1 #define _FILE_OFFSET_BITS 64
3 #include <linux/kernel.h>
17 #include "event-parse.h"
18 #include "perf_regs.h"
21 static int perf_session__open(struct perf_session
*self
, bool force
)
23 struct stat input_stat
;
25 if (!strcmp(self
->filename
, "-")) {
27 self
->fd
= STDIN_FILENO
;
29 if (perf_session__read_header(self
, self
->fd
) < 0)
30 pr_err("incompatible file format (rerun with -v to learn more)");
35 self
->fd
= open(self
->filename
, O_RDONLY
);
39 pr_err("failed to open %s: %s", self
->filename
, strerror(err
));
40 if (err
== ENOENT
&& !strcmp(self
->filename
, "perf.data"))
41 pr_err(" (try 'perf record' first)");
46 if (fstat(self
->fd
, &input_stat
) < 0)
49 if (!force
&& input_stat
.st_uid
&& (input_stat
.st_uid
!= geteuid())) {
50 pr_err("file %s not owned by current user or root\n",
55 if (!input_stat
.st_size
) {
56 pr_info("zero-sized file (%s), nothing to do!\n",
61 if (perf_session__read_header(self
, self
->fd
) < 0) {
62 pr_err("incompatible file format (rerun with -v to learn more)");
66 if (!perf_evlist__valid_sample_type(self
->evlist
)) {
67 pr_err("non matching sample_type");
71 if (!perf_evlist__valid_sample_id_all(self
->evlist
)) {
72 pr_err("non matching sample_id_all");
76 self
->size
= input_stat
.st_size
;
85 void perf_session__set_id_hdr_size(struct perf_session
*session
)
87 u16 id_hdr_size
= perf_evlist__id_hdr_size(session
->evlist
);
89 session
->host_machine
.id_hdr_size
= id_hdr_size
;
90 machines__set_id_hdr_size(&session
->machines
, id_hdr_size
);
93 int perf_session__create_kernel_maps(struct perf_session
*self
)
95 int ret
= machine__create_kernel_maps(&self
->host_machine
);
98 ret
= machines__create_guest_kernel_maps(&self
->machines
);
102 static void perf_session__destroy_kernel_maps(struct perf_session
*self
)
104 machine__destroy_kernel_maps(&self
->host_machine
);
105 machines__destroy_guest_kernel_maps(&self
->machines
);
108 struct perf_session
*perf_session__new(const char *filename
, int mode
,
109 bool force
, bool repipe
,
110 struct perf_tool
*tool
)
112 struct perf_session
*self
;
116 if (!filename
|| !strlen(filename
)) {
117 if (!fstat(STDIN_FILENO
, &st
) && S_ISFIFO(st
.st_mode
))
120 filename
= "perf.data";
123 len
= strlen(filename
);
124 self
= zalloc(sizeof(*self
) + len
);
129 memcpy(self
->filename
, filename
, len
);
130 self
->machines
= RB_ROOT
;
131 self
->repipe
= repipe
;
132 INIT_LIST_HEAD(&self
->ordered_samples
.samples
);
133 INIT_LIST_HEAD(&self
->ordered_samples
.sample_cache
);
134 INIT_LIST_HEAD(&self
->ordered_samples
.to_free
);
135 machine__init(&self
->host_machine
, "", HOST_KERNEL_ID
);
136 hists__init(&self
->hists
);
138 if (mode
== O_RDONLY
) {
139 if (perf_session__open(self
, force
) < 0)
141 perf_session__set_id_hdr_size(self
);
142 } else if (mode
== O_WRONLY
) {
144 * In O_RDONLY mode this will be performed when reading the
145 * kernel MMAP event, in perf_event__process_mmap().
147 if (perf_session__create_kernel_maps(self
) < 0)
151 if (tool
&& tool
->ordering_requires_timestamps
&&
152 tool
->ordered_samples
&& !perf_evlist__sample_id_all(self
->evlist
)) {
153 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
154 tool
->ordered_samples
= false;
160 perf_session__delete(self
);
164 static void perf_session__delete_dead_threads(struct perf_session
*session
)
166 machine__delete_dead_threads(&session
->host_machine
);
169 static void perf_session__delete_threads(struct perf_session
*session
)
171 machine__delete_threads(&session
->host_machine
);
174 static void perf_session_env__delete(struct perf_session_env
*env
)
177 free(env
->os_release
);
184 free(env
->sibling_cores
);
185 free(env
->sibling_threads
);
186 free(env
->numa_nodes
);
187 free(env
->pmu_mappings
);
190 void perf_session__delete(struct perf_session
*self
)
192 perf_session__destroy_kernel_maps(self
);
193 perf_session__delete_dead_threads(self
);
194 perf_session__delete_threads(self
);
195 perf_session_env__delete(&self
->header
.env
);
196 machine__exit(&self
->host_machine
);
202 static int process_event_synth_tracing_data_stub(union perf_event
*event
204 struct perf_session
*session
207 dump_printf(": unhandled!\n");
211 static int process_event_synth_attr_stub(union perf_event
*event __maybe_unused
,
212 struct perf_evlist
**pevlist
215 dump_printf(": unhandled!\n");
219 static int process_event_sample_stub(struct perf_tool
*tool __maybe_unused
,
220 union perf_event
*event __maybe_unused
,
221 struct perf_sample
*sample __maybe_unused
,
222 struct perf_evsel
*evsel __maybe_unused
,
223 struct machine
*machine __maybe_unused
)
225 dump_printf(": unhandled!\n");
229 static int process_event_stub(struct perf_tool
*tool __maybe_unused
,
230 union perf_event
*event __maybe_unused
,
231 struct perf_sample
*sample __maybe_unused
,
232 struct machine
*machine __maybe_unused
)
234 dump_printf(": unhandled!\n");
238 static int process_finished_round_stub(struct perf_tool
*tool __maybe_unused
,
239 union perf_event
*event __maybe_unused
,
240 struct perf_session
*perf_session
243 dump_printf(": unhandled!\n");
247 static int process_event_type_stub(struct perf_tool
*tool __maybe_unused
,
248 union perf_event
*event __maybe_unused
)
250 dump_printf(": unhandled!\n");
254 static int process_finished_round(struct perf_tool
*tool
,
255 union perf_event
*event
,
256 struct perf_session
*session
);
258 static void perf_tool__fill_defaults(struct perf_tool
*tool
)
260 if (tool
->sample
== NULL
)
261 tool
->sample
= process_event_sample_stub
;
262 if (tool
->mmap
== NULL
)
263 tool
->mmap
= process_event_stub
;
264 if (tool
->comm
== NULL
)
265 tool
->comm
= process_event_stub
;
266 if (tool
->fork
== NULL
)
267 tool
->fork
= process_event_stub
;
268 if (tool
->exit
== NULL
)
269 tool
->exit
= process_event_stub
;
270 if (tool
->lost
== NULL
)
271 tool
->lost
= perf_event__process_lost
;
272 if (tool
->read
== NULL
)
273 tool
->read
= process_event_sample_stub
;
274 if (tool
->throttle
== NULL
)
275 tool
->throttle
= process_event_stub
;
276 if (tool
->unthrottle
== NULL
)
277 tool
->unthrottle
= process_event_stub
;
278 if (tool
->attr
== NULL
)
279 tool
->attr
= process_event_synth_attr_stub
;
280 if (tool
->event_type
== NULL
)
281 tool
->event_type
= process_event_type_stub
;
282 if (tool
->tracing_data
== NULL
)
283 tool
->tracing_data
= process_event_synth_tracing_data_stub
;
284 if (tool
->build_id
== NULL
)
285 tool
->build_id
= process_finished_round_stub
;
286 if (tool
->finished_round
== NULL
) {
287 if (tool
->ordered_samples
)
288 tool
->finished_round
= process_finished_round
;
290 tool
->finished_round
= process_finished_round_stub
;
294 void mem_bswap_32(void *src
, int byte_size
)
297 while (byte_size
> 0) {
299 byte_size
-= sizeof(u32
);
304 void mem_bswap_64(void *src
, int byte_size
)
308 while (byte_size
> 0) {
310 byte_size
-= sizeof(u64
);
315 static void swap_sample_id_all(union perf_event
*event
, void *data
)
317 void *end
= (void *) event
+ event
->header
.size
;
318 int size
= end
- data
;
320 BUG_ON(size
% sizeof(u64
));
321 mem_bswap_64(data
, size
);
324 static void perf_event__all64_swap(union perf_event
*event
,
325 bool sample_id_all __maybe_unused
)
327 struct perf_event_header
*hdr
= &event
->header
;
328 mem_bswap_64(hdr
+ 1, event
->header
.size
- sizeof(*hdr
));
331 static void perf_event__comm_swap(union perf_event
*event
, bool sample_id_all
)
333 event
->comm
.pid
= bswap_32(event
->comm
.pid
);
334 event
->comm
.tid
= bswap_32(event
->comm
.tid
);
337 void *data
= &event
->comm
.comm
;
339 data
+= PERF_ALIGN(strlen(data
) + 1, sizeof(u64
));
340 swap_sample_id_all(event
, data
);
344 static void perf_event__mmap_swap(union perf_event
*event
,
347 event
->mmap
.pid
= bswap_32(event
->mmap
.pid
);
348 event
->mmap
.tid
= bswap_32(event
->mmap
.tid
);
349 event
->mmap
.start
= bswap_64(event
->mmap
.start
);
350 event
->mmap
.len
= bswap_64(event
->mmap
.len
);
351 event
->mmap
.pgoff
= bswap_64(event
->mmap
.pgoff
);
354 void *data
= &event
->mmap
.filename
;
356 data
+= PERF_ALIGN(strlen(data
) + 1, sizeof(u64
));
357 swap_sample_id_all(event
, data
);
361 static void perf_event__task_swap(union perf_event
*event
, bool sample_id_all
)
363 event
->fork
.pid
= bswap_32(event
->fork
.pid
);
364 event
->fork
.tid
= bswap_32(event
->fork
.tid
);
365 event
->fork
.ppid
= bswap_32(event
->fork
.ppid
);
366 event
->fork
.ptid
= bswap_32(event
->fork
.ptid
);
367 event
->fork
.time
= bswap_64(event
->fork
.time
);
370 swap_sample_id_all(event
, &event
->fork
+ 1);
373 static void perf_event__read_swap(union perf_event
*event
, bool sample_id_all
)
375 event
->read
.pid
= bswap_32(event
->read
.pid
);
376 event
->read
.tid
= bswap_32(event
->read
.tid
);
377 event
->read
.value
= bswap_64(event
->read
.value
);
378 event
->read
.time_enabled
= bswap_64(event
->read
.time_enabled
);
379 event
->read
.time_running
= bswap_64(event
->read
.time_running
);
380 event
->read
.id
= bswap_64(event
->read
.id
);
383 swap_sample_id_all(event
, &event
->read
+ 1);
386 static u8
revbyte(u8 b
)
388 int rev
= (b
>> 4) | ((b
& 0xf) << 4);
389 rev
= ((rev
& 0xcc) >> 2) | ((rev
& 0x33) << 2);
390 rev
= ((rev
& 0xaa) >> 1) | ((rev
& 0x55) << 1);
395 * XXX this is hack in attempt to carry flags bitfield
396 * throught endian village. ABI says:
398 * Bit-fields are allocated from right to left (least to most significant)
399 * on little-endian implementations and from left to right (most to least
400 * significant) on big-endian implementations.
402 * The above seems to be byte specific, so we need to reverse each
403 * byte of the bitfield. 'Internet' also says this might be implementation
404 * specific and we probably need proper fix and carry perf_event_attr
405 * bitfield flags in separate data file FEAT_ section. Thought this seems
408 static void swap_bitfield(u8
*p
, unsigned len
)
412 for (i
= 0; i
< len
; i
++) {
418 /* exported for swapping attributes in file header */
419 void perf_event__attr_swap(struct perf_event_attr
*attr
)
421 attr
->type
= bswap_32(attr
->type
);
422 attr
->size
= bswap_32(attr
->size
);
423 attr
->config
= bswap_64(attr
->config
);
424 attr
->sample_period
= bswap_64(attr
->sample_period
);
425 attr
->sample_type
= bswap_64(attr
->sample_type
);
426 attr
->read_format
= bswap_64(attr
->read_format
);
427 attr
->wakeup_events
= bswap_32(attr
->wakeup_events
);
428 attr
->bp_type
= bswap_32(attr
->bp_type
);
429 attr
->bp_addr
= bswap_64(attr
->bp_addr
);
430 attr
->bp_len
= bswap_64(attr
->bp_len
);
432 swap_bitfield((u8
*) (&attr
->read_format
+ 1), sizeof(u64
));
435 static void perf_event__hdr_attr_swap(union perf_event
*event
,
436 bool sample_id_all __maybe_unused
)
440 perf_event__attr_swap(&event
->attr
.attr
);
442 size
= event
->header
.size
;
443 size
-= (void *)&event
->attr
.id
- (void *)event
;
444 mem_bswap_64(event
->attr
.id
, size
);
447 static void perf_event__event_type_swap(union perf_event
*event
,
448 bool sample_id_all __maybe_unused
)
450 event
->event_type
.event_type
.event_id
=
451 bswap_64(event
->event_type
.event_type
.event_id
);
454 static void perf_event__tracing_data_swap(union perf_event
*event
,
455 bool sample_id_all __maybe_unused
)
457 event
->tracing_data
.size
= bswap_32(event
->tracing_data
.size
);
460 typedef void (*perf_event__swap_op
)(union perf_event
*event
,
463 static perf_event__swap_op perf_event__swap_ops
[] = {
464 [PERF_RECORD_MMAP
] = perf_event__mmap_swap
,
465 [PERF_RECORD_COMM
] = perf_event__comm_swap
,
466 [PERF_RECORD_FORK
] = perf_event__task_swap
,
467 [PERF_RECORD_EXIT
] = perf_event__task_swap
,
468 [PERF_RECORD_LOST
] = perf_event__all64_swap
,
469 [PERF_RECORD_READ
] = perf_event__read_swap
,
470 [PERF_RECORD_SAMPLE
] = perf_event__all64_swap
,
471 [PERF_RECORD_HEADER_ATTR
] = perf_event__hdr_attr_swap
,
472 [PERF_RECORD_HEADER_EVENT_TYPE
] = perf_event__event_type_swap
,
473 [PERF_RECORD_HEADER_TRACING_DATA
] = perf_event__tracing_data_swap
,
474 [PERF_RECORD_HEADER_BUILD_ID
] = NULL
,
475 [PERF_RECORD_HEADER_MAX
] = NULL
,
478 struct sample_queue
{
481 union perf_event
*event
;
482 struct list_head list
;
485 static void perf_session_free_sample_buffers(struct perf_session
*session
)
487 struct ordered_samples
*os
= &session
->ordered_samples
;
489 while (!list_empty(&os
->to_free
)) {
490 struct sample_queue
*sq
;
492 sq
= list_entry(os
->to_free
.next
, struct sample_queue
, list
);
498 static int perf_session_deliver_event(struct perf_session
*session
,
499 union perf_event
*event
,
500 struct perf_sample
*sample
,
501 struct perf_tool
*tool
,
504 static int flush_sample_queue(struct perf_session
*s
,
505 struct perf_tool
*tool
)
507 struct ordered_samples
*os
= &s
->ordered_samples
;
508 struct list_head
*head
= &os
->samples
;
509 struct sample_queue
*tmp
, *iter
;
510 struct perf_sample sample
;
511 u64 limit
= os
->next_flush
;
512 u64 last_ts
= os
->last_sample
? os
->last_sample
->timestamp
: 0ULL;
513 unsigned idx
= 0, progress_next
= os
->nr_samples
/ 16;
516 if (!tool
->ordered_samples
|| !limit
)
519 list_for_each_entry_safe(iter
, tmp
, head
, list
) {
520 if (iter
->timestamp
> limit
)
523 ret
= perf_evlist__parse_sample(s
->evlist
, iter
->event
, &sample
);
525 pr_err("Can't parse sample, err = %d\n", ret
);
527 ret
= perf_session_deliver_event(s
, iter
->event
, &sample
, tool
,
533 os
->last_flush
= iter
->timestamp
;
534 list_del(&iter
->list
);
535 list_add(&iter
->list
, &os
->sample_cache
);
536 if (++idx
>= progress_next
) {
537 progress_next
+= os
->nr_samples
/ 16;
538 ui_progress__update(idx
, os
->nr_samples
,
539 "Processing time ordered events...");
543 if (list_empty(head
)) {
544 os
->last_sample
= NULL
;
545 } else if (last_ts
<= limit
) {
547 list_entry(head
->prev
, struct sample_queue
, list
);
556 * When perf record finishes a pass on every buffers, it records this pseudo
558 * We record the max timestamp t found in the pass n.
559 * Assuming these timestamps are monotonic across cpus, we know that if
560 * a buffer still has events with timestamps below t, they will be all
561 * available and then read in the pass n + 1.
562 * Hence when we start to read the pass n + 2, we can safely flush every
563 * events with timestamps below t.
565 * ============ PASS n =================
568 * cnt1 timestamps | cnt2 timestamps
571 * - | 4 <--- max recorded
573 * ============ PASS n + 1 ==============
576 * cnt1 timestamps | cnt2 timestamps
579 * 5 | 7 <---- max recorded
581 * Flush every events below timestamp 4
583 * ============ PASS n + 2 ==============
586 * cnt1 timestamps | cnt2 timestamps
591 * Flush every events below timestamp 7
594 static int process_finished_round(struct perf_tool
*tool
,
595 union perf_event
*event __maybe_unused
,
596 struct perf_session
*session
)
598 int ret
= flush_sample_queue(session
, tool
);
600 session
->ordered_samples
.next_flush
= session
->ordered_samples
.max_timestamp
;
605 /* The queue is ordered by time */
606 static void __queue_event(struct sample_queue
*new, struct perf_session
*s
)
608 struct ordered_samples
*os
= &s
->ordered_samples
;
609 struct sample_queue
*sample
= os
->last_sample
;
610 u64 timestamp
= new->timestamp
;
614 os
->last_sample
= new;
617 list_add(&new->list
, &os
->samples
);
618 os
->max_timestamp
= timestamp
;
623 * last_sample might point to some random place in the list as it's
624 * the last queued event. We expect that the new event is close to
627 if (sample
->timestamp
<= timestamp
) {
628 while (sample
->timestamp
<= timestamp
) {
629 p
= sample
->list
.next
;
630 if (p
== &os
->samples
) {
631 list_add_tail(&new->list
, &os
->samples
);
632 os
->max_timestamp
= timestamp
;
635 sample
= list_entry(p
, struct sample_queue
, list
);
637 list_add_tail(&new->list
, &sample
->list
);
639 while (sample
->timestamp
> timestamp
) {
640 p
= sample
->list
.prev
;
641 if (p
== &os
->samples
) {
642 list_add(&new->list
, &os
->samples
);
645 sample
= list_entry(p
, struct sample_queue
, list
);
647 list_add(&new->list
, &sample
->list
);
651 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
653 static int perf_session_queue_event(struct perf_session
*s
, union perf_event
*event
,
654 struct perf_sample
*sample
, u64 file_offset
)
656 struct ordered_samples
*os
= &s
->ordered_samples
;
657 struct list_head
*sc
= &os
->sample_cache
;
658 u64 timestamp
= sample
->time
;
659 struct sample_queue
*new;
661 if (!timestamp
|| timestamp
== ~0ULL)
664 if (timestamp
< s
->ordered_samples
.last_flush
) {
665 printf("Warning: Timestamp below last timeslice flush\n");
669 if (!list_empty(sc
)) {
670 new = list_entry(sc
->next
, struct sample_queue
, list
);
671 list_del(&new->list
);
672 } else if (os
->sample_buffer
) {
673 new = os
->sample_buffer
+ os
->sample_buffer_idx
;
674 if (++os
->sample_buffer_idx
== MAX_SAMPLE_BUFFER
)
675 os
->sample_buffer
= NULL
;
677 os
->sample_buffer
= malloc(MAX_SAMPLE_BUFFER
* sizeof(*new));
678 if (!os
->sample_buffer
)
680 list_add(&os
->sample_buffer
->list
, &os
->to_free
);
681 os
->sample_buffer_idx
= 2;
682 new = os
->sample_buffer
+ 1;
685 new->timestamp
= timestamp
;
686 new->file_offset
= file_offset
;
689 __queue_event(new, s
);
694 static void callchain__printf(struct perf_sample
*sample
)
698 printf("... chain: nr:%" PRIu64
"\n", sample
->callchain
->nr
);
700 for (i
= 0; i
< sample
->callchain
->nr
; i
++)
701 printf("..... %2d: %016" PRIx64
"\n",
702 i
, sample
->callchain
->ips
[i
]);
705 static void branch_stack__printf(struct perf_sample
*sample
)
709 printf("... branch stack: nr:%" PRIu64
"\n", sample
->branch_stack
->nr
);
711 for (i
= 0; i
< sample
->branch_stack
->nr
; i
++)
712 printf("..... %2"PRIu64
": %016" PRIx64
" -> %016" PRIx64
"\n",
713 i
, sample
->branch_stack
->entries
[i
].from
,
714 sample
->branch_stack
->entries
[i
].to
);
717 static void regs_dump__printf(u64 mask
, u64
*regs
)
721 for_each_set_bit(rid
, (unsigned long *) &mask
, sizeof(mask
) * 8) {
724 printf(".... %-5s 0x%" PRIx64
"\n",
725 perf_reg_name(rid
), val
);
729 static void regs_user__printf(struct perf_sample
*sample
, u64 mask
)
731 struct regs_dump
*user_regs
= &sample
->user_regs
;
733 if (user_regs
->regs
) {
734 printf("... user regs: mask 0x%" PRIx64
"\n", mask
);
735 regs_dump__printf(mask
, user_regs
->regs
);
739 static void stack_user__printf(struct stack_dump
*dump
)
741 printf("... ustack: size %" PRIu64
", offset 0x%x\n",
742 dump
->size
, dump
->offset
);
745 static void perf_session__print_tstamp(struct perf_session
*session
,
746 union perf_event
*event
,
747 struct perf_sample
*sample
)
749 u64 sample_type
= perf_evlist__sample_type(session
->evlist
);
751 if (event
->header
.type
!= PERF_RECORD_SAMPLE
&&
752 !perf_evlist__sample_id_all(session
->evlist
)) {
753 fputs("-1 -1 ", stdout
);
757 if ((sample_type
& PERF_SAMPLE_CPU
))
758 printf("%u ", sample
->cpu
);
760 if (sample_type
& PERF_SAMPLE_TIME
)
761 printf("%" PRIu64
" ", sample
->time
);
764 static void dump_event(struct perf_session
*session
, union perf_event
*event
,
765 u64 file_offset
, struct perf_sample
*sample
)
770 printf("\n%#" PRIx64
" [%#x]: event: %d\n",
771 file_offset
, event
->header
.size
, event
->header
.type
);
776 perf_session__print_tstamp(session
, event
, sample
);
778 printf("%#" PRIx64
" [%#x]: PERF_RECORD_%s", file_offset
,
779 event
->header
.size
, perf_event__name(event
->header
.type
));
782 static void dump_sample(struct perf_evsel
*evsel
, union perf_event
*event
,
783 struct perf_sample
*sample
)
790 printf("(IP, %d): %d/%d: %#" PRIx64
" period: %" PRIu64
" addr: %#" PRIx64
"\n",
791 event
->header
.misc
, sample
->pid
, sample
->tid
, sample
->ip
,
792 sample
->period
, sample
->addr
);
794 sample_type
= evsel
->attr
.sample_type
;
796 if (sample_type
& PERF_SAMPLE_CALLCHAIN
)
797 callchain__printf(sample
);
799 if (sample_type
& PERF_SAMPLE_BRANCH_STACK
)
800 branch_stack__printf(sample
);
802 if (sample_type
& PERF_SAMPLE_REGS_USER
)
803 regs_user__printf(sample
, evsel
->attr
.sample_regs_user
);
805 if (sample_type
& PERF_SAMPLE_STACK_USER
)
806 stack_user__printf(&sample
->user_stack
);
809 static struct machine
*
810 perf_session__find_machine_for_cpumode(struct perf_session
*session
,
811 union perf_event
*event
)
813 const u8 cpumode
= event
->header
.misc
& PERF_RECORD_MISC_CPUMODE_MASK
;
816 ((cpumode
== PERF_RECORD_MISC_GUEST_KERNEL
) ||
817 (cpumode
== PERF_RECORD_MISC_GUEST_USER
))) {
820 if (event
->header
.type
== PERF_RECORD_MMAP
)
821 pid
= event
->mmap
.pid
;
825 return perf_session__findnew_machine(session
, pid
);
828 return perf_session__find_host_machine(session
);
831 static int perf_session_deliver_event(struct perf_session
*session
,
832 union perf_event
*event
,
833 struct perf_sample
*sample
,
834 struct perf_tool
*tool
,
837 struct perf_evsel
*evsel
;
838 struct machine
*machine
;
840 dump_event(session
, event
, file_offset
, sample
);
842 evsel
= perf_evlist__id2evsel(session
->evlist
, sample
->id
);
843 if (evsel
!= NULL
&& event
->header
.type
!= PERF_RECORD_SAMPLE
) {
845 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
846 * because the tools right now may apply filters, discarding
847 * some of the samples. For consistency, in the future we
848 * should have something like nr_filtered_samples and remove
849 * the sample->period from total_sample_period, etc, KISS for
852 * Also testing against NULL allows us to handle files without
853 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
854 * future probably it'll be a good idea to restrict event
855 * processing via perf_session to files with both set.
857 hists__inc_nr_events(&evsel
->hists
, event
->header
.type
);
860 machine
= perf_session__find_machine_for_cpumode(session
, event
);
862 switch (event
->header
.type
) {
863 case PERF_RECORD_SAMPLE
:
864 dump_sample(evsel
, event
, sample
);
866 ++session
->hists
.stats
.nr_unknown_id
;
869 if (machine
== NULL
) {
870 ++session
->hists
.stats
.nr_unprocessable_samples
;
873 return tool
->sample(tool
, event
, sample
, evsel
, machine
);
874 case PERF_RECORD_MMAP
:
875 return tool
->mmap(tool
, event
, sample
, machine
);
876 case PERF_RECORD_COMM
:
877 return tool
->comm(tool
, event
, sample
, machine
);
878 case PERF_RECORD_FORK
:
879 return tool
->fork(tool
, event
, sample
, machine
);
880 case PERF_RECORD_EXIT
:
881 return tool
->exit(tool
, event
, sample
, machine
);
882 case PERF_RECORD_LOST
:
883 if (tool
->lost
== perf_event__process_lost
)
884 session
->hists
.stats
.total_lost
+= event
->lost
.lost
;
885 return tool
->lost(tool
, event
, sample
, machine
);
886 case PERF_RECORD_READ
:
887 return tool
->read(tool
, event
, sample
, evsel
, machine
);
888 case PERF_RECORD_THROTTLE
:
889 return tool
->throttle(tool
, event
, sample
, machine
);
890 case PERF_RECORD_UNTHROTTLE
:
891 return tool
->unthrottle(tool
, event
, sample
, machine
);
893 ++session
->hists
.stats
.nr_unknown_events
;
898 static int perf_session__preprocess_sample(struct perf_session
*session
,
899 union perf_event
*event
, struct perf_sample
*sample
)
901 if (event
->header
.type
!= PERF_RECORD_SAMPLE
||
902 !(perf_evlist__sample_type(session
->evlist
) & PERF_SAMPLE_CALLCHAIN
))
905 if (!ip_callchain__valid(sample
->callchain
, event
)) {
906 pr_debug("call-chain problem with event, skipping it.\n");
907 ++session
->hists
.stats
.nr_invalid_chains
;
908 session
->hists
.stats
.total_invalid_chains
+= sample
->period
;
914 static int perf_session__process_user_event(struct perf_session
*session
, union perf_event
*event
,
915 struct perf_tool
*tool
, u64 file_offset
)
919 dump_event(session
, event
, file_offset
, NULL
);
921 /* These events are processed right away */
922 switch (event
->header
.type
) {
923 case PERF_RECORD_HEADER_ATTR
:
924 err
= tool
->attr(event
, &session
->evlist
);
926 perf_session__set_id_hdr_size(session
);
928 case PERF_RECORD_HEADER_EVENT_TYPE
:
929 return tool
->event_type(tool
, event
);
930 case PERF_RECORD_HEADER_TRACING_DATA
:
931 /* setup for reading amidst mmap */
932 lseek(session
->fd
, file_offset
, SEEK_SET
);
933 return tool
->tracing_data(event
, session
);
934 case PERF_RECORD_HEADER_BUILD_ID
:
935 return tool
->build_id(tool
, event
, session
);
936 case PERF_RECORD_FINISHED_ROUND
:
937 return tool
->finished_round(tool
, event
, session
);
943 static void event_swap(union perf_event
*event
, bool sample_id_all
)
945 perf_event__swap_op swap
;
947 swap
= perf_event__swap_ops
[event
->header
.type
];
949 swap(event
, sample_id_all
);
952 static int perf_session__process_event(struct perf_session
*session
,
953 union perf_event
*event
,
954 struct perf_tool
*tool
,
957 struct perf_sample sample
;
960 if (session
->header
.needs_swap
)
961 event_swap(event
, perf_evlist__sample_id_all(session
->evlist
));
963 if (event
->header
.type
>= PERF_RECORD_HEADER_MAX
)
966 hists__inc_nr_events(&session
->hists
, event
->header
.type
);
968 if (event
->header
.type
>= PERF_RECORD_USER_TYPE_START
)
969 return perf_session__process_user_event(session
, event
, tool
, file_offset
);
972 * For all kernel events we get the sample data
974 ret
= perf_evlist__parse_sample(session
->evlist
, event
, &sample
);
978 /* Preprocess sample records - precheck callchains */
979 if (perf_session__preprocess_sample(session
, event
, &sample
))
982 if (tool
->ordered_samples
) {
983 ret
= perf_session_queue_event(session
, event
, &sample
,
989 return perf_session_deliver_event(session
, event
, &sample
, tool
,
993 void perf_event_header__bswap(struct perf_event_header
*self
)
995 self
->type
= bswap_32(self
->type
);
996 self
->misc
= bswap_16(self
->misc
);
997 self
->size
= bswap_16(self
->size
);
1000 struct thread
*perf_session__findnew(struct perf_session
*session
, pid_t pid
)
1002 return machine__findnew_thread(&session
->host_machine
, pid
);
1005 static struct thread
*perf_session__register_idle_thread(struct perf_session
*self
)
1007 struct thread
*thread
= perf_session__findnew(self
, 0);
1009 if (thread
== NULL
|| thread__set_comm(thread
, "swapper")) {
1010 pr_err("problem inserting idle task.\n");
1017 static void perf_session__warn_about_errors(const struct perf_session
*session
,
1018 const struct perf_tool
*tool
)
1020 if (tool
->lost
== perf_event__process_lost
&&
1021 session
->hists
.stats
.nr_events
[PERF_RECORD_LOST
] != 0) {
1022 ui__warning("Processed %d events and lost %d chunks!\n\n"
1023 "Check IO/CPU overload!\n\n",
1024 session
->hists
.stats
.nr_events
[0],
1025 session
->hists
.stats
.nr_events
[PERF_RECORD_LOST
]);
1028 if (session
->hists
.stats
.nr_unknown_events
!= 0) {
1029 ui__warning("Found %u unknown events!\n\n"
1030 "Is this an older tool processing a perf.data "
1031 "file generated by a more recent tool?\n\n"
1032 "If that is not the case, consider "
1033 "reporting to linux-kernel@vger.kernel.org.\n\n",
1034 session
->hists
.stats
.nr_unknown_events
);
1037 if (session
->hists
.stats
.nr_unknown_id
!= 0) {
1038 ui__warning("%u samples with id not present in the header\n",
1039 session
->hists
.stats
.nr_unknown_id
);
1042 if (session
->hists
.stats
.nr_invalid_chains
!= 0) {
1043 ui__warning("Found invalid callchains!\n\n"
1044 "%u out of %u events were discarded for this reason.\n\n"
1045 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1046 session
->hists
.stats
.nr_invalid_chains
,
1047 session
->hists
.stats
.nr_events
[PERF_RECORD_SAMPLE
]);
1050 if (session
->hists
.stats
.nr_unprocessable_samples
!= 0) {
1051 ui__warning("%u unprocessable samples recorded.\n"
1052 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1053 session
->hists
.stats
.nr_unprocessable_samples
);
1057 #define session_done() (*(volatile int *)(&session_done))
1058 volatile int session_done
;
1060 static int __perf_session__process_pipe_events(struct perf_session
*self
,
1061 struct perf_tool
*tool
)
1063 union perf_event
*event
;
1064 uint32_t size
, cur_size
= 0;
1071 perf_tool__fill_defaults(tool
);
1074 cur_size
= sizeof(union perf_event
);
1076 buf
= malloc(cur_size
);
1081 err
= readn(self
->fd
, event
, sizeof(struct perf_event_header
));
1086 pr_err("failed to read event header\n");
1090 if (self
->header
.needs_swap
)
1091 perf_event_header__bswap(&event
->header
);
1093 size
= event
->header
.size
;
1097 if (size
> cur_size
) {
1098 void *new = realloc(buf
, size
);
1100 pr_err("failed to allocate memory to read event\n");
1108 p
+= sizeof(struct perf_event_header
);
1110 if (size
- sizeof(struct perf_event_header
)) {
1111 err
= readn(self
->fd
, p
, size
- sizeof(struct perf_event_header
));
1114 pr_err("unexpected end of event stream\n");
1118 pr_err("failed to read event data\n");
1123 if ((skip
= perf_session__process_event(self
, event
, tool
, head
)) < 0) {
1124 pr_err("%#" PRIx64
" [%#x]: failed to process type: %d\n",
1125 head
, event
->header
.size
, event
->header
.type
);
1135 if (!session_done())
1141 perf_session__warn_about_errors(self
, tool
);
1142 perf_session_free_sample_buffers(self
);
1146 static union perf_event
*
1147 fetch_mmaped_event(struct perf_session
*session
,
1148 u64 head
, size_t mmap_size
, char *buf
)
1150 union perf_event
*event
;
1153 * Ensure we have enough space remaining to read
1154 * the size of the event in the headers.
1156 if (head
+ sizeof(event
->header
) > mmap_size
)
1159 event
= (union perf_event
*)(buf
+ head
);
1161 if (session
->header
.needs_swap
)
1162 perf_event_header__bswap(&event
->header
);
1164 if (head
+ event
->header
.size
> mmap_size
)
1171 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1172 * slices. On 32bit we use 32MB.
1174 #if BITS_PER_LONG == 64
1175 #define MMAP_SIZE ULLONG_MAX
1178 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1179 #define NUM_MMAPS 128
1182 int __perf_session__process_events(struct perf_session
*session
,
1183 u64 data_offset
, u64 data_size
,
1184 u64 file_size
, struct perf_tool
*tool
)
1186 u64 head
, page_offset
, file_offset
, file_pos
, progress_next
;
1187 int err
, mmap_prot
, mmap_flags
, map_idx
= 0;
1189 char *buf
, *mmaps
[NUM_MMAPS
];
1190 union perf_event
*event
;
1193 perf_tool__fill_defaults(tool
);
1195 page_offset
= page_size
* (data_offset
/ page_size
);
1196 file_offset
= page_offset
;
1197 head
= data_offset
- page_offset
;
1199 if (data_offset
+ data_size
< file_size
)
1200 file_size
= data_offset
+ data_size
;
1202 progress_next
= file_size
/ 16;
1204 mmap_size
= MMAP_SIZE
;
1205 if (mmap_size
> file_size
)
1206 mmap_size
= file_size
;
1208 memset(mmaps
, 0, sizeof(mmaps
));
1210 mmap_prot
= PROT_READ
;
1211 mmap_flags
= MAP_SHARED
;
1213 if (session
->header
.needs_swap
) {
1214 mmap_prot
|= PROT_WRITE
;
1215 mmap_flags
= MAP_PRIVATE
;
1218 buf
= mmap(NULL
, mmap_size
, mmap_prot
, mmap_flags
, session
->fd
,
1220 if (buf
== MAP_FAILED
) {
1221 pr_err("failed to mmap file\n");
1225 mmaps
[map_idx
] = buf
;
1226 map_idx
= (map_idx
+ 1) & (ARRAY_SIZE(mmaps
) - 1);
1227 file_pos
= file_offset
+ head
;
1230 event
= fetch_mmaped_event(session
, head
, mmap_size
, buf
);
1232 if (mmaps
[map_idx
]) {
1233 munmap(mmaps
[map_idx
], mmap_size
);
1234 mmaps
[map_idx
] = NULL
;
1237 page_offset
= page_size
* (head
/ page_size
);
1238 file_offset
+= page_offset
;
1239 head
-= page_offset
;
1243 size
= event
->header
.size
;
1246 perf_session__process_event(session
, event
, tool
, file_pos
) < 0) {
1247 pr_err("%#" PRIx64
" [%#x]: failed to process type: %d\n",
1248 file_offset
+ head
, event
->header
.size
,
1249 event
->header
.type
);
1257 if (file_pos
>= progress_next
) {
1258 progress_next
+= file_size
/ 16;
1259 ui_progress__update(file_pos
, file_size
,
1260 "Processing events...");
1263 if (file_pos
< file_size
)
1267 /* do the final flush for ordered samples */
1268 session
->ordered_samples
.next_flush
= ULLONG_MAX
;
1269 err
= flush_sample_queue(session
, tool
);
1271 ui_progress__finish();
1272 perf_session__warn_about_errors(session
, tool
);
1273 perf_session_free_sample_buffers(session
);
1277 int perf_session__process_events(struct perf_session
*self
,
1278 struct perf_tool
*tool
)
1282 if (perf_session__register_idle_thread(self
) == NULL
)
1286 err
= __perf_session__process_events(self
,
1287 self
->header
.data_offset
,
1288 self
->header
.data_size
,
1291 err
= __perf_session__process_pipe_events(self
, tool
);
1296 bool perf_session__has_traces(struct perf_session
*session
, const char *msg
)
1298 if (!(perf_evlist__sample_type(session
->evlist
) & PERF_SAMPLE_RAW
)) {
1299 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg
);
1306 int maps__set_kallsyms_ref_reloc_sym(struct map
**maps
,
1307 const char *symbol_name
, u64 addr
)
1311 struct ref_reloc_sym
*ref
;
1313 ref
= zalloc(sizeof(struct ref_reloc_sym
));
1317 ref
->name
= strdup(symbol_name
);
1318 if (ref
->name
== NULL
) {
1323 bracket
= strchr(ref
->name
, ']');
1329 for (i
= 0; i
< MAP__NR_TYPES
; ++i
) {
1330 struct kmap
*kmap
= map__kmap(maps
[i
]);
1331 kmap
->ref_reloc_sym
= ref
;
1337 size_t perf_session__fprintf_dsos(struct perf_session
*self
, FILE *fp
)
1339 return __dsos__fprintf(&self
->host_machine
.kernel_dsos
, fp
) +
1340 __dsos__fprintf(&self
->host_machine
.user_dsos
, fp
) +
1341 machines__fprintf_dsos(&self
->machines
, fp
);
1344 size_t perf_session__fprintf_dsos_buildid(struct perf_session
*self
, FILE *fp
,
1345 bool (skip
)(struct dso
*dso
, int parm
), int parm
)
1347 size_t ret
= machine__fprintf_dsos_buildid(&self
->host_machine
, fp
, skip
, parm
);
1348 return ret
+ machines__fprintf_dsos_buildid(&self
->machines
, fp
, skip
, parm
);
1351 size_t perf_session__fprintf_nr_events(struct perf_session
*session
, FILE *fp
)
1353 struct perf_evsel
*pos
;
1354 size_t ret
= fprintf(fp
, "Aggregated stats:\n");
1356 ret
+= hists__fprintf_nr_events(&session
->hists
, fp
);
1358 list_for_each_entry(pos
, &session
->evlist
->entries
, node
) {
1359 ret
+= fprintf(fp
, "%s stats:\n", perf_evsel__name(pos
));
1360 ret
+= hists__fprintf_nr_events(&pos
->hists
, fp
);
1366 size_t perf_session__fprintf(struct perf_session
*session
, FILE *fp
)
1369 * FIXME: Here we have to actually print all the machines in this
1370 * session, not just the host...
1372 return machine__fprintf(&session
->host_machine
, fp
);
1375 void perf_session__remove_thread(struct perf_session
*session
,
1379 * FIXME: This one makes no sense, we need to remove the thread from
1380 * the machine it belongs to, perf_session can have many machines, so
1381 * doing it always on ->host_machine is wrong. Fix when auditing all
1382 * the 'perf kvm' code.
1384 machine__remove_thread(&session
->host_machine
, th
);
1387 struct perf_evsel
*perf_session__find_first_evtype(struct perf_session
*session
,
1390 struct perf_evsel
*pos
;
1392 list_for_each_entry(pos
, &session
->evlist
->entries
, node
) {
1393 if (pos
->attr
.type
== type
)
1399 void perf_evsel__print_ip(struct perf_evsel
*evsel
, union perf_event
*event
,
1400 struct perf_sample
*sample
, struct machine
*machine
,
1401 int print_sym
, int print_dso
, int print_symoffset
)
1403 struct addr_location al
;
1404 struct callchain_cursor_node
*node
;
1406 if (perf_event__preprocess_sample(event
, machine
, &al
, sample
,
1408 error("problem processing %d event, skipping it.\n",
1409 event
->header
.type
);
1413 if (symbol_conf
.use_callchain
&& sample
->callchain
) {
1416 if (machine__resolve_callchain(machine
, evsel
, al
.thread
,
1417 sample
, NULL
) != 0) {
1419 error("Failed to resolve callchain. Skipping\n");
1422 callchain_cursor_commit(&callchain_cursor
);
1425 node
= callchain_cursor_current(&callchain_cursor
);
1429 printf("\t%16" PRIx64
, node
->ip
);
1432 symbol__fprintf_symname(node
->sym
, stdout
);
1436 map__fprintf_dsoname(node
->map
, stdout
);
1441 callchain_cursor_advance(&callchain_cursor
);
1445 printf("%16" PRIx64
, sample
->ip
);
1448 if (print_symoffset
)
1449 symbol__fprintf_symname_offs(al
.sym
, &al
,
1452 symbol__fprintf_symname(al
.sym
, stdout
);
1457 map__fprintf_dsoname(al
.map
, stdout
);
1463 int perf_session__cpu_bitmap(struct perf_session
*session
,
1464 const char *cpu_list
, unsigned long *cpu_bitmap
)
1467 struct cpu_map
*map
;
1469 for (i
= 0; i
< PERF_TYPE_MAX
; ++i
) {
1470 struct perf_evsel
*evsel
;
1472 evsel
= perf_session__find_first_evtype(session
, i
);
1476 if (!(evsel
->attr
.sample_type
& PERF_SAMPLE_CPU
)) {
1477 pr_err("File does not contain CPU events. "
1478 "Remove -c option to proceed.\n");
1483 map
= cpu_map__new(cpu_list
);
1485 pr_err("Invalid cpu_list\n");
1489 for (i
= 0; i
< map
->nr
; i
++) {
1490 int cpu
= map
->map
[i
];
1492 if (cpu
>= MAX_NR_CPUS
) {
1493 pr_err("Requested CPU %d too large. "
1494 "Consider raising MAX_NR_CPUS\n", cpu
);
1498 set_bit(cpu
, cpu_bitmap
);
1504 void perf_session__fprintf_info(struct perf_session
*session
, FILE *fp
,
1510 if (session
== NULL
|| fp
== NULL
)
1513 ret
= fstat(session
->fd
, &st
);
1517 fprintf(fp
, "# ========\n");
1518 fprintf(fp
, "# captured on: %s", ctime(&st
.st_ctime
));
1519 perf_header__fprintf_info(session
, fp
, full
);
1520 fprintf(fp
, "# ========\n#\n");
1524 int __perf_session__set_tracepoints_handlers(struct perf_session
*session
,
1525 const struct perf_evsel_str_handler
*assocs
,
1528 struct perf_evlist
*evlist
= session
->evlist
;
1529 struct event_format
*format
;
1530 struct perf_evsel
*evsel
;
1531 char *tracepoint
, *name
;
1535 for (i
= 0; i
< nr_assocs
; i
++) {
1537 tracepoint
= strdup(assocs
[i
].name
);
1538 if (tracepoint
== NULL
)
1542 name
= strchr(tracepoint
, ':');
1547 format
= pevent_find_event_by_name(session
->pevent
,
1549 if (format
== NULL
) {
1551 * Adding a handler for an event not in the session,
1557 evsel
= perf_evlist__find_tracepoint_by_id(evlist
, format
->id
);
1562 if (evsel
->handler
.func
!= NULL
)
1564 evsel
->handler
.func
= assocs
[i
].handler
;