]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - tools/perf/util/evsel.c
projects / mirror_ubuntu-artful-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge remote-tracking branches 'spi/topic/adi-v3', 'spi/topic/atmel', 'spi/topic...
[mirror_ubuntu-artful-kernel.git] / tools / perf / util / evsel.c
1 /*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9
10 #include <byteswap.h>
11 #include <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 } perf_missing_features;
33
34 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
35
36 int __perf_evsel__sample_size(u64 sample_type)
37 {
38 u64 mask = sample_type & PERF_SAMPLE_MASK;
39 int size = 0;
40 int i;
41
42 for (i = 0; i < 64; i++) {
43 if (mask & (1ULL << i))
44 size++;
45 }
46
47 size *= sizeof(u64);
48
49 return size;
50 }
51
52 /**
53 * __perf_evsel__calc_id_pos - calculate id_pos.
54 * @sample_type: sample type
55 *
56 * This function returns the position of the event id (PERF_SAMPLE_ID or
57 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
58 * sample_event.
59 */
60 static int __perf_evsel__calc_id_pos(u64 sample_type)
61 {
62 int idx = 0;
63
64 if (sample_type & PERF_SAMPLE_IDENTIFIER)
65 return 0;
66
67 if (!(sample_type & PERF_SAMPLE_ID))
68 return -1;
69
70 if (sample_type & PERF_SAMPLE_IP)
71 idx += 1;
72
73 if (sample_type & PERF_SAMPLE_TID)
74 idx += 1;
75
76 if (sample_type & PERF_SAMPLE_TIME)
77 idx += 1;
78
79 if (sample_type & PERF_SAMPLE_ADDR)
80 idx += 1;
81
82 return idx;
83 }
84
85 /**
86 * __perf_evsel__calc_is_pos - calculate is_pos.
87 * @sample_type: sample type
88 *
89 * This function returns the position (counting backwards) of the event id
90 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
91 * sample_id_all is used there is an id sample appended to non-sample events.
92 */
93 static int __perf_evsel__calc_is_pos(u64 sample_type)
94 {
95 int idx = 1;
96
97 if (sample_type & PERF_SAMPLE_IDENTIFIER)
98 return 1;
99
100 if (!(sample_type & PERF_SAMPLE_ID))
101 return -1;
102
103 if (sample_type & PERF_SAMPLE_CPU)
104 idx += 1;
105
106 if (sample_type & PERF_SAMPLE_STREAM_ID)
107 idx += 1;
108
109 return idx;
110 }
111
112 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
113 {
114 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
115 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
116 }
117
118 void hists__init(struct hists *hists)
119 {
120 memset(hists, 0, sizeof(*hists));
121 hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
122 hists->entries_in = &hists->entries_in_array[0];
123 hists->entries_collapsed = RB_ROOT;
124 hists->entries = RB_ROOT;
125 pthread_mutex_init(&hists->lock, NULL);
126 }
127
128 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
129 enum perf_event_sample_format bit)
130 {
131 if (!(evsel->attr.sample_type & bit)) {
132 evsel->attr.sample_type |= bit;
133 evsel->sample_size += sizeof(u64);
134 perf_evsel__calc_id_pos(evsel);
135 }
136 }
137
138 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
139 enum perf_event_sample_format bit)
140 {
141 if (evsel->attr.sample_type & bit) {
142 evsel->attr.sample_type &= ~bit;
143 evsel->sample_size -= sizeof(u64);
144 perf_evsel__calc_id_pos(evsel);
145 }
146 }
147
148 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
149 bool can_sample_identifier)
150 {
151 if (can_sample_identifier) {
152 perf_evsel__reset_sample_bit(evsel, ID);
153 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
154 } else {
155 perf_evsel__set_sample_bit(evsel, ID);
156 }
157 evsel->attr.read_format |= PERF_FORMAT_ID;
158 }
159
160 void perf_evsel__init(struct perf_evsel *evsel,
161 struct perf_event_attr *attr, int idx)
162 {
163 evsel->idx = idx;
164 evsel->attr = *attr;
165 evsel->leader = evsel;
166 evsel->unit = "";
167 evsel->scale = 1.0;
168 INIT_LIST_HEAD(&evsel->node);
169 hists__init(&evsel->hists);
170 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
171 perf_evsel__calc_id_pos(evsel);
172 }
173
174 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
175 {
176 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
177
178 if (evsel != NULL)
179 perf_evsel__init(evsel, attr, idx);
180
181 return evsel;
182 }
183
184 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
185 {
186 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
187
188 if (evsel != NULL) {
189 struct perf_event_attr attr = {
190 .type = PERF_TYPE_TRACEPOINT,
191 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
192 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
193 };
194
195 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
196 goto out_free;
197
198 evsel->tp_format = trace_event__tp_format(sys, name);
199 if (evsel->tp_format == NULL)
200 goto out_free;
201
202 event_attr_init(&attr);
203 attr.config = evsel->tp_format->id;
204 attr.sample_period = 1;
205 perf_evsel__init(evsel, &attr, idx);
206 }
207
208 return evsel;
209
210 out_free:
211 zfree(&evsel->name);
212 free(evsel);
213 return NULL;
214 }
215
216 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
217 "cycles",
218 "instructions",
219 "cache-references",
220 "cache-misses",
221 "branches",
222 "branch-misses",
223 "bus-cycles",
224 "stalled-cycles-frontend",
225 "stalled-cycles-backend",
226 "ref-cycles",
227 };
228
229 static const char *__perf_evsel__hw_name(u64 config)
230 {
231 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
232 return perf_evsel__hw_names[config];
233
234 return "unknown-hardware";
235 }
236
237 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
238 {
239 int colon = 0, r = 0;
240 struct perf_event_attr *attr = &evsel->attr;
241 bool exclude_guest_default = false;
242
243 #define MOD_PRINT(context, mod) do { \
244 if (!attr->exclude_##context) { \
245 if (!colon) colon = ++r; \
246 r += scnprintf(bf + r, size - r, "%c", mod); \
247 } } while(0)
248
249 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
250 MOD_PRINT(kernel, 'k');
251 MOD_PRINT(user, 'u');
252 MOD_PRINT(hv, 'h');
253 exclude_guest_default = true;
254 }
255
256 if (attr->precise_ip) {
257 if (!colon)
258 colon = ++r;
259 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
260 exclude_guest_default = true;
261 }
262
263 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
264 MOD_PRINT(host, 'H');
265 MOD_PRINT(guest, 'G');
266 }
267 #undef MOD_PRINT
268 if (colon)
269 bf[colon - 1] = ':';
270 return r;
271 }
272
273 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
274 {
275 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
276 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
277 }
278
279 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
280 "cpu-clock",
281 "task-clock",
282 "page-faults",
283 "context-switches",
284 "cpu-migrations",
285 "minor-faults",
286 "major-faults",
287 "alignment-faults",
288 "emulation-faults",
289 "dummy",
290 };
291
292 static const char *__perf_evsel__sw_name(u64 config)
293 {
294 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
295 return perf_evsel__sw_names[config];
296 return "unknown-software";
297 }
298
299 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
300 {
301 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
302 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
303 }
304
305 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
306 {
307 int r;
308
309 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
310
311 if (type & HW_BREAKPOINT_R)
312 r += scnprintf(bf + r, size - r, "r");
313
314 if (type & HW_BREAKPOINT_W)
315 r += scnprintf(bf + r, size - r, "w");
316
317 if (type & HW_BREAKPOINT_X)
318 r += scnprintf(bf + r, size - r, "x");
319
320 return r;
321 }
322
323 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
324 {
325 struct perf_event_attr *attr = &evsel->attr;
326 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
327 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
328 }
329
330 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
331 [PERF_EVSEL__MAX_ALIASES] = {
332 { "L1-dcache", "l1-d", "l1d", "L1-data", },
333 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
334 { "LLC", "L2", },
335 { "dTLB", "d-tlb", "Data-TLB", },
336 { "iTLB", "i-tlb", "Instruction-TLB", },
337 { "branch", "branches", "bpu", "btb", "bpc", },
338 { "node", },
339 };
340
341 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
342 [PERF_EVSEL__MAX_ALIASES] = {
343 { "load", "loads", "read", },
344 { "store", "stores", "write", },
345 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
346 };
347
348 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
349 [PERF_EVSEL__MAX_ALIASES] = {
350 { "refs", "Reference", "ops", "access", },
351 { "misses", "miss", },
352 };
353
354 #define C(x) PERF_COUNT_HW_CACHE_##x
355 #define CACHE_READ (1 << C(OP_READ))
356 #define CACHE_WRITE (1 << C(OP_WRITE))
357 #define CACHE_PREFETCH (1 << C(OP_PREFETCH))
358 #define COP(x) (1 << x)
359
360 /*
361 * cache operartion stat
362 * L1I : Read and prefetch only
363 * ITLB and BPU : Read-only
364 */
365 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
366 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
367 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
368 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
369 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
370 [C(ITLB)] = (CACHE_READ),
371 [C(BPU)] = (CACHE_READ),
372 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
373 };
374
375 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
376 {
377 if (perf_evsel__hw_cache_stat[type] & COP(op))
378 return true; /* valid */
379 else
380 return false; /* invalid */
381 }
382
383 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
384 char *bf, size_t size)
385 {
386 if (result) {
387 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
388 perf_evsel__hw_cache_op[op][0],
389 perf_evsel__hw_cache_result[result][0]);
390 }
391
392 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
393 perf_evsel__hw_cache_op[op][1]);
394 }
395
396 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
397 {
398 u8 op, result, type = (config >> 0) & 0xff;
399 const char *err = "unknown-ext-hardware-cache-type";
400
401 if (type > PERF_COUNT_HW_CACHE_MAX)
402 goto out_err;
403
404 op = (config >> 8) & 0xff;
405 err = "unknown-ext-hardware-cache-op";
406 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
407 goto out_err;
408
409 result = (config >> 16) & 0xff;
410 err = "unknown-ext-hardware-cache-result";
411 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
412 goto out_err;
413
414 err = "invalid-cache";
415 if (!perf_evsel__is_cache_op_valid(type, op))
416 goto out_err;
417
418 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
419 out_err:
420 return scnprintf(bf, size, "%s", err);
421 }
422
423 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
424 {
425 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
426 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
427 }
428
429 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
430 {
431 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
432 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
433 }
434
435 const char *perf_evsel__name(struct perf_evsel *evsel)
436 {
437 char bf[128];
438
439 if (evsel->name)
440 return evsel->name;
441
442 switch (evsel->attr.type) {
443 case PERF_TYPE_RAW:
444 perf_evsel__raw_name(evsel, bf, sizeof(bf));
445 break;
446
447 case PERF_TYPE_HARDWARE:
448 perf_evsel__hw_name(evsel, bf, sizeof(bf));
449 break;
450
451 case PERF_TYPE_HW_CACHE:
452 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
453 break;
454
455 case PERF_TYPE_SOFTWARE:
456 perf_evsel__sw_name(evsel, bf, sizeof(bf));
457 break;
458
459 case PERF_TYPE_TRACEPOINT:
460 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
461 break;
462
463 case PERF_TYPE_BREAKPOINT:
464 perf_evsel__bp_name(evsel, bf, sizeof(bf));
465 break;
466
467 default:
468 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
469 evsel->attr.type);
470 break;
471 }
472
473 evsel->name = strdup(bf);
474
475 return evsel->name ?: "unknown";
476 }
477
478 const char *perf_evsel__group_name(struct perf_evsel *evsel)
479 {
480 return evsel->group_name ?: "anon group";
481 }
482
483 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
484 {
485 int ret;
486 struct perf_evsel *pos;
487 const char *group_name = perf_evsel__group_name(evsel);
488
489 ret = scnprintf(buf, size, "%s", group_name);
490
491 ret += scnprintf(buf + ret, size - ret, " { %s",
492 perf_evsel__name(evsel));
493
494 for_each_group_member(pos, evsel)
495 ret += scnprintf(buf + ret, size - ret, ", %s",
496 perf_evsel__name(pos));
497
498 ret += scnprintf(buf + ret, size - ret, " }");
499
500 return ret;
501 }
502
503 static void
504 perf_evsel__config_callgraph(struct perf_evsel *evsel,
505 struct record_opts *opts)
506 {
507 bool function = perf_evsel__is_function_event(evsel);
508 struct perf_event_attr *attr = &evsel->attr;
509
510 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
511
512 if (opts->call_graph == CALLCHAIN_DWARF) {
513 if (!function) {
514 perf_evsel__set_sample_bit(evsel, REGS_USER);
515 perf_evsel__set_sample_bit(evsel, STACK_USER);
516 attr->sample_regs_user = PERF_REGS_MASK;
517 attr->sample_stack_user = opts->stack_dump_size;
518 attr->exclude_callchain_user = 1;
519 } else {
520 pr_info("Cannot use DWARF unwind for function trace event,"
521 " falling back to framepointers.\n");
522 }
523 }
524
525 if (function) {
526 pr_info("Disabling user space callchains for function trace event.\n");
527 attr->exclude_callchain_user = 1;
528 }
529 }
530
531 /*
532 * The enable_on_exec/disabled value strategy:
533 *
534 * 1) For any type of traced program:
535 * - all independent events and group leaders are disabled
536 * - all group members are enabled
537 *
538 * Group members are ruled by group leaders. They need to
539 * be enabled, because the group scheduling relies on that.
540 *
541 * 2) For traced programs executed by perf:
542 * - all independent events and group leaders have
543 * enable_on_exec set
544 * - we don't specifically enable or disable any event during
545 * the record command
546 *
547 * Independent events and group leaders are initially disabled
548 * and get enabled by exec. Group members are ruled by group
549 * leaders as stated in 1).
550 *
551 * 3) For traced programs attached by perf (pid/tid):
552 * - we specifically enable or disable all events during
553 * the record command
554 *
555 * When attaching events to already running traced we
556 * enable/disable events specifically, as there's no
557 * initial traced exec call.
558 */
559 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
560 {
561 struct perf_evsel *leader = evsel->leader;
562 struct perf_event_attr *attr = &evsel->attr;
563 int track = !evsel->idx; /* only the first counter needs these */
564 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
565
566 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
567 attr->inherit = !opts->no_inherit;
568
569 perf_evsel__set_sample_bit(evsel, IP);
570 perf_evsel__set_sample_bit(evsel, TID);
571
572 if (evsel->sample_read) {
573 perf_evsel__set_sample_bit(evsel, READ);
574
575 /*
576 * We need ID even in case of single event, because
577 * PERF_SAMPLE_READ process ID specific data.
578 */
579 perf_evsel__set_sample_id(evsel, false);
580
581 /*
582 * Apply group format only if we belong to group
583 * with more than one members.
584 */
585 if (leader->nr_members > 1) {
586 attr->read_format |= PERF_FORMAT_GROUP;
587 attr->inherit = 0;
588 }
589 }
590
591 /*
592 * We default some events to have a default interval. But keep
593 * it a weak assumption overridable by the user.
594 */
595 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
596 opts->user_interval != ULLONG_MAX)) {
597 if (opts->freq) {
598 perf_evsel__set_sample_bit(evsel, PERIOD);
599 attr->freq = 1;
600 attr->sample_freq = opts->freq;
601 } else {
602 attr->sample_period = opts->default_interval;
603 }
604 }
605
606 /*
607 * Disable sampling for all group members other
608 * than leader in case leader 'leads' the sampling.
609 */
610 if ((leader != evsel) && leader->sample_read) {
611 attr->sample_freq = 0;
612 attr->sample_period = 0;
613 }
614
615 if (opts->no_samples)
616 attr->sample_freq = 0;
617
618 if (opts->inherit_stat)
619 attr->inherit_stat = 1;
620
621 if (opts->sample_address) {
622 perf_evsel__set_sample_bit(evsel, ADDR);
623 attr->mmap_data = track;
624 }
625
626 if (opts->call_graph_enabled)
627 perf_evsel__config_callgraph(evsel, opts);
628
629 if (target__has_cpu(&opts->target))
630 perf_evsel__set_sample_bit(evsel, CPU);
631
632 if (opts->period)
633 perf_evsel__set_sample_bit(evsel, PERIOD);
634
635 if (!perf_missing_features.sample_id_all &&
636 (opts->sample_time || !opts->no_inherit ||
637 target__has_cpu(&opts->target) || per_cpu))
638 perf_evsel__set_sample_bit(evsel, TIME);
639
640 if (opts->raw_samples) {
641 perf_evsel__set_sample_bit(evsel, TIME);
642 perf_evsel__set_sample_bit(evsel, RAW);
643 perf_evsel__set_sample_bit(evsel, CPU);
644 }
645
646 if (opts->sample_address)
647 perf_evsel__set_sample_bit(evsel, DATA_SRC);
648
649 if (opts->no_buffering) {
650 attr->watermark = 0;
651 attr->wakeup_events = 1;
652 }
653 if (opts->branch_stack) {
654 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
655 attr->branch_sample_type = opts->branch_stack;
656 }
657
658 if (opts->sample_weight)
659 perf_evsel__set_sample_bit(evsel, WEIGHT);
660
661 attr->mmap = track;
662 attr->mmap2 = track && !perf_missing_features.mmap2;
663 attr->comm = track;
664
665 if (opts->sample_transaction)
666 perf_evsel__set_sample_bit(evsel, TRANSACTION);
667
668 /*
669 * XXX see the function comment above
670 *
671 * Disabling only independent events or group leaders,
672 * keeping group members enabled.
673 */
674 if (perf_evsel__is_group_leader(evsel))
675 attr->disabled = 1;
676
677 /*
678 * Setting enable_on_exec for independent events and
679 * group leaders for traced executed by perf.
680 */
681 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
682 !opts->initial_delay)
683 attr->enable_on_exec = 1;
684 }
685
686 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
687 {
688 int cpu, thread;
689 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
690
691 if (evsel->fd) {
692 for (cpu = 0; cpu < ncpus; cpu++) {
693 for (thread = 0; thread < nthreads; thread++) {
694 FD(evsel, cpu, thread) = -1;
695 }
696 }
697 }
698
699 return evsel->fd != NULL ? 0 : -ENOMEM;
700 }
701
702 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
703 int ioc, void *arg)
704 {
705 int cpu, thread;
706
707 for (cpu = 0; cpu < ncpus; cpu++) {
708 for (thread = 0; thread < nthreads; thread++) {
709 int fd = FD(evsel, cpu, thread),
710 err = ioctl(fd, ioc, arg);
711
712 if (err)
713 return err;
714 }
715 }
716
717 return 0;
718 }
719
720 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
721 const char *filter)
722 {
723 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
724 PERF_EVENT_IOC_SET_FILTER,
725 (void *)filter);
726 }
727
728 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
729 {
730 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
731 PERF_EVENT_IOC_ENABLE,
732 0);
733 }
734
735 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
736 {
737 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
738 if (evsel->sample_id == NULL)
739 return -ENOMEM;
740
741 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
742 if (evsel->id == NULL) {
743 xyarray__delete(evsel->sample_id);
744 evsel->sample_id = NULL;
745 return -ENOMEM;
746 }
747
748 return 0;
749 }
750
751 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
752 {
753 memset(evsel->counts, 0, (sizeof(*evsel->counts) +
754 (ncpus * sizeof(struct perf_counts_values))));
755 }
756
757 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
758 {
759 evsel->counts = zalloc((sizeof(*evsel->counts) +
760 (ncpus * sizeof(struct perf_counts_values))));
761 return evsel->counts != NULL ? 0 : -ENOMEM;
762 }
763
764 void perf_evsel__free_fd(struct perf_evsel *evsel)
765 {
766 xyarray__delete(evsel->fd);
767 evsel->fd = NULL;
768 }
769
770 void perf_evsel__free_id(struct perf_evsel *evsel)
771 {
772 xyarray__delete(evsel->sample_id);
773 evsel->sample_id = NULL;
774 zfree(&evsel->id);
775 }
776
777 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
778 {
779 int cpu, thread;
780
781 for (cpu = 0; cpu < ncpus; cpu++)
782 for (thread = 0; thread < nthreads; ++thread) {
783 close(FD(evsel, cpu, thread));
784 FD(evsel, cpu, thread) = -1;
785 }
786 }
787
788 void perf_evsel__free_counts(struct perf_evsel *evsel)
789 {
790 zfree(&evsel->counts);
791 }
792
793 void perf_evsel__exit(struct perf_evsel *evsel)
794 {
795 assert(list_empty(&evsel->node));
796 perf_evsel__free_fd(evsel);
797 perf_evsel__free_id(evsel);
798 }
799
800 void perf_evsel__delete(struct perf_evsel *evsel)
801 {
802 perf_evsel__exit(evsel);
803 close_cgroup(evsel->cgrp);
804 zfree(&evsel->group_name);
805 if (evsel->tp_format)
806 pevent_free_format(evsel->tp_format);
807 zfree(&evsel->name);
808 free(evsel);
809 }
810
811 static inline void compute_deltas(struct perf_evsel *evsel,
812 int cpu,
813 struct perf_counts_values *count)
814 {
815 struct perf_counts_values tmp;
816
817 if (!evsel->prev_raw_counts)
818 return;
819
820 if (cpu == -1) {
821 tmp = evsel->prev_raw_counts->aggr;
822 evsel->prev_raw_counts->aggr = *count;
823 } else {
824 tmp = evsel->prev_raw_counts->cpu[cpu];
825 evsel->prev_raw_counts->cpu[cpu] = *count;
826 }
827
828 count->val = count->val - tmp.val;
829 count->ena = count->ena - tmp.ena;
830 count->run = count->run - tmp.run;
831 }
832
833 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
834 int cpu, int thread, bool scale)
835 {
836 struct perf_counts_values count;
837 size_t nv = scale ? 3 : 1;
838
839 if (FD(evsel, cpu, thread) < 0)
840 return -EINVAL;
841
842 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
843 return -ENOMEM;
844
845 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
846 return -errno;
847
848 compute_deltas(evsel, cpu, &count);
849
850 if (scale) {
851 if (count.run == 0)
852 count.val = 0;
853 else if (count.run < count.ena)
854 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
855 } else
856 count.ena = count.run = 0;
857
858 evsel->counts->cpu[cpu] = count;
859 return 0;
860 }
861
862 int __perf_evsel__read(struct perf_evsel *evsel,
863 int ncpus, int nthreads, bool scale)
864 {
865 size_t nv = scale ? 3 : 1;
866 int cpu, thread;
867 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
868
869 aggr->val = aggr->ena = aggr->run = 0;
870
871 for (cpu = 0; cpu < ncpus; cpu++) {
872 for (thread = 0; thread < nthreads; thread++) {
873 if (FD(evsel, cpu, thread) < 0)
874 continue;
875
876 if (readn(FD(evsel, cpu, thread),
877 &count, nv * sizeof(u64)) < 0)
878 return -errno;
879
880 aggr->val += count.val;
881 if (scale) {
882 aggr->ena += count.ena;
883 aggr->run += count.run;
884 }
885 }
886 }
887
888 compute_deltas(evsel, -1, aggr);
889
890 evsel->counts->scaled = 0;
891 if (scale) {
892 if (aggr->run == 0) {
893 evsel->counts->scaled = -1;
894 aggr->val = 0;
895 return 0;
896 }
897
898 if (aggr->run < aggr->ena) {
899 evsel->counts->scaled = 1;
900 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
901 }
902 } else
903 aggr->ena = aggr->run = 0;
904
905 return 0;
906 }
907
908 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
909 {
910 struct perf_evsel *leader = evsel->leader;
911 int fd;
912
913 if (perf_evsel__is_group_leader(evsel))
914 return -1;
915
916 /*
917 * Leader must be already processed/open,
918 * if not it's a bug.
919 */
920 BUG_ON(!leader->fd);
921
922 fd = FD(leader, cpu, thread);
923 BUG_ON(fd == -1);
924
925 return fd;
926 }
927
928 #define __PRINT_ATTR(fmt, cast, field) \
929 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
930
931 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
932 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
933 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
934 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
935
936 #define PRINT_ATTR2N(name1, field1, name2, field2) \
937 fprintf(fp, " %-19s %u %-19s %u\n", \
938 name1, attr->field1, name2, attr->field2)
939
940 #define PRINT_ATTR2(field1, field2) \
941 PRINT_ATTR2N(#field1, field1, #field2, field2)
942
943 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
944 {
945 size_t ret = 0;
946
947 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
948 ret += fprintf(fp, "perf_event_attr:\n");
949
950 ret += PRINT_ATTR_U32(type);
951 ret += PRINT_ATTR_U32(size);
952 ret += PRINT_ATTR_X64(config);
953 ret += PRINT_ATTR_U64(sample_period);
954 ret += PRINT_ATTR_U64(sample_freq);
955 ret += PRINT_ATTR_X64(sample_type);
956 ret += PRINT_ATTR_X64(read_format);
957
958 ret += PRINT_ATTR2(disabled, inherit);
959 ret += PRINT_ATTR2(pinned, exclusive);
960 ret += PRINT_ATTR2(exclude_user, exclude_kernel);
961 ret += PRINT_ATTR2(exclude_hv, exclude_idle);
962 ret += PRINT_ATTR2(mmap, comm);
963 ret += PRINT_ATTR2(freq, inherit_stat);
964 ret += PRINT_ATTR2(enable_on_exec, task);
965 ret += PRINT_ATTR2(watermark, precise_ip);
966 ret += PRINT_ATTR2(mmap_data, sample_id_all);
967 ret += PRINT_ATTR2(exclude_host, exclude_guest);
968 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
969 "excl.callchain_user", exclude_callchain_user);
970 ret += PRINT_ATTR_U32(mmap2);
971
972 ret += PRINT_ATTR_U32(wakeup_events);
973 ret += PRINT_ATTR_U32(wakeup_watermark);
974 ret += PRINT_ATTR_X32(bp_type);
975 ret += PRINT_ATTR_X64(bp_addr);
976 ret += PRINT_ATTR_X64(config1);
977 ret += PRINT_ATTR_U64(bp_len);
978 ret += PRINT_ATTR_X64(config2);
979 ret += PRINT_ATTR_X64(branch_sample_type);
980 ret += PRINT_ATTR_X64(sample_regs_user);
981 ret += PRINT_ATTR_U32(sample_stack_user);
982
983 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
984
985 return ret;
986 }
987
988 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
989 struct thread_map *threads)
990 {
991 int cpu, thread;
992 unsigned long flags = 0;
993 int pid = -1, err;
994 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
995
996 if (evsel->fd == NULL &&
997 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
998 return -ENOMEM;
999
1000 if (evsel->cgrp) {
1001 flags = PERF_FLAG_PID_CGROUP;
1002 pid = evsel->cgrp->fd;
1003 }
1004
1005 fallback_missing_features:
1006 if (perf_missing_features.mmap2)
1007 evsel->attr.mmap2 = 0;
1008 if (perf_missing_features.exclude_guest)
1009 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1010 retry_sample_id:
1011 if (perf_missing_features.sample_id_all)
1012 evsel->attr.sample_id_all = 0;
1013
1014 if (verbose >= 2)
1015 perf_event_attr__fprintf(&evsel->attr, stderr);
1016
1017 for (cpu = 0; cpu < cpus->nr; cpu++) {
1018
1019 for (thread = 0; thread < threads->nr; thread++) {
1020 int group_fd;
1021
1022 if (!evsel->cgrp)
1023 pid = threads->map[thread];
1024
1025 group_fd = get_group_fd(evsel, cpu, thread);
1026 retry_open:
1027 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1028 pid, cpus->map[cpu], group_fd, flags);
1029
1030 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1031 pid,
1032 cpus->map[cpu],
1033 group_fd, flags);
1034 if (FD(evsel, cpu, thread) < 0) {
1035 err = -errno;
1036 pr_debug2("sys_perf_event_open failed, error %d\n",
1037 err);
1038 goto try_fallback;
1039 }
1040 set_rlimit = NO_CHANGE;
1041 }
1042 }
1043
1044 return 0;
1045
1046 try_fallback:
1047 /*
1048 * perf stat needs between 5 and 22 fds per CPU. When we run out
1049 * of them try to increase the limits.
1050 */
1051 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1052 struct rlimit l;
1053 int old_errno = errno;
1054
1055 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1056 if (set_rlimit == NO_CHANGE)
1057 l.rlim_cur = l.rlim_max;
1058 else {
1059 l.rlim_cur = l.rlim_max + 1000;
1060 l.rlim_max = l.rlim_cur;
1061 }
1062 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1063 set_rlimit++;
1064 errno = old_errno;
1065 goto retry_open;
1066 }
1067 }
1068 errno = old_errno;
1069 }
1070
1071 if (err != -EINVAL || cpu > 0 || thread > 0)
1072 goto out_close;
1073
1074 if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1075 perf_missing_features.mmap2 = true;
1076 goto fallback_missing_features;
1077 } else if (!perf_missing_features.exclude_guest &&
1078 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1079 perf_missing_features.exclude_guest = true;
1080 goto fallback_missing_features;
1081 } else if (!perf_missing_features.sample_id_all) {
1082 perf_missing_features.sample_id_all = true;
1083 goto retry_sample_id;
1084 }
1085
1086 out_close:
1087 do {
1088 while (--thread >= 0) {
1089 close(FD(evsel, cpu, thread));
1090 FD(evsel, cpu, thread) = -1;
1091 }
1092 thread = threads->nr;
1093 } while (--cpu >= 0);
1094 return err;
1095 }
1096
1097 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1098 {
1099 if (evsel->fd == NULL)
1100 return;
1101
1102 perf_evsel__close_fd(evsel, ncpus, nthreads);
1103 perf_evsel__free_fd(evsel);
1104 }
1105
1106 static struct {
1107 struct cpu_map map;
1108 int cpus[1];
1109 } empty_cpu_map = {
1110 .map.nr = 1,
1111 .cpus = { -1, },
1112 };
1113
1114 static struct {
1115 struct thread_map map;
1116 int threads[1];
1117 } empty_thread_map = {
1118 .map.nr = 1,
1119 .threads = { -1, },
1120 };
1121
1122 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1123 struct thread_map *threads)
1124 {
1125 if (cpus == NULL) {
1126 /* Work around old compiler warnings about strict aliasing */
1127 cpus = &empty_cpu_map.map;
1128 }
1129
1130 if (threads == NULL)
1131 threads = &empty_thread_map.map;
1132
1133 return __perf_evsel__open(evsel, cpus, threads);
1134 }
1135
1136 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1137 struct cpu_map *cpus)
1138 {
1139 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1140 }
1141
1142 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1143 struct thread_map *threads)
1144 {
1145 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1146 }
1147
1148 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1149 const union perf_event *event,
1150 struct perf_sample *sample)
1151 {
1152 u64 type = evsel->attr.sample_type;
1153 const u64 *array = event->sample.array;
1154 bool swapped = evsel->needs_swap;
1155 union u64_swap u;
1156
1157 array += ((event->header.size -
1158 sizeof(event->header)) / sizeof(u64)) - 1;
1159
1160 if (type & PERF_SAMPLE_IDENTIFIER) {
1161 sample->id = *array;
1162 array--;
1163 }
1164
1165 if (type & PERF_SAMPLE_CPU) {
1166 u.val64 = *array;
1167 if (swapped) {
1168 /* undo swap of u64, then swap on individual u32s */
1169 u.val64 = bswap_64(u.val64);
1170 u.val32[0] = bswap_32(u.val32[0]);
1171 }
1172
1173 sample->cpu = u.val32[0];
1174 array--;
1175 }
1176
1177 if (type & PERF_SAMPLE_STREAM_ID) {
1178 sample->stream_id = *array;
1179 array--;
1180 }
1181
1182 if (type & PERF_SAMPLE_ID) {
1183 sample->id = *array;
1184 array--;
1185 }
1186
1187 if (type & PERF_SAMPLE_TIME) {
1188 sample->time = *array;
1189 array--;
1190 }
1191
1192 if (type & PERF_SAMPLE_TID) {
1193 u.val64 = *array;
1194 if (swapped) {
1195 /* undo swap of u64, then swap on individual u32s */
1196 u.val64 = bswap_64(u.val64);
1197 u.val32[0] = bswap_32(u.val32[0]);
1198 u.val32[1] = bswap_32(u.val32[1]);
1199 }
1200
1201 sample->pid = u.val32[0];
1202 sample->tid = u.val32[1];
1203 array--;
1204 }
1205
1206 return 0;
1207 }
1208
1209 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1210 u64 size)
1211 {
1212 return size > max_size || offset + size > endp;
1213 }
1214
1215 #define OVERFLOW_CHECK(offset, size, max_size) \
1216 do { \
1217 if (overflow(endp, (max_size), (offset), (size))) \
1218 return -EFAULT; \
1219 } while (0)
1220
1221 #define OVERFLOW_CHECK_u64(offset) \
1222 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1223
1224 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1225 struct perf_sample *data)
1226 {
1227 u64 type = evsel->attr.sample_type;
1228 bool swapped = evsel->needs_swap;
1229 const u64 *array;
1230 u16 max_size = event->header.size;
1231 const void *endp = (void *)event + max_size;
1232 u64 sz;
1233
1234 /*
1235 * used for cross-endian analysis. See git commit 65014ab3
1236 * for why this goofiness is needed.
1237 */
1238 union u64_swap u;
1239
1240 memset(data, 0, sizeof(*data));
1241 data->cpu = data->pid = data->tid = -1;
1242 data->stream_id = data->id = data->time = -1ULL;
1243 data->period = evsel->attr.sample_period;
1244 data->weight = 0;
1245
1246 if (event->header.type != PERF_RECORD_SAMPLE) {
1247 if (!evsel->attr.sample_id_all)
1248 return 0;
1249 return perf_evsel__parse_id_sample(evsel, event, data);
1250 }
1251
1252 array = event->sample.array;
1253
1254 /*
1255 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1256 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1257 * check the format does not go past the end of the event.
1258 */
1259 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1260 return -EFAULT;
1261
1262 data->id = -1ULL;
1263 if (type & PERF_SAMPLE_IDENTIFIER) {
1264 data->id = *array;
1265 array++;
1266 }
1267
1268 if (type & PERF_SAMPLE_IP) {
1269 data->ip = *array;
1270 array++;
1271 }
1272
1273 if (type & PERF_SAMPLE_TID) {
1274 u.val64 = *array;
1275 if (swapped) {
1276 /* undo swap of u64, then swap on individual u32s */
1277 u.val64 = bswap_64(u.val64);
1278 u.val32[0] = bswap_32(u.val32[0]);
1279 u.val32[1] = bswap_32(u.val32[1]);
1280 }
1281
1282 data->pid = u.val32[0];
1283 data->tid = u.val32[1];
1284 array++;
1285 }
1286
1287 if (type & PERF_SAMPLE_TIME) {
1288 data->time = *array;
1289 array++;
1290 }
1291
1292 data->addr = 0;
1293 if (type & PERF_SAMPLE_ADDR) {
1294 data->addr = *array;
1295 array++;
1296 }
1297
1298 if (type & PERF_SAMPLE_ID) {
1299 data->id = *array;
1300 array++;
1301 }
1302
1303 if (type & PERF_SAMPLE_STREAM_ID) {
1304 data->stream_id = *array;
1305 array++;
1306 }
1307
1308 if (type & PERF_SAMPLE_CPU) {
1309
1310 u.val64 = *array;
1311 if (swapped) {
1312 /* undo swap of u64, then swap on individual u32s */
1313 u.val64 = bswap_64(u.val64);
1314 u.val32[0] = bswap_32(u.val32[0]);
1315 }
1316
1317 data->cpu = u.val32[0];
1318 array++;
1319 }
1320
1321 if (type & PERF_SAMPLE_PERIOD) {
1322 data->period = *array;
1323 array++;
1324 }
1325
1326 if (type & PERF_SAMPLE_READ) {
1327 u64 read_format = evsel->attr.read_format;
1328
1329 OVERFLOW_CHECK_u64(array);
1330 if (read_format & PERF_FORMAT_GROUP)
1331 data->read.group.nr = *array;
1332 else
1333 data->read.one.value = *array;
1334
1335 array++;
1336
1337 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1338 OVERFLOW_CHECK_u64(array);
1339 data->read.time_enabled = *array;
1340 array++;
1341 }
1342
1343 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1344 OVERFLOW_CHECK_u64(array);
1345 data->read.time_running = *array;
1346 array++;
1347 }
1348
1349 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1350 if (read_format & PERF_FORMAT_GROUP) {
1351 const u64 max_group_nr = UINT64_MAX /
1352 sizeof(struct sample_read_value);
1353
1354 if (data->read.group.nr > max_group_nr)
1355 return -EFAULT;
1356 sz = data->read.group.nr *
1357 sizeof(struct sample_read_value);
1358 OVERFLOW_CHECK(array, sz, max_size);
1359 data->read.group.values =
1360 (struct sample_read_value *)array;
1361 array = (void *)array + sz;
1362 } else {
1363 OVERFLOW_CHECK_u64(array);
1364 data->read.one.id = *array;
1365 array++;
1366 }
1367 }
1368
1369 if (type & PERF_SAMPLE_CALLCHAIN) {
1370 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1371
1372 OVERFLOW_CHECK_u64(array);
1373 data->callchain = (struct ip_callchain *)array++;
1374 if (data->callchain->nr > max_callchain_nr)
1375 return -EFAULT;
1376 sz = data->callchain->nr * sizeof(u64);
1377 OVERFLOW_CHECK(array, sz, max_size);
1378 array = (void *)array + sz;
1379 }
1380
1381 if (type & PERF_SAMPLE_RAW) {
1382 OVERFLOW_CHECK_u64(array);
1383 u.val64 = *array;
1384 if (WARN_ONCE(swapped,
1385 "Endianness of raw data not corrected!\n")) {
1386 /* undo swap of u64, then swap on individual u32s */
1387 u.val64 = bswap_64(u.val64);
1388 u.val32[0] = bswap_32(u.val32[0]);
1389 u.val32[1] = bswap_32(u.val32[1]);
1390 }
1391 data->raw_size = u.val32[0];
1392 array = (void *)array + sizeof(u32);
1393
1394 OVERFLOW_CHECK(array, data->raw_size, max_size);
1395 data->raw_data = (void *)array;
1396 array = (void *)array + data->raw_size;
1397 }
1398
1399 if (type & PERF_SAMPLE_BRANCH_STACK) {
1400 const u64 max_branch_nr = UINT64_MAX /
1401 sizeof(struct branch_entry);
1402
1403 OVERFLOW_CHECK_u64(array);
1404 data->branch_stack = (struct branch_stack *)array++;
1405
1406 if (data->branch_stack->nr > max_branch_nr)
1407 return -EFAULT;
1408 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1409 OVERFLOW_CHECK(array, sz, max_size);
1410 array = (void *)array + sz;
1411 }
1412
1413 if (type & PERF_SAMPLE_REGS_USER) {
1414 OVERFLOW_CHECK_u64(array);
1415 data->user_regs.abi = *array;
1416 array++;
1417
1418 if (data->user_regs.abi) {
1419 u64 mask = evsel->attr.sample_regs_user;
1420
1421 sz = hweight_long(mask) * sizeof(u64);
1422 OVERFLOW_CHECK(array, sz, max_size);
1423 data->user_regs.mask = mask;
1424 data->user_regs.regs = (u64 *)array;
1425 array = (void *)array + sz;
1426 }
1427 }
1428
1429 if (type & PERF_SAMPLE_STACK_USER) {
1430 OVERFLOW_CHECK_u64(array);
1431 sz = *array++;
1432
1433 data->user_stack.offset = ((char *)(array - 1)
1434 - (char *) event);
1435
1436 if (!sz) {
1437 data->user_stack.size = 0;
1438 } else {
1439 OVERFLOW_CHECK(array, sz, max_size);
1440 data->user_stack.data = (char *)array;
1441 array = (void *)array + sz;
1442 OVERFLOW_CHECK_u64(array);
1443 data->user_stack.size = *array++;
1444 if (WARN_ONCE(data->user_stack.size > sz,
1445 "user stack dump failure\n"))
1446 return -EFAULT;
1447 }
1448 }
1449
1450 data->weight = 0;
1451 if (type & PERF_SAMPLE_WEIGHT) {
1452 OVERFLOW_CHECK_u64(array);
1453 data->weight = *array;
1454 array++;
1455 }
1456
1457 data->data_src = PERF_MEM_DATA_SRC_NONE;
1458 if (type & PERF_SAMPLE_DATA_SRC) {
1459 OVERFLOW_CHECK_u64(array);
1460 data->data_src = *array;
1461 array++;
1462 }
1463
1464 data->transaction = 0;
1465 if (type & PERF_SAMPLE_TRANSACTION) {
1466 OVERFLOW_CHECK_u64(array);
1467 data->transaction = *array;
1468 array++;
1469 }
1470
1471 return 0;
1472 }
1473
1474 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1475 u64 read_format)
1476 {
1477 size_t sz, result = sizeof(struct sample_event);
1478
1479 if (type & PERF_SAMPLE_IDENTIFIER)
1480 result += sizeof(u64);
1481
1482 if (type & PERF_SAMPLE_IP)
1483 result += sizeof(u64);
1484
1485 if (type & PERF_SAMPLE_TID)
1486 result += sizeof(u64);
1487
1488 if (type & PERF_SAMPLE_TIME)
1489 result += sizeof(u64);
1490
1491 if (type & PERF_SAMPLE_ADDR)
1492 result += sizeof(u64);
1493
1494 if (type & PERF_SAMPLE_ID)
1495 result += sizeof(u64);
1496
1497 if (type & PERF_SAMPLE_STREAM_ID)
1498 result += sizeof(u64);
1499
1500 if (type & PERF_SAMPLE_CPU)
1501 result += sizeof(u64);
1502
1503 if (type & PERF_SAMPLE_PERIOD)
1504 result += sizeof(u64);
1505
1506 if (type & PERF_SAMPLE_READ) {
1507 result += sizeof(u64);
1508 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1509 result += sizeof(u64);
1510 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1511 result += sizeof(u64);
1512 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1513 if (read_format & PERF_FORMAT_GROUP) {
1514 sz = sample->read.group.nr *
1515 sizeof(struct sample_read_value);
1516 result += sz;
1517 } else {
1518 result += sizeof(u64);
1519 }
1520 }
1521
1522 if (type & PERF_SAMPLE_CALLCHAIN) {
1523 sz = (sample->callchain->nr + 1) * sizeof(u64);
1524 result += sz;
1525 }
1526
1527 if (type & PERF_SAMPLE_RAW) {
1528 result += sizeof(u32);
1529 result += sample->raw_size;
1530 }
1531
1532 if (type & PERF_SAMPLE_BRANCH_STACK) {
1533 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1534 sz += sizeof(u64);
1535 result += sz;
1536 }
1537
1538 if (type & PERF_SAMPLE_REGS_USER) {
1539 if (sample->user_regs.abi) {
1540 result += sizeof(u64);
1541 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1542 result += sz;
1543 } else {
1544 result += sizeof(u64);
1545 }
1546 }
1547
1548 if (type & PERF_SAMPLE_STACK_USER) {
1549 sz = sample->user_stack.size;
1550 result += sizeof(u64);
1551 if (sz) {
1552 result += sz;
1553 result += sizeof(u64);
1554 }
1555 }
1556
1557 if (type & PERF_SAMPLE_WEIGHT)
1558 result += sizeof(u64);
1559
1560 if (type & PERF_SAMPLE_DATA_SRC)
1561 result += sizeof(u64);
1562
1563 if (type & PERF_SAMPLE_TRANSACTION)
1564 result += sizeof(u64);
1565
1566 return result;
1567 }
1568
1569 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1570 u64 read_format,
1571 const struct perf_sample *sample,
1572 bool swapped)
1573 {
1574 u64 *array;
1575 size_t sz;
1576 /*
1577 * used for cross-endian analysis. See git commit 65014ab3
1578 * for why this goofiness is needed.
1579 */
1580 union u64_swap u;
1581
1582 array = event->sample.array;
1583
1584 if (type & PERF_SAMPLE_IDENTIFIER) {
1585 *array = sample->id;
1586 array++;
1587 }
1588
1589 if (type & PERF_SAMPLE_IP) {
1590 *array = sample->ip;
1591 array++;
1592 }
1593
1594 if (type & PERF_SAMPLE_TID) {
1595 u.val32[0] = sample->pid;
1596 u.val32[1] = sample->tid;
1597 if (swapped) {
1598 /*
1599 * Inverse of what is done in perf_evsel__parse_sample
1600 */
1601 u.val32[0] = bswap_32(u.val32[0]);
1602 u.val32[1] = bswap_32(u.val32[1]);
1603 u.val64 = bswap_64(u.val64);
1604 }
1605
1606 *array = u.val64;
1607 array++;
1608 }
1609
1610 if (type & PERF_SAMPLE_TIME) {
1611 *array = sample->time;
1612 array++;
1613 }
1614
1615 if (type & PERF_SAMPLE_ADDR) {
1616 *array = sample->addr;
1617 array++;
1618 }
1619
1620 if (type & PERF_SAMPLE_ID) {
1621 *array = sample->id;
1622 array++;
1623 }
1624
1625 if (type & PERF_SAMPLE_STREAM_ID) {
1626 *array = sample->stream_id;
1627 array++;
1628 }
1629
1630 if (type & PERF_SAMPLE_CPU) {
1631 u.val32[0] = sample->cpu;
1632 if (swapped) {
1633 /*
1634 * Inverse of what is done in perf_evsel__parse_sample
1635 */
1636 u.val32[0] = bswap_32(u.val32[0]);
1637 u.val64 = bswap_64(u.val64);
1638 }
1639 *array = u.val64;
1640 array++;
1641 }
1642
1643 if (type & PERF_SAMPLE_PERIOD) {
1644 *array = sample->period;
1645 array++;
1646 }
1647
1648 if (type & PERF_SAMPLE_READ) {
1649 if (read_format & PERF_FORMAT_GROUP)
1650 *array = sample->read.group.nr;
1651 else
1652 *array = sample->read.one.value;
1653 array++;
1654
1655 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1656 *array = sample->read.time_enabled;
1657 array++;
1658 }
1659
1660 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1661 *array = sample->read.time_running;
1662 array++;
1663 }
1664
1665 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1666 if (read_format & PERF_FORMAT_GROUP) {
1667 sz = sample->read.group.nr *
1668 sizeof(struct sample_read_value);
1669 memcpy(array, sample->read.group.values, sz);
1670 array = (void *)array + sz;
1671 } else {
1672 *array = sample->read.one.id;
1673 array++;
1674 }
1675 }
1676
1677 if (type & PERF_SAMPLE_CALLCHAIN) {
1678 sz = (sample->callchain->nr + 1) * sizeof(u64);
1679 memcpy(array, sample->callchain, sz);
1680 array = (void *)array + sz;
1681 }
1682
1683 if (type & PERF_SAMPLE_RAW) {
1684 u.val32[0] = sample->raw_size;
1685 if (WARN_ONCE(swapped,
1686 "Endianness of raw data not corrected!\n")) {
1687 /*
1688 * Inverse of what is done in perf_evsel__parse_sample
1689 */
1690 u.val32[0] = bswap_32(u.val32[0]);
1691 u.val32[1] = bswap_32(u.val32[1]);
1692 u.val64 = bswap_64(u.val64);
1693 }
1694 *array = u.val64;
1695 array = (void *)array + sizeof(u32);
1696
1697 memcpy(array, sample->raw_data, sample->raw_size);
1698 array = (void *)array + sample->raw_size;
1699 }
1700
1701 if (type & PERF_SAMPLE_BRANCH_STACK) {
1702 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1703 sz += sizeof(u64);
1704 memcpy(array, sample->branch_stack, sz);
1705 array = (void *)array + sz;
1706 }
1707
1708 if (type & PERF_SAMPLE_REGS_USER) {
1709 if (sample->user_regs.abi) {
1710 *array++ = sample->user_regs.abi;
1711 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1712 memcpy(array, sample->user_regs.regs, sz);
1713 array = (void *)array + sz;
1714 } else {
1715 *array++ = 0;
1716 }
1717 }
1718
1719 if (type & PERF_SAMPLE_STACK_USER) {
1720 sz = sample->user_stack.size;
1721 *array++ = sz;
1722 if (sz) {
1723 memcpy(array, sample->user_stack.data, sz);
1724 array = (void *)array + sz;
1725 *array++ = sz;
1726 }
1727 }
1728
1729 if (type & PERF_SAMPLE_WEIGHT) {
1730 *array = sample->weight;
1731 array++;
1732 }
1733
1734 if (type & PERF_SAMPLE_DATA_SRC) {
1735 *array = sample->data_src;
1736 array++;
1737 }
1738
1739 if (type & PERF_SAMPLE_TRANSACTION) {
1740 *array = sample->transaction;
1741 array++;
1742 }
1743
1744 return 0;
1745 }
1746
1747 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1748 {
1749 return pevent_find_field(evsel->tp_format, name);
1750 }
1751
1752 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1753 const char *name)
1754 {
1755 struct format_field *field = perf_evsel__field(evsel, name);
1756 int offset;
1757
1758 if (!field)
1759 return NULL;
1760
1761 offset = field->offset;
1762
1763 if (field->flags & FIELD_IS_DYNAMIC) {
1764 offset = *(int *)(sample->raw_data + field->offset);
1765 offset &= 0xffff;
1766 }
1767
1768 return sample->raw_data + offset;
1769 }
1770
1771 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1772 const char *name)
1773 {
1774 struct format_field *field = perf_evsel__field(evsel, name);
1775 void *ptr;
1776 u64 value;
1777
1778 if (!field)
1779 return 0;
1780
1781 ptr = sample->raw_data + field->offset;
1782
1783 switch (field->size) {
1784 case 1:
1785 return *(u8 *)ptr;
1786 case 2:
1787 value = *(u16 *)ptr;
1788 break;
1789 case 4:
1790 value = *(u32 *)ptr;
1791 break;
1792 case 8:
1793 value = *(u64 *)ptr;
1794 break;
1795 default:
1796 return 0;
1797 }
1798
1799 if (!evsel->needs_swap)
1800 return value;
1801
1802 switch (field->size) {
1803 case 2:
1804 return bswap_16(value);
1805 case 4:
1806 return bswap_32(value);
1807 case 8:
1808 return bswap_64(value);
1809 default:
1810 return 0;
1811 }
1812
1813 return 0;
1814 }
1815
1816 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1817 {
1818 va_list args;
1819 int ret = 0;
1820
1821 if (!*first) {
1822 ret += fprintf(fp, ",");
1823 } else {
1824 ret += fprintf(fp, ":");
1825 *first = false;
1826 }
1827
1828 va_start(args, fmt);
1829 ret += vfprintf(fp, fmt, args);
1830 va_end(args);
1831 return ret;
1832 }
1833
1834 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1835 {
1836 if (value == 0)
1837 return 0;
1838
1839 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1840 }
1841
1842 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1843
1844 struct bit_names {
1845 int bit;
1846 const char *name;
1847 };
1848
1849 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1850 struct bit_names *bits, bool *first)
1851 {
1852 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1853 bool first_bit = true;
1854
1855 do {
1856 if (value & bits[i].bit) {
1857 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1858 first_bit = false;
1859 }
1860 } while (bits[++i].name != NULL);
1861
1862 return printed;
1863 }
1864
1865 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1866 {
1867 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1868 struct bit_names bits[] = {
1869 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1870 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1871 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1872 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1873 bit_name(IDENTIFIER),
1874 { .name = NULL, }
1875 };
1876 #undef bit_name
1877 return bits__fprintf(fp, "sample_type", value, bits, first);
1878 }
1879
1880 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1881 {
1882 #define bit_name(n) { PERF_FORMAT_##n, #n }
1883 struct bit_names bits[] = {
1884 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1885 bit_name(ID), bit_name(GROUP),
1886 { .name = NULL, }
1887 };
1888 #undef bit_name
1889 return bits__fprintf(fp, "read_format", value, bits, first);
1890 }
1891
1892 int perf_evsel__fprintf(struct perf_evsel *evsel,
1893 struct perf_attr_details *details, FILE *fp)
1894 {
1895 bool first = true;
1896 int printed = 0;
1897
1898 if (details->event_group) {
1899 struct perf_evsel *pos;
1900
1901 if (!perf_evsel__is_group_leader(evsel))
1902 return 0;
1903
1904 if (evsel->nr_members > 1)
1905 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1906
1907 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1908 for_each_group_member(pos, evsel)
1909 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1910
1911 if (evsel->nr_members > 1)
1912 printed += fprintf(fp, "}");
1913 goto out;
1914 }
1915
1916 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1917
1918 if (details->verbose || details->freq) {
1919 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1920 (u64)evsel->attr.sample_freq);
1921 }
1922
1923 if (details->verbose) {
1924 if_print(type);
1925 if_print(config);
1926 if_print(config1);
1927 if_print(config2);
1928 if_print(size);
1929 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1930 if (evsel->attr.read_format)
1931 printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1932 if_print(disabled);
1933 if_print(inherit);
1934 if_print(pinned);
1935 if_print(exclusive);
1936 if_print(exclude_user);
1937 if_print(exclude_kernel);
1938 if_print(exclude_hv);
1939 if_print(exclude_idle);
1940 if_print(mmap);
1941 if_print(mmap2);
1942 if_print(comm);
1943 if_print(freq);
1944 if_print(inherit_stat);
1945 if_print(enable_on_exec);
1946 if_print(task);
1947 if_print(watermark);
1948 if_print(precise_ip);
1949 if_print(mmap_data);
1950 if_print(sample_id_all);
1951 if_print(exclude_host);
1952 if_print(exclude_guest);
1953 if_print(__reserved_1);
1954 if_print(wakeup_events);
1955 if_print(bp_type);
1956 if_print(branch_sample_type);
1957 }
1958 out:
1959 fputc('\n', fp);
1960 return ++printed;
1961 }
1962
1963 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1964 char *msg, size_t msgsize)
1965 {
1966 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
1967 evsel->attr.type == PERF_TYPE_HARDWARE &&
1968 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1969 /*
1970 * If it's cycles then fall back to hrtimer based
1971 * cpu-clock-tick sw counter, which is always available even if
1972 * no PMU support.
1973 *
1974 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1975 * b0a873e).
1976 */
1977 scnprintf(msg, msgsize, "%s",
1978 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1979
1980 evsel->attr.type = PERF_TYPE_SOFTWARE;
1981 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
1982
1983 zfree(&evsel->name);
1984 return true;
1985 }
1986
1987 return false;
1988 }
1989
1990 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
1991 int err, char *msg, size_t size)
1992 {
1993 switch (err) {
1994 case EPERM:
1995 case EACCES:
1996 return scnprintf(msg, size,
1997 "You may not have permission to collect %sstats.\n"
1998 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
1999 " -1 - Not paranoid at all\n"
2000 " 0 - Disallow raw tracepoint access for unpriv\n"
2001 " 1 - Disallow cpu events for unpriv\n"
2002 " 2 - Disallow kernel profiling for unpriv",
2003 target->system_wide ? "system-wide " : "");
2004 case ENOENT:
2005 return scnprintf(msg, size, "The %s event is not supported.",
2006 perf_evsel__name(evsel));
2007 case EMFILE:
2008 return scnprintf(msg, size, "%s",
2009 "Too many events are opened.\n"
2010 "Try again after reducing the number of events.");
2011 case ENODEV:
2012 if (target->cpu_list)
2013 return scnprintf(msg, size, "%s",
2014 "No such device - did you specify an out-of-range profile CPU?\n");
2015 break;
2016 case EOPNOTSUPP:
2017 if (evsel->attr.precise_ip)
2018 return scnprintf(msg, size, "%s",
2019 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2020 #if defined(__i386__) || defined(__x86_64__)
2021 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2022 return scnprintf(msg, size, "%s",
2023 "No hardware sampling interrupt available.\n"
2024 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2025 #endif
2026 break;
2027 default:
2028 break;
2029 }
2030
2031 return scnprintf(msg, size,
2032 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2033 "/bin/dmesg may provide additional information.\n"
2034 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2035 err, strerror(err), perf_evsel__name(evsel));
2036 }
Ubuntu Artful kernel mirror