1 #include <linux/kernel.h>
2 #include <traceevent/event-parse.h>
16 #include "perf_regs.h"
19 static int perf_session__open(struct perf_session *session)
21 struct perf_data_file *file = session->file;
23 if (perf_session__read_header(session) < 0) {
24 pr_err("incompatible file format (rerun with -v to learn more)");
28 if (perf_data_file__is_pipe(file))
31 if (!perf_evlist__valid_sample_type(session->evlist)) {
32 pr_err("non matching sample_type");
36 if (!perf_evlist__valid_sample_id_all(session->evlist)) {
37 pr_err("non matching sample_id_all");
41 if (!perf_evlist__valid_read_format(session->evlist)) {
42 pr_err("non matching read_format");
49 void perf_session__set_id_hdr_size(struct perf_session *session)
51 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
53 machines__set_id_hdr_size(&session->machines, id_hdr_size);
56 int perf_session__create_kernel_maps(struct perf_session *session)
58 int ret = machine__create_kernel_maps(&session->machines.host);
61 ret = machines__create_guest_kernel_maps(&session->machines);
65 static void perf_session__destroy_kernel_maps(struct perf_session *session)
67 machines__destroy_kernel_maps(&session->machines);
70 static bool perf_session__has_comm_exec(struct perf_session *session)
72 struct perf_evsel *evsel;
74 evlist__for_each(session->evlist, evsel) {
75 if (evsel->attr.comm_exec)
82 static void perf_session__set_comm_exec(struct perf_session *session)
84 bool comm_exec = perf_session__has_comm_exec(session);
86 machines__set_comm_exec(&session->machines, comm_exec);
89 struct perf_session *perf_session__new(struct perf_data_file *file,
90 bool repipe, struct perf_tool *tool)
92 struct perf_session *session = zalloc(sizeof(*session));
97 session->repipe = repipe;
98 ordered_events__init(&session->ordered_events);
99 machines__init(&session->machines);
102 if (perf_data_file__open(file))
105 session->file = file;
107 if (perf_data_file__is_read(file)) {
108 if (perf_session__open(session) < 0)
111 perf_session__set_id_hdr_size(session);
112 perf_session__set_comm_exec(session);
116 if (!file || perf_data_file__is_write(file)) {
118 * In O_RDONLY mode this will be performed when reading the
119 * kernel MMAP event, in perf_event__process_mmap().
121 if (perf_session__create_kernel_maps(session) < 0)
122 pr_warning("Cannot read kernel map\n");
125 if (tool && tool->ordering_requires_timestamps &&
126 tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) {
127 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
128 tool->ordered_events = false;
134 perf_data_file__close(file);
136 perf_session__delete(session);
141 static void perf_session__delete_dead_threads(struct perf_session *session)
143 machine__delete_dead_threads(&session->machines.host);
146 static void perf_session__delete_threads(struct perf_session *session)
148 machine__delete_threads(&session->machines.host);
151 static void perf_session_env__delete(struct perf_session_env *env)
153 zfree(&env->hostname);
154 zfree(&env->os_release);
155 zfree(&env->version);
157 zfree(&env->cpu_desc);
160 zfree(&env->cmdline);
161 zfree(&env->sibling_cores);
162 zfree(&env->sibling_threads);
163 zfree(&env->numa_nodes);
164 zfree(&env->pmu_mappings);
167 void perf_session__delete(struct perf_session *session)
169 perf_session__destroy_kernel_maps(session);
170 perf_session__delete_dead_threads(session);
171 perf_session__delete_threads(session);
172 perf_session_env__delete(&session->header.env);
173 machines__exit(&session->machines);
175 perf_data_file__close(session->file);
179 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
181 union perf_event *event
183 struct perf_session *session
186 dump_printf(": unhandled!\n");
190 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
191 union perf_event *event __maybe_unused,
192 struct perf_evlist **pevlist
195 dump_printf(": unhandled!\n");
199 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
200 union perf_event *event __maybe_unused,
201 struct perf_sample *sample __maybe_unused,
202 struct perf_evsel *evsel __maybe_unused,
203 struct machine *machine __maybe_unused)
205 dump_printf(": unhandled!\n");
209 static int process_event_stub(struct perf_tool *tool __maybe_unused,
210 union perf_event *event __maybe_unused,
211 struct perf_sample *sample __maybe_unused,
212 struct machine *machine __maybe_unused)
214 dump_printf(": unhandled!\n");
218 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
219 union perf_event *event __maybe_unused,
220 struct perf_session *perf_session
223 dump_printf(": unhandled!\n");
227 static int process_finished_round(struct perf_tool *tool,
228 union perf_event *event,
229 struct perf_session *session);
231 static int process_id_index_stub(struct perf_tool *tool __maybe_unused,
232 union perf_event *event __maybe_unused,
233 struct perf_session *perf_session
236 dump_printf(": unhandled!\n");
240 void perf_tool__fill_defaults(struct perf_tool *tool)
242 if (tool->sample == NULL)
243 tool->sample = process_event_sample_stub;
244 if (tool->mmap == NULL)
245 tool->mmap = process_event_stub;
246 if (tool->mmap2 == NULL)
247 tool->mmap2 = process_event_stub;
248 if (tool->comm == NULL)
249 tool->comm = process_event_stub;
250 if (tool->fork == NULL)
251 tool->fork = process_event_stub;
252 if (tool->exit == NULL)
253 tool->exit = process_event_stub;
254 if (tool->lost == NULL)
255 tool->lost = perf_event__process_lost;
256 if (tool->read == NULL)
257 tool->read = process_event_sample_stub;
258 if (tool->throttle == NULL)
259 tool->throttle = process_event_stub;
260 if (tool->unthrottle == NULL)
261 tool->unthrottle = process_event_stub;
262 if (tool->attr == NULL)
263 tool->attr = process_event_synth_attr_stub;
264 if (tool->tracing_data == NULL)
265 tool->tracing_data = process_event_synth_tracing_data_stub;
266 if (tool->build_id == NULL)
267 tool->build_id = process_finished_round_stub;
268 if (tool->finished_round == NULL) {
269 if (tool->ordered_events)
270 tool->finished_round = process_finished_round;
272 tool->finished_round = process_finished_round_stub;
274 if (tool->id_index == NULL)
275 tool->id_index = process_id_index_stub;
278 static void swap_sample_id_all(union perf_event *event, void *data)
280 void *end = (void *) event + event->header.size;
281 int size = end - data;
283 BUG_ON(size % sizeof(u64));
284 mem_bswap_64(data, size);
287 static void perf_event__all64_swap(union perf_event *event,
288 bool sample_id_all __maybe_unused)
290 struct perf_event_header *hdr = &event->header;
291 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
294 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
296 event->comm.pid = bswap_32(event->comm.pid);
297 event->comm.tid = bswap_32(event->comm.tid);
300 void *data = &event->comm.comm;
302 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
303 swap_sample_id_all(event, data);
307 static void perf_event__mmap_swap(union perf_event *event,
310 event->mmap.pid = bswap_32(event->mmap.pid);
311 event->mmap.tid = bswap_32(event->mmap.tid);
312 event->mmap.start = bswap_64(event->mmap.start);
313 event->mmap.len = bswap_64(event->mmap.len);
314 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
317 void *data = &event->mmap.filename;
319 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
320 swap_sample_id_all(event, data);
324 static void perf_event__mmap2_swap(union perf_event *event,
327 event->mmap2.pid = bswap_32(event->mmap2.pid);
328 event->mmap2.tid = bswap_32(event->mmap2.tid);
329 event->mmap2.start = bswap_64(event->mmap2.start);
330 event->mmap2.len = bswap_64(event->mmap2.len);
331 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
332 event->mmap2.maj = bswap_32(event->mmap2.maj);
333 event->mmap2.min = bswap_32(event->mmap2.min);
334 event->mmap2.ino = bswap_64(event->mmap2.ino);
337 void *data = &event->mmap2.filename;
339 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
340 swap_sample_id_all(event, data);
343 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
345 event->fork.pid = bswap_32(event->fork.pid);
346 event->fork.tid = bswap_32(event->fork.tid);
347 event->fork.ppid = bswap_32(event->fork.ppid);
348 event->fork.ptid = bswap_32(event->fork.ptid);
349 event->fork.time = bswap_64(event->fork.time);
352 swap_sample_id_all(event, &event->fork + 1);
355 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
357 event->read.pid = bswap_32(event->read.pid);
358 event->read.tid = bswap_32(event->read.tid);
359 event->read.value = bswap_64(event->read.value);
360 event->read.time_enabled = bswap_64(event->read.time_enabled);
361 event->read.time_running = bswap_64(event->read.time_running);
362 event->read.id = bswap_64(event->read.id);
365 swap_sample_id_all(event, &event->read + 1);
368 static void perf_event__throttle_swap(union perf_event *event,
371 event->throttle.time = bswap_64(event->throttle.time);
372 event->throttle.id = bswap_64(event->throttle.id);
373 event->throttle.stream_id = bswap_64(event->throttle.stream_id);
376 swap_sample_id_all(event, &event->throttle + 1);
379 static u8 revbyte(u8 b)
381 int rev = (b >> 4) | ((b & 0xf) << 4);
382 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
383 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
388 * XXX this is hack in attempt to carry flags bitfield
389 * throught endian village. ABI says:
391 * Bit-fields are allocated from right to left (least to most significant)
392 * on little-endian implementations and from left to right (most to least
393 * significant) on big-endian implementations.
395 * The above seems to be byte specific, so we need to reverse each
396 * byte of the bitfield. 'Internet' also says this might be implementation
397 * specific and we probably need proper fix and carry perf_event_attr
398 * bitfield flags in separate data file FEAT_ section. Thought this seems
401 static void swap_bitfield(u8 *p, unsigned len)
405 for (i = 0; i < len; i++) {
411 /* exported for swapping attributes in file header */
412 void perf_event__attr_swap(struct perf_event_attr *attr)
414 attr->type = bswap_32(attr->type);
415 attr->size = bswap_32(attr->size);
416 attr->config = bswap_64(attr->config);
417 attr->sample_period = bswap_64(attr->sample_period);
418 attr->sample_type = bswap_64(attr->sample_type);
419 attr->read_format = bswap_64(attr->read_format);
420 attr->wakeup_events = bswap_32(attr->wakeup_events);
421 attr->bp_type = bswap_32(attr->bp_type);
422 attr->bp_addr = bswap_64(attr->bp_addr);
423 attr->bp_len = bswap_64(attr->bp_len);
424 attr->branch_sample_type = bswap_64(attr->branch_sample_type);
425 attr->sample_regs_user = bswap_64(attr->sample_regs_user);
426 attr->sample_stack_user = bswap_32(attr->sample_stack_user);
428 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
431 static void perf_event__hdr_attr_swap(union perf_event *event,
432 bool sample_id_all __maybe_unused)
436 perf_event__attr_swap(&event->attr.attr);
438 size = event->header.size;
439 size -= (void *)&event->attr.id - (void *)event;
440 mem_bswap_64(event->attr.id, size);
443 static void perf_event__event_type_swap(union perf_event *event,
444 bool sample_id_all __maybe_unused)
446 event->event_type.event_type.event_id =
447 bswap_64(event->event_type.event_type.event_id);
450 static void perf_event__tracing_data_swap(union perf_event *event,
451 bool sample_id_all __maybe_unused)
453 event->tracing_data.size = bswap_32(event->tracing_data.size);
456 typedef void (*perf_event__swap_op)(union perf_event *event,
459 static perf_event__swap_op perf_event__swap_ops[] = {
460 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
461 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap,
462 [PERF_RECORD_COMM] = perf_event__comm_swap,
463 [PERF_RECORD_FORK] = perf_event__task_swap,
464 [PERF_RECORD_EXIT] = perf_event__task_swap,
465 [PERF_RECORD_LOST] = perf_event__all64_swap,
466 [PERF_RECORD_READ] = perf_event__read_swap,
467 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap,
468 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap,
469 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
470 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
471 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
472 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
473 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
474 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap,
475 [PERF_RECORD_HEADER_MAX] = NULL,
479 * When perf record finishes a pass on every buffers, it records this pseudo
481 * We record the max timestamp t found in the pass n.
482 * Assuming these timestamps are monotonic across cpus, we know that if
483 * a buffer still has events with timestamps below t, they will be all
484 * available and then read in the pass n + 1.
485 * Hence when we start to read the pass n + 2, we can safely flush every
486 * events with timestamps below t.
488 * ============ PASS n =================
491 * cnt1 timestamps | cnt2 timestamps
494 * - | 4 <--- max recorded
496 * ============ PASS n + 1 ==============
499 * cnt1 timestamps | cnt2 timestamps
502 * 5 | 7 <---- max recorded
504 * Flush every events below timestamp 4
506 * ============ PASS n + 2 ==============
509 * cnt1 timestamps | cnt2 timestamps
514 * Flush every events below timestamp 7
517 static int process_finished_round(struct perf_tool *tool,
518 union perf_event *event __maybe_unused,
519 struct perf_session *session)
521 return ordered_events__flush(session, tool, OE_FLUSH__ROUND);
524 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
525 struct perf_tool *tool, struct perf_sample *sample,
528 struct ordered_events *oe = &s->ordered_events;
529 u64 timestamp = sample->time;
530 struct ordered_event *new;
532 if (!timestamp || timestamp == ~0ULL)
535 if (timestamp < oe->last_flush) {
536 pr_oe_time(timestamp, "out of order event\n");
537 pr_oe_time(oe->last_flush, "last flush, last_flush_type %d\n",
538 oe->last_flush_type);
540 s->evlist->stats.nr_unordered_events++;
543 new = ordered_events__new(oe, timestamp, event);
545 ordered_events__flush(s, tool, OE_FLUSH__HALF);
546 new = ordered_events__new(oe, timestamp, event);
552 new->file_offset = file_offset;
556 static void callchain__lbr_callstack_printf(struct perf_sample *sample)
558 struct ip_callchain *callchain = sample->callchain;
559 struct branch_stack *lbr_stack = sample->branch_stack;
560 u64 kernel_callchain_nr = callchain->nr;
563 for (i = 0; i < kernel_callchain_nr; i++) {
564 if (callchain->ips[i] == PERF_CONTEXT_USER)
568 if ((i != kernel_callchain_nr) && lbr_stack->nr) {
571 * LBR callstack can only get user call chain,
572 * i is kernel call chain number,
573 * 1 is PERF_CONTEXT_USER.
575 * The user call chain is stored in LBR registers.
576 * LBR are pair registers. The caller is stored
577 * in "from" register, while the callee is stored
579 * For example, there is a call stack
580 * "A"->"B"->"C"->"D".
581 * The LBR registers will recorde like
582 * "C"->"D", "B"->"C", "A"->"B".
583 * So only the first "to" register and all "from"
584 * registers are needed to construct the whole stack.
586 total_nr = i + 1 + lbr_stack->nr + 1;
587 kernel_callchain_nr = i + 1;
589 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr);
591 for (i = 0; i < kernel_callchain_nr; i++)
592 printf("..... %2d: %016" PRIx64 "\n",
593 i, callchain->ips[i]);
595 printf("..... %2d: %016" PRIx64 "\n",
596 (int)(kernel_callchain_nr), lbr_stack->entries[0].to);
597 for (i = 0; i < lbr_stack->nr; i++)
598 printf("..... %2d: %016" PRIx64 "\n",
599 (int)(i + kernel_callchain_nr + 1), lbr_stack->entries[i].from);
603 static void callchain__printf(struct perf_evsel *evsel,
604 struct perf_sample *sample)
607 struct ip_callchain *callchain = sample->callchain;
609 if (has_branch_callstack(evsel))
610 callchain__lbr_callstack_printf(sample);
612 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr);
614 for (i = 0; i < callchain->nr; i++)
615 printf("..... %2d: %016" PRIx64 "\n",
616 i, callchain->ips[i]);
619 static void branch_stack__printf(struct perf_sample *sample)
623 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
625 for (i = 0; i < sample->branch_stack->nr; i++)
626 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
627 i, sample->branch_stack->entries[i].from,
628 sample->branch_stack->entries[i].to);
631 static void regs_dump__printf(u64 mask, u64 *regs)
635 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
638 printf(".... %-5s 0x%" PRIx64 "\n",
639 perf_reg_name(rid), val);
643 static const char *regs_abi[] = {
644 [PERF_SAMPLE_REGS_ABI_NONE] = "none",
645 [PERF_SAMPLE_REGS_ABI_32] = "32-bit",
646 [PERF_SAMPLE_REGS_ABI_64] = "64-bit",
649 static inline const char *regs_dump_abi(struct regs_dump *d)
651 if (d->abi > PERF_SAMPLE_REGS_ABI_64)
654 return regs_abi[d->abi];
657 static void regs__printf(const char *type, struct regs_dump *regs)
659 u64 mask = regs->mask;
661 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n",
664 regs_dump_abi(regs));
666 regs_dump__printf(mask, regs->regs);
669 static void regs_user__printf(struct perf_sample *sample)
671 struct regs_dump *user_regs = &sample->user_regs;
674 regs__printf("user", user_regs);
677 static void regs_intr__printf(struct perf_sample *sample)
679 struct regs_dump *intr_regs = &sample->intr_regs;
682 regs__printf("intr", intr_regs);
685 static void stack_user__printf(struct stack_dump *dump)
687 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
688 dump->size, dump->offset);
691 static void perf_session__print_tstamp(struct perf_session *session,
692 union perf_event *event,
693 struct perf_sample *sample)
695 u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
697 if (event->header.type != PERF_RECORD_SAMPLE &&
698 !perf_evlist__sample_id_all(session->evlist)) {
699 fputs("-1 -1 ", stdout);
703 if ((sample_type & PERF_SAMPLE_CPU))
704 printf("%u ", sample->cpu);
706 if (sample_type & PERF_SAMPLE_TIME)
707 printf("%" PRIu64 " ", sample->time);
710 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
712 printf("... sample_read:\n");
714 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
715 printf("...... time enabled %016" PRIx64 "\n",
716 sample->read.time_enabled);
718 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
719 printf("...... time running %016" PRIx64 "\n",
720 sample->read.time_running);
722 if (read_format & PERF_FORMAT_GROUP) {
725 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
727 for (i = 0; i < sample->read.group.nr; i++) {
728 struct sample_read_value *value;
730 value = &sample->read.group.values[i];
731 printf("..... id %016" PRIx64
732 ", value %016" PRIx64 "\n",
733 value->id, value->value);
736 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
737 sample->read.one.id, sample->read.one.value);
740 static void dump_event(struct perf_session *session, union perf_event *event,
741 u64 file_offset, struct perf_sample *sample)
746 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
747 file_offset, event->header.size, event->header.type);
752 perf_session__print_tstamp(session, event, sample);
754 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
755 event->header.size, perf_event__name(event->header.type));
758 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
759 struct perf_sample *sample)
766 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
767 event->header.misc, sample->pid, sample->tid, sample->ip,
768 sample->period, sample->addr);
770 sample_type = evsel->attr.sample_type;
772 if (sample_type & PERF_SAMPLE_CALLCHAIN)
773 callchain__printf(evsel, sample);
775 if ((sample_type & PERF_SAMPLE_BRANCH_STACK) && !has_branch_callstack(evsel))
776 branch_stack__printf(sample);
778 if (sample_type & PERF_SAMPLE_REGS_USER)
779 regs_user__printf(sample);
781 if (sample_type & PERF_SAMPLE_REGS_INTR)
782 regs_intr__printf(sample);
784 if (sample_type & PERF_SAMPLE_STACK_USER)
785 stack_user__printf(&sample->user_stack);
787 if (sample_type & PERF_SAMPLE_WEIGHT)
788 printf("... weight: %" PRIu64 "\n", sample->weight);
790 if (sample_type & PERF_SAMPLE_DATA_SRC)
791 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
793 if (sample_type & PERF_SAMPLE_TRANSACTION)
794 printf("... transaction: %" PRIx64 "\n", sample->transaction);
796 if (sample_type & PERF_SAMPLE_READ)
797 sample_read__printf(sample, evsel->attr.read_format);
800 static struct machine *machines__find_for_cpumode(struct machines *machines,
801 union perf_event *event,
802 struct perf_sample *sample)
804 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
805 struct machine *machine;
808 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
809 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
812 if (event->header.type == PERF_RECORD_MMAP
813 || event->header.type == PERF_RECORD_MMAP2)
814 pid = event->mmap.pid;
818 machine = machines__find(machines, pid);
820 machine = machines__find(machines, DEFAULT_GUEST_KERNEL_ID);
824 return &machines->host;
827 static int deliver_sample_value(struct perf_evlist *evlist,
828 struct perf_tool *tool,
829 union perf_event *event,
830 struct perf_sample *sample,
831 struct sample_read_value *v,
832 struct machine *machine)
834 struct perf_sample_id *sid = perf_evlist__id2sid(evlist, v->id);
838 sample->period = v->value - sid->period;
839 sid->period = v->value;
842 if (!sid || sid->evsel == NULL) {
843 ++evlist->stats.nr_unknown_id;
847 return tool->sample(tool, event, sample, sid->evsel, machine);
850 static int deliver_sample_group(struct perf_evlist *evlist,
851 struct perf_tool *tool,
852 union perf_event *event,
853 struct perf_sample *sample,
854 struct machine *machine)
859 for (i = 0; i < sample->read.group.nr; i++) {
860 ret = deliver_sample_value(evlist, tool, event, sample,
861 &sample->read.group.values[i],
871 perf_evlist__deliver_sample(struct perf_evlist *evlist,
872 struct perf_tool *tool,
873 union perf_event *event,
874 struct perf_sample *sample,
875 struct perf_evsel *evsel,
876 struct machine *machine)
878 /* We know evsel != NULL. */
879 u64 sample_type = evsel->attr.sample_type;
880 u64 read_format = evsel->attr.read_format;
882 /* Standard sample delievery. */
883 if (!(sample_type & PERF_SAMPLE_READ))
884 return tool->sample(tool, event, sample, evsel, machine);
886 /* For PERF_SAMPLE_READ we have either single or group mode. */
887 if (read_format & PERF_FORMAT_GROUP)
888 return deliver_sample_group(evlist, tool, event, sample,
891 return deliver_sample_value(evlist, tool, event, sample,
892 &sample->read.one, machine);
895 int perf_session__deliver_event(struct perf_session *session,
896 union perf_event *event,
897 struct perf_sample *sample,
898 struct perf_tool *tool, u64 file_offset)
900 struct perf_evlist *evlist = session->evlist;
901 struct perf_evsel *evsel;
902 struct machine *machine;
904 dump_event(session, event, file_offset, sample);
906 evsel = perf_evlist__id2evsel(evlist, sample->id);
908 machine = machines__find_for_cpumode(&session->machines, event, sample);
910 switch (event->header.type) {
911 case PERF_RECORD_SAMPLE:
912 dump_sample(evsel, event, sample);
914 ++evlist->stats.nr_unknown_id;
917 if (machine == NULL) {
918 ++evlist->stats.nr_unprocessable_samples;
921 return perf_evlist__deliver_sample(evlist, tool, event, sample, evsel, machine);
922 case PERF_RECORD_MMAP:
923 return tool->mmap(tool, event, sample, machine);
924 case PERF_RECORD_MMAP2:
925 return tool->mmap2(tool, event, sample, machine);
926 case PERF_RECORD_COMM:
927 return tool->comm(tool, event, sample, machine);
928 case PERF_RECORD_FORK:
929 return tool->fork(tool, event, sample, machine);
930 case PERF_RECORD_EXIT:
931 return tool->exit(tool, event, sample, machine);
932 case PERF_RECORD_LOST:
933 if (tool->lost == perf_event__process_lost)
934 evlist->stats.total_lost += event->lost.lost;
935 return tool->lost(tool, event, sample, machine);
936 case PERF_RECORD_READ:
937 return tool->read(tool, event, sample, evsel, machine);
938 case PERF_RECORD_THROTTLE:
939 return tool->throttle(tool, event, sample, machine);
940 case PERF_RECORD_UNTHROTTLE:
941 return tool->unthrottle(tool, event, sample, machine);
943 ++evlist->stats.nr_unknown_events;
948 static s64 perf_session__process_user_event(struct perf_session *session,
949 union perf_event *event,
950 struct perf_tool *tool,
953 int fd = perf_data_file__fd(session->file);
956 dump_event(session, event, file_offset, NULL);
958 /* These events are processed right away */
959 switch (event->header.type) {
960 case PERF_RECORD_HEADER_ATTR:
961 err = tool->attr(tool, event, &session->evlist);
963 perf_session__set_id_hdr_size(session);
964 perf_session__set_comm_exec(session);
967 case PERF_RECORD_HEADER_EVENT_TYPE:
969 * Depreceated, but we need to handle it for sake
970 * of old data files create in pipe mode.
973 case PERF_RECORD_HEADER_TRACING_DATA:
974 /* setup for reading amidst mmap */
975 lseek(fd, file_offset, SEEK_SET);
976 return tool->tracing_data(tool, event, session);
977 case PERF_RECORD_HEADER_BUILD_ID:
978 return tool->build_id(tool, event, session);
979 case PERF_RECORD_FINISHED_ROUND:
980 return tool->finished_round(tool, event, session);
981 case PERF_RECORD_ID_INDEX:
982 return tool->id_index(tool, event, session);
988 int perf_session__deliver_synth_event(struct perf_session *session,
989 union perf_event *event,
990 struct perf_sample *sample,
991 struct perf_tool *tool)
993 events_stats__inc(&session->evlist->stats, event->header.type);
995 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
996 return perf_session__process_user_event(session, event, tool, 0);
998 return perf_session__deliver_event(session, event, sample, tool, 0);
1001 static void event_swap(union perf_event *event, bool sample_id_all)
1003 perf_event__swap_op swap;
1005 swap = perf_event__swap_ops[event->header.type];
1007 swap(event, sample_id_all);
1010 int perf_session__peek_event(struct perf_session *session, off_t file_offset,
1011 void *buf, size_t buf_sz,
1012 union perf_event **event_ptr,
1013 struct perf_sample *sample)
1015 union perf_event *event;
1016 size_t hdr_sz, rest;
1019 if (session->one_mmap && !session->header.needs_swap) {
1020 event = file_offset - session->one_mmap_offset +
1021 session->one_mmap_addr;
1022 goto out_parse_sample;
1025 if (perf_data_file__is_pipe(session->file))
1028 fd = perf_data_file__fd(session->file);
1029 hdr_sz = sizeof(struct perf_event_header);
1031 if (buf_sz < hdr_sz)
1034 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 ||
1035 readn(fd, &buf, hdr_sz) != (ssize_t)hdr_sz)
1038 event = (union perf_event *)buf;
1040 if (session->header.needs_swap)
1041 perf_event_header__bswap(&event->header);
1043 if (event->header.size < hdr_sz)
1046 rest = event->header.size - hdr_sz;
1048 if (readn(fd, &buf, rest) != (ssize_t)rest)
1051 if (session->header.needs_swap)
1052 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1056 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START &&
1057 perf_evlist__parse_sample(session->evlist, event, sample))
1065 static s64 perf_session__process_event(struct perf_session *session,
1066 union perf_event *event,
1067 struct perf_tool *tool,
1070 struct perf_evlist *evlist = session->evlist;
1071 struct perf_sample sample;
1074 if (session->header.needs_swap)
1075 event_swap(event, perf_evlist__sample_id_all(evlist));
1077 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1080 events_stats__inc(&evlist->stats, event->header.type);
1082 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1083 return perf_session__process_user_event(session, event, tool, file_offset);
1086 * For all kernel events we get the sample data
1088 ret = perf_evlist__parse_sample(evlist, event, &sample);
1092 if (tool->ordered_events) {
1093 ret = perf_session_queue_event(session, event, tool, &sample,
1099 return perf_session__deliver_event(session, event, &sample, tool,
1103 void perf_event_header__bswap(struct perf_event_header *hdr)
1105 hdr->type = bswap_32(hdr->type);
1106 hdr->misc = bswap_16(hdr->misc);
1107 hdr->size = bswap_16(hdr->size);
1110 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1112 return machine__findnew_thread(&session->machines.host, -1, pid);
1115 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1117 struct thread *thread;
1119 thread = machine__findnew_thread(&session->machines.host, 0, 0);
1120 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1121 pr_err("problem inserting idle task.\n");
1128 static void perf_tool__warn_about_errors(const struct perf_tool *tool,
1129 const struct events_stats *stats)
1131 if (tool->lost == perf_event__process_lost &&
1132 stats->nr_events[PERF_RECORD_LOST] != 0) {
1133 ui__warning("Processed %d events and lost %d chunks!\n\n"
1134 "Check IO/CPU overload!\n\n",
1135 stats->nr_events[0],
1136 stats->nr_events[PERF_RECORD_LOST]);
1139 if (stats->nr_unknown_events != 0) {
1140 ui__warning("Found %u unknown events!\n\n"
1141 "Is this an older tool processing a perf.data "
1142 "file generated by a more recent tool?\n\n"
1143 "If that is not the case, consider "
1144 "reporting to linux-kernel@vger.kernel.org.\n\n",
1145 stats->nr_unknown_events);
1148 if (stats->nr_unknown_id != 0) {
1149 ui__warning("%u samples with id not present in the header\n",
1150 stats->nr_unknown_id);
1153 if (stats->nr_invalid_chains != 0) {
1154 ui__warning("Found invalid callchains!\n\n"
1155 "%u out of %u events were discarded for this reason.\n\n"
1156 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1157 stats->nr_invalid_chains,
1158 stats->nr_events[PERF_RECORD_SAMPLE]);
1161 if (stats->nr_unprocessable_samples != 0) {
1162 ui__warning("%u unprocessable samples recorded.\n"
1163 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1164 stats->nr_unprocessable_samples);
1167 if (stats->nr_unordered_events != 0)
1168 ui__warning("%u out of order events recorded.\n", stats->nr_unordered_events);
1171 volatile int session_done;
1173 static int __perf_session__process_pipe_events(struct perf_session *session,
1174 struct perf_tool *tool)
1176 int fd = perf_data_file__fd(session->file);
1177 union perf_event *event;
1178 uint32_t size, cur_size = 0;
1185 perf_tool__fill_defaults(tool);
1188 cur_size = sizeof(union perf_event);
1190 buf = malloc(cur_size);
1195 err = readn(fd, event, sizeof(struct perf_event_header));
1200 pr_err("failed to read event header\n");
1204 if (session->header.needs_swap)
1205 perf_event_header__bswap(&event->header);
1207 size = event->header.size;
1208 if (size < sizeof(struct perf_event_header)) {
1209 pr_err("bad event header size\n");
1213 if (size > cur_size) {
1214 void *new = realloc(buf, size);
1216 pr_err("failed to allocate memory to read event\n");
1224 p += sizeof(struct perf_event_header);
1226 if (size - sizeof(struct perf_event_header)) {
1227 err = readn(fd, p, size - sizeof(struct perf_event_header));
1230 pr_err("unexpected end of event stream\n");
1234 pr_err("failed to read event data\n");
1239 if ((skip = perf_session__process_event(session, event, tool, head)) < 0) {
1240 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1241 head, event->header.size, event->header.type);
1251 if (!session_done())
1254 /* do the final flush for ordered samples */
1255 err = ordered_events__flush(session, tool, OE_FLUSH__FINAL);
1258 perf_tool__warn_about_errors(tool, &session->evlist->stats);
1259 ordered_events__free(&session->ordered_events);
1263 static union perf_event *
1264 fetch_mmaped_event(struct perf_session *session,
1265 u64 head, size_t mmap_size, char *buf)
1267 union perf_event *event;
1270 * Ensure we have enough space remaining to read
1271 * the size of the event in the headers.
1273 if (head + sizeof(event->header) > mmap_size)
1276 event = (union perf_event *)(buf + head);
1278 if (session->header.needs_swap)
1279 perf_event_header__bswap(&event->header);
1281 if (head + event->header.size > mmap_size) {
1282 /* We're not fetching the event so swap back again */
1283 if (session->header.needs_swap)
1284 perf_event_header__bswap(&event->header);
1292 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1293 * slices. On 32bit we use 32MB.
1295 #if BITS_PER_LONG == 64
1296 #define MMAP_SIZE ULLONG_MAX
1299 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1300 #define NUM_MMAPS 128
1303 static int __perf_session__process_events(struct perf_session *session,
1304 u64 data_offset, u64 data_size,
1305 u64 file_size, struct perf_tool *tool)
1307 int fd = perf_data_file__fd(session->file);
1308 u64 head, page_offset, file_offset, file_pos, size;
1309 int err, mmap_prot, mmap_flags, map_idx = 0;
1311 char *buf, *mmaps[NUM_MMAPS];
1312 union perf_event *event;
1313 struct ui_progress prog;
1316 perf_tool__fill_defaults(tool);
1318 page_offset = page_size * (data_offset / page_size);
1319 file_offset = page_offset;
1320 head = data_offset - page_offset;
1322 if (data_size && (data_offset + data_size < file_size))
1323 file_size = data_offset + data_size;
1325 ui_progress__init(&prog, file_size, "Processing events...");
1327 mmap_size = MMAP_SIZE;
1328 if (mmap_size > file_size) {
1329 mmap_size = file_size;
1330 session->one_mmap = true;
1333 memset(mmaps, 0, sizeof(mmaps));
1335 mmap_prot = PROT_READ;
1336 mmap_flags = MAP_SHARED;
1338 if (session->header.needs_swap) {
1339 mmap_prot |= PROT_WRITE;
1340 mmap_flags = MAP_PRIVATE;
1343 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1345 if (buf == MAP_FAILED) {
1346 pr_err("failed to mmap file\n");
1350 mmaps[map_idx] = buf;
1351 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1352 file_pos = file_offset + head;
1353 if (session->one_mmap) {
1354 session->one_mmap_addr = buf;
1355 session->one_mmap_offset = file_offset;
1359 event = fetch_mmaped_event(session, head, mmap_size, buf);
1361 if (mmaps[map_idx]) {
1362 munmap(mmaps[map_idx], mmap_size);
1363 mmaps[map_idx] = NULL;
1366 page_offset = page_size * (head / page_size);
1367 file_offset += page_offset;
1368 head -= page_offset;
1372 size = event->header.size;
1374 if (size < sizeof(struct perf_event_header) ||
1375 (skip = perf_session__process_event(session, event, tool, file_pos))
1377 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1378 file_offset + head, event->header.size,
1379 event->header.type);
1390 ui_progress__update(&prog, size);
1395 if (file_pos < file_size)
1399 /* do the final flush for ordered samples */
1400 err = ordered_events__flush(session, tool, OE_FLUSH__FINAL);
1402 ui_progress__finish();
1403 perf_tool__warn_about_errors(tool, &session->evlist->stats);
1404 ordered_events__free(&session->ordered_events);
1405 session->one_mmap = false;
1409 int perf_session__process_events(struct perf_session *session,
1410 struct perf_tool *tool)
1412 u64 size = perf_data_file__size(session->file);
1415 if (perf_session__register_idle_thread(session) == NULL)
1418 if (!perf_data_file__is_pipe(session->file))
1419 err = __perf_session__process_events(session,
1420 session->header.data_offset,
1421 session->header.data_size,
1424 err = __perf_session__process_pipe_events(session, tool);
1429 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1431 struct perf_evsel *evsel;
1433 evlist__for_each(session->evlist, evsel) {
1434 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1438 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1442 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1443 const char *symbol_name, u64 addr)
1447 struct ref_reloc_sym *ref;
1449 ref = zalloc(sizeof(struct ref_reloc_sym));
1453 ref->name = strdup(symbol_name);
1454 if (ref->name == NULL) {
1459 bracket = strchr(ref->name, ']');
1465 for (i = 0; i < MAP__NR_TYPES; ++i) {
1466 struct kmap *kmap = map__kmap(maps[i]);
1467 kmap->ref_reloc_sym = ref;
1473 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1475 return machines__fprintf_dsos(&session->machines, fp);
1478 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1479 bool (skip)(struct dso *dso, int parm), int parm)
1481 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1484 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1486 size_t ret = fprintf(fp, "Aggregated stats:\n");
1488 ret += events_stats__fprintf(&session->evlist->stats, fp);
1492 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1495 * FIXME: Here we have to actually print all the machines in this
1496 * session, not just the host...
1498 return machine__fprintf(&session->machines.host, fp);
1501 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1504 struct perf_evsel *pos;
1506 evlist__for_each(session->evlist, pos) {
1507 if (pos->attr.type == type)
1513 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1514 struct addr_location *al,
1515 unsigned int print_opts, unsigned int stack_depth)
1517 struct callchain_cursor_node *node;
1518 int print_ip = print_opts & PRINT_IP_OPT_IP;
1519 int print_sym = print_opts & PRINT_IP_OPT_SYM;
1520 int print_dso = print_opts & PRINT_IP_OPT_DSO;
1521 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1522 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1523 int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1524 char s = print_oneline ? ' ' : '\t';
1526 if (symbol_conf.use_callchain && sample->callchain) {
1527 struct addr_location node_al;
1529 if (thread__resolve_callchain(al->thread, evsel,
1531 PERF_MAX_STACK_DEPTH) != 0) {
1533 error("Failed to resolve callchain. Skipping\n");
1536 callchain_cursor_commit(&callchain_cursor);
1538 if (print_symoffset)
1541 while (stack_depth) {
1544 node = callchain_cursor_current(&callchain_cursor);
1548 if (node->sym && node->sym->ignore)
1552 printf("%c%16" PRIx64, s, node->ip);
1555 addr = node->map->map_ip(node->map, node->ip);
1559 if (print_symoffset) {
1560 node_al.addr = addr;
1561 node_al.map = node->map;
1562 symbol__fprintf_symname_offs(node->sym, &node_al, stdout);
1564 symbol__fprintf_symname(node->sym, stdout);
1569 map__fprintf_dsoname(node->map, stdout);
1574 map__fprintf_srcline(node->map, addr, "\n ",
1582 callchain_cursor_advance(&callchain_cursor);
1586 if (al->sym && al->sym->ignore)
1590 printf("%16" PRIx64, sample->ip);
1594 if (print_symoffset)
1595 symbol__fprintf_symname_offs(al->sym, al,
1598 symbol__fprintf_symname(al->sym, stdout);
1603 map__fprintf_dsoname(al->map, stdout);
1608 map__fprintf_srcline(al->map, al->addr, "\n ", stdout);
1612 int perf_session__cpu_bitmap(struct perf_session *session,
1613 const char *cpu_list, unsigned long *cpu_bitmap)
1616 struct cpu_map *map;
1618 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1619 struct perf_evsel *evsel;
1621 evsel = perf_session__find_first_evtype(session, i);
1625 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1626 pr_err("File does not contain CPU events. "
1627 "Remove -c option to proceed.\n");
1632 map = cpu_map__new(cpu_list);
1634 pr_err("Invalid cpu_list\n");
1638 for (i = 0; i < map->nr; i++) {
1639 int cpu = map->map[i];
1641 if (cpu >= MAX_NR_CPUS) {
1642 pr_err("Requested CPU %d too large. "
1643 "Consider raising MAX_NR_CPUS\n", cpu);
1644 goto out_delete_map;
1647 set_bit(cpu, cpu_bitmap);
1653 cpu_map__delete(map);
1657 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1663 if (session == NULL || fp == NULL)
1666 fd = perf_data_file__fd(session->file);
1668 ret = fstat(fd, &st);
1672 fprintf(fp, "# ========\n");
1673 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1674 perf_header__fprintf_info(session, fp, full);
1675 fprintf(fp, "# ========\n#\n");
1679 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1680 const struct perf_evsel_str_handler *assocs,
1683 struct perf_evsel *evsel;
1687 for (i = 0; i < nr_assocs; i++) {
1689 * Adding a handler for an event not in the session,
1692 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1697 if (evsel->handler != NULL)
1699 evsel->handler = assocs[i].handler;
1707 int perf_event__process_id_index(struct perf_tool *tool __maybe_unused,
1708 union perf_event *event,
1709 struct perf_session *session)
1711 struct perf_evlist *evlist = session->evlist;
1712 struct id_index_event *ie = &event->id_index;
1713 size_t i, nr, max_nr;
1715 max_nr = (ie->header.size - sizeof(struct id_index_event)) /
1716 sizeof(struct id_index_entry);
1722 fprintf(stdout, " nr: %zu\n", nr);
1724 for (i = 0; i < nr; i++) {
1725 struct id_index_entry *e = &ie->entries[i];
1726 struct perf_sample_id *sid;
1729 fprintf(stdout, " ... id: %"PRIu64, e->id);
1730 fprintf(stdout, " idx: %"PRIu64, e->idx);
1731 fprintf(stdout, " cpu: %"PRId64, e->cpu);
1732 fprintf(stdout, " tid: %"PRId64"\n", e->tid);
1735 sid = perf_evlist__id2sid(evlist, e->id);
1745 int perf_event__synthesize_id_index(struct perf_tool *tool,
1746 perf_event__handler_t process,
1747 struct perf_evlist *evlist,
1748 struct machine *machine)
1750 union perf_event *ev;
1751 struct perf_evsel *evsel;
1752 size_t nr = 0, i = 0, sz, max_nr, n;
1755 pr_debug2("Synthesizing id index\n");
1757 max_nr = (UINT16_MAX - sizeof(struct id_index_event)) /
1758 sizeof(struct id_index_entry);
1760 evlist__for_each(evlist, evsel)
1763 n = nr > max_nr ? max_nr : nr;
1764 sz = sizeof(struct id_index_event) + n * sizeof(struct id_index_entry);
1769 ev->id_index.header.type = PERF_RECORD_ID_INDEX;
1770 ev->id_index.header.size = sz;
1771 ev->id_index.nr = n;
1773 evlist__for_each(evlist, evsel) {
1776 for (j = 0; j < evsel->ids; j++) {
1777 struct id_index_entry *e;
1778 struct perf_sample_id *sid;
1781 err = process(tool, ev, NULL, machine);
1788 e = &ev->id_index.entries[i++];
1790 e->id = evsel->id[j];
1792 sid = perf_evlist__id2sid(evlist, e->id);
1804 sz = sizeof(struct id_index_event) + nr * sizeof(struct id_index_entry);
1805 ev->id_index.header.size = sz;
1806 ev->id_index.nr = nr;
1808 err = process(tool, ev, NULL, machine);