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