(name, data, len);
2942 free(data);
2943 }
2944
2945 closedir(dir);
2946 return ret ? false : true;
2947 }
2948
2949 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target,
2950 int err, char *msg, size_t size)
2951 {
2952 char sbuf[STRERR_BUFSIZE];
2953 int printed = 0;
2954
2955 switch (err) {
2956 case EPERM:
2957 case EACCES:
2958 if (err == EPERM)
2959 printed = scnprintf(msg, size,
2960 "No permission to enable %s event.\n\n",
2961 perf_evsel__name(evsel));
2962
2963 return scnprintf(msg + printed, size - printed,
2964 "You may not have permission to collect %sstats.\n\n"
2965 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2966 "which controls use of the performance events system by\n"
2967 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2968 "The current value is %d:\n\n"
2969 " -1: Allow use of (almost) all events by all users\n"
2970 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2971 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2972 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2973 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2974 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2975 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2976 " kernel.perf_event_paranoid = -1\n" ,
2977 target->system_wide ? "system-wide " : "",
2978 perf_event_paranoid());
2979 case ENOENT:
2980 return scnprintf(msg, size, "The %s event is not supported.",
2981 perf_evsel__name(evsel));
2982 case EMFILE:
2983 return scnprintf(msg, size, "%s",
2984 "Too many events are opened.\n"
2985 "Probably the maximum number of open file descriptors has been reached.\n"
2986 "Hint: Try again after reducing the number of events.\n"
2987 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2988 case ENOMEM:
2989 if (evsel__has_callchain(evsel) &&
2990 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2991 return scnprintf(msg, size,
2992 "Not enough memory to setup event with callchain.\n"
2993 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2994 "Hint: Current value: %d", sysctl__max_stack());
2995 break;
2996 case ENODEV:
2997 if (target->cpu_list)
2998 return scnprintf(msg, size, "%s",
2999 "No such device - did you specify an out-of-range profile CPU?");
3000 break;
3001 case EOPNOTSUPP:
3002 if (evsel->core.attr.sample_period != 0)
3003 return scnprintf(msg, size,
3004 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3005 perf_evsel__name(evsel));
3006 if (evsel->core.attr.precise_ip)
3007 return scnprintf(msg, size, "%s",
3008 "\'precise\' request may not be supported. Try removing 'p' modifier.");
3009 #if defined(__i386__) || defined(__x86_64__)
3010 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3011 return scnprintf(msg, size, "%s",
3012 "No hardware sampling interrupt available.\n");
3013 #endif
3014 break;
3015 case EBUSY:
3016 if (find_process("oprofiled"))
3017 return scnprintf(msg, size,
3018 "The PMU counters are busy/taken by another profiler.\n"
3019 "We found oprofile daemon running, please stop it and try again.");
3020 break;
3021 case EINVAL:
3022 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3023 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3024 if (perf_missing_features.clockid)
3025 return scnprintf(msg, size, "clockid feature not supported.");
3026 if (perf_missing_features.clockid_wrong)
3027 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3028 break;
3029 default:
3030 break;
3031 }
3032
3033 return scnprintf(msg, size,
3034 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3035 "/bin/dmesg | grep -i perf may provide additional information.\n",
3036 err, str_error_r(err, sbuf, sizeof(sbuf)),
3037 perf_evsel__name(evsel));
3038 }
3039
3040 struct perf_env *perf_evsel__env(struct evsel *evsel)
3041 {
3042 if (evsel && evsel->evlist)
3043 return evsel->evlist->env;
3044 return NULL;
3045 }
3046
3047 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3048 {
3049 int cpu, thread;
3050
3051 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
3052 for (thread = 0; thread < xyarray__max_y(evsel->fd);
3053 thread++) {
3054 int fd = FD(evsel, cpu, thread);
3055
3056 if (perf_evlist__id_add_fd(evlist, evsel,
3057 cpu, thread, fd) < 0)
3058 return -1;
3059 }
3060 }
3061
3062 return 0;
3063 }
3064
3065 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3066 {
3067 struct perf_cpu_map *cpus = evsel->core.cpus;
3068 struct perf_thread_map *threads = evsel->core.threads;
3069
3070 if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
3071 return -ENOMEM;
3072
3073 return store_evsel_ids(evsel, evlist);
3074 }
Ubuntu Focal Linux kernel mirror