perf tools: Fix mmap limitations on 32-bit
[~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         self->machines = RB_ROOT;
132         self->repipe = repipe;
133         INIT_LIST_HEAD(&self->ordered_samples.samples);
134         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
135         INIT_LIST_HEAD(&self->ordered_samples.to_free);
136         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
137         hists__init(&self->hists);
138
139         if (mode == O_RDONLY) {
140                 if (perf_session__open(self, force) < 0)
141                         goto out_delete;
142                 perf_session__set_id_hdr_size(self);
143         } else if (mode == O_WRONLY) {
144                 /*
145                  * In O_RDONLY mode this will be performed when reading the
146                  * kernel MMAP event, in perf_event__process_mmap().
147                  */
148                 if (perf_session__create_kernel_maps(self) < 0)
149                         goto out_delete;
150         }
151
152         if (tool && tool->ordering_requires_timestamps &&
153             tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
154                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
155                 tool->ordered_samples = false;
156         }
157
158 out:
159         return self;
160 out_delete:
161         perf_session__delete(self);
162         return NULL;
163 }
164
165 static void machine__delete_dead_threads(struct machine *machine)
166 {
167         struct thread *n, *t;
168
169         list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
170                 list_del(&t->node);
171                 thread__delete(t);
172         }
173 }
174
175 static void perf_session__delete_dead_threads(struct perf_session *session)
176 {
177         machine__delete_dead_threads(&session->host_machine);
178 }
179
180 static void machine__delete_threads(struct machine *self)
181 {
182         struct rb_node *nd = rb_first(&self->threads);
183
184         while (nd) {
185                 struct thread *t = rb_entry(nd, struct thread, rb_node);
186
187                 rb_erase(&t->rb_node, &self->threads);
188                 nd = rb_next(nd);
189                 thread__delete(t);
190         }
191 }
192
193 static void perf_session__delete_threads(struct perf_session *session)
194 {
195         machine__delete_threads(&session->host_machine);
196 }
197
198 static void perf_session_env__delete(struct perf_session_env *env)
199 {
200         free(env->hostname);
201         free(env->os_release);
202         free(env->version);
203         free(env->arch);
204         free(env->cpu_desc);
205         free(env->cpuid);
206
207         free(env->cmdline);
208         free(env->sibling_cores);
209         free(env->sibling_threads);
210         free(env->numa_nodes);
211         free(env->pmu_mappings);
212 }
213
214 void perf_session__delete(struct perf_session *self)
215 {
216         perf_session__destroy_kernel_maps(self);
217         perf_session__delete_dead_threads(self);
218         perf_session__delete_threads(self);
219         perf_session_env__delete(&self->header.env);
220         machine__exit(&self->host_machine);
221         close(self->fd);
222         free(self);
223         vdso__exit();
224 }
225
226 void machine__remove_thread(struct machine *self, struct thread *th)
227 {
228         self->last_match = NULL;
229         rb_erase(&th->rb_node, &self->threads);
230         /*
231          * We may have references to this thread, for instance in some hist_entry
232          * instances, so just move them to a separate list.
233          */
234         list_add_tail(&th->node, &self->dead_threads);
235 }
236
237 static bool symbol__match_parent_regex(struct symbol *sym)
238 {
239         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
240                 return 1;
241
242         return 0;
243 }
244
245 static const u8 cpumodes[] = {
246         PERF_RECORD_MISC_USER,
247         PERF_RECORD_MISC_KERNEL,
248         PERF_RECORD_MISC_GUEST_USER,
249         PERF_RECORD_MISC_GUEST_KERNEL
250 };
251 #define NCPUMODES (sizeof(cpumodes)/sizeof(u8))
252
253 static void ip__resolve_ams(struct machine *self, struct thread *thread,
254                             struct addr_map_symbol *ams,
255                             u64 ip)
256 {
257         struct addr_location al;
258         size_t i;
259         u8 m;
260
261         memset(&al, 0, sizeof(al));
262
263         for (i = 0; i < NCPUMODES; i++) {
264                 m = cpumodes[i];
265                 /*
266                  * We cannot use the header.misc hint to determine whether a
267                  * branch stack address is user, kernel, guest, hypervisor.
268                  * Branches may straddle the kernel/user/hypervisor boundaries.
269                  * Thus, we have to try consecutively until we find a match
270                  * or else, the symbol is unknown
271                  */
272                 thread__find_addr_location(thread, self, m, MAP__FUNCTION,
273                                 ip, &al, NULL);
274                 if (al.sym)
275                         goto found;
276         }
277 found:
278         ams->addr = ip;
279         ams->al_addr = al.addr;
280         ams->sym = al.sym;
281         ams->map = al.map;
282 }
283
284 struct branch_info *machine__resolve_bstack(struct machine *self,
285                                             struct thread *thr,
286                                             struct branch_stack *bs)
287 {
288         struct branch_info *bi;
289         unsigned int i;
290
291         bi = calloc(bs->nr, sizeof(struct branch_info));
292         if (!bi)
293                 return NULL;
294
295         for (i = 0; i < bs->nr; i++) {
296                 ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to);
297                 ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from);
298                 bi[i].flags = bs->entries[i].flags;
299         }
300         return bi;
301 }
302
303 static int machine__resolve_callchain_sample(struct machine *machine,
304                                              struct thread *thread,
305                                              struct ip_callchain *chain,
306                                              struct symbol **parent)
307
308 {
309         u8 cpumode = PERF_RECORD_MISC_USER;
310         unsigned int i;
311         int err;
312
313         callchain_cursor_reset(&callchain_cursor);
314
315         if (chain->nr > PERF_MAX_STACK_DEPTH) {
316                 pr_warning("corrupted callchain. skipping...\n");
317                 return 0;
318         }
319
320         for (i = 0; i < chain->nr; i++) {
321                 u64 ip;
322                 struct addr_location al;
323
324                 if (callchain_param.order == ORDER_CALLEE)
325                         ip = chain->ips[i];
326                 else
327                         ip = chain->ips[chain->nr - i - 1];
328
329                 if (ip >= PERF_CONTEXT_MAX) {
330                         switch (ip) {
331                         case PERF_CONTEXT_HV:
332                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;
333                                 break;
334                         case PERF_CONTEXT_KERNEL:
335                                 cpumode = PERF_RECORD_MISC_KERNEL;
336                                 break;
337                         case PERF_CONTEXT_USER:
338                                 cpumode = PERF_RECORD_MISC_USER;
339                                 break;
340                         default:
341                                 pr_debug("invalid callchain context: "
342                                          "%"PRId64"\n", (s64) ip);
343                                 /*
344                                  * It seems the callchain is corrupted.
345                                  * Discard all.
346                                  */
347                                 callchain_cursor_reset(&callchain_cursor);
348                                 return 0;
349                         }
350                         continue;
351                 }
352
353                 al.filtered = false;
354                 thread__find_addr_location(thread, machine, cpumode,
355                                            MAP__FUNCTION, ip, &al, NULL);
356                 if (al.sym != NULL) {
357                         if (sort__has_parent && !*parent &&
358                             symbol__match_parent_regex(al.sym))
359                                 *parent = al.sym;
360                         if (!symbol_conf.use_callchain)
361                                 break;
362                 }
363
364                 err = callchain_cursor_append(&callchain_cursor,
365                                               ip, al.map, al.sym);
366                 if (err)
367                         return err;
368         }
369
370         return 0;
371 }
372
373 static int unwind_entry(struct unwind_entry *entry, void *arg)
374 {
375         struct callchain_cursor *cursor = arg;
376         return callchain_cursor_append(cursor, entry->ip,
377                                        entry->map, entry->sym);
378 }
379
380 int machine__resolve_callchain(struct machine *machine,
381                                struct perf_evsel *evsel,
382                                struct thread *thread,
383                                struct perf_sample *sample,
384                                struct symbol **parent)
385
386 {
387         int ret;
388
389         callchain_cursor_reset(&callchain_cursor);
390
391         ret = machine__resolve_callchain_sample(machine, thread,
392                                                 sample->callchain, parent);
393         if (ret)
394                 return ret;
395
396         /* Can we do dwarf post unwind? */
397         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
398               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
399                 return 0;
400
401         /* Bail out if nothing was captured. */
402         if ((!sample->user_regs.regs) ||
403             (!sample->user_stack.size))
404                 return 0;
405
406         return unwind__get_entries(unwind_entry, &callchain_cursor, machine,
407                                    thread, evsel->attr.sample_regs_user,
408                                    sample);
409
410 }
411
412 static int process_event_synth_tracing_data_stub(union perf_event *event
413                                                  __maybe_unused,
414                                                  struct perf_session *session
415                                                 __maybe_unused)
416 {
417         dump_printf(": unhandled!\n");
418         return 0;
419 }
420
421 static int process_event_synth_attr_stub(union perf_event *event __maybe_unused,
422                                          struct perf_evlist **pevlist
423                                          __maybe_unused)
424 {
425         dump_printf(": unhandled!\n");
426         return 0;
427 }
428
429 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
430                                      union perf_event *event __maybe_unused,
431                                      struct perf_sample *sample __maybe_unused,
432                                      struct perf_evsel *evsel __maybe_unused,
433                                      struct machine *machine __maybe_unused)
434 {
435         dump_printf(": unhandled!\n");
436         return 0;
437 }
438
439 static int process_event_stub(struct perf_tool *tool __maybe_unused,
440                               union perf_event *event __maybe_unused,
441                               struct perf_sample *sample __maybe_unused,
442                               struct machine *machine __maybe_unused)
443 {
444         dump_printf(": unhandled!\n");
445         return 0;
446 }
447
448 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
449                                        union perf_event *event __maybe_unused,
450                                        struct perf_session *perf_session
451                                        __maybe_unused)
452 {
453         dump_printf(": unhandled!\n");
454         return 0;
455 }
456
457 static int process_event_type_stub(struct perf_tool *tool __maybe_unused,
458                                    union perf_event *event __maybe_unused)
459 {
460         dump_printf(": unhandled!\n");
461         return 0;
462 }
463
464 static int process_finished_round(struct perf_tool *tool,
465                                   union perf_event *event,
466                                   struct perf_session *session);
467
468 static void perf_tool__fill_defaults(struct perf_tool *tool)
469 {
470         if (tool->sample == NULL)
471                 tool->sample = process_event_sample_stub;
472         if (tool->mmap == NULL)
473                 tool->mmap = process_event_stub;
474         if (tool->comm == NULL)
475                 tool->comm = process_event_stub;
476         if (tool->fork == NULL)
477                 tool->fork = process_event_stub;
478         if (tool->exit == NULL)
479                 tool->exit = process_event_stub;
480         if (tool->lost == NULL)
481                 tool->lost = perf_event__process_lost;
482         if (tool->read == NULL)
483                 tool->read = process_event_sample_stub;
484         if (tool->throttle == NULL)
485                 tool->throttle = process_event_stub;
486         if (tool->unthrottle == NULL)
487                 tool->unthrottle = process_event_stub;
488         if (tool->attr == NULL)
489                 tool->attr = process_event_synth_attr_stub;
490         if (tool->event_type == NULL)
491                 tool->event_type = process_event_type_stub;
492         if (tool->tracing_data == NULL)
493                 tool->tracing_data = process_event_synth_tracing_data_stub;
494         if (tool->build_id == NULL)
495                 tool->build_id = process_finished_round_stub;
496         if (tool->finished_round == NULL) {
497                 if (tool->ordered_samples)
498                         tool->finished_round = process_finished_round;
499                 else
500                         tool->finished_round = process_finished_round_stub;
501         }
502 }
503  
504 void mem_bswap_32(void *src, int byte_size)
505 {
506         u32 *m = src;
507         while (byte_size > 0) {
508                 *m = bswap_32(*m);
509                 byte_size -= sizeof(u32);
510                 ++m;
511         }
512 }
513
514 void mem_bswap_64(void *src, int byte_size)
515 {
516         u64 *m = src;
517
518         while (byte_size > 0) {
519                 *m = bswap_64(*m);
520                 byte_size -= sizeof(u64);
521                 ++m;
522         }
523 }
524
525 static void swap_sample_id_all(union perf_event *event, void *data)
526 {
527         void *end = (void *) event + event->header.size;
528         int size = end - data;
529
530         BUG_ON(size % sizeof(u64));
531         mem_bswap_64(data, size);
532 }
533
534 static void perf_event__all64_swap(union perf_event *event,
535                                    bool sample_id_all __maybe_unused)
536 {
537         struct perf_event_header *hdr = &event->header;
538         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
539 }
540
541 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
542 {
543         event->comm.pid = bswap_32(event->comm.pid);
544         event->comm.tid = bswap_32(event->comm.tid);
545
546         if (sample_id_all) {
547                 void *data = &event->comm.comm;
548
549                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
550                 swap_sample_id_all(event, data);
551         }
552 }
553
554 static void perf_event__mmap_swap(union perf_event *event,
555                                   bool sample_id_all)
556 {
557         event->mmap.pid   = bswap_32(event->mmap.pid);
558         event->mmap.tid   = bswap_32(event->mmap.tid);
559         event->mmap.start = bswap_64(event->mmap.start);
560         event->mmap.len   = bswap_64(event->mmap.len);
561         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
562
563         if (sample_id_all) {
564                 void *data = &event->mmap.filename;
565
566                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
567                 swap_sample_id_all(event, data);
568         }
569 }
570
571 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
572 {
573         event->fork.pid  = bswap_32(event->fork.pid);
574         event->fork.tid  = bswap_32(event->fork.tid);
575         event->fork.ppid = bswap_32(event->fork.ppid);
576         event->fork.ptid = bswap_32(event->fork.ptid);
577         event->fork.time = bswap_64(event->fork.time);
578
579         if (sample_id_all)
580                 swap_sample_id_all(event, &event->fork + 1);
581 }
582
583 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
584 {
585         event->read.pid          = bswap_32(event->read.pid);
586         event->read.tid          = bswap_32(event->read.tid);
587         event->read.value        = bswap_64(event->read.value);
588         event->read.time_enabled = bswap_64(event->read.time_enabled);
589         event->read.time_running = bswap_64(event->read.time_running);
590         event->read.id           = bswap_64(event->read.id);
591
592         if (sample_id_all)
593                 swap_sample_id_all(event, &event->read + 1);
594 }
595
596 static u8 revbyte(u8 b)
597 {
598         int rev = (b >> 4) | ((b & 0xf) << 4);
599         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
600         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
601         return (u8) rev;
602 }
603
604 /*
605  * XXX this is hack in attempt to carry flags bitfield
606  * throught endian village. ABI says:
607  *
608  * Bit-fields are allocated from right to left (least to most significant)
609  * on little-endian implementations and from left to right (most to least
610  * significant) on big-endian implementations.
611  *
612  * The above seems to be byte specific, so we need to reverse each
613  * byte of the bitfield. 'Internet' also says this might be implementation
614  * specific and we probably need proper fix and carry perf_event_attr
615  * bitfield flags in separate data file FEAT_ section. Thought this seems
616  * to work for now.
617  */
618 static void swap_bitfield(u8 *p, unsigned len)
619 {
620         unsigned i;
621
622         for (i = 0; i < len; i++) {
623                 *p = revbyte(*p);
624                 p++;
625         }
626 }
627
628 /* exported for swapping attributes in file header */
629 void perf_event__attr_swap(struct perf_event_attr *attr)
630 {
631         attr->type              = bswap_32(attr->type);
632         attr->size              = bswap_32(attr->size);
633         attr->config            = bswap_64(attr->config);
634         attr->sample_period     = bswap_64(attr->sample_period);
635         attr->sample_type       = bswap_64(attr->sample_type);
636         attr->read_format       = bswap_64(attr->read_format);
637         attr->wakeup_events     = bswap_32(attr->wakeup_events);
638         attr->bp_type           = bswap_32(attr->bp_type);
639         attr->bp_addr           = bswap_64(attr->bp_addr);
640         attr->bp_len            = bswap_64(attr->bp_len);
641
642         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
643 }
644
645 static void perf_event__hdr_attr_swap(union perf_event *event,
646                                       bool sample_id_all __maybe_unused)
647 {
648         size_t size;
649
650         perf_event__attr_swap(&event->attr.attr);
651
652         size = event->header.size;
653         size -= (void *)&event->attr.id - (void *)event;
654         mem_bswap_64(event->attr.id, size);
655 }
656
657 static void perf_event__event_type_swap(union perf_event *event,
658                                         bool sample_id_all __maybe_unused)
659 {
660         event->event_type.event_type.event_id =
661                 bswap_64(event->event_type.event_type.event_id);
662 }
663
664 static void perf_event__tracing_data_swap(union perf_event *event,
665                                           bool sample_id_all __maybe_unused)
666 {
667         event->tracing_data.size = bswap_32(event->tracing_data.size);
668 }
669
670 typedef void (*perf_event__swap_op)(union perf_event *event,
671                                     bool sample_id_all);
672
673 static perf_event__swap_op perf_event__swap_ops[] = {
674         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
675         [PERF_RECORD_COMM]                = perf_event__comm_swap,
676         [PERF_RECORD_FORK]                = perf_event__task_swap,
677         [PERF_RECORD_EXIT]                = perf_event__task_swap,
678         [PERF_RECORD_LOST]                = perf_event__all64_swap,
679         [PERF_RECORD_READ]                = perf_event__read_swap,
680         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
681         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
682         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
683         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
684         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
685         [PERF_RECORD_HEADER_MAX]          = NULL,
686 };
687
688 struct sample_queue {
689         u64                     timestamp;
690         u64                     file_offset;
691         union perf_event        *event;
692         struct list_head        list;
693 };
694
695 static void perf_session_free_sample_buffers(struct perf_session *session)
696 {
697         struct ordered_samples *os = &session->ordered_samples;
698
699         while (!list_empty(&os->to_free)) {
700                 struct sample_queue *sq;
701
702                 sq = list_entry(os->to_free.next, struct sample_queue, list);
703                 list_del(&sq->list);
704                 free(sq);
705         }
706 }
707
708 static int perf_session_deliver_event(struct perf_session *session,
709                                       union perf_event *event,
710                                       struct perf_sample *sample,
711                                       struct perf_tool *tool,
712                                       u64 file_offset);
713
714 static int flush_sample_queue(struct perf_session *s,
715                                struct perf_tool *tool)
716 {
717         struct ordered_samples *os = &s->ordered_samples;
718         struct list_head *head = &os->samples;
719         struct sample_queue *tmp, *iter;
720         struct perf_sample sample;
721         u64 limit = os->next_flush;
722         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
723         unsigned idx = 0, progress_next = os->nr_samples / 16;
724         int ret;
725
726         if (!tool->ordered_samples || !limit)
727                 return 0;
728
729         list_for_each_entry_safe(iter, tmp, head, list) {
730                 if (iter->timestamp > limit)
731                         break;
732
733                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
734                 if (ret)
735                         pr_err("Can't parse sample, err = %d\n", ret);
736                 else {
737                         ret = perf_session_deliver_event(s, iter->event, &sample, tool,
738                                                          iter->file_offset);
739                         if (ret)
740                                 return ret;
741                 }
742
743                 os->last_flush = iter->timestamp;
744                 list_del(&iter->list);
745                 list_add(&iter->list, &os->sample_cache);
746                 if (++idx >= progress_next) {
747                         progress_next += os->nr_samples / 16;
748                         ui_progress__update(idx, os->nr_samples,
749                                             "Processing time ordered events...");
750                 }
751         }
752
753         if (list_empty(head)) {
754                 os->last_sample = NULL;
755         } else if (last_ts <= limit) {
756                 os->last_sample =
757                         list_entry(head->prev, struct sample_queue, list);
758         }
759
760         os->nr_samples = 0;
761
762         return 0;
763 }
764
765 /*
766  * When perf record finishes a pass on every buffers, it records this pseudo
767  * event.
768  * We record the max timestamp t found in the pass n.
769  * Assuming these timestamps are monotonic across cpus, we know that if
770  * a buffer still has events with timestamps below t, they will be all
771  * available and then read in the pass n + 1.
772  * Hence when we start to read the pass n + 2, we can safely flush every
773  * events with timestamps below t.
774  *
775  *    ============ PASS n =================
776  *       CPU 0         |   CPU 1
777  *                     |
778  *    cnt1 timestamps  |   cnt2 timestamps
779  *          1          |         2
780  *          2          |         3
781  *          -          |         4  <--- max recorded
782  *
783  *    ============ PASS n + 1 ==============
784  *       CPU 0         |   CPU 1
785  *                     |
786  *    cnt1 timestamps  |   cnt2 timestamps
787  *          3          |         5
788  *          4          |         6
789  *          5          |         7 <---- max recorded
790  *
791  *      Flush every events below timestamp 4
792  *
793  *    ============ PASS n + 2 ==============
794  *       CPU 0         |   CPU 1
795  *                     |
796  *    cnt1 timestamps  |   cnt2 timestamps
797  *          6          |         8
798  *          7          |         9
799  *          -          |         10
800  *
801  *      Flush every events below timestamp 7
802  *      etc...
803  */
804 static int process_finished_round(struct perf_tool *tool,
805                                   union perf_event *event __maybe_unused,
806                                   struct perf_session *session)
807 {
808         int ret = flush_sample_queue(session, tool);
809         if (!ret)
810                 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
811
812         return ret;
813 }
814
815 /* The queue is ordered by time */
816 static void __queue_event(struct sample_queue *new, struct perf_session *s)
817 {
818         struct ordered_samples *os = &s->ordered_samples;
819         struct sample_queue *sample = os->last_sample;
820         u64 timestamp = new->timestamp;
821         struct list_head *p;
822
823         ++os->nr_samples;
824         os->last_sample = new;
825
826         if (!sample) {
827                 list_add(&new->list, &os->samples);
828                 os->max_timestamp = timestamp;
829                 return;
830         }
831
832         /*
833          * last_sample might point to some random place in the list as it's
834          * the last queued event. We expect that the new event is close to
835          * this.
836          */
837         if (sample->timestamp <= timestamp) {
838                 while (sample->timestamp <= timestamp) {
839                         p = sample->list.next;
840                         if (p == &os->samples) {
841                                 list_add_tail(&new->list, &os->samples);
842                                 os->max_timestamp = timestamp;
843                                 return;
844                         }
845                         sample = list_entry(p, struct sample_queue, list);
846                 }
847                 list_add_tail(&new->list, &sample->list);
848         } else {
849                 while (sample->timestamp > timestamp) {
850                         p = sample->list.prev;
851                         if (p == &os->samples) {
852                                 list_add(&new->list, &os->samples);
853                                 return;
854                         }
855                         sample = list_entry(p, struct sample_queue, list);
856                 }
857                 list_add(&new->list, &sample->list);
858         }
859 }
860
861 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
862
863 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
864                                     struct perf_sample *sample, u64 file_offset)
865 {
866         struct ordered_samples *os = &s->ordered_samples;
867         struct list_head *sc = &os->sample_cache;
868         u64 timestamp = sample->time;
869         struct sample_queue *new;
870
871         if (!timestamp || timestamp == ~0ULL)
872                 return -ETIME;
873
874         if (timestamp < s->ordered_samples.last_flush) {
875                 printf("Warning: Timestamp below last timeslice flush\n");
876                 return -EINVAL;
877         }
878
879         if (!list_empty(sc)) {
880                 new = list_entry(sc->next, struct sample_queue, list);
881                 list_del(&new->list);
882         } else if (os->sample_buffer) {
883                 new = os->sample_buffer + os->sample_buffer_idx;
884                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
885                         os->sample_buffer = NULL;
886         } else {
887                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
888                 if (!os->sample_buffer)
889                         return -ENOMEM;
890                 list_add(&os->sample_buffer->list, &os->to_free);
891                 os->sample_buffer_idx = 2;
892                 new = os->sample_buffer + 1;
893         }
894
895         new->timestamp = timestamp;
896         new->file_offset = file_offset;
897         new->event = event;
898
899         __queue_event(new, s);
900
901         return 0;
902 }
903
904 static void callchain__printf(struct perf_sample *sample)
905 {
906         unsigned int i;
907
908         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
909
910         for (i = 0; i < sample->callchain->nr; i++)
911                 printf("..... %2d: %016" PRIx64 "\n",
912                        i, sample->callchain->ips[i]);
913 }
914
915 static void branch_stack__printf(struct perf_sample *sample)
916 {
917         uint64_t i;
918
919         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
920
921         for (i = 0; i < sample->branch_stack->nr; i++)
922                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
923                         i, sample->branch_stack->entries[i].from,
924                         sample->branch_stack->entries[i].to);
925 }
926
927 static void regs_dump__printf(u64 mask, u64 *regs)
928 {
929         unsigned rid, i = 0;
930
931         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
932                 u64 val = regs[i++];
933
934                 printf(".... %-5s 0x%" PRIx64 "\n",
935                        perf_reg_name(rid), val);
936         }
937 }
938
939 static void regs_user__printf(struct perf_sample *sample, u64 mask)
940 {
941         struct regs_dump *user_regs = &sample->user_regs;
942
943         if (user_regs->regs) {
944                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
945                 regs_dump__printf(mask, user_regs->regs);
946         }
947 }
948
949 static void stack_user__printf(struct stack_dump *dump)
950 {
951         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
952                dump->size, dump->offset);
953 }
954
955 static void perf_session__print_tstamp(struct perf_session *session,
956                                        union perf_event *event,
957                                        struct perf_sample *sample)
958 {
959         u64 sample_type = perf_evlist__sample_type(session->evlist);
960
961         if (event->header.type != PERF_RECORD_SAMPLE &&
962             !perf_evlist__sample_id_all(session->evlist)) {
963                 fputs("-1 -1 ", stdout);
964                 return;
965         }
966
967         if ((sample_type & PERF_SAMPLE_CPU))
968                 printf("%u ", sample->cpu);
969
970         if (sample_type & PERF_SAMPLE_TIME)
971                 printf("%" PRIu64 " ", sample->time);
972 }
973
974 static void dump_event(struct perf_session *session, union perf_event *event,
975                        u64 file_offset, struct perf_sample *sample)
976 {
977         if (!dump_trace)
978                 return;
979
980         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
981                file_offset, event->header.size, event->header.type);
982
983         trace_event(event);
984
985         if (sample)
986                 perf_session__print_tstamp(session, event, sample);
987
988         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
989                event->header.size, perf_event__name(event->header.type));
990 }
991
992 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
993                         struct perf_sample *sample)
994 {
995         u64 sample_type;
996
997         if (!dump_trace)
998                 return;
999
1000         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
1001                event->header.misc, sample->pid, sample->tid, sample->ip,
1002                sample->period, sample->addr);
1003
1004         sample_type = evsel->attr.sample_type;
1005
1006         if (sample_type & PERF_SAMPLE_CALLCHAIN)
1007                 callchain__printf(sample);
1008
1009         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
1010                 branch_stack__printf(sample);
1011
1012         if (sample_type & PERF_SAMPLE_REGS_USER)
1013                 regs_user__printf(sample, evsel->attr.sample_regs_user);
1014
1015         if (sample_type & PERF_SAMPLE_STACK_USER)
1016                 stack_user__printf(&sample->user_stack);
1017 }
1018
1019 static struct machine *
1020         perf_session__find_machine_for_cpumode(struct perf_session *session,
1021                                                union perf_event *event)
1022 {
1023         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1024
1025         if (perf_guest &&
1026             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
1027              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
1028                 u32 pid;
1029
1030                 if (event->header.type == PERF_RECORD_MMAP)
1031                         pid = event->mmap.pid;
1032                 else
1033                         pid = event->ip.pid;
1034
1035                 return perf_session__findnew_machine(session, pid);
1036         }
1037
1038         return perf_session__find_host_machine(session);
1039 }
1040
1041 static int perf_session_deliver_event(struct perf_session *session,
1042                                       union perf_event *event,
1043                                       struct perf_sample *sample,
1044                                       struct perf_tool *tool,
1045                                       u64 file_offset)
1046 {
1047         struct perf_evsel *evsel;
1048         struct machine *machine;
1049
1050         dump_event(session, event, file_offset, sample);
1051
1052         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
1053         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
1054                 /*
1055                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
1056                  * because the tools right now may apply filters, discarding
1057                  * some of the samples. For consistency, in the future we
1058                  * should have something like nr_filtered_samples and remove
1059                  * the sample->period from total_sample_period, etc, KISS for
1060                  * now tho.
1061                  *
1062                  * Also testing against NULL allows us to handle files without
1063                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
1064                  * future probably it'll be a good idea to restrict event
1065                  * processing via perf_session to files with both set.
1066                  */
1067                 hists__inc_nr_events(&evsel->hists, event->header.type);
1068         }
1069
1070         machine = perf_session__find_machine_for_cpumode(session, event);
1071
1072         switch (event->header.type) {
1073         case PERF_RECORD_SAMPLE:
1074                 dump_sample(evsel, event, sample);
1075                 if (evsel == NULL) {
1076                         ++session->hists.stats.nr_unknown_id;
1077                         return 0;
1078                 }
1079                 if (machine == NULL) {
1080                         ++session->hists.stats.nr_unprocessable_samples;
1081                         return 0;
1082                 }
1083                 return tool->sample(tool, event, sample, evsel, machine);
1084         case PERF_RECORD_MMAP:
1085                 return tool->mmap(tool, event, sample, machine);
1086         case PERF_RECORD_COMM:
1087                 return tool->comm(tool, event, sample, machine);
1088         case PERF_RECORD_FORK:
1089                 return tool->fork(tool, event, sample, machine);
1090         case PERF_RECORD_EXIT:
1091                 return tool->exit(tool, event, sample, machine);
1092         case PERF_RECORD_LOST:
1093                 if (tool->lost == perf_event__process_lost)
1094                         session->hists.stats.total_lost += event->lost.lost;
1095                 return tool->lost(tool, event, sample, machine);
1096         case PERF_RECORD_READ:
1097                 return tool->read(tool, event, sample, evsel, machine);
1098         case PERF_RECORD_THROTTLE:
1099                 return tool->throttle(tool, event, sample, machine);
1100         case PERF_RECORD_UNTHROTTLE:
1101                 return tool->unthrottle(tool, event, sample, machine);
1102         default:
1103                 ++session->hists.stats.nr_unknown_events;
1104                 return -1;
1105         }
1106 }
1107
1108 static int perf_session__preprocess_sample(struct perf_session *session,
1109                                            union perf_event *event, struct perf_sample *sample)
1110 {
1111         if (event->header.type != PERF_RECORD_SAMPLE ||
1112             !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN))
1113                 return 0;
1114
1115         if (!ip_callchain__valid(sample->callchain, event)) {
1116                 pr_debug("call-chain problem with event, skipping it.\n");
1117                 ++session->hists.stats.nr_invalid_chains;
1118                 session->hists.stats.total_invalid_chains += sample->period;
1119                 return -EINVAL;
1120         }
1121         return 0;
1122 }
1123
1124 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1125                                             struct perf_tool *tool, u64 file_offset)
1126 {
1127         int err;
1128
1129         dump_event(session, event, file_offset, NULL);
1130
1131         /* These events are processed right away */
1132         switch (event->header.type) {
1133         case PERF_RECORD_HEADER_ATTR:
1134                 err = tool->attr(event, &session->evlist);
1135                 if (err == 0)
1136                         perf_session__set_id_hdr_size(session);
1137                 return err;
1138         case PERF_RECORD_HEADER_EVENT_TYPE:
1139                 return tool->event_type(tool, event);
1140         case PERF_RECORD_HEADER_TRACING_DATA:
1141                 /* setup for reading amidst mmap */
1142                 lseek(session->fd, file_offset, SEEK_SET);
1143                 return tool->tracing_data(event, session);
1144         case PERF_RECORD_HEADER_BUILD_ID:
1145                 return tool->build_id(tool, event, session);
1146         case PERF_RECORD_FINISHED_ROUND:
1147                 return tool->finished_round(tool, event, session);
1148         default:
1149                 return -EINVAL;
1150         }
1151 }
1152
1153 static void event_swap(union perf_event *event, bool sample_id_all)
1154 {
1155         perf_event__swap_op swap;
1156
1157         swap = perf_event__swap_ops[event->header.type];
1158         if (swap)
1159                 swap(event, sample_id_all);
1160 }
1161
1162 static int perf_session__process_event(struct perf_session *session,
1163                                        union perf_event *event,
1164                                        struct perf_tool *tool,
1165                                        u64 file_offset)
1166 {
1167         struct perf_sample sample;
1168         int ret;
1169
1170         if (session->header.needs_swap)
1171                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1172
1173         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1174                 return -EINVAL;
1175
1176         hists__inc_nr_events(&session->hists, event->header.type);
1177
1178         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1179                 return perf_session__process_user_event(session, event, tool, file_offset);
1180
1181         /*
1182          * For all kernel events we get the sample data
1183          */
1184         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1185         if (ret)
1186                 return ret;
1187
1188         /* Preprocess sample records - precheck callchains */
1189         if (perf_session__preprocess_sample(session, event, &sample))
1190                 return 0;
1191
1192         if (tool->ordered_samples) {
1193                 ret = perf_session_queue_event(session, event, &sample,
1194                                                file_offset);
1195                 if (ret != -ETIME)
1196                         return ret;
1197         }
1198
1199         return perf_session_deliver_event(session, event, &sample, tool,
1200                                           file_offset);
1201 }
1202
1203 void perf_event_header__bswap(struct perf_event_header *self)
1204 {
1205         self->type = bswap_32(self->type);
1206         self->misc = bswap_16(self->misc);
1207         self->size = bswap_16(self->size);
1208 }
1209
1210 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1211 {
1212         return machine__findnew_thread(&session->host_machine, pid);
1213 }
1214
1215 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1216 {
1217         struct thread *thread = perf_session__findnew(self, 0);
1218
1219         if (thread == NULL || thread__set_comm(thread, "swapper")) {
1220                 pr_err("problem inserting idle task.\n");
1221                 thread = NULL;
1222         }
1223
1224         return thread;
1225 }
1226
1227 static void perf_session__warn_about_errors(const struct perf_session *session,
1228                                             const struct perf_tool *tool)
1229 {
1230         if (tool->lost == perf_event__process_lost &&
1231             session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1232                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1233                             "Check IO/CPU overload!\n\n",
1234                             session->hists.stats.nr_events[0],
1235                             session->hists.stats.nr_events[PERF_RECORD_LOST]);
1236         }
1237
1238         if (session->hists.stats.nr_unknown_events != 0) {
1239                 ui__warning("Found %u unknown events!\n\n"
1240                             "Is this an older tool processing a perf.data "
1241                             "file generated by a more recent tool?\n\n"
1242                             "If that is not the case, consider "
1243                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1244                             session->hists.stats.nr_unknown_events);
1245         }
1246
1247         if (session->hists.stats.nr_unknown_id != 0) {
1248                 ui__warning("%u samples with id not present in the header\n",
1249                             session->hists.stats.nr_unknown_id);
1250         }
1251
1252         if (session->hists.stats.nr_invalid_chains != 0) {
1253                 ui__warning("Found invalid callchains!\n\n"
1254                             "%u out of %u events were discarded for this reason.\n\n"
1255                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1256                             session->hists.stats.nr_invalid_chains,
1257                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1258         }
1259
1260         if (session->hists.stats.nr_unprocessable_samples != 0) {
1261                 ui__warning("%u unprocessable samples recorded.\n"
1262                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1263                             session->hists.stats.nr_unprocessable_samples);
1264         }
1265 }
1266
1267 #define session_done()  (*(volatile int *)(&session_done))
1268 volatile int session_done;
1269
1270 static int __perf_session__process_pipe_events(struct perf_session *self,
1271                                                struct perf_tool *tool)
1272 {
1273         union perf_event *event;
1274         uint32_t size, cur_size = 0;
1275         void *buf = NULL;
1276         int skip = 0;
1277         u64 head;
1278         int err;
1279         void *p;
1280
1281         perf_tool__fill_defaults(tool);
1282
1283         head = 0;
1284         cur_size = sizeof(union perf_event);
1285
1286         buf = malloc(cur_size);
1287         if (!buf)
1288                 return -errno;
1289 more:
1290         event = buf;
1291         err = readn(self->fd, event, sizeof(struct perf_event_header));
1292         if (err <= 0) {
1293                 if (err == 0)
1294                         goto done;
1295
1296                 pr_err("failed to read event header\n");
1297                 goto out_err;
1298         }
1299
1300         if (self->header.needs_swap)
1301                 perf_event_header__bswap(&event->header);
1302
1303         size = event->header.size;
1304         if (size == 0)
1305                 size = 8;
1306
1307         if (size > cur_size) {
1308                 void *new = realloc(buf, size);
1309                 if (!new) {
1310                         pr_err("failed to allocate memory to read event\n");
1311                         goto out_err;
1312                 }
1313                 buf = new;
1314                 cur_size = size;
1315                 event = buf;
1316         }
1317         p = event;
1318         p += sizeof(struct perf_event_header);
1319
1320         if (size - sizeof(struct perf_event_header)) {
1321                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1322                 if (err <= 0) {
1323                         if (err == 0) {
1324                                 pr_err("unexpected end of event stream\n");
1325                                 goto done;
1326                         }
1327
1328                         pr_err("failed to read event data\n");
1329                         goto out_err;
1330                 }
1331         }
1332
1333         if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1334                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1335                        head, event->header.size, event->header.type);
1336                 err = -EINVAL;
1337                 goto out_err;
1338         }
1339
1340         head += size;
1341
1342         if (skip > 0)
1343                 head += skip;
1344
1345         if (!session_done())
1346                 goto more;
1347 done:
1348         err = 0;
1349 out_err:
1350         free(buf);
1351         perf_session__warn_about_errors(self, tool);
1352         perf_session_free_sample_buffers(self);
1353         return err;
1354 }
1355
1356 static union perf_event *
1357 fetch_mmaped_event(struct perf_session *session,
1358                    u64 head, size_t mmap_size, char *buf)
1359 {
1360         union perf_event *event;
1361
1362         /*
1363          * Ensure we have enough space remaining to read
1364          * the size of the event in the headers.
1365          */
1366         if (head + sizeof(event->header) > mmap_size)
1367                 return NULL;
1368
1369         event = (union perf_event *)(buf + head);
1370
1371         if (session->header.needs_swap)
1372                 perf_event_header__bswap(&event->header);
1373
1374         if (head + event->header.size > mmap_size)
1375                 return NULL;
1376
1377         return event;
1378 }
1379
1380 /*
1381  * On 64bit we can mmap the data file in one go. No need for tiny mmap
1382  * slices. On 32bit we use 32MB.
1383  */
1384 #if BITS_PER_LONG == 64
1385 #define MMAP_SIZE ULLONG_MAX
1386 #define NUM_MMAPS 1
1387 #else
1388 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1389 #define NUM_MMAPS 128
1390 #endif
1391
1392 int __perf_session__process_events(struct perf_session *session,
1393                                    u64 data_offset, u64 data_size,
1394                                    u64 file_size, struct perf_tool *tool)
1395 {
1396         u64 head, page_offset, file_offset, file_pos, progress_next;
1397         int err, mmap_prot, mmap_flags, map_idx = 0;
1398         size_t  mmap_size;
1399         char *buf, *mmaps[NUM_MMAPS];
1400         union perf_event *event;
1401         uint32_t size;
1402
1403         perf_tool__fill_defaults(tool);
1404
1405         page_offset = page_size * (data_offset / page_size);
1406         file_offset = page_offset;
1407         head = data_offset - page_offset;
1408
1409         if (data_offset + data_size < file_size)
1410                 file_size = data_offset + data_size;
1411
1412         progress_next = file_size / 16;
1413
1414         mmap_size = MMAP_SIZE;
1415         if (mmap_size > file_size)
1416                 mmap_size = file_size;
1417
1418         memset(mmaps, 0, sizeof(mmaps));
1419
1420         mmap_prot  = PROT_READ;
1421         mmap_flags = MAP_SHARED;
1422
1423         if (session->header.needs_swap) {
1424                 mmap_prot  |= PROT_WRITE;
1425                 mmap_flags = MAP_PRIVATE;
1426         }
1427 remap:
1428         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1429                    file_offset);
1430         if (buf == MAP_FAILED) {
1431                 pr_err("failed to mmap file\n");
1432                 err = -errno;
1433                 goto out_err;
1434         }
1435         mmaps[map_idx] = buf;
1436         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1437         file_pos = file_offset + head;
1438
1439 more:
1440         event = fetch_mmaped_event(session, head, mmap_size, buf);
1441         if (!event) {
1442                 if (mmaps[map_idx]) {
1443                         munmap(mmaps[map_idx], mmap_size);
1444                         mmaps[map_idx] = NULL;
1445                 }
1446
1447                 page_offset = page_size * (head / page_size);
1448                 file_offset += page_offset;
1449                 head -= page_offset;
1450                 goto remap;
1451         }
1452
1453         size = event->header.size;
1454
1455         if (size == 0 ||
1456             perf_session__process_event(session, event, tool, file_pos) < 0) {
1457                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1458                        file_offset + head, event->header.size,
1459                        event->header.type);
1460                 err = -EINVAL;
1461                 goto out_err;
1462         }
1463
1464         head += size;
1465         file_pos += size;
1466
1467         if (file_pos >= progress_next) {
1468                 progress_next += file_size / 16;
1469                 ui_progress__update(file_pos, file_size,
1470                                     "Processing events...");
1471         }
1472
1473         if (file_pos < file_size)
1474                 goto more;
1475
1476         err = 0;
1477         /* do the final flush for ordered samples */
1478         session->ordered_samples.next_flush = ULLONG_MAX;
1479         err = flush_sample_queue(session, tool);
1480 out_err:
1481         ui_progress__finish();
1482         perf_session__warn_about_errors(session, tool);
1483         perf_session_free_sample_buffers(session);
1484         return err;
1485 }
1486
1487 int perf_session__process_events(struct perf_session *self,
1488                                  struct perf_tool *tool)
1489 {
1490         int err;
1491
1492         if (perf_session__register_idle_thread(self) == NULL)
1493                 return -ENOMEM;
1494
1495         if (!self->fd_pipe)
1496                 err = __perf_session__process_events(self,
1497                                                      self->header.data_offset,
1498                                                      self->header.data_size,
1499                                                      self->size, tool);
1500         else
1501                 err = __perf_session__process_pipe_events(self, tool);
1502
1503         return err;
1504 }
1505
1506 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1507 {
1508         if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) {
1509                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1510                 return false;
1511         }
1512
1513         return true;
1514 }
1515
1516 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1517                                      const char *symbol_name, u64 addr)
1518 {
1519         char *bracket;
1520         enum map_type i;
1521         struct ref_reloc_sym *ref;
1522
1523         ref = zalloc(sizeof(struct ref_reloc_sym));
1524         if (ref == NULL)
1525                 return -ENOMEM;
1526
1527         ref->name = strdup(symbol_name);
1528         if (ref->name == NULL) {
1529                 free(ref);
1530                 return -ENOMEM;
1531         }
1532
1533         bracket = strchr(ref->name, ']');
1534         if (bracket)
1535                 *bracket = '\0';
1536
1537         ref->addr = addr;
1538
1539         for (i = 0; i < MAP__NR_TYPES; ++i) {
1540                 struct kmap *kmap = map__kmap(maps[i]);
1541                 kmap->ref_reloc_sym = ref;
1542         }
1543
1544         return 0;
1545 }
1546
1547 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1548 {
1549         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1550                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1551                machines__fprintf_dsos(&self->machines, fp);
1552 }
1553
1554 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1555                                           bool with_hits)
1556 {
1557         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1558         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1559 }
1560
1561 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1562 {
1563         struct perf_evsel *pos;
1564         size_t ret = fprintf(fp, "Aggregated stats:\n");
1565
1566         ret += hists__fprintf_nr_events(&session->hists, fp);
1567
1568         list_for_each_entry(pos, &session->evlist->entries, node) {
1569                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1570                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1571         }
1572
1573         return ret;
1574 }
1575
1576 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1577 {
1578         /*
1579          * FIXME: Here we have to actually print all the machines in this
1580          * session, not just the host...
1581          */
1582         return machine__fprintf(&session->host_machine, fp);
1583 }
1584
1585 void perf_session__remove_thread(struct perf_session *session,
1586                                  struct thread *th)
1587 {
1588         /*
1589          * FIXME: This one makes no sense, we need to remove the thread from
1590          * the machine it belongs to, perf_session can have many machines, so
1591          * doing it always on ->host_machine is wrong.  Fix when auditing all
1592          * the 'perf kvm' code.
1593          */
1594         machine__remove_thread(&session->host_machine, th);
1595 }
1596
1597 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1598                                               unsigned int type)
1599 {
1600         struct perf_evsel *pos;
1601
1602         list_for_each_entry(pos, &session->evlist->entries, node) {
1603                 if (pos->attr.type == type)
1604                         return pos;
1605         }
1606         return NULL;
1607 }
1608
1609 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1610                           struct perf_sample *sample, struct machine *machine,
1611                           int print_sym, int print_dso, int print_symoffset)
1612 {
1613         struct addr_location al;
1614         struct callchain_cursor_node *node;
1615
1616         if (perf_event__preprocess_sample(event, machine, &al, sample,
1617                                           NULL) < 0) {
1618                 error("problem processing %d event, skipping it.\n",
1619                         event->header.type);
1620                 return;
1621         }
1622
1623         if (symbol_conf.use_callchain && sample->callchain) {
1624
1625
1626                 if (machine__resolve_callchain(machine, evsel, al.thread,
1627                                                sample, NULL) != 0) {
1628                         if (verbose)
1629                                 error("Failed to resolve callchain. Skipping\n");
1630                         return;
1631                 }
1632                 callchain_cursor_commit(&callchain_cursor);
1633
1634                 while (1) {
1635                         node = callchain_cursor_current(&callchain_cursor);
1636                         if (!node)
1637                                 break;
1638
1639                         printf("\t%16" PRIx64, node->ip);
1640                         if (print_sym) {
1641                                 printf(" ");
1642                                 symbol__fprintf_symname(node->sym, stdout);
1643                         }
1644                         if (print_dso) {
1645                                 printf(" (");
1646                                 map__fprintf_dsoname(node->map, stdout);
1647                                 printf(")");
1648                         }
1649                         printf("\n");
1650
1651                         callchain_cursor_advance(&callchain_cursor);
1652                 }
1653
1654         } else {
1655                 printf("%16" PRIx64, sample->ip);
1656                 if (print_sym) {
1657                         printf(" ");
1658                         if (print_symoffset)
1659                                 symbol__fprintf_symname_offs(al.sym, &al,
1660                                                              stdout);
1661                         else
1662                                 symbol__fprintf_symname(al.sym, stdout);
1663                 }
1664
1665                 if (print_dso) {
1666                         printf(" (");
1667                         map__fprintf_dsoname(al.map, stdout);
1668                         printf(")");
1669                 }
1670         }
1671 }
1672
1673 int perf_session__cpu_bitmap(struct perf_session *session,
1674                              const char *cpu_list, unsigned long *cpu_bitmap)
1675 {
1676         int i;
1677         struct cpu_map *map;
1678
1679         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1680                 struct perf_evsel *evsel;
1681
1682                 evsel = perf_session__find_first_evtype(session, i);
1683                 if (!evsel)
1684                         continue;
1685
1686                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1687                         pr_err("File does not contain CPU events. "
1688                                "Remove -c option to proceed.\n");
1689                         return -1;
1690                 }
1691         }
1692
1693         map = cpu_map__new(cpu_list);
1694         if (map == NULL) {
1695                 pr_err("Invalid cpu_list\n");
1696                 return -1;
1697         }
1698
1699         for (i = 0; i < map->nr; i++) {
1700                 int cpu = map->map[i];
1701
1702                 if (cpu >= MAX_NR_CPUS) {
1703                         pr_err("Requested CPU %d too large. "
1704                                "Consider raising MAX_NR_CPUS\n", cpu);
1705                         return -1;
1706                 }
1707
1708                 set_bit(cpu, cpu_bitmap);
1709         }
1710
1711         return 0;
1712 }
1713
1714 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1715                                 bool full)
1716 {
1717         struct stat st;
1718         int ret;
1719
1720         if (session == NULL || fp == NULL)
1721                 return;
1722
1723         ret = fstat(session->fd, &st);
1724         if (ret == -1)
1725                 return;
1726
1727         fprintf(fp, "# ========\n");
1728         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1729         perf_header__fprintf_info(session, fp, full);
1730         fprintf(fp, "# ========\n#\n");
1731 }
1732
1733
1734 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1735                                              const struct perf_evsel_str_handler *assocs,
1736                                              size_t nr_assocs)
1737 {
1738         struct perf_evlist *evlist = session->evlist;
1739         struct event_format *format;
1740         struct perf_evsel *evsel;
1741         char *tracepoint, *name;
1742         size_t i;
1743         int err;
1744
1745         for (i = 0; i < nr_assocs; i++) {
1746                 err = -ENOMEM;
1747                 tracepoint = strdup(assocs[i].name);
1748                 if (tracepoint == NULL)
1749                         goto out;
1750
1751                 err = -ENOENT;
1752                 name = strchr(tracepoint, ':');
1753                 if (name == NULL)
1754                         goto out_free;
1755
1756                 *name++ = '\0';
1757                 format = pevent_find_event_by_name(session->pevent,
1758                                                    tracepoint, name);
1759                 if (format == NULL) {
1760                         /*
1761                          * Adding a handler for an event not in the session,
1762                          * just ignore it.
1763                          */
1764                         goto next;
1765                 }
1766
1767                 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1768                 if (evsel == NULL)
1769                         goto next;
1770
1771                 err = -EEXIST;
1772                 if (evsel->handler.func != NULL)
1773                         goto out_free;
1774                 evsel->handler.func = assocs[i].handler;
1775 next:
1776                 free(tracepoint);
1777         }
1778
1779         err = 0;
1780 out:
1781         return err;
1782
1783 out_free:
1784         free(tracepoint);
1785         goto out;
1786 }