Loading...
1/*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9
10#include <byteswap.h>
11#include <errno.h>
12#include <inttypes.h>
13#include <linux/bitops.h>
14#include <api/fs/fs.h>
15#include <api/fs/tracing_path.h>
16#include <traceevent/event-parse.h>
17#include <linux/hw_breakpoint.h>
18#include <linux/perf_event.h>
19#include <linux/compiler.h>
20#include <linux/err.h>
21#include <sys/ioctl.h>
22#include <sys/resource.h>
23#include <sys/types.h>
24#include <dirent.h>
25#include "asm/bug.h"
26#include "callchain.h"
27#include "cgroup.h"
28#include "event.h"
29#include "evsel.h"
30#include "evlist.h"
31#include "util.h"
32#include "cpumap.h"
33#include "thread_map.h"
34#include "target.h"
35#include "perf_regs.h"
36#include "debug.h"
37#include "trace-event.h"
38#include "stat.h"
39#include "memswap.h"
40#include "util/parse-branch-options.h"
41
42#include "sane_ctype.h"
43
44struct perf_missing_features perf_missing_features;
45
46static clockid_t clockid;
47
48static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
49{
50 return 0;
51}
52
53void __weak test_attr__ready(void) { }
54
55static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
56{
57}
58
59static struct {
60 size_t size;
61 int (*init)(struct perf_evsel *evsel);
62 void (*fini)(struct perf_evsel *evsel);
63} perf_evsel__object = {
64 .size = sizeof(struct perf_evsel),
65 .init = perf_evsel__no_extra_init,
66 .fini = perf_evsel__no_extra_fini,
67};
68
69int perf_evsel__object_config(size_t object_size,
70 int (*init)(struct perf_evsel *evsel),
71 void (*fini)(struct perf_evsel *evsel))
72{
73
74 if (object_size == 0)
75 goto set_methods;
76
77 if (perf_evsel__object.size > object_size)
78 return -EINVAL;
79
80 perf_evsel__object.size = object_size;
81
82set_methods:
83 if (init != NULL)
84 perf_evsel__object.init = init;
85
86 if (fini != NULL)
87 perf_evsel__object.fini = fini;
88
89 return 0;
90}
91
92#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
93
94int __perf_evsel__sample_size(u64 sample_type)
95{
96 u64 mask = sample_type & PERF_SAMPLE_MASK;
97 int size = 0;
98 int i;
99
100 for (i = 0; i < 64; i++) {
101 if (mask & (1ULL << i))
102 size++;
103 }
104
105 size *= sizeof(u64);
106
107 return size;
108}
109
110/**
111 * __perf_evsel__calc_id_pos - calculate id_pos.
112 * @sample_type: sample type
113 *
114 * This function returns the position of the event id (PERF_SAMPLE_ID or
115 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
116 * sample_event.
117 */
118static int __perf_evsel__calc_id_pos(u64 sample_type)
119{
120 int idx = 0;
121
122 if (sample_type & PERF_SAMPLE_IDENTIFIER)
123 return 0;
124
125 if (!(sample_type & PERF_SAMPLE_ID))
126 return -1;
127
128 if (sample_type & PERF_SAMPLE_IP)
129 idx += 1;
130
131 if (sample_type & PERF_SAMPLE_TID)
132 idx += 1;
133
134 if (sample_type & PERF_SAMPLE_TIME)
135 idx += 1;
136
137 if (sample_type & PERF_SAMPLE_ADDR)
138 idx += 1;
139
140 return idx;
141}
142
143/**
144 * __perf_evsel__calc_is_pos - calculate is_pos.
145 * @sample_type: sample type
146 *
147 * This function returns the position (counting backwards) of the event id
148 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
149 * sample_id_all is used there is an id sample appended to non-sample events.
150 */
151static int __perf_evsel__calc_is_pos(u64 sample_type)
152{
153 int idx = 1;
154
155 if (sample_type & PERF_SAMPLE_IDENTIFIER)
156 return 1;
157
158 if (!(sample_type & PERF_SAMPLE_ID))
159 return -1;
160
161 if (sample_type & PERF_SAMPLE_CPU)
162 idx += 1;
163
164 if (sample_type & PERF_SAMPLE_STREAM_ID)
165 idx += 1;
166
167 return idx;
168}
169
170void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
171{
172 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
173 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
174}
175
176void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
177 enum perf_event_sample_format bit)
178{
179 if (!(evsel->attr.sample_type & bit)) {
180 evsel->attr.sample_type |= bit;
181 evsel->sample_size += sizeof(u64);
182 perf_evsel__calc_id_pos(evsel);
183 }
184}
185
186void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
187 enum perf_event_sample_format bit)
188{
189 if (evsel->attr.sample_type & bit) {
190 evsel->attr.sample_type &= ~bit;
191 evsel->sample_size -= sizeof(u64);
192 perf_evsel__calc_id_pos(evsel);
193 }
194}
195
196void perf_evsel__set_sample_id(struct perf_evsel *evsel,
197 bool can_sample_identifier)
198{
199 if (can_sample_identifier) {
200 perf_evsel__reset_sample_bit(evsel, ID);
201 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
202 } else {
203 perf_evsel__set_sample_bit(evsel, ID);
204 }
205 evsel->attr.read_format |= PERF_FORMAT_ID;
206}
207
208/**
209 * perf_evsel__is_function_event - Return whether given evsel is a function
210 * trace event
211 *
212 * @evsel - evsel selector to be tested
213 *
214 * Return %true if event is function trace event
215 */
216bool perf_evsel__is_function_event(struct perf_evsel *evsel)
217{
218#define FUNCTION_EVENT "ftrace:function"
219
220 return evsel->name &&
221 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
222
223#undef FUNCTION_EVENT
224}
225
226void perf_evsel__init(struct perf_evsel *evsel,
227 struct perf_event_attr *attr, int idx)
228{
229 evsel->idx = idx;
230 evsel->tracking = !idx;
231 evsel->attr = *attr;
232 evsel->leader = evsel;
233 evsel->unit = "";
234 evsel->scale = 1.0;
235 evsel->evlist = NULL;
236 evsel->bpf_fd = -1;
237 INIT_LIST_HEAD(&evsel->node);
238 INIT_LIST_HEAD(&evsel->config_terms);
239 perf_evsel__object.init(evsel);
240 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
241 perf_evsel__calc_id_pos(evsel);
242 evsel->cmdline_group_boundary = false;
243 evsel->metric_expr = NULL;
244 evsel->metric_name = NULL;
245 evsel->metric_events = NULL;
246 evsel->collect_stat = false;
247 evsel->pmu_name = NULL;
248}
249
250struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
251{
252 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
253
254 if (evsel != NULL)
255 perf_evsel__init(evsel, attr, idx);
256
257 if (perf_evsel__is_bpf_output(evsel)) {
258 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
259 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
260 evsel->attr.sample_period = 1;
261 }
262
263 return evsel;
264}
265
266static bool perf_event_can_profile_kernel(void)
267{
268 return geteuid() == 0 || perf_event_paranoid() == -1;
269}
270
271struct perf_evsel *perf_evsel__new_cycles(bool precise)
272{
273 struct perf_event_attr attr = {
274 .type = PERF_TYPE_HARDWARE,
275 .config = PERF_COUNT_HW_CPU_CYCLES,
276 .exclude_kernel = !perf_event_can_profile_kernel(),
277 };
278 struct perf_evsel *evsel;
279
280 event_attr_init(&attr);
281
282 if (!precise)
283 goto new_event;
284 /*
285 * Unnamed union member, not supported as struct member named
286 * initializer in older compilers such as gcc 4.4.7
287 *
288 * Just for probing the precise_ip:
289 */
290 attr.sample_period = 1;
291
292 perf_event_attr__set_max_precise_ip(&attr);
293 /*
294 * Now let the usual logic to set up the perf_event_attr defaults
295 * to kick in when we return and before perf_evsel__open() is called.
296 */
297 attr.sample_period = 0;
298new_event:
299 evsel = perf_evsel__new(&attr);
300 if (evsel == NULL)
301 goto out;
302
303 /* use asprintf() because free(evsel) assumes name is allocated */
304 if (asprintf(&evsel->name, "cycles%s%s%.*s",
305 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
306 attr.exclude_kernel ? "u" : "",
307 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
308 goto error_free;
309out:
310 return evsel;
311error_free:
312 perf_evsel__delete(evsel);
313 evsel = NULL;
314 goto out;
315}
316
317/*
318 * Returns pointer with encoded error via <linux/err.h> interface.
319 */
320struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
321{
322 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
323 int err = -ENOMEM;
324
325 if (evsel == NULL) {
326 goto out_err;
327 } else {
328 struct perf_event_attr attr = {
329 .type = PERF_TYPE_TRACEPOINT,
330 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
331 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
332 };
333
334 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
335 goto out_free;
336
337 evsel->tp_format = trace_event__tp_format(sys, name);
338 if (IS_ERR(evsel->tp_format)) {
339 err = PTR_ERR(evsel->tp_format);
340 goto out_free;
341 }
342
343 event_attr_init(&attr);
344 attr.config = evsel->tp_format->id;
345 attr.sample_period = 1;
346 perf_evsel__init(evsel, &attr, idx);
347 }
348
349 return evsel;
350
351out_free:
352 zfree(&evsel->name);
353 free(evsel);
354out_err:
355 return ERR_PTR(err);
356}
357
358const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
359 "cycles",
360 "instructions",
361 "cache-references",
362 "cache-misses",
363 "branches",
364 "branch-misses",
365 "bus-cycles",
366 "stalled-cycles-frontend",
367 "stalled-cycles-backend",
368 "ref-cycles",
369};
370
371static const char *__perf_evsel__hw_name(u64 config)
372{
373 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
374 return perf_evsel__hw_names[config];
375
376 return "unknown-hardware";
377}
378
379static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
380{
381 int colon = 0, r = 0;
382 struct perf_event_attr *attr = &evsel->attr;
383 bool exclude_guest_default = false;
384
385#define MOD_PRINT(context, mod) do { \
386 if (!attr->exclude_##context) { \
387 if (!colon) colon = ++r; \
388 r += scnprintf(bf + r, size - r, "%c", mod); \
389 } } while(0)
390
391 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
392 MOD_PRINT(kernel, 'k');
393 MOD_PRINT(user, 'u');
394 MOD_PRINT(hv, 'h');
395 exclude_guest_default = true;
396 }
397
398 if (attr->precise_ip) {
399 if (!colon)
400 colon = ++r;
401 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
402 exclude_guest_default = true;
403 }
404
405 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
406 MOD_PRINT(host, 'H');
407 MOD_PRINT(guest, 'G');
408 }
409#undef MOD_PRINT
410 if (colon)
411 bf[colon - 1] = ':';
412 return r;
413}
414
415static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
416{
417 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
418 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
419}
420
421const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
422 "cpu-clock",
423 "task-clock",
424 "page-faults",
425 "context-switches",
426 "cpu-migrations",
427 "minor-faults",
428 "major-faults",
429 "alignment-faults",
430 "emulation-faults",
431 "dummy",
432};
433
434static const char *__perf_evsel__sw_name(u64 config)
435{
436 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
437 return perf_evsel__sw_names[config];
438 return "unknown-software";
439}
440
441static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
442{
443 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
444 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
445}
446
447static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
448{
449 int r;
450
451 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
452
453 if (type & HW_BREAKPOINT_R)
454 r += scnprintf(bf + r, size - r, "r");
455
456 if (type & HW_BREAKPOINT_W)
457 r += scnprintf(bf + r, size - r, "w");
458
459 if (type & HW_BREAKPOINT_X)
460 r += scnprintf(bf + r, size - r, "x");
461
462 return r;
463}
464
465static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
466{
467 struct perf_event_attr *attr = &evsel->attr;
468 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
469 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
470}
471
472const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
473 [PERF_EVSEL__MAX_ALIASES] = {
474 { "L1-dcache", "l1-d", "l1d", "L1-data", },
475 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
476 { "LLC", "L2", },
477 { "dTLB", "d-tlb", "Data-TLB", },
478 { "iTLB", "i-tlb", "Instruction-TLB", },
479 { "branch", "branches", "bpu", "btb", "bpc", },
480 { "node", },
481};
482
483const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
484 [PERF_EVSEL__MAX_ALIASES] = {
485 { "load", "loads", "read", },
486 { "store", "stores", "write", },
487 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
488};
489
490const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
491 [PERF_EVSEL__MAX_ALIASES] = {
492 { "refs", "Reference", "ops", "access", },
493 { "misses", "miss", },
494};
495
496#define C(x) PERF_COUNT_HW_CACHE_##x
497#define CACHE_READ (1 << C(OP_READ))
498#define CACHE_WRITE (1 << C(OP_WRITE))
499#define CACHE_PREFETCH (1 << C(OP_PREFETCH))
500#define COP(x) (1 << x)
501
502/*
503 * cache operartion stat
504 * L1I : Read and prefetch only
505 * ITLB and BPU : Read-only
506 */
507static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
508 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
509 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
510 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
511 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
512 [C(ITLB)] = (CACHE_READ),
513 [C(BPU)] = (CACHE_READ),
514 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
515};
516
517bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
518{
519 if (perf_evsel__hw_cache_stat[type] & COP(op))
520 return true; /* valid */
521 else
522 return false; /* invalid */
523}
524
525int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
526 char *bf, size_t size)
527{
528 if (result) {
529 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
530 perf_evsel__hw_cache_op[op][0],
531 perf_evsel__hw_cache_result[result][0]);
532 }
533
534 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
535 perf_evsel__hw_cache_op[op][1]);
536}
537
538static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
539{
540 u8 op, result, type = (config >> 0) & 0xff;
541 const char *err = "unknown-ext-hardware-cache-type";
542
543 if (type >= PERF_COUNT_HW_CACHE_MAX)
544 goto out_err;
545
546 op = (config >> 8) & 0xff;
547 err = "unknown-ext-hardware-cache-op";
548 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
549 goto out_err;
550
551 result = (config >> 16) & 0xff;
552 err = "unknown-ext-hardware-cache-result";
553 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
554 goto out_err;
555
556 err = "invalid-cache";
557 if (!perf_evsel__is_cache_op_valid(type, op))
558 goto out_err;
559
560 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
561out_err:
562 return scnprintf(bf, size, "%s", err);
563}
564
565static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
566{
567 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
568 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
569}
570
571static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
572{
573 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
574 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
575}
576
577const char *perf_evsel__name(struct perf_evsel *evsel)
578{
579 char bf[128];
580
581 if (evsel->name)
582 return evsel->name;
583
584 switch (evsel->attr.type) {
585 case PERF_TYPE_RAW:
586 perf_evsel__raw_name(evsel, bf, sizeof(bf));
587 break;
588
589 case PERF_TYPE_HARDWARE:
590 perf_evsel__hw_name(evsel, bf, sizeof(bf));
591 break;
592
593 case PERF_TYPE_HW_CACHE:
594 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
595 break;
596
597 case PERF_TYPE_SOFTWARE:
598 perf_evsel__sw_name(evsel, bf, sizeof(bf));
599 break;
600
601 case PERF_TYPE_TRACEPOINT:
602 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
603 break;
604
605 case PERF_TYPE_BREAKPOINT:
606 perf_evsel__bp_name(evsel, bf, sizeof(bf));
607 break;
608
609 default:
610 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
611 evsel->attr.type);
612 break;
613 }
614
615 evsel->name = strdup(bf);
616
617 return evsel->name ?: "unknown";
618}
619
620const char *perf_evsel__group_name(struct perf_evsel *evsel)
621{
622 return evsel->group_name ?: "anon group";
623}
624
625/*
626 * Returns the group details for the specified leader,
627 * with following rules.
628 *
629 * For record -e '{cycles,instructions}'
630 * 'anon group { cycles:u, instructions:u }'
631 *
632 * For record -e 'cycles,instructions' and report --group
633 * 'cycles:u, instructions:u'
634 */
635int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
636{
637 int ret = 0;
638 struct perf_evsel *pos;
639 const char *group_name = perf_evsel__group_name(evsel);
640
641 if (!evsel->forced_leader)
642 ret = scnprintf(buf, size, "%s { ", group_name);
643
644 ret += scnprintf(buf + ret, size - ret, "%s",
645 perf_evsel__name(evsel));
646
647 for_each_group_member(pos, evsel)
648 ret += scnprintf(buf + ret, size - ret, ", %s",
649 perf_evsel__name(pos));
650
651 if (!evsel->forced_leader)
652 ret += scnprintf(buf + ret, size - ret, " }");
653
654 return ret;
655}
656
657static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
658 struct record_opts *opts,
659 struct callchain_param *param)
660{
661 bool function = perf_evsel__is_function_event(evsel);
662 struct perf_event_attr *attr = &evsel->attr;
663
664 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
665
666 attr->sample_max_stack = param->max_stack;
667
668 if (param->record_mode == CALLCHAIN_LBR) {
669 if (!opts->branch_stack) {
670 if (attr->exclude_user) {
671 pr_warning("LBR callstack option is only available "
672 "to get user callchain information. "
673 "Falling back to framepointers.\n");
674 } else {
675 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
676 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
677 PERF_SAMPLE_BRANCH_CALL_STACK |
678 PERF_SAMPLE_BRANCH_NO_CYCLES |
679 PERF_SAMPLE_BRANCH_NO_FLAGS;
680 }
681 } else
682 pr_warning("Cannot use LBR callstack with branch stack. "
683 "Falling back to framepointers.\n");
684 }
685
686 if (param->record_mode == CALLCHAIN_DWARF) {
687 if (!function) {
688 perf_evsel__set_sample_bit(evsel, REGS_USER);
689 perf_evsel__set_sample_bit(evsel, STACK_USER);
690 attr->sample_regs_user |= PERF_REGS_MASK;
691 attr->sample_stack_user = param->dump_size;
692 attr->exclude_callchain_user = 1;
693 } else {
694 pr_info("Cannot use DWARF unwind for function trace event,"
695 " falling back to framepointers.\n");
696 }
697 }
698
699 if (function) {
700 pr_info("Disabling user space callchains for function trace event.\n");
701 attr->exclude_callchain_user = 1;
702 }
703}
704
705void perf_evsel__config_callchain(struct perf_evsel *evsel,
706 struct record_opts *opts,
707 struct callchain_param *param)
708{
709 if (param->enabled)
710 return __perf_evsel__config_callchain(evsel, opts, param);
711}
712
713static void
714perf_evsel__reset_callgraph(struct perf_evsel *evsel,
715 struct callchain_param *param)
716{
717 struct perf_event_attr *attr = &evsel->attr;
718
719 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
720 if (param->record_mode == CALLCHAIN_LBR) {
721 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
722 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
723 PERF_SAMPLE_BRANCH_CALL_STACK);
724 }
725 if (param->record_mode == CALLCHAIN_DWARF) {
726 perf_evsel__reset_sample_bit(evsel, REGS_USER);
727 perf_evsel__reset_sample_bit(evsel, STACK_USER);
728 }
729}
730
731static void apply_config_terms(struct perf_evsel *evsel,
732 struct record_opts *opts, bool track)
733{
734 struct perf_evsel_config_term *term;
735 struct list_head *config_terms = &evsel->config_terms;
736 struct perf_event_attr *attr = &evsel->attr;
737 /* callgraph default */
738 struct callchain_param param = {
739 .record_mode = callchain_param.record_mode,
740 };
741 u32 dump_size = 0;
742 int max_stack = 0;
743 const char *callgraph_buf = NULL;
744
745 list_for_each_entry(term, config_terms, list) {
746 switch (term->type) {
747 case PERF_EVSEL__CONFIG_TERM_PERIOD:
748 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
749 attr->sample_period = term->val.period;
750 attr->freq = 0;
751 perf_evsel__reset_sample_bit(evsel, PERIOD);
752 }
753 break;
754 case PERF_EVSEL__CONFIG_TERM_FREQ:
755 if (!(term->weak && opts->user_freq != UINT_MAX)) {
756 attr->sample_freq = term->val.freq;
757 attr->freq = 1;
758 perf_evsel__set_sample_bit(evsel, PERIOD);
759 }
760 break;
761 case PERF_EVSEL__CONFIG_TERM_TIME:
762 if (term->val.time)
763 perf_evsel__set_sample_bit(evsel, TIME);
764 else
765 perf_evsel__reset_sample_bit(evsel, TIME);
766 break;
767 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
768 callgraph_buf = term->val.callgraph;
769 break;
770 case PERF_EVSEL__CONFIG_TERM_BRANCH:
771 if (term->val.branch && strcmp(term->val.branch, "no")) {
772 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
773 parse_branch_str(term->val.branch,
774 &attr->branch_sample_type);
775 } else
776 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
777 break;
778 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
779 dump_size = term->val.stack_user;
780 break;
781 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
782 max_stack = term->val.max_stack;
783 break;
784 case PERF_EVSEL__CONFIG_TERM_INHERIT:
785 /*
786 * attr->inherit should has already been set by
787 * perf_evsel__config. If user explicitly set
788 * inherit using config terms, override global
789 * opt->no_inherit setting.
790 */
791 attr->inherit = term->val.inherit ? 1 : 0;
792 break;
793 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
794 attr->write_backward = term->val.overwrite ? 1 : 0;
795 break;
796 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
797 break;
798 default:
799 break;
800 }
801 }
802
803 /* User explicitly set per-event callgraph, clear the old setting and reset. */
804 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
805 bool sample_address = false;
806
807 if (max_stack) {
808 param.max_stack = max_stack;
809 if (callgraph_buf == NULL)
810 callgraph_buf = "fp";
811 }
812
813 /* parse callgraph parameters */
814 if (callgraph_buf != NULL) {
815 if (!strcmp(callgraph_buf, "no")) {
816 param.enabled = false;
817 param.record_mode = CALLCHAIN_NONE;
818 } else {
819 param.enabled = true;
820 if (parse_callchain_record(callgraph_buf, ¶m)) {
821 pr_err("per-event callgraph setting for %s failed. "
822 "Apply callgraph global setting for it\n",
823 evsel->name);
824 return;
825 }
826 if (param.record_mode == CALLCHAIN_DWARF)
827 sample_address = true;
828 }
829 }
830 if (dump_size > 0) {
831 dump_size = round_up(dump_size, sizeof(u64));
832 param.dump_size = dump_size;
833 }
834
835 /* If global callgraph set, clear it */
836 if (callchain_param.enabled)
837 perf_evsel__reset_callgraph(evsel, &callchain_param);
838
839 /* set perf-event callgraph */
840 if (param.enabled) {
841 if (sample_address) {
842 perf_evsel__set_sample_bit(evsel, ADDR);
843 perf_evsel__set_sample_bit(evsel, DATA_SRC);
844 evsel->attr.mmap_data = track;
845 }
846 perf_evsel__config_callchain(evsel, opts, ¶m);
847 }
848 }
849}
850
851/*
852 * The enable_on_exec/disabled value strategy:
853 *
854 * 1) For any type of traced program:
855 * - all independent events and group leaders are disabled
856 * - all group members are enabled
857 *
858 * Group members are ruled by group leaders. They need to
859 * be enabled, because the group scheduling relies on that.
860 *
861 * 2) For traced programs executed by perf:
862 * - all independent events and group leaders have
863 * enable_on_exec set
864 * - we don't specifically enable or disable any event during
865 * the record command
866 *
867 * Independent events and group leaders are initially disabled
868 * and get enabled by exec. Group members are ruled by group
869 * leaders as stated in 1).
870 *
871 * 3) For traced programs attached by perf (pid/tid):
872 * - we specifically enable or disable all events during
873 * the record command
874 *
875 * When attaching events to already running traced we
876 * enable/disable events specifically, as there's no
877 * initial traced exec call.
878 */
879void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
880 struct callchain_param *callchain)
881{
882 struct perf_evsel *leader = evsel->leader;
883 struct perf_event_attr *attr = &evsel->attr;
884 int track = evsel->tracking;
885 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
886
887 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
888 attr->inherit = !opts->no_inherit;
889 attr->write_backward = opts->overwrite ? 1 : 0;
890
891 perf_evsel__set_sample_bit(evsel, IP);
892 perf_evsel__set_sample_bit(evsel, TID);
893
894 if (evsel->sample_read) {
895 perf_evsel__set_sample_bit(evsel, READ);
896
897 /*
898 * We need ID even in case of single event, because
899 * PERF_SAMPLE_READ process ID specific data.
900 */
901 perf_evsel__set_sample_id(evsel, false);
902
903 /*
904 * Apply group format only if we belong to group
905 * with more than one members.
906 */
907 if (leader->nr_members > 1) {
908 attr->read_format |= PERF_FORMAT_GROUP;
909 attr->inherit = 0;
910 }
911 }
912
913 /*
914 * We default some events to have a default interval. But keep
915 * it a weak assumption overridable by the user.
916 */
917 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
918 opts->user_interval != ULLONG_MAX)) {
919 if (opts->freq) {
920 perf_evsel__set_sample_bit(evsel, PERIOD);
921 attr->freq = 1;
922 attr->sample_freq = opts->freq;
923 } else {
924 attr->sample_period = opts->default_interval;
925 }
926 }
927
928 /*
929 * Disable sampling for all group members other
930 * than leader in case leader 'leads' the sampling.
931 */
932 if ((leader != evsel) && leader->sample_read) {
933 attr->freq = 0;
934 attr->sample_freq = 0;
935 attr->sample_period = 0;
936 attr->write_backward = 0;
937 attr->sample_id_all = 0;
938 }
939
940 if (opts->no_samples)
941 attr->sample_freq = 0;
942
943 if (opts->inherit_stat) {
944 evsel->attr.read_format |=
945 PERF_FORMAT_TOTAL_TIME_ENABLED |
946 PERF_FORMAT_TOTAL_TIME_RUNNING |
947 PERF_FORMAT_ID;
948 attr->inherit_stat = 1;
949 }
950
951 if (opts->sample_address) {
952 perf_evsel__set_sample_bit(evsel, ADDR);
953 attr->mmap_data = track;
954 }
955
956 /*
957 * We don't allow user space callchains for function trace
958 * event, due to issues with page faults while tracing page
959 * fault handler and its overall trickiness nature.
960 */
961 if (perf_evsel__is_function_event(evsel))
962 evsel->attr.exclude_callchain_user = 1;
963
964 if (callchain && callchain->enabled && !evsel->no_aux_samples)
965 perf_evsel__config_callchain(evsel, opts, callchain);
966
967 if (opts->sample_intr_regs) {
968 attr->sample_regs_intr = opts->sample_intr_regs;
969 perf_evsel__set_sample_bit(evsel, REGS_INTR);
970 }
971
972 if (opts->sample_user_regs) {
973 attr->sample_regs_user |= opts->sample_user_regs;
974 perf_evsel__set_sample_bit(evsel, REGS_USER);
975 }
976
977 if (target__has_cpu(&opts->target) || opts->sample_cpu)
978 perf_evsel__set_sample_bit(evsel, CPU);
979
980 /*
981 * When the user explicitly disabled time don't force it here.
982 */
983 if (opts->sample_time &&
984 (!perf_missing_features.sample_id_all &&
985 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
986 opts->sample_time_set)))
987 perf_evsel__set_sample_bit(evsel, TIME);
988
989 if (opts->raw_samples && !evsel->no_aux_samples) {
990 perf_evsel__set_sample_bit(evsel, TIME);
991 perf_evsel__set_sample_bit(evsel, RAW);
992 perf_evsel__set_sample_bit(evsel, CPU);
993 }
994
995 if (opts->sample_address)
996 perf_evsel__set_sample_bit(evsel, DATA_SRC);
997
998 if (opts->sample_phys_addr)
999 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1000
1001 if (opts->no_buffering) {
1002 attr->watermark = 0;
1003 attr->wakeup_events = 1;
1004 }
1005 if (opts->branch_stack && !evsel->no_aux_samples) {
1006 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1007 attr->branch_sample_type = opts->branch_stack;
1008 }
1009
1010 if (opts->sample_weight)
1011 perf_evsel__set_sample_bit(evsel, WEIGHT);
1012
1013 attr->task = track;
1014 attr->mmap = track;
1015 attr->mmap2 = track && !perf_missing_features.mmap2;
1016 attr->comm = track;
1017
1018 if (opts->record_namespaces)
1019 attr->namespaces = track;
1020
1021 if (opts->record_switch_events)
1022 attr->context_switch = track;
1023
1024 if (opts->sample_transaction)
1025 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1026
1027 if (opts->running_time) {
1028 evsel->attr.read_format |=
1029 PERF_FORMAT_TOTAL_TIME_ENABLED |
1030 PERF_FORMAT_TOTAL_TIME_RUNNING;
1031 }
1032
1033 /*
1034 * XXX see the function comment above
1035 *
1036 * Disabling only independent events or group leaders,
1037 * keeping group members enabled.
1038 */
1039 if (perf_evsel__is_group_leader(evsel))
1040 attr->disabled = 1;
1041
1042 /*
1043 * Setting enable_on_exec for independent events and
1044 * group leaders for traced executed by perf.
1045 */
1046 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1047 !opts->initial_delay)
1048 attr->enable_on_exec = 1;
1049
1050 if (evsel->immediate) {
1051 attr->disabled = 0;
1052 attr->enable_on_exec = 0;
1053 }
1054
1055 clockid = opts->clockid;
1056 if (opts->use_clockid) {
1057 attr->use_clockid = 1;
1058 attr->clockid = opts->clockid;
1059 }
1060
1061 if (evsel->precise_max)
1062 perf_event_attr__set_max_precise_ip(attr);
1063
1064 if (opts->all_user) {
1065 attr->exclude_kernel = 1;
1066 attr->exclude_user = 0;
1067 }
1068
1069 if (opts->all_kernel) {
1070 attr->exclude_kernel = 0;
1071 attr->exclude_user = 1;
1072 }
1073
1074 /*
1075 * Apply event specific term settings,
1076 * it overloads any global configuration.
1077 */
1078 apply_config_terms(evsel, opts, track);
1079
1080 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1081
1082 /* The --period option takes the precedence. */
1083 if (opts->period_set) {
1084 if (opts->period)
1085 perf_evsel__set_sample_bit(evsel, PERIOD);
1086 else
1087 perf_evsel__reset_sample_bit(evsel, PERIOD);
1088 }
1089}
1090
1091static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1092{
1093 if (evsel->system_wide)
1094 nthreads = 1;
1095
1096 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1097
1098 if (evsel->fd) {
1099 int cpu, thread;
1100 for (cpu = 0; cpu < ncpus; cpu++) {
1101 for (thread = 0; thread < nthreads; thread++) {
1102 FD(evsel, cpu, thread) = -1;
1103 }
1104 }
1105 }
1106
1107 return evsel->fd != NULL ? 0 : -ENOMEM;
1108}
1109
1110static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1111 int ioc, void *arg)
1112{
1113 int cpu, thread;
1114
1115 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1116 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1117 int fd = FD(evsel, cpu, thread),
1118 err = ioctl(fd, ioc, arg);
1119
1120 if (err)
1121 return err;
1122 }
1123 }
1124
1125 return 0;
1126}
1127
1128int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1129{
1130 return perf_evsel__run_ioctl(evsel,
1131 PERF_EVENT_IOC_SET_FILTER,
1132 (void *)filter);
1133}
1134
1135int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1136{
1137 char *new_filter = strdup(filter);
1138
1139 if (new_filter != NULL) {
1140 free(evsel->filter);
1141 evsel->filter = new_filter;
1142 return 0;
1143 }
1144
1145 return -1;
1146}
1147
1148static int perf_evsel__append_filter(struct perf_evsel *evsel,
1149 const char *fmt, const char *filter)
1150{
1151 char *new_filter;
1152
1153 if (evsel->filter == NULL)
1154 return perf_evsel__set_filter(evsel, filter);
1155
1156 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1157 free(evsel->filter);
1158 evsel->filter = new_filter;
1159 return 0;
1160 }
1161
1162 return -1;
1163}
1164
1165int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1166{
1167 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1168}
1169
1170int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1171{
1172 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1173}
1174
1175int perf_evsel__enable(struct perf_evsel *evsel)
1176{
1177 return perf_evsel__run_ioctl(evsel,
1178 PERF_EVENT_IOC_ENABLE,
1179 0);
1180}
1181
1182int perf_evsel__disable(struct perf_evsel *evsel)
1183{
1184 return perf_evsel__run_ioctl(evsel,
1185 PERF_EVENT_IOC_DISABLE,
1186 0);
1187}
1188
1189int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1190{
1191 if (ncpus == 0 || nthreads == 0)
1192 return 0;
1193
1194 if (evsel->system_wide)
1195 nthreads = 1;
1196
1197 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1198 if (evsel->sample_id == NULL)
1199 return -ENOMEM;
1200
1201 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1202 if (evsel->id == NULL) {
1203 xyarray__delete(evsel->sample_id);
1204 evsel->sample_id = NULL;
1205 return -ENOMEM;
1206 }
1207
1208 return 0;
1209}
1210
1211static void perf_evsel__free_fd(struct perf_evsel *evsel)
1212{
1213 xyarray__delete(evsel->fd);
1214 evsel->fd = NULL;
1215}
1216
1217static void perf_evsel__free_id(struct perf_evsel *evsel)
1218{
1219 xyarray__delete(evsel->sample_id);
1220 evsel->sample_id = NULL;
1221 zfree(&evsel->id);
1222}
1223
1224static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1225{
1226 struct perf_evsel_config_term *term, *h;
1227
1228 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1229 list_del(&term->list);
1230 free(term);
1231 }
1232}
1233
1234void perf_evsel__close_fd(struct perf_evsel *evsel)
1235{
1236 int cpu, thread;
1237
1238 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1239 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1240 close(FD(evsel, cpu, thread));
1241 FD(evsel, cpu, thread) = -1;
1242 }
1243}
1244
1245void perf_evsel__exit(struct perf_evsel *evsel)
1246{
1247 assert(list_empty(&evsel->node));
1248 assert(evsel->evlist == NULL);
1249 perf_evsel__free_fd(evsel);
1250 perf_evsel__free_id(evsel);
1251 perf_evsel__free_config_terms(evsel);
1252 cgroup__put(evsel->cgrp);
1253 cpu_map__put(evsel->cpus);
1254 cpu_map__put(evsel->own_cpus);
1255 thread_map__put(evsel->threads);
1256 zfree(&evsel->group_name);
1257 zfree(&evsel->name);
1258 perf_evsel__object.fini(evsel);
1259}
1260
1261void perf_evsel__delete(struct perf_evsel *evsel)
1262{
1263 perf_evsel__exit(evsel);
1264 free(evsel);
1265}
1266
1267void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1268 struct perf_counts_values *count)
1269{
1270 struct perf_counts_values tmp;
1271
1272 if (!evsel->prev_raw_counts)
1273 return;
1274
1275 if (cpu == -1) {
1276 tmp = evsel->prev_raw_counts->aggr;
1277 evsel->prev_raw_counts->aggr = *count;
1278 } else {
1279 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1280 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1281 }
1282
1283 count->val = count->val - tmp.val;
1284 count->ena = count->ena - tmp.ena;
1285 count->run = count->run - tmp.run;
1286}
1287
1288void perf_counts_values__scale(struct perf_counts_values *count,
1289 bool scale, s8 *pscaled)
1290{
1291 s8 scaled = 0;
1292
1293 if (scale) {
1294 if (count->run == 0) {
1295 scaled = -1;
1296 count->val = 0;
1297 } else if (count->run < count->ena) {
1298 scaled = 1;
1299 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1300 }
1301 } else
1302 count->ena = count->run = 0;
1303
1304 if (pscaled)
1305 *pscaled = scaled;
1306}
1307
1308static int perf_evsel__read_size(struct perf_evsel *evsel)
1309{
1310 u64 read_format = evsel->attr.read_format;
1311 int entry = sizeof(u64); /* value */
1312 int size = 0;
1313 int nr = 1;
1314
1315 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1316 size += sizeof(u64);
1317
1318 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1319 size += sizeof(u64);
1320
1321 if (read_format & PERF_FORMAT_ID)
1322 entry += sizeof(u64);
1323
1324 if (read_format & PERF_FORMAT_GROUP) {
1325 nr = evsel->nr_members;
1326 size += sizeof(u64);
1327 }
1328
1329 size += entry * nr;
1330 return size;
1331}
1332
1333int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1334 struct perf_counts_values *count)
1335{
1336 size_t size = perf_evsel__read_size(evsel);
1337
1338 memset(count, 0, sizeof(*count));
1339
1340 if (FD(evsel, cpu, thread) < 0)
1341 return -EINVAL;
1342
1343 if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1344 return -errno;
1345
1346 return 0;
1347}
1348
1349static int
1350perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1351{
1352 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1353
1354 return perf_evsel__read(evsel, cpu, thread, count);
1355}
1356
1357static void
1358perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1359 u64 val, u64 ena, u64 run)
1360{
1361 struct perf_counts_values *count;
1362
1363 count = perf_counts(counter->counts, cpu, thread);
1364
1365 count->val = val;
1366 count->ena = ena;
1367 count->run = run;
1368 count->loaded = true;
1369}
1370
1371static int
1372perf_evsel__process_group_data(struct perf_evsel *leader,
1373 int cpu, int thread, u64 *data)
1374{
1375 u64 read_format = leader->attr.read_format;
1376 struct sample_read_value *v;
1377 u64 nr, ena = 0, run = 0, i;
1378
1379 nr = *data++;
1380
1381 if (nr != (u64) leader->nr_members)
1382 return -EINVAL;
1383
1384 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1385 ena = *data++;
1386
1387 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1388 run = *data++;
1389
1390 v = (struct sample_read_value *) data;
1391
1392 perf_evsel__set_count(leader, cpu, thread,
1393 v[0].value, ena, run);
1394
1395 for (i = 1; i < nr; i++) {
1396 struct perf_evsel *counter;
1397
1398 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1399 if (!counter)
1400 return -EINVAL;
1401
1402 perf_evsel__set_count(counter, cpu, thread,
1403 v[i].value, ena, run);
1404 }
1405
1406 return 0;
1407}
1408
1409static int
1410perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1411{
1412 struct perf_stat_evsel *ps = leader->stats;
1413 u64 read_format = leader->attr.read_format;
1414 int size = perf_evsel__read_size(leader);
1415 u64 *data = ps->group_data;
1416
1417 if (!(read_format & PERF_FORMAT_ID))
1418 return -EINVAL;
1419
1420 if (!perf_evsel__is_group_leader(leader))
1421 return -EINVAL;
1422
1423 if (!data) {
1424 data = zalloc(size);
1425 if (!data)
1426 return -ENOMEM;
1427
1428 ps->group_data = data;
1429 }
1430
1431 if (FD(leader, cpu, thread) < 0)
1432 return -EINVAL;
1433
1434 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1435 return -errno;
1436
1437 return perf_evsel__process_group_data(leader, cpu, thread, data);
1438}
1439
1440int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1441{
1442 u64 read_format = evsel->attr.read_format;
1443
1444 if (read_format & PERF_FORMAT_GROUP)
1445 return perf_evsel__read_group(evsel, cpu, thread);
1446 else
1447 return perf_evsel__read_one(evsel, cpu, thread);
1448}
1449
1450int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1451 int cpu, int thread, bool scale)
1452{
1453 struct perf_counts_values count;
1454 size_t nv = scale ? 3 : 1;
1455
1456 if (FD(evsel, cpu, thread) < 0)
1457 return -EINVAL;
1458
1459 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1460 return -ENOMEM;
1461
1462 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1463 return -errno;
1464
1465 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1466 perf_counts_values__scale(&count, scale, NULL);
1467 *perf_counts(evsel->counts, cpu, thread) = count;
1468 return 0;
1469}
1470
1471static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1472{
1473 struct perf_evsel *leader = evsel->leader;
1474 int fd;
1475
1476 if (perf_evsel__is_group_leader(evsel))
1477 return -1;
1478
1479 /*
1480 * Leader must be already processed/open,
1481 * if not it's a bug.
1482 */
1483 BUG_ON(!leader->fd);
1484
1485 fd = FD(leader, cpu, thread);
1486 BUG_ON(fd == -1);
1487
1488 return fd;
1489}
1490
1491struct bit_names {
1492 int bit;
1493 const char *name;
1494};
1495
1496static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1497{
1498 bool first_bit = true;
1499 int i = 0;
1500
1501 do {
1502 if (value & bits[i].bit) {
1503 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1504 first_bit = false;
1505 }
1506 } while (bits[++i].name != NULL);
1507}
1508
1509static void __p_sample_type(char *buf, size_t size, u64 value)
1510{
1511#define bit_name(n) { PERF_SAMPLE_##n, #n }
1512 struct bit_names bits[] = {
1513 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1514 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1515 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1516 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1517 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1518 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1519 { .name = NULL, }
1520 };
1521#undef bit_name
1522 __p_bits(buf, size, value, bits);
1523}
1524
1525static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1526{
1527#define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1528 struct bit_names bits[] = {
1529 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1530 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1531 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1532 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1533 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1534 { .name = NULL, }
1535 };
1536#undef bit_name
1537 __p_bits(buf, size, value, bits);
1538}
1539
1540static void __p_read_format(char *buf, size_t size, u64 value)
1541{
1542#define bit_name(n) { PERF_FORMAT_##n, #n }
1543 struct bit_names bits[] = {
1544 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1545 bit_name(ID), bit_name(GROUP),
1546 { .name = NULL, }
1547 };
1548#undef bit_name
1549 __p_bits(buf, size, value, bits);
1550}
1551
1552#define BUF_SIZE 1024
1553
1554#define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1555#define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1556#define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1557#define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1558#define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1559#define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1560
1561#define PRINT_ATTRn(_n, _f, _p) \
1562do { \
1563 if (attr->_f) { \
1564 _p(attr->_f); \
1565 ret += attr__fprintf(fp, _n, buf, priv);\
1566 } \
1567} while (0)
1568
1569#define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1570
1571int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1572 attr__fprintf_f attr__fprintf, void *priv)
1573{
1574 char buf[BUF_SIZE];
1575 int ret = 0;
1576
1577 PRINT_ATTRf(type, p_unsigned);
1578 PRINT_ATTRf(size, p_unsigned);
1579 PRINT_ATTRf(config, p_hex);
1580 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1581 PRINT_ATTRf(sample_type, p_sample_type);
1582 PRINT_ATTRf(read_format, p_read_format);
1583
1584 PRINT_ATTRf(disabled, p_unsigned);
1585 PRINT_ATTRf(inherit, p_unsigned);
1586 PRINT_ATTRf(pinned, p_unsigned);
1587 PRINT_ATTRf(exclusive, p_unsigned);
1588 PRINT_ATTRf(exclude_user, p_unsigned);
1589 PRINT_ATTRf(exclude_kernel, p_unsigned);
1590 PRINT_ATTRf(exclude_hv, p_unsigned);
1591 PRINT_ATTRf(exclude_idle, p_unsigned);
1592 PRINT_ATTRf(mmap, p_unsigned);
1593 PRINT_ATTRf(comm, p_unsigned);
1594 PRINT_ATTRf(freq, p_unsigned);
1595 PRINT_ATTRf(inherit_stat, p_unsigned);
1596 PRINT_ATTRf(enable_on_exec, p_unsigned);
1597 PRINT_ATTRf(task, p_unsigned);
1598 PRINT_ATTRf(watermark, p_unsigned);
1599 PRINT_ATTRf(precise_ip, p_unsigned);
1600 PRINT_ATTRf(mmap_data, p_unsigned);
1601 PRINT_ATTRf(sample_id_all, p_unsigned);
1602 PRINT_ATTRf(exclude_host, p_unsigned);
1603 PRINT_ATTRf(exclude_guest, p_unsigned);
1604 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1605 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1606 PRINT_ATTRf(mmap2, p_unsigned);
1607 PRINT_ATTRf(comm_exec, p_unsigned);
1608 PRINT_ATTRf(use_clockid, p_unsigned);
1609 PRINT_ATTRf(context_switch, p_unsigned);
1610 PRINT_ATTRf(write_backward, p_unsigned);
1611 PRINT_ATTRf(namespaces, p_unsigned);
1612
1613 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1614 PRINT_ATTRf(bp_type, p_unsigned);
1615 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1616 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1617 PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1618 PRINT_ATTRf(sample_regs_user, p_hex);
1619 PRINT_ATTRf(sample_stack_user, p_unsigned);
1620 PRINT_ATTRf(clockid, p_signed);
1621 PRINT_ATTRf(sample_regs_intr, p_hex);
1622 PRINT_ATTRf(aux_watermark, p_unsigned);
1623 PRINT_ATTRf(sample_max_stack, p_unsigned);
1624
1625 return ret;
1626}
1627
1628static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1629 void *priv __maybe_unused)
1630{
1631 return fprintf(fp, " %-32s %s\n", name, val);
1632}
1633
1634static void perf_evsel__remove_fd(struct perf_evsel *pos,
1635 int nr_cpus, int nr_threads,
1636 int thread_idx)
1637{
1638 for (int cpu = 0; cpu < nr_cpus; cpu++)
1639 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1640 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1641}
1642
1643static int update_fds(struct perf_evsel *evsel,
1644 int nr_cpus, int cpu_idx,
1645 int nr_threads, int thread_idx)
1646{
1647 struct perf_evsel *pos;
1648
1649 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1650 return -EINVAL;
1651
1652 evlist__for_each_entry(evsel->evlist, pos) {
1653 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1654
1655 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1656
1657 /*
1658 * Since fds for next evsel has not been created,
1659 * there is no need to iterate whole event list.
1660 */
1661 if (pos == evsel)
1662 break;
1663 }
1664 return 0;
1665}
1666
1667static bool ignore_missing_thread(struct perf_evsel *evsel,
1668 int nr_cpus, int cpu,
1669 struct thread_map *threads,
1670 int thread, int err)
1671{
1672 pid_t ignore_pid = thread_map__pid(threads, thread);
1673
1674 if (!evsel->ignore_missing_thread)
1675 return false;
1676
1677 /* The system wide setup does not work with threads. */
1678 if (evsel->system_wide)
1679 return false;
1680
1681 /* The -ESRCH is perf event syscall errno for pid's not found. */
1682 if (err != -ESRCH)
1683 return false;
1684
1685 /* If there's only one thread, let it fail. */
1686 if (threads->nr == 1)
1687 return false;
1688
1689 /*
1690 * We should remove fd for missing_thread first
1691 * because thread_map__remove() will decrease threads->nr.
1692 */
1693 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1694 return false;
1695
1696 if (thread_map__remove(threads, thread))
1697 return false;
1698
1699 pr_warning("WARNING: Ignored open failure for pid %d\n",
1700 ignore_pid);
1701 return true;
1702}
1703
1704int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1705 struct thread_map *threads)
1706{
1707 int cpu, thread, nthreads;
1708 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1709 int pid = -1, err;
1710 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1711
1712 if (perf_missing_features.write_backward && evsel->attr.write_backward)
1713 return -EINVAL;
1714
1715 if (cpus == NULL) {
1716 static struct cpu_map *empty_cpu_map;
1717
1718 if (empty_cpu_map == NULL) {
1719 empty_cpu_map = cpu_map__dummy_new();
1720 if (empty_cpu_map == NULL)
1721 return -ENOMEM;
1722 }
1723
1724 cpus = empty_cpu_map;
1725 }
1726
1727 if (threads == NULL) {
1728 static struct thread_map *empty_thread_map;
1729
1730 if (empty_thread_map == NULL) {
1731 empty_thread_map = thread_map__new_by_tid(-1);
1732 if (empty_thread_map == NULL)
1733 return -ENOMEM;
1734 }
1735
1736 threads = empty_thread_map;
1737 }
1738
1739 if (evsel->system_wide)
1740 nthreads = 1;
1741 else
1742 nthreads = threads->nr;
1743
1744 if (evsel->fd == NULL &&
1745 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1746 return -ENOMEM;
1747
1748 if (evsel->cgrp) {
1749 flags |= PERF_FLAG_PID_CGROUP;
1750 pid = evsel->cgrp->fd;
1751 }
1752
1753fallback_missing_features:
1754 if (perf_missing_features.clockid_wrong)
1755 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1756 if (perf_missing_features.clockid) {
1757 evsel->attr.use_clockid = 0;
1758 evsel->attr.clockid = 0;
1759 }
1760 if (perf_missing_features.cloexec)
1761 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1762 if (perf_missing_features.mmap2)
1763 evsel->attr.mmap2 = 0;
1764 if (perf_missing_features.exclude_guest)
1765 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1766 if (perf_missing_features.lbr_flags)
1767 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1768 PERF_SAMPLE_BRANCH_NO_CYCLES);
1769 if (perf_missing_features.group_read && evsel->attr.inherit)
1770 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1771retry_sample_id:
1772 if (perf_missing_features.sample_id_all)
1773 evsel->attr.sample_id_all = 0;
1774
1775 if (verbose >= 2) {
1776 fprintf(stderr, "%.60s\n", graph_dotted_line);
1777 fprintf(stderr, "perf_event_attr:\n");
1778 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1779 fprintf(stderr, "%.60s\n", graph_dotted_line);
1780 }
1781
1782 for (cpu = 0; cpu < cpus->nr; cpu++) {
1783
1784 for (thread = 0; thread < nthreads; thread++) {
1785 int fd, group_fd;
1786
1787 if (!evsel->cgrp && !evsel->system_wide)
1788 pid = thread_map__pid(threads, thread);
1789
1790 group_fd = get_group_fd(evsel, cpu, thread);
1791retry_open:
1792 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1793 pid, cpus->map[cpu], group_fd, flags);
1794
1795 test_attr__ready();
1796
1797 fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1798 group_fd, flags);
1799
1800 FD(evsel, cpu, thread) = fd;
1801
1802 if (fd < 0) {
1803 err = -errno;
1804
1805 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1806 /*
1807 * We just removed 1 thread, so take a step
1808 * back on thread index and lower the upper
1809 * nthreads limit.
1810 */
1811 nthreads--;
1812 thread--;
1813
1814 /* ... and pretend like nothing have happened. */
1815 err = 0;
1816 continue;
1817 }
1818
1819 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1820 err);
1821 goto try_fallback;
1822 }
1823
1824 pr_debug2(" = %d\n", fd);
1825
1826 if (evsel->bpf_fd >= 0) {
1827 int evt_fd = fd;
1828 int bpf_fd = evsel->bpf_fd;
1829
1830 err = ioctl(evt_fd,
1831 PERF_EVENT_IOC_SET_BPF,
1832 bpf_fd);
1833 if (err && errno != EEXIST) {
1834 pr_err("failed to attach bpf fd %d: %s\n",
1835 bpf_fd, strerror(errno));
1836 err = -EINVAL;
1837 goto out_close;
1838 }
1839 }
1840
1841 set_rlimit = NO_CHANGE;
1842
1843 /*
1844 * If we succeeded but had to kill clockid, fail and
1845 * have perf_evsel__open_strerror() print us a nice
1846 * error.
1847 */
1848 if (perf_missing_features.clockid ||
1849 perf_missing_features.clockid_wrong) {
1850 err = -EINVAL;
1851 goto out_close;
1852 }
1853 }
1854 }
1855
1856 return 0;
1857
1858try_fallback:
1859 /*
1860 * perf stat needs between 5 and 22 fds per CPU. When we run out
1861 * of them try to increase the limits.
1862 */
1863 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1864 struct rlimit l;
1865 int old_errno = errno;
1866
1867 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1868 if (set_rlimit == NO_CHANGE)
1869 l.rlim_cur = l.rlim_max;
1870 else {
1871 l.rlim_cur = l.rlim_max + 1000;
1872 l.rlim_max = l.rlim_cur;
1873 }
1874 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1875 set_rlimit++;
1876 errno = old_errno;
1877 goto retry_open;
1878 }
1879 }
1880 errno = old_errno;
1881 }
1882
1883 if (err != -EINVAL || cpu > 0 || thread > 0)
1884 goto out_close;
1885
1886 /*
1887 * Must probe features in the order they were added to the
1888 * perf_event_attr interface.
1889 */
1890 if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1891 perf_missing_features.write_backward = true;
1892 pr_debug2("switching off write_backward\n");
1893 goto out_close;
1894 } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1895 perf_missing_features.clockid_wrong = true;
1896 pr_debug2("switching off clockid\n");
1897 goto fallback_missing_features;
1898 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1899 perf_missing_features.clockid = true;
1900 pr_debug2("switching off use_clockid\n");
1901 goto fallback_missing_features;
1902 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1903 perf_missing_features.cloexec = true;
1904 pr_debug2("switching off cloexec flag\n");
1905 goto fallback_missing_features;
1906 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1907 perf_missing_features.mmap2 = true;
1908 pr_debug2("switching off mmap2\n");
1909 goto fallback_missing_features;
1910 } else if (!perf_missing_features.exclude_guest &&
1911 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1912 perf_missing_features.exclude_guest = true;
1913 pr_debug2("switching off exclude_guest, exclude_host\n");
1914 goto fallback_missing_features;
1915 } else if (!perf_missing_features.sample_id_all) {
1916 perf_missing_features.sample_id_all = true;
1917 pr_debug2("switching off sample_id_all\n");
1918 goto retry_sample_id;
1919 } else if (!perf_missing_features.lbr_flags &&
1920 (evsel->attr.branch_sample_type &
1921 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1922 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1923 perf_missing_features.lbr_flags = true;
1924 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1925 goto fallback_missing_features;
1926 } else if (!perf_missing_features.group_read &&
1927 evsel->attr.inherit &&
1928 (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
1929 perf_evsel__is_group_leader(evsel)) {
1930 perf_missing_features.group_read = true;
1931 pr_debug2("switching off group read\n");
1932 goto fallback_missing_features;
1933 }
1934out_close:
1935 if (err)
1936 threads->err_thread = thread;
1937
1938 do {
1939 while (--thread >= 0) {
1940 close(FD(evsel, cpu, thread));
1941 FD(evsel, cpu, thread) = -1;
1942 }
1943 thread = nthreads;
1944 } while (--cpu >= 0);
1945 return err;
1946}
1947
1948void perf_evsel__close(struct perf_evsel *evsel)
1949{
1950 if (evsel->fd == NULL)
1951 return;
1952
1953 perf_evsel__close_fd(evsel);
1954 perf_evsel__free_fd(evsel);
1955}
1956
1957int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1958 struct cpu_map *cpus)
1959{
1960 return perf_evsel__open(evsel, cpus, NULL);
1961}
1962
1963int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1964 struct thread_map *threads)
1965{
1966 return perf_evsel__open(evsel, NULL, threads);
1967}
1968
1969static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1970 const union perf_event *event,
1971 struct perf_sample *sample)
1972{
1973 u64 type = evsel->attr.sample_type;
1974 const u64 *array = event->sample.array;
1975 bool swapped = evsel->needs_swap;
1976 union u64_swap u;
1977
1978 array += ((event->header.size -
1979 sizeof(event->header)) / sizeof(u64)) - 1;
1980
1981 if (type & PERF_SAMPLE_IDENTIFIER) {
1982 sample->id = *array;
1983 array--;
1984 }
1985
1986 if (type & PERF_SAMPLE_CPU) {
1987 u.val64 = *array;
1988 if (swapped) {
1989 /* undo swap of u64, then swap on individual u32s */
1990 u.val64 = bswap_64(u.val64);
1991 u.val32[0] = bswap_32(u.val32[0]);
1992 }
1993
1994 sample->cpu = u.val32[0];
1995 array--;
1996 }
1997
1998 if (type & PERF_SAMPLE_STREAM_ID) {
1999 sample->stream_id = *array;
2000 array--;
2001 }
2002
2003 if (type & PERF_SAMPLE_ID) {
2004 sample->id = *array;
2005 array--;
2006 }
2007
2008 if (type & PERF_SAMPLE_TIME) {
2009 sample->time = *array;
2010 array--;
2011 }
2012
2013 if (type & PERF_SAMPLE_TID) {
2014 u.val64 = *array;
2015 if (swapped) {
2016 /* undo swap of u64, then swap on individual u32s */
2017 u.val64 = bswap_64(u.val64);
2018 u.val32[0] = bswap_32(u.val32[0]);
2019 u.val32[1] = bswap_32(u.val32[1]);
2020 }
2021
2022 sample->pid = u.val32[0];
2023 sample->tid = u.val32[1];
2024 array--;
2025 }
2026
2027 return 0;
2028}
2029
2030static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2031 u64 size)
2032{
2033 return size > max_size || offset + size > endp;
2034}
2035
2036#define OVERFLOW_CHECK(offset, size, max_size) \
2037 do { \
2038 if (overflow(endp, (max_size), (offset), (size))) \
2039 return -EFAULT; \
2040 } while (0)
2041
2042#define OVERFLOW_CHECK_u64(offset) \
2043 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2044
2045static int
2046perf_event__check_size(union perf_event *event, unsigned int sample_size)
2047{
2048 /*
2049 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2050 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2051 * check the format does not go past the end of the event.
2052 */
2053 if (sample_size + sizeof(event->header) > event->header.size)
2054 return -EFAULT;
2055
2056 return 0;
2057}
2058
2059int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2060 struct perf_sample *data)
2061{
2062 u64 type = evsel->attr.sample_type;
2063 bool swapped = evsel->needs_swap;
2064 const u64 *array;
2065 u16 max_size = event->header.size;
2066 const void *endp = (void *)event + max_size;
2067 u64 sz;
2068
2069 /*
2070 * used for cross-endian analysis. See git commit 65014ab3
2071 * for why this goofiness is needed.
2072 */
2073 union u64_swap u;
2074
2075 memset(data, 0, sizeof(*data));
2076 data->cpu = data->pid = data->tid = -1;
2077 data->stream_id = data->id = data->time = -1ULL;
2078 data->period = evsel->attr.sample_period;
2079 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2080 data->misc = event->header.misc;
2081 data->id = -1ULL;
2082 data->data_src = PERF_MEM_DATA_SRC_NONE;
2083
2084 if (event->header.type != PERF_RECORD_SAMPLE) {
2085 if (!evsel->attr.sample_id_all)
2086 return 0;
2087 return perf_evsel__parse_id_sample(evsel, event, data);
2088 }
2089
2090 array = event->sample.array;
2091
2092 if (perf_event__check_size(event, evsel->sample_size))
2093 return -EFAULT;
2094
2095 if (type & PERF_SAMPLE_IDENTIFIER) {
2096 data->id = *array;
2097 array++;
2098 }
2099
2100 if (type & PERF_SAMPLE_IP) {
2101 data->ip = *array;
2102 array++;
2103 }
2104
2105 if (type & PERF_SAMPLE_TID) {
2106 u.val64 = *array;
2107 if (swapped) {
2108 /* undo swap of u64, then swap on individual u32s */
2109 u.val64 = bswap_64(u.val64);
2110 u.val32[0] = bswap_32(u.val32[0]);
2111 u.val32[1] = bswap_32(u.val32[1]);
2112 }
2113
2114 data->pid = u.val32[0];
2115 data->tid = u.val32[1];
2116 array++;
2117 }
2118
2119 if (type & PERF_SAMPLE_TIME) {
2120 data->time = *array;
2121 array++;
2122 }
2123
2124 if (type & PERF_SAMPLE_ADDR) {
2125 data->addr = *array;
2126 array++;
2127 }
2128
2129 if (type & PERF_SAMPLE_ID) {
2130 data->id = *array;
2131 array++;
2132 }
2133
2134 if (type & PERF_SAMPLE_STREAM_ID) {
2135 data->stream_id = *array;
2136 array++;
2137 }
2138
2139 if (type & PERF_SAMPLE_CPU) {
2140
2141 u.val64 = *array;
2142 if (swapped) {
2143 /* undo swap of u64, then swap on individual u32s */
2144 u.val64 = bswap_64(u.val64);
2145 u.val32[0] = bswap_32(u.val32[0]);
2146 }
2147
2148 data->cpu = u.val32[0];
2149 array++;
2150 }
2151
2152 if (type & PERF_SAMPLE_PERIOD) {
2153 data->period = *array;
2154 array++;
2155 }
2156
2157 if (type & PERF_SAMPLE_READ) {
2158 u64 read_format = evsel->attr.read_format;
2159
2160 OVERFLOW_CHECK_u64(array);
2161 if (read_format & PERF_FORMAT_GROUP)
2162 data->read.group.nr = *array;
2163 else
2164 data->read.one.value = *array;
2165
2166 array++;
2167
2168 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2169 OVERFLOW_CHECK_u64(array);
2170 data->read.time_enabled = *array;
2171 array++;
2172 }
2173
2174 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2175 OVERFLOW_CHECK_u64(array);
2176 data->read.time_running = *array;
2177 array++;
2178 }
2179
2180 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2181 if (read_format & PERF_FORMAT_GROUP) {
2182 const u64 max_group_nr = UINT64_MAX /
2183 sizeof(struct sample_read_value);
2184
2185 if (data->read.group.nr > max_group_nr)
2186 return -EFAULT;
2187 sz = data->read.group.nr *
2188 sizeof(struct sample_read_value);
2189 OVERFLOW_CHECK(array, sz, max_size);
2190 data->read.group.values =
2191 (struct sample_read_value *)array;
2192 array = (void *)array + sz;
2193 } else {
2194 OVERFLOW_CHECK_u64(array);
2195 data->read.one.id = *array;
2196 array++;
2197 }
2198 }
2199
2200 if (type & PERF_SAMPLE_CALLCHAIN) {
2201 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2202
2203 OVERFLOW_CHECK_u64(array);
2204 data->callchain = (struct ip_callchain *)array++;
2205 if (data->callchain->nr > max_callchain_nr)
2206 return -EFAULT;
2207 sz = data->callchain->nr * sizeof(u64);
2208 OVERFLOW_CHECK(array, sz, max_size);
2209 array = (void *)array + sz;
2210 }
2211
2212 if (type & PERF_SAMPLE_RAW) {
2213 OVERFLOW_CHECK_u64(array);
2214 u.val64 = *array;
2215
2216 /*
2217 * Undo swap of u64, then swap on individual u32s,
2218 * get the size of the raw area and undo all of the
2219 * swap. The pevent interface handles endianity by
2220 * itself.
2221 */
2222 if (swapped) {
2223 u.val64 = bswap_64(u.val64);
2224 u.val32[0] = bswap_32(u.val32[0]);
2225 u.val32[1] = bswap_32(u.val32[1]);
2226 }
2227 data->raw_size = u.val32[0];
2228
2229 /*
2230 * The raw data is aligned on 64bits including the
2231 * u32 size, so it's safe to use mem_bswap_64.
2232 */
2233 if (swapped)
2234 mem_bswap_64((void *) array, data->raw_size);
2235
2236 array = (void *)array + sizeof(u32);
2237
2238 OVERFLOW_CHECK(array, data->raw_size, max_size);
2239 data->raw_data = (void *)array;
2240 array = (void *)array + data->raw_size;
2241 }
2242
2243 if (type & PERF_SAMPLE_BRANCH_STACK) {
2244 const u64 max_branch_nr = UINT64_MAX /
2245 sizeof(struct branch_entry);
2246
2247 OVERFLOW_CHECK_u64(array);
2248 data->branch_stack = (struct branch_stack *)array++;
2249
2250 if (data->branch_stack->nr > max_branch_nr)
2251 return -EFAULT;
2252 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2253 OVERFLOW_CHECK(array, sz, max_size);
2254 array = (void *)array + sz;
2255 }
2256
2257 if (type & PERF_SAMPLE_REGS_USER) {
2258 OVERFLOW_CHECK_u64(array);
2259 data->user_regs.abi = *array;
2260 array++;
2261
2262 if (data->user_regs.abi) {
2263 u64 mask = evsel->attr.sample_regs_user;
2264
2265 sz = hweight_long(mask) * sizeof(u64);
2266 OVERFLOW_CHECK(array, sz, max_size);
2267 data->user_regs.mask = mask;
2268 data->user_regs.regs = (u64 *)array;
2269 array = (void *)array + sz;
2270 }
2271 }
2272
2273 if (type & PERF_SAMPLE_STACK_USER) {
2274 OVERFLOW_CHECK_u64(array);
2275 sz = *array++;
2276
2277 data->user_stack.offset = ((char *)(array - 1)
2278 - (char *) event);
2279
2280 if (!sz) {
2281 data->user_stack.size = 0;
2282 } else {
2283 OVERFLOW_CHECK(array, sz, max_size);
2284 data->user_stack.data = (char *)array;
2285 array = (void *)array + sz;
2286 OVERFLOW_CHECK_u64(array);
2287 data->user_stack.size = *array++;
2288 if (WARN_ONCE(data->user_stack.size > sz,
2289 "user stack dump failure\n"))
2290 return -EFAULT;
2291 }
2292 }
2293
2294 if (type & PERF_SAMPLE_WEIGHT) {
2295 OVERFLOW_CHECK_u64(array);
2296 data->weight = *array;
2297 array++;
2298 }
2299
2300 if (type & PERF_SAMPLE_DATA_SRC) {
2301 OVERFLOW_CHECK_u64(array);
2302 data->data_src = *array;
2303 array++;
2304 }
2305
2306 if (type & PERF_SAMPLE_TRANSACTION) {
2307 OVERFLOW_CHECK_u64(array);
2308 data->transaction = *array;
2309 array++;
2310 }
2311
2312 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2313 if (type & PERF_SAMPLE_REGS_INTR) {
2314 OVERFLOW_CHECK_u64(array);
2315 data->intr_regs.abi = *array;
2316 array++;
2317
2318 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2319 u64 mask = evsel->attr.sample_regs_intr;
2320
2321 sz = hweight_long(mask) * sizeof(u64);
2322 OVERFLOW_CHECK(array, sz, max_size);
2323 data->intr_regs.mask = mask;
2324 data->intr_regs.regs = (u64 *)array;
2325 array = (void *)array + sz;
2326 }
2327 }
2328
2329 data->phys_addr = 0;
2330 if (type & PERF_SAMPLE_PHYS_ADDR) {
2331 data->phys_addr = *array;
2332 array++;
2333 }
2334
2335 return 0;
2336}
2337
2338int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2339 union perf_event *event,
2340 u64 *timestamp)
2341{
2342 u64 type = evsel->attr.sample_type;
2343 const u64 *array;
2344
2345 if (!(type & PERF_SAMPLE_TIME))
2346 return -1;
2347
2348 if (event->header.type != PERF_RECORD_SAMPLE) {
2349 struct perf_sample data = {
2350 .time = -1ULL,
2351 };
2352
2353 if (!evsel->attr.sample_id_all)
2354 return -1;
2355 if (perf_evsel__parse_id_sample(evsel, event, &data))
2356 return -1;
2357
2358 *timestamp = data.time;
2359 return 0;
2360 }
2361
2362 array = event->sample.array;
2363
2364 if (perf_event__check_size(event, evsel->sample_size))
2365 return -EFAULT;
2366
2367 if (type & PERF_SAMPLE_IDENTIFIER)
2368 array++;
2369
2370 if (type & PERF_SAMPLE_IP)
2371 array++;
2372
2373 if (type & PERF_SAMPLE_TID)
2374 array++;
2375
2376 if (type & PERF_SAMPLE_TIME)
2377 *timestamp = *array;
2378
2379 return 0;
2380}
2381
2382size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2383 u64 read_format)
2384{
2385 size_t sz, result = sizeof(struct sample_event);
2386
2387 if (type & PERF_SAMPLE_IDENTIFIER)
2388 result += sizeof(u64);
2389
2390 if (type & PERF_SAMPLE_IP)
2391 result += sizeof(u64);
2392
2393 if (type & PERF_SAMPLE_TID)
2394 result += sizeof(u64);
2395
2396 if (type & PERF_SAMPLE_TIME)
2397 result += sizeof(u64);
2398
2399 if (type & PERF_SAMPLE_ADDR)
2400 result += sizeof(u64);
2401
2402 if (type & PERF_SAMPLE_ID)
2403 result += sizeof(u64);
2404
2405 if (type & PERF_SAMPLE_STREAM_ID)
2406 result += sizeof(u64);
2407
2408 if (type & PERF_SAMPLE_CPU)
2409 result += sizeof(u64);
2410
2411 if (type & PERF_SAMPLE_PERIOD)
2412 result += sizeof(u64);
2413
2414 if (type & PERF_SAMPLE_READ) {
2415 result += sizeof(u64);
2416 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2417 result += sizeof(u64);
2418 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2419 result += sizeof(u64);
2420 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2421 if (read_format & PERF_FORMAT_GROUP) {
2422 sz = sample->read.group.nr *
2423 sizeof(struct sample_read_value);
2424 result += sz;
2425 } else {
2426 result += sizeof(u64);
2427 }
2428 }
2429
2430 if (type & PERF_SAMPLE_CALLCHAIN) {
2431 sz = (sample->callchain->nr + 1) * sizeof(u64);
2432 result += sz;
2433 }
2434
2435 if (type & PERF_SAMPLE_RAW) {
2436 result += sizeof(u32);
2437 result += sample->raw_size;
2438 }
2439
2440 if (type & PERF_SAMPLE_BRANCH_STACK) {
2441 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2442 sz += sizeof(u64);
2443 result += sz;
2444 }
2445
2446 if (type & PERF_SAMPLE_REGS_USER) {
2447 if (sample->user_regs.abi) {
2448 result += sizeof(u64);
2449 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2450 result += sz;
2451 } else {
2452 result += sizeof(u64);
2453 }
2454 }
2455
2456 if (type & PERF_SAMPLE_STACK_USER) {
2457 sz = sample->user_stack.size;
2458 result += sizeof(u64);
2459 if (sz) {
2460 result += sz;
2461 result += sizeof(u64);
2462 }
2463 }
2464
2465 if (type & PERF_SAMPLE_WEIGHT)
2466 result += sizeof(u64);
2467
2468 if (type & PERF_SAMPLE_DATA_SRC)
2469 result += sizeof(u64);
2470
2471 if (type & PERF_SAMPLE_TRANSACTION)
2472 result += sizeof(u64);
2473
2474 if (type & PERF_SAMPLE_REGS_INTR) {
2475 if (sample->intr_regs.abi) {
2476 result += sizeof(u64);
2477 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2478 result += sz;
2479 } else {
2480 result += sizeof(u64);
2481 }
2482 }
2483
2484 if (type & PERF_SAMPLE_PHYS_ADDR)
2485 result += sizeof(u64);
2486
2487 return result;
2488}
2489
2490int perf_event__synthesize_sample(union perf_event *event, u64 type,
2491 u64 read_format,
2492 const struct perf_sample *sample)
2493{
2494 u64 *array;
2495 size_t sz;
2496 /*
2497 * used for cross-endian analysis. See git commit 65014ab3
2498 * for why this goofiness is needed.
2499 */
2500 union u64_swap u;
2501
2502 array = event->sample.array;
2503
2504 if (type & PERF_SAMPLE_IDENTIFIER) {
2505 *array = sample->id;
2506 array++;
2507 }
2508
2509 if (type & PERF_SAMPLE_IP) {
2510 *array = sample->ip;
2511 array++;
2512 }
2513
2514 if (type & PERF_SAMPLE_TID) {
2515 u.val32[0] = sample->pid;
2516 u.val32[1] = sample->tid;
2517 *array = u.val64;
2518 array++;
2519 }
2520
2521 if (type & PERF_SAMPLE_TIME) {
2522 *array = sample->time;
2523 array++;
2524 }
2525
2526 if (type & PERF_SAMPLE_ADDR) {
2527 *array = sample->addr;
2528 array++;
2529 }
2530
2531 if (type & PERF_SAMPLE_ID) {
2532 *array = sample->id;
2533 array++;
2534 }
2535
2536 if (type & PERF_SAMPLE_STREAM_ID) {
2537 *array = sample->stream_id;
2538 array++;
2539 }
2540
2541 if (type & PERF_SAMPLE_CPU) {
2542 u.val32[0] = sample->cpu;
2543 u.val32[1] = 0;
2544 *array = u.val64;
2545 array++;
2546 }
2547
2548 if (type & PERF_SAMPLE_PERIOD) {
2549 *array = sample->period;
2550 array++;
2551 }
2552
2553 if (type & PERF_SAMPLE_READ) {
2554 if (read_format & PERF_FORMAT_GROUP)
2555 *array = sample->read.group.nr;
2556 else
2557 *array = sample->read.one.value;
2558 array++;
2559
2560 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2561 *array = sample->read.time_enabled;
2562 array++;
2563 }
2564
2565 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2566 *array = sample->read.time_running;
2567 array++;
2568 }
2569
2570 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2571 if (read_format & PERF_FORMAT_GROUP) {
2572 sz = sample->read.group.nr *
2573 sizeof(struct sample_read_value);
2574 memcpy(array, sample->read.group.values, sz);
2575 array = (void *)array + sz;
2576 } else {
2577 *array = sample->read.one.id;
2578 array++;
2579 }
2580 }
2581
2582 if (type & PERF_SAMPLE_CALLCHAIN) {
2583 sz = (sample->callchain->nr + 1) * sizeof(u64);
2584 memcpy(array, sample->callchain, sz);
2585 array = (void *)array + sz;
2586 }
2587
2588 if (type & PERF_SAMPLE_RAW) {
2589 u.val32[0] = sample->raw_size;
2590 *array = u.val64;
2591 array = (void *)array + sizeof(u32);
2592
2593 memcpy(array, sample->raw_data, sample->raw_size);
2594 array = (void *)array + sample->raw_size;
2595 }
2596
2597 if (type & PERF_SAMPLE_BRANCH_STACK) {
2598 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2599 sz += sizeof(u64);
2600 memcpy(array, sample->branch_stack, sz);
2601 array = (void *)array + sz;
2602 }
2603
2604 if (type & PERF_SAMPLE_REGS_USER) {
2605 if (sample->user_regs.abi) {
2606 *array++ = sample->user_regs.abi;
2607 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2608 memcpy(array, sample->user_regs.regs, sz);
2609 array = (void *)array + sz;
2610 } else {
2611 *array++ = 0;
2612 }
2613 }
2614
2615 if (type & PERF_SAMPLE_STACK_USER) {
2616 sz = sample->user_stack.size;
2617 *array++ = sz;
2618 if (sz) {
2619 memcpy(array, sample->user_stack.data, sz);
2620 array = (void *)array + sz;
2621 *array++ = sz;
2622 }
2623 }
2624
2625 if (type & PERF_SAMPLE_WEIGHT) {
2626 *array = sample->weight;
2627 array++;
2628 }
2629
2630 if (type & PERF_SAMPLE_DATA_SRC) {
2631 *array = sample->data_src;
2632 array++;
2633 }
2634
2635 if (type & PERF_SAMPLE_TRANSACTION) {
2636 *array = sample->transaction;
2637 array++;
2638 }
2639
2640 if (type & PERF_SAMPLE_REGS_INTR) {
2641 if (sample->intr_regs.abi) {
2642 *array++ = sample->intr_regs.abi;
2643 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2644 memcpy(array, sample->intr_regs.regs, sz);
2645 array = (void *)array + sz;
2646 } else {
2647 *array++ = 0;
2648 }
2649 }
2650
2651 if (type & PERF_SAMPLE_PHYS_ADDR) {
2652 *array = sample->phys_addr;
2653 array++;
2654 }
2655
2656 return 0;
2657}
2658
2659struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2660{
2661 return pevent_find_field(evsel->tp_format, name);
2662}
2663
2664void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2665 const char *name)
2666{
2667 struct format_field *field = perf_evsel__field(evsel, name);
2668 int offset;
2669
2670 if (!field)
2671 return NULL;
2672
2673 offset = field->offset;
2674
2675 if (field->flags & FIELD_IS_DYNAMIC) {
2676 offset = *(int *)(sample->raw_data + field->offset);
2677 offset &= 0xffff;
2678 }
2679
2680 return sample->raw_data + offset;
2681}
2682
2683u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2684 bool needs_swap)
2685{
2686 u64 value;
2687 void *ptr = sample->raw_data + field->offset;
2688
2689 switch (field->size) {
2690 case 1:
2691 return *(u8 *)ptr;
2692 case 2:
2693 value = *(u16 *)ptr;
2694 break;
2695 case 4:
2696 value = *(u32 *)ptr;
2697 break;
2698 case 8:
2699 memcpy(&value, ptr, sizeof(u64));
2700 break;
2701 default:
2702 return 0;
2703 }
2704
2705 if (!needs_swap)
2706 return value;
2707
2708 switch (field->size) {
2709 case 2:
2710 return bswap_16(value);
2711 case 4:
2712 return bswap_32(value);
2713 case 8:
2714 return bswap_64(value);
2715 default:
2716 return 0;
2717 }
2718
2719 return 0;
2720}
2721
2722u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2723 const char *name)
2724{
2725 struct format_field *field = perf_evsel__field(evsel, name);
2726
2727 if (!field)
2728 return 0;
2729
2730 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2731}
2732
2733bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2734 char *msg, size_t msgsize)
2735{
2736 int paranoid;
2737
2738 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2739 evsel->attr.type == PERF_TYPE_HARDWARE &&
2740 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2741 /*
2742 * If it's cycles then fall back to hrtimer based
2743 * cpu-clock-tick sw counter, which is always available even if
2744 * no PMU support.
2745 *
2746 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2747 * b0a873e).
2748 */
2749 scnprintf(msg, msgsize, "%s",
2750"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2751
2752 evsel->attr.type = PERF_TYPE_SOFTWARE;
2753 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2754
2755 zfree(&evsel->name);
2756 return true;
2757 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2758 (paranoid = perf_event_paranoid()) > 1) {
2759 const char *name = perf_evsel__name(evsel);
2760 char *new_name;
2761 const char *sep = ":";
2762
2763 /* Is there already the separator in the name. */
2764 if (strchr(name, '/') ||
2765 strchr(name, ':'))
2766 sep = "";
2767
2768 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2769 return false;
2770
2771 if (evsel->name)
2772 free(evsel->name);
2773 evsel->name = new_name;
2774 scnprintf(msg, msgsize,
2775"kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2776 evsel->attr.exclude_kernel = 1;
2777
2778 return true;
2779 }
2780
2781 return false;
2782}
2783
2784static bool find_process(const char *name)
2785{
2786 size_t len = strlen(name);
2787 DIR *dir;
2788 struct dirent *d;
2789 int ret = -1;
2790
2791 dir = opendir(procfs__mountpoint());
2792 if (!dir)
2793 return false;
2794
2795 /* Walk through the directory. */
2796 while (ret && (d = readdir(dir)) != NULL) {
2797 char path[PATH_MAX];
2798 char *data;
2799 size_t size;
2800
2801 if ((d->d_type != DT_DIR) ||
2802 !strcmp(".", d->d_name) ||
2803 !strcmp("..", d->d_name))
2804 continue;
2805
2806 scnprintf(path, sizeof(path), "%s/%s/comm",
2807 procfs__mountpoint(), d->d_name);
2808
2809 if (filename__read_str(path, &data, &size))
2810 continue;
2811
2812 ret = strncmp(name, data, len);
2813 free(data);
2814 }
2815
2816 closedir(dir);
2817 return ret ? false : true;
2818}
2819
2820int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2821 int err, char *msg, size_t size)
2822{
2823 char sbuf[STRERR_BUFSIZE];
2824 int printed = 0;
2825
2826 switch (err) {
2827 case EPERM:
2828 case EACCES:
2829 if (err == EPERM)
2830 printed = scnprintf(msg, size,
2831 "No permission to enable %s event.\n\n",
2832 perf_evsel__name(evsel));
2833
2834 return scnprintf(msg + printed, size - printed,
2835 "You may not have permission to collect %sstats.\n\n"
2836 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2837 "which controls use of the performance events system by\n"
2838 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2839 "The current value is %d:\n\n"
2840 " -1: Allow use of (almost) all events by all users\n"
2841 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2842 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2843 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2844 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2845 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2846 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2847 " kernel.perf_event_paranoid = -1\n" ,
2848 target->system_wide ? "system-wide " : "",
2849 perf_event_paranoid());
2850 case ENOENT:
2851 return scnprintf(msg, size, "The %s event is not supported.",
2852 perf_evsel__name(evsel));
2853 case EMFILE:
2854 return scnprintf(msg, size, "%s",
2855 "Too many events are opened.\n"
2856 "Probably the maximum number of open file descriptors has been reached.\n"
2857 "Hint: Try again after reducing the number of events.\n"
2858 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2859 case ENOMEM:
2860 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2861 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2862 return scnprintf(msg, size,
2863 "Not enough memory to setup event with callchain.\n"
2864 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2865 "Hint: Current value: %d", sysctl_perf_event_max_stack);
2866 break;
2867 case ENODEV:
2868 if (target->cpu_list)
2869 return scnprintf(msg, size, "%s",
2870 "No such device - did you specify an out-of-range profile CPU?");
2871 break;
2872 case EOPNOTSUPP:
2873 if (evsel->attr.sample_period != 0)
2874 return scnprintf(msg, size,
2875 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2876 perf_evsel__name(evsel));
2877 if (evsel->attr.precise_ip)
2878 return scnprintf(msg, size, "%s",
2879 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2880#if defined(__i386__) || defined(__x86_64__)
2881 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2882 return scnprintf(msg, size, "%s",
2883 "No hardware sampling interrupt available.\n");
2884#endif
2885 break;
2886 case EBUSY:
2887 if (find_process("oprofiled"))
2888 return scnprintf(msg, size,
2889 "The PMU counters are busy/taken by another profiler.\n"
2890 "We found oprofile daemon running, please stop it and try again.");
2891 break;
2892 case EINVAL:
2893 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2894 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2895 if (perf_missing_features.clockid)
2896 return scnprintf(msg, size, "clockid feature not supported.");
2897 if (perf_missing_features.clockid_wrong)
2898 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2899 break;
2900 default:
2901 break;
2902 }
2903
2904 return scnprintf(msg, size,
2905 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2906 "/bin/dmesg | grep -i perf may provide additional information.\n",
2907 err, str_error_r(err, sbuf, sizeof(sbuf)),
2908 perf_evsel__name(evsel));
2909}
2910
2911struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2912{
2913 if (evsel && evsel->evlist)
2914 return evsel->evlist->env;
2915 return NULL;
2916}
1/*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9
10#include <byteswap.h>
11#include <linux/bitops.h>
12#include <api/fs/tracing_path.h>
13#include <traceevent/event-parse.h>
14#include <linux/hw_breakpoint.h>
15#include <linux/perf_event.h>
16#include <linux/err.h>
17#include <sys/resource.h>
18#include "asm/bug.h"
19#include "callchain.h"
20#include "cgroup.h"
21#include "evsel.h"
22#include "evlist.h"
23#include "util.h"
24#include "cpumap.h"
25#include "thread_map.h"
26#include "target.h"
27#include "perf_regs.h"
28#include "debug.h"
29#include "trace-event.h"
30#include "stat.h"
31#include "util/parse-branch-options.h"
32
33static struct {
34 bool sample_id_all;
35 bool exclude_guest;
36 bool mmap2;
37 bool cloexec;
38 bool clockid;
39 bool clockid_wrong;
40 bool lbr_flags;
41 bool write_backward;
42} perf_missing_features;
43
44static clockid_t clockid;
45
46static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
47{
48 return 0;
49}
50
51static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
52{
53}
54
55static struct {
56 size_t size;
57 int (*init)(struct perf_evsel *evsel);
58 void (*fini)(struct perf_evsel *evsel);
59} perf_evsel__object = {
60 .size = sizeof(struct perf_evsel),
61 .init = perf_evsel__no_extra_init,
62 .fini = perf_evsel__no_extra_fini,
63};
64
65int perf_evsel__object_config(size_t object_size,
66 int (*init)(struct perf_evsel *evsel),
67 void (*fini)(struct perf_evsel *evsel))
68{
69
70 if (object_size == 0)
71 goto set_methods;
72
73 if (perf_evsel__object.size > object_size)
74 return -EINVAL;
75
76 perf_evsel__object.size = object_size;
77
78set_methods:
79 if (init != NULL)
80 perf_evsel__object.init = init;
81
82 if (fini != NULL)
83 perf_evsel__object.fini = fini;
84
85 return 0;
86}
87
88#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
89
90int __perf_evsel__sample_size(u64 sample_type)
91{
92 u64 mask = sample_type & PERF_SAMPLE_MASK;
93 int size = 0;
94 int i;
95
96 for (i = 0; i < 64; i++) {
97 if (mask & (1ULL << i))
98 size++;
99 }
100
101 size *= sizeof(u64);
102
103 return size;
104}
105
106/**
107 * __perf_evsel__calc_id_pos - calculate id_pos.
108 * @sample_type: sample type
109 *
110 * This function returns the position of the event id (PERF_SAMPLE_ID or
111 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
112 * sample_event.
113 */
114static int __perf_evsel__calc_id_pos(u64 sample_type)
115{
116 int idx = 0;
117
118 if (sample_type & PERF_SAMPLE_IDENTIFIER)
119 return 0;
120
121 if (!(sample_type & PERF_SAMPLE_ID))
122 return -1;
123
124 if (sample_type & PERF_SAMPLE_IP)
125 idx += 1;
126
127 if (sample_type & PERF_SAMPLE_TID)
128 idx += 1;
129
130 if (sample_type & PERF_SAMPLE_TIME)
131 idx += 1;
132
133 if (sample_type & PERF_SAMPLE_ADDR)
134 idx += 1;
135
136 return idx;
137}
138
139/**
140 * __perf_evsel__calc_is_pos - calculate is_pos.
141 * @sample_type: sample type
142 *
143 * This function returns the position (counting backwards) of the event id
144 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
145 * sample_id_all is used there is an id sample appended to non-sample events.
146 */
147static int __perf_evsel__calc_is_pos(u64 sample_type)
148{
149 int idx = 1;
150
151 if (sample_type & PERF_SAMPLE_IDENTIFIER)
152 return 1;
153
154 if (!(sample_type & PERF_SAMPLE_ID))
155 return -1;
156
157 if (sample_type & PERF_SAMPLE_CPU)
158 idx += 1;
159
160 if (sample_type & PERF_SAMPLE_STREAM_ID)
161 idx += 1;
162
163 return idx;
164}
165
166void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
167{
168 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
169 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
170}
171
172void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
173 enum perf_event_sample_format bit)
174{
175 if (!(evsel->attr.sample_type & bit)) {
176 evsel->attr.sample_type |= bit;
177 evsel->sample_size += sizeof(u64);
178 perf_evsel__calc_id_pos(evsel);
179 }
180}
181
182void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
183 enum perf_event_sample_format bit)
184{
185 if (evsel->attr.sample_type & bit) {
186 evsel->attr.sample_type &= ~bit;
187 evsel->sample_size -= sizeof(u64);
188 perf_evsel__calc_id_pos(evsel);
189 }
190}
191
192void perf_evsel__set_sample_id(struct perf_evsel *evsel,
193 bool can_sample_identifier)
194{
195 if (can_sample_identifier) {
196 perf_evsel__reset_sample_bit(evsel, ID);
197 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
198 } else {
199 perf_evsel__set_sample_bit(evsel, ID);
200 }
201 evsel->attr.read_format |= PERF_FORMAT_ID;
202}
203
204/**
205 * perf_evsel__is_function_event - Return whether given evsel is a function
206 * trace event
207 *
208 * @evsel - evsel selector to be tested
209 *
210 * Return %true if event is function trace event
211 */
212bool perf_evsel__is_function_event(struct perf_evsel *evsel)
213{
214#define FUNCTION_EVENT "ftrace:function"
215
216 return evsel->name &&
217 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
218
219#undef FUNCTION_EVENT
220}
221
222void perf_evsel__init(struct perf_evsel *evsel,
223 struct perf_event_attr *attr, int idx)
224{
225 evsel->idx = idx;
226 evsel->tracking = !idx;
227 evsel->attr = *attr;
228 evsel->leader = evsel;
229 evsel->unit = "";
230 evsel->scale = 1.0;
231 evsel->evlist = NULL;
232 evsel->bpf_fd = -1;
233 INIT_LIST_HEAD(&evsel->node);
234 INIT_LIST_HEAD(&evsel->config_terms);
235 perf_evsel__object.init(evsel);
236 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
237 perf_evsel__calc_id_pos(evsel);
238 evsel->cmdline_group_boundary = false;
239}
240
241struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
242{
243 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
244
245 if (evsel != NULL)
246 perf_evsel__init(evsel, attr, idx);
247
248 if (perf_evsel__is_bpf_output(evsel)) {
249 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
250 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
251 evsel->attr.sample_period = 1;
252 }
253
254 return evsel;
255}
256
257struct perf_evsel *perf_evsel__new_cycles(void)
258{
259 struct perf_event_attr attr = {
260 .type = PERF_TYPE_HARDWARE,
261 .config = PERF_COUNT_HW_CPU_CYCLES,
262 };
263 struct perf_evsel *evsel;
264
265 event_attr_init(&attr);
266
267 perf_event_attr__set_max_precise_ip(&attr);
268
269 evsel = perf_evsel__new(&attr);
270 if (evsel == NULL)
271 goto out;
272
273 /* use asprintf() because free(evsel) assumes name is allocated */
274 if (asprintf(&evsel->name, "cycles%.*s",
275 attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
276 goto error_free;
277out:
278 return evsel;
279error_free:
280 perf_evsel__delete(evsel);
281 evsel = NULL;
282 goto out;
283}
284
285/*
286 * Returns pointer with encoded error via <linux/err.h> interface.
287 */
288struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
289{
290 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
291 int err = -ENOMEM;
292
293 if (evsel == NULL) {
294 goto out_err;
295 } else {
296 struct perf_event_attr attr = {
297 .type = PERF_TYPE_TRACEPOINT,
298 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
299 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
300 };
301
302 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
303 goto out_free;
304
305 evsel->tp_format = trace_event__tp_format(sys, name);
306 if (IS_ERR(evsel->tp_format)) {
307 err = PTR_ERR(evsel->tp_format);
308 goto out_free;
309 }
310
311 event_attr_init(&attr);
312 attr.config = evsel->tp_format->id;
313 attr.sample_period = 1;
314 perf_evsel__init(evsel, &attr, idx);
315 }
316
317 return evsel;
318
319out_free:
320 zfree(&evsel->name);
321 free(evsel);
322out_err:
323 return ERR_PTR(err);
324}
325
326const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
327 "cycles",
328 "instructions",
329 "cache-references",
330 "cache-misses",
331 "branches",
332 "branch-misses",
333 "bus-cycles",
334 "stalled-cycles-frontend",
335 "stalled-cycles-backend",
336 "ref-cycles",
337};
338
339static const char *__perf_evsel__hw_name(u64 config)
340{
341 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
342 return perf_evsel__hw_names[config];
343
344 return "unknown-hardware";
345}
346
347static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
348{
349 int colon = 0, r = 0;
350 struct perf_event_attr *attr = &evsel->attr;
351 bool exclude_guest_default = false;
352
353#define MOD_PRINT(context, mod) do { \
354 if (!attr->exclude_##context) { \
355 if (!colon) colon = ++r; \
356 r += scnprintf(bf + r, size - r, "%c", mod); \
357 } } while(0)
358
359 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
360 MOD_PRINT(kernel, 'k');
361 MOD_PRINT(user, 'u');
362 MOD_PRINT(hv, 'h');
363 exclude_guest_default = true;
364 }
365
366 if (attr->precise_ip) {
367 if (!colon)
368 colon = ++r;
369 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
370 exclude_guest_default = true;
371 }
372
373 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
374 MOD_PRINT(host, 'H');
375 MOD_PRINT(guest, 'G');
376 }
377#undef MOD_PRINT
378 if (colon)
379 bf[colon - 1] = ':';
380 return r;
381}
382
383static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
384{
385 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
386 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
387}
388
389const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
390 "cpu-clock",
391 "task-clock",
392 "page-faults",
393 "context-switches",
394 "cpu-migrations",
395 "minor-faults",
396 "major-faults",
397 "alignment-faults",
398 "emulation-faults",
399 "dummy",
400};
401
402static const char *__perf_evsel__sw_name(u64 config)
403{
404 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
405 return perf_evsel__sw_names[config];
406 return "unknown-software";
407}
408
409static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
410{
411 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
412 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
413}
414
415static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
416{
417 int r;
418
419 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
420
421 if (type & HW_BREAKPOINT_R)
422 r += scnprintf(bf + r, size - r, "r");
423
424 if (type & HW_BREAKPOINT_W)
425 r += scnprintf(bf + r, size - r, "w");
426
427 if (type & HW_BREAKPOINT_X)
428 r += scnprintf(bf + r, size - r, "x");
429
430 return r;
431}
432
433static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
434{
435 struct perf_event_attr *attr = &evsel->attr;
436 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
437 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
438}
439
440const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
441 [PERF_EVSEL__MAX_ALIASES] = {
442 { "L1-dcache", "l1-d", "l1d", "L1-data", },
443 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
444 { "LLC", "L2", },
445 { "dTLB", "d-tlb", "Data-TLB", },
446 { "iTLB", "i-tlb", "Instruction-TLB", },
447 { "branch", "branches", "bpu", "btb", "bpc", },
448 { "node", },
449};
450
451const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
452 [PERF_EVSEL__MAX_ALIASES] = {
453 { "load", "loads", "read", },
454 { "store", "stores", "write", },
455 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
456};
457
458const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
459 [PERF_EVSEL__MAX_ALIASES] = {
460 { "refs", "Reference", "ops", "access", },
461 { "misses", "miss", },
462};
463
464#define C(x) PERF_COUNT_HW_CACHE_##x
465#define CACHE_READ (1 << C(OP_READ))
466#define CACHE_WRITE (1 << C(OP_WRITE))
467#define CACHE_PREFETCH (1 << C(OP_PREFETCH))
468#define COP(x) (1 << x)
469
470/*
471 * cache operartion stat
472 * L1I : Read and prefetch only
473 * ITLB and BPU : Read-only
474 */
475static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
476 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
477 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
478 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
479 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
480 [C(ITLB)] = (CACHE_READ),
481 [C(BPU)] = (CACHE_READ),
482 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
483};
484
485bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
486{
487 if (perf_evsel__hw_cache_stat[type] & COP(op))
488 return true; /* valid */
489 else
490 return false; /* invalid */
491}
492
493int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
494 char *bf, size_t size)
495{
496 if (result) {
497 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
498 perf_evsel__hw_cache_op[op][0],
499 perf_evsel__hw_cache_result[result][0]);
500 }
501
502 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
503 perf_evsel__hw_cache_op[op][1]);
504}
505
506static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
507{
508 u8 op, result, type = (config >> 0) & 0xff;
509 const char *err = "unknown-ext-hardware-cache-type";
510
511 if (type >= PERF_COUNT_HW_CACHE_MAX)
512 goto out_err;
513
514 op = (config >> 8) & 0xff;
515 err = "unknown-ext-hardware-cache-op";
516 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
517 goto out_err;
518
519 result = (config >> 16) & 0xff;
520 err = "unknown-ext-hardware-cache-result";
521 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
522 goto out_err;
523
524 err = "invalid-cache";
525 if (!perf_evsel__is_cache_op_valid(type, op))
526 goto out_err;
527
528 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
529out_err:
530 return scnprintf(bf, size, "%s", err);
531}
532
533static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
534{
535 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
536 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
537}
538
539static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
540{
541 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
542 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
543}
544
545const char *perf_evsel__name(struct perf_evsel *evsel)
546{
547 char bf[128];
548
549 if (evsel->name)
550 return evsel->name;
551
552 switch (evsel->attr.type) {
553 case PERF_TYPE_RAW:
554 perf_evsel__raw_name(evsel, bf, sizeof(bf));
555 break;
556
557 case PERF_TYPE_HARDWARE:
558 perf_evsel__hw_name(evsel, bf, sizeof(bf));
559 break;
560
561 case PERF_TYPE_HW_CACHE:
562 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
563 break;
564
565 case PERF_TYPE_SOFTWARE:
566 perf_evsel__sw_name(evsel, bf, sizeof(bf));
567 break;
568
569 case PERF_TYPE_TRACEPOINT:
570 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
571 break;
572
573 case PERF_TYPE_BREAKPOINT:
574 perf_evsel__bp_name(evsel, bf, sizeof(bf));
575 break;
576
577 default:
578 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
579 evsel->attr.type);
580 break;
581 }
582
583 evsel->name = strdup(bf);
584
585 return evsel->name ?: "unknown";
586}
587
588const char *perf_evsel__group_name(struct perf_evsel *evsel)
589{
590 return evsel->group_name ?: "anon group";
591}
592
593int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
594{
595 int ret;
596 struct perf_evsel *pos;
597 const char *group_name = perf_evsel__group_name(evsel);
598
599 ret = scnprintf(buf, size, "%s", group_name);
600
601 ret += scnprintf(buf + ret, size - ret, " { %s",
602 perf_evsel__name(evsel));
603
604 for_each_group_member(pos, evsel)
605 ret += scnprintf(buf + ret, size - ret, ", %s",
606 perf_evsel__name(pos));
607
608 ret += scnprintf(buf + ret, size - ret, " }");
609
610 return ret;
611}
612
613void perf_evsel__config_callchain(struct perf_evsel *evsel,
614 struct record_opts *opts,
615 struct callchain_param *param)
616{
617 bool function = perf_evsel__is_function_event(evsel);
618 struct perf_event_attr *attr = &evsel->attr;
619
620 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
621
622 attr->sample_max_stack = param->max_stack;
623
624 if (param->record_mode == CALLCHAIN_LBR) {
625 if (!opts->branch_stack) {
626 if (attr->exclude_user) {
627 pr_warning("LBR callstack option is only available "
628 "to get user callchain information. "
629 "Falling back to framepointers.\n");
630 } else {
631 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
632 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
633 PERF_SAMPLE_BRANCH_CALL_STACK |
634 PERF_SAMPLE_BRANCH_NO_CYCLES |
635 PERF_SAMPLE_BRANCH_NO_FLAGS;
636 }
637 } else
638 pr_warning("Cannot use LBR callstack with branch stack. "
639 "Falling back to framepointers.\n");
640 }
641
642 if (param->record_mode == CALLCHAIN_DWARF) {
643 if (!function) {
644 perf_evsel__set_sample_bit(evsel, REGS_USER);
645 perf_evsel__set_sample_bit(evsel, STACK_USER);
646 attr->sample_regs_user = PERF_REGS_MASK;
647 attr->sample_stack_user = param->dump_size;
648 attr->exclude_callchain_user = 1;
649 } else {
650 pr_info("Cannot use DWARF unwind for function trace event,"
651 " falling back to framepointers.\n");
652 }
653 }
654
655 if (function) {
656 pr_info("Disabling user space callchains for function trace event.\n");
657 attr->exclude_callchain_user = 1;
658 }
659}
660
661static void
662perf_evsel__reset_callgraph(struct perf_evsel *evsel,
663 struct callchain_param *param)
664{
665 struct perf_event_attr *attr = &evsel->attr;
666
667 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
668 if (param->record_mode == CALLCHAIN_LBR) {
669 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
670 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
671 PERF_SAMPLE_BRANCH_CALL_STACK);
672 }
673 if (param->record_mode == CALLCHAIN_DWARF) {
674 perf_evsel__reset_sample_bit(evsel, REGS_USER);
675 perf_evsel__reset_sample_bit(evsel, STACK_USER);
676 }
677}
678
679static void apply_config_terms(struct perf_evsel *evsel,
680 struct record_opts *opts)
681{
682 struct perf_evsel_config_term *term;
683 struct list_head *config_terms = &evsel->config_terms;
684 struct perf_event_attr *attr = &evsel->attr;
685 struct callchain_param param;
686 u32 dump_size = 0;
687 int max_stack = 0;
688 const char *callgraph_buf = NULL;
689
690 /* callgraph default */
691 param.record_mode = callchain_param.record_mode;
692
693 list_for_each_entry(term, config_terms, list) {
694 switch (term->type) {
695 case PERF_EVSEL__CONFIG_TERM_PERIOD:
696 attr->sample_period = term->val.period;
697 attr->freq = 0;
698 break;
699 case PERF_EVSEL__CONFIG_TERM_FREQ:
700 attr->sample_freq = term->val.freq;
701 attr->freq = 1;
702 break;
703 case PERF_EVSEL__CONFIG_TERM_TIME:
704 if (term->val.time)
705 perf_evsel__set_sample_bit(evsel, TIME);
706 else
707 perf_evsel__reset_sample_bit(evsel, TIME);
708 break;
709 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
710 callgraph_buf = term->val.callgraph;
711 break;
712 case PERF_EVSEL__CONFIG_TERM_BRANCH:
713 if (term->val.branch && strcmp(term->val.branch, "no")) {
714 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
715 parse_branch_str(term->val.branch,
716 &attr->branch_sample_type);
717 } else
718 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
719 break;
720 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
721 dump_size = term->val.stack_user;
722 break;
723 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
724 max_stack = term->val.max_stack;
725 break;
726 case PERF_EVSEL__CONFIG_TERM_INHERIT:
727 /*
728 * attr->inherit should has already been set by
729 * perf_evsel__config. If user explicitly set
730 * inherit using config terms, override global
731 * opt->no_inherit setting.
732 */
733 attr->inherit = term->val.inherit ? 1 : 0;
734 break;
735 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
736 attr->write_backward = term->val.overwrite ? 1 : 0;
737 break;
738 default:
739 break;
740 }
741 }
742
743 /* User explicitly set per-event callgraph, clear the old setting and reset. */
744 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
745 if (max_stack) {
746 param.max_stack = max_stack;
747 if (callgraph_buf == NULL)
748 callgraph_buf = "fp";
749 }
750
751 /* parse callgraph parameters */
752 if (callgraph_buf != NULL) {
753 if (!strcmp(callgraph_buf, "no")) {
754 param.enabled = false;
755 param.record_mode = CALLCHAIN_NONE;
756 } else {
757 param.enabled = true;
758 if (parse_callchain_record(callgraph_buf, ¶m)) {
759 pr_err("per-event callgraph setting for %s failed. "
760 "Apply callgraph global setting for it\n",
761 evsel->name);
762 return;
763 }
764 }
765 }
766 if (dump_size > 0) {
767 dump_size = round_up(dump_size, sizeof(u64));
768 param.dump_size = dump_size;
769 }
770
771 /* If global callgraph set, clear it */
772 if (callchain_param.enabled)
773 perf_evsel__reset_callgraph(evsel, &callchain_param);
774
775 /* set perf-event callgraph */
776 if (param.enabled)
777 perf_evsel__config_callchain(evsel, opts, ¶m);
778 }
779}
780
781/*
782 * The enable_on_exec/disabled value strategy:
783 *
784 * 1) For any type of traced program:
785 * - all independent events and group leaders are disabled
786 * - all group members are enabled
787 *
788 * Group members are ruled by group leaders. They need to
789 * be enabled, because the group scheduling relies on that.
790 *
791 * 2) For traced programs executed by perf:
792 * - all independent events and group leaders have
793 * enable_on_exec set
794 * - we don't specifically enable or disable any event during
795 * the record command
796 *
797 * Independent events and group leaders are initially disabled
798 * and get enabled by exec. Group members are ruled by group
799 * leaders as stated in 1).
800 *
801 * 3) For traced programs attached by perf (pid/tid):
802 * - we specifically enable or disable all events during
803 * the record command
804 *
805 * When attaching events to already running traced we
806 * enable/disable events specifically, as there's no
807 * initial traced exec call.
808 */
809void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
810 struct callchain_param *callchain)
811{
812 struct perf_evsel *leader = evsel->leader;
813 struct perf_event_attr *attr = &evsel->attr;
814 int track = evsel->tracking;
815 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
816
817 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
818 attr->inherit = !opts->no_inherit;
819 attr->write_backward = opts->overwrite ? 1 : 0;
820
821 perf_evsel__set_sample_bit(evsel, IP);
822 perf_evsel__set_sample_bit(evsel, TID);
823
824 if (evsel->sample_read) {
825 perf_evsel__set_sample_bit(evsel, READ);
826
827 /*
828 * We need ID even in case of single event, because
829 * PERF_SAMPLE_READ process ID specific data.
830 */
831 perf_evsel__set_sample_id(evsel, false);
832
833 /*
834 * Apply group format only if we belong to group
835 * with more than one members.
836 */
837 if (leader->nr_members > 1) {
838 attr->read_format |= PERF_FORMAT_GROUP;
839 attr->inherit = 0;
840 }
841 }
842
843 /*
844 * We default some events to have a default interval. But keep
845 * it a weak assumption overridable by the user.
846 */
847 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
848 opts->user_interval != ULLONG_MAX)) {
849 if (opts->freq) {
850 perf_evsel__set_sample_bit(evsel, PERIOD);
851 attr->freq = 1;
852 attr->sample_freq = opts->freq;
853 } else {
854 attr->sample_period = opts->default_interval;
855 }
856 }
857
858 /*
859 * Disable sampling for all group members other
860 * than leader in case leader 'leads' the sampling.
861 */
862 if ((leader != evsel) && leader->sample_read) {
863 attr->sample_freq = 0;
864 attr->sample_period = 0;
865 }
866
867 if (opts->no_samples)
868 attr->sample_freq = 0;
869
870 if (opts->inherit_stat)
871 attr->inherit_stat = 1;
872
873 if (opts->sample_address) {
874 perf_evsel__set_sample_bit(evsel, ADDR);
875 attr->mmap_data = track;
876 }
877
878 /*
879 * We don't allow user space callchains for function trace
880 * event, due to issues with page faults while tracing page
881 * fault handler and its overall trickiness nature.
882 */
883 if (perf_evsel__is_function_event(evsel))
884 evsel->attr.exclude_callchain_user = 1;
885
886 if (callchain && callchain->enabled && !evsel->no_aux_samples)
887 perf_evsel__config_callchain(evsel, opts, callchain);
888
889 if (opts->sample_intr_regs) {
890 attr->sample_regs_intr = opts->sample_intr_regs;
891 perf_evsel__set_sample_bit(evsel, REGS_INTR);
892 }
893
894 if (target__has_cpu(&opts->target) || opts->sample_cpu)
895 perf_evsel__set_sample_bit(evsel, CPU);
896
897 if (opts->period)
898 perf_evsel__set_sample_bit(evsel, PERIOD);
899
900 /*
901 * When the user explicitly disabled time don't force it here.
902 */
903 if (opts->sample_time &&
904 (!perf_missing_features.sample_id_all &&
905 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
906 opts->sample_time_set)))
907 perf_evsel__set_sample_bit(evsel, TIME);
908
909 if (opts->raw_samples && !evsel->no_aux_samples) {
910 perf_evsel__set_sample_bit(evsel, TIME);
911 perf_evsel__set_sample_bit(evsel, RAW);
912 perf_evsel__set_sample_bit(evsel, CPU);
913 }
914
915 if (opts->sample_address)
916 perf_evsel__set_sample_bit(evsel, DATA_SRC);
917
918 if (opts->no_buffering) {
919 attr->watermark = 0;
920 attr->wakeup_events = 1;
921 }
922 if (opts->branch_stack && !evsel->no_aux_samples) {
923 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
924 attr->branch_sample_type = opts->branch_stack;
925 }
926
927 if (opts->sample_weight)
928 perf_evsel__set_sample_bit(evsel, WEIGHT);
929
930 attr->task = track;
931 attr->mmap = track;
932 attr->mmap2 = track && !perf_missing_features.mmap2;
933 attr->comm = track;
934
935 if (opts->record_switch_events)
936 attr->context_switch = track;
937
938 if (opts->sample_transaction)
939 perf_evsel__set_sample_bit(evsel, TRANSACTION);
940
941 if (opts->running_time) {
942 evsel->attr.read_format |=
943 PERF_FORMAT_TOTAL_TIME_ENABLED |
944 PERF_FORMAT_TOTAL_TIME_RUNNING;
945 }
946
947 /*
948 * XXX see the function comment above
949 *
950 * Disabling only independent events or group leaders,
951 * keeping group members enabled.
952 */
953 if (perf_evsel__is_group_leader(evsel))
954 attr->disabled = 1;
955
956 /*
957 * Setting enable_on_exec for independent events and
958 * group leaders for traced executed by perf.
959 */
960 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
961 !opts->initial_delay)
962 attr->enable_on_exec = 1;
963
964 if (evsel->immediate) {
965 attr->disabled = 0;
966 attr->enable_on_exec = 0;
967 }
968
969 clockid = opts->clockid;
970 if (opts->use_clockid) {
971 attr->use_clockid = 1;
972 attr->clockid = opts->clockid;
973 }
974
975 if (evsel->precise_max)
976 perf_event_attr__set_max_precise_ip(attr);
977
978 if (opts->all_user) {
979 attr->exclude_kernel = 1;
980 attr->exclude_user = 0;
981 }
982
983 if (opts->all_kernel) {
984 attr->exclude_kernel = 0;
985 attr->exclude_user = 1;
986 }
987
988 /*
989 * Apply event specific term settings,
990 * it overloads any global configuration.
991 */
992 apply_config_terms(evsel, opts);
993
994 evsel->ignore_missing_thread = opts->ignore_missing_thread;
995}
996
997static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
998{
999 if (evsel->system_wide)
1000 nthreads = 1;
1001
1002 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1003
1004 if (evsel->fd) {
1005 int cpu, thread;
1006 for (cpu = 0; cpu < ncpus; cpu++) {
1007 for (thread = 0; thread < nthreads; thread++) {
1008 FD(evsel, cpu, thread) = -1;
1009 }
1010 }
1011 }
1012
1013 return evsel->fd != NULL ? 0 : -ENOMEM;
1014}
1015
1016static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
1017 int ioc, void *arg)
1018{
1019 int cpu, thread;
1020
1021 if (evsel->system_wide)
1022 nthreads = 1;
1023
1024 for (cpu = 0; cpu < ncpus; cpu++) {
1025 for (thread = 0; thread < nthreads; thread++) {
1026 int fd = FD(evsel, cpu, thread),
1027 err = ioctl(fd, ioc, arg);
1028
1029 if (err)
1030 return err;
1031 }
1032 }
1033
1034 return 0;
1035}
1036
1037int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
1038 const char *filter)
1039{
1040 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1041 PERF_EVENT_IOC_SET_FILTER,
1042 (void *)filter);
1043}
1044
1045int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1046{
1047 char *new_filter = strdup(filter);
1048
1049 if (new_filter != NULL) {
1050 free(evsel->filter);
1051 evsel->filter = new_filter;
1052 return 0;
1053 }
1054
1055 return -1;
1056}
1057
1058static int perf_evsel__append_filter(struct perf_evsel *evsel,
1059 const char *fmt, const char *filter)
1060{
1061 char *new_filter;
1062
1063 if (evsel->filter == NULL)
1064 return perf_evsel__set_filter(evsel, filter);
1065
1066 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1067 free(evsel->filter);
1068 evsel->filter = new_filter;
1069 return 0;
1070 }
1071
1072 return -1;
1073}
1074
1075int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1076{
1077 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1078}
1079
1080int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1081{
1082 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1083}
1084
1085int perf_evsel__enable(struct perf_evsel *evsel)
1086{
1087 int nthreads = thread_map__nr(evsel->threads);
1088 int ncpus = cpu_map__nr(evsel->cpus);
1089
1090 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1091 PERF_EVENT_IOC_ENABLE,
1092 0);
1093}
1094
1095int perf_evsel__disable(struct perf_evsel *evsel)
1096{
1097 int nthreads = thread_map__nr(evsel->threads);
1098 int ncpus = cpu_map__nr(evsel->cpus);
1099
1100 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1101 PERF_EVENT_IOC_DISABLE,
1102 0);
1103}
1104
1105int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1106{
1107 if (ncpus == 0 || nthreads == 0)
1108 return 0;
1109
1110 if (evsel->system_wide)
1111 nthreads = 1;
1112
1113 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1114 if (evsel->sample_id == NULL)
1115 return -ENOMEM;
1116
1117 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1118 if (evsel->id == NULL) {
1119 xyarray__delete(evsel->sample_id);
1120 evsel->sample_id = NULL;
1121 return -ENOMEM;
1122 }
1123
1124 return 0;
1125}
1126
1127static void perf_evsel__free_fd(struct perf_evsel *evsel)
1128{
1129 xyarray__delete(evsel->fd);
1130 evsel->fd = NULL;
1131}
1132
1133static void perf_evsel__free_id(struct perf_evsel *evsel)
1134{
1135 xyarray__delete(evsel->sample_id);
1136 evsel->sample_id = NULL;
1137 zfree(&evsel->id);
1138}
1139
1140static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1141{
1142 struct perf_evsel_config_term *term, *h;
1143
1144 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1145 list_del(&term->list);
1146 free(term);
1147 }
1148}
1149
1150void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1151{
1152 int cpu, thread;
1153
1154 if (evsel->system_wide)
1155 nthreads = 1;
1156
1157 for (cpu = 0; cpu < ncpus; cpu++)
1158 for (thread = 0; thread < nthreads; ++thread) {
1159 close(FD(evsel, cpu, thread));
1160 FD(evsel, cpu, thread) = -1;
1161 }
1162}
1163
1164void perf_evsel__exit(struct perf_evsel *evsel)
1165{
1166 assert(list_empty(&evsel->node));
1167 assert(evsel->evlist == NULL);
1168 perf_evsel__free_fd(evsel);
1169 perf_evsel__free_id(evsel);
1170 perf_evsel__free_config_terms(evsel);
1171 close_cgroup(evsel->cgrp);
1172 cpu_map__put(evsel->cpus);
1173 cpu_map__put(evsel->own_cpus);
1174 thread_map__put(evsel->threads);
1175 zfree(&evsel->group_name);
1176 zfree(&evsel->name);
1177 perf_evsel__object.fini(evsel);
1178}
1179
1180void perf_evsel__delete(struct perf_evsel *evsel)
1181{
1182 perf_evsel__exit(evsel);
1183 free(evsel);
1184}
1185
1186void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1187 struct perf_counts_values *count)
1188{
1189 struct perf_counts_values tmp;
1190
1191 if (!evsel->prev_raw_counts)
1192 return;
1193
1194 if (cpu == -1) {
1195 tmp = evsel->prev_raw_counts->aggr;
1196 evsel->prev_raw_counts->aggr = *count;
1197 } else {
1198 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1199 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1200 }
1201
1202 count->val = count->val - tmp.val;
1203 count->ena = count->ena - tmp.ena;
1204 count->run = count->run - tmp.run;
1205}
1206
1207void perf_counts_values__scale(struct perf_counts_values *count,
1208 bool scale, s8 *pscaled)
1209{
1210 s8 scaled = 0;
1211
1212 if (scale) {
1213 if (count->run == 0) {
1214 scaled = -1;
1215 count->val = 0;
1216 } else if (count->run < count->ena) {
1217 scaled = 1;
1218 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1219 }
1220 } else
1221 count->ena = count->run = 0;
1222
1223 if (pscaled)
1224 *pscaled = scaled;
1225}
1226
1227int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1228 struct perf_counts_values *count)
1229{
1230 memset(count, 0, sizeof(*count));
1231
1232 if (FD(evsel, cpu, thread) < 0)
1233 return -EINVAL;
1234
1235 if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1236 return -errno;
1237
1238 return 0;
1239}
1240
1241int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1242 int cpu, int thread, bool scale)
1243{
1244 struct perf_counts_values count;
1245 size_t nv = scale ? 3 : 1;
1246
1247 if (FD(evsel, cpu, thread) < 0)
1248 return -EINVAL;
1249
1250 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1251 return -ENOMEM;
1252
1253 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1254 return -errno;
1255
1256 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1257 perf_counts_values__scale(&count, scale, NULL);
1258 *perf_counts(evsel->counts, cpu, thread) = count;
1259 return 0;
1260}
1261
1262static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1263{
1264 struct perf_evsel *leader = evsel->leader;
1265 int fd;
1266
1267 if (perf_evsel__is_group_leader(evsel))
1268 return -1;
1269
1270 /*
1271 * Leader must be already processed/open,
1272 * if not it's a bug.
1273 */
1274 BUG_ON(!leader->fd);
1275
1276 fd = FD(leader, cpu, thread);
1277 BUG_ON(fd == -1);
1278
1279 return fd;
1280}
1281
1282struct bit_names {
1283 int bit;
1284 const char *name;
1285};
1286
1287static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1288{
1289 bool first_bit = true;
1290 int i = 0;
1291
1292 do {
1293 if (value & bits[i].bit) {
1294 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1295 first_bit = false;
1296 }
1297 } while (bits[++i].name != NULL);
1298}
1299
1300static void __p_sample_type(char *buf, size_t size, u64 value)
1301{
1302#define bit_name(n) { PERF_SAMPLE_##n, #n }
1303 struct bit_names bits[] = {
1304 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1305 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1306 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1307 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1308 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1309 bit_name(WEIGHT),
1310 { .name = NULL, }
1311 };
1312#undef bit_name
1313 __p_bits(buf, size, value, bits);
1314}
1315
1316static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1317{
1318#define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1319 struct bit_names bits[] = {
1320 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1321 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1322 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1323 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1324 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1325 { .name = NULL, }
1326 };
1327#undef bit_name
1328 __p_bits(buf, size, value, bits);
1329}
1330
1331static void __p_read_format(char *buf, size_t size, u64 value)
1332{
1333#define bit_name(n) { PERF_FORMAT_##n, #n }
1334 struct bit_names bits[] = {
1335 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1336 bit_name(ID), bit_name(GROUP),
1337 { .name = NULL, }
1338 };
1339#undef bit_name
1340 __p_bits(buf, size, value, bits);
1341}
1342
1343#define BUF_SIZE 1024
1344
1345#define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1346#define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1347#define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1348#define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1349#define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1350#define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1351
1352#define PRINT_ATTRn(_n, _f, _p) \
1353do { \
1354 if (attr->_f) { \
1355 _p(attr->_f); \
1356 ret += attr__fprintf(fp, _n, buf, priv);\
1357 } \
1358} while (0)
1359
1360#define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1361
1362int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1363 attr__fprintf_f attr__fprintf, void *priv)
1364{
1365 char buf[BUF_SIZE];
1366 int ret = 0;
1367
1368 PRINT_ATTRf(type, p_unsigned);
1369 PRINT_ATTRf(size, p_unsigned);
1370 PRINT_ATTRf(config, p_hex);
1371 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1372 PRINT_ATTRf(sample_type, p_sample_type);
1373 PRINT_ATTRf(read_format, p_read_format);
1374
1375 PRINT_ATTRf(disabled, p_unsigned);
1376 PRINT_ATTRf(inherit, p_unsigned);
1377 PRINT_ATTRf(pinned, p_unsigned);
1378 PRINT_ATTRf(exclusive, p_unsigned);
1379 PRINT_ATTRf(exclude_user, p_unsigned);
1380 PRINT_ATTRf(exclude_kernel, p_unsigned);
1381 PRINT_ATTRf(exclude_hv, p_unsigned);
1382 PRINT_ATTRf(exclude_idle, p_unsigned);
1383 PRINT_ATTRf(mmap, p_unsigned);
1384 PRINT_ATTRf(comm, p_unsigned);
1385 PRINT_ATTRf(freq, p_unsigned);
1386 PRINT_ATTRf(inherit_stat, p_unsigned);
1387 PRINT_ATTRf(enable_on_exec, p_unsigned);
1388 PRINT_ATTRf(task, p_unsigned);
1389 PRINT_ATTRf(watermark, p_unsigned);
1390 PRINT_ATTRf(precise_ip, p_unsigned);
1391 PRINT_ATTRf(mmap_data, p_unsigned);
1392 PRINT_ATTRf(sample_id_all, p_unsigned);
1393 PRINT_ATTRf(exclude_host, p_unsigned);
1394 PRINT_ATTRf(exclude_guest, p_unsigned);
1395 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1396 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1397 PRINT_ATTRf(mmap2, p_unsigned);
1398 PRINT_ATTRf(comm_exec, p_unsigned);
1399 PRINT_ATTRf(use_clockid, p_unsigned);
1400 PRINT_ATTRf(context_switch, p_unsigned);
1401 PRINT_ATTRf(write_backward, p_unsigned);
1402
1403 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1404 PRINT_ATTRf(bp_type, p_unsigned);
1405 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1406 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1407 PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1408 PRINT_ATTRf(sample_regs_user, p_hex);
1409 PRINT_ATTRf(sample_stack_user, p_unsigned);
1410 PRINT_ATTRf(clockid, p_signed);
1411 PRINT_ATTRf(sample_regs_intr, p_hex);
1412 PRINT_ATTRf(aux_watermark, p_unsigned);
1413 PRINT_ATTRf(sample_max_stack, p_unsigned);
1414
1415 return ret;
1416}
1417
1418static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1419 void *priv __attribute__((unused)))
1420{
1421 return fprintf(fp, " %-32s %s\n", name, val);
1422}
1423
1424static bool ignore_missing_thread(struct perf_evsel *evsel,
1425 struct thread_map *threads,
1426 int thread, int err)
1427{
1428 if (!evsel->ignore_missing_thread)
1429 return false;
1430
1431 /* The system wide setup does not work with threads. */
1432 if (evsel->system_wide)
1433 return false;
1434
1435 /* The -ESRCH is perf event syscall errno for pid's not found. */
1436 if (err != -ESRCH)
1437 return false;
1438
1439 /* If there's only one thread, let it fail. */
1440 if (threads->nr == 1)
1441 return false;
1442
1443 if (thread_map__remove(threads, thread))
1444 return false;
1445
1446 pr_warning("WARNING: Ignored open failure for pid %d\n",
1447 thread_map__pid(threads, thread));
1448 return true;
1449}
1450
1451static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1452 struct thread_map *threads)
1453{
1454 int cpu, thread, nthreads;
1455 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1456 int pid = -1, err;
1457 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1458
1459 if (perf_missing_features.write_backward && evsel->attr.write_backward)
1460 return -EINVAL;
1461
1462 if (evsel->system_wide)
1463 nthreads = 1;
1464 else
1465 nthreads = threads->nr;
1466
1467 if (evsel->fd == NULL &&
1468 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1469 return -ENOMEM;
1470
1471 if (evsel->cgrp) {
1472 flags |= PERF_FLAG_PID_CGROUP;
1473 pid = evsel->cgrp->fd;
1474 }
1475
1476fallback_missing_features:
1477 if (perf_missing_features.clockid_wrong)
1478 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1479 if (perf_missing_features.clockid) {
1480 evsel->attr.use_clockid = 0;
1481 evsel->attr.clockid = 0;
1482 }
1483 if (perf_missing_features.cloexec)
1484 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1485 if (perf_missing_features.mmap2)
1486 evsel->attr.mmap2 = 0;
1487 if (perf_missing_features.exclude_guest)
1488 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1489 if (perf_missing_features.lbr_flags)
1490 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1491 PERF_SAMPLE_BRANCH_NO_CYCLES);
1492retry_sample_id:
1493 if (perf_missing_features.sample_id_all)
1494 evsel->attr.sample_id_all = 0;
1495
1496 if (verbose >= 2) {
1497 fprintf(stderr, "%.60s\n", graph_dotted_line);
1498 fprintf(stderr, "perf_event_attr:\n");
1499 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1500 fprintf(stderr, "%.60s\n", graph_dotted_line);
1501 }
1502
1503 for (cpu = 0; cpu < cpus->nr; cpu++) {
1504
1505 for (thread = 0; thread < nthreads; thread++) {
1506 int fd, group_fd;
1507
1508 if (!evsel->cgrp && !evsel->system_wide)
1509 pid = thread_map__pid(threads, thread);
1510
1511 group_fd = get_group_fd(evsel, cpu, thread);
1512retry_open:
1513 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1514 pid, cpus->map[cpu], group_fd, flags);
1515
1516 fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1517 group_fd, flags);
1518
1519 FD(evsel, cpu, thread) = fd;
1520
1521 if (fd < 0) {
1522 err = -errno;
1523
1524 if (ignore_missing_thread(evsel, threads, thread, err)) {
1525 /*
1526 * We just removed 1 thread, so take a step
1527 * back on thread index and lower the upper
1528 * nthreads limit.
1529 */
1530 nthreads--;
1531 thread--;
1532
1533 /* ... and pretend like nothing have happened. */
1534 err = 0;
1535 continue;
1536 }
1537
1538 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1539 err);
1540 goto try_fallback;
1541 }
1542
1543 pr_debug2(" = %d\n", fd);
1544
1545 if (evsel->bpf_fd >= 0) {
1546 int evt_fd = fd;
1547 int bpf_fd = evsel->bpf_fd;
1548
1549 err = ioctl(evt_fd,
1550 PERF_EVENT_IOC_SET_BPF,
1551 bpf_fd);
1552 if (err && errno != EEXIST) {
1553 pr_err("failed to attach bpf fd %d: %s\n",
1554 bpf_fd, strerror(errno));
1555 err = -EINVAL;
1556 goto out_close;
1557 }
1558 }
1559
1560 set_rlimit = NO_CHANGE;
1561
1562 /*
1563 * If we succeeded but had to kill clockid, fail and
1564 * have perf_evsel__open_strerror() print us a nice
1565 * error.
1566 */
1567 if (perf_missing_features.clockid ||
1568 perf_missing_features.clockid_wrong) {
1569 err = -EINVAL;
1570 goto out_close;
1571 }
1572 }
1573 }
1574
1575 return 0;
1576
1577try_fallback:
1578 /*
1579 * perf stat needs between 5 and 22 fds per CPU. When we run out
1580 * of them try to increase the limits.
1581 */
1582 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1583 struct rlimit l;
1584 int old_errno = errno;
1585
1586 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1587 if (set_rlimit == NO_CHANGE)
1588 l.rlim_cur = l.rlim_max;
1589 else {
1590 l.rlim_cur = l.rlim_max + 1000;
1591 l.rlim_max = l.rlim_cur;
1592 }
1593 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1594 set_rlimit++;
1595 errno = old_errno;
1596 goto retry_open;
1597 }
1598 }
1599 errno = old_errno;
1600 }
1601
1602 if (err != -EINVAL || cpu > 0 || thread > 0)
1603 goto out_close;
1604
1605 /*
1606 * Must probe features in the order they were added to the
1607 * perf_event_attr interface.
1608 */
1609 if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1610 perf_missing_features.write_backward = true;
1611 goto out_close;
1612 } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1613 perf_missing_features.clockid_wrong = true;
1614 goto fallback_missing_features;
1615 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1616 perf_missing_features.clockid = true;
1617 goto fallback_missing_features;
1618 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1619 perf_missing_features.cloexec = true;
1620 goto fallback_missing_features;
1621 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1622 perf_missing_features.mmap2 = true;
1623 goto fallback_missing_features;
1624 } else if (!perf_missing_features.exclude_guest &&
1625 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1626 perf_missing_features.exclude_guest = true;
1627 goto fallback_missing_features;
1628 } else if (!perf_missing_features.sample_id_all) {
1629 perf_missing_features.sample_id_all = true;
1630 goto retry_sample_id;
1631 } else if (!perf_missing_features.lbr_flags &&
1632 (evsel->attr.branch_sample_type &
1633 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1634 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1635 perf_missing_features.lbr_flags = true;
1636 goto fallback_missing_features;
1637 }
1638out_close:
1639 do {
1640 while (--thread >= 0) {
1641 close(FD(evsel, cpu, thread));
1642 FD(evsel, cpu, thread) = -1;
1643 }
1644 thread = nthreads;
1645 } while (--cpu >= 0);
1646 return err;
1647}
1648
1649void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1650{
1651 if (evsel->fd == NULL)
1652 return;
1653
1654 perf_evsel__close_fd(evsel, ncpus, nthreads);
1655 perf_evsel__free_fd(evsel);
1656}
1657
1658static struct {
1659 struct cpu_map map;
1660 int cpus[1];
1661} empty_cpu_map = {
1662 .map.nr = 1,
1663 .cpus = { -1, },
1664};
1665
1666static struct {
1667 struct thread_map map;
1668 int threads[1];
1669} empty_thread_map = {
1670 .map.nr = 1,
1671 .threads = { -1, },
1672};
1673
1674int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1675 struct thread_map *threads)
1676{
1677 if (cpus == NULL) {
1678 /* Work around old compiler warnings about strict aliasing */
1679 cpus = &empty_cpu_map.map;
1680 }
1681
1682 if (threads == NULL)
1683 threads = &empty_thread_map.map;
1684
1685 return __perf_evsel__open(evsel, cpus, threads);
1686}
1687
1688int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1689 struct cpu_map *cpus)
1690{
1691 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1692}
1693
1694int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1695 struct thread_map *threads)
1696{
1697 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1698}
1699
1700static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1701 const union perf_event *event,
1702 struct perf_sample *sample)
1703{
1704 u64 type = evsel->attr.sample_type;
1705 const u64 *array = event->sample.array;
1706 bool swapped = evsel->needs_swap;
1707 union u64_swap u;
1708
1709 array += ((event->header.size -
1710 sizeof(event->header)) / sizeof(u64)) - 1;
1711
1712 if (type & PERF_SAMPLE_IDENTIFIER) {
1713 sample->id = *array;
1714 array--;
1715 }
1716
1717 if (type & PERF_SAMPLE_CPU) {
1718 u.val64 = *array;
1719 if (swapped) {
1720 /* undo swap of u64, then swap on individual u32s */
1721 u.val64 = bswap_64(u.val64);
1722 u.val32[0] = bswap_32(u.val32[0]);
1723 }
1724
1725 sample->cpu = u.val32[0];
1726 array--;
1727 }
1728
1729 if (type & PERF_SAMPLE_STREAM_ID) {
1730 sample->stream_id = *array;
1731 array--;
1732 }
1733
1734 if (type & PERF_SAMPLE_ID) {
1735 sample->id = *array;
1736 array--;
1737 }
1738
1739 if (type & PERF_SAMPLE_TIME) {
1740 sample->time = *array;
1741 array--;
1742 }
1743
1744 if (type & PERF_SAMPLE_TID) {
1745 u.val64 = *array;
1746 if (swapped) {
1747 /* undo swap of u64, then swap on individual u32s */
1748 u.val64 = bswap_64(u.val64);
1749 u.val32[0] = bswap_32(u.val32[0]);
1750 u.val32[1] = bswap_32(u.val32[1]);
1751 }
1752
1753 sample->pid = u.val32[0];
1754 sample->tid = u.val32[1];
1755 array--;
1756 }
1757
1758 return 0;
1759}
1760
1761static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1762 u64 size)
1763{
1764 return size > max_size || offset + size > endp;
1765}
1766
1767#define OVERFLOW_CHECK(offset, size, max_size) \
1768 do { \
1769 if (overflow(endp, (max_size), (offset), (size))) \
1770 return -EFAULT; \
1771 } while (0)
1772
1773#define OVERFLOW_CHECK_u64(offset) \
1774 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1775
1776int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1777 struct perf_sample *data)
1778{
1779 u64 type = evsel->attr.sample_type;
1780 bool swapped = evsel->needs_swap;
1781 const u64 *array;
1782 u16 max_size = event->header.size;
1783 const void *endp = (void *)event + max_size;
1784 u64 sz;
1785
1786 /*
1787 * used for cross-endian analysis. See git commit 65014ab3
1788 * for why this goofiness is needed.
1789 */
1790 union u64_swap u;
1791
1792 memset(data, 0, sizeof(*data));
1793 data->cpu = data->pid = data->tid = -1;
1794 data->stream_id = data->id = data->time = -1ULL;
1795 data->period = evsel->attr.sample_period;
1796 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1797
1798 if (event->header.type != PERF_RECORD_SAMPLE) {
1799 if (!evsel->attr.sample_id_all)
1800 return 0;
1801 return perf_evsel__parse_id_sample(evsel, event, data);
1802 }
1803
1804 array = event->sample.array;
1805
1806 /*
1807 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1808 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1809 * check the format does not go past the end of the event.
1810 */
1811 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1812 return -EFAULT;
1813
1814 data->id = -1ULL;
1815 if (type & PERF_SAMPLE_IDENTIFIER) {
1816 data->id = *array;
1817 array++;
1818 }
1819
1820 if (type & PERF_SAMPLE_IP) {
1821 data->ip = *array;
1822 array++;
1823 }
1824
1825 if (type & PERF_SAMPLE_TID) {
1826 u.val64 = *array;
1827 if (swapped) {
1828 /* undo swap of u64, then swap on individual u32s */
1829 u.val64 = bswap_64(u.val64);
1830 u.val32[0] = bswap_32(u.val32[0]);
1831 u.val32[1] = bswap_32(u.val32[1]);
1832 }
1833
1834 data->pid = u.val32[0];
1835 data->tid = u.val32[1];
1836 array++;
1837 }
1838
1839 if (type & PERF_SAMPLE_TIME) {
1840 data->time = *array;
1841 array++;
1842 }
1843
1844 data->addr = 0;
1845 if (type & PERF_SAMPLE_ADDR) {
1846 data->addr = *array;
1847 array++;
1848 }
1849
1850 if (type & PERF_SAMPLE_ID) {
1851 data->id = *array;
1852 array++;
1853 }
1854
1855 if (type & PERF_SAMPLE_STREAM_ID) {
1856 data->stream_id = *array;
1857 array++;
1858 }
1859
1860 if (type & PERF_SAMPLE_CPU) {
1861
1862 u.val64 = *array;
1863 if (swapped) {
1864 /* undo swap of u64, then swap on individual u32s */
1865 u.val64 = bswap_64(u.val64);
1866 u.val32[0] = bswap_32(u.val32[0]);
1867 }
1868
1869 data->cpu = u.val32[0];
1870 array++;
1871 }
1872
1873 if (type & PERF_SAMPLE_PERIOD) {
1874 data->period = *array;
1875 array++;
1876 }
1877
1878 if (type & PERF_SAMPLE_READ) {
1879 u64 read_format = evsel->attr.read_format;
1880
1881 OVERFLOW_CHECK_u64(array);
1882 if (read_format & PERF_FORMAT_GROUP)
1883 data->read.group.nr = *array;
1884 else
1885 data->read.one.value = *array;
1886
1887 array++;
1888
1889 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1890 OVERFLOW_CHECK_u64(array);
1891 data->read.time_enabled = *array;
1892 array++;
1893 }
1894
1895 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1896 OVERFLOW_CHECK_u64(array);
1897 data->read.time_running = *array;
1898 array++;
1899 }
1900
1901 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1902 if (read_format & PERF_FORMAT_GROUP) {
1903 const u64 max_group_nr = UINT64_MAX /
1904 sizeof(struct sample_read_value);
1905
1906 if (data->read.group.nr > max_group_nr)
1907 return -EFAULT;
1908 sz = data->read.group.nr *
1909 sizeof(struct sample_read_value);
1910 OVERFLOW_CHECK(array, sz, max_size);
1911 data->read.group.values =
1912 (struct sample_read_value *)array;
1913 array = (void *)array + sz;
1914 } else {
1915 OVERFLOW_CHECK_u64(array);
1916 data->read.one.id = *array;
1917 array++;
1918 }
1919 }
1920
1921 if (type & PERF_SAMPLE_CALLCHAIN) {
1922 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1923
1924 OVERFLOW_CHECK_u64(array);
1925 data->callchain = (struct ip_callchain *)array++;
1926 if (data->callchain->nr > max_callchain_nr)
1927 return -EFAULT;
1928 sz = data->callchain->nr * sizeof(u64);
1929 OVERFLOW_CHECK(array, sz, max_size);
1930 array = (void *)array + sz;
1931 }
1932
1933 if (type & PERF_SAMPLE_RAW) {
1934 OVERFLOW_CHECK_u64(array);
1935 u.val64 = *array;
1936 if (WARN_ONCE(swapped,
1937 "Endianness of raw data not corrected!\n")) {
1938 /* undo swap of u64, then swap on individual u32s */
1939 u.val64 = bswap_64(u.val64);
1940 u.val32[0] = bswap_32(u.val32[0]);
1941 u.val32[1] = bswap_32(u.val32[1]);
1942 }
1943 data->raw_size = u.val32[0];
1944 array = (void *)array + sizeof(u32);
1945
1946 OVERFLOW_CHECK(array, data->raw_size, max_size);
1947 data->raw_data = (void *)array;
1948 array = (void *)array + data->raw_size;
1949 }
1950
1951 if (type & PERF_SAMPLE_BRANCH_STACK) {
1952 const u64 max_branch_nr = UINT64_MAX /
1953 sizeof(struct branch_entry);
1954
1955 OVERFLOW_CHECK_u64(array);
1956 data->branch_stack = (struct branch_stack *)array++;
1957
1958 if (data->branch_stack->nr > max_branch_nr)
1959 return -EFAULT;
1960 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1961 OVERFLOW_CHECK(array, sz, max_size);
1962 array = (void *)array + sz;
1963 }
1964
1965 if (type & PERF_SAMPLE_REGS_USER) {
1966 OVERFLOW_CHECK_u64(array);
1967 data->user_regs.abi = *array;
1968 array++;
1969
1970 if (data->user_regs.abi) {
1971 u64 mask = evsel->attr.sample_regs_user;
1972
1973 sz = hweight_long(mask) * sizeof(u64);
1974 OVERFLOW_CHECK(array, sz, max_size);
1975 data->user_regs.mask = mask;
1976 data->user_regs.regs = (u64 *)array;
1977 array = (void *)array + sz;
1978 }
1979 }
1980
1981 if (type & PERF_SAMPLE_STACK_USER) {
1982 OVERFLOW_CHECK_u64(array);
1983 sz = *array++;
1984
1985 data->user_stack.offset = ((char *)(array - 1)
1986 - (char *) event);
1987
1988 if (!sz) {
1989 data->user_stack.size = 0;
1990 } else {
1991 OVERFLOW_CHECK(array, sz, max_size);
1992 data->user_stack.data = (char *)array;
1993 array = (void *)array + sz;
1994 OVERFLOW_CHECK_u64(array);
1995 data->user_stack.size = *array++;
1996 if (WARN_ONCE(data->user_stack.size > sz,
1997 "user stack dump failure\n"))
1998 return -EFAULT;
1999 }
2000 }
2001
2002 if (type & PERF_SAMPLE_WEIGHT) {
2003 OVERFLOW_CHECK_u64(array);
2004 data->weight = *array;
2005 array++;
2006 }
2007
2008 data->data_src = PERF_MEM_DATA_SRC_NONE;
2009 if (type & PERF_SAMPLE_DATA_SRC) {
2010 OVERFLOW_CHECK_u64(array);
2011 data->data_src = *array;
2012 array++;
2013 }
2014
2015 data->transaction = 0;
2016 if (type & PERF_SAMPLE_TRANSACTION) {
2017 OVERFLOW_CHECK_u64(array);
2018 data->transaction = *array;
2019 array++;
2020 }
2021
2022 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2023 if (type & PERF_SAMPLE_REGS_INTR) {
2024 OVERFLOW_CHECK_u64(array);
2025 data->intr_regs.abi = *array;
2026 array++;
2027
2028 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2029 u64 mask = evsel->attr.sample_regs_intr;
2030
2031 sz = hweight_long(mask) * sizeof(u64);
2032 OVERFLOW_CHECK(array, sz, max_size);
2033 data->intr_regs.mask = mask;
2034 data->intr_regs.regs = (u64 *)array;
2035 array = (void *)array + sz;
2036 }
2037 }
2038
2039 return 0;
2040}
2041
2042size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2043 u64 read_format)
2044{
2045 size_t sz, result = sizeof(struct sample_event);
2046
2047 if (type & PERF_SAMPLE_IDENTIFIER)
2048 result += sizeof(u64);
2049
2050 if (type & PERF_SAMPLE_IP)
2051 result += sizeof(u64);
2052
2053 if (type & PERF_SAMPLE_TID)
2054 result += sizeof(u64);
2055
2056 if (type & PERF_SAMPLE_TIME)
2057 result += sizeof(u64);
2058
2059 if (type & PERF_SAMPLE_ADDR)
2060 result += sizeof(u64);
2061
2062 if (type & PERF_SAMPLE_ID)
2063 result += sizeof(u64);
2064
2065 if (type & PERF_SAMPLE_STREAM_ID)
2066 result += sizeof(u64);
2067
2068 if (type & PERF_SAMPLE_CPU)
2069 result += sizeof(u64);
2070
2071 if (type & PERF_SAMPLE_PERIOD)
2072 result += sizeof(u64);
2073
2074 if (type & PERF_SAMPLE_READ) {
2075 result += sizeof(u64);
2076 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2077 result += sizeof(u64);
2078 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2079 result += sizeof(u64);
2080 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2081 if (read_format & PERF_FORMAT_GROUP) {
2082 sz = sample->read.group.nr *
2083 sizeof(struct sample_read_value);
2084 result += sz;
2085 } else {
2086 result += sizeof(u64);
2087 }
2088 }
2089
2090 if (type & PERF_SAMPLE_CALLCHAIN) {
2091 sz = (sample->callchain->nr + 1) * sizeof(u64);
2092 result += sz;
2093 }
2094
2095 if (type & PERF_SAMPLE_RAW) {
2096 result += sizeof(u32);
2097 result += sample->raw_size;
2098 }
2099
2100 if (type & PERF_SAMPLE_BRANCH_STACK) {
2101 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2102 sz += sizeof(u64);
2103 result += sz;
2104 }
2105
2106 if (type & PERF_SAMPLE_REGS_USER) {
2107 if (sample->user_regs.abi) {
2108 result += sizeof(u64);
2109 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2110 result += sz;
2111 } else {
2112 result += sizeof(u64);
2113 }
2114 }
2115
2116 if (type & PERF_SAMPLE_STACK_USER) {
2117 sz = sample->user_stack.size;
2118 result += sizeof(u64);
2119 if (sz) {
2120 result += sz;
2121 result += sizeof(u64);
2122 }
2123 }
2124
2125 if (type & PERF_SAMPLE_WEIGHT)
2126 result += sizeof(u64);
2127
2128 if (type & PERF_SAMPLE_DATA_SRC)
2129 result += sizeof(u64);
2130
2131 if (type & PERF_SAMPLE_TRANSACTION)
2132 result += sizeof(u64);
2133
2134 if (type & PERF_SAMPLE_REGS_INTR) {
2135 if (sample->intr_regs.abi) {
2136 result += sizeof(u64);
2137 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2138 result += sz;
2139 } else {
2140 result += sizeof(u64);
2141 }
2142 }
2143
2144 return result;
2145}
2146
2147int perf_event__synthesize_sample(union perf_event *event, u64 type,
2148 u64 read_format,
2149 const struct perf_sample *sample,
2150 bool swapped)
2151{
2152 u64 *array;
2153 size_t sz;
2154 /*
2155 * used for cross-endian analysis. See git commit 65014ab3
2156 * for why this goofiness is needed.
2157 */
2158 union u64_swap u;
2159
2160 array = event->sample.array;
2161
2162 if (type & PERF_SAMPLE_IDENTIFIER) {
2163 *array = sample->id;
2164 array++;
2165 }
2166
2167 if (type & PERF_SAMPLE_IP) {
2168 *array = sample->ip;
2169 array++;
2170 }
2171
2172 if (type & PERF_SAMPLE_TID) {
2173 u.val32[0] = sample->pid;
2174 u.val32[1] = sample->tid;
2175 if (swapped) {
2176 /*
2177 * Inverse of what is done in perf_evsel__parse_sample
2178 */
2179 u.val32[0] = bswap_32(u.val32[0]);
2180 u.val32[1] = bswap_32(u.val32[1]);
2181 u.val64 = bswap_64(u.val64);
2182 }
2183
2184 *array = u.val64;
2185 array++;
2186 }
2187
2188 if (type & PERF_SAMPLE_TIME) {
2189 *array = sample->time;
2190 array++;
2191 }
2192
2193 if (type & PERF_SAMPLE_ADDR) {
2194 *array = sample->addr;
2195 array++;
2196 }
2197
2198 if (type & PERF_SAMPLE_ID) {
2199 *array = sample->id;
2200 array++;
2201 }
2202
2203 if (type & PERF_SAMPLE_STREAM_ID) {
2204 *array = sample->stream_id;
2205 array++;
2206 }
2207
2208 if (type & PERF_SAMPLE_CPU) {
2209 u.val32[0] = sample->cpu;
2210 if (swapped) {
2211 /*
2212 * Inverse of what is done in perf_evsel__parse_sample
2213 */
2214 u.val32[0] = bswap_32(u.val32[0]);
2215 u.val64 = bswap_64(u.val64);
2216 }
2217 *array = u.val64;
2218 array++;
2219 }
2220
2221 if (type & PERF_SAMPLE_PERIOD) {
2222 *array = sample->period;
2223 array++;
2224 }
2225
2226 if (type & PERF_SAMPLE_READ) {
2227 if (read_format & PERF_FORMAT_GROUP)
2228 *array = sample->read.group.nr;
2229 else
2230 *array = sample->read.one.value;
2231 array++;
2232
2233 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2234 *array = sample->read.time_enabled;
2235 array++;
2236 }
2237
2238 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2239 *array = sample->read.time_running;
2240 array++;
2241 }
2242
2243 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2244 if (read_format & PERF_FORMAT_GROUP) {
2245 sz = sample->read.group.nr *
2246 sizeof(struct sample_read_value);
2247 memcpy(array, sample->read.group.values, sz);
2248 array = (void *)array + sz;
2249 } else {
2250 *array = sample->read.one.id;
2251 array++;
2252 }
2253 }
2254
2255 if (type & PERF_SAMPLE_CALLCHAIN) {
2256 sz = (sample->callchain->nr + 1) * sizeof(u64);
2257 memcpy(array, sample->callchain, sz);
2258 array = (void *)array + sz;
2259 }
2260
2261 if (type & PERF_SAMPLE_RAW) {
2262 u.val32[0] = sample->raw_size;
2263 if (WARN_ONCE(swapped,
2264 "Endianness of raw data not corrected!\n")) {
2265 /*
2266 * Inverse of what is done in perf_evsel__parse_sample
2267 */
2268 u.val32[0] = bswap_32(u.val32[0]);
2269 u.val32[1] = bswap_32(u.val32[1]);
2270 u.val64 = bswap_64(u.val64);
2271 }
2272 *array = u.val64;
2273 array = (void *)array + sizeof(u32);
2274
2275 memcpy(array, sample->raw_data, sample->raw_size);
2276 array = (void *)array + sample->raw_size;
2277 }
2278
2279 if (type & PERF_SAMPLE_BRANCH_STACK) {
2280 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2281 sz += sizeof(u64);
2282 memcpy(array, sample->branch_stack, sz);
2283 array = (void *)array + sz;
2284 }
2285
2286 if (type & PERF_SAMPLE_REGS_USER) {
2287 if (sample->user_regs.abi) {
2288 *array++ = sample->user_regs.abi;
2289 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2290 memcpy(array, sample->user_regs.regs, sz);
2291 array = (void *)array + sz;
2292 } else {
2293 *array++ = 0;
2294 }
2295 }
2296
2297 if (type & PERF_SAMPLE_STACK_USER) {
2298 sz = sample->user_stack.size;
2299 *array++ = sz;
2300 if (sz) {
2301 memcpy(array, sample->user_stack.data, sz);
2302 array = (void *)array + sz;
2303 *array++ = sz;
2304 }
2305 }
2306
2307 if (type & PERF_SAMPLE_WEIGHT) {
2308 *array = sample->weight;
2309 array++;
2310 }
2311
2312 if (type & PERF_SAMPLE_DATA_SRC) {
2313 *array = sample->data_src;
2314 array++;
2315 }
2316
2317 if (type & PERF_SAMPLE_TRANSACTION) {
2318 *array = sample->transaction;
2319 array++;
2320 }
2321
2322 if (type & PERF_SAMPLE_REGS_INTR) {
2323 if (sample->intr_regs.abi) {
2324 *array++ = sample->intr_regs.abi;
2325 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2326 memcpy(array, sample->intr_regs.regs, sz);
2327 array = (void *)array + sz;
2328 } else {
2329 *array++ = 0;
2330 }
2331 }
2332
2333 return 0;
2334}
2335
2336struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2337{
2338 return pevent_find_field(evsel->tp_format, name);
2339}
2340
2341void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2342 const char *name)
2343{
2344 struct format_field *field = perf_evsel__field(evsel, name);
2345 int offset;
2346
2347 if (!field)
2348 return NULL;
2349
2350 offset = field->offset;
2351
2352 if (field->flags & FIELD_IS_DYNAMIC) {
2353 offset = *(int *)(sample->raw_data + field->offset);
2354 offset &= 0xffff;
2355 }
2356
2357 return sample->raw_data + offset;
2358}
2359
2360u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2361 bool needs_swap)
2362{
2363 u64 value;
2364 void *ptr = sample->raw_data + field->offset;
2365
2366 switch (field->size) {
2367 case 1:
2368 return *(u8 *)ptr;
2369 case 2:
2370 value = *(u16 *)ptr;
2371 break;
2372 case 4:
2373 value = *(u32 *)ptr;
2374 break;
2375 case 8:
2376 memcpy(&value, ptr, sizeof(u64));
2377 break;
2378 default:
2379 return 0;
2380 }
2381
2382 if (!needs_swap)
2383 return value;
2384
2385 switch (field->size) {
2386 case 2:
2387 return bswap_16(value);
2388 case 4:
2389 return bswap_32(value);
2390 case 8:
2391 return bswap_64(value);
2392 default:
2393 return 0;
2394 }
2395
2396 return 0;
2397}
2398
2399u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2400 const char *name)
2401{
2402 struct format_field *field = perf_evsel__field(evsel, name);
2403
2404 if (!field)
2405 return 0;
2406
2407 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2408}
2409
2410bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2411 char *msg, size_t msgsize)
2412{
2413 int paranoid;
2414
2415 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2416 evsel->attr.type == PERF_TYPE_HARDWARE &&
2417 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2418 /*
2419 * If it's cycles then fall back to hrtimer based
2420 * cpu-clock-tick sw counter, which is always available even if
2421 * no PMU support.
2422 *
2423 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2424 * b0a873e).
2425 */
2426 scnprintf(msg, msgsize, "%s",
2427"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2428
2429 evsel->attr.type = PERF_TYPE_SOFTWARE;
2430 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2431
2432 zfree(&evsel->name);
2433 return true;
2434 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2435 (paranoid = perf_event_paranoid()) > 1) {
2436 const char *name = perf_evsel__name(evsel);
2437 char *new_name;
2438
2439 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2440 return false;
2441
2442 if (evsel->name)
2443 free(evsel->name);
2444 evsel->name = new_name;
2445 scnprintf(msg, msgsize,
2446"kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2447 evsel->attr.exclude_kernel = 1;
2448
2449 return true;
2450 }
2451
2452 return false;
2453}
2454
2455int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2456 int err, char *msg, size_t size)
2457{
2458 char sbuf[STRERR_BUFSIZE];
2459
2460 switch (err) {
2461 case EPERM:
2462 case EACCES:
2463 return scnprintf(msg, size,
2464 "You may not have permission to collect %sstats.\n\n"
2465 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2466 "which controls use of the performance events system by\n"
2467 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2468 "The current value is %d:\n\n"
2469 " -1: Allow use of (almost) all events by all users\n"
2470 ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2471 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2472 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2473 target->system_wide ? "system-wide " : "",
2474 perf_event_paranoid());
2475 case ENOENT:
2476 return scnprintf(msg, size, "The %s event is not supported.",
2477 perf_evsel__name(evsel));
2478 case EMFILE:
2479 return scnprintf(msg, size, "%s",
2480 "Too many events are opened.\n"
2481 "Probably the maximum number of open file descriptors has been reached.\n"
2482 "Hint: Try again after reducing the number of events.\n"
2483 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2484 case ENOMEM:
2485 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2486 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2487 return scnprintf(msg, size,
2488 "Not enough memory to setup event with callchain.\n"
2489 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2490 "Hint: Current value: %d", sysctl_perf_event_max_stack);
2491 break;
2492 case ENODEV:
2493 if (target->cpu_list)
2494 return scnprintf(msg, size, "%s",
2495 "No such device - did you specify an out-of-range profile CPU?");
2496 break;
2497 case EOPNOTSUPP:
2498 if (evsel->attr.sample_period != 0)
2499 return scnprintf(msg, size, "%s",
2500 "PMU Hardware doesn't support sampling/overflow-interrupts.");
2501 if (evsel->attr.precise_ip)
2502 return scnprintf(msg, size, "%s",
2503 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2504#if defined(__i386__) || defined(__x86_64__)
2505 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2506 return scnprintf(msg, size, "%s",
2507 "No hardware sampling interrupt available.\n"
2508 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2509#endif
2510 break;
2511 case EBUSY:
2512 if (find_process("oprofiled"))
2513 return scnprintf(msg, size,
2514 "The PMU counters are busy/taken by another profiler.\n"
2515 "We found oprofile daemon running, please stop it and try again.");
2516 break;
2517 case EINVAL:
2518 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2519 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2520 if (perf_missing_features.clockid)
2521 return scnprintf(msg, size, "clockid feature not supported.");
2522 if (perf_missing_features.clockid_wrong)
2523 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2524 break;
2525 default:
2526 break;
2527 }
2528
2529 return scnprintf(msg, size,
2530 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2531 "/bin/dmesg may provide additional information.\n"
2532 "No CONFIG_PERF_EVENTS=y kernel support configured?",
2533 err, str_error_r(err, sbuf, sizeof(sbuf)),
2534 perf_evsel__name(evsel));
2535}
2536
2537char *perf_evsel__env_arch(struct perf_evsel *evsel)
2538{
2539 if (evsel && evsel->evlist && evsel->evlist->env)
2540 return evsel->evlist->env->arch;
2541 return NULL;
2542}