]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - tools/perf/util/evsel.c
projects / mirror_ubuntu-bionic-kernel.git / blob
? search:


37ee1aebbf8a41fafe5fea0d4e7a066d384abc12
[mirror_ubuntu-bionic-kernel.git] / tools / perf / util / evsel.c
1 /*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9
10 #include <byteswap.h>
11 #include <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 static struct {
45 bool sample_id_all;
46 bool exclude_guest;
47 bool mmap2;
48 bool cloexec;
49 bool clockid;
50 bool clockid_wrong;
51 bool lbr_flags;
52 bool write_backward;
53 bool group_read;
54 } perf_missing_features;
55
56 static clockid_t clockid;
57
58 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
59 {
60 return 0;
61 }
62
63 void __weak test_attr__ready(void) { }
64
65 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
66 {
67 }
68
69 static struct {
70 size_t size;
71 int (*init)(struct perf_evsel *evsel);
72 void (*fini)(struct perf_evsel *evsel);
73 } perf_evsel__object = {
74 .size = sizeof(struct perf_evsel),
75 .init = perf_evsel__no_extra_init,
76 .fini = perf_evsel__no_extra_fini,
77 };
78
79 int perf_evsel__object_config(size_t object_size,
80 int (*init)(struct perf_evsel *evsel),
81 void (*fini)(struct perf_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->fd, x, y))
103
104 int __perf_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 * sample_event.
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 perf_evsel__calc_id_pos(struct perf_evsel *evsel)
181 {
182 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
183 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
184 }
185
186 void __perf_evsel__set_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__reset_sample_bit(struct perf_evsel *evsel,
197 enum perf_event_sample_format bit)
198 {
199 if (evsel->attr.sample_type & bit) {
200 evsel->attr.sample_type &= ~bit;
201 evsel->sample_size -= sizeof(u64);
202 perf_evsel__calc_id_pos(evsel);
203 }
204 }
205
206 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
207 bool can_sample_identifier)
208 {
209 if (can_sample_identifier) {
210 perf_evsel__reset_sample_bit(evsel, ID);
211 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
212 } else {
213 perf_evsel__set_sample_bit(evsel, ID);
214 }
215 evsel->attr.read_format |= PERF_FORMAT_ID;
216 }
217
218 /**
219 * perf_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 perf_evsel__is_function_event(struct perf_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 perf_evsel__init(struct perf_evsel *evsel,
237 struct perf_event_attr *attr, int idx)
238 {
239 evsel->idx = idx;
240 evsel->tracking = !idx;
241 evsel->attr = *attr;
242 evsel->leader = evsel;
243 evsel->unit = "";
244 evsel->scale = 1.0;
245 evsel->evlist = NULL;
246 evsel->bpf_fd = -1;
247 INIT_LIST_HEAD(&evsel->node);
248 INIT_LIST_HEAD(&evsel->config_terms);
249 perf_evsel__object.init(evsel);
250 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
251 perf_evsel__calc_id_pos(evsel);
252 evsel->cmdline_group_boundary = false;
253 evsel->metric_expr = NULL;
254 evsel->metric_name = NULL;
255 evsel->metric_events = NULL;
256 evsel->collect_stat = false;
257 }
258
259 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
260 {
261 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
262
263 if (evsel != NULL)
264 perf_evsel__init(evsel, attr, idx);
265
266 if (perf_evsel__is_bpf_output(evsel)) {
267 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
268 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
269 evsel->attr.sample_period = 1;
270 }
271
272 return evsel;
273 }
274
275 static bool perf_event_can_profile_kernel(void)
276 {
277 return geteuid() == 0 || perf_event_paranoid() == -1;
278 }
279
280 struct perf_evsel *perf_evsel__new_cycles(bool precise)
281 {
282 struct perf_event_attr attr = {
283 .type = PERF_TYPE_HARDWARE,
284 .config = PERF_COUNT_HW_CPU_CYCLES,
285 .exclude_kernel = !perf_event_can_profile_kernel(),
286 };
287 struct perf_evsel *evsel;
288
289 event_attr_init(&attr);
290
291 if (!precise)
292 goto new_event;
293 /*
294 * Unnamed union member, not supported as struct member named
295 * initializer in older compilers such as gcc 4.4.7
296 *
297 * Just for probing the precise_ip:
298 */
299 attr.sample_period = 1;
300
301 perf_event_attr__set_max_precise_ip(&attr);
302 /*
303 * Now let the usual logic to set up the perf_event_attr defaults
304 * to kick in when we return and before perf_evsel__open() is called.
305 */
306 attr.sample_period = 0;
307 new_event:
308 evsel = perf_evsel__new(&attr);
309 if (evsel == NULL)
310 goto out;
311
312 /* use asprintf() because free(evsel) assumes name is allocated */
313 if (asprintf(&evsel->name, "cycles%s%s%.*s",
314 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
315 attr.exclude_kernel ? "u" : "",
316 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
317 goto error_free;
318 out:
319 return evsel;
320 error_free:
321 perf_evsel__delete(evsel);
322 evsel = NULL;
323 goto out;
324 }
325
326 /*
327 * Returns pointer with encoded error via <linux/err.h> interface.
328 */
329 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
330 {
331 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
332 int err = -ENOMEM;
333
334 if (evsel == NULL) {
335 goto out_err;
336 } else {
337 struct perf_event_attr attr = {
338 .type = PERF_TYPE_TRACEPOINT,
339 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
340 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
341 };
342
343 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
344 goto out_free;
345
346 evsel->tp_format = trace_event__tp_format(sys, name);
347 if (IS_ERR(evsel->tp_format)) {
348 err = PTR_ERR(evsel->tp_format);
349 goto out_free;
350 }
351
352 event_attr_init(&attr);
353 attr.config = evsel->tp_format->id;
354 attr.sample_period = 1;
355 perf_evsel__init(evsel, &attr, idx);
356 }
357
358 return evsel;
359
360 out_free:
361 zfree(&evsel->name);
362 free(evsel);
363 out_err:
364 return ERR_PTR(err);
365 }
366
367 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
368 "cycles",
369 "instructions",
370 "cache-references",
371 "cache-misses",
372 "branches",
373 "branch-misses",
374 "bus-cycles",
375 "stalled-cycles-frontend",
376 "stalled-cycles-backend",
377 "ref-cycles",
378 };
379
380 static const char *__perf_evsel__hw_name(u64 config)
381 {
382 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
383 return perf_evsel__hw_names[config];
384
385 return "unknown-hardware";
386 }
387
388 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
389 {
390 int colon = 0, r = 0;
391 struct perf_event_attr *attr = &evsel->attr;
392 bool exclude_guest_default = false;
393
394 #define MOD_PRINT(context, mod) do { \
395 if (!attr->exclude_##context) { \
396 if (!colon) colon = ++r; \
397 r += scnprintf(bf + r, size - r, "%c", mod); \
398 } } while(0)
399
400 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
401 MOD_PRINT(kernel, 'k');
402 MOD_PRINT(user, 'u');
403 MOD_PRINT(hv, 'h');
404 exclude_guest_default = true;
405 }
406
407 if (attr->precise_ip) {
408 if (!colon)
409 colon = ++r;
410 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
411 exclude_guest_default = true;
412 }
413
414 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
415 MOD_PRINT(host, 'H');
416 MOD_PRINT(guest, 'G');
417 }
418 #undef MOD_PRINT
419 if (colon)
420 bf[colon - 1] = ':';
421 return r;
422 }
423
424 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
425 {
426 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
427 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
428 }
429
430 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
431 "cpu-clock",
432 "task-clock",
433 "page-faults",
434 "context-switches",
435 "cpu-migrations",
436 "minor-faults",
437 "major-faults",
438 "alignment-faults",
439 "emulation-faults",
440 "dummy",
441 };
442
443 static const char *__perf_evsel__sw_name(u64 config)
444 {
445 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
446 return perf_evsel__sw_names[config];
447 return "unknown-software";
448 }
449
450 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
451 {
452 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
453 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
454 }
455
456 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
457 {
458 int r;
459
460 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
461
462 if (type & HW_BREAKPOINT_R)
463 r += scnprintf(bf + r, size - r, "r");
464
465 if (type & HW_BREAKPOINT_W)
466 r += scnprintf(bf + r, size - r, "w");
467
468 if (type & HW_BREAKPOINT_X)
469 r += scnprintf(bf + r, size - r, "x");
470
471 return r;
472 }
473
474 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
475 {
476 struct perf_event_attr *attr = &evsel->attr;
477 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
478 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
479 }
480
481 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
482 [PERF_EVSEL__MAX_ALIASES] = {
483 { "L1-dcache", "l1-d", "l1d", "L1-data", },
484 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
485 { "LLC", "L2", },
486 { "dTLB", "d-tlb", "Data-TLB", },
487 { "iTLB", "i-tlb", "Instruction-TLB", },
488 { "branch", "branches", "bpu", "btb", "bpc", },
489 { "node", },
490 };
491
492 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
493 [PERF_EVSEL__MAX_ALIASES] = {
494 { "load", "loads", "read", },
495 { "store", "stores", "write", },
496 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
497 };
498
499 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
500 [PERF_EVSEL__MAX_ALIASES] = {
501 { "refs", "Reference", "ops", "access", },
502 { "misses", "miss", },
503 };
504
505 #define C(x) PERF_COUNT_HW_CACHE_##x
506 #define CACHE_READ (1 << C(OP_READ))
507 #define CACHE_WRITE (1 << C(OP_WRITE))
508 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
509 #define COP(x) (1 << x)
510
511 /*
512 * cache operartion stat
513 * L1I : Read and prefetch only
514 * ITLB and BPU : Read-only
515 */
516 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
517 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
518 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
519 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
520 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
521 [C(ITLB)] = (CACHE_READ),
522 [C(BPU)] = (CACHE_READ),
523 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
524 };
525
526 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
527 {
528 if (perf_evsel__hw_cache_stat[type] & COP(op))
529 return true; /* valid */
530 else
531 return false; /* invalid */
532 }
533
534 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
535 char *bf, size_t size)
536 {
537 if (result) {
538 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
539 perf_evsel__hw_cache_op[op][0],
540 perf_evsel__hw_cache_result[result][0]);
541 }
542
543 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
544 perf_evsel__hw_cache_op[op][1]);
545 }
546
547 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
548 {
549 u8 op, result, type = (config >> 0) & 0xff;
550 const char *err = "unknown-ext-hardware-cache-type";
551
552 if (type >= PERF_COUNT_HW_CACHE_MAX)
553 goto out_err;
554
555 op = (config >> 8) & 0xff;
556 err = "unknown-ext-hardware-cache-op";
557 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
558 goto out_err;
559
560 result = (config >> 16) & 0xff;
561 err = "unknown-ext-hardware-cache-result";
562 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
563 goto out_err;
564
565 err = "invalid-cache";
566 if (!perf_evsel__is_cache_op_valid(type, op))
567 goto out_err;
568
569 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
570 out_err:
571 return scnprintf(bf, size, "%s", err);
572 }
573
574 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
575 {
576 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
577 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
578 }
579
580 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
581 {
582 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
583 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
584 }
585
586 const char *perf_evsel__name(struct perf_evsel *evsel)
587 {
588 char bf[128];
589
590 if (evsel->name)
591 return evsel->name;
592
593 switch (evsel->attr.type) {
594 case PERF_TYPE_RAW:
595 perf_evsel__raw_name(evsel, bf, sizeof(bf));
596 break;
597
598 case PERF_TYPE_HARDWARE:
599 perf_evsel__hw_name(evsel, bf, sizeof(bf));
600 break;
601
602 case PERF_TYPE_HW_CACHE:
603 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
604 break;
605
606 case PERF_TYPE_SOFTWARE:
607 perf_evsel__sw_name(evsel, bf, sizeof(bf));
608 break;
609
610 case PERF_TYPE_TRACEPOINT:
611 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
612 break;
613
614 case PERF_TYPE_BREAKPOINT:
615 perf_evsel__bp_name(evsel, bf, sizeof(bf));
616 break;
617
618 default:
619 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
620 evsel->attr.type);
621 break;
622 }
623
624 evsel->name = strdup(bf);
625
626 return evsel->name ?: "unknown";
627 }
628
629 const char *perf_evsel__group_name(struct perf_evsel *evsel)
630 {
631 return evsel->group_name ?: "anon group";
632 }
633
634 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
635 {
636 int ret;
637 struct perf_evsel *pos;
638 const char *group_name = perf_evsel__group_name(evsel);
639
640 ret = scnprintf(buf, size, "%s", group_name);
641
642 ret += scnprintf(buf + ret, size - ret, " { %s",
643 perf_evsel__name(evsel));
644
645 for_each_group_member(pos, evsel)
646 ret += scnprintf(buf + ret, size - ret, ", %s",
647 perf_evsel__name(pos));
648
649 ret += scnprintf(buf + ret, size - ret, " }");
650
651 return ret;
652 }
653
654 void perf_evsel__config_callchain(struct perf_evsel *evsel,
655 struct record_opts *opts,
656 struct callchain_param *param)
657 {
658 bool function = perf_evsel__is_function_event(evsel);
659 struct perf_event_attr *attr = &evsel->attr;
660
661 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
662
663 attr->sample_max_stack = param->max_stack;
664
665 if (param->record_mode == CALLCHAIN_LBR) {
666 if (!opts->branch_stack) {
667 if (attr->exclude_user) {
668 pr_warning("LBR callstack option is only available "
669 "to get user callchain information. "
670 "Falling back to framepointers.\n");
671 } else {
672 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
673 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
674 PERF_SAMPLE_BRANCH_CALL_STACK |
675 PERF_SAMPLE_BRANCH_NO_CYCLES |
676 PERF_SAMPLE_BRANCH_NO_FLAGS;
677 }
678 } else
679 pr_warning("Cannot use LBR callstack with branch stack. "
680 "Falling back to framepointers.\n");
681 }
682
683 if (param->record_mode == CALLCHAIN_DWARF) {
684 if (!function) {
685 perf_evsel__set_sample_bit(evsel, REGS_USER);
686 perf_evsel__set_sample_bit(evsel, STACK_USER);
687 attr->sample_regs_user |= PERF_REGS_MASK;
688 attr->sample_stack_user = param->dump_size;
689 attr->exclude_callchain_user = 1;
690 } else {
691 pr_info("Cannot use DWARF unwind for function trace event,"
692 " falling back to framepointers.\n");
693 }
694 }
695
696 if (function) {
697 pr_info("Disabling user space callchains for function trace event.\n");
698 attr->exclude_callchain_user = 1;
699 }
700 }
701
702 static void
703 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
704 struct callchain_param *param)
705 {
706 struct perf_event_attr *attr = &evsel->attr;
707
708 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
709 if (param->record_mode == CALLCHAIN_LBR) {
710 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
711 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
712 PERF_SAMPLE_BRANCH_CALL_STACK);
713 }
714 if (param->record_mode == CALLCHAIN_DWARF) {
715 perf_evsel__reset_sample_bit(evsel, REGS_USER);
716 perf_evsel__reset_sample_bit(evsel, STACK_USER);
717 }
718 }
719
720 static void apply_config_terms(struct perf_evsel *evsel,
721 struct record_opts *opts)
722 {
723 struct perf_evsel_config_term *term;
724 struct list_head *config_terms = &evsel->config_terms;
725 struct perf_event_attr *attr = &evsel->attr;
726 /* callgraph default */
727 struct callchain_param param = {
728 .record_mode = callchain_param.record_mode,
729 };
730 u32 dump_size = 0;
731 int max_stack = 0;
732 const char *callgraph_buf = NULL;
733
734 list_for_each_entry(term, config_terms, list) {
735 switch (term->type) {
736 case PERF_EVSEL__CONFIG_TERM_PERIOD:
737 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
738 attr->sample_period = term->val.period;
739 attr->freq = 0;
740 }
741 break;
742 case PERF_EVSEL__CONFIG_TERM_FREQ:
743 if (!(term->weak && opts->user_freq != UINT_MAX)) {
744 attr->sample_freq = term->val.freq;
745 attr->freq = 1;
746 }
747 break;
748 case PERF_EVSEL__CONFIG_TERM_TIME:
749 if (term->val.time)
750 perf_evsel__set_sample_bit(evsel, TIME);
751 else
752 perf_evsel__reset_sample_bit(evsel, TIME);
753 break;
754 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
755 callgraph_buf = term->val.callgraph;
756 break;
757 case PERF_EVSEL__CONFIG_TERM_BRANCH:
758 if (term->val.branch && strcmp(term->val.branch, "no")) {
759 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
760 parse_branch_str(term->val.branch,
761 &attr->branch_sample_type);
762 } else
763 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
764 break;
765 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
766 dump_size = term->val.stack_user;
767 break;
768 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
769 max_stack = term->val.max_stack;
770 break;
771 case PERF_EVSEL__CONFIG_TERM_INHERIT:
772 /*
773 * attr->inherit should has already been set by
774 * perf_evsel__config. If user explicitly set
775 * inherit using config terms, override global
776 * opt->no_inherit setting.
777 */
778 attr->inherit = term->val.inherit ? 1 : 0;
779 break;
780 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
781 attr->write_backward = term->val.overwrite ? 1 : 0;
782 break;
783 default:
784 break;
785 }
786 }
787
788 /* User explicitly set per-event callgraph, clear the old setting and reset. */
789 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
790 if (max_stack) {
791 param.max_stack = max_stack;
792 if (callgraph_buf == NULL)
793 callgraph_buf = "fp";
794 }
795
796 /* parse callgraph parameters */
797 if (callgraph_buf != NULL) {
798 if (!strcmp(callgraph_buf, "no")) {
799 param.enabled = false;
800 param.record_mode = CALLCHAIN_NONE;
801 } else {
802 param.enabled = true;
803 if (parse_callchain_record(callgraph_buf, &param)) {
804 pr_err("per-event callgraph setting for %s failed. "
805 "Apply callgraph global setting for it\n",
806 evsel->name);
807 return;
808 }
809 }
810 }
811 if (dump_size > 0) {
812 dump_size = round_up(dump_size, sizeof(u64));
813 param.dump_size = dump_size;
814 }
815
816 /* If global callgraph set, clear it */
817 if (callchain_param.enabled)
818 perf_evsel__reset_callgraph(evsel, &callchain_param);
819
820 /* set perf-event callgraph */
821 if (param.enabled)
822 perf_evsel__config_callchain(evsel, opts, &param);
823 }
824 }
825
826 /*
827 * The enable_on_exec/disabled value strategy:
828 *
829 * 1) For any type of traced program:
830 * - all independent events and group leaders are disabled
831 * - all group members are enabled
832 *
833 * Group members are ruled by group leaders. They need to
834 * be enabled, because the group scheduling relies on that.
835 *
836 * 2) For traced programs executed by perf:
837 * - all independent events and group leaders have
838 * enable_on_exec set
839 * - we don't specifically enable or disable any event during
840 * the record command
841 *
842 * Independent events and group leaders are initially disabled
843 * and get enabled by exec. Group members are ruled by group
844 * leaders as stated in 1).
845 *
846 * 3) For traced programs attached by perf (pid/tid):
847 * - we specifically enable or disable all events during
848 * the record command
849 *
850 * When attaching events to already running traced we
851 * enable/disable events specifically, as there's no
852 * initial traced exec call.
853 */
854 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
855 struct callchain_param *callchain)
856 {
857 struct perf_evsel *leader = evsel->leader;
858 struct perf_event_attr *attr = &evsel->attr;
859 int track = evsel->tracking;
860 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
861
862 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
863 attr->inherit = !opts->no_inherit;
864 attr->write_backward = opts->overwrite ? 1 : 0;
865
866 perf_evsel__set_sample_bit(evsel, IP);
867 perf_evsel__set_sample_bit(evsel, TID);
868
869 if (evsel->sample_read) {
870 perf_evsel__set_sample_bit(evsel, READ);
871
872 /*
873 * We need ID even in case of single event, because
874 * PERF_SAMPLE_READ process ID specific data.
875 */
876 perf_evsel__set_sample_id(evsel, false);
877
878 /*
879 * Apply group format only if we belong to group
880 * with more than one members.
881 */
882 if (leader->nr_members > 1) {
883 attr->read_format |= PERF_FORMAT_GROUP;
884 attr->inherit = 0;
885 }
886 }
887
888 /*
889 * We default some events to have a default interval. But keep
890 * it a weak assumption overridable by the user.
891 */
892 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
893 opts->user_interval != ULLONG_MAX)) {
894 if (opts->freq) {
895 perf_evsel__set_sample_bit(evsel, PERIOD);
896 attr->freq = 1;
897 attr->sample_freq = opts->freq;
898 } else {
899 attr->sample_period = opts->default_interval;
900 }
901 }
902
903 /*
904 * Disable sampling for all group members other
905 * than leader in case leader 'leads' the sampling.
906 */
907 if ((leader != evsel) && leader->sample_read) {
908 attr->sample_freq = 0;
909 attr->sample_period = 0;
910 }
911
912 if (opts->no_samples)
913 attr->sample_freq = 0;
914
915 if (opts->inherit_stat) {
916 evsel->attr.read_format |=
917 PERF_FORMAT_TOTAL_TIME_ENABLED |
918 PERF_FORMAT_TOTAL_TIME_RUNNING |
919 PERF_FORMAT_ID;
920 attr->inherit_stat = 1;
921 }
922
923 if (opts->sample_address) {
924 perf_evsel__set_sample_bit(evsel, ADDR);
925 attr->mmap_data = track;
926 }
927
928 /*
929 * We don't allow user space callchains for function trace
930 * event, due to issues with page faults while tracing page
931 * fault handler and its overall trickiness nature.
932 */
933 if (perf_evsel__is_function_event(evsel))
934 evsel->attr.exclude_callchain_user = 1;
935
936 if (callchain && callchain->enabled && !evsel->no_aux_samples)
937 perf_evsel__config_callchain(evsel, opts, callchain);
938
939 if (opts->sample_intr_regs) {
940 attr->sample_regs_intr = opts->sample_intr_regs;
941 perf_evsel__set_sample_bit(evsel, REGS_INTR);
942 }
943
944 if (opts->sample_user_regs) {
945 attr->sample_regs_user |= opts->sample_user_regs;
946 perf_evsel__set_sample_bit(evsel, REGS_USER);
947 }
948
949 if (target__has_cpu(&opts->target) || opts->sample_cpu)
950 perf_evsel__set_sample_bit(evsel, CPU);
951
952 if (opts->period)
953 perf_evsel__set_sample_bit(evsel, PERIOD);
954
955 /*
956 * When the user explicitly disabled time don't force it here.
957 */
958 if (opts->sample_time &&
959 (!perf_missing_features.sample_id_all &&
960 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
961 opts->sample_time_set)))
962 perf_evsel__set_sample_bit(evsel, TIME);
963
964 if (opts->raw_samples && !evsel->no_aux_samples) {
965 perf_evsel__set_sample_bit(evsel, TIME);
966 perf_evsel__set_sample_bit(evsel, RAW);
967 perf_evsel__set_sample_bit(evsel, CPU);
968 }
969
970 if (opts->sample_address)
971 perf_evsel__set_sample_bit(evsel, DATA_SRC);
972
973 if (opts->sample_phys_addr)
974 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
975
976 if (opts->no_buffering) {
977 attr->watermark = 0;
978 attr->wakeup_events = 1;
979 }
980 if (opts->branch_stack && !evsel->no_aux_samples) {
981 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
982 attr->branch_sample_type = opts->branch_stack;
983 }
984
985 if (opts->sample_weight)
986 perf_evsel__set_sample_bit(evsel, WEIGHT);
987
988 attr->task = track;
989 attr->mmap = track;
990 attr->mmap2 = track && !perf_missing_features.mmap2;
991 attr->comm = track;
992
993 if (opts->record_namespaces)
994 attr->namespaces = track;
995
996 if (opts->record_switch_events)
997 attr->context_switch = track;
998
999 if (opts->sample_transaction)
1000 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1001
1002 if (opts->running_time) {
1003 evsel->attr.read_format |=
1004 PERF_FORMAT_TOTAL_TIME_ENABLED |
1005 PERF_FORMAT_TOTAL_TIME_RUNNING;
1006 }
1007
1008 /*
1009 * XXX see the function comment above
1010 *
1011 * Disabling only independent events or group leaders,
1012 * keeping group members enabled.
1013 */
1014 if (perf_evsel__is_group_leader(evsel))
1015 attr->disabled = 1;
1016
1017 /*
1018 * Setting enable_on_exec for independent events and
1019 * group leaders for traced executed by perf.
1020 */
1021 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1022 !opts->initial_delay)
1023 attr->enable_on_exec = 1;
1024
1025 if (evsel->immediate) {
1026 attr->disabled = 0;
1027 attr->enable_on_exec = 0;
1028 }
1029
1030 clockid = opts->clockid;
1031 if (opts->use_clockid) {
1032 attr->use_clockid = 1;
1033 attr->clockid = opts->clockid;
1034 }
1035
1036 if (evsel->precise_max)
1037 perf_event_attr__set_max_precise_ip(attr);
1038
1039 if (opts->all_user) {
1040 attr->exclude_kernel = 1;
1041 attr->exclude_user = 0;
1042 }
1043
1044 if (opts->all_kernel) {
1045 attr->exclude_kernel = 0;
1046 attr->exclude_user = 1;
1047 }
1048
1049 /*
1050 * Apply event specific term settings,
1051 * it overloads any global configuration.
1052 */
1053 apply_config_terms(evsel, opts);
1054
1055 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1056 }
1057
1058 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1059 {
1060 if (evsel->system_wide)
1061 nthreads = 1;
1062
1063 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1064
1065 if (evsel->fd) {
1066 int cpu, thread;
1067 for (cpu = 0; cpu < ncpus; cpu++) {
1068 for (thread = 0; thread < nthreads; thread++) {
1069 FD(evsel, cpu, thread) = -1;
1070 }
1071 }
1072 }
1073
1074 return evsel->fd != NULL ? 0 : -ENOMEM;
1075 }
1076
1077 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1078 int ioc, void *arg)
1079 {
1080 int cpu, thread;
1081
1082 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1083 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1084 int fd = FD(evsel, cpu, thread),
1085 err = ioctl(fd, ioc, arg);
1086
1087 if (err)
1088 return err;
1089 }
1090 }
1091
1092 return 0;
1093 }
1094
1095 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1096 {
1097 return perf_evsel__run_ioctl(evsel,
1098 PERF_EVENT_IOC_SET_FILTER,
1099 (void *)filter);
1100 }
1101
1102 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1103 {
1104 char *new_filter = strdup(filter);
1105
1106 if (new_filter != NULL) {
1107 free(evsel->filter);
1108 evsel->filter = new_filter;
1109 return 0;
1110 }
1111
1112 return -1;
1113 }
1114
1115 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1116 const char *fmt, const char *filter)
1117 {
1118 char *new_filter;
1119
1120 if (evsel->filter == NULL)
1121 return perf_evsel__set_filter(evsel, filter);
1122
1123 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1124 free(evsel->filter);
1125 evsel->filter = new_filter;
1126 return 0;
1127 }
1128
1129 return -1;
1130 }
1131
1132 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1133 {
1134 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1135 }
1136
1137 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1138 {
1139 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1140 }
1141
1142 int perf_evsel__enable(struct perf_evsel *evsel)
1143 {
1144 return perf_evsel__run_ioctl(evsel,
1145 PERF_EVENT_IOC_ENABLE,
1146 0);
1147 }
1148
1149 int perf_evsel__disable(struct perf_evsel *evsel)
1150 {
1151 return perf_evsel__run_ioctl(evsel,
1152 PERF_EVENT_IOC_DISABLE,
1153 0);
1154 }
1155
1156 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1157 {
1158 if (ncpus == 0 || nthreads == 0)
1159 return 0;
1160
1161 if (evsel->system_wide)
1162 nthreads = 1;
1163
1164 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1165 if (evsel->sample_id == NULL)
1166 return -ENOMEM;
1167
1168 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1169 if (evsel->id == NULL) {
1170 xyarray__delete(evsel->sample_id);
1171 evsel->sample_id = NULL;
1172 return -ENOMEM;
1173 }
1174
1175 return 0;
1176 }
1177
1178 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1179 {
1180 xyarray__delete(evsel->fd);
1181 evsel->fd = NULL;
1182 }
1183
1184 static void perf_evsel__free_id(struct perf_evsel *evsel)
1185 {
1186 xyarray__delete(evsel->sample_id);
1187 evsel->sample_id = NULL;
1188 zfree(&evsel->id);
1189 }
1190
1191 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1192 {
1193 struct perf_evsel_config_term *term, *h;
1194
1195 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1196 list_del(&term->list);
1197 free(term);
1198 }
1199 }
1200
1201 void perf_evsel__close_fd(struct perf_evsel *evsel)
1202 {
1203 int cpu, thread;
1204
1205 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1206 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1207 close(FD(evsel, cpu, thread));
1208 FD(evsel, cpu, thread) = -1;
1209 }
1210 }
1211
1212 void perf_evsel__exit(struct perf_evsel *evsel)
1213 {
1214 assert(list_empty(&evsel->node));
1215 assert(evsel->evlist == NULL);
1216 perf_evsel__free_fd(evsel);
1217 perf_evsel__free_id(evsel);
1218 perf_evsel__free_config_terms(evsel);
1219 close_cgroup(evsel->cgrp);
1220 cpu_map__put(evsel->cpus);
1221 cpu_map__put(evsel->own_cpus);
1222 thread_map__put(evsel->threads);
1223 zfree(&evsel->group_name);
1224 zfree(&evsel->name);
1225 perf_evsel__object.fini(evsel);
1226 }
1227
1228 void perf_evsel__delete(struct perf_evsel *evsel)
1229 {
1230 perf_evsel__exit(evsel);
1231 free(evsel);
1232 }
1233
1234 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1235 struct perf_counts_values *count)
1236 {
1237 struct perf_counts_values tmp;
1238
1239 if (!evsel->prev_raw_counts)
1240 return;
1241
1242 if (cpu == -1) {
1243 tmp = evsel->prev_raw_counts->aggr;
1244 evsel->prev_raw_counts->aggr = *count;
1245 } else {
1246 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1247 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1248 }
1249
1250 count->val = count->val - tmp.val;
1251 count->ena = count->ena - tmp.ena;
1252 count->run = count->run - tmp.run;
1253 }
1254
1255 void perf_counts_values__scale(struct perf_counts_values *count,
1256 bool scale, s8 *pscaled)
1257 {
1258 s8 scaled = 0;
1259
1260 if (scale) {
1261 if (count->run == 0) {
1262 scaled = -1;
1263 count->val = 0;
1264 } else if (count->run < count->ena) {
1265 scaled = 1;
1266 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1267 }
1268 } else
1269 count->ena = count->run = 0;
1270
1271 if (pscaled)
1272 *pscaled = scaled;
1273 }
1274
1275 static int perf_evsel__read_size(struct perf_evsel *evsel)
1276 {
1277 u64 read_format = evsel->attr.read_format;
1278 int entry = sizeof(u64); /* value */
1279 int size = 0;
1280 int nr = 1;
1281
1282 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1283 size += sizeof(u64);
1284
1285 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1286 size += sizeof(u64);
1287
1288 if (read_format & PERF_FORMAT_ID)
1289 entry += sizeof(u64);
1290
1291 if (read_format & PERF_FORMAT_GROUP) {
1292 nr = evsel->nr_members;
1293 size += sizeof(u64);
1294 }
1295
1296 size += entry * nr;
1297 return size;
1298 }
1299
1300 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1301 struct perf_counts_values *count)
1302 {
1303 size_t size = perf_evsel__read_size(evsel);
1304
1305 memset(count, 0, sizeof(*count));
1306
1307 if (FD(evsel, cpu, thread) < 0)
1308 return -EINVAL;
1309
1310 if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1311 return -errno;
1312
1313 return 0;
1314 }
1315
1316 static int
1317 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1318 {
1319 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1320
1321 return perf_evsel__read(evsel, cpu, thread, count);
1322 }
1323
1324 static void
1325 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1326 u64 val, u64 ena, u64 run)
1327 {
1328 struct perf_counts_values *count;
1329
1330 count = perf_counts(counter->counts, cpu, thread);
1331
1332 count->val = val;
1333 count->ena = ena;
1334 count->run = run;
1335 count->loaded = true;
1336 }
1337
1338 static int
1339 perf_evsel__process_group_data(struct perf_evsel *leader,
1340 int cpu, int thread, u64 *data)
1341 {
1342 u64 read_format = leader->attr.read_format;
1343 struct sample_read_value *v;
1344 u64 nr, ena = 0, run = 0, i;
1345
1346 nr = *data++;
1347
1348 if (nr != (u64) leader->nr_members)
1349 return -EINVAL;
1350
1351 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1352 ena = *data++;
1353
1354 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1355 run = *data++;
1356
1357 v = (struct sample_read_value *) data;
1358
1359 perf_evsel__set_count(leader, cpu, thread,
1360 v[0].value, ena, run);
1361
1362 for (i = 1; i < nr; i++) {
1363 struct perf_evsel *counter;
1364
1365 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1366 if (!counter)
1367 return -EINVAL;
1368
1369 perf_evsel__set_count(counter, cpu, thread,
1370 v[i].value, ena, run);
1371 }
1372
1373 return 0;
1374 }
1375
1376 static int
1377 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1378 {
1379 struct perf_stat_evsel *ps = leader->stats;
1380 u64 read_format = leader->attr.read_format;
1381 int size = perf_evsel__read_size(leader);
1382 u64 *data = ps->group_data;
1383
1384 if (!(read_format & PERF_FORMAT_ID))
1385 return -EINVAL;
1386
1387 if (!perf_evsel__is_group_leader(leader))
1388 return -EINVAL;
1389
1390 if (!data) {
1391 data = zalloc(size);
1392 if (!data)
1393 return -ENOMEM;
1394
1395 ps->group_data = data;
1396 }
1397
1398 if (FD(leader, cpu, thread) < 0)
1399 return -EINVAL;
1400
1401 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1402 return -errno;
1403
1404 return perf_evsel__process_group_data(leader, cpu, thread, data);
1405 }
1406
1407 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1408 {
1409 u64 read_format = evsel->attr.read_format;
1410
1411 if (read_format & PERF_FORMAT_GROUP)
1412 return perf_evsel__read_group(evsel, cpu, thread);
1413 else
1414 return perf_evsel__read_one(evsel, cpu, thread);
1415 }
1416
1417 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1418 int cpu, int thread, bool scale)
1419 {
1420 struct perf_counts_values count;
1421 size_t nv = scale ? 3 : 1;
1422
1423 if (FD(evsel, cpu, thread) < 0)
1424 return -EINVAL;
1425
1426 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1427 return -ENOMEM;
1428
1429 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1430 return -errno;
1431
1432 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1433 perf_counts_values__scale(&count, scale, NULL);
1434 *perf_counts(evsel->counts, cpu, thread) = count;
1435 return 0;
1436 }
1437
1438 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1439 {
1440 struct perf_evsel *leader = evsel->leader;
1441 int fd;
1442
1443 if (perf_evsel__is_group_leader(evsel))
1444 return -1;
1445
1446 /*
1447 * Leader must be already processed/open,
1448 * if not it's a bug.
1449 */
1450 BUG_ON(!leader->fd);
1451
1452 fd = FD(leader, cpu, thread);
1453 BUG_ON(fd == -1);
1454
1455 return fd;
1456 }
1457
1458 struct bit_names {
1459 int bit;
1460 const char *name;
1461 };
1462
1463 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1464 {
1465 bool first_bit = true;
1466 int i = 0;
1467
1468 do {
1469 if (value & bits[i].bit) {
1470 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1471 first_bit = false;
1472 }
1473 } while (bits[++i].name != NULL);
1474 }
1475
1476 static void __p_sample_type(char *buf, size_t size, u64 value)
1477 {
1478 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1479 struct bit_names bits[] = {
1480 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1481 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1482 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1483 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1484 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1485 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1486 { .name = NULL, }
1487 };
1488 #undef bit_name
1489 __p_bits(buf, size, value, bits);
1490 }
1491
1492 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1493 {
1494 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1495 struct bit_names bits[] = {
1496 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1497 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1498 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1499 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1500 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1501 { .name = NULL, }
1502 };
1503 #undef bit_name
1504 __p_bits(buf, size, value, bits);
1505 }
1506
1507 static void __p_read_format(char *buf, size_t size, u64 value)
1508 {
1509 #define bit_name(n) { PERF_FORMAT_##n, #n }
1510 struct bit_names bits[] = {
1511 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1512 bit_name(ID), bit_name(GROUP),
1513 { .name = NULL, }
1514 };
1515 #undef bit_name
1516 __p_bits(buf, size, value, bits);
1517 }
1518
1519 #define BUF_SIZE 1024
1520
1521 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1522 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1523 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1524 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1525 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1526 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1527
1528 #define PRINT_ATTRn(_n, _f, _p) \
1529 do { \
1530 if (attr->_f) { \
1531 _p(attr->_f); \
1532 ret += attr__fprintf(fp, _n, buf, priv);\
1533 } \
1534 } while (0)
1535
1536 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1537
1538 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1539 attr__fprintf_f attr__fprintf, void *priv)
1540 {
1541 char buf[BUF_SIZE];
1542 int ret = 0;
1543
1544 PRINT_ATTRf(type, p_unsigned);
1545 PRINT_ATTRf(size, p_unsigned);
1546 PRINT_ATTRf(config, p_hex);
1547 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1548 PRINT_ATTRf(sample_type, p_sample_type);
1549 PRINT_ATTRf(read_format, p_read_format);
1550
1551 PRINT_ATTRf(disabled, p_unsigned);
1552 PRINT_ATTRf(inherit, p_unsigned);
1553 PRINT_ATTRf(pinned, p_unsigned);
1554 PRINT_ATTRf(exclusive, p_unsigned);
1555 PRINT_ATTRf(exclude_user, p_unsigned);
1556 PRINT_ATTRf(exclude_kernel, p_unsigned);
1557 PRINT_ATTRf(exclude_hv, p_unsigned);
1558 PRINT_ATTRf(exclude_idle, p_unsigned);
1559 PRINT_ATTRf(mmap, p_unsigned);
1560 PRINT_ATTRf(comm, p_unsigned);
1561 PRINT_ATTRf(freq, p_unsigned);
1562 PRINT_ATTRf(inherit_stat, p_unsigned);
1563 PRINT_ATTRf(enable_on_exec, p_unsigned);
1564 PRINT_ATTRf(task, p_unsigned);
1565 PRINT_ATTRf(watermark, p_unsigned);
1566 PRINT_ATTRf(precise_ip, p_unsigned);
1567 PRINT_ATTRf(mmap_data, p_unsigned);
1568 PRINT_ATTRf(sample_id_all, p_unsigned);
1569 PRINT_ATTRf(exclude_host, p_unsigned);
1570 PRINT_ATTRf(exclude_guest, p_unsigned);
1571 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1572 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1573 PRINT_ATTRf(mmap2, p_unsigned);
1574 PRINT_ATTRf(comm_exec, p_unsigned);
1575 PRINT_ATTRf(use_clockid, p_unsigned);
1576 PRINT_ATTRf(context_switch, p_unsigned);
1577 PRINT_ATTRf(write_backward, p_unsigned);
1578
1579 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1580 PRINT_ATTRf(bp_type, p_unsigned);
1581 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1582 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1583 PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1584 PRINT_ATTRf(sample_regs_user, p_hex);
1585 PRINT_ATTRf(sample_stack_user, p_unsigned);
1586 PRINT_ATTRf(clockid, p_signed);
1587 PRINT_ATTRf(sample_regs_intr, p_hex);
1588 PRINT_ATTRf(aux_watermark, p_unsigned);
1589 PRINT_ATTRf(sample_max_stack, p_unsigned);
1590
1591 return ret;
1592 }
1593
1594 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1595 void *priv __maybe_unused)
1596 {
1597 return fprintf(fp, " %-32s %s\n", name, val);
1598 }
1599
1600 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1601 int nr_cpus, int nr_threads,
1602 int thread_idx)
1603 {
1604 for (int cpu = 0; cpu < nr_cpus; cpu++)
1605 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1606 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1607 }
1608
1609 static int update_fds(struct perf_evsel *evsel,
1610 int nr_cpus, int cpu_idx,
1611 int nr_threads, int thread_idx)
1612 {
1613 struct perf_evsel *pos;
1614
1615 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1616 return -EINVAL;
1617
1618 evlist__for_each_entry(evsel->evlist, pos) {
1619 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1620
1621 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1622
1623 /*
1624 * Since fds for next evsel has not been created,
1625 * there is no need to iterate whole event list.
1626 */
1627 if (pos == evsel)
1628 break;
1629 }
1630 return 0;
1631 }
1632
1633 static bool ignore_missing_thread(struct perf_evsel *evsel,
1634 int nr_cpus, int cpu,
1635 struct thread_map *threads,
1636 int thread, int err)
1637 {
1638 pid_t ignore_pid = thread_map__pid(threads, thread);
1639
1640 if (!evsel->ignore_missing_thread)
1641 return false;
1642
1643 /* The system wide setup does not work with threads. */
1644 if (evsel->system_wide)
1645 return false;
1646
1647 /* The -ESRCH is perf event syscall errno for pid's not found. */
1648 if (err != -ESRCH)
1649 return false;
1650
1651 /* If there's only one thread, let it fail. */
1652 if (threads->nr == 1)
1653 return false;
1654
1655 /*
1656 * We should remove fd for missing_thread first
1657 * because thread_map__remove() will decrease threads->nr.
1658 */
1659 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1660 return false;
1661
1662 if (thread_map__remove(threads, thread))
1663 return false;
1664
1665 pr_warning("WARNING: Ignored open failure for pid %d\n",
1666 ignore_pid);
1667 return true;
1668 }
1669
1670 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1671 struct thread_map *threads)
1672 {
1673 int cpu, thread, nthreads;
1674 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1675 int pid = -1, err;
1676 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1677
1678 if (perf_missing_features.write_backward && evsel->attr.write_backward)
1679 return -EINVAL;
1680
1681 if (cpus == NULL) {
1682 static struct cpu_map *empty_cpu_map;
1683
1684 if (empty_cpu_map == NULL) {
1685 empty_cpu_map = cpu_map__dummy_new();
1686 if (empty_cpu_map == NULL)
1687 return -ENOMEM;
1688 }
1689
1690 cpus = empty_cpu_map;
1691 }
1692
1693 if (threads == NULL) {
1694 static struct thread_map *empty_thread_map;
1695
1696 if (empty_thread_map == NULL) {
1697 empty_thread_map = thread_map__new_by_tid(-1);
1698 if (empty_thread_map == NULL)
1699 return -ENOMEM;
1700 }
1701
1702 threads = empty_thread_map;
1703 }
1704
1705 if (evsel->system_wide)
1706 nthreads = 1;
1707 else
1708 nthreads = threads->nr;
1709
1710 if (evsel->fd == NULL &&
1711 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1712 return -ENOMEM;
1713
1714 if (evsel->cgrp) {
1715 flags |= PERF_FLAG_PID_CGROUP;
1716 pid = evsel->cgrp->fd;
1717 }
1718
1719 fallback_missing_features:
1720 if (perf_missing_features.clockid_wrong)
1721 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1722 if (perf_missing_features.clockid) {
1723 evsel->attr.use_clockid = 0;
1724 evsel->attr.clockid = 0;
1725 }
1726 if (perf_missing_features.cloexec)
1727 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1728 if (perf_missing_features.mmap2)
1729 evsel->attr.mmap2 = 0;
1730 if (perf_missing_features.exclude_guest)
1731 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1732 if (perf_missing_features.lbr_flags)
1733 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1734 PERF_SAMPLE_BRANCH_NO_CYCLES);
1735 if (perf_missing_features.group_read && evsel->attr.inherit)
1736 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1737 retry_sample_id:
1738 if (perf_missing_features.sample_id_all)
1739 evsel->attr.sample_id_all = 0;
1740
1741 if (verbose >= 2) {
1742 fprintf(stderr, "%.60s\n", graph_dotted_line);
1743 fprintf(stderr, "perf_event_attr:\n");
1744 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1745 fprintf(stderr, "%.60s\n", graph_dotted_line);
1746 }
1747
1748 for (cpu = 0; cpu < cpus->nr; cpu++) {
1749
1750 for (thread = 0; thread < nthreads; thread++) {
1751 int fd, group_fd;
1752
1753 if (!evsel->cgrp && !evsel->system_wide)
1754 pid = thread_map__pid(threads, thread);
1755
1756 group_fd = get_group_fd(evsel, cpu, thread);
1757 retry_open:
1758 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1759 pid, cpus->map[cpu], group_fd, flags);
1760
1761 test_attr__ready();
1762
1763 fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1764 group_fd, flags);
1765
1766 FD(evsel, cpu, thread) = fd;
1767
1768 if (fd < 0) {
1769 err = -errno;
1770
1771 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1772 /*
1773 * We just removed 1 thread, so take a step
1774 * back on thread index and lower the upper
1775 * nthreads limit.
1776 */
1777 nthreads--;
1778 thread--;
1779
1780 /* ... and pretend like nothing have happened. */
1781 err = 0;
1782 continue;
1783 }
1784
1785 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1786 err);
1787 goto try_fallback;
1788 }
1789
1790 pr_debug2(" = %d\n", fd);
1791
1792 if (evsel->bpf_fd >= 0) {
1793 int evt_fd = fd;
1794 int bpf_fd = evsel->bpf_fd;
1795
1796 err = ioctl(evt_fd,
1797 PERF_EVENT_IOC_SET_BPF,
1798 bpf_fd);
1799 if (err && errno != EEXIST) {
1800 pr_err("failed to attach bpf fd %d: %s\n",
1801 bpf_fd, strerror(errno));
1802 err = -EINVAL;
1803 goto out_close;
1804 }
1805 }
1806
1807 set_rlimit = NO_CHANGE;
1808
1809 /*
1810 * If we succeeded but had to kill clockid, fail and
1811 * have perf_evsel__open_strerror() print us a nice
1812 * error.
1813 */
1814 if (perf_missing_features.clockid ||
1815 perf_missing_features.clockid_wrong) {
1816 err = -EINVAL;
1817 goto out_close;
1818 }
1819 }
1820 }
1821
1822 return 0;
1823
1824 try_fallback:
1825 /*
1826 * perf stat needs between 5 and 22 fds per CPU. When we run out
1827 * of them try to increase the limits.
1828 */
1829 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1830 struct rlimit l;
1831 int old_errno = errno;
1832
1833 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1834 if (set_rlimit == NO_CHANGE)
1835 l.rlim_cur = l.rlim_max;
1836 else {
1837 l.rlim_cur = l.rlim_max + 1000;
1838 l.rlim_max = l.rlim_cur;
1839 }
1840 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1841 set_rlimit++;
1842 errno = old_errno;
1843 goto retry_open;
1844 }
1845 }
1846 errno = old_errno;
1847 }
1848
1849 if (err != -EINVAL || cpu > 0 || thread > 0)
1850 goto out_close;
1851
1852 /*
1853 * Must probe features in the order they were added to the
1854 * perf_event_attr interface.
1855 */
1856 if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1857 perf_missing_features.write_backward = true;
1858 pr_debug2("switching off write_backward\n");
1859 goto out_close;
1860 } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1861 perf_missing_features.clockid_wrong = true;
1862 pr_debug2("switching off clockid\n");
1863 goto fallback_missing_features;
1864 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1865 perf_missing_features.clockid = true;
1866 pr_debug2("switching off use_clockid\n");
1867 goto fallback_missing_features;
1868 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1869 perf_missing_features.cloexec = true;
1870 pr_debug2("switching off cloexec flag\n");
1871 goto fallback_missing_features;
1872 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1873 perf_missing_features.mmap2 = true;
1874 pr_debug2("switching off mmap2\n");
1875 goto fallback_missing_features;
1876 } else if (!perf_missing_features.exclude_guest &&
1877 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1878 perf_missing_features.exclude_guest = true;
1879 pr_debug2("switching off exclude_guest, exclude_host\n");
1880 goto fallback_missing_features;
1881 } else if (!perf_missing_features.sample_id_all) {
1882 perf_missing_features.sample_id_all = true;
1883 pr_debug2("switching off sample_id_all\n");
1884 goto retry_sample_id;
1885 } else if (!perf_missing_features.lbr_flags &&
1886 (evsel->attr.branch_sample_type &
1887 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1888 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1889 perf_missing_features.lbr_flags = true;
1890 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1891 goto fallback_missing_features;
1892 } else if (!perf_missing_features.group_read &&
1893 evsel->attr.inherit &&
1894 (evsel->attr.read_format & PERF_FORMAT_GROUP)) {
1895 perf_missing_features.group_read = true;
1896 pr_debug2("switching off group read\n");
1897 goto fallback_missing_features;
1898 }
1899 out_close:
1900 do {
1901 while (--thread >= 0) {
1902 close(FD(evsel, cpu, thread));
1903 FD(evsel, cpu, thread) = -1;
1904 }
1905 thread = nthreads;
1906 } while (--cpu >= 0);
1907 return err;
1908 }
1909
1910 void perf_evsel__close(struct perf_evsel *evsel)
1911 {
1912 if (evsel->fd == NULL)
1913 return;
1914
1915 perf_evsel__close_fd(evsel);
1916 perf_evsel__free_fd(evsel);
1917 }
1918
1919 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1920 struct cpu_map *cpus)
1921 {
1922 return perf_evsel__open(evsel, cpus, NULL);
1923 }
1924
1925 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1926 struct thread_map *threads)
1927 {
1928 return perf_evsel__open(evsel, NULL, threads);
1929 }
1930
1931 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1932 const union perf_event *event,
1933 struct perf_sample *sample)
1934 {
1935 u64 type = evsel->attr.sample_type;
1936 const u64 *array = event->sample.array;
1937 bool swapped = evsel->needs_swap;
1938 union u64_swap u;
1939
1940 array += ((event->header.size -
1941 sizeof(event->header)) / sizeof(u64)) - 1;
1942
1943 if (type & PERF_SAMPLE_IDENTIFIER) {
1944 sample->id = *array;
1945 array--;
1946 }
1947
1948 if (type & PERF_SAMPLE_CPU) {
1949 u.val64 = *array;
1950 if (swapped) {
1951 /* undo swap of u64, then swap on individual u32s */
1952 u.val64 = bswap_64(u.val64);
1953 u.val32[0] = bswap_32(u.val32[0]);
1954 }
1955
1956 sample->cpu = u.val32[0];
1957 array--;
1958 }
1959
1960 if (type & PERF_SAMPLE_STREAM_ID) {
1961 sample->stream_id = *array;
1962 array--;
1963 }
1964
1965 if (type & PERF_SAMPLE_ID) {
1966 sample->id = *array;
1967 array--;
1968 }
1969
1970 if (type & PERF_SAMPLE_TIME) {
1971 sample->time = *array;
1972 array--;
1973 }
1974
1975 if (type & PERF_SAMPLE_TID) {
1976 u.val64 = *array;
1977 if (swapped) {
1978 /* undo swap of u64, then swap on individual u32s */
1979 u.val64 = bswap_64(u.val64);
1980 u.val32[0] = bswap_32(u.val32[0]);
1981 u.val32[1] = bswap_32(u.val32[1]);
1982 }
1983
1984 sample->pid = u.val32[0];
1985 sample->tid = u.val32[1];
1986 array--;
1987 }
1988
1989 return 0;
1990 }
1991
1992 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1993 u64 size)
1994 {
1995 return size > max_size || offset + size > endp;
1996 }
1997
1998 #define OVERFLOW_CHECK(offset, size, max_size) \
1999 do { \
2000 if (overflow(endp, (max_size), (offset), (size))) \
2001 return -EFAULT; \
2002 } while (0)
2003
2004 #define OVERFLOW_CHECK_u64(offset) \
2005 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2006
2007 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2008 struct perf_sample *data)
2009 {
2010 u64 type = evsel->attr.sample_type;
2011 bool swapped = evsel->needs_swap;
2012 const u64 *array;
2013 u16 max_size = event->header.size;
2014 const void *endp = (void *)event + max_size;
2015 u64 sz;
2016
2017 /*
2018 * used for cross-endian analysis. See git commit 65014ab3
2019 * for why this goofiness is needed.
2020 */
2021 union u64_swap u;
2022
2023 memset(data, 0, sizeof(*data));
2024 data->cpu = data->pid = data->tid = -1;
2025 data->stream_id = data->id = data->time = -1ULL;
2026 data->period = evsel->attr.sample_period;
2027 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2028
2029 if (event->header.type != PERF_RECORD_SAMPLE) {
2030 if (!evsel->attr.sample_id_all)
2031 return 0;
2032 return perf_evsel__parse_id_sample(evsel, event, data);
2033 }
2034
2035 array = event->sample.array;
2036
2037 /*
2038 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2039 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2040 * check the format does not go past the end of the event.
2041 */
2042 if (evsel->sample_size + sizeof(event->header) > event->header.size)
2043 return -EFAULT;
2044
2045 data->id = -1ULL;
2046 if (type & PERF_SAMPLE_IDENTIFIER) {
2047 data->id = *array;
2048 array++;
2049 }
2050
2051 if (type & PERF_SAMPLE_IP) {
2052 data->ip = *array;
2053 array++;
2054 }
2055
2056 if (type & PERF_SAMPLE_TID) {
2057 u.val64 = *array;
2058 if (swapped) {
2059 /* undo swap of u64, then swap on individual u32s */
2060 u.val64 = bswap_64(u.val64);
2061 u.val32[0] = bswap_32(u.val32[0]);
2062 u.val32[1] = bswap_32(u.val32[1]);
2063 }
2064
2065 data->pid = u.val32[0];
2066 data->tid = u.val32[1];
2067 array++;
2068 }
2069
2070 if (type & PERF_SAMPLE_TIME) {
2071 data->time = *array;
2072 array++;
2073 }
2074
2075 data->addr = 0;
2076 if (type & PERF_SAMPLE_ADDR) {
2077 data->addr = *array;
2078 array++;
2079 }
2080
2081 if (type & PERF_SAMPLE_ID) {
2082 data->id = *array;
2083 array++;
2084 }
2085
2086 if (type & PERF_SAMPLE_STREAM_ID) {
2087 data->stream_id = *array;
2088 array++;
2089 }
2090
2091 if (type & PERF_SAMPLE_CPU) {
2092
2093 u.val64 = *array;
2094 if (swapped) {
2095 /* undo swap of u64, then swap on individual u32s */
2096 u.val64 = bswap_64(u.val64);
2097 u.val32[0] = bswap_32(u.val32[0]);
2098 }
2099
2100 data->cpu = u.val32[0];
2101 array++;
2102 }
2103
2104 if (type & PERF_SAMPLE_PERIOD) {
2105 data->period = *array;
2106 array++;
2107 }
2108
2109 if (type & PERF_SAMPLE_READ) {
2110 u64 read_format = evsel->attr.read_format;
2111
2112 OVERFLOW_CHECK_u64(array);
2113 if (read_format & PERF_FORMAT_GROUP)
2114 data->read.group.nr = *array;
2115 else
2116 data->read.one.value = *array;
2117
2118 array++;
2119
2120 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2121 OVERFLOW_CHECK_u64(array);
2122 data->read.time_enabled = *array;
2123 array++;
2124 }
2125
2126 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2127 OVERFLOW_CHECK_u64(array);
2128 data->read.time_running = *array;
2129 array++;
2130 }
2131
2132 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2133 if (read_format & PERF_FORMAT_GROUP) {
2134 const u64 max_group_nr = UINT64_MAX /
2135 sizeof(struct sample_read_value);
2136
2137 if (data->read.group.nr > max_group_nr)
2138 return -EFAULT;
2139 sz = data->read.group.nr *
2140 sizeof(struct sample_read_value);
2141 OVERFLOW_CHECK(array, sz, max_size);
2142 data->read.group.values =
2143 (struct sample_read_value *)array;
2144 array = (void *)array + sz;
2145 } else {
2146 OVERFLOW_CHECK_u64(array);
2147 data->read.one.id = *array;
2148 array++;
2149 }
2150 }
2151
2152 if (type & PERF_SAMPLE_CALLCHAIN) {
2153 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2154
2155 OVERFLOW_CHECK_u64(array);
2156 data->callchain = (struct ip_callchain *)array++;
2157 if (data->callchain->nr > max_callchain_nr)
2158 return -EFAULT;
2159 sz = data->callchain->nr * sizeof(u64);
2160 OVERFLOW_CHECK(array, sz, max_size);
2161 array = (void *)array + sz;
2162 }
2163
2164 if (type & PERF_SAMPLE_RAW) {
2165 OVERFLOW_CHECK_u64(array);
2166 u.val64 = *array;
2167
2168 /*
2169 * Undo swap of u64, then swap on individual u32s,
2170 * get the size of the raw area and undo all of the
2171 * swap. The pevent interface handles endianity by
2172 * itself.
2173 */
2174 if (swapped) {
2175 u.val64 = bswap_64(u.val64);
2176 u.val32[0] = bswap_32(u.val32[0]);
2177 u.val32[1] = bswap_32(u.val32[1]);
2178 }
2179 data->raw_size = u.val32[0];
2180
2181 /*
2182 * The raw data is aligned on 64bits including the
2183 * u32 size, so it's safe to use mem_bswap_64.
2184 */
2185 if (swapped)
2186 mem_bswap_64((void *) array, data->raw_size);
2187
2188 array = (void *)array + sizeof(u32);
2189
2190 OVERFLOW_CHECK(array, data->raw_size, max_size);
2191 data->raw_data = (void *)array;
2192 array = (void *)array + data->raw_size;
2193 }
2194
2195 if (type & PERF_SAMPLE_BRANCH_STACK) {
2196 const u64 max_branch_nr = UINT64_MAX /
2197 sizeof(struct branch_entry);
2198
2199 OVERFLOW_CHECK_u64(array);
2200 data->branch_stack = (struct branch_stack *)array++;
2201
2202 if (data->branch_stack->nr > max_branch_nr)
2203 return -EFAULT;
2204 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2205 OVERFLOW_CHECK(array, sz, max_size);
2206 array = (void *)array + sz;
2207 }
2208
2209 if (type & PERF_SAMPLE_REGS_USER) {
2210 OVERFLOW_CHECK_u64(array);
2211 data->user_regs.abi = *array;
2212 array++;
2213
2214 if (data->user_regs.abi) {
2215 u64 mask = evsel->attr.sample_regs_user;
2216
2217 sz = hweight_long(mask) * sizeof(u64);
2218 OVERFLOW_CHECK(array, sz, max_size);
2219 data->user_regs.mask = mask;
2220 data->user_regs.regs = (u64 *)array;
2221 array = (void *)array + sz;
2222 }
2223 }
2224
2225 if (type & PERF_SAMPLE_STACK_USER) {
2226 OVERFLOW_CHECK_u64(array);
2227 sz = *array++;
2228
2229 data->user_stack.offset = ((char *)(array - 1)
2230 - (char *) event);
2231
2232 if (!sz) {
2233 data->user_stack.size = 0;
2234 } else {
2235 OVERFLOW_CHECK(array, sz, max_size);
2236 data->user_stack.data = (char *)array;
2237 array = (void *)array + sz;
2238 OVERFLOW_CHECK_u64(array);
2239 data->user_stack.size = *array++;
2240 if (WARN_ONCE(data->user_stack.size > sz,
2241 "user stack dump failure\n"))
2242 return -EFAULT;
2243 }
2244 }
2245
2246 if (type & PERF_SAMPLE_WEIGHT) {
2247 OVERFLOW_CHECK_u64(array);
2248 data->weight = *array;
2249 array++;
2250 }
2251
2252 data->data_src = PERF_MEM_DATA_SRC_NONE;
2253 if (type & PERF_SAMPLE_DATA_SRC) {
2254 OVERFLOW_CHECK_u64(array);
2255 data->data_src = *array;
2256 array++;
2257 }
2258
2259 data->transaction = 0;
2260 if (type & PERF_SAMPLE_TRANSACTION) {
2261 OVERFLOW_CHECK_u64(array);
2262 data->transaction = *array;
2263 array++;
2264 }
2265
2266 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2267 if (type & PERF_SAMPLE_REGS_INTR) {
2268 OVERFLOW_CHECK_u64(array);
2269 data->intr_regs.abi = *array;
2270 array++;
2271
2272 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2273 u64 mask = evsel->attr.sample_regs_intr;
2274
2275 sz = hweight_long(mask) * sizeof(u64);
2276 OVERFLOW_CHECK(array, sz, max_size);
2277 data->intr_regs.mask = mask;
2278 data->intr_regs.regs = (u64 *)array;
2279 array = (void *)array + sz;
2280 }
2281 }
2282
2283 data->phys_addr = 0;
2284 if (type & PERF_SAMPLE_PHYS_ADDR) {
2285 data->phys_addr = *array;
2286 array++;
2287 }
2288
2289 return 0;
2290 }
2291
2292 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2293 u64 read_format)
2294 {
2295 size_t sz, result = sizeof(struct sample_event);
2296
2297 if (type & PERF_SAMPLE_IDENTIFIER)
2298 result += sizeof(u64);
2299
2300 if (type & PERF_SAMPLE_IP)
2301 result += sizeof(u64);
2302
2303 if (type & PERF_SAMPLE_TID)
2304 result += sizeof(u64);
2305
2306 if (type & PERF_SAMPLE_TIME)
2307 result += sizeof(u64);
2308
2309 if (type & PERF_SAMPLE_ADDR)
2310 result += sizeof(u64);
2311
2312 if (type & PERF_SAMPLE_ID)
2313 result += sizeof(u64);
2314
2315 if (type & PERF_SAMPLE_STREAM_ID)
2316 result += sizeof(u64);
2317
2318 if (type & PERF_SAMPLE_CPU)
2319 result += sizeof(u64);
2320
2321 if (type & PERF_SAMPLE_PERIOD)
2322 result += sizeof(u64);
2323
2324 if (type & PERF_SAMPLE_READ) {
2325 result += sizeof(u64);
2326 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2327 result += sizeof(u64);
2328 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2329 result += sizeof(u64);
2330 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2331 if (read_format & PERF_FORMAT_GROUP) {
2332 sz = sample->read.group.nr *
2333 sizeof(struct sample_read_value);
2334 result += sz;
2335 } else {
2336 result += sizeof(u64);
2337 }
2338 }
2339
2340 if (type & PERF_SAMPLE_CALLCHAIN) {
2341 sz = (sample->callchain->nr + 1) * sizeof(u64);
2342 result += sz;
2343 }
2344
2345 if (type & PERF_SAMPLE_RAW) {
2346 result += sizeof(u32);
2347 result += sample->raw_size;
2348 }
2349
2350 if (type & PERF_SAMPLE_BRANCH_STACK) {
2351 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2352 sz += sizeof(u64);
2353 result += sz;
2354 }
2355
2356 if (type & PERF_SAMPLE_REGS_USER) {
2357 if (sample->user_regs.abi) {
2358 result += sizeof(u64);
2359 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2360 result += sz;
2361 } else {
2362 result += sizeof(u64);
2363 }
2364 }
2365
2366 if (type & PERF_SAMPLE_STACK_USER) {
2367 sz = sample->user_stack.size;
2368 result += sizeof(u64);
2369 if (sz) {
2370 result += sz;
2371 result += sizeof(u64);
2372 }
2373 }
2374
2375 if (type & PERF_SAMPLE_WEIGHT)
2376 result += sizeof(u64);
2377
2378 if (type & PERF_SAMPLE_DATA_SRC)
2379 result += sizeof(u64);
2380
2381 if (type & PERF_SAMPLE_TRANSACTION)
2382 result += sizeof(u64);
2383
2384 if (type & PERF_SAMPLE_REGS_INTR) {
2385 if (sample->intr_regs.abi) {
2386 result += sizeof(u64);
2387 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2388 result += sz;
2389 } else {
2390 result += sizeof(u64);
2391 }
2392 }
2393
2394 if (type & PERF_SAMPLE_PHYS_ADDR)
2395 result += sizeof(u64);
2396
2397 return result;
2398 }
2399
2400 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2401 u64 read_format,
2402 const struct perf_sample *sample,
2403 bool swapped)
2404 {
2405 u64 *array;
2406 size_t sz;
2407 /*
2408 * used for cross-endian analysis. See git commit 65014ab3
2409 * for why this goofiness is needed.
2410 */
2411 union u64_swap u;
2412
2413 array = event->sample.array;
2414
2415 if (type & PERF_SAMPLE_IDENTIFIER) {
2416 *array = sample->id;
2417 array++;
2418 }
2419
2420 if (type & PERF_SAMPLE_IP) {
2421 *array = sample->ip;
2422 array++;
2423 }
2424
2425 if (type & PERF_SAMPLE_TID) {
2426 u.val32[0] = sample->pid;
2427 u.val32[1] = sample->tid;
2428 if (swapped) {
2429 /*
2430 * Inverse of what is done in perf_evsel__parse_sample
2431 */
2432 u.val32[0] = bswap_32(u.val32[0]);
2433 u.val32[1] = bswap_32(u.val32[1]);
2434 u.val64 = bswap_64(u.val64);
2435 }
2436
2437 *array = u.val64;
2438 array++;
2439 }
2440
2441 if (type & PERF_SAMPLE_TIME) {
2442 *array = sample->time;
2443 array++;
2444 }
2445
2446 if (type & PERF_SAMPLE_ADDR) {
2447 *array = sample->addr;
2448 array++;
2449 }
2450
2451 if (type & PERF_SAMPLE_ID) {
2452 *array = sample->id;
2453 array++;
2454 }
2455
2456 if (type & PERF_SAMPLE_STREAM_ID) {
2457 *array = sample->stream_id;
2458 array++;
2459 }
2460
2461 if (type & PERF_SAMPLE_CPU) {
2462 u.val32[0] = sample->cpu;
2463 if (swapped) {
2464 /*
2465 * Inverse of what is done in perf_evsel__parse_sample
2466 */
2467 u.val32[0] = bswap_32(u.val32[0]);
2468 u.val64 = bswap_64(u.val64);
2469 }
2470 *array = u.val64;
2471 array++;
2472 }
2473
2474 if (type & PERF_SAMPLE_PERIOD) {
2475 *array = sample->period;
2476 array++;
2477 }
2478
2479 if (type & PERF_SAMPLE_READ) {
2480 if (read_format & PERF_FORMAT_GROUP)
2481 *array = sample->read.group.nr;
2482 else
2483 *array = sample->read.one.value;
2484 array++;
2485
2486 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2487 *array = sample->read.time_enabled;
2488 array++;
2489 }
2490
2491 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2492 *array = sample->read.time_running;
2493 array++;
2494 }
2495
2496 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2497 if (read_format & PERF_FORMAT_GROUP) {
2498 sz = sample->read.group.nr *
2499 sizeof(struct sample_read_value);
2500 memcpy(array, sample->read.group.values, sz);
2501 array = (void *)array + sz;
2502 } else {
2503 *array = sample->read.one.id;
2504 array++;
2505 }
2506 }
2507
2508 if (type & PERF_SAMPLE_CALLCHAIN) {
2509 sz = (sample->callchain->nr + 1) * sizeof(u64);
2510 memcpy(array, sample->callchain, sz);
2511 array = (void *)array + sz;
2512 }
2513
2514 if (type & PERF_SAMPLE_RAW) {
2515 u.val32[0] = sample->raw_size;
2516 if (WARN_ONCE(swapped,
2517 "Endianness of raw data not corrected!\n")) {
2518 /*
2519 * Inverse of what is done in perf_evsel__parse_sample
2520 */
2521 u.val32[0] = bswap_32(u.val32[0]);
2522 u.val32[1] = bswap_32(u.val32[1]);
2523 u.val64 = bswap_64(u.val64);
2524 }
2525 *array = u.val64;
2526 array = (void *)array + sizeof(u32);
2527
2528 memcpy(array, sample->raw_data, sample->raw_size);
2529 array = (void *)array + sample->raw_size;
2530 }
2531
2532 if (type & PERF_SAMPLE_BRANCH_STACK) {
2533 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2534 sz += sizeof(u64);
2535 memcpy(array, sample->branch_stack, sz);
2536 array = (void *)array + sz;
2537 }
2538
2539 if (type & PERF_SAMPLE_REGS_USER) {
2540 if (sample->user_regs.abi) {
2541 *array++ = sample->user_regs.abi;
2542 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2543 memcpy(array, sample->user_regs.regs, sz);
2544 array = (void *)array + sz;
2545 } else {
2546 *array++ = 0;
2547 }
2548 }
2549
2550 if (type & PERF_SAMPLE_STACK_USER) {
2551 sz = sample->user_stack.size;
2552 *array++ = sz;
2553 if (sz) {
2554 memcpy(array, sample->user_stack.data, sz);
2555 array = (void *)array + sz;
2556 *array++ = sz;
2557 }
2558 }
2559
2560 if (type & PERF_SAMPLE_WEIGHT) {
2561 *array = sample->weight;
2562 array++;
2563 }
2564
2565 if (type & PERF_SAMPLE_DATA_SRC) {
2566 *array = sample->data_src;
2567 array++;
2568 }
2569
2570 if (type & PERF_SAMPLE_TRANSACTION) {
2571 *array = sample->transaction;
2572 array++;
2573 }
2574
2575 if (type & PERF_SAMPLE_REGS_INTR) {
2576 if (sample->intr_regs.abi) {
2577 *array++ = sample->intr_regs.abi;
2578 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2579 memcpy(array, sample->intr_regs.regs, sz);
2580 array = (void *)array + sz;
2581 } else {
2582 *array++ = 0;
2583 }
2584 }
2585
2586 if (type & PERF_SAMPLE_PHYS_ADDR) {
2587 *array = sample->phys_addr;
2588 array++;
2589 }
2590
2591 return 0;
2592 }
2593
2594 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2595 {
2596 return pevent_find_field(evsel->tp_format, name);
2597 }
2598
2599 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2600 const char *name)
2601 {
2602 struct format_field *field = perf_evsel__field(evsel, name);
2603 int offset;
2604
2605 if (!field)
2606 return NULL;
2607
2608 offset = field->offset;
2609
2610 if (field->flags & FIELD_IS_DYNAMIC) {
2611 offset = *(int *)(sample->raw_data + field->offset);
2612 offset &= 0xffff;
2613 }
2614
2615 return sample->raw_data + offset;
2616 }
2617
2618 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2619 bool needs_swap)
2620 {
2621 u64 value;
2622 void *ptr = sample->raw_data + field->offset;
2623
2624 switch (field->size) {
2625 case 1:
2626 return *(u8 *)ptr;
2627 case 2:
2628 value = *(u16 *)ptr;
2629 break;
2630 case 4:
2631 value = *(u32 *)ptr;
2632 break;
2633 case 8:
2634 memcpy(&value, ptr, sizeof(u64));
2635 break;
2636 default:
2637 return 0;
2638 }
2639
2640 if (!needs_swap)
2641 return value;
2642
2643 switch (field->size) {
2644 case 2:
2645 return bswap_16(value);
2646 case 4:
2647 return bswap_32(value);
2648 case 8:
2649 return bswap_64(value);
2650 default:
2651 return 0;
2652 }
2653
2654 return 0;
2655 }
2656
2657 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2658 const char *name)
2659 {
2660 struct format_field *field = perf_evsel__field(evsel, name);
2661
2662 if (!field)
2663 return 0;
2664
2665 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2666 }
2667
2668 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2669 char *msg, size_t msgsize)
2670 {
2671 int paranoid;
2672
2673 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2674 evsel->attr.type == PERF_TYPE_HARDWARE &&
2675 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2676 /*
2677 * If it's cycles then fall back to hrtimer based
2678 * cpu-clock-tick sw counter, which is always available even if
2679 * no PMU support.
2680 *
2681 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2682 * b0a873e).
2683 */
2684 scnprintf(msg, msgsize, "%s",
2685 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2686
2687 evsel->attr.type = PERF_TYPE_SOFTWARE;
2688 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2689
2690 zfree(&evsel->name);
2691 return true;
2692 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2693 (paranoid = perf_event_paranoid()) > 1) {
2694 const char *name = perf_evsel__name(evsel);
2695 char *new_name;
2696
2697 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2698 return false;
2699
2700 if (evsel->name)
2701 free(evsel->name);
2702 evsel->name = new_name;
2703 scnprintf(msg, msgsize,
2704 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2705 evsel->attr.exclude_kernel = 1;
2706
2707 return true;
2708 }
2709
2710 return false;
2711 }
2712
2713 static bool find_process(const char *name)
2714 {
2715 size_t len = strlen(name);
2716 DIR *dir;
2717 struct dirent *d;
2718 int ret = -1;
2719
2720 dir = opendir(procfs__mountpoint());
2721 if (!dir)
2722 return false;
2723
2724 /* Walk through the directory. */
2725 while (ret && (d = readdir(dir)) != NULL) {
2726 char path[PATH_MAX];
2727 char *data;
2728 size_t size;
2729
2730 if ((d->d_type != DT_DIR) ||
2731 !strcmp(".", d->d_name) ||
2732 !strcmp("..", d->d_name))
2733 continue;
2734
2735 scnprintf(path, sizeof(path), "%s/%s/comm",
2736 procfs__mountpoint(), d->d_name);
2737
2738 if (filename__read_str(path, &data, &size))
2739 continue;
2740
2741 ret = strncmp(name, data, len);
2742 free(data);
2743 }
2744
2745 closedir(dir);
2746 return ret ? false : true;
2747 }
2748
2749 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2750 int err, char *msg, size_t size)
2751 {
2752 char sbuf[STRERR_BUFSIZE];
2753 int printed = 0;
2754
2755 switch (err) {
2756 case EPERM:
2757 case EACCES:
2758 if (err == EPERM)
2759 printed = scnprintf(msg, size,
2760 "No permission to enable %s event.\n\n",
2761 perf_evsel__name(evsel));
2762
2763 return scnprintf(msg + printed, size - printed,
2764 "You may not have permission to collect %sstats.\n\n"
2765 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2766 "which controls use of the performance events system by\n"
2767 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2768 "The current value is %d:\n\n"
2769 " -1: Allow use of (almost) all events by all users\n"
2770 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2771 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2772 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2773 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2774 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2775 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2776 " kernel.perf_event_paranoid = -1\n" ,
2777 target->system_wide ? "system-wide " : "",
2778 perf_event_paranoid());
2779 case ENOENT:
2780 return scnprintf(msg, size, "The %s event is not supported.",
2781 perf_evsel__name(evsel));
2782 case EMFILE:
2783 return scnprintf(msg, size, "%s",
2784 "Too many events are opened.\n"
2785 "Probably the maximum number of open file descriptors has been reached.\n"
2786 "Hint: Try again after reducing the number of events.\n"
2787 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2788 case ENOMEM:
2789 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2790 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2791 return scnprintf(msg, size,
2792 "Not enough memory to setup event with callchain.\n"
2793 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2794 "Hint: Current value: %d", sysctl_perf_event_max_stack);
2795 break;
2796 case ENODEV:
2797 if (target->cpu_list)
2798 return scnprintf(msg, size, "%s",
2799 "No such device - did you specify an out-of-range profile CPU?");
2800 break;
2801 case EOPNOTSUPP:
2802 if (evsel->attr.sample_period != 0)
2803 return scnprintf(msg, size, "%s",
2804 "PMU Hardware doesn't support sampling/overflow-interrupts.");
2805 if (evsel->attr.precise_ip)
2806 return scnprintf(msg, size, "%s",
2807 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2808 #if defined(__i386__) || defined(__x86_64__)
2809 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2810 return scnprintf(msg, size, "%s",
2811 "No hardware sampling interrupt available.\n"
2812 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2813 #endif
2814 break;
2815 case EBUSY:
2816 if (find_process("oprofiled"))
2817 return scnprintf(msg, size,
2818 "The PMU counters are busy/taken by another profiler.\n"
2819 "We found oprofile daemon running, please stop it and try again.");
2820 break;
2821 case EINVAL:
2822 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2823 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2824 if (perf_missing_features.clockid)
2825 return scnprintf(msg, size, "clockid feature not supported.");
2826 if (perf_missing_features.clockid_wrong)
2827 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2828 break;
2829 default:
2830 break;
2831 }
2832
2833 return scnprintf(msg, size,
2834 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2835 "/bin/dmesg may provide additional information.\n"
2836 "No CONFIG_PERF_EVENTS=y kernel support configured?",
2837 err, str_error_r(err, sbuf, sizeof(sbuf)),
2838 perf_evsel__name(evsel));
2839 }
2840
2841 char *perf_evsel__env_arch(struct perf_evsel *evsel)
2842 {
2843 if (evsel && evsel->evlist && evsel->evlist->env)
2844 return evsel->evlist->env->arch;
2845 return NULL;
2846 }
2847
2848 char *perf_evsel__env_cpuid(struct perf_evsel *evsel)
2849 {
2850 if (evsel && evsel->evlist && evsel->evlist->env)
2851 return evsel->evlist->env->cpuid;
2852 return NULL;
2853 }
ubuntu-bionic-kernel mirror