?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
fix a kmap leak in virtio_console
[linux/fpc-iii.git] / tools / perf / util / evsel.c
blob55407c594b87f1657e8ff06f8e64ee27ba2662c2
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 /*
504 * The enable_on_exec/disabled value strategy:
505 *
506 * 1) For any type of traced program:
507 * - all independent events and group leaders are disabled
508 * - all group members are enabled
509 *
510 * Group members are ruled by group leaders. They need to
511 * be enabled, because the group scheduling relies on that.
512 *
513 * 2) For traced programs executed by perf:
514 * - all independent events and group leaders have
515 * enable_on_exec set
516 * - we don't specifically enable or disable any event during
517 * the record command
518 *
519 * Independent events and group leaders are initially disabled
520 * and get enabled by exec. Group members are ruled by group
521 * leaders as stated in 1).
522 *
523 * 3) For traced programs attached by perf (pid/tid):
524 * - we specifically enable or disable all events during
525 * the record command
526 *
527 * When attaching events to already running traced we
528 * enable/disable events specifically, as there's no
529 * initial traced exec call.
530 */
531 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
532 {
533 struct perf_evsel *leader = evsel->leader;
534 struct perf_event_attr *attr = &evsel->attr;
535 int track = !evsel->idx; /* only the first counter needs these */
536 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
537
538 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
539 attr->inherit = !opts->no_inherit;
540
541 perf_evsel__set_sample_bit(evsel, IP);
542 perf_evsel__set_sample_bit(evsel, TID);
543
544 if (evsel->sample_read) {
545 perf_evsel__set_sample_bit(evsel, READ);
546
547 /*
548 * We need ID even in case of single event, because
549 * PERF_SAMPLE_READ process ID specific data.
550 */
551 perf_evsel__set_sample_id(evsel, false);
552
553 /*
554 * Apply group format only if we belong to group
555 * with more than one members.
556 */
557 if (leader->nr_members > 1) {
558 attr->read_format |= PERF_FORMAT_GROUP;
559 attr->inherit = 0;
560 }
561 }
562
563 /*
564 * We default some events to a 1 default interval. But keep
565 * it a weak assumption overridable by the user.
566 */
567 if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
568 opts->user_interval != ULLONG_MAX)) {
569 if (opts->freq) {
570 perf_evsel__set_sample_bit(evsel, PERIOD);
571 attr->freq = 1;
572 attr->sample_freq = opts->freq;
573 } else {
574 attr->sample_period = opts->default_interval;
575 }
576 }
577
578 /*
579 * Disable sampling for all group members other
580 * than leader in case leader 'leads' the sampling.
581 */
582 if ((leader != evsel) && leader->sample_read) {
583 attr->sample_freq = 0;
584 attr->sample_period = 0;
585 }
586
587 if (opts->no_samples)
588 attr->sample_freq = 0;
589
590 if (opts->inherit_stat)
591 attr->inherit_stat = 1;
592
593 if (opts->sample_address) {
594 perf_evsel__set_sample_bit(evsel, ADDR);
595 attr->mmap_data = track;
596 }
597
598 if (opts->call_graph) {
599 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
600
601 if (opts->call_graph == CALLCHAIN_DWARF) {
602 perf_evsel__set_sample_bit(evsel, REGS_USER);
603 perf_evsel__set_sample_bit(evsel, STACK_USER);
604 attr->sample_regs_user = PERF_REGS_MASK;
605 attr->sample_stack_user = opts->stack_dump_size;
606 attr->exclude_callchain_user = 1;
607 }
608 }
609
610 if (target__has_cpu(&opts->target))
611 perf_evsel__set_sample_bit(evsel, CPU);
612
613 if (opts->period)
614 perf_evsel__set_sample_bit(evsel, PERIOD);
615
616 if (!perf_missing_features.sample_id_all &&
617 (opts->sample_time || !opts->no_inherit ||
618 target__has_cpu(&opts->target) || per_cpu))
619 perf_evsel__set_sample_bit(evsel, TIME);
620
621 if (opts->raw_samples) {
622 perf_evsel__set_sample_bit(evsel, TIME);
623 perf_evsel__set_sample_bit(evsel, RAW);
624 perf_evsel__set_sample_bit(evsel, CPU);
625 }
626
627 if (opts->sample_address)
628 perf_evsel__set_sample_bit(evsel, DATA_SRC);
629
630 if (opts->no_buffering) {
631 attr->watermark = 0;
632 attr->wakeup_events = 1;
633 }
634 if (opts->branch_stack) {
635 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
636 attr->branch_sample_type = opts->branch_stack;
637 }
638
639 if (opts->sample_weight)
640 perf_evsel__set_sample_bit(evsel, WEIGHT);
641
642 attr->mmap = track;
643 attr->comm = track;
644
645 if (opts->sample_transaction)
646 perf_evsel__set_sample_bit(evsel, TRANSACTION);
647
648 /*
649 * XXX see the function comment above
650 *
651 * Disabling only independent events or group leaders,
652 * keeping group members enabled.
653 */
654 if (perf_evsel__is_group_leader(evsel))
655 attr->disabled = 1;
656
657 /*
658 * Setting enable_on_exec for independent events and
659 * group leaders for traced executed by perf.
660 */
661 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
662 !opts->initial_delay)
663 attr->enable_on_exec = 1;
664 }
665
666 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
667 {
668 int cpu, thread;
669 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
670
671 if (evsel->fd) {
672 for (cpu = 0; cpu < ncpus; cpu++) {
673 for (thread = 0; thread < nthreads; thread++) {
674 FD(evsel, cpu, thread) = -1;
675 }
676 }
677 }
678
679 return evsel->fd != NULL ? 0 : -ENOMEM;
680 }
681
682 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
683 int ioc, void *arg)
684 {
685 int cpu, thread;
686
687 for (cpu = 0; cpu < ncpus; cpu++) {
688 for (thread = 0; thread < nthreads; thread++) {
689 int fd = FD(evsel, cpu, thread),
690 err = ioctl(fd, ioc, arg);
691
692 if (err)
693 return err;
694 }
695 }
696
697 return 0;
698 }
699
700 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
701 const char *filter)
702 {
703 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
704 PERF_EVENT_IOC_SET_FILTER,
705 (void *)filter);
706 }
707
708 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
709 {
710 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
711 PERF_EVENT_IOC_ENABLE,
712 0);
713 }
714
715 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
716 {
717 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
718 if (evsel->sample_id == NULL)
719 return -ENOMEM;
720
721 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
722 if (evsel->id == NULL) {
723 xyarray__delete(evsel->sample_id);
724 evsel->sample_id = NULL;
725 return -ENOMEM;
726 }
727
728 return 0;
729 }
730
731 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
732 {
733 memset(evsel->counts, 0, (sizeof(*evsel->counts) +
734 (ncpus * sizeof(struct perf_counts_values))));
735 }
736
737 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
738 {
739 evsel->counts = zalloc((sizeof(*evsel->counts) +
740 (ncpus * sizeof(struct perf_counts_values))));
741 return evsel->counts != NULL ? 0 : -ENOMEM;
742 }
743
744 void perf_evsel__free_fd(struct perf_evsel *evsel)
745 {
746 xyarray__delete(evsel->fd);
747 evsel->fd = NULL;
748 }
749
750 void perf_evsel__free_id(struct perf_evsel *evsel)
751 {
752 xyarray__delete(evsel->sample_id);
753 evsel->sample_id = NULL;
754 zfree(&evsel->id);
755 }
756
757 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
758 {
759 int cpu, thread;
760
761 for (cpu = 0; cpu < ncpus; cpu++)
762 for (thread = 0; thread < nthreads; ++thread) {
763 close(FD(evsel, cpu, thread));
764 FD(evsel, cpu, thread) = -1;
765 }
766 }
767
768 void perf_evsel__free_counts(struct perf_evsel *evsel)
769 {
770 zfree(&evsel->counts);
771 }
772
773 void perf_evsel__exit(struct perf_evsel *evsel)
774 {
775 assert(list_empty(&evsel->node));
776 perf_evsel__free_fd(evsel);
777 perf_evsel__free_id(evsel);
778 }
779
780 void perf_evsel__delete(struct perf_evsel *evsel)
781 {
782 perf_evsel__exit(evsel);
783 close_cgroup(evsel->cgrp);
784 zfree(&evsel->group_name);
785 if (evsel->tp_format)
786 pevent_free_format(evsel->tp_format);
787 zfree(&evsel->name);
788 free(evsel);
789 }
790
791 static inline void compute_deltas(struct perf_evsel *evsel,
792 int cpu,
793 struct perf_counts_values *count)
794 {
795 struct perf_counts_values tmp;
796
797 if (!evsel->prev_raw_counts)
798 return;
799
800 if (cpu == -1) {
801 tmp = evsel->prev_raw_counts->aggr;
802 evsel->prev_raw_counts->aggr = *count;
803 } else {
804 tmp = evsel->prev_raw_counts->cpu[cpu];
805 evsel->prev_raw_counts->cpu[cpu] = *count;
806 }
807
808 count->val = count->val - tmp.val;
809 count->ena = count->ena - tmp.ena;
810 count->run = count->run - tmp.run;
811 }
812
813 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
814 int cpu, int thread, bool scale)
815 {
816 struct perf_counts_values count;
817 size_t nv = scale ? 3 : 1;
818
819 if (FD(evsel, cpu, thread) < 0)
820 return -EINVAL;
821
822 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
823 return -ENOMEM;
824
825 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
826 return -errno;
827
828 compute_deltas(evsel, cpu, &count);
829
830 if (scale) {
831 if (count.run == 0)
832 count.val = 0;
833 else if (count.run < count.ena)
834 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
835 } else
836 count.ena = count.run = 0;
837
838 evsel->counts->cpu[cpu] = count;
839 return 0;
840 }
841
842 int __perf_evsel__read(struct perf_evsel *evsel,
843 int ncpus, int nthreads, bool scale)
844 {
845 size_t nv = scale ? 3 : 1;
846 int cpu, thread;
847 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
848
849 aggr->val = aggr->ena = aggr->run = 0;
850
851 for (cpu = 0; cpu < ncpus; cpu++) {
852 for (thread = 0; thread < nthreads; thread++) {
853 if (FD(evsel, cpu, thread) < 0)
854 continue;
855
856 if (readn(FD(evsel, cpu, thread),
857 &count, nv * sizeof(u64)) < 0)
858 return -errno;
859
860 aggr->val += count.val;
861 if (scale) {
862 aggr->ena += count.ena;
863 aggr->run += count.run;
864 }
865 }
866 }
867
868 compute_deltas(evsel, -1, aggr);
869
870 evsel->counts->scaled = 0;
871 if (scale) {
872 if (aggr->run == 0) {
873 evsel->counts->scaled = -1;
874 aggr->val = 0;
875 return 0;
876 }
877
878 if (aggr->run < aggr->ena) {
879 evsel->counts->scaled = 1;
880 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
881 }
882 } else
883 aggr->ena = aggr->run = 0;
884
885 return 0;
886 }
887
888 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
889 {
890 struct perf_evsel *leader = evsel->leader;
891 int fd;
892
893 if (perf_evsel__is_group_leader(evsel))
894 return -1;
895
896 /*
897 * Leader must be already processed/open,
898 * if not it's a bug.
899 */
900 BUG_ON(!leader->fd);
901
902 fd = FD(leader, cpu, thread);
903 BUG_ON(fd == -1);
904
905 return fd;
906 }
907
908 #define __PRINT_ATTR(fmt, cast, field) \
909 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field)
910
911 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field)
912 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field)
913 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field)
914 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
915
916 #define PRINT_ATTR2N(name1, field1, name2, field2) \
917 fprintf(fp, " %-19s %u %-19s %u\n", \
918 name1, attr->field1, name2, attr->field2)
919
920 #define PRINT_ATTR2(field1, field2) \
921 PRINT_ATTR2N(#field1, field1, #field2, field2)
922
923 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
924 {
925 size_t ret = 0;
926
927 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
928 ret += fprintf(fp, "perf_event_attr:\n");
929
930 ret += PRINT_ATTR_U32(type);
931 ret += PRINT_ATTR_U32(size);
932 ret += PRINT_ATTR_X64(config);
933 ret += PRINT_ATTR_U64(sample_period);
934 ret += PRINT_ATTR_U64(sample_freq);
935 ret += PRINT_ATTR_X64(sample_type);
936 ret += PRINT_ATTR_X64(read_format);
937
938 ret += PRINT_ATTR2(disabled, inherit);
939 ret += PRINT_ATTR2(pinned, exclusive);
940 ret += PRINT_ATTR2(exclude_user, exclude_kernel);
941 ret += PRINT_ATTR2(exclude_hv, exclude_idle);
942 ret += PRINT_ATTR2(mmap, comm);
943 ret += PRINT_ATTR2(freq, inherit_stat);
944 ret += PRINT_ATTR2(enable_on_exec, task);
945 ret += PRINT_ATTR2(watermark, precise_ip);
946 ret += PRINT_ATTR2(mmap_data, sample_id_all);
947 ret += PRINT_ATTR2(exclude_host, exclude_guest);
948 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
949 "excl.callchain_user", exclude_callchain_user);
950 ret += PRINT_ATTR_U32(mmap2);
951
952 ret += PRINT_ATTR_U32(wakeup_events);
953 ret += PRINT_ATTR_U32(wakeup_watermark);
954 ret += PRINT_ATTR_X32(bp_type);
955 ret += PRINT_ATTR_X64(bp_addr);
956 ret += PRINT_ATTR_X64(config1);
957 ret += PRINT_ATTR_U64(bp_len);
958 ret += PRINT_ATTR_X64(config2);
959 ret += PRINT_ATTR_X64(branch_sample_type);
960 ret += PRINT_ATTR_X64(sample_regs_user);
961 ret += PRINT_ATTR_U32(sample_stack_user);
962
963 ret += fprintf(fp, "%.60s\n", graph_dotted_line);
964
965 return ret;
966 }
967
968 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
969 struct thread_map *threads)
970 {
971 int cpu, thread;
972 unsigned long flags = 0;
973 int pid = -1, err;
974 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
975
976 if (evsel->fd == NULL &&
977 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
978 return -ENOMEM;
979
980 if (evsel->cgrp) {
981 flags = PERF_FLAG_PID_CGROUP;
982 pid = evsel->cgrp->fd;
983 }
984
985 fallback_missing_features:
986 if (perf_missing_features.mmap2)
987 evsel->attr.mmap2 = 0;
988 if (perf_missing_features.exclude_guest)
989 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
990 retry_sample_id:
991 if (perf_missing_features.sample_id_all)
992 evsel->attr.sample_id_all = 0;
993
994 if (verbose >= 2)
995 perf_event_attr__fprintf(&evsel->attr, stderr);
996
997 for (cpu = 0; cpu < cpus->nr; cpu++) {
998
999 for (thread = 0; thread < threads->nr; thread++) {
1000 int group_fd;
1001
1002 if (!evsel->cgrp)
1003 pid = threads->map[thread];
1004
1005 group_fd = get_group_fd(evsel, cpu, thread);
1006 retry_open:
1007 pr_debug2("perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1008 pid, cpus->map[cpu], group_fd, flags);
1009
1010 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1011 pid,
1012 cpus->map[cpu],
1013 group_fd, flags);
1014 if (FD(evsel, cpu, thread) < 0) {
1015 err = -errno;
1016 pr_debug2("perf_event_open failed, error %d\n",
1017 err);
1018 goto try_fallback;
1019 }
1020 set_rlimit = NO_CHANGE;
1021 }
1022 }
1023
1024 return 0;
1025
1026 try_fallback:
1027 /*
1028 * perf stat needs between 5 and 22 fds per CPU. When we run out
1029 * of them try to increase the limits.
1030 */
1031 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1032 struct rlimit l;
1033 int old_errno = errno;
1034
1035 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1036 if (set_rlimit == NO_CHANGE)
1037 l.rlim_cur = l.rlim_max;
1038 else {
1039 l.rlim_cur = l.rlim_max + 1000;
1040 l.rlim_max = l.rlim_cur;
1041 }
1042 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1043 set_rlimit++;
1044 errno = old_errno;
1045 goto retry_open;
1046 }
1047 }
1048 errno = old_errno;
1049 }
1050
1051 if (err != -EINVAL || cpu > 0 || thread > 0)
1052 goto out_close;
1053
1054 if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1055 perf_missing_features.mmap2 = true;
1056 goto fallback_missing_features;
1057 } else if (!perf_missing_features.exclude_guest &&
1058 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1059 perf_missing_features.exclude_guest = true;
1060 goto fallback_missing_features;
1061 } else if (!perf_missing_features.sample_id_all) {
1062 perf_missing_features.sample_id_all = true;
1063 goto retry_sample_id;
1064 }
1065
1066 out_close:
1067 do {
1068 while (--thread >= 0) {
1069 close(FD(evsel, cpu, thread));
1070 FD(evsel, cpu, thread) = -1;
1071 }
1072 thread = threads->nr;
1073 } while (--cpu >= 0);
1074 return err;
1075 }
1076
1077 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1078 {
1079 if (evsel->fd == NULL)
1080 return;
1081
1082 perf_evsel__close_fd(evsel, ncpus, nthreads);
1083 perf_evsel__free_fd(evsel);
1084 }
1085
1086 static struct {
1087 struct cpu_map map;
1088 int cpus[1];
1089 } empty_cpu_map = {
1090 .map.nr = 1,
1091 .cpus = { -1, },
1092 };
1093
1094 static struct {
1095 struct thread_map map;
1096 int threads[1];
1097 } empty_thread_map = {
1098 .map.nr = 1,
1099 .threads = { -1, },
1100 };
1101
1102 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1103 struct thread_map *threads)
1104 {
1105 if (cpus == NULL) {
1106 /* Work around old compiler warnings about strict aliasing */
1107 cpus = &empty_cpu_map.map;
1108 }
1109
1110 if (threads == NULL)
1111 threads = &empty_thread_map.map;
1112
1113 return __perf_evsel__open(evsel, cpus, threads);
1114 }
1115
1116 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1117 struct cpu_map *cpus)
1118 {
1119 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1120 }
1121
1122 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1123 struct thread_map *threads)
1124 {
1125 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1126 }
1127
1128 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1129 const union perf_event *event,
1130 struct perf_sample *sample)
1131 {
1132 u64 type = evsel->attr.sample_type;
1133 const u64 *array = event->sample.array;
1134 bool swapped = evsel->needs_swap;
1135 union u64_swap u;
1136
1137 array += ((event->header.size -
1138 sizeof(event->header)) / sizeof(u64)) - 1;
1139
1140 if (type & PERF_SAMPLE_IDENTIFIER) {
1141 sample->id = *array;
1142 array--;
1143 }
1144
1145 if (type & PERF_SAMPLE_CPU) {
1146 u.val64 = *array;
1147 if (swapped) {
1148 /* undo swap of u64, then swap on individual u32s */
1149 u.val64 = bswap_64(u.val64);
1150 u.val32[0] = bswap_32(u.val32[0]);
1151 }
1152
1153 sample->cpu = u.val32[0];
1154 array--;
1155 }
1156
1157 if (type & PERF_SAMPLE_STREAM_ID) {
1158 sample->stream_id = *array;
1159 array--;
1160 }
1161
1162 if (type & PERF_SAMPLE_ID) {
1163 sample->id = *array;
1164 array--;
1165 }
1166
1167 if (type & PERF_SAMPLE_TIME) {
1168 sample->time = *array;
1169 array--;
1170 }
1171
1172 if (type & PERF_SAMPLE_TID) {
1173 u.val64 = *array;
1174 if (swapped) {
1175 /* undo swap of u64, then swap on individual u32s */
1176 u.val64 = bswap_64(u.val64);
1177 u.val32[0] = bswap_32(u.val32[0]);
1178 u.val32[1] = bswap_32(u.val32[1]);
1179 }
1180
1181 sample->pid = u.val32[0];
1182 sample->tid = u.val32[1];
1183 array--;
1184 }
1185
1186 return 0;
1187 }
1188
1189 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1190 u64 size)
1191 {
1192 return size > max_size || offset + size > endp;
1193 }
1194
1195 #define OVERFLOW_CHECK(offset, size, max_size) \
1196 do { \
1197 if (overflow(endp, (max_size), (offset), (size))) \
1198 return -EFAULT; \
1199 } while (0)
1200
1201 #define OVERFLOW_CHECK_u64(offset) \
1202 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1203
1204 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1205 struct perf_sample *data)
1206 {
1207 u64 type = evsel->attr.sample_type;
1208 bool swapped = evsel->needs_swap;
1209 const u64 *array;
1210 u16 max_size = event->header.size;
1211 const void *endp = (void *)event + max_size;
1212 u64 sz;
1213
1214 /*
1215 * used for cross-endian analysis. See git commit 65014ab3
1216 * for why this goofiness is needed.
1217 */
1218 union u64_swap u;
1219
1220 memset(data, 0, sizeof(*data));
1221 data->cpu = data->pid = data->tid = -1;
1222 data->stream_id = data->id = data->time = -1ULL;
1223 data->period = 1;
1224 data->weight = 0;
1225
1226 if (event->header.type != PERF_RECORD_SAMPLE) {
1227 if (!evsel->attr.sample_id_all)
1228 return 0;
1229 return perf_evsel__parse_id_sample(evsel, event, data);
1230 }
1231
1232 array = event->sample.array;
1233
1234 /*
1235 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1236 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1237 * check the format does not go past the end of the event.
1238 */
1239 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1240 return -EFAULT;
1241
1242 data->id = -1ULL;
1243 if (type & PERF_SAMPLE_IDENTIFIER) {
1244 data->id = *array;
1245 array++;
1246 }
1247
1248 if (type & PERF_SAMPLE_IP) {
1249 data->ip = *array;
1250 array++;
1251 }
1252
1253 if (type & PERF_SAMPLE_TID) {
1254 u.val64 = *array;
1255 if (swapped) {
1256 /* undo swap of u64, then swap on individual u32s */
1257 u.val64 = bswap_64(u.val64);
1258 u.val32[0] = bswap_32(u.val32[0]);
1259 u.val32[1] = bswap_32(u.val32[1]);
1260 }
1261
1262 data->pid = u.val32[0];
1263 data->tid = u.val32[1];
1264 array++;
1265 }
1266
1267 if (type & PERF_SAMPLE_TIME) {
1268 data->time = *array;
1269 array++;
1270 }
1271
1272 data->addr = 0;
1273 if (type & PERF_SAMPLE_ADDR) {
1274 data->addr = *array;
1275 array++;
1276 }
1277
1278 if (type & PERF_SAMPLE_ID) {
1279 data->id = *array;
1280 array++;
1281 }
1282
1283 if (type & PERF_SAMPLE_STREAM_ID) {
1284 data->stream_id = *array;
1285 array++;
1286 }
1287
1288 if (type & PERF_SAMPLE_CPU) {
1289
1290 u.val64 = *array;
1291 if (swapped) {
1292 /* undo swap of u64, then swap on individual u32s */
1293 u.val64 = bswap_64(u.val64);
1294 u.val32[0] = bswap_32(u.val32[0]);
1295 }
1296
1297 data->cpu = u.val32[0];
1298 array++;
1299 }
1300
1301 if (type & PERF_SAMPLE_PERIOD) {
1302 data->period = *array;
1303 array++;
1304 }
1305
1306 if (type & PERF_SAMPLE_READ) {
1307 u64 read_format = evsel->attr.read_format;
1308
1309 OVERFLOW_CHECK_u64(array);
1310 if (read_format & PERF_FORMAT_GROUP)
1311 data->read.group.nr = *array;
1312 else
1313 data->read.one.value = *array;
1314
1315 array++;
1316
1317 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1318 OVERFLOW_CHECK_u64(array);
1319 data->read.time_enabled = *array;
1320 array++;
1321 }
1322
1323 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1324 OVERFLOW_CHECK_u64(array);
1325 data->read.time_running = *array;
1326 array++;
1327 }
1328
1329 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1330 if (read_format & PERF_FORMAT_GROUP) {
1331 const u64 max_group_nr = UINT64_MAX /
1332 sizeof(struct sample_read_value);
1333
1334 if (data->read.group.nr > max_group_nr)
1335 return -EFAULT;
1336 sz = data->read.group.nr *
1337 sizeof(struct sample_read_value);
1338 OVERFLOW_CHECK(array, sz, max_size);
1339 data->read.group.values =
1340 (struct sample_read_value *)array;
1341 array = (void *)array + sz;
1342 } else {
1343 OVERFLOW_CHECK_u64(array);
1344 data->read.one.id = *array;
1345 array++;
1346 }
1347 }
1348
1349 if (type & PERF_SAMPLE_CALLCHAIN) {
1350 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1351
1352 OVERFLOW_CHECK_u64(array);
1353 data->callchain = (struct ip_callchain *)array++;
1354 if (data->callchain->nr > max_callchain_nr)
1355 return -EFAULT;
1356 sz = data->callchain->nr * sizeof(u64);
1357 OVERFLOW_CHECK(array, sz, max_size);
1358 array = (void *)array + sz;
1359 }
1360
1361 if (type & PERF_SAMPLE_RAW) {
1362 OVERFLOW_CHECK_u64(array);
1363 u.val64 = *array;
1364 if (WARN_ONCE(swapped,
1365 "Endianness of raw data not corrected!\n")) {
1366 /* undo swap of u64, then swap on individual u32s */
1367 u.val64 = bswap_64(u.val64);
1368 u.val32[0] = bswap_32(u.val32[0]);
1369 u.val32[1] = bswap_32(u.val32[1]);
1370 }
1371 data->raw_size = u.val32[0];
1372 array = (void *)array + sizeof(u32);
1373
1374 OVERFLOW_CHECK(array, data->raw_size, max_size);
1375 data->raw_data = (void *)array;
1376 array = (void *)array + data->raw_size;
1377 }
1378
1379 if (type & PERF_SAMPLE_BRANCH_STACK) {
1380 const u64 max_branch_nr = UINT64_MAX /
1381 sizeof(struct branch_entry);
1382
1383 OVERFLOW_CHECK_u64(array);
1384 data->branch_stack = (struct branch_stack *)array++;
1385
1386 if (data->branch_stack->nr > max_branch_nr)
1387 return -EFAULT;
1388 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1389 OVERFLOW_CHECK(array, sz, max_size);
1390 array = (void *)array + sz;
1391 }
1392
1393 if (type & PERF_SAMPLE_REGS_USER) {
1394 OVERFLOW_CHECK_u64(array);
1395 data->user_regs.abi = *array;
1396 array++;
1397
1398 if (data->user_regs.abi) {
1399 u64 regs_user = evsel->attr.sample_regs_user;
1400
1401 sz = hweight_long(regs_user) * sizeof(u64);
1402 OVERFLOW_CHECK(array, sz, max_size);
1403 data->user_regs.regs = (u64 *)array;
1404 array = (void *)array + sz;
1405 }
1406 }
1407
1408 if (type & PERF_SAMPLE_STACK_USER) {
1409 OVERFLOW_CHECK_u64(array);
1410 sz = *array++;
1411
1412 data->user_stack.offset = ((char *)(array - 1)
1413 - (char *) event);
1414
1415 if (!sz) {
1416 data->user_stack.size = 0;
1417 } else {
1418 OVERFLOW_CHECK(array, sz, max_size);
1419 data->user_stack.data = (char *)array;
1420 array = (void *)array + sz;
1421 OVERFLOW_CHECK_u64(array);
1422 data->user_stack.size = *array++;
1423 if (WARN_ONCE(data->user_stack.size > sz,
1424 "user stack dump failure\n"))
1425 return -EFAULT;
1426 }
1427 }
1428
1429 data->weight = 0;
1430 if (type & PERF_SAMPLE_WEIGHT) {
1431 OVERFLOW_CHECK_u64(array);
1432 data->weight = *array;
1433 array++;
1434 }
1435
1436 data->data_src = PERF_MEM_DATA_SRC_NONE;
1437 if (type & PERF_SAMPLE_DATA_SRC) {
1438 OVERFLOW_CHECK_u64(array);
1439 data->data_src = *array;
1440 array++;
1441 }
1442
1443 data->transaction = 0;
1444 if (type & PERF_SAMPLE_TRANSACTION) {
1445 OVERFLOW_CHECK_u64(array);
1446 data->transaction = *array;
1447 array++;
1448 }
1449
1450 return 0;
1451 }
1452
1453 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1454 u64 sample_regs_user, u64 read_format)
1455 {
1456 size_t sz, result = sizeof(struct sample_event);
1457
1458 if (type & PERF_SAMPLE_IDENTIFIER)
1459 result += sizeof(u64);
1460
1461 if (type & PERF_SAMPLE_IP)
1462 result += sizeof(u64);
1463
1464 if (type & PERF_SAMPLE_TID)
1465 result += sizeof(u64);
1466
1467 if (type & PERF_SAMPLE_TIME)
1468 result += sizeof(u64);
1469
1470 if (type & PERF_SAMPLE_ADDR)
1471 result += sizeof(u64);
1472
1473 if (type & PERF_SAMPLE_ID)
1474 result += sizeof(u64);
1475
1476 if (type & PERF_SAMPLE_STREAM_ID)
1477 result += sizeof(u64);
1478
1479 if (type & PERF_SAMPLE_CPU)
1480 result += sizeof(u64);
1481
1482 if (type & PERF_SAMPLE_PERIOD)
1483 result += sizeof(u64);
1484
1485 if (type & PERF_SAMPLE_READ) {
1486 result += sizeof(u64);
1487 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1488 result += sizeof(u64);
1489 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1490 result += sizeof(u64);
1491 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1492 if (read_format & PERF_FORMAT_GROUP) {
1493 sz = sample->read.group.nr *
1494 sizeof(struct sample_read_value);
1495 result += sz;
1496 } else {
1497 result += sizeof(u64);
1498 }
1499 }
1500
1501 if (type & PERF_SAMPLE_CALLCHAIN) {
1502 sz = (sample->callchain->nr + 1) * sizeof(u64);
1503 result += sz;
1504 }
1505
1506 if (type & PERF_SAMPLE_RAW) {
1507 result += sizeof(u32);
1508 result += sample->raw_size;
1509 }
1510
1511 if (type & PERF_SAMPLE_BRANCH_STACK) {
1512 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1513 sz += sizeof(u64);
1514 result += sz;
1515 }
1516
1517 if (type & PERF_SAMPLE_REGS_USER) {
1518 if (sample->user_regs.abi) {
1519 result += sizeof(u64);
1520 sz = hweight_long(sample_regs_user) * sizeof(u64);
1521 result += sz;
1522 } else {
1523 result += sizeof(u64);
1524 }
1525 }
1526
1527 if (type & PERF_SAMPLE_STACK_USER) {
1528 sz = sample->user_stack.size;
1529 result += sizeof(u64);
1530 if (sz) {
1531 result += sz;
1532 result += sizeof(u64);
1533 }
1534 }
1535
1536 if (type & PERF_SAMPLE_WEIGHT)
1537 result += sizeof(u64);
1538
1539 if (type & PERF_SAMPLE_DATA_SRC)
1540 result += sizeof(u64);
1541
1542 if (type & PERF_SAMPLE_TRANSACTION)
1543 result += sizeof(u64);
1544
1545 return result;
1546 }
1547
1548 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1549 u64 sample_regs_user, u64 read_format,
1550 const struct perf_sample *sample,
1551 bool swapped)
1552 {
1553 u64 *array;
1554 size_t sz;
1555 /*
1556 * used for cross-endian analysis. See git commit 65014ab3
1557 * for why this goofiness is needed.
1558 */
1559 union u64_swap u;
1560
1561 array = event->sample.array;
1562
1563 if (type & PERF_SAMPLE_IDENTIFIER) {
1564 *array = sample->id;
1565 array++;
1566 }
1567
1568 if (type & PERF_SAMPLE_IP) {
1569 *array = sample->ip;
1570 array++;
1571 }
1572
1573 if (type & PERF_SAMPLE_TID) {
1574 u.val32[0] = sample->pid;
1575 u.val32[1] = sample->tid;
1576 if (swapped) {
1577 /*
1578 * Inverse of what is done in perf_evsel__parse_sample
1579 */
1580 u.val32[0] = bswap_32(u.val32[0]);
1581 u.val32[1] = bswap_32(u.val32[1]);
1582 u.val64 = bswap_64(u.val64);
1583 }
1584
1585 *array = u.val64;
1586 array++;
1587 }
1588
1589 if (type & PERF_SAMPLE_TIME) {
1590 *array = sample->time;
1591 array++;
1592 }
1593
1594 if (type & PERF_SAMPLE_ADDR) {
1595 *array = sample->addr;
1596 array++;
1597 }
1598
1599 if (type & PERF_SAMPLE_ID) {
1600 *array = sample->id;
1601 array++;
1602 }
1603
1604 if (type & PERF_SAMPLE_STREAM_ID) {
1605 *array = sample->stream_id;
1606 array++;
1607 }
1608
1609 if (type & PERF_SAMPLE_CPU) {
1610 u.val32[0] = sample->cpu;
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.val64 = bswap_64(u.val64);
1617 }
1618 *array = u.val64;
1619 array++;
1620 }
1621
1622 if (type & PERF_SAMPLE_PERIOD) {
1623 *array = sample->period;
1624 array++;
1625 }
1626
1627 if (type & PERF_SAMPLE_READ) {
1628 if (read_format & PERF_FORMAT_GROUP)
1629 *array = sample->read.group.nr;
1630 else
1631 *array = sample->read.one.value;
1632 array++;
1633
1634 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1635 *array = sample->read.time_enabled;
1636 array++;
1637 }
1638
1639 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1640 *array = sample->read.time_running;
1641 array++;
1642 }
1643
1644 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1645 if (read_format & PERF_FORMAT_GROUP) {
1646 sz = sample->read.group.nr *
1647 sizeof(struct sample_read_value);
1648 memcpy(array, sample->read.group.values, sz);
1649 array = (void *)array + sz;
1650 } else {
1651 *array = sample->read.one.id;
1652 array++;
1653 }
1654 }
1655
1656 if (type & PERF_SAMPLE_CALLCHAIN) {
1657 sz = (sample->callchain->nr + 1) * sizeof(u64);
1658 memcpy(array, sample->callchain, sz);
1659 array = (void *)array + sz;
1660 }
1661
1662 if (type & PERF_SAMPLE_RAW) {
1663 u.val32[0] = sample->raw_size;
1664 if (WARN_ONCE(swapped,
1665 "Endianness of raw data not corrected!\n")) {
1666 /*
1667 * Inverse of what is done in perf_evsel__parse_sample
1668 */
1669 u.val32[0] = bswap_32(u.val32[0]);
1670 u.val32[1] = bswap_32(u.val32[1]);
1671 u.val64 = bswap_64(u.val64);
1672 }
1673 *array = u.val64;
1674 array = (void *)array + sizeof(u32);
1675
1676 memcpy(array, sample->raw_data, sample->raw_size);
1677 array = (void *)array + sample->raw_size;
1678 }
1679
1680 if (type & PERF_SAMPLE_BRANCH_STACK) {
1681 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1682 sz += sizeof(u64);
1683 memcpy(array, sample->branch_stack, sz);
1684 array = (void *)array + sz;
1685 }
1686
1687 if (type & PERF_SAMPLE_REGS_USER) {
1688 if (sample->user_regs.abi) {
1689 *array++ = sample->user_regs.abi;
1690 sz = hweight_long(sample_regs_user) * sizeof(u64);
1691 memcpy(array, sample->user_regs.regs, sz);
1692 array = (void *)array + sz;
1693 } else {
1694 *array++ = 0;
1695 }
1696 }
1697
1698 if (type & PERF_SAMPLE_STACK_USER) {
1699 sz = sample->user_stack.size;
1700 *array++ = sz;
1701 if (sz) {
1702 memcpy(array, sample->user_stack.data, sz);
1703 array = (void *)array + sz;
1704 *array++ = sz;
1705 }
1706 }
1707
1708 if (type & PERF_SAMPLE_WEIGHT) {
1709 *array = sample->weight;
1710 array++;
1711 }
1712
1713 if (type & PERF_SAMPLE_DATA_SRC) {
1714 *array = sample->data_src;
1715 array++;
1716 }
1717
1718 if (type & PERF_SAMPLE_TRANSACTION) {
1719 *array = sample->transaction;
1720 array++;
1721 }
1722
1723 return 0;
1724 }
1725
1726 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1727 {
1728 return pevent_find_field(evsel->tp_format, name);
1729 }
1730
1731 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1732 const char *name)
1733 {
1734 struct format_field *field = perf_evsel__field(evsel, name);
1735 int offset;
1736
1737 if (!field)
1738 return NULL;
1739
1740 offset = field->offset;
1741
1742 if (field->flags & FIELD_IS_DYNAMIC) {
1743 offset = *(int *)(sample->raw_data + field->offset);
1744 offset &= 0xffff;
1745 }
1746
1747 return sample->raw_data + offset;
1748 }
1749
1750 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1751 const char *name)
1752 {
1753 struct format_field *field = perf_evsel__field(evsel, name);
1754 void *ptr;
1755 u64 value;
1756
1757 if (!field)
1758 return 0;
1759
1760 ptr = sample->raw_data + field->offset;
1761
1762 switch (field->size) {
1763 case 1:
1764 return *(u8 *)ptr;
1765 case 2:
1766 value = *(u16 *)ptr;
1767 break;
1768 case 4:
1769 value = *(u32 *)ptr;
1770 break;
1771 case 8:
1772 value = *(u64 *)ptr;
1773 break;
1774 default:
1775 return 0;
1776 }
1777
1778 if (!evsel->needs_swap)
1779 return value;
1780
1781 switch (field->size) {
1782 case 2:
1783 return bswap_16(value);
1784 case 4:
1785 return bswap_32(value);
1786 case 8:
1787 return bswap_64(value);
1788 default:
1789 return 0;
1790 }
1791
1792 return 0;
1793 }
1794
1795 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1796 {
1797 va_list args;
1798 int ret = 0;
1799
1800 if (!*first) {
1801 ret += fprintf(fp, ",");
1802 } else {
1803 ret += fprintf(fp, ":");
1804 *first = false;
1805 }
1806
1807 va_start(args, fmt);
1808 ret += vfprintf(fp, fmt, args);
1809 va_end(args);
1810 return ret;
1811 }
1812
1813 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1814 {
1815 if (value == 0)
1816 return 0;
1817
1818 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1819 }
1820
1821 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1822
1823 struct bit_names {
1824 int bit;
1825 const char *name;
1826 };
1827
1828 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1829 struct bit_names *bits, bool *first)
1830 {
1831 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1832 bool first_bit = true;
1833
1834 do {
1835 if (value & bits[i].bit) {
1836 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1837 first_bit = false;
1838 }
1839 } while (bits[++i].name != NULL);
1840
1841 return printed;
1842 }
1843
1844 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1845 {
1846 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1847 struct bit_names bits[] = {
1848 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1849 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1850 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1851 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1852 bit_name(IDENTIFIER),
1853 { .name = NULL, }
1854 };
1855 #undef bit_name
1856 return bits__fprintf(fp, "sample_type", value, bits, first);
1857 }
1858
1859 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1860 {
1861 #define bit_name(n) { PERF_FORMAT_##n, #n }
1862 struct bit_names bits[] = {
1863 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1864 bit_name(ID), bit_name(GROUP),
1865 { .name = NULL, }
1866 };
1867 #undef bit_name
1868 return bits__fprintf(fp, "read_format", value, bits, first);
1869 }
1870
1871 int perf_evsel__fprintf(struct perf_evsel *evsel,
1872 struct perf_attr_details *details, FILE *fp)
1873 {
1874 bool first = true;
1875 int printed = 0;
1876
1877 if (details->event_group) {
1878 struct perf_evsel *pos;
1879
1880 if (!perf_evsel__is_group_leader(evsel))
1881 return 0;
1882
1883 if (evsel->nr_members > 1)
1884 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1885
1886 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1887 for_each_group_member(pos, evsel)
1888 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1889
1890 if (evsel->nr_members > 1)
1891 printed += fprintf(fp, "}");
1892 goto out;
1893 }
1894
1895 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1896
1897 if (details->verbose || details->freq) {
1898 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1899 (u64)evsel->attr.sample_freq);
1900 }
1901
1902 if (details->verbose) {
1903 if_print(type);
1904 if_print(config);
1905 if_print(config1);
1906 if_print(config2);
1907 if_print(size);
1908 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1909 if (evsel->attr.read_format)
1910 printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1911 if_print(disabled);
1912 if_print(inherit);
1913 if_print(pinned);
1914 if_print(exclusive);
1915 if_print(exclude_user);
1916 if_print(exclude_kernel);
1917 if_print(exclude_hv);
1918 if_print(exclude_idle);
1919 if_print(mmap);
1920 if_print(mmap2);
1921 if_print(comm);
1922 if_print(freq);
1923 if_print(inherit_stat);
1924 if_print(enable_on_exec);
1925 if_print(task);
1926 if_print(watermark);
1927 if_print(precise_ip);
1928 if_print(mmap_data);
1929 if_print(sample_id_all);
1930 if_print(exclude_host);
1931 if_print(exclude_guest);
1932 if_print(__reserved_1);
1933 if_print(wakeup_events);
1934 if_print(bp_type);
1935 if_print(branch_sample_type);
1936 }
1937 out:
1938 fputc('\n', fp);
1939 return ++printed;
1940 }
1941
1942 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1943 char *msg, size_t msgsize)
1944 {
1945 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
1946 evsel->attr.type == PERF_TYPE_HARDWARE &&
1947 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1948 /*
1949 * If it's cycles then fall back to hrtimer based
1950 * cpu-clock-tick sw counter, which is always available even if
1951 * no PMU support.
1952 *
1953 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1954 * b0a873e).
1955 */
1956 scnprintf(msg, msgsize, "%s",
1957 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1958
1959 evsel->attr.type = PERF_TYPE_SOFTWARE;
1960 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
1961
1962 zfree(&evsel->name);
1963 return true;
1964 }
1965
1966 return false;
1967 }
1968
1969 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
1970 int err, char *msg, size_t size)
1971 {
1972 switch (err) {
1973 case EPERM:
1974 case EACCES:
1975 return scnprintf(msg, size,
1976 "You may not have permission to collect %sstats.\n"
1977 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
1978 " -1 - Not paranoid at all\n"
1979 " 0 - Disallow raw tracepoint access for unpriv\n"
1980 " 1 - Disallow cpu events for unpriv\n"
1981 " 2 - Disallow kernel profiling for unpriv",
1982 target->system_wide ? "system-wide " : "");
1983 case ENOENT:
1984 return scnprintf(msg, size, "The %s event is not supported.",
1985 perf_evsel__name(evsel));
1986 case EMFILE:
1987 return scnprintf(msg, size, "%s",
1988 "Too many events are opened.\n"
1989 "Try again after reducing the number of events.");
1990 case ENODEV:
1991 if (target->cpu_list)
1992 return scnprintf(msg, size, "%s",
1993 "No such device - did you specify an out-of-range profile CPU?\n");
1994 break;
1995 case EOPNOTSUPP:
1996 if (evsel->attr.precise_ip)
1997 return scnprintf(msg, size, "%s",
1998 "\'precise\' request may not be supported. Try removing 'p' modifier.");
1999 #if defined(__i386__) || defined(__x86_64__)
2000 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2001 return scnprintf(msg, size, "%s",
2002 "No hardware sampling interrupt available.\n"
2003 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2004 #endif
2005 break;
2006 default:
2007 break;
2008 }
2009
2010 return scnprintf(msg, size,
2011 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s). \n"
2012 "/bin/dmesg may provide additional information.\n"
2013 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2014 err, strerror(err), perf_evsel__name(evsel));
2015 }
Linux with added FPC-III support