?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'block-5.9-2020-08-14' of git://git.kernel.dk/linux-block
[linux/fpc-iii.git] / tools / perf / util / evsel.c
blobfd865002cbbd444c2eee3eed8f092dc1ea7d3c07
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 <stdlib.h>
26 #include <perf/evsel.h>
27 #include "asm/bug.h"
28 #include "callchain.h"
29 #include "cgroup.h"
30 #include "counts.h"
31 #include "event.h"
32 #include "evsel.h"
33 #include "util/env.h"
34 #include "util/evsel_config.h"
35 #include "util/evsel_fprintf.h"
36 #include "evlist.h"
37 #include <perf/cpumap.h>
38 #include "thread_map.h"
39 #include "target.h"
40 #include "perf_regs.h"
41 #include "record.h"
42 #include "debug.h"
43 #include "trace-event.h"
44 #include "stat.h"
45 #include "string2.h"
46 #include "memswap.h"
47 #include "util.h"
48 #include "../perf-sys.h"
49 #include "util/parse-branch-options.h"
50 #include <internal/xyarray.h>
51 #include <internal/lib.h>
52
53 #include <linux/ctype.h>
54
55 struct perf_missing_features perf_missing_features;
56
57 static clockid_t clockid;
58
59 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
60 {
61 return 0;
62 }
63
64 void __weak test_attr__ready(void) { }
65
66 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
67 {
68 }
69
70 static struct {
71 size_t size;
72 int (*init)(struct evsel *evsel);
73 void (*fini)(struct evsel *evsel);
74 } perf_evsel__object = {
75 .size = sizeof(struct evsel),
76 .init = evsel__no_extra_init,
77 .fini = evsel__no_extra_fini,
78 };
79
80 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
81 void (*fini)(struct evsel *evsel))
82 {
83
84 if (object_size == 0)
85 goto set_methods;
86
87 if (perf_evsel__object.size > object_size)
88 return -EINVAL;
89
90 perf_evsel__object.size = object_size;
91
92 set_methods:
93 if (init != NULL)
94 perf_evsel__object.init = init;
95
96 if (fini != NULL)
97 perf_evsel__object.fini = fini;
98
99 return 0;
100 }
101
102 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
103
104 int __evsel__sample_size(u64 sample_type)
105 {
106 u64 mask = sample_type & PERF_SAMPLE_MASK;
107 int size = 0;
108 int i;
109
110 for (i = 0; i < 64; i++) {
111 if (mask & (1ULL << i))
112 size++;
113 }
114
115 size *= sizeof(u64);
116
117 return size;
118 }
119
120 /**
121 * __perf_evsel__calc_id_pos - calculate id_pos.
122 * @sample_type: sample type
123 *
124 * This function returns the position of the event id (PERF_SAMPLE_ID or
125 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
126 * perf_record_sample.
127 */
128 static int __perf_evsel__calc_id_pos(u64 sample_type)
129 {
130 int idx = 0;
131
132 if (sample_type & PERF_SAMPLE_IDENTIFIER)
133 return 0;
134
135 if (!(sample_type & PERF_SAMPLE_ID))
136 return -1;
137
138 if (sample_type & PERF_SAMPLE_IP)
139 idx += 1;
140
141 if (sample_type & PERF_SAMPLE_TID)
142 idx += 1;
143
144 if (sample_type & PERF_SAMPLE_TIME)
145 idx += 1;
146
147 if (sample_type & PERF_SAMPLE_ADDR)
148 idx += 1;
149
150 return idx;
151 }
152
153 /**
154 * __perf_evsel__calc_is_pos - calculate is_pos.
155 * @sample_type: sample type
156 *
157 * This function returns the position (counting backwards) of the event id
158 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
159 * sample_id_all is used there is an id sample appended to non-sample events.
160 */
161 static int __perf_evsel__calc_is_pos(u64 sample_type)
162 {
163 int idx = 1;
164
165 if (sample_type & PERF_SAMPLE_IDENTIFIER)
166 return 1;
167
168 if (!(sample_type & PERF_SAMPLE_ID))
169 return -1;
170
171 if (sample_type & PERF_SAMPLE_CPU)
172 idx += 1;
173
174 if (sample_type & PERF_SAMPLE_STREAM_ID)
175 idx += 1;
176
177 return idx;
178 }
179
180 void evsel__calc_id_pos(struct evsel *evsel)
181 {
182 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
183 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
184 }
185
186 void __evsel__set_sample_bit(struct evsel *evsel,
187 enum perf_event_sample_format bit)
188 {
189 if (!(evsel->core.attr.sample_type & bit)) {
190 evsel->core.attr.sample_type |= bit;
191 evsel->sample_size += sizeof(u64);
192 evsel__calc_id_pos(evsel);
193 }
194 }
195
196 void __evsel__reset_sample_bit(struct evsel *evsel,
197 enum perf_event_sample_format bit)
198 {
199 if (evsel->core.attr.sample_type & bit) {
200 evsel->core.attr.sample_type &= ~bit;
201 evsel->sample_size -= sizeof(u64);
202 evsel__calc_id_pos(evsel);
203 }
204 }
205
206 void evsel__set_sample_id(struct evsel *evsel,
207 bool can_sample_identifier)
208 {
209 if (can_sample_identifier) {
210 evsel__reset_sample_bit(evsel, ID);
211 evsel__set_sample_bit(evsel, IDENTIFIER);
212 } else {
213 evsel__set_sample_bit(evsel, ID);
214 }
215 evsel->core.attr.read_format |= PERF_FORMAT_ID;
216 }
217
218 /**
219 * evsel__is_function_event - Return whether given evsel is a function
220 * trace event
221 *
222 * @evsel - evsel selector to be tested
223 *
224 * Return %true if event is function trace event
225 */
226 bool evsel__is_function_event(struct evsel *evsel)
227 {
228 #define FUNCTION_EVENT "ftrace:function"
229
230 return evsel->name &&
231 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
232
233 #undef FUNCTION_EVENT
234 }
235
236 void evsel__init(struct evsel *evsel,
237 struct perf_event_attr *attr, int idx)
238 {
239 perf_evsel__init(&evsel->core, attr);
240 evsel->idx = idx;
241 evsel->tracking = !idx;
242 evsel->leader = evsel;
243 evsel->unit = "";
244 evsel->scale = 1.0;
245 evsel->max_events = ULONG_MAX;
246 evsel->evlist = NULL;
247 evsel->bpf_obj = NULL;
248 evsel->bpf_fd = -1;
249 INIT_LIST_HEAD(&evsel->config_terms);
250 perf_evsel__object.init(evsel);
251 evsel->sample_size = __evsel__sample_size(attr->sample_type);
252 evsel__calc_id_pos(evsel);
253 evsel->cmdline_group_boundary = false;
254 evsel->metric_expr = NULL;
255 evsel->metric_name = NULL;
256 evsel->metric_events = NULL;
257 evsel->per_pkg_mask = NULL;
258 evsel->collect_stat = false;
259 evsel->pmu_name = NULL;
260 }
261
262 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
263 {
264 struct evsel *evsel = zalloc(perf_evsel__object.size);
265
266 if (!evsel)
267 return NULL;
268 evsel__init(evsel, attr, idx);
269
270 if (evsel__is_bpf_output(evsel)) {
271 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
272 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
273 evsel->core.attr.sample_period = 1;
274 }
275
276 if (evsel__is_clock(evsel)) {
277 /*
278 * The evsel->unit points to static alias->unit
279 * so it's ok to use static string in here.
280 */
281 static const char *unit = "msec";
282
283 evsel->unit = unit;
284 evsel->scale = 1e-6;
285 }
286
287 return evsel;
288 }
289
290 static bool perf_event_can_profile_kernel(void)
291 {
292 return perf_event_paranoid_check(1);
293 }
294
295 struct evsel *evsel__new_cycles(bool precise)
296 {
297 struct perf_event_attr attr = {
298 .type = PERF_TYPE_HARDWARE,
299 .config = PERF_COUNT_HW_CPU_CYCLES,
300 .exclude_kernel = !perf_event_can_profile_kernel(),
301 };
302 struct evsel *evsel;
303
304 event_attr_init(&attr);
305
306 if (!precise)
307 goto new_event;
308
309 /*
310 * Now let the usual logic to set up the perf_event_attr defaults
311 * to kick in when we return and before perf_evsel__open() is called.
312 */
313 new_event:
314 evsel = evsel__new(&attr);
315 if (evsel == NULL)
316 goto out;
317
318 evsel->precise_max = true;
319
320 /* use asprintf() because free(evsel) assumes name is allocated */
321 if (asprintf(&evsel->name, "cycles%s%s%.*s",
322 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
323 attr.exclude_kernel ? "u" : "",
324 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
325 goto error_free;
326 out:
327 return evsel;
328 error_free:
329 evsel__delete(evsel);
330 evsel = NULL;
331 goto out;
332 }
333
334 /*
335 * Returns pointer with encoded error via <linux/err.h> interface.
336 */
337 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
338 {
339 struct evsel *evsel = zalloc(perf_evsel__object.size);
340 int err = -ENOMEM;
341
342 if (evsel == NULL) {
343 goto out_err;
344 } else {
345 struct perf_event_attr attr = {
346 .type = PERF_TYPE_TRACEPOINT,
347 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
348 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
349 };
350
351 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
352 goto out_free;
353
354 evsel->tp_format = trace_event__tp_format(sys, name);
355 if (IS_ERR(evsel->tp_format)) {
356 err = PTR_ERR(evsel->tp_format);
357 goto out_free;
358 }
359
360 event_attr_init(&attr);
361 attr.config = evsel->tp_format->id;
362 attr.sample_period = 1;
363 evsel__init(evsel, &attr, idx);
364 }
365
366 return evsel;
367
368 out_free:
369 zfree(&evsel->name);
370 free(evsel);
371 out_err:
372 return ERR_PTR(err);
373 }
374
375 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
376 "cycles",
377 "instructions",
378 "cache-references",
379 "cache-misses",
380 "branches",
381 "branch-misses",
382 "bus-cycles",
383 "stalled-cycles-frontend",
384 "stalled-cycles-backend",
385 "ref-cycles",
386 };
387
388 static const char *__evsel__hw_name(u64 config)
389 {
390 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
391 return evsel__hw_names[config];
392
393 return "unknown-hardware";
394 }
395
396 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
397 {
398 int colon = 0, r = 0;
399 struct perf_event_attr *attr = &evsel->core.attr;
400 bool exclude_guest_default = false;
401
402 #define MOD_PRINT(context, mod) do { \
403 if (!attr->exclude_##context) { \
404 if (!colon) colon = ++r; \
405 r += scnprintf(bf + r, size - r, "%c", mod); \
406 } } while(0)
407
408 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
409 MOD_PRINT(kernel, 'k');
410 MOD_PRINT(user, 'u');
411 MOD_PRINT(hv, 'h');
412 exclude_guest_default = true;
413 }
414
415 if (attr->precise_ip) {
416 if (!colon)
417 colon = ++r;
418 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
419 exclude_guest_default = true;
420 }
421
422 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
423 MOD_PRINT(host, 'H');
424 MOD_PRINT(guest, 'G');
425 }
426 #undef MOD_PRINT
427 if (colon)
428 bf[colon - 1] = ':';
429 return r;
430 }
431
432 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
433 {
434 int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
435 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
436 }
437
438 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
439 "cpu-clock",
440 "task-clock",
441 "page-faults",
442 "context-switches",
443 "cpu-migrations",
444 "minor-faults",
445 "major-faults",
446 "alignment-faults",
447 "emulation-faults",
448 "dummy",
449 };
450
451 static const char *__evsel__sw_name(u64 config)
452 {
453 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
454 return evsel__sw_names[config];
455 return "unknown-software";
456 }
457
458 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
459 {
460 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
461 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
462 }
463
464 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
465 {
466 int r;
467
468 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
469
470 if (type & HW_BREAKPOINT_R)
471 r += scnprintf(bf + r, size - r, "r");
472
473 if (type & HW_BREAKPOINT_W)
474 r += scnprintf(bf + r, size - r, "w");
475
476 if (type & HW_BREAKPOINT_X)
477 r += scnprintf(bf + r, size - r, "x");
478
479 return r;
480 }
481
482 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
483 {
484 struct perf_event_attr *attr = &evsel->core.attr;
485 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
486 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
487 }
488
489 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
490 { "L1-dcache", "l1-d", "l1d", "L1-data", },
491 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
492 { "LLC", "L2", },
493 { "dTLB", "d-tlb", "Data-TLB", },
494 { "iTLB", "i-tlb", "Instruction-TLB", },
495 { "branch", "branches", "bpu", "btb", "bpc", },
496 { "node", },
497 };
498
499 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
500 { "load", "loads", "read", },
501 { "store", "stores", "write", },
502 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
503 };
504
505 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
506 { "refs", "Reference", "ops", "access", },
507 { "misses", "miss", },
508 };
509
510 #define C(x) PERF_COUNT_HW_CACHE_##x
511 #define CACHE_READ (1 << C(OP_READ))
512 #define CACHE_WRITE (1 << C(OP_WRITE))
513 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
514 #define COP(x) (1 << x)
515
516 /*
517 * cache operartion stat
518 * L1I : Read and prefetch only
519 * ITLB and BPU : Read-only
520 */
521 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
522 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
523 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
524 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
525 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
526 [C(ITLB)] = (CACHE_READ),
527 [C(BPU)] = (CACHE_READ),
528 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
529 };
530
531 bool evsel__is_cache_op_valid(u8 type, u8 op)
532 {
533 if (evsel__hw_cache_stat[type] & COP(op))
534 return true; /* valid */
535 else
536 return false; /* invalid */
537 }
538
539 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
540 {
541 if (result) {
542 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
543 evsel__hw_cache_op[op][0],
544 evsel__hw_cache_result[result][0]);
545 }
546
547 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
548 evsel__hw_cache_op[op][1]);
549 }
550
551 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
552 {
553 u8 op, result, type = (config >> 0) & 0xff;
554 const char *err = "unknown-ext-hardware-cache-type";
555
556 if (type >= PERF_COUNT_HW_CACHE_MAX)
557 goto out_err;
558
559 op = (config >> 8) & 0xff;
560 err = "unknown-ext-hardware-cache-op";
561 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
562 goto out_err;
563
564 result = (config >> 16) & 0xff;
565 err = "unknown-ext-hardware-cache-result";
566 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
567 goto out_err;
568
569 err = "invalid-cache";
570 if (!evsel__is_cache_op_valid(type, op))
571 goto out_err;
572
573 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
574 out_err:
575 return scnprintf(bf, size, "%s", err);
576 }
577
578 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
579 {
580 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
581 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
582 }
583
584 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
585 {
586 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
587 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
588 }
589
590 static int evsel__tool_name(char *bf, size_t size)
591 {
592 int ret = scnprintf(bf, size, "duration_time");
593 return ret;
594 }
595
596 const char *evsel__name(struct evsel *evsel)
597 {
598 char bf[128];
599
600 if (!evsel)
601 goto out_unknown;
602
603 if (evsel->name)
604 return evsel->name;
605
606 switch (evsel->core.attr.type) {
607 case PERF_TYPE_RAW:
608 evsel__raw_name(evsel, bf, sizeof(bf));
609 break;
610
611 case PERF_TYPE_HARDWARE:
612 evsel__hw_name(evsel, bf, sizeof(bf));
613 break;
614
615 case PERF_TYPE_HW_CACHE:
616 evsel__hw_cache_name(evsel, bf, sizeof(bf));
617 break;
618
619 case PERF_TYPE_SOFTWARE:
620 if (evsel->tool_event)
621 evsel__tool_name(bf, sizeof(bf));
622 else
623 evsel__sw_name(evsel, bf, sizeof(bf));
624 break;
625
626 case PERF_TYPE_TRACEPOINT:
627 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
628 break;
629
630 case PERF_TYPE_BREAKPOINT:
631 evsel__bp_name(evsel, bf, sizeof(bf));
632 break;
633
634 default:
635 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
636 evsel->core.attr.type);
637 break;
638 }
639
640 evsel->name = strdup(bf);
641
642 if (evsel->name)
643 return evsel->name;
644 out_unknown:
645 return "unknown";
646 }
647
648 const char *evsel__group_name(struct evsel *evsel)
649 {
650 return evsel->group_name ?: "anon group";
651 }
652
653 /*
654 * Returns the group details for the specified leader,
655 * with following rules.
656 *
657 * For record -e '{cycles,instructions}'
658 * 'anon group { cycles:u, instructions:u }'
659 *
660 * For record -e 'cycles,instructions' and report --group
661 * 'cycles:u, instructions:u'
662 */
663 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
664 {
665 int ret = 0;
666 struct evsel *pos;
667 const char *group_name = evsel__group_name(evsel);
668
669 if (!evsel->forced_leader)
670 ret = scnprintf(buf, size, "%s { ", group_name);
671
672 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
673
674 for_each_group_member(pos, evsel)
675 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
676
677 if (!evsel->forced_leader)
678 ret += scnprintf(buf + ret, size - ret, " }");
679
680 return ret;
681 }
682
683 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
684 struct callchain_param *param)
685 {
686 bool function = evsel__is_function_event(evsel);
687 struct perf_event_attr *attr = &evsel->core.attr;
688
689 evsel__set_sample_bit(evsel, CALLCHAIN);
690
691 attr->sample_max_stack = param->max_stack;
692
693 if (opts->kernel_callchains)
694 attr->exclude_callchain_user = 1;
695 if (opts->user_callchains)
696 attr->exclude_callchain_kernel = 1;
697 if (param->record_mode == CALLCHAIN_LBR) {
698 if (!opts->branch_stack) {
699 if (attr->exclude_user) {
700 pr_warning("LBR callstack option is only available "
701 "to get user callchain information. "
702 "Falling back to framepointers.\n");
703 } else {
704 evsel__set_sample_bit(evsel, BRANCH_STACK);
705 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
706 PERF_SAMPLE_BRANCH_CALL_STACK |
707 PERF_SAMPLE_BRANCH_NO_CYCLES |
708 PERF_SAMPLE_BRANCH_NO_FLAGS |
709 PERF_SAMPLE_BRANCH_HW_INDEX;
710 }
711 } else
712 pr_warning("Cannot use LBR callstack with branch stack. "
713 "Falling back to framepointers.\n");
714 }
715
716 if (param->record_mode == CALLCHAIN_DWARF) {
717 if (!function) {
718 evsel__set_sample_bit(evsel, REGS_USER);
719 evsel__set_sample_bit(evsel, STACK_USER);
720 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
721 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
722 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
723 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
724 "so the minimal registers set (IP, SP) is explicitly forced.\n");
725 } else {
726 attr->sample_regs_user |= PERF_REGS_MASK;
727 }
728 attr->sample_stack_user = param->dump_size;
729 attr->exclude_callchain_user = 1;
730 } else {
731 pr_info("Cannot use DWARF unwind for function trace event,"
732 " falling back to framepointers.\n");
733 }
734 }
735
736 if (function) {
737 pr_info("Disabling user space callchains for function trace event.\n");
738 attr->exclude_callchain_user = 1;
739 }
740 }
741
742 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
743 struct callchain_param *param)
744 {
745 if (param->enabled)
746 return __evsel__config_callchain(evsel, opts, param);
747 }
748
749 static void
750 perf_evsel__reset_callgraph(struct evsel *evsel,
751 struct callchain_param *param)
752 {
753 struct perf_event_attr *attr = &evsel->core.attr;
754
755 evsel__reset_sample_bit(evsel, CALLCHAIN);
756 if (param->record_mode == CALLCHAIN_LBR) {
757 evsel__reset_sample_bit(evsel, BRANCH_STACK);
758 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
759 PERF_SAMPLE_BRANCH_CALL_STACK |
760 PERF_SAMPLE_BRANCH_HW_INDEX);
761 }
762 if (param->record_mode == CALLCHAIN_DWARF) {
763 evsel__reset_sample_bit(evsel, REGS_USER);
764 evsel__reset_sample_bit(evsel, STACK_USER);
765 }
766 }
767
768 static void evsel__apply_config_terms(struct evsel *evsel,
769 struct record_opts *opts, bool track)
770 {
771 struct evsel_config_term *term;
772 struct list_head *config_terms = &evsel->config_terms;
773 struct perf_event_attr *attr = &evsel->core.attr;
774 /* callgraph default */
775 struct callchain_param param = {
776 .record_mode = callchain_param.record_mode,
777 };
778 u32 dump_size = 0;
779 int max_stack = 0;
780 const char *callgraph_buf = NULL;
781
782 list_for_each_entry(term, config_terms, list) {
783 switch (term->type) {
784 case EVSEL__CONFIG_TERM_PERIOD:
785 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
786 attr->sample_period = term->val.period;
787 attr->freq = 0;
788 evsel__reset_sample_bit(evsel, PERIOD);
789 }
790 break;
791 case EVSEL__CONFIG_TERM_FREQ:
792 if (!(term->weak && opts->user_freq != UINT_MAX)) {
793 attr->sample_freq = term->val.freq;
794 attr->freq = 1;
795 evsel__set_sample_bit(evsel, PERIOD);
796 }
797 break;
798 case EVSEL__CONFIG_TERM_TIME:
799 if (term->val.time)
800 evsel__set_sample_bit(evsel, TIME);
801 else
802 evsel__reset_sample_bit(evsel, TIME);
803 break;
804 case EVSEL__CONFIG_TERM_CALLGRAPH:
805 callgraph_buf = term->val.str;
806 break;
807 case EVSEL__CONFIG_TERM_BRANCH:
808 if (term->val.str && strcmp(term->val.str, "no")) {
809 evsel__set_sample_bit(evsel, BRANCH_STACK);
810 parse_branch_str(term->val.str,
811 &attr->branch_sample_type);
812 } else
813 evsel__reset_sample_bit(evsel, BRANCH_STACK);
814 break;
815 case EVSEL__CONFIG_TERM_STACK_USER:
816 dump_size = term->val.stack_user;
817 break;
818 case EVSEL__CONFIG_TERM_MAX_STACK:
819 max_stack = term->val.max_stack;
820 break;
821 case EVSEL__CONFIG_TERM_MAX_EVENTS:
822 evsel->max_events = term->val.max_events;
823 break;
824 case EVSEL__CONFIG_TERM_INHERIT:
825 /*
826 * attr->inherit should has already been set by
827 * evsel__config. If user explicitly set
828 * inherit using config terms, override global
829 * opt->no_inherit setting.
830 */
831 attr->inherit = term->val.inherit ? 1 : 0;
832 break;
833 case EVSEL__CONFIG_TERM_OVERWRITE:
834 attr->write_backward = term->val.overwrite ? 1 : 0;
835 break;
836 case EVSEL__CONFIG_TERM_DRV_CFG:
837 break;
838 case EVSEL__CONFIG_TERM_PERCORE:
839 break;
840 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
841 attr->aux_output = term->val.aux_output ? 1 : 0;
842 break;
843 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
844 /* Already applied by auxtrace */
845 break;
846 case EVSEL__CONFIG_TERM_CFG_CHG:
847 break;
848 default:
849 break;
850 }
851 }
852
853 /* User explicitly set per-event callgraph, clear the old setting and reset. */
854 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
855 bool sample_address = false;
856
857 if (max_stack) {
858 param.max_stack = max_stack;
859 if (callgraph_buf == NULL)
860 callgraph_buf = "fp";
861 }
862
863 /* parse callgraph parameters */
864 if (callgraph_buf != NULL) {
865 if (!strcmp(callgraph_buf, "no")) {
866 param.enabled = false;
867 param.record_mode = CALLCHAIN_NONE;
868 } else {
869 param.enabled = true;
870 if (parse_callchain_record(callgraph_buf, &param)) {
871 pr_err("per-event callgraph setting for %s failed. "
872 "Apply callgraph global setting for it\n",
873 evsel->name);
874 return;
875 }
876 if (param.record_mode == CALLCHAIN_DWARF)
877 sample_address = true;
878 }
879 }
880 if (dump_size > 0) {
881 dump_size = round_up(dump_size, sizeof(u64));
882 param.dump_size = dump_size;
883 }
884
885 /* If global callgraph set, clear it */
886 if (callchain_param.enabled)
887 perf_evsel__reset_callgraph(evsel, &callchain_param);
888
889 /* set perf-event callgraph */
890 if (param.enabled) {
891 if (sample_address) {
892 evsel__set_sample_bit(evsel, ADDR);
893 evsel__set_sample_bit(evsel, DATA_SRC);
894 evsel->core.attr.mmap_data = track;
895 }
896 evsel__config_callchain(evsel, opts, &param);
897 }
898 }
899 }
900
901 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
902 {
903 struct evsel_config_term *term, *found_term = NULL;
904
905 list_for_each_entry(term, &evsel->config_terms, list) {
906 if (term->type == type)
907 found_term = term;
908 }
909
910 return found_term;
911 }
912
913 /*
914 * The enable_on_exec/disabled value strategy:
915 *
916 * 1) For any type of traced program:
917 * - all independent events and group leaders are disabled
918 * - all group members are enabled
919 *
920 * Group members are ruled by group leaders. They need to
921 * be enabled, because the group scheduling relies on that.
922 *
923 * 2) For traced programs executed by perf:
924 * - all independent events and group leaders have
925 * enable_on_exec set
926 * - we don't specifically enable or disable any event during
927 * the record command
928 *
929 * Independent events and group leaders are initially disabled
930 * and get enabled by exec. Group members are ruled by group
931 * leaders as stated in 1).
932 *
933 * 3) For traced programs attached by perf (pid/tid):
934 * - we specifically enable or disable all events during
935 * the record command
936 *
937 * When attaching events to already running traced we
938 * enable/disable events specifically, as there's no
939 * initial traced exec call.
940 */
941 void evsel__config(struct evsel *evsel, struct record_opts *opts,
942 struct callchain_param *callchain)
943 {
944 struct evsel *leader = evsel->leader;
945 struct perf_event_attr *attr = &evsel->core.attr;
946 int track = evsel->tracking;
947 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
948
949 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
950 attr->inherit = !opts->no_inherit;
951 attr->write_backward = opts->overwrite ? 1 : 0;
952
953 evsel__set_sample_bit(evsel, IP);
954 evsel__set_sample_bit(evsel, TID);
955
956 if (evsel->sample_read) {
957 evsel__set_sample_bit(evsel, READ);
958
959 /*
960 * We need ID even in case of single event, because
961 * PERF_SAMPLE_READ process ID specific data.
962 */
963 evsel__set_sample_id(evsel, false);
964
965 /*
966 * Apply group format only if we belong to group
967 * with more than one members.
968 */
969 if (leader->core.nr_members > 1) {
970 attr->read_format |= PERF_FORMAT_GROUP;
971 attr->inherit = 0;
972 }
973 }
974
975 /*
976 * We default some events to have a default interval. But keep
977 * it a weak assumption overridable by the user.
978 */
979 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
980 opts->user_interval != ULLONG_MAX)) {
981 if (opts->freq) {
982 evsel__set_sample_bit(evsel, PERIOD);
983 attr->freq = 1;
984 attr->sample_freq = opts->freq;
985 } else {
986 attr->sample_period = opts->default_interval;
987 }
988 }
989
990 if (opts->no_samples)
991 attr->sample_freq = 0;
992
993 if (opts->inherit_stat) {
994 evsel->core.attr.read_format |=
995 PERF_FORMAT_TOTAL_TIME_ENABLED |
996 PERF_FORMAT_TOTAL_TIME_RUNNING |
997 PERF_FORMAT_ID;
998 attr->inherit_stat = 1;
999 }
1000
1001 if (opts->sample_address) {
1002 evsel__set_sample_bit(evsel, ADDR);
1003 attr->mmap_data = track;
1004 }
1005
1006 /*
1007 * We don't allow user space callchains for function trace
1008 * event, due to issues with page faults while tracing page
1009 * fault handler and its overall trickiness nature.
1010 */
1011 if (evsel__is_function_event(evsel))
1012 evsel->core.attr.exclude_callchain_user = 1;
1013
1014 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1015 evsel__config_callchain(evsel, opts, callchain);
1016
1017 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1018 !evsel__is_dummy_event(evsel)) {
1019 attr->sample_regs_intr = opts->sample_intr_regs;
1020 evsel__set_sample_bit(evsel, REGS_INTR);
1021 }
1022
1023 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1024 !evsel__is_dummy_event(evsel)) {
1025 attr->sample_regs_user |= opts->sample_user_regs;
1026 evsel__set_sample_bit(evsel, REGS_USER);
1027 }
1028
1029 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1030 evsel__set_sample_bit(evsel, CPU);
1031
1032 /*
1033 * When the user explicitly disabled time don't force it here.
1034 */
1035 if (opts->sample_time &&
1036 (!perf_missing_features.sample_id_all &&
1037 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1038 opts->sample_time_set)))
1039 evsel__set_sample_bit(evsel, TIME);
1040
1041 if (opts->raw_samples && !evsel->no_aux_samples) {
1042 evsel__set_sample_bit(evsel, TIME);
1043 evsel__set_sample_bit(evsel, RAW);
1044 evsel__set_sample_bit(evsel, CPU);
1045 }
1046
1047 if (opts->sample_address)
1048 evsel__set_sample_bit(evsel, DATA_SRC);
1049
1050 if (opts->sample_phys_addr)
1051 evsel__set_sample_bit(evsel, PHYS_ADDR);
1052
1053 if (opts->no_buffering) {
1054 attr->watermark = 0;
1055 attr->wakeup_events = 1;
1056 }
1057 if (opts->branch_stack && !evsel->no_aux_samples) {
1058 evsel__set_sample_bit(evsel, BRANCH_STACK);
1059 attr->branch_sample_type = opts->branch_stack;
1060 }
1061
1062 if (opts->sample_weight)
1063 evsel__set_sample_bit(evsel, WEIGHT);
1064
1065 attr->task = track;
1066 attr->mmap = track;
1067 attr->mmap2 = track && !perf_missing_features.mmap2;
1068 attr->comm = track;
1069 /*
1070 * ksymbol is tracked separately with text poke because it needs to be
1071 * system wide and enabled immediately.
1072 */
1073 if (!opts->text_poke)
1074 attr->ksymbol = track && !perf_missing_features.ksymbol;
1075 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1076
1077 if (opts->record_namespaces)
1078 attr->namespaces = track;
1079
1080 if (opts->record_cgroup) {
1081 attr->cgroup = track && !perf_missing_features.cgroup;
1082 evsel__set_sample_bit(evsel, CGROUP);
1083 }
1084
1085 if (opts->record_switch_events)
1086 attr->context_switch = track;
1087
1088 if (opts->sample_transaction)
1089 evsel__set_sample_bit(evsel, TRANSACTION);
1090
1091 if (opts->running_time) {
1092 evsel->core.attr.read_format |=
1093 PERF_FORMAT_TOTAL_TIME_ENABLED |
1094 PERF_FORMAT_TOTAL_TIME_RUNNING;
1095 }
1096
1097 /*
1098 * XXX see the function comment above
1099 *
1100 * Disabling only independent events or group leaders,
1101 * keeping group members enabled.
1102 */
1103 if (evsel__is_group_leader(evsel))
1104 attr->disabled = 1;
1105
1106 /*
1107 * Setting enable_on_exec for independent events and
1108 * group leaders for traced executed by perf.
1109 */
1110 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1111 !opts->initial_delay)
1112 attr->enable_on_exec = 1;
1113
1114 if (evsel->immediate) {
1115 attr->disabled = 0;
1116 attr->enable_on_exec = 0;
1117 }
1118
1119 clockid = opts->clockid;
1120 if (opts->use_clockid) {
1121 attr->use_clockid = 1;
1122 attr->clockid = opts->clockid;
1123 }
1124
1125 if (evsel->precise_max)
1126 attr->precise_ip = 3;
1127
1128 if (opts->all_user) {
1129 attr->exclude_kernel = 1;
1130 attr->exclude_user = 0;
1131 }
1132
1133 if (opts->all_kernel) {
1134 attr->exclude_kernel = 0;
1135 attr->exclude_user = 1;
1136 }
1137
1138 if (evsel->core.own_cpus || evsel->unit)
1139 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1140
1141 /*
1142 * Apply event specific term settings,
1143 * it overloads any global configuration.
1144 */
1145 evsel__apply_config_terms(evsel, opts, track);
1146
1147 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1148
1149 /* The --period option takes the precedence. */
1150 if (opts->period_set) {
1151 if (opts->period)
1152 evsel__set_sample_bit(evsel, PERIOD);
1153 else
1154 evsel__reset_sample_bit(evsel, PERIOD);
1155 }
1156
1157 /*
1158 * A dummy event never triggers any actual counter and therefore
1159 * cannot be used with branch_stack.
1160 *
1161 * For initial_delay, a dummy event is added implicitly.
1162 * The software event will trigger -EOPNOTSUPP error out,
1163 * if BRANCH_STACK bit is set.
1164 */
1165 if (evsel__is_dummy_event(evsel))
1166 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1167 }
1168
1169 int evsel__set_filter(struct evsel *evsel, const char *filter)
1170 {
1171 char *new_filter = strdup(filter);
1172
1173 if (new_filter != NULL) {
1174 free(evsel->filter);
1175 evsel->filter = new_filter;
1176 return 0;
1177 }
1178
1179 return -1;
1180 }
1181
1182 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1183 {
1184 char *new_filter;
1185
1186 if (evsel->filter == NULL)
1187 return evsel__set_filter(evsel, filter);
1188
1189 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1190 free(evsel->filter);
1191 evsel->filter = new_filter;
1192 return 0;
1193 }
1194
1195 return -1;
1196 }
1197
1198 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1199 {
1200 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1201 }
1202
1203 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1204 {
1205 return evsel__append_filter(evsel, "%s,%s", filter);
1206 }
1207
1208 /* Caller has to clear disabled after going through all CPUs. */
1209 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1210 {
1211 return perf_evsel__enable_cpu(&evsel->core, cpu);
1212 }
1213
1214 int evsel__enable(struct evsel *evsel)
1215 {
1216 int err = perf_evsel__enable(&evsel->core);
1217
1218 if (!err)
1219 evsel->disabled = false;
1220 return err;
1221 }
1222
1223 /* Caller has to set disabled after going through all CPUs. */
1224 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1225 {
1226 return perf_evsel__disable_cpu(&evsel->core, cpu);
1227 }
1228
1229 int evsel__disable(struct evsel *evsel)
1230 {
1231 int err = perf_evsel__disable(&evsel->core);
1232 /*
1233 * We mark it disabled here so that tools that disable a event can
1234 * ignore events after they disable it. I.e. the ring buffer may have
1235 * already a few more events queued up before the kernel got the stop
1236 * request.
1237 */
1238 if (!err)
1239 evsel->disabled = true;
1240
1241 return err;
1242 }
1243
1244 static void evsel__free_config_terms(struct evsel *evsel)
1245 {
1246 struct evsel_config_term *term, *h;
1247
1248 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1249 list_del_init(&term->list);
1250 if (term->free_str)
1251 zfree(&term->val.str);
1252 free(term);
1253 }
1254 }
1255
1256 void evsel__exit(struct evsel *evsel)
1257 {
1258 assert(list_empty(&evsel->core.node));
1259 assert(evsel->evlist == NULL);
1260 evsel__free_counts(evsel);
1261 perf_evsel__free_fd(&evsel->core);
1262 perf_evsel__free_id(&evsel->core);
1263 evsel__free_config_terms(evsel);
1264 cgroup__put(evsel->cgrp);
1265 perf_cpu_map__put(evsel->core.cpus);
1266 perf_cpu_map__put(evsel->core.own_cpus);
1267 perf_thread_map__put(evsel->core.threads);
1268 zfree(&evsel->group_name);
1269 zfree(&evsel->name);
1270 zfree(&evsel->pmu_name);
1271 zfree(&evsel->per_pkg_mask);
1272 zfree(&evsel->metric_events);
1273 perf_evsel__object.fini(evsel);
1274 }
1275
1276 void evsel__delete(struct evsel *evsel)
1277 {
1278 evsel__exit(evsel);
1279 free(evsel);
1280 }
1281
1282 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1283 struct perf_counts_values *count)
1284 {
1285 struct perf_counts_values tmp;
1286
1287 if (!evsel->prev_raw_counts)
1288 return;
1289
1290 if (cpu == -1) {
1291 tmp = evsel->prev_raw_counts->aggr;
1292 evsel->prev_raw_counts->aggr = *count;
1293 } else {
1294 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1295 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1296 }
1297
1298 count->val = count->val - tmp.val;
1299 count->ena = count->ena - tmp.ena;
1300 count->run = count->run - tmp.run;
1301 }
1302
1303 void perf_counts_values__scale(struct perf_counts_values *count,
1304 bool scale, s8 *pscaled)
1305 {
1306 s8 scaled = 0;
1307
1308 if (scale) {
1309 if (count->run == 0) {
1310 scaled = -1;
1311 count->val = 0;
1312 } else if (count->run < count->ena) {
1313 scaled = 1;
1314 count->val = (u64)((double) count->val * count->ena / count->run);
1315 }
1316 }
1317
1318 if (pscaled)
1319 *pscaled = scaled;
1320 }
1321
1322 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1323 {
1324 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1325
1326 return perf_evsel__read(&evsel->core, cpu, thread, count);
1327 }
1328
1329 static void
1330 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1331 u64 val, u64 ena, u64 run)
1332 {
1333 struct perf_counts_values *count;
1334
1335 count = perf_counts(counter->counts, cpu, thread);
1336
1337 count->val = val;
1338 count->ena = ena;
1339 count->run = run;
1340
1341 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1342 }
1343
1344 static int
1345 perf_evsel__process_group_data(struct evsel *leader,
1346 int cpu, int thread, u64 *data)
1347 {
1348 u64 read_format = leader->core.attr.read_format;
1349 struct sample_read_value *v;
1350 u64 nr, ena = 0, run = 0, i;
1351
1352 nr = *data++;
1353
1354 if (nr != (u64) leader->core.nr_members)
1355 return -EINVAL;
1356
1357 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1358 ena = *data++;
1359
1360 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1361 run = *data++;
1362
1363 v = (struct sample_read_value *) data;
1364
1365 perf_evsel__set_count(leader, cpu, thread,
1366 v[0].value, ena, run);
1367
1368 for (i = 1; i < nr; i++) {
1369 struct evsel *counter;
1370
1371 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1372 if (!counter)
1373 return -EINVAL;
1374
1375 perf_evsel__set_count(counter, cpu, thread,
1376 v[i].value, ena, run);
1377 }
1378
1379 return 0;
1380 }
1381
1382 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1383 {
1384 struct perf_stat_evsel *ps = leader->stats;
1385 u64 read_format = leader->core.attr.read_format;
1386 int size = perf_evsel__read_size(&leader->core);
1387 u64 *data = ps->group_data;
1388
1389 if (!(read_format & PERF_FORMAT_ID))
1390 return -EINVAL;
1391
1392 if (!evsel__is_group_leader(leader))
1393 return -EINVAL;
1394
1395 if (!data) {
1396 data = zalloc(size);
1397 if (!data)
1398 return -ENOMEM;
1399
1400 ps->group_data = data;
1401 }
1402
1403 if (FD(leader, cpu, thread) < 0)
1404 return -EINVAL;
1405
1406 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1407 return -errno;
1408
1409 return perf_evsel__process_group_data(leader, cpu, thread, data);
1410 }
1411
1412 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1413 {
1414 u64 read_format = evsel->core.attr.read_format;
1415
1416 if (read_format & PERF_FORMAT_GROUP)
1417 return evsel__read_group(evsel, cpu, thread);
1418
1419 return evsel__read_one(evsel, cpu, thread);
1420 }
1421
1422 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1423 {
1424 struct perf_counts_values count;
1425 size_t nv = scale ? 3 : 1;
1426
1427 if (FD(evsel, cpu, thread) < 0)
1428 return -EINVAL;
1429
1430 if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1431 return -ENOMEM;
1432
1433 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1434 return -errno;
1435
1436 evsel__compute_deltas(evsel, cpu, thread, &count);
1437 perf_counts_values__scale(&count, scale, NULL);
1438 *perf_counts(evsel->counts, cpu, thread) = count;
1439 return 0;
1440 }
1441
1442 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1443 {
1444 struct evsel *leader = evsel->leader;
1445 int fd;
1446
1447 if (evsel__is_group_leader(evsel))
1448 return -1;
1449
1450 /*
1451 * Leader must be already processed/open,
1452 * if not it's a bug.
1453 */
1454 BUG_ON(!leader->core.fd);
1455
1456 fd = FD(leader, cpu, thread);
1457 BUG_ON(fd == -1);
1458
1459 return fd;
1460 }
1461
1462 static void perf_evsel__remove_fd(struct evsel *pos,
1463 int nr_cpus, int nr_threads,
1464 int thread_idx)
1465 {
1466 for (int cpu = 0; cpu < nr_cpus; cpu++)
1467 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1468 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1469 }
1470
1471 static int update_fds(struct evsel *evsel,
1472 int nr_cpus, int cpu_idx,
1473 int nr_threads, int thread_idx)
1474 {
1475 struct evsel *pos;
1476
1477 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1478 return -EINVAL;
1479
1480 evlist__for_each_entry(evsel->evlist, pos) {
1481 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1482
1483 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1484
1485 /*
1486 * Since fds for next evsel has not been created,
1487 * there is no need to iterate whole event list.
1488 */
1489 if (pos == evsel)
1490 break;
1491 }
1492 return 0;
1493 }
1494
1495 static bool ignore_missing_thread(struct evsel *evsel,
1496 int nr_cpus, int cpu,
1497 struct perf_thread_map *threads,
1498 int thread, int err)
1499 {
1500 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1501
1502 if (!evsel->ignore_missing_thread)
1503 return false;
1504
1505 /* The system wide setup does not work with threads. */
1506 if (evsel->core.system_wide)
1507 return false;
1508
1509 /* The -ESRCH is perf event syscall errno for pid's not found. */
1510 if (err != -ESRCH)
1511 return false;
1512
1513 /* If there's only one thread, let it fail. */
1514 if (threads->nr == 1)
1515 return false;
1516
1517 /*
1518 * We should remove fd for missing_thread first
1519 * because thread_map__remove() will decrease threads->nr.
1520 */
1521 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1522 return false;
1523
1524 if (thread_map__remove(threads, thread))
1525 return false;
1526
1527 pr_warning("WARNING: Ignored open failure for pid %d\n",
1528 ignore_pid);
1529 return true;
1530 }
1531
1532 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1533 void *priv __maybe_unused)
1534 {
1535 return fprintf(fp, " %-32s %s\n", name, val);
1536 }
1537
1538 static void display_attr(struct perf_event_attr *attr)
1539 {
1540 if (verbose >= 2 || debug_peo_args) {
1541 fprintf(stderr, "%.60s\n", graph_dotted_line);
1542 fprintf(stderr, "perf_event_attr:\n");
1543 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1544 fprintf(stderr, "%.60s\n", graph_dotted_line);
1545 }
1546 }
1547
1548 static int perf_event_open(struct evsel *evsel,
1549 pid_t pid, int cpu, int group_fd,
1550 unsigned long flags)
1551 {
1552 int precise_ip = evsel->core.attr.precise_ip;
1553 int fd;
1554
1555 while (1) {
1556 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1557 pid, cpu, group_fd, flags);
1558
1559 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1560 if (fd >= 0)
1561 break;
1562
1563 /* Do not try less precise if not requested. */
1564 if (!evsel->precise_max)
1565 break;
1566
1567 /*
1568 * We tried all the precise_ip values, and it's
1569 * still failing, so leave it to standard fallback.
1570 */
1571 if (!evsel->core.attr.precise_ip) {
1572 evsel->core.attr.precise_ip = precise_ip;
1573 break;
1574 }
1575
1576 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1577 evsel->core.attr.precise_ip--;
1578 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1579 display_attr(&evsel->core.attr);
1580 }
1581
1582 return fd;
1583 }
1584
1585 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1586 struct perf_thread_map *threads,
1587 int start_cpu, int end_cpu)
1588 {
1589 int cpu, thread, nthreads;
1590 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1591 int pid = -1, err, old_errno;
1592 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1593
1594 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1595 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1596 return -EINVAL;
1597
1598 if (cpus == NULL) {
1599 static struct perf_cpu_map *empty_cpu_map;
1600
1601 if (empty_cpu_map == NULL) {
1602 empty_cpu_map = perf_cpu_map__dummy_new();
1603 if (empty_cpu_map == NULL)
1604 return -ENOMEM;
1605 }
1606
1607 cpus = empty_cpu_map;
1608 }
1609
1610 if (threads == NULL) {
1611 static struct perf_thread_map *empty_thread_map;
1612
1613 if (empty_thread_map == NULL) {
1614 empty_thread_map = thread_map__new_by_tid(-1);
1615 if (empty_thread_map == NULL)
1616 return -ENOMEM;
1617 }
1618
1619 threads = empty_thread_map;
1620 }
1621
1622 if (evsel->core.system_wide)
1623 nthreads = 1;
1624 else
1625 nthreads = threads->nr;
1626
1627 if (evsel->core.fd == NULL &&
1628 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1629 return -ENOMEM;
1630
1631 if (evsel->cgrp) {
1632 flags |= PERF_FLAG_PID_CGROUP;
1633 pid = evsel->cgrp->fd;
1634 }
1635
1636 fallback_missing_features:
1637 if (perf_missing_features.clockid_wrong)
1638 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1639 if (perf_missing_features.clockid) {
1640 evsel->core.attr.use_clockid = 0;
1641 evsel->core.attr.clockid = 0;
1642 }
1643 if (perf_missing_features.cloexec)
1644 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1645 if (perf_missing_features.mmap2)
1646 evsel->core.attr.mmap2 = 0;
1647 if (perf_missing_features.exclude_guest)
1648 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1649 if (perf_missing_features.lbr_flags)
1650 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1651 PERF_SAMPLE_BRANCH_NO_CYCLES);
1652 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1653 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1654 if (perf_missing_features.ksymbol)
1655 evsel->core.attr.ksymbol = 0;
1656 if (perf_missing_features.bpf)
1657 evsel->core.attr.bpf_event = 0;
1658 if (perf_missing_features.branch_hw_idx)
1659 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1660 retry_sample_id:
1661 if (perf_missing_features.sample_id_all)
1662 evsel->core.attr.sample_id_all = 0;
1663
1664 display_attr(&evsel->core.attr);
1665
1666 for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1667
1668 for (thread = 0; thread < nthreads; thread++) {
1669 int fd, group_fd;
1670
1671 if (!evsel->cgrp && !evsel->core.system_wide)
1672 pid = perf_thread_map__pid(threads, thread);
1673
1674 group_fd = get_group_fd(evsel, cpu, thread);
1675 retry_open:
1676 test_attr__ready();
1677
1678 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1679 group_fd, flags);
1680
1681 FD(evsel, cpu, thread) = fd;
1682
1683 if (fd < 0) {
1684 err = -errno;
1685
1686 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1687 /*
1688 * We just removed 1 thread, so take a step
1689 * back on thread index and lower the upper
1690 * nthreads limit.
1691 */
1692 nthreads--;
1693 thread--;
1694
1695 /* ... and pretend like nothing have happened. */
1696 err = 0;
1697 continue;
1698 }
1699
1700 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1701 err);
1702 goto try_fallback;
1703 }
1704
1705 pr_debug2_peo(" = %d\n", fd);
1706
1707 if (evsel->bpf_fd >= 0) {
1708 int evt_fd = fd;
1709 int bpf_fd = evsel->bpf_fd;
1710
1711 err = ioctl(evt_fd,
1712 PERF_EVENT_IOC_SET_BPF,
1713 bpf_fd);
1714 if (err && errno != EEXIST) {
1715 pr_err("failed to attach bpf fd %d: %s\n",
1716 bpf_fd, strerror(errno));
1717 err = -EINVAL;
1718 goto out_close;
1719 }
1720 }
1721
1722 set_rlimit = NO_CHANGE;
1723
1724 /*
1725 * If we succeeded but had to kill clockid, fail and
1726 * have evsel__open_strerror() print us a nice error.
1727 */
1728 if (perf_missing_features.clockid ||
1729 perf_missing_features.clockid_wrong) {
1730 err = -EINVAL;
1731 goto out_close;
1732 }
1733 }
1734 }
1735
1736 return 0;
1737
1738 try_fallback:
1739 /*
1740 * perf stat needs between 5 and 22 fds per CPU. When we run out
1741 * of them try to increase the limits.
1742 */
1743 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1744 struct rlimit l;
1745
1746 old_errno = errno;
1747 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1748 if (set_rlimit == NO_CHANGE)
1749 l.rlim_cur = l.rlim_max;
1750 else {
1751 l.rlim_cur = l.rlim_max + 1000;
1752 l.rlim_max = l.rlim_cur;
1753 }
1754 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1755 set_rlimit++;
1756 errno = old_errno;
1757 goto retry_open;
1758 }
1759 }
1760 errno = old_errno;
1761 }
1762
1763 if (err != -EINVAL || cpu > 0 || thread > 0)
1764 goto out_close;
1765
1766 /*
1767 * Must probe features in the order they were added to the
1768 * perf_event_attr interface.
1769 */
1770 if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1771 perf_missing_features.cgroup = true;
1772 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1773 goto out_close;
1774 } else if (!perf_missing_features.branch_hw_idx &&
1775 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1776 perf_missing_features.branch_hw_idx = true;
1777 pr_debug2("switching off branch HW index support\n");
1778 goto fallback_missing_features;
1779 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1780 perf_missing_features.aux_output = true;
1781 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1782 goto out_close;
1783 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1784 perf_missing_features.bpf = true;
1785 pr_debug2_peo("switching off bpf_event\n");
1786 goto fallback_missing_features;
1787 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1788 perf_missing_features.ksymbol = true;
1789 pr_debug2_peo("switching off ksymbol\n");
1790 goto fallback_missing_features;
1791 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1792 perf_missing_features.write_backward = true;
1793 pr_debug2_peo("switching off write_backward\n");
1794 goto out_close;
1795 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1796 perf_missing_features.clockid_wrong = true;
1797 pr_debug2_peo("switching off clockid\n");
1798 goto fallback_missing_features;
1799 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1800 perf_missing_features.clockid = true;
1801 pr_debug2_peo("switching off use_clockid\n");
1802 goto fallback_missing_features;
1803 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1804 perf_missing_features.cloexec = true;
1805 pr_debug2_peo("switching off cloexec flag\n");
1806 goto fallback_missing_features;
1807 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1808 perf_missing_features.mmap2 = true;
1809 pr_debug2_peo("switching off mmap2\n");
1810 goto fallback_missing_features;
1811 } else if (!perf_missing_features.exclude_guest &&
1812 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1813 perf_missing_features.exclude_guest = true;
1814 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1815 goto fallback_missing_features;
1816 } else if (!perf_missing_features.sample_id_all) {
1817 perf_missing_features.sample_id_all = true;
1818 pr_debug2_peo("switching off sample_id_all\n");
1819 goto retry_sample_id;
1820 } else if (!perf_missing_features.lbr_flags &&
1821 (evsel->core.attr.branch_sample_type &
1822 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1823 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1824 perf_missing_features.lbr_flags = true;
1825 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1826 goto fallback_missing_features;
1827 } else if (!perf_missing_features.group_read &&
1828 evsel->core.attr.inherit &&
1829 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1830 evsel__is_group_leader(evsel)) {
1831 perf_missing_features.group_read = true;
1832 pr_debug2_peo("switching off group read\n");
1833 goto fallback_missing_features;
1834 }
1835 out_close:
1836 if (err)
1837 threads->err_thread = thread;
1838
1839 old_errno = errno;
1840 do {
1841 while (--thread >= 0) {
1842 if (FD(evsel, cpu, thread) >= 0)
1843 close(FD(evsel, cpu, thread));
1844 FD(evsel, cpu, thread) = -1;
1845 }
1846 thread = nthreads;
1847 } while (--cpu >= 0);
1848 errno = old_errno;
1849 return err;
1850 }
1851
1852 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1853 struct perf_thread_map *threads)
1854 {
1855 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1856 }
1857
1858 void evsel__close(struct evsel *evsel)
1859 {
1860 perf_evsel__close(&evsel->core);
1861 perf_evsel__free_id(&evsel->core);
1862 }
1863
1864 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
1865 {
1866 if (cpu == -1)
1867 return evsel__open_cpu(evsel, cpus, NULL, 0,
1868 cpus ? cpus->nr : 1);
1869
1870 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1871 }
1872
1873 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
1874 {
1875 return evsel__open(evsel, NULL, threads);
1876 }
1877
1878 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1879 const union perf_event *event,
1880 struct perf_sample *sample)
1881 {
1882 u64 type = evsel->core.attr.sample_type;
1883 const __u64 *array = event->sample.array;
1884 bool swapped = evsel->needs_swap;
1885 union u64_swap u;
1886
1887 array += ((event->header.size -
1888 sizeof(event->header)) / sizeof(u64)) - 1;
1889
1890 if (type & PERF_SAMPLE_IDENTIFIER) {
1891 sample->id = *array;
1892 array--;
1893 }
1894
1895 if (type & PERF_SAMPLE_CPU) {
1896 u.val64 = *array;
1897 if (swapped) {
1898 /* undo swap of u64, then swap on individual u32s */
1899 u.val64 = bswap_64(u.val64);
1900 u.val32[0] = bswap_32(u.val32[0]);
1901 }
1902
1903 sample->cpu = u.val32[0];
1904 array--;
1905 }
1906
1907 if (type & PERF_SAMPLE_STREAM_ID) {
1908 sample->stream_id = *array;
1909 array--;
1910 }
1911
1912 if (type & PERF_SAMPLE_ID) {
1913 sample->id = *array;
1914 array--;
1915 }
1916
1917 if (type & PERF_SAMPLE_TIME) {
1918 sample->time = *array;
1919 array--;
1920 }
1921
1922 if (type & PERF_SAMPLE_TID) {
1923 u.val64 = *array;
1924 if (swapped) {
1925 /* undo swap of u64, then swap on individual u32s */
1926 u.val64 = bswap_64(u.val64);
1927 u.val32[0] = bswap_32(u.val32[0]);
1928 u.val32[1] = bswap_32(u.val32[1]);
1929 }
1930
1931 sample->pid = u.val32[0];
1932 sample->tid = u.val32[1];
1933 array--;
1934 }
1935
1936 return 0;
1937 }
1938
1939 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1940 u64 size)
1941 {
1942 return size > max_size || offset + size > endp;
1943 }
1944
1945 #define OVERFLOW_CHECK(offset, size, max_size) \
1946 do { \
1947 if (overflow(endp, (max_size), (offset), (size))) \
1948 return -EFAULT; \
1949 } while (0)
1950
1951 #define OVERFLOW_CHECK_u64(offset) \
1952 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1953
1954 static int
1955 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1956 {
1957 /*
1958 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1959 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1960 * check the format does not go past the end of the event.
1961 */
1962 if (sample_size + sizeof(event->header) > event->header.size)
1963 return -EFAULT;
1964
1965 return 0;
1966 }
1967
1968 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1969 struct perf_sample *data)
1970 {
1971 u64 type = evsel->core.attr.sample_type;
1972 bool swapped = evsel->needs_swap;
1973 const __u64 *array;
1974 u16 max_size = event->header.size;
1975 const void *endp = (void *)event + max_size;
1976 u64 sz;
1977
1978 /*
1979 * used for cross-endian analysis. See git commit 65014ab3
1980 * for why this goofiness is needed.
1981 */
1982 union u64_swap u;
1983
1984 memset(data, 0, sizeof(*data));
1985 data->cpu = data->pid = data->tid = -1;
1986 data->stream_id = data->id = data->time = -1ULL;
1987 data->period = evsel->core.attr.sample_period;
1988 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1989 data->misc = event->header.misc;
1990 data->id = -1ULL;
1991 data->data_src = PERF_MEM_DATA_SRC_NONE;
1992
1993 if (event->header.type != PERF_RECORD_SAMPLE) {
1994 if (!evsel->core.attr.sample_id_all)
1995 return 0;
1996 return perf_evsel__parse_id_sample(evsel, event, data);
1997 }
1998
1999 array = event->sample.array;
2000
2001 if (perf_event__check_size(event, evsel->sample_size))
2002 return -EFAULT;
2003
2004 if (type & PERF_SAMPLE_IDENTIFIER) {
2005 data->id = *array;
2006 array++;
2007 }
2008
2009 if (type & PERF_SAMPLE_IP) {
2010 data->ip = *array;
2011 array++;
2012 }
2013
2014 if (type & PERF_SAMPLE_TID) {
2015 u.val64 = *array;
2016 if (swapped) {
2017 /* undo swap of u64, then swap on individual u32s */
2018 u.val64 = bswap_64(u.val64);
2019 u.val32[0] = bswap_32(u.val32[0]);
2020 u.val32[1] = bswap_32(u.val32[1]);
2021 }
2022
2023 data->pid = u.val32[0];
2024 data->tid = u.val32[1];
2025 array++;
2026 }
2027
2028 if (type & PERF_SAMPLE_TIME) {
2029 data->time = *array;
2030 array++;
2031 }
2032
2033 if (type & PERF_SAMPLE_ADDR) {
2034 data->addr = *array;
2035 array++;
2036 }
2037
2038 if (type & PERF_SAMPLE_ID) {
2039 data->id = *array;
2040 array++;
2041 }
2042
2043 if (type & PERF_SAMPLE_STREAM_ID) {
2044 data->stream_id = *array;
2045 array++;
2046 }
2047
2048 if (type & PERF_SAMPLE_CPU) {
2049
2050 u.val64 = *array;
2051 if (swapped) {
2052 /* undo swap of u64, then swap on individual u32s */
2053 u.val64 = bswap_64(u.val64);
2054 u.val32[0] = bswap_32(u.val32[0]);
2055 }
2056
2057 data->cpu = u.val32[0];
2058 array++;
2059 }
2060
2061 if (type & PERF_SAMPLE_PERIOD) {
2062 data->period = *array;
2063 array++;
2064 }
2065
2066 if (type & PERF_SAMPLE_READ) {
2067 u64 read_format = evsel->core.attr.read_format;
2068
2069 OVERFLOW_CHECK_u64(array);
2070 if (read_format & PERF_FORMAT_GROUP)
2071 data->read.group.nr = *array;
2072 else
2073 data->read.one.value = *array;
2074
2075 array++;
2076
2077 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2078 OVERFLOW_CHECK_u64(array);
2079 data->read.time_enabled = *array;
2080 array++;
2081 }
2082
2083 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2084 OVERFLOW_CHECK_u64(array);
2085 data->read.time_running = *array;
2086 array++;
2087 }
2088
2089 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2090 if (read_format & PERF_FORMAT_GROUP) {
2091 const u64 max_group_nr = UINT64_MAX /
2092 sizeof(struct sample_read_value);
2093
2094 if (data->read.group.nr > max_group_nr)
2095 return -EFAULT;
2096 sz = data->read.group.nr *
2097 sizeof(struct sample_read_value);
2098 OVERFLOW_CHECK(array, sz, max_size);
2099 data->read.group.values =
2100 (struct sample_read_value *)array;
2101 array = (void *)array + sz;
2102 } else {
2103 OVERFLOW_CHECK_u64(array);
2104 data->read.one.id = *array;
2105 array++;
2106 }
2107 }
2108
2109 if (type & PERF_SAMPLE_CALLCHAIN) {
2110 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2111
2112 OVERFLOW_CHECK_u64(array);
2113 data->callchain = (struct ip_callchain *)array++;
2114 if (data->callchain->nr > max_callchain_nr)
2115 return -EFAULT;
2116 sz = data->callchain->nr * sizeof(u64);
2117 OVERFLOW_CHECK(array, sz, max_size);
2118 array = (void *)array + sz;
2119 }
2120
2121 if (type & PERF_SAMPLE_RAW) {
2122 OVERFLOW_CHECK_u64(array);
2123 u.val64 = *array;
2124
2125 /*
2126 * Undo swap of u64, then swap on individual u32s,
2127 * get the size of the raw area and undo all of the
2128 * swap. The pevent interface handles endianity by
2129 * itself.
2130 */
2131 if (swapped) {
2132 u.val64 = bswap_64(u.val64);
2133 u.val32[0] = bswap_32(u.val32[0]);
2134 u.val32[1] = bswap_32(u.val32[1]);
2135 }
2136 data->raw_size = u.val32[0];
2137
2138 /*
2139 * The raw data is aligned on 64bits including the
2140 * u32 size, so it's safe to use mem_bswap_64.
2141 */
2142 if (swapped)
2143 mem_bswap_64((void *) array, data->raw_size);
2144
2145 array = (void *)array + sizeof(u32);
2146
2147 OVERFLOW_CHECK(array, data->raw_size, max_size);
2148 data->raw_data = (void *)array;
2149 array = (void *)array + data->raw_size;
2150 }
2151
2152 if (type & PERF_SAMPLE_BRANCH_STACK) {
2153 const u64 max_branch_nr = UINT64_MAX /
2154 sizeof(struct branch_entry);
2155
2156 OVERFLOW_CHECK_u64(array);
2157 data->branch_stack = (struct branch_stack *)array++;
2158
2159 if (data->branch_stack->nr > max_branch_nr)
2160 return -EFAULT;
2161
2162 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2163 if (evsel__has_branch_hw_idx(evsel))
2164 sz += sizeof(u64);
2165 else
2166 data->no_hw_idx = true;
2167 OVERFLOW_CHECK(array, sz, max_size);
2168 array = (void *)array + sz;
2169 }
2170
2171 if (type & PERF_SAMPLE_REGS_USER) {
2172 OVERFLOW_CHECK_u64(array);
2173 data->user_regs.abi = *array;
2174 array++;
2175
2176 if (data->user_regs.abi) {
2177 u64 mask = evsel->core.attr.sample_regs_user;
2178
2179 sz = hweight64(mask) * sizeof(u64);
2180 OVERFLOW_CHECK(array, sz, max_size);
2181 data->user_regs.mask = mask;
2182 data->user_regs.regs = (u64 *)array;
2183 array = (void *)array + sz;
2184 }
2185 }
2186
2187 if (type & PERF_SAMPLE_STACK_USER) {
2188 OVERFLOW_CHECK_u64(array);
2189 sz = *array++;
2190
2191 data->user_stack.offset = ((char *)(array - 1)
2192 - (char *) event);
2193
2194 if (!sz) {
2195 data->user_stack.size = 0;
2196 } else {
2197 OVERFLOW_CHECK(array, sz, max_size);
2198 data->user_stack.data = (char *)array;
2199 array = (void *)array + sz;
2200 OVERFLOW_CHECK_u64(array);
2201 data->user_stack.size = *array++;
2202 if (WARN_ONCE(data->user_stack.size > sz,
2203 "user stack dump failure\n"))
2204 return -EFAULT;
2205 }
2206 }
2207
2208 if (type & PERF_SAMPLE_WEIGHT) {
2209 OVERFLOW_CHECK_u64(array);
2210 data->weight = *array;
2211 array++;
2212 }
2213
2214 if (type & PERF_SAMPLE_DATA_SRC) {
2215 OVERFLOW_CHECK_u64(array);
2216 data->data_src = *array;
2217 array++;
2218 }
2219
2220 if (type & PERF_SAMPLE_TRANSACTION) {
2221 OVERFLOW_CHECK_u64(array);
2222 data->transaction = *array;
2223 array++;
2224 }
2225
2226 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2227 if (type & PERF_SAMPLE_REGS_INTR) {
2228 OVERFLOW_CHECK_u64(array);
2229 data->intr_regs.abi = *array;
2230 array++;
2231
2232 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2233 u64 mask = evsel->core.attr.sample_regs_intr;
2234
2235 sz = hweight64(mask) * sizeof(u64);
2236 OVERFLOW_CHECK(array, sz, max_size);
2237 data->intr_regs.mask = mask;
2238 data->intr_regs.regs = (u64 *)array;
2239 array = (void *)array + sz;
2240 }
2241 }
2242
2243 data->phys_addr = 0;
2244 if (type & PERF_SAMPLE_PHYS_ADDR) {
2245 data->phys_addr = *array;
2246 array++;
2247 }
2248
2249 data->cgroup = 0;
2250 if (type & PERF_SAMPLE_CGROUP) {
2251 data->cgroup = *array;
2252 array++;
2253 }
2254
2255 if (type & PERF_SAMPLE_AUX) {
2256 OVERFLOW_CHECK_u64(array);
2257 sz = *array++;
2258
2259 OVERFLOW_CHECK(array, sz, max_size);
2260 /* Undo swap of data */
2261 if (swapped)
2262 mem_bswap_64((char *)array, sz);
2263 data->aux_sample.size = sz;
2264 data->aux_sample.data = (char *)array;
2265 array = (void *)array + sz;
2266 }
2267
2268 return 0;
2269 }
2270
2271 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2272 u64 *timestamp)
2273 {
2274 u64 type = evsel->core.attr.sample_type;
2275 const __u64 *array;
2276
2277 if (!(type & PERF_SAMPLE_TIME))
2278 return -1;
2279
2280 if (event->header.type != PERF_RECORD_SAMPLE) {
2281 struct perf_sample data = {
2282 .time = -1ULL,
2283 };
2284
2285 if (!evsel->core.attr.sample_id_all)
2286 return -1;
2287 if (perf_evsel__parse_id_sample(evsel, event, &data))
2288 return -1;
2289
2290 *timestamp = data.time;
2291 return 0;
2292 }
2293
2294 array = event->sample.array;
2295
2296 if (perf_event__check_size(event, evsel->sample_size))
2297 return -EFAULT;
2298
2299 if (type & PERF_SAMPLE_IDENTIFIER)
2300 array++;
2301
2302 if (type & PERF_SAMPLE_IP)
2303 array++;
2304
2305 if (type & PERF_SAMPLE_TID)
2306 array++;
2307
2308 if (type & PERF_SAMPLE_TIME)
2309 *timestamp = *array;
2310
2311 return 0;
2312 }
2313
2314 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2315 {
2316 return tep_find_field(evsel->tp_format, name);
2317 }
2318
2319 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2320 {
2321 struct tep_format_field *field = evsel__field(evsel, name);
2322 int offset;
2323
2324 if (!field)
2325 return NULL;
2326
2327 offset = field->offset;
2328
2329 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2330 offset = *(int *)(sample->raw_data + field->offset);
2331 offset &= 0xffff;
2332 }
2333
2334 return sample->raw_data + offset;
2335 }
2336
2337 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2338 bool needs_swap)
2339 {
2340 u64 value;
2341 void *ptr = sample->raw_data + field->offset;
2342
2343 switch (field->size) {
2344 case 1:
2345 return *(u8 *)ptr;
2346 case 2:
2347 value = *(u16 *)ptr;
2348 break;
2349 case 4:
2350 value = *(u32 *)ptr;
2351 break;
2352 case 8:
2353 memcpy(&value, ptr, sizeof(u64));
2354 break;
2355 default:
2356 return 0;
2357 }
2358
2359 if (!needs_swap)
2360 return value;
2361
2362 switch (field->size) {
2363 case 2:
2364 return bswap_16(value);
2365 case 4:
2366 return bswap_32(value);
2367 case 8:
2368 return bswap_64(value);
2369 default:
2370 return 0;
2371 }
2372
2373 return 0;
2374 }
2375
2376 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2377 {
2378 struct tep_format_field *field = evsel__field(evsel, name);
2379
2380 if (!field)
2381 return 0;
2382
2383 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2384 }
2385
2386 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2387 {
2388 int paranoid;
2389
2390 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2391 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2392 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2393 /*
2394 * If it's cycles then fall back to hrtimer based
2395 * cpu-clock-tick sw counter, which is always available even if
2396 * no PMU support.
2397 *
2398 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2399 * b0a873e).
2400 */
2401 scnprintf(msg, msgsize, "%s",
2402 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2403
2404 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2405 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2406
2407 zfree(&evsel->name);
2408 return true;
2409 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2410 (paranoid = perf_event_paranoid()) > 1) {
2411 const char *name = evsel__name(evsel);
2412 char *new_name;
2413 const char *sep = ":";
2414
2415 /* If event has exclude user then don't exclude kernel. */
2416 if (evsel->core.attr.exclude_user)
2417 return false;
2418
2419 /* Is there already the separator in the name. */
2420 if (strchr(name, '/') ||
2421 (strchr(name, ':') && !evsel->is_libpfm_event))
2422 sep = "";
2423
2424 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2425 return false;
2426
2427 if (evsel->name)
2428 free(evsel->name);
2429 evsel->name = new_name;
2430 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2431 "to fall back to excluding kernel and hypervisor "
2432 " samples", paranoid);
2433 evsel->core.attr.exclude_kernel = 1;
2434 evsel->core.attr.exclude_hv = 1;
2435
2436 return true;
2437 }
2438
2439 return false;
2440 }
2441
2442 static bool find_process(const char *name)
2443 {
2444 size_t len = strlen(name);
2445 DIR *dir;
2446 struct dirent *d;
2447 int ret = -1;
2448
2449 dir = opendir(procfs__mountpoint());
2450 if (!dir)
2451 return false;
2452
2453 /* Walk through the directory. */
2454 while (ret && (d = readdir(dir)) != NULL) {
2455 char path[PATH_MAX];
2456 char *data;
2457 size_t size;
2458
2459 if ((d->d_type != DT_DIR) ||
2460 !strcmp(".", d->d_name) ||
2461 !strcmp("..", d->d_name))
2462 continue;
2463
2464 scnprintf(path, sizeof(path), "%s/%s/comm",
2465 procfs__mountpoint(), d->d_name);
2466
2467 if (filename__read_str(path, &data, &size))
2468 continue;
2469
2470 ret = strncmp(name, data, len);
2471 free(data);
2472 }
2473
2474 closedir(dir);
2475 return ret ? false : true;
2476 }
2477
2478 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2479 int err, char *msg, size_t size)
2480 {
2481 char sbuf[STRERR_BUFSIZE];
2482 int printed = 0, enforced = 0;
2483
2484 switch (err) {
2485 case EPERM:
2486 case EACCES:
2487 printed += scnprintf(msg + printed, size - printed,
2488 "Access to performance monitoring and observability operations is limited.\n");
2489
2490 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2491 if (enforced) {
2492 printed += scnprintf(msg + printed, size - printed,
2493 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2494 "monitoring and observability operations. Inspect system audit records for\n"
2495 "more perf_event access control information and adjusting the policy.\n");
2496 }
2497 }
2498
2499 if (err == EPERM)
2500 printed += scnprintf(msg, size,
2501 "No permission to enable %s event.\n\n", evsel__name(evsel));
2502
2503 return scnprintf(msg + printed, size - printed,
2504 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2505 "access to performance monitoring and observability operations for processes\n"
2506 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2507 "More information can be found at 'Perf events and tool security' document:\n"
2508 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2509 "perf_event_paranoid setting is %d:\n"
2510 " -1: Allow use of (almost) all events by all users\n"
2511 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2512 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2513 ">= 1: Disallow CPU event access\n"
2514 ">= 2: Disallow kernel profiling\n"
2515 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2516 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2517 perf_event_paranoid());
2518 case ENOENT:
2519 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2520 case EMFILE:
2521 return scnprintf(msg, size, "%s",
2522 "Too many events are opened.\n"
2523 "Probably the maximum number of open file descriptors has been reached.\n"
2524 "Hint: Try again after reducing the number of events.\n"
2525 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2526 case ENOMEM:
2527 if (evsel__has_callchain(evsel) &&
2528 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2529 return scnprintf(msg, size,
2530 "Not enough memory to setup event with callchain.\n"
2531 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2532 "Hint: Current value: %d", sysctl__max_stack());
2533 break;
2534 case ENODEV:
2535 if (target->cpu_list)
2536 return scnprintf(msg, size, "%s",
2537 "No such device - did you specify an out-of-range profile CPU?");
2538 break;
2539 case EOPNOTSUPP:
2540 if (evsel->core.attr.aux_output)
2541 return scnprintf(msg, size,
2542 "%s: PMU Hardware doesn't support 'aux_output' feature",
2543 evsel__name(evsel));
2544 if (evsel->core.attr.sample_period != 0)
2545 return scnprintf(msg, size,
2546 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2547 evsel__name(evsel));
2548 if (evsel->core.attr.precise_ip)
2549 return scnprintf(msg, size, "%s",
2550 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2551 #if defined(__i386__) || defined(__x86_64__)
2552 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2553 return scnprintf(msg, size, "%s",
2554 "No hardware sampling interrupt available.\n");
2555 #endif
2556 break;
2557 case EBUSY:
2558 if (find_process("oprofiled"))
2559 return scnprintf(msg, size,
2560 "The PMU counters are busy/taken by another profiler.\n"
2561 "We found oprofile daemon running, please stop it and try again.");
2562 break;
2563 case EINVAL:
2564 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2565 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2566 if (perf_missing_features.clockid)
2567 return scnprintf(msg, size, "clockid feature not supported.");
2568 if (perf_missing_features.clockid_wrong)
2569 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2570 if (perf_missing_features.aux_output)
2571 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2572 break;
2573 default:
2574 break;
2575 }
2576
2577 return scnprintf(msg, size,
2578 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2579 "/bin/dmesg | grep -i perf may provide additional information.\n",
2580 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2581 }
2582
2583 struct perf_env *evsel__env(struct evsel *evsel)
2584 {
2585 if (evsel && evsel->evlist)
2586 return evsel->evlist->env;
2587 return &perf_env;
2588 }
2589
2590 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2591 {
2592 int cpu, thread;
2593
2594 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2595 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2596 thread++) {
2597 int fd = FD(evsel, cpu, thread);
2598
2599 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2600 cpu, thread, fd) < 0)
2601 return -1;
2602 }
2603 }
2604
2605 return 0;
2606 }
2607
2608 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2609 {
2610 struct perf_cpu_map *cpus = evsel->core.cpus;
2611 struct perf_thread_map *threads = evsel->core.threads;
2612
2613 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2614 return -ENOMEM;
2615
2616 return store_evsel_ids(evsel, evlist);
2617 }
Linux with added FPC-III support