d5fb60760bacd6e9b5b6a18ccecb86962e4623f4
[~shefty/rdma-dev.git] / tools / perf / util / session.c
1 #define _FILE_OFFSET_BITS 64
2
3 #include <linux/kernel.h>
4
5 #include <byteswap.h>
6 #include <unistd.h>
7 #include <sys/types.h>
8 #include <sys/mman.h>
9
10 #include "evlist.h"
11 #include "evsel.h"
12 #include "session.h"
13 #include "tool.h"
14 #include "sort.h"
15 #include "util.h"
16 #include "cpumap.h"
17 #include "event-parse.h"
18 #include "perf_regs.h"
19 #include "unwind.h"
20 #include "vdso.h"
21
22 static int perf_session__open(struct perf_session *self, bool force)
23 {
24         struct stat input_stat;
25
26         if (!strcmp(self->filename, "-")) {
27                 self->fd_pipe = true;
28                 self->fd = STDIN_FILENO;
29
30                 if (perf_session__read_header(self, self->fd) < 0)
31                         pr_err("incompatible file format (rerun with -v to learn more)");
32
33                 return 0;
34         }
35
36         self->fd = open(self->filename, O_RDONLY);
37         if (self->fd < 0) {
38                 int err = errno;
39
40                 pr_err("failed to open %s: %s", self->filename, strerror(err));
41                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
42                         pr_err("  (try 'perf record' first)");
43                 pr_err("\n");
44                 return -errno;
45         }
46
47         if (fstat(self->fd, &input_stat) < 0)
48                 goto out_close;
49
50         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
51                 pr_err("file %s not owned by current user or root\n",
52                        self->filename);
53                 goto out_close;
54         }
55
56         if (!input_stat.st_size) {
57                 pr_info("zero-sized file (%s), nothing to do!\n",
58                         self->filename);
59                 goto out_close;
60         }
61
62         if (perf_session__read_header(self, self->fd) < 0) {
63                 pr_err("incompatible file format (rerun with -v to learn more)");
64                 goto out_close;
65         }
66
67         if (!perf_evlist__valid_sample_type(self->evlist)) {
68                 pr_err("non matching sample_type");
69                 goto out_close;
70         }
71
72         if (!perf_evlist__valid_sample_id_all(self->evlist)) {
73                 pr_err("non matching sample_id_all");
74                 goto out_close;
75         }
76
77         self->size = input_stat.st_size;
78         return 0;
79
80 out_close:
81         close(self->fd);
82         self->fd = -1;
83         return -1;
84 }
85
86 void perf_session__set_id_hdr_size(struct perf_session *session)
87 {
88         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
89
90         session->host_machine.id_hdr_size = id_hdr_size;
91         machines__set_id_hdr_size(&session->machines, id_hdr_size);
92 }
93
94 int perf_session__create_kernel_maps(struct perf_session *self)
95 {
96         int ret = machine__create_kernel_maps(&self->host_machine);
97
98         if (ret >= 0)
99                 ret = machines__create_guest_kernel_maps(&self->machines);
100         return ret;
101 }
102
103 static void perf_session__destroy_kernel_maps(struct perf_session *self)
104 {
105         machine__destroy_kernel_maps(&self->host_machine);
106         machines__destroy_guest_kernel_maps(&self->machines);
107 }
108
109 struct perf_session *perf_session__new(const char *filename, int mode,
110                                        bool force, bool repipe,
111                                        struct perf_tool *tool)
112 {
113         struct perf_session *self;
114         struct stat st;
115         size_t len;
116
117         if (!filename || !strlen(filename)) {
118                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
119                         filename = "-";
120                 else
121                         filename = "perf.data";
122         }
123
124         len = strlen(filename);
125         self = zalloc(sizeof(*self) + len);
126
127         if (self == NULL)
128                 goto out;
129
130         memcpy(self->filename, filename, len);
131         /*
132          * On 64bit we can mmap the data file in one go. No need for tiny mmap
133          * slices. On 32bit we use 32MB.
134          */
135 #if BITS_PER_LONG == 64
136         self->mmap_window = ULLONG_MAX;
137 #else
138         self->mmap_window = 32 * 1024 * 1024ULL;
139 #endif
140         self->machines = RB_ROOT;
141         self->repipe = repipe;
142         INIT_LIST_HEAD(&self->ordered_samples.samples);
143         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
144         INIT_LIST_HEAD(&self->ordered_samples.to_free);
145         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
146         hists__init(&self->hists);
147
148         if (mode == O_RDONLY) {
149                 if (perf_session__open(self, force) < 0)
150                         goto out_delete;
151                 perf_session__set_id_hdr_size(self);
152         } else if (mode == O_WRONLY) {
153                 /*
154                  * In O_RDONLY mode this will be performed when reading the
155                  * kernel MMAP event, in perf_event__process_mmap().
156                  */
157                 if (perf_session__create_kernel_maps(self) < 0)
158                         goto out_delete;
159         }
160
161         if (tool && tool->ordering_requires_timestamps &&
162             tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
163                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
164                 tool->ordered_samples = false;
165         }
166
167 out:
168         return self;
169 out_delete:
170         perf_session__delete(self);
171         return NULL;
172 }
173
174 static void machine__delete_dead_threads(struct machine *machine)
175 {
176         struct thread *n, *t;
177
178         list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
179                 list_del(&t->node);
180                 thread__delete(t);
181         }
182 }
183
184 static void perf_session__delete_dead_threads(struct perf_session *session)
185 {
186         machine__delete_dead_threads(&session->host_machine);
187 }
188
189 static void machine__delete_threads(struct machine *self)
190 {
191         struct rb_node *nd = rb_first(&self->threads);
192
193         while (nd) {
194                 struct thread *t = rb_entry(nd, struct thread, rb_node);
195
196                 rb_erase(&t->rb_node, &self->threads);
197                 nd = rb_next(nd);
198                 thread__delete(t);
199         }
200 }
201
202 static void perf_session__delete_threads(struct perf_session *session)
203 {
204         machine__delete_threads(&session->host_machine);
205 }
206
207 static void perf_session_env__delete(struct perf_session_env *env)
208 {
209         free(env->hostname);
210         free(env->os_release);
211         free(env->version);
212         free(env->arch);
213         free(env->cpu_desc);
214         free(env->cpuid);
215
216         free(env->cmdline);
217         free(env->sibling_cores);
218         free(env->sibling_threads);
219         free(env->numa_nodes);
220         free(env->pmu_mappings);
221 }
222
223 void perf_session__delete(struct perf_session *self)
224 {
225         perf_session__destroy_kernel_maps(self);
226         perf_session__delete_dead_threads(self);
227         perf_session__delete_threads(self);
228         perf_session_env__delete(&self->header.env);
229         machine__exit(&self->host_machine);
230         close(self->fd);
231         free(self);
232         vdso__exit();
233 }
234
235 void machine__remove_thread(struct machine *self, struct thread *th)
236 {
237         self->last_match = NULL;
238         rb_erase(&th->rb_node, &self->threads);
239         /*
240          * We may have references to this thread, for instance in some hist_entry
241          * instances, so just move them to a separate list.
242          */
243         list_add_tail(&th->node, &self->dead_threads);
244 }
245
246 static bool symbol__match_parent_regex(struct symbol *sym)
247 {
248         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
249                 return 1;
250
251         return 0;
252 }
253
254 static const u8 cpumodes[] = {
255         PERF_RECORD_MISC_USER,
256         PERF_RECORD_MISC_KERNEL,
257         PERF_RECORD_MISC_GUEST_USER,
258         PERF_RECORD_MISC_GUEST_KERNEL
259 };
260 #define NCPUMODES (sizeof(cpumodes)/sizeof(u8))
261
262 static void ip__resolve_ams(struct machine *self, struct thread *thread,
263                             struct addr_map_symbol *ams,
264                             u64 ip)
265 {
266         struct addr_location al;
267         size_t i;
268         u8 m;
269
270         memset(&al, 0, sizeof(al));
271
272         for (i = 0; i < NCPUMODES; i++) {
273                 m = cpumodes[i];
274                 /*
275                  * We cannot use the header.misc hint to determine whether a
276                  * branch stack address is user, kernel, guest, hypervisor.
277                  * Branches may straddle the kernel/user/hypervisor boundaries.
278                  * Thus, we have to try consecutively until we find a match
279                  * or else, the symbol is unknown
280                  */
281                 thread__find_addr_location(thread, self, m, MAP__FUNCTION,
282                                 ip, &al, NULL);
283                 if (al.sym)
284                         goto found;
285         }
286 found:
287         ams->addr = ip;
288         ams->al_addr = al.addr;
289         ams->sym = al.sym;
290         ams->map = al.map;
291 }
292
293 struct branch_info *machine__resolve_bstack(struct machine *self,
294                                             struct thread *thr,
295                                             struct branch_stack *bs)
296 {
297         struct branch_info *bi;
298         unsigned int i;
299
300         bi = calloc(bs->nr, sizeof(struct branch_info));
301         if (!bi)
302                 return NULL;
303
304         for (i = 0; i < bs->nr; i++) {
305                 ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to);
306                 ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from);
307                 bi[i].flags = bs->entries[i].flags;
308         }
309         return bi;
310 }
311
312 static int machine__resolve_callchain_sample(struct machine *machine,
313                                              struct thread *thread,
314                                              struct ip_callchain *chain,
315                                              struct symbol **parent)
316
317 {
318         u8 cpumode = PERF_RECORD_MISC_USER;
319         unsigned int i;
320         int err;
321
322         callchain_cursor_reset(&callchain_cursor);
323
324         if (chain->nr > PERF_MAX_STACK_DEPTH) {
325                 pr_warning("corrupted callchain. skipping...\n");
326                 return 0;
327         }
328
329         for (i = 0; i < chain->nr; i++) {
330                 u64 ip;
331                 struct addr_location al;
332
333                 if (callchain_param.order == ORDER_CALLEE)
334                         ip = chain->ips[i];
335                 else
336                         ip = chain->ips[chain->nr - i - 1];
337
338                 if (ip >= PERF_CONTEXT_MAX) {
339                         switch (ip) {
340                         case PERF_CONTEXT_HV:
341                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;
342                                 break;
343                         case PERF_CONTEXT_KERNEL:
344                                 cpumode = PERF_RECORD_MISC_KERNEL;
345                                 break;
346                         case PERF_CONTEXT_USER:
347                                 cpumode = PERF_RECORD_MISC_USER;
348                                 break;
349                         default:
350                                 pr_debug("invalid callchain context: "
351                                          "%"PRId64"\n", (s64) ip);
352                                 /*
353                                  * It seems the callchain is corrupted.
354                                  * Discard all.
355                                  */
356                                 callchain_cursor_reset(&callchain_cursor);
357                                 return 0;
358                         }
359                         continue;
360                 }
361
362                 al.filtered = false;
363                 thread__find_addr_location(thread, machine, cpumode,
364                                            MAP__FUNCTION, ip, &al, NULL);
365                 if (al.sym != NULL) {
366                         if (sort__has_parent && !*parent &&
367                             symbol__match_parent_regex(al.sym))
368                                 *parent = al.sym;
369                         if (!symbol_conf.use_callchain)
370                                 break;
371                 }
372
373                 err = callchain_cursor_append(&callchain_cursor,
374                                               ip, al.map, al.sym);
375                 if (err)
376                         return err;
377         }
378
379         return 0;
380 }
381
382 static int unwind_entry(struct unwind_entry *entry, void *arg)
383 {
384         struct callchain_cursor *cursor = arg;
385         return callchain_cursor_append(cursor, entry->ip,
386                                        entry->map, entry->sym);
387 }
388
389 int machine__resolve_callchain(struct machine *machine,
390                                struct perf_evsel *evsel,
391                                struct thread *thread,
392                                struct perf_sample *sample,
393                                struct symbol **parent)
394
395 {
396         int ret;
397
398         callchain_cursor_reset(&callchain_cursor);
399
400         ret = machine__resolve_callchain_sample(machine, thread,
401                                                 sample->callchain, parent);
402         if (ret)
403                 return ret;
404
405         /* Can we do dwarf post unwind? */
406         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
407               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
408                 return 0;
409
410         /* Bail out if nothing was captured. */
411         if ((!sample->user_regs.regs) ||
412             (!sample->user_stack.size))
413                 return 0;
414
415         return unwind__get_entries(unwind_entry, &callchain_cursor, machine,
416                                    thread, evsel->attr.sample_regs_user,
417                                    sample);
418
419 }
420
421 static int process_event_synth_tracing_data_stub(union perf_event *event
422                                                  __maybe_unused,
423                                                  struct perf_session *session
424                                                 __maybe_unused)
425 {
426         dump_printf(": unhandled!\n");
427         return 0;
428 }
429
430 static int process_event_synth_attr_stub(union perf_event *event __maybe_unused,
431                                          struct perf_evlist **pevlist
432                                          __maybe_unused)
433 {
434         dump_printf(": unhandled!\n");
435         return 0;
436 }
437
438 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
439                                      union perf_event *event __maybe_unused,
440                                      struct perf_sample *sample __maybe_unused,
441                                      struct perf_evsel *evsel __maybe_unused,
442                                      struct machine *machine __maybe_unused)
443 {
444         dump_printf(": unhandled!\n");
445         return 0;
446 }
447
448 static int process_event_stub(struct perf_tool *tool __maybe_unused,
449                               union perf_event *event __maybe_unused,
450                               struct perf_sample *sample __maybe_unused,
451                               struct machine *machine __maybe_unused)
452 {
453         dump_printf(": unhandled!\n");
454         return 0;
455 }
456
457 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
458                                        union perf_event *event __maybe_unused,
459                                        struct perf_session *perf_session
460                                        __maybe_unused)
461 {
462         dump_printf(": unhandled!\n");
463         return 0;
464 }
465
466 static int process_event_type_stub(struct perf_tool *tool __maybe_unused,
467                                    union perf_event *event __maybe_unused)
468 {
469         dump_printf(": unhandled!\n");
470         return 0;
471 }
472
473 static int process_finished_round(struct perf_tool *tool,
474                                   union perf_event *event,
475                                   struct perf_session *session);
476
477 static void perf_tool__fill_defaults(struct perf_tool *tool)
478 {
479         if (tool->sample == NULL)
480                 tool->sample = process_event_sample_stub;
481         if (tool->mmap == NULL)
482                 tool->mmap = process_event_stub;
483         if (tool->comm == NULL)
484                 tool->comm = process_event_stub;
485         if (tool->fork == NULL)
486                 tool->fork = process_event_stub;
487         if (tool->exit == NULL)
488                 tool->exit = process_event_stub;
489         if (tool->lost == NULL)
490                 tool->lost = perf_event__process_lost;
491         if (tool->read == NULL)
492                 tool->read = process_event_sample_stub;
493         if (tool->throttle == NULL)
494                 tool->throttle = process_event_stub;
495         if (tool->unthrottle == NULL)
496                 tool->unthrottle = process_event_stub;
497         if (tool->attr == NULL)
498                 tool->attr = process_event_synth_attr_stub;
499         if (tool->event_type == NULL)
500                 tool->event_type = process_event_type_stub;
501         if (tool->tracing_data == NULL)
502                 tool->tracing_data = process_event_synth_tracing_data_stub;
503         if (tool->build_id == NULL)
504                 tool->build_id = process_finished_round_stub;
505         if (tool->finished_round == NULL) {
506                 if (tool->ordered_samples)
507                         tool->finished_round = process_finished_round;
508                 else
509                         tool->finished_round = process_finished_round_stub;
510         }
511 }
512  
513 void mem_bswap_32(void *src, int byte_size)
514 {
515         u32 *m = src;
516         while (byte_size > 0) {
517                 *m = bswap_32(*m);
518                 byte_size -= sizeof(u32);
519                 ++m;
520         }
521 }
522
523 void mem_bswap_64(void *src, int byte_size)
524 {
525         u64 *m = src;
526
527         while (byte_size > 0) {
528                 *m = bswap_64(*m);
529                 byte_size -= sizeof(u64);
530                 ++m;
531         }
532 }
533
534 static void swap_sample_id_all(union perf_event *event, void *data)
535 {
536         void *end = (void *) event + event->header.size;
537         int size = end - data;
538
539         BUG_ON(size % sizeof(u64));
540         mem_bswap_64(data, size);
541 }
542
543 static void perf_event__all64_swap(union perf_event *event,
544                                    bool sample_id_all __maybe_unused)
545 {
546         struct perf_event_header *hdr = &event->header;
547         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
548 }
549
550 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
551 {
552         event->comm.pid = bswap_32(event->comm.pid);
553         event->comm.tid = bswap_32(event->comm.tid);
554
555         if (sample_id_all) {
556                 void *data = &event->comm.comm;
557
558                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
559                 swap_sample_id_all(event, data);
560         }
561 }
562
563 static void perf_event__mmap_swap(union perf_event *event,
564                                   bool sample_id_all)
565 {
566         event->mmap.pid   = bswap_32(event->mmap.pid);
567         event->mmap.tid   = bswap_32(event->mmap.tid);
568         event->mmap.start = bswap_64(event->mmap.start);
569         event->mmap.len   = bswap_64(event->mmap.len);
570         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
571
572         if (sample_id_all) {
573                 void *data = &event->mmap.filename;
574
575                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
576                 swap_sample_id_all(event, data);
577         }
578 }
579
580 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
581 {
582         event->fork.pid  = bswap_32(event->fork.pid);
583         event->fork.tid  = bswap_32(event->fork.tid);
584         event->fork.ppid = bswap_32(event->fork.ppid);
585         event->fork.ptid = bswap_32(event->fork.ptid);
586         event->fork.time = bswap_64(event->fork.time);
587
588         if (sample_id_all)
589                 swap_sample_id_all(event, &event->fork + 1);
590 }
591
592 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
593 {
594         event->read.pid          = bswap_32(event->read.pid);
595         event->read.tid          = bswap_32(event->read.tid);
596         event->read.value        = bswap_64(event->read.value);
597         event->read.time_enabled = bswap_64(event->read.time_enabled);
598         event->read.time_running = bswap_64(event->read.time_running);
599         event->read.id           = bswap_64(event->read.id);
600
601         if (sample_id_all)
602                 swap_sample_id_all(event, &event->read + 1);
603 }
604
605 static u8 revbyte(u8 b)
606 {
607         int rev = (b >> 4) | ((b & 0xf) << 4);
608         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
609         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
610         return (u8) rev;
611 }
612
613 /*
614  * XXX this is hack in attempt to carry flags bitfield
615  * throught endian village. ABI says:
616  *
617  * Bit-fields are allocated from right to left (least to most significant)
618  * on little-endian implementations and from left to right (most to least
619  * significant) on big-endian implementations.
620  *
621  * The above seems to be byte specific, so we need to reverse each
622  * byte of the bitfield. 'Internet' also says this might be implementation
623  * specific and we probably need proper fix and carry perf_event_attr
624  * bitfield flags in separate data file FEAT_ section. Thought this seems
625  * to work for now.
626  */
627 static void swap_bitfield(u8 *p, unsigned len)
628 {
629         unsigned i;
630
631         for (i = 0; i < len; i++) {
632                 *p = revbyte(*p);
633                 p++;
634         }
635 }
636
637 /* exported for swapping attributes in file header */
638 void perf_event__attr_swap(struct perf_event_attr *attr)
639 {
640         attr->type              = bswap_32(attr->type);
641         attr->size              = bswap_32(attr->size);
642         attr->config            = bswap_64(attr->config);
643         attr->sample_period     = bswap_64(attr->sample_period);
644         attr->sample_type       = bswap_64(attr->sample_type);
645         attr->read_format       = bswap_64(attr->read_format);
646         attr->wakeup_events     = bswap_32(attr->wakeup_events);
647         attr->bp_type           = bswap_32(attr->bp_type);
648         attr->bp_addr           = bswap_64(attr->bp_addr);
649         attr->bp_len            = bswap_64(attr->bp_len);
650
651         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
652 }
653
654 static void perf_event__hdr_attr_swap(union perf_event *event,
655                                       bool sample_id_all __maybe_unused)
656 {
657         size_t size;
658
659         perf_event__attr_swap(&event->attr.attr);
660
661         size = event->header.size;
662         size -= (void *)&event->attr.id - (void *)event;
663         mem_bswap_64(event->attr.id, size);
664 }
665
666 static void perf_event__event_type_swap(union perf_event *event,
667                                         bool sample_id_all __maybe_unused)
668 {
669         event->event_type.event_type.event_id =
670                 bswap_64(event->event_type.event_type.event_id);
671 }
672
673 static void perf_event__tracing_data_swap(union perf_event *event,
674                                           bool sample_id_all __maybe_unused)
675 {
676         event->tracing_data.size = bswap_32(event->tracing_data.size);
677 }
678
679 typedef void (*perf_event__swap_op)(union perf_event *event,
680                                     bool sample_id_all);
681
682 static perf_event__swap_op perf_event__swap_ops[] = {
683         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
684         [PERF_RECORD_COMM]                = perf_event__comm_swap,
685         [PERF_RECORD_FORK]                = perf_event__task_swap,
686         [PERF_RECORD_EXIT]                = perf_event__task_swap,
687         [PERF_RECORD_LOST]                = perf_event__all64_swap,
688         [PERF_RECORD_READ]                = perf_event__read_swap,
689         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
690         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
691         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
692         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
693         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
694         [PERF_RECORD_HEADER_MAX]          = NULL,
695 };
696
697 struct sample_queue {
698         u64                     timestamp;
699         u64                     file_offset;
700         union perf_event        *event;
701         struct list_head        list;
702 };
703
704 static void perf_session_free_sample_buffers(struct perf_session *session)
705 {
706         struct ordered_samples *os = &session->ordered_samples;
707
708         while (!list_empty(&os->to_free)) {
709                 struct sample_queue *sq;
710
711                 sq = list_entry(os->to_free.next, struct sample_queue, list);
712                 list_del(&sq->list);
713                 free(sq);
714         }
715 }
716
717 static int perf_session_deliver_event(struct perf_session *session,
718                                       union perf_event *event,
719                                       struct perf_sample *sample,
720                                       struct perf_tool *tool,
721                                       u64 file_offset);
722
723 static int flush_sample_queue(struct perf_session *s,
724                                struct perf_tool *tool)
725 {
726         struct ordered_samples *os = &s->ordered_samples;
727         struct list_head *head = &os->samples;
728         struct sample_queue *tmp, *iter;
729         struct perf_sample sample;
730         u64 limit = os->next_flush;
731         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
732         unsigned idx = 0, progress_next = os->nr_samples / 16;
733         int ret;
734
735         if (!tool->ordered_samples || !limit)
736                 return 0;
737
738         list_for_each_entry_safe(iter, tmp, head, list) {
739                 if (iter->timestamp > limit)
740                         break;
741
742                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
743                 if (ret)
744                         pr_err("Can't parse sample, err = %d\n", ret);
745                 else {
746                         ret = perf_session_deliver_event(s, iter->event, &sample, tool,
747                                                          iter->file_offset);
748                         if (ret)
749                                 return ret;
750                 }
751
752                 os->last_flush = iter->timestamp;
753                 list_del(&iter->list);
754                 list_add(&iter->list, &os->sample_cache);
755                 if (++idx >= progress_next) {
756                         progress_next += os->nr_samples / 16;
757                         ui_progress__update(idx, os->nr_samples,
758                                             "Processing time ordered events...");
759                 }
760         }
761
762         if (list_empty(head)) {
763                 os->last_sample = NULL;
764         } else if (last_ts <= limit) {
765                 os->last_sample =
766                         list_entry(head->prev, struct sample_queue, list);
767         }
768
769         os->nr_samples = 0;
770
771         return 0;
772 }
773
774 /*
775  * When perf record finishes a pass on every buffers, it records this pseudo
776  * event.
777  * We record the max timestamp t found in the pass n.
778  * Assuming these timestamps are monotonic across cpus, we know that if
779  * a buffer still has events with timestamps below t, they will be all
780  * available and then read in the pass n + 1.
781  * Hence when we start to read the pass n + 2, we can safely flush every
782  * events with timestamps below t.
783  *
784  *    ============ PASS n =================
785  *       CPU 0         |   CPU 1
786  *                     |
787  *    cnt1 timestamps  |   cnt2 timestamps
788  *          1          |         2
789  *          2          |         3
790  *          -          |         4  <--- max recorded
791  *
792  *    ============ PASS n + 1 ==============
793  *       CPU 0         |   CPU 1
794  *                     |
795  *    cnt1 timestamps  |   cnt2 timestamps
796  *          3          |         5
797  *          4          |         6
798  *          5          |         7 <---- max recorded
799  *
800  *      Flush every events below timestamp 4
801  *
802  *    ============ PASS n + 2 ==============
803  *       CPU 0         |   CPU 1
804  *                     |
805  *    cnt1 timestamps  |   cnt2 timestamps
806  *          6          |         8
807  *          7          |         9
808  *          -          |         10
809  *
810  *      Flush every events below timestamp 7
811  *      etc...
812  */
813 static int process_finished_round(struct perf_tool *tool,
814                                   union perf_event *event __maybe_unused,
815                                   struct perf_session *session)
816 {
817         int ret = flush_sample_queue(session, tool);
818         if (!ret)
819                 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
820
821         return ret;
822 }
823
824 /* The queue is ordered by time */
825 static void __queue_event(struct sample_queue *new, struct perf_session *s)
826 {
827         struct ordered_samples *os = &s->ordered_samples;
828         struct sample_queue *sample = os->last_sample;
829         u64 timestamp = new->timestamp;
830         struct list_head *p;
831
832         ++os->nr_samples;
833         os->last_sample = new;
834
835         if (!sample) {
836                 list_add(&new->list, &os->samples);
837                 os->max_timestamp = timestamp;
838                 return;
839         }
840
841         /*
842          * last_sample might point to some random place in the list as it's
843          * the last queued event. We expect that the new event is close to
844          * this.
845          */
846         if (sample->timestamp <= timestamp) {
847                 while (sample->timestamp <= timestamp) {
848                         p = sample->list.next;
849                         if (p == &os->samples) {
850                                 list_add_tail(&new->list, &os->samples);
851                                 os->max_timestamp = timestamp;
852                                 return;
853                         }
854                         sample = list_entry(p, struct sample_queue, list);
855                 }
856                 list_add_tail(&new->list, &sample->list);
857         } else {
858                 while (sample->timestamp > timestamp) {
859                         p = sample->list.prev;
860                         if (p == &os->samples) {
861                                 list_add(&new->list, &os->samples);
862                                 return;
863                         }
864                         sample = list_entry(p, struct sample_queue, list);
865                 }
866                 list_add(&new->list, &sample->list);
867         }
868 }
869
870 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
871
872 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
873                                     struct perf_sample *sample, u64 file_offset)
874 {
875         struct ordered_samples *os = &s->ordered_samples;
876         struct list_head *sc = &os->sample_cache;
877         u64 timestamp = sample->time;
878         struct sample_queue *new;
879
880         if (!timestamp || timestamp == ~0ULL)
881                 return -ETIME;
882
883         if (timestamp < s->ordered_samples.last_flush) {
884                 printf("Warning: Timestamp below last timeslice flush\n");
885                 return -EINVAL;
886         }
887
888         if (!list_empty(sc)) {
889                 new = list_entry(sc->next, struct sample_queue, list);
890                 list_del(&new->list);
891         } else if (os->sample_buffer) {
892                 new = os->sample_buffer + os->sample_buffer_idx;
893                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
894                         os->sample_buffer = NULL;
895         } else {
896                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
897                 if (!os->sample_buffer)
898                         return -ENOMEM;
899                 list_add(&os->sample_buffer->list, &os->to_free);
900                 os->sample_buffer_idx = 2;
901                 new = os->sample_buffer + 1;
902         }
903
904         new->timestamp = timestamp;
905         new->file_offset = file_offset;
906         new->event = event;
907
908         __queue_event(new, s);
909
910         return 0;
911 }
912
913 static void callchain__printf(struct perf_sample *sample)
914 {
915         unsigned int i;
916
917         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
918
919         for (i = 0; i < sample->callchain->nr; i++)
920                 printf("..... %2d: %016" PRIx64 "\n",
921                        i, sample->callchain->ips[i]);
922 }
923
924 static void branch_stack__printf(struct perf_sample *sample)
925 {
926         uint64_t i;
927
928         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
929
930         for (i = 0; i < sample->branch_stack->nr; i++)
931                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
932                         i, sample->branch_stack->entries[i].from,
933                         sample->branch_stack->entries[i].to);
934 }
935
936 static void regs_dump__printf(u64 mask, u64 *regs)
937 {
938         unsigned rid, i = 0;
939
940         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
941                 u64 val = regs[i++];
942
943                 printf(".... %-5s 0x%" PRIx64 "\n",
944                        perf_reg_name(rid), val);
945         }
946 }
947
948 static void regs_user__printf(struct perf_sample *sample, u64 mask)
949 {
950         struct regs_dump *user_regs = &sample->user_regs;
951
952         if (user_regs->regs) {
953                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
954                 regs_dump__printf(mask, user_regs->regs);
955         }
956 }
957
958 static void stack_user__printf(struct stack_dump *dump)
959 {
960         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
961                dump->size, dump->offset);
962 }
963
964 static void perf_session__print_tstamp(struct perf_session *session,
965                                        union perf_event *event,
966                                        struct perf_sample *sample)
967 {
968         u64 sample_type = perf_evlist__sample_type(session->evlist);
969
970         if (event->header.type != PERF_RECORD_SAMPLE &&
971             !perf_evlist__sample_id_all(session->evlist)) {
972                 fputs("-1 -1 ", stdout);
973                 return;
974         }
975
976         if ((sample_type & PERF_SAMPLE_CPU))
977                 printf("%u ", sample->cpu);
978
979         if (sample_type & PERF_SAMPLE_TIME)
980                 printf("%" PRIu64 " ", sample->time);
981 }
982
983 static void dump_event(struct perf_session *session, union perf_event *event,
984                        u64 file_offset, struct perf_sample *sample)
985 {
986         if (!dump_trace)
987                 return;
988
989         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
990                file_offset, event->header.size, event->header.type);
991
992         trace_event(event);
993
994         if (sample)
995                 perf_session__print_tstamp(session, event, sample);
996
997         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
998                event->header.size, perf_event__name(event->header.type));
999 }
1000
1001 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
1002                         struct perf_sample *sample)
1003 {
1004         u64 sample_type;
1005
1006         if (!dump_trace)
1007                 return;
1008
1009         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
1010                event->header.misc, sample->pid, sample->tid, sample->ip,
1011                sample->period, sample->addr);
1012
1013         sample_type = evsel->attr.sample_type;
1014
1015         if (sample_type & PERF_SAMPLE_CALLCHAIN)
1016                 callchain__printf(sample);
1017
1018         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
1019                 branch_stack__printf(sample);
1020
1021         if (sample_type & PERF_SAMPLE_REGS_USER)
1022                 regs_user__printf(sample, evsel->attr.sample_regs_user);
1023
1024         if (sample_type & PERF_SAMPLE_STACK_USER)
1025                 stack_user__printf(&sample->user_stack);
1026 }
1027
1028 static struct machine *
1029         perf_session__find_machine_for_cpumode(struct perf_session *session,
1030                                                union perf_event *event)
1031 {
1032         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1033
1034         if (perf_guest &&
1035             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
1036              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
1037                 u32 pid;
1038
1039                 if (event->header.type == PERF_RECORD_MMAP)
1040                         pid = event->mmap.pid;
1041                 else
1042                         pid = event->ip.pid;
1043
1044                 return perf_session__findnew_machine(session, pid);
1045         }
1046
1047         return perf_session__find_host_machine(session);
1048 }
1049
1050 static int perf_session_deliver_event(struct perf_session *session,
1051                                       union perf_event *event,
1052                                       struct perf_sample *sample,
1053                                       struct perf_tool *tool,
1054                                       u64 file_offset)
1055 {
1056         struct perf_evsel *evsel;
1057         struct machine *machine;
1058
1059         dump_event(session, event, file_offset, sample);
1060
1061         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
1062         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
1063                 /*
1064                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
1065                  * because the tools right now may apply filters, discarding
1066                  * some of the samples. For consistency, in the future we
1067                  * should have something like nr_filtered_samples and remove
1068                  * the sample->period from total_sample_period, etc, KISS for
1069                  * now tho.
1070                  *
1071                  * Also testing against NULL allows us to handle files without
1072                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
1073                  * future probably it'll be a good idea to restrict event
1074                  * processing via perf_session to files with both set.
1075                  */
1076                 hists__inc_nr_events(&evsel->hists, event->header.type);
1077         }
1078
1079         machine = perf_session__find_machine_for_cpumode(session, event);
1080
1081         switch (event->header.type) {
1082         case PERF_RECORD_SAMPLE:
1083                 dump_sample(evsel, event, sample);
1084                 if (evsel == NULL) {
1085                         ++session->hists.stats.nr_unknown_id;
1086                         return 0;
1087                 }
1088                 if (machine == NULL) {
1089                         ++session->hists.stats.nr_unprocessable_samples;
1090                         return 0;
1091                 }
1092                 return tool->sample(tool, event, sample, evsel, machine);
1093         case PERF_RECORD_MMAP:
1094                 return tool->mmap(tool, event, sample, machine);
1095         case PERF_RECORD_COMM:
1096                 return tool->comm(tool, event, sample, machine);
1097         case PERF_RECORD_FORK:
1098                 return tool->fork(tool, event, sample, machine);
1099         case PERF_RECORD_EXIT:
1100                 return tool->exit(tool, event, sample, machine);
1101         case PERF_RECORD_LOST:
1102                 if (tool->lost == perf_event__process_lost)
1103                         session->hists.stats.total_lost += event->lost.lost;
1104                 return tool->lost(tool, event, sample, machine);
1105         case PERF_RECORD_READ:
1106                 return tool->read(tool, event, sample, evsel, machine);
1107         case PERF_RECORD_THROTTLE:
1108                 return tool->throttle(tool, event, sample, machine);
1109         case PERF_RECORD_UNTHROTTLE:
1110                 return tool->unthrottle(tool, event, sample, machine);
1111         default:
1112                 ++session->hists.stats.nr_unknown_events;
1113                 return -1;
1114         }
1115 }
1116
1117 static int perf_session__preprocess_sample(struct perf_session *session,
1118                                            union perf_event *event, struct perf_sample *sample)
1119 {
1120         if (event->header.type != PERF_RECORD_SAMPLE ||
1121             !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN))
1122                 return 0;
1123
1124         if (!ip_callchain__valid(sample->callchain, event)) {
1125                 pr_debug("call-chain problem with event, skipping it.\n");
1126                 ++session->hists.stats.nr_invalid_chains;
1127                 session->hists.stats.total_invalid_chains += sample->period;
1128                 return -EINVAL;
1129         }
1130         return 0;
1131 }
1132
1133 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1134                                             struct perf_tool *tool, u64 file_offset)
1135 {
1136         int err;
1137
1138         dump_event(session, event, file_offset, NULL);
1139
1140         /* These events are processed right away */
1141         switch (event->header.type) {
1142         case PERF_RECORD_HEADER_ATTR:
1143                 err = tool->attr(event, &session->evlist);
1144                 if (err == 0)
1145                         perf_session__set_id_hdr_size(session);
1146                 return err;
1147         case PERF_RECORD_HEADER_EVENT_TYPE:
1148                 return tool->event_type(tool, event);
1149         case PERF_RECORD_HEADER_TRACING_DATA:
1150                 /* setup for reading amidst mmap */
1151                 lseek(session->fd, file_offset, SEEK_SET);
1152                 return tool->tracing_data(event, session);
1153         case PERF_RECORD_HEADER_BUILD_ID:
1154                 return tool->build_id(tool, event, session);
1155         case PERF_RECORD_FINISHED_ROUND:
1156                 return tool->finished_round(tool, event, session);
1157         default:
1158                 return -EINVAL;
1159         }
1160 }
1161
1162 static void event_swap(union perf_event *event, bool sample_id_all)
1163 {
1164         perf_event__swap_op swap;
1165
1166         swap = perf_event__swap_ops[event->header.type];
1167         if (swap)
1168                 swap(event, sample_id_all);
1169 }
1170
1171 static int perf_session__process_event(struct perf_session *session,
1172                                        union perf_event *event,
1173                                        struct perf_tool *tool,
1174                                        u64 file_offset)
1175 {
1176         struct perf_sample sample;
1177         int ret;
1178
1179         if (session->header.needs_swap)
1180                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1181
1182         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1183                 return -EINVAL;
1184
1185         hists__inc_nr_events(&session->hists, event->header.type);
1186
1187         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1188                 return perf_session__process_user_event(session, event, tool, file_offset);
1189
1190         /*
1191          * For all kernel events we get the sample data
1192          */
1193         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1194         if (ret)
1195                 return ret;
1196
1197         /* Preprocess sample records - precheck callchains */
1198         if (perf_session__preprocess_sample(session, event, &sample))
1199                 return 0;
1200
1201         if (tool->ordered_samples) {
1202                 ret = perf_session_queue_event(session, event, &sample,
1203                                                file_offset);
1204                 if (ret != -ETIME)
1205                         return ret;
1206         }
1207
1208         return perf_session_deliver_event(session, event, &sample, tool,
1209                                           file_offset);
1210 }
1211
1212 void perf_event_header__bswap(struct perf_event_header *self)
1213 {
1214         self->type = bswap_32(self->type);
1215         self->misc = bswap_16(self->misc);
1216         self->size = bswap_16(self->size);
1217 }
1218
1219 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1220 {
1221         return machine__findnew_thread(&session->host_machine, pid);
1222 }
1223
1224 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1225 {
1226         struct thread *thread = perf_session__findnew(self, 0);
1227
1228         if (thread == NULL || thread__set_comm(thread, "swapper")) {
1229                 pr_err("problem inserting idle task.\n");
1230                 thread = NULL;
1231         }
1232
1233         return thread;
1234 }
1235
1236 static void perf_session__warn_about_errors(const struct perf_session *session,
1237                                             const struct perf_tool *tool)
1238 {
1239         if (tool->lost == perf_event__process_lost &&
1240             session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1241                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1242                             "Check IO/CPU overload!\n\n",
1243                             session->hists.stats.nr_events[0],
1244                             session->hists.stats.nr_events[PERF_RECORD_LOST]);
1245         }
1246
1247         if (session->hists.stats.nr_unknown_events != 0) {
1248                 ui__warning("Found %u unknown events!\n\n"
1249                             "Is this an older tool processing a perf.data "
1250                             "file generated by a more recent tool?\n\n"
1251                             "If that is not the case, consider "
1252                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1253                             session->hists.stats.nr_unknown_events);
1254         }
1255
1256         if (session->hists.stats.nr_unknown_id != 0) {
1257                 ui__warning("%u samples with id not present in the header\n",
1258                             session->hists.stats.nr_unknown_id);
1259         }
1260
1261         if (session->hists.stats.nr_invalid_chains != 0) {
1262                 ui__warning("Found invalid callchains!\n\n"
1263                             "%u out of %u events were discarded for this reason.\n\n"
1264                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1265                             session->hists.stats.nr_invalid_chains,
1266                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1267         }
1268
1269         if (session->hists.stats.nr_unprocessable_samples != 0) {
1270                 ui__warning("%u unprocessable samples recorded.\n"
1271                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1272                             session->hists.stats.nr_unprocessable_samples);
1273         }
1274 }
1275
1276 #define session_done()  (*(volatile int *)(&session_done))
1277 volatile int session_done;
1278
1279 static int __perf_session__process_pipe_events(struct perf_session *self,
1280                                                struct perf_tool *tool)
1281 {
1282         union perf_event *event;
1283         uint32_t size, cur_size = 0;
1284         void *buf = NULL;
1285         int skip = 0;
1286         u64 head;
1287         int err;
1288         void *p;
1289
1290         perf_tool__fill_defaults(tool);
1291
1292         head = 0;
1293         cur_size = sizeof(union perf_event);
1294
1295         buf = malloc(cur_size);
1296         if (!buf)
1297                 return -errno;
1298 more:
1299         event = buf;
1300         err = readn(self->fd, event, sizeof(struct perf_event_header));
1301         if (err <= 0) {
1302                 if (err == 0)
1303                         goto done;
1304
1305                 pr_err("failed to read event header\n");
1306                 goto out_err;
1307         }
1308
1309         if (self->header.needs_swap)
1310                 perf_event_header__bswap(&event->header);
1311
1312         size = event->header.size;
1313         if (size == 0)
1314                 size = 8;
1315
1316         if (size > cur_size) {
1317                 void *new = realloc(buf, size);
1318                 if (!new) {
1319                         pr_err("failed to allocate memory to read event\n");
1320                         goto out_err;
1321                 }
1322                 buf = new;
1323                 cur_size = size;
1324                 event = buf;
1325         }
1326         p = event;
1327         p += sizeof(struct perf_event_header);
1328
1329         if (size - sizeof(struct perf_event_header)) {
1330                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1331                 if (err <= 0) {
1332                         if (err == 0) {
1333                                 pr_err("unexpected end of event stream\n");
1334                                 goto done;
1335                         }
1336
1337                         pr_err("failed to read event data\n");
1338                         goto out_err;
1339                 }
1340         }
1341
1342         if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1343                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1344                        head, event->header.size, event->header.type);
1345                 err = -EINVAL;
1346                 goto out_err;
1347         }
1348
1349         head += size;
1350
1351         if (skip > 0)
1352                 head += skip;
1353
1354         if (!session_done())
1355                 goto more;
1356 done:
1357         err = 0;
1358 out_err:
1359         free(buf);
1360         perf_session__warn_about_errors(self, tool);
1361         perf_session_free_sample_buffers(self);
1362         return err;
1363 }
1364
1365 static union perf_event *
1366 fetch_mmaped_event(struct perf_session *session,
1367                    u64 head, size_t mmap_size, char *buf)
1368 {
1369         union perf_event *event;
1370
1371         /*
1372          * Ensure we have enough space remaining to read
1373          * the size of the event in the headers.
1374          */
1375         if (head + sizeof(event->header) > mmap_size)
1376                 return NULL;
1377
1378         event = (union perf_event *)(buf + head);
1379
1380         if (session->header.needs_swap)
1381                 perf_event_header__bswap(&event->header);
1382
1383         if (head + event->header.size > mmap_size)
1384                 return NULL;
1385
1386         return event;
1387 }
1388
1389 int __perf_session__process_events(struct perf_session *session,
1390                                    u64 data_offset, u64 data_size,
1391                                    u64 file_size, struct perf_tool *tool)
1392 {
1393         u64 head, page_offset, file_offset, file_pos, progress_next;
1394         int err, mmap_prot, mmap_flags, map_idx = 0;
1395         size_t  mmap_size;
1396         char *buf, *mmaps[8];
1397         union perf_event *event;
1398         uint32_t size;
1399
1400         perf_tool__fill_defaults(tool);
1401
1402         page_offset = page_size * (data_offset / page_size);
1403         file_offset = page_offset;
1404         head = data_offset - page_offset;
1405
1406         if (data_offset + data_size < file_size)
1407                 file_size = data_offset + data_size;
1408
1409         progress_next = file_size / 16;
1410
1411         mmap_size = session->mmap_window;
1412         if (mmap_size > file_size)
1413                 mmap_size = file_size;
1414
1415         memset(mmaps, 0, sizeof(mmaps));
1416
1417         mmap_prot  = PROT_READ;
1418         mmap_flags = MAP_SHARED;
1419
1420         if (session->header.needs_swap) {
1421                 mmap_prot  |= PROT_WRITE;
1422                 mmap_flags = MAP_PRIVATE;
1423         }
1424 remap:
1425         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1426                    file_offset);
1427         if (buf == MAP_FAILED) {
1428                 pr_err("failed to mmap file\n");
1429                 err = -errno;
1430                 goto out_err;
1431         }
1432         mmaps[map_idx] = buf;
1433         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1434         file_pos = file_offset + head;
1435
1436 more:
1437         event = fetch_mmaped_event(session, head, mmap_size, buf);
1438         if (!event) {
1439                 if (mmaps[map_idx]) {
1440                         munmap(mmaps[map_idx], mmap_size);
1441                         mmaps[map_idx] = NULL;
1442                 }
1443
1444                 page_offset = page_size * (head / page_size);
1445                 file_offset += page_offset;
1446                 head -= page_offset;
1447                 goto remap;
1448         }
1449
1450         size = event->header.size;
1451
1452         if (size == 0 ||
1453             perf_session__process_event(session, event, tool, file_pos) < 0) {
1454                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1455                        file_offset + head, event->header.size,
1456                        event->header.type);
1457                 err = -EINVAL;
1458                 goto out_err;
1459         }
1460
1461         head += size;
1462         file_pos += size;
1463
1464         if (file_pos >= progress_next) {
1465                 progress_next += file_size / 16;
1466                 ui_progress__update(file_pos, file_size,
1467                                     "Processing events...");
1468         }
1469
1470         if (file_pos < file_size)
1471                 goto more;
1472
1473         err = 0;
1474         /* do the final flush for ordered samples */
1475         session->ordered_samples.next_flush = ULLONG_MAX;
1476         err = flush_sample_queue(session, tool);
1477 out_err:
1478         ui_progress__finish();
1479         perf_session__warn_about_errors(session, tool);
1480         perf_session_free_sample_buffers(session);
1481         return err;
1482 }
1483
1484 int perf_session__process_events(struct perf_session *self,
1485                                  struct perf_tool *tool)
1486 {
1487         int err;
1488
1489         if (perf_session__register_idle_thread(self) == NULL)
1490                 return -ENOMEM;
1491
1492         if (!self->fd_pipe)
1493                 err = __perf_session__process_events(self,
1494                                                      self->header.data_offset,
1495                                                      self->header.data_size,
1496                                                      self->size, tool);
1497         else
1498                 err = __perf_session__process_pipe_events(self, tool);
1499
1500         return err;
1501 }
1502
1503 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1504 {
1505         if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) {
1506                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1507                 return false;
1508         }
1509
1510         return true;
1511 }
1512
1513 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1514                                      const char *symbol_name, u64 addr)
1515 {
1516         char *bracket;
1517         enum map_type i;
1518         struct ref_reloc_sym *ref;
1519
1520         ref = zalloc(sizeof(struct ref_reloc_sym));
1521         if (ref == NULL)
1522                 return -ENOMEM;
1523
1524         ref->name = strdup(symbol_name);
1525         if (ref->name == NULL) {
1526                 free(ref);
1527                 return -ENOMEM;
1528         }
1529
1530         bracket = strchr(ref->name, ']');
1531         if (bracket)
1532                 *bracket = '\0';
1533
1534         ref->addr = addr;
1535
1536         for (i = 0; i < MAP__NR_TYPES; ++i) {
1537                 struct kmap *kmap = map__kmap(maps[i]);
1538                 kmap->ref_reloc_sym = ref;
1539         }
1540
1541         return 0;
1542 }
1543
1544 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1545 {
1546         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1547                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1548                machines__fprintf_dsos(&self->machines, fp);
1549 }
1550
1551 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1552                                           bool with_hits)
1553 {
1554         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1555         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1556 }
1557
1558 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1559 {
1560         struct perf_evsel *pos;
1561         size_t ret = fprintf(fp, "Aggregated stats:\n");
1562
1563         ret += hists__fprintf_nr_events(&session->hists, fp);
1564
1565         list_for_each_entry(pos, &session->evlist->entries, node) {
1566                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1567                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1568         }
1569
1570         return ret;
1571 }
1572
1573 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1574 {
1575         /*
1576          * FIXME: Here we have to actually print all the machines in this
1577          * session, not just the host...
1578          */
1579         return machine__fprintf(&session->host_machine, fp);
1580 }
1581
1582 void perf_session__remove_thread(struct perf_session *session,
1583                                  struct thread *th)
1584 {
1585         /*
1586          * FIXME: This one makes no sense, we need to remove the thread from
1587          * the machine it belongs to, perf_session can have many machines, so
1588          * doing it always on ->host_machine is wrong.  Fix when auditing all
1589          * the 'perf kvm' code.
1590          */
1591         machine__remove_thread(&session->host_machine, th);
1592 }
1593
1594 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1595                                               unsigned int type)
1596 {
1597         struct perf_evsel *pos;
1598
1599         list_for_each_entry(pos, &session->evlist->entries, node) {
1600                 if (pos->attr.type == type)
1601                         return pos;
1602         }
1603         return NULL;
1604 }
1605
1606 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1607                           struct perf_sample *sample, struct machine *machine,
1608                           int print_sym, int print_dso, int print_symoffset)
1609 {
1610         struct addr_location al;
1611         struct callchain_cursor_node *node;
1612
1613         if (perf_event__preprocess_sample(event, machine, &al, sample,
1614                                           NULL) < 0) {
1615                 error("problem processing %d event, skipping it.\n",
1616                         event->header.type);
1617                 return;
1618         }
1619
1620         if (symbol_conf.use_callchain && sample->callchain) {
1621
1622
1623                 if (machine__resolve_callchain(machine, evsel, al.thread,
1624                                                sample, NULL) != 0) {
1625                         if (verbose)
1626                                 error("Failed to resolve callchain. Skipping\n");
1627                         return;
1628                 }
1629                 callchain_cursor_commit(&callchain_cursor);
1630
1631                 while (1) {
1632                         node = callchain_cursor_current(&callchain_cursor);
1633                         if (!node)
1634                                 break;
1635
1636                         printf("\t%16" PRIx64, node->ip);
1637                         if (print_sym) {
1638                                 printf(" ");
1639                                 symbol__fprintf_symname(node->sym, stdout);
1640                         }
1641                         if (print_dso) {
1642                                 printf(" (");
1643                                 map__fprintf_dsoname(node->map, stdout);
1644                                 printf(")");
1645                         }
1646                         printf("\n");
1647
1648                         callchain_cursor_advance(&callchain_cursor);
1649                 }
1650
1651         } else {
1652                 printf("%16" PRIx64, sample->ip);
1653                 if (print_sym) {
1654                         printf(" ");
1655                         if (print_symoffset)
1656                                 symbol__fprintf_symname_offs(al.sym, &al,
1657                                                              stdout);
1658                         else
1659                                 symbol__fprintf_symname(al.sym, stdout);
1660                 }
1661
1662                 if (print_dso) {
1663                         printf(" (");
1664                         map__fprintf_dsoname(al.map, stdout);
1665                         printf(")");
1666                 }
1667         }
1668 }
1669
1670 int perf_session__cpu_bitmap(struct perf_session *session,
1671                              const char *cpu_list, unsigned long *cpu_bitmap)
1672 {
1673         int i;
1674         struct cpu_map *map;
1675
1676         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1677                 struct perf_evsel *evsel;
1678
1679                 evsel = perf_session__find_first_evtype(session, i);
1680                 if (!evsel)
1681                         continue;
1682
1683                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1684                         pr_err("File does not contain CPU events. "
1685                                "Remove -c option to proceed.\n");
1686                         return -1;
1687                 }
1688         }
1689
1690         map = cpu_map__new(cpu_list);
1691         if (map == NULL) {
1692                 pr_err("Invalid cpu_list\n");
1693                 return -1;
1694         }
1695
1696         for (i = 0; i < map->nr; i++) {
1697                 int cpu = map->map[i];
1698
1699                 if (cpu >= MAX_NR_CPUS) {
1700                         pr_err("Requested CPU %d too large. "
1701                                "Consider raising MAX_NR_CPUS\n", cpu);
1702                         return -1;
1703                 }
1704
1705                 set_bit(cpu, cpu_bitmap);
1706         }
1707
1708         return 0;
1709 }
1710
1711 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1712                                 bool full)
1713 {
1714         struct stat st;
1715         int ret;
1716
1717         if (session == NULL || fp == NULL)
1718                 return;
1719
1720         ret = fstat(session->fd, &st);
1721         if (ret == -1)
1722                 return;
1723
1724         fprintf(fp, "# ========\n");
1725         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1726         perf_header__fprintf_info(session, fp, full);
1727         fprintf(fp, "# ========\n#\n");
1728 }
1729
1730
1731 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1732                                              const struct perf_evsel_str_handler *assocs,
1733                                              size_t nr_assocs)
1734 {
1735         struct perf_evlist *evlist = session->evlist;
1736         struct event_format *format;
1737         struct perf_evsel *evsel;
1738         char *tracepoint, *name;
1739         size_t i;
1740         int err;
1741
1742         for (i = 0; i < nr_assocs; i++) {
1743                 err = -ENOMEM;
1744                 tracepoint = strdup(assocs[i].name);
1745                 if (tracepoint == NULL)
1746                         goto out;
1747
1748                 err = -ENOENT;
1749                 name = strchr(tracepoint, ':');
1750                 if (name == NULL)
1751                         goto out_free;
1752
1753                 *name++ = '\0';
1754                 format = pevent_find_event_by_name(session->pevent,
1755                                                    tracepoint, name);
1756                 if (format == NULL) {
1757                         /*
1758                          * Adding a handler for an event not in the session,
1759                          * just ignore it.
1760                          */
1761                         goto next;
1762                 }
1763
1764                 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1765                 if (evsel == NULL)
1766                         goto next;
1767
1768                 err = -EEXIST;
1769                 if (evsel->handler.func != NULL)
1770                         goto out_free;
1771                 evsel->handler.func = assocs[i].handler;
1772 next:
1773                 free(tracepoint);
1774         }
1775
1776         err = 0;
1777 out:
1778         return err;
1779
1780 out_free:
1781         free(tracepoint);
1782         goto out;
1783 }