Merge branch 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[sfrench/cifs-2.6.git] / tools / perf / util / evsel.c
1 /*
2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3  *
4  * Parts came from builtin-{top,stat,record}.c, see those files for further
5  * copyright notes.
6  *
7  * Released under the GPL v2. (and only v2, not any later version)
8  */
9
10 #include <byteswap.h>
11 #include <errno.h>
12 #include <inttypes.h>
13 #include <linux/bitops.h>
14 #include <api/fs/fs.h>
15 #include <api/fs/tracing_path.h>
16 #include <traceevent/event-parse.h>
17 #include <linux/hw_breakpoint.h>
18 #include <linux/perf_event.h>
19 #include <linux/compiler.h>
20 #include <linux/err.h>
21 #include <sys/ioctl.h>
22 #include <sys/resource.h>
23 #include <sys/types.h>
24 #include <dirent.h>
25 #include "asm/bug.h"
26 #include "callchain.h"
27 #include "cgroup.h"
28 #include "event.h"
29 #include "evsel.h"
30 #include "evlist.h"
31 #include "util.h"
32 #include "cpumap.h"
33 #include "thread_map.h"
34 #include "target.h"
35 #include "perf_regs.h"
36 #include "debug.h"
37 #include "trace-event.h"
38 #include "stat.h"
39 #include "memswap.h"
40 #include "util/parse-branch-options.h"
41
42 #include "sane_ctype.h"
43
44 struct perf_missing_features perf_missing_features;
45
46 static clockid_t clockid;
47
48 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
49 {
50         return 0;
51 }
52
53 void __weak test_attr__ready(void) { }
54
55 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
56 {
57 }
58
59 static struct {
60         size_t  size;
61         int     (*init)(struct perf_evsel *evsel);
62         void    (*fini)(struct perf_evsel *evsel);
63 } perf_evsel__object = {
64         .size = sizeof(struct perf_evsel),
65         .init = perf_evsel__no_extra_init,
66         .fini = perf_evsel__no_extra_fini,
67 };
68
69 int perf_evsel__object_config(size_t object_size,
70                               int (*init)(struct perf_evsel *evsel),
71                               void (*fini)(struct perf_evsel *evsel))
72 {
73
74         if (object_size == 0)
75                 goto set_methods;
76
77         if (perf_evsel__object.size > object_size)
78                 return -EINVAL;
79
80         perf_evsel__object.size = object_size;
81
82 set_methods:
83         if (init != NULL)
84                 perf_evsel__object.init = init;
85
86         if (fini != NULL)
87                 perf_evsel__object.fini = fini;
88
89         return 0;
90 }
91
92 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
93
94 int __perf_evsel__sample_size(u64 sample_type)
95 {
96         u64 mask = sample_type & PERF_SAMPLE_MASK;
97         int size = 0;
98         int i;
99
100         for (i = 0; i < 64; i++) {
101                 if (mask & (1ULL << i))
102                         size++;
103         }
104
105         size *= sizeof(u64);
106
107         return size;
108 }
109
110 /**
111  * __perf_evsel__calc_id_pos - calculate id_pos.
112  * @sample_type: sample type
113  *
114  * This function returns the position of the event id (PERF_SAMPLE_ID or
115  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
116  * sample_event.
117  */
118 static int __perf_evsel__calc_id_pos(u64 sample_type)
119 {
120         int idx = 0;
121
122         if (sample_type & PERF_SAMPLE_IDENTIFIER)
123                 return 0;
124
125         if (!(sample_type & PERF_SAMPLE_ID))
126                 return -1;
127
128         if (sample_type & PERF_SAMPLE_IP)
129                 idx += 1;
130
131         if (sample_type & PERF_SAMPLE_TID)
132                 idx += 1;
133
134         if (sample_type & PERF_SAMPLE_TIME)
135                 idx += 1;
136
137         if (sample_type & PERF_SAMPLE_ADDR)
138                 idx += 1;
139
140         return idx;
141 }
142
143 /**
144  * __perf_evsel__calc_is_pos - calculate is_pos.
145  * @sample_type: sample type
146  *
147  * This function returns the position (counting backwards) of the event id
148  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
149  * sample_id_all is used there is an id sample appended to non-sample events.
150  */
151 static int __perf_evsel__calc_is_pos(u64 sample_type)
152 {
153         int idx = 1;
154
155         if (sample_type & PERF_SAMPLE_IDENTIFIER)
156                 return 1;
157
158         if (!(sample_type & PERF_SAMPLE_ID))
159                 return -1;
160
161         if (sample_type & PERF_SAMPLE_CPU)
162                 idx += 1;
163
164         if (sample_type & PERF_SAMPLE_STREAM_ID)
165                 idx += 1;
166
167         return idx;
168 }
169
170 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
171 {
172         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
173         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
174 }
175
176 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
177                                   enum perf_event_sample_format bit)
178 {
179         if (!(evsel->attr.sample_type & bit)) {
180                 evsel->attr.sample_type |= bit;
181                 evsel->sample_size += sizeof(u64);
182                 perf_evsel__calc_id_pos(evsel);
183         }
184 }
185
186 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
187                                     enum perf_event_sample_format bit)
188 {
189         if (evsel->attr.sample_type & bit) {
190                 evsel->attr.sample_type &= ~bit;
191                 evsel->sample_size -= sizeof(u64);
192                 perf_evsel__calc_id_pos(evsel);
193         }
194 }
195
196 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
197                                bool can_sample_identifier)
198 {
199         if (can_sample_identifier) {
200                 perf_evsel__reset_sample_bit(evsel, ID);
201                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
202         } else {
203                 perf_evsel__set_sample_bit(evsel, ID);
204         }
205         evsel->attr.read_format |= PERF_FORMAT_ID;
206 }
207
208 /**
209  * perf_evsel__is_function_event - Return whether given evsel is a function
210  * trace event
211  *
212  * @evsel - evsel selector to be tested
213  *
214  * Return %true if event is function trace event
215  */
216 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
217 {
218 #define FUNCTION_EVENT "ftrace:function"
219
220         return evsel->name &&
221                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
222
223 #undef FUNCTION_EVENT
224 }
225
226 void perf_evsel__init(struct perf_evsel *evsel,
227                       struct perf_event_attr *attr, int idx)
228 {
229         evsel->idx         = idx;
230         evsel->tracking    = !idx;
231         evsel->attr        = *attr;
232         evsel->leader      = evsel;
233         evsel->unit        = "";
234         evsel->scale       = 1.0;
235         evsel->max_events  = ULONG_MAX;
236         evsel->evlist      = NULL;
237         evsel->bpf_fd      = -1;
238         INIT_LIST_HEAD(&evsel->node);
239         INIT_LIST_HEAD(&evsel->config_terms);
240         perf_evsel__object.init(evsel);
241         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
242         perf_evsel__calc_id_pos(evsel);
243         evsel->cmdline_group_boundary = false;
244         evsel->metric_expr   = NULL;
245         evsel->metric_name   = NULL;
246         evsel->metric_events = NULL;
247         evsel->collect_stat  = false;
248         evsel->pmu_name      = NULL;
249 }
250
251 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
252 {
253         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
254
255         if (!evsel)
256                 return NULL;
257         perf_evsel__init(evsel, attr, idx);
258
259         if (perf_evsel__is_bpf_output(evsel)) {
260                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
261                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
262                 evsel->attr.sample_period = 1;
263         }
264
265         if (perf_evsel__is_clock(evsel)) {
266                 /*
267                  * The evsel->unit points to static alias->unit
268                  * so it's ok to use static string in here.
269                  */
270                 static const char *unit = "msec";
271
272                 evsel->unit = unit;
273                 evsel->scale = 1e-6;
274         }
275
276         return evsel;
277 }
278
279 static bool perf_event_can_profile_kernel(void)
280 {
281         return geteuid() == 0 || perf_event_paranoid() == -1;
282 }
283
284 struct perf_evsel *perf_evsel__new_cycles(bool precise)
285 {
286         struct perf_event_attr attr = {
287                 .type   = PERF_TYPE_HARDWARE,
288                 .config = PERF_COUNT_HW_CPU_CYCLES,
289                 .exclude_kernel = !perf_event_can_profile_kernel(),
290         };
291         struct perf_evsel *evsel;
292
293         event_attr_init(&attr);
294
295         if (!precise)
296                 goto new_event;
297
298         /*
299          * Now let the usual logic to set up the perf_event_attr defaults
300          * to kick in when we return and before perf_evsel__open() is called.
301          */
302 new_event:
303         evsel = perf_evsel__new(&attr);
304         if (evsel == NULL)
305                 goto out;
306
307         evsel->precise_max = true;
308
309         /* use asprintf() because free(evsel) assumes name is allocated */
310         if (asprintf(&evsel->name, "cycles%s%s%.*s",
311                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
312                      attr.exclude_kernel ? "u" : "",
313                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
314                 goto error_free;
315 out:
316         return evsel;
317 error_free:
318         perf_evsel__delete(evsel);
319         evsel = NULL;
320         goto out;
321 }
322
323 /*
324  * Returns pointer with encoded error via <linux/err.h> interface.
325  */
326 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
327 {
328         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
329         int err = -ENOMEM;
330
331         if (evsel == NULL) {
332                 goto out_err;
333         } else {
334                 struct perf_event_attr attr = {
335                         .type          = PERF_TYPE_TRACEPOINT,
336                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
337                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
338                 };
339
340                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
341                         goto out_free;
342
343                 evsel->tp_format = trace_event__tp_format(sys, name);
344                 if (IS_ERR(evsel->tp_format)) {
345                         err = PTR_ERR(evsel->tp_format);
346                         goto out_free;
347                 }
348
349                 event_attr_init(&attr);
350                 attr.config = evsel->tp_format->id;
351                 attr.sample_period = 1;
352                 perf_evsel__init(evsel, &attr, idx);
353         }
354
355         return evsel;
356
357 out_free:
358         zfree(&evsel->name);
359         free(evsel);
360 out_err:
361         return ERR_PTR(err);
362 }
363
364 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
365         "cycles",
366         "instructions",
367         "cache-references",
368         "cache-misses",
369         "branches",
370         "branch-misses",
371         "bus-cycles",
372         "stalled-cycles-frontend",
373         "stalled-cycles-backend",
374         "ref-cycles",
375 };
376
377 static const char *__perf_evsel__hw_name(u64 config)
378 {
379         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
380                 return perf_evsel__hw_names[config];
381
382         return "unknown-hardware";
383 }
384
385 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
386 {
387         int colon = 0, r = 0;
388         struct perf_event_attr *attr = &evsel->attr;
389         bool exclude_guest_default = false;
390
391 #define MOD_PRINT(context, mod) do {                                    \
392                 if (!attr->exclude_##context) {                         \
393                         if (!colon) colon = ++r;                        \
394                         r += scnprintf(bf + r, size - r, "%c", mod);    \
395                 } } while(0)
396
397         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
398                 MOD_PRINT(kernel, 'k');
399                 MOD_PRINT(user, 'u');
400                 MOD_PRINT(hv, 'h');
401                 exclude_guest_default = true;
402         }
403
404         if (attr->precise_ip) {
405                 if (!colon)
406                         colon = ++r;
407                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
408                 exclude_guest_default = true;
409         }
410
411         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
412                 MOD_PRINT(host, 'H');
413                 MOD_PRINT(guest, 'G');
414         }
415 #undef MOD_PRINT
416         if (colon)
417                 bf[colon - 1] = ':';
418         return r;
419 }
420
421 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
422 {
423         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
424         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
425 }
426
427 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
428         "cpu-clock",
429         "task-clock",
430         "page-faults",
431         "context-switches",
432         "cpu-migrations",
433         "minor-faults",
434         "major-faults",
435         "alignment-faults",
436         "emulation-faults",
437         "dummy",
438 };
439
440 static const char *__perf_evsel__sw_name(u64 config)
441 {
442         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
443                 return perf_evsel__sw_names[config];
444         return "unknown-software";
445 }
446
447 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
448 {
449         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
450         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
451 }
452
453 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
454 {
455         int r;
456
457         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
458
459         if (type & HW_BREAKPOINT_R)
460                 r += scnprintf(bf + r, size - r, "r");
461
462         if (type & HW_BREAKPOINT_W)
463                 r += scnprintf(bf + r, size - r, "w");
464
465         if (type & HW_BREAKPOINT_X)
466                 r += scnprintf(bf + r, size - r, "x");
467
468         return r;
469 }
470
471 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
472 {
473         struct perf_event_attr *attr = &evsel->attr;
474         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
475         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
476 }
477
478 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
479                                 [PERF_EVSEL__MAX_ALIASES] = {
480  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
481  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
482  { "LLC",       "L2",                                                   },
483  { "dTLB",      "d-tlb",        "Data-TLB",                             },
484  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
485  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
486  { "node",                                                              },
487 };
488
489 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
490                                    [PERF_EVSEL__MAX_ALIASES] = {
491  { "load",      "loads",        "read",                                 },
492  { "store",     "stores",       "write",                                },
493  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
494 };
495
496 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
497                                        [PERF_EVSEL__MAX_ALIASES] = {
498  { "refs",      "Reference",    "ops",          "access",               },
499  { "misses",    "miss",                                                 },
500 };
501
502 #define C(x)            PERF_COUNT_HW_CACHE_##x
503 #define CACHE_READ      (1 << C(OP_READ))
504 #define CACHE_WRITE     (1 << C(OP_WRITE))
505 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
506 #define COP(x)          (1 << x)
507
508 /*
509  * cache operartion stat
510  * L1I : Read and prefetch only
511  * ITLB and BPU : Read-only
512  */
513 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
514  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
515  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
516  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
517  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
518  [C(ITLB)]      = (CACHE_READ),
519  [C(BPU)]       = (CACHE_READ),
520  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
521 };
522
523 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
524 {
525         if (perf_evsel__hw_cache_stat[type] & COP(op))
526                 return true;    /* valid */
527         else
528                 return false;   /* invalid */
529 }
530
531 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
532                                             char *bf, size_t size)
533 {
534         if (result) {
535                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
536                                  perf_evsel__hw_cache_op[op][0],
537                                  perf_evsel__hw_cache_result[result][0]);
538         }
539
540         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
541                          perf_evsel__hw_cache_op[op][1]);
542 }
543
544 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
545 {
546         u8 op, result, type = (config >>  0) & 0xff;
547         const char *err = "unknown-ext-hardware-cache-type";
548
549         if (type >= PERF_COUNT_HW_CACHE_MAX)
550                 goto out_err;
551
552         op = (config >>  8) & 0xff;
553         err = "unknown-ext-hardware-cache-op";
554         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
555                 goto out_err;
556
557         result = (config >> 16) & 0xff;
558         err = "unknown-ext-hardware-cache-result";
559         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
560                 goto out_err;
561
562         err = "invalid-cache";
563         if (!perf_evsel__is_cache_op_valid(type, op))
564                 goto out_err;
565
566         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
567 out_err:
568         return scnprintf(bf, size, "%s", err);
569 }
570
571 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
572 {
573         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
574         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
575 }
576
577 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
578 {
579         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
580         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
581 }
582
583 const char *perf_evsel__name(struct perf_evsel *evsel)
584 {
585         char bf[128];
586
587         if (evsel->name)
588                 return evsel->name;
589
590         switch (evsel->attr.type) {
591         case PERF_TYPE_RAW:
592                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
593                 break;
594
595         case PERF_TYPE_HARDWARE:
596                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
597                 break;
598
599         case PERF_TYPE_HW_CACHE:
600                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
601                 break;
602
603         case PERF_TYPE_SOFTWARE:
604                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
605                 break;
606
607         case PERF_TYPE_TRACEPOINT:
608                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
609                 break;
610
611         case PERF_TYPE_BREAKPOINT:
612                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
613                 break;
614
615         default:
616                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
617                           evsel->attr.type);
618                 break;
619         }
620
621         evsel->name = strdup(bf);
622
623         return evsel->name ?: "unknown";
624 }
625
626 const char *perf_evsel__group_name(struct perf_evsel *evsel)
627 {
628         return evsel->group_name ?: "anon group";
629 }
630
631 /*
632  * Returns the group details for the specified leader,
633  * with following rules.
634  *
635  *  For record -e '{cycles,instructions}'
636  *    'anon group { cycles:u, instructions:u }'
637  *
638  *  For record -e 'cycles,instructions' and report --group
639  *    'cycles:u, instructions:u'
640  */
641 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
642 {
643         int ret = 0;
644         struct perf_evsel *pos;
645         const char *group_name = perf_evsel__group_name(evsel);
646
647         if (!evsel->forced_leader)
648                 ret = scnprintf(buf, size, "%s { ", group_name);
649
650         ret += scnprintf(buf + ret, size - ret, "%s",
651                          perf_evsel__name(evsel));
652
653         for_each_group_member(pos, evsel)
654                 ret += scnprintf(buf + ret, size - ret, ", %s",
655                                  perf_evsel__name(pos));
656
657         if (!evsel->forced_leader)
658                 ret += scnprintf(buf + ret, size - ret, " }");
659
660         return ret;
661 }
662
663 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
664                                            struct record_opts *opts,
665                                            struct callchain_param *param)
666 {
667         bool function = perf_evsel__is_function_event(evsel);
668         struct perf_event_attr *attr = &evsel->attr;
669
670         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
671
672         attr->sample_max_stack = param->max_stack;
673
674         if (param->record_mode == CALLCHAIN_LBR) {
675                 if (!opts->branch_stack) {
676                         if (attr->exclude_user) {
677                                 pr_warning("LBR callstack option is only available "
678                                            "to get user callchain information. "
679                                            "Falling back to framepointers.\n");
680                         } else {
681                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
682                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
683                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
684                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
685                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
686                         }
687                 } else
688                          pr_warning("Cannot use LBR callstack with branch stack. "
689                                     "Falling back to framepointers.\n");
690         }
691
692         if (param->record_mode == CALLCHAIN_DWARF) {
693                 if (!function) {
694                         perf_evsel__set_sample_bit(evsel, REGS_USER);
695                         perf_evsel__set_sample_bit(evsel, STACK_USER);
696                         attr->sample_regs_user |= PERF_REGS_MASK;
697                         attr->sample_stack_user = param->dump_size;
698                         attr->exclude_callchain_user = 1;
699                 } else {
700                         pr_info("Cannot use DWARF unwind for function trace event,"
701                                 " falling back to framepointers.\n");
702                 }
703         }
704
705         if (function) {
706                 pr_info("Disabling user space callchains for function trace event.\n");
707                 attr->exclude_callchain_user = 1;
708         }
709 }
710
711 void perf_evsel__config_callchain(struct perf_evsel *evsel,
712                                   struct record_opts *opts,
713                                   struct callchain_param *param)
714 {
715         if (param->enabled)
716                 return __perf_evsel__config_callchain(evsel, opts, param);
717 }
718
719 static void
720 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
721                             struct callchain_param *param)
722 {
723         struct perf_event_attr *attr = &evsel->attr;
724
725         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
726         if (param->record_mode == CALLCHAIN_LBR) {
727                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
728                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
729                                               PERF_SAMPLE_BRANCH_CALL_STACK);
730         }
731         if (param->record_mode == CALLCHAIN_DWARF) {
732                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
733                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
734         }
735 }
736
737 static void apply_config_terms(struct perf_evsel *evsel,
738                                struct record_opts *opts, bool track)
739 {
740         struct perf_evsel_config_term *term;
741         struct list_head *config_terms = &evsel->config_terms;
742         struct perf_event_attr *attr = &evsel->attr;
743         /* callgraph default */
744         struct callchain_param param = {
745                 .record_mode = callchain_param.record_mode,
746         };
747         u32 dump_size = 0;
748         int max_stack = 0;
749         const char *callgraph_buf = NULL;
750
751         list_for_each_entry(term, config_terms, list) {
752                 switch (term->type) {
753                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
754                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
755                                 attr->sample_period = term->val.period;
756                                 attr->freq = 0;
757                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
758                         }
759                         break;
760                 case PERF_EVSEL__CONFIG_TERM_FREQ:
761                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
762                                 attr->sample_freq = term->val.freq;
763                                 attr->freq = 1;
764                                 perf_evsel__set_sample_bit(evsel, PERIOD);
765                         }
766                         break;
767                 case PERF_EVSEL__CONFIG_TERM_TIME:
768                         if (term->val.time)
769                                 perf_evsel__set_sample_bit(evsel, TIME);
770                         else
771                                 perf_evsel__reset_sample_bit(evsel, TIME);
772                         break;
773                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
774                         callgraph_buf = term->val.callgraph;
775                         break;
776                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
777                         if (term->val.branch && strcmp(term->val.branch, "no")) {
778                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
779                                 parse_branch_str(term->val.branch,
780                                                  &attr->branch_sample_type);
781                         } else
782                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
783                         break;
784                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
785                         dump_size = term->val.stack_user;
786                         break;
787                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
788                         max_stack = term->val.max_stack;
789                         break;
790                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
791                         evsel->max_events = term->val.max_events;
792                         break;
793                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
794                         /*
795                          * attr->inherit should has already been set by
796                          * perf_evsel__config. If user explicitly set
797                          * inherit using config terms, override global
798                          * opt->no_inherit setting.
799                          */
800                         attr->inherit = term->val.inherit ? 1 : 0;
801                         break;
802                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
803                         attr->write_backward = term->val.overwrite ? 1 : 0;
804                         break;
805                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
806                         break;
807                 default:
808                         break;
809                 }
810         }
811
812         /* User explicitly set per-event callgraph, clear the old setting and reset. */
813         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
814                 bool sample_address = false;
815
816                 if (max_stack) {
817                         param.max_stack = max_stack;
818                         if (callgraph_buf == NULL)
819                                 callgraph_buf = "fp";
820                 }
821
822                 /* parse callgraph parameters */
823                 if (callgraph_buf != NULL) {
824                         if (!strcmp(callgraph_buf, "no")) {
825                                 param.enabled = false;
826                                 param.record_mode = CALLCHAIN_NONE;
827                         } else {
828                                 param.enabled = true;
829                                 if (parse_callchain_record(callgraph_buf, &param)) {
830                                         pr_err("per-event callgraph setting for %s failed. "
831                                                "Apply callgraph global setting for it\n",
832                                                evsel->name);
833                                         return;
834                                 }
835                                 if (param.record_mode == CALLCHAIN_DWARF)
836                                         sample_address = true;
837                         }
838                 }
839                 if (dump_size > 0) {
840                         dump_size = round_up(dump_size, sizeof(u64));
841                         param.dump_size = dump_size;
842                 }
843
844                 /* If global callgraph set, clear it */
845                 if (callchain_param.enabled)
846                         perf_evsel__reset_callgraph(evsel, &callchain_param);
847
848                 /* set perf-event callgraph */
849                 if (param.enabled) {
850                         if (sample_address) {
851                                 perf_evsel__set_sample_bit(evsel, ADDR);
852                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
853                                 evsel->attr.mmap_data = track;
854                         }
855                         perf_evsel__config_callchain(evsel, opts, &param);
856                 }
857         }
858 }
859
860 static bool is_dummy_event(struct perf_evsel *evsel)
861 {
862         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
863                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
864 }
865
866 /*
867  * The enable_on_exec/disabled value strategy:
868  *
869  *  1) For any type of traced program:
870  *    - all independent events and group leaders are disabled
871  *    - all group members are enabled
872  *
873  *     Group members are ruled by group leaders. They need to
874  *     be enabled, because the group scheduling relies on that.
875  *
876  *  2) For traced programs executed by perf:
877  *     - all independent events and group leaders have
878  *       enable_on_exec set
879  *     - we don't specifically enable or disable any event during
880  *       the record command
881  *
882  *     Independent events and group leaders are initially disabled
883  *     and get enabled by exec. Group members are ruled by group
884  *     leaders as stated in 1).
885  *
886  *  3) For traced programs attached by perf (pid/tid):
887  *     - we specifically enable or disable all events during
888  *       the record command
889  *
890  *     When attaching events to already running traced we
891  *     enable/disable events specifically, as there's no
892  *     initial traced exec call.
893  */
894 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
895                         struct callchain_param *callchain)
896 {
897         struct perf_evsel *leader = evsel->leader;
898         struct perf_event_attr *attr = &evsel->attr;
899         int track = evsel->tracking;
900         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
901
902         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
903         attr->inherit       = !opts->no_inherit;
904         attr->write_backward = opts->overwrite ? 1 : 0;
905
906         perf_evsel__set_sample_bit(evsel, IP);
907         perf_evsel__set_sample_bit(evsel, TID);
908
909         if (evsel->sample_read) {
910                 perf_evsel__set_sample_bit(evsel, READ);
911
912                 /*
913                  * We need ID even in case of single event, because
914                  * PERF_SAMPLE_READ process ID specific data.
915                  */
916                 perf_evsel__set_sample_id(evsel, false);
917
918                 /*
919                  * Apply group format only if we belong to group
920                  * with more than one members.
921                  */
922                 if (leader->nr_members > 1) {
923                         attr->read_format |= PERF_FORMAT_GROUP;
924                         attr->inherit = 0;
925                 }
926         }
927
928         /*
929          * We default some events to have a default interval. But keep
930          * it a weak assumption overridable by the user.
931          */
932         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
933                                      opts->user_interval != ULLONG_MAX)) {
934                 if (opts->freq) {
935                         perf_evsel__set_sample_bit(evsel, PERIOD);
936                         attr->freq              = 1;
937                         attr->sample_freq       = opts->freq;
938                 } else {
939                         attr->sample_period = opts->default_interval;
940                 }
941         }
942
943         /*
944          * Disable sampling for all group members other
945          * than leader in case leader 'leads' the sampling.
946          */
947         if ((leader != evsel) && leader->sample_read) {
948                 attr->freq           = 0;
949                 attr->sample_freq    = 0;
950                 attr->sample_period  = 0;
951                 attr->write_backward = 0;
952
953                 /*
954                  * We don't get sample for slave events, we make them
955                  * when delivering group leader sample. Set the slave
956                  * event to follow the master sample_type to ease up
957                  * report.
958                  */
959                 attr->sample_type = leader->attr.sample_type;
960         }
961
962         if (opts->no_samples)
963                 attr->sample_freq = 0;
964
965         if (opts->inherit_stat) {
966                 evsel->attr.read_format |=
967                         PERF_FORMAT_TOTAL_TIME_ENABLED |
968                         PERF_FORMAT_TOTAL_TIME_RUNNING |
969                         PERF_FORMAT_ID;
970                 attr->inherit_stat = 1;
971         }
972
973         if (opts->sample_address) {
974                 perf_evsel__set_sample_bit(evsel, ADDR);
975                 attr->mmap_data = track;
976         }
977
978         /*
979          * We don't allow user space callchains for  function trace
980          * event, due to issues with page faults while tracing page
981          * fault handler and its overall trickiness nature.
982          */
983         if (perf_evsel__is_function_event(evsel))
984                 evsel->attr.exclude_callchain_user = 1;
985
986         if (callchain && callchain->enabled && !evsel->no_aux_samples)
987                 perf_evsel__config_callchain(evsel, opts, callchain);
988
989         if (opts->sample_intr_regs) {
990                 attr->sample_regs_intr = opts->sample_intr_regs;
991                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
992         }
993
994         if (opts->sample_user_regs) {
995                 attr->sample_regs_user |= opts->sample_user_regs;
996                 perf_evsel__set_sample_bit(evsel, REGS_USER);
997         }
998
999         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1000                 perf_evsel__set_sample_bit(evsel, CPU);
1001
1002         /*
1003          * When the user explicitly disabled time don't force it here.
1004          */
1005         if (opts->sample_time &&
1006             (!perf_missing_features.sample_id_all &&
1007             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1008              opts->sample_time_set)))
1009                 perf_evsel__set_sample_bit(evsel, TIME);
1010
1011         if (opts->raw_samples && !evsel->no_aux_samples) {
1012                 perf_evsel__set_sample_bit(evsel, TIME);
1013                 perf_evsel__set_sample_bit(evsel, RAW);
1014                 perf_evsel__set_sample_bit(evsel, CPU);
1015         }
1016
1017         if (opts->sample_address)
1018                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1019
1020         if (opts->sample_phys_addr)
1021                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1022
1023         if (opts->no_buffering) {
1024                 attr->watermark = 0;
1025                 attr->wakeup_events = 1;
1026         }
1027         if (opts->branch_stack && !evsel->no_aux_samples) {
1028                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1029                 attr->branch_sample_type = opts->branch_stack;
1030         }
1031
1032         if (opts->sample_weight)
1033                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1034
1035         attr->task  = track;
1036         attr->mmap  = track;
1037         attr->mmap2 = track && !perf_missing_features.mmap2;
1038         attr->comm  = track;
1039         attr->ksymbol = track && !perf_missing_features.ksymbol;
1040         attr->bpf_event = track && !opts->no_bpf_event &&
1041                 !perf_missing_features.bpf_event;
1042
1043         if (opts->record_namespaces)
1044                 attr->namespaces  = track;
1045
1046         if (opts->record_switch_events)
1047                 attr->context_switch = track;
1048
1049         if (opts->sample_transaction)
1050                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1051
1052         if (opts->running_time) {
1053                 evsel->attr.read_format |=
1054                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1055                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1056         }
1057
1058         /*
1059          * XXX see the function comment above
1060          *
1061          * Disabling only independent events or group leaders,
1062          * keeping group members enabled.
1063          */
1064         if (perf_evsel__is_group_leader(evsel))
1065                 attr->disabled = 1;
1066
1067         /*
1068          * Setting enable_on_exec for independent events and
1069          * group leaders for traced executed by perf.
1070          */
1071         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1072                 !opts->initial_delay)
1073                 attr->enable_on_exec = 1;
1074
1075         if (evsel->immediate) {
1076                 attr->disabled = 0;
1077                 attr->enable_on_exec = 0;
1078         }
1079
1080         clockid = opts->clockid;
1081         if (opts->use_clockid) {
1082                 attr->use_clockid = 1;
1083                 attr->clockid = opts->clockid;
1084         }
1085
1086         if (evsel->precise_max)
1087                 attr->precise_ip = 3;
1088
1089         if (opts->all_user) {
1090                 attr->exclude_kernel = 1;
1091                 attr->exclude_user   = 0;
1092         }
1093
1094         if (opts->all_kernel) {
1095                 attr->exclude_kernel = 0;
1096                 attr->exclude_user   = 1;
1097         }
1098
1099         if (evsel->own_cpus || evsel->unit)
1100                 evsel->attr.read_format |= PERF_FORMAT_ID;
1101
1102         /*
1103          * Apply event specific term settings,
1104          * it overloads any global configuration.
1105          */
1106         apply_config_terms(evsel, opts, track);
1107
1108         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1109
1110         /* The --period option takes the precedence. */
1111         if (opts->period_set) {
1112                 if (opts->period)
1113                         perf_evsel__set_sample_bit(evsel, PERIOD);
1114                 else
1115                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1116         }
1117
1118         /*
1119          * For initial_delay, a dummy event is added implicitly.
1120          * The software event will trigger -EOPNOTSUPP error out,
1121          * if BRANCH_STACK bit is set.
1122          */
1123         if (opts->initial_delay && is_dummy_event(evsel))
1124                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1125 }
1126
1127 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1128 {
1129         if (evsel->system_wide)
1130                 nthreads = 1;
1131
1132         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1133
1134         if (evsel->fd) {
1135                 int cpu, thread;
1136                 for (cpu = 0; cpu < ncpus; cpu++) {
1137                         for (thread = 0; thread < nthreads; thread++) {
1138                                 FD(evsel, cpu, thread) = -1;
1139                         }
1140                 }
1141         }
1142
1143         return evsel->fd != NULL ? 0 : -ENOMEM;
1144 }
1145
1146 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1147                           int ioc,  void *arg)
1148 {
1149         int cpu, thread;
1150
1151         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1152                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1153                         int fd = FD(evsel, cpu, thread),
1154                             err = ioctl(fd, ioc, arg);
1155
1156                         if (err)
1157                                 return err;
1158                 }
1159         }
1160
1161         return 0;
1162 }
1163
1164 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1165 {
1166         return perf_evsel__run_ioctl(evsel,
1167                                      PERF_EVENT_IOC_SET_FILTER,
1168                                      (void *)filter);
1169 }
1170
1171 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1172 {
1173         char *new_filter = strdup(filter);
1174
1175         if (new_filter != NULL) {
1176                 free(evsel->filter);
1177                 evsel->filter = new_filter;
1178                 return 0;
1179         }
1180
1181         return -1;
1182 }
1183
1184 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1185                                      const char *fmt, const char *filter)
1186 {
1187         char *new_filter;
1188
1189         if (evsel->filter == NULL)
1190                 return perf_evsel__set_filter(evsel, filter);
1191
1192         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1193                 free(evsel->filter);
1194                 evsel->filter = new_filter;
1195                 return 0;
1196         }
1197
1198         return -1;
1199 }
1200
1201 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1202 {
1203         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1204 }
1205
1206 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1207 {
1208         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1209 }
1210
1211 int perf_evsel__enable(struct perf_evsel *evsel)
1212 {
1213         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1214
1215         if (!err)
1216                 evsel->disabled = false;
1217
1218         return err;
1219 }
1220
1221 int perf_evsel__disable(struct perf_evsel *evsel)
1222 {
1223         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1224         /*
1225          * We mark it disabled here so that tools that disable a event can
1226          * ignore events after they disable it. I.e. the ring buffer may have
1227          * already a few more events queued up before the kernel got the stop
1228          * request.
1229          */
1230         if (!err)
1231                 evsel->disabled = true;
1232
1233         return err;
1234 }
1235
1236 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1237 {
1238         if (ncpus == 0 || nthreads == 0)
1239                 return 0;
1240
1241         if (evsel->system_wide)
1242                 nthreads = 1;
1243
1244         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1245         if (evsel->sample_id == NULL)
1246                 return -ENOMEM;
1247
1248         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1249         if (evsel->id == NULL) {
1250                 xyarray__delete(evsel->sample_id);
1251                 evsel->sample_id = NULL;
1252                 return -ENOMEM;
1253         }
1254
1255         return 0;
1256 }
1257
1258 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1259 {
1260         xyarray__delete(evsel->fd);
1261         evsel->fd = NULL;
1262 }
1263
1264 static void perf_evsel__free_id(struct perf_evsel *evsel)
1265 {
1266         xyarray__delete(evsel->sample_id);
1267         evsel->sample_id = NULL;
1268         zfree(&evsel->id);
1269 }
1270
1271 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1272 {
1273         struct perf_evsel_config_term *term, *h;
1274
1275         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1276                 list_del(&term->list);
1277                 free(term);
1278         }
1279 }
1280
1281 void perf_evsel__close_fd(struct perf_evsel *evsel)
1282 {
1283         int cpu, thread;
1284
1285         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1286                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1287                         close(FD(evsel, cpu, thread));
1288                         FD(evsel, cpu, thread) = -1;
1289                 }
1290 }
1291
1292 void perf_evsel__exit(struct perf_evsel *evsel)
1293 {
1294         assert(list_empty(&evsel->node));
1295         assert(evsel->evlist == NULL);
1296         perf_evsel__free_counts(evsel);
1297         perf_evsel__free_fd(evsel);
1298         perf_evsel__free_id(evsel);
1299         perf_evsel__free_config_terms(evsel);
1300         cgroup__put(evsel->cgrp);
1301         cpu_map__put(evsel->cpus);
1302         cpu_map__put(evsel->own_cpus);
1303         thread_map__put(evsel->threads);
1304         zfree(&evsel->group_name);
1305         zfree(&evsel->name);
1306         perf_evsel__object.fini(evsel);
1307 }
1308
1309 void perf_evsel__delete(struct perf_evsel *evsel)
1310 {
1311         perf_evsel__exit(evsel);
1312         free(evsel);
1313 }
1314
1315 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1316                                 struct perf_counts_values *count)
1317 {
1318         struct perf_counts_values tmp;
1319
1320         if (!evsel->prev_raw_counts)
1321                 return;
1322
1323         if (cpu == -1) {
1324                 tmp = evsel->prev_raw_counts->aggr;
1325                 evsel->prev_raw_counts->aggr = *count;
1326         } else {
1327                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1328                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1329         }
1330
1331         count->val = count->val - tmp.val;
1332         count->ena = count->ena - tmp.ena;
1333         count->run = count->run - tmp.run;
1334 }
1335
1336 void perf_counts_values__scale(struct perf_counts_values *count,
1337                                bool scale, s8 *pscaled)
1338 {
1339         s8 scaled = 0;
1340
1341         if (scale) {
1342                 if (count->run == 0) {
1343                         scaled = -1;
1344                         count->val = 0;
1345                 } else if (count->run < count->ena) {
1346                         scaled = 1;
1347                         count->val = (u64)((double) count->val * count->ena / count->run);
1348                 }
1349         }
1350
1351         if (pscaled)
1352                 *pscaled = scaled;
1353 }
1354
1355 static int perf_evsel__read_size(struct perf_evsel *evsel)
1356 {
1357         u64 read_format = evsel->attr.read_format;
1358         int entry = sizeof(u64); /* value */
1359         int size = 0;
1360         int nr = 1;
1361
1362         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1363                 size += sizeof(u64);
1364
1365         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1366                 size += sizeof(u64);
1367
1368         if (read_format & PERF_FORMAT_ID)
1369                 entry += sizeof(u64);
1370
1371         if (read_format & PERF_FORMAT_GROUP) {
1372                 nr = evsel->nr_members;
1373                 size += sizeof(u64);
1374         }
1375
1376         size += entry * nr;
1377         return size;
1378 }
1379
1380 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1381                      struct perf_counts_values *count)
1382 {
1383         size_t size = perf_evsel__read_size(evsel);
1384
1385         memset(count, 0, sizeof(*count));
1386
1387         if (FD(evsel, cpu, thread) < 0)
1388                 return -EINVAL;
1389
1390         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1391                 return -errno;
1392
1393         return 0;
1394 }
1395
1396 static int
1397 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1398 {
1399         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1400
1401         return perf_evsel__read(evsel, cpu, thread, count);
1402 }
1403
1404 static void
1405 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1406                       u64 val, u64 ena, u64 run)
1407 {
1408         struct perf_counts_values *count;
1409
1410         count = perf_counts(counter->counts, cpu, thread);
1411
1412         count->val    = val;
1413         count->ena    = ena;
1414         count->run    = run;
1415         count->loaded = true;
1416 }
1417
1418 static int
1419 perf_evsel__process_group_data(struct perf_evsel *leader,
1420                                int cpu, int thread, u64 *data)
1421 {
1422         u64 read_format = leader->attr.read_format;
1423         struct sample_read_value *v;
1424         u64 nr, ena = 0, run = 0, i;
1425
1426         nr = *data++;
1427
1428         if (nr != (u64) leader->nr_members)
1429                 return -EINVAL;
1430
1431         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1432                 ena = *data++;
1433
1434         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1435                 run = *data++;
1436
1437         v = (struct sample_read_value *) data;
1438
1439         perf_evsel__set_count(leader, cpu, thread,
1440                               v[0].value, ena, run);
1441
1442         for (i = 1; i < nr; i++) {
1443                 struct perf_evsel *counter;
1444
1445                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1446                 if (!counter)
1447                         return -EINVAL;
1448
1449                 perf_evsel__set_count(counter, cpu, thread,
1450                                       v[i].value, ena, run);
1451         }
1452
1453         return 0;
1454 }
1455
1456 static int
1457 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1458 {
1459         struct perf_stat_evsel *ps = leader->stats;
1460         u64 read_format = leader->attr.read_format;
1461         int size = perf_evsel__read_size(leader);
1462         u64 *data = ps->group_data;
1463
1464         if (!(read_format & PERF_FORMAT_ID))
1465                 return -EINVAL;
1466
1467         if (!perf_evsel__is_group_leader(leader))
1468                 return -EINVAL;
1469
1470         if (!data) {
1471                 data = zalloc(size);
1472                 if (!data)
1473                         return -ENOMEM;
1474
1475                 ps->group_data = data;
1476         }
1477
1478         if (FD(leader, cpu, thread) < 0)
1479                 return -EINVAL;
1480
1481         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1482                 return -errno;
1483
1484         return perf_evsel__process_group_data(leader, cpu, thread, data);
1485 }
1486
1487 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1488 {
1489         u64 read_format = evsel->attr.read_format;
1490
1491         if (read_format & PERF_FORMAT_GROUP)
1492                 return perf_evsel__read_group(evsel, cpu, thread);
1493         else
1494                 return perf_evsel__read_one(evsel, cpu, thread);
1495 }
1496
1497 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1498                               int cpu, int thread, bool scale)
1499 {
1500         struct perf_counts_values count;
1501         size_t nv = scale ? 3 : 1;
1502
1503         if (FD(evsel, cpu, thread) < 0)
1504                 return -EINVAL;
1505
1506         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1507                 return -ENOMEM;
1508
1509         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1510                 return -errno;
1511
1512         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1513         perf_counts_values__scale(&count, scale, NULL);
1514         *perf_counts(evsel->counts, cpu, thread) = count;
1515         return 0;
1516 }
1517
1518 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1519 {
1520         struct perf_evsel *leader = evsel->leader;
1521         int fd;
1522
1523         if (perf_evsel__is_group_leader(evsel))
1524                 return -1;
1525
1526         /*
1527          * Leader must be already processed/open,
1528          * if not it's a bug.
1529          */
1530         BUG_ON(!leader->fd);
1531
1532         fd = FD(leader, cpu, thread);
1533         BUG_ON(fd == -1);
1534
1535         return fd;
1536 }
1537
1538 struct bit_names {
1539         int bit;
1540         const char *name;
1541 };
1542
1543 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1544 {
1545         bool first_bit = true;
1546         int i = 0;
1547
1548         do {
1549                 if (value & bits[i].bit) {
1550                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1551                         first_bit = false;
1552                 }
1553         } while (bits[++i].name != NULL);
1554 }
1555
1556 static void __p_sample_type(char *buf, size_t size, u64 value)
1557 {
1558 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1559         struct bit_names bits[] = {
1560                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1561                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1562                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1563                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1564                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1565                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1566                 { .name = NULL, }
1567         };
1568 #undef bit_name
1569         __p_bits(buf, size, value, bits);
1570 }
1571
1572 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1573 {
1574 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1575         struct bit_names bits[] = {
1576                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1577                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1578                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1579                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1580                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1581                 { .name = NULL, }
1582         };
1583 #undef bit_name
1584         __p_bits(buf, size, value, bits);
1585 }
1586
1587 static void __p_read_format(char *buf, size_t size, u64 value)
1588 {
1589 #define bit_name(n) { PERF_FORMAT_##n, #n }
1590         struct bit_names bits[] = {
1591                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1592                 bit_name(ID), bit_name(GROUP),
1593                 { .name = NULL, }
1594         };
1595 #undef bit_name
1596         __p_bits(buf, size, value, bits);
1597 }
1598
1599 #define BUF_SIZE                1024
1600
1601 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1602 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1603 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1604 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1605 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1606 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1607
1608 #define PRINT_ATTRn(_n, _f, _p)                         \
1609 do {                                                    \
1610         if (attr->_f) {                                 \
1611                 _p(attr->_f);                           \
1612                 ret += attr__fprintf(fp, _n, buf, priv);\
1613         }                                               \
1614 } while (0)
1615
1616 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1617
1618 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1619                              attr__fprintf_f attr__fprintf, void *priv)
1620 {
1621         char buf[BUF_SIZE];
1622         int ret = 0;
1623
1624         PRINT_ATTRf(type, p_unsigned);
1625         PRINT_ATTRf(size, p_unsigned);
1626         PRINT_ATTRf(config, p_hex);
1627         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1628         PRINT_ATTRf(sample_type, p_sample_type);
1629         PRINT_ATTRf(read_format, p_read_format);
1630
1631         PRINT_ATTRf(disabled, p_unsigned);
1632         PRINT_ATTRf(inherit, p_unsigned);
1633         PRINT_ATTRf(pinned, p_unsigned);
1634         PRINT_ATTRf(exclusive, p_unsigned);
1635         PRINT_ATTRf(exclude_user, p_unsigned);
1636         PRINT_ATTRf(exclude_kernel, p_unsigned);
1637         PRINT_ATTRf(exclude_hv, p_unsigned);
1638         PRINT_ATTRf(exclude_idle, p_unsigned);
1639         PRINT_ATTRf(mmap, p_unsigned);
1640         PRINT_ATTRf(comm, p_unsigned);
1641         PRINT_ATTRf(freq, p_unsigned);
1642         PRINT_ATTRf(inherit_stat, p_unsigned);
1643         PRINT_ATTRf(enable_on_exec, p_unsigned);
1644         PRINT_ATTRf(task, p_unsigned);
1645         PRINT_ATTRf(watermark, p_unsigned);
1646         PRINT_ATTRf(precise_ip, p_unsigned);
1647         PRINT_ATTRf(mmap_data, p_unsigned);
1648         PRINT_ATTRf(sample_id_all, p_unsigned);
1649         PRINT_ATTRf(exclude_host, p_unsigned);
1650         PRINT_ATTRf(exclude_guest, p_unsigned);
1651         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1652         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1653         PRINT_ATTRf(mmap2, p_unsigned);
1654         PRINT_ATTRf(comm_exec, p_unsigned);
1655         PRINT_ATTRf(use_clockid, p_unsigned);
1656         PRINT_ATTRf(context_switch, p_unsigned);
1657         PRINT_ATTRf(write_backward, p_unsigned);
1658         PRINT_ATTRf(namespaces, p_unsigned);
1659         PRINT_ATTRf(ksymbol, p_unsigned);
1660         PRINT_ATTRf(bpf_event, p_unsigned);
1661
1662         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1663         PRINT_ATTRf(bp_type, p_unsigned);
1664         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1665         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1666         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1667         PRINT_ATTRf(sample_regs_user, p_hex);
1668         PRINT_ATTRf(sample_stack_user, p_unsigned);
1669         PRINT_ATTRf(clockid, p_signed);
1670         PRINT_ATTRf(sample_regs_intr, p_hex);
1671         PRINT_ATTRf(aux_watermark, p_unsigned);
1672         PRINT_ATTRf(sample_max_stack, p_unsigned);
1673
1674         return ret;
1675 }
1676
1677 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1678                                 void *priv __maybe_unused)
1679 {
1680         return fprintf(fp, "  %-32s %s\n", name, val);
1681 }
1682
1683 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1684                                   int nr_cpus, int nr_threads,
1685                                   int thread_idx)
1686 {
1687         for (int cpu = 0; cpu < nr_cpus; cpu++)
1688                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1689                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1690 }
1691
1692 static int update_fds(struct perf_evsel *evsel,
1693                       int nr_cpus, int cpu_idx,
1694                       int nr_threads, int thread_idx)
1695 {
1696         struct perf_evsel *pos;
1697
1698         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1699                 return -EINVAL;
1700
1701         evlist__for_each_entry(evsel->evlist, pos) {
1702                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1703
1704                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1705
1706                 /*
1707                  * Since fds for next evsel has not been created,
1708                  * there is no need to iterate whole event list.
1709                  */
1710                 if (pos == evsel)
1711                         break;
1712         }
1713         return 0;
1714 }
1715
1716 static bool ignore_missing_thread(struct perf_evsel *evsel,
1717                                   int nr_cpus, int cpu,
1718                                   struct thread_map *threads,
1719                                   int thread, int err)
1720 {
1721         pid_t ignore_pid = thread_map__pid(threads, thread);
1722
1723         if (!evsel->ignore_missing_thread)
1724                 return false;
1725
1726         /* The system wide setup does not work with threads. */
1727         if (evsel->system_wide)
1728                 return false;
1729
1730         /* The -ESRCH is perf event syscall errno for pid's not found. */
1731         if (err != -ESRCH)
1732                 return false;
1733
1734         /* If there's only one thread, let it fail. */
1735         if (threads->nr == 1)
1736                 return false;
1737
1738         /*
1739          * We should remove fd for missing_thread first
1740          * because thread_map__remove() will decrease threads->nr.
1741          */
1742         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1743                 return false;
1744
1745         if (thread_map__remove(threads, thread))
1746                 return false;
1747
1748         pr_warning("WARNING: Ignored open failure for pid %d\n",
1749                    ignore_pid);
1750         return true;
1751 }
1752
1753 static void display_attr(struct perf_event_attr *attr)
1754 {
1755         if (verbose >= 2) {
1756                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1757                 fprintf(stderr, "perf_event_attr:\n");
1758                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1759                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1760         }
1761 }
1762
1763 static int perf_event_open(struct perf_evsel *evsel,
1764                            pid_t pid, int cpu, int group_fd,
1765                            unsigned long flags)
1766 {
1767         int precise_ip = evsel->attr.precise_ip;
1768         int fd;
1769
1770         while (1) {
1771                 pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1772                           pid, cpu, group_fd, flags);
1773
1774                 fd = sys_perf_event_open(&evsel->attr, pid, cpu, group_fd, flags);
1775                 if (fd >= 0)
1776                         break;
1777
1778                 /*
1779                  * Do quick precise_ip fallback if:
1780                  *  - there is precise_ip set in perf_event_attr
1781                  *  - maximum precise is requested
1782                  *  - sys_perf_event_open failed with ENOTSUP error,
1783                  *    which is associated with wrong precise_ip
1784                  */
1785                 if (!precise_ip || !evsel->precise_max || (errno != ENOTSUP))
1786                         break;
1787
1788                 /*
1789                  * We tried all the precise_ip values, and it's
1790                  * still failing, so leave it to standard fallback.
1791                  */
1792                 if (!evsel->attr.precise_ip) {
1793                         evsel->attr.precise_ip = precise_ip;
1794                         break;
1795                 }
1796
1797                 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1798                 evsel->attr.precise_ip--;
1799                 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->attr.precise_ip);
1800                 display_attr(&evsel->attr);
1801         }
1802
1803         return fd;
1804 }
1805
1806 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1807                      struct thread_map *threads)
1808 {
1809         int cpu, thread, nthreads;
1810         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1811         int pid = -1, err;
1812         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1813
1814         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1815                 return -EINVAL;
1816
1817         if (cpus == NULL) {
1818                 static struct cpu_map *empty_cpu_map;
1819
1820                 if (empty_cpu_map == NULL) {
1821                         empty_cpu_map = cpu_map__dummy_new();
1822                         if (empty_cpu_map == NULL)
1823                                 return -ENOMEM;
1824                 }
1825
1826                 cpus = empty_cpu_map;
1827         }
1828
1829         if (threads == NULL) {
1830                 static struct thread_map *empty_thread_map;
1831
1832                 if (empty_thread_map == NULL) {
1833                         empty_thread_map = thread_map__new_by_tid(-1);
1834                         if (empty_thread_map == NULL)
1835                                 return -ENOMEM;
1836                 }
1837
1838                 threads = empty_thread_map;
1839         }
1840
1841         if (evsel->system_wide)
1842                 nthreads = 1;
1843         else
1844                 nthreads = threads->nr;
1845
1846         if (evsel->fd == NULL &&
1847             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1848                 return -ENOMEM;
1849
1850         if (evsel->cgrp) {
1851                 flags |= PERF_FLAG_PID_CGROUP;
1852                 pid = evsel->cgrp->fd;
1853         }
1854
1855 fallback_missing_features:
1856         if (perf_missing_features.clockid_wrong)
1857                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1858         if (perf_missing_features.clockid) {
1859                 evsel->attr.use_clockid = 0;
1860                 evsel->attr.clockid = 0;
1861         }
1862         if (perf_missing_features.cloexec)
1863                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1864         if (perf_missing_features.mmap2)
1865                 evsel->attr.mmap2 = 0;
1866         if (perf_missing_features.exclude_guest)
1867                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1868         if (perf_missing_features.lbr_flags)
1869                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1870                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1871         if (perf_missing_features.group_read && evsel->attr.inherit)
1872                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1873         if (perf_missing_features.ksymbol)
1874                 evsel->attr.ksymbol = 0;
1875         if (perf_missing_features.bpf_event)
1876                 evsel->attr.bpf_event = 0;
1877 retry_sample_id:
1878         if (perf_missing_features.sample_id_all)
1879                 evsel->attr.sample_id_all = 0;
1880
1881         display_attr(&evsel->attr);
1882
1883         for (cpu = 0; cpu < cpus->nr; cpu++) {
1884
1885                 for (thread = 0; thread < nthreads; thread++) {
1886                         int fd, group_fd;
1887
1888                         if (!evsel->cgrp && !evsel->system_wide)
1889                                 pid = thread_map__pid(threads, thread);
1890
1891                         group_fd = get_group_fd(evsel, cpu, thread);
1892 retry_open:
1893                         test_attr__ready();
1894
1895                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1896                                              group_fd, flags);
1897
1898                         FD(evsel, cpu, thread) = fd;
1899
1900                         if (fd < 0) {
1901                                 err = -errno;
1902
1903                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1904                                         /*
1905                                          * We just removed 1 thread, so take a step
1906                                          * back on thread index and lower the upper
1907                                          * nthreads limit.
1908                                          */
1909                                         nthreads--;
1910                                         thread--;
1911
1912                                         /* ... and pretend like nothing have happened. */
1913                                         err = 0;
1914                                         continue;
1915                                 }
1916
1917                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1918                                           err);
1919                                 goto try_fallback;
1920                         }
1921
1922                         pr_debug2(" = %d\n", fd);
1923
1924                         if (evsel->bpf_fd >= 0) {
1925                                 int evt_fd = fd;
1926                                 int bpf_fd = evsel->bpf_fd;
1927
1928                                 err = ioctl(evt_fd,
1929                                             PERF_EVENT_IOC_SET_BPF,
1930                                             bpf_fd);
1931                                 if (err && errno != EEXIST) {
1932                                         pr_err("failed to attach bpf fd %d: %s\n",
1933                                                bpf_fd, strerror(errno));
1934                                         err = -EINVAL;
1935                                         goto out_close;
1936                                 }
1937                         }
1938
1939                         set_rlimit = NO_CHANGE;
1940
1941                         /*
1942                          * If we succeeded but had to kill clockid, fail and
1943                          * have perf_evsel__open_strerror() print us a nice
1944                          * error.
1945                          */
1946                         if (perf_missing_features.clockid ||
1947                             perf_missing_features.clockid_wrong) {
1948                                 err = -EINVAL;
1949                                 goto out_close;
1950                         }
1951                 }
1952         }
1953
1954         return 0;
1955
1956 try_fallback:
1957         /*
1958          * perf stat needs between 5 and 22 fds per CPU. When we run out
1959          * of them try to increase the limits.
1960          */
1961         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1962                 struct rlimit l;
1963                 int old_errno = errno;
1964
1965                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1966                         if (set_rlimit == NO_CHANGE)
1967                                 l.rlim_cur = l.rlim_max;
1968                         else {
1969                                 l.rlim_cur = l.rlim_max + 1000;
1970                                 l.rlim_max = l.rlim_cur;
1971                         }
1972                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1973                                 set_rlimit++;
1974                                 errno = old_errno;
1975                                 goto retry_open;
1976                         }
1977                 }
1978                 errno = old_errno;
1979         }
1980
1981         if (err != -EINVAL || cpu > 0 || thread > 0)
1982                 goto out_close;
1983
1984         /*
1985          * Must probe features in the order they were added to the
1986          * perf_event_attr interface.
1987          */
1988         if (!perf_missing_features.bpf_event && evsel->attr.bpf_event) {
1989                 perf_missing_features.bpf_event = true;
1990                 pr_debug2("switching off bpf_event\n");
1991                 goto fallback_missing_features;
1992         } else if (!perf_missing_features.ksymbol && evsel->attr.ksymbol) {
1993                 perf_missing_features.ksymbol = true;
1994                 pr_debug2("switching off ksymbol\n");
1995                 goto fallback_missing_features;
1996         } else if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1997                 perf_missing_features.write_backward = true;
1998                 pr_debug2("switching off write_backward\n");
1999                 goto out_close;
2000         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
2001                 perf_missing_features.clockid_wrong = true;
2002                 pr_debug2("switching off clockid\n");
2003                 goto fallback_missing_features;
2004         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
2005                 perf_missing_features.clockid = true;
2006                 pr_debug2("switching off use_clockid\n");
2007                 goto fallback_missing_features;
2008         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
2009                 perf_missing_features.cloexec = true;
2010                 pr_debug2("switching off cloexec flag\n");
2011                 goto fallback_missing_features;
2012         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
2013                 perf_missing_features.mmap2 = true;
2014                 pr_debug2("switching off mmap2\n");
2015                 goto fallback_missing_features;
2016         } else if (!perf_missing_features.exclude_guest &&
2017                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
2018                 perf_missing_features.exclude_guest = true;
2019                 pr_debug2("switching off exclude_guest, exclude_host\n");
2020                 goto fallback_missing_features;
2021         } else if (!perf_missing_features.sample_id_all) {
2022                 perf_missing_features.sample_id_all = true;
2023                 pr_debug2("switching off sample_id_all\n");
2024                 goto retry_sample_id;
2025         } else if (!perf_missing_features.lbr_flags &&
2026                         (evsel->attr.branch_sample_type &
2027                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2028                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2029                 perf_missing_features.lbr_flags = true;
2030                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
2031                 goto fallback_missing_features;
2032         } else if (!perf_missing_features.group_read &&
2033                     evsel->attr.inherit &&
2034                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
2035                    perf_evsel__is_group_leader(evsel)) {
2036                 perf_missing_features.group_read = true;
2037                 pr_debug2("switching off group read\n");
2038                 goto fallback_missing_features;
2039         }
2040 out_close:
2041         if (err)
2042                 threads->err_thread = thread;
2043
2044         do {
2045                 while (--thread >= 0) {
2046                         close(FD(evsel, cpu, thread));
2047                         FD(evsel, cpu, thread) = -1;
2048                 }
2049                 thread = nthreads;
2050         } while (--cpu >= 0);
2051         return err;
2052 }
2053
2054 void perf_evsel__close(struct perf_evsel *evsel)
2055 {
2056         if (evsel->fd == NULL)
2057                 return;
2058
2059         perf_evsel__close_fd(evsel);
2060         perf_evsel__free_fd(evsel);
2061 }
2062
2063 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2064                              struct cpu_map *cpus)
2065 {
2066         return perf_evsel__open(evsel, cpus, NULL);
2067 }
2068
2069 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2070                                 struct thread_map *threads)
2071 {
2072         return perf_evsel__open(evsel, NULL, threads);
2073 }
2074
2075 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2076                                        const union perf_event *event,
2077                                        struct perf_sample *sample)
2078 {
2079         u64 type = evsel->attr.sample_type;
2080         const u64 *array = event->sample.array;
2081         bool swapped = evsel->needs_swap;
2082         union u64_swap u;
2083
2084         array += ((event->header.size -
2085                    sizeof(event->header)) / sizeof(u64)) - 1;
2086
2087         if (type & PERF_SAMPLE_IDENTIFIER) {
2088                 sample->id = *array;
2089                 array--;
2090         }
2091
2092         if (type & PERF_SAMPLE_CPU) {
2093                 u.val64 = *array;
2094                 if (swapped) {
2095                         /* undo swap of u64, then swap on individual u32s */
2096                         u.val64 = bswap_64(u.val64);
2097                         u.val32[0] = bswap_32(u.val32[0]);
2098                 }
2099
2100                 sample->cpu = u.val32[0];
2101                 array--;
2102         }
2103
2104         if (type & PERF_SAMPLE_STREAM_ID) {
2105                 sample->stream_id = *array;
2106                 array--;
2107         }
2108
2109         if (type & PERF_SAMPLE_ID) {
2110                 sample->id = *array;
2111                 array--;
2112         }
2113
2114         if (type & PERF_SAMPLE_TIME) {
2115                 sample->time = *array;
2116                 array--;
2117         }
2118
2119         if (type & PERF_SAMPLE_TID) {
2120                 u.val64 = *array;
2121                 if (swapped) {
2122                         /* undo swap of u64, then swap on individual u32s */
2123                         u.val64 = bswap_64(u.val64);
2124                         u.val32[0] = bswap_32(u.val32[0]);
2125                         u.val32[1] = bswap_32(u.val32[1]);
2126                 }
2127
2128                 sample->pid = u.val32[0];
2129                 sample->tid = u.val32[1];
2130                 array--;
2131         }
2132
2133         return 0;
2134 }
2135
2136 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2137                             u64 size)
2138 {
2139         return size > max_size || offset + size > endp;
2140 }
2141
2142 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2143         do {                                                            \
2144                 if (overflow(endp, (max_size), (offset), (size)))       \
2145                         return -EFAULT;                                 \
2146         } while (0)
2147
2148 #define OVERFLOW_CHECK_u64(offset) \
2149         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2150
2151 static int
2152 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2153 {
2154         /*
2155          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2156          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2157          * check the format does not go past the end of the event.
2158          */
2159         if (sample_size + sizeof(event->header) > event->header.size)
2160                 return -EFAULT;
2161
2162         return 0;
2163 }
2164
2165 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2166                              struct perf_sample *data)
2167 {
2168         u64 type = evsel->attr.sample_type;
2169         bool swapped = evsel->needs_swap;
2170         const u64 *array;
2171         u16 max_size = event->header.size;
2172         const void *endp = (void *)event + max_size;
2173         u64 sz;
2174
2175         /*
2176          * used for cross-endian analysis. See git commit 65014ab3
2177          * for why this goofiness is needed.
2178          */
2179         union u64_swap u;
2180
2181         memset(data, 0, sizeof(*data));
2182         data->cpu = data->pid = data->tid = -1;
2183         data->stream_id = data->id = data->time = -1ULL;
2184         data->period = evsel->attr.sample_period;
2185         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2186         data->misc    = event->header.misc;
2187         data->id = -1ULL;
2188         data->data_src = PERF_MEM_DATA_SRC_NONE;
2189
2190         if (event->header.type != PERF_RECORD_SAMPLE) {
2191                 if (!evsel->attr.sample_id_all)
2192                         return 0;
2193                 return perf_evsel__parse_id_sample(evsel, event, data);
2194         }
2195
2196         array = event->sample.array;
2197
2198         if (perf_event__check_size(event, evsel->sample_size))
2199                 return -EFAULT;
2200
2201         if (type & PERF_SAMPLE_IDENTIFIER) {
2202                 data->id = *array;
2203                 array++;
2204         }
2205
2206         if (type & PERF_SAMPLE_IP) {
2207                 data->ip = *array;
2208                 array++;
2209         }
2210
2211         if (type & PERF_SAMPLE_TID) {
2212                 u.val64 = *array;
2213                 if (swapped) {
2214                         /* undo swap of u64, then swap on individual u32s */
2215                         u.val64 = bswap_64(u.val64);
2216                         u.val32[0] = bswap_32(u.val32[0]);
2217                         u.val32[1] = bswap_32(u.val32[1]);
2218                 }
2219
2220                 data->pid = u.val32[0];
2221                 data->tid = u.val32[1];
2222                 array++;
2223         }
2224
2225         if (type & PERF_SAMPLE_TIME) {
2226                 data->time = *array;
2227                 array++;
2228         }
2229
2230         if (type & PERF_SAMPLE_ADDR) {
2231                 data->addr = *array;
2232                 array++;
2233         }
2234
2235         if (type & PERF_SAMPLE_ID) {
2236                 data->id = *array;
2237                 array++;
2238         }
2239
2240         if (type & PERF_SAMPLE_STREAM_ID) {
2241                 data->stream_id = *array;
2242                 array++;
2243         }
2244
2245         if (type & PERF_SAMPLE_CPU) {
2246
2247                 u.val64 = *array;
2248                 if (swapped) {
2249                         /* undo swap of u64, then swap on individual u32s */
2250                         u.val64 = bswap_64(u.val64);
2251                         u.val32[0] = bswap_32(u.val32[0]);
2252                 }
2253
2254                 data->cpu = u.val32[0];
2255                 array++;
2256         }
2257
2258         if (type & PERF_SAMPLE_PERIOD) {
2259                 data->period = *array;
2260                 array++;
2261         }
2262
2263         if (type & PERF_SAMPLE_READ) {
2264                 u64 read_format = evsel->attr.read_format;
2265
2266                 OVERFLOW_CHECK_u64(array);
2267                 if (read_format & PERF_FORMAT_GROUP)
2268                         data->read.group.nr = *array;
2269                 else
2270                         data->read.one.value = *array;
2271
2272                 array++;
2273
2274                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2275                         OVERFLOW_CHECK_u64(array);
2276                         data->read.time_enabled = *array;
2277                         array++;
2278                 }
2279
2280                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2281                         OVERFLOW_CHECK_u64(array);
2282                         data->read.time_running = *array;
2283                         array++;
2284                 }
2285
2286                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2287                 if (read_format & PERF_FORMAT_GROUP) {
2288                         const u64 max_group_nr = UINT64_MAX /
2289                                         sizeof(struct sample_read_value);
2290
2291                         if (data->read.group.nr > max_group_nr)
2292                                 return -EFAULT;
2293                         sz = data->read.group.nr *
2294                              sizeof(struct sample_read_value);
2295                         OVERFLOW_CHECK(array, sz, max_size);
2296                         data->read.group.values =
2297                                         (struct sample_read_value *)array;
2298                         array = (void *)array + sz;
2299                 } else {
2300                         OVERFLOW_CHECK_u64(array);
2301                         data->read.one.id = *array;
2302                         array++;
2303                 }
2304         }
2305
2306         if (evsel__has_callchain(evsel)) {
2307                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2308
2309                 OVERFLOW_CHECK_u64(array);
2310                 data->callchain = (struct ip_callchain *)array++;
2311                 if (data->callchain->nr > max_callchain_nr)
2312                         return -EFAULT;
2313                 sz = data->callchain->nr * sizeof(u64);
2314                 OVERFLOW_CHECK(array, sz, max_size);
2315                 array = (void *)array + sz;
2316         }
2317
2318         if (type & PERF_SAMPLE_RAW) {
2319                 OVERFLOW_CHECK_u64(array);
2320                 u.val64 = *array;
2321
2322                 /*
2323                  * Undo swap of u64, then swap on individual u32s,
2324                  * get the size of the raw area and undo all of the
2325                  * swap. The pevent interface handles endianity by
2326                  * itself.
2327                  */
2328                 if (swapped) {
2329                         u.val64 = bswap_64(u.val64);
2330                         u.val32[0] = bswap_32(u.val32[0]);
2331                         u.val32[1] = bswap_32(u.val32[1]);
2332                 }
2333                 data->raw_size = u.val32[0];
2334
2335                 /*
2336                  * The raw data is aligned on 64bits including the
2337                  * u32 size, so it's safe to use mem_bswap_64.
2338                  */
2339                 if (swapped)
2340                         mem_bswap_64((void *) array, data->raw_size);
2341
2342                 array = (void *)array + sizeof(u32);
2343
2344                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2345                 data->raw_data = (void *)array;
2346                 array = (void *)array + data->raw_size;
2347         }
2348
2349         if (type & PERF_SAMPLE_BRANCH_STACK) {
2350                 const u64 max_branch_nr = UINT64_MAX /
2351                                           sizeof(struct branch_entry);
2352
2353                 OVERFLOW_CHECK_u64(array);
2354                 data->branch_stack = (struct branch_stack *)array++;
2355
2356                 if (data->branch_stack->nr > max_branch_nr)
2357                         return -EFAULT;
2358                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2359                 OVERFLOW_CHECK(array, sz, max_size);
2360                 array = (void *)array + sz;
2361         }
2362
2363         if (type & PERF_SAMPLE_REGS_USER) {
2364                 OVERFLOW_CHECK_u64(array);
2365                 data->user_regs.abi = *array;
2366                 array++;
2367
2368                 if (data->user_regs.abi) {
2369                         u64 mask = evsel->attr.sample_regs_user;
2370
2371                         sz = hweight_long(mask) * sizeof(u64);
2372                         OVERFLOW_CHECK(array, sz, max_size);
2373                         data->user_regs.mask = mask;
2374                         data->user_regs.regs = (u64 *)array;
2375                         array = (void *)array + sz;
2376                 }
2377         }
2378
2379         if (type & PERF_SAMPLE_STACK_USER) {
2380                 OVERFLOW_CHECK_u64(array);
2381                 sz = *array++;
2382
2383                 data->user_stack.offset = ((char *)(array - 1)
2384                                           - (char *) event);
2385
2386                 if (!sz) {
2387                         data->user_stack.size = 0;
2388                 } else {
2389                         OVERFLOW_CHECK(array, sz, max_size);
2390                         data->user_stack.data = (char *)array;
2391                         array = (void *)array + sz;
2392                         OVERFLOW_CHECK_u64(array);
2393                         data->user_stack.size = *array++;
2394                         if (WARN_ONCE(data->user_stack.size > sz,
2395                                       "user stack dump failure\n"))
2396                                 return -EFAULT;
2397                 }
2398         }
2399
2400         if (type & PERF_SAMPLE_WEIGHT) {
2401                 OVERFLOW_CHECK_u64(array);
2402                 data->weight = *array;
2403                 array++;
2404         }
2405
2406         if (type & PERF_SAMPLE_DATA_SRC) {
2407                 OVERFLOW_CHECK_u64(array);
2408                 data->data_src = *array;
2409                 array++;
2410         }
2411
2412         if (type & PERF_SAMPLE_TRANSACTION) {
2413                 OVERFLOW_CHECK_u64(array);
2414                 data->transaction = *array;
2415                 array++;
2416         }
2417
2418         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2419         if (type & PERF_SAMPLE_REGS_INTR) {
2420                 OVERFLOW_CHECK_u64(array);
2421                 data->intr_regs.abi = *array;
2422                 array++;
2423
2424                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2425                         u64 mask = evsel->attr.sample_regs_intr;
2426
2427                         sz = hweight_long(mask) * sizeof(u64);
2428                         OVERFLOW_CHECK(array, sz, max_size);
2429                         data->intr_regs.mask = mask;
2430                         data->intr_regs.regs = (u64 *)array;
2431                         array = (void *)array + sz;
2432                 }
2433         }
2434
2435         data->phys_addr = 0;
2436         if (type & PERF_SAMPLE_PHYS_ADDR) {
2437                 data->phys_addr = *array;
2438                 array++;
2439         }
2440
2441         return 0;
2442 }
2443
2444 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2445                                        union perf_event *event,
2446                                        u64 *timestamp)
2447 {
2448         u64 type = evsel->attr.sample_type;
2449         const u64 *array;
2450
2451         if (!(type & PERF_SAMPLE_TIME))
2452                 return -1;
2453
2454         if (event->header.type != PERF_RECORD_SAMPLE) {
2455                 struct perf_sample data = {
2456                         .time = -1ULL,
2457                 };
2458
2459                 if (!evsel->attr.sample_id_all)
2460                         return -1;
2461                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2462                         return -1;
2463
2464                 *timestamp = data.time;
2465                 return 0;
2466         }
2467
2468         array = event->sample.array;
2469
2470         if (perf_event__check_size(event, evsel->sample_size))
2471                 return -EFAULT;
2472
2473         if (type & PERF_SAMPLE_IDENTIFIER)
2474                 array++;
2475
2476         if (type & PERF_SAMPLE_IP)
2477                 array++;
2478
2479         if (type & PERF_SAMPLE_TID)
2480                 array++;
2481
2482         if (type & PERF_SAMPLE_TIME)
2483                 *timestamp = *array;
2484
2485         return 0;
2486 }
2487
2488 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2489                                      u64 read_format)
2490 {
2491         size_t sz, result = sizeof(struct sample_event);
2492
2493         if (type & PERF_SAMPLE_IDENTIFIER)
2494                 result += sizeof(u64);
2495
2496         if (type & PERF_SAMPLE_IP)
2497                 result += sizeof(u64);
2498
2499         if (type & PERF_SAMPLE_TID)
2500                 result += sizeof(u64);
2501
2502         if (type & PERF_SAMPLE_TIME)
2503                 result += sizeof(u64);
2504
2505         if (type & PERF_SAMPLE_ADDR)
2506                 result += sizeof(u64);
2507
2508         if (type & PERF_SAMPLE_ID)
2509                 result += sizeof(u64);
2510
2511         if (type & PERF_SAMPLE_STREAM_ID)
2512                 result += sizeof(u64);
2513
2514         if (type & PERF_SAMPLE_CPU)
2515                 result += sizeof(u64);
2516
2517         if (type & PERF_SAMPLE_PERIOD)
2518                 result += sizeof(u64);
2519
2520         if (type & PERF_SAMPLE_READ) {
2521                 result += sizeof(u64);
2522                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2523                         result += sizeof(u64);
2524                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2525                         result += sizeof(u64);
2526                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2527                 if (read_format & PERF_FORMAT_GROUP) {
2528                         sz = sample->read.group.nr *
2529                              sizeof(struct sample_read_value);
2530                         result += sz;
2531                 } else {
2532                         result += sizeof(u64);
2533                 }
2534         }
2535
2536         if (type & PERF_SAMPLE_CALLCHAIN) {
2537                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2538                 result += sz;
2539         }
2540
2541         if (type & PERF_SAMPLE_RAW) {
2542                 result += sizeof(u32);
2543                 result += sample->raw_size;
2544         }
2545
2546         if (type & PERF_SAMPLE_BRANCH_STACK) {
2547                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2548                 sz += sizeof(u64);
2549                 result += sz;
2550         }
2551
2552         if (type & PERF_SAMPLE_REGS_USER) {
2553                 if (sample->user_regs.abi) {
2554                         result += sizeof(u64);
2555                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2556                         result += sz;
2557                 } else {
2558                         result += sizeof(u64);
2559                 }
2560         }
2561
2562         if (type & PERF_SAMPLE_STACK_USER) {
2563                 sz = sample->user_stack.size;
2564                 result += sizeof(u64);
2565                 if (sz) {
2566                         result += sz;
2567                         result += sizeof(u64);
2568                 }
2569         }
2570
2571         if (type & PERF_SAMPLE_WEIGHT)
2572                 result += sizeof(u64);
2573
2574         if (type & PERF_SAMPLE_DATA_SRC)
2575                 result += sizeof(u64);
2576
2577         if (type & PERF_SAMPLE_TRANSACTION)
2578                 result += sizeof(u64);
2579
2580         if (type & PERF_SAMPLE_REGS_INTR) {
2581                 if (sample->intr_regs.abi) {
2582                         result += sizeof(u64);
2583                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2584                         result += sz;
2585                 } else {
2586                         result += sizeof(u64);
2587                 }
2588         }
2589
2590         if (type & PERF_SAMPLE_PHYS_ADDR)
2591                 result += sizeof(u64);
2592
2593         return result;
2594 }
2595
2596 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2597                                   u64 read_format,
2598                                   const struct perf_sample *sample)
2599 {
2600         u64 *array;
2601         size_t sz;
2602         /*
2603          * used for cross-endian analysis. See git commit 65014ab3
2604          * for why this goofiness is needed.
2605          */
2606         union u64_swap u;
2607
2608         array = event->sample.array;
2609
2610         if (type & PERF_SAMPLE_IDENTIFIER) {
2611                 *array = sample->id;
2612                 array++;
2613         }
2614
2615         if (type & PERF_SAMPLE_IP) {
2616                 *array = sample->ip;
2617                 array++;
2618         }
2619
2620         if (type & PERF_SAMPLE_TID) {
2621                 u.val32[0] = sample->pid;
2622                 u.val32[1] = sample->tid;
2623                 *array = u.val64;
2624                 array++;
2625         }
2626
2627         if (type & PERF_SAMPLE_TIME) {
2628                 *array = sample->time;
2629                 array++;
2630         }
2631
2632         if (type & PERF_SAMPLE_ADDR) {
2633                 *array = sample->addr;
2634                 array++;
2635         }
2636
2637         if (type & PERF_SAMPLE_ID) {
2638                 *array = sample->id;
2639                 array++;
2640         }
2641
2642         if (type & PERF_SAMPLE_STREAM_ID) {
2643                 *array = sample->stream_id;
2644                 array++;
2645         }
2646
2647         if (type & PERF_SAMPLE_CPU) {
2648                 u.val32[0] = sample->cpu;
2649                 u.val32[1] = 0;
2650                 *array = u.val64;
2651                 array++;
2652         }
2653
2654         if (type & PERF_SAMPLE_PERIOD) {
2655                 *array = sample->period;
2656                 array++;
2657         }
2658
2659         if (type & PERF_SAMPLE_READ) {
2660                 if (read_format & PERF_FORMAT_GROUP)
2661                         *array = sample->read.group.nr;
2662                 else
2663                         *array = sample->read.one.value;
2664                 array++;
2665
2666                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2667                         *array = sample->read.time_enabled;
2668                         array++;
2669                 }
2670
2671                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2672                         *array = sample->read.time_running;
2673                         array++;
2674                 }
2675
2676                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2677                 if (read_format & PERF_FORMAT_GROUP) {
2678                         sz = sample->read.group.nr *
2679                              sizeof(struct sample_read_value);
2680                         memcpy(array, sample->read.group.values, sz);
2681                         array = (void *)array + sz;
2682                 } else {
2683                         *array = sample->read.one.id;
2684                         array++;
2685                 }
2686         }
2687
2688         if (type & PERF_SAMPLE_CALLCHAIN) {
2689                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2690                 memcpy(array, sample->callchain, sz);
2691                 array = (void *)array + sz;
2692         }
2693
2694         if (type & PERF_SAMPLE_RAW) {
2695                 u.val32[0] = sample->raw_size;
2696                 *array = u.val64;
2697                 array = (void *)array + sizeof(u32);
2698
2699                 memcpy(array, sample->raw_data, sample->raw_size);
2700                 array = (void *)array + sample->raw_size;
2701         }
2702
2703         if (type & PERF_SAMPLE_BRANCH_STACK) {
2704                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2705                 sz += sizeof(u64);
2706                 memcpy(array, sample->branch_stack, sz);
2707                 array = (void *)array + sz;
2708         }
2709
2710         if (type & PERF_SAMPLE_REGS_USER) {
2711                 if (sample->user_regs.abi) {
2712                         *array++ = sample->user_regs.abi;
2713                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2714                         memcpy(array, sample->user_regs.regs, sz);
2715                         array = (void *)array + sz;
2716                 } else {
2717                         *array++ = 0;
2718                 }
2719         }
2720
2721         if (type & PERF_SAMPLE_STACK_USER) {
2722                 sz = sample->user_stack.size;
2723                 *array++ = sz;
2724                 if (sz) {
2725                         memcpy(array, sample->user_stack.data, sz);
2726                         array = (void *)array + sz;
2727                         *array++ = sz;
2728                 }
2729         }
2730
2731         if (type & PERF_SAMPLE_WEIGHT) {
2732                 *array = sample->weight;
2733                 array++;
2734         }
2735
2736         if (type & PERF_SAMPLE_DATA_SRC) {
2737                 *array = sample->data_src;
2738                 array++;
2739         }
2740
2741         if (type & PERF_SAMPLE_TRANSACTION) {
2742                 *array = sample->transaction;
2743                 array++;
2744         }
2745
2746         if (type & PERF_SAMPLE_REGS_INTR) {
2747                 if (sample->intr_regs.abi) {
2748                         *array++ = sample->intr_regs.abi;
2749                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2750                         memcpy(array, sample->intr_regs.regs, sz);
2751                         array = (void *)array + sz;
2752                 } else {
2753                         *array++ = 0;
2754                 }
2755         }
2756
2757         if (type & PERF_SAMPLE_PHYS_ADDR) {
2758                 *array = sample->phys_addr;
2759                 array++;
2760         }
2761
2762         return 0;
2763 }
2764
2765 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2766 {
2767         return tep_find_field(evsel->tp_format, name);
2768 }
2769
2770 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2771                          const char *name)
2772 {
2773         struct tep_format_field *field = perf_evsel__field(evsel, name);
2774         int offset;
2775
2776         if (!field)
2777                 return NULL;
2778
2779         offset = field->offset;
2780
2781         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2782                 offset = *(int *)(sample->raw_data + field->offset);
2783                 offset &= 0xffff;
2784         }
2785
2786         return sample->raw_data + offset;
2787 }
2788
2789 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2790                          bool needs_swap)
2791 {
2792         u64 value;
2793         void *ptr = sample->raw_data + field->offset;
2794
2795         switch (field->size) {
2796         case 1:
2797                 return *(u8 *)ptr;
2798         case 2:
2799                 value = *(u16 *)ptr;
2800                 break;
2801         case 4:
2802                 value = *(u32 *)ptr;
2803                 break;
2804         case 8:
2805                 memcpy(&value, ptr, sizeof(u64));
2806                 break;
2807         default:
2808                 return 0;
2809         }
2810
2811         if (!needs_swap)
2812                 return value;
2813
2814         switch (field->size) {
2815         case 2:
2816                 return bswap_16(value);
2817         case 4:
2818                 return bswap_32(value);
2819         case 8:
2820                 return bswap_64(value);
2821         default:
2822                 return 0;
2823         }
2824
2825         return 0;
2826 }
2827
2828 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2829                        const char *name)
2830 {
2831         struct tep_format_field *field = perf_evsel__field(evsel, name);
2832
2833         if (!field)
2834                 return 0;
2835
2836         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2837 }
2838
2839 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2840                           char *msg, size_t msgsize)
2841 {
2842         int paranoid;
2843
2844         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2845             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2846             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2847                 /*
2848                  * If it's cycles then fall back to hrtimer based
2849                  * cpu-clock-tick sw counter, which is always available even if
2850                  * no PMU support.
2851                  *
2852                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2853                  * b0a873e).
2854                  */
2855                 scnprintf(msg, msgsize, "%s",
2856 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2857
2858                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2859                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2860
2861                 zfree(&evsel->name);
2862                 return true;
2863         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2864                    (paranoid = perf_event_paranoid()) > 1) {
2865                 const char *name = perf_evsel__name(evsel);
2866                 char *new_name;
2867                 const char *sep = ":";
2868
2869                 /* Is there already the separator in the name. */
2870                 if (strchr(name, '/') ||
2871                     strchr(name, ':'))
2872                         sep = "";
2873
2874                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2875                         return false;
2876
2877                 if (evsel->name)
2878                         free(evsel->name);
2879                 evsel->name = new_name;
2880                 scnprintf(msg, msgsize,
2881 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2882                 evsel->attr.exclude_kernel = 1;
2883
2884                 return true;
2885         }
2886
2887         return false;
2888 }
2889
2890 static bool find_process(const char *name)
2891 {
2892         size_t len = strlen(name);
2893         DIR *dir;
2894         struct dirent *d;
2895         int ret = -1;
2896
2897         dir = opendir(procfs__mountpoint());
2898         if (!dir)
2899                 return false;
2900
2901         /* Walk through the directory. */
2902         while (ret && (d = readdir(dir)) != NULL) {
2903                 char path[PATH_MAX];
2904                 char *data;
2905                 size_t size;
2906
2907                 if ((d->d_type != DT_DIR) ||
2908                      !strcmp(".", d->d_name) ||
2909                      !strcmp("..", d->d_name))
2910                         continue;
2911
2912                 scnprintf(path, sizeof(path), "%s/%s/comm",
2913                           procfs__mountpoint(), d->d_name);
2914
2915                 if (filename__read_str(path, &data, &size))
2916                         continue;
2917
2918                 ret = strncmp(name, data, len);
2919                 free(data);
2920         }
2921
2922         closedir(dir);
2923         return ret ? false : true;
2924 }
2925
2926 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2927                               int err, char *msg, size_t size)
2928 {
2929         char sbuf[STRERR_BUFSIZE];
2930         int printed = 0;
2931
2932         switch (err) {
2933         case EPERM:
2934         case EACCES:
2935                 if (err == EPERM)
2936                         printed = scnprintf(msg, size,
2937                                 "No permission to enable %s event.\n\n",
2938                                 perf_evsel__name(evsel));
2939
2940                 return scnprintf(msg + printed, size - printed,
2941                  "You may not have permission to collect %sstats.\n\n"
2942                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2943                  "which controls use of the performance events system by\n"
2944                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2945                  "The current value is %d:\n\n"
2946                  "  -1: Allow use of (almost) all events by all users\n"
2947                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2948                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2949                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2950                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2951                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2952                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2953                  "      kernel.perf_event_paranoid = -1\n" ,
2954                                  target->system_wide ? "system-wide " : "",
2955                                  perf_event_paranoid());
2956         case ENOENT:
2957                 return scnprintf(msg, size, "The %s event is not supported.",
2958                                  perf_evsel__name(evsel));
2959         case EMFILE:
2960                 return scnprintf(msg, size, "%s",
2961                          "Too many events are opened.\n"
2962                          "Probably the maximum number of open file descriptors has been reached.\n"
2963                          "Hint: Try again after reducing the number of events.\n"
2964                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2965         case ENOMEM:
2966                 if (evsel__has_callchain(evsel) &&
2967                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2968                         return scnprintf(msg, size,
2969                                          "Not enough memory to setup event with callchain.\n"
2970                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2971                                          "Hint: Current value: %d", sysctl__max_stack());
2972                 break;
2973         case ENODEV:
2974                 if (target->cpu_list)
2975                         return scnprintf(msg, size, "%s",
2976          "No such device - did you specify an out-of-range profile CPU?");
2977                 break;
2978         case EOPNOTSUPP:
2979                 if (evsel->attr.sample_period != 0)
2980                         return scnprintf(msg, size,
2981         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2982                                          perf_evsel__name(evsel));
2983                 if (evsel->attr.precise_ip)
2984                         return scnprintf(msg, size, "%s",
2985         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2986 #if defined(__i386__) || defined(__x86_64__)
2987                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2988                         return scnprintf(msg, size, "%s",
2989         "No hardware sampling interrupt available.\n");
2990 #endif
2991                 break;
2992         case EBUSY:
2993                 if (find_process("oprofiled"))
2994                         return scnprintf(msg, size,
2995         "The PMU counters are busy/taken by another profiler.\n"
2996         "We found oprofile daemon running, please stop it and try again.");
2997                 break;
2998         case EINVAL:
2999                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
3000                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3001                 if (perf_missing_features.clockid)
3002                         return scnprintf(msg, size, "clockid feature not supported.");
3003                 if (perf_missing_features.clockid_wrong)
3004                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3005                 break;
3006         default:
3007                 break;
3008         }
3009
3010         return scnprintf(msg, size,
3011         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3012         "/bin/dmesg | grep -i perf may provide additional information.\n",
3013                          err, str_error_r(err, sbuf, sizeof(sbuf)),
3014                          perf_evsel__name(evsel));
3015 }
3016
3017 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
3018 {
3019         if (evsel && evsel->evlist)
3020                 return evsel->evlist->env;
3021         return NULL;
3022 }
3023
3024 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3025 {
3026         int cpu, thread;
3027
3028         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
3029                 for (thread = 0; thread < xyarray__max_y(evsel->fd);
3030                      thread++) {
3031                         int fd = FD(evsel, cpu, thread);
3032
3033                         if (perf_evlist__id_add_fd(evlist, evsel,
3034                                                    cpu, thread, fd) < 0)
3035                                 return -1;
3036                 }
3037         }
3038
3039         return 0;
3040 }
3041
3042 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3043 {
3044         struct cpu_map *cpus = evsel->cpus;
3045         struct thread_map *threads = evsel->threads;
3046
3047         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
3048                 return -ENOMEM;
3049
3050         return store_evsel_ids(evsel, evlist);
3051 }