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 <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
32static struct {
33 bool sample_id_all;
34 bool exclude_guest;
35 bool mmap2;
36 bool cloexec;
37 bool clockid;
38 bool clockid_wrong;
39 bool lbr_flags;
40} perf_missing_features;
41
42static clockid_t clockid;
43
44static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
45{
46 return 0;
47}
48
49static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
50{
51}
52
53static struct {
54 size_t size;
55 int (*init)(struct perf_evsel *evsel);
56 void (*fini)(struct perf_evsel *evsel);
57} perf_evsel__object = {
58 .size = sizeof(struct perf_evsel),
59 .init = perf_evsel__no_extra_init,
60 .fini = perf_evsel__no_extra_fini,
61};
62
63int perf_evsel__object_config(size_t object_size,
64 int (*init)(struct perf_evsel *evsel),
65 void (*fini)(struct perf_evsel *evsel))
66{
67
68 if (object_size == 0)
69 goto set_methods;
70
71 if (perf_evsel__object.size > object_size)
72 return -EINVAL;
73
74 perf_evsel__object.size = object_size;
75
76set_methods:
77 if (init != NULL)
78 perf_evsel__object.init = init;
79
80 if (fini != NULL)
81 perf_evsel__object.fini = fini;
82
83 return 0;
84}
85
86#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
87
88int __perf_evsel__sample_size(u64 sample_type)
89{
90 u64 mask = sample_type & PERF_SAMPLE_MASK;
91 int size = 0;
92 int i;
93
94 for (i = 0; i < 64; i++) {
95 if (mask & (1ULL << i))
96 size++;
97 }
98
99 size *= sizeof(u64);
100
101 return size;
102}
103
104/**
105 * __perf_evsel__calc_id_pos - calculate id_pos.
106 * @sample_type: sample type
107 *
108 * This function returns the position of the event id (PERF_SAMPLE_ID or
109 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
110 * sample_event.
111 */
112static int __perf_evsel__calc_id_pos(u64 sample_type)
113{
114 int idx = 0;
115
116 if (sample_type & PERF_SAMPLE_IDENTIFIER)
117 return 0;
118
119 if (!(sample_type & PERF_SAMPLE_ID))
120 return -1;
121
122 if (sample_type & PERF_SAMPLE_IP)
123 idx += 1;
124
125 if (sample_type & PERF_SAMPLE_TID)
126 idx += 1;
127
128 if (sample_type & PERF_SAMPLE_TIME)
129 idx += 1;
130
131 if (sample_type & PERF_SAMPLE_ADDR)
132 idx += 1;
133
134 return idx;
135}
136
137/**
138 * __perf_evsel__calc_is_pos - calculate is_pos.
139 * @sample_type: sample type
140 *
141 * This function returns the position (counting backwards) of the event id
142 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
143 * sample_id_all is used there is an id sample appended to non-sample events.
144 */
145static int __perf_evsel__calc_is_pos(u64 sample_type)
146{
147 int idx = 1;
148
149 if (sample_type & PERF_SAMPLE_IDENTIFIER)
150 return 1;
151
152 if (!(sample_type & PERF_SAMPLE_ID))
153 return -1;
154
155 if (sample_type & PERF_SAMPLE_CPU)
156 idx += 1;
157
158 if (sample_type & PERF_SAMPLE_STREAM_ID)
159 idx += 1;
160
161 return idx;
162}
163
164void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
165{
166 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
167 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
168}
169
170void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
171 enum perf_event_sample_format bit)
172{
173 if (!(evsel->attr.sample_type & bit)) {
174 evsel->attr.sample_type |= bit;
175 evsel->sample_size += sizeof(u64);
176 perf_evsel__calc_id_pos(evsel);
177 }
178}
179
180void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
181 enum perf_event_sample_format bit)
182{
183 if (evsel->attr.sample_type & bit) {
184 evsel->attr.sample_type &= ~bit;
185 evsel->sample_size -= sizeof(u64);
186 perf_evsel__calc_id_pos(evsel);
187 }
188}
189
190void perf_evsel__set_sample_id(struct perf_evsel *evsel,
191 bool can_sample_identifier)
192{
193 if (can_sample_identifier) {
194 perf_evsel__reset_sample_bit(evsel, ID);
195 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
196 } else {
197 perf_evsel__set_sample_bit(evsel, ID);
198 }
199 evsel->attr.read_format |= PERF_FORMAT_ID;
200}
201
202void perf_evsel__init(struct perf_evsel *evsel,
203 struct perf_event_attr *attr, int idx)
204{
205 evsel->idx = idx;
206 evsel->tracking = !idx;
207 evsel->attr = *attr;
208 evsel->leader = evsel;
209 evsel->unit = "";
210 evsel->scale = 1.0;
211 evsel->evlist = NULL;
212 evsel->bpf_fd = -1;
213 INIT_LIST_HEAD(&evsel->node);
214 INIT_LIST_HEAD(&evsel->config_terms);
215 perf_evsel__object.init(evsel);
216 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
217 perf_evsel__calc_id_pos(evsel);
218 evsel->cmdline_group_boundary = false;
219}
220
221struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
222{
223 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
224
225 if (evsel != NULL)
226 perf_evsel__init(evsel, attr, idx);
227
228 if (perf_evsel__is_bpf_output(evsel)) {
229 evsel->attr.sample_type |= PERF_SAMPLE_RAW;
230 evsel->attr.sample_period = 1;
231 }
232
233 return evsel;
234}
235
236/*
237 * Returns pointer with encoded error via <linux/err.h> interface.
238 */
239struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
240{
241 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
242 int err = -ENOMEM;
243
244 if (evsel == NULL) {
245 goto out_err;
246 } else {
247 struct perf_event_attr attr = {
248 .type = PERF_TYPE_TRACEPOINT,
249 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
250 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
251 };
252
253 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
254 goto out_free;
255
256 evsel->tp_format = trace_event__tp_format(sys, name);
257 if (IS_ERR(evsel->tp_format)) {
258 err = PTR_ERR(evsel->tp_format);
259 goto out_free;
260 }
261
262 event_attr_init(&attr);
263 attr.config = evsel->tp_format->id;
264 attr.sample_period = 1;
265 perf_evsel__init(evsel, &attr, idx);
266 }
267
268 return evsel;
269
270out_free:
271 zfree(&evsel->name);
272 free(evsel);
273out_err:
274 return ERR_PTR(err);
275}
276
277const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
278 "cycles",
279 "instructions",
280 "cache-references",
281 "cache-misses",
282 "branches",
283 "branch-misses",
284 "bus-cycles",
285 "stalled-cycles-frontend",
286 "stalled-cycles-backend",
287 "ref-cycles",
288};
289
290static const char *__perf_evsel__hw_name(u64 config)
291{
292 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
293 return perf_evsel__hw_names[config];
294
295 return "unknown-hardware";
296}
297
298static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
299{
300 int colon = 0, r = 0;
301 struct perf_event_attr *attr = &evsel->attr;
302 bool exclude_guest_default = false;
303
304#define MOD_PRINT(context, mod) do { \
305 if (!attr->exclude_##context) { \
306 if (!colon) colon = ++r; \
307 r += scnprintf(bf + r, size - r, "%c", mod); \
308 } } while(0)
309
310 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
311 MOD_PRINT(kernel, 'k');
312 MOD_PRINT(user, 'u');
313 MOD_PRINT(hv, 'h');
314 exclude_guest_default = true;
315 }
316
317 if (attr->precise_ip) {
318 if (!colon)
319 colon = ++r;
320 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
321 exclude_guest_default = true;
322 }
323
324 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
325 MOD_PRINT(host, 'H');
326 MOD_PRINT(guest, 'G');
327 }
328#undef MOD_PRINT
329 if (colon)
330 bf[colon - 1] = ':';
331 return r;
332}
333
334static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
335{
336 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
337 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
338}
339
340const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
341 "cpu-clock",
342 "task-clock",
343 "page-faults",
344 "context-switches",
345 "cpu-migrations",
346 "minor-faults",
347 "major-faults",
348 "alignment-faults",
349 "emulation-faults",
350 "dummy",
351};
352
353static const char *__perf_evsel__sw_name(u64 config)
354{
355 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
356 return perf_evsel__sw_names[config];
357 return "unknown-software";
358}
359
360static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
361{
362 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
363 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
364}
365
366static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
367{
368 int r;
369
370 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
371
372 if (type & HW_BREAKPOINT_R)
373 r += scnprintf(bf + r, size - r, "r");
374
375 if (type & HW_BREAKPOINT_W)
376 r += scnprintf(bf + r, size - r, "w");
377
378 if (type & HW_BREAKPOINT_X)
379 r += scnprintf(bf + r, size - r, "x");
380
381 return r;
382}
383
384static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
385{
386 struct perf_event_attr *attr = &evsel->attr;
387 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
388 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
389}
390
391const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
392 [PERF_EVSEL__MAX_ALIASES] = {
393 { "L1-dcache", "l1-d", "l1d", "L1-data", },
394 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
395 { "LLC", "L2", },
396 { "dTLB", "d-tlb", "Data-TLB", },
397 { "iTLB", "i-tlb", "Instruction-TLB", },
398 { "branch", "branches", "bpu", "btb", "bpc", },
399 { "node", },
400};
401
402const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
403 [PERF_EVSEL__MAX_ALIASES] = {
404 { "load", "loads", "read", },
405 { "store", "stores", "write", },
406 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
407};
408
409const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
410 [PERF_EVSEL__MAX_ALIASES] = {
411 { "refs", "Reference", "ops", "access", },
412 { "misses", "miss", },
413};
414
415#define C(x) PERF_COUNT_HW_CACHE_##x
416#define CACHE_READ (1 << C(OP_READ))
417#define CACHE_WRITE (1 << C(OP_WRITE))
418#define CACHE_PREFETCH (1 << C(OP_PREFETCH))
419#define COP(x) (1 << x)
420
421/*
422 * cache operartion stat
423 * L1I : Read and prefetch only
424 * ITLB and BPU : Read-only
425 */
426static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
427 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
428 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
429 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
430 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
431 [C(ITLB)] = (CACHE_READ),
432 [C(BPU)] = (CACHE_READ),
433 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
434};
435
436bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
437{
438 if (perf_evsel__hw_cache_stat[type] & COP(op))
439 return true; /* valid */
440 else
441 return false; /* invalid */
442}
443
444int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
445 char *bf, size_t size)
446{
447 if (result) {
448 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
449 perf_evsel__hw_cache_op[op][0],
450 perf_evsel__hw_cache_result[result][0]);
451 }
452
453 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
454 perf_evsel__hw_cache_op[op][1]);
455}
456
457static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
458{
459 u8 op, result, type = (config >> 0) & 0xff;
460 const char *err = "unknown-ext-hardware-cache-type";
461
462 if (type > PERF_COUNT_HW_CACHE_MAX)
463 goto out_err;
464
465 op = (config >> 8) & 0xff;
466 err = "unknown-ext-hardware-cache-op";
467 if (op > PERF_COUNT_HW_CACHE_OP_MAX)
468 goto out_err;
469
470 result = (config >> 16) & 0xff;
471 err = "unknown-ext-hardware-cache-result";
472 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
473 goto out_err;
474
475 err = "invalid-cache";
476 if (!perf_evsel__is_cache_op_valid(type, op))
477 goto out_err;
478
479 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
480out_err:
481 return scnprintf(bf, size, "%s", err);
482}
483
484static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
485{
486 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
487 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
488}
489
490static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
491{
492 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
493 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
494}
495
496const char *perf_evsel__name(struct perf_evsel *evsel)
497{
498 char bf[128];
499
500 if (evsel->name)
501 return evsel->name;
502
503 switch (evsel->attr.type) {
504 case PERF_TYPE_RAW:
505 perf_evsel__raw_name(evsel, bf, sizeof(bf));
506 break;
507
508 case PERF_TYPE_HARDWARE:
509 perf_evsel__hw_name(evsel, bf, sizeof(bf));
510 break;
511
512 case PERF_TYPE_HW_CACHE:
513 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
514 break;
515
516 case PERF_TYPE_SOFTWARE:
517 perf_evsel__sw_name(evsel, bf, sizeof(bf));
518 break;
519
520 case PERF_TYPE_TRACEPOINT:
521 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
522 break;
523
524 case PERF_TYPE_BREAKPOINT:
525 perf_evsel__bp_name(evsel, bf, sizeof(bf));
526 break;
527
528 default:
529 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
530 evsel->attr.type);
531 break;
532 }
533
534 evsel->name = strdup(bf);
535
536 return evsel->name ?: "unknown";
537}
538
539const char *perf_evsel__group_name(struct perf_evsel *evsel)
540{
541 return evsel->group_name ?: "anon group";
542}
543
544int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
545{
546 int ret;
547 struct perf_evsel *pos;
548 const char *group_name = perf_evsel__group_name(evsel);
549
550 ret = scnprintf(buf, size, "%s", group_name);
551
552 ret += scnprintf(buf + ret, size - ret, " { %s",
553 perf_evsel__name(evsel));
554
555 for_each_group_member(pos, evsel)
556 ret += scnprintf(buf + ret, size - ret, ", %s",
557 perf_evsel__name(pos));
558
559 ret += scnprintf(buf + ret, size - ret, " }");
560
561 return ret;
562}
563
564static void
565perf_evsel__config_callgraph(struct perf_evsel *evsel,
566 struct record_opts *opts,
567 struct callchain_param *param)
568{
569 bool function = perf_evsel__is_function_event(evsel);
570 struct perf_event_attr *attr = &evsel->attr;
571
572 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
573
574 if (param->record_mode == CALLCHAIN_LBR) {
575 if (!opts->branch_stack) {
576 if (attr->exclude_user) {
577 pr_warning("LBR callstack option is only available "
578 "to get user callchain information. "
579 "Falling back to framepointers.\n");
580 } else {
581 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
582 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
583 PERF_SAMPLE_BRANCH_CALL_STACK |
584 PERF_SAMPLE_BRANCH_NO_CYCLES |
585 PERF_SAMPLE_BRANCH_NO_FLAGS;
586 }
587 } else
588 pr_warning("Cannot use LBR callstack with branch stack. "
589 "Falling back to framepointers.\n");
590 }
591
592 if (param->record_mode == CALLCHAIN_DWARF) {
593 if (!function) {
594 perf_evsel__set_sample_bit(evsel, REGS_USER);
595 perf_evsel__set_sample_bit(evsel, STACK_USER);
596 attr->sample_regs_user = PERF_REGS_MASK;
597 attr->sample_stack_user = param->dump_size;
598 attr->exclude_callchain_user = 1;
599 } else {
600 pr_info("Cannot use DWARF unwind for function trace event,"
601 " falling back to framepointers.\n");
602 }
603 }
604
605 if (function) {
606 pr_info("Disabling user space callchains for function trace event.\n");
607 attr->exclude_callchain_user = 1;
608 }
609}
610
611static void
612perf_evsel__reset_callgraph(struct perf_evsel *evsel,
613 struct callchain_param *param)
614{
615 struct perf_event_attr *attr = &evsel->attr;
616
617 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
618 if (param->record_mode == CALLCHAIN_LBR) {
619 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
620 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
621 PERF_SAMPLE_BRANCH_CALL_STACK);
622 }
623 if (param->record_mode == CALLCHAIN_DWARF) {
624 perf_evsel__reset_sample_bit(evsel, REGS_USER);
625 perf_evsel__reset_sample_bit(evsel, STACK_USER);
626 }
627}
628
629static void apply_config_terms(struct perf_evsel *evsel,
630 struct record_opts *opts)
631{
632 struct perf_evsel_config_term *term;
633 struct list_head *config_terms = &evsel->config_terms;
634 struct perf_event_attr *attr = &evsel->attr;
635 struct callchain_param param;
636 u32 dump_size = 0;
637 char *callgraph_buf = NULL;
638
639 /* callgraph default */
640 param.record_mode = callchain_param.record_mode;
641
642 list_for_each_entry(term, config_terms, list) {
643 switch (term->type) {
644 case PERF_EVSEL__CONFIG_TERM_PERIOD:
645 attr->sample_period = term->val.period;
646 attr->freq = 0;
647 break;
648 case PERF_EVSEL__CONFIG_TERM_FREQ:
649 attr->sample_freq = term->val.freq;
650 attr->freq = 1;
651 break;
652 case PERF_EVSEL__CONFIG_TERM_TIME:
653 if (term->val.time)
654 perf_evsel__set_sample_bit(evsel, TIME);
655 else
656 perf_evsel__reset_sample_bit(evsel, TIME);
657 break;
658 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
659 callgraph_buf = term->val.callgraph;
660 break;
661 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
662 dump_size = term->val.stack_user;
663 break;
664 case PERF_EVSEL__CONFIG_TERM_INHERIT:
665 /*
666 * attr->inherit should has already been set by
667 * perf_evsel__config. If user explicitly set
668 * inherit using config terms, override global
669 * opt->no_inherit setting.
670 */
671 attr->inherit = term->val.inherit ? 1 : 0;
672 break;
673 default:
674 break;
675 }
676 }
677
678 /* User explicitly set per-event callgraph, clear the old setting and reset. */
679 if ((callgraph_buf != NULL) || (dump_size > 0)) {
680
681 /* parse callgraph parameters */
682 if (callgraph_buf != NULL) {
683 if (!strcmp(callgraph_buf, "no")) {
684 param.enabled = false;
685 param.record_mode = CALLCHAIN_NONE;
686 } else {
687 param.enabled = true;
688 if (parse_callchain_record(callgraph_buf, ¶m)) {
689 pr_err("per-event callgraph setting for %s failed. "
690 "Apply callgraph global setting for it\n",
691 evsel->name);
692 return;
693 }
694 }
695 }
696 if (dump_size > 0) {
697 dump_size = round_up(dump_size, sizeof(u64));
698 param.dump_size = dump_size;
699 }
700
701 /* If global callgraph set, clear it */
702 if (callchain_param.enabled)
703 perf_evsel__reset_callgraph(evsel, &callchain_param);
704
705 /* set perf-event callgraph */
706 if (param.enabled)
707 perf_evsel__config_callgraph(evsel, opts, ¶m);
708 }
709}
710
711/*
712 * The enable_on_exec/disabled value strategy:
713 *
714 * 1) For any type of traced program:
715 * - all independent events and group leaders are disabled
716 * - all group members are enabled
717 *
718 * Group members are ruled by group leaders. They need to
719 * be enabled, because the group scheduling relies on that.
720 *
721 * 2) For traced programs executed by perf:
722 * - all independent events and group leaders have
723 * enable_on_exec set
724 * - we don't specifically enable or disable any event during
725 * the record command
726 *
727 * Independent events and group leaders are initially disabled
728 * and get enabled by exec. Group members are ruled by group
729 * leaders as stated in 1).
730 *
731 * 3) For traced programs attached by perf (pid/tid):
732 * - we specifically enable or disable all events during
733 * the record command
734 *
735 * When attaching events to already running traced we
736 * enable/disable events specifically, as there's no
737 * initial traced exec call.
738 */
739void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
740{
741 struct perf_evsel *leader = evsel->leader;
742 struct perf_event_attr *attr = &evsel->attr;
743 int track = evsel->tracking;
744 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
745
746 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
747 attr->inherit = !opts->no_inherit;
748
749 perf_evsel__set_sample_bit(evsel, IP);
750 perf_evsel__set_sample_bit(evsel, TID);
751
752 if (evsel->sample_read) {
753 perf_evsel__set_sample_bit(evsel, READ);
754
755 /*
756 * We need ID even in case of single event, because
757 * PERF_SAMPLE_READ process ID specific data.
758 */
759 perf_evsel__set_sample_id(evsel, false);
760
761 /*
762 * Apply group format only if we belong to group
763 * with more than one members.
764 */
765 if (leader->nr_members > 1) {
766 attr->read_format |= PERF_FORMAT_GROUP;
767 attr->inherit = 0;
768 }
769 }
770
771 /*
772 * We default some events to have a default interval. But keep
773 * it a weak assumption overridable by the user.
774 */
775 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
776 opts->user_interval != ULLONG_MAX)) {
777 if (opts->freq) {
778 perf_evsel__set_sample_bit(evsel, PERIOD);
779 attr->freq = 1;
780 attr->sample_freq = opts->freq;
781 } else {
782 attr->sample_period = opts->default_interval;
783 }
784 }
785
786 /*
787 * Disable sampling for all group members other
788 * than leader in case leader 'leads' the sampling.
789 */
790 if ((leader != evsel) && leader->sample_read) {
791 attr->sample_freq = 0;
792 attr->sample_period = 0;
793 }
794
795 if (opts->no_samples)
796 attr->sample_freq = 0;
797
798 if (opts->inherit_stat)
799 attr->inherit_stat = 1;
800
801 if (opts->sample_address) {
802 perf_evsel__set_sample_bit(evsel, ADDR);
803 attr->mmap_data = track;
804 }
805
806 /*
807 * We don't allow user space callchains for function trace
808 * event, due to issues with page faults while tracing page
809 * fault handler and its overall trickiness nature.
810 */
811 if (perf_evsel__is_function_event(evsel))
812 evsel->attr.exclude_callchain_user = 1;
813
814 if (callchain_param.enabled && !evsel->no_aux_samples)
815 perf_evsel__config_callgraph(evsel, opts, &callchain_param);
816
817 if (opts->sample_intr_regs) {
818 attr->sample_regs_intr = opts->sample_intr_regs;
819 perf_evsel__set_sample_bit(evsel, REGS_INTR);
820 }
821
822 if (target__has_cpu(&opts->target))
823 perf_evsel__set_sample_bit(evsel, CPU);
824
825 if (opts->period)
826 perf_evsel__set_sample_bit(evsel, PERIOD);
827
828 /*
829 * When the user explicitely disabled time don't force it here.
830 */
831 if (opts->sample_time &&
832 (!perf_missing_features.sample_id_all &&
833 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
834 opts->sample_time_set)))
835 perf_evsel__set_sample_bit(evsel, TIME);
836
837 if (opts->raw_samples && !evsel->no_aux_samples) {
838 perf_evsel__set_sample_bit(evsel, TIME);
839 perf_evsel__set_sample_bit(evsel, RAW);
840 perf_evsel__set_sample_bit(evsel, CPU);
841 }
842
843 if (opts->sample_address)
844 perf_evsel__set_sample_bit(evsel, DATA_SRC);
845
846 if (opts->no_buffering) {
847 attr->watermark = 0;
848 attr->wakeup_events = 1;
849 }
850 if (opts->branch_stack && !evsel->no_aux_samples) {
851 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
852 attr->branch_sample_type = opts->branch_stack;
853 }
854
855 if (opts->sample_weight)
856 perf_evsel__set_sample_bit(evsel, WEIGHT);
857
858 attr->task = track;
859 attr->mmap = track;
860 attr->mmap2 = track && !perf_missing_features.mmap2;
861 attr->comm = track;
862
863 if (opts->record_switch_events)
864 attr->context_switch = track;
865
866 if (opts->sample_transaction)
867 perf_evsel__set_sample_bit(evsel, TRANSACTION);
868
869 if (opts->running_time) {
870 evsel->attr.read_format |=
871 PERF_FORMAT_TOTAL_TIME_ENABLED |
872 PERF_FORMAT_TOTAL_TIME_RUNNING;
873 }
874
875 /*
876 * XXX see the function comment above
877 *
878 * Disabling only independent events or group leaders,
879 * keeping group members enabled.
880 */
881 if (perf_evsel__is_group_leader(evsel))
882 attr->disabled = 1;
883
884 /*
885 * Setting enable_on_exec for independent events and
886 * group leaders for traced executed by perf.
887 */
888 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
889 !opts->initial_delay)
890 attr->enable_on_exec = 1;
891
892 if (evsel->immediate) {
893 attr->disabled = 0;
894 attr->enable_on_exec = 0;
895 }
896
897 clockid = opts->clockid;
898 if (opts->use_clockid) {
899 attr->use_clockid = 1;
900 attr->clockid = opts->clockid;
901 }
902
903 if (evsel->precise_max)
904 perf_event_attr__set_max_precise_ip(attr);
905
906 if (opts->all_user) {
907 attr->exclude_kernel = 1;
908 attr->exclude_user = 0;
909 }
910
911 if (opts->all_kernel) {
912 attr->exclude_kernel = 0;
913 attr->exclude_user = 1;
914 }
915
916 /*
917 * Apply event specific term settings,
918 * it overloads any global configuration.
919 */
920 apply_config_terms(evsel, opts);
921}
922
923static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
924{
925 int cpu, thread;
926
927 if (evsel->system_wide)
928 nthreads = 1;
929
930 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
931
932 if (evsel->fd) {
933 for (cpu = 0; cpu < ncpus; cpu++) {
934 for (thread = 0; thread < nthreads; thread++) {
935 FD(evsel, cpu, thread) = -1;
936 }
937 }
938 }
939
940 return evsel->fd != NULL ? 0 : -ENOMEM;
941}
942
943static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
944 int ioc, void *arg)
945{
946 int cpu, thread;
947
948 if (evsel->system_wide)
949 nthreads = 1;
950
951 for (cpu = 0; cpu < ncpus; cpu++) {
952 for (thread = 0; thread < nthreads; thread++) {
953 int fd = FD(evsel, cpu, thread),
954 err = ioctl(fd, ioc, arg);
955
956 if (err)
957 return err;
958 }
959 }
960
961 return 0;
962}
963
964int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
965 const char *filter)
966{
967 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
968 PERF_EVENT_IOC_SET_FILTER,
969 (void *)filter);
970}
971
972int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
973{
974 char *new_filter = strdup(filter);
975
976 if (new_filter != NULL) {
977 free(evsel->filter);
978 evsel->filter = new_filter;
979 return 0;
980 }
981
982 return -1;
983}
984
985int perf_evsel__append_filter(struct perf_evsel *evsel,
986 const char *op, const char *filter)
987{
988 char *new_filter;
989
990 if (evsel->filter == NULL)
991 return perf_evsel__set_filter(evsel, filter);
992
993 if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
994 free(evsel->filter);
995 evsel->filter = new_filter;
996 return 0;
997 }
998
999 return -1;
1000}
1001
1002int perf_evsel__enable(struct perf_evsel *evsel)
1003{
1004 int nthreads = thread_map__nr(evsel->threads);
1005 int ncpus = cpu_map__nr(evsel->cpus);
1006
1007 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1008 PERF_EVENT_IOC_ENABLE,
1009 0);
1010}
1011
1012int perf_evsel__disable(struct perf_evsel *evsel)
1013{
1014 int nthreads = thread_map__nr(evsel->threads);
1015 int ncpus = cpu_map__nr(evsel->cpus);
1016
1017 return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1018 PERF_EVENT_IOC_DISABLE,
1019 0);
1020}
1021
1022int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1023{
1024 if (ncpus == 0 || nthreads == 0)
1025 return 0;
1026
1027 if (evsel->system_wide)
1028 nthreads = 1;
1029
1030 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1031 if (evsel->sample_id == NULL)
1032 return -ENOMEM;
1033
1034 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1035 if (evsel->id == NULL) {
1036 xyarray__delete(evsel->sample_id);
1037 evsel->sample_id = NULL;
1038 return -ENOMEM;
1039 }
1040
1041 return 0;
1042}
1043
1044static void perf_evsel__free_fd(struct perf_evsel *evsel)
1045{
1046 xyarray__delete(evsel->fd);
1047 evsel->fd = NULL;
1048}
1049
1050static void perf_evsel__free_id(struct perf_evsel *evsel)
1051{
1052 xyarray__delete(evsel->sample_id);
1053 evsel->sample_id = NULL;
1054 zfree(&evsel->id);
1055}
1056
1057static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1058{
1059 struct perf_evsel_config_term *term, *h;
1060
1061 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1062 list_del(&term->list);
1063 free(term);
1064 }
1065}
1066
1067void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1068{
1069 int cpu, thread;
1070
1071 if (evsel->system_wide)
1072 nthreads = 1;
1073
1074 for (cpu = 0; cpu < ncpus; cpu++)
1075 for (thread = 0; thread < nthreads; ++thread) {
1076 close(FD(evsel, cpu, thread));
1077 FD(evsel, cpu, thread) = -1;
1078 }
1079}
1080
1081void perf_evsel__exit(struct perf_evsel *evsel)
1082{
1083 assert(list_empty(&evsel->node));
1084 assert(evsel->evlist == NULL);
1085 perf_evsel__free_fd(evsel);
1086 perf_evsel__free_id(evsel);
1087 perf_evsel__free_config_terms(evsel);
1088 close_cgroup(evsel->cgrp);
1089 cpu_map__put(evsel->cpus);
1090 cpu_map__put(evsel->own_cpus);
1091 thread_map__put(evsel->threads);
1092 zfree(&evsel->group_name);
1093 zfree(&evsel->name);
1094 perf_evsel__object.fini(evsel);
1095}
1096
1097void perf_evsel__delete(struct perf_evsel *evsel)
1098{
1099 perf_evsel__exit(evsel);
1100 free(evsel);
1101}
1102
1103void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1104 struct perf_counts_values *count)
1105{
1106 struct perf_counts_values tmp;
1107
1108 if (!evsel->prev_raw_counts)
1109 return;
1110
1111 if (cpu == -1) {
1112 tmp = evsel->prev_raw_counts->aggr;
1113 evsel->prev_raw_counts->aggr = *count;
1114 } else {
1115 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1116 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1117 }
1118
1119 count->val = count->val - tmp.val;
1120 count->ena = count->ena - tmp.ena;
1121 count->run = count->run - tmp.run;
1122}
1123
1124void perf_counts_values__scale(struct perf_counts_values *count,
1125 bool scale, s8 *pscaled)
1126{
1127 s8 scaled = 0;
1128
1129 if (scale) {
1130 if (count->run == 0) {
1131 scaled = -1;
1132 count->val = 0;
1133 } else if (count->run < count->ena) {
1134 scaled = 1;
1135 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1136 }
1137 } else
1138 count->ena = count->run = 0;
1139
1140 if (pscaled)
1141 *pscaled = scaled;
1142}
1143
1144int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1145 struct perf_counts_values *count)
1146{
1147 memset(count, 0, sizeof(*count));
1148
1149 if (FD(evsel, cpu, thread) < 0)
1150 return -EINVAL;
1151
1152 if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1153 return -errno;
1154
1155 return 0;
1156}
1157
1158int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1159 int cpu, int thread, bool scale)
1160{
1161 struct perf_counts_values count;
1162 size_t nv = scale ? 3 : 1;
1163
1164 if (FD(evsel, cpu, thread) < 0)
1165 return -EINVAL;
1166
1167 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1168 return -ENOMEM;
1169
1170 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1171 return -errno;
1172
1173 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1174 perf_counts_values__scale(&count, scale, NULL);
1175 *perf_counts(evsel->counts, cpu, thread) = count;
1176 return 0;
1177}
1178
1179static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1180{
1181 struct perf_evsel *leader = evsel->leader;
1182 int fd;
1183
1184 if (perf_evsel__is_group_leader(evsel))
1185 return -1;
1186
1187 /*
1188 * Leader must be already processed/open,
1189 * if not it's a bug.
1190 */
1191 BUG_ON(!leader->fd);
1192
1193 fd = FD(leader, cpu, thread);
1194 BUG_ON(fd == -1);
1195
1196 return fd;
1197}
1198
1199struct bit_names {
1200 int bit;
1201 const char *name;
1202};
1203
1204static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1205{
1206 bool first_bit = true;
1207 int i = 0;
1208
1209 do {
1210 if (value & bits[i].bit) {
1211 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1212 first_bit = false;
1213 }
1214 } while (bits[++i].name != NULL);
1215}
1216
1217static void __p_sample_type(char *buf, size_t size, u64 value)
1218{
1219#define bit_name(n) { PERF_SAMPLE_##n, #n }
1220 struct bit_names bits[] = {
1221 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1222 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1223 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1224 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1225 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1226 bit_name(WEIGHT),
1227 { .name = NULL, }
1228 };
1229#undef bit_name
1230 __p_bits(buf, size, value, bits);
1231}
1232
1233static void __p_read_format(char *buf, size_t size, u64 value)
1234{
1235#define bit_name(n) { PERF_FORMAT_##n, #n }
1236 struct bit_names bits[] = {
1237 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1238 bit_name(ID), bit_name(GROUP),
1239 { .name = NULL, }
1240 };
1241#undef bit_name
1242 __p_bits(buf, size, value, bits);
1243}
1244
1245#define BUF_SIZE 1024
1246
1247#define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1248#define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1249#define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1250#define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1251#define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1252
1253#define PRINT_ATTRn(_n, _f, _p) \
1254do { \
1255 if (attr->_f) { \
1256 _p(attr->_f); \
1257 ret += attr__fprintf(fp, _n, buf, priv);\
1258 } \
1259} while (0)
1260
1261#define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1262
1263int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1264 attr__fprintf_f attr__fprintf, void *priv)
1265{
1266 char buf[BUF_SIZE];
1267 int ret = 0;
1268
1269 PRINT_ATTRf(type, p_unsigned);
1270 PRINT_ATTRf(size, p_unsigned);
1271 PRINT_ATTRf(config, p_hex);
1272 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1273 PRINT_ATTRf(sample_type, p_sample_type);
1274 PRINT_ATTRf(read_format, p_read_format);
1275
1276 PRINT_ATTRf(disabled, p_unsigned);
1277 PRINT_ATTRf(inherit, p_unsigned);
1278 PRINT_ATTRf(pinned, p_unsigned);
1279 PRINT_ATTRf(exclusive, p_unsigned);
1280 PRINT_ATTRf(exclude_user, p_unsigned);
1281 PRINT_ATTRf(exclude_kernel, p_unsigned);
1282 PRINT_ATTRf(exclude_hv, p_unsigned);
1283 PRINT_ATTRf(exclude_idle, p_unsigned);
1284 PRINT_ATTRf(mmap, p_unsigned);
1285 PRINT_ATTRf(comm, p_unsigned);
1286 PRINT_ATTRf(freq, p_unsigned);
1287 PRINT_ATTRf(inherit_stat, p_unsigned);
1288 PRINT_ATTRf(enable_on_exec, p_unsigned);
1289 PRINT_ATTRf(task, p_unsigned);
1290 PRINT_ATTRf(watermark, p_unsigned);
1291 PRINT_ATTRf(precise_ip, p_unsigned);
1292 PRINT_ATTRf(mmap_data, p_unsigned);
1293 PRINT_ATTRf(sample_id_all, p_unsigned);
1294 PRINT_ATTRf(exclude_host, p_unsigned);
1295 PRINT_ATTRf(exclude_guest, p_unsigned);
1296 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1297 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1298 PRINT_ATTRf(mmap2, p_unsigned);
1299 PRINT_ATTRf(comm_exec, p_unsigned);
1300 PRINT_ATTRf(use_clockid, p_unsigned);
1301 PRINT_ATTRf(context_switch, p_unsigned);
1302
1303 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1304 PRINT_ATTRf(bp_type, p_unsigned);
1305 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1306 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1307 PRINT_ATTRf(branch_sample_type, p_unsigned);
1308 PRINT_ATTRf(sample_regs_user, p_hex);
1309 PRINT_ATTRf(sample_stack_user, p_unsigned);
1310 PRINT_ATTRf(clockid, p_signed);
1311 PRINT_ATTRf(sample_regs_intr, p_hex);
1312 PRINT_ATTRf(aux_watermark, p_unsigned);
1313
1314 return ret;
1315}
1316
1317static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1318 void *priv __attribute__((unused)))
1319{
1320 return fprintf(fp, " %-32s %s\n", name, val);
1321}
1322
1323static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1324 struct thread_map *threads)
1325{
1326 int cpu, thread, nthreads;
1327 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1328 int pid = -1, err;
1329 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1330
1331 if (evsel->system_wide)
1332 nthreads = 1;
1333 else
1334 nthreads = threads->nr;
1335
1336 if (evsel->fd == NULL &&
1337 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1338 return -ENOMEM;
1339
1340 if (evsel->cgrp) {
1341 flags |= PERF_FLAG_PID_CGROUP;
1342 pid = evsel->cgrp->fd;
1343 }
1344
1345fallback_missing_features:
1346 if (perf_missing_features.clockid_wrong)
1347 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1348 if (perf_missing_features.clockid) {
1349 evsel->attr.use_clockid = 0;
1350 evsel->attr.clockid = 0;
1351 }
1352 if (perf_missing_features.cloexec)
1353 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1354 if (perf_missing_features.mmap2)
1355 evsel->attr.mmap2 = 0;
1356 if (perf_missing_features.exclude_guest)
1357 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1358 if (perf_missing_features.lbr_flags)
1359 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1360 PERF_SAMPLE_BRANCH_NO_CYCLES);
1361retry_sample_id:
1362 if (perf_missing_features.sample_id_all)
1363 evsel->attr.sample_id_all = 0;
1364
1365 if (verbose >= 2) {
1366 fprintf(stderr, "%.60s\n", graph_dotted_line);
1367 fprintf(stderr, "perf_event_attr:\n");
1368 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1369 fprintf(stderr, "%.60s\n", graph_dotted_line);
1370 }
1371
1372 for (cpu = 0; cpu < cpus->nr; cpu++) {
1373
1374 for (thread = 0; thread < nthreads; thread++) {
1375 int group_fd;
1376
1377 if (!evsel->cgrp && !evsel->system_wide)
1378 pid = thread_map__pid(threads, thread);
1379
1380 group_fd = get_group_fd(evsel, cpu, thread);
1381retry_open:
1382 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n",
1383 pid, cpus->map[cpu], group_fd, flags);
1384
1385 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1386 pid,
1387 cpus->map[cpu],
1388 group_fd, flags);
1389 if (FD(evsel, cpu, thread) < 0) {
1390 err = -errno;
1391 pr_debug2("sys_perf_event_open failed, error %d\n",
1392 err);
1393 goto try_fallback;
1394 }
1395
1396 if (evsel->bpf_fd >= 0) {
1397 int evt_fd = FD(evsel, cpu, thread);
1398 int bpf_fd = evsel->bpf_fd;
1399
1400 err = ioctl(evt_fd,
1401 PERF_EVENT_IOC_SET_BPF,
1402 bpf_fd);
1403 if (err && errno != EEXIST) {
1404 pr_err("failed to attach bpf fd %d: %s\n",
1405 bpf_fd, strerror(errno));
1406 err = -EINVAL;
1407 goto out_close;
1408 }
1409 }
1410
1411 set_rlimit = NO_CHANGE;
1412
1413 /*
1414 * If we succeeded but had to kill clockid, fail and
1415 * have perf_evsel__open_strerror() print us a nice
1416 * error.
1417 */
1418 if (perf_missing_features.clockid ||
1419 perf_missing_features.clockid_wrong) {
1420 err = -EINVAL;
1421 goto out_close;
1422 }
1423 }
1424 }
1425
1426 return 0;
1427
1428try_fallback:
1429 /*
1430 * perf stat needs between 5 and 22 fds per CPU. When we run out
1431 * of them try to increase the limits.
1432 */
1433 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1434 struct rlimit l;
1435 int old_errno = errno;
1436
1437 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1438 if (set_rlimit == NO_CHANGE)
1439 l.rlim_cur = l.rlim_max;
1440 else {
1441 l.rlim_cur = l.rlim_max + 1000;
1442 l.rlim_max = l.rlim_cur;
1443 }
1444 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1445 set_rlimit++;
1446 errno = old_errno;
1447 goto retry_open;
1448 }
1449 }
1450 errno = old_errno;
1451 }
1452
1453 if (err != -EINVAL || cpu > 0 || thread > 0)
1454 goto out_close;
1455
1456 /*
1457 * Must probe features in the order they were added to the
1458 * perf_event_attr interface.
1459 */
1460 if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1461 perf_missing_features.clockid_wrong = true;
1462 goto fallback_missing_features;
1463 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1464 perf_missing_features.clockid = true;
1465 goto fallback_missing_features;
1466 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1467 perf_missing_features.cloexec = true;
1468 goto fallback_missing_features;
1469 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1470 perf_missing_features.mmap2 = true;
1471 goto fallback_missing_features;
1472 } else if (!perf_missing_features.exclude_guest &&
1473 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1474 perf_missing_features.exclude_guest = true;
1475 goto fallback_missing_features;
1476 } else if (!perf_missing_features.sample_id_all) {
1477 perf_missing_features.sample_id_all = true;
1478 goto retry_sample_id;
1479 } else if (!perf_missing_features.lbr_flags &&
1480 (evsel->attr.branch_sample_type &
1481 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1482 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1483 perf_missing_features.lbr_flags = true;
1484 goto fallback_missing_features;
1485 }
1486
1487out_close:
1488 do {
1489 while (--thread >= 0) {
1490 close(FD(evsel, cpu, thread));
1491 FD(evsel, cpu, thread) = -1;
1492 }
1493 thread = nthreads;
1494 } while (--cpu >= 0);
1495 return err;
1496}
1497
1498void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1499{
1500 if (evsel->fd == NULL)
1501 return;
1502
1503 perf_evsel__close_fd(evsel, ncpus, nthreads);
1504 perf_evsel__free_fd(evsel);
1505}
1506
1507static struct {
1508 struct cpu_map map;
1509 int cpus[1];
1510} empty_cpu_map = {
1511 .map.nr = 1,
1512 .cpus = { -1, },
1513};
1514
1515static struct {
1516 struct thread_map map;
1517 int threads[1];
1518} empty_thread_map = {
1519 .map.nr = 1,
1520 .threads = { -1, },
1521};
1522
1523int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1524 struct thread_map *threads)
1525{
1526 if (cpus == NULL) {
1527 /* Work around old compiler warnings about strict aliasing */
1528 cpus = &empty_cpu_map.map;
1529 }
1530
1531 if (threads == NULL)
1532 threads = &empty_thread_map.map;
1533
1534 return __perf_evsel__open(evsel, cpus, threads);
1535}
1536
1537int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1538 struct cpu_map *cpus)
1539{
1540 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1541}
1542
1543int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1544 struct thread_map *threads)
1545{
1546 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1547}
1548
1549static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1550 const union perf_event *event,
1551 struct perf_sample *sample)
1552{
1553 u64 type = evsel->attr.sample_type;
1554 const u64 *array = event->sample.array;
1555 bool swapped = evsel->needs_swap;
1556 union u64_swap u;
1557
1558 array += ((event->header.size -
1559 sizeof(event->header)) / sizeof(u64)) - 1;
1560
1561 if (type & PERF_SAMPLE_IDENTIFIER) {
1562 sample->id = *array;
1563 array--;
1564 }
1565
1566 if (type & PERF_SAMPLE_CPU) {
1567 u.val64 = *array;
1568 if (swapped) {
1569 /* undo swap of u64, then swap on individual u32s */
1570 u.val64 = bswap_64(u.val64);
1571 u.val32[0] = bswap_32(u.val32[0]);
1572 }
1573
1574 sample->cpu = u.val32[0];
1575 array--;
1576 }
1577
1578 if (type & PERF_SAMPLE_STREAM_ID) {
1579 sample->stream_id = *array;
1580 array--;
1581 }
1582
1583 if (type & PERF_SAMPLE_ID) {
1584 sample->id = *array;
1585 array--;
1586 }
1587
1588 if (type & PERF_SAMPLE_TIME) {
1589 sample->time = *array;
1590 array--;
1591 }
1592
1593 if (type & PERF_SAMPLE_TID) {
1594 u.val64 = *array;
1595 if (swapped) {
1596 /* undo swap of u64, then swap on individual u32s */
1597 u.val64 = bswap_64(u.val64);
1598 u.val32[0] = bswap_32(u.val32[0]);
1599 u.val32[1] = bswap_32(u.val32[1]);
1600 }
1601
1602 sample->pid = u.val32[0];
1603 sample->tid = u.val32[1];
1604 array--;
1605 }
1606
1607 return 0;
1608}
1609
1610static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1611 u64 size)
1612{
1613 return size > max_size || offset + size > endp;
1614}
1615
1616#define OVERFLOW_CHECK(offset, size, max_size) \
1617 do { \
1618 if (overflow(endp, (max_size), (offset), (size))) \
1619 return -EFAULT; \
1620 } while (0)
1621
1622#define OVERFLOW_CHECK_u64(offset) \
1623 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1624
1625int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1626 struct perf_sample *data)
1627{
1628 u64 type = evsel->attr.sample_type;
1629 bool swapped = evsel->needs_swap;
1630 const u64 *array;
1631 u16 max_size = event->header.size;
1632 const void *endp = (void *)event + max_size;
1633 u64 sz;
1634
1635 /*
1636 * used for cross-endian analysis. See git commit 65014ab3
1637 * for why this goofiness is needed.
1638 */
1639 union u64_swap u;
1640
1641 memset(data, 0, sizeof(*data));
1642 data->cpu = data->pid = data->tid = -1;
1643 data->stream_id = data->id = data->time = -1ULL;
1644 data->period = evsel->attr.sample_period;
1645 data->weight = 0;
1646 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1647
1648 if (event->header.type != PERF_RECORD_SAMPLE) {
1649 if (!evsel->attr.sample_id_all)
1650 return 0;
1651 return perf_evsel__parse_id_sample(evsel, event, data);
1652 }
1653
1654 array = event->sample.array;
1655
1656 /*
1657 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1658 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1659 * check the format does not go past the end of the event.
1660 */
1661 if (evsel->sample_size + sizeof(event->header) > event->header.size)
1662 return -EFAULT;
1663
1664 data->id = -1ULL;
1665 if (type & PERF_SAMPLE_IDENTIFIER) {
1666 data->id = *array;
1667 array++;
1668 }
1669
1670 if (type & PERF_SAMPLE_IP) {
1671 data->ip = *array;
1672 array++;
1673 }
1674
1675 if (type & PERF_SAMPLE_TID) {
1676 u.val64 = *array;
1677 if (swapped) {
1678 /* undo swap of u64, then swap on individual u32s */
1679 u.val64 = bswap_64(u.val64);
1680 u.val32[0] = bswap_32(u.val32[0]);
1681 u.val32[1] = bswap_32(u.val32[1]);
1682 }
1683
1684 data->pid = u.val32[0];
1685 data->tid = u.val32[1];
1686 array++;
1687 }
1688
1689 if (type & PERF_SAMPLE_TIME) {
1690 data->time = *array;
1691 array++;
1692 }
1693
1694 data->addr = 0;
1695 if (type & PERF_SAMPLE_ADDR) {
1696 data->addr = *array;
1697 array++;
1698 }
1699
1700 if (type & PERF_SAMPLE_ID) {
1701 data->id = *array;
1702 array++;
1703 }
1704
1705 if (type & PERF_SAMPLE_STREAM_ID) {
1706 data->stream_id = *array;
1707 array++;
1708 }
1709
1710 if (type & PERF_SAMPLE_CPU) {
1711
1712 u.val64 = *array;
1713 if (swapped) {
1714 /* undo swap of u64, then swap on individual u32s */
1715 u.val64 = bswap_64(u.val64);
1716 u.val32[0] = bswap_32(u.val32[0]);
1717 }
1718
1719 data->cpu = u.val32[0];
1720 array++;
1721 }
1722
1723 if (type & PERF_SAMPLE_PERIOD) {
1724 data->period = *array;
1725 array++;
1726 }
1727
1728 if (type & PERF_SAMPLE_READ) {
1729 u64 read_format = evsel->attr.read_format;
1730
1731 OVERFLOW_CHECK_u64(array);
1732 if (read_format & PERF_FORMAT_GROUP)
1733 data->read.group.nr = *array;
1734 else
1735 data->read.one.value = *array;
1736
1737 array++;
1738
1739 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1740 OVERFLOW_CHECK_u64(array);
1741 data->read.time_enabled = *array;
1742 array++;
1743 }
1744
1745 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1746 OVERFLOW_CHECK_u64(array);
1747 data->read.time_running = *array;
1748 array++;
1749 }
1750
1751 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1752 if (read_format & PERF_FORMAT_GROUP) {
1753 const u64 max_group_nr = UINT64_MAX /
1754 sizeof(struct sample_read_value);
1755
1756 if (data->read.group.nr > max_group_nr)
1757 return -EFAULT;
1758 sz = data->read.group.nr *
1759 sizeof(struct sample_read_value);
1760 OVERFLOW_CHECK(array, sz, max_size);
1761 data->read.group.values =
1762 (struct sample_read_value *)array;
1763 array = (void *)array + sz;
1764 } else {
1765 OVERFLOW_CHECK_u64(array);
1766 data->read.one.id = *array;
1767 array++;
1768 }
1769 }
1770
1771 if (type & PERF_SAMPLE_CALLCHAIN) {
1772 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1773
1774 OVERFLOW_CHECK_u64(array);
1775 data->callchain = (struct ip_callchain *)array++;
1776 if (data->callchain->nr > max_callchain_nr)
1777 return -EFAULT;
1778 sz = data->callchain->nr * sizeof(u64);
1779 OVERFLOW_CHECK(array, sz, max_size);
1780 array = (void *)array + sz;
1781 }
1782
1783 if (type & PERF_SAMPLE_RAW) {
1784 OVERFLOW_CHECK_u64(array);
1785 u.val64 = *array;
1786 if (WARN_ONCE(swapped,
1787 "Endianness of raw data not corrected!\n")) {
1788 /* undo swap of u64, then swap on individual u32s */
1789 u.val64 = bswap_64(u.val64);
1790 u.val32[0] = bswap_32(u.val32[0]);
1791 u.val32[1] = bswap_32(u.val32[1]);
1792 }
1793 data->raw_size = u.val32[0];
1794 array = (void *)array + sizeof(u32);
1795
1796 OVERFLOW_CHECK(array, data->raw_size, max_size);
1797 data->raw_data = (void *)array;
1798 array = (void *)array + data->raw_size;
1799 }
1800
1801 if (type & PERF_SAMPLE_BRANCH_STACK) {
1802 const u64 max_branch_nr = UINT64_MAX /
1803 sizeof(struct branch_entry);
1804
1805 OVERFLOW_CHECK_u64(array);
1806 data->branch_stack = (struct branch_stack *)array++;
1807
1808 if (data->branch_stack->nr > max_branch_nr)
1809 return -EFAULT;
1810 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1811 OVERFLOW_CHECK(array, sz, max_size);
1812 array = (void *)array + sz;
1813 }
1814
1815 if (type & PERF_SAMPLE_REGS_USER) {
1816 OVERFLOW_CHECK_u64(array);
1817 data->user_regs.abi = *array;
1818 array++;
1819
1820 if (data->user_regs.abi) {
1821 u64 mask = evsel->attr.sample_regs_user;
1822
1823 sz = hweight_long(mask) * sizeof(u64);
1824 OVERFLOW_CHECK(array, sz, max_size);
1825 data->user_regs.mask = mask;
1826 data->user_regs.regs = (u64 *)array;
1827 array = (void *)array + sz;
1828 }
1829 }
1830
1831 if (type & PERF_SAMPLE_STACK_USER) {
1832 OVERFLOW_CHECK_u64(array);
1833 sz = *array++;
1834
1835 data->user_stack.offset = ((char *)(array - 1)
1836 - (char *) event);
1837
1838 if (!sz) {
1839 data->user_stack.size = 0;
1840 } else {
1841 OVERFLOW_CHECK(array, sz, max_size);
1842 data->user_stack.data = (char *)array;
1843 array = (void *)array + sz;
1844 OVERFLOW_CHECK_u64(array);
1845 data->user_stack.size = *array++;
1846 if (WARN_ONCE(data->user_stack.size > sz,
1847 "user stack dump failure\n"))
1848 return -EFAULT;
1849 }
1850 }
1851
1852 data->weight = 0;
1853 if (type & PERF_SAMPLE_WEIGHT) {
1854 OVERFLOW_CHECK_u64(array);
1855 data->weight = *array;
1856 array++;
1857 }
1858
1859 data->data_src = PERF_MEM_DATA_SRC_NONE;
1860 if (type & PERF_SAMPLE_DATA_SRC) {
1861 OVERFLOW_CHECK_u64(array);
1862 data->data_src = *array;
1863 array++;
1864 }
1865
1866 data->transaction = 0;
1867 if (type & PERF_SAMPLE_TRANSACTION) {
1868 OVERFLOW_CHECK_u64(array);
1869 data->transaction = *array;
1870 array++;
1871 }
1872
1873 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1874 if (type & PERF_SAMPLE_REGS_INTR) {
1875 OVERFLOW_CHECK_u64(array);
1876 data->intr_regs.abi = *array;
1877 array++;
1878
1879 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1880 u64 mask = evsel->attr.sample_regs_intr;
1881
1882 sz = hweight_long(mask) * sizeof(u64);
1883 OVERFLOW_CHECK(array, sz, max_size);
1884 data->intr_regs.mask = mask;
1885 data->intr_regs.regs = (u64 *)array;
1886 array = (void *)array + sz;
1887 }
1888 }
1889
1890 return 0;
1891}
1892
1893size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1894 u64 read_format)
1895{
1896 size_t sz, result = sizeof(struct sample_event);
1897
1898 if (type & PERF_SAMPLE_IDENTIFIER)
1899 result += sizeof(u64);
1900
1901 if (type & PERF_SAMPLE_IP)
1902 result += sizeof(u64);
1903
1904 if (type & PERF_SAMPLE_TID)
1905 result += sizeof(u64);
1906
1907 if (type & PERF_SAMPLE_TIME)
1908 result += sizeof(u64);
1909
1910 if (type & PERF_SAMPLE_ADDR)
1911 result += sizeof(u64);
1912
1913 if (type & PERF_SAMPLE_ID)
1914 result += sizeof(u64);
1915
1916 if (type & PERF_SAMPLE_STREAM_ID)
1917 result += sizeof(u64);
1918
1919 if (type & PERF_SAMPLE_CPU)
1920 result += sizeof(u64);
1921
1922 if (type & PERF_SAMPLE_PERIOD)
1923 result += sizeof(u64);
1924
1925 if (type & PERF_SAMPLE_READ) {
1926 result += sizeof(u64);
1927 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1928 result += sizeof(u64);
1929 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1930 result += sizeof(u64);
1931 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1932 if (read_format & PERF_FORMAT_GROUP) {
1933 sz = sample->read.group.nr *
1934 sizeof(struct sample_read_value);
1935 result += sz;
1936 } else {
1937 result += sizeof(u64);
1938 }
1939 }
1940
1941 if (type & PERF_SAMPLE_CALLCHAIN) {
1942 sz = (sample->callchain->nr + 1) * sizeof(u64);
1943 result += sz;
1944 }
1945
1946 if (type & PERF_SAMPLE_RAW) {
1947 result += sizeof(u32);
1948 result += sample->raw_size;
1949 }
1950
1951 if (type & PERF_SAMPLE_BRANCH_STACK) {
1952 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1953 sz += sizeof(u64);
1954 result += sz;
1955 }
1956
1957 if (type & PERF_SAMPLE_REGS_USER) {
1958 if (sample->user_regs.abi) {
1959 result += sizeof(u64);
1960 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1961 result += sz;
1962 } else {
1963 result += sizeof(u64);
1964 }
1965 }
1966
1967 if (type & PERF_SAMPLE_STACK_USER) {
1968 sz = sample->user_stack.size;
1969 result += sizeof(u64);
1970 if (sz) {
1971 result += sz;
1972 result += sizeof(u64);
1973 }
1974 }
1975
1976 if (type & PERF_SAMPLE_WEIGHT)
1977 result += sizeof(u64);
1978
1979 if (type & PERF_SAMPLE_DATA_SRC)
1980 result += sizeof(u64);
1981
1982 if (type & PERF_SAMPLE_TRANSACTION)
1983 result += sizeof(u64);
1984
1985 if (type & PERF_SAMPLE_REGS_INTR) {
1986 if (sample->intr_regs.abi) {
1987 result += sizeof(u64);
1988 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1989 result += sz;
1990 } else {
1991 result += sizeof(u64);
1992 }
1993 }
1994
1995 return result;
1996}
1997
1998int perf_event__synthesize_sample(union perf_event *event, u64 type,
1999 u64 read_format,
2000 const struct perf_sample *sample,
2001 bool swapped)
2002{
2003 u64 *array;
2004 size_t sz;
2005 /*
2006 * used for cross-endian analysis. See git commit 65014ab3
2007 * for why this goofiness is needed.
2008 */
2009 union u64_swap u;
2010
2011 array = event->sample.array;
2012
2013 if (type & PERF_SAMPLE_IDENTIFIER) {
2014 *array = sample->id;
2015 array++;
2016 }
2017
2018 if (type & PERF_SAMPLE_IP) {
2019 *array = sample->ip;
2020 array++;
2021 }
2022
2023 if (type & PERF_SAMPLE_TID) {
2024 u.val32[0] = sample->pid;
2025 u.val32[1] = sample->tid;
2026 if (swapped) {
2027 /*
2028 * Inverse of what is done in perf_evsel__parse_sample
2029 */
2030 u.val32[0] = bswap_32(u.val32[0]);
2031 u.val32[1] = bswap_32(u.val32[1]);
2032 u.val64 = bswap_64(u.val64);
2033 }
2034
2035 *array = u.val64;
2036 array++;
2037 }
2038
2039 if (type & PERF_SAMPLE_TIME) {
2040 *array = sample->time;
2041 array++;
2042 }
2043
2044 if (type & PERF_SAMPLE_ADDR) {
2045 *array = sample->addr;
2046 array++;
2047 }
2048
2049 if (type & PERF_SAMPLE_ID) {
2050 *array = sample->id;
2051 array++;
2052 }
2053
2054 if (type & PERF_SAMPLE_STREAM_ID) {
2055 *array = sample->stream_id;
2056 array++;
2057 }
2058
2059 if (type & PERF_SAMPLE_CPU) {
2060 u.val32[0] = sample->cpu;
2061 if (swapped) {
2062 /*
2063 * Inverse of what is done in perf_evsel__parse_sample
2064 */
2065 u.val32[0] = bswap_32(u.val32[0]);
2066 u.val64 = bswap_64(u.val64);
2067 }
2068 *array = u.val64;
2069 array++;
2070 }
2071
2072 if (type & PERF_SAMPLE_PERIOD) {
2073 *array = sample->period;
2074 array++;
2075 }
2076
2077 if (type & PERF_SAMPLE_READ) {
2078 if (read_format & PERF_FORMAT_GROUP)
2079 *array = sample->read.group.nr;
2080 else
2081 *array = sample->read.one.value;
2082 array++;
2083
2084 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2085 *array = sample->read.time_enabled;
2086 array++;
2087 }
2088
2089 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2090 *array = sample->read.time_running;
2091 array++;
2092 }
2093
2094 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2095 if (read_format & PERF_FORMAT_GROUP) {
2096 sz = sample->read.group.nr *
2097 sizeof(struct sample_read_value);
2098 memcpy(array, sample->read.group.values, sz);
2099 array = (void *)array + sz;
2100 } else {
2101 *array = sample->read.one.id;
2102 array++;
2103 }
2104 }
2105
2106 if (type & PERF_SAMPLE_CALLCHAIN) {
2107 sz = (sample->callchain->nr + 1) * sizeof(u64);
2108 memcpy(array, sample->callchain, sz);
2109 array = (void *)array + sz;
2110 }
2111
2112 if (type & PERF_SAMPLE_RAW) {
2113 u.val32[0] = sample->raw_size;
2114 if (WARN_ONCE(swapped,
2115 "Endianness of raw data not corrected!\n")) {
2116 /*
2117 * Inverse of what is done in perf_evsel__parse_sample
2118 */
2119 u.val32[0] = bswap_32(u.val32[0]);
2120 u.val32[1] = bswap_32(u.val32[1]);
2121 u.val64 = bswap_64(u.val64);
2122 }
2123 *array = u.val64;
2124 array = (void *)array + sizeof(u32);
2125
2126 memcpy(array, sample->raw_data, sample->raw_size);
2127 array = (void *)array + sample->raw_size;
2128 }
2129
2130 if (type & PERF_SAMPLE_BRANCH_STACK) {
2131 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2132 sz += sizeof(u64);
2133 memcpy(array, sample->branch_stack, sz);
2134 array = (void *)array + sz;
2135 }
2136
2137 if (type & PERF_SAMPLE_REGS_USER) {
2138 if (sample->user_regs.abi) {
2139 *array++ = sample->user_regs.abi;
2140 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2141 memcpy(array, sample->user_regs.regs, sz);
2142 array = (void *)array + sz;
2143 } else {
2144 *array++ = 0;
2145 }
2146 }
2147
2148 if (type & PERF_SAMPLE_STACK_USER) {
2149 sz = sample->user_stack.size;
2150 *array++ = sz;
2151 if (sz) {
2152 memcpy(array, sample->user_stack.data, sz);
2153 array = (void *)array + sz;
2154 *array++ = sz;
2155 }
2156 }
2157
2158 if (type & PERF_SAMPLE_WEIGHT) {
2159 *array = sample->weight;
2160 array++;
2161 }
2162
2163 if (type & PERF_SAMPLE_DATA_SRC) {
2164 *array = sample->data_src;
2165 array++;
2166 }
2167
2168 if (type & PERF_SAMPLE_TRANSACTION) {
2169 *array = sample->transaction;
2170 array++;
2171 }
2172
2173 if (type & PERF_SAMPLE_REGS_INTR) {
2174 if (sample->intr_regs.abi) {
2175 *array++ = sample->intr_regs.abi;
2176 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2177 memcpy(array, sample->intr_regs.regs, sz);
2178 array = (void *)array + sz;
2179 } else {
2180 *array++ = 0;
2181 }
2182 }
2183
2184 return 0;
2185}
2186
2187struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2188{
2189 return pevent_find_field(evsel->tp_format, name);
2190}
2191
2192void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2193 const char *name)
2194{
2195 struct format_field *field = perf_evsel__field(evsel, name);
2196 int offset;
2197
2198 if (!field)
2199 return NULL;
2200
2201 offset = field->offset;
2202
2203 if (field->flags & FIELD_IS_DYNAMIC) {
2204 offset = *(int *)(sample->raw_data + field->offset);
2205 offset &= 0xffff;
2206 }
2207
2208 return sample->raw_data + offset;
2209}
2210
2211u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2212 const char *name)
2213{
2214 struct format_field *field = perf_evsel__field(evsel, name);
2215 void *ptr;
2216 u64 value;
2217
2218 if (!field)
2219 return 0;
2220
2221 ptr = sample->raw_data + field->offset;
2222
2223 switch (field->size) {
2224 case 1:
2225 return *(u8 *)ptr;
2226 case 2:
2227 value = *(u16 *)ptr;
2228 break;
2229 case 4:
2230 value = *(u32 *)ptr;
2231 break;
2232 case 8:
2233 memcpy(&value, ptr, sizeof(u64));
2234 break;
2235 default:
2236 return 0;
2237 }
2238
2239 if (!evsel->needs_swap)
2240 return value;
2241
2242 switch (field->size) {
2243 case 2:
2244 return bswap_16(value);
2245 case 4:
2246 return bswap_32(value);
2247 case 8:
2248 return bswap_64(value);
2249 default:
2250 return 0;
2251 }
2252
2253 return 0;
2254}
2255
2256static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2257{
2258 va_list args;
2259 int ret = 0;
2260
2261 if (!*first) {
2262 ret += fprintf(fp, ",");
2263 } else {
2264 ret += fprintf(fp, ":");
2265 *first = false;
2266 }
2267
2268 va_start(args, fmt);
2269 ret += vfprintf(fp, fmt, args);
2270 va_end(args);
2271 return ret;
2272}
2273
2274static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2275{
2276 return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2277}
2278
2279int perf_evsel__fprintf(struct perf_evsel *evsel,
2280 struct perf_attr_details *details, FILE *fp)
2281{
2282 bool first = true;
2283 int printed = 0;
2284
2285 if (details->event_group) {
2286 struct perf_evsel *pos;
2287
2288 if (!perf_evsel__is_group_leader(evsel))
2289 return 0;
2290
2291 if (evsel->nr_members > 1)
2292 printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2293
2294 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2295 for_each_group_member(pos, evsel)
2296 printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2297
2298 if (evsel->nr_members > 1)
2299 printed += fprintf(fp, "}");
2300 goto out;
2301 }
2302
2303 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2304
2305 if (details->verbose) {
2306 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2307 __print_attr__fprintf, &first);
2308 } else if (details->freq) {
2309 const char *term = "sample_freq";
2310
2311 if (!evsel->attr.freq)
2312 term = "sample_period";
2313
2314 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2315 term, (u64)evsel->attr.sample_freq);
2316 }
2317
2318 if (details->trace_fields) {
2319 struct format_field *field;
2320
2321 if (evsel->attr.type != PERF_TYPE_TRACEPOINT) {
2322 printed += comma_fprintf(fp, &first, " (not a tracepoint)");
2323 goto out;
2324 }
2325
2326 field = evsel->tp_format->format.fields;
2327 if (field == NULL) {
2328 printed += comma_fprintf(fp, &first, " (no trace field)");
2329 goto out;
2330 }
2331
2332 printed += comma_fprintf(fp, &first, " trace_fields: %s", field->name);
2333
2334 field = field->next;
2335 while (field) {
2336 printed += comma_fprintf(fp, &first, "%s", field->name);
2337 field = field->next;
2338 }
2339 }
2340out:
2341 fputc('\n', fp);
2342 return ++printed;
2343}
2344
2345bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2346 char *msg, size_t msgsize)
2347{
2348 int paranoid;
2349
2350 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2351 evsel->attr.type == PERF_TYPE_HARDWARE &&
2352 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2353 /*
2354 * If it's cycles then fall back to hrtimer based
2355 * cpu-clock-tick sw counter, which is always available even if
2356 * no PMU support.
2357 *
2358 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2359 * b0a873e).
2360 */
2361 scnprintf(msg, msgsize, "%s",
2362"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2363
2364 evsel->attr.type = PERF_TYPE_SOFTWARE;
2365 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2366
2367 zfree(&evsel->name);
2368 return true;
2369 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2370 (paranoid = perf_event_paranoid()) > 1) {
2371 const char *name = perf_evsel__name(evsel);
2372 char *new_name;
2373
2374 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2375 return false;
2376
2377 if (evsel->name)
2378 free(evsel->name);
2379 evsel->name = new_name;
2380 scnprintf(msg, msgsize,
2381"kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2382 evsel->attr.exclude_kernel = 1;
2383
2384 return true;
2385 }
2386
2387 return false;
2388}
2389
2390int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2391 int err, char *msg, size_t size)
2392{
2393 char sbuf[STRERR_BUFSIZE];
2394
2395 switch (err) {
2396 case EPERM:
2397 case EACCES:
2398 return scnprintf(msg, size,
2399 "You may not have permission to collect %sstats.\n\n"
2400 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2401 "which controls use of the performance events system by\n"
2402 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2403 "The current value is %d:\n\n"
2404 " -1: Allow use of (almost) all events by all users\n"
2405 ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2406 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2407 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2408 target->system_wide ? "system-wide " : "",
2409 perf_event_paranoid());
2410 case ENOENT:
2411 return scnprintf(msg, size, "The %s event is not supported.",
2412 perf_evsel__name(evsel));
2413 case EMFILE:
2414 return scnprintf(msg, size, "%s",
2415 "Too many events are opened.\n"
2416 "Probably the maximum number of open file descriptors has been reached.\n"
2417 "Hint: Try again after reducing the number of events.\n"
2418 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2419 case ENODEV:
2420 if (target->cpu_list)
2421 return scnprintf(msg, size, "%s",
2422 "No such device - did you specify an out-of-range profile CPU?\n");
2423 break;
2424 case EOPNOTSUPP:
2425 if (evsel->attr.precise_ip)
2426 return scnprintf(msg, size, "%s",
2427 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2428#if defined(__i386__) || defined(__x86_64__)
2429 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2430 return scnprintf(msg, size, "%s",
2431 "No hardware sampling interrupt available.\n"
2432 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2433#endif
2434 break;
2435 case EBUSY:
2436 if (find_process("oprofiled"))
2437 return scnprintf(msg, size,
2438 "The PMU counters are busy/taken by another profiler.\n"
2439 "We found oprofile daemon running, please stop it and try again.");
2440 break;
2441 case EINVAL:
2442 if (perf_missing_features.clockid)
2443 return scnprintf(msg, size, "clockid feature not supported.");
2444 if (perf_missing_features.clockid_wrong)
2445 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2446 break;
2447 default:
2448 break;
2449 }
2450
2451 return scnprintf(msg, size,
2452 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2453 "/bin/dmesg may provide additional information.\n"
2454 "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2455 err, strerror_r(err, sbuf, sizeof(sbuf)),
2456 perf_evsel__name(evsel));
2457}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4 *
5 * Parts came from builtin-{top,stat,record}.c, see those files for further
6 * copyright notes.
7 */
8
9#include <byteswap.h>
10#include <errno.h>
11#include <inttypes.h>
12#include <linux/bitops.h>
13#include <api/fs/fs.h>
14#include <api/fs/tracing_path.h>
15#include <traceevent/event-parse.h>
16#include <linux/hw_breakpoint.h>
17#include <linux/perf_event.h>
18#include <linux/compiler.h>
19#include <linux/err.h>
20#include <linux/zalloc.h>
21#include <sys/ioctl.h>
22#include <sys/resource.h>
23#include <sys/types.h>
24#include <dirent.h>
25#include <stdlib.h>
26#include <perf/evsel.h>
27#include "asm/bug.h"
28#include "callchain.h"
29#include "cgroup.h"
30#include "counts.h"
31#include "event.h"
32#include "evsel.h"
33#include "util/env.h"
34#include "util/evsel_config.h"
35#include "util/evsel_fprintf.h"
36#include "evlist.h"
37#include <perf/cpumap.h>
38#include "thread_map.h"
39#include "target.h"
40#include "perf_regs.h"
41#include "record.h"
42#include "debug.h"
43#include "trace-event.h"
44#include "stat.h"
45#include "string2.h"
46#include "memswap.h"
47#include "util.h"
48#include "../perf-sys.h"
49#include "util/parse-branch-options.h"
50#include <internal/xyarray.h>
51#include <internal/lib.h>
52
53#include <linux/ctype.h>
54
55struct perf_missing_features perf_missing_features;
56
57static clockid_t clockid;
58
59static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
60{
61 return 0;
62}
63
64void __weak test_attr__ready(void) { }
65
66static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
67{
68}
69
70static struct {
71 size_t size;
72 int (*init)(struct evsel *evsel);
73 void (*fini)(struct evsel *evsel);
74} perf_evsel__object = {
75 .size = sizeof(struct evsel),
76 .init = evsel__no_extra_init,
77 .fini = evsel__no_extra_fini,
78};
79
80int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
81 void (*fini)(struct evsel *evsel))
82{
83
84 if (object_size == 0)
85 goto set_methods;
86
87 if (perf_evsel__object.size > object_size)
88 return -EINVAL;
89
90 perf_evsel__object.size = object_size;
91
92set_methods:
93 if (init != NULL)
94 perf_evsel__object.init = init;
95
96 if (fini != NULL)
97 perf_evsel__object.fini = fini;
98
99 return 0;
100}
101
102#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
103
104int __evsel__sample_size(u64 sample_type)
105{
106 u64 mask = sample_type & PERF_SAMPLE_MASK;
107 int size = 0;
108 int i;
109
110 for (i = 0; i < 64; i++) {
111 if (mask & (1ULL << i))
112 size++;
113 }
114
115 size *= sizeof(u64);
116
117 return size;
118}
119
120/**
121 * __perf_evsel__calc_id_pos - calculate id_pos.
122 * @sample_type: sample type
123 *
124 * This function returns the position of the event id (PERF_SAMPLE_ID or
125 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
126 * perf_record_sample.
127 */
128static int __perf_evsel__calc_id_pos(u64 sample_type)
129{
130 int idx = 0;
131
132 if (sample_type & PERF_SAMPLE_IDENTIFIER)
133 return 0;
134
135 if (!(sample_type & PERF_SAMPLE_ID))
136 return -1;
137
138 if (sample_type & PERF_SAMPLE_IP)
139 idx += 1;
140
141 if (sample_type & PERF_SAMPLE_TID)
142 idx += 1;
143
144 if (sample_type & PERF_SAMPLE_TIME)
145 idx += 1;
146
147 if (sample_type & PERF_SAMPLE_ADDR)
148 idx += 1;
149
150 return idx;
151}
152
153/**
154 * __perf_evsel__calc_is_pos - calculate is_pos.
155 * @sample_type: sample type
156 *
157 * This function returns the position (counting backwards) of the event id
158 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
159 * sample_id_all is used there is an id sample appended to non-sample events.
160 */
161static int __perf_evsel__calc_is_pos(u64 sample_type)
162{
163 int idx = 1;
164
165 if (sample_type & PERF_SAMPLE_IDENTIFIER)
166 return 1;
167
168 if (!(sample_type & PERF_SAMPLE_ID))
169 return -1;
170
171 if (sample_type & PERF_SAMPLE_CPU)
172 idx += 1;
173
174 if (sample_type & PERF_SAMPLE_STREAM_ID)
175 idx += 1;
176
177 return idx;
178}
179
180void evsel__calc_id_pos(struct evsel *evsel)
181{
182 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
183 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
184}
185
186void __evsel__set_sample_bit(struct evsel *evsel,
187 enum perf_event_sample_format bit)
188{
189 if (!(evsel->core.attr.sample_type & bit)) {
190 evsel->core.attr.sample_type |= bit;
191 evsel->sample_size += sizeof(u64);
192 evsel__calc_id_pos(evsel);
193 }
194}
195
196void __evsel__reset_sample_bit(struct evsel *evsel,
197 enum perf_event_sample_format bit)
198{
199 if (evsel->core.attr.sample_type & bit) {
200 evsel->core.attr.sample_type &= ~bit;
201 evsel->sample_size -= sizeof(u64);
202 evsel__calc_id_pos(evsel);
203 }
204}
205
206void evsel__set_sample_id(struct evsel *evsel,
207 bool can_sample_identifier)
208{
209 if (can_sample_identifier) {
210 evsel__reset_sample_bit(evsel, ID);
211 evsel__set_sample_bit(evsel, IDENTIFIER);
212 } else {
213 evsel__set_sample_bit(evsel, ID);
214 }
215 evsel->core.attr.read_format |= PERF_FORMAT_ID;
216}
217
218/**
219 * evsel__is_function_event - Return whether given evsel is a function
220 * trace event
221 *
222 * @evsel - evsel selector to be tested
223 *
224 * Return %true if event is function trace event
225 */
226bool evsel__is_function_event(struct evsel *evsel)
227{
228#define FUNCTION_EVENT "ftrace:function"
229
230 return evsel->name &&
231 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
232
233#undef FUNCTION_EVENT
234}
235
236void evsel__init(struct evsel *evsel,
237 struct perf_event_attr *attr, int idx)
238{
239 perf_evsel__init(&evsel->core, attr);
240 evsel->idx = idx;
241 evsel->tracking = !idx;
242 evsel->leader = evsel;
243 evsel->unit = "";
244 evsel->scale = 1.0;
245 evsel->max_events = ULONG_MAX;
246 evsel->evlist = NULL;
247 evsel->bpf_obj = NULL;
248 evsel->bpf_fd = -1;
249 INIT_LIST_HEAD(&evsel->config_terms);
250 perf_evsel__object.init(evsel);
251 evsel->sample_size = __evsel__sample_size(attr->sample_type);
252 evsel__calc_id_pos(evsel);
253 evsel->cmdline_group_boundary = false;
254 evsel->metric_expr = NULL;
255 evsel->metric_name = NULL;
256 evsel->metric_events = NULL;
257 evsel->per_pkg_mask = NULL;
258 evsel->collect_stat = false;
259 evsel->pmu_name = NULL;
260}
261
262struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
263{
264 struct evsel *evsel = zalloc(perf_evsel__object.size);
265
266 if (!evsel)
267 return NULL;
268 evsel__init(evsel, attr, idx);
269
270 if (evsel__is_bpf_output(evsel)) {
271 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
272 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
273 evsel->core.attr.sample_period = 1;
274 }
275
276 if (evsel__is_clock(evsel)) {
277 /*
278 * The evsel->unit points to static alias->unit
279 * so it's ok to use static string in here.
280 */
281 static const char *unit = "msec";
282
283 evsel->unit = unit;
284 evsel->scale = 1e-6;
285 }
286
287 return evsel;
288}
289
290static bool perf_event_can_profile_kernel(void)
291{
292 return perf_event_paranoid_check(1);
293}
294
295struct evsel *evsel__new_cycles(bool precise)
296{
297 struct perf_event_attr attr = {
298 .type = PERF_TYPE_HARDWARE,
299 .config = PERF_COUNT_HW_CPU_CYCLES,
300 .exclude_kernel = !perf_event_can_profile_kernel(),
301 };
302 struct evsel *evsel;
303
304 event_attr_init(&attr);
305
306 if (!precise)
307 goto new_event;
308
309 /*
310 * Now let the usual logic to set up the perf_event_attr defaults
311 * to kick in when we return and before perf_evsel__open() is called.
312 */
313new_event:
314 evsel = evsel__new(&attr);
315 if (evsel == NULL)
316 goto out;
317
318 evsel->precise_max = true;
319
320 /* use asprintf() because free(evsel) assumes name is allocated */
321 if (asprintf(&evsel->name, "cycles%s%s%.*s",
322 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
323 attr.exclude_kernel ? "u" : "",
324 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
325 goto error_free;
326out:
327 return evsel;
328error_free:
329 evsel__delete(evsel);
330 evsel = NULL;
331 goto out;
332}
333
334/*
335 * Returns pointer with encoded error via <linux/err.h> interface.
336 */
337struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
338{
339 struct evsel *evsel = zalloc(perf_evsel__object.size);
340 int err = -ENOMEM;
341
342 if (evsel == NULL) {
343 goto out_err;
344 } else {
345 struct perf_event_attr attr = {
346 .type = PERF_TYPE_TRACEPOINT,
347 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
348 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
349 };
350
351 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
352 goto out_free;
353
354 evsel->tp_format = trace_event__tp_format(sys, name);
355 if (IS_ERR(evsel->tp_format)) {
356 err = PTR_ERR(evsel->tp_format);
357 goto out_free;
358 }
359
360 event_attr_init(&attr);
361 attr.config = evsel->tp_format->id;
362 attr.sample_period = 1;
363 evsel__init(evsel, &attr, idx);
364 }
365
366 return evsel;
367
368out_free:
369 zfree(&evsel->name);
370 free(evsel);
371out_err:
372 return ERR_PTR(err);
373}
374
375const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
376 "cycles",
377 "instructions",
378 "cache-references",
379 "cache-misses",
380 "branches",
381 "branch-misses",
382 "bus-cycles",
383 "stalled-cycles-frontend",
384 "stalled-cycles-backend",
385 "ref-cycles",
386};
387
388static const char *__evsel__hw_name(u64 config)
389{
390 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
391 return evsel__hw_names[config];
392
393 return "unknown-hardware";
394}
395
396static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
397{
398 int colon = 0, r = 0;
399 struct perf_event_attr *attr = &evsel->core.attr;
400 bool exclude_guest_default = false;
401
402#define MOD_PRINT(context, mod) do { \
403 if (!attr->exclude_##context) { \
404 if (!colon) colon = ++r; \
405 r += scnprintf(bf + r, size - r, "%c", mod); \
406 } } while(0)
407
408 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
409 MOD_PRINT(kernel, 'k');
410 MOD_PRINT(user, 'u');
411 MOD_PRINT(hv, 'h');
412 exclude_guest_default = true;
413 }
414
415 if (attr->precise_ip) {
416 if (!colon)
417 colon = ++r;
418 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
419 exclude_guest_default = true;
420 }
421
422 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
423 MOD_PRINT(host, 'H');
424 MOD_PRINT(guest, 'G');
425 }
426#undef MOD_PRINT
427 if (colon)
428 bf[colon - 1] = ':';
429 return r;
430}
431
432static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
433{
434 int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
435 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
436}
437
438const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
439 "cpu-clock",
440 "task-clock",
441 "page-faults",
442 "context-switches",
443 "cpu-migrations",
444 "minor-faults",
445 "major-faults",
446 "alignment-faults",
447 "emulation-faults",
448 "dummy",
449};
450
451static const char *__evsel__sw_name(u64 config)
452{
453 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
454 return evsel__sw_names[config];
455 return "unknown-software";
456}
457
458static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
459{
460 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
461 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
462}
463
464static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
465{
466 int r;
467
468 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
469
470 if (type & HW_BREAKPOINT_R)
471 r += scnprintf(bf + r, size - r, "r");
472
473 if (type & HW_BREAKPOINT_W)
474 r += scnprintf(bf + r, size - r, "w");
475
476 if (type & HW_BREAKPOINT_X)
477 r += scnprintf(bf + r, size - r, "x");
478
479 return r;
480}
481
482static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
483{
484 struct perf_event_attr *attr = &evsel->core.attr;
485 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
486 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
487}
488
489const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
490 { "L1-dcache", "l1-d", "l1d", "L1-data", },
491 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
492 { "LLC", "L2", },
493 { "dTLB", "d-tlb", "Data-TLB", },
494 { "iTLB", "i-tlb", "Instruction-TLB", },
495 { "branch", "branches", "bpu", "btb", "bpc", },
496 { "node", },
497};
498
499const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
500 { "load", "loads", "read", },
501 { "store", "stores", "write", },
502 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
503};
504
505const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
506 { "refs", "Reference", "ops", "access", },
507 { "misses", "miss", },
508};
509
510#define C(x) PERF_COUNT_HW_CACHE_##x
511#define CACHE_READ (1 << C(OP_READ))
512#define CACHE_WRITE (1 << C(OP_WRITE))
513#define CACHE_PREFETCH (1 << C(OP_PREFETCH))
514#define COP(x) (1 << x)
515
516/*
517 * cache operartion stat
518 * L1I : Read and prefetch only
519 * ITLB and BPU : Read-only
520 */
521static unsigned long evsel__hw_cache_stat[C(MAX)] = {
522 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
523 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
524 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
525 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
526 [C(ITLB)] = (CACHE_READ),
527 [C(BPU)] = (CACHE_READ),
528 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
529};
530
531bool evsel__is_cache_op_valid(u8 type, u8 op)
532{
533 if (evsel__hw_cache_stat[type] & COP(op))
534 return true; /* valid */
535 else
536 return false; /* invalid */
537}
538
539int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
540{
541 if (result) {
542 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
543 evsel__hw_cache_op[op][0],
544 evsel__hw_cache_result[result][0]);
545 }
546
547 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
548 evsel__hw_cache_op[op][1]);
549}
550
551static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
552{
553 u8 op, result, type = (config >> 0) & 0xff;
554 const char *err = "unknown-ext-hardware-cache-type";
555
556 if (type >= PERF_COUNT_HW_CACHE_MAX)
557 goto out_err;
558
559 op = (config >> 8) & 0xff;
560 err = "unknown-ext-hardware-cache-op";
561 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
562 goto out_err;
563
564 result = (config >> 16) & 0xff;
565 err = "unknown-ext-hardware-cache-result";
566 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
567 goto out_err;
568
569 err = "invalid-cache";
570 if (!evsel__is_cache_op_valid(type, op))
571 goto out_err;
572
573 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
574out_err:
575 return scnprintf(bf, size, "%s", err);
576}
577
578static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
579{
580 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
581 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
582}
583
584static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
585{
586 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
587 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
588}
589
590static int evsel__tool_name(char *bf, size_t size)
591{
592 int ret = scnprintf(bf, size, "duration_time");
593 return ret;
594}
595
596const char *evsel__name(struct evsel *evsel)
597{
598 char bf[128];
599
600 if (!evsel)
601 goto out_unknown;
602
603 if (evsel->name)
604 return evsel->name;
605
606 switch (evsel->core.attr.type) {
607 case PERF_TYPE_RAW:
608 evsel__raw_name(evsel, bf, sizeof(bf));
609 break;
610
611 case PERF_TYPE_HARDWARE:
612 evsel__hw_name(evsel, bf, sizeof(bf));
613 break;
614
615 case PERF_TYPE_HW_CACHE:
616 evsel__hw_cache_name(evsel, bf, sizeof(bf));
617 break;
618
619 case PERF_TYPE_SOFTWARE:
620 if (evsel->tool_event)
621 evsel__tool_name(bf, sizeof(bf));
622 else
623 evsel__sw_name(evsel, bf, sizeof(bf));
624 break;
625
626 case PERF_TYPE_TRACEPOINT:
627 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
628 break;
629
630 case PERF_TYPE_BREAKPOINT:
631 evsel__bp_name(evsel, bf, sizeof(bf));
632 break;
633
634 default:
635 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
636 evsel->core.attr.type);
637 break;
638 }
639
640 evsel->name = strdup(bf);
641
642 if (evsel->name)
643 return evsel->name;
644out_unknown:
645 return "unknown";
646}
647
648const char *evsel__group_name(struct evsel *evsel)
649{
650 return evsel->group_name ?: "anon group";
651}
652
653/*
654 * Returns the group details for the specified leader,
655 * with following rules.
656 *
657 * For record -e '{cycles,instructions}'
658 * 'anon group { cycles:u, instructions:u }'
659 *
660 * For record -e 'cycles,instructions' and report --group
661 * 'cycles:u, instructions:u'
662 */
663int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
664{
665 int ret = 0;
666 struct evsel *pos;
667 const char *group_name = evsel__group_name(evsel);
668
669 if (!evsel->forced_leader)
670 ret = scnprintf(buf, size, "%s { ", group_name);
671
672 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
673
674 for_each_group_member(pos, evsel)
675 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
676
677 if (!evsel->forced_leader)
678 ret += scnprintf(buf + ret, size - ret, " }");
679
680 return ret;
681}
682
683static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
684 struct callchain_param *param)
685{
686 bool function = evsel__is_function_event(evsel);
687 struct perf_event_attr *attr = &evsel->core.attr;
688
689 evsel__set_sample_bit(evsel, CALLCHAIN);
690
691 attr->sample_max_stack = param->max_stack;
692
693 if (opts->kernel_callchains)
694 attr->exclude_callchain_user = 1;
695 if (opts->user_callchains)
696 attr->exclude_callchain_kernel = 1;
697 if (param->record_mode == CALLCHAIN_LBR) {
698 if (!opts->branch_stack) {
699 if (attr->exclude_user) {
700 pr_warning("LBR callstack option is only available "
701 "to get user callchain information. "
702 "Falling back to framepointers.\n");
703 } else {
704 evsel__set_sample_bit(evsel, BRANCH_STACK);
705 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
706 PERF_SAMPLE_BRANCH_CALL_STACK |
707 PERF_SAMPLE_BRANCH_NO_CYCLES |
708 PERF_SAMPLE_BRANCH_NO_FLAGS |
709 PERF_SAMPLE_BRANCH_HW_INDEX;
710 }
711 } else
712 pr_warning("Cannot use LBR callstack with branch stack. "
713 "Falling back to framepointers.\n");
714 }
715
716 if (param->record_mode == CALLCHAIN_DWARF) {
717 if (!function) {
718 evsel__set_sample_bit(evsel, REGS_USER);
719 evsel__set_sample_bit(evsel, STACK_USER);
720 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
721 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
722 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
723 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
724 "so the minimal registers set (IP, SP) is explicitly forced.\n");
725 } else {
726 attr->sample_regs_user |= PERF_REGS_MASK;
727 }
728 attr->sample_stack_user = param->dump_size;
729 attr->exclude_callchain_user = 1;
730 } else {
731 pr_info("Cannot use DWARF unwind for function trace event,"
732 " falling back to framepointers.\n");
733 }
734 }
735
736 if (function) {
737 pr_info("Disabling user space callchains for function trace event.\n");
738 attr->exclude_callchain_user = 1;
739 }
740}
741
742void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
743 struct callchain_param *param)
744{
745 if (param->enabled)
746 return __evsel__config_callchain(evsel, opts, param);
747}
748
749static void
750perf_evsel__reset_callgraph(struct evsel *evsel,
751 struct callchain_param *param)
752{
753 struct perf_event_attr *attr = &evsel->core.attr;
754
755 evsel__reset_sample_bit(evsel, CALLCHAIN);
756 if (param->record_mode == CALLCHAIN_LBR) {
757 evsel__reset_sample_bit(evsel, BRANCH_STACK);
758 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
759 PERF_SAMPLE_BRANCH_CALL_STACK |
760 PERF_SAMPLE_BRANCH_HW_INDEX);
761 }
762 if (param->record_mode == CALLCHAIN_DWARF) {
763 evsel__reset_sample_bit(evsel, REGS_USER);
764 evsel__reset_sample_bit(evsel, STACK_USER);
765 }
766}
767
768static void evsel__apply_config_terms(struct evsel *evsel,
769 struct record_opts *opts, bool track)
770{
771 struct evsel_config_term *term;
772 struct list_head *config_terms = &evsel->config_terms;
773 struct perf_event_attr *attr = &evsel->core.attr;
774 /* callgraph default */
775 struct callchain_param param = {
776 .record_mode = callchain_param.record_mode,
777 };
778 u32 dump_size = 0;
779 int max_stack = 0;
780 const char *callgraph_buf = NULL;
781
782 list_for_each_entry(term, config_terms, list) {
783 switch (term->type) {
784 case EVSEL__CONFIG_TERM_PERIOD:
785 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
786 attr->sample_period = term->val.period;
787 attr->freq = 0;
788 evsel__reset_sample_bit(evsel, PERIOD);
789 }
790 break;
791 case EVSEL__CONFIG_TERM_FREQ:
792 if (!(term->weak && opts->user_freq != UINT_MAX)) {
793 attr->sample_freq = term->val.freq;
794 attr->freq = 1;
795 evsel__set_sample_bit(evsel, PERIOD);
796 }
797 break;
798 case EVSEL__CONFIG_TERM_TIME:
799 if (term->val.time)
800 evsel__set_sample_bit(evsel, TIME);
801 else
802 evsel__reset_sample_bit(evsel, TIME);
803 break;
804 case EVSEL__CONFIG_TERM_CALLGRAPH:
805 callgraph_buf = term->val.str;
806 break;
807 case EVSEL__CONFIG_TERM_BRANCH:
808 if (term->val.str && strcmp(term->val.str, "no")) {
809 evsel__set_sample_bit(evsel, BRANCH_STACK);
810 parse_branch_str(term->val.str,
811 &attr->branch_sample_type);
812 } else
813 evsel__reset_sample_bit(evsel, BRANCH_STACK);
814 break;
815 case EVSEL__CONFIG_TERM_STACK_USER:
816 dump_size = term->val.stack_user;
817 break;
818 case EVSEL__CONFIG_TERM_MAX_STACK:
819 max_stack = term->val.max_stack;
820 break;
821 case EVSEL__CONFIG_TERM_MAX_EVENTS:
822 evsel->max_events = term->val.max_events;
823 break;
824 case EVSEL__CONFIG_TERM_INHERIT:
825 /*
826 * attr->inherit should has already been set by
827 * evsel__config. If user explicitly set
828 * inherit using config terms, override global
829 * opt->no_inherit setting.
830 */
831 attr->inherit = term->val.inherit ? 1 : 0;
832 break;
833 case EVSEL__CONFIG_TERM_OVERWRITE:
834 attr->write_backward = term->val.overwrite ? 1 : 0;
835 break;
836 case EVSEL__CONFIG_TERM_DRV_CFG:
837 break;
838 case EVSEL__CONFIG_TERM_PERCORE:
839 break;
840 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
841 attr->aux_output = term->val.aux_output ? 1 : 0;
842 break;
843 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
844 /* Already applied by auxtrace */
845 break;
846 case EVSEL__CONFIG_TERM_CFG_CHG:
847 break;
848 default:
849 break;
850 }
851 }
852
853 /* User explicitly set per-event callgraph, clear the old setting and reset. */
854 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
855 bool sample_address = false;
856
857 if (max_stack) {
858 param.max_stack = max_stack;
859 if (callgraph_buf == NULL)
860 callgraph_buf = "fp";
861 }
862
863 /* parse callgraph parameters */
864 if (callgraph_buf != NULL) {
865 if (!strcmp(callgraph_buf, "no")) {
866 param.enabled = false;
867 param.record_mode = CALLCHAIN_NONE;
868 } else {
869 param.enabled = true;
870 if (parse_callchain_record(callgraph_buf, ¶m)) {
871 pr_err("per-event callgraph setting for %s failed. "
872 "Apply callgraph global setting for it\n",
873 evsel->name);
874 return;
875 }
876 if (param.record_mode == CALLCHAIN_DWARF)
877 sample_address = true;
878 }
879 }
880 if (dump_size > 0) {
881 dump_size = round_up(dump_size, sizeof(u64));
882 param.dump_size = dump_size;
883 }
884
885 /* If global callgraph set, clear it */
886 if (callchain_param.enabled)
887 perf_evsel__reset_callgraph(evsel, &callchain_param);
888
889 /* set perf-event callgraph */
890 if (param.enabled) {
891 if (sample_address) {
892 evsel__set_sample_bit(evsel, ADDR);
893 evsel__set_sample_bit(evsel, DATA_SRC);
894 evsel->core.attr.mmap_data = track;
895 }
896 evsel__config_callchain(evsel, opts, ¶m);
897 }
898 }
899}
900
901struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
902{
903 struct evsel_config_term *term, *found_term = NULL;
904
905 list_for_each_entry(term, &evsel->config_terms, list) {
906 if (term->type == type)
907 found_term = term;
908 }
909
910 return found_term;
911}
912
913/*
914 * The enable_on_exec/disabled value strategy:
915 *
916 * 1) For any type of traced program:
917 * - all independent events and group leaders are disabled
918 * - all group members are enabled
919 *
920 * Group members are ruled by group leaders. They need to
921 * be enabled, because the group scheduling relies on that.
922 *
923 * 2) For traced programs executed by perf:
924 * - all independent events and group leaders have
925 * enable_on_exec set
926 * - we don't specifically enable or disable any event during
927 * the record command
928 *
929 * Independent events and group leaders are initially disabled
930 * and get enabled by exec. Group members are ruled by group
931 * leaders as stated in 1).
932 *
933 * 3) For traced programs attached by perf (pid/tid):
934 * - we specifically enable or disable all events during
935 * the record command
936 *
937 * When attaching events to already running traced we
938 * enable/disable events specifically, as there's no
939 * initial traced exec call.
940 */
941void evsel__config(struct evsel *evsel, struct record_opts *opts,
942 struct callchain_param *callchain)
943{
944 struct evsel *leader = evsel->leader;
945 struct perf_event_attr *attr = &evsel->core.attr;
946 int track = evsel->tracking;
947 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
948
949 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
950 attr->inherit = !opts->no_inherit;
951 attr->write_backward = opts->overwrite ? 1 : 0;
952
953 evsel__set_sample_bit(evsel, IP);
954 evsel__set_sample_bit(evsel, TID);
955
956 if (evsel->sample_read) {
957 evsel__set_sample_bit(evsel, READ);
958
959 /*
960 * We need ID even in case of single event, because
961 * PERF_SAMPLE_READ process ID specific data.
962 */
963 evsel__set_sample_id(evsel, false);
964
965 /*
966 * Apply group format only if we belong to group
967 * with more than one members.
968 */
969 if (leader->core.nr_members > 1) {
970 attr->read_format |= PERF_FORMAT_GROUP;
971 attr->inherit = 0;
972 }
973 }
974
975 /*
976 * We default some events to have a default interval. But keep
977 * it a weak assumption overridable by the user.
978 */
979 if (!attr->sample_period) {
980 if (opts->freq) {
981 attr->freq = 1;
982 attr->sample_freq = opts->freq;
983 } else {
984 attr->sample_period = opts->default_interval;
985 }
986 }
987 /*
988 * If attr->freq was set (here or earlier), ask for period
989 * to be sampled.
990 */
991 if (attr->freq)
992 evsel__set_sample_bit(evsel, PERIOD);
993
994 if (opts->no_samples)
995 attr->sample_freq = 0;
996
997 if (opts->inherit_stat) {
998 evsel->core.attr.read_format |=
999 PERF_FORMAT_TOTAL_TIME_ENABLED |
1000 PERF_FORMAT_TOTAL_TIME_RUNNING |
1001 PERF_FORMAT_ID;
1002 attr->inherit_stat = 1;
1003 }
1004
1005 if (opts->sample_address) {
1006 evsel__set_sample_bit(evsel, ADDR);
1007 attr->mmap_data = track;
1008 }
1009
1010 /*
1011 * We don't allow user space callchains for function trace
1012 * event, due to issues with page faults while tracing page
1013 * fault handler and its overall trickiness nature.
1014 */
1015 if (evsel__is_function_event(evsel))
1016 evsel->core.attr.exclude_callchain_user = 1;
1017
1018 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1019 evsel__config_callchain(evsel, opts, callchain);
1020
1021 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1022 !evsel__is_dummy_event(evsel)) {
1023 attr->sample_regs_intr = opts->sample_intr_regs;
1024 evsel__set_sample_bit(evsel, REGS_INTR);
1025 }
1026
1027 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1028 !evsel__is_dummy_event(evsel)) {
1029 attr->sample_regs_user |= opts->sample_user_regs;
1030 evsel__set_sample_bit(evsel, REGS_USER);
1031 }
1032
1033 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1034 evsel__set_sample_bit(evsel, CPU);
1035
1036 /*
1037 * When the user explicitly disabled time don't force it here.
1038 */
1039 if (opts->sample_time &&
1040 (!perf_missing_features.sample_id_all &&
1041 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1042 opts->sample_time_set)))
1043 evsel__set_sample_bit(evsel, TIME);
1044
1045 if (opts->raw_samples && !evsel->no_aux_samples) {
1046 evsel__set_sample_bit(evsel, TIME);
1047 evsel__set_sample_bit(evsel, RAW);
1048 evsel__set_sample_bit(evsel, CPU);
1049 }
1050
1051 if (opts->sample_address)
1052 evsel__set_sample_bit(evsel, DATA_SRC);
1053
1054 if (opts->sample_phys_addr)
1055 evsel__set_sample_bit(evsel, PHYS_ADDR);
1056
1057 if (opts->no_buffering) {
1058 attr->watermark = 0;
1059 attr->wakeup_events = 1;
1060 }
1061 if (opts->branch_stack && !evsel->no_aux_samples) {
1062 evsel__set_sample_bit(evsel, BRANCH_STACK);
1063 attr->branch_sample_type = opts->branch_stack;
1064 }
1065
1066 if (opts->sample_weight)
1067 evsel__set_sample_bit(evsel, WEIGHT);
1068
1069 attr->task = track;
1070 attr->mmap = track;
1071 attr->mmap2 = track && !perf_missing_features.mmap2;
1072 attr->comm = track;
1073 /*
1074 * ksymbol is tracked separately with text poke because it needs to be
1075 * system wide and enabled immediately.
1076 */
1077 if (!opts->text_poke)
1078 attr->ksymbol = track && !perf_missing_features.ksymbol;
1079 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1080
1081 if (opts->record_namespaces)
1082 attr->namespaces = track;
1083
1084 if (opts->record_cgroup) {
1085 attr->cgroup = track && !perf_missing_features.cgroup;
1086 evsel__set_sample_bit(evsel, CGROUP);
1087 }
1088
1089 if (opts->record_switch_events)
1090 attr->context_switch = track;
1091
1092 if (opts->sample_transaction)
1093 evsel__set_sample_bit(evsel, TRANSACTION);
1094
1095 if (opts->running_time) {
1096 evsel->core.attr.read_format |=
1097 PERF_FORMAT_TOTAL_TIME_ENABLED |
1098 PERF_FORMAT_TOTAL_TIME_RUNNING;
1099 }
1100
1101 /*
1102 * XXX see the function comment above
1103 *
1104 * Disabling only independent events or group leaders,
1105 * keeping group members enabled.
1106 */
1107 if (evsel__is_group_leader(evsel))
1108 attr->disabled = 1;
1109
1110 /*
1111 * Setting enable_on_exec for independent events and
1112 * group leaders for traced executed by perf.
1113 */
1114 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1115 !opts->initial_delay)
1116 attr->enable_on_exec = 1;
1117
1118 if (evsel->immediate) {
1119 attr->disabled = 0;
1120 attr->enable_on_exec = 0;
1121 }
1122
1123 clockid = opts->clockid;
1124 if (opts->use_clockid) {
1125 attr->use_clockid = 1;
1126 attr->clockid = opts->clockid;
1127 }
1128
1129 if (evsel->precise_max)
1130 attr->precise_ip = 3;
1131
1132 if (opts->all_user) {
1133 attr->exclude_kernel = 1;
1134 attr->exclude_user = 0;
1135 }
1136
1137 if (opts->all_kernel) {
1138 attr->exclude_kernel = 0;
1139 attr->exclude_user = 1;
1140 }
1141
1142 if (evsel->core.own_cpus || evsel->unit)
1143 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1144
1145 /*
1146 * Apply event specific term settings,
1147 * it overloads any global configuration.
1148 */
1149 evsel__apply_config_terms(evsel, opts, track);
1150
1151 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1152
1153 /* The --period option takes the precedence. */
1154 if (opts->period_set) {
1155 if (opts->period)
1156 evsel__set_sample_bit(evsel, PERIOD);
1157 else
1158 evsel__reset_sample_bit(evsel, PERIOD);
1159 }
1160
1161 /*
1162 * A dummy event never triggers any actual counter and therefore
1163 * cannot be used with branch_stack.
1164 *
1165 * For initial_delay, a dummy event is added implicitly.
1166 * The software event will trigger -EOPNOTSUPP error out,
1167 * if BRANCH_STACK bit is set.
1168 */
1169 if (evsel__is_dummy_event(evsel))
1170 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1171}
1172
1173int evsel__set_filter(struct evsel *evsel, const char *filter)
1174{
1175 char *new_filter = strdup(filter);
1176
1177 if (new_filter != NULL) {
1178 free(evsel->filter);
1179 evsel->filter = new_filter;
1180 return 0;
1181 }
1182
1183 return -1;
1184}
1185
1186static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1187{
1188 char *new_filter;
1189
1190 if (evsel->filter == NULL)
1191 return evsel__set_filter(evsel, filter);
1192
1193 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1194 free(evsel->filter);
1195 evsel->filter = new_filter;
1196 return 0;
1197 }
1198
1199 return -1;
1200}
1201
1202int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1203{
1204 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1205}
1206
1207int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1208{
1209 return evsel__append_filter(evsel, "%s,%s", filter);
1210}
1211
1212/* Caller has to clear disabled after going through all CPUs. */
1213int evsel__enable_cpu(struct evsel *evsel, int cpu)
1214{
1215 return perf_evsel__enable_cpu(&evsel->core, cpu);
1216}
1217
1218int evsel__enable(struct evsel *evsel)
1219{
1220 int err = perf_evsel__enable(&evsel->core);
1221
1222 if (!err)
1223 evsel->disabled = false;
1224 return err;
1225}
1226
1227/* Caller has to set disabled after going through all CPUs. */
1228int evsel__disable_cpu(struct evsel *evsel, int cpu)
1229{
1230 return perf_evsel__disable_cpu(&evsel->core, cpu);
1231}
1232
1233int evsel__disable(struct evsel *evsel)
1234{
1235 int err = perf_evsel__disable(&evsel->core);
1236 /*
1237 * We mark it disabled here so that tools that disable a event can
1238 * ignore events after they disable it. I.e. the ring buffer may have
1239 * already a few more events queued up before the kernel got the stop
1240 * request.
1241 */
1242 if (!err)
1243 evsel->disabled = true;
1244
1245 return err;
1246}
1247
1248static void evsel__free_config_terms(struct evsel *evsel)
1249{
1250 struct evsel_config_term *term, *h;
1251
1252 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1253 list_del_init(&term->list);
1254 if (term->free_str)
1255 zfree(&term->val.str);
1256 free(term);
1257 }
1258}
1259
1260void evsel__exit(struct evsel *evsel)
1261{
1262 assert(list_empty(&evsel->core.node));
1263 assert(evsel->evlist == NULL);
1264 evsel__free_counts(evsel);
1265 perf_evsel__free_fd(&evsel->core);
1266 perf_evsel__free_id(&evsel->core);
1267 evsel__free_config_terms(evsel);
1268 cgroup__put(evsel->cgrp);
1269 perf_cpu_map__put(evsel->core.cpus);
1270 perf_cpu_map__put(evsel->core.own_cpus);
1271 perf_thread_map__put(evsel->core.threads);
1272 zfree(&evsel->group_name);
1273 zfree(&evsel->name);
1274 zfree(&evsel->pmu_name);
1275 zfree(&evsel->per_pkg_mask);
1276 zfree(&evsel->metric_events);
1277 perf_evsel__object.fini(evsel);
1278}
1279
1280void evsel__delete(struct evsel *evsel)
1281{
1282 evsel__exit(evsel);
1283 free(evsel);
1284}
1285
1286void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1287 struct perf_counts_values *count)
1288{
1289 struct perf_counts_values tmp;
1290
1291 if (!evsel->prev_raw_counts)
1292 return;
1293
1294 if (cpu == -1) {
1295 tmp = evsel->prev_raw_counts->aggr;
1296 evsel->prev_raw_counts->aggr = *count;
1297 } else {
1298 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1299 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1300 }
1301
1302 count->val = count->val - tmp.val;
1303 count->ena = count->ena - tmp.ena;
1304 count->run = count->run - tmp.run;
1305}
1306
1307void perf_counts_values__scale(struct perf_counts_values *count,
1308 bool scale, s8 *pscaled)
1309{
1310 s8 scaled = 0;
1311
1312 if (scale) {
1313 if (count->run == 0) {
1314 scaled = -1;
1315 count->val = 0;
1316 } else if (count->run < count->ena) {
1317 scaled = 1;
1318 count->val = (u64)((double) count->val * count->ena / count->run);
1319 }
1320 }
1321
1322 if (pscaled)
1323 *pscaled = scaled;
1324}
1325
1326static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1327{
1328 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1329
1330 return perf_evsel__read(&evsel->core, cpu, thread, count);
1331}
1332
1333static void
1334perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1335 u64 val, u64 ena, u64 run)
1336{
1337 struct perf_counts_values *count;
1338
1339 count = perf_counts(counter->counts, cpu, thread);
1340
1341 count->val = val;
1342 count->ena = ena;
1343 count->run = run;
1344
1345 perf_counts__set_loaded(counter->counts, cpu, thread, true);
1346}
1347
1348static int
1349perf_evsel__process_group_data(struct evsel *leader,
1350 int cpu, int thread, u64 *data)
1351{
1352 u64 read_format = leader->core.attr.read_format;
1353 struct sample_read_value *v;
1354 u64 nr, ena = 0, run = 0, i;
1355
1356 nr = *data++;
1357
1358 if (nr != (u64) leader->core.nr_members)
1359 return -EINVAL;
1360
1361 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1362 ena = *data++;
1363
1364 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1365 run = *data++;
1366
1367 v = (struct sample_read_value *) data;
1368
1369 perf_evsel__set_count(leader, cpu, thread,
1370 v[0].value, ena, run);
1371
1372 for (i = 1; i < nr; i++) {
1373 struct evsel *counter;
1374
1375 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1376 if (!counter)
1377 return -EINVAL;
1378
1379 perf_evsel__set_count(counter, cpu, thread,
1380 v[i].value, ena, run);
1381 }
1382
1383 return 0;
1384}
1385
1386static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1387{
1388 struct perf_stat_evsel *ps = leader->stats;
1389 u64 read_format = leader->core.attr.read_format;
1390 int size = perf_evsel__read_size(&leader->core);
1391 u64 *data = ps->group_data;
1392
1393 if (!(read_format & PERF_FORMAT_ID))
1394 return -EINVAL;
1395
1396 if (!evsel__is_group_leader(leader))
1397 return -EINVAL;
1398
1399 if (!data) {
1400 data = zalloc(size);
1401 if (!data)
1402 return -ENOMEM;
1403
1404 ps->group_data = data;
1405 }
1406
1407 if (FD(leader, cpu, thread) < 0)
1408 return -EINVAL;
1409
1410 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1411 return -errno;
1412
1413 return perf_evsel__process_group_data(leader, cpu, thread, data);
1414}
1415
1416int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1417{
1418 u64 read_format = evsel->core.attr.read_format;
1419
1420 if (read_format & PERF_FORMAT_GROUP)
1421 return evsel__read_group(evsel, cpu, thread);
1422
1423 return evsel__read_one(evsel, cpu, thread);
1424}
1425
1426int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1427{
1428 struct perf_counts_values count;
1429 size_t nv = scale ? 3 : 1;
1430
1431 if (FD(evsel, cpu, thread) < 0)
1432 return -EINVAL;
1433
1434 if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1435 return -ENOMEM;
1436
1437 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1438 return -errno;
1439
1440 evsel__compute_deltas(evsel, cpu, thread, &count);
1441 perf_counts_values__scale(&count, scale, NULL);
1442 *perf_counts(evsel->counts, cpu, thread) = count;
1443 return 0;
1444}
1445
1446static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1447{
1448 struct evsel *leader = evsel->leader;
1449 int fd;
1450
1451 if (evsel__is_group_leader(evsel))
1452 return -1;
1453
1454 /*
1455 * Leader must be already processed/open,
1456 * if not it's a bug.
1457 */
1458 BUG_ON(!leader->core.fd);
1459
1460 fd = FD(leader, cpu, thread);
1461 BUG_ON(fd == -1);
1462
1463 return fd;
1464}
1465
1466static void perf_evsel__remove_fd(struct evsel *pos,
1467 int nr_cpus, int nr_threads,
1468 int thread_idx)
1469{
1470 for (int cpu = 0; cpu < nr_cpus; cpu++)
1471 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1472 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1473}
1474
1475static int update_fds(struct evsel *evsel,
1476 int nr_cpus, int cpu_idx,
1477 int nr_threads, int thread_idx)
1478{
1479 struct evsel *pos;
1480
1481 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1482 return -EINVAL;
1483
1484 evlist__for_each_entry(evsel->evlist, pos) {
1485 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1486
1487 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1488
1489 /*
1490 * Since fds for next evsel has not been created,
1491 * there is no need to iterate whole event list.
1492 */
1493 if (pos == evsel)
1494 break;
1495 }
1496 return 0;
1497}
1498
1499static bool ignore_missing_thread(struct evsel *evsel,
1500 int nr_cpus, int cpu,
1501 struct perf_thread_map *threads,
1502 int thread, int err)
1503{
1504 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1505
1506 if (!evsel->ignore_missing_thread)
1507 return false;
1508
1509 /* The system wide setup does not work with threads. */
1510 if (evsel->core.system_wide)
1511 return false;
1512
1513 /* The -ESRCH is perf event syscall errno for pid's not found. */
1514 if (err != -ESRCH)
1515 return false;
1516
1517 /* If there's only one thread, let it fail. */
1518 if (threads->nr == 1)
1519 return false;
1520
1521 /*
1522 * We should remove fd for missing_thread first
1523 * because thread_map__remove() will decrease threads->nr.
1524 */
1525 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1526 return false;
1527
1528 if (thread_map__remove(threads, thread))
1529 return false;
1530
1531 pr_warning("WARNING: Ignored open failure for pid %d\n",
1532 ignore_pid);
1533 return true;
1534}
1535
1536static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1537 void *priv __maybe_unused)
1538{
1539 return fprintf(fp, " %-32s %s\n", name, val);
1540}
1541
1542static void display_attr(struct perf_event_attr *attr)
1543{
1544 if (verbose >= 2 || debug_peo_args) {
1545 fprintf(stderr, "%.60s\n", graph_dotted_line);
1546 fprintf(stderr, "perf_event_attr:\n");
1547 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1548 fprintf(stderr, "%.60s\n", graph_dotted_line);
1549 }
1550}
1551
1552static int perf_event_open(struct evsel *evsel,
1553 pid_t pid, int cpu, int group_fd,
1554 unsigned long flags)
1555{
1556 int precise_ip = evsel->core.attr.precise_ip;
1557 int fd;
1558
1559 while (1) {
1560 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1561 pid, cpu, group_fd, flags);
1562
1563 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1564 if (fd >= 0)
1565 break;
1566
1567 /* Do not try less precise if not requested. */
1568 if (!evsel->precise_max)
1569 break;
1570
1571 /*
1572 * We tried all the precise_ip values, and it's
1573 * still failing, so leave it to standard fallback.
1574 */
1575 if (!evsel->core.attr.precise_ip) {
1576 evsel->core.attr.precise_ip = precise_ip;
1577 break;
1578 }
1579
1580 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1581 evsel->core.attr.precise_ip--;
1582 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1583 display_attr(&evsel->core.attr);
1584 }
1585
1586 return fd;
1587}
1588
1589static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1590 struct perf_thread_map *threads,
1591 int start_cpu, int end_cpu)
1592{
1593 int cpu, thread, nthreads;
1594 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1595 int pid = -1, err, old_errno;
1596 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1597
1598 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1599 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1600 return -EINVAL;
1601
1602 if (cpus == NULL) {
1603 static struct perf_cpu_map *empty_cpu_map;
1604
1605 if (empty_cpu_map == NULL) {
1606 empty_cpu_map = perf_cpu_map__dummy_new();
1607 if (empty_cpu_map == NULL)
1608 return -ENOMEM;
1609 }
1610
1611 cpus = empty_cpu_map;
1612 }
1613
1614 if (threads == NULL) {
1615 static struct perf_thread_map *empty_thread_map;
1616
1617 if (empty_thread_map == NULL) {
1618 empty_thread_map = thread_map__new_by_tid(-1);
1619 if (empty_thread_map == NULL)
1620 return -ENOMEM;
1621 }
1622
1623 threads = empty_thread_map;
1624 }
1625
1626 if (evsel->core.system_wide)
1627 nthreads = 1;
1628 else
1629 nthreads = threads->nr;
1630
1631 if (evsel->core.fd == NULL &&
1632 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1633 return -ENOMEM;
1634
1635 if (evsel->cgrp) {
1636 flags |= PERF_FLAG_PID_CGROUP;
1637 pid = evsel->cgrp->fd;
1638 }
1639
1640fallback_missing_features:
1641 if (perf_missing_features.clockid_wrong)
1642 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1643 if (perf_missing_features.clockid) {
1644 evsel->core.attr.use_clockid = 0;
1645 evsel->core.attr.clockid = 0;
1646 }
1647 if (perf_missing_features.cloexec)
1648 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1649 if (perf_missing_features.mmap2)
1650 evsel->core.attr.mmap2 = 0;
1651 if (perf_missing_features.exclude_guest)
1652 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1653 if (perf_missing_features.lbr_flags)
1654 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1655 PERF_SAMPLE_BRANCH_NO_CYCLES);
1656 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1657 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1658 if (perf_missing_features.ksymbol)
1659 evsel->core.attr.ksymbol = 0;
1660 if (perf_missing_features.bpf)
1661 evsel->core.attr.bpf_event = 0;
1662 if (perf_missing_features.branch_hw_idx)
1663 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1664retry_sample_id:
1665 if (perf_missing_features.sample_id_all)
1666 evsel->core.attr.sample_id_all = 0;
1667
1668 display_attr(&evsel->core.attr);
1669
1670 for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1671
1672 for (thread = 0; thread < nthreads; thread++) {
1673 int fd, group_fd;
1674
1675 if (!evsel->cgrp && !evsel->core.system_wide)
1676 pid = perf_thread_map__pid(threads, thread);
1677
1678 group_fd = get_group_fd(evsel, cpu, thread);
1679retry_open:
1680 test_attr__ready();
1681
1682 fd = perf_event_open(evsel, pid, cpus->map[cpu],
1683 group_fd, flags);
1684
1685 FD(evsel, cpu, thread) = fd;
1686
1687 if (fd < 0) {
1688 err = -errno;
1689
1690 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1691 /*
1692 * We just removed 1 thread, so take a step
1693 * back on thread index and lower the upper
1694 * nthreads limit.
1695 */
1696 nthreads--;
1697 thread--;
1698
1699 /* ... and pretend like nothing have happened. */
1700 err = 0;
1701 continue;
1702 }
1703
1704 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1705 err);
1706 goto try_fallback;
1707 }
1708
1709 pr_debug2_peo(" = %d\n", fd);
1710
1711 if (evsel->bpf_fd >= 0) {
1712 int evt_fd = fd;
1713 int bpf_fd = evsel->bpf_fd;
1714
1715 err = ioctl(evt_fd,
1716 PERF_EVENT_IOC_SET_BPF,
1717 bpf_fd);
1718 if (err && errno != EEXIST) {
1719 pr_err("failed to attach bpf fd %d: %s\n",
1720 bpf_fd, strerror(errno));
1721 err = -EINVAL;
1722 goto out_close;
1723 }
1724 }
1725
1726 set_rlimit = NO_CHANGE;
1727
1728 /*
1729 * If we succeeded but had to kill clockid, fail and
1730 * have evsel__open_strerror() print us a nice error.
1731 */
1732 if (perf_missing_features.clockid ||
1733 perf_missing_features.clockid_wrong) {
1734 err = -EINVAL;
1735 goto out_close;
1736 }
1737 }
1738 }
1739
1740 return 0;
1741
1742try_fallback:
1743 /*
1744 * perf stat needs between 5 and 22 fds per CPU. When we run out
1745 * of them try to increase the limits.
1746 */
1747 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1748 struct rlimit l;
1749
1750 old_errno = errno;
1751 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1752 if (set_rlimit == NO_CHANGE)
1753 l.rlim_cur = l.rlim_max;
1754 else {
1755 l.rlim_cur = l.rlim_max + 1000;
1756 l.rlim_max = l.rlim_cur;
1757 }
1758 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1759 set_rlimit++;
1760 errno = old_errno;
1761 goto retry_open;
1762 }
1763 }
1764 errno = old_errno;
1765 }
1766
1767 if (err != -EINVAL || cpu > 0 || thread > 0)
1768 goto out_close;
1769
1770 /*
1771 * Must probe features in the order they were added to the
1772 * perf_event_attr interface.
1773 */
1774 if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1775 perf_missing_features.cgroup = true;
1776 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1777 goto out_close;
1778 } else if (!perf_missing_features.branch_hw_idx &&
1779 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1780 perf_missing_features.branch_hw_idx = true;
1781 pr_debug2("switching off branch HW index support\n");
1782 goto fallback_missing_features;
1783 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1784 perf_missing_features.aux_output = true;
1785 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1786 goto out_close;
1787 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1788 perf_missing_features.bpf = true;
1789 pr_debug2_peo("switching off bpf_event\n");
1790 goto fallback_missing_features;
1791 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1792 perf_missing_features.ksymbol = true;
1793 pr_debug2_peo("switching off ksymbol\n");
1794 goto fallback_missing_features;
1795 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1796 perf_missing_features.write_backward = true;
1797 pr_debug2_peo("switching off write_backward\n");
1798 goto out_close;
1799 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1800 perf_missing_features.clockid_wrong = true;
1801 pr_debug2_peo("switching off clockid\n");
1802 goto fallback_missing_features;
1803 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1804 perf_missing_features.clockid = true;
1805 pr_debug2_peo("switching off use_clockid\n");
1806 goto fallback_missing_features;
1807 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1808 perf_missing_features.cloexec = true;
1809 pr_debug2_peo("switching off cloexec flag\n");
1810 goto fallback_missing_features;
1811 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1812 perf_missing_features.mmap2 = true;
1813 pr_debug2_peo("switching off mmap2\n");
1814 goto fallback_missing_features;
1815 } else if (!perf_missing_features.exclude_guest &&
1816 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1817 perf_missing_features.exclude_guest = true;
1818 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1819 goto fallback_missing_features;
1820 } else if (!perf_missing_features.sample_id_all) {
1821 perf_missing_features.sample_id_all = true;
1822 pr_debug2_peo("switching off sample_id_all\n");
1823 goto retry_sample_id;
1824 } else if (!perf_missing_features.lbr_flags &&
1825 (evsel->core.attr.branch_sample_type &
1826 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1827 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1828 perf_missing_features.lbr_flags = true;
1829 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1830 goto fallback_missing_features;
1831 } else if (!perf_missing_features.group_read &&
1832 evsel->core.attr.inherit &&
1833 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1834 evsel__is_group_leader(evsel)) {
1835 perf_missing_features.group_read = true;
1836 pr_debug2_peo("switching off group read\n");
1837 goto fallback_missing_features;
1838 }
1839out_close:
1840 if (err)
1841 threads->err_thread = thread;
1842
1843 old_errno = errno;
1844 do {
1845 while (--thread >= 0) {
1846 if (FD(evsel, cpu, thread) >= 0)
1847 close(FD(evsel, cpu, thread));
1848 FD(evsel, cpu, thread) = -1;
1849 }
1850 thread = nthreads;
1851 } while (--cpu >= 0);
1852 errno = old_errno;
1853 return err;
1854}
1855
1856int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1857 struct perf_thread_map *threads)
1858{
1859 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1860}
1861
1862void evsel__close(struct evsel *evsel)
1863{
1864 perf_evsel__close(&evsel->core);
1865 perf_evsel__free_id(&evsel->core);
1866}
1867
1868int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
1869{
1870 if (cpu == -1)
1871 return evsel__open_cpu(evsel, cpus, NULL, 0,
1872 cpus ? cpus->nr : 1);
1873
1874 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1875}
1876
1877int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
1878{
1879 return evsel__open(evsel, NULL, threads);
1880}
1881
1882static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1883 const union perf_event *event,
1884 struct perf_sample *sample)
1885{
1886 u64 type = evsel->core.attr.sample_type;
1887 const __u64 *array = event->sample.array;
1888 bool swapped = evsel->needs_swap;
1889 union u64_swap u;
1890
1891 array += ((event->header.size -
1892 sizeof(event->header)) / sizeof(u64)) - 1;
1893
1894 if (type & PERF_SAMPLE_IDENTIFIER) {
1895 sample->id = *array;
1896 array--;
1897 }
1898
1899 if (type & PERF_SAMPLE_CPU) {
1900 u.val64 = *array;
1901 if (swapped) {
1902 /* undo swap of u64, then swap on individual u32s */
1903 u.val64 = bswap_64(u.val64);
1904 u.val32[0] = bswap_32(u.val32[0]);
1905 }
1906
1907 sample->cpu = u.val32[0];
1908 array--;
1909 }
1910
1911 if (type & PERF_SAMPLE_STREAM_ID) {
1912 sample->stream_id = *array;
1913 array--;
1914 }
1915
1916 if (type & PERF_SAMPLE_ID) {
1917 sample->id = *array;
1918 array--;
1919 }
1920
1921 if (type & PERF_SAMPLE_TIME) {
1922 sample->time = *array;
1923 array--;
1924 }
1925
1926 if (type & PERF_SAMPLE_TID) {
1927 u.val64 = *array;
1928 if (swapped) {
1929 /* undo swap of u64, then swap on individual u32s */
1930 u.val64 = bswap_64(u.val64);
1931 u.val32[0] = bswap_32(u.val32[0]);
1932 u.val32[1] = bswap_32(u.val32[1]);
1933 }
1934
1935 sample->pid = u.val32[0];
1936 sample->tid = u.val32[1];
1937 array--;
1938 }
1939
1940 return 0;
1941}
1942
1943static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1944 u64 size)
1945{
1946 return size > max_size || offset + size > endp;
1947}
1948
1949#define OVERFLOW_CHECK(offset, size, max_size) \
1950 do { \
1951 if (overflow(endp, (max_size), (offset), (size))) \
1952 return -EFAULT; \
1953 } while (0)
1954
1955#define OVERFLOW_CHECK_u64(offset) \
1956 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1957
1958static int
1959perf_event__check_size(union perf_event *event, unsigned int sample_size)
1960{
1961 /*
1962 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1963 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
1964 * check the format does not go past the end of the event.
1965 */
1966 if (sample_size + sizeof(event->header) > event->header.size)
1967 return -EFAULT;
1968
1969 return 0;
1970}
1971
1972int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1973 struct perf_sample *data)
1974{
1975 u64 type = evsel->core.attr.sample_type;
1976 bool swapped = evsel->needs_swap;
1977 const __u64 *array;
1978 u16 max_size = event->header.size;
1979 const void *endp = (void *)event + max_size;
1980 u64 sz;
1981
1982 /*
1983 * used for cross-endian analysis. See git commit 65014ab3
1984 * for why this goofiness is needed.
1985 */
1986 union u64_swap u;
1987
1988 memset(data, 0, sizeof(*data));
1989 data->cpu = data->pid = data->tid = -1;
1990 data->stream_id = data->id = data->time = -1ULL;
1991 data->period = evsel->core.attr.sample_period;
1992 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1993 data->misc = event->header.misc;
1994 data->id = -1ULL;
1995 data->data_src = PERF_MEM_DATA_SRC_NONE;
1996
1997 if (event->header.type != PERF_RECORD_SAMPLE) {
1998 if (!evsel->core.attr.sample_id_all)
1999 return 0;
2000 return perf_evsel__parse_id_sample(evsel, event, data);
2001 }
2002
2003 array = event->sample.array;
2004
2005 if (perf_event__check_size(event, evsel->sample_size))
2006 return -EFAULT;
2007
2008 if (type & PERF_SAMPLE_IDENTIFIER) {
2009 data->id = *array;
2010 array++;
2011 }
2012
2013 if (type & PERF_SAMPLE_IP) {
2014 data->ip = *array;
2015 array++;
2016 }
2017
2018 if (type & PERF_SAMPLE_TID) {
2019 u.val64 = *array;
2020 if (swapped) {
2021 /* undo swap of u64, then swap on individual u32s */
2022 u.val64 = bswap_64(u.val64);
2023 u.val32[0] = bswap_32(u.val32[0]);
2024 u.val32[1] = bswap_32(u.val32[1]);
2025 }
2026
2027 data->pid = u.val32[0];
2028 data->tid = u.val32[1];
2029 array++;
2030 }
2031
2032 if (type & PERF_SAMPLE_TIME) {
2033 data->time = *array;
2034 array++;
2035 }
2036
2037 if (type & PERF_SAMPLE_ADDR) {
2038 data->addr = *array;
2039 array++;
2040 }
2041
2042 if (type & PERF_SAMPLE_ID) {
2043 data->id = *array;
2044 array++;
2045 }
2046
2047 if (type & PERF_SAMPLE_STREAM_ID) {
2048 data->stream_id = *array;
2049 array++;
2050 }
2051
2052 if (type & PERF_SAMPLE_CPU) {
2053
2054 u.val64 = *array;
2055 if (swapped) {
2056 /* undo swap of u64, then swap on individual u32s */
2057 u.val64 = bswap_64(u.val64);
2058 u.val32[0] = bswap_32(u.val32[0]);
2059 }
2060
2061 data->cpu = u.val32[0];
2062 array++;
2063 }
2064
2065 if (type & PERF_SAMPLE_PERIOD) {
2066 data->period = *array;
2067 array++;
2068 }
2069
2070 if (type & PERF_SAMPLE_READ) {
2071 u64 read_format = evsel->core.attr.read_format;
2072
2073 OVERFLOW_CHECK_u64(array);
2074 if (read_format & PERF_FORMAT_GROUP)
2075 data->read.group.nr = *array;
2076 else
2077 data->read.one.value = *array;
2078
2079 array++;
2080
2081 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2082 OVERFLOW_CHECK_u64(array);
2083 data->read.time_enabled = *array;
2084 array++;
2085 }
2086
2087 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2088 OVERFLOW_CHECK_u64(array);
2089 data->read.time_running = *array;
2090 array++;
2091 }
2092
2093 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2094 if (read_format & PERF_FORMAT_GROUP) {
2095 const u64 max_group_nr = UINT64_MAX /
2096 sizeof(struct sample_read_value);
2097
2098 if (data->read.group.nr > max_group_nr)
2099 return -EFAULT;
2100 sz = data->read.group.nr *
2101 sizeof(struct sample_read_value);
2102 OVERFLOW_CHECK(array, sz, max_size);
2103 data->read.group.values =
2104 (struct sample_read_value *)array;
2105 array = (void *)array + sz;
2106 } else {
2107 OVERFLOW_CHECK_u64(array);
2108 data->read.one.id = *array;
2109 array++;
2110 }
2111 }
2112
2113 if (type & PERF_SAMPLE_CALLCHAIN) {
2114 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2115
2116 OVERFLOW_CHECK_u64(array);
2117 data->callchain = (struct ip_callchain *)array++;
2118 if (data->callchain->nr > max_callchain_nr)
2119 return -EFAULT;
2120 sz = data->callchain->nr * sizeof(u64);
2121 OVERFLOW_CHECK(array, sz, max_size);
2122 array = (void *)array + sz;
2123 }
2124
2125 if (type & PERF_SAMPLE_RAW) {
2126 OVERFLOW_CHECK_u64(array);
2127 u.val64 = *array;
2128
2129 /*
2130 * Undo swap of u64, then swap on individual u32s,
2131 * get the size of the raw area and undo all of the
2132 * swap. The pevent interface handles endianity by
2133 * itself.
2134 */
2135 if (swapped) {
2136 u.val64 = bswap_64(u.val64);
2137 u.val32[0] = bswap_32(u.val32[0]);
2138 u.val32[1] = bswap_32(u.val32[1]);
2139 }
2140 data->raw_size = u.val32[0];
2141
2142 /*
2143 * The raw data is aligned on 64bits including the
2144 * u32 size, so it's safe to use mem_bswap_64.
2145 */
2146 if (swapped)
2147 mem_bswap_64((void *) array, data->raw_size);
2148
2149 array = (void *)array + sizeof(u32);
2150
2151 OVERFLOW_CHECK(array, data->raw_size, max_size);
2152 data->raw_data = (void *)array;
2153 array = (void *)array + data->raw_size;
2154 }
2155
2156 if (type & PERF_SAMPLE_BRANCH_STACK) {
2157 const u64 max_branch_nr = UINT64_MAX /
2158 sizeof(struct branch_entry);
2159
2160 OVERFLOW_CHECK_u64(array);
2161 data->branch_stack = (struct branch_stack *)array++;
2162
2163 if (data->branch_stack->nr > max_branch_nr)
2164 return -EFAULT;
2165
2166 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2167 if (evsel__has_branch_hw_idx(evsel))
2168 sz += sizeof(u64);
2169 else
2170 data->no_hw_idx = true;
2171 OVERFLOW_CHECK(array, sz, max_size);
2172 array = (void *)array + sz;
2173 }
2174
2175 if (type & PERF_SAMPLE_REGS_USER) {
2176 OVERFLOW_CHECK_u64(array);
2177 data->user_regs.abi = *array;
2178 array++;
2179
2180 if (data->user_regs.abi) {
2181 u64 mask = evsel->core.attr.sample_regs_user;
2182
2183 sz = hweight64(mask) * sizeof(u64);
2184 OVERFLOW_CHECK(array, sz, max_size);
2185 data->user_regs.mask = mask;
2186 data->user_regs.regs = (u64 *)array;
2187 array = (void *)array + sz;
2188 }
2189 }
2190
2191 if (type & PERF_SAMPLE_STACK_USER) {
2192 OVERFLOW_CHECK_u64(array);
2193 sz = *array++;
2194
2195 data->user_stack.offset = ((char *)(array - 1)
2196 - (char *) event);
2197
2198 if (!sz) {
2199 data->user_stack.size = 0;
2200 } else {
2201 OVERFLOW_CHECK(array, sz, max_size);
2202 data->user_stack.data = (char *)array;
2203 array = (void *)array + sz;
2204 OVERFLOW_CHECK_u64(array);
2205 data->user_stack.size = *array++;
2206 if (WARN_ONCE(data->user_stack.size > sz,
2207 "user stack dump failure\n"))
2208 return -EFAULT;
2209 }
2210 }
2211
2212 if (type & PERF_SAMPLE_WEIGHT) {
2213 OVERFLOW_CHECK_u64(array);
2214 data->weight = *array;
2215 array++;
2216 }
2217
2218 if (type & PERF_SAMPLE_DATA_SRC) {
2219 OVERFLOW_CHECK_u64(array);
2220 data->data_src = *array;
2221 array++;
2222 }
2223
2224 if (type & PERF_SAMPLE_TRANSACTION) {
2225 OVERFLOW_CHECK_u64(array);
2226 data->transaction = *array;
2227 array++;
2228 }
2229
2230 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2231 if (type & PERF_SAMPLE_REGS_INTR) {
2232 OVERFLOW_CHECK_u64(array);
2233 data->intr_regs.abi = *array;
2234 array++;
2235
2236 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2237 u64 mask = evsel->core.attr.sample_regs_intr;
2238
2239 sz = hweight64(mask) * sizeof(u64);
2240 OVERFLOW_CHECK(array, sz, max_size);
2241 data->intr_regs.mask = mask;
2242 data->intr_regs.regs = (u64 *)array;
2243 array = (void *)array + sz;
2244 }
2245 }
2246
2247 data->phys_addr = 0;
2248 if (type & PERF_SAMPLE_PHYS_ADDR) {
2249 data->phys_addr = *array;
2250 array++;
2251 }
2252
2253 data->cgroup = 0;
2254 if (type & PERF_SAMPLE_CGROUP) {
2255 data->cgroup = *array;
2256 array++;
2257 }
2258
2259 if (type & PERF_SAMPLE_AUX) {
2260 OVERFLOW_CHECK_u64(array);
2261 sz = *array++;
2262
2263 OVERFLOW_CHECK(array, sz, max_size);
2264 /* Undo swap of data */
2265 if (swapped)
2266 mem_bswap_64((char *)array, sz);
2267 data->aux_sample.size = sz;
2268 data->aux_sample.data = (char *)array;
2269 array = (void *)array + sz;
2270 }
2271
2272 return 0;
2273}
2274
2275int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2276 u64 *timestamp)
2277{
2278 u64 type = evsel->core.attr.sample_type;
2279 const __u64 *array;
2280
2281 if (!(type & PERF_SAMPLE_TIME))
2282 return -1;
2283
2284 if (event->header.type != PERF_RECORD_SAMPLE) {
2285 struct perf_sample data = {
2286 .time = -1ULL,
2287 };
2288
2289 if (!evsel->core.attr.sample_id_all)
2290 return -1;
2291 if (perf_evsel__parse_id_sample(evsel, event, &data))
2292 return -1;
2293
2294 *timestamp = data.time;
2295 return 0;
2296 }
2297
2298 array = event->sample.array;
2299
2300 if (perf_event__check_size(event, evsel->sample_size))
2301 return -EFAULT;
2302
2303 if (type & PERF_SAMPLE_IDENTIFIER)
2304 array++;
2305
2306 if (type & PERF_SAMPLE_IP)
2307 array++;
2308
2309 if (type & PERF_SAMPLE_TID)
2310 array++;
2311
2312 if (type & PERF_SAMPLE_TIME)
2313 *timestamp = *array;
2314
2315 return 0;
2316}
2317
2318struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2319{
2320 return tep_find_field(evsel->tp_format, name);
2321}
2322
2323void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2324{
2325 struct tep_format_field *field = evsel__field(evsel, name);
2326 int offset;
2327
2328 if (!field)
2329 return NULL;
2330
2331 offset = field->offset;
2332
2333 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2334 offset = *(int *)(sample->raw_data + field->offset);
2335 offset &= 0xffff;
2336 }
2337
2338 return sample->raw_data + offset;
2339}
2340
2341u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2342 bool needs_swap)
2343{
2344 u64 value;
2345 void *ptr = sample->raw_data + field->offset;
2346
2347 switch (field->size) {
2348 case 1:
2349 return *(u8 *)ptr;
2350 case 2:
2351 value = *(u16 *)ptr;
2352 break;
2353 case 4:
2354 value = *(u32 *)ptr;
2355 break;
2356 case 8:
2357 memcpy(&value, ptr, sizeof(u64));
2358 break;
2359 default:
2360 return 0;
2361 }
2362
2363 if (!needs_swap)
2364 return value;
2365
2366 switch (field->size) {
2367 case 2:
2368 return bswap_16(value);
2369 case 4:
2370 return bswap_32(value);
2371 case 8:
2372 return bswap_64(value);
2373 default:
2374 return 0;
2375 }
2376
2377 return 0;
2378}
2379
2380u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2381{
2382 struct tep_format_field *field = evsel__field(evsel, name);
2383
2384 if (!field)
2385 return 0;
2386
2387 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2388}
2389
2390bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2391{
2392 int paranoid;
2393
2394 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2395 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2396 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2397 /*
2398 * If it's cycles then fall back to hrtimer based
2399 * cpu-clock-tick sw counter, which is always available even if
2400 * no PMU support.
2401 *
2402 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2403 * b0a873e).
2404 */
2405 scnprintf(msg, msgsize, "%s",
2406"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2407
2408 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2409 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2410
2411 zfree(&evsel->name);
2412 return true;
2413 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2414 (paranoid = perf_event_paranoid()) > 1) {
2415 const char *name = evsel__name(evsel);
2416 char *new_name;
2417 const char *sep = ":";
2418
2419 /* If event has exclude user then don't exclude kernel. */
2420 if (evsel->core.attr.exclude_user)
2421 return false;
2422
2423 /* Is there already the separator in the name. */
2424 if (strchr(name, '/') ||
2425 (strchr(name, ':') && !evsel->is_libpfm_event))
2426 sep = "";
2427
2428 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2429 return false;
2430
2431 if (evsel->name)
2432 free(evsel->name);
2433 evsel->name = new_name;
2434 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2435 "to fall back to excluding kernel and hypervisor "
2436 " samples", paranoid);
2437 evsel->core.attr.exclude_kernel = 1;
2438 evsel->core.attr.exclude_hv = 1;
2439
2440 return true;
2441 }
2442
2443 return false;
2444}
2445
2446static bool find_process(const char *name)
2447{
2448 size_t len = strlen(name);
2449 DIR *dir;
2450 struct dirent *d;
2451 int ret = -1;
2452
2453 dir = opendir(procfs__mountpoint());
2454 if (!dir)
2455 return false;
2456
2457 /* Walk through the directory. */
2458 while (ret && (d = readdir(dir)) != NULL) {
2459 char path[PATH_MAX];
2460 char *data;
2461 size_t size;
2462
2463 if ((d->d_type != DT_DIR) ||
2464 !strcmp(".", d->d_name) ||
2465 !strcmp("..", d->d_name))
2466 continue;
2467
2468 scnprintf(path, sizeof(path), "%s/%s/comm",
2469 procfs__mountpoint(), d->d_name);
2470
2471 if (filename__read_str(path, &data, &size))
2472 continue;
2473
2474 ret = strncmp(name, data, len);
2475 free(data);
2476 }
2477
2478 closedir(dir);
2479 return ret ? false : true;
2480}
2481
2482int evsel__open_strerror(struct evsel *evsel, struct target *target,
2483 int err, char *msg, size_t size)
2484{
2485 char sbuf[STRERR_BUFSIZE];
2486 int printed = 0, enforced = 0;
2487
2488 switch (err) {
2489 case EPERM:
2490 case EACCES:
2491 printed += scnprintf(msg + printed, size - printed,
2492 "Access to performance monitoring and observability operations is limited.\n");
2493
2494 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2495 if (enforced) {
2496 printed += scnprintf(msg + printed, size - printed,
2497 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2498 "monitoring and observability operations. Inspect system audit records for\n"
2499 "more perf_event access control information and adjusting the policy.\n");
2500 }
2501 }
2502
2503 if (err == EPERM)
2504 printed += scnprintf(msg, size,
2505 "No permission to enable %s event.\n\n", evsel__name(evsel));
2506
2507 return scnprintf(msg + printed, size - printed,
2508 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2509 "access to performance monitoring and observability operations for processes\n"
2510 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2511 "More information can be found at 'Perf events and tool security' document:\n"
2512 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2513 "perf_event_paranoid setting is %d:\n"
2514 " -1: Allow use of (almost) all events by all users\n"
2515 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2516 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2517 ">= 1: Disallow CPU event access\n"
2518 ">= 2: Disallow kernel profiling\n"
2519 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2520 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2521 perf_event_paranoid());
2522 case ENOENT:
2523 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2524 case EMFILE:
2525 return scnprintf(msg, size, "%s",
2526 "Too many events are opened.\n"
2527 "Probably the maximum number of open file descriptors has been reached.\n"
2528 "Hint: Try again after reducing the number of events.\n"
2529 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2530 case ENOMEM:
2531 if (evsel__has_callchain(evsel) &&
2532 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2533 return scnprintf(msg, size,
2534 "Not enough memory to setup event with callchain.\n"
2535 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2536 "Hint: Current value: %d", sysctl__max_stack());
2537 break;
2538 case ENODEV:
2539 if (target->cpu_list)
2540 return scnprintf(msg, size, "%s",
2541 "No such device - did you specify an out-of-range profile CPU?");
2542 break;
2543 case EOPNOTSUPP:
2544 if (evsel->core.attr.aux_output)
2545 return scnprintf(msg, size,
2546 "%s: PMU Hardware doesn't support 'aux_output' feature",
2547 evsel__name(evsel));
2548 if (evsel->core.attr.sample_period != 0)
2549 return scnprintf(msg, size,
2550 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2551 evsel__name(evsel));
2552 if (evsel->core.attr.precise_ip)
2553 return scnprintf(msg, size, "%s",
2554 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2555#if defined(__i386__) || defined(__x86_64__)
2556 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2557 return scnprintf(msg, size, "%s",
2558 "No hardware sampling interrupt available.\n");
2559#endif
2560 break;
2561 case EBUSY:
2562 if (find_process("oprofiled"))
2563 return scnprintf(msg, size,
2564 "The PMU counters are busy/taken by another profiler.\n"
2565 "We found oprofile daemon running, please stop it and try again.");
2566 break;
2567 case EINVAL:
2568 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2569 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2570 if (perf_missing_features.clockid)
2571 return scnprintf(msg, size, "clockid feature not supported.");
2572 if (perf_missing_features.clockid_wrong)
2573 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2574 if (perf_missing_features.aux_output)
2575 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2576 break;
2577 default:
2578 break;
2579 }
2580
2581 return scnprintf(msg, size,
2582 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2583 "/bin/dmesg | grep -i perf may provide additional information.\n",
2584 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2585}
2586
2587struct perf_env *evsel__env(struct evsel *evsel)
2588{
2589 if (evsel && evsel->evlist)
2590 return evsel->evlist->env;
2591 return &perf_env;
2592}
2593
2594static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2595{
2596 int cpu, thread;
2597
2598 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2599 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2600 thread++) {
2601 int fd = FD(evsel, cpu, thread);
2602
2603 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2604 cpu, thread, fd) < 0)
2605 return -1;
2606 }
2607 }
2608
2609 return 0;
2610}
2611
2612int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2613{
2614 struct perf_cpu_map *cpus = evsel->core.cpus;
2615 struct perf_thread_map *threads = evsel->core.threads;
2616
2617 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2618 return -ENOMEM;
2619
2620 return store_evsel_ids(evsel, evlist);
2621}