ASoC: codecs: rt5670: add jack detection quirk for Dell Venue 5585
[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 "util/parse-branch-options.h"
40
41 #include "sane_ctype.h"
42
43 static struct {
44         bool sample_id_all;
45         bool exclude_guest;
46         bool mmap2;
47         bool cloexec;
48         bool clockid;
49         bool clockid_wrong;
50         bool lbr_flags;
51         bool write_backward;
52 } perf_missing_features;
53
54 static clockid_t clockid;
55
56 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
57 {
58         return 0;
59 }
60
61 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
62 {
63 }
64
65 static struct {
66         size_t  size;
67         int     (*init)(struct perf_evsel *evsel);
68         void    (*fini)(struct perf_evsel *evsel);
69 } perf_evsel__object = {
70         .size = sizeof(struct perf_evsel),
71         .init = perf_evsel__no_extra_init,
72         .fini = perf_evsel__no_extra_fini,
73 };
74
75 int perf_evsel__object_config(size_t object_size,
76                               int (*init)(struct perf_evsel *evsel),
77                               void (*fini)(struct perf_evsel *evsel))
78 {
79
80         if (object_size == 0)
81                 goto set_methods;
82
83         if (perf_evsel__object.size > object_size)
84                 return -EINVAL;
85
86         perf_evsel__object.size = object_size;
87
88 set_methods:
89         if (init != NULL)
90                 perf_evsel__object.init = init;
91
92         if (fini != NULL)
93                 perf_evsel__object.fini = fini;
94
95         return 0;
96 }
97
98 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
99
100 int __perf_evsel__sample_size(u64 sample_type)
101 {
102         u64 mask = sample_type & PERF_SAMPLE_MASK;
103         int size = 0;
104         int i;
105
106         for (i = 0; i < 64; i++) {
107                 if (mask & (1ULL << i))
108                         size++;
109         }
110
111         size *= sizeof(u64);
112
113         return size;
114 }
115
116 /**
117  * __perf_evsel__calc_id_pos - calculate id_pos.
118  * @sample_type: sample type
119  *
120  * This function returns the position of the event id (PERF_SAMPLE_ID or
121  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
122  * sample_event.
123  */
124 static int __perf_evsel__calc_id_pos(u64 sample_type)
125 {
126         int idx = 0;
127
128         if (sample_type & PERF_SAMPLE_IDENTIFIER)
129                 return 0;
130
131         if (!(sample_type & PERF_SAMPLE_ID))
132                 return -1;
133
134         if (sample_type & PERF_SAMPLE_IP)
135                 idx += 1;
136
137         if (sample_type & PERF_SAMPLE_TID)
138                 idx += 1;
139
140         if (sample_type & PERF_SAMPLE_TIME)
141                 idx += 1;
142
143         if (sample_type & PERF_SAMPLE_ADDR)
144                 idx += 1;
145
146         return idx;
147 }
148
149 /**
150  * __perf_evsel__calc_is_pos - calculate is_pos.
151  * @sample_type: sample type
152  *
153  * This function returns the position (counting backwards) of the event id
154  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
155  * sample_id_all is used there is an id sample appended to non-sample events.
156  */
157 static int __perf_evsel__calc_is_pos(u64 sample_type)
158 {
159         int idx = 1;
160
161         if (sample_type & PERF_SAMPLE_IDENTIFIER)
162                 return 1;
163
164         if (!(sample_type & PERF_SAMPLE_ID))
165                 return -1;
166
167         if (sample_type & PERF_SAMPLE_CPU)
168                 idx += 1;
169
170         if (sample_type & PERF_SAMPLE_STREAM_ID)
171                 idx += 1;
172
173         return idx;
174 }
175
176 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
177 {
178         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
179         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
180 }
181
182 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
183                                   enum perf_event_sample_format bit)
184 {
185         if (!(evsel->attr.sample_type & bit)) {
186                 evsel->attr.sample_type |= bit;
187                 evsel->sample_size += sizeof(u64);
188                 perf_evsel__calc_id_pos(evsel);
189         }
190 }
191
192 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
193                                     enum perf_event_sample_format bit)
194 {
195         if (evsel->attr.sample_type & bit) {
196                 evsel->attr.sample_type &= ~bit;
197                 evsel->sample_size -= sizeof(u64);
198                 perf_evsel__calc_id_pos(evsel);
199         }
200 }
201
202 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
203                                bool can_sample_identifier)
204 {
205         if (can_sample_identifier) {
206                 perf_evsel__reset_sample_bit(evsel, ID);
207                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
208         } else {
209                 perf_evsel__set_sample_bit(evsel, ID);
210         }
211         evsel->attr.read_format |= PERF_FORMAT_ID;
212 }
213
214 /**
215  * perf_evsel__is_function_event - Return whether given evsel is a function
216  * trace event
217  *
218  * @evsel - evsel selector to be tested
219  *
220  * Return %true if event is function trace event
221  */
222 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
223 {
224 #define FUNCTION_EVENT "ftrace:function"
225
226         return evsel->name &&
227                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
228
229 #undef FUNCTION_EVENT
230 }
231
232 void perf_evsel__init(struct perf_evsel *evsel,
233                       struct perf_event_attr *attr, int idx)
234 {
235         evsel->idx         = idx;
236         evsel->tracking    = !idx;
237         evsel->attr        = *attr;
238         evsel->leader      = evsel;
239         evsel->unit        = "";
240         evsel->scale       = 1.0;
241         evsel->evlist      = NULL;
242         evsel->bpf_fd      = -1;
243         INIT_LIST_HEAD(&evsel->node);
244         INIT_LIST_HEAD(&evsel->config_terms);
245         perf_evsel__object.init(evsel);
246         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
247         perf_evsel__calc_id_pos(evsel);
248         evsel->cmdline_group_boundary = false;
249         evsel->metric_expr   = NULL;
250         evsel->metric_name   = NULL;
251         evsel->metric_events = NULL;
252         evsel->collect_stat  = false;
253 }
254
255 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
256 {
257         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
258
259         if (evsel != NULL)
260                 perf_evsel__init(evsel, attr, idx);
261
262         if (perf_evsel__is_bpf_output(evsel)) {
263                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
264                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
265                 evsel->attr.sample_period = 1;
266         }
267
268         return evsel;
269 }
270
271 struct perf_evsel *perf_evsel__new_cycles(void)
272 {
273         struct perf_event_attr attr = {
274                 .type   = PERF_TYPE_HARDWARE,
275                 .config = PERF_COUNT_HW_CPU_CYCLES,
276                 .exclude_kernel = 1,
277         };
278         struct perf_evsel *evsel;
279
280         event_attr_init(&attr);
281         /*
282          * Unnamed union member, not supported as struct member named
283          * initializer in older compilers such as gcc 4.4.7
284          *
285          * Just for probing the precise_ip:
286          */
287         attr.sample_period = 1;
288
289         perf_event_attr__set_max_precise_ip(&attr);
290         /*
291          * Now let the usual logic to set up the perf_event_attr defaults
292          * to kick in when we return and before perf_evsel__open() is called.
293          */
294         attr.sample_period = 0;
295
296         evsel = perf_evsel__new(&attr);
297         if (evsel == NULL)
298                 goto out;
299
300         /* use asprintf() because free(evsel) assumes name is allocated */
301         if (asprintf(&evsel->name, "cycles%.*s",
302                      attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
303                 goto error_free;
304 out:
305         return evsel;
306 error_free:
307         perf_evsel__delete(evsel);
308         evsel = NULL;
309         goto out;
310 }
311
312 /*
313  * Returns pointer with encoded error via <linux/err.h> interface.
314  */
315 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
316 {
317         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
318         int err = -ENOMEM;
319
320         if (evsel == NULL) {
321                 goto out_err;
322         } else {
323                 struct perf_event_attr attr = {
324                         .type          = PERF_TYPE_TRACEPOINT,
325                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
326                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
327                 };
328
329                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
330                         goto out_free;
331
332                 evsel->tp_format = trace_event__tp_format(sys, name);
333                 if (IS_ERR(evsel->tp_format)) {
334                         err = PTR_ERR(evsel->tp_format);
335                         goto out_free;
336                 }
337
338                 event_attr_init(&attr);
339                 attr.config = evsel->tp_format->id;
340                 attr.sample_period = 1;
341                 perf_evsel__init(evsel, &attr, idx);
342         }
343
344         return evsel;
345
346 out_free:
347         zfree(&evsel->name);
348         free(evsel);
349 out_err:
350         return ERR_PTR(err);
351 }
352
353 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
354         "cycles",
355         "instructions",
356         "cache-references",
357         "cache-misses",
358         "branches",
359         "branch-misses",
360         "bus-cycles",
361         "stalled-cycles-frontend",
362         "stalled-cycles-backend",
363         "ref-cycles",
364 };
365
366 static const char *__perf_evsel__hw_name(u64 config)
367 {
368         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
369                 return perf_evsel__hw_names[config];
370
371         return "unknown-hardware";
372 }
373
374 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
375 {
376         int colon = 0, r = 0;
377         struct perf_event_attr *attr = &evsel->attr;
378         bool exclude_guest_default = false;
379
380 #define MOD_PRINT(context, mod) do {                                    \
381                 if (!attr->exclude_##context) {                         \
382                         if (!colon) colon = ++r;                        \
383                         r += scnprintf(bf + r, size - r, "%c", mod);    \
384                 } } while(0)
385
386         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
387                 MOD_PRINT(kernel, 'k');
388                 MOD_PRINT(user, 'u');
389                 MOD_PRINT(hv, 'h');
390                 exclude_guest_default = true;
391         }
392
393         if (attr->precise_ip) {
394                 if (!colon)
395                         colon = ++r;
396                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
397                 exclude_guest_default = true;
398         }
399
400         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
401                 MOD_PRINT(host, 'H');
402                 MOD_PRINT(guest, 'G');
403         }
404 #undef MOD_PRINT
405         if (colon)
406                 bf[colon - 1] = ':';
407         return r;
408 }
409
410 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
411 {
412         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
413         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
414 }
415
416 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
417         "cpu-clock",
418         "task-clock",
419         "page-faults",
420         "context-switches",
421         "cpu-migrations",
422         "minor-faults",
423         "major-faults",
424         "alignment-faults",
425         "emulation-faults",
426         "dummy",
427 };
428
429 static const char *__perf_evsel__sw_name(u64 config)
430 {
431         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
432                 return perf_evsel__sw_names[config];
433         return "unknown-software";
434 }
435
436 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
437 {
438         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
439         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
440 }
441
442 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
443 {
444         int r;
445
446         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
447
448         if (type & HW_BREAKPOINT_R)
449                 r += scnprintf(bf + r, size - r, "r");
450
451         if (type & HW_BREAKPOINT_W)
452                 r += scnprintf(bf + r, size - r, "w");
453
454         if (type & HW_BREAKPOINT_X)
455                 r += scnprintf(bf + r, size - r, "x");
456
457         return r;
458 }
459
460 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
461 {
462         struct perf_event_attr *attr = &evsel->attr;
463         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
464         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
465 }
466
467 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
468                                 [PERF_EVSEL__MAX_ALIASES] = {
469  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
470  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
471  { "LLC",       "L2",                                                   },
472  { "dTLB",      "d-tlb",        "Data-TLB",                             },
473  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
474  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
475  { "node",                                                              },
476 };
477
478 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
479                                    [PERF_EVSEL__MAX_ALIASES] = {
480  { "load",      "loads",        "read",                                 },
481  { "store",     "stores",       "write",                                },
482  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
483 };
484
485 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
486                                        [PERF_EVSEL__MAX_ALIASES] = {
487  { "refs",      "Reference",    "ops",          "access",               },
488  { "misses",    "miss",                                                 },
489 };
490
491 #define C(x)            PERF_COUNT_HW_CACHE_##x
492 #define CACHE_READ      (1 << C(OP_READ))
493 #define CACHE_WRITE     (1 << C(OP_WRITE))
494 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
495 #define COP(x)          (1 << x)
496
497 /*
498  * cache operartion stat
499  * L1I : Read and prefetch only
500  * ITLB and BPU : Read-only
501  */
502 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
503  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
504  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
505  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
506  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
507  [C(ITLB)]      = (CACHE_READ),
508  [C(BPU)]       = (CACHE_READ),
509  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
510 };
511
512 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
513 {
514         if (perf_evsel__hw_cache_stat[type] & COP(op))
515                 return true;    /* valid */
516         else
517                 return false;   /* invalid */
518 }
519
520 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
521                                             char *bf, size_t size)
522 {
523         if (result) {
524                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
525                                  perf_evsel__hw_cache_op[op][0],
526                                  perf_evsel__hw_cache_result[result][0]);
527         }
528
529         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
530                          perf_evsel__hw_cache_op[op][1]);
531 }
532
533 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
534 {
535         u8 op, result, type = (config >>  0) & 0xff;
536         const char *err = "unknown-ext-hardware-cache-type";
537
538         if (type >= PERF_COUNT_HW_CACHE_MAX)
539                 goto out_err;
540
541         op = (config >>  8) & 0xff;
542         err = "unknown-ext-hardware-cache-op";
543         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
544                 goto out_err;
545
546         result = (config >> 16) & 0xff;
547         err = "unknown-ext-hardware-cache-result";
548         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
549                 goto out_err;
550
551         err = "invalid-cache";
552         if (!perf_evsel__is_cache_op_valid(type, op))
553                 goto out_err;
554
555         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
556 out_err:
557         return scnprintf(bf, size, "%s", err);
558 }
559
560 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
561 {
562         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
563         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
564 }
565
566 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
567 {
568         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
569         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
570 }
571
572 const char *perf_evsel__name(struct perf_evsel *evsel)
573 {
574         char bf[128];
575
576         if (evsel->name)
577                 return evsel->name;
578
579         switch (evsel->attr.type) {
580         case PERF_TYPE_RAW:
581                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
582                 break;
583
584         case PERF_TYPE_HARDWARE:
585                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
586                 break;
587
588         case PERF_TYPE_HW_CACHE:
589                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
590                 break;
591
592         case PERF_TYPE_SOFTWARE:
593                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
594                 break;
595
596         case PERF_TYPE_TRACEPOINT:
597                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
598                 break;
599
600         case PERF_TYPE_BREAKPOINT:
601                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
602                 break;
603
604         default:
605                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
606                           evsel->attr.type);
607                 break;
608         }
609
610         evsel->name = strdup(bf);
611
612         return evsel->name ?: "unknown";
613 }
614
615 const char *perf_evsel__group_name(struct perf_evsel *evsel)
616 {
617         return evsel->group_name ?: "anon group";
618 }
619
620 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
621 {
622         int ret;
623         struct perf_evsel *pos;
624         const char *group_name = perf_evsel__group_name(evsel);
625
626         ret = scnprintf(buf, size, "%s", group_name);
627
628         ret += scnprintf(buf + ret, size - ret, " { %s",
629                          perf_evsel__name(evsel));
630
631         for_each_group_member(pos, evsel)
632                 ret += scnprintf(buf + ret, size - ret, ", %s",
633                                  perf_evsel__name(pos));
634
635         ret += scnprintf(buf + ret, size - ret, " }");
636
637         return ret;
638 }
639
640 void perf_evsel__config_callchain(struct perf_evsel *evsel,
641                                   struct record_opts *opts,
642                                   struct callchain_param *param)
643 {
644         bool function = perf_evsel__is_function_event(evsel);
645         struct perf_event_attr *attr = &evsel->attr;
646
647         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
648
649         attr->sample_max_stack = param->max_stack;
650
651         if (param->record_mode == CALLCHAIN_LBR) {
652                 if (!opts->branch_stack) {
653                         if (attr->exclude_user) {
654                                 pr_warning("LBR callstack option is only available "
655                                            "to get user callchain information. "
656                                            "Falling back to framepointers.\n");
657                         } else {
658                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
659                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
660                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
661                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
662                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
663                         }
664                 } else
665                          pr_warning("Cannot use LBR callstack with branch stack. "
666                                     "Falling back to framepointers.\n");
667         }
668
669         if (param->record_mode == CALLCHAIN_DWARF) {
670                 if (!function) {
671                         perf_evsel__set_sample_bit(evsel, REGS_USER);
672                         perf_evsel__set_sample_bit(evsel, STACK_USER);
673                         attr->sample_regs_user = PERF_REGS_MASK;
674                         attr->sample_stack_user = param->dump_size;
675                         attr->exclude_callchain_user = 1;
676                 } else {
677                         pr_info("Cannot use DWARF unwind for function trace event,"
678                                 " falling back to framepointers.\n");
679                 }
680         }
681
682         if (function) {
683                 pr_info("Disabling user space callchains for function trace event.\n");
684                 attr->exclude_callchain_user = 1;
685         }
686 }
687
688 static void
689 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
690                             struct callchain_param *param)
691 {
692         struct perf_event_attr *attr = &evsel->attr;
693
694         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
695         if (param->record_mode == CALLCHAIN_LBR) {
696                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
697                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
698                                               PERF_SAMPLE_BRANCH_CALL_STACK);
699         }
700         if (param->record_mode == CALLCHAIN_DWARF) {
701                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
702                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
703         }
704 }
705
706 static void apply_config_terms(struct perf_evsel *evsel,
707                                struct record_opts *opts)
708 {
709         struct perf_evsel_config_term *term;
710         struct list_head *config_terms = &evsel->config_terms;
711         struct perf_event_attr *attr = &evsel->attr;
712         struct callchain_param param;
713         u32 dump_size = 0;
714         int max_stack = 0;
715         const char *callgraph_buf = NULL;
716
717         /* callgraph default */
718         param.record_mode = callchain_param.record_mode;
719
720         list_for_each_entry(term, config_terms, list) {
721                 switch (term->type) {
722                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
723                         attr->sample_period = term->val.period;
724                         attr->freq = 0;
725                         break;
726                 case PERF_EVSEL__CONFIG_TERM_FREQ:
727                         attr->sample_freq = term->val.freq;
728                         attr->freq = 1;
729                         break;
730                 case PERF_EVSEL__CONFIG_TERM_TIME:
731                         if (term->val.time)
732                                 perf_evsel__set_sample_bit(evsel, TIME);
733                         else
734                                 perf_evsel__reset_sample_bit(evsel, TIME);
735                         break;
736                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
737                         callgraph_buf = term->val.callgraph;
738                         break;
739                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
740                         if (term->val.branch && strcmp(term->val.branch, "no")) {
741                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
742                                 parse_branch_str(term->val.branch,
743                                                  &attr->branch_sample_type);
744                         } else
745                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
746                         break;
747                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
748                         dump_size = term->val.stack_user;
749                         break;
750                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
751                         max_stack = term->val.max_stack;
752                         break;
753                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
754                         /*
755                          * attr->inherit should has already been set by
756                          * perf_evsel__config. If user explicitly set
757                          * inherit using config terms, override global
758                          * opt->no_inherit setting.
759                          */
760                         attr->inherit = term->val.inherit ? 1 : 0;
761                         break;
762                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
763                         attr->write_backward = term->val.overwrite ? 1 : 0;
764                         break;
765                 default:
766                         break;
767                 }
768         }
769
770         /* User explicitly set per-event callgraph, clear the old setting and reset. */
771         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
772                 if (max_stack) {
773                         param.max_stack = max_stack;
774                         if (callgraph_buf == NULL)
775                                 callgraph_buf = "fp";
776                 }
777
778                 /* parse callgraph parameters */
779                 if (callgraph_buf != NULL) {
780                         if (!strcmp(callgraph_buf, "no")) {
781                                 param.enabled = false;
782                                 param.record_mode = CALLCHAIN_NONE;
783                         } else {
784                                 param.enabled = true;
785                                 if (parse_callchain_record(callgraph_buf, &param)) {
786                                         pr_err("per-event callgraph setting for %s failed. "
787                                                "Apply callgraph global setting for it\n",
788                                                evsel->name);
789                                         return;
790                                 }
791                         }
792                 }
793                 if (dump_size > 0) {
794                         dump_size = round_up(dump_size, sizeof(u64));
795                         param.dump_size = dump_size;
796                 }
797
798                 /* If global callgraph set, clear it */
799                 if (callchain_param.enabled)
800                         perf_evsel__reset_callgraph(evsel, &callchain_param);
801
802                 /* set perf-event callgraph */
803                 if (param.enabled)
804                         perf_evsel__config_callchain(evsel, opts, &param);
805         }
806 }
807
808 /*
809  * The enable_on_exec/disabled value strategy:
810  *
811  *  1) For any type of traced program:
812  *    - all independent events and group leaders are disabled
813  *    - all group members are enabled
814  *
815  *     Group members are ruled by group leaders. They need to
816  *     be enabled, because the group scheduling relies on that.
817  *
818  *  2) For traced programs executed by perf:
819  *     - all independent events and group leaders have
820  *       enable_on_exec set
821  *     - we don't specifically enable or disable any event during
822  *       the record command
823  *
824  *     Independent events and group leaders are initially disabled
825  *     and get enabled by exec. Group members are ruled by group
826  *     leaders as stated in 1).
827  *
828  *  3) For traced programs attached by perf (pid/tid):
829  *     - we specifically enable or disable all events during
830  *       the record command
831  *
832  *     When attaching events to already running traced we
833  *     enable/disable events specifically, as there's no
834  *     initial traced exec call.
835  */
836 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
837                         struct callchain_param *callchain)
838 {
839         struct perf_evsel *leader = evsel->leader;
840         struct perf_event_attr *attr = &evsel->attr;
841         int track = evsel->tracking;
842         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
843
844         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
845         attr->inherit       = !opts->no_inherit;
846         attr->write_backward = opts->overwrite ? 1 : 0;
847
848         perf_evsel__set_sample_bit(evsel, IP);
849         perf_evsel__set_sample_bit(evsel, TID);
850
851         if (evsel->sample_read) {
852                 perf_evsel__set_sample_bit(evsel, READ);
853
854                 /*
855                  * We need ID even in case of single event, because
856                  * PERF_SAMPLE_READ process ID specific data.
857                  */
858                 perf_evsel__set_sample_id(evsel, false);
859
860                 /*
861                  * Apply group format only if we belong to group
862                  * with more than one members.
863                  */
864                 if (leader->nr_members > 1) {
865                         attr->read_format |= PERF_FORMAT_GROUP;
866                         attr->inherit = 0;
867                 }
868         }
869
870         /*
871          * We default some events to have a default interval. But keep
872          * it a weak assumption overridable by the user.
873          */
874         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
875                                      opts->user_interval != ULLONG_MAX)) {
876                 if (opts->freq) {
877                         perf_evsel__set_sample_bit(evsel, PERIOD);
878                         attr->freq              = 1;
879                         attr->sample_freq       = opts->freq;
880                 } else {
881                         attr->sample_period = opts->default_interval;
882                 }
883         }
884
885         /*
886          * Disable sampling for all group members other
887          * than leader in case leader 'leads' the sampling.
888          */
889         if ((leader != evsel) && leader->sample_read) {
890                 attr->sample_freq   = 0;
891                 attr->sample_period = 0;
892         }
893
894         if (opts->no_samples)
895                 attr->sample_freq = 0;
896
897         if (opts->inherit_stat)
898                 attr->inherit_stat = 1;
899
900         if (opts->sample_address) {
901                 perf_evsel__set_sample_bit(evsel, ADDR);
902                 attr->mmap_data = track;
903         }
904
905         /*
906          * We don't allow user space callchains for  function trace
907          * event, due to issues with page faults while tracing page
908          * fault handler and its overall trickiness nature.
909          */
910         if (perf_evsel__is_function_event(evsel))
911                 evsel->attr.exclude_callchain_user = 1;
912
913         if (callchain && callchain->enabled && !evsel->no_aux_samples)
914                 perf_evsel__config_callchain(evsel, opts, callchain);
915
916         if (opts->sample_intr_regs) {
917                 attr->sample_regs_intr = opts->sample_intr_regs;
918                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
919         }
920
921         if (target__has_cpu(&opts->target) || opts->sample_cpu)
922                 perf_evsel__set_sample_bit(evsel, CPU);
923
924         if (opts->period)
925                 perf_evsel__set_sample_bit(evsel, PERIOD);
926
927         /*
928          * When the user explicitly disabled time don't force it here.
929          */
930         if (opts->sample_time &&
931             (!perf_missing_features.sample_id_all &&
932             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
933              opts->sample_time_set)))
934                 perf_evsel__set_sample_bit(evsel, TIME);
935
936         if (opts->raw_samples && !evsel->no_aux_samples) {
937                 perf_evsel__set_sample_bit(evsel, TIME);
938                 perf_evsel__set_sample_bit(evsel, RAW);
939                 perf_evsel__set_sample_bit(evsel, CPU);
940         }
941
942         if (opts->sample_address)
943                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
944
945         if (opts->no_buffering) {
946                 attr->watermark = 0;
947                 attr->wakeup_events = 1;
948         }
949         if (opts->branch_stack && !evsel->no_aux_samples) {
950                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
951                 attr->branch_sample_type = opts->branch_stack;
952         }
953
954         if (opts->sample_weight)
955                 perf_evsel__set_sample_bit(evsel, WEIGHT);
956
957         attr->task  = track;
958         attr->mmap  = track;
959         attr->mmap2 = track && !perf_missing_features.mmap2;
960         attr->comm  = track;
961
962         if (opts->record_namespaces)
963                 attr->namespaces  = track;
964
965         if (opts->record_switch_events)
966                 attr->context_switch = track;
967
968         if (opts->sample_transaction)
969                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
970
971         if (opts->running_time) {
972                 evsel->attr.read_format |=
973                         PERF_FORMAT_TOTAL_TIME_ENABLED |
974                         PERF_FORMAT_TOTAL_TIME_RUNNING;
975         }
976
977         /*
978          * XXX see the function comment above
979          *
980          * Disabling only independent events or group leaders,
981          * keeping group members enabled.
982          */
983         if (perf_evsel__is_group_leader(evsel))
984                 attr->disabled = 1;
985
986         /*
987          * Setting enable_on_exec for independent events and
988          * group leaders for traced executed by perf.
989          */
990         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
991                 !opts->initial_delay)
992                 attr->enable_on_exec = 1;
993
994         if (evsel->immediate) {
995                 attr->disabled = 0;
996                 attr->enable_on_exec = 0;
997         }
998
999         clockid = opts->clockid;
1000         if (opts->use_clockid) {
1001                 attr->use_clockid = 1;
1002                 attr->clockid = opts->clockid;
1003         }
1004
1005         if (evsel->precise_max)
1006                 perf_event_attr__set_max_precise_ip(attr);
1007
1008         if (opts->all_user) {
1009                 attr->exclude_kernel = 1;
1010                 attr->exclude_user   = 0;
1011         }
1012
1013         if (opts->all_kernel) {
1014                 attr->exclude_kernel = 0;
1015                 attr->exclude_user   = 1;
1016         }
1017
1018         /*
1019          * Apply event specific term settings,
1020          * it overloads any global configuration.
1021          */
1022         apply_config_terms(evsel, opts);
1023
1024         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1025 }
1026
1027 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1028 {
1029         if (evsel->system_wide)
1030                 nthreads = 1;
1031
1032         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1033
1034         if (evsel->fd) {
1035                 int cpu, thread;
1036                 for (cpu = 0; cpu < ncpus; cpu++) {
1037                         for (thread = 0; thread < nthreads; thread++) {
1038                                 FD(evsel, cpu, thread) = -1;
1039                         }
1040                 }
1041         }
1042
1043         return evsel->fd != NULL ? 0 : -ENOMEM;
1044 }
1045
1046 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
1047                           int ioc,  void *arg)
1048 {
1049         int cpu, thread;
1050
1051         if (evsel->system_wide)
1052                 nthreads = 1;
1053
1054         for (cpu = 0; cpu < ncpus; cpu++) {
1055                 for (thread = 0; thread < nthreads; thread++) {
1056                         int fd = FD(evsel, cpu, thread),
1057                             err = ioctl(fd, ioc, arg);
1058
1059                         if (err)
1060                                 return err;
1061                 }
1062         }
1063
1064         return 0;
1065 }
1066
1067 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
1068                              const char *filter)
1069 {
1070         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1071                                      PERF_EVENT_IOC_SET_FILTER,
1072                                      (void *)filter);
1073 }
1074
1075 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1076 {
1077         char *new_filter = strdup(filter);
1078
1079         if (new_filter != NULL) {
1080                 free(evsel->filter);
1081                 evsel->filter = new_filter;
1082                 return 0;
1083         }
1084
1085         return -1;
1086 }
1087
1088 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1089                                      const char *fmt, const char *filter)
1090 {
1091         char *new_filter;
1092
1093         if (evsel->filter == NULL)
1094                 return perf_evsel__set_filter(evsel, filter);
1095
1096         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1097                 free(evsel->filter);
1098                 evsel->filter = new_filter;
1099                 return 0;
1100         }
1101
1102         return -1;
1103 }
1104
1105 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1106 {
1107         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1108 }
1109
1110 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1111 {
1112         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1113 }
1114
1115 int perf_evsel__enable(struct perf_evsel *evsel)
1116 {
1117         int nthreads = thread_map__nr(evsel->threads);
1118         int ncpus = cpu_map__nr(evsel->cpus);
1119
1120         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1121                                      PERF_EVENT_IOC_ENABLE,
1122                                      0);
1123 }
1124
1125 int perf_evsel__disable(struct perf_evsel *evsel)
1126 {
1127         int nthreads = thread_map__nr(evsel->threads);
1128         int ncpus = cpu_map__nr(evsel->cpus);
1129
1130         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1131                                      PERF_EVENT_IOC_DISABLE,
1132                                      0);
1133 }
1134
1135 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1136 {
1137         if (ncpus == 0 || nthreads == 0)
1138                 return 0;
1139
1140         if (evsel->system_wide)
1141                 nthreads = 1;
1142
1143         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1144         if (evsel->sample_id == NULL)
1145                 return -ENOMEM;
1146
1147         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1148         if (evsel->id == NULL) {
1149                 xyarray__delete(evsel->sample_id);
1150                 evsel->sample_id = NULL;
1151                 return -ENOMEM;
1152         }
1153
1154         return 0;
1155 }
1156
1157 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1158 {
1159         xyarray__delete(evsel->fd);
1160         evsel->fd = NULL;
1161 }
1162
1163 static void perf_evsel__free_id(struct perf_evsel *evsel)
1164 {
1165         xyarray__delete(evsel->sample_id);
1166         evsel->sample_id = NULL;
1167         zfree(&evsel->id);
1168 }
1169
1170 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1171 {
1172         struct perf_evsel_config_term *term, *h;
1173
1174         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1175                 list_del(&term->list);
1176                 free(term);
1177         }
1178 }
1179
1180 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1181 {
1182         int cpu, thread;
1183
1184         if (evsel->system_wide)
1185                 nthreads = 1;
1186
1187         for (cpu = 0; cpu < ncpus; cpu++)
1188                 for (thread = 0; thread < nthreads; ++thread) {
1189                         close(FD(evsel, cpu, thread));
1190                         FD(evsel, cpu, thread) = -1;
1191                 }
1192 }
1193
1194 void perf_evsel__exit(struct perf_evsel *evsel)
1195 {
1196         assert(list_empty(&evsel->node));
1197         assert(evsel->evlist == NULL);
1198         perf_evsel__free_fd(evsel);
1199         perf_evsel__free_id(evsel);
1200         perf_evsel__free_config_terms(evsel);
1201         close_cgroup(evsel->cgrp);
1202         cpu_map__put(evsel->cpus);
1203         cpu_map__put(evsel->own_cpus);
1204         thread_map__put(evsel->threads);
1205         zfree(&evsel->group_name);
1206         zfree(&evsel->name);
1207         perf_evsel__object.fini(evsel);
1208 }
1209
1210 void perf_evsel__delete(struct perf_evsel *evsel)
1211 {
1212         perf_evsel__exit(evsel);
1213         free(evsel);
1214 }
1215
1216 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1217                                 struct perf_counts_values *count)
1218 {
1219         struct perf_counts_values tmp;
1220
1221         if (!evsel->prev_raw_counts)
1222                 return;
1223
1224         if (cpu == -1) {
1225                 tmp = evsel->prev_raw_counts->aggr;
1226                 evsel->prev_raw_counts->aggr = *count;
1227         } else {
1228                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1229                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1230         }
1231
1232         count->val = count->val - tmp.val;
1233         count->ena = count->ena - tmp.ena;
1234         count->run = count->run - tmp.run;
1235 }
1236
1237 void perf_counts_values__scale(struct perf_counts_values *count,
1238                                bool scale, s8 *pscaled)
1239 {
1240         s8 scaled = 0;
1241
1242         if (scale) {
1243                 if (count->run == 0) {
1244                         scaled = -1;
1245                         count->val = 0;
1246                 } else if (count->run < count->ena) {
1247                         scaled = 1;
1248                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1249                 }
1250         } else
1251                 count->ena = count->run = 0;
1252
1253         if (pscaled)
1254                 *pscaled = scaled;
1255 }
1256
1257 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1258                      struct perf_counts_values *count)
1259 {
1260         memset(count, 0, sizeof(*count));
1261
1262         if (FD(evsel, cpu, thread) < 0)
1263                 return -EINVAL;
1264
1265         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) <= 0)
1266                 return -errno;
1267
1268         return 0;
1269 }
1270
1271 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1272                               int cpu, int thread, bool scale)
1273 {
1274         struct perf_counts_values count;
1275         size_t nv = scale ? 3 : 1;
1276
1277         if (FD(evsel, cpu, thread) < 0)
1278                 return -EINVAL;
1279
1280         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1281                 return -ENOMEM;
1282
1283         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1284                 return -errno;
1285
1286         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1287         perf_counts_values__scale(&count, scale, NULL);
1288         *perf_counts(evsel->counts, cpu, thread) = count;
1289         return 0;
1290 }
1291
1292 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1293 {
1294         struct perf_evsel *leader = evsel->leader;
1295         int fd;
1296
1297         if (perf_evsel__is_group_leader(evsel))
1298                 return -1;
1299
1300         /*
1301          * Leader must be already processed/open,
1302          * if not it's a bug.
1303          */
1304         BUG_ON(!leader->fd);
1305
1306         fd = FD(leader, cpu, thread);
1307         BUG_ON(fd == -1);
1308
1309         return fd;
1310 }
1311
1312 struct bit_names {
1313         int bit;
1314         const char *name;
1315 };
1316
1317 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1318 {
1319         bool first_bit = true;
1320         int i = 0;
1321
1322         do {
1323                 if (value & bits[i].bit) {
1324                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1325                         first_bit = false;
1326                 }
1327         } while (bits[++i].name != NULL);
1328 }
1329
1330 static void __p_sample_type(char *buf, size_t size, u64 value)
1331 {
1332 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1333         struct bit_names bits[] = {
1334                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1335                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1336                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1337                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1338                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1339                 bit_name(WEIGHT),
1340                 { .name = NULL, }
1341         };
1342 #undef bit_name
1343         __p_bits(buf, size, value, bits);
1344 }
1345
1346 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1347 {
1348 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1349         struct bit_names bits[] = {
1350                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1351                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1352                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1353                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1354                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1355                 { .name = NULL, }
1356         };
1357 #undef bit_name
1358         __p_bits(buf, size, value, bits);
1359 }
1360
1361 static void __p_read_format(char *buf, size_t size, u64 value)
1362 {
1363 #define bit_name(n) { PERF_FORMAT_##n, #n }
1364         struct bit_names bits[] = {
1365                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1366                 bit_name(ID), bit_name(GROUP),
1367                 { .name = NULL, }
1368         };
1369 #undef bit_name
1370         __p_bits(buf, size, value, bits);
1371 }
1372
1373 #define BUF_SIZE                1024
1374
1375 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1376 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1377 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1378 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1379 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1380 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1381
1382 #define PRINT_ATTRn(_n, _f, _p)                         \
1383 do {                                                    \
1384         if (attr->_f) {                                 \
1385                 _p(attr->_f);                           \
1386                 ret += attr__fprintf(fp, _n, buf, priv);\
1387         }                                               \
1388 } while (0)
1389
1390 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1391
1392 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1393                              attr__fprintf_f attr__fprintf, void *priv)
1394 {
1395         char buf[BUF_SIZE];
1396         int ret = 0;
1397
1398         PRINT_ATTRf(type, p_unsigned);
1399         PRINT_ATTRf(size, p_unsigned);
1400         PRINT_ATTRf(config, p_hex);
1401         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1402         PRINT_ATTRf(sample_type, p_sample_type);
1403         PRINT_ATTRf(read_format, p_read_format);
1404
1405         PRINT_ATTRf(disabled, p_unsigned);
1406         PRINT_ATTRf(inherit, p_unsigned);
1407         PRINT_ATTRf(pinned, p_unsigned);
1408         PRINT_ATTRf(exclusive, p_unsigned);
1409         PRINT_ATTRf(exclude_user, p_unsigned);
1410         PRINT_ATTRf(exclude_kernel, p_unsigned);
1411         PRINT_ATTRf(exclude_hv, p_unsigned);
1412         PRINT_ATTRf(exclude_idle, p_unsigned);
1413         PRINT_ATTRf(mmap, p_unsigned);
1414         PRINT_ATTRf(comm, p_unsigned);
1415         PRINT_ATTRf(freq, p_unsigned);
1416         PRINT_ATTRf(inherit_stat, p_unsigned);
1417         PRINT_ATTRf(enable_on_exec, p_unsigned);
1418         PRINT_ATTRf(task, p_unsigned);
1419         PRINT_ATTRf(watermark, p_unsigned);
1420         PRINT_ATTRf(precise_ip, p_unsigned);
1421         PRINT_ATTRf(mmap_data, p_unsigned);
1422         PRINT_ATTRf(sample_id_all, p_unsigned);
1423         PRINT_ATTRf(exclude_host, p_unsigned);
1424         PRINT_ATTRf(exclude_guest, p_unsigned);
1425         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1426         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1427         PRINT_ATTRf(mmap2, p_unsigned);
1428         PRINT_ATTRf(comm_exec, p_unsigned);
1429         PRINT_ATTRf(use_clockid, p_unsigned);
1430         PRINT_ATTRf(context_switch, p_unsigned);
1431         PRINT_ATTRf(write_backward, p_unsigned);
1432
1433         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1434         PRINT_ATTRf(bp_type, p_unsigned);
1435         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1436         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1437         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1438         PRINT_ATTRf(sample_regs_user, p_hex);
1439         PRINT_ATTRf(sample_stack_user, p_unsigned);
1440         PRINT_ATTRf(clockid, p_signed);
1441         PRINT_ATTRf(sample_regs_intr, p_hex);
1442         PRINT_ATTRf(aux_watermark, p_unsigned);
1443         PRINT_ATTRf(sample_max_stack, p_unsigned);
1444
1445         return ret;
1446 }
1447
1448 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1449                                 void *priv __maybe_unused)
1450 {
1451         return fprintf(fp, "  %-32s %s\n", name, val);
1452 }
1453
1454 static bool ignore_missing_thread(struct perf_evsel *evsel,
1455                                   struct thread_map *threads,
1456                                   int thread, int err)
1457 {
1458         if (!evsel->ignore_missing_thread)
1459                 return false;
1460
1461         /* The system wide setup does not work with threads. */
1462         if (evsel->system_wide)
1463                 return false;
1464
1465         /* The -ESRCH is perf event syscall errno for pid's not found. */
1466         if (err != -ESRCH)
1467                 return false;
1468
1469         /* If there's only one thread, let it fail. */
1470         if (threads->nr == 1)
1471                 return false;
1472
1473         if (thread_map__remove(threads, thread))
1474                 return false;
1475
1476         pr_warning("WARNING: Ignored open failure for pid %d\n",
1477                    thread_map__pid(threads, thread));
1478         return true;
1479 }
1480
1481 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1482                      struct thread_map *threads)
1483 {
1484         int cpu, thread, nthreads;
1485         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1486         int pid = -1, err;
1487         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1488
1489         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1490                 return -EINVAL;
1491
1492         if (cpus == NULL) {
1493                 static struct cpu_map *empty_cpu_map;
1494
1495                 if (empty_cpu_map == NULL) {
1496                         empty_cpu_map = cpu_map__dummy_new();
1497                         if (empty_cpu_map == NULL)
1498                                 return -ENOMEM;
1499                 }
1500
1501                 cpus = empty_cpu_map;
1502         }
1503
1504         if (threads == NULL) {
1505                 static struct thread_map *empty_thread_map;
1506
1507                 if (empty_thread_map == NULL) {
1508                         empty_thread_map = thread_map__new_by_tid(-1);
1509                         if (empty_thread_map == NULL)
1510                                 return -ENOMEM;
1511                 }
1512
1513                 threads = empty_thread_map;
1514         }
1515
1516         if (evsel->system_wide)
1517                 nthreads = 1;
1518         else
1519                 nthreads = threads->nr;
1520
1521         if (evsel->fd == NULL &&
1522             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1523                 return -ENOMEM;
1524
1525         if (evsel->cgrp) {
1526                 flags |= PERF_FLAG_PID_CGROUP;
1527                 pid = evsel->cgrp->fd;
1528         }
1529
1530 fallback_missing_features:
1531         if (perf_missing_features.clockid_wrong)
1532                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1533         if (perf_missing_features.clockid) {
1534                 evsel->attr.use_clockid = 0;
1535                 evsel->attr.clockid = 0;
1536         }
1537         if (perf_missing_features.cloexec)
1538                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1539         if (perf_missing_features.mmap2)
1540                 evsel->attr.mmap2 = 0;
1541         if (perf_missing_features.exclude_guest)
1542                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1543         if (perf_missing_features.lbr_flags)
1544                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1545                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1546 retry_sample_id:
1547         if (perf_missing_features.sample_id_all)
1548                 evsel->attr.sample_id_all = 0;
1549
1550         if (verbose >= 2) {
1551                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1552                 fprintf(stderr, "perf_event_attr:\n");
1553                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1554                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1555         }
1556
1557         for (cpu = 0; cpu < cpus->nr; cpu++) {
1558
1559                 for (thread = 0; thread < nthreads; thread++) {
1560                         int fd, group_fd;
1561
1562                         if (!evsel->cgrp && !evsel->system_wide)
1563                                 pid = thread_map__pid(threads, thread);
1564
1565                         group_fd = get_group_fd(evsel, cpu, thread);
1566 retry_open:
1567                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1568                                   pid, cpus->map[cpu], group_fd, flags);
1569
1570                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1571                                                  group_fd, flags);
1572
1573                         FD(evsel, cpu, thread) = fd;
1574
1575                         if (fd < 0) {
1576                                 err = -errno;
1577
1578                                 if (ignore_missing_thread(evsel, threads, thread, err)) {
1579                                         /*
1580                                          * We just removed 1 thread, so take a step
1581                                          * back on thread index and lower the upper
1582                                          * nthreads limit.
1583                                          */
1584                                         nthreads--;
1585                                         thread--;
1586
1587                                         /* ... and pretend like nothing have happened. */
1588                                         err = 0;
1589                                         continue;
1590                                 }
1591
1592                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1593                                           err);
1594                                 goto try_fallback;
1595                         }
1596
1597                         pr_debug2(" = %d\n", fd);
1598
1599                         if (evsel->bpf_fd >= 0) {
1600                                 int evt_fd = fd;
1601                                 int bpf_fd = evsel->bpf_fd;
1602
1603                                 err = ioctl(evt_fd,
1604                                             PERF_EVENT_IOC_SET_BPF,
1605                                             bpf_fd);
1606                                 if (err && errno != EEXIST) {
1607                                         pr_err("failed to attach bpf fd %d: %s\n",
1608                                                bpf_fd, strerror(errno));
1609                                         err = -EINVAL;
1610                                         goto out_close;
1611                                 }
1612                         }
1613
1614                         set_rlimit = NO_CHANGE;
1615
1616                         /*
1617                          * If we succeeded but had to kill clockid, fail and
1618                          * have perf_evsel__open_strerror() print us a nice
1619                          * error.
1620                          */
1621                         if (perf_missing_features.clockid ||
1622                             perf_missing_features.clockid_wrong) {
1623                                 err = -EINVAL;
1624                                 goto out_close;
1625                         }
1626                 }
1627         }
1628
1629         return 0;
1630
1631 try_fallback:
1632         /*
1633          * perf stat needs between 5 and 22 fds per CPU. When we run out
1634          * of them try to increase the limits.
1635          */
1636         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1637                 struct rlimit l;
1638                 int old_errno = errno;
1639
1640                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1641                         if (set_rlimit == NO_CHANGE)
1642                                 l.rlim_cur = l.rlim_max;
1643                         else {
1644                                 l.rlim_cur = l.rlim_max + 1000;
1645                                 l.rlim_max = l.rlim_cur;
1646                         }
1647                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1648                                 set_rlimit++;
1649                                 errno = old_errno;
1650                                 goto retry_open;
1651                         }
1652                 }
1653                 errno = old_errno;
1654         }
1655
1656         if (err != -EINVAL || cpu > 0 || thread > 0)
1657                 goto out_close;
1658
1659         /*
1660          * Must probe features in the order they were added to the
1661          * perf_event_attr interface.
1662          */
1663         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1664                 perf_missing_features.write_backward = true;
1665                 goto out_close;
1666         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1667                 perf_missing_features.clockid_wrong = true;
1668                 goto fallback_missing_features;
1669         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1670                 perf_missing_features.clockid = true;
1671                 goto fallback_missing_features;
1672         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1673                 perf_missing_features.cloexec = true;
1674                 goto fallback_missing_features;
1675         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1676                 perf_missing_features.mmap2 = true;
1677                 goto fallback_missing_features;
1678         } else if (!perf_missing_features.exclude_guest &&
1679                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1680                 perf_missing_features.exclude_guest = true;
1681                 goto fallback_missing_features;
1682         } else if (!perf_missing_features.sample_id_all) {
1683                 perf_missing_features.sample_id_all = true;
1684                 goto retry_sample_id;
1685         } else if (!perf_missing_features.lbr_flags &&
1686                         (evsel->attr.branch_sample_type &
1687                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1688                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1689                 perf_missing_features.lbr_flags = true;
1690                 goto fallback_missing_features;
1691         }
1692 out_close:
1693         do {
1694                 while (--thread >= 0) {
1695                         close(FD(evsel, cpu, thread));
1696                         FD(evsel, cpu, thread) = -1;
1697                 }
1698                 thread = nthreads;
1699         } while (--cpu >= 0);
1700         return err;
1701 }
1702
1703 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1704 {
1705         if (evsel->fd == NULL)
1706                 return;
1707
1708         perf_evsel__close_fd(evsel, ncpus, nthreads);
1709         perf_evsel__free_fd(evsel);
1710 }
1711
1712 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1713                              struct cpu_map *cpus)
1714 {
1715         return perf_evsel__open(evsel, cpus, NULL);
1716 }
1717
1718 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1719                                 struct thread_map *threads)
1720 {
1721         return perf_evsel__open(evsel, NULL, threads);
1722 }
1723
1724 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1725                                        const union perf_event *event,
1726                                        struct perf_sample *sample)
1727 {
1728         u64 type = evsel->attr.sample_type;
1729         const u64 *array = event->sample.array;
1730         bool swapped = evsel->needs_swap;
1731         union u64_swap u;
1732
1733         array += ((event->header.size -
1734                    sizeof(event->header)) / sizeof(u64)) - 1;
1735
1736         if (type & PERF_SAMPLE_IDENTIFIER) {
1737                 sample->id = *array;
1738                 array--;
1739         }
1740
1741         if (type & PERF_SAMPLE_CPU) {
1742                 u.val64 = *array;
1743                 if (swapped) {
1744                         /* undo swap of u64, then swap on individual u32s */
1745                         u.val64 = bswap_64(u.val64);
1746                         u.val32[0] = bswap_32(u.val32[0]);
1747                 }
1748
1749                 sample->cpu = u.val32[0];
1750                 array--;
1751         }
1752
1753         if (type & PERF_SAMPLE_STREAM_ID) {
1754                 sample->stream_id = *array;
1755                 array--;
1756         }
1757
1758         if (type & PERF_SAMPLE_ID) {
1759                 sample->id = *array;
1760                 array--;
1761         }
1762
1763         if (type & PERF_SAMPLE_TIME) {
1764                 sample->time = *array;
1765                 array--;
1766         }
1767
1768         if (type & PERF_SAMPLE_TID) {
1769                 u.val64 = *array;
1770                 if (swapped) {
1771                         /* undo swap of u64, then swap on individual u32s */
1772                         u.val64 = bswap_64(u.val64);
1773                         u.val32[0] = bswap_32(u.val32[0]);
1774                         u.val32[1] = bswap_32(u.val32[1]);
1775                 }
1776
1777                 sample->pid = u.val32[0];
1778                 sample->tid = u.val32[1];
1779                 array--;
1780         }
1781
1782         return 0;
1783 }
1784
1785 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1786                             u64 size)
1787 {
1788         return size > max_size || offset + size > endp;
1789 }
1790
1791 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1792         do {                                                            \
1793                 if (overflow(endp, (max_size), (offset), (size)))       \
1794                         return -EFAULT;                                 \
1795         } while (0)
1796
1797 #define OVERFLOW_CHECK_u64(offset) \
1798         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1799
1800 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1801                              struct perf_sample *data)
1802 {
1803         u64 type = evsel->attr.sample_type;
1804         bool swapped = evsel->needs_swap;
1805         const u64 *array;
1806         u16 max_size = event->header.size;
1807         const void *endp = (void *)event + max_size;
1808         u64 sz;
1809
1810         /*
1811          * used for cross-endian analysis. See git commit 65014ab3
1812          * for why this goofiness is needed.
1813          */
1814         union u64_swap u;
1815
1816         memset(data, 0, sizeof(*data));
1817         data->cpu = data->pid = data->tid = -1;
1818         data->stream_id = data->id = data->time = -1ULL;
1819         data->period = evsel->attr.sample_period;
1820         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1821
1822         if (event->header.type != PERF_RECORD_SAMPLE) {
1823                 if (!evsel->attr.sample_id_all)
1824                         return 0;
1825                 return perf_evsel__parse_id_sample(evsel, event, data);
1826         }
1827
1828         array = event->sample.array;
1829
1830         /*
1831          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1832          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1833          * check the format does not go past the end of the event.
1834          */
1835         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1836                 return -EFAULT;
1837
1838         data->id = -1ULL;
1839         if (type & PERF_SAMPLE_IDENTIFIER) {
1840                 data->id = *array;
1841                 array++;
1842         }
1843
1844         if (type & PERF_SAMPLE_IP) {
1845                 data->ip = *array;
1846                 array++;
1847         }
1848
1849         if (type & PERF_SAMPLE_TID) {
1850                 u.val64 = *array;
1851                 if (swapped) {
1852                         /* undo swap of u64, then swap on individual u32s */
1853                         u.val64 = bswap_64(u.val64);
1854                         u.val32[0] = bswap_32(u.val32[0]);
1855                         u.val32[1] = bswap_32(u.val32[1]);
1856                 }
1857
1858                 data->pid = u.val32[0];
1859                 data->tid = u.val32[1];
1860                 array++;
1861         }
1862
1863         if (type & PERF_SAMPLE_TIME) {
1864                 data->time = *array;
1865                 array++;
1866         }
1867
1868         data->addr = 0;
1869         if (type & PERF_SAMPLE_ADDR) {
1870                 data->addr = *array;
1871                 array++;
1872         }
1873
1874         if (type & PERF_SAMPLE_ID) {
1875                 data->id = *array;
1876                 array++;
1877         }
1878
1879         if (type & PERF_SAMPLE_STREAM_ID) {
1880                 data->stream_id = *array;
1881                 array++;
1882         }
1883
1884         if (type & PERF_SAMPLE_CPU) {
1885
1886                 u.val64 = *array;
1887                 if (swapped) {
1888                         /* undo swap of u64, then swap on individual u32s */
1889                         u.val64 = bswap_64(u.val64);
1890                         u.val32[0] = bswap_32(u.val32[0]);
1891                 }
1892
1893                 data->cpu = u.val32[0];
1894                 array++;
1895         }
1896
1897         if (type & PERF_SAMPLE_PERIOD) {
1898                 data->period = *array;
1899                 array++;
1900         }
1901
1902         if (type & PERF_SAMPLE_READ) {
1903                 u64 read_format = evsel->attr.read_format;
1904
1905                 OVERFLOW_CHECK_u64(array);
1906                 if (read_format & PERF_FORMAT_GROUP)
1907                         data->read.group.nr = *array;
1908                 else
1909                         data->read.one.value = *array;
1910
1911                 array++;
1912
1913                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1914                         OVERFLOW_CHECK_u64(array);
1915                         data->read.time_enabled = *array;
1916                         array++;
1917                 }
1918
1919                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1920                         OVERFLOW_CHECK_u64(array);
1921                         data->read.time_running = *array;
1922                         array++;
1923                 }
1924
1925                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1926                 if (read_format & PERF_FORMAT_GROUP) {
1927                         const u64 max_group_nr = UINT64_MAX /
1928                                         sizeof(struct sample_read_value);
1929
1930                         if (data->read.group.nr > max_group_nr)
1931                                 return -EFAULT;
1932                         sz = data->read.group.nr *
1933                              sizeof(struct sample_read_value);
1934                         OVERFLOW_CHECK(array, sz, max_size);
1935                         data->read.group.values =
1936                                         (struct sample_read_value *)array;
1937                         array = (void *)array + sz;
1938                 } else {
1939                         OVERFLOW_CHECK_u64(array);
1940                         data->read.one.id = *array;
1941                         array++;
1942                 }
1943         }
1944
1945         if (type & PERF_SAMPLE_CALLCHAIN) {
1946                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1947
1948                 OVERFLOW_CHECK_u64(array);
1949                 data->callchain = (struct ip_callchain *)array++;
1950                 if (data->callchain->nr > max_callchain_nr)
1951                         return -EFAULT;
1952                 sz = data->callchain->nr * sizeof(u64);
1953                 OVERFLOW_CHECK(array, sz, max_size);
1954                 array = (void *)array + sz;
1955         }
1956
1957         if (type & PERF_SAMPLE_RAW) {
1958                 OVERFLOW_CHECK_u64(array);
1959                 u.val64 = *array;
1960                 if (WARN_ONCE(swapped,
1961                               "Endianness of raw data not corrected!\n")) {
1962                         /* undo swap of u64, then swap on individual u32s */
1963                         u.val64 = bswap_64(u.val64);
1964                         u.val32[0] = bswap_32(u.val32[0]);
1965                         u.val32[1] = bswap_32(u.val32[1]);
1966                 }
1967                 data->raw_size = u.val32[0];
1968                 array = (void *)array + sizeof(u32);
1969
1970                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1971                 data->raw_data = (void *)array;
1972                 array = (void *)array + data->raw_size;
1973         }
1974
1975         if (type & PERF_SAMPLE_BRANCH_STACK) {
1976                 const u64 max_branch_nr = UINT64_MAX /
1977                                           sizeof(struct branch_entry);
1978
1979                 OVERFLOW_CHECK_u64(array);
1980                 data->branch_stack = (struct branch_stack *)array++;
1981
1982                 if (data->branch_stack->nr > max_branch_nr)
1983                         return -EFAULT;
1984                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1985                 OVERFLOW_CHECK(array, sz, max_size);
1986                 array = (void *)array + sz;
1987         }
1988
1989         if (type & PERF_SAMPLE_REGS_USER) {
1990                 OVERFLOW_CHECK_u64(array);
1991                 data->user_regs.abi = *array;
1992                 array++;
1993
1994                 if (data->user_regs.abi) {
1995                         u64 mask = evsel->attr.sample_regs_user;
1996
1997                         sz = hweight_long(mask) * sizeof(u64);
1998                         OVERFLOW_CHECK(array, sz, max_size);
1999                         data->user_regs.mask = mask;
2000                         data->user_regs.regs = (u64 *)array;
2001                         array = (void *)array + sz;
2002                 }
2003         }
2004
2005         if (type & PERF_SAMPLE_STACK_USER) {
2006                 OVERFLOW_CHECK_u64(array);
2007                 sz = *array++;
2008
2009                 data->user_stack.offset = ((char *)(array - 1)
2010                                           - (char *) event);
2011
2012                 if (!sz) {
2013                         data->user_stack.size = 0;
2014                 } else {
2015                         OVERFLOW_CHECK(array, sz, max_size);
2016                         data->user_stack.data = (char *)array;
2017                         array = (void *)array + sz;
2018                         OVERFLOW_CHECK_u64(array);
2019                         data->user_stack.size = *array++;
2020                         if (WARN_ONCE(data->user_stack.size > sz,
2021                                       "user stack dump failure\n"))
2022                                 return -EFAULT;
2023                 }
2024         }
2025
2026         if (type & PERF_SAMPLE_WEIGHT) {
2027                 OVERFLOW_CHECK_u64(array);
2028                 data->weight = *array;
2029                 array++;
2030         }
2031
2032         data->data_src = PERF_MEM_DATA_SRC_NONE;
2033         if (type & PERF_SAMPLE_DATA_SRC) {
2034                 OVERFLOW_CHECK_u64(array);
2035                 data->data_src = *array;
2036                 array++;
2037         }
2038
2039         data->transaction = 0;
2040         if (type & PERF_SAMPLE_TRANSACTION) {
2041                 OVERFLOW_CHECK_u64(array);
2042                 data->transaction = *array;
2043                 array++;
2044         }
2045
2046         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2047         if (type & PERF_SAMPLE_REGS_INTR) {
2048                 OVERFLOW_CHECK_u64(array);
2049                 data->intr_regs.abi = *array;
2050                 array++;
2051
2052                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2053                         u64 mask = evsel->attr.sample_regs_intr;
2054
2055                         sz = hweight_long(mask) * sizeof(u64);
2056                         OVERFLOW_CHECK(array, sz, max_size);
2057                         data->intr_regs.mask = mask;
2058                         data->intr_regs.regs = (u64 *)array;
2059                         array = (void *)array + sz;
2060                 }
2061         }
2062
2063         return 0;
2064 }
2065
2066 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2067                                      u64 read_format)
2068 {
2069         size_t sz, result = sizeof(struct sample_event);
2070
2071         if (type & PERF_SAMPLE_IDENTIFIER)
2072                 result += sizeof(u64);
2073
2074         if (type & PERF_SAMPLE_IP)
2075                 result += sizeof(u64);
2076
2077         if (type & PERF_SAMPLE_TID)
2078                 result += sizeof(u64);
2079
2080         if (type & PERF_SAMPLE_TIME)
2081                 result += sizeof(u64);
2082
2083         if (type & PERF_SAMPLE_ADDR)
2084                 result += sizeof(u64);
2085
2086         if (type & PERF_SAMPLE_ID)
2087                 result += sizeof(u64);
2088
2089         if (type & PERF_SAMPLE_STREAM_ID)
2090                 result += sizeof(u64);
2091
2092         if (type & PERF_SAMPLE_CPU)
2093                 result += sizeof(u64);
2094
2095         if (type & PERF_SAMPLE_PERIOD)
2096                 result += sizeof(u64);
2097
2098         if (type & PERF_SAMPLE_READ) {
2099                 result += sizeof(u64);
2100                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2101                         result += sizeof(u64);
2102                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2103                         result += sizeof(u64);
2104                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2105                 if (read_format & PERF_FORMAT_GROUP) {
2106                         sz = sample->read.group.nr *
2107                              sizeof(struct sample_read_value);
2108                         result += sz;
2109                 } else {
2110                         result += sizeof(u64);
2111                 }
2112         }
2113
2114         if (type & PERF_SAMPLE_CALLCHAIN) {
2115                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2116                 result += sz;
2117         }
2118
2119         if (type & PERF_SAMPLE_RAW) {
2120                 result += sizeof(u32);
2121                 result += sample->raw_size;
2122         }
2123
2124         if (type & PERF_SAMPLE_BRANCH_STACK) {
2125                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2126                 sz += sizeof(u64);
2127                 result += sz;
2128         }
2129
2130         if (type & PERF_SAMPLE_REGS_USER) {
2131                 if (sample->user_regs.abi) {
2132                         result += sizeof(u64);
2133                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2134                         result += sz;
2135                 } else {
2136                         result += sizeof(u64);
2137                 }
2138         }
2139
2140         if (type & PERF_SAMPLE_STACK_USER) {
2141                 sz = sample->user_stack.size;
2142                 result += sizeof(u64);
2143                 if (sz) {
2144                         result += sz;
2145                         result += sizeof(u64);
2146                 }
2147         }
2148
2149         if (type & PERF_SAMPLE_WEIGHT)
2150                 result += sizeof(u64);
2151
2152         if (type & PERF_SAMPLE_DATA_SRC)
2153                 result += sizeof(u64);
2154
2155         if (type & PERF_SAMPLE_TRANSACTION)
2156                 result += sizeof(u64);
2157
2158         if (type & PERF_SAMPLE_REGS_INTR) {
2159                 if (sample->intr_regs.abi) {
2160                         result += sizeof(u64);
2161                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2162                         result += sz;
2163                 } else {
2164                         result += sizeof(u64);
2165                 }
2166         }
2167
2168         return result;
2169 }
2170
2171 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2172                                   u64 read_format,
2173                                   const struct perf_sample *sample,
2174                                   bool swapped)
2175 {
2176         u64 *array;
2177         size_t sz;
2178         /*
2179          * used for cross-endian analysis. See git commit 65014ab3
2180          * for why this goofiness is needed.
2181          */
2182         union u64_swap u;
2183
2184         array = event->sample.array;
2185
2186         if (type & PERF_SAMPLE_IDENTIFIER) {
2187                 *array = sample->id;
2188                 array++;
2189         }
2190
2191         if (type & PERF_SAMPLE_IP) {
2192                 *array = sample->ip;
2193                 array++;
2194         }
2195
2196         if (type & PERF_SAMPLE_TID) {
2197                 u.val32[0] = sample->pid;
2198                 u.val32[1] = sample->tid;
2199                 if (swapped) {
2200                         /*
2201                          * Inverse of what is done in perf_evsel__parse_sample
2202                          */
2203                         u.val32[0] = bswap_32(u.val32[0]);
2204                         u.val32[1] = bswap_32(u.val32[1]);
2205                         u.val64 = bswap_64(u.val64);
2206                 }
2207
2208                 *array = u.val64;
2209                 array++;
2210         }
2211
2212         if (type & PERF_SAMPLE_TIME) {
2213                 *array = sample->time;
2214                 array++;
2215         }
2216
2217         if (type & PERF_SAMPLE_ADDR) {
2218                 *array = sample->addr;
2219                 array++;
2220         }
2221
2222         if (type & PERF_SAMPLE_ID) {
2223                 *array = sample->id;
2224                 array++;
2225         }
2226
2227         if (type & PERF_SAMPLE_STREAM_ID) {
2228                 *array = sample->stream_id;
2229                 array++;
2230         }
2231
2232         if (type & PERF_SAMPLE_CPU) {
2233                 u.val32[0] = sample->cpu;
2234                 if (swapped) {
2235                         /*
2236                          * Inverse of what is done in perf_evsel__parse_sample
2237                          */
2238                         u.val32[0] = bswap_32(u.val32[0]);
2239                         u.val64 = bswap_64(u.val64);
2240                 }
2241                 *array = u.val64;
2242                 array++;
2243         }
2244
2245         if (type & PERF_SAMPLE_PERIOD) {
2246                 *array = sample->period;
2247                 array++;
2248         }
2249
2250         if (type & PERF_SAMPLE_READ) {
2251                 if (read_format & PERF_FORMAT_GROUP)
2252                         *array = sample->read.group.nr;
2253                 else
2254                         *array = sample->read.one.value;
2255                 array++;
2256
2257                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2258                         *array = sample->read.time_enabled;
2259                         array++;
2260                 }
2261
2262                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2263                         *array = sample->read.time_running;
2264                         array++;
2265                 }
2266
2267                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2268                 if (read_format & PERF_FORMAT_GROUP) {
2269                         sz = sample->read.group.nr *
2270                              sizeof(struct sample_read_value);
2271                         memcpy(array, sample->read.group.values, sz);
2272                         array = (void *)array + sz;
2273                 } else {
2274                         *array = sample->read.one.id;
2275                         array++;
2276                 }
2277         }
2278
2279         if (type & PERF_SAMPLE_CALLCHAIN) {
2280                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2281                 memcpy(array, sample->callchain, sz);
2282                 array = (void *)array + sz;
2283         }
2284
2285         if (type & PERF_SAMPLE_RAW) {
2286                 u.val32[0] = sample->raw_size;
2287                 if (WARN_ONCE(swapped,
2288                               "Endianness of raw data not corrected!\n")) {
2289                         /*
2290                          * Inverse of what is done in perf_evsel__parse_sample
2291                          */
2292                         u.val32[0] = bswap_32(u.val32[0]);
2293                         u.val32[1] = bswap_32(u.val32[1]);
2294                         u.val64 = bswap_64(u.val64);
2295                 }
2296                 *array = u.val64;
2297                 array = (void *)array + sizeof(u32);
2298
2299                 memcpy(array, sample->raw_data, sample->raw_size);
2300                 array = (void *)array + sample->raw_size;
2301         }
2302
2303         if (type & PERF_SAMPLE_BRANCH_STACK) {
2304                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2305                 sz += sizeof(u64);
2306                 memcpy(array, sample->branch_stack, sz);
2307                 array = (void *)array + sz;
2308         }
2309
2310         if (type & PERF_SAMPLE_REGS_USER) {
2311                 if (sample->user_regs.abi) {
2312                         *array++ = sample->user_regs.abi;
2313                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2314                         memcpy(array, sample->user_regs.regs, sz);
2315                         array = (void *)array + sz;
2316                 } else {
2317                         *array++ = 0;
2318                 }
2319         }
2320
2321         if (type & PERF_SAMPLE_STACK_USER) {
2322                 sz = sample->user_stack.size;
2323                 *array++ = sz;
2324                 if (sz) {
2325                         memcpy(array, sample->user_stack.data, sz);
2326                         array = (void *)array + sz;
2327                         *array++ = sz;
2328                 }
2329         }
2330
2331         if (type & PERF_SAMPLE_WEIGHT) {
2332                 *array = sample->weight;
2333                 array++;
2334         }
2335
2336         if (type & PERF_SAMPLE_DATA_SRC) {
2337                 *array = sample->data_src;
2338                 array++;
2339         }
2340
2341         if (type & PERF_SAMPLE_TRANSACTION) {
2342                 *array = sample->transaction;
2343                 array++;
2344         }
2345
2346         if (type & PERF_SAMPLE_REGS_INTR) {
2347                 if (sample->intr_regs.abi) {
2348                         *array++ = sample->intr_regs.abi;
2349                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2350                         memcpy(array, sample->intr_regs.regs, sz);
2351                         array = (void *)array + sz;
2352                 } else {
2353                         *array++ = 0;
2354                 }
2355         }
2356
2357         return 0;
2358 }
2359
2360 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2361 {
2362         return pevent_find_field(evsel->tp_format, name);
2363 }
2364
2365 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2366                          const char *name)
2367 {
2368         struct format_field *field = perf_evsel__field(evsel, name);
2369         int offset;
2370
2371         if (!field)
2372                 return NULL;
2373
2374         offset = field->offset;
2375
2376         if (field->flags & FIELD_IS_DYNAMIC) {
2377                 offset = *(int *)(sample->raw_data + field->offset);
2378                 offset &= 0xffff;
2379         }
2380
2381         return sample->raw_data + offset;
2382 }
2383
2384 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2385                          bool needs_swap)
2386 {
2387         u64 value;
2388         void *ptr = sample->raw_data + field->offset;
2389
2390         switch (field->size) {
2391         case 1:
2392                 return *(u8 *)ptr;
2393         case 2:
2394                 value = *(u16 *)ptr;
2395                 break;
2396         case 4:
2397                 value = *(u32 *)ptr;
2398                 break;
2399         case 8:
2400                 memcpy(&value, ptr, sizeof(u64));
2401                 break;
2402         default:
2403                 return 0;
2404         }
2405
2406         if (!needs_swap)
2407                 return value;
2408
2409         switch (field->size) {
2410         case 2:
2411                 return bswap_16(value);
2412         case 4:
2413                 return bswap_32(value);
2414         case 8:
2415                 return bswap_64(value);
2416         default:
2417                 return 0;
2418         }
2419
2420         return 0;
2421 }
2422
2423 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2424                        const char *name)
2425 {
2426         struct format_field *field = perf_evsel__field(evsel, name);
2427
2428         if (!field)
2429                 return 0;
2430
2431         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2432 }
2433
2434 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2435                           char *msg, size_t msgsize)
2436 {
2437         int paranoid;
2438
2439         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2440             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2441             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2442                 /*
2443                  * If it's cycles then fall back to hrtimer based
2444                  * cpu-clock-tick sw counter, which is always available even if
2445                  * no PMU support.
2446                  *
2447                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2448                  * b0a873e).
2449                  */
2450                 scnprintf(msg, msgsize, "%s",
2451 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2452
2453                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2454                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2455
2456                 zfree(&evsel->name);
2457                 return true;
2458         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2459                    (paranoid = perf_event_paranoid()) > 1) {
2460                 const char *name = perf_evsel__name(evsel);
2461                 char *new_name;
2462
2463                 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2464                         return false;
2465
2466                 if (evsel->name)
2467                         free(evsel->name);
2468                 evsel->name = new_name;
2469                 scnprintf(msg, msgsize,
2470 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2471                 evsel->attr.exclude_kernel = 1;
2472
2473                 return true;
2474         }
2475
2476         return false;
2477 }
2478
2479 static bool find_process(const char *name)
2480 {
2481         size_t len = strlen(name);
2482         DIR *dir;
2483         struct dirent *d;
2484         int ret = -1;
2485
2486         dir = opendir(procfs__mountpoint());
2487         if (!dir)
2488                 return false;
2489
2490         /* Walk through the directory. */
2491         while (ret && (d = readdir(dir)) != NULL) {
2492                 char path[PATH_MAX];
2493                 char *data;
2494                 size_t size;
2495
2496                 if ((d->d_type != DT_DIR) ||
2497                      !strcmp(".", d->d_name) ||
2498                      !strcmp("..", d->d_name))
2499                         continue;
2500
2501                 scnprintf(path, sizeof(path), "%s/%s/comm",
2502                           procfs__mountpoint(), d->d_name);
2503
2504                 if (filename__read_str(path, &data, &size))
2505                         continue;
2506
2507                 ret = strncmp(name, data, len);
2508                 free(data);
2509         }
2510
2511         closedir(dir);
2512         return ret ? false : true;
2513 }
2514
2515 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2516                               int err, char *msg, size_t size)
2517 {
2518         char sbuf[STRERR_BUFSIZE];
2519         int printed = 0;
2520
2521         switch (err) {
2522         case EPERM:
2523         case EACCES:
2524                 if (err == EPERM)
2525                         printed = scnprintf(msg, size,
2526                                 "No permission to enable %s event.\n\n",
2527                                 perf_evsel__name(evsel));
2528
2529                 return scnprintf(msg + printed, size - printed,
2530                  "You may not have permission to collect %sstats.\n\n"
2531                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2532                  "which controls use of the performance events system by\n"
2533                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2534                  "The current value is %d:\n\n"
2535                  "  -1: Allow use of (almost) all events by all users\n"
2536                  ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2537                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2538                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2539                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2540                  "      kernel.perf_event_paranoid = -1\n" ,
2541                                  target->system_wide ? "system-wide " : "",
2542                                  perf_event_paranoid());
2543         case ENOENT:
2544                 return scnprintf(msg, size, "The %s event is not supported.",
2545                                  perf_evsel__name(evsel));
2546         case EMFILE:
2547                 return scnprintf(msg, size, "%s",
2548                          "Too many events are opened.\n"
2549                          "Probably the maximum number of open file descriptors has been reached.\n"
2550                          "Hint: Try again after reducing the number of events.\n"
2551                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2552         case ENOMEM:
2553                 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2554                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2555                         return scnprintf(msg, size,
2556                                          "Not enough memory to setup event with callchain.\n"
2557                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2558                                          "Hint: Current value: %d", sysctl_perf_event_max_stack);
2559                 break;
2560         case ENODEV:
2561                 if (target->cpu_list)
2562                         return scnprintf(msg, size, "%s",
2563          "No such device - did you specify an out-of-range profile CPU?");
2564                 break;
2565         case EOPNOTSUPP:
2566                 if (evsel->attr.sample_period != 0)
2567                         return scnprintf(msg, size, "%s",
2568         "PMU Hardware doesn't support sampling/overflow-interrupts.");
2569                 if (evsel->attr.precise_ip)
2570                         return scnprintf(msg, size, "%s",
2571         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2572 #if defined(__i386__) || defined(__x86_64__)
2573                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2574                         return scnprintf(msg, size, "%s",
2575         "No hardware sampling interrupt available.\n"
2576         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2577 #endif
2578                 break;
2579         case EBUSY:
2580                 if (find_process("oprofiled"))
2581                         return scnprintf(msg, size,
2582         "The PMU counters are busy/taken by another profiler.\n"
2583         "We found oprofile daemon running, please stop it and try again.");
2584                 break;
2585         case EINVAL:
2586                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2587                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2588                 if (perf_missing_features.clockid)
2589                         return scnprintf(msg, size, "clockid feature not supported.");
2590                 if (perf_missing_features.clockid_wrong)
2591                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2592                 break;
2593         default:
2594                 break;
2595         }
2596
2597         return scnprintf(msg, size,
2598         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2599         "/bin/dmesg may provide additional information.\n"
2600         "No CONFIG_PERF_EVENTS=y kernel support configured?",
2601                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2602                          perf_evsel__name(evsel));
2603 }
2604
2605 char *perf_evsel__env_arch(struct perf_evsel *evsel)
2606 {
2607         if (evsel && evsel->evlist && evsel->evlist->env)
2608                 return evsel->evlist->env->arch;
2609         return NULL;
2610 }