projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
perf tools: Fix one of the probe events to exclude kernel
[deliverable/linux.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 <linux/bitops.h>
12 #include <api/fs/debugfs.h>
13 #include <traceevent/event-parse.h>
14 #include <linux/hw_breakpoint.h>
15 #include <linux/perf_event.h>
16 #include <sys/resource.h>
17 #include "asm/bug.h"
18 #include "evsel.h"
19 #include "evlist.h"
20 #include "util.h"
21 #include "cpumap.h"
22 #include "thread_map.h"
23 #include "target.h"
24 #include "perf_regs.h"
25 #include "debug.h"
26 #include "trace-event.h"
27
28 static struct {
29 bool sample_id_all;
30 bool exclude_guest;
31 bool mmap2;
32 bool cloexec;
33 } perf_missing_features;
34
35 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
36
37 int __perf_evsel__sample_size(u64 sample_type)
38 {
39 u64 mask = sample_type & PERF_SAMPLE_MASK;
40 int size = 0;
41 int i;
42
43 for (i = 0; i < 64; i++) {
44 if (mask & (1ULL << i))
45 size++;
46 }
47
48 size *= sizeof(u64);
49
50 return size;
51 }
52
53 /**
54 * __perf_evsel__calc_id_pos - calculate id_pos.
55 * @sample_type: sample type
56 *
57 * This function returns the position of the event id (PERF_SAMPLE_ID or
58 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
59 * sample_event.
60 */
61 static int __perf_evsel__calc_id_pos(u64 sample_type)
62 {
63 int idx = 0;
64
65 if (sample_type & PERF_SAMPLE_IDENTIFIER)
66 return 0;
67
68 if (!(sample_type & PERF_SAMPLE_ID))
69 return -1;
70
71 if (sample_type & PERF_SAMPLE_IP)
72 idx += 1;
73
74 if (sample_type & PERF_SAMPLE_TID)
75 idx += 1;
76
77 if (sample_type & PERF_SAMPLE_TIME)
78 idx += 1;
79
80 if (sample_type & PERF_SAMPLE_ADDR)
81 idx += 1;
82
83 return idx;
84 }
85
86 /**
87 * __perf_evsel__calc_is_pos - calculate is_pos.
88 * @sample_type: sample type
89 *
90 * This function returns the position (counting backwards) of the event id
91 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
92 * sample_id_all is used there is an id sample appended to non-sample events.
93 */
94 static int __perf_evsel__calc_is_pos(u64 sample_type)
95 {
96 int idx = 1;
97
98 if (sample_type & PERF_SAMPLE_IDENTIFIER)
99 return 1;
100
101 if (!(sample_type & PERF_SAMPLE_ID))
102 return -1;
103
104 if (sample_type & PERF_SAMPLE_CPU)
105 idx += 1;
106
107 if (sample_type & PERF_SAMPLE_STREAM_ID)
108 idx += 1;
109
110 return idx;
111 }
112
113 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
114 {
115 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
116 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
117 }
118
119 void hists__init(struct hists *hists)
120 {
121 memset(hists, 0, sizeof(*hists));
122 hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
123 hists->entries_in = &hists->entries_in_array[0];
124 hists->entries_collapsed = RB_ROOT;
125 hists->entries = RB_ROOT;
126 pthread_mutex_init(&hists->lock, NULL);
127 }
128
129 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
130 enum perf_event_sample_format bit)
131 {
132 if (!(evsel->attr.sample_type & bit)) {
133 evsel->attr.sample_type |= bit;
134 evsel->sample_size += sizeof(u64);
135 perf_evsel__calc_id_pos(evsel);
136 }
137 }
138
139 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
140 enum perf_event_sample_format bit)
141 {
142 if (evsel->attr.sample_type & bit) {
143 evsel->attr.sample_type &= ~bit;
144 evsel->sample_size -= sizeof(u64);
145 perf_evsel__calc_id_pos(evsel);
146 }
147 }
148
149 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
150 bool can_sample_identifier)
151 {
152 if (can_sample_identifier) {
153 perf_evsel__reset_sample_bit(evsel, ID);
154 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
155 } else {
156 perf_evsel__set_sample_bit(evsel, ID);
157 }
158 evsel->attr.read_format |= PERF_FORMAT_ID;
159 }
160
161 void perf_evsel__init(struct perf_evsel *evsel,
162 struct perf_event_attr *attr, int idx)
163 {
164 evsel->idx = idx;
165 evsel->attr = *attr;
166 evsel->leader = evsel;
167 evsel->unit = "";
168 evsel->scale = 1.0;
169 INIT_LIST_HEAD(&evsel->node);
170 hists__init(&evsel->hists);
171 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
172 perf_evsel__calc_id_pos(evsel);
173 }
174
175 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
176 {
177 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
178
179 if (evsel != NULL)
180 perf_evsel__init(evsel, attr, idx);
181
182 return evsel;
183 }
184
185 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
186 {
187 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
188
189 if (evsel != NULL) {
190 struct perf_event_attr attr = {
191 .type = PERF_TYPE_TRACEPOINT,
192 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
193 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
194 };
195
196 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
197 goto out_free;
198
199 evsel->tp_format = trace_event__tp_format(sys, name);
200 if (evsel->tp_format == NULL)
201 goto out_free;
202
203 event_attr_init(&attr);
204 attr.config = evsel->tp_format->id;
205 attr.sample_period = 1;
206 perf_evsel__init(evsel, &attr, idx);
207 }
208
209 return evsel;
210
211 out_free:
212 zfree(&evsel->name);
213 free(evsel);
214 return NULL;
215 }
216
217 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
218 "cycles",
219 "instructions",
220 "cache-references",
221 "cache-misses",
222 "branches",
223 "branch-misses",
224 "bus-cycles",
225 "stalled-cycles-frontend",
226 "stalled-cycles-backend",
227 "ref-cycles",
228 };
229
230 static const char *__perf_evsel__hw_name(u64 config)
231 {
232 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
233 return perf_evsel__hw_names[config];
234
235 return "unknown-hardware";
236 }
237
238 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
239 {
240 int colon = 0, r = 0;
241 struct perf_event_attr *attr = &evsel->attr;
242 bool exclude_guest_default = false;
243
244 #define MOD_PRINT(context, mod) do { \
245 if (!attr->exclude_##context) { \
246 if (!colon) colon = ++r; \
247 r += scnprintf(bf + r, size - r, "%c", mod); \
248 } } while(0)
249
250 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
251 MOD_PRINT(kernel, 'k');
252 MOD_PRINT(user, 'u');
253 MOD_PRINT(hv, 'h');
254 exclude_guest_default = true;
255 }
256
257 if (attr->precise_ip) {
258 if (!colon)
259 colon = ++r;
260 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
261 exclude_guest_default = true;
262 }
263
264 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
265 MOD_PRINT(host, 'H');
266 MOD_PRINT(guest, 'G');
267 }
268 #undef MOD_PRINT
269 if (colon)
270 bf[colon - 1] = ':';
271 return r;
272 }
273
274 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
275 {
276 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
277 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
278 }
279
280 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
281 "cpu-clock",
282 "task-clock",
283 "page-faults",
284 "context-switches",
285 "cpu-migrations",
286 "minor-faults",
287 "major-faults",
288 "alignment-faults",
289 "emulation-faults",
290 "dummy",
291 };
292
293 static const char *__perf_evsel__sw_name(u64 config)
294 {
295 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
296 return perf_evsel__sw_names[config];
297 return "unknown-software";
298 }
299
300 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
301 {
302 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
303 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
304 }
305
306 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
307 {
308 int r;
309
310 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
311
312 if (type & HW_BREAKPOINT_R)
313 r += scnprintf(bf + r, size - r, "r");
314
315 if (type & HW_BREAKPOINT_W)
316 r += scnprintf(bf + r, size - r, "w");
317
318 if (type & HW_BREAKPOINT_X)
319 r += scnprintf(bf + r, size - r, "x");
320
321 return r;
322 }
323
324 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
325 {
326 struct perf_event_attr *attr = &evsel->attr;
327 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
328 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
329 }
330
331 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
332 [PERF_EVSEL__MAX_ALIASES] = {
333 { "L1-dcache", "l1-d", "l1d", "L1-data", },
334 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
335 { "LLC", "L2", },
336 { "dTLB", "d-tlb", "Data-TLB", },
337 { "iTLB", "i-tlb", "Instruction-TLB", },
338 { "branch", "branches", "bpu", "btb", "bpc", },
339 { "node", },
340 };
341
342 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
343 [PERF_EVSEL__MAX_ALIASES] = {
344 { "load", "loads", "read", },
345 { "store", "stores", "write", },
346 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
347 };
348
349 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
350 [PERF_EVSEL__MAX_ALIASES] = {
351 { "refs", "Reference", "ops", "access", },
352 { "misses", "miss", },
353 };
354
355 #define C(x) PERF_COUNT_HW_CACHE_##x
356 #define CACHE_READ (1 << C(OP_READ))
357 #define CACHE_WRITE (1 << C(OP_WRITE))
358 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
359 #define COP(x) (1 << x)
360
361 /*
362 * cache operartion stat
363 * L1I : Read and prefetch only
364 * ITLB and BPU : Read-only
365 */
366 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
367 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
368 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
369 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
370 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
371 [C(ITLB)] = (CACHE_READ),
372 [C(BPU)] = (CACHE_READ),
373 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
374 };
375
376 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
377 {
378 if (perf_evsel__hw_cache_stat[type] & COP(op))
379 return true; /* valid */
380 else
381 return false; /* invalid */
382 }
383
384 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
385 char *bf, size_t size)
386 {
387 if (result) {
388 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
389 perf_evsel__hw_cache_op[op][0],
390 perf_evsel__hw_cache_result[result][0]);
391 }
392
393 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
394 perf_evsel__hw_cache_op[op][1]);
395 }
396
397 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
398 {
399 u8 op, result, type = (config >> 0) & 0xff;
400 const char *err = "unknown-ext-hardware-cache-type";
401
402 if (type > PERF_COUNT_HW_CACHE_MAX)
403 goto out_err;
404
405 op = (config >> 8) & 0xff;
406 err = "unknown-ext-hardware-cache-op";
407 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
408 goto out_err;
409
410 result = (config >> 16) & 0xff;
411 err = "unknown-ext-hardware-cache-result";
412 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
413 goto out_err;
414
415 err = "invalid-cache";
416 if (!perf_evsel__is_cache_op_valid(type, op))
417 goto out_err;
418
419 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
420 out_err:
421 return scnprintf(bf, size, "%s", err);
422 }
423
424 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
425 {
426 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
427 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
428 }
429
430 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
431 {
432 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
433 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
434 }
435
436 const char *perf_evsel__name(struct perf_evsel *evsel)
437 {
438 char bf[128];
439
440 if (evsel->name)
441 return evsel->name;
442
443 switch (evsel->attr.type) {
444 case PERF_TYPE_RAW:
445 perf_evsel__raw_name(evsel, bf, sizeof(bf));
446 break;
447
448 case PERF_TYPE_HARDWARE:
449 perf_evsel__hw_name(evsel, bf, sizeof(bf));
450 break;
451
452 case PERF_TYPE_HW_CACHE:
453 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
454 break;
455
456 case PERF_TYPE_SOFTWARE:
457 perf_evsel__sw_name(evsel, bf, sizeof(bf));
458 break;
459
460 case PERF_TYPE_TRACEPOINT:
461 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
462 break;
463
464 case PERF_TYPE_BREAKPOINT:
465 perf_evsel__bp_name(evsel, bf, sizeof(bf));
466 break;
467
468 default:
469 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
470 evsel->attr.type);
471 break;
472 }
473
474 evsel->name = strdup(bf);
475
476 return evsel->name ?: "unknown";
477 }
478
479 const char *perf_evsel__group_name(struct perf_evsel *evsel)
480 {
481 return evsel->group_name ?: "anon group";
482 }
483
484 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
485 {
486 int ret;
487 struct perf_evsel *pos;
488 const char *group_name = perf_evsel__group_name(evsel);
489
490 ret = scnprintf(buf, size, "%s", group_name);
491
492 ret += scnprintf(buf + ret, size - ret, " { %s",
493 perf_evsel__name(evsel));
494
495 for_each_group_member(pos, evsel)
496 ret += scnprintf(buf + ret, size - ret, ", %s",
497 perf_evsel__name(pos));
498
499 ret += scnprintf(buf + ret, size - ret, " }");
500
501 return ret;
502 }
503
504 static void
505 perf_evsel__config_callgraph(struct perf_evsel *evsel,
506 struct record_opts *opts)
507 {
508 bool function = perf_evsel__is_function_event(evsel);
509 struct perf_event_attr *attr = &evsel->attr;
510
511 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
512
513 if (opts->call_graph == CALLCHAIN_DWARF) {
514 if (!function) {
515 perf_evsel__set_sample_bit(evsel, REGS_USER);
516 perf_evsel__set_sample_bit(evsel, STACK_USER);
517 attr->sample_regs_user = PERF_REGS_MASK;
518 attr->sample_stack_user = opts->stack_dump_size;
519 attr->exclude_callchain_user = 1;
520 } else {
521 pr_info("Cannot use DWARF unwind for function trace event,"
522 " falling back to framepointers.\n");
523 }
524 }
525
526 if (function) {
527 pr_info("Disabling user space callchains for function trace event.\n");
528 attr->exclude_callchain_user = 1;
529 }
530 }
531
532 /*
533 * The enable_on_exec/disabled value strategy:
534 *
535 * 1) For any type of traced program:
536 * - all independent events and group leaders are disabled
537 * - all group members are enabled
538 *
539 * Group members are ruled by group leaders. They need to
540 * be enabled, because the group scheduling relies on that.
541 *
542 * 2) For traced programs executed by perf:
543 * - all independent events and group leaders have
544 * enable_on_exec set
545 * - we don't specifically enable or disable any event during
546 * the record command
547 *
548 * Independent events and group leaders are initially disabled
549 * and get enabled by exec. Group members are ruled by group
550 * leaders as stated in 1).
551 *
552 * 3) For traced programs attached by perf (pid/tid):
553 * - we specifically enable or disable all events during
554 * the record command
555 *
556 * When attaching events to already running traced we
557 * enable/disable events specifically, as there's no
558 * initial traced exec call.
559 */
560 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
561 {
562 struct perf_evsel *leader = evsel->leader;
563 struct perf_event_attr *attr = &evsel->attr;
564 int track = !evsel->idx; /* only the first counter needs these */
565 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
566
567 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
568 attr->inherit = !opts->no_inherit;
569
570 perf_evsel__set_sample_bit(evsel, IP);
571 perf_evsel__set_sample_bit(evsel, TID);
572
573 if (evsel->sample_read) {
574 perf_evsel__set_sample_bit(evsel, READ);
575
576 /*
577 * We need ID even in case of single event, because
578 * PERF_SAMPLE_READ process ID specific data.
579 */
580 perf_evsel__set_sample_id(evsel, false);
581
582 /*
583 * Apply group format only if we belong to group
584 * with more than one members.
585 */
586 if (leader->nr_members > 1) {
587 attr->read_format |= PERF_FORMAT_GROUP;
588 attr->inherit = 0;
589 }
590 }
591
592 /*
593 * We default some events to have a default interval. But keep
594 * it a weak assumption overridable by the user.
595 */
596 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
597 opts->user_interval != ULLONG_MAX)) {
598 if (opts->freq) {
599 perf_evsel__set_sample_bit(evsel, PERIOD);
600 attr->freq = 1;
601 attr->sample_freq = opts->freq;
602 } else {
603 attr->sample_period = opts->default_interval;
604 }
605 }
606
607 /*
608 * Disable sampling for all group members other
609 * than leader in case leader 'leads' the sampling.
610 */
611 if ((leader != evsel) && leader->sample_read) {
612 attr->sample_freq = 0;
613 attr->sample_period = 0;
614 }
615
616 if (opts->no_samples)
617 attr->sample_freq = 0;
618
619 if (opts->inherit_stat)
620 attr->inherit_stat = 1;
621
622 if (opts->sample_address) {
623 perf_evsel__set_sample_bit(evsel, ADDR);
624 attr->mmap_data = track;
625 }
626
627 if (opts->call_graph_enabled && !evsel->no_aux_samples)
628 perf_evsel__config_callgraph(evsel, opts);
629
630 if (target__has_cpu(&opts->target))
631 perf_evsel__set_sample_bit(evsel, CPU);
632
633 if (opts->period)
634 perf_evsel__set_sample_bit(evsel, PERIOD);
635
636 /*
637 * When the user explicitely disabled time don't force it here.
638 */
639 if (opts->sample_time &&
640 (!perf_missing_features.sample_id_all &&
641 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu)))
642 perf_evsel__set_sample_bit(evsel, TIME);
643
644 if (opts->raw_samples && !evsel->no_aux_samples) {
645 perf_evsel__set_sample_bit(evsel, TIME);
646 perf_evsel__set_sample_bit(evsel, RAW);
647 perf_evsel__set_sample_bit(evsel, CPU);
648 }
649
650 if (opts->sample_address)
651 perf_evsel__set_sample_bit(evsel, DATA_SRC);
652
653 if (opts->no_buffering) {
654 attr->watermark = 0;
655 attr->wakeup_events = 1;
656 }
657 if (opts->branch_stack && !evsel->no_aux_samples) {
658 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
659 attr->branch_sample_type = opts->branch_stack;
660 }
661
662 if (opts->sample_weight)
663 perf_evsel__set_sample_bit(evsel, WEIGHT);
664
665 attr->mmap = track;
666 attr->mmap2 = track && !perf_missing_features.mmap2;
667 attr->comm = track;
668
669 if (opts->sample_transaction)
670 perf_evsel__set_sample_bit(evsel, TRANSACTION);
671
672 /*
673 * XXX see the function comment above
674 *
675 * Disabling only independent events or group leaders,
676 * keeping group members enabled.
677 */
678 if (perf_evsel__is_group_leader(evsel))
679 attr->disabled = 1;
680
681 /*
682 * Setting enable_on_exec for independent events and
683 * group leaders for traced executed by perf.
684 */
685 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
686 !opts->initial_delay)
687 attr->enable_on_exec = 1;
688
689 if (evsel->immediate) {
690 attr->disabled = 0;
691 attr->enable_on_exec = 0;
692 }
693 }
694
695 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
696 {
697 int cpu, thread;
698 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
699
700 if (evsel->fd) {
701 for (cpu = 0; cpu < ncpus; cpu++) {
702 for (thread = 0; thread < nthreads; thread++) {
703 FD(evsel, cpu, thread) = -1;
704 }
705 }
706 }
707
708 return evsel->fd != NULL ? 0 : -ENOMEM;
709 }
710
711 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
712 int ioc, void *arg)
713 {
714 int cpu, thread;
715
716 for (cpu = 0; cpu < ncpus; cpu++) {
717 for (thread = 0; thread < nthreads; thread++) {
718 int fd = FD(evsel, cpu, thread),
719 err = ioctl(fd, ioc, arg);
720
721 if (err)
722 return err;
723 }
724 }
725
726 return 0;
727 }
728
729 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
730 const char *filter)
731 {
732 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
733 PERF_EVENT_IOC_SET_FILTER,
734 (void *)filter);
735 }
736
737 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
738 {
739 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
740 PERF_EVENT_IOC_ENABLE,
741 0);
742 }
743
744 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
745 {
746 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
747 if (evsel->sample_id == NULL)
748 return -ENOMEM;
749
750 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
751 if (evsel->id == NULL) {
752 xyarray__delete(evsel->sample_id);
753 evsel->sample_id = NULL;
754 return -ENOMEM;
755 }
756
757 return 0;
758 }
759
760 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
761 {
762 memset(evsel->counts, 0, (sizeof(*evsel->counts) +
763 (ncpus * sizeof(struct perf_counts_values))));
764 }
765
766 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
767 {
768 evsel->counts = zalloc((sizeof(*evsel->counts) +
769 (ncpus * sizeof(struct perf_counts_values))));
770 return evsel->counts != NULL ? 0 : -ENOMEM;
771 }
772
773 void perf_evsel__free_fd(struct perf_evsel *evsel)
774 {
775 xyarray__delete(evsel->fd);
776 evsel->fd = NULL;
777 }
778
779 void perf_evsel__free_id(struct perf_evsel *evsel)
780 {
781 xyarray__delete(evsel->sample_id);
782 evsel->sample_id = NULL;
783 zfree(&evsel->id);
784 }
785
786 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
787 {
788 int cpu, thread;
789
790 for (cpu = 0; cpu < ncpus; cpu++)
791 for (thread = 0; thread < nthreads; ++thread) {
792 close(FD(evsel, cpu, thread));
793 FD(evsel, cpu, thread) = -1;
794 }
795 }
796
797 void perf_evsel__free_counts(struct perf_evsel *evsel)
798 {
799 zfree(&evsel->counts);
800 }
801
802 void perf_evsel__exit(struct perf_evsel *evsel)
803 {
804 assert(list_empty(&evsel->node));
805 perf_evsel__free_fd(evsel);
806 perf_evsel__free_id(evsel);
807 }
808
809 void perf_evsel__delete(struct perf_evsel *evsel)
810 {
811 perf_evsel__exit(evsel);
812 close_cgroup(evsel->cgrp);
813 zfree(&evsel->group_name);
814 if (evsel->tp_format)
815 pevent_free_format(evsel->tp_format);
816 zfree(&evsel->name);
817 free(evsel);
818 }
819
820 static inline void compute_deltas(struct perf_evsel *evsel,
821 int cpu,
822 struct perf_counts_values *count)
823 {
824 struct perf_counts_values tmp;
825
826 if (!evsel->prev_raw_counts)
827 return;
828
829 if (cpu == -1) {
830 tmp = evsel->prev_raw_counts->aggr;
831 evsel->prev_raw_counts->aggr = *count;
832 } else {
833 tmp = evsel->prev_raw_counts->cpu[cpu];
834 evsel->prev_raw_counts->cpu[cpu] = *count;
835 }
836
837 count->val = count->val - tmp.val;
838 count->ena = count->ena - tmp.ena;
839 count->run = count->run - tmp.run;
840 }
841
842 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
843 int cpu, int thread, bool scale)
844 {
845 struct perf_counts_values count;
846 size_t nv = scale ? 3 : 1;
847
848 if (FD(evsel, cpu, thread) < 0)
849 return -EINVAL;
850
851 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
852 return -ENOMEM;
853
854 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
855 return -errno;
856
857 compute_deltas(evsel, cpu, &count);
858
859 if (scale) {
860 if (count.run == 0)
861 count.val = 0;
862 else if (count.run < count.ena)
863 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
864 } else
865 count.ena = count.run = 0;
866
867 evsel->counts->cpu[cpu] = count;
868 return 0;
869 }
870
871 int __perf_evsel__read(struct perf_evsel *evsel,
872 int ncpus, int nthreads, bool scale)
873 {
874 size_t nv = scale ? 3 : 1;
875 int cpu, thread;
876 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
877
878 aggr->val = aggr->ena = aggr->run = 0;
879
880 for (cpu = 0; cpu < ncpus; cpu++) {
881 for (thread = 0; thread < nthreads; thread++) {
882 if (FD(evsel, cpu, thread) < 0)
883 continue;
884
885 if (readn(FD(evsel, cpu, thread),
886 &count, nv * sizeof(u64)) < 0)
887 return -errno;
888
889 aggr->val += count.val;
890 if (scale) {
891 aggr->ena += count.ena;
892 aggr->run += count.run;
893 }
894 }
895 }
896
897 compute_deltas(evsel, -1, aggr);
898
899 evsel->counts->scaled = 0;
900 if (scale) {
901 if (aggr->run == 0) {
902 evsel->counts->scaled = -1;
903 aggr->val = 0;
904 return 0;
905 }
906
907 if (aggr->run < aggr->ena) {
908 evsel->counts->scaled = 1;
909 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
910 }
911 } else
912 aggr->ena = aggr->run = 0;
913
914 return 0;
915 }
916
917 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
918 {
919 struct perf_evsel *leader = evsel->leader;
920 int fd;
921
922 if (perf_evsel__is_group_leader(evsel))
923 return -1;
924
925 /*
926 * Leader must be already processed/open,
927 * if not it's a bug.
928 */
929 BUG_ON(!leader->fd);
930
931 fd = FD(leader, cpu, thread);
932 BUG_ON(fd == -1);
933
934 return fd;
935 }
936
937 #define __PRINT_ATTR(fmt, cast, field) \
938 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
939
940 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
941 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
942 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
943 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
944
945 #define PRINT_ATTR2N(name1, field1, name2, field2) \
946 fprintf(fp, " %-19s %u %-19s %u\n", \
947 name1, attr->field1, name2, attr->field2)
948
949 #define PRINT_ATTR2(field1, field2) \
950 PRINT_ATTR2N(#field1, field1, #field2, field2)
951
952 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
953 {
954 size_t ret = 0;
955
956 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
957 ret += fprintf(fp, "perf_event_attr:\n");
958
959 ret += PRINT_ATTR_U32(type);
960 ret += PRINT_ATTR_U32(size);
961 ret += PRINT_ATTR_X64(config);
962 ret += PRINT_ATTR_U64(sample_period);
963 ret += PRINT_ATTR_U64(sample_freq);
964 ret += PRINT_ATTR_X64(sample_type);
965 ret += PRINT_ATTR_X64(read_format);
966
967 ret += PRINT_ATTR2(disabled, inherit);
968 ret += PRINT_ATTR2(pinned, exclusive);
969 ret += PRINT_ATTR2(exclude_user, exclude_kernel);
970 ret += PRINT_ATTR2(exclude_hv, exclude_idle);
971 ret += PRINT_ATTR2(mmap, comm);
972 ret += PRINT_ATTR2(mmap2, comm_exec);
973 ret += PRINT_ATTR2(freq, inherit_stat);
974 ret += PRINT_ATTR2(enable_on_exec, task);
975 ret += PRINT_ATTR2(watermark, precise_ip);
976 ret += PRINT_ATTR2(mmap_data, sample_id_all);
977 ret += PRINT_ATTR2(exclude_host, exclude_guest);
978 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
979 "excl.callchain_user", exclude_callchain_user);
980
981 ret += PRINT_ATTR_U32(wakeup_events);
982 ret += PRINT_ATTR_U32(wakeup_watermark);
983 ret += PRINT_ATTR_X32(bp_type);
984 ret += PRINT_ATTR_X64(bp_addr);
985 ret += PRINT_ATTR_X64(config1);
986 ret += PRINT_ATTR_U64(bp_len);
987 ret += PRINT_ATTR_X64(config2);
988 ret += PRINT_ATTR_X64(branch_sample_type);
989 ret += PRINT_ATTR_X64(sample_regs_user);
990 ret += PRINT_ATTR_U32(sample_stack_user);
991
992 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
993
994 return ret;
995 }
996
997 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
998 struct thread_map *threads)
999 {
1000 int cpu, thread;
1001 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1002 int pid = -1, err;
1003 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1004
1005 if (evsel->fd == NULL &&
1006 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
1007 return -ENOMEM;
1008
1009 if (evsel->cgrp) {
1010 flags |= PERF_FLAG_PID_CGROUP;
1011 pid = evsel->cgrp->fd;
1012 }
1013
1014 fallback_missing_features:
1015 if (perf_missing_features.cloexec)
1016 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1017 if (perf_missing_features.mmap2)
1018 evsel->attr.mmap2 = 0;
1019 if (perf_missing_features.exclude_guest)
1020 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1021 retry_sample_id:
1022 if (perf_missing_features.sample_id_all)
1023 evsel->attr.sample_id_all = 0;
1024
1025 if (verbose >= 2)
1026 perf_event_attr__fprintf(&evsel->attr, stderr);
1027
1028 for (cpu = 0; cpu < cpus->nr; cpu++) {
1029
1030 for (thread = 0; thread < threads->nr; thread++) {
1031 int group_fd;
1032
1033 if (!evsel->cgrp)
1034 pid = threads->map[thread];
1035
1036 group_fd = get_group_fd(evsel, cpu, thread);
1037 retry_open:
1038 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1039 pid, cpus->map[cpu], group_fd, flags);
1040
1041 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1042 pid,
1043 cpus->map[cpu],
1044 group_fd, flags);
1045 if (FD(evsel, cpu, thread) < 0) {
1046 err = -errno;
1047 pr_debug2("sys_perf_event_open failed, error %d\n",
1048 err);
1049 goto try_fallback;
1050 }
1051 set_rlimit = NO_CHANGE;
1052 }
1053 }
1054
1055 return 0;
1056
1057 try_fallback:
1058 /*
1059 * perf stat needs between 5 and 22 fds per CPU. When we run out
1060 * of them try to increase the limits.
1061 */
1062 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1063 struct rlimit l;
1064 int old_errno = errno;
1065
1066 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1067 if (set_rlimit == NO_CHANGE)
1068 l.rlim_cur = l.rlim_max;
1069 else {
1070 l.rlim_cur = l.rlim_max + 1000;
1071 l.rlim_max = l.rlim_cur;
1072 }
1073 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1074 set_rlimit++;
1075 errno = old_errno;
1076 goto retry_open;
1077 }
1078 }
1079 errno = old_errno;
1080 }
1081
1082 if (err != -EINVAL || cpu > 0 || thread > 0)
1083 goto out_close;
1084
1085 if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1086 perf_missing_features.cloexec = true;
1087 goto fallback_missing_features;
1088 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1089 perf_missing_features.mmap2 = true;
1090 goto fallback_missing_features;
1091 } else if (!perf_missing_features.exclude_guest &&
1092 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1093 perf_missing_features.exclude_guest = true;
1094 goto fallback_missing_features;
1095 } else if (!perf_missing_features.sample_id_all) {
1096 perf_missing_features.sample_id_all = true;
1097 goto retry_sample_id;
1098 }
1099
1100 out_close:
1101 do {
1102 while (--thread >= 0) {
1103 close(FD(evsel, cpu, thread));
1104 FD(evsel, cpu, thread) = -1;
1105 }
1106 thread = threads->nr;
1107 } while (--cpu >= 0);
1108 return err;
1109 }
1110
1111 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1112 {
1113 if (evsel->fd == NULL)
1114 return;
1115
1116 perf_evsel__close_fd(evsel, ncpus, nthreads);
1117 perf_evsel__free_fd(evsel);
1118 }
1119
1120 static struct {
1121 struct cpu_map map;
1122 int cpus[1];
1123 } empty_cpu_map = {
1124 .map.nr = 1,
1125 .cpus = { -1, },
1126 };
1127
1128 static struct {
1129 struct thread_map map;
1130 int threads[1];
1131 } empty_thread_map = {
1132 .map.nr = 1,
1133 .threads = { -1, },
1134 };
1135
1136 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1137 struct thread_map *threads)
1138 {
1139 if (cpus == NULL) {
1140 /* Work around old compiler warnings about strict aliasing */
1141 cpus = &empty_cpu_map.map;
1142 }
1143
1144 if (threads == NULL)
1145 threads = &empty_thread_map.map;
1146
1147 return __perf_evsel__open(evsel, cpus, threads);
1148 }
1149
1150 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1151 struct cpu_map *cpus)
1152 {
1153 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1154 }
1155
1156 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1157 struct thread_map *threads)
1158 {
1159 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1160 }
1161
1162 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1163 const union perf_event *event,
1164 struct perf_sample *sample)
1165 {
1166 u64 type = evsel->attr.sample_type;
1167 const u64 *array = event->sample.array;
1168 bool swapped = evsel->needs_swap;
1169 union u64_swap u;
1170
1171 array += ((event->header.size -
1172 sizeof(event->header)) / sizeof(u64)) - 1;
1173
1174 if (type & PERF_SAMPLE_IDENTIFIER) {
1175 sample->id = *array;
1176 array--;
1177 }
1178
1179 if (type & PERF_SAMPLE_CPU) {
1180 u.val64 = *array;
1181 if (swapped) {
1182 /* undo swap of u64, then swap on individual u32s */
1183 u.val64 = bswap_64(u.val64);
1184 u.val32[0] = bswap_32(u.val32[0]);
1185 }
1186
1187 sample->cpu = u.val32[0];
1188 array--;
1189 }
1190
1191 if (type & PERF_SAMPLE_STREAM_ID) {
1192 sample->stream_id = *array;
1193 array--;
1194 }
1195
1196 if (type & PERF_SAMPLE_ID) {
1197 sample->id = *array;
1198 array--;
1199 }
1200
1201 if (type & PERF_SAMPLE_TIME) {
1202 sample->time = *array;
1203 array--;
1204 }
1205
1206 if (type & PERF_SAMPLE_TID) {
1207 u.val64 = *array;
1208 if (swapped) {
1209 /* undo swap of u64, then swap on individual u32s */
1210 u.val64 = bswap_64(u.val64);
1211 u.val32[0] = bswap_32(u.val32[0]);
1212 u.val32[1] = bswap_32(u.val32[1]);
1213 }
1214
1215 sample->pid = u.val32[0];
1216 sample->tid = u.val32[1];
1217 array--;
1218 }
1219
1220 return 0;
1221 }
1222
1223 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1224 u64 size)
1225 {
1226 return size > max_size || offset + size > endp;
1227 }
1228
1229 #define OVERFLOW_CHECK(offset, size, max_size) \
1230 do { \
1231 if (overflow(endp, (max_size), (offset), (size))) \
1232 return -EFAULT; \
1233 } while (0)
1234
1235 #define OVERFLOW_CHECK_u64(offset) \
1236 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1237
1238 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1239 struct perf_sample *data)
1240 {
1241 u64 type = evsel->attr.sample_type;
1242 bool swapped = evsel->needs_swap;
1243 const u64 *array;
1244 u16 max_size = event->header.size;
1245 const void *endp = (void *)event + max_size;
1246 u64 sz;
1247
1248 /*
1249 * used for cross-endian analysis. See git commit 65014ab3
1250 * for why this goofiness is needed.
1251 */
1252 union u64_swap u;
1253
1254 memset(data, 0, sizeof(*data));
1255 data->cpu = data->pid = data->tid = -1;
1256 data->stream_id = data->id = data->time = -1ULL;
1257 data->period = evsel->attr.sample_period;
1258 data->weight = 0;
1259
1260 if (event->header.type != PERF_RECORD_SAMPLE) {
1261 if (!evsel->attr.sample_id_all)
1262 return 0;
1263 return perf_evsel__parse_id_sample(evsel, event, data);
1264 }
1265
1266 array = event->sample.array;
1267
1268 /*
1269 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1270 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1271 * check the format does not go past the end of the event.
1272 */
1273 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1274 return -EFAULT;
1275
1276 data->id = -1ULL;
1277 if (type & PERF_SAMPLE_IDENTIFIER) {
1278 data->id = *array;
1279 array++;
1280 }
1281
1282 if (type & PERF_SAMPLE_IP) {
1283 data->ip = *array;
1284 array++;
1285 }
1286
1287 if (type & PERF_SAMPLE_TID) {
1288 u.val64 = *array;
1289 if (swapped) {
1290 /* undo swap of u64, then swap on individual u32s */
1291 u.val64 = bswap_64(u.val64);
1292 u.val32[0] = bswap_32(u.val32[0]);
1293 u.val32[1] = bswap_32(u.val32[1]);
1294 }
1295
1296 data->pid = u.val32[0];
1297 data->tid = u.val32[1];
1298 array++;
1299 }
1300
1301 if (type & PERF_SAMPLE_TIME) {
1302 data->time = *array;
1303 array++;
1304 }
1305
1306 data->addr = 0;
1307 if (type & PERF_SAMPLE_ADDR) {
1308 data->addr = *array;
1309 array++;
1310 }
1311
1312 if (type & PERF_SAMPLE_ID) {
1313 data->id = *array;
1314 array++;
1315 }
1316
1317 if (type & PERF_SAMPLE_STREAM_ID) {
1318 data->stream_id = *array;
1319 array++;
1320 }
1321
1322 if (type & PERF_SAMPLE_CPU) {
1323
1324 u.val64 = *array;
1325 if (swapped) {
1326 /* undo swap of u64, then swap on individual u32s */
1327 u.val64 = bswap_64(u.val64);
1328 u.val32[0] = bswap_32(u.val32[0]);
1329 }
1330
1331 data->cpu = u.val32[0];
1332 array++;
1333 }
1334
1335 if (type & PERF_SAMPLE_PERIOD) {
1336 data->period = *array;
1337 array++;
1338 }
1339
1340 if (type & PERF_SAMPLE_READ) {
1341 u64 read_format = evsel->attr.read_format;
1342
1343 OVERFLOW_CHECK_u64(array);
1344 if (read_format & PERF_FORMAT_GROUP)
1345 data->read.group.nr = *array;
1346 else
1347 data->read.one.value = *array;
1348
1349 array++;
1350
1351 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1352 OVERFLOW_CHECK_u64(array);
1353 data->read.time_enabled = *array;
1354 array++;
1355 }
1356
1357 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1358 OVERFLOW_CHECK_u64(array);
1359 data->read.time_running = *array;
1360 array++;
1361 }
1362
1363 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1364 if (read_format & PERF_FORMAT_GROUP) {
1365 const u64 max_group_nr = UINT64_MAX /
1366 sizeof(struct sample_read_value);
1367
1368 if (data->read.group.nr > max_group_nr)
1369 return -EFAULT;
1370 sz = data->read.group.nr *
1371 sizeof(struct sample_read_value);
1372 OVERFLOW_CHECK(array, sz, max_size);
1373 data->read.group.values =
1374 (struct sample_read_value *)array;
1375 array = (void *)array + sz;
1376 } else {
1377 OVERFLOW_CHECK_u64(array);
1378 data->read.one.id = *array;
1379 array++;
1380 }
1381 }
1382
1383 if (type & PERF_SAMPLE_CALLCHAIN) {
1384 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1385
1386 OVERFLOW_CHECK_u64(array);
1387 data->callchain = (struct ip_callchain *)array++;
1388 if (data->callchain->nr > max_callchain_nr)
1389 return -EFAULT;
1390 sz = data->callchain->nr * sizeof(u64);
1391 OVERFLOW_CHECK(array, sz, max_size);
1392 array = (void *)array + sz;
1393 }
1394
1395 if (type & PERF_SAMPLE_RAW) {
1396 OVERFLOW_CHECK_u64(array);
1397 u.val64 = *array;
1398 if (WARN_ONCE(swapped,
1399 "Endianness of raw data not corrected!\n")) {
1400 /* undo swap of u64, then swap on individual u32s */
1401 u.val64 = bswap_64(u.val64);
1402 u.val32[0] = bswap_32(u.val32[0]);
1403 u.val32[1] = bswap_32(u.val32[1]);
1404 }
1405 data->raw_size = u.val32[0];
1406 array = (void *)array + sizeof(u32);
1407
1408 OVERFLOW_CHECK(array, data->raw_size, max_size);
1409 data->raw_data = (void *)array;
1410 array = (void *)array + data->raw_size;
1411 }
1412
1413 if (type & PERF_SAMPLE_BRANCH_STACK) {
1414 const u64 max_branch_nr = UINT64_MAX /
1415 sizeof(struct branch_entry);
1416
1417 OVERFLOW_CHECK_u64(array);
1418 data->branch_stack = (struct branch_stack *)array++;
1419
1420 if (data->branch_stack->nr > max_branch_nr)
1421 return -EFAULT;
1422 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1423 OVERFLOW_CHECK(array, sz, max_size);
1424 array = (void *)array + sz;
1425 }
1426
1427 if (type & PERF_SAMPLE_REGS_USER) {
1428 OVERFLOW_CHECK_u64(array);
1429 data->user_regs.abi = *array;
1430 array++;
1431
1432 if (data->user_regs.abi) {
1433 u64 mask = evsel->attr.sample_regs_user;
1434
1435 sz = hweight_long(mask) * sizeof(u64);
1436 OVERFLOW_CHECK(array, sz, max_size);
1437 data->user_regs.mask = mask;
1438 data->user_regs.regs = (u64 *)array;
1439 array = (void *)array + sz;
1440 }
1441 }
1442
1443 if (type & PERF_SAMPLE_STACK_USER) {
1444 OVERFLOW_CHECK_u64(array);
1445 sz = *array++;
1446
1447 data->user_stack.offset = ((char *)(array - 1)
1448 - (char *) event);
1449
1450 if (!sz) {
1451 data->user_stack.size = 0;
1452 } else {
1453 OVERFLOW_CHECK(array, sz, max_size);
1454 data->user_stack.data = (char *)array;
1455 array = (void *)array + sz;
1456 OVERFLOW_CHECK_u64(array);
1457 data->user_stack.size = *array++;
1458 if (WARN_ONCE(data->user_stack.size > sz,
1459 "user stack dump failure\n"))
1460 return -EFAULT;
1461 }
1462 }
1463
1464 data->weight = 0;
1465 if (type & PERF_SAMPLE_WEIGHT) {
1466 OVERFLOW_CHECK_u64(array);
1467 data->weight = *array;
1468 array++;
1469 }
1470
1471 data->data_src = PERF_MEM_DATA_SRC_NONE;
1472 if (type & PERF_SAMPLE_DATA_SRC) {
1473 OVERFLOW_CHECK_u64(array);
1474 data->data_src = *array;
1475 array++;
1476 }
1477
1478 data->transaction = 0;
1479 if (type & PERF_SAMPLE_TRANSACTION) {
1480 OVERFLOW_CHECK_u64(array);
1481 data->transaction = *array;
1482 array++;
1483 }
1484
1485 return 0;
1486 }
1487
1488 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1489 u64 read_format)
1490 {
1491 size_t sz, result = sizeof(struct sample_event);
1492
1493 if (type & PERF_SAMPLE_IDENTIFIER)
1494 result += sizeof(u64);
1495
1496 if (type & PERF_SAMPLE_IP)
1497 result += sizeof(u64);
1498
1499 if (type & PERF_SAMPLE_TID)
1500 result += sizeof(u64);
1501
1502 if (type & PERF_SAMPLE_TIME)
1503 result += sizeof(u64);
1504
1505 if (type & PERF_SAMPLE_ADDR)
1506 result += sizeof(u64);
1507
1508 if (type & PERF_SAMPLE_ID)
1509 result += sizeof(u64);
1510
1511 if (type & PERF_SAMPLE_STREAM_ID)
1512 result += sizeof(u64);
1513
1514 if (type & PERF_SAMPLE_CPU)
1515 result += sizeof(u64);
1516
1517 if (type & PERF_SAMPLE_PERIOD)
1518 result += sizeof(u64);
1519
1520 if (type & PERF_SAMPLE_READ) {
1521 result += sizeof(u64);
1522 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1523 result += sizeof(u64);
1524 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1525 result += sizeof(u64);
1526 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1527 if (read_format & PERF_FORMAT_GROUP) {
1528 sz = sample->read.group.nr *
1529 sizeof(struct sample_read_value);
1530 result += sz;
1531 } else {
1532 result += sizeof(u64);
1533 }
1534 }
1535
1536 if (type & PERF_SAMPLE_CALLCHAIN) {
1537 sz = (sample->callchain->nr + 1) * sizeof(u64);
1538 result += sz;
1539 }
1540
1541 if (type & PERF_SAMPLE_RAW) {
1542 result += sizeof(u32);
1543 result += sample->raw_size;
1544 }
1545
1546 if (type & PERF_SAMPLE_BRANCH_STACK) {
1547 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1548 sz += sizeof(u64);
1549 result += sz;
1550 }
1551
1552 if (type & PERF_SAMPLE_REGS_USER) {
1553 if (sample->user_regs.abi) {
1554 result += sizeof(u64);
1555 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1556 result += sz;
1557 } else {
1558 result += sizeof(u64);
1559 }
1560 }
1561
1562 if (type & PERF_SAMPLE_STACK_USER) {
1563 sz = sample->user_stack.size;
1564 result += sizeof(u64);
1565 if (sz) {
1566 result += sz;
1567 result += sizeof(u64);
1568 }
1569 }
1570
1571 if (type & PERF_SAMPLE_WEIGHT)
1572 result += sizeof(u64);
1573
1574 if (type & PERF_SAMPLE_DATA_SRC)
1575 result += sizeof(u64);
1576
1577 if (type & PERF_SAMPLE_TRANSACTION)
1578 result += sizeof(u64);
1579
1580 return result;
1581 }
1582
1583 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1584 u64 read_format,
1585 const struct perf_sample *sample,
1586 bool swapped)
1587 {
1588 u64 *array;
1589 size_t sz;
1590 /*
1591 * used for cross-endian analysis. See git commit 65014ab3
1592 * for why this goofiness is needed.
1593 */
1594 union u64_swap u;
1595
1596 array = event->sample.array;
1597
1598 if (type & PERF_SAMPLE_IDENTIFIER) {
1599 *array = sample->id;
1600 array++;
1601 }
1602
1603 if (type & PERF_SAMPLE_IP) {
1604 *array = sample->ip;
1605 array++;
1606 }
1607
1608 if (type & PERF_SAMPLE_TID) {
1609 u.val32[0] = sample->pid;
1610 u.val32[1] = sample->tid;
1611 if (swapped) {
1612 /*
1613 * Inverse of what is done in perf_evsel__parse_sample
1614 */
1615 u.val32[0] = bswap_32(u.val32[0]);
1616 u.val32[1] = bswap_32(u.val32[1]);
1617 u.val64 = bswap_64(u.val64);
1618 }
1619
1620 *array = u.val64;
1621 array++;
1622 }
1623
1624 if (type & PERF_SAMPLE_TIME) {
1625 *array = sample->time;
1626 array++;
1627 }
1628
1629 if (type & PERF_SAMPLE_ADDR) {
1630 *array = sample->addr;
1631 array++;
1632 }
1633
1634 if (type & PERF_SAMPLE_ID) {
1635 *array = sample->id;
1636 array++;
1637 }
1638
1639 if (type & PERF_SAMPLE_STREAM_ID) {
1640 *array = sample->stream_id;
1641 array++;
1642 }
1643
1644 if (type & PERF_SAMPLE_CPU) {
1645 u.val32[0] = sample->cpu;
1646 if (swapped) {
1647 /*
1648 * Inverse of what is done in perf_evsel__parse_sample
1649 */
1650 u.val32[0] = bswap_32(u.val32[0]);
1651 u.val64 = bswap_64(u.val64);
1652 }
1653 *array = u.val64;
1654 array++;
1655 }
1656
1657 if (type & PERF_SAMPLE_PERIOD) {
1658 *array = sample->period;
1659 array++;
1660 }
1661
1662 if (type & PERF_SAMPLE_READ) {
1663 if (read_format & PERF_FORMAT_GROUP)
1664 *array = sample->read.group.nr;
1665 else
1666 *array = sample->read.one.value;
1667 array++;
1668
1669 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1670 *array = sample->read.time_enabled;
1671 array++;
1672 }
1673
1674 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1675 *array = sample->read.time_running;
1676 array++;
1677 }
1678
1679 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1680 if (read_format & PERF_FORMAT_GROUP) {
1681 sz = sample->read.group.nr *
1682 sizeof(struct sample_read_value);
1683 memcpy(array, sample->read.group.values, sz);
1684 array = (void *)array + sz;
1685 } else {
1686 *array = sample->read.one.id;
1687 array++;
1688 }
1689 }
1690
1691 if (type & PERF_SAMPLE_CALLCHAIN) {
1692 sz = (sample->callchain->nr + 1) * sizeof(u64);
1693 memcpy(array, sample->callchain, sz);
1694 array = (void *)array + sz;
1695 }
1696
1697 if (type & PERF_SAMPLE_RAW) {
1698 u.val32[0] = sample->raw_size;
1699 if (WARN_ONCE(swapped,
1700 "Endianness of raw data not corrected!\n")) {
1701 /*
1702 * Inverse of what is done in perf_evsel__parse_sample
1703 */
1704 u.val32[0] = bswap_32(u.val32[0]);
1705 u.val32[1] = bswap_32(u.val32[1]);
1706 u.val64 = bswap_64(u.val64);
1707 }
1708 *array = u.val64;
1709 array = (void *)array + sizeof(u32);
1710
1711 memcpy(array, sample->raw_data, sample->raw_size);
1712 array = (void *)array + sample->raw_size;
1713 }
1714
1715 if (type & PERF_SAMPLE_BRANCH_STACK) {
1716 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1717 sz += sizeof(u64);
1718 memcpy(array, sample->branch_stack, sz);
1719 array = (void *)array + sz;
1720 }
1721
1722 if (type & PERF_SAMPLE_REGS_USER) {
1723 if (sample->user_regs.abi) {
1724 *array++ = sample->user_regs.abi;
1725 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1726 memcpy(array, sample->user_regs.regs, sz);
1727 array = (void *)array + sz;
1728 } else {
1729 *array++ = 0;
1730 }
1731 }
1732
1733 if (type & PERF_SAMPLE_STACK_USER) {
1734 sz = sample->user_stack.size;
1735 *array++ = sz;
1736 if (sz) {
1737 memcpy(array, sample->user_stack.data, sz);
1738 array = (void *)array + sz;
1739 *array++ = sz;
1740 }
1741 }
1742
1743 if (type & PERF_SAMPLE_WEIGHT) {
1744 *array = sample->weight;
1745 array++;
1746 }
1747
1748 if (type & PERF_SAMPLE_DATA_SRC) {
1749 *array = sample->data_src;
1750 array++;
1751 }
1752
1753 if (type & PERF_SAMPLE_TRANSACTION) {
1754 *array = sample->transaction;
1755 array++;
1756 }
1757
1758 return 0;
1759 }
1760
1761 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1762 {
1763 return pevent_find_field(evsel->tp_format, name);
1764 }
1765
1766 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1767 const char *name)
1768 {
1769 struct format_field *field = perf_evsel__field(evsel, name);
1770 int offset;
1771
1772 if (!field)
1773 return NULL;
1774
1775 offset = field->offset;
1776
1777 if (field->flags & FIELD_IS_DYNAMIC) {
1778 offset = *(int *)(sample->raw_data + field->offset);
1779 offset &= 0xffff;
1780 }
1781
1782 return sample->raw_data + offset;
1783 }
1784
1785 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1786 const char *name)
1787 {
1788 struct format_field *field = perf_evsel__field(evsel, name);
1789 void *ptr;
1790 u64 value;
1791
1792 if (!field)
1793 return 0;
1794
1795 ptr = sample->raw_data + field->offset;
1796
1797 switch (field->size) {
1798 case 1:
1799 return *(u8 *)ptr;
1800 case 2:
1801 value = *(u16 *)ptr;
1802 break;
1803 case 4:
1804 value = *(u32 *)ptr;
1805 break;
1806 case 8:
1807 value = *(u64 *)ptr;
1808 break;
1809 default:
1810 return 0;
1811 }
1812
1813 if (!evsel->needs_swap)
1814 return value;
1815
1816 switch (field->size) {
1817 case 2:
1818 return bswap_16(value);
1819 case 4:
1820 return bswap_32(value);
1821 case 8:
1822 return bswap_64(value);
1823 default:
1824 return 0;
1825 }
1826
1827 return 0;
1828 }
1829
1830 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1831 {
1832 va_list args;
1833 int ret = 0;
1834
1835 if (!*first) {
1836 ret += fprintf(fp, ",");
1837 } else {
1838 ret += fprintf(fp, ":");
1839 *first = false;
1840 }
1841
1842 va_start(args, fmt);
1843 ret += vfprintf(fp, fmt, args);
1844 va_end(args);
1845 return ret;
1846 }
1847
1848 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1849 {
1850 if (value == 0)
1851 return 0;
1852
1853 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1854 }
1855
1856 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1857
1858 struct bit_names {
1859 int bit;
1860 const char *name;
1861 };
1862
1863 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1864 struct bit_names *bits, bool *first)
1865 {
1866 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1867 bool first_bit = true;
1868
1869 do {
1870 if (value & bits[i].bit) {
1871 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1872 first_bit = false;
1873 }
1874 } while (bits[++i].name != NULL);
1875
1876 return printed;
1877 }
1878
1879 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1880 {
1881 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1882 struct bit_names bits[] = {
1883 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1884 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1885 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1886 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1887 bit_name(IDENTIFIER),
1888 { .name = NULL, }
1889 };
1890 #undef bit_name
1891 return bits__fprintf(fp, "sample_type", value, bits, first);
1892 }
1893
1894 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1895 {
1896 #define bit_name(n) { PERF_FORMAT_##n, #n }
1897 struct bit_names bits[] = {
1898 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1899 bit_name(ID), bit_name(GROUP),
1900 { .name = NULL, }
1901 };
1902 #undef bit_name
1903 return bits__fprintf(fp, "read_format", value, bits, first);
1904 }
1905
1906 int perf_evsel__fprintf(struct perf_evsel *evsel,
1907 struct perf_attr_details *details, FILE *fp)
1908 {
1909 bool first = true;
1910 int printed = 0;
1911
1912 if (details->event_group) {
1913 struct perf_evsel *pos;
1914
1915 if (!perf_evsel__is_group_leader(evsel))
1916 return 0;
1917
1918 if (evsel->nr_members > 1)
1919 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1920
1921 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1922 for_each_group_member(pos, evsel)
1923 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1924
1925 if (evsel->nr_members > 1)
1926 printed += fprintf(fp, "}");
1927 goto out;
1928 }
1929
1930 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1931
1932 if (details->verbose || details->freq) {
1933 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1934 (u64)evsel->attr.sample_freq);
1935 }
1936
1937 if (details->verbose) {
1938 if_print(type);
1939 if_print(config);
1940 if_print(config1);
1941 if_print(config2);
1942 if_print(size);
1943 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1944 if (evsel->attr.read_format)
1945 printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1946 if_print(disabled);
1947 if_print(inherit);
1948 if_print(pinned);
1949 if_print(exclusive);
1950 if_print(exclude_user);
1951 if_print(exclude_kernel);
1952 if_print(exclude_hv);
1953 if_print(exclude_idle);
1954 if_print(mmap);
1955 if_print(mmap2);
1956 if_print(comm);
1957 if_print(comm_exec);
1958 if_print(freq);
1959 if_print(inherit_stat);
1960 if_print(enable_on_exec);
1961 if_print(task);
1962 if_print(watermark);
1963 if_print(precise_ip);
1964 if_print(mmap_data);
1965 if_print(sample_id_all);
1966 if_print(exclude_host);
1967 if_print(exclude_guest);
1968 if_print(__reserved_1);
1969 if_print(wakeup_events);
1970 if_print(bp_type);
1971 if_print(branch_sample_type);
1972 }
1973 out:
1974 fputc('\n', fp);
1975 return ++printed;
1976 }
1977
1978 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1979 char *msg, size_t msgsize)
1980 {
1981 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
1982 evsel->attr.type == PERF_TYPE_HARDWARE &&
1983 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1984 /*
1985 * If it's cycles then fall back to hrtimer based
1986 * cpu-clock-tick sw counter, which is always available even if
1987 * no PMU support.
1988 *
1989 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1990 * b0a873e).
1991 */
1992 scnprintf(msg, msgsize, "%s",
1993 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1994
1995 evsel->attr.type = PERF_TYPE_SOFTWARE;
1996 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
1997
1998 zfree(&evsel->name);
1999 return true;
2000 }
2001
2002 return false;
2003 }
2004
2005 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2006 int err, char *msg, size_t size)
2007 {
2008 switch (err) {
2009 case EPERM:
2010 case EACCES:
2011 return scnprintf(msg, size,
2012 "You may not have permission to collect %sstats.\n"
2013 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2014 " -1 - Not paranoid at all\n"
2015 " 0 - Disallow raw tracepoint access for unpriv\n"
2016 " 1 - Disallow cpu events for unpriv\n"
2017 " 2 - Disallow kernel profiling for unpriv",
2018 target->system_wide ? "system-wide " : "");
2019 case ENOENT:
2020 return scnprintf(msg, size, "The %s event is not supported.",
2021 perf_evsel__name(evsel));
2022 case EMFILE:
2023 return scnprintf(msg, size, "%s",
2024 "Too many events are opened.\n"
2025 "Try again after reducing the number of events.");
2026 case ENODEV:
2027 if (target->cpu_list)
2028 return scnprintf(msg, size, "%s",
2029 "No such device - did you specify an out-of-range profile CPU?\n");
2030 break;
2031 case EOPNOTSUPP:
2032 if (evsel->attr.precise_ip)
2033 return scnprintf(msg, size, "%s",
2034 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2035 #if defined(__i386__) || defined(__x86_64__)
2036 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2037 return scnprintf(msg, size, "%s",
2038 "No hardware sampling interrupt available.\n"
2039 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2040 #endif
2041 break;
2042 case EBUSY:
2043 if (find_process("oprofiled"))
2044 return scnprintf(msg, size,
2045 "The PMU counters are busy/taken by another profiler.\n"
2046 "We found oprofile daemon running, please stop it and try again.");
2047 break;
2048 default:
2049 break;
2050 }
2051
2052 return scnprintf(msg, size,
2053 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2054 "/bin/dmesg may provide additional information.\n"
2055 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2056 err, strerror(err), perf_evsel__name(evsel));
2057 }
Linux kernel - Deliverables
This page took 0.073756 seconds and 5 git commands to generate.