?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf intel-pt: Add lookahead callback
[linux/fpc-iii.git] / tools / perf / util / evsel.c
blob68beef8f47ffa9ea0ebc02a38d3d4706df420f56
1 /*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9
10 #include <byteswap.h>
11 #include <errno.h>
12 #include <inttypes.h>
13 #include <linux/bitops.h>
14 #include <api/fs/fs.h>
15 #include <api/fs/tracing_path.h>
16 #include <traceevent/event-parse.h>
17 #include <linux/hw_breakpoint.h>
18 #include <linux/perf_event.h>
19 #include <linux/compiler.h>
20 #include <linux/err.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
24 #include <dirent.h>
25 #include "asm/bug.h"
26 #include "callchain.h"
27 #include "cgroup.h"
28 #include "event.h"
29 #include "evsel.h"
30 #include "evlist.h"
31 #include "util.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 "memswap.h"
40 #include "util/parse-branch-options.h"
41
42 #include "sane_ctype.h"
43
44 struct perf_missing_features perf_missing_features;
45
46 static clockid_t clockid;
47
48 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
49 {
50 return 0;
51 }
52
53 void __weak test_attr__ready(void) { }
54
55 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
56 {
57 }
58
59 static struct {
60 size_t size;
61 int (*init)(struct perf_evsel *evsel);
62 void (*fini)(struct perf_evsel *evsel);
63 } perf_evsel__object = {
64 .size = sizeof(struct perf_evsel),
65 .init = perf_evsel__no_extra_init,
66 .fini = perf_evsel__no_extra_fini,
67 };
68
69 int perf_evsel__object_config(size_t object_size,
70 int (*init)(struct perf_evsel *evsel),
71 void (*fini)(struct perf_evsel *evsel))
72 {
73
74 if (object_size == 0)
75 goto set_methods;
76
77 if (perf_evsel__object.size > object_size)
78 return -EINVAL;
79
80 perf_evsel__object.size = object_size;
81
82 set_methods:
83 if (init != NULL)
84 perf_evsel__object.init = init;
85
86 if (fini != NULL)
87 perf_evsel__object.fini = fini;
88
89 return 0;
90 }
91
92 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
93
94 int __perf_evsel__sample_size(u64 sample_type)
95 {
96 u64 mask = sample_type & PERF_SAMPLE_MASK;
97 int size = 0;
98 int i;
99
100 for (i = 0; i < 64; i++) {
101 if (mask & (1ULL << i))
102 size++;
103 }
104
105 size *= sizeof(u64);
106
107 return size;
108 }
109
110 /**
111 * __perf_evsel__calc_id_pos - calculate id_pos.
112 * @sample_type: sample type
113 *
114 * This function returns the position of the event id (PERF_SAMPLE_ID or
115 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
116 * sample_event.
117 */
118 static int __perf_evsel__calc_id_pos(u64 sample_type)
119 {
120 int idx = 0;
121
122 if (sample_type & PERF_SAMPLE_IDENTIFIER)
123 return 0;
124
125 if (!(sample_type & PERF_SAMPLE_ID))
126 return -1;
127
128 if (sample_type & PERF_SAMPLE_IP)
129 idx += 1;
130
131 if (sample_type & PERF_SAMPLE_TID)
132 idx += 1;
133
134 if (sample_type & PERF_SAMPLE_TIME)
135 idx += 1;
136
137 if (sample_type & PERF_SAMPLE_ADDR)
138 idx += 1;
139
140 return idx;
141 }
142
143 /**
144 * __perf_evsel__calc_is_pos - calculate is_pos.
145 * @sample_type: sample type
146 *
147 * This function returns the position (counting backwards) of the event id
148 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
149 * sample_id_all is used there is an id sample appended to non-sample events.
150 */
151 static int __perf_evsel__calc_is_pos(u64 sample_type)
152 {
153 int idx = 1;
154
155 if (sample_type & PERF_SAMPLE_IDENTIFIER)
156 return 1;
157
158 if (!(sample_type & PERF_SAMPLE_ID))
159 return -1;
160
161 if (sample_type & PERF_SAMPLE_CPU)
162 idx += 1;
163
164 if (sample_type & PERF_SAMPLE_STREAM_ID)
165 idx += 1;
166
167 return idx;
168 }
169
170 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
171 {
172 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
173 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
174 }
175
176 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
177 enum perf_event_sample_format bit)
178 {
179 if (!(evsel->attr.sample_type & bit)) {
180 evsel->attr.sample_type |= bit;
181 evsel->sample_size += sizeof(u64);
182 perf_evsel__calc_id_pos(evsel);
183 }
184 }
185
186 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
187 enum perf_event_sample_format bit)
188 {
189 if (evsel->attr.sample_type & bit) {
190 evsel->attr.sample_type &= ~bit;
191 evsel->sample_size -= sizeof(u64);
192 perf_evsel__calc_id_pos(evsel);
193 }
194 }
195
196 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
197 bool can_sample_identifier)
198 {
199 if (can_sample_identifier) {
200 perf_evsel__reset_sample_bit(evsel, ID);
201 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
202 } else {
203 perf_evsel__set_sample_bit(evsel, ID);
204 }
205 evsel->attr.read_format |= PERF_FORMAT_ID;
206 }
207
208 /**
209 * perf_evsel__is_function_event - Return whether given evsel is a function
210 * trace event
211 *
212 * @evsel - evsel selector to be tested
213 *
214 * Return %true if event is function trace event
215 */
216 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
217 {
218 #define FUNCTION_EVENT "ftrace:function"
219
220 return evsel->name &&
221 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
222
223 #undef FUNCTION_EVENT
224 }
225
226 void perf_evsel__init(struct perf_evsel *evsel,
227 struct perf_event_attr *attr, int idx)
228 {
229 evsel->idx = idx;
230 evsel->tracking = !idx;
231 evsel->attr = *attr;
232 evsel->leader = evsel;
233 evsel->unit = "";
234 evsel->scale = 1.0;
235 evsel->max_events = ULONG_MAX;
236 evsel->evlist = NULL;
237 evsel->bpf_fd = -1;
238 INIT_LIST_HEAD(&evsel->node);
239 INIT_LIST_HEAD(&evsel->config_terms);
240 perf_evsel__object.init(evsel);
241 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
242 perf_evsel__calc_id_pos(evsel);
243 evsel->cmdline_group_boundary = false;
244 evsel->metric_expr = NULL;
245 evsel->metric_name = NULL;
246 evsel->metric_events = NULL;
247 evsel->collect_stat = false;
248 evsel->pmu_name = NULL;
249 }
250
251 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
252 {
253 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
254
255 if (!evsel)
256 return NULL;
257 perf_evsel__init(evsel, attr, idx);
258
259 if (perf_evsel__is_bpf_output(evsel)) {
260 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
261 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
262 evsel->attr.sample_period = 1;
263 }
264
265 if (perf_evsel__is_clock(evsel)) {
266 /*
267 * The evsel->unit points to static alias->unit
268 * so it's ok to use static string in here.
269 */
270 static const char *unit = "msec";
271
272 evsel->unit = unit;
273 evsel->scale = 1e-6;
274 }
275
276 return evsel;
277 }
278
279 static bool perf_event_can_profile_kernel(void)
280 {
281 return geteuid() == 0 || perf_event_paranoid() == -1;
282 }
283
284 struct perf_evsel *perf_evsel__new_cycles(bool precise)
285 {
286 struct perf_event_attr attr = {
287 .type = PERF_TYPE_HARDWARE,
288 .config = PERF_COUNT_HW_CPU_CYCLES,
289 .exclude_kernel = !perf_event_can_profile_kernel(),
290 };
291 struct perf_evsel *evsel;
292
293 event_attr_init(&attr);
294
295 if (!precise)
296 goto new_event;
297
298 /*
299 * Now let the usual logic to set up the perf_event_attr defaults
300 * to kick in when we return and before perf_evsel__open() is called.
301 */
302 new_event:
303 evsel = perf_evsel__new(&attr);
304 if (evsel == NULL)
305 goto out;
306
307 evsel->precise_max = true;
308
309 /* use asprintf() because free(evsel) assumes name is allocated */
310 if (asprintf(&evsel->name, "cycles%s%s%.*s",
311 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
312 attr.exclude_kernel ? "u" : "",
313 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
314 goto error_free;
315 out:
316 return evsel;
317 error_free:
318 perf_evsel__delete(evsel);
319 evsel = NULL;
320 goto out;
321 }
322
323 /*
324 * Returns pointer with encoded error via <linux/err.h> interface.
325 */
326 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
327 {
328 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
329 int err = -ENOMEM;
330
331 if (evsel == NULL) {
332 goto out_err;
333 } else {
334 struct perf_event_attr attr = {
335 .type = PERF_TYPE_TRACEPOINT,
336 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
337 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
338 };
339
340 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
341 goto out_free;
342
343 evsel->tp_format = trace_event__tp_format(sys, name);
344 if (IS_ERR(evsel->tp_format)) {
345 err = PTR_ERR(evsel->tp_format);
346 goto out_free;
347 }
348
349 event_attr_init(&attr);
350 attr.config = evsel->tp_format->id;
351 attr.sample_period = 1;
352 perf_evsel__init(evsel, &attr, idx);
353 }
354
355 return evsel;
356
357 out_free:
358 zfree(&evsel->name);
359 free(evsel);
360 out_err:
361 return ERR_PTR(err);
362 }
363
364 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
365 "cycles",
366 "instructions",
367 "cache-references",
368 "cache-misses",
369 "branches",
370 "branch-misses",
371 "bus-cycles",
372 "stalled-cycles-frontend",
373 "stalled-cycles-backend",
374 "ref-cycles",
375 };
376
377 static const char *__perf_evsel__hw_name(u64 config)
378 {
379 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
380 return perf_evsel__hw_names[config];
381
382 return "unknown-hardware";
383 }
384
385 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
386 {
387 int colon = 0, r = 0;
388 struct perf_event_attr *attr = &evsel->attr;
389 bool exclude_guest_default = false;
390
391 #define MOD_PRINT(context, mod) do { \
392 if (!attr->exclude_##context) { \
393 if (!colon) colon = ++r; \
394 r += scnprintf(bf + r, size - r, "%c", mod); \
395 } } while(0)
396
397 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
398 MOD_PRINT(kernel, 'k');
399 MOD_PRINT(user, 'u');
400 MOD_PRINT(hv, 'h');
401 exclude_guest_default = true;
402 }
403
404 if (attr->precise_ip) {
405 if (!colon)
406 colon = ++r;
407 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
408 exclude_guest_default = true;
409 }
410
411 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
412 MOD_PRINT(host, 'H');
413 MOD_PRINT(guest, 'G');
414 }
415 #undef MOD_PRINT
416 if (colon)
417 bf[colon - 1] = ':';
418 return r;
419 }
420
421 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
422 {
423 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
424 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
425 }
426
427 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
428 "cpu-clock",
429 "task-clock",
430 "page-faults",
431 "context-switches",
432 "cpu-migrations",
433 "minor-faults",
434 "major-faults",
435 "alignment-faults",
436 "emulation-faults",
437 "dummy",
438 };
439
440 static const char *__perf_evsel__sw_name(u64 config)
441 {
442 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
443 return perf_evsel__sw_names[config];
444 return "unknown-software";
445 }
446
447 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
448 {
449 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
450 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
451 }
452
453 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
454 {
455 int r;
456
457 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
458
459 if (type & HW_BREAKPOINT_R)
460 r += scnprintf(bf + r, size - r, "r");
461
462 if (type & HW_BREAKPOINT_W)
463 r += scnprintf(bf + r, size - r, "w");
464
465 if (type & HW_BREAKPOINT_X)
466 r += scnprintf(bf + r, size - r, "x");
467
468 return r;
469 }
470
471 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
472 {
473 struct perf_event_attr *attr = &evsel->attr;
474 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
475 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
476 }
477
478 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
479 [PERF_EVSEL__MAX_ALIASES] = {
480 { "L1-dcache", "l1-d", "l1d", "L1-data", },
481 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
482 { "LLC", "L2", },
483 { "dTLB", "d-tlb", "Data-TLB", },
484 { "iTLB", "i-tlb", "Instruction-TLB", },
485 { "branch", "branches", "bpu", "btb", "bpc", },
486 { "node", },
487 };
488
489 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
490 [PERF_EVSEL__MAX_ALIASES] = {
491 { "load", "loads", "read", },
492 { "store", "stores", "write", },
493 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
494 };
495
496 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
497 [PERF_EVSEL__MAX_ALIASES] = {
498 { "refs", "Reference", "ops", "access", },
499 { "misses", "miss", },
500 };
501
502 #define C(x) PERF_COUNT_HW_CACHE_##x
503 #define CACHE_READ (1 << C(OP_READ))
504 #define CACHE_WRITE (1 << C(OP_WRITE))
505 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
506 #define COP(x) (1 << x)
507
508 /*
509 * cache operartion stat
510 * L1I : Read and prefetch only
511 * ITLB and BPU : Read-only
512 */
513 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
514 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
515 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
516 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
517 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
518 [C(ITLB)] = (CACHE_READ),
519 [C(BPU)] = (CACHE_READ),
520 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
521 };
522
523 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
524 {
525 if (perf_evsel__hw_cache_stat[type] & COP(op))
526 return true; /* valid */
527 else
528 return false; /* invalid */
529 }
530
531 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
532 char *bf, size_t size)
533 {
534 if (result) {
535 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
536 perf_evsel__hw_cache_op[op][0],
537 perf_evsel__hw_cache_result[result][0]);
538 }
539
540 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
541 perf_evsel__hw_cache_op[op][1]);
542 }
543
544 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
545 {
546 u8 op, result, type = (config >> 0) & 0xff;
547 const char *err = "unknown-ext-hardware-cache-type";
548
549 if (type >= PERF_COUNT_HW_CACHE_MAX)
550 goto out_err;
551
552 op = (config >> 8) & 0xff;
553 err = "unknown-ext-hardware-cache-op";
554 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
555 goto out_err;
556
557 result = (config >> 16) & 0xff;
558 err = "unknown-ext-hardware-cache-result";
559 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
560 goto out_err;
561
562 err = "invalid-cache";
563 if (!perf_evsel__is_cache_op_valid(type, op))
564 goto out_err;
565
566 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
567 out_err:
568 return scnprintf(bf, size, "%s", err);
569 }
570
571 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
572 {
573 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
574 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
575 }
576
577 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
578 {
579 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
580 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
581 }
582
583 static int perf_evsel__tool_name(char *bf, size_t size)
584 {
585 int ret = scnprintf(bf, size, "duration_time");
586 return ret;
587 }
588
589 const char *perf_evsel__name(struct perf_evsel *evsel)
590 {
591 char bf[128];
592
593 if (evsel->name)
594 return evsel->name;
595
596 switch (evsel->attr.type) {
597 case PERF_TYPE_RAW:
598 perf_evsel__raw_name(evsel, bf, sizeof(bf));
599 break;
600
601 case PERF_TYPE_HARDWARE:
602 perf_evsel__hw_name(evsel, bf, sizeof(bf));
603 break;
604
605 case PERF_TYPE_HW_CACHE:
606 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
607 break;
608
609 case PERF_TYPE_SOFTWARE:
610 if (evsel->tool_event)
611 perf_evsel__tool_name(bf, sizeof(bf));
612 else
613 perf_evsel__sw_name(evsel, bf, sizeof(bf));
614 break;
615
616 case PERF_TYPE_TRACEPOINT:
617 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
618 break;
619
620 case PERF_TYPE_BREAKPOINT:
621 perf_evsel__bp_name(evsel, bf, sizeof(bf));
622 break;
623
624 default:
625 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
626 evsel->attr.type);
627 break;
628 }
629
630 evsel->name = strdup(bf);
631
632 return evsel->name ?: "unknown";
633 }
634
635 const char *perf_evsel__group_name(struct perf_evsel *evsel)
636 {
637 return evsel->group_name ?: "anon group";
638 }
639
640 /*
641 * Returns the group details for the specified leader,
642 * with following rules.
643 *
644 * For record -e '{cycles,instructions}'
645 * 'anon group { cycles:u, instructions:u }'
646 *
647 * For record -e 'cycles,instructions' and report --group
648 * 'cycles:u, instructions:u'
649 */
650 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
651 {
652 int ret = 0;
653 struct perf_evsel *pos;
654 const char *group_name = perf_evsel__group_name(evsel);
655
656 if (!evsel->forced_leader)
657 ret = scnprintf(buf, size, "%s { ", group_name);
658
659 ret += scnprintf(buf + ret, size - ret, "%s",
660 perf_evsel__name(evsel));
661
662 for_each_group_member(pos, evsel)
663 ret += scnprintf(buf + ret, size - ret, ", %s",
664 perf_evsel__name(pos));
665
666 if (!evsel->forced_leader)
667 ret += scnprintf(buf + ret, size - ret, " }");
668
669 return ret;
670 }
671
672 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
673 struct record_opts *opts,
674 struct callchain_param *param)
675 {
676 bool function = perf_evsel__is_function_event(evsel);
677 struct perf_event_attr *attr = &evsel->attr;
678
679 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
680
681 attr->sample_max_stack = param->max_stack;
682
683 if (opts->kernel_callchains)
684 attr->exclude_callchain_user = 1;
685 if (opts->user_callchains)
686 attr->exclude_callchain_kernel = 1;
687 if (param->record_mode == CALLCHAIN_LBR) {
688 if (!opts->branch_stack) {
689 if (attr->exclude_user) {
690 pr_warning("LBR callstack option is only available "
691 "to get user callchain information. "
692 "Falling back to framepointers.\n");
693 } else {
694 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
695 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
696 PERF_SAMPLE_BRANCH_CALL_STACK |
697 PERF_SAMPLE_BRANCH_NO_CYCLES |
698 PERF_SAMPLE_BRANCH_NO_FLAGS;
699 }
700 } else
701 pr_warning("Cannot use LBR callstack with branch stack. "
702 "Falling back to framepointers.\n");
703 }
704
705 if (param->record_mode == CALLCHAIN_DWARF) {
706 if (!function) {
707 perf_evsel__set_sample_bit(evsel, REGS_USER);
708 perf_evsel__set_sample_bit(evsel, STACK_USER);
709 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
710 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
711 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
712 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
713 "so the minimal registers set (IP, SP) is explicitly forced.\n");
714 } else {
715 attr->sample_regs_user |= PERF_REGS_MASK;
716 }
717 attr->sample_stack_user = param->dump_size;
718 attr->exclude_callchain_user = 1;
719 } else {
720 pr_info("Cannot use DWARF unwind for function trace event,"
721 " falling back to framepointers.\n");
722 }
723 }
724
725 if (function) {
726 pr_info("Disabling user space callchains for function trace event.\n");
727 attr->exclude_callchain_user = 1;
728 }
729 }
730
731 void perf_evsel__config_callchain(struct perf_evsel *evsel,
732 struct record_opts *opts,
733 struct callchain_param *param)
734 {
735 if (param->enabled)
736 return __perf_evsel__config_callchain(evsel, opts, param);
737 }
738
739 static void
740 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
741 struct callchain_param *param)
742 {
743 struct perf_event_attr *attr = &evsel->attr;
744
745 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
746 if (param->record_mode == CALLCHAIN_LBR) {
747 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
748 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
749 PERF_SAMPLE_BRANCH_CALL_STACK);
750 }
751 if (param->record_mode == CALLCHAIN_DWARF) {
752 perf_evsel__reset_sample_bit(evsel, REGS_USER);
753 perf_evsel__reset_sample_bit(evsel, STACK_USER);
754 }
755 }
756
757 static void apply_config_terms(struct perf_evsel *evsel,
758 struct record_opts *opts, bool track)
759 {
760 struct perf_evsel_config_term *term;
761 struct list_head *config_terms = &evsel->config_terms;
762 struct perf_event_attr *attr = &evsel->attr;
763 /* callgraph default */
764 struct callchain_param param = {
765 .record_mode = callchain_param.record_mode,
766 };
767 u32 dump_size = 0;
768 int max_stack = 0;
769 const char *callgraph_buf = NULL;
770
771 list_for_each_entry(term, config_terms, list) {
772 switch (term->type) {
773 case PERF_EVSEL__CONFIG_TERM_PERIOD:
774 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
775 attr->sample_period = term->val.period;
776 attr->freq = 0;
777 perf_evsel__reset_sample_bit(evsel, PERIOD);
778 }
779 break;
780 case PERF_EVSEL__CONFIG_TERM_FREQ:
781 if (!(term->weak && opts->user_freq != UINT_MAX)) {
782 attr->sample_freq = term->val.freq;
783 attr->freq = 1;
784 perf_evsel__set_sample_bit(evsel, PERIOD);
785 }
786 break;
787 case PERF_EVSEL__CONFIG_TERM_TIME:
788 if (term->val.time)
789 perf_evsel__set_sample_bit(evsel, TIME);
790 else
791 perf_evsel__reset_sample_bit(evsel, TIME);
792 break;
793 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
794 callgraph_buf = term->val.callgraph;
795 break;
796 case PERF_EVSEL__CONFIG_TERM_BRANCH:
797 if (term->val.branch && strcmp(term->val.branch, "no")) {
798 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
799 parse_branch_str(term->val.branch,
800 &attr->branch_sample_type);
801 } else
802 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
803 break;
804 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
805 dump_size = term->val.stack_user;
806 break;
807 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
808 max_stack = term->val.max_stack;
809 break;
810 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
811 evsel->max_events = term->val.max_events;
812 break;
813 case PERF_EVSEL__CONFIG_TERM_INHERIT:
814 /*
815 * attr->inherit should has already been set by
816 * perf_evsel__config. If user explicitly set
817 * inherit using config terms, override global
818 * opt->no_inherit setting.
819 */
820 attr->inherit = term->val.inherit ? 1 : 0;
821 break;
822 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
823 attr->write_backward = term->val.overwrite ? 1 : 0;
824 break;
825 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
826 break;
827 case PERF_EVSEL__CONFIG_TERM_PERCORE:
828 break;
829 default:
830 break;
831 }
832 }
833
834 /* User explicitly set per-event callgraph, clear the old setting and reset. */
835 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
836 bool sample_address = false;
837
838 if (max_stack) {
839 param.max_stack = max_stack;
840 if (callgraph_buf == NULL)
841 callgraph_buf = "fp";
842 }
843
844 /* parse callgraph parameters */
845 if (callgraph_buf != NULL) {
846 if (!strcmp(callgraph_buf, "no")) {
847 param.enabled = false;
848 param.record_mode = CALLCHAIN_NONE;
849 } else {
850 param.enabled = true;
851 if (parse_callchain_record(callgraph_buf, &param)) {
852 pr_err("per-event callgraph setting for %s failed. "
853 "Apply callgraph global setting for it\n",
854 evsel->name);
855 return;
856 }
857 if (param.record_mode == CALLCHAIN_DWARF)
858 sample_address = true;
859 }
860 }
861 if (dump_size > 0) {
862 dump_size = round_up(dump_size, sizeof(u64));
863 param.dump_size = dump_size;
864 }
865
866 /* If global callgraph set, clear it */
867 if (callchain_param.enabled)
868 perf_evsel__reset_callgraph(evsel, &callchain_param);
869
870 /* set perf-event callgraph */
871 if (param.enabled) {
872 if (sample_address) {
873 perf_evsel__set_sample_bit(evsel, ADDR);
874 perf_evsel__set_sample_bit(evsel, DATA_SRC);
875 evsel->attr.mmap_data = track;
876 }
877 perf_evsel__config_callchain(evsel, opts, &param);
878 }
879 }
880 }
881
882 static bool is_dummy_event(struct perf_evsel *evsel)
883 {
884 return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
885 (evsel->attr.config == PERF_COUNT_SW_DUMMY);
886 }
887
888 /*
889 * The enable_on_exec/disabled value strategy:
890 *
891 * 1) For any type of traced program:
892 * - all independent events and group leaders are disabled
893 * - all group members are enabled
894 *
895 * Group members are ruled by group leaders. They need to
896 * be enabled, because the group scheduling relies on that.
897 *
898 * 2) For traced programs executed by perf:
899 * - all independent events and group leaders have
900 * enable_on_exec set
901 * - we don't specifically enable or disable any event during
902 * the record command
903 *
904 * Independent events and group leaders are initially disabled
905 * and get enabled by exec. Group members are ruled by group
906 * leaders as stated in 1).
907 *
908 * 3) For traced programs attached by perf (pid/tid):
909 * - we specifically enable or disable all events during
910 * the record command
911 *
912 * When attaching events to already running traced we
913 * enable/disable events specifically, as there's no
914 * initial traced exec call.
915 */
916 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
917 struct callchain_param *callchain)
918 {
919 struct perf_evsel *leader = evsel->leader;
920 struct perf_event_attr *attr = &evsel->attr;
921 int track = evsel->tracking;
922 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
923
924 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
925 attr->inherit = !opts->no_inherit;
926 attr->write_backward = opts->overwrite ? 1 : 0;
927
928 perf_evsel__set_sample_bit(evsel, IP);
929 perf_evsel__set_sample_bit(evsel, TID);
930
931 if (evsel->sample_read) {
932 perf_evsel__set_sample_bit(evsel, READ);
933
934 /*
935 * We need ID even in case of single event, because
936 * PERF_SAMPLE_READ process ID specific data.
937 */
938 perf_evsel__set_sample_id(evsel, false);
939
940 /*
941 * Apply group format only if we belong to group
942 * with more than one members.
943 */
944 if (leader->nr_members > 1) {
945 attr->read_format |= PERF_FORMAT_GROUP;
946 attr->inherit = 0;
947 }
948 }
949
950 /*
951 * We default some events to have a default interval. But keep
952 * it a weak assumption overridable by the user.
953 */
954 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
955 opts->user_interval != ULLONG_MAX)) {
956 if (opts->freq) {
957 perf_evsel__set_sample_bit(evsel, PERIOD);
958 attr->freq = 1;
959 attr->sample_freq = opts->freq;
960 } else {
961 attr->sample_period = opts->default_interval;
962 }
963 }
964
965 /*
966 * Disable sampling for all group members other
967 * than leader in case leader 'leads' the sampling.
968 */
969 if ((leader != evsel) && leader->sample_read) {
970 attr->freq = 0;
971 attr->sample_freq = 0;
972 attr->sample_period = 0;
973 attr->write_backward = 0;
974
975 /*
976 * We don't get sample for slave events, we make them
977 * when delivering group leader sample. Set the slave
978 * event to follow the master sample_type to ease up
979 * report.
980 */
981 attr->sample_type = leader->attr.sample_type;
982 }
983
984 if (opts->no_samples)
985 attr->sample_freq = 0;
986
987 if (opts->inherit_stat) {
988 evsel->attr.read_format |=
989 PERF_FORMAT_TOTAL_TIME_ENABLED |
990 PERF_FORMAT_TOTAL_TIME_RUNNING |
991 PERF_FORMAT_ID;
992 attr->inherit_stat = 1;
993 }
994
995 if (opts->sample_address) {
996 perf_evsel__set_sample_bit(evsel, ADDR);
997 attr->mmap_data = track;
998 }
999
1000 /*
1001 * We don't allow user space callchains for function trace
1002 * event, due to issues with page faults while tracing page
1003 * fault handler and its overall trickiness nature.
1004 */
1005 if (perf_evsel__is_function_event(evsel))
1006 evsel->attr.exclude_callchain_user = 1;
1007
1008 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1009 perf_evsel__config_callchain(evsel, opts, callchain);
1010
1011 if (opts->sample_intr_regs) {
1012 attr->sample_regs_intr = opts->sample_intr_regs;
1013 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1014 }
1015
1016 if (opts->sample_user_regs) {
1017 attr->sample_regs_user |= opts->sample_user_regs;
1018 perf_evsel__set_sample_bit(evsel, REGS_USER);
1019 }
1020
1021 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1022 perf_evsel__set_sample_bit(evsel, CPU);
1023
1024 /*
1025 * When the user explicitly disabled time don't force it here.
1026 */
1027 if (opts->sample_time &&
1028 (!perf_missing_features.sample_id_all &&
1029 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1030 opts->sample_time_set)))
1031 perf_evsel__set_sample_bit(evsel, TIME);
1032
1033 if (opts->raw_samples && !evsel->no_aux_samples) {
1034 perf_evsel__set_sample_bit(evsel, TIME);
1035 perf_evsel__set_sample_bit(evsel, RAW);
1036 perf_evsel__set_sample_bit(evsel, CPU);
1037 }
1038
1039 if (opts->sample_address)
1040 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1041
1042 if (opts->sample_phys_addr)
1043 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1044
1045 if (opts->no_buffering) {
1046 attr->watermark = 0;
1047 attr->wakeup_events = 1;
1048 }
1049 if (opts->branch_stack && !evsel->no_aux_samples) {
1050 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1051 attr->branch_sample_type = opts->branch_stack;
1052 }
1053
1054 if (opts->sample_weight)
1055 perf_evsel__set_sample_bit(evsel, WEIGHT);
1056
1057 attr->task = track;
1058 attr->mmap = track;
1059 attr->mmap2 = track && !perf_missing_features.mmap2;
1060 attr->comm = track;
1061 attr->ksymbol = track && !perf_missing_features.ksymbol;
1062 attr->bpf_event = track && !opts->no_bpf_event &&
1063 !perf_missing_features.bpf_event;
1064
1065 if (opts->record_namespaces)
1066 attr->namespaces = track;
1067
1068 if (opts->record_switch_events)
1069 attr->context_switch = track;
1070
1071 if (opts->sample_transaction)
1072 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1073
1074 if (opts->running_time) {
1075 evsel->attr.read_format |=
1076 PERF_FORMAT_TOTAL_TIME_ENABLED |
1077 PERF_FORMAT_TOTAL_TIME_RUNNING;
1078 }
1079
1080 /*
1081 * XXX see the function comment above
1082 *
1083 * Disabling only independent events or group leaders,
1084 * keeping group members enabled.
1085 */
1086 if (perf_evsel__is_group_leader(evsel))
1087 attr->disabled = 1;
1088
1089 /*
1090 * Setting enable_on_exec for independent events and
1091 * group leaders for traced executed by perf.
1092 */
1093 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1094 !opts->initial_delay)
1095 attr->enable_on_exec = 1;
1096
1097 if (evsel->immediate) {
1098 attr->disabled = 0;
1099 attr->enable_on_exec = 0;
1100 }
1101
1102 clockid = opts->clockid;
1103 if (opts->use_clockid) {
1104 attr->use_clockid = 1;
1105 attr->clockid = opts->clockid;
1106 }
1107
1108 if (evsel->precise_max)
1109 attr->precise_ip = 3;
1110
1111 if (opts->all_user) {
1112 attr->exclude_kernel = 1;
1113 attr->exclude_user = 0;
1114 }
1115
1116 if (opts->all_kernel) {
1117 attr->exclude_kernel = 0;
1118 attr->exclude_user = 1;
1119 }
1120
1121 if (evsel->own_cpus || evsel->unit)
1122 evsel->attr.read_format |= PERF_FORMAT_ID;
1123
1124 /*
1125 * Apply event specific term settings,
1126 * it overloads any global configuration.
1127 */
1128 apply_config_terms(evsel, opts, track);
1129
1130 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1131
1132 /* The --period option takes the precedence. */
1133 if (opts->period_set) {
1134 if (opts->period)
1135 perf_evsel__set_sample_bit(evsel, PERIOD);
1136 else
1137 perf_evsel__reset_sample_bit(evsel, PERIOD);
1138 }
1139
1140 /*
1141 * For initial_delay, a dummy event is added implicitly.
1142 * The software event will trigger -EOPNOTSUPP error out,
1143 * if BRANCH_STACK bit is set.
1144 */
1145 if (opts->initial_delay && is_dummy_event(evsel))
1146 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1147 }
1148
1149 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1150 {
1151 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1152
1153 if (evsel->fd) {
1154 int cpu, thread;
1155 for (cpu = 0; cpu < ncpus; cpu++) {
1156 for (thread = 0; thread < nthreads; thread++) {
1157 FD(evsel, cpu, thread) = -1;
1158 }
1159 }
1160 }
1161
1162 return evsel->fd != NULL ? 0 : -ENOMEM;
1163 }
1164
1165 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1166 int ioc, void *arg)
1167 {
1168 int cpu, thread;
1169
1170 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1171 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1172 int fd = FD(evsel, cpu, thread),
1173 err = ioctl(fd, ioc, arg);
1174
1175 if (err)
1176 return err;
1177 }
1178 }
1179
1180 return 0;
1181 }
1182
1183 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1184 {
1185 return perf_evsel__run_ioctl(evsel,
1186 PERF_EVENT_IOC_SET_FILTER,
1187 (void *)filter);
1188 }
1189
1190 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1191 {
1192 char *new_filter = strdup(filter);
1193
1194 if (new_filter != NULL) {
1195 free(evsel->filter);
1196 evsel->filter = new_filter;
1197 return 0;
1198 }
1199
1200 return -1;
1201 }
1202
1203 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1204 const char *fmt, const char *filter)
1205 {
1206 char *new_filter;
1207
1208 if (evsel->filter == NULL)
1209 return perf_evsel__set_filter(evsel, filter);
1210
1211 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1212 free(evsel->filter);
1213 evsel->filter = new_filter;
1214 return 0;
1215 }
1216
1217 return -1;
1218 }
1219
1220 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1221 {
1222 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1223 }
1224
1225 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1226 {
1227 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1228 }
1229
1230 int perf_evsel__enable(struct perf_evsel *evsel)
1231 {
1232 int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1233
1234 if (!err)
1235 evsel->disabled = false;
1236
1237 return err;
1238 }
1239
1240 int perf_evsel__disable(struct perf_evsel *evsel)
1241 {
1242 int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1243 /*
1244 * We mark it disabled here so that tools that disable a event can
1245 * ignore events after they disable it. I.e. the ring buffer may have
1246 * already a few more events queued up before the kernel got the stop
1247 * request.
1248 */
1249 if (!err)
1250 evsel->disabled = true;
1251
1252 return err;
1253 }
1254
1255 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1256 {
1257 if (ncpus == 0 || nthreads == 0)
1258 return 0;
1259
1260 if (evsel->system_wide)
1261 nthreads = 1;
1262
1263 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1264 if (evsel->sample_id == NULL)
1265 return -ENOMEM;
1266
1267 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1268 if (evsel->id == NULL) {
1269 xyarray__delete(evsel->sample_id);
1270 evsel->sample_id = NULL;
1271 return -ENOMEM;
1272 }
1273
1274 return 0;
1275 }
1276
1277 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1278 {
1279 xyarray__delete(evsel->fd);
1280 evsel->fd = NULL;
1281 }
1282
1283 static void perf_evsel__free_id(struct perf_evsel *evsel)
1284 {
1285 xyarray__delete(evsel->sample_id);
1286 evsel->sample_id = NULL;
1287 zfree(&evsel->id);
1288 }
1289
1290 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1291 {
1292 struct perf_evsel_config_term *term, *h;
1293
1294 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1295 list_del(&term->list);
1296 free(term);
1297 }
1298 }
1299
1300 void perf_evsel__close_fd(struct perf_evsel *evsel)
1301 {
1302 int cpu, thread;
1303
1304 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1305 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1306 close(FD(evsel, cpu, thread));
1307 FD(evsel, cpu, thread) = -1;
1308 }
1309 }
1310
1311 void perf_evsel__exit(struct perf_evsel *evsel)
1312 {
1313 assert(list_empty(&evsel->node));
1314 assert(evsel->evlist == NULL);
1315 perf_evsel__free_counts(evsel);
1316 perf_evsel__free_fd(evsel);
1317 perf_evsel__free_id(evsel);
1318 perf_evsel__free_config_terms(evsel);
1319 cgroup__put(evsel->cgrp);
1320 cpu_map__put(evsel->cpus);
1321 cpu_map__put(evsel->own_cpus);
1322 thread_map__put(evsel->threads);
1323 zfree(&evsel->group_name);
1324 zfree(&evsel->name);
1325 perf_evsel__object.fini(evsel);
1326 }
1327
1328 void perf_evsel__delete(struct perf_evsel *evsel)
1329 {
1330 perf_evsel__exit(evsel);
1331 free(evsel);
1332 }
1333
1334 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1335 struct perf_counts_values *count)
1336 {
1337 struct perf_counts_values tmp;
1338
1339 if (!evsel->prev_raw_counts)
1340 return;
1341
1342 if (cpu == -1) {
1343 tmp = evsel->prev_raw_counts->aggr;
1344 evsel->prev_raw_counts->aggr = *count;
1345 } else {
1346 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1347 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1348 }
1349
1350 count->val = count->val - tmp.val;
1351 count->ena = count->ena - tmp.ena;
1352 count->run = count->run - tmp.run;
1353 }
1354
1355 void perf_counts_values__scale(struct perf_counts_values *count,
1356 bool scale, s8 *pscaled)
1357 {
1358 s8 scaled = 0;
1359
1360 if (scale) {
1361 if (count->run == 0) {
1362 scaled = -1;
1363 count->val = 0;
1364 } else if (count->run < count->ena) {
1365 scaled = 1;
1366 count->val = (u64)((double) count->val * count->ena / count->run);
1367 }
1368 }
1369
1370 if (pscaled)
1371 *pscaled = scaled;
1372 }
1373
1374 static int perf_evsel__read_size(struct perf_evsel *evsel)
1375 {
1376 u64 read_format = evsel->attr.read_format;
1377 int entry = sizeof(u64); /* value */
1378 int size = 0;
1379 int nr = 1;
1380
1381 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1382 size += sizeof(u64);
1383
1384 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1385 size += sizeof(u64);
1386
1387 if (read_format & PERF_FORMAT_ID)
1388 entry += sizeof(u64);
1389
1390 if (read_format & PERF_FORMAT_GROUP) {
1391 nr = evsel->nr_members;
1392 size += sizeof(u64);
1393 }
1394
1395 size += entry * nr;
1396 return size;
1397 }
1398
1399 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1400 struct perf_counts_values *count)
1401 {
1402 size_t size = perf_evsel__read_size(evsel);
1403
1404 memset(count, 0, sizeof(*count));
1405
1406 if (FD(evsel, cpu, thread) < 0)
1407 return -EINVAL;
1408
1409 if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1410 return -errno;
1411
1412 return 0;
1413 }
1414
1415 static int
1416 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1417 {
1418 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1419
1420 return perf_evsel__read(evsel, cpu, thread, count);
1421 }
1422
1423 static void
1424 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1425 u64 val, u64 ena, u64 run)
1426 {
1427 struct perf_counts_values *count;
1428
1429 count = perf_counts(counter->counts, cpu, thread);
1430
1431 count->val = val;
1432 count->ena = ena;
1433 count->run = run;
1434 count->loaded = true;
1435 }
1436
1437 static int
1438 perf_evsel__process_group_data(struct perf_evsel *leader,
1439 int cpu, int thread, u64 *data)
1440 {
1441 u64 read_format = leader->attr.read_format;
1442 struct sample_read_value *v;
1443 u64 nr, ena = 0, run = 0, i;
1444
1445 nr = *data++;
1446
1447 if (nr != (u64) leader->nr_members)
1448 return -EINVAL;
1449
1450 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1451 ena = *data++;
1452
1453 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1454 run = *data++;
1455
1456 v = (struct sample_read_value *) data;
1457
1458 perf_evsel__set_count(leader, cpu, thread,
1459 v[0].value, ena, run);
1460
1461 for (i = 1; i < nr; i++) {
1462 struct perf_evsel *counter;
1463
1464 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1465 if (!counter)
1466 return -EINVAL;
1467
1468 perf_evsel__set_count(counter, cpu, thread,
1469 v[i].value, ena, run);
1470 }
1471
1472 return 0;
1473 }
1474
1475 static int
1476 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1477 {
1478 struct perf_stat_evsel *ps = leader->stats;
1479 u64 read_format = leader->attr.read_format;
1480 int size = perf_evsel__read_size(leader);
1481 u64 *data = ps->group_data;
1482
1483 if (!(read_format & PERF_FORMAT_ID))
1484 return -EINVAL;
1485
1486 if (!perf_evsel__is_group_leader(leader))
1487 return -EINVAL;
1488
1489 if (!data) {
1490 data = zalloc(size);
1491 if (!data)
1492 return -ENOMEM;
1493
1494 ps->group_data = data;
1495 }
1496
1497 if (FD(leader, cpu, thread) < 0)
1498 return -EINVAL;
1499
1500 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1501 return -errno;
1502
1503 return perf_evsel__process_group_data(leader, cpu, thread, data);
1504 }
1505
1506 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1507 {
1508 u64 read_format = evsel->attr.read_format;
1509
1510 if (read_format & PERF_FORMAT_GROUP)
1511 return perf_evsel__read_group(evsel, cpu, thread);
1512 else
1513 return perf_evsel__read_one(evsel, cpu, thread);
1514 }
1515
1516 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1517 int cpu, int thread, bool scale)
1518 {
1519 struct perf_counts_values count;
1520 size_t nv = scale ? 3 : 1;
1521
1522 if (FD(evsel, cpu, thread) < 0)
1523 return -EINVAL;
1524
1525 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1526 return -ENOMEM;
1527
1528 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1529 return -errno;
1530
1531 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1532 perf_counts_values__scale(&count, scale, NULL);
1533 *perf_counts(evsel->counts, cpu, thread) = count;
1534 return 0;
1535 }
1536
1537 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1538 {
1539 struct perf_evsel *leader = evsel->leader;
1540 int fd;
1541
1542 if (perf_evsel__is_group_leader(evsel))
1543 return -1;
1544
1545 /*
1546 * Leader must be already processed/open,
1547 * if not it's a bug.
1548 */
1549 BUG_ON(!leader->fd);
1550
1551 fd = FD(leader, cpu, thread);
1552 BUG_ON(fd == -1);
1553
1554 return fd;
1555 }
1556
1557 struct bit_names {
1558 int bit;
1559 const char *name;
1560 };
1561
1562 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1563 {
1564 bool first_bit = true;
1565 int i = 0;
1566
1567 do {
1568 if (value & bits[i].bit) {
1569 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1570 first_bit = false;
1571 }
1572 } while (bits[++i].name != NULL);
1573 }
1574
1575 static void __p_sample_type(char *buf, size_t size, u64 value)
1576 {
1577 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1578 struct bit_names bits[] = {
1579 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1580 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1581 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1582 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1583 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1584 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1585 { .name = NULL, }
1586 };
1587 #undef bit_name
1588 __p_bits(buf, size, value, bits);
1589 }
1590
1591 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1592 {
1593 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1594 struct bit_names bits[] = {
1595 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1596 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1597 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1598 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1599 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1600 { .name = NULL, }
1601 };
1602 #undef bit_name
1603 __p_bits(buf, size, value, bits);
1604 }
1605
1606 static void __p_read_format(char *buf, size_t size, u64 value)
1607 {
1608 #define bit_name(n) { PERF_FORMAT_##n, #n }
1609 struct bit_names bits[] = {
1610 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1611 bit_name(ID), bit_name(GROUP),
1612 { .name = NULL, }
1613 };
1614 #undef bit_name
1615 __p_bits(buf, size, value, bits);
1616 }
1617
1618 #define BUF_SIZE 1024
1619
1620 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1621 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1622 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1623 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1624 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1625 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1626
1627 #define PRINT_ATTRn(_n, _f, _p) \
1628 do { \
1629 if (attr->_f) { \
1630 _p(attr->_f); \
1631 ret += attr__fprintf(fp, _n, buf, priv);\
1632 } \
1633 } while (0)
1634
1635 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1636
1637 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1638 attr__fprintf_f attr__fprintf, void *priv)
1639 {
1640 char buf[BUF_SIZE];
1641 int ret = 0;
1642
1643 PRINT_ATTRf(type, p_unsigned);
1644 PRINT_ATTRf(size, p_unsigned);
1645 PRINT_ATTRf(config, p_hex);
1646 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1647 PRINT_ATTRf(sample_type, p_sample_type);
1648 PRINT_ATTRf(read_format, p_read_format);
1649
1650 PRINT_ATTRf(disabled, p_unsigned);
1651 PRINT_ATTRf(inherit, p_unsigned);
1652 PRINT_ATTRf(pinned, p_unsigned);
1653 PRINT_ATTRf(exclusive, p_unsigned);
1654 PRINT_ATTRf(exclude_user, p_unsigned);
1655 PRINT_ATTRf(exclude_kernel, p_unsigned);
1656 PRINT_ATTRf(exclude_hv, p_unsigned);
1657 PRINT_ATTRf(exclude_idle, p_unsigned);
1658 PRINT_ATTRf(mmap, p_unsigned);
1659 PRINT_ATTRf(comm, p_unsigned);
1660 PRINT_ATTRf(freq, p_unsigned);
1661 PRINT_ATTRf(inherit_stat, p_unsigned);
1662 PRINT_ATTRf(enable_on_exec, p_unsigned);
1663 PRINT_ATTRf(task, p_unsigned);
1664 PRINT_ATTRf(watermark, p_unsigned);
1665 PRINT_ATTRf(precise_ip, p_unsigned);
1666 PRINT_ATTRf(mmap_data, p_unsigned);
1667 PRINT_ATTRf(sample_id_all, p_unsigned);
1668 PRINT_ATTRf(exclude_host, p_unsigned);
1669 PRINT_ATTRf(exclude_guest, p_unsigned);
1670 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1671 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1672 PRINT_ATTRf(mmap2, p_unsigned);
1673 PRINT_ATTRf(comm_exec, p_unsigned);
1674 PRINT_ATTRf(use_clockid, p_unsigned);
1675 PRINT_ATTRf(context_switch, p_unsigned);
1676 PRINT_ATTRf(write_backward, p_unsigned);
1677 PRINT_ATTRf(namespaces, p_unsigned);
1678 PRINT_ATTRf(ksymbol, p_unsigned);
1679 PRINT_ATTRf(bpf_event, p_unsigned);
1680
1681 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1682 PRINT_ATTRf(bp_type, p_unsigned);
1683 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1684 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1685 PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1686 PRINT_ATTRf(sample_regs_user, p_hex);
1687 PRINT_ATTRf(sample_stack_user, p_unsigned);
1688 PRINT_ATTRf(clockid, p_signed);
1689 PRINT_ATTRf(sample_regs_intr, p_hex);
1690 PRINT_ATTRf(aux_watermark, p_unsigned);
1691 PRINT_ATTRf(sample_max_stack, p_unsigned);
1692
1693 return ret;
1694 }
1695
1696 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1697 void *priv __maybe_unused)
1698 {
1699 return fprintf(fp, " %-32s %s\n", name, val);
1700 }
1701
1702 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1703 int nr_cpus, int nr_threads,
1704 int thread_idx)
1705 {
1706 for (int cpu = 0; cpu < nr_cpus; cpu++)
1707 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1708 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1709 }
1710
1711 static int update_fds(struct perf_evsel *evsel,
1712 int nr_cpus, int cpu_idx,
1713 int nr_threads, int thread_idx)
1714 {
1715 struct perf_evsel *pos;
1716
1717 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1718 return -EINVAL;
1719
1720 evlist__for_each_entry(evsel->evlist, pos) {
1721 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1722
1723 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1724
1725 /*
1726 * Since fds for next evsel has not been created,
1727 * there is no need to iterate whole event list.
1728 */
1729 if (pos == evsel)
1730 break;
1731 }
1732 return 0;
1733 }
1734
1735 static bool ignore_missing_thread(struct perf_evsel *evsel,
1736 int nr_cpus, int cpu,
1737 struct thread_map *threads,
1738 int thread, int err)
1739 {
1740 pid_t ignore_pid = thread_map__pid(threads, thread);
1741
1742 if (!evsel->ignore_missing_thread)
1743 return false;
1744
1745 /* The system wide setup does not work with threads. */
1746 if (evsel->system_wide)
1747 return false;
1748
1749 /* The -ESRCH is perf event syscall errno for pid's not found. */
1750 if (err != -ESRCH)
1751 return false;
1752
1753 /* If there's only one thread, let it fail. */
1754 if (threads->nr == 1)
1755 return false;
1756
1757 /*
1758 * We should remove fd for missing_thread first
1759 * because thread_map__remove() will decrease threads->nr.
1760 */
1761 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1762 return false;
1763
1764 if (thread_map__remove(threads, thread))
1765 return false;
1766
1767 pr_warning("WARNING: Ignored open failure for pid %d\n",
1768 ignore_pid);
1769 return true;
1770 }
1771
1772 static void display_attr(struct perf_event_attr *attr)
1773 {
1774 if (verbose >= 2) {
1775 fprintf(stderr, "%.60s\n", graph_dotted_line);
1776 fprintf(stderr, "perf_event_attr:\n");
1777 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1778 fprintf(stderr, "%.60s\n", graph_dotted_line);
1779 }
1780 }
1781
1782 static int perf_event_open(struct perf_evsel *evsel,
1783 pid_t pid, int cpu, int group_fd,
1784 unsigned long flags)
1785 {
1786 int precise_ip = evsel->attr.precise_ip;
1787 int fd;
1788
1789 while (1) {
1790 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1791 pid, cpu, group_fd, flags);
1792
1793 fd = sys_perf_event_open(&evsel->attr, pid, cpu, group_fd, flags);
1794 if (fd >= 0)
1795 break;
1796
1797 /*
1798 * Do quick precise_ip fallback if:
1799 * - there is precise_ip set in perf_event_attr
1800 * - maximum precise is requested
1801 * - sys_perf_event_open failed with ENOTSUP error,
1802 * which is associated with wrong precise_ip
1803 */
1804 if (!precise_ip || !evsel->precise_max || (errno != ENOTSUP))
1805 break;
1806
1807 /*
1808 * We tried all the precise_ip values, and it's
1809 * still failing, so leave it to standard fallback.
1810 */
1811 if (!evsel->attr.precise_ip) {
1812 evsel->attr.precise_ip = precise_ip;
1813 break;
1814 }
1815
1816 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1817 evsel->attr.precise_ip--;
1818 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->attr.precise_ip);
1819 display_attr(&evsel->attr);
1820 }
1821
1822 return fd;
1823 }
1824
1825 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1826 struct thread_map *threads)
1827 {
1828 int cpu, thread, nthreads;
1829 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1830 int pid = -1, err;
1831 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1832
1833 if (perf_missing_features.write_backward && evsel->attr.write_backward)
1834 return -EINVAL;
1835
1836 if (cpus == NULL) {
1837 static struct cpu_map *empty_cpu_map;
1838
1839 if (empty_cpu_map == NULL) {
1840 empty_cpu_map = cpu_map__dummy_new();
1841 if (empty_cpu_map == NULL)
1842 return -ENOMEM;
1843 }
1844
1845 cpus = empty_cpu_map;
1846 }
1847
1848 if (threads == NULL) {
1849 static struct thread_map *empty_thread_map;
1850
1851 if (empty_thread_map == NULL) {
1852 empty_thread_map = thread_map__new_by_tid(-1);
1853 if (empty_thread_map == NULL)
1854 return -ENOMEM;
1855 }
1856
1857 threads = empty_thread_map;
1858 }
1859
1860 if (evsel->system_wide)
1861 nthreads = 1;
1862 else
1863 nthreads = threads->nr;
1864
1865 if (evsel->fd == NULL &&
1866 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1867 return -ENOMEM;
1868
1869 if (evsel->cgrp) {
1870 flags |= PERF_FLAG_PID_CGROUP;
1871 pid = evsel->cgrp->fd;
1872 }
1873
1874 fallback_missing_features:
1875 if (perf_missing_features.clockid_wrong)
1876 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1877 if (perf_missing_features.clockid) {
1878 evsel->attr.use_clockid = 0;
1879 evsel->attr.clockid = 0;
1880 }
1881 if (perf_missing_features.cloexec)
1882 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1883 if (perf_missing_features.mmap2)
1884 evsel->attr.mmap2 = 0;
1885 if (perf_missing_features.exclude_guest)
1886 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1887 if (perf_missing_features.lbr_flags)
1888 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1889 PERF_SAMPLE_BRANCH_NO_CYCLES);
1890 if (perf_missing_features.group_read && evsel->attr.inherit)
1891 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1892 if (perf_missing_features.ksymbol)
1893 evsel->attr.ksymbol = 0;
1894 if (perf_missing_features.bpf_event)
1895 evsel->attr.bpf_event = 0;
1896 retry_sample_id:
1897 if (perf_missing_features.sample_id_all)
1898 evsel->attr.sample_id_all = 0;
1899
1900 display_attr(&evsel->attr);
1901
1902 for (cpu = 0; cpu < cpus->nr; cpu++) {
1903
1904 for (thread = 0; thread < nthreads; thread++) {
1905 int fd, group_fd;
1906
1907 if (!evsel->cgrp && !evsel->system_wide)
1908 pid = thread_map__pid(threads, thread);
1909
1910 group_fd = get_group_fd(evsel, cpu, thread);
1911 retry_open:
1912 test_attr__ready();
1913
1914 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1915 group_fd, flags);
1916
1917 FD(evsel, cpu, thread) = fd;
1918
1919 if (fd < 0) {
1920 err = -errno;
1921
1922 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1923 /*
1924 * We just removed 1 thread, so take a step
1925 * back on thread index and lower the upper
1926 * nthreads limit.
1927 */
1928 nthreads--;
1929 thread--;
1930
1931 /* ... and pretend like nothing have happened. */
1932 err = 0;
1933 continue;
1934 }
1935
1936 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1937 err);
1938 goto try_fallback;
1939 }
1940
1941 pr_debug2(" = %d\n", fd);
1942
1943 if (evsel->bpf_fd >= 0) {
1944 int evt_fd = fd;
1945 int bpf_fd = evsel->bpf_fd;
1946
1947 err = ioctl(evt_fd,
1948 PERF_EVENT_IOC_SET_BPF,
1949 bpf_fd);
1950 if (err && errno != EEXIST) {
1951 pr_err("failed to attach bpf fd %d: %s\n",
1952 bpf_fd, strerror(errno));
1953 err = -EINVAL;
1954 goto out_close;
1955 }
1956 }
1957
1958 set_rlimit = NO_CHANGE;
1959
1960 /*
1961 * If we succeeded but had to kill clockid, fail and
1962 * have perf_evsel__open_strerror() print us a nice
1963 * error.
1964 */
1965 if (perf_missing_features.clockid ||
1966 perf_missing_features.clockid_wrong) {
1967 err = -EINVAL;
1968 goto out_close;
1969 }
1970 }
1971 }
1972
1973 return 0;
1974
1975 try_fallback:
1976 /*
1977 * perf stat needs between 5 and 22 fds per CPU. When we run out
1978 * of them try to increase the limits.
1979 */
1980 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1981 struct rlimit l;
1982 int old_errno = errno;
1983
1984 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1985 if (set_rlimit == NO_CHANGE)
1986 l.rlim_cur = l.rlim_max;
1987 else {
1988 l.rlim_cur = l.rlim_max + 1000;
1989 l.rlim_max = l.rlim_cur;
1990 }
1991 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1992 set_rlimit++;
1993 errno = old_errno;
1994 goto retry_open;
1995 }
1996 }
1997 errno = old_errno;
1998 }
1999
2000 if (err != -EINVAL || cpu > 0 || thread > 0)
2001 goto out_close;
2002
2003 /*
2004 * Must probe features in the order they were added to the
2005 * perf_event_attr interface.
2006 */
2007 if (!perf_missing_features.bpf_event && evsel->attr.bpf_event) {
2008 perf_missing_features.bpf_event = true;
2009 pr_debug2("switching off bpf_event\n");
2010 goto fallback_missing_features;
2011 } else if (!perf_missing_features.ksymbol && evsel->attr.ksymbol) {
2012 perf_missing_features.ksymbol = true;
2013 pr_debug2("switching off ksymbol\n");
2014 goto fallback_missing_features;
2015 } else if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
2016 perf_missing_features.write_backward = true;
2017 pr_debug2("switching off write_backward\n");
2018 goto out_close;
2019 } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
2020 perf_missing_features.clockid_wrong = true;
2021 pr_debug2("switching off clockid\n");
2022 goto fallback_missing_features;
2023 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
2024 perf_missing_features.clockid = true;
2025 pr_debug2("switching off use_clockid\n");
2026 goto fallback_missing_features;
2027 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
2028 perf_missing_features.cloexec = true;
2029 pr_debug2("switching off cloexec flag\n");
2030 goto fallback_missing_features;
2031 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
2032 perf_missing_features.mmap2 = true;
2033 pr_debug2("switching off mmap2\n");
2034 goto fallback_missing_features;
2035 } else if (!perf_missing_features.exclude_guest &&
2036 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
2037 perf_missing_features.exclude_guest = true;
2038 pr_debug2("switching off exclude_guest, exclude_host\n");
2039 goto fallback_missing_features;
2040 } else if (!perf_missing_features.sample_id_all) {
2041 perf_missing_features.sample_id_all = true;
2042 pr_debug2("switching off sample_id_all\n");
2043 goto retry_sample_id;
2044 } else if (!perf_missing_features.lbr_flags &&
2045 (evsel->attr.branch_sample_type &
2046 (PERF_SAMPLE_BRANCH_NO_CYCLES |
2047 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2048 perf_missing_features.lbr_flags = true;
2049 pr_debug2("switching off branch sample type no (cycles/flags)\n");
2050 goto fallback_missing_features;
2051 } else if (!perf_missing_features.group_read &&
2052 evsel->attr.inherit &&
2053 (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
2054 perf_evsel__is_group_leader(evsel)) {
2055 perf_missing_features.group_read = true;
2056 pr_debug2("switching off group read\n");
2057 goto fallback_missing_features;
2058 }
2059 out_close:
2060 if (err)
2061 threads->err_thread = thread;
2062
2063 do {
2064 while (--thread >= 0) {
2065 close(FD(evsel, cpu, thread));
2066 FD(evsel, cpu, thread) = -1;
2067 }
2068 thread = nthreads;
2069 } while (--cpu >= 0);
2070 return err;
2071 }
2072
2073 void perf_evsel__close(struct perf_evsel *evsel)
2074 {
2075 if (evsel->fd == NULL)
2076 return;
2077
2078 perf_evsel__close_fd(evsel);
2079 perf_evsel__free_fd(evsel);
2080 }
2081
2082 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2083 struct cpu_map *cpus)
2084 {
2085 return perf_evsel__open(evsel, cpus, NULL);
2086 }
2087
2088 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2089 struct thread_map *threads)
2090 {
2091 return perf_evsel__open(evsel, NULL, threads);
2092 }
2093
2094 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2095 const union perf_event *event,
2096 struct perf_sample *sample)
2097 {
2098 u64 type = evsel->attr.sample_type;
2099 const u64 *array = event->sample.array;
2100 bool swapped = evsel->needs_swap;
2101 union u64_swap u;
2102
2103 array += ((event->header.size -
2104 sizeof(event->header)) / sizeof(u64)) - 1;
2105
2106 if (type & PERF_SAMPLE_IDENTIFIER) {
2107 sample->id = *array;
2108 array--;
2109 }
2110
2111 if (type & PERF_SAMPLE_CPU) {
2112 u.val64 = *array;
2113 if (swapped) {
2114 /* undo swap of u64, then swap on individual u32s */
2115 u.val64 = bswap_64(u.val64);
2116 u.val32[0] = bswap_32(u.val32[0]);
2117 }
2118
2119 sample->cpu = u.val32[0];
2120 array--;
2121 }
2122
2123 if (type & PERF_SAMPLE_STREAM_ID) {
2124 sample->stream_id = *array;
2125 array--;
2126 }
2127
2128 if (type & PERF_SAMPLE_ID) {
2129 sample->id = *array;
2130 array--;
2131 }
2132
2133 if (type & PERF_SAMPLE_TIME) {
2134 sample->time = *array;
2135 array--;
2136 }
2137
2138 if (type & PERF_SAMPLE_TID) {
2139 u.val64 = *array;
2140 if (swapped) {
2141 /* undo swap of u64, then swap on individual u32s */
2142 u.val64 = bswap_64(u.val64);
2143 u.val32[0] = bswap_32(u.val32[0]);
2144 u.val32[1] = bswap_32(u.val32[1]);
2145 }
2146
2147 sample->pid = u.val32[0];
2148 sample->tid = u.val32[1];
2149 array--;
2150 }
2151
2152 return 0;
2153 }
2154
2155 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2156 u64 size)
2157 {
2158 return size > max_size || offset + size > endp;
2159 }
2160
2161 #define OVERFLOW_CHECK(offset, size, max_size) \
2162 do { \
2163 if (overflow(endp, (max_size), (offset), (size))) \
2164 return -EFAULT; \
2165 } while (0)
2166
2167 #define OVERFLOW_CHECK_u64(offset) \
2168 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2169
2170 static int
2171 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2172 {
2173 /*
2174 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2175 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2176 * check the format does not go past the end of the event.
2177 */
2178 if (sample_size + sizeof(event->header) > event->header.size)
2179 return -EFAULT;
2180
2181 return 0;
2182 }
2183
2184 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2185 struct perf_sample *data)
2186 {
2187 u64 type = evsel->attr.sample_type;
2188 bool swapped = evsel->needs_swap;
2189 const u64 *array;
2190 u16 max_size = event->header.size;
2191 const void *endp = (void *)event + max_size;
2192 u64 sz;
2193
2194 /*
2195 * used for cross-endian analysis. See git commit 65014ab3
2196 * for why this goofiness is needed.
2197 */
2198 union u64_swap u;
2199
2200 memset(data, 0, sizeof(*data));
2201 data->cpu = data->pid = data->tid = -1;
2202 data->stream_id = data->id = data->time = -1ULL;
2203 data->period = evsel->attr.sample_period;
2204 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2205 data->misc = event->header.misc;
2206 data->id = -1ULL;
2207 data->data_src = PERF_MEM_DATA_SRC_NONE;
2208
2209 if (event->header.type != PERF_RECORD_SAMPLE) {
2210 if (!evsel->attr.sample_id_all)
2211 return 0;
2212 return perf_evsel__parse_id_sample(evsel, event, data);
2213 }
2214
2215 array = event->sample.array;
2216
2217 if (perf_event__check_size(event, evsel->sample_size))
2218 return -EFAULT;
2219
2220 if (type & PERF_SAMPLE_IDENTIFIER) {
2221 data->id = *array;
2222 array++;
2223 }
2224
2225 if (type & PERF_SAMPLE_IP) {
2226 data->ip = *array;
2227 array++;
2228 }
2229
2230 if (type & PERF_SAMPLE_TID) {
2231 u.val64 = *array;
2232 if (swapped) {
2233 /* undo swap of u64, then swap on individual u32s */
2234 u.val64 = bswap_64(u.val64);
2235 u.val32[0] = bswap_32(u.val32[0]);
2236 u.val32[1] = bswap_32(u.val32[1]);
2237 }
2238
2239 data->pid = u.val32[0];
2240 data->tid = u.val32[1];
2241 array++;
2242 }
2243
2244 if (type & PERF_SAMPLE_TIME) {
2245 data->time = *array;
2246 array++;
2247 }
2248
2249 if (type & PERF_SAMPLE_ADDR) {
2250 data->addr = *array;
2251 array++;
2252 }
2253
2254 if (type & PERF_SAMPLE_ID) {
2255 data->id = *array;
2256 array++;
2257 }
2258
2259 if (type & PERF_SAMPLE_STREAM_ID) {
2260 data->stream_id = *array;
2261 array++;
2262 }
2263
2264 if (type & PERF_SAMPLE_CPU) {
2265
2266 u.val64 = *array;
2267 if (swapped) {
2268 /* undo swap of u64, then swap on individual u32s */
2269 u.val64 = bswap_64(u.val64);
2270 u.val32[0] = bswap_32(u.val32[0]);
2271 }
2272
2273 data->cpu = u.val32[0];
2274 array++;
2275 }
2276
2277 if (type & PERF_SAMPLE_PERIOD) {
2278 data->period = *array;
2279 array++;
2280 }
2281
2282 if (type & PERF_SAMPLE_READ) {
2283 u64 read_format = evsel->attr.read_format;
2284
2285 OVERFLOW_CHECK_u64(array);
2286 if (read_format & PERF_FORMAT_GROUP)
2287 data->read.group.nr = *array;
2288 else
2289 data->read.one.value = *array;
2290
2291 array++;
2292
2293 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2294 OVERFLOW_CHECK_u64(array);
2295 data->read.time_enabled = *array;
2296 array++;
2297 }
2298
2299 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2300 OVERFLOW_CHECK_u64(array);
2301 data->read.time_running = *array;
2302 array++;
2303 }
2304
2305 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2306 if (read_format & PERF_FORMAT_GROUP) {
2307 const u64 max_group_nr = UINT64_MAX /
2308 sizeof(struct sample_read_value);
2309
2310 if (data->read.group.nr > max_group_nr)
2311 return -EFAULT;
2312 sz = data->read.group.nr *
2313 sizeof(struct sample_read_value);
2314 OVERFLOW_CHECK(array, sz, max_size);
2315 data->read.group.values =
2316 (struct sample_read_value *)array;
2317 array = (void *)array + sz;
2318 } else {
2319 OVERFLOW_CHECK_u64(array);
2320 data->read.one.id = *array;
2321 array++;
2322 }
2323 }
2324
2325 if (evsel__has_callchain(evsel)) {
2326 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2327
2328 OVERFLOW_CHECK_u64(array);
2329 data->callchain = (struct ip_callchain *)array++;
2330 if (data->callchain->nr > max_callchain_nr)
2331 return -EFAULT;
2332 sz = data->callchain->nr * sizeof(u64);
2333 OVERFLOW_CHECK(array, sz, max_size);
2334 array = (void *)array + sz;
2335 }
2336
2337 if (type & PERF_SAMPLE_RAW) {
2338 OVERFLOW_CHECK_u64(array);
2339 u.val64 = *array;
2340
2341 /*
2342 * Undo swap of u64, then swap on individual u32s,
2343 * get the size of the raw area and undo all of the
2344 * swap. The pevent interface handles endianity by
2345 * itself.
2346 */
2347 if (swapped) {
2348 u.val64 = bswap_64(u.val64);
2349 u.val32[0] = bswap_32(u.val32[0]);
2350 u.val32[1] = bswap_32(u.val32[1]);
2351 }
2352 data->raw_size = u.val32[0];
2353
2354 /*
2355 * The raw data is aligned on 64bits including the
2356 * u32 size, so it's safe to use mem_bswap_64.
2357 */
2358 if (swapped)
2359 mem_bswap_64((void *) array, data->raw_size);
2360
2361 array = (void *)array + sizeof(u32);
2362
2363 OVERFLOW_CHECK(array, data->raw_size, max_size);
2364 data->raw_data = (void *)array;
2365 array = (void *)array + data->raw_size;
2366 }
2367
2368 if (type & PERF_SAMPLE_BRANCH_STACK) {
2369 const u64 max_branch_nr = UINT64_MAX /
2370 sizeof(struct branch_entry);
2371
2372 OVERFLOW_CHECK_u64(array);
2373 data->branch_stack = (struct branch_stack *)array++;
2374
2375 if (data->branch_stack->nr > max_branch_nr)
2376 return -EFAULT;
2377 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2378 OVERFLOW_CHECK(array, sz, max_size);
2379 array = (void *)array + sz;
2380 }
2381
2382 if (type & PERF_SAMPLE_REGS_USER) {
2383 OVERFLOW_CHECK_u64(array);
2384 data->user_regs.abi = *array;
2385 array++;
2386
2387 if (data->user_regs.abi) {
2388 u64 mask = evsel->attr.sample_regs_user;
2389
2390 sz = hweight64(mask) * sizeof(u64);
2391 OVERFLOW_CHECK(array, sz, max_size);
2392 data->user_regs.mask = mask;
2393 data->user_regs.regs = (u64 *)array;
2394 array = (void *)array + sz;
2395 }
2396 }
2397
2398 if (type & PERF_SAMPLE_STACK_USER) {
2399 OVERFLOW_CHECK_u64(array);
2400 sz = *array++;
2401
2402 data->user_stack.offset = ((char *)(array - 1)
2403 - (char *) event);
2404
2405 if (!sz) {
2406 data->user_stack.size = 0;
2407 } else {
2408 OVERFLOW_CHECK(array, sz, max_size);
2409 data->user_stack.data = (char *)array;
2410 array = (void *)array + sz;
2411 OVERFLOW_CHECK_u64(array);
2412 data->user_stack.size = *array++;
2413 if (WARN_ONCE(data->user_stack.size > sz,
2414 "user stack dump failure\n"))
2415 return -EFAULT;
2416 }
2417 }
2418
2419 if (type & PERF_SAMPLE_WEIGHT) {
2420 OVERFLOW_CHECK_u64(array);
2421 data->weight = *array;
2422 array++;
2423 }
2424
2425 if (type & PERF_SAMPLE_DATA_SRC) {
2426 OVERFLOW_CHECK_u64(array);
2427 data->data_src = *array;
2428 array++;
2429 }
2430
2431 if (type & PERF_SAMPLE_TRANSACTION) {
2432 OVERFLOW_CHECK_u64(array);
2433 data->transaction = *array;
2434 array++;
2435 }
2436
2437 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2438 if (type & PERF_SAMPLE_REGS_INTR) {
2439 OVERFLOW_CHECK_u64(array);
2440 data->intr_regs.abi = *array;
2441 array++;
2442
2443 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2444 u64 mask = evsel->attr.sample_regs_intr;
2445
2446 sz = hweight64(mask) * sizeof(u64);
2447 OVERFLOW_CHECK(array, sz, max_size);
2448 data->intr_regs.mask = mask;
2449 data->intr_regs.regs = (u64 *)array;
2450 array = (void *)array + sz;
2451 }
2452 }
2453
2454 data->phys_addr = 0;
2455 if (type & PERF_SAMPLE_PHYS_ADDR) {
2456 data->phys_addr = *array;
2457 array++;
2458 }
2459
2460 return 0;
2461 }
2462
2463 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2464 union perf_event *event,
2465 u64 *timestamp)
2466 {
2467 u64 type = evsel->attr.sample_type;
2468 const u64 *array;
2469
2470 if (!(type & PERF_SAMPLE_TIME))
2471 return -1;
2472
2473 if (event->header.type != PERF_RECORD_SAMPLE) {
2474 struct perf_sample data = {
2475 .time = -1ULL,
2476 };
2477
2478 if (!evsel->attr.sample_id_all)
2479 return -1;
2480 if (perf_evsel__parse_id_sample(evsel, event, &data))
2481 return -1;
2482
2483 *timestamp = data.time;
2484 return 0;
2485 }
2486
2487 array = event->sample.array;
2488
2489 if (perf_event__check_size(event, evsel->sample_size))
2490 return -EFAULT;
2491
2492 if (type & PERF_SAMPLE_IDENTIFIER)
2493 array++;
2494
2495 if (type & PERF_SAMPLE_IP)
2496 array++;
2497
2498 if (type & PERF_SAMPLE_TID)
2499 array++;
2500
2501 if (type & PERF_SAMPLE_TIME)
2502 *timestamp = *array;
2503
2504 return 0;
2505 }
2506
2507 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2508 u64 read_format)
2509 {
2510 size_t sz, result = sizeof(struct sample_event);
2511
2512 if (type & PERF_SAMPLE_IDENTIFIER)
2513 result += sizeof(u64);
2514
2515 if (type & PERF_SAMPLE_IP)
2516 result += sizeof(u64);
2517
2518 if (type & PERF_SAMPLE_TID)
2519 result += sizeof(u64);
2520
2521 if (type & PERF_SAMPLE_TIME)
2522 result += sizeof(u64);
2523
2524 if (type & PERF_SAMPLE_ADDR)
2525 result += sizeof(u64);
2526
2527 if (type & PERF_SAMPLE_ID)
2528 result += sizeof(u64);
2529
2530 if (type & PERF_SAMPLE_STREAM_ID)
2531 result += sizeof(u64);
2532
2533 if (type & PERF_SAMPLE_CPU)
2534 result += sizeof(u64);
2535
2536 if (type & PERF_SAMPLE_PERIOD)
2537 result += sizeof(u64);
2538
2539 if (type & PERF_SAMPLE_READ) {
2540 result += sizeof(u64);
2541 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2542 result += sizeof(u64);
2543 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2544 result += sizeof(u64);
2545 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2546 if (read_format & PERF_FORMAT_GROUP) {
2547 sz = sample->read.group.nr *
2548 sizeof(struct sample_read_value);
2549 result += sz;
2550 } else {
2551 result += sizeof(u64);
2552 }
2553 }
2554
2555 if (type & PERF_SAMPLE_CALLCHAIN) {
2556 sz = (sample->callchain->nr + 1) * sizeof(u64);
2557 result += sz;
2558 }
2559
2560 if (type & PERF_SAMPLE_RAW) {
2561 result += sizeof(u32);
2562 result += sample->raw_size;
2563 }
2564
2565 if (type & PERF_SAMPLE_BRANCH_STACK) {
2566 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2567 sz += sizeof(u64);
2568 result += sz;
2569 }
2570
2571 if (type & PERF_SAMPLE_REGS_USER) {
2572 if (sample->user_regs.abi) {
2573 result += sizeof(u64);
2574 sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2575 result += sz;
2576 } else {
2577 result += sizeof(u64);
2578 }
2579 }
2580
2581 if (type & PERF_SAMPLE_STACK_USER) {
2582 sz = sample->user_stack.size;
2583 result += sizeof(u64);
2584 if (sz) {
2585 result += sz;
2586 result += sizeof(u64);
2587 }
2588 }
2589
2590 if (type & PERF_SAMPLE_WEIGHT)
2591 result += sizeof(u64);
2592
2593 if (type & PERF_SAMPLE_DATA_SRC)
2594 result += sizeof(u64);
2595
2596 if (type & PERF_SAMPLE_TRANSACTION)
2597 result += sizeof(u64);
2598
2599 if (type & PERF_SAMPLE_REGS_INTR) {
2600 if (sample->intr_regs.abi) {
2601 result += sizeof(u64);
2602 sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2603 result += sz;
2604 } else {
2605 result += sizeof(u64);
2606 }
2607 }
2608
2609 if (type & PERF_SAMPLE_PHYS_ADDR)
2610 result += sizeof(u64);
2611
2612 return result;
2613 }
2614
2615 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2616 u64 read_format,
2617 const struct perf_sample *sample)
2618 {
2619 u64 *array;
2620 size_t sz;
2621 /*
2622 * used for cross-endian analysis. See git commit 65014ab3
2623 * for why this goofiness is needed.
2624 */
2625 union u64_swap u;
2626
2627 array = event->sample.array;
2628
2629 if (type & PERF_SAMPLE_IDENTIFIER) {
2630 *array = sample->id;
2631 array++;
2632 }
2633
2634 if (type & PERF_SAMPLE_IP) {
2635 *array = sample->ip;
2636 array++;
2637 }
2638
2639 if (type & PERF_SAMPLE_TID) {
2640 u.val32[0] = sample->pid;
2641 u.val32[1] = sample->tid;
2642 *array = u.val64;
2643 array++;
2644 }
2645
2646 if (type & PERF_SAMPLE_TIME) {
2647 *array = sample->time;
2648 array++;
2649 }
2650
2651 if (type & PERF_SAMPLE_ADDR) {
2652 *array = sample->addr;
2653 array++;
2654 }
2655
2656 if (type & PERF_SAMPLE_ID) {
2657 *array = sample->id;
2658 array++;
2659 }
2660
2661 if (type & PERF_SAMPLE_STREAM_ID) {
2662 *array = sample->stream_id;
2663 array++;
2664 }
2665
2666 if (type & PERF_SAMPLE_CPU) {
2667 u.val32[0] = sample->cpu;
2668 u.val32[1] = 0;
2669 *array = u.val64;
2670 array++;
2671 }
2672
2673 if (type & PERF_SAMPLE_PERIOD) {
2674 *array = sample->period;
2675 array++;
2676 }
2677
2678 if (type & PERF_SAMPLE_READ) {
2679 if (read_format & PERF_FORMAT_GROUP)
2680 *array = sample->read.group.nr;
2681 else
2682 *array = sample->read.one.value;
2683 array++;
2684
2685 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2686 *array = sample->read.time_enabled;
2687 array++;
2688 }
2689
2690 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2691 *array = sample->read.time_running;
2692 array++;
2693 }
2694
2695 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2696 if (read_format & PERF_FORMAT_GROUP) {
2697 sz = sample->read.group.nr *
2698 sizeof(struct sample_read_value);
2699 memcpy(array, sample->read.group.values, sz);
2700 array = (void *)array + sz;
2701 } else {
2702 *array = sample->read.one.id;
2703 array++;
2704 }
2705 }
2706
2707 if (type & PERF_SAMPLE_CALLCHAIN) {
2708 sz = (sample->callchain->nr + 1) * sizeof(u64);
2709 memcpy(array, sample->callchain, sz);
2710 array = (void *)array + sz;
2711 }
2712
2713 if (type & PERF_SAMPLE_RAW) {
2714 u.val32[0] = sample->raw_size;
2715 *array = u.val64;
2716 array = (void *)array + sizeof(u32);
2717
2718 memcpy(array, sample->raw_data, sample->raw_size);
2719 array = (void *)array + sample->raw_size;
2720 }
2721
2722 if (type & PERF_SAMPLE_BRANCH_STACK) {
2723 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2724 sz += sizeof(u64);
2725 memcpy(array, sample->branch_stack, sz);
2726 array = (void *)array + sz;
2727 }
2728
2729 if (type & PERF_SAMPLE_REGS_USER) {
2730 if (sample->user_regs.abi) {
2731 *array++ = sample->user_regs.abi;
2732 sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2733 memcpy(array, sample->user_regs.regs, sz);
2734 array = (void *)array + sz;
2735 } else {
2736 *array++ = 0;
2737 }
2738 }
2739
2740 if (type & PERF_SAMPLE_STACK_USER) {
2741 sz = sample->user_stack.size;
2742 *array++ = sz;
2743 if (sz) {
2744 memcpy(array, sample->user_stack.data, sz);
2745 array = (void *)array + sz;
2746 *array++ = sz;
2747 }
2748 }
2749
2750 if (type & PERF_SAMPLE_WEIGHT) {
2751 *array = sample->weight;
2752 array++;
2753 }
2754
2755 if (type & PERF_SAMPLE_DATA_SRC) {
2756 *array = sample->data_src;
2757 array++;
2758 }
2759
2760 if (type & PERF_SAMPLE_TRANSACTION) {
2761 *array = sample->transaction;
2762 array++;
2763 }
2764
2765 if (type & PERF_SAMPLE_REGS_INTR) {
2766 if (sample->intr_regs.abi) {
2767 *array++ = sample->intr_regs.abi;
2768 sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2769 memcpy(array, sample->intr_regs.regs, sz);
2770 array = (void *)array + sz;
2771 } else {
2772 *array++ = 0;
2773 }
2774 }
2775
2776 if (type & PERF_SAMPLE_PHYS_ADDR) {
2777 *array = sample->phys_addr;
2778 array++;
2779 }
2780
2781 return 0;
2782 }
2783
2784 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2785 {
2786 return tep_find_field(evsel->tp_format, name);
2787 }
2788
2789 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2790 const char *name)
2791 {
2792 struct tep_format_field *field = perf_evsel__field(evsel, name);
2793 int offset;
2794
2795 if (!field)
2796 return NULL;
2797
2798 offset = field->offset;
2799
2800 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2801 offset = *(int *)(sample->raw_data + field->offset);
2802 offset &= 0xffff;
2803 }
2804
2805 return sample->raw_data + offset;
2806 }
2807
2808 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2809 bool needs_swap)
2810 {
2811 u64 value;
2812 void *ptr = sample->raw_data + field->offset;
2813
2814 switch (field->size) {
2815 case 1:
2816 return *(u8 *)ptr;
2817 case 2:
2818 value = *(u16 *)ptr;
2819 break;
2820 case 4:
2821 value = *(u32 *)ptr;
2822 break;
2823 case 8:
2824 memcpy(&value, ptr, sizeof(u64));
2825 break;
2826 default:
2827 return 0;
2828 }
2829
2830 if (!needs_swap)
2831 return value;
2832
2833 switch (field->size) {
2834 case 2:
2835 return bswap_16(value);
2836 case 4:
2837 return bswap_32(value);
2838 case 8:
2839 return bswap_64(value);
2840 default:
2841 return 0;
2842 }
2843
2844 return 0;
2845 }
2846
2847 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2848 const char *name)
2849 {
2850 struct tep_format_field *field = perf_evsel__field(evsel, name);
2851
2852 if (!field)
2853 return 0;
2854
2855 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2856 }
2857
2858 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2859 char *msg, size_t msgsize)
2860 {
2861 int paranoid;
2862
2863 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2864 evsel->attr.type == PERF_TYPE_HARDWARE &&
2865 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2866 /*
2867 * If it's cycles then fall back to hrtimer based
2868 * cpu-clock-tick sw counter, which is always available even if
2869 * no PMU support.
2870 *
2871 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2872 * b0a873e).
2873 */
2874 scnprintf(msg, msgsize, "%s",
2875 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2876
2877 evsel->attr.type = PERF_TYPE_SOFTWARE;
2878 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2879
2880 zfree(&evsel->name);
2881 return true;
2882 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2883 (paranoid = perf_event_paranoid()) > 1) {
2884 const char *name = perf_evsel__name(evsel);
2885 char *new_name;
2886 const char *sep = ":";
2887
2888 /* Is there already the separator in the name. */
2889 if (strchr(name, '/') ||
2890 strchr(name, ':'))
2891 sep = "";
2892
2893 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2894 return false;
2895
2896 if (evsel->name)
2897 free(evsel->name);
2898 evsel->name = new_name;
2899 scnprintf(msg, msgsize,
2900 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2901 evsel->attr.exclude_kernel = 1;
2902
2903 return true;
2904 }
2905
2906 return false;
2907 }
2908
2909 static bool find_process(const char *name)
2910 {
2911 size_t len = strlen(name);
2912 DIR *dir;
2913 struct dirent *d;
2914 int ret = -1;
2915
2916 dir = opendir(procfs__mountpoint());
2917 if (!dir)
2918 return false;
2919
2920 /* Walk through the directory. */
2921 while (ret && (d = readdir(dir)) != NULL) {
2922 char path[PATH_MAX];
2923 char *data;
2924 size_t size;
2925
2926 if ((d->d_type != DT_DIR) ||
2927 !strcmp(".", d->d_name) ||
2928 !strcmp("..", d->d_name))
2929 continue;
2930
2931 scnprintf(path, sizeof(path), "%s/%s/comm",
2932 procfs__mountpoint(), d->d_name);
2933
2934 if (filename__read_str(path, &data, &size))
2935 continue;
2936
2937 ret = strncmp(name, data, len);
2938 free(data);
2939 }
2940
2941 closedir(dir);
2942 return ret ? false : true;
2943 }
2944
2945 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2946 int err, char *msg, size_t size)
2947 {
2948 char sbuf[STRERR_BUFSIZE];
2949 int printed = 0;
2950
2951 switch (err) {
2952 case EPERM:
2953 case EACCES:
2954 if (err == EPERM)
2955 printed = scnprintf(msg, size,
2956 "No permission to enable %s event.\n\n",
2957 perf_evsel__name(evsel));
2958
2959 return scnprintf(msg + printed, size - printed,
2960 "You may not have permission to collect %sstats.\n\n"
2961 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2962 "which controls use of the performance events system by\n"
2963 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2964 "The current value is %d:\n\n"
2965 " -1: Allow use of (almost) all events by all users\n"
2966 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2967 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2968 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2969 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2970 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2971 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2972 " kernel.perf_event_paranoid = -1\n" ,
2973 target->system_wide ? "system-wide " : "",
2974 perf_event_paranoid());
2975 case ENOENT:
2976 return scnprintf(msg, size, "The %s event is not supported.",
2977 perf_evsel__name(evsel));
2978 case EMFILE:
2979 return scnprintf(msg, size, "%s",
2980 "Too many events are opened.\n"
2981 "Probably the maximum number of open file descriptors has been reached.\n"
2982 "Hint: Try again after reducing the number of events.\n"
2983 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2984 case ENOMEM:
2985 if (evsel__has_callchain(evsel) &&
2986 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2987 return scnprintf(msg, size,
2988 "Not enough memory to setup event with callchain.\n"
2989 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2990 "Hint: Current value: %d", sysctl__max_stack());
2991 break;
2992 case ENODEV:
2993 if (target->cpu_list)
2994 return scnprintf(msg, size, "%s",
2995 "No such device - did you specify an out-of-range profile CPU?");
2996 break;
2997 case EOPNOTSUPP:
2998 if (evsel->attr.sample_period != 0)
2999 return scnprintf(msg, size,
3000 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3001 perf_evsel__name(evsel));
3002 if (evsel->attr.precise_ip)
3003 return scnprintf(msg, size, "%s",
3004 "\'precise\' request may not be supported. Try removing 'p' modifier.");
3005 #if defined(__i386__) || defined(__x86_64__)
3006 if (evsel->attr.type == PERF_TYPE_HARDWARE)
3007 return scnprintf(msg, size, "%s",
3008 "No hardware sampling interrupt available.\n");
3009 #endif
3010 break;
3011 case EBUSY:
3012 if (find_process("oprofiled"))
3013 return scnprintf(msg, size,
3014 "The PMU counters are busy/taken by another profiler.\n"
3015 "We found oprofile daemon running, please stop it and try again.");
3016 break;
3017 case EINVAL:
3018 if (evsel->attr.write_backward && perf_missing_features.write_backward)
3019 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3020 if (perf_missing_features.clockid)
3021 return scnprintf(msg, size, "clockid feature not supported.");
3022 if (perf_missing_features.clockid_wrong)
3023 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3024 break;
3025 default:
3026 break;
3027 }
3028
3029 return scnprintf(msg, size,
3030 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3031 "/bin/dmesg | grep -i perf may provide additional information.\n",
3032 err, str_error_r(err, sbuf, sizeof(sbuf)),
3033 perf_evsel__name(evsel));
3034 }
3035
3036 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
3037 {
3038 if (evsel && evsel->evlist)
3039 return evsel->evlist->env;
3040 return NULL;
3041 }
3042
3043 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3044 {
3045 int cpu, thread;
3046
3047 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
3048 for (thread = 0; thread < xyarray__max_y(evsel->fd);
3049 thread++) {
3050 int fd = FD(evsel, cpu, thread);
3051
3052 if (perf_evlist__id_add_fd(evlist, evsel,
3053 cpu, thread, fd) < 0)
3054 return -1;
3055 }
3056 }
3057
3058 return 0;
3059 }
3060
3061 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3062 {
3063 struct cpu_map *cpus = evsel->cpus;
3064 struct thread_map *threads = evsel->threads;
3065
3066 if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
3067 return -ENOMEM;
3068
3069 return store_evsel_ids(evsel, evlist);
3070 }
Linux with added FPC-III support