11e754746984567285e31f27b91d7d601e63a455
[linux-2.6-microblaze.git] / tools / perf / util / evsel.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
4  *
5  * Parts came from builtin-{top,stat,record}.c, see those files for further
6  * copyright notes.
7  */
8
9 #include <byteswap.h>
10 #include <errno.h>
11 #include <inttypes.h>
12 #include <linux/bitops.h>
13 #include <api/fs/fs.h>
14 #include <api/fs/tracing_path.h>
15 #include <traceevent/event-parse.h>
16 #include <linux/hw_breakpoint.h>
17 #include <linux/perf_event.h>
18 #include <linux/compiler.h>
19 #include <linux/err.h>
20 #include <linux/zalloc.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
24 #include <dirent.h>
25 #include <stdlib.h>
26 #include <perf/evsel.h>
27 #include "asm/bug.h"
28 #include "callchain.h"
29 #include "cgroup.h"
30 #include "counts.h"
31 #include "event.h"
32 #include "evsel.h"
33 #include "util/env.h"
34 #include "util/evsel_config.h"
35 #include "util/evsel_fprintf.h"
36 #include "evlist.h"
37 #include <perf/cpumap.h>
38 #include "thread_map.h"
39 #include "target.h"
40 #include "perf_regs.h"
41 #include "record.h"
42 #include "debug.h"
43 #include "trace-event.h"
44 #include "stat.h"
45 #include "string2.h"
46 #include "memswap.h"
47 #include "util.h"
48 #include "../perf-sys.h"
49 #include "util/parse-branch-options.h"
50 #include <internal/xyarray.h>
51 #include <internal/lib.h>
52
53 #include <linux/ctype.h>
54
55 struct perf_missing_features perf_missing_features;
56
57 static clockid_t clockid;
58
59 static int perf_evsel__no_extra_init(struct evsel *evsel __maybe_unused)
60 {
61         return 0;
62 }
63
64 void __weak test_attr__ready(void) { }
65
66 static void perf_evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
67 {
68 }
69
70 static struct {
71         size_t  size;
72         int     (*init)(struct evsel *evsel);
73         void    (*fini)(struct evsel *evsel);
74 } perf_evsel__object = {
75         .size = sizeof(struct evsel),
76         .init = perf_evsel__no_extra_init,
77         .fini = perf_evsel__no_extra_fini,
78 };
79
80 int perf_evsel__object_config(size_t object_size,
81                               int (*init)(struct evsel *evsel),
82                               void (*fini)(struct evsel *evsel))
83 {
84
85         if (object_size == 0)
86                 goto set_methods;
87
88         if (perf_evsel__object.size > object_size)
89                 return -EINVAL;
90
91         perf_evsel__object.size = object_size;
92
93 set_methods:
94         if (init != NULL)
95                 perf_evsel__object.init = init;
96
97         if (fini != NULL)
98                 perf_evsel__object.fini = fini;
99
100         return 0;
101 }
102
103 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
104
105 int __perf_evsel__sample_size(u64 sample_type)
106 {
107         u64 mask = sample_type & PERF_SAMPLE_MASK;
108         int size = 0;
109         int i;
110
111         for (i = 0; i < 64; i++) {
112                 if (mask & (1ULL << i))
113                         size++;
114         }
115
116         size *= sizeof(u64);
117
118         return size;
119 }
120
121 /**
122  * __perf_evsel__calc_id_pos - calculate id_pos.
123  * @sample_type: sample type
124  *
125  * This function returns the position of the event id (PERF_SAMPLE_ID or
126  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
127  * perf_record_sample.
128  */
129 static int __perf_evsel__calc_id_pos(u64 sample_type)
130 {
131         int idx = 0;
132
133         if (sample_type & PERF_SAMPLE_IDENTIFIER)
134                 return 0;
135
136         if (!(sample_type & PERF_SAMPLE_ID))
137                 return -1;
138
139         if (sample_type & PERF_SAMPLE_IP)
140                 idx += 1;
141
142         if (sample_type & PERF_SAMPLE_TID)
143                 idx += 1;
144
145         if (sample_type & PERF_SAMPLE_TIME)
146                 idx += 1;
147
148         if (sample_type & PERF_SAMPLE_ADDR)
149                 idx += 1;
150
151         return idx;
152 }
153
154 /**
155  * __perf_evsel__calc_is_pos - calculate is_pos.
156  * @sample_type: sample type
157  *
158  * This function returns the position (counting backwards) of the event id
159  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
160  * sample_id_all is used there is an id sample appended to non-sample events.
161  */
162 static int __perf_evsel__calc_is_pos(u64 sample_type)
163 {
164         int idx = 1;
165
166         if (sample_type & PERF_SAMPLE_IDENTIFIER)
167                 return 1;
168
169         if (!(sample_type & PERF_SAMPLE_ID))
170                 return -1;
171
172         if (sample_type & PERF_SAMPLE_CPU)
173                 idx += 1;
174
175         if (sample_type & PERF_SAMPLE_STREAM_ID)
176                 idx += 1;
177
178         return idx;
179 }
180
181 void perf_evsel__calc_id_pos(struct evsel *evsel)
182 {
183         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
184         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
185 }
186
187 void __perf_evsel__set_sample_bit(struct evsel *evsel,
188                                   enum perf_event_sample_format bit)
189 {
190         if (!(evsel->core.attr.sample_type & bit)) {
191                 evsel->core.attr.sample_type |= bit;
192                 evsel->sample_size += sizeof(u64);
193                 perf_evsel__calc_id_pos(evsel);
194         }
195 }
196
197 void __perf_evsel__reset_sample_bit(struct evsel *evsel,
198                                     enum perf_event_sample_format bit)
199 {
200         if (evsel->core.attr.sample_type & bit) {
201                 evsel->core.attr.sample_type &= ~bit;
202                 evsel->sample_size -= sizeof(u64);
203                 perf_evsel__calc_id_pos(evsel);
204         }
205 }
206
207 void perf_evsel__set_sample_id(struct evsel *evsel,
208                                bool can_sample_identifier)
209 {
210         if (can_sample_identifier) {
211                 perf_evsel__reset_sample_bit(evsel, ID);
212                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
213         } else {
214                 perf_evsel__set_sample_bit(evsel, ID);
215         }
216         evsel->core.attr.read_format |= PERF_FORMAT_ID;
217 }
218
219 /**
220  * perf_evsel__is_function_event - Return whether given evsel is a function
221  * trace event
222  *
223  * @evsel - evsel selector to be tested
224  *
225  * Return %true if event is function trace event
226  */
227 bool perf_evsel__is_function_event(struct evsel *evsel)
228 {
229 #define FUNCTION_EVENT "ftrace:function"
230
231         return evsel->name &&
232                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
233
234 #undef FUNCTION_EVENT
235 }
236
237 void evsel__init(struct evsel *evsel,
238                  struct perf_event_attr *attr, int idx)
239 {
240         perf_evsel__init(&evsel->core, attr);
241         evsel->idx         = idx;
242         evsel->tracking    = !idx;
243         evsel->leader      = evsel;
244         evsel->unit        = "";
245         evsel->scale       = 1.0;
246         evsel->max_events  = ULONG_MAX;
247         evsel->evlist      = NULL;
248         evsel->bpf_obj     = NULL;
249         evsel->bpf_fd      = -1;
250         INIT_LIST_HEAD(&evsel->config_terms);
251         perf_evsel__object.init(evsel);
252         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
253         perf_evsel__calc_id_pos(evsel);
254         evsel->cmdline_group_boundary = false;
255         evsel->metric_expr   = NULL;
256         evsel->metric_name   = NULL;
257         evsel->metric_events = NULL;
258         evsel->collect_stat  = false;
259         evsel->pmu_name      = NULL;
260 }
261
262 struct evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
263 {
264         struct evsel *evsel = zalloc(perf_evsel__object.size);
265
266         if (!evsel)
267                 return NULL;
268         evsel__init(evsel, attr, idx);
269
270         if (perf_evsel__is_bpf_output(evsel)) {
271                 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
272                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
273                 evsel->core.attr.sample_period = 1;
274         }
275
276         if (perf_evsel__is_clock(evsel)) {
277                 /*
278                  * The evsel->unit points to static alias->unit
279                  * so it's ok to use static string in here.
280                  */
281                 static const char *unit = "msec";
282
283                 evsel->unit = unit;
284                 evsel->scale = 1e-6;
285         }
286
287         return evsel;
288 }
289
290 static bool perf_event_can_profile_kernel(void)
291 {
292         return perf_event_paranoid_check(1);
293 }
294
295 struct evsel *perf_evsel__new_cycles(bool precise)
296 {
297         struct perf_event_attr attr = {
298                 .type   = PERF_TYPE_HARDWARE,
299                 .config = PERF_COUNT_HW_CPU_CYCLES,
300                 .exclude_kernel = !perf_event_can_profile_kernel(),
301         };
302         struct evsel *evsel;
303
304         event_attr_init(&attr);
305
306         if (!precise)
307                 goto new_event;
308
309         /*
310          * Now let the usual logic to set up the perf_event_attr defaults
311          * to kick in when we return and before perf_evsel__open() is called.
312          */
313 new_event:
314         evsel = evsel__new(&attr);
315         if (evsel == NULL)
316                 goto out;
317
318         evsel->precise_max = true;
319
320         /* use asprintf() because free(evsel) assumes name is allocated */
321         if (asprintf(&evsel->name, "cycles%s%s%.*s",
322                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
323                      attr.exclude_kernel ? "u" : "",
324                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
325                 goto error_free;
326 out:
327         return evsel;
328 error_free:
329         evsel__delete(evsel);
330         evsel = NULL;
331         goto out;
332 }
333
334 /*
335  * Returns pointer with encoded error via <linux/err.h> interface.
336  */
337 struct evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
338 {
339         struct evsel *evsel = zalloc(perf_evsel__object.size);
340         int err = -ENOMEM;
341
342         if (evsel == NULL) {
343                 goto out_err;
344         } else {
345                 struct perf_event_attr attr = {
346                         .type          = PERF_TYPE_TRACEPOINT,
347                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
348                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
349                 };
350
351                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
352                         goto out_free;
353
354                 evsel->tp_format = trace_event__tp_format(sys, name);
355                 if (IS_ERR(evsel->tp_format)) {
356                         err = PTR_ERR(evsel->tp_format);
357                         goto out_free;
358                 }
359
360                 event_attr_init(&attr);
361                 attr.config = evsel->tp_format->id;
362                 attr.sample_period = 1;
363                 evsel__init(evsel, &attr, idx);
364         }
365
366         return evsel;
367
368 out_free:
369         zfree(&evsel->name);
370         free(evsel);
371 out_err:
372         return ERR_PTR(err);
373 }
374
375 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
376         "cycles",
377         "instructions",
378         "cache-references",
379         "cache-misses",
380         "branches",
381         "branch-misses",
382         "bus-cycles",
383         "stalled-cycles-frontend",
384         "stalled-cycles-backend",
385         "ref-cycles",
386 };
387
388 static const char *__perf_evsel__hw_name(u64 config)
389 {
390         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
391                 return perf_evsel__hw_names[config];
392
393         return "unknown-hardware";
394 }
395
396 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
397 {
398         int colon = 0, r = 0;
399         struct perf_event_attr *attr = &evsel->core.attr;
400         bool exclude_guest_default = false;
401
402 #define MOD_PRINT(context, mod) do {                                    \
403                 if (!attr->exclude_##context) {                         \
404                         if (!colon) colon = ++r;                        \
405                         r += scnprintf(bf + r, size - r, "%c", mod);    \
406                 } } while(0)
407
408         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
409                 MOD_PRINT(kernel, 'k');
410                 MOD_PRINT(user, 'u');
411                 MOD_PRINT(hv, 'h');
412                 exclude_guest_default = true;
413         }
414
415         if (attr->precise_ip) {
416                 if (!colon)
417                         colon = ++r;
418                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
419                 exclude_guest_default = true;
420         }
421
422         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
423                 MOD_PRINT(host, 'H');
424                 MOD_PRINT(guest, 'G');
425         }
426 #undef MOD_PRINT
427         if (colon)
428                 bf[colon - 1] = ':';
429         return r;
430 }
431
432 static int perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
433 {
434         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config));
435         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
436 }
437
438 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
439         "cpu-clock",
440         "task-clock",
441         "page-faults",
442         "context-switches",
443         "cpu-migrations",
444         "minor-faults",
445         "major-faults",
446         "alignment-faults",
447         "emulation-faults",
448         "dummy",
449 };
450
451 static const char *__perf_evsel__sw_name(u64 config)
452 {
453         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
454                 return perf_evsel__sw_names[config];
455         return "unknown-software";
456 }
457
458 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
459 {
460         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config));
461         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
462 }
463
464 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
465 {
466         int r;
467
468         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
469
470         if (type & HW_BREAKPOINT_R)
471                 r += scnprintf(bf + r, size - r, "r");
472
473         if (type & HW_BREAKPOINT_W)
474                 r += scnprintf(bf + r, size - r, "w");
475
476         if (type & HW_BREAKPOINT_X)
477                 r += scnprintf(bf + r, size - r, "x");
478
479         return r;
480 }
481
482 static int perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
483 {
484         struct perf_event_attr *attr = &evsel->core.attr;
485         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
486         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
487 }
488
489 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
490                                 [PERF_EVSEL__MAX_ALIASES] = {
491  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
492  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
493  { "LLC",       "L2",                                                   },
494  { "dTLB",      "d-tlb",        "Data-TLB",                             },
495  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
496  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
497  { "node",                                                              },
498 };
499
500 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
501                                    [PERF_EVSEL__MAX_ALIASES] = {
502  { "load",      "loads",        "read",                                 },
503  { "store",     "stores",       "write",                                },
504  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
505 };
506
507 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
508                                        [PERF_EVSEL__MAX_ALIASES] = {
509  { "refs",      "Reference",    "ops",          "access",               },
510  { "misses",    "miss",                                                 },
511 };
512
513 #define C(x)            PERF_COUNT_HW_CACHE_##x
514 #define CACHE_READ      (1 << C(OP_READ))
515 #define CACHE_WRITE     (1 << C(OP_WRITE))
516 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
517 #define COP(x)          (1 << x)
518
519 /*
520  * cache operartion stat
521  * L1I : Read and prefetch only
522  * ITLB and BPU : Read-only
523  */
524 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
525  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
526  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
527  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
528  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
529  [C(ITLB)]      = (CACHE_READ),
530  [C(BPU)]       = (CACHE_READ),
531  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
532 };
533
534 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
535 {
536         if (perf_evsel__hw_cache_stat[type] & COP(op))
537                 return true;    /* valid */
538         else
539                 return false;   /* invalid */
540 }
541
542 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
543                                             char *bf, size_t size)
544 {
545         if (result) {
546                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
547                                  perf_evsel__hw_cache_op[op][0],
548                                  perf_evsel__hw_cache_result[result][0]);
549         }
550
551         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
552                          perf_evsel__hw_cache_op[op][1]);
553 }
554
555 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
556 {
557         u8 op, result, type = (config >>  0) & 0xff;
558         const char *err = "unknown-ext-hardware-cache-type";
559
560         if (type >= PERF_COUNT_HW_CACHE_MAX)
561                 goto out_err;
562
563         op = (config >>  8) & 0xff;
564         err = "unknown-ext-hardware-cache-op";
565         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
566                 goto out_err;
567
568         result = (config >> 16) & 0xff;
569         err = "unknown-ext-hardware-cache-result";
570         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
571                 goto out_err;
572
573         err = "invalid-cache";
574         if (!perf_evsel__is_cache_op_valid(type, op))
575                 goto out_err;
576
577         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
578 out_err:
579         return scnprintf(bf, size, "%s", err);
580 }
581
582 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
583 {
584         int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size);
585         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
586 }
587
588 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
589 {
590         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
591         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
592 }
593
594 static int perf_evsel__tool_name(char *bf, size_t size)
595 {
596         int ret = scnprintf(bf, size, "duration_time");
597         return ret;
598 }
599
600 const char *perf_evsel__name(struct evsel *evsel)
601 {
602         char bf[128];
603
604         if (!evsel)
605                 goto out_unknown;
606
607         if (evsel->name)
608                 return evsel->name;
609
610         switch (evsel->core.attr.type) {
611         case PERF_TYPE_RAW:
612                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
613                 break;
614
615         case PERF_TYPE_HARDWARE:
616                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
617                 break;
618
619         case PERF_TYPE_HW_CACHE:
620                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
621                 break;
622
623         case PERF_TYPE_SOFTWARE:
624                 if (evsel->tool_event)
625                         perf_evsel__tool_name(bf, sizeof(bf));
626                 else
627                         perf_evsel__sw_name(evsel, bf, sizeof(bf));
628                 break;
629
630         case PERF_TYPE_TRACEPOINT:
631                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
632                 break;
633
634         case PERF_TYPE_BREAKPOINT:
635                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
636                 break;
637
638         default:
639                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
640                           evsel->core.attr.type);
641                 break;
642         }
643
644         evsel->name = strdup(bf);
645
646         if (evsel->name)
647                 return evsel->name;
648 out_unknown:
649         return "unknown";
650 }
651
652 const char *perf_evsel__group_name(struct evsel *evsel)
653 {
654         return evsel->group_name ?: "anon group";
655 }
656
657 /*
658  * Returns the group details for the specified leader,
659  * with following rules.
660  *
661  *  For record -e '{cycles,instructions}'
662  *    'anon group { cycles:u, instructions:u }'
663  *
664  *  For record -e 'cycles,instructions' and report --group
665  *    'cycles:u, instructions:u'
666  */
667 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
668 {
669         int ret = 0;
670         struct evsel *pos;
671         const char *group_name = perf_evsel__group_name(evsel);
672
673         if (!evsel->forced_leader)
674                 ret = scnprintf(buf, size, "%s { ", group_name);
675
676         ret += scnprintf(buf + ret, size - ret, "%s",
677                          perf_evsel__name(evsel));
678
679         for_each_group_member(pos, evsel)
680                 ret += scnprintf(buf + ret, size - ret, ", %s",
681                                  perf_evsel__name(pos));
682
683         if (!evsel->forced_leader)
684                 ret += scnprintf(buf + ret, size - ret, " }");
685
686         return ret;
687 }
688
689 static void __perf_evsel__config_callchain(struct evsel *evsel,
690                                            struct record_opts *opts,
691                                            struct callchain_param *param)
692 {
693         bool function = perf_evsel__is_function_event(evsel);
694         struct perf_event_attr *attr = &evsel->core.attr;
695
696         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
697
698         attr->sample_max_stack = param->max_stack;
699
700         if (opts->kernel_callchains)
701                 attr->exclude_callchain_user = 1;
702         if (opts->user_callchains)
703                 attr->exclude_callchain_kernel = 1;
704         if (param->record_mode == CALLCHAIN_LBR) {
705                 if (!opts->branch_stack) {
706                         if (attr->exclude_user) {
707                                 pr_warning("LBR callstack option is only available "
708                                            "to get user callchain information. "
709                                            "Falling back to framepointers.\n");
710                         } else {
711                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
712                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
713                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
714                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
715                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
716                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
717                         }
718                 } else
719                          pr_warning("Cannot use LBR callstack with branch stack. "
720                                     "Falling back to framepointers.\n");
721         }
722
723         if (param->record_mode == CALLCHAIN_DWARF) {
724                 if (!function) {
725                         perf_evsel__set_sample_bit(evsel, REGS_USER);
726                         perf_evsel__set_sample_bit(evsel, STACK_USER);
727                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
728                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
729                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
730                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
731                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
732                         } else {
733                                 attr->sample_regs_user |= PERF_REGS_MASK;
734                         }
735                         attr->sample_stack_user = param->dump_size;
736                         attr->exclude_callchain_user = 1;
737                 } else {
738                         pr_info("Cannot use DWARF unwind for function trace event,"
739                                 " falling back to framepointers.\n");
740                 }
741         }
742
743         if (function) {
744                 pr_info("Disabling user space callchains for function trace event.\n");
745                 attr->exclude_callchain_user = 1;
746         }
747 }
748
749 void perf_evsel__config_callchain(struct evsel *evsel,
750                                   struct record_opts *opts,
751                                   struct callchain_param *param)
752 {
753         if (param->enabled)
754                 return __perf_evsel__config_callchain(evsel, opts, param);
755 }
756
757 static void
758 perf_evsel__reset_callgraph(struct evsel *evsel,
759                             struct callchain_param *param)
760 {
761         struct perf_event_attr *attr = &evsel->core.attr;
762
763         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
764         if (param->record_mode == CALLCHAIN_LBR) {
765                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
766                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
767                                               PERF_SAMPLE_BRANCH_CALL_STACK |
768                                               PERF_SAMPLE_BRANCH_HW_INDEX);
769         }
770         if (param->record_mode == CALLCHAIN_DWARF) {
771                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
772                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
773         }
774 }
775
776 static void apply_config_terms(struct evsel *evsel,
777                                struct record_opts *opts, bool track)
778 {
779         struct perf_evsel_config_term *term;
780         struct list_head *config_terms = &evsel->config_terms;
781         struct perf_event_attr *attr = &evsel->core.attr;
782         /* callgraph default */
783         struct callchain_param param = {
784                 .record_mode = callchain_param.record_mode,
785         };
786         u32 dump_size = 0;
787         int max_stack = 0;
788         const char *callgraph_buf = NULL;
789
790         list_for_each_entry(term, config_terms, list) {
791                 switch (term->type) {
792                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
793                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
794                                 attr->sample_period = term->val.period;
795                                 attr->freq = 0;
796                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
797                         }
798                         break;
799                 case PERF_EVSEL__CONFIG_TERM_FREQ:
800                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
801                                 attr->sample_freq = term->val.freq;
802                                 attr->freq = 1;
803                                 perf_evsel__set_sample_bit(evsel, PERIOD);
804                         }
805                         break;
806                 case PERF_EVSEL__CONFIG_TERM_TIME:
807                         if (term->val.time)
808                                 perf_evsel__set_sample_bit(evsel, TIME);
809                         else
810                                 perf_evsel__reset_sample_bit(evsel, TIME);
811                         break;
812                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
813                         callgraph_buf = term->val.str;
814                         break;
815                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
816                         if (term->val.str && strcmp(term->val.str, "no")) {
817                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
818                                 parse_branch_str(term->val.str,
819                                                  &attr->branch_sample_type);
820                         } else
821                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
822                         break;
823                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
824                         dump_size = term->val.stack_user;
825                         break;
826                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
827                         max_stack = term->val.max_stack;
828                         break;
829                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
830                         evsel->max_events = term->val.max_events;
831                         break;
832                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
833                         /*
834                          * attr->inherit should has already been set by
835                          * perf_evsel__config. If user explicitly set
836                          * inherit using config terms, override global
837                          * opt->no_inherit setting.
838                          */
839                         attr->inherit = term->val.inherit ? 1 : 0;
840                         break;
841                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
842                         attr->write_backward = term->val.overwrite ? 1 : 0;
843                         break;
844                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
845                         break;
846                 case PERF_EVSEL__CONFIG_TERM_PERCORE:
847                         break;
848                 case PERF_EVSEL__CONFIG_TERM_AUX_OUTPUT:
849                         attr->aux_output = term->val.aux_output ? 1 : 0;
850                         break;
851                 case PERF_EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
852                         /* Already applied by auxtrace */
853                         break;
854                 case PERF_EVSEL__CONFIG_TERM_CFG_CHG:
855                         break;
856                 default:
857                         break;
858                 }
859         }
860
861         /* User explicitly set per-event callgraph, clear the old setting and reset. */
862         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
863                 bool sample_address = false;
864
865                 if (max_stack) {
866                         param.max_stack = max_stack;
867                         if (callgraph_buf == NULL)
868                                 callgraph_buf = "fp";
869                 }
870
871                 /* parse callgraph parameters */
872                 if (callgraph_buf != NULL) {
873                         if (!strcmp(callgraph_buf, "no")) {
874                                 param.enabled = false;
875                                 param.record_mode = CALLCHAIN_NONE;
876                         } else {
877                                 param.enabled = true;
878                                 if (parse_callchain_record(callgraph_buf, &param)) {
879                                         pr_err("per-event callgraph setting for %s failed. "
880                                                "Apply callgraph global setting for it\n",
881                                                evsel->name);
882                                         return;
883                                 }
884                                 if (param.record_mode == CALLCHAIN_DWARF)
885                                         sample_address = true;
886                         }
887                 }
888                 if (dump_size > 0) {
889                         dump_size = round_up(dump_size, sizeof(u64));
890                         param.dump_size = dump_size;
891                 }
892
893                 /* If global callgraph set, clear it */
894                 if (callchain_param.enabled)
895                         perf_evsel__reset_callgraph(evsel, &callchain_param);
896
897                 /* set perf-event callgraph */
898                 if (param.enabled) {
899                         if (sample_address) {
900                                 perf_evsel__set_sample_bit(evsel, ADDR);
901                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
902                                 evsel->core.attr.mmap_data = track;
903                         }
904                         perf_evsel__config_callchain(evsel, opts, &param);
905                 }
906         }
907 }
908
909 static bool is_dummy_event(struct evsel *evsel)
910 {
911         return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
912                (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
913 }
914
915 struct perf_evsel_config_term *__perf_evsel__get_config_term(struct evsel *evsel,
916                                                              enum evsel_term_type type)
917 {
918         struct perf_evsel_config_term *term, *found_term = NULL;
919
920         list_for_each_entry(term, &evsel->config_terms, list) {
921                 if (term->type == type)
922                         found_term = term;
923         }
924
925         return found_term;
926 }
927
928 /*
929  * The enable_on_exec/disabled value strategy:
930  *
931  *  1) For any type of traced program:
932  *    - all independent events and group leaders are disabled
933  *    - all group members are enabled
934  *
935  *     Group members are ruled by group leaders. They need to
936  *     be enabled, because the group scheduling relies on that.
937  *
938  *  2) For traced programs executed by perf:
939  *     - all independent events and group leaders have
940  *       enable_on_exec set
941  *     - we don't specifically enable or disable any event during
942  *       the record command
943  *
944  *     Independent events and group leaders are initially disabled
945  *     and get enabled by exec. Group members are ruled by group
946  *     leaders as stated in 1).
947  *
948  *  3) For traced programs attached by perf (pid/tid):
949  *     - we specifically enable or disable all events during
950  *       the record command
951  *
952  *     When attaching events to already running traced we
953  *     enable/disable events specifically, as there's no
954  *     initial traced exec call.
955  */
956 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
957                         struct callchain_param *callchain)
958 {
959         struct evsel *leader = evsel->leader;
960         struct perf_event_attr *attr = &evsel->core.attr;
961         int track = evsel->tracking;
962         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
963
964         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
965         attr->inherit       = !opts->no_inherit;
966         attr->write_backward = opts->overwrite ? 1 : 0;
967
968         perf_evsel__set_sample_bit(evsel, IP);
969         perf_evsel__set_sample_bit(evsel, TID);
970
971         if (evsel->sample_read) {
972                 perf_evsel__set_sample_bit(evsel, READ);
973
974                 /*
975                  * We need ID even in case of single event, because
976                  * PERF_SAMPLE_READ process ID specific data.
977                  */
978                 perf_evsel__set_sample_id(evsel, false);
979
980                 /*
981                  * Apply group format only if we belong to group
982                  * with more than one members.
983                  */
984                 if (leader->core.nr_members > 1) {
985                         attr->read_format |= PERF_FORMAT_GROUP;
986                         attr->inherit = 0;
987                 }
988         }
989
990         /*
991          * We default some events to have a default interval. But keep
992          * it a weak assumption overridable by the user.
993          */
994         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
995                                      opts->user_interval != ULLONG_MAX)) {
996                 if (opts->freq) {
997                         perf_evsel__set_sample_bit(evsel, PERIOD);
998                         attr->freq              = 1;
999                         attr->sample_freq       = opts->freq;
1000                 } else {
1001                         attr->sample_period = opts->default_interval;
1002                 }
1003         }
1004
1005         if (opts->no_samples)
1006                 attr->sample_freq = 0;
1007
1008         if (opts->inherit_stat) {
1009                 evsel->core.attr.read_format |=
1010                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1011                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1012                         PERF_FORMAT_ID;
1013                 attr->inherit_stat = 1;
1014         }
1015
1016         if (opts->sample_address) {
1017                 perf_evsel__set_sample_bit(evsel, ADDR);
1018                 attr->mmap_data = track;
1019         }
1020
1021         /*
1022          * We don't allow user space callchains for  function trace
1023          * event, due to issues with page faults while tracing page
1024          * fault handler and its overall trickiness nature.
1025          */
1026         if (perf_evsel__is_function_event(evsel))
1027                 evsel->core.attr.exclude_callchain_user = 1;
1028
1029         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1030                 perf_evsel__config_callchain(evsel, opts, callchain);
1031
1032         if (opts->sample_intr_regs) {
1033                 attr->sample_regs_intr = opts->sample_intr_regs;
1034                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1035         }
1036
1037         if (opts->sample_user_regs) {
1038                 attr->sample_regs_user |= opts->sample_user_regs;
1039                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1040         }
1041
1042         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1043                 perf_evsel__set_sample_bit(evsel, CPU);
1044
1045         /*
1046          * When the user explicitly disabled time don't force it here.
1047          */
1048         if (opts->sample_time &&
1049             (!perf_missing_features.sample_id_all &&
1050             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1051              opts->sample_time_set)))
1052                 perf_evsel__set_sample_bit(evsel, TIME);
1053
1054         if (opts->raw_samples && !evsel->no_aux_samples) {
1055                 perf_evsel__set_sample_bit(evsel, TIME);
1056                 perf_evsel__set_sample_bit(evsel, RAW);
1057                 perf_evsel__set_sample_bit(evsel, CPU);
1058         }
1059
1060         if (opts->sample_address)
1061                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1062
1063         if (opts->sample_phys_addr)
1064                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1065
1066         if (opts->no_buffering) {
1067                 attr->watermark = 0;
1068                 attr->wakeup_events = 1;
1069         }
1070         if (opts->branch_stack && !evsel->no_aux_samples) {
1071                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1072                 attr->branch_sample_type = opts->branch_stack;
1073         }
1074
1075         if (opts->sample_weight)
1076                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1077
1078         attr->task  = track;
1079         attr->mmap  = track;
1080         attr->mmap2 = track && !perf_missing_features.mmap2;
1081         attr->comm  = track;
1082         attr->ksymbol = track && !perf_missing_features.ksymbol;
1083         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1084
1085         if (opts->record_namespaces)
1086                 attr->namespaces  = track;
1087
1088         if (opts->record_cgroup) {
1089                 attr->cgroup = track && !perf_missing_features.cgroup;
1090                 perf_evsel__set_sample_bit(evsel, CGROUP);
1091         }
1092
1093         if (opts->record_switch_events)
1094                 attr->context_switch = track;
1095
1096         if (opts->sample_transaction)
1097                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1098
1099         if (opts->running_time) {
1100                 evsel->core.attr.read_format |=
1101                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1102                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1103         }
1104
1105         /*
1106          * XXX see the function comment above
1107          *
1108          * Disabling only independent events or group leaders,
1109          * keeping group members enabled.
1110          */
1111         if (perf_evsel__is_group_leader(evsel))
1112                 attr->disabled = 1;
1113
1114         /*
1115          * Setting enable_on_exec for independent events and
1116          * group leaders for traced executed by perf.
1117          */
1118         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1119                 !opts->initial_delay)
1120                 attr->enable_on_exec = 1;
1121
1122         if (evsel->immediate) {
1123                 attr->disabled = 0;
1124                 attr->enable_on_exec = 0;
1125         }
1126
1127         clockid = opts->clockid;
1128         if (opts->use_clockid) {
1129                 attr->use_clockid = 1;
1130                 attr->clockid = opts->clockid;
1131         }
1132
1133         if (evsel->precise_max)
1134                 attr->precise_ip = 3;
1135
1136         if (opts->all_user) {
1137                 attr->exclude_kernel = 1;
1138                 attr->exclude_user   = 0;
1139         }
1140
1141         if (opts->all_kernel) {
1142                 attr->exclude_kernel = 0;
1143                 attr->exclude_user   = 1;
1144         }
1145
1146         if (evsel->core.own_cpus || evsel->unit)
1147                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1148
1149         /*
1150          * Apply event specific term settings,
1151          * it overloads any global configuration.
1152          */
1153         apply_config_terms(evsel, opts, track);
1154
1155         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1156
1157         /* The --period option takes the precedence. */
1158         if (opts->period_set) {
1159                 if (opts->period)
1160                         perf_evsel__set_sample_bit(evsel, PERIOD);
1161                 else
1162                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1163         }
1164
1165         /*
1166          * For initial_delay, a dummy event is added implicitly.
1167          * The software event will trigger -EOPNOTSUPP error out,
1168          * if BRANCH_STACK bit is set.
1169          */
1170         if (opts->initial_delay && is_dummy_event(evsel))
1171                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1172 }
1173
1174 int perf_evsel__set_filter(struct evsel *evsel, const char *filter)
1175 {
1176         char *new_filter = strdup(filter);
1177
1178         if (new_filter != NULL) {
1179                 free(evsel->filter);
1180                 evsel->filter = new_filter;
1181                 return 0;
1182         }
1183
1184         return -1;
1185 }
1186
1187 static int perf_evsel__append_filter(struct evsel *evsel,
1188                                      const char *fmt, const char *filter)
1189 {
1190         char *new_filter;
1191
1192         if (evsel->filter == NULL)
1193                 return perf_evsel__set_filter(evsel, filter);
1194
1195         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1196                 free(evsel->filter);
1197                 evsel->filter = new_filter;
1198                 return 0;
1199         }
1200
1201         return -1;
1202 }
1203
1204 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1205 {
1206         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1207 }
1208
1209 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1210 {
1211         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1212 }
1213
1214 /* Caller has to clear disabled after going through all CPUs. */
1215 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1216 {
1217         return perf_evsel__enable_cpu(&evsel->core, cpu);
1218 }
1219
1220 int evsel__enable(struct evsel *evsel)
1221 {
1222         int err = perf_evsel__enable(&evsel->core);
1223
1224         if (!err)
1225                 evsel->disabled = false;
1226         return err;
1227 }
1228
1229 /* Caller has to set disabled after going through all CPUs. */
1230 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1231 {
1232         return perf_evsel__disable_cpu(&evsel->core, cpu);
1233 }
1234
1235 int evsel__disable(struct evsel *evsel)
1236 {
1237         int err = perf_evsel__disable(&evsel->core);
1238         /*
1239          * We mark it disabled here so that tools that disable a event can
1240          * ignore events after they disable it. I.e. the ring buffer may have
1241          * already a few more events queued up before the kernel got the stop
1242          * request.
1243          */
1244         if (!err)
1245                 evsel->disabled = true;
1246
1247         return err;
1248 }
1249
1250 static void perf_evsel__free_config_terms(struct evsel *evsel)
1251 {
1252         struct perf_evsel_config_term *term, *h;
1253
1254         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1255                 list_del_init(&term->list);
1256                 if (term->free_str)
1257                         zfree(&term->val.str);
1258                 free(term);
1259         }
1260 }
1261
1262 void evsel__exit(struct evsel *evsel)
1263 {
1264         assert(list_empty(&evsel->core.node));
1265         assert(evsel->evlist == NULL);
1266         perf_evsel__free_counts(evsel);
1267         perf_evsel__free_fd(&evsel->core);
1268         perf_evsel__free_id(&evsel->core);
1269         perf_evsel__free_config_terms(evsel);
1270         cgroup__put(evsel->cgrp);
1271         perf_cpu_map__put(evsel->core.cpus);
1272         perf_cpu_map__put(evsel->core.own_cpus);
1273         perf_thread_map__put(evsel->core.threads);
1274         zfree(&evsel->group_name);
1275         zfree(&evsel->name);
1276         zfree(&evsel->pmu_name);
1277         perf_evsel__object.fini(evsel);
1278 }
1279
1280 void evsel__delete(struct evsel *evsel)
1281 {
1282         evsel__exit(evsel);
1283         free(evsel);
1284 }
1285
1286 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1287                            struct perf_counts_values *count)
1288 {
1289         struct perf_counts_values tmp;
1290
1291         if (!evsel->prev_raw_counts)
1292                 return;
1293
1294         if (cpu == -1) {
1295                 tmp = evsel->prev_raw_counts->aggr;
1296                 evsel->prev_raw_counts->aggr = *count;
1297         } else {
1298                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1299                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1300         }
1301
1302         count->val = count->val - tmp.val;
1303         count->ena = count->ena - tmp.ena;
1304         count->run = count->run - tmp.run;
1305 }
1306
1307 void perf_counts_values__scale(struct perf_counts_values *count,
1308                                bool scale, s8 *pscaled)
1309 {
1310         s8 scaled = 0;
1311
1312         if (scale) {
1313                 if (count->run == 0) {
1314                         scaled = -1;
1315                         count->val = 0;
1316                 } else if (count->run < count->ena) {
1317                         scaled = 1;
1318                         count->val = (u64)((double) count->val * count->ena / count->run);
1319                 }
1320         }
1321
1322         if (pscaled)
1323                 *pscaled = scaled;
1324 }
1325
1326 static int
1327 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread)
1328 {
1329         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1330
1331         return perf_evsel__read(&evsel->core, cpu, thread, count);
1332 }
1333
1334 static void
1335 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1336                       u64 val, u64 ena, u64 run)
1337 {
1338         struct perf_counts_values *count;
1339
1340         count = perf_counts(counter->counts, cpu, thread);
1341
1342         count->val    = val;
1343         count->ena    = ena;
1344         count->run    = run;
1345
1346         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1347 }
1348
1349 static int
1350 perf_evsel__process_group_data(struct evsel *leader,
1351                                int cpu, int thread, u64 *data)
1352 {
1353         u64 read_format = leader->core.attr.read_format;
1354         struct sample_read_value *v;
1355         u64 nr, ena = 0, run = 0, i;
1356
1357         nr = *data++;
1358
1359         if (nr != (u64) leader->core.nr_members)
1360                 return -EINVAL;
1361
1362         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1363                 ena = *data++;
1364
1365         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1366                 run = *data++;
1367
1368         v = (struct sample_read_value *) data;
1369
1370         perf_evsel__set_count(leader, cpu, thread,
1371                               v[0].value, ena, run);
1372
1373         for (i = 1; i < nr; i++) {
1374                 struct evsel *counter;
1375
1376                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1377                 if (!counter)
1378                         return -EINVAL;
1379
1380                 perf_evsel__set_count(counter, cpu, thread,
1381                                       v[i].value, ena, run);
1382         }
1383
1384         return 0;
1385 }
1386
1387 static int
1388 perf_evsel__read_group(struct evsel *leader, int cpu, int thread)
1389 {
1390         struct perf_stat_evsel *ps = leader->stats;
1391         u64 read_format = leader->core.attr.read_format;
1392         int size = perf_evsel__read_size(&leader->core);
1393         u64 *data = ps->group_data;
1394
1395         if (!(read_format & PERF_FORMAT_ID))
1396                 return -EINVAL;
1397
1398         if (!perf_evsel__is_group_leader(leader))
1399                 return -EINVAL;
1400
1401         if (!data) {
1402                 data = zalloc(size);
1403                 if (!data)
1404                         return -ENOMEM;
1405
1406                 ps->group_data = data;
1407         }
1408
1409         if (FD(leader, cpu, thread) < 0)
1410                 return -EINVAL;
1411
1412         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1413                 return -errno;
1414
1415         return perf_evsel__process_group_data(leader, cpu, thread, data);
1416 }
1417
1418 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1419 {
1420         u64 read_format = evsel->core.attr.read_format;
1421
1422         if (read_format & PERF_FORMAT_GROUP)
1423                 return perf_evsel__read_group(evsel, cpu, thread);
1424         else
1425                 return perf_evsel__read_one(evsel, cpu, thread);
1426 }
1427
1428 int __perf_evsel__read_on_cpu(struct evsel *evsel,
1429                               int cpu, int thread, bool scale)
1430 {
1431         struct perf_counts_values count;
1432         size_t nv = scale ? 3 : 1;
1433
1434         if (FD(evsel, cpu, thread) < 0)
1435                 return -EINVAL;
1436
1437         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1438                 return -ENOMEM;
1439
1440         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1441                 return -errno;
1442
1443         evsel__compute_deltas(evsel, cpu, thread, &count);
1444         perf_counts_values__scale(&count, scale, NULL);
1445         *perf_counts(evsel->counts, cpu, thread) = count;
1446         return 0;
1447 }
1448
1449 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1450 {
1451         struct evsel *leader = evsel->leader;
1452         int fd;
1453
1454         if (perf_evsel__is_group_leader(evsel))
1455                 return -1;
1456
1457         /*
1458          * Leader must be already processed/open,
1459          * if not it's a bug.
1460          */
1461         BUG_ON(!leader->core.fd);
1462
1463         fd = FD(leader, cpu, thread);
1464         BUG_ON(fd == -1);
1465
1466         return fd;
1467 }
1468
1469 static void perf_evsel__remove_fd(struct evsel *pos,
1470                                   int nr_cpus, int nr_threads,
1471                                   int thread_idx)
1472 {
1473         for (int cpu = 0; cpu < nr_cpus; cpu++)
1474                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1475                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1476 }
1477
1478 static int update_fds(struct evsel *evsel,
1479                       int nr_cpus, int cpu_idx,
1480                       int nr_threads, int thread_idx)
1481 {
1482         struct evsel *pos;
1483
1484         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1485                 return -EINVAL;
1486
1487         evlist__for_each_entry(evsel->evlist, pos) {
1488                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1489
1490                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1491
1492                 /*
1493                  * Since fds for next evsel has not been created,
1494                  * there is no need to iterate whole event list.
1495                  */
1496                 if (pos == evsel)
1497                         break;
1498         }
1499         return 0;
1500 }
1501
1502 static bool ignore_missing_thread(struct evsel *evsel,
1503                                   int nr_cpus, int cpu,
1504                                   struct perf_thread_map *threads,
1505                                   int thread, int err)
1506 {
1507         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1508
1509         if (!evsel->ignore_missing_thread)
1510                 return false;
1511
1512         /* The system wide setup does not work with threads. */
1513         if (evsel->core.system_wide)
1514                 return false;
1515
1516         /* The -ESRCH is perf event syscall errno for pid's not found. */
1517         if (err != -ESRCH)
1518                 return false;
1519
1520         /* If there's only one thread, let it fail. */
1521         if (threads->nr == 1)
1522                 return false;
1523
1524         /*
1525          * We should remove fd for missing_thread first
1526          * because thread_map__remove() will decrease threads->nr.
1527          */
1528         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1529                 return false;
1530
1531         if (thread_map__remove(threads, thread))
1532                 return false;
1533
1534         pr_warning("WARNING: Ignored open failure for pid %d\n",
1535                    ignore_pid);
1536         return true;
1537 }
1538
1539 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1540                                 void *priv __maybe_unused)
1541 {
1542         return fprintf(fp, "  %-32s %s\n", name, val);
1543 }
1544
1545 static void display_attr(struct perf_event_attr *attr)
1546 {
1547         if (verbose >= 2 || debug_peo_args) {
1548                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1549                 fprintf(stderr, "perf_event_attr:\n");
1550                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1551                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1552         }
1553 }
1554
1555 static int perf_event_open(struct evsel *evsel,
1556                            pid_t pid, int cpu, int group_fd,
1557                            unsigned long flags)
1558 {
1559         int precise_ip = evsel->core.attr.precise_ip;
1560         int fd;
1561
1562         while (1) {
1563                 pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1564                           pid, cpu, group_fd, flags);
1565
1566                 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1567                 if (fd >= 0)
1568                         break;
1569
1570                 /* Do not try less precise if not requested. */
1571                 if (!evsel->precise_max)
1572                         break;
1573
1574                 /*
1575                  * We tried all the precise_ip values, and it's
1576                  * still failing, so leave it to standard fallback.
1577                  */
1578                 if (!evsel->core.attr.precise_ip) {
1579                         evsel->core.attr.precise_ip = precise_ip;
1580                         break;
1581                 }
1582
1583                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1584                 evsel->core.attr.precise_ip--;
1585                 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1586                 display_attr(&evsel->core.attr);
1587         }
1588
1589         return fd;
1590 }
1591
1592 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1593                 struct perf_thread_map *threads,
1594                 int start_cpu, int end_cpu)
1595 {
1596         int cpu, thread, nthreads;
1597         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1598         int pid = -1, err, old_errno;
1599         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1600
1601         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1602             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1603                 return -EINVAL;
1604
1605         if (cpus == NULL) {
1606                 static struct perf_cpu_map *empty_cpu_map;
1607
1608                 if (empty_cpu_map == NULL) {
1609                         empty_cpu_map = perf_cpu_map__dummy_new();
1610                         if (empty_cpu_map == NULL)
1611                                 return -ENOMEM;
1612                 }
1613
1614                 cpus = empty_cpu_map;
1615         }
1616
1617         if (threads == NULL) {
1618                 static struct perf_thread_map *empty_thread_map;
1619
1620                 if (empty_thread_map == NULL) {
1621                         empty_thread_map = thread_map__new_by_tid(-1);
1622                         if (empty_thread_map == NULL)
1623                                 return -ENOMEM;
1624                 }
1625
1626                 threads = empty_thread_map;
1627         }
1628
1629         if (evsel->core.system_wide)
1630                 nthreads = 1;
1631         else
1632                 nthreads = threads->nr;
1633
1634         if (evsel->core.fd == NULL &&
1635             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1636                 return -ENOMEM;
1637
1638         if (evsel->cgrp) {
1639                 flags |= PERF_FLAG_PID_CGROUP;
1640                 pid = evsel->cgrp->fd;
1641         }
1642
1643 fallback_missing_features:
1644         if (perf_missing_features.clockid_wrong)
1645                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1646         if (perf_missing_features.clockid) {
1647                 evsel->core.attr.use_clockid = 0;
1648                 evsel->core.attr.clockid = 0;
1649         }
1650         if (perf_missing_features.cloexec)
1651                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1652         if (perf_missing_features.mmap2)
1653                 evsel->core.attr.mmap2 = 0;
1654         if (perf_missing_features.exclude_guest)
1655                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1656         if (perf_missing_features.lbr_flags)
1657                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1658                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1659         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1660                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1661         if (perf_missing_features.ksymbol)
1662                 evsel->core.attr.ksymbol = 0;
1663         if (perf_missing_features.bpf)
1664                 evsel->core.attr.bpf_event = 0;
1665         if (perf_missing_features.branch_hw_idx)
1666                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1667 retry_sample_id:
1668         if (perf_missing_features.sample_id_all)
1669                 evsel->core.attr.sample_id_all = 0;
1670
1671         display_attr(&evsel->core.attr);
1672
1673         for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1674
1675                 for (thread = 0; thread < nthreads; thread++) {
1676                         int fd, group_fd;
1677
1678                         if (!evsel->cgrp && !evsel->core.system_wide)
1679                                 pid = perf_thread_map__pid(threads, thread);
1680
1681                         group_fd = get_group_fd(evsel, cpu, thread);
1682 retry_open:
1683                         test_attr__ready();
1684
1685                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1686                                              group_fd, flags);
1687
1688                         FD(evsel, cpu, thread) = fd;
1689
1690                         if (fd < 0) {
1691                                 err = -errno;
1692
1693                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1694                                         /*
1695                                          * We just removed 1 thread, so take a step
1696                                          * back on thread index and lower the upper
1697                                          * nthreads limit.
1698                                          */
1699                                         nthreads--;
1700                                         thread--;
1701
1702                                         /* ... and pretend like nothing have happened. */
1703                                         err = 0;
1704                                         continue;
1705                                 }
1706
1707                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1708                                           err);
1709                                 goto try_fallback;
1710                         }
1711
1712                         pr_debug2_peo(" = %d\n", fd);
1713
1714                         if (evsel->bpf_fd >= 0) {
1715                                 int evt_fd = fd;
1716                                 int bpf_fd = evsel->bpf_fd;
1717
1718                                 err = ioctl(evt_fd,
1719                                             PERF_EVENT_IOC_SET_BPF,
1720                                             bpf_fd);
1721                                 if (err && errno != EEXIST) {
1722                                         pr_err("failed to attach bpf fd %d: %s\n",
1723                                                bpf_fd, strerror(errno));
1724                                         err = -EINVAL;
1725                                         goto out_close;
1726                                 }
1727                         }
1728
1729                         set_rlimit = NO_CHANGE;
1730
1731                         /*
1732                          * If we succeeded but had to kill clockid, fail and
1733                          * have perf_evsel__open_strerror() print us a nice
1734                          * error.
1735                          */
1736                         if (perf_missing_features.clockid ||
1737                             perf_missing_features.clockid_wrong) {
1738                                 err = -EINVAL;
1739                                 goto out_close;
1740                         }
1741                 }
1742         }
1743
1744         return 0;
1745
1746 try_fallback:
1747         /*
1748          * perf stat needs between 5 and 22 fds per CPU. When we run out
1749          * of them try to increase the limits.
1750          */
1751         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1752                 struct rlimit l;
1753
1754                 old_errno = errno;
1755                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1756                         if (set_rlimit == NO_CHANGE)
1757                                 l.rlim_cur = l.rlim_max;
1758                         else {
1759                                 l.rlim_cur = l.rlim_max + 1000;
1760                                 l.rlim_max = l.rlim_cur;
1761                         }
1762                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1763                                 set_rlimit++;
1764                                 errno = old_errno;
1765                                 goto retry_open;
1766                         }
1767                 }
1768                 errno = old_errno;
1769         }
1770
1771         if (err != -EINVAL || cpu > 0 || thread > 0)
1772                 goto out_close;
1773
1774         /*
1775          * Must probe features in the order they were added to the
1776          * perf_event_attr interface.
1777          */
1778         if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1779                 perf_missing_features.cgroup = true;
1780                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1781                 goto out_close;
1782         } else if (!perf_missing_features.branch_hw_idx &&
1783             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1784                 perf_missing_features.branch_hw_idx = true;
1785                 pr_debug2("switching off branch HW index support\n");
1786                 goto fallback_missing_features;
1787         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1788                 perf_missing_features.aux_output = true;
1789                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1790                 goto out_close;
1791         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1792                 perf_missing_features.bpf = true;
1793                 pr_debug2_peo("switching off bpf_event\n");
1794                 goto fallback_missing_features;
1795         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1796                 perf_missing_features.ksymbol = true;
1797                 pr_debug2_peo("switching off ksymbol\n");
1798                 goto fallback_missing_features;
1799         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1800                 perf_missing_features.write_backward = true;
1801                 pr_debug2_peo("switching off write_backward\n");
1802                 goto out_close;
1803         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1804                 perf_missing_features.clockid_wrong = true;
1805                 pr_debug2_peo("switching off clockid\n");
1806                 goto fallback_missing_features;
1807         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1808                 perf_missing_features.clockid = true;
1809                 pr_debug2_peo("switching off use_clockid\n");
1810                 goto fallback_missing_features;
1811         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1812                 perf_missing_features.cloexec = true;
1813                 pr_debug2_peo("switching off cloexec flag\n");
1814                 goto fallback_missing_features;
1815         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1816                 perf_missing_features.mmap2 = true;
1817                 pr_debug2_peo("switching off mmap2\n");
1818                 goto fallback_missing_features;
1819         } else if (!perf_missing_features.exclude_guest &&
1820                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1821                 perf_missing_features.exclude_guest = true;
1822                 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1823                 goto fallback_missing_features;
1824         } else if (!perf_missing_features.sample_id_all) {
1825                 perf_missing_features.sample_id_all = true;
1826                 pr_debug2_peo("switching off sample_id_all\n");
1827                 goto retry_sample_id;
1828         } else if (!perf_missing_features.lbr_flags &&
1829                         (evsel->core.attr.branch_sample_type &
1830                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1831                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1832                 perf_missing_features.lbr_flags = true;
1833                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1834                 goto fallback_missing_features;
1835         } else if (!perf_missing_features.group_read &&
1836                     evsel->core.attr.inherit &&
1837                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1838                    perf_evsel__is_group_leader(evsel)) {
1839                 perf_missing_features.group_read = true;
1840                 pr_debug2_peo("switching off group read\n");
1841                 goto fallback_missing_features;
1842         }
1843 out_close:
1844         if (err)
1845                 threads->err_thread = thread;
1846
1847         old_errno = errno;
1848         do {
1849                 while (--thread >= 0) {
1850                         if (FD(evsel, cpu, thread) >= 0)
1851                                 close(FD(evsel, cpu, thread));
1852                         FD(evsel, cpu, thread) = -1;
1853                 }
1854                 thread = nthreads;
1855         } while (--cpu >= 0);
1856         errno = old_errno;
1857         return err;
1858 }
1859
1860 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1861                 struct perf_thread_map *threads)
1862 {
1863         return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1864 }
1865
1866 void evsel__close(struct evsel *evsel)
1867 {
1868         perf_evsel__close(&evsel->core);
1869         perf_evsel__free_id(&evsel->core);
1870 }
1871
1872 int perf_evsel__open_per_cpu(struct evsel *evsel,
1873                              struct perf_cpu_map *cpus,
1874                              int cpu)
1875 {
1876         if (cpu == -1)
1877                 return evsel__open_cpu(evsel, cpus, NULL, 0,
1878                                         cpus ? cpus->nr : 1);
1879
1880         return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1881 }
1882
1883 int perf_evsel__open_per_thread(struct evsel *evsel,
1884                                 struct perf_thread_map *threads)
1885 {
1886         return evsel__open(evsel, NULL, threads);
1887 }
1888
1889 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1890                                        const union perf_event *event,
1891                                        struct perf_sample *sample)
1892 {
1893         u64 type = evsel->core.attr.sample_type;
1894         const __u64 *array = event->sample.array;
1895         bool swapped = evsel->needs_swap;
1896         union u64_swap u;
1897
1898         array += ((event->header.size -
1899                    sizeof(event->header)) / sizeof(u64)) - 1;
1900
1901         if (type & PERF_SAMPLE_IDENTIFIER) {
1902                 sample->id = *array;
1903                 array--;
1904         }
1905
1906         if (type & PERF_SAMPLE_CPU) {
1907                 u.val64 = *array;
1908                 if (swapped) {
1909                         /* undo swap of u64, then swap on individual u32s */
1910                         u.val64 = bswap_64(u.val64);
1911                         u.val32[0] = bswap_32(u.val32[0]);
1912                 }
1913
1914                 sample->cpu = u.val32[0];
1915                 array--;
1916         }
1917
1918         if (type & PERF_SAMPLE_STREAM_ID) {
1919                 sample->stream_id = *array;
1920                 array--;
1921         }
1922
1923         if (type & PERF_SAMPLE_ID) {
1924                 sample->id = *array;
1925                 array--;
1926         }
1927
1928         if (type & PERF_SAMPLE_TIME) {
1929                 sample->time = *array;
1930                 array--;
1931         }
1932
1933         if (type & PERF_SAMPLE_TID) {
1934                 u.val64 = *array;
1935                 if (swapped) {
1936                         /* undo swap of u64, then swap on individual u32s */
1937                         u.val64 = bswap_64(u.val64);
1938                         u.val32[0] = bswap_32(u.val32[0]);
1939                         u.val32[1] = bswap_32(u.val32[1]);
1940                 }
1941
1942                 sample->pid = u.val32[0];
1943                 sample->tid = u.val32[1];
1944                 array--;
1945         }
1946
1947         return 0;
1948 }
1949
1950 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1951                             u64 size)
1952 {
1953         return size > max_size || offset + size > endp;
1954 }
1955
1956 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1957         do {                                                            \
1958                 if (overflow(endp, (max_size), (offset), (size)))       \
1959                         return -EFAULT;                                 \
1960         } while (0)
1961
1962 #define OVERFLOW_CHECK_u64(offset) \
1963         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1964
1965 static int
1966 perf_event__check_size(union perf_event *event, unsigned int sample_size)
1967 {
1968         /*
1969          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1970          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1971          * check the format does not go past the end of the event.
1972          */
1973         if (sample_size + sizeof(event->header) > event->header.size)
1974                 return -EFAULT;
1975
1976         return 0;
1977 }
1978
1979 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event,
1980                              struct perf_sample *data)
1981 {
1982         u64 type = evsel->core.attr.sample_type;
1983         bool swapped = evsel->needs_swap;
1984         const __u64 *array;
1985         u16 max_size = event->header.size;
1986         const void *endp = (void *)event + max_size;
1987         u64 sz;
1988
1989         /*
1990          * used for cross-endian analysis. See git commit 65014ab3
1991          * for why this goofiness is needed.
1992          */
1993         union u64_swap u;
1994
1995         memset(data, 0, sizeof(*data));
1996         data->cpu = data->pid = data->tid = -1;
1997         data->stream_id = data->id = data->time = -1ULL;
1998         data->period = evsel->core.attr.sample_period;
1999         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2000         data->misc    = event->header.misc;
2001         data->id = -1ULL;
2002         data->data_src = PERF_MEM_DATA_SRC_NONE;
2003
2004         if (event->header.type != PERF_RECORD_SAMPLE) {
2005                 if (!evsel->core.attr.sample_id_all)
2006                         return 0;
2007                 return perf_evsel__parse_id_sample(evsel, event, data);
2008         }
2009
2010         array = event->sample.array;
2011
2012         if (perf_event__check_size(event, evsel->sample_size))
2013                 return -EFAULT;
2014
2015         if (type & PERF_SAMPLE_IDENTIFIER) {
2016                 data->id = *array;
2017                 array++;
2018         }
2019
2020         if (type & PERF_SAMPLE_IP) {
2021                 data->ip = *array;
2022                 array++;
2023         }
2024
2025         if (type & PERF_SAMPLE_TID) {
2026                 u.val64 = *array;
2027                 if (swapped) {
2028                         /* undo swap of u64, then swap on individual u32s */
2029                         u.val64 = bswap_64(u.val64);
2030                         u.val32[0] = bswap_32(u.val32[0]);
2031                         u.val32[1] = bswap_32(u.val32[1]);
2032                 }
2033
2034                 data->pid = u.val32[0];
2035                 data->tid = u.val32[1];
2036                 array++;
2037         }
2038
2039         if (type & PERF_SAMPLE_TIME) {
2040                 data->time = *array;
2041                 array++;
2042         }
2043
2044         if (type & PERF_SAMPLE_ADDR) {
2045                 data->addr = *array;
2046                 array++;
2047         }
2048
2049         if (type & PERF_SAMPLE_ID) {
2050                 data->id = *array;
2051                 array++;
2052         }
2053
2054         if (type & PERF_SAMPLE_STREAM_ID) {
2055                 data->stream_id = *array;
2056                 array++;
2057         }
2058
2059         if (type & PERF_SAMPLE_CPU) {
2060
2061                 u.val64 = *array;
2062                 if (swapped) {
2063                         /* undo swap of u64, then swap on individual u32s */
2064                         u.val64 = bswap_64(u.val64);
2065                         u.val32[0] = bswap_32(u.val32[0]);
2066                 }
2067
2068                 data->cpu = u.val32[0];
2069                 array++;
2070         }
2071
2072         if (type & PERF_SAMPLE_PERIOD) {
2073                 data->period = *array;
2074                 array++;
2075         }
2076
2077         if (type & PERF_SAMPLE_READ) {
2078                 u64 read_format = evsel->core.attr.read_format;
2079
2080                 OVERFLOW_CHECK_u64(array);
2081                 if (read_format & PERF_FORMAT_GROUP)
2082                         data->read.group.nr = *array;
2083                 else
2084                         data->read.one.value = *array;
2085
2086                 array++;
2087
2088                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2089                         OVERFLOW_CHECK_u64(array);
2090                         data->read.time_enabled = *array;
2091                         array++;
2092                 }
2093
2094                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2095                         OVERFLOW_CHECK_u64(array);
2096                         data->read.time_running = *array;
2097                         array++;
2098                 }
2099
2100                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2101                 if (read_format & PERF_FORMAT_GROUP) {
2102                         const u64 max_group_nr = UINT64_MAX /
2103                                         sizeof(struct sample_read_value);
2104
2105                         if (data->read.group.nr > max_group_nr)
2106                                 return -EFAULT;
2107                         sz = data->read.group.nr *
2108                              sizeof(struct sample_read_value);
2109                         OVERFLOW_CHECK(array, sz, max_size);
2110                         data->read.group.values =
2111                                         (struct sample_read_value *)array;
2112                         array = (void *)array + sz;
2113                 } else {
2114                         OVERFLOW_CHECK_u64(array);
2115                         data->read.one.id = *array;
2116                         array++;
2117                 }
2118         }
2119
2120         if (type & PERF_SAMPLE_CALLCHAIN) {
2121                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2122
2123                 OVERFLOW_CHECK_u64(array);
2124                 data->callchain = (struct ip_callchain *)array++;
2125                 if (data->callchain->nr > max_callchain_nr)
2126                         return -EFAULT;
2127                 sz = data->callchain->nr * sizeof(u64);
2128                 OVERFLOW_CHECK(array, sz, max_size);
2129                 array = (void *)array + sz;
2130         }
2131
2132         if (type & PERF_SAMPLE_RAW) {
2133                 OVERFLOW_CHECK_u64(array);
2134                 u.val64 = *array;
2135
2136                 /*
2137                  * Undo swap of u64, then swap on individual u32s,
2138                  * get the size of the raw area and undo all of the
2139                  * swap. The pevent interface handles endianity by
2140                  * itself.
2141                  */
2142                 if (swapped) {
2143                         u.val64 = bswap_64(u.val64);
2144                         u.val32[0] = bswap_32(u.val32[0]);
2145                         u.val32[1] = bswap_32(u.val32[1]);
2146                 }
2147                 data->raw_size = u.val32[0];
2148
2149                 /*
2150                  * The raw data is aligned on 64bits including the
2151                  * u32 size, so it's safe to use mem_bswap_64.
2152                  */
2153                 if (swapped)
2154                         mem_bswap_64((void *) array, data->raw_size);
2155
2156                 array = (void *)array + sizeof(u32);
2157
2158                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2159                 data->raw_data = (void *)array;
2160                 array = (void *)array + data->raw_size;
2161         }
2162
2163         if (type & PERF_SAMPLE_BRANCH_STACK) {
2164                 const u64 max_branch_nr = UINT64_MAX /
2165                                           sizeof(struct branch_entry);
2166
2167                 OVERFLOW_CHECK_u64(array);
2168                 data->branch_stack = (struct branch_stack *)array++;
2169
2170                 if (data->branch_stack->nr > max_branch_nr)
2171                         return -EFAULT;
2172
2173                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2174                 if (perf_evsel__has_branch_hw_idx(evsel))
2175                         sz += sizeof(u64);
2176                 else
2177                         data->no_hw_idx = true;
2178                 OVERFLOW_CHECK(array, sz, max_size);
2179                 array = (void *)array + sz;
2180         }
2181
2182         if (type & PERF_SAMPLE_REGS_USER) {
2183                 OVERFLOW_CHECK_u64(array);
2184                 data->user_regs.abi = *array;
2185                 array++;
2186
2187                 if (data->user_regs.abi) {
2188                         u64 mask = evsel->core.attr.sample_regs_user;
2189
2190                         sz = hweight64(mask) * sizeof(u64);
2191                         OVERFLOW_CHECK(array, sz, max_size);
2192                         data->user_regs.mask = mask;
2193                         data->user_regs.regs = (u64 *)array;
2194                         array = (void *)array + sz;
2195                 }
2196         }
2197
2198         if (type & PERF_SAMPLE_STACK_USER) {
2199                 OVERFLOW_CHECK_u64(array);
2200                 sz = *array++;
2201
2202                 data->user_stack.offset = ((char *)(array - 1)
2203                                           - (char *) event);
2204
2205                 if (!sz) {
2206                         data->user_stack.size = 0;
2207                 } else {
2208                         OVERFLOW_CHECK(array, sz, max_size);
2209                         data->user_stack.data = (char *)array;
2210                         array = (void *)array + sz;
2211                         OVERFLOW_CHECK_u64(array);
2212                         data->user_stack.size = *array++;
2213                         if (WARN_ONCE(data->user_stack.size > sz,
2214                                       "user stack dump failure\n"))
2215                                 return -EFAULT;
2216                 }
2217         }
2218
2219         if (type & PERF_SAMPLE_WEIGHT) {
2220                 OVERFLOW_CHECK_u64(array);
2221                 data->weight = *array;
2222                 array++;
2223         }
2224
2225         if (type & PERF_SAMPLE_DATA_SRC) {
2226                 OVERFLOW_CHECK_u64(array);
2227                 data->data_src = *array;
2228                 array++;
2229         }
2230
2231         if (type & PERF_SAMPLE_TRANSACTION) {
2232                 OVERFLOW_CHECK_u64(array);
2233                 data->transaction = *array;
2234                 array++;
2235         }
2236
2237         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2238         if (type & PERF_SAMPLE_REGS_INTR) {
2239                 OVERFLOW_CHECK_u64(array);
2240                 data->intr_regs.abi = *array;
2241                 array++;
2242
2243                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2244                         u64 mask = evsel->core.attr.sample_regs_intr;
2245
2246                         sz = hweight64(mask) * sizeof(u64);
2247                         OVERFLOW_CHECK(array, sz, max_size);
2248                         data->intr_regs.mask = mask;
2249                         data->intr_regs.regs = (u64 *)array;
2250                         array = (void *)array + sz;
2251                 }
2252         }
2253
2254         data->phys_addr = 0;
2255         if (type & PERF_SAMPLE_PHYS_ADDR) {
2256                 data->phys_addr = *array;
2257                 array++;
2258         }
2259
2260         data->cgroup = 0;
2261         if (type & PERF_SAMPLE_CGROUP) {
2262                 data->cgroup = *array;
2263                 array++;
2264         }
2265
2266         if (type & PERF_SAMPLE_AUX) {
2267                 OVERFLOW_CHECK_u64(array);
2268                 sz = *array++;
2269
2270                 OVERFLOW_CHECK(array, sz, max_size);
2271                 /* Undo swap of data */
2272                 if (swapped)
2273                         mem_bswap_64((char *)array, sz);
2274                 data->aux_sample.size = sz;
2275                 data->aux_sample.data = (char *)array;
2276                 array = (void *)array + sz;
2277         }
2278
2279         return 0;
2280 }
2281
2282 int perf_evsel__parse_sample_timestamp(struct evsel *evsel,
2283                                        union perf_event *event,
2284                                        u64 *timestamp)
2285 {
2286         u64 type = evsel->core.attr.sample_type;
2287         const __u64 *array;
2288
2289         if (!(type & PERF_SAMPLE_TIME))
2290                 return -1;
2291
2292         if (event->header.type != PERF_RECORD_SAMPLE) {
2293                 struct perf_sample data = {
2294                         .time = -1ULL,
2295                 };
2296
2297                 if (!evsel->core.attr.sample_id_all)
2298                         return -1;
2299                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2300                         return -1;
2301
2302                 *timestamp = data.time;
2303                 return 0;
2304         }
2305
2306         array = event->sample.array;
2307
2308         if (perf_event__check_size(event, evsel->sample_size))
2309                 return -EFAULT;
2310
2311         if (type & PERF_SAMPLE_IDENTIFIER)
2312                 array++;
2313
2314         if (type & PERF_SAMPLE_IP)
2315                 array++;
2316
2317         if (type & PERF_SAMPLE_TID)
2318                 array++;
2319
2320         if (type & PERF_SAMPLE_TIME)
2321                 *timestamp = *array;
2322
2323         return 0;
2324 }
2325
2326 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name)
2327 {
2328         return tep_find_field(evsel->tp_format, name);
2329 }
2330
2331 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample,
2332                          const char *name)
2333 {
2334         struct tep_format_field *field = perf_evsel__field(evsel, name);
2335         int offset;
2336
2337         if (!field)
2338                 return NULL;
2339
2340         offset = field->offset;
2341
2342         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2343                 offset = *(int *)(sample->raw_data + field->offset);
2344                 offset &= 0xffff;
2345         }
2346
2347         return sample->raw_data + offset;
2348 }
2349
2350 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2351                          bool needs_swap)
2352 {
2353         u64 value;
2354         void *ptr = sample->raw_data + field->offset;
2355
2356         switch (field->size) {
2357         case 1:
2358                 return *(u8 *)ptr;
2359         case 2:
2360                 value = *(u16 *)ptr;
2361                 break;
2362         case 4:
2363                 value = *(u32 *)ptr;
2364                 break;
2365         case 8:
2366                 memcpy(&value, ptr, sizeof(u64));
2367                 break;
2368         default:
2369                 return 0;
2370         }
2371
2372         if (!needs_swap)
2373                 return value;
2374
2375         switch (field->size) {
2376         case 2:
2377                 return bswap_16(value);
2378         case 4:
2379                 return bswap_32(value);
2380         case 8:
2381                 return bswap_64(value);
2382         default:
2383                 return 0;
2384         }
2385
2386         return 0;
2387 }
2388
2389 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample,
2390                        const char *name)
2391 {
2392         struct tep_format_field *field = perf_evsel__field(evsel, name);
2393
2394         if (!field)
2395                 return 0;
2396
2397         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2398 }
2399
2400 bool perf_evsel__fallback(struct evsel *evsel, int err,
2401                           char *msg, size_t msgsize)
2402 {
2403         int paranoid;
2404
2405         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2406             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2407             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2408                 /*
2409                  * If it's cycles then fall back to hrtimer based
2410                  * cpu-clock-tick sw counter, which is always available even if
2411                  * no PMU support.
2412                  *
2413                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2414                  * b0a873e).
2415                  */
2416                 scnprintf(msg, msgsize, "%s",
2417 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2418
2419                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2420                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2421
2422                 zfree(&evsel->name);
2423                 return true;
2424         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2425                    (paranoid = perf_event_paranoid()) > 1) {
2426                 const char *name = perf_evsel__name(evsel);
2427                 char *new_name;
2428                 const char *sep = ":";
2429
2430                 /* If event has exclude user then don't exclude kernel. */
2431                 if (evsel->core.attr.exclude_user)
2432                         return false;
2433
2434                 /* Is there already the separator in the name. */
2435                 if (strchr(name, '/') ||
2436                     strchr(name, ':'))
2437                         sep = "";
2438
2439                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2440                         return false;
2441
2442                 if (evsel->name)
2443                         free(evsel->name);
2444                 evsel->name = new_name;
2445                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2446                           "to fall back to excluding kernel and hypervisor "
2447                           " samples", paranoid);
2448                 evsel->core.attr.exclude_kernel = 1;
2449                 evsel->core.attr.exclude_hv     = 1;
2450
2451                 return true;
2452         }
2453
2454         return false;
2455 }
2456
2457 static bool find_process(const char *name)
2458 {
2459         size_t len = strlen(name);
2460         DIR *dir;
2461         struct dirent *d;
2462         int ret = -1;
2463
2464         dir = opendir(procfs__mountpoint());
2465         if (!dir)
2466                 return false;
2467
2468         /* Walk through the directory. */
2469         while (ret && (d = readdir(dir)) != NULL) {
2470                 char path[PATH_MAX];
2471                 char *data;
2472                 size_t size;
2473
2474                 if ((d->d_type != DT_DIR) ||
2475                      !strcmp(".", d->d_name) ||
2476                      !strcmp("..", d->d_name))
2477                         continue;
2478
2479                 scnprintf(path, sizeof(path), "%s/%s/comm",
2480                           procfs__mountpoint(), d->d_name);
2481
2482                 if (filename__read_str(path, &data, &size))
2483                         continue;
2484
2485                 ret = strncmp(name, data, len);
2486                 free(data);
2487         }
2488
2489         closedir(dir);
2490         return ret ? false : true;
2491 }
2492
2493 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target,
2494                               int err, char *msg, size_t size)
2495 {
2496         char sbuf[STRERR_BUFSIZE];
2497         int printed = 0;
2498
2499         switch (err) {
2500         case EPERM:
2501         case EACCES:
2502                 if (err == EPERM)
2503                         printed = scnprintf(msg, size,
2504                                 "No permission to enable %s event.\n\n",
2505                                 perf_evsel__name(evsel));
2506
2507                 return scnprintf(msg + printed, size - printed,
2508                  "You may not have permission to collect %sstats.\n\n"
2509                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2510                  "which controls use of the performance events system by\n"
2511                  "unprivileged users (without CAP_PERFMON or CAP_SYS_ADMIN).\n\n"
2512                  "The current value is %d:\n\n"
2513                  "  -1: Allow use of (almost) all events by all users\n"
2514                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2515                  ">= 0: Disallow ftrace function tracepoint by users without CAP_PERFMON or CAP_SYS_ADMIN\n"
2516                  "      Disallow raw tracepoint access by users without CAP_SYS_PERFMON or CAP_SYS_ADMIN\n"
2517                  ">= 1: Disallow CPU event access by users without CAP_PERFMON or CAP_SYS_ADMIN\n"
2518                  ">= 2: Disallow kernel profiling by users without CAP_PERFMON or CAP_SYS_ADMIN\n\n"
2519                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2520                  "      kernel.perf_event_paranoid = -1\n" ,
2521                                  target->system_wide ? "system-wide " : "",
2522                                  perf_event_paranoid());
2523         case ENOENT:
2524                 return scnprintf(msg, size, "The %s event is not supported.",
2525                                  perf_evsel__name(evsel));
2526         case EMFILE:
2527                 return scnprintf(msg, size, "%s",
2528                          "Too many events are opened.\n"
2529                          "Probably the maximum number of open file descriptors has been reached.\n"
2530                          "Hint: Try again after reducing the number of events.\n"
2531                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2532         case ENOMEM:
2533                 if (evsel__has_callchain(evsel) &&
2534                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2535                         return scnprintf(msg, size,
2536                                          "Not enough memory to setup event with callchain.\n"
2537                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2538                                          "Hint: Current value: %d", sysctl__max_stack());
2539                 break;
2540         case ENODEV:
2541                 if (target->cpu_list)
2542                         return scnprintf(msg, size, "%s",
2543          "No such device - did you specify an out-of-range profile CPU?");
2544                 break;
2545         case EOPNOTSUPP:
2546                 if (evsel->core.attr.sample_period != 0)
2547                         return scnprintf(msg, size,
2548         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2549                                          perf_evsel__name(evsel));
2550                 if (evsel->core.attr.precise_ip)
2551                         return scnprintf(msg, size, "%s",
2552         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2553 #if defined(__i386__) || defined(__x86_64__)
2554                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2555                         return scnprintf(msg, size, "%s",
2556         "No hardware sampling interrupt available.\n");
2557 #endif
2558                 break;
2559         case EBUSY:
2560                 if (find_process("oprofiled"))
2561                         return scnprintf(msg, size,
2562         "The PMU counters are busy/taken by another profiler.\n"
2563         "We found oprofile daemon running, please stop it and try again.");
2564                 break;
2565         case EINVAL:
2566                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2567                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2568                 if (perf_missing_features.clockid)
2569                         return scnprintf(msg, size, "clockid feature not supported.");
2570                 if (perf_missing_features.clockid_wrong)
2571                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2572                 if (perf_missing_features.aux_output)
2573                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2574                 break;
2575         default:
2576                 break;
2577         }
2578
2579         return scnprintf(msg, size,
2580         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2581         "/bin/dmesg | grep -i perf may provide additional information.\n",
2582                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2583                          perf_evsel__name(evsel));
2584 }
2585
2586 struct perf_env *perf_evsel__env(struct evsel *evsel)
2587 {
2588         if (evsel && evsel->evlist)
2589                 return evsel->evlist->env;
2590         return &perf_env;
2591 }
2592
2593 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2594 {
2595         int cpu, thread;
2596
2597         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2598                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2599                      thread++) {
2600                         int fd = FD(evsel, cpu, thread);
2601
2602                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2603                                                    cpu, thread, fd) < 0)
2604                                 return -1;
2605                 }
2606         }
2607
2608         return 0;
2609 }
2610
2611 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2612 {
2613         struct perf_cpu_map *cpus = evsel->core.cpus;
2614         struct perf_thread_map *threads = evsel->core.threads;
2615
2616         if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2617                 return -ENOMEM;
2618
2619         return store_evsel_ids(evsel, evlist);
2620 }