?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.8-rc8
[linux/fpc-iii.git] / tools / perf / util / evsel.c
blob21fd573106edd90ee4c33b3a0a9b46e43ecd052b
1 /*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9
10 #include <byteswap.h>
11 #include <linux/bitops.h>
12 #include <api/fs/tracing_path.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <linux/err.h>
17 #include <sys/resource.h>
18 #include "asm/bug.h"
19 #include "callchain.h"
20 #include "cgroup.h"
21 #include "evsel.h"
22 #include "evlist.h"
23 #include "util.h"
24 #include "cpumap.h"
25 #include "thread_map.h"
26 #include "target.h"
27 #include "perf_regs.h"
28 #include "debug.h"
29 #include "trace-event.h"
30 #include "stat.h"
31
32 static struct {
33 bool sample_id_all;
34 bool exclude_guest;
35 bool mmap2;
36 bool cloexec;
37 bool clockid;
38 bool clockid_wrong;
39 bool lbr_flags;
40 bool write_backward;
41 } perf_missing_features;
42
43 static clockid_t clockid;
44
45 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
46 {
47 return 0;
48 }
49
50 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
51 {
52 }
53
54 static struct {
55 size_t size;
56 int (*init)(struct perf_evsel *evsel);
57 void (*fini)(struct perf_evsel *evsel);
58 } perf_evsel__object = {
59 .size = sizeof(struct perf_evsel),
60 .init = perf_evsel__no_extra_init,
61 .fini = perf_evsel__no_extra_fini,
62 };
63
64 int perf_evsel__object_config(size_t object_size,
65 int (*init)(struct perf_evsel *evsel),
66 void (*fini)(struct perf_evsel *evsel))
67 {
68
69 if (object_size == 0)
70 goto set_methods;
71
72 if (perf_evsel__object.size > object_size)
73 return -EINVAL;
74
75 perf_evsel__object.size = object_size;
76
77 set_methods:
78 if (init != NULL)
79 perf_evsel__object.init = init;
80
81 if (fini != NULL)
82 perf_evsel__object.fini = fini;
83
84 return 0;
85 }
86
87 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
88
89 int __perf_evsel__sample_size(u64 sample_type)
90 {
91 u64 mask = sample_type & PERF_SAMPLE_MASK;
92 int size = 0;
93 int i;
94
95 for (i = 0; i < 64; i++) {
96 if (mask & (1ULL << i))
97 size++;
98 }
99
100 size *= sizeof(u64);
101
102 return size;
103 }
104
105 /**
106 * __perf_evsel__calc_id_pos - calculate id_pos.
107 * @sample_type: sample type
108 *
109 * This function returns the position of the event id (PERF_SAMPLE_ID or
110 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
111 * sample_event.
112 */
113 static int __perf_evsel__calc_id_pos(u64 sample_type)
114 {
115 int idx = 0;
116
117 if (sample_type & PERF_SAMPLE_IDENTIFIER)
118 return 0;
119
120 if (!(sample_type & PERF_SAMPLE_ID))
121 return -1;
122
123 if (sample_type & PERF_SAMPLE_IP)
124 idx += 1;
125
126 if (sample_type & PERF_SAMPLE_TID)
127 idx += 1;
128
129 if (sample_type & PERF_SAMPLE_TIME)
130 idx += 1;
131
132 if (sample_type & PERF_SAMPLE_ADDR)
133 idx += 1;
134
135 return idx;
136 }
137
138 /**
139 * __perf_evsel__calc_is_pos - calculate is_pos.
140 * @sample_type: sample type
141 *
142 * This function returns the position (counting backwards) of the event id
143 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
144 * sample_id_all is used there is an id sample appended to non-sample events.
145 */
146 static int __perf_evsel__calc_is_pos(u64 sample_type)
147 {
148 int idx = 1;
149
150 if (sample_type & PERF_SAMPLE_IDENTIFIER)
151 return 1;
152
153 if (!(sample_type & PERF_SAMPLE_ID))
154 return -1;
155
156 if (sample_type & PERF_SAMPLE_CPU)
157 idx += 1;
158
159 if (sample_type & PERF_SAMPLE_STREAM_ID)
160 idx += 1;
161
162 return idx;
163 }
164
165 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
166 {
167 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
168 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
169 }
170
171 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
172 enum perf_event_sample_format bit)
173 {
174 if (!(evsel->attr.sample_type & bit)) {
175 evsel->attr.sample_type |= bit;
176 evsel->sample_size += sizeof(u64);
177 perf_evsel__calc_id_pos(evsel);
178 }
179 }
180
181 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
182 enum perf_event_sample_format bit)
183 {
184 if (evsel->attr.sample_type & bit) {
185 evsel->attr.sample_type &= ~bit;
186 evsel->sample_size -= sizeof(u64);
187 perf_evsel__calc_id_pos(evsel);
188 }
189 }
190
191 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
192 bool can_sample_identifier)
193 {
194 if (can_sample_identifier) {
195 perf_evsel__reset_sample_bit(evsel, ID);
196 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
197 } else {
198 perf_evsel__set_sample_bit(evsel, ID);
199 }
200 evsel->attr.read_format |= PERF_FORMAT_ID;
201 }
202
203 /**
204 * perf_evsel__is_function_event - Return whether given evsel is a function
205 * trace event
206 *
207 * @evsel - evsel selector to be tested
208 *
209 * Return %true if event is function trace event
210 */
211 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
212 {
213 #define FUNCTION_EVENT "ftrace:function"
214
215 return evsel->name &&
216 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
217
218 #undef FUNCTION_EVENT
219 }
220
221 void perf_evsel__init(struct perf_evsel *evsel,
222 struct perf_event_attr *attr, int idx)
223 {
224 evsel->idx = idx;
225 evsel->tracking = !idx;
226 evsel->attr = *attr;
227 evsel->leader = evsel;
228 evsel->unit = "";
229 evsel->scale = 1.0;
230 evsel->evlist = NULL;
231 evsel->bpf_fd = -1;
232 INIT_LIST_HEAD(&evsel->node);
233 INIT_LIST_HEAD(&evsel->config_terms);
234 perf_evsel__object.init(evsel);
235 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
236 perf_evsel__calc_id_pos(evsel);
237 evsel->cmdline_group_boundary = false;
238 }
239
240 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
241 {
242 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
243
244 if (evsel != NULL)
245 perf_evsel__init(evsel, attr, idx);
246
247 if (perf_evsel__is_bpf_output(evsel)) {
248 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
249 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
250 evsel->attr.sample_period = 1;
251 }
252
253 return evsel;
254 }
255
256 struct perf_evsel *perf_evsel__new_cycles(void)
257 {
258 struct perf_event_attr attr = {
259 .type = PERF_TYPE_HARDWARE,
260 .config = PERF_COUNT_HW_CPU_CYCLES,
261 };
262 struct perf_evsel *evsel;
263
264 event_attr_init(&attr);
265
266 perf_event_attr__set_max_precise_ip(&attr);
267
268 evsel = perf_evsel__new(&attr);
269 if (evsel == NULL)
270 goto out;
271
272 /* use asprintf() because free(evsel) assumes name is allocated */
273 if (asprintf(&evsel->name, "cycles%.*s",
274 attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
275 goto error_free;
276 out:
277 return evsel;
278 error_free:
279 perf_evsel__delete(evsel);
280 evsel = NULL;
281 goto out;
282 }
283
284 /*
285 * Returns pointer with encoded error via <linux/err.h> interface.
286 */
287 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
288 {
289 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
290 int err = -ENOMEM;
291
292 if (evsel == NULL) {
293 goto out_err;
294 } else {
295 struct perf_event_attr attr = {
296 .type = PERF_TYPE_TRACEPOINT,
297 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
298 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
299 };
300
301 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
302 goto out_free;
303
304 evsel->tp_format = trace_event__tp_format(sys, name);
305 if (IS_ERR(evsel->tp_format)) {
306 err = PTR_ERR(evsel->tp_format);
307 goto out_free;
308 }
309
310 event_attr_init(&attr);
311 attr.config = evsel->tp_format->id;
312 attr.sample_period = 1;
313 perf_evsel__init(evsel, &attr, idx);
314 }
315
316 return evsel;
317
318 out_free:
319 zfree(&evsel->name);
320 free(evsel);
321 out_err:
322 return ERR_PTR(err);
323 }
324
325 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
326 "cycles",
327 "instructions",
328 "cache-references",
329 "cache-misses",
330 "branches",
331 "branch-misses",
332 "bus-cycles",
333 "stalled-cycles-frontend",
334 "stalled-cycles-backend",
335 "ref-cycles",
336 };
337
338 static const char *__perf_evsel__hw_name(u64 config)
339 {
340 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
341 return perf_evsel__hw_names[config];
342
343 return "unknown-hardware";
344 }
345
346 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
347 {
348 int colon = 0, r = 0;
349 struct perf_event_attr *attr = &evsel->attr;
350 bool exclude_guest_default = false;
351
352 #define MOD_PRINT(context, mod) do { \
353 if (!attr->exclude_##context) { \
354 if (!colon) colon = ++r; \
355 r += scnprintf(bf + r, size - r, "%c", mod); \
356 } } while(0)
357
358 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
359 MOD_PRINT(kernel, 'k');
360 MOD_PRINT(user, 'u');
361 MOD_PRINT(hv, 'h');
362 exclude_guest_default = true;
363 }
364
365 if (attr->precise_ip) {
366 if (!colon)
367 colon = ++r;
368 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
369 exclude_guest_default = true;
370 }
371
372 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
373 MOD_PRINT(host, 'H');
374 MOD_PRINT(guest, 'G');
375 }
376 #undef MOD_PRINT
377 if (colon)
378 bf[colon - 1] = ':';
379 return r;
380 }
381
382 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
383 {
384 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
385 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
386 }
387
388 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
389 "cpu-clock",
390 "task-clock",
391 "page-faults",
392 "context-switches",
393 "cpu-migrations",
394 "minor-faults",
395 "major-faults",
396 "alignment-faults",
397 "emulation-faults",
398 "dummy",
399 };
400
401 static const char *__perf_evsel__sw_name(u64 config)
402 {
403 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
404 return perf_evsel__sw_names[config];
405 return "unknown-software";
406 }
407
408 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
409 {
410 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
411 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
412 }
413
414 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
415 {
416 int r;
417
418 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
419
420 if (type & HW_BREAKPOINT_R)
421 r += scnprintf(bf + r, size - r, "r");
422
423 if (type & HW_BREAKPOINT_W)
424 r += scnprintf(bf + r, size - r, "w");
425
426 if (type & HW_BREAKPOINT_X)
427 r += scnprintf(bf + r, size - r, "x");
428
429 return r;
430 }
431
432 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
433 {
434 struct perf_event_attr *attr = &evsel->attr;
435 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
436 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
437 }
438
439 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
440 [PERF_EVSEL__MAX_ALIASES] = {
441 { "L1-dcache", "l1-d", "l1d", "L1-data", },
442 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
443 { "LLC", "L2", },
444 { "dTLB", "d-tlb", "Data-TLB", },
445 { "iTLB", "i-tlb", "Instruction-TLB", },
446 { "branch", "branches", "bpu", "btb", "bpc", },
447 { "node", },
448 };
449
450 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
451 [PERF_EVSEL__MAX_ALIASES] = {
452 { "load", "loads", "read", },
453 { "store", "stores", "write", },
454 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
455 };
456
457 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
458 [PERF_EVSEL__MAX_ALIASES] = {
459 { "refs", "Reference", "ops", "access", },
460 { "misses", "miss", },
461 };
462
463 #define C(x) PERF_COUNT_HW_CACHE_##x
464 #define CACHE_READ (1 << C(OP_READ))
465 #define CACHE_WRITE (1 << C(OP_WRITE))
466 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
467 #define COP(x) (1 << x)
468
469 /*
470 * cache operartion stat
471 * L1I : Read and prefetch only
472 * ITLB and BPU : Read-only
473 */
474 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
475 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
476 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
477 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
478 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
479 [C(ITLB)] = (CACHE_READ),
480 [C(BPU)] = (CACHE_READ),
481 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
482 };
483
484 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
485 {
486 if (perf_evsel__hw_cache_stat[type] & COP(op))
487 return true; /* valid */
488 else
489 return false; /* invalid */
490 }
491
492 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
493 char *bf, size_t size)
494 {
495 if (result) {
496 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
497 perf_evsel__hw_cache_op[op][0],
498 perf_evsel__hw_cache_result[result][0]);
499 }
500
501 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
502 perf_evsel__hw_cache_op[op][1]);
503 }
504
505 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
506 {
507 u8 op, result, type = (config >> 0) & 0xff;
508 const char *err = "unknown-ext-hardware-cache-type";
509
510 if (type >= PERF_COUNT_HW_CACHE_MAX)
511 goto out_err;
512
513 op = (config >> 8) & 0xff;
514 err = "unknown-ext-hardware-cache-op";
515 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
516 goto out_err;
517
518 result = (config >> 16) & 0xff;
519 err = "unknown-ext-hardware-cache-result";
520 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
521 goto out_err;
522
523 err = "invalid-cache";
524 if (!perf_evsel__is_cache_op_valid(type, op))
525 goto out_err;
526
527 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
528 out_err:
529 return scnprintf(bf, size, "%s", err);
530 }
531
532 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
533 {
534 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
535 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
536 }
537
538 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
539 {
540 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
541 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
542 }
543
544 const char *perf_evsel__name(struct perf_evsel *evsel)
545 {
546 char bf[128];
547
548 if (evsel->name)
549 return evsel->name;
550
551 switch (evsel->attr.type) {
552 case PERF_TYPE_RAW:
553 perf_evsel__raw_name(evsel, bf, sizeof(bf));
554 break;
555
556 case PERF_TYPE_HARDWARE:
557 perf_evsel__hw_name(evsel, bf, sizeof(bf));
558 break;
559
560 case PERF_TYPE_HW_CACHE:
561 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
562 break;
563
564 case PERF_TYPE_SOFTWARE:
565 perf_evsel__sw_name(evsel, bf, sizeof(bf));
566 break;
567
568 case PERF_TYPE_TRACEPOINT:
569 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
570 break;
571
572 case PERF_TYPE_BREAKPOINT:
573 perf_evsel__bp_name(evsel, bf, sizeof(bf));
574 break;
575
576 default:
577 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
578 evsel->attr.type);
579 break;
580 }
581
582 evsel->name = strdup(bf);
583
584 return evsel->name ?: "unknown";
585 }
586
587 const char *perf_evsel__group_name(struct perf_evsel *evsel)
588 {
589 return evsel->group_name ?: "anon group";
590 }
591
592 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
593 {
594 int ret;
595 struct perf_evsel *pos;
596 const char *group_name = perf_evsel__group_name(evsel);
597
598 ret = scnprintf(buf, size, "%s", group_name);
599
600 ret += scnprintf(buf + ret, size - ret, " { %s",
601 perf_evsel__name(evsel));
602
603 for_each_group_member(pos, evsel)
604 ret += scnprintf(buf + ret, size - ret, ", %s",
605 perf_evsel__name(pos));
606
607 ret += scnprintf(buf + ret, size - ret, " }");
608
609 return ret;
610 }
611
612 void perf_evsel__config_callchain(struct perf_evsel *evsel,
613 struct record_opts *opts,
614 struct callchain_param *param)
615 {
616 bool function = perf_evsel__is_function_event(evsel);
617 struct perf_event_attr *attr = &evsel->attr;
618
619 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
620
621 attr->sample_max_stack = param->max_stack;
622
623 if (param->record_mode == CALLCHAIN_LBR) {
624 if (!opts->branch_stack) {
625 if (attr->exclude_user) {
626 pr_warning("LBR callstack option is only available "
627 "to get user callchain information. "
628 "Falling back to framepointers.\n");
629 } else {
630 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
631 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
632 PERF_SAMPLE_BRANCH_CALL_STACK |
633 PERF_SAMPLE_BRANCH_NO_CYCLES |
634 PERF_SAMPLE_BRANCH_NO_FLAGS;
635 }
636 } else
637 pr_warning("Cannot use LBR callstack with branch stack. "
638 "Falling back to framepointers.\n");
639 }
640
641 if (param->record_mode == CALLCHAIN_DWARF) {
642 if (!function) {
643 perf_evsel__set_sample_bit(evsel, REGS_USER);
644 perf_evsel__set_sample_bit(evsel, STACK_USER);
645 attr->sample_regs_user = PERF_REGS_MASK;
646 attr->sample_stack_user = param->dump_size;
647 attr->exclude_callchain_user = 1;
648 } else {
649 pr_info("Cannot use DWARF unwind for function trace event,"
650 " falling back to framepointers.\n");
651 }
652 }
653
654 if (function) {
655 pr_info("Disabling user space callchains for function trace event.\n");
656 attr->exclude_callchain_user = 1;
657 }
658 }
659
660 static void
661 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
662 struct callchain_param *param)
663 {
664 struct perf_event_attr *attr = &evsel->attr;
665
666 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
667 if (param->record_mode == CALLCHAIN_LBR) {
668 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
669 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
670 PERF_SAMPLE_BRANCH_CALL_STACK);
671 }
672 if (param->record_mode == CALLCHAIN_DWARF) {
673 perf_evsel__reset_sample_bit(evsel, REGS_USER);
674 perf_evsel__reset_sample_bit(evsel, STACK_USER);
675 }
676 }
677
678 static void apply_config_terms(struct perf_evsel *evsel,
679 struct record_opts *opts)
680 {
681 struct perf_evsel_config_term *term;
682 struct list_head *config_terms = &evsel->config_terms;
683 struct perf_event_attr *attr = &evsel->attr;
684 struct callchain_param param;
685 u32 dump_size = 0;
686 int max_stack = 0;
687 const char *callgraph_buf = NULL;
688
689 /* callgraph default */
690 param.record_mode = callchain_param.record_mode;
691
692 list_for_each_entry(term, config_terms, list) {
693 switch (term->type) {
694 case PERF_EVSEL__CONFIG_TERM_PERIOD:
695 attr->sample_period = term->val.period;
696 attr->freq = 0;
697 break;
698 case PERF_EVSEL__CONFIG_TERM_FREQ:
699 attr->sample_freq = term->val.freq;
700 attr->freq = 1;
701 break;
702 case PERF_EVSEL__CONFIG_TERM_TIME:
703 if (term->val.time)
704 perf_evsel__set_sample_bit(evsel, TIME);
705 else
706 perf_evsel__reset_sample_bit(evsel, TIME);
707 break;
708 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
709 callgraph_buf = term->val.callgraph;
710 break;
711 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
712 dump_size = term->val.stack_user;
713 break;
714 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
715 max_stack = term->val.max_stack;
716 break;
717 case PERF_EVSEL__CONFIG_TERM_INHERIT:
718 /*
719 * attr->inherit should has already been set by
720 * perf_evsel__config. If user explicitly set
721 * inherit using config terms, override global
722 * opt->no_inherit setting.
723 */
724 attr->inherit = term->val.inherit ? 1 : 0;
725 break;
726 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
727 attr->write_backward = term->val.overwrite ? 1 : 0;
728 break;
729 default:
730 break;
731 }
732 }
733
734 /* User explicitly set per-event callgraph, clear the old setting and reset. */
735 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
736 if (max_stack) {
737 param.max_stack = max_stack;
738 if (callgraph_buf == NULL)
739 callgraph_buf = "fp";
740 }
741
742 /* parse callgraph parameters */
743 if (callgraph_buf != NULL) {
744 if (!strcmp(callgraph_buf, "no")) {
745 param.enabled = false;
746 param.record_mode = CALLCHAIN_NONE;
747 } else {
748 param.enabled = true;
749 if (parse_callchain_record(callgraph_buf, &param)) {
750 pr_err("per-event callgraph setting for %s failed. "
751 "Apply callgraph global setting for it\n",
752 evsel->name);
753 return;
754 }
755 }
756 }
757 if (dump_size > 0) {
758 dump_size = round_up(dump_size, sizeof(u64));
759 param.dump_size = dump_size;
760 }
761
762 /* If global callgraph set, clear it */
763 if (callchain_param.enabled)
764 perf_evsel__reset_callgraph(evsel, &callchain_param);
765
766 /* set perf-event callgraph */
767 if (param.enabled)
768 perf_evsel__config_callchain(evsel, opts, &param);
769 }
770 }
771
772 /*
773 * The enable_on_exec/disabled value strategy:
774 *
775 * 1) For any type of traced program:
776 * - all independent events and group leaders are disabled
777 * - all group members are enabled
778 *
779 * Group members are ruled by group leaders. They need to
780 * be enabled, because the group scheduling relies on that.
781 *
782 * 2) For traced programs executed by perf:
783 * - all independent events and group leaders have
784 * enable_on_exec set
785 * - we don't specifically enable or disable any event during
786 * the record command
787 *
788 * Independent events and group leaders are initially disabled
789 * and get enabled by exec. Group members are ruled by group
790 * leaders as stated in 1).
791 *
792 * 3) For traced programs attached by perf (pid/tid):
793 * - we specifically enable or disable all events during
794 * the record command
795 *
796 * When attaching events to already running traced we
797 * enable/disable events specifically, as there's no
798 * initial traced exec call.
799 */
800 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
801 struct callchain_param *callchain)
802 {
803 struct perf_evsel *leader = evsel->leader;
804 struct perf_event_attr *attr = &evsel->attr;
805 int track = evsel->tracking;
806 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
807
808 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
809 attr->inherit = !opts->no_inherit;
810 attr->write_backward = opts->overwrite ? 1 : 0;
811
812 perf_evsel__set_sample_bit(evsel, IP);
813 perf_evsel__set_sample_bit(evsel, TID);
814
815 if (evsel->sample_read) {
816 perf_evsel__set_sample_bit(evsel, READ);
817
818 /*
819 * We need ID even in case of single event, because
820 * PERF_SAMPLE_READ process ID specific data.
821 */
822 perf_evsel__set_sample_id(evsel, false);
823
824 /*
825 * Apply group format only if we belong to group
826 * with more than one members.
827 */
828 if (leader->nr_members > 1) {
829 attr->read_format |= PERF_FORMAT_GROUP;
830 attr->inherit = 0;
831 }
832 }
833
834 /*
835 * We default some events to have a default interval. But keep
836 * it a weak assumption overridable by the user.
837 */
838 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
839 opts->user_interval != ULLONG_MAX)) {
840 if (opts->freq) {
841 perf_evsel__set_sample_bit(evsel, PERIOD);
842 attr->freq = 1;
843 attr->sample_freq = opts->freq;
844 } else {
845 attr->sample_period = opts->default_interval;
846 }
847 }
848
849 /*
850 * Disable sampling for all group members other
851 * than leader in case leader 'leads' the sampling.
852 */
853 if ((leader != evsel) && leader->sample_read) {
854 attr->sample_freq = 0;
855 attr->sample_period = 0;
856 }
857
858 if (opts->no_samples)
859 attr->sample_freq = 0;
860
861 if (opts->inherit_stat)
862 attr->inherit_stat = 1;
863
864 if (opts->sample_address) {
865 perf_evsel__set_sample_bit(evsel, ADDR);
866 attr->mmap_data = track;
867 }
868
869 /*
870 * We don't allow user space callchains for function trace
871 * event, due to issues with page faults while tracing page
872 * fault handler and its overall trickiness nature.
873 */
874 if (perf_evsel__is_function_event(evsel))
875 evsel->attr.exclude_callchain_user = 1;
876
877 if (callchain && callchain->enabled && !evsel->no_aux_samples)
878 perf_evsel__config_callchain(evsel, opts, callchain);
879
880 if (opts->sample_intr_regs) {
881 attr->sample_regs_intr = opts->sample_intr_regs;
882 perf_evsel__set_sample_bit(evsel, REGS_INTR);
883 }
884
885 if (target__has_cpu(&opts->target) || opts->sample_cpu)
886 perf_evsel__set_sample_bit(evsel, CPU);
887
888 if (opts->period)
889 perf_evsel__set_sample_bit(evsel, PERIOD);
890
891 /*
892 * When the user explicitly disabled time don't force it here.
893 */
894 if (opts->sample_time &&
895 (!perf_missing_features.sample_id_all &&
896 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
897 opts->sample_time_set)))
898 perf_evsel__set_sample_bit(evsel, TIME);
899
900 if (opts->raw_samples && !evsel->no_aux_samples) {
901 perf_evsel__set_sample_bit(evsel, TIME);
902 perf_evsel__set_sample_bit(evsel, RAW);
903 perf_evsel__set_sample_bit(evsel, CPU);
904 }
905
906 if (opts->sample_address)
907 perf_evsel__set_sample_bit(evsel, DATA_SRC);
908
909 if (opts->no_buffering) {
910 attr->watermark = 0;
911 attr->wakeup_events = 1;
912 }
913 if (opts->branch_stack && !evsel->no_aux_samples) {
914 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
915 attr->branch_sample_type = opts->branch_stack;
916 }
917
918 if (opts->sample_weight)
919 perf_evsel__set_sample_bit(evsel, WEIGHT);
920
921 attr->task = track;
922 attr->mmap = track;
923 attr->mmap2 = track && !perf_missing_features.mmap2;
924 attr->comm = track;
925
926 if (opts->record_switch_events)
927 attr->context_switch = track;
928
929 if (opts->sample_transaction)
930 perf_evsel__set_sample_bit(evsel, TRANSACTION);
931
932 if (opts->running_time) {
933 evsel->attr.read_format |=
934 PERF_FORMAT_TOTAL_TIME_ENABLED |
935 PERF_FORMAT_TOTAL_TIME_RUNNING;
936 }
937
938 /*
939 * XXX see the function comment above
940 *
941 * Disabling only independent events or group leaders,
942 * keeping group members enabled.
943 */
944 if (perf_evsel__is_group_leader(evsel))
945 attr->disabled = 1;
946
947 /*
948 * Setting enable_on_exec for independent events and
949 * group leaders for traced executed by perf.
950 */
951 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
952 !opts->initial_delay)
953 attr->enable_on_exec = 1;
954
955 if (evsel->immediate) {
956 attr->disabled = 0;
957 attr->enable_on_exec = 0;
958 }
959
960 clockid = opts->clockid;
961 if (opts->use_clockid) {
962 attr->use_clockid = 1;
963 attr->clockid = opts->clockid;
964 }
965
966 if (evsel->precise_max)
967 perf_event_attr__set_max_precise_ip(attr);
968
969 if (opts->all_user) {
970 attr->exclude_kernel = 1;
971 attr->exclude_user = 0;
972 }
973
974 if (opts->all_kernel) {
975 attr->exclude_kernel = 0;
976 attr->exclude_user = 1;
977 }
978
979 /*
980 * Apply event specific term settings,
981 * it overloads any global configuration.
982 */
983 apply_config_terms(evsel, opts);
984 }
985
986 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
987 {
988 int cpu, thread;
989
990 if (evsel->system_wide)
991 nthreads = 1;
992
993 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
994
995 if (evsel->fd) {
996 for (cpu = 0; cpu < ncpus; cpu++) {
997 for (thread = 0; thread < nthreads; thread++) {
998 FD(evsel, cpu, thread) = -1;
999 }
1000 }
1001 }
1002
1003 return evsel->fd != NULL ? 0 : -ENOMEM;
1004 }
1005
1006 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
1007 int ioc, void *arg)
1008 {
1009 int cpu, thread;
1010
1011 if (evsel->system_wide)
1012 nthreads = 1;
1013
1014 for (cpu = 0; cpu < ncpus; cpu++) {
1015 for (thread = 0; thread < nthreads; thread++) {
1016 int fd = FD(evsel, cpu, thread),
1017 err = ioctl(fd, ioc, arg);
1018
1019 if (err)
1020 return err;
1021 }
1022 }
1023
1024 return 0;
1025 }
1026
1027 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
1028 const char *filter)
1029 {
1030 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1031 PERF_EVENT_IOC_SET_FILTER,
1032 (void *)filter);
1033 }
1034
1035 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1036 {
1037 char *new_filter = strdup(filter);
1038
1039 if (new_filter != NULL) {
1040 free(evsel->filter);
1041 evsel->filter = new_filter;
1042 return 0;
1043 }
1044
1045 return -1;
1046 }
1047
1048 int perf_evsel__append_filter(struct perf_evsel *evsel,
1049 const char *op, const char *filter)
1050 {
1051 char *new_filter;
1052
1053 if (evsel->filter == NULL)
1054 return perf_evsel__set_filter(evsel, filter);
1055
1056 if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
1057 free(evsel->filter);
1058 evsel->filter = new_filter;
1059 return 0;
1060 }
1061
1062 return -1;
1063 }
1064
1065 int perf_evsel__enable(struct perf_evsel *evsel)
1066 {
1067 int nthreads = thread_map__nr(evsel->threads);
1068 int ncpus = cpu_map__nr(evsel->cpus);
1069
1070 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1071 PERF_EVENT_IOC_ENABLE,
1072 0);
1073 }
1074
1075 int perf_evsel__disable(struct perf_evsel *evsel)
1076 {
1077 int nthreads = thread_map__nr(evsel->threads);
1078 int ncpus = cpu_map__nr(evsel->cpus);
1079
1080 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1081 PERF_EVENT_IOC_DISABLE,
1082 0);
1083 }
1084
1085 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1086 {
1087 if (ncpus == 0 || nthreads == 0)
1088 return 0;
1089
1090 if (evsel->system_wide)
1091 nthreads = 1;
1092
1093 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1094 if (evsel->sample_id == NULL)
1095 return -ENOMEM;
1096
1097 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1098 if (evsel->id == NULL) {
1099 xyarray__delete(evsel->sample_id);
1100 evsel->sample_id = NULL;
1101 return -ENOMEM;
1102 }
1103
1104 return 0;
1105 }
1106
1107 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1108 {
1109 xyarray__delete(evsel->fd);
1110 evsel->fd = NULL;
1111 }
1112
1113 static void perf_evsel__free_id(struct perf_evsel *evsel)
1114 {
1115 xyarray__delete(evsel->sample_id);
1116 evsel->sample_id = NULL;
1117 zfree(&evsel->id);
1118 }
1119
1120 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1121 {
1122 struct perf_evsel_config_term *term, *h;
1123
1124 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1125 list_del(&term->list);
1126 free(term);
1127 }
1128 }
1129
1130 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1131 {
1132 int cpu, thread;
1133
1134 if (evsel->system_wide)
1135 nthreads = 1;
1136
1137 for (cpu = 0; cpu < ncpus; cpu++)
1138 for (thread = 0; thread < nthreads; ++thread) {
1139 close(FD(evsel, cpu, thread));
1140 FD(evsel, cpu, thread) = -1;
1141 }
1142 }
1143
1144 void perf_evsel__exit(struct perf_evsel *evsel)
1145 {
1146 assert(list_empty(&evsel->node));
1147 assert(evsel->evlist == NULL);
1148 perf_evsel__free_fd(evsel);
1149 perf_evsel__free_id(evsel);
1150 perf_evsel__free_config_terms(evsel);
1151 close_cgroup(evsel->cgrp);
1152 cpu_map__put(evsel->cpus);
1153 cpu_map__put(evsel->own_cpus);
1154 thread_map__put(evsel->threads);
1155 zfree(&evsel->group_name);
1156 zfree(&evsel->name);
1157 perf_evsel__object.fini(evsel);
1158 }
1159
1160 void perf_evsel__delete(struct perf_evsel *evsel)
1161 {
1162 perf_evsel__exit(evsel);
1163 free(evsel);
1164 }
1165
1166 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1167 struct perf_counts_values *count)
1168 {
1169 struct perf_counts_values tmp;
1170
1171 if (!evsel->prev_raw_counts)
1172 return;
1173
1174 if (cpu == -1) {
1175 tmp = evsel->prev_raw_counts->aggr;
1176 evsel->prev_raw_counts->aggr = *count;
1177 } else {
1178 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1179 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1180 }
1181
1182 count->val = count->val - tmp.val;
1183 count->ena = count->ena - tmp.ena;
1184 count->run = count->run - tmp.run;
1185 }
1186
1187 void perf_counts_values__scale(struct perf_counts_values *count,
1188 bool scale, s8 *pscaled)
1189 {
1190 s8 scaled = 0;
1191
1192 if (scale) {
1193 if (count->run == 0) {
1194 scaled = -1;
1195 count->val = 0;
1196 } else if (count->run < count->ena) {
1197 scaled = 1;
1198 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1199 }
1200 } else
1201 count->ena = count->run = 0;
1202
1203 if (pscaled)
1204 *pscaled = scaled;
1205 }
1206
1207 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1208 struct perf_counts_values *count)
1209 {
1210 memset(count, 0, sizeof(*count));
1211
1212 if (FD(evsel, cpu, thread) < 0)
1213 return -EINVAL;
1214
1215 if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1216 return -errno;
1217
1218 return 0;
1219 }
1220
1221 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1222 int cpu, int thread, bool scale)
1223 {
1224 struct perf_counts_values count;
1225 size_t nv = scale ? 3 : 1;
1226
1227 if (FD(evsel, cpu, thread) < 0)
1228 return -EINVAL;
1229
1230 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1231 return -ENOMEM;
1232
1233 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1234 return -errno;
1235
1236 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1237 perf_counts_values__scale(&count, scale, NULL);
1238 *perf_counts(evsel->counts, cpu, thread) = count;
1239 return 0;
1240 }
1241
1242 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1243 {
1244 struct perf_evsel *leader = evsel->leader;
1245 int fd;
1246
1247 if (perf_evsel__is_group_leader(evsel))
1248 return -1;
1249
1250 /*
1251 * Leader must be already processed/open,
1252 * if not it's a bug.
1253 */
1254 BUG_ON(!leader->fd);
1255
1256 fd = FD(leader, cpu, thread);
1257 BUG_ON(fd == -1);
1258
1259 return fd;
1260 }
1261
1262 struct bit_names {
1263 int bit;
1264 const char *name;
1265 };
1266
1267 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1268 {
1269 bool first_bit = true;
1270 int i = 0;
1271
1272 do {
1273 if (value & bits[i].bit) {
1274 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1275 first_bit = false;
1276 }
1277 } while (bits[++i].name != NULL);
1278 }
1279
1280 static void __p_sample_type(char *buf, size_t size, u64 value)
1281 {
1282 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1283 struct bit_names bits[] = {
1284 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1285 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1286 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1287 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1288 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1289 bit_name(WEIGHT),
1290 { .name = NULL, }
1291 };
1292 #undef bit_name
1293 __p_bits(buf, size, value, bits);
1294 }
1295
1296 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1297 {
1298 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1299 struct bit_names bits[] = {
1300 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1301 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1302 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1303 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1304 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1305 { .name = NULL, }
1306 };
1307 #undef bit_name
1308 __p_bits(buf, size, value, bits);
1309 }
1310
1311 static void __p_read_format(char *buf, size_t size, u64 value)
1312 {
1313 #define bit_name(n) { PERF_FORMAT_##n, #n }
1314 struct bit_names bits[] = {
1315 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1316 bit_name(ID), bit_name(GROUP),
1317 { .name = NULL, }
1318 };
1319 #undef bit_name
1320 __p_bits(buf, size, value, bits);
1321 }
1322
1323 #define BUF_SIZE 1024
1324
1325 #define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1326 #define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1327 #define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1328 #define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1329 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1330 #define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1331
1332 #define PRINT_ATTRn(_n, _f, _p) \
1333 do { \
1334 if (attr->_f) { \
1335 _p(attr->_f); \
1336 ret += attr__fprintf(fp, _n, buf, priv);\
1337 } \
1338 } while (0)
1339
1340 #define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1341
1342 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1343 attr__fprintf_f attr__fprintf, void *priv)
1344 {
1345 char buf[BUF_SIZE];
1346 int ret = 0;
1347
1348 PRINT_ATTRf(type, p_unsigned);
1349 PRINT_ATTRf(size, p_unsigned);
1350 PRINT_ATTRf(config, p_hex);
1351 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1352 PRINT_ATTRf(sample_type, p_sample_type);
1353 PRINT_ATTRf(read_format, p_read_format);
1354
1355 PRINT_ATTRf(disabled, p_unsigned);
1356 PRINT_ATTRf(inherit, p_unsigned);
1357 PRINT_ATTRf(pinned, p_unsigned);
1358 PRINT_ATTRf(exclusive, p_unsigned);
1359 PRINT_ATTRf(exclude_user, p_unsigned);
1360 PRINT_ATTRf(exclude_kernel, p_unsigned);
1361 PRINT_ATTRf(exclude_hv, p_unsigned);
1362 PRINT_ATTRf(exclude_idle, p_unsigned);
1363 PRINT_ATTRf(mmap, p_unsigned);
1364 PRINT_ATTRf(comm, p_unsigned);
1365 PRINT_ATTRf(freq, p_unsigned);
1366 PRINT_ATTRf(inherit_stat, p_unsigned);
1367 PRINT_ATTRf(enable_on_exec, p_unsigned);
1368 PRINT_ATTRf(task, p_unsigned);
1369 PRINT_ATTRf(watermark, p_unsigned);
1370 PRINT_ATTRf(precise_ip, p_unsigned);
1371 PRINT_ATTRf(mmap_data, p_unsigned);
1372 PRINT_ATTRf(sample_id_all, p_unsigned);
1373 PRINT_ATTRf(exclude_host, p_unsigned);
1374 PRINT_ATTRf(exclude_guest, p_unsigned);
1375 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1376 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1377 PRINT_ATTRf(mmap2, p_unsigned);
1378 PRINT_ATTRf(comm_exec, p_unsigned);
1379 PRINT_ATTRf(use_clockid, p_unsigned);
1380 PRINT_ATTRf(context_switch, p_unsigned);
1381 PRINT_ATTRf(write_backward, p_unsigned);
1382
1383 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1384 PRINT_ATTRf(bp_type, p_unsigned);
1385 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1386 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1387 PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1388 PRINT_ATTRf(sample_regs_user, p_hex);
1389 PRINT_ATTRf(sample_stack_user, p_unsigned);
1390 PRINT_ATTRf(clockid, p_signed);
1391 PRINT_ATTRf(sample_regs_intr, p_hex);
1392 PRINT_ATTRf(aux_watermark, p_unsigned);
1393 PRINT_ATTRf(sample_max_stack, p_unsigned);
1394
1395 return ret;
1396 }
1397
1398 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1399 void *priv __attribute__((unused)))
1400 {
1401 return fprintf(fp, " %-32s %s\n", name, val);
1402 }
1403
1404 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1405 struct thread_map *threads)
1406 {
1407 int cpu, thread, nthreads;
1408 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1409 int pid = -1, err;
1410 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1411
1412 if (perf_missing_features.write_backward && evsel->attr.write_backward)
1413 return -EINVAL;
1414
1415 if (evsel->system_wide)
1416 nthreads = 1;
1417 else
1418 nthreads = threads->nr;
1419
1420 if (evsel->fd == NULL &&
1421 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1422 return -ENOMEM;
1423
1424 if (evsel->cgrp) {
1425 flags |= PERF_FLAG_PID_CGROUP;
1426 pid = evsel->cgrp->fd;
1427 }
1428
1429 fallback_missing_features:
1430 if (perf_missing_features.clockid_wrong)
1431 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1432 if (perf_missing_features.clockid) {
1433 evsel->attr.use_clockid = 0;
1434 evsel->attr.clockid = 0;
1435 }
1436 if (perf_missing_features.cloexec)
1437 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1438 if (perf_missing_features.mmap2)
1439 evsel->attr.mmap2 = 0;
1440 if (perf_missing_features.exclude_guest)
1441 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1442 if (perf_missing_features.lbr_flags)
1443 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1444 PERF_SAMPLE_BRANCH_NO_CYCLES);
1445 retry_sample_id:
1446 if (perf_missing_features.sample_id_all)
1447 evsel->attr.sample_id_all = 0;
1448
1449 if (verbose >= 2) {
1450 fprintf(stderr, "%.60s\n", graph_dotted_line);
1451 fprintf(stderr, "perf_event_attr:\n");
1452 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1453 fprintf(stderr, "%.60s\n", graph_dotted_line);
1454 }
1455
1456 for (cpu = 0; cpu < cpus->nr; cpu++) {
1457
1458 for (thread = 0; thread < nthreads; thread++) {
1459 int group_fd;
1460
1461 if (!evsel->cgrp && !evsel->system_wide)
1462 pid = thread_map__pid(threads, thread);
1463
1464 group_fd = get_group_fd(evsel, cpu, thread);
1465 retry_open:
1466 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1467 pid, cpus->map[cpu], group_fd, flags);
1468
1469 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1470 pid,
1471 cpus->map[cpu],
1472 group_fd, flags);
1473 if (FD(evsel, cpu, thread) < 0) {
1474 err = -errno;
1475 pr_debug2("sys_perf_event_open failed, error %d\n",
1476 err);
1477 goto try_fallback;
1478 }
1479
1480 if (evsel->bpf_fd >= 0) {
1481 int evt_fd = FD(evsel, cpu, thread);
1482 int bpf_fd = evsel->bpf_fd;
1483
1484 err = ioctl(evt_fd,
1485 PERF_EVENT_IOC_SET_BPF,
1486 bpf_fd);
1487 if (err && errno != EEXIST) {
1488 pr_err("failed to attach bpf fd %d: %s\n",
1489 bpf_fd, strerror(errno));
1490 err = -EINVAL;
1491 goto out_close;
1492 }
1493 }
1494
1495 set_rlimit = NO_CHANGE;
1496
1497 /*
1498 * If we succeeded but had to kill clockid, fail and
1499 * have perf_evsel__open_strerror() print us a nice
1500 * error.
1501 */
1502 if (perf_missing_features.clockid ||
1503 perf_missing_features.clockid_wrong) {
1504 err = -EINVAL;
1505 goto out_close;
1506 }
1507 }
1508 }
1509
1510 return 0;
1511
1512 try_fallback:
1513 /*
1514 * perf stat needs between 5 and 22 fds per CPU. When we run out
1515 * of them try to increase the limits.
1516 */
1517 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1518 struct rlimit l;
1519 int old_errno = errno;
1520
1521 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1522 if (set_rlimit == NO_CHANGE)
1523 l.rlim_cur = l.rlim_max;
1524 else {
1525 l.rlim_cur = l.rlim_max + 1000;
1526 l.rlim_max = l.rlim_cur;
1527 }
1528 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1529 set_rlimit++;
1530 errno = old_errno;
1531 goto retry_open;
1532 }
1533 }
1534 errno = old_errno;
1535 }
1536
1537 if (err != -EINVAL || cpu > 0 || thread > 0)
1538 goto out_close;
1539
1540 /*
1541 * Must probe features in the order they were added to the
1542 * perf_event_attr interface.
1543 */
1544 if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1545 perf_missing_features.write_backward = true;
1546 goto out_close;
1547 } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1548 perf_missing_features.clockid_wrong = true;
1549 goto fallback_missing_features;
1550 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1551 perf_missing_features.clockid = true;
1552 goto fallback_missing_features;
1553 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1554 perf_missing_features.cloexec = true;
1555 goto fallback_missing_features;
1556 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1557 perf_missing_features.mmap2 = true;
1558 goto fallback_missing_features;
1559 } else if (!perf_missing_features.exclude_guest &&
1560 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1561 perf_missing_features.exclude_guest = true;
1562 goto fallback_missing_features;
1563 } else if (!perf_missing_features.sample_id_all) {
1564 perf_missing_features.sample_id_all = true;
1565 goto retry_sample_id;
1566 } else if (!perf_missing_features.lbr_flags &&
1567 (evsel->attr.branch_sample_type &
1568 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1569 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1570 perf_missing_features.lbr_flags = true;
1571 goto fallback_missing_features;
1572 }
1573 out_close:
1574 do {
1575 while (--thread >= 0) {
1576 close(FD(evsel, cpu, thread));
1577 FD(evsel, cpu, thread) = -1;
1578 }
1579 thread = nthreads;
1580 } while (--cpu >= 0);
1581 return err;
1582 }
1583
1584 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1585 {
1586 if (evsel->fd == NULL)
1587 return;
1588
1589 perf_evsel__close_fd(evsel, ncpus, nthreads);
1590 perf_evsel__free_fd(evsel);
1591 }
1592
1593 static struct {
1594 struct cpu_map map;
1595 int cpus[1];
1596 } empty_cpu_map = {
1597 .map.nr = 1,
1598 .cpus = { -1, },
1599 };
1600
1601 static struct {
1602 struct thread_map map;
1603 int threads[1];
1604 } empty_thread_map = {
1605 .map.nr = 1,
1606 .threads = { -1, },
1607 };
1608
1609 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1610 struct thread_map *threads)
1611 {
1612 if (cpus == NULL) {
1613 /* Work around old compiler warnings about strict aliasing */
1614 cpus = &empty_cpu_map.map;
1615 }
1616
1617 if (threads == NULL)
1618 threads = &empty_thread_map.map;
1619
1620 return __perf_evsel__open(evsel, cpus, threads);
1621 }
1622
1623 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1624 struct cpu_map *cpus)
1625 {
1626 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1627 }
1628
1629 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1630 struct thread_map *threads)
1631 {
1632 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1633 }
1634
1635 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1636 const union perf_event *event,
1637 struct perf_sample *sample)
1638 {
1639 u64 type = evsel->attr.sample_type;
1640 const u64 *array = event->sample.array;
1641 bool swapped = evsel->needs_swap;
1642 union u64_swap u;
1643
1644 array += ((event->header.size -
1645 sizeof(event->header)) / sizeof(u64)) - 1;
1646
1647 if (type & PERF_SAMPLE_IDENTIFIER) {
1648 sample->id = *array;
1649 array--;
1650 }
1651
1652 if (type & PERF_SAMPLE_CPU) {
1653 u.val64 = *array;
1654 if (swapped) {
1655 /* undo swap of u64, then swap on individual u32s */
1656 u.val64 = bswap_64(u.val64);
1657 u.val32[0] = bswap_32(u.val32[0]);
1658 }
1659
1660 sample->cpu = u.val32[0];
1661 array--;
1662 }
1663
1664 if (type & PERF_SAMPLE_STREAM_ID) {
1665 sample->stream_id = *array;
1666 array--;
1667 }
1668
1669 if (type & PERF_SAMPLE_ID) {
1670 sample->id = *array;
1671 array--;
1672 }
1673
1674 if (type & PERF_SAMPLE_TIME) {
1675 sample->time = *array;
1676 array--;
1677 }
1678
1679 if (type & PERF_SAMPLE_TID) {
1680 u.val64 = *array;
1681 if (swapped) {
1682 /* undo swap of u64, then swap on individual u32s */
1683 u.val64 = bswap_64(u.val64);
1684 u.val32[0] = bswap_32(u.val32[0]);
1685 u.val32[1] = bswap_32(u.val32[1]);
1686 }
1687
1688 sample->pid = u.val32[0];
1689 sample->tid = u.val32[1];
1690 array--;
1691 }
1692
1693 return 0;
1694 }
1695
1696 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1697 u64 size)
1698 {
1699 return size > max_size || offset + size > endp;
1700 }
1701
1702 #define OVERFLOW_CHECK(offset, size, max_size) \
1703 do { \
1704 if (overflow(endp, (max_size), (offset), (size))) \
1705 return -EFAULT; \
1706 } while (0)
1707
1708 #define OVERFLOW_CHECK_u64(offset) \
1709 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1710
1711 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1712 struct perf_sample *data)
1713 {
1714 u64 type = evsel->attr.sample_type;
1715 bool swapped = evsel->needs_swap;
1716 const u64 *array;
1717 u16 max_size = event->header.size;
1718 const void *endp = (void *)event + max_size;
1719 u64 sz;
1720
1721 /*
1722 * used for cross-endian analysis. See git commit 65014ab3
1723 * for why this goofiness is needed.
1724 */
1725 union u64_swap u;
1726
1727 memset(data, 0, sizeof(*data));
1728 data->cpu = data->pid = data->tid = -1;
1729 data->stream_id = data->id = data->time = -1ULL;
1730 data->period = evsel->attr.sample_period;
1731 data->weight = 0;
1732 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1733
1734 if (event->header.type != PERF_RECORD_SAMPLE) {
1735 if (!evsel->attr.sample_id_all)
1736 return 0;
1737 return perf_evsel__parse_id_sample(evsel, event, data);
1738 }
1739
1740 array = event->sample.array;
1741
1742 /*
1743 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1744 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1745 * check the format does not go past the end of the event.
1746 */
1747 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1748 return -EFAULT;
1749
1750 data->id = -1ULL;
1751 if (type & PERF_SAMPLE_IDENTIFIER) {
1752 data->id = *array;
1753 array++;
1754 }
1755
1756 if (type & PERF_SAMPLE_IP) {
1757 data->ip = *array;
1758 array++;
1759 }
1760
1761 if (type & PERF_SAMPLE_TID) {
1762 u.val64 = *array;
1763 if (swapped) {
1764 /* undo swap of u64, then swap on individual u32s */
1765 u.val64 = bswap_64(u.val64);
1766 u.val32[0] = bswap_32(u.val32[0]);
1767 u.val32[1] = bswap_32(u.val32[1]);
1768 }
1769
1770 data->pid = u.val32[0];
1771 data->tid = u.val32[1];
1772 array++;
1773 }
1774
1775 if (type & PERF_SAMPLE_TIME) {
1776 data->time = *array;
1777 array++;
1778 }
1779
1780 data->addr = 0;
1781 if (type & PERF_SAMPLE_ADDR) {
1782 data->addr = *array;
1783 array++;
1784 }
1785
1786 if (type & PERF_SAMPLE_ID) {
1787 data->id = *array;
1788 array++;
1789 }
1790
1791 if (type & PERF_SAMPLE_STREAM_ID) {
1792 data->stream_id = *array;
1793 array++;
1794 }
1795
1796 if (type & PERF_SAMPLE_CPU) {
1797
1798 u.val64 = *array;
1799 if (swapped) {
1800 /* undo swap of u64, then swap on individual u32s */
1801 u.val64 = bswap_64(u.val64);
1802 u.val32[0] = bswap_32(u.val32[0]);
1803 }
1804
1805 data->cpu = u.val32[0];
1806 array++;
1807 }
1808
1809 if (type & PERF_SAMPLE_PERIOD) {
1810 data->period = *array;
1811 array++;
1812 }
1813
1814 if (type & PERF_SAMPLE_READ) {
1815 u64 read_format = evsel->attr.read_format;
1816
1817 OVERFLOW_CHECK_u64(array);
1818 if (read_format & PERF_FORMAT_GROUP)
1819 data->read.group.nr = *array;
1820 else
1821 data->read.one.value = *array;
1822
1823 array++;
1824
1825 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1826 OVERFLOW_CHECK_u64(array);
1827 data->read.time_enabled = *array;
1828 array++;
1829 }
1830
1831 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1832 OVERFLOW_CHECK_u64(array);
1833 data->read.time_running = *array;
1834 array++;
1835 }
1836
1837 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1838 if (read_format & PERF_FORMAT_GROUP) {
1839 const u64 max_group_nr = UINT64_MAX /
1840 sizeof(struct sample_read_value);
1841
1842 if (data->read.group.nr > max_group_nr)
1843 return -EFAULT;
1844 sz = data->read.group.nr *
1845 sizeof(struct sample_read_value);
1846 OVERFLOW_CHECK(array, sz, max_size);
1847 data->read.group.values =
1848 (struct sample_read_value *)array;
1849 array = (void *)array + sz;
1850 } else {
1851 OVERFLOW_CHECK_u64(array);
1852 data->read.one.id = *array;
1853 array++;
1854 }
1855 }
1856
1857 if (type & PERF_SAMPLE_CALLCHAIN) {
1858 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1859
1860 OVERFLOW_CHECK_u64(array);
1861 data->callchain = (struct ip_callchain *)array++;
1862 if (data->callchain->nr > max_callchain_nr)
1863 return -EFAULT;
1864 sz = data->callchain->nr * sizeof(u64);
1865 OVERFLOW_CHECK(array, sz, max_size);
1866 array = (void *)array + sz;
1867 }
1868
1869 if (type & PERF_SAMPLE_RAW) {
1870 OVERFLOW_CHECK_u64(array);
1871 u.val64 = *array;
1872 if (WARN_ONCE(swapped,
1873 "Endianness of raw data not corrected!\n")) {
1874 /* undo swap of u64, then swap on individual u32s */
1875 u.val64 = bswap_64(u.val64);
1876 u.val32[0] = bswap_32(u.val32[0]);
1877 u.val32[1] = bswap_32(u.val32[1]);
1878 }
1879 data->raw_size = u.val32[0];
1880 array = (void *)array + sizeof(u32);
1881
1882 OVERFLOW_CHECK(array, data->raw_size, max_size);
1883 data->raw_data = (void *)array;
1884 array = (void *)array + data->raw_size;
1885 }
1886
1887 if (type & PERF_SAMPLE_BRANCH_STACK) {
1888 const u64 max_branch_nr = UINT64_MAX /
1889 sizeof(struct branch_entry);
1890
1891 OVERFLOW_CHECK_u64(array);
1892 data->branch_stack = (struct branch_stack *)array++;
1893
1894 if (data->branch_stack->nr > max_branch_nr)
1895 return -EFAULT;
1896 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1897 OVERFLOW_CHECK(array, sz, max_size);
1898 array = (void *)array + sz;
1899 }
1900
1901 if (type & PERF_SAMPLE_REGS_USER) {
1902 OVERFLOW_CHECK_u64(array);
1903 data->user_regs.abi = *array;
1904 array++;
1905
1906 if (data->user_regs.abi) {
1907 u64 mask = evsel->attr.sample_regs_user;
1908
1909 sz = hweight_long(mask) * sizeof(u64);
1910 OVERFLOW_CHECK(array, sz, max_size);
1911 data->user_regs.mask = mask;
1912 data->user_regs.regs = (u64 *)array;
1913 array = (void *)array + sz;
1914 }
1915 }
1916
1917 if (type & PERF_SAMPLE_STACK_USER) {
1918 OVERFLOW_CHECK_u64(array);
1919 sz = *array++;
1920
1921 data->user_stack.offset = ((char *)(array - 1)
1922 - (char *) event);
1923
1924 if (!sz) {
1925 data->user_stack.size = 0;
1926 } else {
1927 OVERFLOW_CHECK(array, sz, max_size);
1928 data->user_stack.data = (char *)array;
1929 array = (void *)array + sz;
1930 OVERFLOW_CHECK_u64(array);
1931 data->user_stack.size = *array++;
1932 if (WARN_ONCE(data->user_stack.size > sz,
1933 "user stack dump failure\n"))
1934 return -EFAULT;
1935 }
1936 }
1937
1938 data->weight = 0;
1939 if (type & PERF_SAMPLE_WEIGHT) {
1940 OVERFLOW_CHECK_u64(array);
1941 data->weight = *array;
1942 array++;
1943 }
1944
1945 data->data_src = PERF_MEM_DATA_SRC_NONE;
1946 if (type & PERF_SAMPLE_DATA_SRC) {
1947 OVERFLOW_CHECK_u64(array);
1948 data->data_src = *array;
1949 array++;
1950 }
1951
1952 data->transaction = 0;
1953 if (type & PERF_SAMPLE_TRANSACTION) {
1954 OVERFLOW_CHECK_u64(array);
1955 data->transaction = *array;
1956 array++;
1957 }
1958
1959 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1960 if (type & PERF_SAMPLE_REGS_INTR) {
1961 OVERFLOW_CHECK_u64(array);
1962 data->intr_regs.abi = *array;
1963 array++;
1964
1965 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1966 u64 mask = evsel->attr.sample_regs_intr;
1967
1968 sz = hweight_long(mask) * sizeof(u64);
1969 OVERFLOW_CHECK(array, sz, max_size);
1970 data->intr_regs.mask = mask;
1971 data->intr_regs.regs = (u64 *)array;
1972 array = (void *)array + sz;
1973 }
1974 }
1975
1976 return 0;
1977 }
1978
1979 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1980 u64 read_format)
1981 {
1982 size_t sz, result = sizeof(struct sample_event);
1983
1984 if (type & PERF_SAMPLE_IDENTIFIER)
1985 result += sizeof(u64);
1986
1987 if (type & PERF_SAMPLE_IP)
1988 result += sizeof(u64);
1989
1990 if (type & PERF_SAMPLE_TID)
1991 result += sizeof(u64);
1992
1993 if (type & PERF_SAMPLE_TIME)
1994 result += sizeof(u64);
1995
1996 if (type & PERF_SAMPLE_ADDR)
1997 result += sizeof(u64);
1998
1999 if (type & PERF_SAMPLE_ID)
2000 result += sizeof(u64);
2001
2002 if (type & PERF_SAMPLE_STREAM_ID)
2003 result += sizeof(u64);
2004
2005 if (type & PERF_SAMPLE_CPU)
2006 result += sizeof(u64);
2007
2008 if (type & PERF_SAMPLE_PERIOD)
2009 result += sizeof(u64);
2010
2011 if (type & PERF_SAMPLE_READ) {
2012 result += sizeof(u64);
2013 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2014 result += sizeof(u64);
2015 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2016 result += sizeof(u64);
2017 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2018 if (read_format & PERF_FORMAT_GROUP) {
2019 sz = sample->read.group.nr *
2020 sizeof(struct sample_read_value);
2021 result += sz;
2022 } else {
2023 result += sizeof(u64);
2024 }
2025 }
2026
2027 if (type & PERF_SAMPLE_CALLCHAIN) {
2028 sz = (sample->callchain->nr + 1) * sizeof(u64);
2029 result += sz;
2030 }
2031
2032 if (type & PERF_SAMPLE_RAW) {
2033 result += sizeof(u32);
2034 result += sample->raw_size;
2035 }
2036
2037 if (type & PERF_SAMPLE_BRANCH_STACK) {
2038 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2039 sz += sizeof(u64);
2040 result += sz;
2041 }
2042
2043 if (type & PERF_SAMPLE_REGS_USER) {
2044 if (sample->user_regs.abi) {
2045 result += sizeof(u64);
2046 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2047 result += sz;
2048 } else {
2049 result += sizeof(u64);
2050 }
2051 }
2052
2053 if (type & PERF_SAMPLE_STACK_USER) {
2054 sz = sample->user_stack.size;
2055 result += sizeof(u64);
2056 if (sz) {
2057 result += sz;
2058 result += sizeof(u64);
2059 }
2060 }
2061
2062 if (type & PERF_SAMPLE_WEIGHT)
2063 result += sizeof(u64);
2064
2065 if (type & PERF_SAMPLE_DATA_SRC)
2066 result += sizeof(u64);
2067
2068 if (type & PERF_SAMPLE_TRANSACTION)
2069 result += sizeof(u64);
2070
2071 if (type & PERF_SAMPLE_REGS_INTR) {
2072 if (sample->intr_regs.abi) {
2073 result += sizeof(u64);
2074 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2075 result += sz;
2076 } else {
2077 result += sizeof(u64);
2078 }
2079 }
2080
2081 return result;
2082 }
2083
2084 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2085 u64 read_format,
2086 const struct perf_sample *sample,
2087 bool swapped)
2088 {
2089 u64 *array;
2090 size_t sz;
2091 /*
2092 * used for cross-endian analysis. See git commit 65014ab3
2093 * for why this goofiness is needed.
2094 */
2095 union u64_swap u;
2096
2097 array = event->sample.array;
2098
2099 if (type & PERF_SAMPLE_IDENTIFIER) {
2100 *array = sample->id;
2101 array++;
2102 }
2103
2104 if (type & PERF_SAMPLE_IP) {
2105 *array = sample->ip;
2106 array++;
2107 }
2108
2109 if (type & PERF_SAMPLE_TID) {
2110 u.val32[0] = sample->pid;
2111 u.val32[1] = sample->tid;
2112 if (swapped) {
2113 /*
2114 * Inverse of what is done in perf_evsel__parse_sample
2115 */
2116 u.val32[0] = bswap_32(u.val32[0]);
2117 u.val32[1] = bswap_32(u.val32[1]);
2118 u.val64 = bswap_64(u.val64);
2119 }
2120
2121 *array = u.val64;
2122 array++;
2123 }
2124
2125 if (type & PERF_SAMPLE_TIME) {
2126 *array = sample->time;
2127 array++;
2128 }
2129
2130 if (type & PERF_SAMPLE_ADDR) {
2131 *array = sample->addr;
2132 array++;
2133 }
2134
2135 if (type & PERF_SAMPLE_ID) {
2136 *array = sample->id;
2137 array++;
2138 }
2139
2140 if (type & PERF_SAMPLE_STREAM_ID) {
2141 *array = sample->stream_id;
2142 array++;
2143 }
2144
2145 if (type & PERF_SAMPLE_CPU) {
2146 u.val32[0] = sample->cpu;
2147 if (swapped) {
2148 /*
2149 * Inverse of what is done in perf_evsel__parse_sample
2150 */
2151 u.val32[0] = bswap_32(u.val32[0]);
2152 u.val64 = bswap_64(u.val64);
2153 }
2154 *array = u.val64;
2155 array++;
2156 }
2157
2158 if (type & PERF_SAMPLE_PERIOD) {
2159 *array = sample->period;
2160 array++;
2161 }
2162
2163 if (type & PERF_SAMPLE_READ) {
2164 if (read_format & PERF_FORMAT_GROUP)
2165 *array = sample->read.group.nr;
2166 else
2167 *array = sample->read.one.value;
2168 array++;
2169
2170 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2171 *array = sample->read.time_enabled;
2172 array++;
2173 }
2174
2175 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2176 *array = sample->read.time_running;
2177 array++;
2178 }
2179
2180 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2181 if (read_format & PERF_FORMAT_GROUP) {
2182 sz = sample->read.group.nr *
2183 sizeof(struct sample_read_value);
2184 memcpy(array, sample->read.group.values, sz);
2185 array = (void *)array + sz;
2186 } else {
2187 *array = sample->read.one.id;
2188 array++;
2189 }
2190 }
2191
2192 if (type & PERF_SAMPLE_CALLCHAIN) {
2193 sz = (sample->callchain->nr + 1) * sizeof(u64);
2194 memcpy(array, sample->callchain, sz);
2195 array = (void *)array + sz;
2196 }
2197
2198 if (type & PERF_SAMPLE_RAW) {
2199 u.val32[0] = sample->raw_size;
2200 if (WARN_ONCE(swapped,
2201 "Endianness of raw data not corrected!\n")) {
2202 /*
2203 * Inverse of what is done in perf_evsel__parse_sample
2204 */
2205 u.val32[0] = bswap_32(u.val32[0]);
2206 u.val32[1] = bswap_32(u.val32[1]);
2207 u.val64 = bswap_64(u.val64);
2208 }
2209 *array = u.val64;
2210 array = (void *)array + sizeof(u32);
2211
2212 memcpy(array, sample->raw_data, sample->raw_size);
2213 array = (void *)array + sample->raw_size;
2214 }
2215
2216 if (type & PERF_SAMPLE_BRANCH_STACK) {
2217 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2218 sz += sizeof(u64);
2219 memcpy(array, sample->branch_stack, sz);
2220 array = (void *)array + sz;
2221 }
2222
2223 if (type & PERF_SAMPLE_REGS_USER) {
2224 if (sample->user_regs.abi) {
2225 *array++ = sample->user_regs.abi;
2226 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2227 memcpy(array, sample->user_regs.regs, sz);
2228 array = (void *)array + sz;
2229 } else {
2230 *array++ = 0;
2231 }
2232 }
2233
2234 if (type & PERF_SAMPLE_STACK_USER) {
2235 sz = sample->user_stack.size;
2236 *array++ = sz;
2237 if (sz) {
2238 memcpy(array, sample->user_stack.data, sz);
2239 array = (void *)array + sz;
2240 *array++ = sz;
2241 }
2242 }
2243
2244 if (type & PERF_SAMPLE_WEIGHT) {
2245 *array = sample->weight;
2246 array++;
2247 }
2248
2249 if (type & PERF_SAMPLE_DATA_SRC) {
2250 *array = sample->data_src;
2251 array++;
2252 }
2253
2254 if (type & PERF_SAMPLE_TRANSACTION) {
2255 *array = sample->transaction;
2256 array++;
2257 }
2258
2259 if (type & PERF_SAMPLE_REGS_INTR) {
2260 if (sample->intr_regs.abi) {
2261 *array++ = sample->intr_regs.abi;
2262 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2263 memcpy(array, sample->intr_regs.regs, sz);
2264 array = (void *)array + sz;
2265 } else {
2266 *array++ = 0;
2267 }
2268 }
2269
2270 return 0;
2271 }
2272
2273 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2274 {
2275 return pevent_find_field(evsel->tp_format, name);
2276 }
2277
2278 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2279 const char *name)
2280 {
2281 struct format_field *field = perf_evsel__field(evsel, name);
2282 int offset;
2283
2284 if (!field)
2285 return NULL;
2286
2287 offset = field->offset;
2288
2289 if (field->flags & FIELD_IS_DYNAMIC) {
2290 offset = *(int *)(sample->raw_data + field->offset);
2291 offset &= 0xffff;
2292 }
2293
2294 return sample->raw_data + offset;
2295 }
2296
2297 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2298 bool needs_swap)
2299 {
2300 u64 value;
2301 void *ptr = sample->raw_data + field->offset;
2302
2303 switch (field->size) {
2304 case 1:
2305 return *(u8 *)ptr;
2306 case 2:
2307 value = *(u16 *)ptr;
2308 break;
2309 case 4:
2310 value = *(u32 *)ptr;
2311 break;
2312 case 8:
2313 memcpy(&value, ptr, sizeof(u64));
2314 break;
2315 default:
2316 return 0;
2317 }
2318
2319 if (!needs_swap)
2320 return value;
2321
2322 switch (field->size) {
2323 case 2:
2324 return bswap_16(value);
2325 case 4:
2326 return bswap_32(value);
2327 case 8:
2328 return bswap_64(value);
2329 default:
2330 return 0;
2331 }
2332
2333 return 0;
2334 }
2335
2336 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2337 const char *name)
2338 {
2339 struct format_field *field = perf_evsel__field(evsel, name);
2340
2341 if (!field)
2342 return 0;
2343
2344 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2345 }
2346
2347 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2348 char *msg, size_t msgsize)
2349 {
2350 int paranoid;
2351
2352 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2353 evsel->attr.type == PERF_TYPE_HARDWARE &&
2354 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2355 /*
2356 * If it's cycles then fall back to hrtimer based
2357 * cpu-clock-tick sw counter, which is always available even if
2358 * no PMU support.
2359 *
2360 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2361 * b0a873e).
2362 */
2363 scnprintf(msg, msgsize, "%s",
2364 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2365
2366 evsel->attr.type = PERF_TYPE_SOFTWARE;
2367 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2368
2369 zfree(&evsel->name);
2370 return true;
2371 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2372 (paranoid = perf_event_paranoid()) > 1) {
2373 const char *name = perf_evsel__name(evsel);
2374 char *new_name;
2375
2376 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2377 return false;
2378
2379 if (evsel->name)
2380 free(evsel->name);
2381 evsel->name = new_name;
2382 scnprintf(msg, msgsize,
2383 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2384 evsel->attr.exclude_kernel = 1;
2385
2386 return true;
2387 }
2388
2389 return false;
2390 }
2391
2392 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2393 int err, char *msg, size_t size)
2394 {
2395 char sbuf[STRERR_BUFSIZE];
2396
2397 switch (err) {
2398 case EPERM:
2399 case EACCES:
2400 return scnprintf(msg, size,
2401 "You may not have permission to collect %sstats.\n\n"
2402 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2403 "which controls use of the performance events system by\n"
2404 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2405 "The current value is %d:\n\n"
2406 " -1: Allow use of (almost) all events by all users\n"
2407 ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2408 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2409 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2410 target->system_wide ? "system-wide " : "",
2411 perf_event_paranoid());
2412 case ENOENT:
2413 return scnprintf(msg, size, "The %s event is not supported.",
2414 perf_evsel__name(evsel));
2415 case EMFILE:
2416 return scnprintf(msg, size, "%s",
2417 "Too many events are opened.\n"
2418 "Probably the maximum number of open file descriptors has been reached.\n"
2419 "Hint: Try again after reducing the number of events.\n"
2420 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2421 case ENOMEM:
2422 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2423 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2424 return scnprintf(msg, size,
2425 "Not enough memory to setup event with callchain.\n"
2426 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2427 "Hint: Current value: %d", sysctl_perf_event_max_stack);
2428 break;
2429 case ENODEV:
2430 if (target->cpu_list)
2431 return scnprintf(msg, size, "%s",
2432 "No such device - did you specify an out-of-range profile CPU?");
2433 break;
2434 case EOPNOTSUPP:
2435 if (evsel->attr.sample_period != 0)
2436 return scnprintf(msg, size, "%s",
2437 "PMU Hardware doesn't support sampling/overflow-interrupts.");
2438 if (evsel->attr.precise_ip)
2439 return scnprintf(msg, size, "%s",
2440 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2441 #if defined(__i386__) || defined(__x86_64__)
2442 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2443 return scnprintf(msg, size, "%s",
2444 "No hardware sampling interrupt available.\n"
2445 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2446 #endif
2447 break;
2448 case EBUSY:
2449 if (find_process("oprofiled"))
2450 return scnprintf(msg, size,
2451 "The PMU counters are busy/taken by another profiler.\n"
2452 "We found oprofile daemon running, please stop it and try again.");
2453 break;
2454 case EINVAL:
2455 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2456 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2457 if (perf_missing_features.clockid)
2458 return scnprintf(msg, size, "clockid feature not supported.");
2459 if (perf_missing_features.clockid_wrong)
2460 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2461 break;
2462 default:
2463 break;
2464 }
2465
2466 return scnprintf(msg, size,
2467 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2468 "/bin/dmesg may provide additional information.\n"
2469 "No CONFIG_PERF_EVENTS=y kernel support configured?",
2470 err, str_error_r(err, sbuf, sizeof(sbuf)),
2471 perf_evsel__name(evsel));
2472 }
2473
2474 char *perf_evsel__env_arch(struct perf_evsel *evsel)
2475 {
2476 if (evsel && evsel->evlist && evsel->evlist->env)
2477 return evsel->evlist->env->arch;
2478 return NULL;
2479 }
Linux with added FPC-III support