Loading...
1/*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9
10#include <byteswap.h>
11#include <errno.h>
12#include <inttypes.h>
13#include <linux/bitops.h>
14#include <api/fs/fs.h>
15#include <api/fs/tracing_path.h>
16#include <traceevent/event-parse.h>
17#include <linux/hw_breakpoint.h>
18#include <linux/perf_event.h>
19#include <linux/compiler.h>
20#include <linux/err.h>
21#include <sys/ioctl.h>
22#include <sys/resource.h>
23#include <sys/types.h>
24#include <dirent.h>
25#include "asm/bug.h"
26#include "callchain.h"
27#include "cgroup.h"
28#include "event.h"
29#include "evsel.h"
30#include "evlist.h"
31#include "util.h"
32#include "cpumap.h"
33#include "thread_map.h"
34#include "target.h"
35#include "perf_regs.h"
36#include "debug.h"
37#include "trace-event.h"
38#include "stat.h"
39#include "memswap.h"
40#include "util/parse-branch-options.h"
41
42#include "sane_ctype.h"
43
44struct perf_missing_features perf_missing_features;
45
46static clockid_t clockid;
47
48static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
49{
50 return 0;
51}
52
53void __weak test_attr__ready(void) { }
54
55static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
56{
57}
58
59static struct {
60 size_t size;
61 int (*init)(struct perf_evsel *evsel);
62 void (*fini)(struct perf_evsel *evsel);
63} perf_evsel__object = {
64 .size = sizeof(struct perf_evsel),
65 .init = perf_evsel__no_extra_init,
66 .fini = perf_evsel__no_extra_fini,
67};
68
69int perf_evsel__object_config(size_t object_size,
70 int (*init)(struct perf_evsel *evsel),
71 void (*fini)(struct perf_evsel *evsel))
72{
73
74 if (object_size == 0)
75 goto set_methods;
76
77 if (perf_evsel__object.size > object_size)
78 return -EINVAL;
79
80 perf_evsel__object.size = object_size;
81
82set_methods:
83 if (init != NULL)
84 perf_evsel__object.init = init;
85
86 if (fini != NULL)
87 perf_evsel__object.fini = fini;
88
89 return 0;
90}
91
92#define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
93
94int __perf_evsel__sample_size(u64 sample_type)
95{
96 u64 mask = sample_type & PERF_SAMPLE_MASK;
97 int size = 0;
98 int i;
99
100 for (i = 0; i < 64; i++) {
101 if (mask & (1ULL << i))
102 size++;
103 }
104
105 size *= sizeof(u64);
106
107 return size;
108}
109
110/**
111 * __perf_evsel__calc_id_pos - calculate id_pos.
112 * @sample_type: sample type
113 *
114 * This function returns the position of the event id (PERF_SAMPLE_ID or
115 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
116 * sample_event.
117 */
118static int __perf_evsel__calc_id_pos(u64 sample_type)
119{
120 int idx = 0;
121
122 if (sample_type & PERF_SAMPLE_IDENTIFIER)
123 return 0;
124
125 if (!(sample_type & PERF_SAMPLE_ID))
126 return -1;
127
128 if (sample_type & PERF_SAMPLE_IP)
129 idx += 1;
130
131 if (sample_type & PERF_SAMPLE_TID)
132 idx += 1;
133
134 if (sample_type & PERF_SAMPLE_TIME)
135 idx += 1;
136
137 if (sample_type & PERF_SAMPLE_ADDR)
138 idx += 1;
139
140 return idx;
141}
142
143/**
144 * __perf_evsel__calc_is_pos - calculate is_pos.
145 * @sample_type: sample type
146 *
147 * This function returns the position (counting backwards) of the event id
148 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
149 * sample_id_all is used there is an id sample appended to non-sample events.
150 */
151static int __perf_evsel__calc_is_pos(u64 sample_type)
152{
153 int idx = 1;
154
155 if (sample_type & PERF_SAMPLE_IDENTIFIER)
156 return 1;
157
158 if (!(sample_type & PERF_SAMPLE_ID))
159 return -1;
160
161 if (sample_type & PERF_SAMPLE_CPU)
162 idx += 1;
163
164 if (sample_type & PERF_SAMPLE_STREAM_ID)
165 idx += 1;
166
167 return idx;
168}
169
170void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
171{
172 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
173 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
174}
175
176void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
177 enum perf_event_sample_format bit)
178{
179 if (!(evsel->attr.sample_type & bit)) {
180 evsel->attr.sample_type |= bit;
181 evsel->sample_size += sizeof(u64);
182 perf_evsel__calc_id_pos(evsel);
183 }
184}
185
186void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
187 enum perf_event_sample_format bit)
188{
189 if (evsel->attr.sample_type & bit) {
190 evsel->attr.sample_type &= ~bit;
191 evsel->sample_size -= sizeof(u64);
192 perf_evsel__calc_id_pos(evsel);
193 }
194}
195
196void perf_evsel__set_sample_id(struct perf_evsel *evsel,
197 bool can_sample_identifier)
198{
199 if (can_sample_identifier) {
200 perf_evsel__reset_sample_bit(evsel, ID);
201 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
202 } else {
203 perf_evsel__set_sample_bit(evsel, ID);
204 }
205 evsel->attr.read_format |= PERF_FORMAT_ID;
206}
207
208/**
209 * perf_evsel__is_function_event - Return whether given evsel is a function
210 * trace event
211 *
212 * @evsel - evsel selector to be tested
213 *
214 * Return %true if event is function trace event
215 */
216bool perf_evsel__is_function_event(struct perf_evsel *evsel)
217{
218#define FUNCTION_EVENT "ftrace:function"
219
220 return evsel->name &&
221 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
222
223#undef FUNCTION_EVENT
224}
225
226void perf_evsel__init(struct perf_evsel *evsel,
227 struct perf_event_attr *attr, int idx)
228{
229 evsel->idx = idx;
230 evsel->tracking = !idx;
231 evsel->attr = *attr;
232 evsel->leader = evsel;
233 evsel->unit = "";
234 evsel->scale = 1.0;
235 evsel->evlist = NULL;
236 evsel->bpf_fd = -1;
237 INIT_LIST_HEAD(&evsel->node);
238 INIT_LIST_HEAD(&evsel->config_terms);
239 perf_evsel__object.init(evsel);
240 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
241 perf_evsel__calc_id_pos(evsel);
242 evsel->cmdline_group_boundary = false;
243 evsel->metric_expr = NULL;
244 evsel->metric_name = NULL;
245 evsel->metric_events = NULL;
246 evsel->collect_stat = false;
247 evsel->pmu_name = NULL;
248}
249
250struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
251{
252 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
253
254 if (evsel != NULL)
255 perf_evsel__init(evsel, attr, idx);
256
257 if (perf_evsel__is_bpf_output(evsel)) {
258 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
259 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
260 evsel->attr.sample_period = 1;
261 }
262
263 return evsel;
264}
265
266static bool perf_event_can_profile_kernel(void)
267{
268 return geteuid() == 0 || perf_event_paranoid() == -1;
269}
270
271struct perf_evsel *perf_evsel__new_cycles(bool precise)
272{
273 struct perf_event_attr attr = {
274 .type = PERF_TYPE_HARDWARE,
275 .config = PERF_COUNT_HW_CPU_CYCLES,
276 .exclude_kernel = !perf_event_can_profile_kernel(),
277 };
278 struct perf_evsel *evsel;
279
280 event_attr_init(&attr);
281
282 if (!precise)
283 goto new_event;
284 /*
285 * Unnamed union member, not supported as struct member named
286 * initializer in older compilers such as gcc 4.4.7
287 *
288 * Just for probing the precise_ip:
289 */
290 attr.sample_period = 1;
291
292 perf_event_attr__set_max_precise_ip(&attr);
293 /*
294 * Now let the usual logic to set up the perf_event_attr defaults
295 * to kick in when we return and before perf_evsel__open() is called.
296 */
297 attr.sample_period = 0;
298new_event:
299 evsel = perf_evsel__new(&attr);
300 if (evsel == NULL)
301 goto out;
302
303 /* use asprintf() because free(evsel) assumes name is allocated */
304 if (asprintf(&evsel->name, "cycles%s%s%.*s",
305 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
306 attr.exclude_kernel ? "u" : "",
307 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
308 goto error_free;
309out:
310 return evsel;
311error_free:
312 perf_evsel__delete(evsel);
313 evsel = NULL;
314 goto out;
315}
316
317/*
318 * Returns pointer with encoded error via <linux/err.h> interface.
319 */
320struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
321{
322 struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
323 int err = -ENOMEM;
324
325 if (evsel == NULL) {
326 goto out_err;
327 } else {
328 struct perf_event_attr attr = {
329 .type = PERF_TYPE_TRACEPOINT,
330 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
331 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
332 };
333
334 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
335 goto out_free;
336
337 evsel->tp_format = trace_event__tp_format(sys, name);
338 if (IS_ERR(evsel->tp_format)) {
339 err = PTR_ERR(evsel->tp_format);
340 goto out_free;
341 }
342
343 event_attr_init(&attr);
344 attr.config = evsel->tp_format->id;
345 attr.sample_period = 1;
346 perf_evsel__init(evsel, &attr, idx);
347 }
348
349 return evsel;
350
351out_free:
352 zfree(&evsel->name);
353 free(evsel);
354out_err:
355 return ERR_PTR(err);
356}
357
358const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
359 "cycles",
360 "instructions",
361 "cache-references",
362 "cache-misses",
363 "branches",
364 "branch-misses",
365 "bus-cycles",
366 "stalled-cycles-frontend",
367 "stalled-cycles-backend",
368 "ref-cycles",
369};
370
371static const char *__perf_evsel__hw_name(u64 config)
372{
373 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
374 return perf_evsel__hw_names[config];
375
376 return "unknown-hardware";
377}
378
379static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
380{
381 int colon = 0, r = 0;
382 struct perf_event_attr *attr = &evsel->attr;
383 bool exclude_guest_default = false;
384
385#define MOD_PRINT(context, mod) do { \
386 if (!attr->exclude_##context) { \
387 if (!colon) colon = ++r; \
388 r += scnprintf(bf + r, size - r, "%c", mod); \
389 } } while(0)
390
391 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
392 MOD_PRINT(kernel, 'k');
393 MOD_PRINT(user, 'u');
394 MOD_PRINT(hv, 'h');
395 exclude_guest_default = true;
396 }
397
398 if (attr->precise_ip) {
399 if (!colon)
400 colon = ++r;
401 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
402 exclude_guest_default = true;
403 }
404
405 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
406 MOD_PRINT(host, 'H');
407 MOD_PRINT(guest, 'G');
408 }
409#undef MOD_PRINT
410 if (colon)
411 bf[colon - 1] = ':';
412 return r;
413}
414
415static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
416{
417 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
418 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
419}
420
421const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
422 "cpu-clock",
423 "task-clock",
424 "page-faults",
425 "context-switches",
426 "cpu-migrations",
427 "minor-faults",
428 "major-faults",
429 "alignment-faults",
430 "emulation-faults",
431 "dummy",
432};
433
434static const char *__perf_evsel__sw_name(u64 config)
435{
436 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
437 return perf_evsel__sw_names[config];
438 return "unknown-software";
439}
440
441static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
442{
443 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
444 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
445}
446
447static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
448{
449 int r;
450
451 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
452
453 if (type & HW_BREAKPOINT_R)
454 r += scnprintf(bf + r, size - r, "r");
455
456 if (type & HW_BREAKPOINT_W)
457 r += scnprintf(bf + r, size - r, "w");
458
459 if (type & HW_BREAKPOINT_X)
460 r += scnprintf(bf + r, size - r, "x");
461
462 return r;
463}
464
465static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
466{
467 struct perf_event_attr *attr = &evsel->attr;
468 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
469 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
470}
471
472const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
473 [PERF_EVSEL__MAX_ALIASES] = {
474 { "L1-dcache", "l1-d", "l1d", "L1-data", },
475 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
476 { "LLC", "L2", },
477 { "dTLB", "d-tlb", "Data-TLB", },
478 { "iTLB", "i-tlb", "Instruction-TLB", },
479 { "branch", "branches", "bpu", "btb", "bpc", },
480 { "node", },
481};
482
483const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
484 [PERF_EVSEL__MAX_ALIASES] = {
485 { "load", "loads", "read", },
486 { "store", "stores", "write", },
487 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
488};
489
490const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
491 [PERF_EVSEL__MAX_ALIASES] = {
492 { "refs", "Reference", "ops", "access", },
493 { "misses", "miss", },
494};
495
496#define C(x) PERF_COUNT_HW_CACHE_##x
497#define CACHE_READ (1 << C(OP_READ))
498#define CACHE_WRITE (1 << C(OP_WRITE))
499#define CACHE_PREFETCH (1 << C(OP_PREFETCH))
500#define COP(x) (1 << x)
501
502/*
503 * cache operartion stat
504 * L1I : Read and prefetch only
505 * ITLB and BPU : Read-only
506 */
507static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
508 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
509 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
510 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
511 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
512 [C(ITLB)] = (CACHE_READ),
513 [C(BPU)] = (CACHE_READ),
514 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
515};
516
517bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
518{
519 if (perf_evsel__hw_cache_stat[type] & COP(op))
520 return true; /* valid */
521 else
522 return false; /* invalid */
523}
524
525int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
526 char *bf, size_t size)
527{
528 if (result) {
529 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
530 perf_evsel__hw_cache_op[op][0],
531 perf_evsel__hw_cache_result[result][0]);
532 }
533
534 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
535 perf_evsel__hw_cache_op[op][1]);
536}
537
538static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
539{
540 u8 op, result, type = (config >> 0) & 0xff;
541 const char *err = "unknown-ext-hardware-cache-type";
542
543 if (type >= PERF_COUNT_HW_CACHE_MAX)
544 goto out_err;
545
546 op = (config >> 8) & 0xff;
547 err = "unknown-ext-hardware-cache-op";
548 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
549 goto out_err;
550
551 result = (config >> 16) & 0xff;
552 err = "unknown-ext-hardware-cache-result";
553 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
554 goto out_err;
555
556 err = "invalid-cache";
557 if (!perf_evsel__is_cache_op_valid(type, op))
558 goto out_err;
559
560 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
561out_err:
562 return scnprintf(bf, size, "%s", err);
563}
564
565static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
566{
567 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
568 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
569}
570
571static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
572{
573 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
574 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
575}
576
577const char *perf_evsel__name(struct perf_evsel *evsel)
578{
579 char bf[128];
580
581 if (evsel->name)
582 return evsel->name;
583
584 switch (evsel->attr.type) {
585 case PERF_TYPE_RAW:
586 perf_evsel__raw_name(evsel, bf, sizeof(bf));
587 break;
588
589 case PERF_TYPE_HARDWARE:
590 perf_evsel__hw_name(evsel, bf, sizeof(bf));
591 break;
592
593 case PERF_TYPE_HW_CACHE:
594 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
595 break;
596
597 case PERF_TYPE_SOFTWARE:
598 perf_evsel__sw_name(evsel, bf, sizeof(bf));
599 break;
600
601 case PERF_TYPE_TRACEPOINT:
602 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
603 break;
604
605 case PERF_TYPE_BREAKPOINT:
606 perf_evsel__bp_name(evsel, bf, sizeof(bf));
607 break;
608
609 default:
610 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
611 evsel->attr.type);
612 break;
613 }
614
615 evsel->name = strdup(bf);
616
617 return evsel->name ?: "unknown";
618}
619
620const char *perf_evsel__group_name(struct perf_evsel *evsel)
621{
622 return evsel->group_name ?: "anon group";
623}
624
625/*
626 * Returns the group details for the specified leader,
627 * with following rules.
628 *
629 * For record -e '{cycles,instructions}'
630 * 'anon group { cycles:u, instructions:u }'
631 *
632 * For record -e 'cycles,instructions' and report --group
633 * 'cycles:u, instructions:u'
634 */
635int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
636{
637 int ret = 0;
638 struct perf_evsel *pos;
639 const char *group_name = perf_evsel__group_name(evsel);
640
641 if (!evsel->forced_leader)
642 ret = scnprintf(buf, size, "%s { ", group_name);
643
644 ret += scnprintf(buf + ret, size - ret, "%s",
645 perf_evsel__name(evsel));
646
647 for_each_group_member(pos, evsel)
648 ret += scnprintf(buf + ret, size - ret, ", %s",
649 perf_evsel__name(pos));
650
651 if (!evsel->forced_leader)
652 ret += scnprintf(buf + ret, size - ret, " }");
653
654 return ret;
655}
656
657static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
658 struct record_opts *opts,
659 struct callchain_param *param)
660{
661 bool function = perf_evsel__is_function_event(evsel);
662 struct perf_event_attr *attr = &evsel->attr;
663
664 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
665
666 attr->sample_max_stack = param->max_stack;
667
668 if (param->record_mode == CALLCHAIN_LBR) {
669 if (!opts->branch_stack) {
670 if (attr->exclude_user) {
671 pr_warning("LBR callstack option is only available "
672 "to get user callchain information. "
673 "Falling back to framepointers.\n");
674 } else {
675 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
676 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
677 PERF_SAMPLE_BRANCH_CALL_STACK |
678 PERF_SAMPLE_BRANCH_NO_CYCLES |
679 PERF_SAMPLE_BRANCH_NO_FLAGS;
680 }
681 } else
682 pr_warning("Cannot use LBR callstack with branch stack. "
683 "Falling back to framepointers.\n");
684 }
685
686 if (param->record_mode == CALLCHAIN_DWARF) {
687 if (!function) {
688 perf_evsel__set_sample_bit(evsel, REGS_USER);
689 perf_evsel__set_sample_bit(evsel, STACK_USER);
690 attr->sample_regs_user |= PERF_REGS_MASK;
691 attr->sample_stack_user = param->dump_size;
692 attr->exclude_callchain_user = 1;
693 } else {
694 pr_info("Cannot use DWARF unwind for function trace event,"
695 " falling back to framepointers.\n");
696 }
697 }
698
699 if (function) {
700 pr_info("Disabling user space callchains for function trace event.\n");
701 attr->exclude_callchain_user = 1;
702 }
703}
704
705void perf_evsel__config_callchain(struct perf_evsel *evsel,
706 struct record_opts *opts,
707 struct callchain_param *param)
708{
709 if (param->enabled)
710 return __perf_evsel__config_callchain(evsel, opts, param);
711}
712
713static void
714perf_evsel__reset_callgraph(struct perf_evsel *evsel,
715 struct callchain_param *param)
716{
717 struct perf_event_attr *attr = &evsel->attr;
718
719 perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
720 if (param->record_mode == CALLCHAIN_LBR) {
721 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
722 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
723 PERF_SAMPLE_BRANCH_CALL_STACK);
724 }
725 if (param->record_mode == CALLCHAIN_DWARF) {
726 perf_evsel__reset_sample_bit(evsel, REGS_USER);
727 perf_evsel__reset_sample_bit(evsel, STACK_USER);
728 }
729}
730
731static void apply_config_terms(struct perf_evsel *evsel,
732 struct record_opts *opts, bool track)
733{
734 struct perf_evsel_config_term *term;
735 struct list_head *config_terms = &evsel->config_terms;
736 struct perf_event_attr *attr = &evsel->attr;
737 /* callgraph default */
738 struct callchain_param param = {
739 .record_mode = callchain_param.record_mode,
740 };
741 u32 dump_size = 0;
742 int max_stack = 0;
743 const char *callgraph_buf = NULL;
744
745 list_for_each_entry(term, config_terms, list) {
746 switch (term->type) {
747 case PERF_EVSEL__CONFIG_TERM_PERIOD:
748 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
749 attr->sample_period = term->val.period;
750 attr->freq = 0;
751 perf_evsel__reset_sample_bit(evsel, PERIOD);
752 }
753 break;
754 case PERF_EVSEL__CONFIG_TERM_FREQ:
755 if (!(term->weak && opts->user_freq != UINT_MAX)) {
756 attr->sample_freq = term->val.freq;
757 attr->freq = 1;
758 perf_evsel__set_sample_bit(evsel, PERIOD);
759 }
760 break;
761 case PERF_EVSEL__CONFIG_TERM_TIME:
762 if (term->val.time)
763 perf_evsel__set_sample_bit(evsel, TIME);
764 else
765 perf_evsel__reset_sample_bit(evsel, TIME);
766 break;
767 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
768 callgraph_buf = term->val.callgraph;
769 break;
770 case PERF_EVSEL__CONFIG_TERM_BRANCH:
771 if (term->val.branch && strcmp(term->val.branch, "no")) {
772 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
773 parse_branch_str(term->val.branch,
774 &attr->branch_sample_type);
775 } else
776 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
777 break;
778 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
779 dump_size = term->val.stack_user;
780 break;
781 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
782 max_stack = term->val.max_stack;
783 break;
784 case PERF_EVSEL__CONFIG_TERM_INHERIT:
785 /*
786 * attr->inherit should has already been set by
787 * perf_evsel__config. If user explicitly set
788 * inherit using config terms, override global
789 * opt->no_inherit setting.
790 */
791 attr->inherit = term->val.inherit ? 1 : 0;
792 break;
793 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
794 attr->write_backward = term->val.overwrite ? 1 : 0;
795 break;
796 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
797 break;
798 default:
799 break;
800 }
801 }
802
803 /* User explicitly set per-event callgraph, clear the old setting and reset. */
804 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
805 bool sample_address = false;
806
807 if (max_stack) {
808 param.max_stack = max_stack;
809 if (callgraph_buf == NULL)
810 callgraph_buf = "fp";
811 }
812
813 /* parse callgraph parameters */
814 if (callgraph_buf != NULL) {
815 if (!strcmp(callgraph_buf, "no")) {
816 param.enabled = false;
817 param.record_mode = CALLCHAIN_NONE;
818 } else {
819 param.enabled = true;
820 if (parse_callchain_record(callgraph_buf, ¶m)) {
821 pr_err("per-event callgraph setting for %s failed. "
822 "Apply callgraph global setting for it\n",
823 evsel->name);
824 return;
825 }
826 if (param.record_mode == CALLCHAIN_DWARF)
827 sample_address = true;
828 }
829 }
830 if (dump_size > 0) {
831 dump_size = round_up(dump_size, sizeof(u64));
832 param.dump_size = dump_size;
833 }
834
835 /* If global callgraph set, clear it */
836 if (callchain_param.enabled)
837 perf_evsel__reset_callgraph(evsel, &callchain_param);
838
839 /* set perf-event callgraph */
840 if (param.enabled) {
841 if (sample_address) {
842 perf_evsel__set_sample_bit(evsel, ADDR);
843 perf_evsel__set_sample_bit(evsel, DATA_SRC);
844 evsel->attr.mmap_data = track;
845 }
846 perf_evsel__config_callchain(evsel, opts, ¶m);
847 }
848 }
849}
850
851/*
852 * The enable_on_exec/disabled value strategy:
853 *
854 * 1) For any type of traced program:
855 * - all independent events and group leaders are disabled
856 * - all group members are enabled
857 *
858 * Group members are ruled by group leaders. They need to
859 * be enabled, because the group scheduling relies on that.
860 *
861 * 2) For traced programs executed by perf:
862 * - all independent events and group leaders have
863 * enable_on_exec set
864 * - we don't specifically enable or disable any event during
865 * the record command
866 *
867 * Independent events and group leaders are initially disabled
868 * and get enabled by exec. Group members are ruled by group
869 * leaders as stated in 1).
870 *
871 * 3) For traced programs attached by perf (pid/tid):
872 * - we specifically enable or disable all events during
873 * the record command
874 *
875 * When attaching events to already running traced we
876 * enable/disable events specifically, as there's no
877 * initial traced exec call.
878 */
879void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
880 struct callchain_param *callchain)
881{
882 struct perf_evsel *leader = evsel->leader;
883 struct perf_event_attr *attr = &evsel->attr;
884 int track = evsel->tracking;
885 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
886
887 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
888 attr->inherit = !opts->no_inherit;
889 attr->write_backward = opts->overwrite ? 1 : 0;
890
891 perf_evsel__set_sample_bit(evsel, IP);
892 perf_evsel__set_sample_bit(evsel, TID);
893
894 if (evsel->sample_read) {
895 perf_evsel__set_sample_bit(evsel, READ);
896
897 /*
898 * We need ID even in case of single event, because
899 * PERF_SAMPLE_READ process ID specific data.
900 */
901 perf_evsel__set_sample_id(evsel, false);
902
903 /*
904 * Apply group format only if we belong to group
905 * with more than one members.
906 */
907 if (leader->nr_members > 1) {
908 attr->read_format |= PERF_FORMAT_GROUP;
909 attr->inherit = 0;
910 }
911 }
912
913 /*
914 * We default some events to have a default interval. But keep
915 * it a weak assumption overridable by the user.
916 */
917 if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
918 opts->user_interval != ULLONG_MAX)) {
919 if (opts->freq) {
920 perf_evsel__set_sample_bit(evsel, PERIOD);
921 attr->freq = 1;
922 attr->sample_freq = opts->freq;
923 } else {
924 attr->sample_period = opts->default_interval;
925 }
926 }
927
928 /*
929 * Disable sampling for all group members other
930 * than leader in case leader 'leads' the sampling.
931 */
932 if ((leader != evsel) && leader->sample_read) {
933 attr->freq = 0;
934 attr->sample_freq = 0;
935 attr->sample_period = 0;
936 attr->write_backward = 0;
937 attr->sample_id_all = 0;
938 }
939
940 if (opts->no_samples)
941 attr->sample_freq = 0;
942
943 if (opts->inherit_stat) {
944 evsel->attr.read_format |=
945 PERF_FORMAT_TOTAL_TIME_ENABLED |
946 PERF_FORMAT_TOTAL_TIME_RUNNING |
947 PERF_FORMAT_ID;
948 attr->inherit_stat = 1;
949 }
950
951 if (opts->sample_address) {
952 perf_evsel__set_sample_bit(evsel, ADDR);
953 attr->mmap_data = track;
954 }
955
956 /*
957 * We don't allow user space callchains for function trace
958 * event, due to issues with page faults while tracing page
959 * fault handler and its overall trickiness nature.
960 */
961 if (perf_evsel__is_function_event(evsel))
962 evsel->attr.exclude_callchain_user = 1;
963
964 if (callchain && callchain->enabled && !evsel->no_aux_samples)
965 perf_evsel__config_callchain(evsel, opts, callchain);
966
967 if (opts->sample_intr_regs) {
968 attr->sample_regs_intr = opts->sample_intr_regs;
969 perf_evsel__set_sample_bit(evsel, REGS_INTR);
970 }
971
972 if (opts->sample_user_regs) {
973 attr->sample_regs_user |= opts->sample_user_regs;
974 perf_evsel__set_sample_bit(evsel, REGS_USER);
975 }
976
977 if (target__has_cpu(&opts->target) || opts->sample_cpu)
978 perf_evsel__set_sample_bit(evsel, CPU);
979
980 /*
981 * When the user explicitly disabled time don't force it here.
982 */
983 if (opts->sample_time &&
984 (!perf_missing_features.sample_id_all &&
985 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
986 opts->sample_time_set)))
987 perf_evsel__set_sample_bit(evsel, TIME);
988
989 if (opts->raw_samples && !evsel->no_aux_samples) {
990 perf_evsel__set_sample_bit(evsel, TIME);
991 perf_evsel__set_sample_bit(evsel, RAW);
992 perf_evsel__set_sample_bit(evsel, CPU);
993 }
994
995 if (opts->sample_address)
996 perf_evsel__set_sample_bit(evsel, DATA_SRC);
997
998 if (opts->sample_phys_addr)
999 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1000
1001 if (opts->no_buffering) {
1002 attr->watermark = 0;
1003 attr->wakeup_events = 1;
1004 }
1005 if (opts->branch_stack && !evsel->no_aux_samples) {
1006 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1007 attr->branch_sample_type = opts->branch_stack;
1008 }
1009
1010 if (opts->sample_weight)
1011 perf_evsel__set_sample_bit(evsel, WEIGHT);
1012
1013 attr->task = track;
1014 attr->mmap = track;
1015 attr->mmap2 = track && !perf_missing_features.mmap2;
1016 attr->comm = track;
1017
1018 if (opts->record_namespaces)
1019 attr->namespaces = track;
1020
1021 if (opts->record_switch_events)
1022 attr->context_switch = track;
1023
1024 if (opts->sample_transaction)
1025 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1026
1027 if (opts->running_time) {
1028 evsel->attr.read_format |=
1029 PERF_FORMAT_TOTAL_TIME_ENABLED |
1030 PERF_FORMAT_TOTAL_TIME_RUNNING;
1031 }
1032
1033 /*
1034 * XXX see the function comment above
1035 *
1036 * Disabling only independent events or group leaders,
1037 * keeping group members enabled.
1038 */
1039 if (perf_evsel__is_group_leader(evsel))
1040 attr->disabled = 1;
1041
1042 /*
1043 * Setting enable_on_exec for independent events and
1044 * group leaders for traced executed by perf.
1045 */
1046 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1047 !opts->initial_delay)
1048 attr->enable_on_exec = 1;
1049
1050 if (evsel->immediate) {
1051 attr->disabled = 0;
1052 attr->enable_on_exec = 0;
1053 }
1054
1055 clockid = opts->clockid;
1056 if (opts->use_clockid) {
1057 attr->use_clockid = 1;
1058 attr->clockid = opts->clockid;
1059 }
1060
1061 if (evsel->precise_max)
1062 perf_event_attr__set_max_precise_ip(attr);
1063
1064 if (opts->all_user) {
1065 attr->exclude_kernel = 1;
1066 attr->exclude_user = 0;
1067 }
1068
1069 if (opts->all_kernel) {
1070 attr->exclude_kernel = 0;
1071 attr->exclude_user = 1;
1072 }
1073
1074 /*
1075 * Apply event specific term settings,
1076 * it overloads any global configuration.
1077 */
1078 apply_config_terms(evsel, opts, track);
1079
1080 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1081
1082 /* The --period option takes the precedence. */
1083 if (opts->period_set) {
1084 if (opts->period)
1085 perf_evsel__set_sample_bit(evsel, PERIOD);
1086 else
1087 perf_evsel__reset_sample_bit(evsel, PERIOD);
1088 }
1089}
1090
1091static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1092{
1093 if (evsel->system_wide)
1094 nthreads = 1;
1095
1096 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1097
1098 if (evsel->fd) {
1099 int cpu, thread;
1100 for (cpu = 0; cpu < ncpus; cpu++) {
1101 for (thread = 0; thread < nthreads; thread++) {
1102 FD(evsel, cpu, thread) = -1;
1103 }
1104 }
1105 }
1106
1107 return evsel->fd != NULL ? 0 : -ENOMEM;
1108}
1109
1110static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1111 int ioc, void *arg)
1112{
1113 int cpu, thread;
1114
1115 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1116 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1117 int fd = FD(evsel, cpu, thread),
1118 err = ioctl(fd, ioc, arg);
1119
1120 if (err)
1121 return err;
1122 }
1123 }
1124
1125 return 0;
1126}
1127
1128int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1129{
1130 return perf_evsel__run_ioctl(evsel,
1131 PERF_EVENT_IOC_SET_FILTER,
1132 (void *)filter);
1133}
1134
1135int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1136{
1137 char *new_filter = strdup(filter);
1138
1139 if (new_filter != NULL) {
1140 free(evsel->filter);
1141 evsel->filter = new_filter;
1142 return 0;
1143 }
1144
1145 return -1;
1146}
1147
1148static int perf_evsel__append_filter(struct perf_evsel *evsel,
1149 const char *fmt, const char *filter)
1150{
1151 char *new_filter;
1152
1153 if (evsel->filter == NULL)
1154 return perf_evsel__set_filter(evsel, filter);
1155
1156 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1157 free(evsel->filter);
1158 evsel->filter = new_filter;
1159 return 0;
1160 }
1161
1162 return -1;
1163}
1164
1165int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1166{
1167 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1168}
1169
1170int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1171{
1172 return perf_evsel__append_filter(evsel, "%s,%s", filter);
1173}
1174
1175int perf_evsel__enable(struct perf_evsel *evsel)
1176{
1177 return perf_evsel__run_ioctl(evsel,
1178 PERF_EVENT_IOC_ENABLE,
1179 0);
1180}
1181
1182int perf_evsel__disable(struct perf_evsel *evsel)
1183{
1184 return perf_evsel__run_ioctl(evsel,
1185 PERF_EVENT_IOC_DISABLE,
1186 0);
1187}
1188
1189int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1190{
1191 if (ncpus == 0 || nthreads == 0)
1192 return 0;
1193
1194 if (evsel->system_wide)
1195 nthreads = 1;
1196
1197 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1198 if (evsel->sample_id == NULL)
1199 return -ENOMEM;
1200
1201 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1202 if (evsel->id == NULL) {
1203 xyarray__delete(evsel->sample_id);
1204 evsel->sample_id = NULL;
1205 return -ENOMEM;
1206 }
1207
1208 return 0;
1209}
1210
1211static void perf_evsel__free_fd(struct perf_evsel *evsel)
1212{
1213 xyarray__delete(evsel->fd);
1214 evsel->fd = NULL;
1215}
1216
1217static void perf_evsel__free_id(struct perf_evsel *evsel)
1218{
1219 xyarray__delete(evsel->sample_id);
1220 evsel->sample_id = NULL;
1221 zfree(&evsel->id);
1222}
1223
1224static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1225{
1226 struct perf_evsel_config_term *term, *h;
1227
1228 list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1229 list_del(&term->list);
1230 free(term);
1231 }
1232}
1233
1234void perf_evsel__close_fd(struct perf_evsel *evsel)
1235{
1236 int cpu, thread;
1237
1238 for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1239 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1240 close(FD(evsel, cpu, thread));
1241 FD(evsel, cpu, thread) = -1;
1242 }
1243}
1244
1245void perf_evsel__exit(struct perf_evsel *evsel)
1246{
1247 assert(list_empty(&evsel->node));
1248 assert(evsel->evlist == NULL);
1249 perf_evsel__free_fd(evsel);
1250 perf_evsel__free_id(evsel);
1251 perf_evsel__free_config_terms(evsel);
1252 cgroup__put(evsel->cgrp);
1253 cpu_map__put(evsel->cpus);
1254 cpu_map__put(evsel->own_cpus);
1255 thread_map__put(evsel->threads);
1256 zfree(&evsel->group_name);
1257 zfree(&evsel->name);
1258 perf_evsel__object.fini(evsel);
1259}
1260
1261void perf_evsel__delete(struct perf_evsel *evsel)
1262{
1263 perf_evsel__exit(evsel);
1264 free(evsel);
1265}
1266
1267void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1268 struct perf_counts_values *count)
1269{
1270 struct perf_counts_values tmp;
1271
1272 if (!evsel->prev_raw_counts)
1273 return;
1274
1275 if (cpu == -1) {
1276 tmp = evsel->prev_raw_counts->aggr;
1277 evsel->prev_raw_counts->aggr = *count;
1278 } else {
1279 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1280 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1281 }
1282
1283 count->val = count->val - tmp.val;
1284 count->ena = count->ena - tmp.ena;
1285 count->run = count->run - tmp.run;
1286}
1287
1288void perf_counts_values__scale(struct perf_counts_values *count,
1289 bool scale, s8 *pscaled)
1290{
1291 s8 scaled = 0;
1292
1293 if (scale) {
1294 if (count->run == 0) {
1295 scaled = -1;
1296 count->val = 0;
1297 } else if (count->run < count->ena) {
1298 scaled = 1;
1299 count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1300 }
1301 } else
1302 count->ena = count->run = 0;
1303
1304 if (pscaled)
1305 *pscaled = scaled;
1306}
1307
1308static int perf_evsel__read_size(struct perf_evsel *evsel)
1309{
1310 u64 read_format = evsel->attr.read_format;
1311 int entry = sizeof(u64); /* value */
1312 int size = 0;
1313 int nr = 1;
1314
1315 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1316 size += sizeof(u64);
1317
1318 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1319 size += sizeof(u64);
1320
1321 if (read_format & PERF_FORMAT_ID)
1322 entry += sizeof(u64);
1323
1324 if (read_format & PERF_FORMAT_GROUP) {
1325 nr = evsel->nr_members;
1326 size += sizeof(u64);
1327 }
1328
1329 size += entry * nr;
1330 return size;
1331}
1332
1333int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1334 struct perf_counts_values *count)
1335{
1336 size_t size = perf_evsel__read_size(evsel);
1337
1338 memset(count, 0, sizeof(*count));
1339
1340 if (FD(evsel, cpu, thread) < 0)
1341 return -EINVAL;
1342
1343 if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1344 return -errno;
1345
1346 return 0;
1347}
1348
1349static int
1350perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1351{
1352 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1353
1354 return perf_evsel__read(evsel, cpu, thread, count);
1355}
1356
1357static void
1358perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1359 u64 val, u64 ena, u64 run)
1360{
1361 struct perf_counts_values *count;
1362
1363 count = perf_counts(counter->counts, cpu, thread);
1364
1365 count->val = val;
1366 count->ena = ena;
1367 count->run = run;
1368 count->loaded = true;
1369}
1370
1371static int
1372perf_evsel__process_group_data(struct perf_evsel *leader,
1373 int cpu, int thread, u64 *data)
1374{
1375 u64 read_format = leader->attr.read_format;
1376 struct sample_read_value *v;
1377 u64 nr, ena = 0, run = 0, i;
1378
1379 nr = *data++;
1380
1381 if (nr != (u64) leader->nr_members)
1382 return -EINVAL;
1383
1384 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1385 ena = *data++;
1386
1387 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1388 run = *data++;
1389
1390 v = (struct sample_read_value *) data;
1391
1392 perf_evsel__set_count(leader, cpu, thread,
1393 v[0].value, ena, run);
1394
1395 for (i = 1; i < nr; i++) {
1396 struct perf_evsel *counter;
1397
1398 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1399 if (!counter)
1400 return -EINVAL;
1401
1402 perf_evsel__set_count(counter, cpu, thread,
1403 v[i].value, ena, run);
1404 }
1405
1406 return 0;
1407}
1408
1409static int
1410perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1411{
1412 struct perf_stat_evsel *ps = leader->stats;
1413 u64 read_format = leader->attr.read_format;
1414 int size = perf_evsel__read_size(leader);
1415 u64 *data = ps->group_data;
1416
1417 if (!(read_format & PERF_FORMAT_ID))
1418 return -EINVAL;
1419
1420 if (!perf_evsel__is_group_leader(leader))
1421 return -EINVAL;
1422
1423 if (!data) {
1424 data = zalloc(size);
1425 if (!data)
1426 return -ENOMEM;
1427
1428 ps->group_data = data;
1429 }
1430
1431 if (FD(leader, cpu, thread) < 0)
1432 return -EINVAL;
1433
1434 if (readn(FD(leader, cpu, thread), data, size) <= 0)
1435 return -errno;
1436
1437 return perf_evsel__process_group_data(leader, cpu, thread, data);
1438}
1439
1440int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1441{
1442 u64 read_format = evsel->attr.read_format;
1443
1444 if (read_format & PERF_FORMAT_GROUP)
1445 return perf_evsel__read_group(evsel, cpu, thread);
1446 else
1447 return perf_evsel__read_one(evsel, cpu, thread);
1448}
1449
1450int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1451 int cpu, int thread, bool scale)
1452{
1453 struct perf_counts_values count;
1454 size_t nv = scale ? 3 : 1;
1455
1456 if (FD(evsel, cpu, thread) < 0)
1457 return -EINVAL;
1458
1459 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1460 return -ENOMEM;
1461
1462 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1463 return -errno;
1464
1465 perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1466 perf_counts_values__scale(&count, scale, NULL);
1467 *perf_counts(evsel->counts, cpu, thread) = count;
1468 return 0;
1469}
1470
1471static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1472{
1473 struct perf_evsel *leader = evsel->leader;
1474 int fd;
1475
1476 if (perf_evsel__is_group_leader(evsel))
1477 return -1;
1478
1479 /*
1480 * Leader must be already processed/open,
1481 * if not it's a bug.
1482 */
1483 BUG_ON(!leader->fd);
1484
1485 fd = FD(leader, cpu, thread);
1486 BUG_ON(fd == -1);
1487
1488 return fd;
1489}
1490
1491struct bit_names {
1492 int bit;
1493 const char *name;
1494};
1495
1496static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1497{
1498 bool first_bit = true;
1499 int i = 0;
1500
1501 do {
1502 if (value & bits[i].bit) {
1503 buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1504 first_bit = false;
1505 }
1506 } while (bits[++i].name != NULL);
1507}
1508
1509static void __p_sample_type(char *buf, size_t size, u64 value)
1510{
1511#define bit_name(n) { PERF_SAMPLE_##n, #n }
1512 struct bit_names bits[] = {
1513 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1514 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1515 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1516 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1517 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1518 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1519 { .name = NULL, }
1520 };
1521#undef bit_name
1522 __p_bits(buf, size, value, bits);
1523}
1524
1525static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1526{
1527#define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1528 struct bit_names bits[] = {
1529 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1530 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1531 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1532 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1533 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1534 { .name = NULL, }
1535 };
1536#undef bit_name
1537 __p_bits(buf, size, value, bits);
1538}
1539
1540static void __p_read_format(char *buf, size_t size, u64 value)
1541{
1542#define bit_name(n) { PERF_FORMAT_##n, #n }
1543 struct bit_names bits[] = {
1544 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1545 bit_name(ID), bit_name(GROUP),
1546 { .name = NULL, }
1547 };
1548#undef bit_name
1549 __p_bits(buf, size, value, bits);
1550}
1551
1552#define BUF_SIZE 1024
1553
1554#define p_hex(val) snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1555#define p_unsigned(val) snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1556#define p_signed(val) snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1557#define p_sample_type(val) __p_sample_type(buf, BUF_SIZE, val)
1558#define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1559#define p_read_format(val) __p_read_format(buf, BUF_SIZE, val)
1560
1561#define PRINT_ATTRn(_n, _f, _p) \
1562do { \
1563 if (attr->_f) { \
1564 _p(attr->_f); \
1565 ret += attr__fprintf(fp, _n, buf, priv);\
1566 } \
1567} while (0)
1568
1569#define PRINT_ATTRf(_f, _p) PRINT_ATTRn(#_f, _f, _p)
1570
1571int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1572 attr__fprintf_f attr__fprintf, void *priv)
1573{
1574 char buf[BUF_SIZE];
1575 int ret = 0;
1576
1577 PRINT_ATTRf(type, p_unsigned);
1578 PRINT_ATTRf(size, p_unsigned);
1579 PRINT_ATTRf(config, p_hex);
1580 PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1581 PRINT_ATTRf(sample_type, p_sample_type);
1582 PRINT_ATTRf(read_format, p_read_format);
1583
1584 PRINT_ATTRf(disabled, p_unsigned);
1585 PRINT_ATTRf(inherit, p_unsigned);
1586 PRINT_ATTRf(pinned, p_unsigned);
1587 PRINT_ATTRf(exclusive, p_unsigned);
1588 PRINT_ATTRf(exclude_user, p_unsigned);
1589 PRINT_ATTRf(exclude_kernel, p_unsigned);
1590 PRINT_ATTRf(exclude_hv, p_unsigned);
1591 PRINT_ATTRf(exclude_idle, p_unsigned);
1592 PRINT_ATTRf(mmap, p_unsigned);
1593 PRINT_ATTRf(comm, p_unsigned);
1594 PRINT_ATTRf(freq, p_unsigned);
1595 PRINT_ATTRf(inherit_stat, p_unsigned);
1596 PRINT_ATTRf(enable_on_exec, p_unsigned);
1597 PRINT_ATTRf(task, p_unsigned);
1598 PRINT_ATTRf(watermark, p_unsigned);
1599 PRINT_ATTRf(precise_ip, p_unsigned);
1600 PRINT_ATTRf(mmap_data, p_unsigned);
1601 PRINT_ATTRf(sample_id_all, p_unsigned);
1602 PRINT_ATTRf(exclude_host, p_unsigned);
1603 PRINT_ATTRf(exclude_guest, p_unsigned);
1604 PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1605 PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1606 PRINT_ATTRf(mmap2, p_unsigned);
1607 PRINT_ATTRf(comm_exec, p_unsigned);
1608 PRINT_ATTRf(use_clockid, p_unsigned);
1609 PRINT_ATTRf(context_switch, p_unsigned);
1610 PRINT_ATTRf(write_backward, p_unsigned);
1611 PRINT_ATTRf(namespaces, p_unsigned);
1612
1613 PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1614 PRINT_ATTRf(bp_type, p_unsigned);
1615 PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1616 PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1617 PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1618 PRINT_ATTRf(sample_regs_user, p_hex);
1619 PRINT_ATTRf(sample_stack_user, p_unsigned);
1620 PRINT_ATTRf(clockid, p_signed);
1621 PRINT_ATTRf(sample_regs_intr, p_hex);
1622 PRINT_ATTRf(aux_watermark, p_unsigned);
1623 PRINT_ATTRf(sample_max_stack, p_unsigned);
1624
1625 return ret;
1626}
1627
1628static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1629 void *priv __maybe_unused)
1630{
1631 return fprintf(fp, " %-32s %s\n", name, val);
1632}
1633
1634static void perf_evsel__remove_fd(struct perf_evsel *pos,
1635 int nr_cpus, int nr_threads,
1636 int thread_idx)
1637{
1638 for (int cpu = 0; cpu < nr_cpus; cpu++)
1639 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1640 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1641}
1642
1643static int update_fds(struct perf_evsel *evsel,
1644 int nr_cpus, int cpu_idx,
1645 int nr_threads, int thread_idx)
1646{
1647 struct perf_evsel *pos;
1648
1649 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1650 return -EINVAL;
1651
1652 evlist__for_each_entry(evsel->evlist, pos) {
1653 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1654
1655 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1656
1657 /*
1658 * Since fds for next evsel has not been created,
1659 * there is no need to iterate whole event list.
1660 */
1661 if (pos == evsel)
1662 break;
1663 }
1664 return 0;
1665}
1666
1667static bool ignore_missing_thread(struct perf_evsel *evsel,
1668 int nr_cpus, int cpu,
1669 struct thread_map *threads,
1670 int thread, int err)
1671{
1672 pid_t ignore_pid = thread_map__pid(threads, thread);
1673
1674 if (!evsel->ignore_missing_thread)
1675 return false;
1676
1677 /* The system wide setup does not work with threads. */
1678 if (evsel->system_wide)
1679 return false;
1680
1681 /* The -ESRCH is perf event syscall errno for pid's not found. */
1682 if (err != -ESRCH)
1683 return false;
1684
1685 /* If there's only one thread, let it fail. */
1686 if (threads->nr == 1)
1687 return false;
1688
1689 /*
1690 * We should remove fd for missing_thread first
1691 * because thread_map__remove() will decrease threads->nr.
1692 */
1693 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1694 return false;
1695
1696 if (thread_map__remove(threads, thread))
1697 return false;
1698
1699 pr_warning("WARNING: Ignored open failure for pid %d\n",
1700 ignore_pid);
1701 return true;
1702}
1703
1704int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1705 struct thread_map *threads)
1706{
1707 int cpu, thread, nthreads;
1708 unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1709 int pid = -1, err;
1710 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1711
1712 if (perf_missing_features.write_backward && evsel->attr.write_backward)
1713 return -EINVAL;
1714
1715 if (cpus == NULL) {
1716 static struct cpu_map *empty_cpu_map;
1717
1718 if (empty_cpu_map == NULL) {
1719 empty_cpu_map = cpu_map__dummy_new();
1720 if (empty_cpu_map == NULL)
1721 return -ENOMEM;
1722 }
1723
1724 cpus = empty_cpu_map;
1725 }
1726
1727 if (threads == NULL) {
1728 static struct thread_map *empty_thread_map;
1729
1730 if (empty_thread_map == NULL) {
1731 empty_thread_map = thread_map__new_by_tid(-1);
1732 if (empty_thread_map == NULL)
1733 return -ENOMEM;
1734 }
1735
1736 threads = empty_thread_map;
1737 }
1738
1739 if (evsel->system_wide)
1740 nthreads = 1;
1741 else
1742 nthreads = threads->nr;
1743
1744 if (evsel->fd == NULL &&
1745 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1746 return -ENOMEM;
1747
1748 if (evsel->cgrp) {
1749 flags |= PERF_FLAG_PID_CGROUP;
1750 pid = evsel->cgrp->fd;
1751 }
1752
1753fallback_missing_features:
1754 if (perf_missing_features.clockid_wrong)
1755 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1756 if (perf_missing_features.clockid) {
1757 evsel->attr.use_clockid = 0;
1758 evsel->attr.clockid = 0;
1759 }
1760 if (perf_missing_features.cloexec)
1761 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1762 if (perf_missing_features.mmap2)
1763 evsel->attr.mmap2 = 0;
1764 if (perf_missing_features.exclude_guest)
1765 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1766 if (perf_missing_features.lbr_flags)
1767 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1768 PERF_SAMPLE_BRANCH_NO_CYCLES);
1769 if (perf_missing_features.group_read && evsel->attr.inherit)
1770 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1771retry_sample_id:
1772 if (perf_missing_features.sample_id_all)
1773 evsel->attr.sample_id_all = 0;
1774
1775 if (verbose >= 2) {
1776 fprintf(stderr, "%.60s\n", graph_dotted_line);
1777 fprintf(stderr, "perf_event_attr:\n");
1778 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1779 fprintf(stderr, "%.60s\n", graph_dotted_line);
1780 }
1781
1782 for (cpu = 0; cpu < cpus->nr; cpu++) {
1783
1784 for (thread = 0; thread < nthreads; thread++) {
1785 int fd, group_fd;
1786
1787 if (!evsel->cgrp && !evsel->system_wide)
1788 pid = thread_map__pid(threads, thread);
1789
1790 group_fd = get_group_fd(evsel, cpu, thread);
1791retry_open:
1792 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
1793 pid, cpus->map[cpu], group_fd, flags);
1794
1795 test_attr__ready();
1796
1797 fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1798 group_fd, flags);
1799
1800 FD(evsel, cpu, thread) = fd;
1801
1802 if (fd < 0) {
1803 err = -errno;
1804
1805 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1806 /*
1807 * We just removed 1 thread, so take a step
1808 * back on thread index and lower the upper
1809 * nthreads limit.
1810 */
1811 nthreads--;
1812 thread--;
1813
1814 /* ... and pretend like nothing have happened. */
1815 err = 0;
1816 continue;
1817 }
1818
1819 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1820 err);
1821 goto try_fallback;
1822 }
1823
1824 pr_debug2(" = %d\n", fd);
1825
1826 if (evsel->bpf_fd >= 0) {
1827 int evt_fd = fd;
1828 int bpf_fd = evsel->bpf_fd;
1829
1830 err = ioctl(evt_fd,
1831 PERF_EVENT_IOC_SET_BPF,
1832 bpf_fd);
1833 if (err && errno != EEXIST) {
1834 pr_err("failed to attach bpf fd %d: %s\n",
1835 bpf_fd, strerror(errno));
1836 err = -EINVAL;
1837 goto out_close;
1838 }
1839 }
1840
1841 set_rlimit = NO_CHANGE;
1842
1843 /*
1844 * If we succeeded but had to kill clockid, fail and
1845 * have perf_evsel__open_strerror() print us a nice
1846 * error.
1847 */
1848 if (perf_missing_features.clockid ||
1849 perf_missing_features.clockid_wrong) {
1850 err = -EINVAL;
1851 goto out_close;
1852 }
1853 }
1854 }
1855
1856 return 0;
1857
1858try_fallback:
1859 /*
1860 * perf stat needs between 5 and 22 fds per CPU. When we run out
1861 * of them try to increase the limits.
1862 */
1863 if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1864 struct rlimit l;
1865 int old_errno = errno;
1866
1867 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1868 if (set_rlimit == NO_CHANGE)
1869 l.rlim_cur = l.rlim_max;
1870 else {
1871 l.rlim_cur = l.rlim_max + 1000;
1872 l.rlim_max = l.rlim_cur;
1873 }
1874 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1875 set_rlimit++;
1876 errno = old_errno;
1877 goto retry_open;
1878 }
1879 }
1880 errno = old_errno;
1881 }
1882
1883 if (err != -EINVAL || cpu > 0 || thread > 0)
1884 goto out_close;
1885
1886 /*
1887 * Must probe features in the order they were added to the
1888 * perf_event_attr interface.
1889 */
1890 if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1891 perf_missing_features.write_backward = true;
1892 pr_debug2("switching off write_backward\n");
1893 goto out_close;
1894 } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1895 perf_missing_features.clockid_wrong = true;
1896 pr_debug2("switching off clockid\n");
1897 goto fallback_missing_features;
1898 } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1899 perf_missing_features.clockid = true;
1900 pr_debug2("switching off use_clockid\n");
1901 goto fallback_missing_features;
1902 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1903 perf_missing_features.cloexec = true;
1904 pr_debug2("switching off cloexec flag\n");
1905 goto fallback_missing_features;
1906 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1907 perf_missing_features.mmap2 = true;
1908 pr_debug2("switching off mmap2\n");
1909 goto fallback_missing_features;
1910 } else if (!perf_missing_features.exclude_guest &&
1911 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1912 perf_missing_features.exclude_guest = true;
1913 pr_debug2("switching off exclude_guest, exclude_host\n");
1914 goto fallback_missing_features;
1915 } else if (!perf_missing_features.sample_id_all) {
1916 perf_missing_features.sample_id_all = true;
1917 pr_debug2("switching off sample_id_all\n");
1918 goto retry_sample_id;
1919 } else if (!perf_missing_features.lbr_flags &&
1920 (evsel->attr.branch_sample_type &
1921 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1922 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1923 perf_missing_features.lbr_flags = true;
1924 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1925 goto fallback_missing_features;
1926 } else if (!perf_missing_features.group_read &&
1927 evsel->attr.inherit &&
1928 (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
1929 perf_evsel__is_group_leader(evsel)) {
1930 perf_missing_features.group_read = true;
1931 pr_debug2("switching off group read\n");
1932 goto fallback_missing_features;
1933 }
1934out_close:
1935 if (err)
1936 threads->err_thread = thread;
1937
1938 do {
1939 while (--thread >= 0) {
1940 close(FD(evsel, cpu, thread));
1941 FD(evsel, cpu, thread) = -1;
1942 }
1943 thread = nthreads;
1944 } while (--cpu >= 0);
1945 return err;
1946}
1947
1948void perf_evsel__close(struct perf_evsel *evsel)
1949{
1950 if (evsel->fd == NULL)
1951 return;
1952
1953 perf_evsel__close_fd(evsel);
1954 perf_evsel__free_fd(evsel);
1955}
1956
1957int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1958 struct cpu_map *cpus)
1959{
1960 return perf_evsel__open(evsel, cpus, NULL);
1961}
1962
1963int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1964 struct thread_map *threads)
1965{
1966 return perf_evsel__open(evsel, NULL, threads);
1967}
1968
1969static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1970 const union perf_event *event,
1971 struct perf_sample *sample)
1972{
1973 u64 type = evsel->attr.sample_type;
1974 const u64 *array = event->sample.array;
1975 bool swapped = evsel->needs_swap;
1976 union u64_swap u;
1977
1978 array += ((event->header.size -
1979 sizeof(event->header)) / sizeof(u64)) - 1;
1980
1981 if (type & PERF_SAMPLE_IDENTIFIER) {
1982 sample->id = *array;
1983 array--;
1984 }
1985
1986 if (type & PERF_SAMPLE_CPU) {
1987 u.val64 = *array;
1988 if (swapped) {
1989 /* undo swap of u64, then swap on individual u32s */
1990 u.val64 = bswap_64(u.val64);
1991 u.val32[0] = bswap_32(u.val32[0]);
1992 }
1993
1994 sample->cpu = u.val32[0];
1995 array--;
1996 }
1997
1998 if (type & PERF_SAMPLE_STREAM_ID) {
1999 sample->stream_id = *array;
2000 array--;
2001 }
2002
2003 if (type & PERF_SAMPLE_ID) {
2004 sample->id = *array;
2005 array--;
2006 }
2007
2008 if (type & PERF_SAMPLE_TIME) {
2009 sample->time = *array;
2010 array--;
2011 }
2012
2013 if (type & PERF_SAMPLE_TID) {
2014 u.val64 = *array;
2015 if (swapped) {
2016 /* undo swap of u64, then swap on individual u32s */
2017 u.val64 = bswap_64(u.val64);
2018 u.val32[0] = bswap_32(u.val32[0]);
2019 u.val32[1] = bswap_32(u.val32[1]);
2020 }
2021
2022 sample->pid = u.val32[0];
2023 sample->tid = u.val32[1];
2024 array--;
2025 }
2026
2027 return 0;
2028}
2029
2030static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2031 u64 size)
2032{
2033 return size > max_size || offset + size > endp;
2034}
2035
2036#define OVERFLOW_CHECK(offset, size, max_size) \
2037 do { \
2038 if (overflow(endp, (max_size), (offset), (size))) \
2039 return -EFAULT; \
2040 } while (0)
2041
2042#define OVERFLOW_CHECK_u64(offset) \
2043 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2044
2045static int
2046perf_event__check_size(union perf_event *event, unsigned int sample_size)
2047{
2048 /*
2049 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2050 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2051 * check the format does not go past the end of the event.
2052 */
2053 if (sample_size + sizeof(event->header) > event->header.size)
2054 return -EFAULT;
2055
2056 return 0;
2057}
2058
2059int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2060 struct perf_sample *data)
2061{
2062 u64 type = evsel->attr.sample_type;
2063 bool swapped = evsel->needs_swap;
2064 const u64 *array;
2065 u16 max_size = event->header.size;
2066 const void *endp = (void *)event + max_size;
2067 u64 sz;
2068
2069 /*
2070 * used for cross-endian analysis. See git commit 65014ab3
2071 * for why this goofiness is needed.
2072 */
2073 union u64_swap u;
2074
2075 memset(data, 0, sizeof(*data));
2076 data->cpu = data->pid = data->tid = -1;
2077 data->stream_id = data->id = data->time = -1ULL;
2078 data->period = evsel->attr.sample_period;
2079 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2080 data->misc = event->header.misc;
2081 data->id = -1ULL;
2082 data->data_src = PERF_MEM_DATA_SRC_NONE;
2083
2084 if (event->header.type != PERF_RECORD_SAMPLE) {
2085 if (!evsel->attr.sample_id_all)
2086 return 0;
2087 return perf_evsel__parse_id_sample(evsel, event, data);
2088 }
2089
2090 array = event->sample.array;
2091
2092 if (perf_event__check_size(event, evsel->sample_size))
2093 return -EFAULT;
2094
2095 if (type & PERF_SAMPLE_IDENTIFIER) {
2096 data->id = *array;
2097 array++;
2098 }
2099
2100 if (type & PERF_SAMPLE_IP) {
2101 data->ip = *array;
2102 array++;
2103 }
2104
2105 if (type & PERF_SAMPLE_TID) {
2106 u.val64 = *array;
2107 if (swapped) {
2108 /* undo swap of u64, then swap on individual u32s */
2109 u.val64 = bswap_64(u.val64);
2110 u.val32[0] = bswap_32(u.val32[0]);
2111 u.val32[1] = bswap_32(u.val32[1]);
2112 }
2113
2114 data->pid = u.val32[0];
2115 data->tid = u.val32[1];
2116 array++;
2117 }
2118
2119 if (type & PERF_SAMPLE_TIME) {
2120 data->time = *array;
2121 array++;
2122 }
2123
2124 if (type & PERF_SAMPLE_ADDR) {
2125 data->addr = *array;
2126 array++;
2127 }
2128
2129 if (type & PERF_SAMPLE_ID) {
2130 data->id = *array;
2131 array++;
2132 }
2133
2134 if (type & PERF_SAMPLE_STREAM_ID) {
2135 data->stream_id = *array;
2136 array++;
2137 }
2138
2139 if (type & PERF_SAMPLE_CPU) {
2140
2141 u.val64 = *array;
2142 if (swapped) {
2143 /* undo swap of u64, then swap on individual u32s */
2144 u.val64 = bswap_64(u.val64);
2145 u.val32[0] = bswap_32(u.val32[0]);
2146 }
2147
2148 data->cpu = u.val32[0];
2149 array++;
2150 }
2151
2152 if (type & PERF_SAMPLE_PERIOD) {
2153 data->period = *array;
2154 array++;
2155 }
2156
2157 if (type & PERF_SAMPLE_READ) {
2158 u64 read_format = evsel->attr.read_format;
2159
2160 OVERFLOW_CHECK_u64(array);
2161 if (read_format & PERF_FORMAT_GROUP)
2162 data->read.group.nr = *array;
2163 else
2164 data->read.one.value = *array;
2165
2166 array++;
2167
2168 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2169 OVERFLOW_CHECK_u64(array);
2170 data->read.time_enabled = *array;
2171 array++;
2172 }
2173
2174 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2175 OVERFLOW_CHECK_u64(array);
2176 data->read.time_running = *array;
2177 array++;
2178 }
2179
2180 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2181 if (read_format & PERF_FORMAT_GROUP) {
2182 const u64 max_group_nr = UINT64_MAX /
2183 sizeof(struct sample_read_value);
2184
2185 if (data->read.group.nr > max_group_nr)
2186 return -EFAULT;
2187 sz = data->read.group.nr *
2188 sizeof(struct sample_read_value);
2189 OVERFLOW_CHECK(array, sz, max_size);
2190 data->read.group.values =
2191 (struct sample_read_value *)array;
2192 array = (void *)array + sz;
2193 } else {
2194 OVERFLOW_CHECK_u64(array);
2195 data->read.one.id = *array;
2196 array++;
2197 }
2198 }
2199
2200 if (type & PERF_SAMPLE_CALLCHAIN) {
2201 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2202
2203 OVERFLOW_CHECK_u64(array);
2204 data->callchain = (struct ip_callchain *)array++;
2205 if (data->callchain->nr > max_callchain_nr)
2206 return -EFAULT;
2207 sz = data->callchain->nr * sizeof(u64);
2208 OVERFLOW_CHECK(array, sz, max_size);
2209 array = (void *)array + sz;
2210 }
2211
2212 if (type & PERF_SAMPLE_RAW) {
2213 OVERFLOW_CHECK_u64(array);
2214 u.val64 = *array;
2215
2216 /*
2217 * Undo swap of u64, then swap on individual u32s,
2218 * get the size of the raw area and undo all of the
2219 * swap. The pevent interface handles endianity by
2220 * itself.
2221 */
2222 if (swapped) {
2223 u.val64 = bswap_64(u.val64);
2224 u.val32[0] = bswap_32(u.val32[0]);
2225 u.val32[1] = bswap_32(u.val32[1]);
2226 }
2227 data->raw_size = u.val32[0];
2228
2229 /*
2230 * The raw data is aligned on 64bits including the
2231 * u32 size, so it's safe to use mem_bswap_64.
2232 */
2233 if (swapped)
2234 mem_bswap_64((void *) array, data->raw_size);
2235
2236 array = (void *)array + sizeof(u32);
2237
2238 OVERFLOW_CHECK(array, data->raw_size, max_size);
2239 data->raw_data = (void *)array;
2240 array = (void *)array + data->raw_size;
2241 }
2242
2243 if (type & PERF_SAMPLE_BRANCH_STACK) {
2244 const u64 max_branch_nr = UINT64_MAX /
2245 sizeof(struct branch_entry);
2246
2247 OVERFLOW_CHECK_u64(array);
2248 data->branch_stack = (struct branch_stack *)array++;
2249
2250 if (data->branch_stack->nr > max_branch_nr)
2251 return -EFAULT;
2252 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2253 OVERFLOW_CHECK(array, sz, max_size);
2254 array = (void *)array + sz;
2255 }
2256
2257 if (type & PERF_SAMPLE_REGS_USER) {
2258 OVERFLOW_CHECK_u64(array);
2259 data->user_regs.abi = *array;
2260 array++;
2261
2262 if (data->user_regs.abi) {
2263 u64 mask = evsel->attr.sample_regs_user;
2264
2265 sz = hweight_long(mask) * sizeof(u64);
2266 OVERFLOW_CHECK(array, sz, max_size);
2267 data->user_regs.mask = mask;
2268 data->user_regs.regs = (u64 *)array;
2269 array = (void *)array + sz;
2270 }
2271 }
2272
2273 if (type & PERF_SAMPLE_STACK_USER) {
2274 OVERFLOW_CHECK_u64(array);
2275 sz = *array++;
2276
2277 data->user_stack.offset = ((char *)(array - 1)
2278 - (char *) event);
2279
2280 if (!sz) {
2281 data->user_stack.size = 0;
2282 } else {
2283 OVERFLOW_CHECK(array, sz, max_size);
2284 data->user_stack.data = (char *)array;
2285 array = (void *)array + sz;
2286 OVERFLOW_CHECK_u64(array);
2287 data->user_stack.size = *array++;
2288 if (WARN_ONCE(data->user_stack.size > sz,
2289 "user stack dump failure\n"))
2290 return -EFAULT;
2291 }
2292 }
2293
2294 if (type & PERF_SAMPLE_WEIGHT) {
2295 OVERFLOW_CHECK_u64(array);
2296 data->weight = *array;
2297 array++;
2298 }
2299
2300 if (type & PERF_SAMPLE_DATA_SRC) {
2301 OVERFLOW_CHECK_u64(array);
2302 data->data_src = *array;
2303 array++;
2304 }
2305
2306 if (type & PERF_SAMPLE_TRANSACTION) {
2307 OVERFLOW_CHECK_u64(array);
2308 data->transaction = *array;
2309 array++;
2310 }
2311
2312 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2313 if (type & PERF_SAMPLE_REGS_INTR) {
2314 OVERFLOW_CHECK_u64(array);
2315 data->intr_regs.abi = *array;
2316 array++;
2317
2318 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2319 u64 mask = evsel->attr.sample_regs_intr;
2320
2321 sz = hweight_long(mask) * sizeof(u64);
2322 OVERFLOW_CHECK(array, sz, max_size);
2323 data->intr_regs.mask = mask;
2324 data->intr_regs.regs = (u64 *)array;
2325 array = (void *)array + sz;
2326 }
2327 }
2328
2329 data->phys_addr = 0;
2330 if (type & PERF_SAMPLE_PHYS_ADDR) {
2331 data->phys_addr = *array;
2332 array++;
2333 }
2334
2335 return 0;
2336}
2337
2338int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2339 union perf_event *event,
2340 u64 *timestamp)
2341{
2342 u64 type = evsel->attr.sample_type;
2343 const u64 *array;
2344
2345 if (!(type & PERF_SAMPLE_TIME))
2346 return -1;
2347
2348 if (event->header.type != PERF_RECORD_SAMPLE) {
2349 struct perf_sample data = {
2350 .time = -1ULL,
2351 };
2352
2353 if (!evsel->attr.sample_id_all)
2354 return -1;
2355 if (perf_evsel__parse_id_sample(evsel, event, &data))
2356 return -1;
2357
2358 *timestamp = data.time;
2359 return 0;
2360 }
2361
2362 array = event->sample.array;
2363
2364 if (perf_event__check_size(event, evsel->sample_size))
2365 return -EFAULT;
2366
2367 if (type & PERF_SAMPLE_IDENTIFIER)
2368 array++;
2369
2370 if (type & PERF_SAMPLE_IP)
2371 array++;
2372
2373 if (type & PERF_SAMPLE_TID)
2374 array++;
2375
2376 if (type & PERF_SAMPLE_TIME)
2377 *timestamp = *array;
2378
2379 return 0;
2380}
2381
2382size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2383 u64 read_format)
2384{
2385 size_t sz, result = sizeof(struct sample_event);
2386
2387 if (type & PERF_SAMPLE_IDENTIFIER)
2388 result += sizeof(u64);
2389
2390 if (type & PERF_SAMPLE_IP)
2391 result += sizeof(u64);
2392
2393 if (type & PERF_SAMPLE_TID)
2394 result += sizeof(u64);
2395
2396 if (type & PERF_SAMPLE_TIME)
2397 result += sizeof(u64);
2398
2399 if (type & PERF_SAMPLE_ADDR)
2400 result += sizeof(u64);
2401
2402 if (type & PERF_SAMPLE_ID)
2403 result += sizeof(u64);
2404
2405 if (type & PERF_SAMPLE_STREAM_ID)
2406 result += sizeof(u64);
2407
2408 if (type & PERF_SAMPLE_CPU)
2409 result += sizeof(u64);
2410
2411 if (type & PERF_SAMPLE_PERIOD)
2412 result += sizeof(u64);
2413
2414 if (type & PERF_SAMPLE_READ) {
2415 result += sizeof(u64);
2416 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2417 result += sizeof(u64);
2418 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2419 result += sizeof(u64);
2420 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2421 if (read_format & PERF_FORMAT_GROUP) {
2422 sz = sample->read.group.nr *
2423 sizeof(struct sample_read_value);
2424 result += sz;
2425 } else {
2426 result += sizeof(u64);
2427 }
2428 }
2429
2430 if (type & PERF_SAMPLE_CALLCHAIN) {
2431 sz = (sample->callchain->nr + 1) * sizeof(u64);
2432 result += sz;
2433 }
2434
2435 if (type & PERF_SAMPLE_RAW) {
2436 result += sizeof(u32);
2437 result += sample->raw_size;
2438 }
2439
2440 if (type & PERF_SAMPLE_BRANCH_STACK) {
2441 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2442 sz += sizeof(u64);
2443 result += sz;
2444 }
2445
2446 if (type & PERF_SAMPLE_REGS_USER) {
2447 if (sample->user_regs.abi) {
2448 result += sizeof(u64);
2449 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2450 result += sz;
2451 } else {
2452 result += sizeof(u64);
2453 }
2454 }
2455
2456 if (type & PERF_SAMPLE_STACK_USER) {
2457 sz = sample->user_stack.size;
2458 result += sizeof(u64);
2459 if (sz) {
2460 result += sz;
2461 result += sizeof(u64);
2462 }
2463 }
2464
2465 if (type & PERF_SAMPLE_WEIGHT)
2466 result += sizeof(u64);
2467
2468 if (type & PERF_SAMPLE_DATA_SRC)
2469 result += sizeof(u64);
2470
2471 if (type & PERF_SAMPLE_TRANSACTION)
2472 result += sizeof(u64);
2473
2474 if (type & PERF_SAMPLE_REGS_INTR) {
2475 if (sample->intr_regs.abi) {
2476 result += sizeof(u64);
2477 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2478 result += sz;
2479 } else {
2480 result += sizeof(u64);
2481 }
2482 }
2483
2484 if (type & PERF_SAMPLE_PHYS_ADDR)
2485 result += sizeof(u64);
2486
2487 return result;
2488}
2489
2490int perf_event__synthesize_sample(union perf_event *event, u64 type,
2491 u64 read_format,
2492 const struct perf_sample *sample)
2493{
2494 u64 *array;
2495 size_t sz;
2496 /*
2497 * used for cross-endian analysis. See git commit 65014ab3
2498 * for why this goofiness is needed.
2499 */
2500 union u64_swap u;
2501
2502 array = event->sample.array;
2503
2504 if (type & PERF_SAMPLE_IDENTIFIER) {
2505 *array = sample->id;
2506 array++;
2507 }
2508
2509 if (type & PERF_SAMPLE_IP) {
2510 *array = sample->ip;
2511 array++;
2512 }
2513
2514 if (type & PERF_SAMPLE_TID) {
2515 u.val32[0] = sample->pid;
2516 u.val32[1] = sample->tid;
2517 *array = u.val64;
2518 array++;
2519 }
2520
2521 if (type & PERF_SAMPLE_TIME) {
2522 *array = sample->time;
2523 array++;
2524 }
2525
2526 if (type & PERF_SAMPLE_ADDR) {
2527 *array = sample->addr;
2528 array++;
2529 }
2530
2531 if (type & PERF_SAMPLE_ID) {
2532 *array = sample->id;
2533 array++;
2534 }
2535
2536 if (type & PERF_SAMPLE_STREAM_ID) {
2537 *array = sample->stream_id;
2538 array++;
2539 }
2540
2541 if (type & PERF_SAMPLE_CPU) {
2542 u.val32[0] = sample->cpu;
2543 u.val32[1] = 0;
2544 *array = u.val64;
2545 array++;
2546 }
2547
2548 if (type & PERF_SAMPLE_PERIOD) {
2549 *array = sample->period;
2550 array++;
2551 }
2552
2553 if (type & PERF_SAMPLE_READ) {
2554 if (read_format & PERF_FORMAT_GROUP)
2555 *array = sample->read.group.nr;
2556 else
2557 *array = sample->read.one.value;
2558 array++;
2559
2560 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2561 *array = sample->read.time_enabled;
2562 array++;
2563 }
2564
2565 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2566 *array = sample->read.time_running;
2567 array++;
2568 }
2569
2570 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2571 if (read_format & PERF_FORMAT_GROUP) {
2572 sz = sample->read.group.nr *
2573 sizeof(struct sample_read_value);
2574 memcpy(array, sample->read.group.values, sz);
2575 array = (void *)array + sz;
2576 } else {
2577 *array = sample->read.one.id;
2578 array++;
2579 }
2580 }
2581
2582 if (type & PERF_SAMPLE_CALLCHAIN) {
2583 sz = (sample->callchain->nr + 1) * sizeof(u64);
2584 memcpy(array, sample->callchain, sz);
2585 array = (void *)array + sz;
2586 }
2587
2588 if (type & PERF_SAMPLE_RAW) {
2589 u.val32[0] = sample->raw_size;
2590 *array = u.val64;
2591 array = (void *)array + sizeof(u32);
2592
2593 memcpy(array, sample->raw_data, sample->raw_size);
2594 array = (void *)array + sample->raw_size;
2595 }
2596
2597 if (type & PERF_SAMPLE_BRANCH_STACK) {
2598 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2599 sz += sizeof(u64);
2600 memcpy(array, sample->branch_stack, sz);
2601 array = (void *)array + sz;
2602 }
2603
2604 if (type & PERF_SAMPLE_REGS_USER) {
2605 if (sample->user_regs.abi) {
2606 *array++ = sample->user_regs.abi;
2607 sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2608 memcpy(array, sample->user_regs.regs, sz);
2609 array = (void *)array + sz;
2610 } else {
2611 *array++ = 0;
2612 }
2613 }
2614
2615 if (type & PERF_SAMPLE_STACK_USER) {
2616 sz = sample->user_stack.size;
2617 *array++ = sz;
2618 if (sz) {
2619 memcpy(array, sample->user_stack.data, sz);
2620 array = (void *)array + sz;
2621 *array++ = sz;
2622 }
2623 }
2624
2625 if (type & PERF_SAMPLE_WEIGHT) {
2626 *array = sample->weight;
2627 array++;
2628 }
2629
2630 if (type & PERF_SAMPLE_DATA_SRC) {
2631 *array = sample->data_src;
2632 array++;
2633 }
2634
2635 if (type & PERF_SAMPLE_TRANSACTION) {
2636 *array = sample->transaction;
2637 array++;
2638 }
2639
2640 if (type & PERF_SAMPLE_REGS_INTR) {
2641 if (sample->intr_regs.abi) {
2642 *array++ = sample->intr_regs.abi;
2643 sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2644 memcpy(array, sample->intr_regs.regs, sz);
2645 array = (void *)array + sz;
2646 } else {
2647 *array++ = 0;
2648 }
2649 }
2650
2651 if (type & PERF_SAMPLE_PHYS_ADDR) {
2652 *array = sample->phys_addr;
2653 array++;
2654 }
2655
2656 return 0;
2657}
2658
2659struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2660{
2661 return pevent_find_field(evsel->tp_format, name);
2662}
2663
2664void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2665 const char *name)
2666{
2667 struct format_field *field = perf_evsel__field(evsel, name);
2668 int offset;
2669
2670 if (!field)
2671 return NULL;
2672
2673 offset = field->offset;
2674
2675 if (field->flags & FIELD_IS_DYNAMIC) {
2676 offset = *(int *)(sample->raw_data + field->offset);
2677 offset &= 0xffff;
2678 }
2679
2680 return sample->raw_data + offset;
2681}
2682
2683u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2684 bool needs_swap)
2685{
2686 u64 value;
2687 void *ptr = sample->raw_data + field->offset;
2688
2689 switch (field->size) {
2690 case 1:
2691 return *(u8 *)ptr;
2692 case 2:
2693 value = *(u16 *)ptr;
2694 break;
2695 case 4:
2696 value = *(u32 *)ptr;
2697 break;
2698 case 8:
2699 memcpy(&value, ptr, sizeof(u64));
2700 break;
2701 default:
2702 return 0;
2703 }
2704
2705 if (!needs_swap)
2706 return value;
2707
2708 switch (field->size) {
2709 case 2:
2710 return bswap_16(value);
2711 case 4:
2712 return bswap_32(value);
2713 case 8:
2714 return bswap_64(value);
2715 default:
2716 return 0;
2717 }
2718
2719 return 0;
2720}
2721
2722u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2723 const char *name)
2724{
2725 struct format_field *field = perf_evsel__field(evsel, name);
2726
2727 if (!field)
2728 return 0;
2729
2730 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2731}
2732
2733bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2734 char *msg, size_t msgsize)
2735{
2736 int paranoid;
2737
2738 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2739 evsel->attr.type == PERF_TYPE_HARDWARE &&
2740 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2741 /*
2742 * If it's cycles then fall back to hrtimer based
2743 * cpu-clock-tick sw counter, which is always available even if
2744 * no PMU support.
2745 *
2746 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2747 * b0a873e).
2748 */
2749 scnprintf(msg, msgsize, "%s",
2750"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2751
2752 evsel->attr.type = PERF_TYPE_SOFTWARE;
2753 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2754
2755 zfree(&evsel->name);
2756 return true;
2757 } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2758 (paranoid = perf_event_paranoid()) > 1) {
2759 const char *name = perf_evsel__name(evsel);
2760 char *new_name;
2761 const char *sep = ":";
2762
2763 /* Is there already the separator in the name. */
2764 if (strchr(name, '/') ||
2765 strchr(name, ':'))
2766 sep = "";
2767
2768 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2769 return false;
2770
2771 if (evsel->name)
2772 free(evsel->name);
2773 evsel->name = new_name;
2774 scnprintf(msg, msgsize,
2775"kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2776 evsel->attr.exclude_kernel = 1;
2777
2778 return true;
2779 }
2780
2781 return false;
2782}
2783
2784static bool find_process(const char *name)
2785{
2786 size_t len = strlen(name);
2787 DIR *dir;
2788 struct dirent *d;
2789 int ret = -1;
2790
2791 dir = opendir(procfs__mountpoint());
2792 if (!dir)
2793 return false;
2794
2795 /* Walk through the directory. */
2796 while (ret && (d = readdir(dir)) != NULL) {
2797 char path[PATH_MAX];
2798 char *data;
2799 size_t size;
2800
2801 if ((d->d_type != DT_DIR) ||
2802 !strcmp(".", d->d_name) ||
2803 !strcmp("..", d->d_name))
2804 continue;
2805
2806 scnprintf(path, sizeof(path), "%s/%s/comm",
2807 procfs__mountpoint(), d->d_name);
2808
2809 if (filename__read_str(path, &data, &size))
2810 continue;
2811
2812 ret = strncmp(name, data, len);
2813 free(data);
2814 }
2815
2816 closedir(dir);
2817 return ret ? false : true;
2818}
2819
2820int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2821 int err, char *msg, size_t size)
2822{
2823 char sbuf[STRERR_BUFSIZE];
2824 int printed = 0;
2825
2826 switch (err) {
2827 case EPERM:
2828 case EACCES:
2829 if (err == EPERM)
2830 printed = scnprintf(msg, size,
2831 "No permission to enable %s event.\n\n",
2832 perf_evsel__name(evsel));
2833
2834 return scnprintf(msg + printed, size - printed,
2835 "You may not have permission to collect %sstats.\n\n"
2836 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2837 "which controls use of the performance events system by\n"
2838 "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2839 "The current value is %d:\n\n"
2840 " -1: Allow use of (almost) all events by all users\n"
2841 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2842 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2843 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2844 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2845 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2846 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2847 " kernel.perf_event_paranoid = -1\n" ,
2848 target->system_wide ? "system-wide " : "",
2849 perf_event_paranoid());
2850 case ENOENT:
2851 return scnprintf(msg, size, "The %s event is not supported.",
2852 perf_evsel__name(evsel));
2853 case EMFILE:
2854 return scnprintf(msg, size, "%s",
2855 "Too many events are opened.\n"
2856 "Probably the maximum number of open file descriptors has been reached.\n"
2857 "Hint: Try again after reducing the number of events.\n"
2858 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2859 case ENOMEM:
2860 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2861 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2862 return scnprintf(msg, size,
2863 "Not enough memory to setup event with callchain.\n"
2864 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2865 "Hint: Current value: %d", sysctl_perf_event_max_stack);
2866 break;
2867 case ENODEV:
2868 if (target->cpu_list)
2869 return scnprintf(msg, size, "%s",
2870 "No such device - did you specify an out-of-range profile CPU?");
2871 break;
2872 case EOPNOTSUPP:
2873 if (evsel->attr.sample_period != 0)
2874 return scnprintf(msg, size,
2875 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2876 perf_evsel__name(evsel));
2877 if (evsel->attr.precise_ip)
2878 return scnprintf(msg, size, "%s",
2879 "\'precise\' request may not be supported. Try removing 'p' modifier.");
2880#if defined(__i386__) || defined(__x86_64__)
2881 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2882 return scnprintf(msg, size, "%s",
2883 "No hardware sampling interrupt available.\n");
2884#endif
2885 break;
2886 case EBUSY:
2887 if (find_process("oprofiled"))
2888 return scnprintf(msg, size,
2889 "The PMU counters are busy/taken by another profiler.\n"
2890 "We found oprofile daemon running, please stop it and try again.");
2891 break;
2892 case EINVAL:
2893 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2894 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2895 if (perf_missing_features.clockid)
2896 return scnprintf(msg, size, "clockid feature not supported.");
2897 if (perf_missing_features.clockid_wrong)
2898 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2899 break;
2900 default:
2901 break;
2902 }
2903
2904 return scnprintf(msg, size,
2905 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2906 "/bin/dmesg | grep -i perf may provide additional information.\n",
2907 err, str_error_r(err, sbuf, sizeof(sbuf)),
2908 perf_evsel__name(evsel));
2909}
2910
2911struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2912{
2913 if (evsel && evsel->evlist)
2914 return evsel->evlist->env;
2915 return NULL;
2916}
1// 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 <linux/hw_breakpoint.h>
16#include <linux/perf_event.h>
17#include <linux/compiler.h>
18#include <linux/err.h>
19#include <linux/zalloc.h>
20#include <sys/ioctl.h>
21#include <sys/resource.h>
22#include <sys/types.h>
23#include <dirent.h>
24#include <stdlib.h>
25#include <perf/evsel.h>
26#include "asm/bug.h"
27#include "bpf_counter.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 "util/hashmap.h"
49#include "pmu-hybrid.h"
50#include "off_cpu.h"
51#include "../perf-sys.h"
52#include "util/parse-branch-options.h"
53#include <internal/xyarray.h>
54#include <internal/lib.h>
55#include <internal/threadmap.h>
56
57#include <linux/ctype.h>
58
59#ifdef HAVE_LIBTRACEEVENT
60#include <traceevent/event-parse.h>
61#endif
62
63struct perf_missing_features perf_missing_features;
64
65static clockid_t clockid;
66
67static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
68 NULL,
69 "duration_time",
70 "user_time",
71 "system_time",
72};
73
74const char *perf_tool_event__to_str(enum perf_tool_event ev)
75{
76 if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
77 return perf_tool_event__tool_names[ev];
78
79 return NULL;
80}
81
82enum perf_tool_event perf_tool_event__from_str(const char *str)
83{
84 int i;
85
86 perf_tool_event__for_each_event(i) {
87 if (!strcmp(str, perf_tool_event__tool_names[i]))
88 return i;
89 }
90 return PERF_TOOL_NONE;
91}
92
93
94static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
95{
96 return 0;
97}
98
99void __weak test_attr__ready(void) { }
100
101static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
102{
103}
104
105static struct {
106 size_t size;
107 int (*init)(struct evsel *evsel);
108 void (*fini)(struct evsel *evsel);
109} perf_evsel__object = {
110 .size = sizeof(struct evsel),
111 .init = evsel__no_extra_init,
112 .fini = evsel__no_extra_fini,
113};
114
115int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
116 void (*fini)(struct evsel *evsel))
117{
118
119 if (object_size == 0)
120 goto set_methods;
121
122 if (perf_evsel__object.size > object_size)
123 return -EINVAL;
124
125 perf_evsel__object.size = object_size;
126
127set_methods:
128 if (init != NULL)
129 perf_evsel__object.init = init;
130
131 if (fini != NULL)
132 perf_evsel__object.fini = fini;
133
134 return 0;
135}
136
137#define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
138
139int __evsel__sample_size(u64 sample_type)
140{
141 u64 mask = sample_type & PERF_SAMPLE_MASK;
142 int size = 0;
143 int i;
144
145 for (i = 0; i < 64; i++) {
146 if (mask & (1ULL << i))
147 size++;
148 }
149
150 size *= sizeof(u64);
151
152 return size;
153}
154
155/**
156 * __perf_evsel__calc_id_pos - calculate id_pos.
157 * @sample_type: sample type
158 *
159 * This function returns the position of the event id (PERF_SAMPLE_ID or
160 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
161 * perf_record_sample.
162 */
163static int __perf_evsel__calc_id_pos(u64 sample_type)
164{
165 int idx = 0;
166
167 if (sample_type & PERF_SAMPLE_IDENTIFIER)
168 return 0;
169
170 if (!(sample_type & PERF_SAMPLE_ID))
171 return -1;
172
173 if (sample_type & PERF_SAMPLE_IP)
174 idx += 1;
175
176 if (sample_type & PERF_SAMPLE_TID)
177 idx += 1;
178
179 if (sample_type & PERF_SAMPLE_TIME)
180 idx += 1;
181
182 if (sample_type & PERF_SAMPLE_ADDR)
183 idx += 1;
184
185 return idx;
186}
187
188/**
189 * __perf_evsel__calc_is_pos - calculate is_pos.
190 * @sample_type: sample type
191 *
192 * This function returns the position (counting backwards) of the event id
193 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
194 * sample_id_all is used there is an id sample appended to non-sample events.
195 */
196static int __perf_evsel__calc_is_pos(u64 sample_type)
197{
198 int idx = 1;
199
200 if (sample_type & PERF_SAMPLE_IDENTIFIER)
201 return 1;
202
203 if (!(sample_type & PERF_SAMPLE_ID))
204 return -1;
205
206 if (sample_type & PERF_SAMPLE_CPU)
207 idx += 1;
208
209 if (sample_type & PERF_SAMPLE_STREAM_ID)
210 idx += 1;
211
212 return idx;
213}
214
215void evsel__calc_id_pos(struct evsel *evsel)
216{
217 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
218 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
219}
220
221void __evsel__set_sample_bit(struct evsel *evsel,
222 enum perf_event_sample_format bit)
223{
224 if (!(evsel->core.attr.sample_type & bit)) {
225 evsel->core.attr.sample_type |= bit;
226 evsel->sample_size += sizeof(u64);
227 evsel__calc_id_pos(evsel);
228 }
229}
230
231void __evsel__reset_sample_bit(struct evsel *evsel,
232 enum perf_event_sample_format bit)
233{
234 if (evsel->core.attr.sample_type & bit) {
235 evsel->core.attr.sample_type &= ~bit;
236 evsel->sample_size -= sizeof(u64);
237 evsel__calc_id_pos(evsel);
238 }
239}
240
241void evsel__set_sample_id(struct evsel *evsel,
242 bool can_sample_identifier)
243{
244 if (can_sample_identifier) {
245 evsel__reset_sample_bit(evsel, ID);
246 evsel__set_sample_bit(evsel, IDENTIFIER);
247 } else {
248 evsel__set_sample_bit(evsel, ID);
249 }
250 evsel->core.attr.read_format |= PERF_FORMAT_ID;
251}
252
253/**
254 * evsel__is_function_event - Return whether given evsel is a function
255 * trace event
256 *
257 * @evsel - evsel selector to be tested
258 *
259 * Return %true if event is function trace event
260 */
261bool evsel__is_function_event(struct evsel *evsel)
262{
263#define FUNCTION_EVENT "ftrace:function"
264
265 return evsel->name &&
266 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
267
268#undef FUNCTION_EVENT
269}
270
271void evsel__init(struct evsel *evsel,
272 struct perf_event_attr *attr, int idx)
273{
274 perf_evsel__init(&evsel->core, attr, idx);
275 evsel->tracking = !idx;
276 evsel->unit = strdup("");
277 evsel->scale = 1.0;
278 evsel->max_events = ULONG_MAX;
279 evsel->evlist = NULL;
280 evsel->bpf_obj = NULL;
281 evsel->bpf_fd = -1;
282 INIT_LIST_HEAD(&evsel->config_terms);
283 INIT_LIST_HEAD(&evsel->bpf_counter_list);
284 perf_evsel__object.init(evsel);
285 evsel->sample_size = __evsel__sample_size(attr->sample_type);
286 evsel__calc_id_pos(evsel);
287 evsel->cmdline_group_boundary = false;
288 evsel->metric_expr = NULL;
289 evsel->metric_name = NULL;
290 evsel->metric_events = NULL;
291 evsel->per_pkg_mask = NULL;
292 evsel->collect_stat = false;
293 evsel->pmu_name = NULL;
294}
295
296struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
297{
298 struct evsel *evsel = zalloc(perf_evsel__object.size);
299
300 if (!evsel)
301 return NULL;
302 evsel__init(evsel, attr, idx);
303
304 if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
305 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
306 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
307 evsel->core.attr.sample_period = 1;
308 }
309
310 if (evsel__is_clock(evsel)) {
311 free((char *)evsel->unit);
312 evsel->unit = strdup("msec");
313 evsel->scale = 1e-6;
314 }
315
316 return evsel;
317}
318
319static bool perf_event_can_profile_kernel(void)
320{
321 return perf_event_paranoid_check(1);
322}
323
324struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
325{
326 struct perf_event_attr attr = {
327 .type = type,
328 .config = config,
329 .exclude_kernel = !perf_event_can_profile_kernel(),
330 };
331 struct evsel *evsel;
332
333 event_attr_init(&attr);
334
335 /*
336 * Now let the usual logic to set up the perf_event_attr defaults
337 * to kick in when we return and before perf_evsel__open() is called.
338 */
339 evsel = evsel__new(&attr);
340 if (evsel == NULL)
341 goto out;
342
343 arch_evsel__fixup_new_cycles(&evsel->core.attr);
344
345 evsel->precise_max = true;
346
347 /* use asprintf() because free(evsel) assumes name is allocated */
348 if (asprintf(&evsel->name, "cycles%s%s%.*s",
349 (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
350 attr.exclude_kernel ? "u" : "",
351 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
352 goto error_free;
353out:
354 return evsel;
355error_free:
356 evsel__delete(evsel);
357 evsel = NULL;
358 goto out;
359}
360
361int copy_config_terms(struct list_head *dst, struct list_head *src)
362{
363 struct evsel_config_term *pos, *tmp;
364
365 list_for_each_entry(pos, src, list) {
366 tmp = malloc(sizeof(*tmp));
367 if (tmp == NULL)
368 return -ENOMEM;
369
370 *tmp = *pos;
371 if (tmp->free_str) {
372 tmp->val.str = strdup(pos->val.str);
373 if (tmp->val.str == NULL) {
374 free(tmp);
375 return -ENOMEM;
376 }
377 }
378 list_add_tail(&tmp->list, dst);
379 }
380 return 0;
381}
382
383static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
384{
385 return copy_config_terms(&dst->config_terms, &src->config_terms);
386}
387
388/**
389 * evsel__clone - create a new evsel copied from @orig
390 * @orig: original evsel
391 *
392 * The assumption is that @orig is not configured nor opened yet.
393 * So we only care about the attributes that can be set while it's parsed.
394 */
395struct evsel *evsel__clone(struct evsel *orig)
396{
397 struct evsel *evsel;
398
399 BUG_ON(orig->core.fd);
400 BUG_ON(orig->counts);
401 BUG_ON(orig->priv);
402 BUG_ON(orig->per_pkg_mask);
403
404 /* cannot handle BPF objects for now */
405 if (orig->bpf_obj)
406 return NULL;
407
408 evsel = evsel__new(&orig->core.attr);
409 if (evsel == NULL)
410 return NULL;
411
412 evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
413 evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
414 evsel->core.threads = perf_thread_map__get(orig->core.threads);
415 evsel->core.nr_members = orig->core.nr_members;
416 evsel->core.system_wide = orig->core.system_wide;
417 evsel->core.requires_cpu = orig->core.requires_cpu;
418
419 if (orig->name) {
420 evsel->name = strdup(orig->name);
421 if (evsel->name == NULL)
422 goto out_err;
423 }
424 if (orig->group_name) {
425 evsel->group_name = strdup(orig->group_name);
426 if (evsel->group_name == NULL)
427 goto out_err;
428 }
429 if (orig->pmu_name) {
430 evsel->pmu_name = strdup(orig->pmu_name);
431 if (evsel->pmu_name == NULL)
432 goto out_err;
433 }
434 if (orig->filter) {
435 evsel->filter = strdup(orig->filter);
436 if (evsel->filter == NULL)
437 goto out_err;
438 }
439 if (orig->metric_id) {
440 evsel->metric_id = strdup(orig->metric_id);
441 if (evsel->metric_id == NULL)
442 goto out_err;
443 }
444 evsel->cgrp = cgroup__get(orig->cgrp);
445#ifdef HAVE_LIBTRACEEVENT
446 evsel->tp_format = orig->tp_format;
447#endif
448 evsel->handler = orig->handler;
449 evsel->core.leader = orig->core.leader;
450
451 evsel->max_events = orig->max_events;
452 evsel->tool_event = orig->tool_event;
453 free((char *)evsel->unit);
454 evsel->unit = strdup(orig->unit);
455 if (evsel->unit == NULL)
456 goto out_err;
457
458 evsel->scale = orig->scale;
459 evsel->snapshot = orig->snapshot;
460 evsel->per_pkg = orig->per_pkg;
461 evsel->percore = orig->percore;
462 evsel->precise_max = orig->precise_max;
463 evsel->use_uncore_alias = orig->use_uncore_alias;
464 evsel->is_libpfm_event = orig->is_libpfm_event;
465
466 evsel->exclude_GH = orig->exclude_GH;
467 evsel->sample_read = orig->sample_read;
468 evsel->auto_merge_stats = orig->auto_merge_stats;
469 evsel->collect_stat = orig->collect_stat;
470 evsel->weak_group = orig->weak_group;
471 evsel->use_config_name = orig->use_config_name;
472 evsel->pmu = orig->pmu;
473
474 if (evsel__copy_config_terms(evsel, orig) < 0)
475 goto out_err;
476
477 return evsel;
478
479out_err:
480 evsel__delete(evsel);
481 return NULL;
482}
483
484/*
485 * Returns pointer with encoded error via <linux/err.h> interface.
486 */
487#ifdef HAVE_LIBTRACEEVENT
488struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
489{
490 struct evsel *evsel = zalloc(perf_evsel__object.size);
491 int err = -ENOMEM;
492
493 if (evsel == NULL) {
494 goto out_err;
495 } else {
496 struct perf_event_attr attr = {
497 .type = PERF_TYPE_TRACEPOINT,
498 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
499 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
500 };
501
502 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
503 goto out_free;
504
505 evsel->tp_format = trace_event__tp_format(sys, name);
506 if (IS_ERR(evsel->tp_format)) {
507 err = PTR_ERR(evsel->tp_format);
508 goto out_free;
509 }
510
511 event_attr_init(&attr);
512 attr.config = evsel->tp_format->id;
513 attr.sample_period = 1;
514 evsel__init(evsel, &attr, idx);
515 }
516
517 return evsel;
518
519out_free:
520 zfree(&evsel->name);
521 free(evsel);
522out_err:
523 return ERR_PTR(err);
524}
525#endif
526
527const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
528 "cycles",
529 "instructions",
530 "cache-references",
531 "cache-misses",
532 "branches",
533 "branch-misses",
534 "bus-cycles",
535 "stalled-cycles-frontend",
536 "stalled-cycles-backend",
537 "ref-cycles",
538};
539
540char *evsel__bpf_counter_events;
541
542bool evsel__match_bpf_counter_events(const char *name)
543{
544 int name_len;
545 bool match;
546 char *ptr;
547
548 if (!evsel__bpf_counter_events)
549 return false;
550
551 ptr = strstr(evsel__bpf_counter_events, name);
552 name_len = strlen(name);
553
554 /* check name matches a full token in evsel__bpf_counter_events */
555 match = (ptr != NULL) &&
556 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
557 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
558
559 return match;
560}
561
562static const char *__evsel__hw_name(u64 config)
563{
564 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
565 return evsel__hw_names[config];
566
567 return "unknown-hardware";
568}
569
570static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
571{
572 int colon = 0, r = 0;
573 struct perf_event_attr *attr = &evsel->core.attr;
574 bool exclude_guest_default = false;
575
576#define MOD_PRINT(context, mod) do { \
577 if (!attr->exclude_##context) { \
578 if (!colon) colon = ++r; \
579 r += scnprintf(bf + r, size - r, "%c", mod); \
580 } } while(0)
581
582 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
583 MOD_PRINT(kernel, 'k');
584 MOD_PRINT(user, 'u');
585 MOD_PRINT(hv, 'h');
586 exclude_guest_default = true;
587 }
588
589 if (attr->precise_ip) {
590 if (!colon)
591 colon = ++r;
592 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
593 exclude_guest_default = true;
594 }
595
596 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
597 MOD_PRINT(host, 'H');
598 MOD_PRINT(guest, 'G');
599 }
600#undef MOD_PRINT
601 if (colon)
602 bf[colon - 1] = ':';
603 return r;
604}
605
606int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
607{
608 return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
609}
610
611static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
612{
613 int r = arch_evsel__hw_name(evsel, bf, size);
614 return r + evsel__add_modifiers(evsel, bf + r, size - r);
615}
616
617const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
618 "cpu-clock",
619 "task-clock",
620 "page-faults",
621 "context-switches",
622 "cpu-migrations",
623 "minor-faults",
624 "major-faults",
625 "alignment-faults",
626 "emulation-faults",
627 "dummy",
628};
629
630static const char *__evsel__sw_name(u64 config)
631{
632 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
633 return evsel__sw_names[config];
634 return "unknown-software";
635}
636
637static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
638{
639 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
640 return r + evsel__add_modifiers(evsel, bf + r, size - r);
641}
642
643static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
644{
645 return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
646}
647
648static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
649{
650 int r;
651
652 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
653
654 if (type & HW_BREAKPOINT_R)
655 r += scnprintf(bf + r, size - r, "r");
656
657 if (type & HW_BREAKPOINT_W)
658 r += scnprintf(bf + r, size - r, "w");
659
660 if (type & HW_BREAKPOINT_X)
661 r += scnprintf(bf + r, size - r, "x");
662
663 return r;
664}
665
666static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
667{
668 struct perf_event_attr *attr = &evsel->core.attr;
669 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
670 return r + evsel__add_modifiers(evsel, bf + r, size - r);
671}
672
673const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
674 { "L1-dcache", "l1-d", "l1d", "L1-data", },
675 { "L1-icache", "l1-i", "l1i", "L1-instruction", },
676 { "LLC", "L2", },
677 { "dTLB", "d-tlb", "Data-TLB", },
678 { "iTLB", "i-tlb", "Instruction-TLB", },
679 { "branch", "branches", "bpu", "btb", "bpc", },
680 { "node", },
681};
682
683const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
684 { "load", "loads", "read", },
685 { "store", "stores", "write", },
686 { "prefetch", "prefetches", "speculative-read", "speculative-load", },
687};
688
689const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
690 { "refs", "Reference", "ops", "access", },
691 { "misses", "miss", },
692};
693
694#define C(x) PERF_COUNT_HW_CACHE_##x
695#define CACHE_READ (1 << C(OP_READ))
696#define CACHE_WRITE (1 << C(OP_WRITE))
697#define CACHE_PREFETCH (1 << C(OP_PREFETCH))
698#define COP(x) (1 << x)
699
700/*
701 * cache operation stat
702 * L1I : Read and prefetch only
703 * ITLB and BPU : Read-only
704 */
705static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
706 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
707 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH),
708 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
709 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
710 [C(ITLB)] = (CACHE_READ),
711 [C(BPU)] = (CACHE_READ),
712 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
713};
714
715bool evsel__is_cache_op_valid(u8 type, u8 op)
716{
717 if (evsel__hw_cache_stat[type] & COP(op))
718 return true; /* valid */
719 else
720 return false; /* invalid */
721}
722
723int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
724{
725 if (result) {
726 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
727 evsel__hw_cache_op[op][0],
728 evsel__hw_cache_result[result][0]);
729 }
730
731 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
732 evsel__hw_cache_op[op][1]);
733}
734
735static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
736{
737 u8 op, result, type = (config >> 0) & 0xff;
738 const char *err = "unknown-ext-hardware-cache-type";
739
740 if (type >= PERF_COUNT_HW_CACHE_MAX)
741 goto out_err;
742
743 op = (config >> 8) & 0xff;
744 err = "unknown-ext-hardware-cache-op";
745 if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
746 goto out_err;
747
748 result = (config >> 16) & 0xff;
749 err = "unknown-ext-hardware-cache-result";
750 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
751 goto out_err;
752
753 err = "invalid-cache";
754 if (!evsel__is_cache_op_valid(type, op))
755 goto out_err;
756
757 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
758out_err:
759 return scnprintf(bf, size, "%s", err);
760}
761
762static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
763{
764 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
765 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
766}
767
768static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
769{
770 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
771 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
772}
773
774const char *evsel__name(struct evsel *evsel)
775{
776 char bf[128];
777
778 if (!evsel)
779 goto out_unknown;
780
781 if (evsel->name)
782 return evsel->name;
783
784 switch (evsel->core.attr.type) {
785 case PERF_TYPE_RAW:
786 evsel__raw_name(evsel, bf, sizeof(bf));
787 break;
788
789 case PERF_TYPE_HARDWARE:
790 evsel__hw_name(evsel, bf, sizeof(bf));
791 break;
792
793 case PERF_TYPE_HW_CACHE:
794 evsel__hw_cache_name(evsel, bf, sizeof(bf));
795 break;
796
797 case PERF_TYPE_SOFTWARE:
798 if (evsel__is_tool(evsel))
799 evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
800 else
801 evsel__sw_name(evsel, bf, sizeof(bf));
802 break;
803
804 case PERF_TYPE_TRACEPOINT:
805 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
806 break;
807
808 case PERF_TYPE_BREAKPOINT:
809 evsel__bp_name(evsel, bf, sizeof(bf));
810 break;
811
812 default:
813 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
814 evsel->core.attr.type);
815 break;
816 }
817
818 evsel->name = strdup(bf);
819
820 if (evsel->name)
821 return evsel->name;
822out_unknown:
823 return "unknown";
824}
825
826const char *evsel__metric_id(const struct evsel *evsel)
827{
828 if (evsel->metric_id)
829 return evsel->metric_id;
830
831 if (evsel__is_tool(evsel))
832 return perf_tool_event__to_str(evsel->tool_event);
833
834 return "unknown";
835}
836
837const char *evsel__group_name(struct evsel *evsel)
838{
839 return evsel->group_name ?: "anon group";
840}
841
842/*
843 * Returns the group details for the specified leader,
844 * with following rules.
845 *
846 * For record -e '{cycles,instructions}'
847 * 'anon group { cycles:u, instructions:u }'
848 *
849 * For record -e 'cycles,instructions' and report --group
850 * 'cycles:u, instructions:u'
851 */
852int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
853{
854 int ret = 0;
855 struct evsel *pos;
856 const char *group_name = evsel__group_name(evsel);
857
858 if (!evsel->forced_leader)
859 ret = scnprintf(buf, size, "%s { ", group_name);
860
861 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
862
863 for_each_group_member(pos, evsel)
864 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
865
866 if (!evsel->forced_leader)
867 ret += scnprintf(buf + ret, size - ret, " }");
868
869 return ret;
870}
871
872static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
873 struct callchain_param *param)
874{
875 bool function = evsel__is_function_event(evsel);
876 struct perf_event_attr *attr = &evsel->core.attr;
877
878 evsel__set_sample_bit(evsel, CALLCHAIN);
879
880 attr->sample_max_stack = param->max_stack;
881
882 if (opts->kernel_callchains)
883 attr->exclude_callchain_user = 1;
884 if (opts->user_callchains)
885 attr->exclude_callchain_kernel = 1;
886 if (param->record_mode == CALLCHAIN_LBR) {
887 if (!opts->branch_stack) {
888 if (attr->exclude_user) {
889 pr_warning("LBR callstack option is only available "
890 "to get user callchain information. "
891 "Falling back to framepointers.\n");
892 } else {
893 evsel__set_sample_bit(evsel, BRANCH_STACK);
894 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
895 PERF_SAMPLE_BRANCH_CALL_STACK |
896 PERF_SAMPLE_BRANCH_NO_CYCLES |
897 PERF_SAMPLE_BRANCH_NO_FLAGS |
898 PERF_SAMPLE_BRANCH_HW_INDEX;
899 }
900 } else
901 pr_warning("Cannot use LBR callstack with branch stack. "
902 "Falling back to framepointers.\n");
903 }
904
905 if (param->record_mode == CALLCHAIN_DWARF) {
906 if (!function) {
907 evsel__set_sample_bit(evsel, REGS_USER);
908 evsel__set_sample_bit(evsel, STACK_USER);
909 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
910 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
911 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
912 "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
913 "so the minimal registers set (IP, SP) is explicitly forced.\n");
914 } else {
915 attr->sample_regs_user |= arch__user_reg_mask();
916 }
917 attr->sample_stack_user = param->dump_size;
918 attr->exclude_callchain_user = 1;
919 } else {
920 pr_info("Cannot use DWARF unwind for function trace event,"
921 " falling back to framepointers.\n");
922 }
923 }
924
925 if (function) {
926 pr_info("Disabling user space callchains for function trace event.\n");
927 attr->exclude_callchain_user = 1;
928 }
929}
930
931void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
932 struct callchain_param *param)
933{
934 if (param->enabled)
935 return __evsel__config_callchain(evsel, opts, param);
936}
937
938static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
939{
940 struct perf_event_attr *attr = &evsel->core.attr;
941
942 evsel__reset_sample_bit(evsel, CALLCHAIN);
943 if (param->record_mode == CALLCHAIN_LBR) {
944 evsel__reset_sample_bit(evsel, BRANCH_STACK);
945 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
946 PERF_SAMPLE_BRANCH_CALL_STACK |
947 PERF_SAMPLE_BRANCH_HW_INDEX);
948 }
949 if (param->record_mode == CALLCHAIN_DWARF) {
950 evsel__reset_sample_bit(evsel, REGS_USER);
951 evsel__reset_sample_bit(evsel, STACK_USER);
952 }
953}
954
955static void evsel__apply_config_terms(struct evsel *evsel,
956 struct record_opts *opts, bool track)
957{
958 struct evsel_config_term *term;
959 struct list_head *config_terms = &evsel->config_terms;
960 struct perf_event_attr *attr = &evsel->core.attr;
961 /* callgraph default */
962 struct callchain_param param = {
963 .record_mode = callchain_param.record_mode,
964 };
965 u32 dump_size = 0;
966 int max_stack = 0;
967 const char *callgraph_buf = NULL;
968
969 list_for_each_entry(term, config_terms, list) {
970 switch (term->type) {
971 case EVSEL__CONFIG_TERM_PERIOD:
972 if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
973 attr->sample_period = term->val.period;
974 attr->freq = 0;
975 evsel__reset_sample_bit(evsel, PERIOD);
976 }
977 break;
978 case EVSEL__CONFIG_TERM_FREQ:
979 if (!(term->weak && opts->user_freq != UINT_MAX)) {
980 attr->sample_freq = term->val.freq;
981 attr->freq = 1;
982 evsel__set_sample_bit(evsel, PERIOD);
983 }
984 break;
985 case EVSEL__CONFIG_TERM_TIME:
986 if (term->val.time)
987 evsel__set_sample_bit(evsel, TIME);
988 else
989 evsel__reset_sample_bit(evsel, TIME);
990 break;
991 case EVSEL__CONFIG_TERM_CALLGRAPH:
992 callgraph_buf = term->val.str;
993 break;
994 case EVSEL__CONFIG_TERM_BRANCH:
995 if (term->val.str && strcmp(term->val.str, "no")) {
996 evsel__set_sample_bit(evsel, BRANCH_STACK);
997 parse_branch_str(term->val.str,
998 &attr->branch_sample_type);
999 } else
1000 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1001 break;
1002 case EVSEL__CONFIG_TERM_STACK_USER:
1003 dump_size = term->val.stack_user;
1004 break;
1005 case EVSEL__CONFIG_TERM_MAX_STACK:
1006 max_stack = term->val.max_stack;
1007 break;
1008 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1009 evsel->max_events = term->val.max_events;
1010 break;
1011 case EVSEL__CONFIG_TERM_INHERIT:
1012 /*
1013 * attr->inherit should has already been set by
1014 * evsel__config. If user explicitly set
1015 * inherit using config terms, override global
1016 * opt->no_inherit setting.
1017 */
1018 attr->inherit = term->val.inherit ? 1 : 0;
1019 break;
1020 case EVSEL__CONFIG_TERM_OVERWRITE:
1021 attr->write_backward = term->val.overwrite ? 1 : 0;
1022 break;
1023 case EVSEL__CONFIG_TERM_DRV_CFG:
1024 break;
1025 case EVSEL__CONFIG_TERM_PERCORE:
1026 break;
1027 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1028 attr->aux_output = term->val.aux_output ? 1 : 0;
1029 break;
1030 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1031 /* Already applied by auxtrace */
1032 break;
1033 case EVSEL__CONFIG_TERM_CFG_CHG:
1034 break;
1035 default:
1036 break;
1037 }
1038 }
1039
1040 /* User explicitly set per-event callgraph, clear the old setting and reset. */
1041 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1042 bool sample_address = false;
1043
1044 if (max_stack) {
1045 param.max_stack = max_stack;
1046 if (callgraph_buf == NULL)
1047 callgraph_buf = "fp";
1048 }
1049
1050 /* parse callgraph parameters */
1051 if (callgraph_buf != NULL) {
1052 if (!strcmp(callgraph_buf, "no")) {
1053 param.enabled = false;
1054 param.record_mode = CALLCHAIN_NONE;
1055 } else {
1056 param.enabled = true;
1057 if (parse_callchain_record(callgraph_buf, ¶m)) {
1058 pr_err("per-event callgraph setting for %s failed. "
1059 "Apply callgraph global setting for it\n",
1060 evsel->name);
1061 return;
1062 }
1063 if (param.record_mode == CALLCHAIN_DWARF)
1064 sample_address = true;
1065 }
1066 }
1067 if (dump_size > 0) {
1068 dump_size = round_up(dump_size, sizeof(u64));
1069 param.dump_size = dump_size;
1070 }
1071
1072 /* If global callgraph set, clear it */
1073 if (callchain_param.enabled)
1074 evsel__reset_callgraph(evsel, &callchain_param);
1075
1076 /* set perf-event callgraph */
1077 if (param.enabled) {
1078 if (sample_address) {
1079 evsel__set_sample_bit(evsel, ADDR);
1080 evsel__set_sample_bit(evsel, DATA_SRC);
1081 evsel->core.attr.mmap_data = track;
1082 }
1083 evsel__config_callchain(evsel, opts, ¶m);
1084 }
1085 }
1086}
1087
1088struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1089{
1090 struct evsel_config_term *term, *found_term = NULL;
1091
1092 list_for_each_entry(term, &evsel->config_terms, list) {
1093 if (term->type == type)
1094 found_term = term;
1095 }
1096
1097 return found_term;
1098}
1099
1100void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1101{
1102 evsel__set_sample_bit(evsel, WEIGHT);
1103}
1104
1105void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1106{
1107}
1108
1109void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1110 struct perf_event_attr *attr __maybe_unused)
1111{
1112}
1113
1114static void evsel__set_default_freq_period(struct record_opts *opts,
1115 struct perf_event_attr *attr)
1116{
1117 if (opts->freq) {
1118 attr->freq = 1;
1119 attr->sample_freq = opts->freq;
1120 } else {
1121 attr->sample_period = opts->default_interval;
1122 }
1123}
1124
1125static bool evsel__is_offcpu_event(struct evsel *evsel)
1126{
1127 return evsel__is_bpf_output(evsel) && !strcmp(evsel->name, OFFCPU_EVENT);
1128}
1129
1130/*
1131 * The enable_on_exec/disabled value strategy:
1132 *
1133 * 1) For any type of traced program:
1134 * - all independent events and group leaders are disabled
1135 * - all group members are enabled
1136 *
1137 * Group members are ruled by group leaders. They need to
1138 * be enabled, because the group scheduling relies on that.
1139 *
1140 * 2) For traced programs executed by perf:
1141 * - all independent events and group leaders have
1142 * enable_on_exec set
1143 * - we don't specifically enable or disable any event during
1144 * the record command
1145 *
1146 * Independent events and group leaders are initially disabled
1147 * and get enabled by exec. Group members are ruled by group
1148 * leaders as stated in 1).
1149 *
1150 * 3) For traced programs attached by perf (pid/tid):
1151 * - we specifically enable or disable all events during
1152 * the record command
1153 *
1154 * When attaching events to already running traced we
1155 * enable/disable events specifically, as there's no
1156 * initial traced exec call.
1157 */
1158void evsel__config(struct evsel *evsel, struct record_opts *opts,
1159 struct callchain_param *callchain)
1160{
1161 struct evsel *leader = evsel__leader(evsel);
1162 struct perf_event_attr *attr = &evsel->core.attr;
1163 int track = evsel->tracking;
1164 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1165
1166 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1167 attr->inherit = !opts->no_inherit;
1168 attr->write_backward = opts->overwrite ? 1 : 0;
1169 attr->read_format = PERF_FORMAT_LOST;
1170
1171 evsel__set_sample_bit(evsel, IP);
1172 evsel__set_sample_bit(evsel, TID);
1173
1174 if (evsel->sample_read) {
1175 evsel__set_sample_bit(evsel, READ);
1176
1177 /*
1178 * We need ID even in case of single event, because
1179 * PERF_SAMPLE_READ process ID specific data.
1180 */
1181 evsel__set_sample_id(evsel, false);
1182
1183 /*
1184 * Apply group format only if we belong to group
1185 * with more than one members.
1186 */
1187 if (leader->core.nr_members > 1) {
1188 attr->read_format |= PERF_FORMAT_GROUP;
1189 attr->inherit = 0;
1190 }
1191 }
1192
1193 /*
1194 * We default some events to have a default interval. But keep
1195 * it a weak assumption overridable by the user.
1196 */
1197 if ((evsel->is_libpfm_event && !attr->sample_period) ||
1198 (!evsel->is_libpfm_event && (!attr->sample_period ||
1199 opts->user_freq != UINT_MAX ||
1200 opts->user_interval != ULLONG_MAX)))
1201 evsel__set_default_freq_period(opts, attr);
1202
1203 /*
1204 * If attr->freq was set (here or earlier), ask for period
1205 * to be sampled.
1206 */
1207 if (attr->freq)
1208 evsel__set_sample_bit(evsel, PERIOD);
1209
1210 if (opts->no_samples)
1211 attr->sample_freq = 0;
1212
1213 if (opts->inherit_stat) {
1214 evsel->core.attr.read_format |=
1215 PERF_FORMAT_TOTAL_TIME_ENABLED |
1216 PERF_FORMAT_TOTAL_TIME_RUNNING |
1217 PERF_FORMAT_ID;
1218 attr->inherit_stat = 1;
1219 }
1220
1221 if (opts->sample_address) {
1222 evsel__set_sample_bit(evsel, ADDR);
1223 attr->mmap_data = track;
1224 }
1225
1226 /*
1227 * We don't allow user space callchains for function trace
1228 * event, due to issues with page faults while tracing page
1229 * fault handler and its overall trickiness nature.
1230 */
1231 if (evsel__is_function_event(evsel))
1232 evsel->core.attr.exclude_callchain_user = 1;
1233
1234 if (callchain && callchain->enabled && !evsel->no_aux_samples)
1235 evsel__config_callchain(evsel, opts, callchain);
1236
1237 if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1238 !evsel__is_dummy_event(evsel)) {
1239 attr->sample_regs_intr = opts->sample_intr_regs;
1240 evsel__set_sample_bit(evsel, REGS_INTR);
1241 }
1242
1243 if (opts->sample_user_regs && !evsel->no_aux_samples &&
1244 !evsel__is_dummy_event(evsel)) {
1245 attr->sample_regs_user |= opts->sample_user_regs;
1246 evsel__set_sample_bit(evsel, REGS_USER);
1247 }
1248
1249 if (target__has_cpu(&opts->target) || opts->sample_cpu)
1250 evsel__set_sample_bit(evsel, CPU);
1251
1252 /*
1253 * When the user explicitly disabled time don't force it here.
1254 */
1255 if (opts->sample_time &&
1256 (!perf_missing_features.sample_id_all &&
1257 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1258 opts->sample_time_set)))
1259 evsel__set_sample_bit(evsel, TIME);
1260
1261 if (opts->raw_samples && !evsel->no_aux_samples) {
1262 evsel__set_sample_bit(evsel, TIME);
1263 evsel__set_sample_bit(evsel, RAW);
1264 evsel__set_sample_bit(evsel, CPU);
1265 }
1266
1267 if (opts->sample_address)
1268 evsel__set_sample_bit(evsel, DATA_SRC);
1269
1270 if (opts->sample_phys_addr)
1271 evsel__set_sample_bit(evsel, PHYS_ADDR);
1272
1273 if (opts->no_buffering) {
1274 attr->watermark = 0;
1275 attr->wakeup_events = 1;
1276 }
1277 if (opts->branch_stack && !evsel->no_aux_samples) {
1278 evsel__set_sample_bit(evsel, BRANCH_STACK);
1279 attr->branch_sample_type = opts->branch_stack;
1280 }
1281
1282 if (opts->sample_weight)
1283 arch_evsel__set_sample_weight(evsel);
1284
1285 attr->task = track;
1286 attr->mmap = track;
1287 attr->mmap2 = track && !perf_missing_features.mmap2;
1288 attr->comm = track;
1289 attr->build_id = track && opts->build_id;
1290
1291 /*
1292 * ksymbol is tracked separately with text poke because it needs to be
1293 * system wide and enabled immediately.
1294 */
1295 if (!opts->text_poke)
1296 attr->ksymbol = track && !perf_missing_features.ksymbol;
1297 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1298
1299 if (opts->record_namespaces)
1300 attr->namespaces = track;
1301
1302 if (opts->record_cgroup) {
1303 attr->cgroup = track && !perf_missing_features.cgroup;
1304 evsel__set_sample_bit(evsel, CGROUP);
1305 }
1306
1307 if (opts->sample_data_page_size)
1308 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1309
1310 if (opts->sample_code_page_size)
1311 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1312
1313 if (opts->record_switch_events)
1314 attr->context_switch = track;
1315
1316 if (opts->sample_transaction)
1317 evsel__set_sample_bit(evsel, TRANSACTION);
1318
1319 if (opts->running_time) {
1320 evsel->core.attr.read_format |=
1321 PERF_FORMAT_TOTAL_TIME_ENABLED |
1322 PERF_FORMAT_TOTAL_TIME_RUNNING;
1323 }
1324
1325 /*
1326 * XXX see the function comment above
1327 *
1328 * Disabling only independent events or group leaders,
1329 * keeping group members enabled.
1330 */
1331 if (evsel__is_group_leader(evsel))
1332 attr->disabled = 1;
1333
1334 /*
1335 * Setting enable_on_exec for independent events and
1336 * group leaders for traced executed by perf.
1337 */
1338 if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1339 !opts->initial_delay)
1340 attr->enable_on_exec = 1;
1341
1342 if (evsel->immediate) {
1343 attr->disabled = 0;
1344 attr->enable_on_exec = 0;
1345 }
1346
1347 clockid = opts->clockid;
1348 if (opts->use_clockid) {
1349 attr->use_clockid = 1;
1350 attr->clockid = opts->clockid;
1351 }
1352
1353 if (evsel->precise_max)
1354 attr->precise_ip = 3;
1355
1356 if (opts->all_user) {
1357 attr->exclude_kernel = 1;
1358 attr->exclude_user = 0;
1359 }
1360
1361 if (opts->all_kernel) {
1362 attr->exclude_kernel = 0;
1363 attr->exclude_user = 1;
1364 }
1365
1366 if (evsel->core.own_cpus || evsel->unit)
1367 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1368
1369 /*
1370 * Apply event specific term settings,
1371 * it overloads any global configuration.
1372 */
1373 evsel__apply_config_terms(evsel, opts, track);
1374
1375 evsel->ignore_missing_thread = opts->ignore_missing_thread;
1376
1377 /* The --period option takes the precedence. */
1378 if (opts->period_set) {
1379 if (opts->period)
1380 evsel__set_sample_bit(evsel, PERIOD);
1381 else
1382 evsel__reset_sample_bit(evsel, PERIOD);
1383 }
1384
1385 /*
1386 * A dummy event never triggers any actual counter and therefore
1387 * cannot be used with branch_stack.
1388 *
1389 * For initial_delay, a dummy event is added implicitly.
1390 * The software event will trigger -EOPNOTSUPP error out,
1391 * if BRANCH_STACK bit is set.
1392 */
1393 if (evsel__is_dummy_event(evsel))
1394 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1395
1396 if (evsel__is_offcpu_event(evsel))
1397 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1398
1399 arch__post_evsel_config(evsel, attr);
1400}
1401
1402int evsel__set_filter(struct evsel *evsel, const char *filter)
1403{
1404 char *new_filter = strdup(filter);
1405
1406 if (new_filter != NULL) {
1407 free(evsel->filter);
1408 evsel->filter = new_filter;
1409 return 0;
1410 }
1411
1412 return -1;
1413}
1414
1415static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1416{
1417 char *new_filter;
1418
1419 if (evsel->filter == NULL)
1420 return evsel__set_filter(evsel, filter);
1421
1422 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1423 free(evsel->filter);
1424 evsel->filter = new_filter;
1425 return 0;
1426 }
1427
1428 return -1;
1429}
1430
1431int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1432{
1433 return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1434}
1435
1436int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1437{
1438 return evsel__append_filter(evsel, "%s,%s", filter);
1439}
1440
1441/* Caller has to clear disabled after going through all CPUs. */
1442int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1443{
1444 return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1445}
1446
1447int evsel__enable(struct evsel *evsel)
1448{
1449 int err = perf_evsel__enable(&evsel->core);
1450
1451 if (!err)
1452 evsel->disabled = false;
1453 return err;
1454}
1455
1456/* Caller has to set disabled after going through all CPUs. */
1457int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1458{
1459 return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1460}
1461
1462int evsel__disable(struct evsel *evsel)
1463{
1464 int err = perf_evsel__disable(&evsel->core);
1465 /*
1466 * We mark it disabled here so that tools that disable a event can
1467 * ignore events after they disable it. I.e. the ring buffer may have
1468 * already a few more events queued up before the kernel got the stop
1469 * request.
1470 */
1471 if (!err)
1472 evsel->disabled = true;
1473
1474 return err;
1475}
1476
1477void free_config_terms(struct list_head *config_terms)
1478{
1479 struct evsel_config_term *term, *h;
1480
1481 list_for_each_entry_safe(term, h, config_terms, list) {
1482 list_del_init(&term->list);
1483 if (term->free_str)
1484 zfree(&term->val.str);
1485 free(term);
1486 }
1487}
1488
1489static void evsel__free_config_terms(struct evsel *evsel)
1490{
1491 free_config_terms(&evsel->config_terms);
1492}
1493
1494void evsel__exit(struct evsel *evsel)
1495{
1496 assert(list_empty(&evsel->core.node));
1497 assert(evsel->evlist == NULL);
1498 bpf_counter__destroy(evsel);
1499 evsel__free_counts(evsel);
1500 perf_evsel__free_fd(&evsel->core);
1501 perf_evsel__free_id(&evsel->core);
1502 evsel__free_config_terms(evsel);
1503 cgroup__put(evsel->cgrp);
1504 perf_cpu_map__put(evsel->core.cpus);
1505 perf_cpu_map__put(evsel->core.own_cpus);
1506 perf_thread_map__put(evsel->core.threads);
1507 zfree(&evsel->group_name);
1508 zfree(&evsel->name);
1509 zfree(&evsel->pmu_name);
1510 zfree(&evsel->unit);
1511 zfree(&evsel->metric_id);
1512 evsel__zero_per_pkg(evsel);
1513 hashmap__free(evsel->per_pkg_mask);
1514 evsel->per_pkg_mask = NULL;
1515 zfree(&evsel->metric_events);
1516 perf_evsel__object.fini(evsel);
1517}
1518
1519void evsel__delete(struct evsel *evsel)
1520{
1521 evsel__exit(evsel);
1522 free(evsel);
1523}
1524
1525void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1526 struct perf_counts_values *count)
1527{
1528 struct perf_counts_values tmp;
1529
1530 if (!evsel->prev_raw_counts)
1531 return;
1532
1533 tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1534 *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1535
1536 count->val = count->val - tmp.val;
1537 count->ena = count->ena - tmp.ena;
1538 count->run = count->run - tmp.run;
1539}
1540
1541static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1542{
1543 struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1544
1545 return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1546}
1547
1548static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1549 u64 val, u64 ena, u64 run, u64 lost)
1550{
1551 struct perf_counts_values *count;
1552
1553 count = perf_counts(counter->counts, cpu_map_idx, thread);
1554
1555 count->val = val;
1556 count->ena = ena;
1557 count->run = run;
1558 count->lost = lost;
1559
1560 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1561}
1562
1563static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1564{
1565 u64 read_format = leader->core.attr.read_format;
1566 struct sample_read_value *v;
1567 u64 nr, ena = 0, run = 0, lost = 0;
1568
1569 nr = *data++;
1570
1571 if (nr != (u64) leader->core.nr_members)
1572 return -EINVAL;
1573
1574 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1575 ena = *data++;
1576
1577 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1578 run = *data++;
1579
1580 v = (void *)data;
1581 sample_read_group__for_each(v, nr, read_format) {
1582 struct evsel *counter;
1583
1584 counter = evlist__id2evsel(leader->evlist, v->id);
1585 if (!counter)
1586 return -EINVAL;
1587
1588 if (read_format & PERF_FORMAT_LOST)
1589 lost = v->lost;
1590
1591 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1592 }
1593
1594 return 0;
1595}
1596
1597static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1598{
1599 struct perf_stat_evsel *ps = leader->stats;
1600 u64 read_format = leader->core.attr.read_format;
1601 int size = perf_evsel__read_size(&leader->core);
1602 u64 *data = ps->group_data;
1603
1604 if (!(read_format & PERF_FORMAT_ID))
1605 return -EINVAL;
1606
1607 if (!evsel__is_group_leader(leader))
1608 return -EINVAL;
1609
1610 if (!data) {
1611 data = zalloc(size);
1612 if (!data)
1613 return -ENOMEM;
1614
1615 ps->group_data = data;
1616 }
1617
1618 if (FD(leader, cpu_map_idx, thread) < 0)
1619 return -EINVAL;
1620
1621 if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1622 return -errno;
1623
1624 return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1625}
1626
1627int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1628{
1629 u64 read_format = evsel->core.attr.read_format;
1630
1631 if (read_format & PERF_FORMAT_GROUP)
1632 return evsel__read_group(evsel, cpu_map_idx, thread);
1633
1634 return evsel__read_one(evsel, cpu_map_idx, thread);
1635}
1636
1637int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1638{
1639 struct perf_counts_values count;
1640 size_t nv = scale ? 3 : 1;
1641
1642 if (FD(evsel, cpu_map_idx, thread) < 0)
1643 return -EINVAL;
1644
1645 if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1646 return -ENOMEM;
1647
1648 if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1649 return -errno;
1650
1651 evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1652 perf_counts_values__scale(&count, scale, NULL);
1653 *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1654 return 0;
1655}
1656
1657static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1658 int cpu_map_idx)
1659{
1660 struct perf_cpu cpu;
1661
1662 cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1663 return perf_cpu_map__idx(other->core.cpus, cpu);
1664}
1665
1666static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1667{
1668 struct evsel *leader = evsel__leader(evsel);
1669
1670 if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1671 (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1672 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1673 }
1674
1675 return cpu_map_idx;
1676}
1677
1678static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1679{
1680 struct evsel *leader = evsel__leader(evsel);
1681 int fd;
1682
1683 if (evsel__is_group_leader(evsel))
1684 return -1;
1685
1686 /*
1687 * Leader must be already processed/open,
1688 * if not it's a bug.
1689 */
1690 BUG_ON(!leader->core.fd);
1691
1692 cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1693 if (cpu_map_idx == -1)
1694 return -1;
1695
1696 fd = FD(leader, cpu_map_idx, thread);
1697 BUG_ON(fd == -1);
1698
1699 return fd;
1700}
1701
1702static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1703{
1704 for (int cpu = 0; cpu < nr_cpus; cpu++)
1705 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1706 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1707}
1708
1709static int update_fds(struct evsel *evsel,
1710 int nr_cpus, int cpu_map_idx,
1711 int nr_threads, int thread_idx)
1712{
1713 struct evsel *pos;
1714
1715 if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1716 return -EINVAL;
1717
1718 evlist__for_each_entry(evsel->evlist, pos) {
1719 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1720
1721 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1722
1723 /*
1724 * Since fds for next evsel has not been created,
1725 * there is no need to iterate whole event list.
1726 */
1727 if (pos == evsel)
1728 break;
1729 }
1730 return 0;
1731}
1732
1733static bool evsel__ignore_missing_thread(struct evsel *evsel,
1734 int nr_cpus, int cpu_map_idx,
1735 struct perf_thread_map *threads,
1736 int thread, int err)
1737{
1738 pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1739
1740 if (!evsel->ignore_missing_thread)
1741 return false;
1742
1743 /* The system wide setup does not work with threads. */
1744 if (evsel->core.system_wide)
1745 return false;
1746
1747 /* The -ESRCH is perf event syscall errno for pid's not found. */
1748 if (err != -ESRCH)
1749 return false;
1750
1751 /* If there's only one thread, let it fail. */
1752 if (threads->nr == 1)
1753 return false;
1754
1755 /*
1756 * We should remove fd for missing_thread first
1757 * because thread_map__remove() will decrease threads->nr.
1758 */
1759 if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1760 return false;
1761
1762 if (thread_map__remove(threads, thread))
1763 return false;
1764
1765 pr_warning("WARNING: Ignored open failure for pid %d\n",
1766 ignore_pid);
1767 return true;
1768}
1769
1770static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1771 void *priv __maybe_unused)
1772{
1773 return fprintf(fp, " %-32s %s\n", name, val);
1774}
1775
1776static void display_attr(struct perf_event_attr *attr)
1777{
1778 if (verbose >= 2 || debug_peo_args) {
1779 fprintf(stderr, "%.60s\n", graph_dotted_line);
1780 fprintf(stderr, "perf_event_attr:\n");
1781 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1782 fprintf(stderr, "%.60s\n", graph_dotted_line);
1783 }
1784}
1785
1786bool evsel__precise_ip_fallback(struct evsel *evsel)
1787{
1788 /* Do not try less precise if not requested. */
1789 if (!evsel->precise_max)
1790 return false;
1791
1792 /*
1793 * We tried all the precise_ip values, and it's
1794 * still failing, so leave it to standard fallback.
1795 */
1796 if (!evsel->core.attr.precise_ip) {
1797 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1798 return false;
1799 }
1800
1801 if (!evsel->precise_ip_original)
1802 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1803
1804 evsel->core.attr.precise_ip--;
1805 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1806 display_attr(&evsel->core.attr);
1807 return true;
1808}
1809
1810static struct perf_cpu_map *empty_cpu_map;
1811static struct perf_thread_map *empty_thread_map;
1812
1813static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1814 struct perf_thread_map *threads)
1815{
1816 int nthreads = perf_thread_map__nr(threads);
1817
1818 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1819 (perf_missing_features.aux_output && evsel->core.attr.aux_output))
1820 return -EINVAL;
1821
1822 if (cpus == NULL) {
1823 if (empty_cpu_map == NULL) {
1824 empty_cpu_map = perf_cpu_map__dummy_new();
1825 if (empty_cpu_map == NULL)
1826 return -ENOMEM;
1827 }
1828
1829 cpus = empty_cpu_map;
1830 }
1831
1832 if (threads == NULL) {
1833 if (empty_thread_map == NULL) {
1834 empty_thread_map = thread_map__new_by_tid(-1);
1835 if (empty_thread_map == NULL)
1836 return -ENOMEM;
1837 }
1838
1839 threads = empty_thread_map;
1840 }
1841
1842 if (evsel->core.fd == NULL &&
1843 perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1844 return -ENOMEM;
1845
1846 evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1847 if (evsel->cgrp)
1848 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1849
1850 return 0;
1851}
1852
1853static void evsel__disable_missing_features(struct evsel *evsel)
1854{
1855 if (perf_missing_features.read_lost)
1856 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1857 if (perf_missing_features.weight_struct) {
1858 evsel__set_sample_bit(evsel, WEIGHT);
1859 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1860 }
1861 if (perf_missing_features.clockid_wrong)
1862 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1863 if (perf_missing_features.clockid) {
1864 evsel->core.attr.use_clockid = 0;
1865 evsel->core.attr.clockid = 0;
1866 }
1867 if (perf_missing_features.cloexec)
1868 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1869 if (perf_missing_features.mmap2)
1870 evsel->core.attr.mmap2 = 0;
1871 if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1872 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1873 if (perf_missing_features.lbr_flags)
1874 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1875 PERF_SAMPLE_BRANCH_NO_CYCLES);
1876 if (perf_missing_features.group_read && evsel->core.attr.inherit)
1877 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1878 if (perf_missing_features.ksymbol)
1879 evsel->core.attr.ksymbol = 0;
1880 if (perf_missing_features.bpf)
1881 evsel->core.attr.bpf_event = 0;
1882 if (perf_missing_features.branch_hw_idx)
1883 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1884 if (perf_missing_features.sample_id_all)
1885 evsel->core.attr.sample_id_all = 0;
1886}
1887
1888int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1889 struct perf_thread_map *threads)
1890{
1891 int err;
1892
1893 err = __evsel__prepare_open(evsel, cpus, threads);
1894 if (err)
1895 return err;
1896
1897 evsel__disable_missing_features(evsel);
1898
1899 return err;
1900}
1901
1902bool evsel__detect_missing_features(struct evsel *evsel)
1903{
1904 /*
1905 * Must probe features in the order they were added to the
1906 * perf_event_attr interface.
1907 */
1908 if (!perf_missing_features.read_lost &&
1909 (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1910 perf_missing_features.read_lost = true;
1911 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1912 return true;
1913 } else if (!perf_missing_features.weight_struct &&
1914 (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1915 perf_missing_features.weight_struct = true;
1916 pr_debug2("switching off weight struct support\n");
1917 return true;
1918 } else if (!perf_missing_features.code_page_size &&
1919 (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1920 perf_missing_features.code_page_size = true;
1921 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1922 return false;
1923 } else if (!perf_missing_features.data_page_size &&
1924 (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1925 perf_missing_features.data_page_size = true;
1926 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1927 return false;
1928 } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1929 perf_missing_features.cgroup = true;
1930 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1931 return false;
1932 } else if (!perf_missing_features.branch_hw_idx &&
1933 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1934 perf_missing_features.branch_hw_idx = true;
1935 pr_debug2("switching off branch HW index support\n");
1936 return true;
1937 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1938 perf_missing_features.aux_output = true;
1939 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1940 return false;
1941 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1942 perf_missing_features.bpf = true;
1943 pr_debug2_peo("switching off bpf_event\n");
1944 return true;
1945 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1946 perf_missing_features.ksymbol = true;
1947 pr_debug2_peo("switching off ksymbol\n");
1948 return true;
1949 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1950 perf_missing_features.write_backward = true;
1951 pr_debug2_peo("switching off write_backward\n");
1952 return false;
1953 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1954 perf_missing_features.clockid_wrong = true;
1955 pr_debug2_peo("switching off clockid\n");
1956 return true;
1957 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1958 perf_missing_features.clockid = true;
1959 pr_debug2_peo("switching off use_clockid\n");
1960 return true;
1961 } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1962 perf_missing_features.cloexec = true;
1963 pr_debug2_peo("switching off cloexec flag\n");
1964 return true;
1965 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1966 perf_missing_features.mmap2 = true;
1967 pr_debug2_peo("switching off mmap2\n");
1968 return true;
1969 } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1970 if (evsel->pmu == NULL)
1971 evsel->pmu = evsel__find_pmu(evsel);
1972
1973 if (evsel->pmu)
1974 evsel->pmu->missing_features.exclude_guest = true;
1975 else {
1976 /* we cannot find PMU, disable attrs now */
1977 evsel->core.attr.exclude_host = false;
1978 evsel->core.attr.exclude_guest = false;
1979 }
1980
1981 if (evsel->exclude_GH) {
1982 pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1983 return false;
1984 }
1985 if (!perf_missing_features.exclude_guest) {
1986 perf_missing_features.exclude_guest = true;
1987 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1988 }
1989 return true;
1990 } else if (!perf_missing_features.sample_id_all) {
1991 perf_missing_features.sample_id_all = true;
1992 pr_debug2_peo("switching off sample_id_all\n");
1993 return true;
1994 } else if (!perf_missing_features.lbr_flags &&
1995 (evsel->core.attr.branch_sample_type &
1996 (PERF_SAMPLE_BRANCH_NO_CYCLES |
1997 PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1998 perf_missing_features.lbr_flags = true;
1999 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2000 return true;
2001 } else if (!perf_missing_features.group_read &&
2002 evsel->core.attr.inherit &&
2003 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2004 evsel__is_group_leader(evsel)) {
2005 perf_missing_features.group_read = true;
2006 pr_debug2_peo("switching off group read\n");
2007 return true;
2008 } else {
2009 return false;
2010 }
2011}
2012
2013bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2014{
2015 int old_errno;
2016 struct rlimit l;
2017
2018 if (*set_rlimit < INCREASED_MAX) {
2019 old_errno = errno;
2020
2021 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2022 if (*set_rlimit == NO_CHANGE) {
2023 l.rlim_cur = l.rlim_max;
2024 } else {
2025 l.rlim_cur = l.rlim_max + 1000;
2026 l.rlim_max = l.rlim_cur;
2027 }
2028 if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2029 (*set_rlimit) += 1;
2030 errno = old_errno;
2031 return true;
2032 }
2033 }
2034 errno = old_errno;
2035 }
2036
2037 return false;
2038}
2039
2040static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2041 struct perf_thread_map *threads,
2042 int start_cpu_map_idx, int end_cpu_map_idx)
2043{
2044 int idx, thread, nthreads;
2045 int pid = -1, err, old_errno;
2046 enum rlimit_action set_rlimit = NO_CHANGE;
2047
2048 err = __evsel__prepare_open(evsel, cpus, threads);
2049 if (err)
2050 return err;
2051
2052 if (cpus == NULL)
2053 cpus = empty_cpu_map;
2054
2055 if (threads == NULL)
2056 threads = empty_thread_map;
2057
2058 nthreads = perf_thread_map__nr(threads);
2059
2060 if (evsel->cgrp)
2061 pid = evsel->cgrp->fd;
2062
2063fallback_missing_features:
2064 evsel__disable_missing_features(evsel);
2065
2066 display_attr(&evsel->core.attr);
2067
2068 for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2069
2070 for (thread = 0; thread < nthreads; thread++) {
2071 int fd, group_fd;
2072retry_open:
2073 if (thread >= nthreads)
2074 break;
2075
2076 if (!evsel->cgrp && !evsel->core.system_wide)
2077 pid = perf_thread_map__pid(threads, thread);
2078
2079 group_fd = get_group_fd(evsel, idx, thread);
2080
2081 test_attr__ready();
2082
2083 /* Debug message used by test scripts */
2084 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx",
2085 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2086
2087 fd = sys_perf_event_open(&evsel->core.attr, pid,
2088 perf_cpu_map__cpu(cpus, idx).cpu,
2089 group_fd, evsel->open_flags);
2090
2091 FD(evsel, idx, thread) = fd;
2092
2093 if (fd < 0) {
2094 err = -errno;
2095
2096 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2097 err);
2098 goto try_fallback;
2099 }
2100
2101 bpf_counter__install_pe(evsel, idx, fd);
2102
2103 if (unlikely(test_attr__enabled)) {
2104 test_attr__open(&evsel->core.attr, pid,
2105 perf_cpu_map__cpu(cpus, idx),
2106 fd, group_fd, evsel->open_flags);
2107 }
2108
2109 /* Debug message used by test scripts */
2110 pr_debug2_peo(" = %d\n", fd);
2111
2112 if (evsel->bpf_fd >= 0) {
2113 int evt_fd = fd;
2114 int bpf_fd = evsel->bpf_fd;
2115
2116 err = ioctl(evt_fd,
2117 PERF_EVENT_IOC_SET_BPF,
2118 bpf_fd);
2119 if (err && errno != EEXIST) {
2120 pr_err("failed to attach bpf fd %d: %s\n",
2121 bpf_fd, strerror(errno));
2122 err = -EINVAL;
2123 goto out_close;
2124 }
2125 }
2126
2127 set_rlimit = NO_CHANGE;
2128
2129 /*
2130 * If we succeeded but had to kill clockid, fail and
2131 * have evsel__open_strerror() print us a nice error.
2132 */
2133 if (perf_missing_features.clockid ||
2134 perf_missing_features.clockid_wrong) {
2135 err = -EINVAL;
2136 goto out_close;
2137 }
2138 }
2139 }
2140
2141 return 0;
2142
2143try_fallback:
2144 if (evsel__precise_ip_fallback(evsel))
2145 goto retry_open;
2146
2147 if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2148 idx, threads, thread, err)) {
2149 /* We just removed 1 thread, so lower the upper nthreads limit. */
2150 nthreads--;
2151
2152 /* ... and pretend like nothing have happened. */
2153 err = 0;
2154 goto retry_open;
2155 }
2156 /*
2157 * perf stat needs between 5 and 22 fds per CPU. When we run out
2158 * of them try to increase the limits.
2159 */
2160 if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2161 goto retry_open;
2162
2163 if (err != -EINVAL || idx > 0 || thread > 0)
2164 goto out_close;
2165
2166 if (evsel__detect_missing_features(evsel))
2167 goto fallback_missing_features;
2168out_close:
2169 if (err)
2170 threads->err_thread = thread;
2171
2172 old_errno = errno;
2173 do {
2174 while (--thread >= 0) {
2175 if (FD(evsel, idx, thread) >= 0)
2176 close(FD(evsel, idx, thread));
2177 FD(evsel, idx, thread) = -1;
2178 }
2179 thread = nthreads;
2180 } while (--idx >= 0);
2181 errno = old_errno;
2182 return err;
2183}
2184
2185int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2186 struct perf_thread_map *threads)
2187{
2188 return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2189}
2190
2191void evsel__close(struct evsel *evsel)
2192{
2193 perf_evsel__close(&evsel->core);
2194 perf_evsel__free_id(&evsel->core);
2195}
2196
2197int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2198{
2199 if (cpu_map_idx == -1)
2200 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2201
2202 return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2203}
2204
2205int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2206{
2207 return evsel__open(evsel, NULL, threads);
2208}
2209
2210static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2211 const union perf_event *event,
2212 struct perf_sample *sample)
2213{
2214 u64 type = evsel->core.attr.sample_type;
2215 const __u64 *array = event->sample.array;
2216 bool swapped = evsel->needs_swap;
2217 union u64_swap u;
2218
2219 array += ((event->header.size -
2220 sizeof(event->header)) / sizeof(u64)) - 1;
2221
2222 if (type & PERF_SAMPLE_IDENTIFIER) {
2223 sample->id = *array;
2224 array--;
2225 }
2226
2227 if (type & PERF_SAMPLE_CPU) {
2228 u.val64 = *array;
2229 if (swapped) {
2230 /* undo swap of u64, then swap on individual u32s */
2231 u.val64 = bswap_64(u.val64);
2232 u.val32[0] = bswap_32(u.val32[0]);
2233 }
2234
2235 sample->cpu = u.val32[0];
2236 array--;
2237 }
2238
2239 if (type & PERF_SAMPLE_STREAM_ID) {
2240 sample->stream_id = *array;
2241 array--;
2242 }
2243
2244 if (type & PERF_SAMPLE_ID) {
2245 sample->id = *array;
2246 array--;
2247 }
2248
2249 if (type & PERF_SAMPLE_TIME) {
2250 sample->time = *array;
2251 array--;
2252 }
2253
2254 if (type & PERF_SAMPLE_TID) {
2255 u.val64 = *array;
2256 if (swapped) {
2257 /* undo swap of u64, then swap on individual u32s */
2258 u.val64 = bswap_64(u.val64);
2259 u.val32[0] = bswap_32(u.val32[0]);
2260 u.val32[1] = bswap_32(u.val32[1]);
2261 }
2262
2263 sample->pid = u.val32[0];
2264 sample->tid = u.val32[1];
2265 array--;
2266 }
2267
2268 return 0;
2269}
2270
2271static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2272 u64 size)
2273{
2274 return size > max_size || offset + size > endp;
2275}
2276
2277#define OVERFLOW_CHECK(offset, size, max_size) \
2278 do { \
2279 if (overflow(endp, (max_size), (offset), (size))) \
2280 return -EFAULT; \
2281 } while (0)
2282
2283#define OVERFLOW_CHECK_u64(offset) \
2284 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2285
2286static int
2287perf_event__check_size(union perf_event *event, unsigned int sample_size)
2288{
2289 /*
2290 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2291 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to
2292 * check the format does not go past the end of the event.
2293 */
2294 if (sample_size + sizeof(event->header) > event->header.size)
2295 return -EFAULT;
2296
2297 return 0;
2298}
2299
2300void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2301 const __u64 *array,
2302 u64 type __maybe_unused)
2303{
2304 data->weight = *array;
2305}
2306
2307u64 evsel__bitfield_swap_branch_flags(u64 value)
2308{
2309 u64 new_val = 0;
2310
2311 /*
2312 * branch_flags
2313 * union {
2314 * u64 values;
2315 * struct {
2316 * mispred:1 //target mispredicted
2317 * predicted:1 //target predicted
2318 * in_tx:1 //in transaction
2319 * abort:1 //transaction abort
2320 * cycles:16 //cycle count to last branch
2321 * type:4 //branch type
2322 * reserved:40
2323 * }
2324 * }
2325 *
2326 * Avoid bswap64() the entire branch_flag.value,
2327 * as it has variable bit-field sizes. Instead the
2328 * macro takes the bit-field position/size,
2329 * swaps it based on the host endianness.
2330 */
2331 if (host_is_bigendian()) {
2332 new_val = bitfield_swap(value, 0, 1);
2333 new_val |= bitfield_swap(value, 1, 1);
2334 new_val |= bitfield_swap(value, 2, 1);
2335 new_val |= bitfield_swap(value, 3, 1);
2336 new_val |= bitfield_swap(value, 4, 16);
2337 new_val |= bitfield_swap(value, 20, 4);
2338 new_val |= bitfield_swap(value, 24, 40);
2339 } else {
2340 new_val = bitfield_swap(value, 63, 1);
2341 new_val |= bitfield_swap(value, 62, 1);
2342 new_val |= bitfield_swap(value, 61, 1);
2343 new_val |= bitfield_swap(value, 60, 1);
2344 new_val |= bitfield_swap(value, 44, 16);
2345 new_val |= bitfield_swap(value, 40, 4);
2346 new_val |= bitfield_swap(value, 0, 40);
2347 }
2348
2349 return new_val;
2350}
2351
2352int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2353 struct perf_sample *data)
2354{
2355 u64 type = evsel->core.attr.sample_type;
2356 bool swapped = evsel->needs_swap;
2357 const __u64 *array;
2358 u16 max_size = event->header.size;
2359 const void *endp = (void *)event + max_size;
2360 u64 sz;
2361
2362 /*
2363 * used for cross-endian analysis. See git commit 65014ab3
2364 * for why this goofiness is needed.
2365 */
2366 union u64_swap u;
2367
2368 memset(data, 0, sizeof(*data));
2369 data->cpu = data->pid = data->tid = -1;
2370 data->stream_id = data->id = data->time = -1ULL;
2371 data->period = evsel->core.attr.sample_period;
2372 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2373 data->misc = event->header.misc;
2374 data->id = -1ULL;
2375 data->data_src = PERF_MEM_DATA_SRC_NONE;
2376 data->vcpu = -1;
2377
2378 if (event->header.type != PERF_RECORD_SAMPLE) {
2379 if (!evsel->core.attr.sample_id_all)
2380 return 0;
2381 return perf_evsel__parse_id_sample(evsel, event, data);
2382 }
2383
2384 array = event->sample.array;
2385
2386 if (perf_event__check_size(event, evsel->sample_size))
2387 return -EFAULT;
2388
2389 if (type & PERF_SAMPLE_IDENTIFIER) {
2390 data->id = *array;
2391 array++;
2392 }
2393
2394 if (type & PERF_SAMPLE_IP) {
2395 data->ip = *array;
2396 array++;
2397 }
2398
2399 if (type & PERF_SAMPLE_TID) {
2400 u.val64 = *array;
2401 if (swapped) {
2402 /* undo swap of u64, then swap on individual u32s */
2403 u.val64 = bswap_64(u.val64);
2404 u.val32[0] = bswap_32(u.val32[0]);
2405 u.val32[1] = bswap_32(u.val32[1]);
2406 }
2407
2408 data->pid = u.val32[0];
2409 data->tid = u.val32[1];
2410 array++;
2411 }
2412
2413 if (type & PERF_SAMPLE_TIME) {
2414 data->time = *array;
2415 array++;
2416 }
2417
2418 if (type & PERF_SAMPLE_ADDR) {
2419 data->addr = *array;
2420 array++;
2421 }
2422
2423 if (type & PERF_SAMPLE_ID) {
2424 data->id = *array;
2425 array++;
2426 }
2427
2428 if (type & PERF_SAMPLE_STREAM_ID) {
2429 data->stream_id = *array;
2430 array++;
2431 }
2432
2433 if (type & PERF_SAMPLE_CPU) {
2434
2435 u.val64 = *array;
2436 if (swapped) {
2437 /* undo swap of u64, then swap on individual u32s */
2438 u.val64 = bswap_64(u.val64);
2439 u.val32[0] = bswap_32(u.val32[0]);
2440 }
2441
2442 data->cpu = u.val32[0];
2443 array++;
2444 }
2445
2446 if (type & PERF_SAMPLE_PERIOD) {
2447 data->period = *array;
2448 array++;
2449 }
2450
2451 if (type & PERF_SAMPLE_READ) {
2452 u64 read_format = evsel->core.attr.read_format;
2453
2454 OVERFLOW_CHECK_u64(array);
2455 if (read_format & PERF_FORMAT_GROUP)
2456 data->read.group.nr = *array;
2457 else
2458 data->read.one.value = *array;
2459
2460 array++;
2461
2462 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2463 OVERFLOW_CHECK_u64(array);
2464 data->read.time_enabled = *array;
2465 array++;
2466 }
2467
2468 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2469 OVERFLOW_CHECK_u64(array);
2470 data->read.time_running = *array;
2471 array++;
2472 }
2473
2474 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2475 if (read_format & PERF_FORMAT_GROUP) {
2476 const u64 max_group_nr = UINT64_MAX /
2477 sizeof(struct sample_read_value);
2478
2479 if (data->read.group.nr > max_group_nr)
2480 return -EFAULT;
2481
2482 sz = data->read.group.nr * sample_read_value_size(read_format);
2483 OVERFLOW_CHECK(array, sz, max_size);
2484 data->read.group.values =
2485 (struct sample_read_value *)array;
2486 array = (void *)array + sz;
2487 } else {
2488 OVERFLOW_CHECK_u64(array);
2489 data->read.one.id = *array;
2490 array++;
2491
2492 if (read_format & PERF_FORMAT_LOST) {
2493 OVERFLOW_CHECK_u64(array);
2494 data->read.one.lost = *array;
2495 array++;
2496 }
2497 }
2498 }
2499
2500 if (type & PERF_SAMPLE_CALLCHAIN) {
2501 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2502
2503 OVERFLOW_CHECK_u64(array);
2504 data->callchain = (struct ip_callchain *)array++;
2505 if (data->callchain->nr > max_callchain_nr)
2506 return -EFAULT;
2507 sz = data->callchain->nr * sizeof(u64);
2508 OVERFLOW_CHECK(array, sz, max_size);
2509 array = (void *)array + sz;
2510 }
2511
2512 if (type & PERF_SAMPLE_RAW) {
2513 OVERFLOW_CHECK_u64(array);
2514 u.val64 = *array;
2515
2516 /*
2517 * Undo swap of u64, then swap on individual u32s,
2518 * get the size of the raw area and undo all of the
2519 * swap. The pevent interface handles endianness by
2520 * itself.
2521 */
2522 if (swapped) {
2523 u.val64 = bswap_64(u.val64);
2524 u.val32[0] = bswap_32(u.val32[0]);
2525 u.val32[1] = bswap_32(u.val32[1]);
2526 }
2527 data->raw_size = u.val32[0];
2528
2529 /*
2530 * The raw data is aligned on 64bits including the
2531 * u32 size, so it's safe to use mem_bswap_64.
2532 */
2533 if (swapped)
2534 mem_bswap_64((void *) array, data->raw_size);
2535
2536 array = (void *)array + sizeof(u32);
2537
2538 OVERFLOW_CHECK(array, data->raw_size, max_size);
2539 data->raw_data = (void *)array;
2540 array = (void *)array + data->raw_size;
2541 }
2542
2543 if (type & PERF_SAMPLE_BRANCH_STACK) {
2544 const u64 max_branch_nr = UINT64_MAX /
2545 sizeof(struct branch_entry);
2546 struct branch_entry *e;
2547 unsigned int i;
2548
2549 OVERFLOW_CHECK_u64(array);
2550 data->branch_stack = (struct branch_stack *)array++;
2551
2552 if (data->branch_stack->nr > max_branch_nr)
2553 return -EFAULT;
2554
2555 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2556 if (evsel__has_branch_hw_idx(evsel)) {
2557 sz += sizeof(u64);
2558 e = &data->branch_stack->entries[0];
2559 } else {
2560 data->no_hw_idx = true;
2561 /*
2562 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2563 * only nr and entries[] will be output by kernel.
2564 */
2565 e = (struct branch_entry *)&data->branch_stack->hw_idx;
2566 }
2567
2568 if (swapped) {
2569 /*
2570 * struct branch_flag does not have endian
2571 * specific bit field definition. And bswap
2572 * will not resolve the issue, since these
2573 * are bit fields.
2574 *
2575 * evsel__bitfield_swap_branch_flags() uses a
2576 * bitfield_swap macro to swap the bit position
2577 * based on the host endians.
2578 */
2579 for (i = 0; i < data->branch_stack->nr; i++, e++)
2580 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2581 }
2582
2583 OVERFLOW_CHECK(array, sz, max_size);
2584 array = (void *)array + sz;
2585 }
2586
2587 if (type & PERF_SAMPLE_REGS_USER) {
2588 OVERFLOW_CHECK_u64(array);
2589 data->user_regs.abi = *array;
2590 array++;
2591
2592 if (data->user_regs.abi) {
2593 u64 mask = evsel->core.attr.sample_regs_user;
2594
2595 sz = hweight64(mask) * sizeof(u64);
2596 OVERFLOW_CHECK(array, sz, max_size);
2597 data->user_regs.mask = mask;
2598 data->user_regs.regs = (u64 *)array;
2599 array = (void *)array + sz;
2600 }
2601 }
2602
2603 if (type & PERF_SAMPLE_STACK_USER) {
2604 OVERFLOW_CHECK_u64(array);
2605 sz = *array++;
2606
2607 data->user_stack.offset = ((char *)(array - 1)
2608 - (char *) event);
2609
2610 if (!sz) {
2611 data->user_stack.size = 0;
2612 } else {
2613 OVERFLOW_CHECK(array, sz, max_size);
2614 data->user_stack.data = (char *)array;
2615 array = (void *)array + sz;
2616 OVERFLOW_CHECK_u64(array);
2617 data->user_stack.size = *array++;
2618 if (WARN_ONCE(data->user_stack.size > sz,
2619 "user stack dump failure\n"))
2620 return -EFAULT;
2621 }
2622 }
2623
2624 if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2625 OVERFLOW_CHECK_u64(array);
2626 arch_perf_parse_sample_weight(data, array, type);
2627 array++;
2628 }
2629
2630 if (type & PERF_SAMPLE_DATA_SRC) {
2631 OVERFLOW_CHECK_u64(array);
2632 data->data_src = *array;
2633 array++;
2634 }
2635
2636 if (type & PERF_SAMPLE_TRANSACTION) {
2637 OVERFLOW_CHECK_u64(array);
2638 data->transaction = *array;
2639 array++;
2640 }
2641
2642 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2643 if (type & PERF_SAMPLE_REGS_INTR) {
2644 OVERFLOW_CHECK_u64(array);
2645 data->intr_regs.abi = *array;
2646 array++;
2647
2648 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2649 u64 mask = evsel->core.attr.sample_regs_intr;
2650
2651 sz = hweight64(mask) * sizeof(u64);
2652 OVERFLOW_CHECK(array, sz, max_size);
2653 data->intr_regs.mask = mask;
2654 data->intr_regs.regs = (u64 *)array;
2655 array = (void *)array + sz;
2656 }
2657 }
2658
2659 data->phys_addr = 0;
2660 if (type & PERF_SAMPLE_PHYS_ADDR) {
2661 data->phys_addr = *array;
2662 array++;
2663 }
2664
2665 data->cgroup = 0;
2666 if (type & PERF_SAMPLE_CGROUP) {
2667 data->cgroup = *array;
2668 array++;
2669 }
2670
2671 data->data_page_size = 0;
2672 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2673 data->data_page_size = *array;
2674 array++;
2675 }
2676
2677 data->code_page_size = 0;
2678 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2679 data->code_page_size = *array;
2680 array++;
2681 }
2682
2683 if (type & PERF_SAMPLE_AUX) {
2684 OVERFLOW_CHECK_u64(array);
2685 sz = *array++;
2686
2687 OVERFLOW_CHECK(array, sz, max_size);
2688 /* Undo swap of data */
2689 if (swapped)
2690 mem_bswap_64((char *)array, sz);
2691 data->aux_sample.size = sz;
2692 data->aux_sample.data = (char *)array;
2693 array = (void *)array + sz;
2694 }
2695
2696 return 0;
2697}
2698
2699int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2700 u64 *timestamp)
2701{
2702 u64 type = evsel->core.attr.sample_type;
2703 const __u64 *array;
2704
2705 if (!(type & PERF_SAMPLE_TIME))
2706 return -1;
2707
2708 if (event->header.type != PERF_RECORD_SAMPLE) {
2709 struct perf_sample data = {
2710 .time = -1ULL,
2711 };
2712
2713 if (!evsel->core.attr.sample_id_all)
2714 return -1;
2715 if (perf_evsel__parse_id_sample(evsel, event, &data))
2716 return -1;
2717
2718 *timestamp = data.time;
2719 return 0;
2720 }
2721
2722 array = event->sample.array;
2723
2724 if (perf_event__check_size(event, evsel->sample_size))
2725 return -EFAULT;
2726
2727 if (type & PERF_SAMPLE_IDENTIFIER)
2728 array++;
2729
2730 if (type & PERF_SAMPLE_IP)
2731 array++;
2732
2733 if (type & PERF_SAMPLE_TID)
2734 array++;
2735
2736 if (type & PERF_SAMPLE_TIME)
2737 *timestamp = *array;
2738
2739 return 0;
2740}
2741
2742u16 evsel__id_hdr_size(struct evsel *evsel)
2743{
2744 u64 sample_type = evsel->core.attr.sample_type;
2745 u16 size = 0;
2746
2747 if (sample_type & PERF_SAMPLE_TID)
2748 size += sizeof(u64);
2749
2750 if (sample_type & PERF_SAMPLE_TIME)
2751 size += sizeof(u64);
2752
2753 if (sample_type & PERF_SAMPLE_ID)
2754 size += sizeof(u64);
2755
2756 if (sample_type & PERF_SAMPLE_STREAM_ID)
2757 size += sizeof(u64);
2758
2759 if (sample_type & PERF_SAMPLE_CPU)
2760 size += sizeof(u64);
2761
2762 if (sample_type & PERF_SAMPLE_IDENTIFIER)
2763 size += sizeof(u64);
2764
2765 return size;
2766}
2767
2768#ifdef HAVE_LIBTRACEEVENT
2769struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2770{
2771 return tep_find_field(evsel->tp_format, name);
2772}
2773
2774void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2775{
2776 struct tep_format_field *field = evsel__field(evsel, name);
2777 int offset;
2778
2779 if (!field)
2780 return NULL;
2781
2782 offset = field->offset;
2783
2784 if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2785 offset = *(int *)(sample->raw_data + field->offset);
2786 offset &= 0xffff;
2787#ifdef HAVE_LIBTRACEEVENT_TEP_FIELD_IS_RELATIVE
2788 if (field->flags & TEP_FIELD_IS_RELATIVE)
2789 offset += field->offset + field->size;
2790#endif
2791 }
2792
2793 return sample->raw_data + offset;
2794}
2795
2796u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2797 bool needs_swap)
2798{
2799 u64 value;
2800 void *ptr = sample->raw_data + field->offset;
2801
2802 switch (field->size) {
2803 case 1:
2804 return *(u8 *)ptr;
2805 case 2:
2806 value = *(u16 *)ptr;
2807 break;
2808 case 4:
2809 value = *(u32 *)ptr;
2810 break;
2811 case 8:
2812 memcpy(&value, ptr, sizeof(u64));
2813 break;
2814 default:
2815 return 0;
2816 }
2817
2818 if (!needs_swap)
2819 return value;
2820
2821 switch (field->size) {
2822 case 2:
2823 return bswap_16(value);
2824 case 4:
2825 return bswap_32(value);
2826 case 8:
2827 return bswap_64(value);
2828 default:
2829 return 0;
2830 }
2831
2832 return 0;
2833}
2834
2835u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2836{
2837 struct tep_format_field *field = evsel__field(evsel, name);
2838
2839 if (!field)
2840 return 0;
2841
2842 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2843}
2844#endif
2845
2846bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2847{
2848 int paranoid;
2849
2850 if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2851 evsel->core.attr.type == PERF_TYPE_HARDWARE &&
2852 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2853 /*
2854 * If it's cycles then fall back to hrtimer based
2855 * cpu-clock-tick sw counter, which is always available even if
2856 * no PMU support.
2857 *
2858 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2859 * b0a873e).
2860 */
2861 scnprintf(msg, msgsize, "%s",
2862"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2863
2864 evsel->core.attr.type = PERF_TYPE_SOFTWARE;
2865 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2866
2867 zfree(&evsel->name);
2868 return true;
2869 } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2870 (paranoid = perf_event_paranoid()) > 1) {
2871 const char *name = evsel__name(evsel);
2872 char *new_name;
2873 const char *sep = ":";
2874
2875 /* If event has exclude user then don't exclude kernel. */
2876 if (evsel->core.attr.exclude_user)
2877 return false;
2878
2879 /* Is there already the separator in the name. */
2880 if (strchr(name, '/') ||
2881 (strchr(name, ':') && !evsel->is_libpfm_event))
2882 sep = "";
2883
2884 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2885 return false;
2886
2887 if (evsel->name)
2888 free(evsel->name);
2889 evsel->name = new_name;
2890 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2891 "to fall back to excluding kernel and hypervisor "
2892 " samples", paranoid);
2893 evsel->core.attr.exclude_kernel = 1;
2894 evsel->core.attr.exclude_hv = 1;
2895
2896 return true;
2897 }
2898
2899 return false;
2900}
2901
2902static bool find_process(const char *name)
2903{
2904 size_t len = strlen(name);
2905 DIR *dir;
2906 struct dirent *d;
2907 int ret = -1;
2908
2909 dir = opendir(procfs__mountpoint());
2910 if (!dir)
2911 return false;
2912
2913 /* Walk through the directory. */
2914 while (ret && (d = readdir(dir)) != NULL) {
2915 char path[PATH_MAX];
2916 char *data;
2917 size_t size;
2918
2919 if ((d->d_type != DT_DIR) ||
2920 !strcmp(".", d->d_name) ||
2921 !strcmp("..", d->d_name))
2922 continue;
2923
2924 scnprintf(path, sizeof(path), "%s/%s/comm",
2925 procfs__mountpoint(), d->d_name);
2926
2927 if (filename__read_str(path, &data, &size))
2928 continue;
2929
2930 ret = strncmp(name, data, len);
2931 free(data);
2932 }
2933
2934 closedir(dir);
2935 return ret ? false : true;
2936}
2937
2938static bool is_amd(const char *arch, const char *cpuid)
2939{
2940 return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2941}
2942
2943static bool is_amd_ibs(struct evsel *evsel)
2944{
2945 return evsel->core.attr.precise_ip
2946 || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2947}
2948
2949int evsel__open_strerror(struct evsel *evsel, struct target *target,
2950 int err, char *msg, size_t size)
2951{
2952 struct perf_env *env = evsel__env(evsel);
2953 const char *arch = perf_env__arch(env);
2954 const char *cpuid = perf_env__cpuid(env);
2955 char sbuf[STRERR_BUFSIZE];
2956 int printed = 0, enforced = 0;
2957
2958 switch (err) {
2959 case EPERM:
2960 case EACCES:
2961 printed += scnprintf(msg + printed, size - printed,
2962 "Access to performance monitoring and observability operations is limited.\n");
2963
2964 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2965 if (enforced) {
2966 printed += scnprintf(msg + printed, size - printed,
2967 "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2968 "monitoring and observability operations. Inspect system audit records for\n"
2969 "more perf_event access control information and adjusting the policy.\n");
2970 }
2971 }
2972
2973 if (err == EPERM)
2974 printed += scnprintf(msg, size,
2975 "No permission to enable %s event.\n\n", evsel__name(evsel));
2976
2977 return scnprintf(msg + printed, size - printed,
2978 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2979 "access to performance monitoring and observability operations for processes\n"
2980 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2981 "More information can be found at 'Perf events and tool security' document:\n"
2982 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2983 "perf_event_paranoid setting is %d:\n"
2984 " -1: Allow use of (almost) all events by all users\n"
2985 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2986 ">= 0: Disallow raw and ftrace function tracepoint access\n"
2987 ">= 1: Disallow CPU event access\n"
2988 ">= 2: Disallow kernel profiling\n"
2989 "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2990 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2991 perf_event_paranoid());
2992 case ENOENT:
2993 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2994 case EMFILE:
2995 return scnprintf(msg, size, "%s",
2996 "Too many events are opened.\n"
2997 "Probably the maximum number of open file descriptors has been reached.\n"
2998 "Hint: Try again after reducing the number of events.\n"
2999 "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3000 case ENOMEM:
3001 if (evsel__has_callchain(evsel) &&
3002 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3003 return scnprintf(msg, size,
3004 "Not enough memory to setup event with callchain.\n"
3005 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3006 "Hint: Current value: %d", sysctl__max_stack());
3007 break;
3008 case ENODEV:
3009 if (target->cpu_list)
3010 return scnprintf(msg, size, "%s",
3011 "No such device - did you specify an out-of-range profile CPU?");
3012 break;
3013 case EOPNOTSUPP:
3014 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3015 return scnprintf(msg, size,
3016 "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3017 evsel__name(evsel));
3018 if (evsel->core.attr.aux_output)
3019 return scnprintf(msg, size,
3020 "%s: PMU Hardware doesn't support 'aux_output' feature",
3021 evsel__name(evsel));
3022 if (evsel->core.attr.sample_period != 0)
3023 return scnprintf(msg, size,
3024 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3025 evsel__name(evsel));
3026 if (evsel->core.attr.precise_ip)
3027 return scnprintf(msg, size, "%s",
3028 "\'precise\' request may not be supported. Try removing 'p' modifier.");
3029#if defined(__i386__) || defined(__x86_64__)
3030 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3031 return scnprintf(msg, size, "%s",
3032 "No hardware sampling interrupt available.\n");
3033#endif
3034 break;
3035 case EBUSY:
3036 if (find_process("oprofiled"))
3037 return scnprintf(msg, size,
3038 "The PMU counters are busy/taken by another profiler.\n"
3039 "We found oprofile daemon running, please stop it and try again.");
3040 break;
3041 case EINVAL:
3042 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3043 return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3044 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3045 return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3046 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3047 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3048 if (perf_missing_features.clockid)
3049 return scnprintf(msg, size, "clockid feature not supported.");
3050 if (perf_missing_features.clockid_wrong)
3051 return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3052 if (perf_missing_features.aux_output)
3053 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3054 if (!target__has_cpu(target))
3055 return scnprintf(msg, size,
3056 "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3057 evsel__name(evsel));
3058 if (is_amd(arch, cpuid)) {
3059 if (is_amd_ibs(evsel)) {
3060 if (evsel->core.attr.exclude_kernel)
3061 return scnprintf(msg, size,
3062 "AMD IBS can't exclude kernel events. Try running at a higher privilege level.");
3063 if (!evsel->core.system_wide)
3064 return scnprintf(msg, size,
3065 "AMD IBS may only be available in system-wide/per-cpu mode. Try using -a, or -C and workload affinity");
3066 }
3067 }
3068
3069 break;
3070 case ENODATA:
3071 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3072 "Please add an auxiliary event in front of the load latency event.");
3073 default:
3074 break;
3075 }
3076
3077 return scnprintf(msg, size,
3078 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3079 "/bin/dmesg | grep -i perf may provide additional information.\n",
3080 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3081}
3082
3083struct perf_env *evsel__env(struct evsel *evsel)
3084{
3085 if (evsel && evsel->evlist && evsel->evlist->env)
3086 return evsel->evlist->env;
3087 return &perf_env;
3088}
3089
3090static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3091{
3092 int cpu_map_idx, thread;
3093
3094 for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3095 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3096 thread++) {
3097 int fd = FD(evsel, cpu_map_idx, thread);
3098
3099 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3100 cpu_map_idx, thread, fd) < 0)
3101 return -1;
3102 }
3103 }
3104
3105 return 0;
3106}
3107
3108int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3109{
3110 struct perf_cpu_map *cpus = evsel->core.cpus;
3111 struct perf_thread_map *threads = evsel->core.threads;
3112
3113 if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3114 return -ENOMEM;
3115
3116 return store_evsel_ids(evsel, evlist);
3117}
3118
3119void evsel__zero_per_pkg(struct evsel *evsel)
3120{
3121 struct hashmap_entry *cur;
3122 size_t bkt;
3123
3124 if (evsel->per_pkg_mask) {
3125 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3126 free((void *)cur->pkey);
3127
3128 hashmap__clear(evsel->per_pkg_mask);
3129 }
3130}
3131
3132bool evsel__is_hybrid(const struct evsel *evsel)
3133{
3134 return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3135}
3136
3137struct evsel *evsel__leader(struct evsel *evsel)
3138{
3139 return container_of(evsel->core.leader, struct evsel, core);
3140}
3141
3142bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3143{
3144 return evsel->core.leader == &leader->core;
3145}
3146
3147bool evsel__is_leader(struct evsel *evsel)
3148{
3149 return evsel__has_leader(evsel, evsel);
3150}
3151
3152void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3153{
3154 evsel->core.leader = &leader->core;
3155}
3156
3157int evsel__source_count(const struct evsel *evsel)
3158{
3159 struct evsel *pos;
3160 int count = 0;
3161
3162 evlist__for_each_entry(evsel->evlist, pos) {
3163 if (pos->metric_leader == evsel)
3164 count++;
3165 }
3166 return count;
3167}
3168
3169bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3170{
3171 return false;
3172}
3173
3174/*
3175 * Remove an event from a given group (leader).
3176 * Some events, e.g., perf metrics Topdown events,
3177 * must always be grouped. Ignore the events.
3178 */
3179void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3180{
3181 if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3182 evsel__set_leader(evsel, evsel);
3183 evsel->core.nr_members = 0;
3184 leader->core.nr_members--;
3185 }
3186}