7fde9fb7996696bd97c875661dfb54d20fd6d706
[linux-3.10.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 <linux/bitops.h>
12 #include "asm/bug.h"
13 #include <lk/debugfs.h>
14 #include "event-parse.h"
15 #include "evsel.h"
16 #include "evlist.h"
17 #include "util.h"
18 #include "cpumap.h"
19 #include "thread_map.h"
20 #include "target.h"
21 #include <linux/hw_breakpoint.h>
22 #include <linux/perf_event.h>
23 #include "perf_regs.h"
24
25 static struct {
26         bool sample_id_all;
27         bool exclude_guest;
28 } perf_missing_features;
29
30 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
31
32 static int __perf_evsel__sample_size(u64 sample_type)
33 {
34         u64 mask = sample_type & PERF_SAMPLE_MASK;
35         int size = 0;
36         int i;
37
38         for (i = 0; i < 64; i++) {
39                 if (mask & (1ULL << i))
40                         size++;
41         }
42
43         size *= sizeof(u64);
44
45         return size;
46 }
47
48 void hists__init(struct hists *hists)
49 {
50         memset(hists, 0, sizeof(*hists));
51         hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
52         hists->entries_in = &hists->entries_in_array[0];
53         hists->entries_collapsed = RB_ROOT;
54         hists->entries = RB_ROOT;
55         pthread_mutex_init(&hists->lock, NULL);
56 }
57
58 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
59                                   enum perf_event_sample_format bit)
60 {
61         if (!(evsel->attr.sample_type & bit)) {
62                 evsel->attr.sample_type |= bit;
63                 evsel->sample_size += sizeof(u64);
64         }
65 }
66
67 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
68                                     enum perf_event_sample_format bit)
69 {
70         if (evsel->attr.sample_type & bit) {
71                 evsel->attr.sample_type &= ~bit;
72                 evsel->sample_size -= sizeof(u64);
73         }
74 }
75
76 void perf_evsel__set_sample_id(struct perf_evsel *evsel)
77 {
78         perf_evsel__set_sample_bit(evsel, ID);
79         evsel->attr.read_format |= PERF_FORMAT_ID;
80 }
81
82 void perf_evsel__init(struct perf_evsel *evsel,
83                       struct perf_event_attr *attr, int idx)
84 {
85         evsel->idx         = idx;
86         evsel->attr        = *attr;
87         evsel->leader      = evsel;
88         INIT_LIST_HEAD(&evsel->node);
89         hists__init(&evsel->hists);
90         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
91 }
92
93 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
94 {
95         struct perf_evsel *evsel = zalloc(sizeof(*evsel));
96
97         if (evsel != NULL)
98                 perf_evsel__init(evsel, attr, idx);
99
100         return evsel;
101 }
102
103 struct event_format *event_format__new(const char *sys, const char *name)
104 {
105         int fd, n;
106         char *filename;
107         void *bf = NULL, *nbf;
108         size_t size = 0, alloc_size = 0;
109         struct event_format *format = NULL;
110
111         if (asprintf(&filename, "%s/%s/%s/format", tracing_events_path, sys, name) < 0)
112                 goto out;
113
114         fd = open(filename, O_RDONLY);
115         if (fd < 0)
116                 goto out_free_filename;
117
118         do {
119                 if (size == alloc_size) {
120                         alloc_size += BUFSIZ;
121                         nbf = realloc(bf, alloc_size);
122                         if (nbf == NULL)
123                                 goto out_free_bf;
124                         bf = nbf;
125                 }
126
127                 n = read(fd, bf + size, BUFSIZ);
128                 if (n < 0)
129                         goto out_free_bf;
130                 size += n;
131         } while (n > 0);
132
133         pevent_parse_format(&format, bf, size, sys);
134
135 out_free_bf:
136         free(bf);
137         close(fd);
138 out_free_filename:
139         free(filename);
140 out:
141         return format;
142 }
143
144 struct perf_evsel *perf_evsel__newtp(const char *sys, const char *name, int idx)
145 {
146         struct perf_evsel *evsel = zalloc(sizeof(*evsel));
147
148         if (evsel != NULL) {
149                 struct perf_event_attr attr = {
150                         .type          = PERF_TYPE_TRACEPOINT,
151                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
152                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
153                 };
154
155                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
156                         goto out_free;
157
158                 evsel->tp_format = event_format__new(sys, name);
159                 if (evsel->tp_format == NULL)
160                         goto out_free;
161
162                 event_attr_init(&attr);
163                 attr.config = evsel->tp_format->id;
164                 attr.sample_period = 1;
165                 perf_evsel__init(evsel, &attr, idx);
166         }
167
168         return evsel;
169
170 out_free:
171         free(evsel->name);
172         free(evsel);
173         return NULL;
174 }
175
176 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
177         "cycles",
178         "instructions",
179         "cache-references",
180         "cache-misses",
181         "branches",
182         "branch-misses",
183         "bus-cycles",
184         "stalled-cycles-frontend",
185         "stalled-cycles-backend",
186         "ref-cycles",
187 };
188
189 static const char *__perf_evsel__hw_name(u64 config)
190 {
191         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
192                 return perf_evsel__hw_names[config];
193
194         return "unknown-hardware";
195 }
196
197 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
198 {
199         int colon = 0, r = 0;
200         struct perf_event_attr *attr = &evsel->attr;
201         bool exclude_guest_default = false;
202
203 #define MOD_PRINT(context, mod) do {                                    \
204                 if (!attr->exclude_##context) {                         \
205                         if (!colon) colon = ++r;                        \
206                         r += scnprintf(bf + r, size - r, "%c", mod);    \
207                 } } while(0)
208
209         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
210                 MOD_PRINT(kernel, 'k');
211                 MOD_PRINT(user, 'u');
212                 MOD_PRINT(hv, 'h');
213                 exclude_guest_default = true;
214         }
215
216         if (attr->precise_ip) {
217                 if (!colon)
218                         colon = ++r;
219                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
220                 exclude_guest_default = true;
221         }
222
223         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
224                 MOD_PRINT(host, 'H');
225                 MOD_PRINT(guest, 'G');
226         }
227 #undef MOD_PRINT
228         if (colon)
229                 bf[colon - 1] = ':';
230         return r;
231 }
232
233 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
234 {
235         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
236         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
237 }
238
239 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
240         "cpu-clock",
241         "task-clock",
242         "page-faults",
243         "context-switches",
244         "cpu-migrations",
245         "minor-faults",
246         "major-faults",
247         "alignment-faults",
248         "emulation-faults",
249 };
250
251 static const char *__perf_evsel__sw_name(u64 config)
252 {
253         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
254                 return perf_evsel__sw_names[config];
255         return "unknown-software";
256 }
257
258 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
259 {
260         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
261         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
262 }
263
264 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
265 {
266         int r;
267
268         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
269
270         if (type & HW_BREAKPOINT_R)
271                 r += scnprintf(bf + r, size - r, "r");
272
273         if (type & HW_BREAKPOINT_W)
274                 r += scnprintf(bf + r, size - r, "w");
275
276         if (type & HW_BREAKPOINT_X)
277                 r += scnprintf(bf + r, size - r, "x");
278
279         return r;
280 }
281
282 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
283 {
284         struct perf_event_attr *attr = &evsel->attr;
285         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
286         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
287 }
288
289 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
290                                 [PERF_EVSEL__MAX_ALIASES] = {
291  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
292  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
293  { "LLC",       "L2",                                                   },
294  { "dTLB",      "d-tlb",        "Data-TLB",                             },
295  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
296  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
297  { "node",                                                              },
298 };
299
300 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
301                                    [PERF_EVSEL__MAX_ALIASES] = {
302  { "load",      "loads",        "read",                                 },
303  { "store",     "stores",       "write",                                },
304  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
305 };
306
307 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
308                                        [PERF_EVSEL__MAX_ALIASES] = {
309  { "refs",      "Reference",    "ops",          "access",               },
310  { "misses",    "miss",                                                 },
311 };
312
313 #define C(x)            PERF_COUNT_HW_CACHE_##x
314 #define CACHE_READ      (1 << C(OP_READ))
315 #define CACHE_WRITE     (1 << C(OP_WRITE))
316 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
317 #define COP(x)          (1 << x)
318
319 /*
320  * cache operartion stat
321  * L1I : Read and prefetch only
322  * ITLB and BPU : Read-only
323  */
324 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
325  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
326  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
327  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
328  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
329  [C(ITLB)]      = (CACHE_READ),
330  [C(BPU)]       = (CACHE_READ),
331  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
332 };
333
334 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
335 {
336         if (perf_evsel__hw_cache_stat[type] & COP(op))
337                 return true;    /* valid */
338         else
339                 return false;   /* invalid */
340 }
341
342 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
343                                             char *bf, size_t size)
344 {
345         if (result) {
346                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
347                                  perf_evsel__hw_cache_op[op][0],
348                                  perf_evsel__hw_cache_result[result][0]);
349         }
350
351         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
352                          perf_evsel__hw_cache_op[op][1]);
353 }
354
355 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
356 {
357         u8 op, result, type = (config >>  0) & 0xff;
358         const char *err = "unknown-ext-hardware-cache-type";
359
360         if (type > PERF_COUNT_HW_CACHE_MAX)
361                 goto out_err;
362
363         op = (config >>  8) & 0xff;
364         err = "unknown-ext-hardware-cache-op";
365         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
366                 goto out_err;
367
368         result = (config >> 16) & 0xff;
369         err = "unknown-ext-hardware-cache-result";
370         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
371                 goto out_err;
372
373         err = "invalid-cache";
374         if (!perf_evsel__is_cache_op_valid(type, op))
375                 goto out_err;
376
377         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
378 out_err:
379         return scnprintf(bf, size, "%s", err);
380 }
381
382 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
383 {
384         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
385         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
386 }
387
388 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
389 {
390         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
391         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
392 }
393
394 const char *perf_evsel__name(struct perf_evsel *evsel)
395 {
396         char bf[128];
397
398         if (evsel->name)
399                 return evsel->name;
400
401         switch (evsel->attr.type) {
402         case PERF_TYPE_RAW:
403                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
404                 break;
405
406         case PERF_TYPE_HARDWARE:
407                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
408                 break;
409
410         case PERF_TYPE_HW_CACHE:
411                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
412                 break;
413
414         case PERF_TYPE_SOFTWARE:
415                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
416                 break;
417
418         case PERF_TYPE_TRACEPOINT:
419                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
420                 break;
421
422         case PERF_TYPE_BREAKPOINT:
423                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
424                 break;
425
426         default:
427                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
428                           evsel->attr.type);
429                 break;
430         }
431
432         evsel->name = strdup(bf);
433
434         return evsel->name ?: "unknown";
435 }
436
437 const char *perf_evsel__group_name(struct perf_evsel *evsel)
438 {
439         return evsel->group_name ?: "anon group";
440 }
441
442 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
443 {
444         int ret;
445         struct perf_evsel *pos;
446         const char *group_name = perf_evsel__group_name(evsel);
447
448         ret = scnprintf(buf, size, "%s", group_name);
449
450         ret += scnprintf(buf + ret, size - ret, " { %s",
451                          perf_evsel__name(evsel));
452
453         for_each_group_member(pos, evsel)
454                 ret += scnprintf(buf + ret, size - ret, ", %s",
455                                  perf_evsel__name(pos));
456
457         ret += scnprintf(buf + ret, size - ret, " }");
458
459         return ret;
460 }
461
462 /*
463  * The enable_on_exec/disabled value strategy:
464  *
465  *  1) For any type of traced program:
466  *    - all independent events and group leaders are disabled
467  *    - all group members are enabled
468  *
469  *     Group members are ruled by group leaders. They need to
470  *     be enabled, because the group scheduling relies on that.
471  *
472  *  2) For traced programs executed by perf:
473  *     - all independent events and group leaders have
474  *       enable_on_exec set
475  *     - we don't specifically enable or disable any event during
476  *       the record command
477  *
478  *     Independent events and group leaders are initially disabled
479  *     and get enabled by exec. Group members are ruled by group
480  *     leaders as stated in 1).
481  *
482  *  3) For traced programs attached by perf (pid/tid):
483  *     - we specifically enable or disable all events during
484  *       the record command
485  *
486  *     When attaching events to already running traced we
487  *     enable/disable events specifically, as there's no
488  *     initial traced exec call.
489  */
490 void perf_evsel__config(struct perf_evsel *evsel,
491                         struct perf_record_opts *opts)
492 {
493         struct perf_event_attr *attr = &evsel->attr;
494         int track = !evsel->idx; /* only the first counter needs these */
495
496         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
497         attr->inherit       = !opts->no_inherit;
498
499         perf_evsel__set_sample_bit(evsel, IP);
500         perf_evsel__set_sample_bit(evsel, TID);
501
502         /*
503          * We default some events to a 1 default interval. But keep
504          * it a weak assumption overridable by the user.
505          */
506         if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
507                                      opts->user_interval != ULLONG_MAX)) {
508                 if (opts->freq) {
509                         perf_evsel__set_sample_bit(evsel, PERIOD);
510                         attr->freq              = 1;
511                         attr->sample_freq       = opts->freq;
512                 } else {
513                         attr->sample_period = opts->default_interval;
514                 }
515         }
516
517         if (opts->no_samples)
518                 attr->sample_freq = 0;
519
520         if (opts->inherit_stat)
521                 attr->inherit_stat = 1;
522
523         if (opts->sample_address) {
524                 perf_evsel__set_sample_bit(evsel, ADDR);
525                 attr->mmap_data = track;
526         }
527
528         if (opts->call_graph) {
529                 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
530
531                 if (opts->call_graph == CALLCHAIN_DWARF) {
532                         perf_evsel__set_sample_bit(evsel, REGS_USER);
533                         perf_evsel__set_sample_bit(evsel, STACK_USER);
534                         attr->sample_regs_user = PERF_REGS_MASK;
535                         attr->sample_stack_user = opts->stack_dump_size;
536                         attr->exclude_callchain_user = 1;
537                 }
538         }
539
540         if (perf_target__has_cpu(&opts->target))
541                 perf_evsel__set_sample_bit(evsel, CPU);
542
543         if (opts->period)
544                 perf_evsel__set_sample_bit(evsel, PERIOD);
545
546         if (!perf_missing_features.sample_id_all &&
547             (opts->sample_time || !opts->no_inherit ||
548              perf_target__has_cpu(&opts->target)))
549                 perf_evsel__set_sample_bit(evsel, TIME);
550
551         if (opts->raw_samples) {
552                 perf_evsel__set_sample_bit(evsel, TIME);
553                 perf_evsel__set_sample_bit(evsel, RAW);
554                 perf_evsel__set_sample_bit(evsel, CPU);
555         }
556
557         if (opts->no_delay) {
558                 attr->watermark = 0;
559                 attr->wakeup_events = 1;
560         }
561         if (opts->branch_stack) {
562                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
563                 attr->branch_sample_type = opts->branch_stack;
564         }
565
566         attr->mmap = track;
567         attr->comm = track;
568
569         /*
570          * XXX see the function comment above
571          *
572          * Disabling only independent events or group leaders,
573          * keeping group members enabled.
574          */
575         if (perf_evsel__is_group_leader(evsel))
576                 attr->disabled = 1;
577
578         /*
579          * Setting enable_on_exec for independent events and
580          * group leaders for traced executed by perf.
581          */
582         if (perf_target__none(&opts->target) && perf_evsel__is_group_leader(evsel))
583                 attr->enable_on_exec = 1;
584 }
585
586 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
587 {
588         int cpu, thread;
589         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
590
591         if (evsel->fd) {
592                 for (cpu = 0; cpu < ncpus; cpu++) {
593                         for (thread = 0; thread < nthreads; thread++) {
594                                 FD(evsel, cpu, thread) = -1;
595                         }
596                 }
597         }
598
599         return evsel->fd != NULL ? 0 : -ENOMEM;
600 }
601
602 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
603                            const char *filter)
604 {
605         int cpu, thread;
606
607         for (cpu = 0; cpu < ncpus; cpu++) {
608                 for (thread = 0; thread < nthreads; thread++) {
609                         int fd = FD(evsel, cpu, thread),
610                             err = ioctl(fd, PERF_EVENT_IOC_SET_FILTER, filter);
611
612                         if (err)
613                                 return err;
614                 }
615         }
616
617         return 0;
618 }
619
620 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
621 {
622         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
623         if (evsel->sample_id == NULL)
624                 return -ENOMEM;
625
626         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
627         if (evsel->id == NULL) {
628                 xyarray__delete(evsel->sample_id);
629                 evsel->sample_id = NULL;
630                 return -ENOMEM;
631         }
632
633         return 0;
634 }
635
636 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
637 {
638         evsel->counts = zalloc((sizeof(*evsel->counts) +
639                                 (ncpus * sizeof(struct perf_counts_values))));
640         return evsel->counts != NULL ? 0 : -ENOMEM;
641 }
642
643 void perf_evsel__free_fd(struct perf_evsel *evsel)
644 {
645         xyarray__delete(evsel->fd);
646         evsel->fd = NULL;
647 }
648
649 void perf_evsel__free_id(struct perf_evsel *evsel)
650 {
651         xyarray__delete(evsel->sample_id);
652         evsel->sample_id = NULL;
653         free(evsel->id);
654         evsel->id = NULL;
655 }
656
657 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
658 {
659         int cpu, thread;
660
661         for (cpu = 0; cpu < ncpus; cpu++)
662                 for (thread = 0; thread < nthreads; ++thread) {
663                         close(FD(evsel, cpu, thread));
664                         FD(evsel, cpu, thread) = -1;
665                 }
666 }
667
668 void perf_evsel__free_counts(struct perf_evsel *evsel)
669 {
670         free(evsel->counts);
671 }
672
673 void perf_evsel__exit(struct perf_evsel *evsel)
674 {
675         assert(list_empty(&evsel->node));
676         perf_evsel__free_fd(evsel);
677         perf_evsel__free_id(evsel);
678 }
679
680 void perf_evsel__delete(struct perf_evsel *evsel)
681 {
682         perf_evsel__exit(evsel);
683         close_cgroup(evsel->cgrp);
684         free(evsel->group_name);
685         if (evsel->tp_format)
686                 pevent_free_format(evsel->tp_format);
687         free(evsel->name);
688         free(evsel);
689 }
690
691 static inline void compute_deltas(struct perf_evsel *evsel,
692                                   int cpu,
693                                   struct perf_counts_values *count)
694 {
695         struct perf_counts_values tmp;
696
697         if (!evsel->prev_raw_counts)
698                 return;
699
700         if (cpu == -1) {
701                 tmp = evsel->prev_raw_counts->aggr;
702                 evsel->prev_raw_counts->aggr = *count;
703         } else {
704                 tmp = evsel->prev_raw_counts->cpu[cpu];
705                 evsel->prev_raw_counts->cpu[cpu] = *count;
706         }
707
708         count->val = count->val - tmp.val;
709         count->ena = count->ena - tmp.ena;
710         count->run = count->run - tmp.run;
711 }
712
713 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
714                               int cpu, int thread, bool scale)
715 {
716         struct perf_counts_values count;
717         size_t nv = scale ? 3 : 1;
718
719         if (FD(evsel, cpu, thread) < 0)
720                 return -EINVAL;
721
722         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
723                 return -ENOMEM;
724
725         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
726                 return -errno;
727
728         compute_deltas(evsel, cpu, &count);
729
730         if (scale) {
731                 if (count.run == 0)
732                         count.val = 0;
733                 else if (count.run < count.ena)
734                         count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
735         } else
736                 count.ena = count.run = 0;
737
738         evsel->counts->cpu[cpu] = count;
739         return 0;
740 }
741
742 int __perf_evsel__read(struct perf_evsel *evsel,
743                        int ncpus, int nthreads, bool scale)
744 {
745         size_t nv = scale ? 3 : 1;
746         int cpu, thread;
747         struct perf_counts_values *aggr = &evsel->counts->aggr, count;
748
749         aggr->val = aggr->ena = aggr->run = 0;
750
751         for (cpu = 0; cpu < ncpus; cpu++) {
752                 for (thread = 0; thread < nthreads; thread++) {
753                         if (FD(evsel, cpu, thread) < 0)
754                                 continue;
755
756                         if (readn(FD(evsel, cpu, thread),
757                                   &count, nv * sizeof(u64)) < 0)
758                                 return -errno;
759
760                         aggr->val += count.val;
761                         if (scale) {
762                                 aggr->ena += count.ena;
763                                 aggr->run += count.run;
764                         }
765                 }
766         }
767
768         compute_deltas(evsel, -1, aggr);
769
770         evsel->counts->scaled = 0;
771         if (scale) {
772                 if (aggr->run == 0) {
773                         evsel->counts->scaled = -1;
774                         aggr->val = 0;
775                         return 0;
776                 }
777
778                 if (aggr->run < aggr->ena) {
779                         evsel->counts->scaled = 1;
780                         aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
781                 }
782         } else
783                 aggr->ena = aggr->run = 0;
784
785         return 0;
786 }
787
788 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
789 {
790         struct perf_evsel *leader = evsel->leader;
791         int fd;
792
793         if (perf_evsel__is_group_leader(evsel))
794                 return -1;
795
796         /*
797          * Leader must be already processed/open,
798          * if not it's a bug.
799          */
800         BUG_ON(!leader->fd);
801
802         fd = FD(leader, cpu, thread);
803         BUG_ON(fd == -1);
804
805         return fd;
806 }
807
808 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
809                               struct thread_map *threads)
810 {
811         int cpu, thread;
812         unsigned long flags = 0;
813         int pid = -1, err;
814
815         if (evsel->fd == NULL &&
816             perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
817                 return -ENOMEM;
818
819         if (evsel->cgrp) {
820                 flags = PERF_FLAG_PID_CGROUP;
821                 pid = evsel->cgrp->fd;
822         }
823
824 fallback_missing_features:
825         if (perf_missing_features.exclude_guest)
826                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
827 retry_sample_id:
828         if (perf_missing_features.sample_id_all)
829                 evsel->attr.sample_id_all = 0;
830
831         for (cpu = 0; cpu < cpus->nr; cpu++) {
832
833                 for (thread = 0; thread < threads->nr; thread++) {
834                         int group_fd;
835
836                         if (!evsel->cgrp)
837                                 pid = threads->map[thread];
838
839                         group_fd = get_group_fd(evsel, cpu, thread);
840
841                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
842                                                                      pid,
843                                                                      cpus->map[cpu],
844                                                                      group_fd, flags);
845                         if (FD(evsel, cpu, thread) < 0) {
846                                 err = -errno;
847                                 goto try_fallback;
848                         }
849                 }
850         }
851
852         return 0;
853
854 try_fallback:
855         if (err != -EINVAL || cpu > 0 || thread > 0)
856                 goto out_close;
857
858         if (!perf_missing_features.exclude_guest &&
859             (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
860                 perf_missing_features.exclude_guest = true;
861                 goto fallback_missing_features;
862         } else if (!perf_missing_features.sample_id_all) {
863                 perf_missing_features.sample_id_all = true;
864                 goto retry_sample_id;
865         }
866
867 out_close:
868         do {
869                 while (--thread >= 0) {
870                         close(FD(evsel, cpu, thread));
871                         FD(evsel, cpu, thread) = -1;
872                 }
873                 thread = threads->nr;
874         } while (--cpu >= 0);
875         return err;
876 }
877
878 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
879 {
880         if (evsel->fd == NULL)
881                 return;
882
883         perf_evsel__close_fd(evsel, ncpus, nthreads);
884         perf_evsel__free_fd(evsel);
885         evsel->fd = NULL;
886 }
887
888 static struct {
889         struct cpu_map map;
890         int cpus[1];
891 } empty_cpu_map = {
892         .map.nr = 1,
893         .cpus   = { -1, },
894 };
895
896 static struct {
897         struct thread_map map;
898         int threads[1];
899 } empty_thread_map = {
900         .map.nr  = 1,
901         .threads = { -1, },
902 };
903
904 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
905                      struct thread_map *threads)
906 {
907         if (cpus == NULL) {
908                 /* Work around old compiler warnings about strict aliasing */
909                 cpus = &empty_cpu_map.map;
910         }
911
912         if (threads == NULL)
913                 threads = &empty_thread_map.map;
914
915         return __perf_evsel__open(evsel, cpus, threads);
916 }
917
918 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
919                              struct cpu_map *cpus)
920 {
921         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
922 }
923
924 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
925                                 struct thread_map *threads)
926 {
927         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
928 }
929
930 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
931                                        const union perf_event *event,
932                                        struct perf_sample *sample)
933 {
934         u64 type = evsel->attr.sample_type;
935         const u64 *array = event->sample.array;
936         bool swapped = evsel->needs_swap;
937         union u64_swap u;
938
939         array += ((event->header.size -
940                    sizeof(event->header)) / sizeof(u64)) - 1;
941
942         if (type & PERF_SAMPLE_CPU) {
943                 u.val64 = *array;
944                 if (swapped) {
945                         /* undo swap of u64, then swap on individual u32s */
946                         u.val64 = bswap_64(u.val64);
947                         u.val32[0] = bswap_32(u.val32[0]);
948                 }
949
950                 sample->cpu = u.val32[0];
951                 array--;
952         }
953
954         if (type & PERF_SAMPLE_STREAM_ID) {
955                 sample->stream_id = *array;
956                 array--;
957         }
958
959         if (type & PERF_SAMPLE_ID) {
960                 sample->id = *array;
961                 array--;
962         }
963
964         if (type & PERF_SAMPLE_TIME) {
965                 sample->time = *array;
966                 array--;
967         }
968
969         if (type & PERF_SAMPLE_TID) {
970                 u.val64 = *array;
971                 if (swapped) {
972                         /* undo swap of u64, then swap on individual u32s */
973                         u.val64 = bswap_64(u.val64);
974                         u.val32[0] = bswap_32(u.val32[0]);
975                         u.val32[1] = bswap_32(u.val32[1]);
976                 }
977
978                 sample->pid = u.val32[0];
979                 sample->tid = u.val32[1];
980         }
981
982         return 0;
983 }
984
985 static bool sample_overlap(const union perf_event *event,
986                            const void *offset, u64 size)
987 {
988         const void *base = event;
989
990         if (offset + size > base + event->header.size)
991                 return true;
992
993         return false;
994 }
995
996 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
997                              struct perf_sample *data)
998 {
999         u64 type = evsel->attr.sample_type;
1000         u64 regs_user = evsel->attr.sample_regs_user;
1001         bool swapped = evsel->needs_swap;
1002         const u64 *array;
1003
1004         /*
1005          * used for cross-endian analysis. See git commit 65014ab3
1006          * for why this goofiness is needed.
1007          */
1008         union u64_swap u;
1009
1010         memset(data, 0, sizeof(*data));
1011         data->cpu = data->pid = data->tid = -1;
1012         data->stream_id = data->id = data->time = -1ULL;
1013         data->period = 1;
1014
1015         if (event->header.type != PERF_RECORD_SAMPLE) {
1016                 if (!evsel->attr.sample_id_all)
1017                         return 0;
1018                 return perf_evsel__parse_id_sample(evsel, event, data);
1019         }
1020
1021         array = event->sample.array;
1022
1023         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1024                 return -EFAULT;
1025
1026         if (type & PERF_SAMPLE_IP) {
1027                 data->ip = event->ip.ip;
1028                 array++;
1029         }
1030
1031         if (type & PERF_SAMPLE_TID) {
1032                 u.val64 = *array;
1033                 if (swapped) {
1034                         /* undo swap of u64, then swap on individual u32s */
1035                         u.val64 = bswap_64(u.val64);
1036                         u.val32[0] = bswap_32(u.val32[0]);
1037                         u.val32[1] = bswap_32(u.val32[1]);
1038                 }
1039
1040                 data->pid = u.val32[0];
1041                 data->tid = u.val32[1];
1042                 array++;
1043         }
1044
1045         if (type & PERF_SAMPLE_TIME) {
1046                 data->time = *array;
1047                 array++;
1048         }
1049
1050         data->addr = 0;
1051         if (type & PERF_SAMPLE_ADDR) {
1052                 data->addr = *array;
1053                 array++;
1054         }
1055
1056         data->id = -1ULL;
1057         if (type & PERF_SAMPLE_ID) {
1058                 data->id = *array;
1059                 array++;
1060         }
1061
1062         if (type & PERF_SAMPLE_STREAM_ID) {
1063                 data->stream_id = *array;
1064                 array++;
1065         }
1066
1067         if (type & PERF_SAMPLE_CPU) {
1068
1069                 u.val64 = *array;
1070                 if (swapped) {
1071                         /* undo swap of u64, then swap on individual u32s */
1072                         u.val64 = bswap_64(u.val64);
1073                         u.val32[0] = bswap_32(u.val32[0]);
1074                 }
1075
1076                 data->cpu = u.val32[0];
1077                 array++;
1078         }
1079
1080         if (type & PERF_SAMPLE_PERIOD) {
1081                 data->period = *array;
1082                 array++;
1083         }
1084
1085         if (type & PERF_SAMPLE_READ) {
1086                 fprintf(stderr, "PERF_SAMPLE_READ is unsupported for now\n");
1087                 return -1;
1088         }
1089
1090         if (type & PERF_SAMPLE_CALLCHAIN) {
1091                 if (sample_overlap(event, array, sizeof(data->callchain->nr)))
1092                         return -EFAULT;
1093
1094                 data->callchain = (struct ip_callchain *)array;
1095
1096                 if (sample_overlap(event, array, data->callchain->nr))
1097                         return -EFAULT;
1098
1099                 array += 1 + data->callchain->nr;
1100         }
1101
1102         if (type & PERF_SAMPLE_RAW) {
1103                 const u64 *pdata;
1104
1105                 u.val64 = *array;
1106                 if (WARN_ONCE(swapped,
1107                               "Endianness of raw data not corrected!\n")) {
1108                         /* undo swap of u64, then swap on individual u32s */
1109                         u.val64 = bswap_64(u.val64);
1110                         u.val32[0] = bswap_32(u.val32[0]);
1111                         u.val32[1] = bswap_32(u.val32[1]);
1112                 }
1113
1114                 if (sample_overlap(event, array, sizeof(u32)))
1115                         return -EFAULT;
1116
1117                 data->raw_size = u.val32[0];
1118                 pdata = (void *) array + sizeof(u32);
1119
1120                 if (sample_overlap(event, pdata, data->raw_size))
1121                         return -EFAULT;
1122
1123                 data->raw_data = (void *) pdata;
1124
1125                 array = (void *)array + data->raw_size + sizeof(u32);
1126         }
1127
1128         if (type & PERF_SAMPLE_BRANCH_STACK) {
1129                 u64 sz;
1130
1131                 data->branch_stack = (struct branch_stack *)array;
1132                 array++; /* nr */
1133
1134                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1135                 sz /= sizeof(u64);
1136                 array += sz;
1137         }
1138
1139         if (type & PERF_SAMPLE_REGS_USER) {
1140                 /* First u64 tells us if we have any regs in sample. */
1141                 u64 avail = *array++;
1142
1143                 if (avail) {
1144                         data->user_regs.regs = (u64 *)array;
1145                         array += hweight_long(regs_user);
1146                 }
1147         }
1148
1149         if (type & PERF_SAMPLE_STACK_USER) {
1150                 u64 size = *array++;
1151
1152                 data->user_stack.offset = ((char *)(array - 1)
1153                                           - (char *) event);
1154
1155                 if (!size) {
1156                         data->user_stack.size = 0;
1157                 } else {
1158                         data->user_stack.data = (char *)array;
1159                         array += size / sizeof(*array);
1160                         data->user_stack.size = *array;
1161                 }
1162         }
1163
1164         return 0;
1165 }
1166
1167 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1168                                   const struct perf_sample *sample,
1169                                   bool swapped)
1170 {
1171         u64 *array;
1172
1173         /*
1174          * used for cross-endian analysis. See git commit 65014ab3
1175          * for why this goofiness is needed.
1176          */
1177         union u64_swap u;
1178
1179         array = event->sample.array;
1180
1181         if (type & PERF_SAMPLE_IP) {
1182                 event->ip.ip = sample->ip;
1183                 array++;
1184         }
1185
1186         if (type & PERF_SAMPLE_TID) {
1187                 u.val32[0] = sample->pid;
1188                 u.val32[1] = sample->tid;
1189                 if (swapped) {
1190                         /*
1191                          * Inverse of what is done in perf_evsel__parse_sample
1192                          */
1193                         u.val32[0] = bswap_32(u.val32[0]);
1194                         u.val32[1] = bswap_32(u.val32[1]);
1195                         u.val64 = bswap_64(u.val64);
1196                 }
1197
1198                 *array = u.val64;
1199                 array++;
1200         }
1201
1202         if (type & PERF_SAMPLE_TIME) {
1203                 *array = sample->time;
1204                 array++;
1205         }
1206
1207         if (type & PERF_SAMPLE_ADDR) {
1208                 *array = sample->addr;
1209                 array++;
1210         }
1211
1212         if (type & PERF_SAMPLE_ID) {
1213                 *array = sample->id;
1214                 array++;
1215         }
1216
1217         if (type & PERF_SAMPLE_STREAM_ID) {
1218                 *array = sample->stream_id;
1219                 array++;
1220         }
1221
1222         if (type & PERF_SAMPLE_CPU) {
1223                 u.val32[0] = sample->cpu;
1224                 if (swapped) {
1225                         /*
1226                          * Inverse of what is done in perf_evsel__parse_sample
1227                          */
1228                         u.val32[0] = bswap_32(u.val32[0]);
1229                         u.val64 = bswap_64(u.val64);
1230                 }
1231                 *array = u.val64;
1232                 array++;
1233         }
1234
1235         if (type & PERF_SAMPLE_PERIOD) {
1236                 *array = sample->period;
1237                 array++;
1238         }
1239
1240         return 0;
1241 }
1242
1243 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1244 {
1245         return pevent_find_field(evsel->tp_format, name);
1246 }
1247
1248 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1249                          const char *name)
1250 {
1251         struct format_field *field = perf_evsel__field(evsel, name);
1252         int offset;
1253
1254         if (!field)
1255                 return NULL;
1256
1257         offset = field->offset;
1258
1259         if (field->flags & FIELD_IS_DYNAMIC) {
1260                 offset = *(int *)(sample->raw_data + field->offset);
1261                 offset &= 0xffff;
1262         }
1263
1264         return sample->raw_data + offset;
1265 }
1266
1267 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1268                        const char *name)
1269 {
1270         struct format_field *field = perf_evsel__field(evsel, name);
1271         void *ptr;
1272         u64 value;
1273
1274         if (!field)
1275                 return 0;
1276
1277         ptr = sample->raw_data + field->offset;
1278
1279         switch (field->size) {
1280         case 1:
1281                 return *(u8 *)ptr;
1282         case 2:
1283                 value = *(u16 *)ptr;
1284                 break;
1285         case 4:
1286                 value = *(u32 *)ptr;
1287                 break;
1288         case 8:
1289                 value = *(u64 *)ptr;
1290                 break;
1291         default:
1292                 return 0;
1293         }
1294
1295         if (!evsel->needs_swap)
1296                 return value;
1297
1298         switch (field->size) {
1299         case 2:
1300                 return bswap_16(value);
1301         case 4:
1302                 return bswap_32(value);
1303         case 8:
1304                 return bswap_64(value);
1305         default:
1306                 return 0;
1307         }
1308
1309         return 0;
1310 }
1311
1312 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1313 {
1314         va_list args;
1315         int ret = 0;
1316
1317         if (!*first) {
1318                 ret += fprintf(fp, ",");
1319         } else {
1320                 ret += fprintf(fp, ":");
1321                 *first = false;
1322         }
1323
1324         va_start(args, fmt);
1325         ret += vfprintf(fp, fmt, args);
1326         va_end(args);
1327         return ret;
1328 }
1329
1330 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1331 {
1332         if (value == 0)
1333                 return 0;
1334
1335         return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1336 }
1337
1338 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1339
1340 struct bit_names {
1341         int bit;
1342         const char *name;
1343 };
1344
1345 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1346                          struct bit_names *bits, bool *first)
1347 {
1348         int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1349         bool first_bit = true;
1350
1351         do {
1352                 if (value & bits[i].bit) {
1353                         printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1354                         first_bit = false;
1355                 }
1356         } while (bits[++i].name != NULL);
1357
1358         return printed;
1359 }
1360
1361 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1362 {
1363 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1364         struct bit_names bits[] = {
1365                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1366                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1367                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1368                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1369                 { .name = NULL, }
1370         };
1371 #undef bit_name
1372         return bits__fprintf(fp, "sample_type", value, bits, first);
1373 }
1374
1375 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1376 {
1377 #define bit_name(n) { PERF_FORMAT_##n, #n }
1378         struct bit_names bits[] = {
1379                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1380                 bit_name(ID), bit_name(GROUP),
1381                 { .name = NULL, }
1382         };
1383 #undef bit_name
1384         return bits__fprintf(fp, "read_format", value, bits, first);
1385 }
1386
1387 int perf_evsel__fprintf(struct perf_evsel *evsel,
1388                         struct perf_attr_details *details, FILE *fp)
1389 {
1390         bool first = true;
1391         int printed = 0;
1392
1393         if (details->event_group) {
1394                 struct perf_evsel *pos;
1395
1396                 if (!perf_evsel__is_group_leader(evsel))
1397                         return 0;
1398
1399                 if (evsel->nr_members > 1)
1400                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1401
1402                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1403                 for_each_group_member(pos, evsel)
1404                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1405
1406                 if (evsel->nr_members > 1)
1407                         printed += fprintf(fp, "}");
1408                 goto out;
1409         }
1410
1411         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1412
1413         if (details->verbose || details->freq) {
1414                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1415                                          (u64)evsel->attr.sample_freq);
1416         }
1417
1418         if (details->verbose) {
1419                 if_print(type);
1420                 if_print(config);
1421                 if_print(config1);
1422                 if_print(config2);
1423                 if_print(size);
1424                 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1425                 if (evsel->attr.read_format)
1426                         printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1427                 if_print(disabled);
1428                 if_print(inherit);
1429                 if_print(pinned);
1430                 if_print(exclusive);
1431                 if_print(exclude_user);
1432                 if_print(exclude_kernel);
1433                 if_print(exclude_hv);
1434                 if_print(exclude_idle);
1435                 if_print(mmap);
1436                 if_print(comm);
1437                 if_print(freq);
1438                 if_print(inherit_stat);
1439                 if_print(enable_on_exec);
1440                 if_print(task);
1441                 if_print(watermark);
1442                 if_print(precise_ip);
1443                 if_print(mmap_data);
1444                 if_print(sample_id_all);
1445                 if_print(exclude_host);
1446                 if_print(exclude_guest);
1447                 if_print(__reserved_1);
1448                 if_print(wakeup_events);
1449                 if_print(bp_type);
1450                 if_print(branch_sample_type);
1451         }
1452 out:
1453         fputc('\n', fp);
1454         return ++printed;
1455 }
1456
1457 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1458                           char *msg, size_t msgsize)
1459 {
1460         if ((err == ENOENT || err == ENXIO) &&
1461             evsel->attr.type   == PERF_TYPE_HARDWARE &&
1462             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1463                 /*
1464                  * If it's cycles then fall back to hrtimer based
1465                  * cpu-clock-tick sw counter, which is always available even if
1466                  * no PMU support.
1467                  *
1468                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1469                  * b0a873e).
1470                  */
1471                 scnprintf(msg, msgsize, "%s",
1472 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1473
1474                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
1475                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
1476
1477                 free(evsel->name);
1478                 evsel->name = NULL;
1479                 return true;
1480         }
1481
1482         return false;
1483 }
1484
1485 int perf_evsel__open_strerror(struct perf_evsel *evsel,
1486                               struct perf_target *target,
1487                               int err, char *msg, size_t size)
1488 {
1489         switch (err) {
1490         case EPERM:
1491         case EACCES:
1492                 return scnprintf(msg, size, "%s",
1493                  "You may not have permission to collect %sstats.\n"
1494                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
1495                  " -1 - Not paranoid at all\n"
1496                  "  0 - Disallow raw tracepoint access for unpriv\n"
1497                  "  1 - Disallow cpu events for unpriv\n"
1498                  "  2 - Disallow kernel profiling for unpriv",
1499                                  target->system_wide ? "system-wide " : "");
1500         case ENOENT:
1501                 return scnprintf(msg, size, "The %s event is not supported.",
1502                                  perf_evsel__name(evsel));
1503         case EMFILE:
1504                 return scnprintf(msg, size, "%s",
1505                          "Too many events are opened.\n"
1506                          "Try again after reducing the number of events.");
1507         case ENODEV:
1508                 if (target->cpu_list)
1509                         return scnprintf(msg, size, "%s",
1510          "No such device - did you specify an out-of-range profile CPU?\n");
1511                 break;
1512         case EOPNOTSUPP:
1513                 if (evsel->attr.precise_ip)
1514                         return scnprintf(msg, size, "%s",
1515         "\'precise\' request may not be supported. Try removing 'p' modifier.");
1516 #if defined(__i386__) || defined(__x86_64__)
1517                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
1518                         return scnprintf(msg, size, "%s",
1519         "No hardware sampling interrupt available.\n"
1520         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
1521 #endif
1522                 break;
1523         default:
1524                 break;
1525         }
1526
1527         return scnprintf(msg, size,
1528         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).  \n"
1529         "/bin/dmesg may provide additional information.\n"
1530         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
1531                          err, strerror(err), perf_evsel__name(evsel));
1532 }