1 // SPDX-License-Identifier: GPL-2.0
4 #include <linux/kernel.h>
5 #include <linux/zalloc.h>
6 #include <traceevent/event-parse.h>
11 #include <sys/types.h>
23 #include "perf_regs.h"
27 #include "thread-stack.h"
28 #include "sample-raw.h"
30 #include "arch/common.h"
32 #ifdef HAVE_ZSTD_SUPPORT
33 static int perf_session__process_compressed_event(struct perf_session *session,
34 union perf_event *event, u64 file_offset)
37 size_t decomp_size, src_size;
38 u64 decomp_last_rem = 0;
39 size_t mmap_len, decomp_len = session->header.env.comp_mmap_len;
40 struct decomp *decomp, *decomp_last = session->decomp_last;
43 decomp_last_rem = decomp_last->size - decomp_last->head;
44 decomp_len += decomp_last_rem;
47 mmap_len = sizeof(struct decomp) + decomp_len;
48 decomp = mmap(NULL, mmap_len, PROT_READ|PROT_WRITE,
49 MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
50 if (decomp == MAP_FAILED) {
51 pr_err("Couldn't allocate memory for decompression\n");
55 decomp->file_pos = file_offset;
56 decomp->mmap_len = mmap_len;
59 if (decomp_last_rem) {
60 memcpy(decomp->data, &(decomp_last->data[decomp_last->head]), decomp_last_rem);
61 decomp->size = decomp_last_rem;
64 src = (void *)event + sizeof(struct compressed_event);
65 src_size = event->pack.header.size - sizeof(struct compressed_event);
67 decomp_size = zstd_decompress_stream(&(session->zstd_data), src, src_size,
68 &(decomp->data[decomp_last_rem]), decomp_len - decomp_last_rem);
70 munmap(decomp, mmap_len);
71 pr_err("Couldn't decompress data\n");
75 decomp->size += decomp_size;
77 if (session->decomp == NULL) {
78 session->decomp = decomp;
79 session->decomp_last = decomp;
81 session->decomp_last->next = decomp;
82 session->decomp_last = decomp;
85 pr_debug("decomp (B): %ld to %ld\n", src_size, decomp_size);
89 #else /* !HAVE_ZSTD_SUPPORT */
90 #define perf_session__process_compressed_event perf_session__process_compressed_event_stub
93 static int perf_session__deliver_event(struct perf_session *session,
94 union perf_event *event,
95 struct perf_tool *tool,
98 static int perf_session__open(struct perf_session *session)
100 struct perf_data *data = session->data;
102 if (perf_session__read_header(session) < 0) {
103 pr_err("incompatible file format (rerun with -v to learn more)\n");
107 if (perf_data__is_pipe(data))
110 if (perf_header__has_feat(&session->header, HEADER_STAT))
113 if (!perf_evlist__valid_sample_type(session->evlist)) {
114 pr_err("non matching sample_type\n");
118 if (!perf_evlist__valid_sample_id_all(session->evlist)) {
119 pr_err("non matching sample_id_all\n");
123 if (!perf_evlist__valid_read_format(session->evlist)) {
124 pr_err("non matching read_format\n");
131 void perf_session__set_id_hdr_size(struct perf_session *session)
133 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
135 machines__set_id_hdr_size(&session->machines, id_hdr_size);
138 int perf_session__create_kernel_maps(struct perf_session *session)
140 int ret = machine__create_kernel_maps(&session->machines.host);
143 ret = machines__create_guest_kernel_maps(&session->machines);
147 static void perf_session__destroy_kernel_maps(struct perf_session *session)
149 machines__destroy_kernel_maps(&session->machines);
152 static bool perf_session__has_comm_exec(struct perf_session *session)
156 evlist__for_each_entry(session->evlist, evsel) {
157 if (evsel->attr.comm_exec)
164 static void perf_session__set_comm_exec(struct perf_session *session)
166 bool comm_exec = perf_session__has_comm_exec(session);
168 machines__set_comm_exec(&session->machines, comm_exec);
171 static int ordered_events__deliver_event(struct ordered_events *oe,
172 struct ordered_event *event)
174 struct perf_session *session = container_of(oe, struct perf_session,
177 return perf_session__deliver_event(session, event->event,
178 session->tool, event->file_offset);
181 struct perf_session *perf_session__new(struct perf_data *data,
182 bool repipe, struct perf_tool *tool)
184 struct perf_session *session = zalloc(sizeof(*session));
189 session->repipe = repipe;
190 session->tool = tool;
191 INIT_LIST_HEAD(&session->auxtrace_index);
192 machines__init(&session->machines);
193 ordered_events__init(&session->ordered_events,
194 ordered_events__deliver_event, NULL);
196 perf_env__init(&session->header.env);
198 if (perf_data__open(data))
201 session->data = data;
203 if (perf_data__is_read(data)) {
204 if (perf_session__open(session) < 0)
208 * set session attributes that are present in perf.data
209 * but not in pipe-mode.
211 if (!data->is_pipe) {
212 perf_session__set_id_hdr_size(session);
213 perf_session__set_comm_exec(session);
216 perf_evlist__init_trace_event_sample_raw(session->evlist);
218 /* Open the directory data. */
219 if (data->is_dir && perf_data__open_dir(data))
223 session->machines.host.env = &perf_env;
226 session->machines.host.single_address_space =
227 perf_env__single_address_space(session->machines.host.env);
229 if (!data || perf_data__is_write(data)) {
231 * In O_RDONLY mode this will be performed when reading the
232 * kernel MMAP event, in perf_event__process_mmap().
234 if (perf_session__create_kernel_maps(session) < 0)
235 pr_warning("Cannot read kernel map\n");
239 * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is
240 * processed, so perf_evlist__sample_id_all is not meaningful here.
242 if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps &&
243 tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) {
244 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
245 tool->ordered_events = false;
251 perf_session__delete(session);
256 static void perf_session__delete_threads(struct perf_session *session)
258 machine__delete_threads(&session->machines.host);
261 static void perf_session__release_decomp_events(struct perf_session *session)
263 struct decomp *next, *decomp;
265 next = session->decomp;
271 mmap_len = decomp->mmap_len;
272 munmap(decomp, mmap_len);
276 void perf_session__delete(struct perf_session *session)
280 auxtrace__free(session);
281 auxtrace_index__free(&session->auxtrace_index);
282 perf_session__destroy_kernel_maps(session);
283 perf_session__delete_threads(session);
284 perf_session__release_decomp_events(session);
285 perf_env__exit(&session->header.env);
286 machines__exit(&session->machines);
288 perf_data__close(session->data);
292 static int process_event_synth_tracing_data_stub(struct perf_session *session
294 union perf_event *event
297 dump_printf(": unhandled!\n");
301 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
302 union perf_event *event __maybe_unused,
303 struct evlist **pevlist
306 dump_printf(": unhandled!\n");
310 static int process_event_synth_event_update_stub(struct perf_tool *tool __maybe_unused,
311 union perf_event *event __maybe_unused,
312 struct evlist **pevlist
316 perf_event__fprintf_event_update(event, stdout);
318 dump_printf(": unhandled!\n");
322 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
323 union perf_event *event __maybe_unused,
324 struct perf_sample *sample __maybe_unused,
325 struct evsel *evsel __maybe_unused,
326 struct machine *machine __maybe_unused)
328 dump_printf(": unhandled!\n");
332 static int process_event_stub(struct perf_tool *tool __maybe_unused,
333 union perf_event *event __maybe_unused,
334 struct perf_sample *sample __maybe_unused,
335 struct machine *machine __maybe_unused)
337 dump_printf(": unhandled!\n");
341 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
342 union perf_event *event __maybe_unused,
343 struct ordered_events *oe __maybe_unused)
345 dump_printf(": unhandled!\n");
349 static int process_finished_round(struct perf_tool *tool,
350 union perf_event *event,
351 struct ordered_events *oe);
353 static int skipn(int fd, off_t n)
359 ret = read(fd, buf, min(n, (off_t)sizeof(buf)));
368 static s64 process_event_auxtrace_stub(struct perf_session *session __maybe_unused,
369 union perf_event *event)
371 dump_printf(": unhandled!\n");
372 if (perf_data__is_pipe(session->data))
373 skipn(perf_data__fd(session->data), event->auxtrace.size);
374 return event->auxtrace.size;
377 static int process_event_op2_stub(struct perf_session *session __maybe_unused,
378 union perf_event *event __maybe_unused)
380 dump_printf(": unhandled!\n");
386 int process_event_thread_map_stub(struct perf_session *session __maybe_unused,
387 union perf_event *event __maybe_unused)
390 perf_event__fprintf_thread_map(event, stdout);
392 dump_printf(": unhandled!\n");
397 int process_event_cpu_map_stub(struct perf_session *session __maybe_unused,
398 union perf_event *event __maybe_unused)
401 perf_event__fprintf_cpu_map(event, stdout);
403 dump_printf(": unhandled!\n");
408 int process_event_stat_config_stub(struct perf_session *session __maybe_unused,
409 union perf_event *event __maybe_unused)
412 perf_event__fprintf_stat_config(event, stdout);
414 dump_printf(": unhandled!\n");
418 static int process_stat_stub(struct perf_session *perf_session __maybe_unused,
419 union perf_event *event)
422 perf_event__fprintf_stat(event, stdout);
424 dump_printf(": unhandled!\n");
428 static int process_stat_round_stub(struct perf_session *perf_session __maybe_unused,
429 union perf_event *event)
432 perf_event__fprintf_stat_round(event, stdout);
434 dump_printf(": unhandled!\n");
438 static int perf_session__process_compressed_event_stub(struct perf_session *session __maybe_unused,
439 union perf_event *event __maybe_unused,
440 u64 file_offset __maybe_unused)
442 dump_printf(": unhandled!\n");
446 void perf_tool__fill_defaults(struct perf_tool *tool)
448 if (tool->sample == NULL)
449 tool->sample = process_event_sample_stub;
450 if (tool->mmap == NULL)
451 tool->mmap = process_event_stub;
452 if (tool->mmap2 == NULL)
453 tool->mmap2 = process_event_stub;
454 if (tool->comm == NULL)
455 tool->comm = process_event_stub;
456 if (tool->namespaces == NULL)
457 tool->namespaces = process_event_stub;
458 if (tool->fork == NULL)
459 tool->fork = process_event_stub;
460 if (tool->exit == NULL)
461 tool->exit = process_event_stub;
462 if (tool->lost == NULL)
463 tool->lost = perf_event__process_lost;
464 if (tool->lost_samples == NULL)
465 tool->lost_samples = perf_event__process_lost_samples;
466 if (tool->aux == NULL)
467 tool->aux = perf_event__process_aux;
468 if (tool->itrace_start == NULL)
469 tool->itrace_start = perf_event__process_itrace_start;
470 if (tool->context_switch == NULL)
471 tool->context_switch = perf_event__process_switch;
472 if (tool->ksymbol == NULL)
473 tool->ksymbol = perf_event__process_ksymbol;
474 if (tool->bpf_event == NULL)
475 tool->bpf_event = perf_event__process_bpf_event;
476 if (tool->read == NULL)
477 tool->read = process_event_sample_stub;
478 if (tool->throttle == NULL)
479 tool->throttle = process_event_stub;
480 if (tool->unthrottle == NULL)
481 tool->unthrottle = process_event_stub;
482 if (tool->attr == NULL)
483 tool->attr = process_event_synth_attr_stub;
484 if (tool->event_update == NULL)
485 tool->event_update = process_event_synth_event_update_stub;
486 if (tool->tracing_data == NULL)
487 tool->tracing_data = process_event_synth_tracing_data_stub;
488 if (tool->build_id == NULL)
489 tool->build_id = process_event_op2_stub;
490 if (tool->finished_round == NULL) {
491 if (tool->ordered_events)
492 tool->finished_round = process_finished_round;
494 tool->finished_round = process_finished_round_stub;
496 if (tool->id_index == NULL)
497 tool->id_index = process_event_op2_stub;
498 if (tool->auxtrace_info == NULL)
499 tool->auxtrace_info = process_event_op2_stub;
500 if (tool->auxtrace == NULL)
501 tool->auxtrace = process_event_auxtrace_stub;
502 if (tool->auxtrace_error == NULL)
503 tool->auxtrace_error = process_event_op2_stub;
504 if (tool->thread_map == NULL)
505 tool->thread_map = process_event_thread_map_stub;
506 if (tool->cpu_map == NULL)
507 tool->cpu_map = process_event_cpu_map_stub;
508 if (tool->stat_config == NULL)
509 tool->stat_config = process_event_stat_config_stub;
510 if (tool->stat == NULL)
511 tool->stat = process_stat_stub;
512 if (tool->stat_round == NULL)
513 tool->stat_round = process_stat_round_stub;
514 if (tool->time_conv == NULL)
515 tool->time_conv = process_event_op2_stub;
516 if (tool->feature == NULL)
517 tool->feature = process_event_op2_stub;
518 if (tool->compressed == NULL)
519 tool->compressed = perf_session__process_compressed_event;
522 static void swap_sample_id_all(union perf_event *event, void *data)
524 void *end = (void *) event + event->header.size;
525 int size = end - data;
527 BUG_ON(size % sizeof(u64));
528 mem_bswap_64(data, size);
531 static void perf_event__all64_swap(union perf_event *event,
532 bool sample_id_all __maybe_unused)
534 struct perf_event_header *hdr = &event->header;
535 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
538 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
540 event->comm.pid = bswap_32(event->comm.pid);
541 event->comm.tid = bswap_32(event->comm.tid);
544 void *data = &event->comm.comm;
546 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
547 swap_sample_id_all(event, data);
551 static void perf_event__mmap_swap(union perf_event *event,
554 event->mmap.pid = bswap_32(event->mmap.pid);
555 event->mmap.tid = bswap_32(event->mmap.tid);
556 event->mmap.start = bswap_64(event->mmap.start);
557 event->mmap.len = bswap_64(event->mmap.len);
558 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
561 void *data = &event->mmap.filename;
563 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
564 swap_sample_id_all(event, data);
568 static void perf_event__mmap2_swap(union perf_event *event,
571 event->mmap2.pid = bswap_32(event->mmap2.pid);
572 event->mmap2.tid = bswap_32(event->mmap2.tid);
573 event->mmap2.start = bswap_64(event->mmap2.start);
574 event->mmap2.len = bswap_64(event->mmap2.len);
575 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
576 event->mmap2.maj = bswap_32(event->mmap2.maj);
577 event->mmap2.min = bswap_32(event->mmap2.min);
578 event->mmap2.ino = bswap_64(event->mmap2.ino);
581 void *data = &event->mmap2.filename;
583 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
584 swap_sample_id_all(event, data);
587 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
589 event->fork.pid = bswap_32(event->fork.pid);
590 event->fork.tid = bswap_32(event->fork.tid);
591 event->fork.ppid = bswap_32(event->fork.ppid);
592 event->fork.ptid = bswap_32(event->fork.ptid);
593 event->fork.time = bswap_64(event->fork.time);
596 swap_sample_id_all(event, &event->fork + 1);
599 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
601 event->read.pid = bswap_32(event->read.pid);
602 event->read.tid = bswap_32(event->read.tid);
603 event->read.value = bswap_64(event->read.value);
604 event->read.time_enabled = bswap_64(event->read.time_enabled);
605 event->read.time_running = bswap_64(event->read.time_running);
606 event->read.id = bswap_64(event->read.id);
609 swap_sample_id_all(event, &event->read + 1);
612 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all)
614 event->aux.aux_offset = bswap_64(event->aux.aux_offset);
615 event->aux.aux_size = bswap_64(event->aux.aux_size);
616 event->aux.flags = bswap_64(event->aux.flags);
619 swap_sample_id_all(event, &event->aux + 1);
622 static void perf_event__itrace_start_swap(union perf_event *event,
625 event->itrace_start.pid = bswap_32(event->itrace_start.pid);
626 event->itrace_start.tid = bswap_32(event->itrace_start.tid);
629 swap_sample_id_all(event, &event->itrace_start + 1);
632 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all)
634 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) {
635 event->context_switch.next_prev_pid =
636 bswap_32(event->context_switch.next_prev_pid);
637 event->context_switch.next_prev_tid =
638 bswap_32(event->context_switch.next_prev_tid);
642 swap_sample_id_all(event, &event->context_switch + 1);
645 static void perf_event__throttle_swap(union perf_event *event,
648 event->throttle.time = bswap_64(event->throttle.time);
649 event->throttle.id = bswap_64(event->throttle.id);
650 event->throttle.stream_id = bswap_64(event->throttle.stream_id);
653 swap_sample_id_all(event, &event->throttle + 1);
656 static void perf_event__namespaces_swap(union perf_event *event,
661 event->namespaces.pid = bswap_32(event->namespaces.pid);
662 event->namespaces.tid = bswap_32(event->namespaces.tid);
663 event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces);
665 for (i = 0; i < event->namespaces.nr_namespaces; i++) {
666 struct perf_ns_link_info *ns = &event->namespaces.link_info[i];
668 ns->dev = bswap_64(ns->dev);
669 ns->ino = bswap_64(ns->ino);
673 swap_sample_id_all(event, &event->namespaces.link_info[i]);
676 static u8 revbyte(u8 b)
678 int rev = (b >> 4) | ((b & 0xf) << 4);
679 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
680 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
685 * XXX this is hack in attempt to carry flags bitfield
686 * through endian village. ABI says:
688 * Bit-fields are allocated from right to left (least to most significant)
689 * on little-endian implementations and from left to right (most to least
690 * significant) on big-endian implementations.
692 * The above seems to be byte specific, so we need to reverse each
693 * byte of the bitfield. 'Internet' also says this might be implementation
694 * specific and we probably need proper fix and carry perf_event_attr
695 * bitfield flags in separate data file FEAT_ section. Thought this seems
698 static void swap_bitfield(u8 *p, unsigned len)
702 for (i = 0; i < len; i++) {
708 /* exported for swapping attributes in file header */
709 void perf_event__attr_swap(struct perf_event_attr *attr)
711 attr->type = bswap_32(attr->type);
712 attr->size = bswap_32(attr->size);
714 #define bswap_safe(f, n) \
715 (attr->size > (offsetof(struct perf_event_attr, f) + \
716 sizeof(attr->f) * (n)))
717 #define bswap_field(f, sz) \
719 if (bswap_safe(f, 0)) \
720 attr->f = bswap_##sz(attr->f); \
722 #define bswap_field_16(f) bswap_field(f, 16)
723 #define bswap_field_32(f) bswap_field(f, 32)
724 #define bswap_field_64(f) bswap_field(f, 64)
726 bswap_field_64(config);
727 bswap_field_64(sample_period);
728 bswap_field_64(sample_type);
729 bswap_field_64(read_format);
730 bswap_field_32(wakeup_events);
731 bswap_field_32(bp_type);
732 bswap_field_64(bp_addr);
733 bswap_field_64(bp_len);
734 bswap_field_64(branch_sample_type);
735 bswap_field_64(sample_regs_user);
736 bswap_field_32(sample_stack_user);
737 bswap_field_32(aux_watermark);
738 bswap_field_16(sample_max_stack);
741 * After read_format are bitfields. Check read_format because
742 * we are unable to use offsetof on bitfield.
744 if (bswap_safe(read_format, 1))
745 swap_bitfield((u8 *) (&attr->read_format + 1),
747 #undef bswap_field_64
748 #undef bswap_field_32
753 static void perf_event__hdr_attr_swap(union perf_event *event,
754 bool sample_id_all __maybe_unused)
758 perf_event__attr_swap(&event->attr.attr);
760 size = event->header.size;
761 size -= (void *)&event->attr.id - (void *)event;
762 mem_bswap_64(event->attr.id, size);
765 static void perf_event__event_update_swap(union perf_event *event,
766 bool sample_id_all __maybe_unused)
768 event->event_update.type = bswap_64(event->event_update.type);
769 event->event_update.id = bswap_64(event->event_update.id);
772 static void perf_event__event_type_swap(union perf_event *event,
773 bool sample_id_all __maybe_unused)
775 event->event_type.event_type.event_id =
776 bswap_64(event->event_type.event_type.event_id);
779 static void perf_event__tracing_data_swap(union perf_event *event,
780 bool sample_id_all __maybe_unused)
782 event->tracing_data.size = bswap_32(event->tracing_data.size);
785 static void perf_event__auxtrace_info_swap(union perf_event *event,
786 bool sample_id_all __maybe_unused)
790 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type);
792 size = event->header.size;
793 size -= (void *)&event->auxtrace_info.priv - (void *)event;
794 mem_bswap_64(event->auxtrace_info.priv, size);
797 static void perf_event__auxtrace_swap(union perf_event *event,
798 bool sample_id_all __maybe_unused)
800 event->auxtrace.size = bswap_64(event->auxtrace.size);
801 event->auxtrace.offset = bswap_64(event->auxtrace.offset);
802 event->auxtrace.reference = bswap_64(event->auxtrace.reference);
803 event->auxtrace.idx = bswap_32(event->auxtrace.idx);
804 event->auxtrace.tid = bswap_32(event->auxtrace.tid);
805 event->auxtrace.cpu = bswap_32(event->auxtrace.cpu);
808 static void perf_event__auxtrace_error_swap(union perf_event *event,
809 bool sample_id_all __maybe_unused)
811 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type);
812 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code);
813 event->auxtrace_error.cpu = bswap_32(event->auxtrace_error.cpu);
814 event->auxtrace_error.pid = bswap_32(event->auxtrace_error.pid);
815 event->auxtrace_error.tid = bswap_32(event->auxtrace_error.tid);
816 event->auxtrace_error.fmt = bswap_32(event->auxtrace_error.fmt);
817 event->auxtrace_error.ip = bswap_64(event->auxtrace_error.ip);
818 if (event->auxtrace_error.fmt)
819 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time);
822 static void perf_event__thread_map_swap(union perf_event *event,
823 bool sample_id_all __maybe_unused)
827 event->thread_map.nr = bswap_64(event->thread_map.nr);
829 for (i = 0; i < event->thread_map.nr; i++)
830 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid);
833 static void perf_event__cpu_map_swap(union perf_event *event,
834 bool sample_id_all __maybe_unused)
836 struct cpu_map_data *data = &event->cpu_map.data;
837 struct cpu_map_entries *cpus;
838 struct cpu_map_mask *mask;
841 data->type = bswap_64(data->type);
843 switch (data->type) {
844 case PERF_CPU_MAP__CPUS:
845 cpus = (struct cpu_map_entries *)data->data;
847 cpus->nr = bswap_16(cpus->nr);
849 for (i = 0; i < cpus->nr; i++)
850 cpus->cpu[i] = bswap_16(cpus->cpu[i]);
852 case PERF_CPU_MAP__MASK:
853 mask = (struct cpu_map_mask *) data->data;
855 mask->nr = bswap_16(mask->nr);
856 mask->long_size = bswap_16(mask->long_size);
858 switch (mask->long_size) {
859 case 4: mem_bswap_32(&mask->mask, mask->nr); break;
860 case 8: mem_bswap_64(&mask->mask, mask->nr); break;
862 pr_err("cpu_map swap: unsupported long size\n");
869 static void perf_event__stat_config_swap(union perf_event *event,
870 bool sample_id_all __maybe_unused)
874 size = event->stat_config.nr * sizeof(event->stat_config.data[0]);
875 size += 1; /* nr item itself */
876 mem_bswap_64(&event->stat_config.nr, size);
879 static void perf_event__stat_swap(union perf_event *event,
880 bool sample_id_all __maybe_unused)
882 event->stat.id = bswap_64(event->stat.id);
883 event->stat.thread = bswap_32(event->stat.thread);
884 event->stat.cpu = bswap_32(event->stat.cpu);
885 event->stat.val = bswap_64(event->stat.val);
886 event->stat.ena = bswap_64(event->stat.ena);
887 event->stat.run = bswap_64(event->stat.run);
890 static void perf_event__stat_round_swap(union perf_event *event,
891 bool sample_id_all __maybe_unused)
893 event->stat_round.type = bswap_64(event->stat_round.type);
894 event->stat_round.time = bswap_64(event->stat_round.time);
897 typedef void (*perf_event__swap_op)(union perf_event *event,
900 static perf_event__swap_op perf_event__swap_ops[] = {
901 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
902 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap,
903 [PERF_RECORD_COMM] = perf_event__comm_swap,
904 [PERF_RECORD_FORK] = perf_event__task_swap,
905 [PERF_RECORD_EXIT] = perf_event__task_swap,
906 [PERF_RECORD_LOST] = perf_event__all64_swap,
907 [PERF_RECORD_READ] = perf_event__read_swap,
908 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap,
909 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap,
910 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
911 [PERF_RECORD_AUX] = perf_event__aux_swap,
912 [PERF_RECORD_ITRACE_START] = perf_event__itrace_start_swap,
913 [PERF_RECORD_LOST_SAMPLES] = perf_event__all64_swap,
914 [PERF_RECORD_SWITCH] = perf_event__switch_swap,
915 [PERF_RECORD_SWITCH_CPU_WIDE] = perf_event__switch_swap,
916 [PERF_RECORD_NAMESPACES] = perf_event__namespaces_swap,
917 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
918 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
919 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
920 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
921 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap,
922 [PERF_RECORD_AUXTRACE_INFO] = perf_event__auxtrace_info_swap,
923 [PERF_RECORD_AUXTRACE] = perf_event__auxtrace_swap,
924 [PERF_RECORD_AUXTRACE_ERROR] = perf_event__auxtrace_error_swap,
925 [PERF_RECORD_THREAD_MAP] = perf_event__thread_map_swap,
926 [PERF_RECORD_CPU_MAP] = perf_event__cpu_map_swap,
927 [PERF_RECORD_STAT_CONFIG] = perf_event__stat_config_swap,
928 [PERF_RECORD_STAT] = perf_event__stat_swap,
929 [PERF_RECORD_STAT_ROUND] = perf_event__stat_round_swap,
930 [PERF_RECORD_EVENT_UPDATE] = perf_event__event_update_swap,
931 [PERF_RECORD_TIME_CONV] = perf_event__all64_swap,
932 [PERF_RECORD_HEADER_MAX] = NULL,
936 * When perf record finishes a pass on every buffers, it records this pseudo
938 * We record the max timestamp t found in the pass n.
939 * Assuming these timestamps are monotonic across cpus, we know that if
940 * a buffer still has events with timestamps below t, they will be all
941 * available and then read in the pass n + 1.
942 * Hence when we start to read the pass n + 2, we can safely flush every
943 * events with timestamps below t.
945 * ============ PASS n =================
948 * cnt1 timestamps | cnt2 timestamps
951 * - | 4 <--- max recorded
953 * ============ PASS n + 1 ==============
956 * cnt1 timestamps | cnt2 timestamps
959 * 5 | 7 <---- max recorded
961 * Flush every events below timestamp 4
963 * ============ PASS n + 2 ==============
966 * cnt1 timestamps | cnt2 timestamps
971 * Flush every events below timestamp 7
974 static int process_finished_round(struct perf_tool *tool __maybe_unused,
975 union perf_event *event __maybe_unused,
976 struct ordered_events *oe)
979 fprintf(stdout, "\n");
980 return ordered_events__flush(oe, OE_FLUSH__ROUND);
983 int perf_session__queue_event(struct perf_session *s, union perf_event *event,
984 u64 timestamp, u64 file_offset)
986 return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset);
989 static void callchain__lbr_callstack_printf(struct perf_sample *sample)
991 struct ip_callchain *callchain = sample->callchain;
992 struct branch_stack *lbr_stack = sample->branch_stack;
993 u64 kernel_callchain_nr = callchain->nr;
996 for (i = 0; i < kernel_callchain_nr; i++) {
997 if (callchain->ips[i] == PERF_CONTEXT_USER)
1001 if ((i != kernel_callchain_nr) && lbr_stack->nr) {
1004 * LBR callstack can only get user call chain,
1005 * i is kernel call chain number,
1006 * 1 is PERF_CONTEXT_USER.
1008 * The user call chain is stored in LBR registers.
1009 * LBR are pair registers. The caller is stored
1010 * in "from" register, while the callee is stored
1012 * For example, there is a call stack
1013 * "A"->"B"->"C"->"D".
1014 * The LBR registers will recorde like
1015 * "C"->"D", "B"->"C", "A"->"B".
1016 * So only the first "to" register and all "from"
1017 * registers are needed to construct the whole stack.
1019 total_nr = i + 1 + lbr_stack->nr + 1;
1020 kernel_callchain_nr = i + 1;
1022 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr);
1024 for (i = 0; i < kernel_callchain_nr; i++)
1025 printf("..... %2d: %016" PRIx64 "\n",
1026 i, callchain->ips[i]);
1028 printf("..... %2d: %016" PRIx64 "\n",
1029 (int)(kernel_callchain_nr), lbr_stack->entries[0].to);
1030 for (i = 0; i < lbr_stack->nr; i++)
1031 printf("..... %2d: %016" PRIx64 "\n",
1032 (int)(i + kernel_callchain_nr + 1), lbr_stack->entries[i].from);
1036 static void callchain__printf(struct evsel *evsel,
1037 struct perf_sample *sample)
1040 struct ip_callchain *callchain = sample->callchain;
1042 if (perf_evsel__has_branch_callstack(evsel))
1043 callchain__lbr_callstack_printf(sample);
1045 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr);
1047 for (i = 0; i < callchain->nr; i++)
1048 printf("..... %2d: %016" PRIx64 "\n",
1049 i, callchain->ips[i]);
1052 static void branch_stack__printf(struct perf_sample *sample)
1056 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
1058 for (i = 0; i < sample->branch_stack->nr; i++) {
1059 struct branch_entry *e = &sample->branch_stack->entries[i];
1061 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x\n",
1063 (unsigned short)e->flags.cycles,
1064 e->flags.mispred ? "M" : " ",
1065 e->flags.predicted ? "P" : " ",
1066 e->flags.abort ? "A" : " ",
1067 e->flags.in_tx ? "T" : " ",
1068 (unsigned)e->flags.reserved);
1072 static void regs_dump__printf(u64 mask, u64 *regs)
1074 unsigned rid, i = 0;
1076 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
1077 u64 val = regs[i++];
1079 printf(".... %-5s 0x%" PRIx64 "\n",
1080 perf_reg_name(rid), val);
1084 static const char *regs_abi[] = {
1085 [PERF_SAMPLE_REGS_ABI_NONE] = "none",
1086 [PERF_SAMPLE_REGS_ABI_32] = "32-bit",
1087 [PERF_SAMPLE_REGS_ABI_64] = "64-bit",
1090 static inline const char *regs_dump_abi(struct regs_dump *d)
1092 if (d->abi > PERF_SAMPLE_REGS_ABI_64)
1095 return regs_abi[d->abi];
1098 static void regs__printf(const char *type, struct regs_dump *regs)
1100 u64 mask = regs->mask;
1102 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n",
1105 regs_dump_abi(regs));
1107 regs_dump__printf(mask, regs->regs);
1110 static void regs_user__printf(struct perf_sample *sample)
1112 struct regs_dump *user_regs = &sample->user_regs;
1114 if (user_regs->regs)
1115 regs__printf("user", user_regs);
1118 static void regs_intr__printf(struct perf_sample *sample)
1120 struct regs_dump *intr_regs = &sample->intr_regs;
1122 if (intr_regs->regs)
1123 regs__printf("intr", intr_regs);
1126 static void stack_user__printf(struct stack_dump *dump)
1128 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
1129 dump->size, dump->offset);
1132 static void perf_evlist__print_tstamp(struct evlist *evlist,
1133 union perf_event *event,
1134 struct perf_sample *sample)
1136 u64 sample_type = __perf_evlist__combined_sample_type(evlist);
1138 if (event->header.type != PERF_RECORD_SAMPLE &&
1139 !perf_evlist__sample_id_all(evlist)) {
1140 fputs("-1 -1 ", stdout);
1144 if ((sample_type & PERF_SAMPLE_CPU))
1145 printf("%u ", sample->cpu);
1147 if (sample_type & PERF_SAMPLE_TIME)
1148 printf("%" PRIu64 " ", sample->time);
1151 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
1153 printf("... sample_read:\n");
1155 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1156 printf("...... time enabled %016" PRIx64 "\n",
1157 sample->read.time_enabled);
1159 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1160 printf("...... time running %016" PRIx64 "\n",
1161 sample->read.time_running);
1163 if (read_format & PERF_FORMAT_GROUP) {
1166 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
1168 for (i = 0; i < sample->read.group.nr; i++) {
1169 struct sample_read_value *value;
1171 value = &sample->read.group.values[i];
1172 printf("..... id %016" PRIx64
1173 ", value %016" PRIx64 "\n",
1174 value->id, value->value);
1177 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
1178 sample->read.one.id, sample->read.one.value);
1181 static void dump_event(struct evlist *evlist, union perf_event *event,
1182 u64 file_offset, struct perf_sample *sample)
1187 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
1188 file_offset, event->header.size, event->header.type);
1191 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw)
1192 evlist->trace_event_sample_raw(evlist, event, sample);
1195 perf_evlist__print_tstamp(evlist, event, sample);
1197 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
1198 event->header.size, perf_event__name(event->header.type));
1201 static void dump_sample(struct evsel *evsel, union perf_event *event,
1202 struct perf_sample *sample)
1209 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
1210 event->header.misc, sample->pid, sample->tid, sample->ip,
1211 sample->period, sample->addr);
1213 sample_type = evsel->attr.sample_type;
1215 if (evsel__has_callchain(evsel))
1216 callchain__printf(evsel, sample);
1218 if ((sample_type & PERF_SAMPLE_BRANCH_STACK) && !perf_evsel__has_branch_callstack(evsel))
1219 branch_stack__printf(sample);
1221 if (sample_type & PERF_SAMPLE_REGS_USER)
1222 regs_user__printf(sample);
1224 if (sample_type & PERF_SAMPLE_REGS_INTR)
1225 regs_intr__printf(sample);
1227 if (sample_type & PERF_SAMPLE_STACK_USER)
1228 stack_user__printf(&sample->user_stack);
1230 if (sample_type & PERF_SAMPLE_WEIGHT)
1231 printf("... weight: %" PRIu64 "\n", sample->weight);
1233 if (sample_type & PERF_SAMPLE_DATA_SRC)
1234 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
1236 if (sample_type & PERF_SAMPLE_PHYS_ADDR)
1237 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr);
1239 if (sample_type & PERF_SAMPLE_TRANSACTION)
1240 printf("... transaction: %" PRIx64 "\n", sample->transaction);
1242 if (sample_type & PERF_SAMPLE_READ)
1243 sample_read__printf(sample, evsel->attr.read_format);
1246 static void dump_read(struct evsel *evsel, union perf_event *event)
1248 struct read_event *read_event = &event->read;
1254 printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
1255 perf_evsel__name(evsel),
1261 read_format = evsel->attr.read_format;
1263 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1264 printf("... time enabled : %" PRIu64 "\n", read_event->time_enabled);
1266 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1267 printf("... time running : %" PRIu64 "\n", read_event->time_running);
1269 if (read_format & PERF_FORMAT_ID)
1270 printf("... id : %" PRIu64 "\n", read_event->id);
1273 static struct machine *machines__find_for_cpumode(struct machines *machines,
1274 union perf_event *event,
1275 struct perf_sample *sample)
1277 struct machine *machine;
1280 ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
1281 (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) {
1284 if (event->header.type == PERF_RECORD_MMAP
1285 || event->header.type == PERF_RECORD_MMAP2)
1286 pid = event->mmap.pid;
1290 machine = machines__find(machines, pid);
1292 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
1296 return &machines->host;
1299 static int deliver_sample_value(struct evlist *evlist,
1300 struct perf_tool *tool,
1301 union perf_event *event,
1302 struct perf_sample *sample,
1303 struct sample_read_value *v,
1304 struct machine *machine)
1306 struct perf_sample_id *sid = perf_evlist__id2sid(evlist, v->id);
1310 sample->period = v->value - sid->period;
1311 sid->period = v->value;
1314 if (!sid || sid->evsel == NULL) {
1315 ++evlist->stats.nr_unknown_id;
1320 * There's no reason to deliver sample
1321 * for zero period, bail out.
1323 if (!sample->period)
1326 return tool->sample(tool, event, sample, sid->evsel, machine);
1329 static int deliver_sample_group(struct evlist *evlist,
1330 struct perf_tool *tool,
1331 union perf_event *event,
1332 struct perf_sample *sample,
1333 struct machine *machine)
1338 for (i = 0; i < sample->read.group.nr; i++) {
1339 ret = deliver_sample_value(evlist, tool, event, sample,
1340 &sample->read.group.values[i],
1350 perf_evlist__deliver_sample(struct evlist *evlist,
1351 struct perf_tool *tool,
1352 union perf_event *event,
1353 struct perf_sample *sample,
1354 struct evsel *evsel,
1355 struct machine *machine)
1357 /* We know evsel != NULL. */
1358 u64 sample_type = evsel->attr.sample_type;
1359 u64 read_format = evsel->attr.read_format;
1361 /* Standard sample delivery. */
1362 if (!(sample_type & PERF_SAMPLE_READ))
1363 return tool->sample(tool, event, sample, evsel, machine);
1365 /* For PERF_SAMPLE_READ we have either single or group mode. */
1366 if (read_format & PERF_FORMAT_GROUP)
1367 return deliver_sample_group(evlist, tool, event, sample,
1370 return deliver_sample_value(evlist, tool, event, sample,
1371 &sample->read.one, machine);
1374 static int machines__deliver_event(struct machines *machines,
1375 struct evlist *evlist,
1376 union perf_event *event,
1377 struct perf_sample *sample,
1378 struct perf_tool *tool, u64 file_offset)
1380 struct evsel *evsel;
1381 struct machine *machine;
1383 dump_event(evlist, event, file_offset, sample);
1385 evsel = perf_evlist__id2evsel(evlist, sample->id);
1387 machine = machines__find_for_cpumode(machines, event, sample);
1389 switch (event->header.type) {
1390 case PERF_RECORD_SAMPLE:
1391 if (evsel == NULL) {
1392 ++evlist->stats.nr_unknown_id;
1395 dump_sample(evsel, event, sample);
1396 if (machine == NULL) {
1397 ++evlist->stats.nr_unprocessable_samples;
1400 return perf_evlist__deliver_sample(evlist, tool, event, sample, evsel, machine);
1401 case PERF_RECORD_MMAP:
1402 return tool->mmap(tool, event, sample, machine);
1403 case PERF_RECORD_MMAP2:
1404 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT)
1405 ++evlist->stats.nr_proc_map_timeout;
1406 return tool->mmap2(tool, event, sample, machine);
1407 case PERF_RECORD_COMM:
1408 return tool->comm(tool, event, sample, machine);
1409 case PERF_RECORD_NAMESPACES:
1410 return tool->namespaces(tool, event, sample, machine);
1411 case PERF_RECORD_FORK:
1412 return tool->fork(tool, event, sample, machine);
1413 case PERF_RECORD_EXIT:
1414 return tool->exit(tool, event, sample, machine);
1415 case PERF_RECORD_LOST:
1416 if (tool->lost == perf_event__process_lost)
1417 evlist->stats.total_lost += event->lost.lost;
1418 return tool->lost(tool, event, sample, machine);
1419 case PERF_RECORD_LOST_SAMPLES:
1420 if (tool->lost_samples == perf_event__process_lost_samples)
1421 evlist->stats.total_lost_samples += event->lost_samples.lost;
1422 return tool->lost_samples(tool, event, sample, machine);
1423 case PERF_RECORD_READ:
1424 dump_read(evsel, event);
1425 return tool->read(tool, event, sample, evsel, machine);
1426 case PERF_RECORD_THROTTLE:
1427 return tool->throttle(tool, event, sample, machine);
1428 case PERF_RECORD_UNTHROTTLE:
1429 return tool->unthrottle(tool, event, sample, machine);
1430 case PERF_RECORD_AUX:
1431 if (tool->aux == perf_event__process_aux) {
1432 if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED)
1433 evlist->stats.total_aux_lost += 1;
1434 if (event->aux.flags & PERF_AUX_FLAG_PARTIAL)
1435 evlist->stats.total_aux_partial += 1;
1437 return tool->aux(tool, event, sample, machine);
1438 case PERF_RECORD_ITRACE_START:
1439 return tool->itrace_start(tool, event, sample, machine);
1440 case PERF_RECORD_SWITCH:
1441 case PERF_RECORD_SWITCH_CPU_WIDE:
1442 return tool->context_switch(tool, event, sample, machine);
1443 case PERF_RECORD_KSYMBOL:
1444 return tool->ksymbol(tool, event, sample, machine);
1445 case PERF_RECORD_BPF_EVENT:
1446 return tool->bpf_event(tool, event, sample, machine);
1448 ++evlist->stats.nr_unknown_events;
1453 static int perf_session__deliver_event(struct perf_session *session,
1454 union perf_event *event,
1455 struct perf_tool *tool,
1458 struct perf_sample sample;
1461 ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1463 pr_err("Can't parse sample, err = %d\n", ret);
1467 ret = auxtrace__process_event(session, event, &sample, tool);
1473 return machines__deliver_event(&session->machines, session->evlist,
1474 event, &sample, tool, file_offset);
1477 static s64 perf_session__process_user_event(struct perf_session *session,
1478 union perf_event *event,
1481 struct ordered_events *oe = &session->ordered_events;
1482 struct perf_tool *tool = session->tool;
1483 struct perf_sample sample = { .time = 0, };
1484 int fd = perf_data__fd(session->data);
1487 if (event->header.type != PERF_RECORD_COMPRESSED ||
1488 tool->compressed == perf_session__process_compressed_event_stub)
1489 dump_event(session->evlist, event, file_offset, &sample);
1491 /* These events are processed right away */
1492 switch (event->header.type) {
1493 case PERF_RECORD_HEADER_ATTR:
1494 err = tool->attr(tool, event, &session->evlist);
1496 perf_session__set_id_hdr_size(session);
1497 perf_session__set_comm_exec(session);
1500 case PERF_RECORD_EVENT_UPDATE:
1501 return tool->event_update(tool, event, &session->evlist);
1502 case PERF_RECORD_HEADER_EVENT_TYPE:
1504 * Depreceated, but we need to handle it for sake
1505 * of old data files create in pipe mode.
1508 case PERF_RECORD_HEADER_TRACING_DATA:
1509 /* setup for reading amidst mmap */
1510 lseek(fd, file_offset, SEEK_SET);
1511 return tool->tracing_data(session, event);
1512 case PERF_RECORD_HEADER_BUILD_ID:
1513 return tool->build_id(session, event);
1514 case PERF_RECORD_FINISHED_ROUND:
1515 return tool->finished_round(tool, event, oe);
1516 case PERF_RECORD_ID_INDEX:
1517 return tool->id_index(session, event);
1518 case PERF_RECORD_AUXTRACE_INFO:
1519 return tool->auxtrace_info(session, event);
1520 case PERF_RECORD_AUXTRACE:
1521 /* setup for reading amidst mmap */
1522 lseek(fd, file_offset + event->header.size, SEEK_SET);
1523 return tool->auxtrace(session, event);
1524 case PERF_RECORD_AUXTRACE_ERROR:
1525 perf_session__auxtrace_error_inc(session, event);
1526 return tool->auxtrace_error(session, event);
1527 case PERF_RECORD_THREAD_MAP:
1528 return tool->thread_map(session, event);
1529 case PERF_RECORD_CPU_MAP:
1530 return tool->cpu_map(session, event);
1531 case PERF_RECORD_STAT_CONFIG:
1532 return tool->stat_config(session, event);
1533 case PERF_RECORD_STAT:
1534 return tool->stat(session, event);
1535 case PERF_RECORD_STAT_ROUND:
1536 return tool->stat_round(session, event);
1537 case PERF_RECORD_TIME_CONV:
1538 session->time_conv = event->time_conv;
1539 return tool->time_conv(session, event);
1540 case PERF_RECORD_HEADER_FEATURE:
1541 return tool->feature(session, event);
1542 case PERF_RECORD_COMPRESSED:
1543 err = tool->compressed(session, event, file_offset);
1545 dump_event(session->evlist, event, file_offset, &sample);
1552 int perf_session__deliver_synth_event(struct perf_session *session,
1553 union perf_event *event,
1554 struct perf_sample *sample)
1556 struct evlist *evlist = session->evlist;
1557 struct perf_tool *tool = session->tool;
1559 events_stats__inc(&evlist->stats, event->header.type);
1561 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1562 return perf_session__process_user_event(session, event, 0);
1564 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0);
1567 static void event_swap(union perf_event *event, bool sample_id_all)
1569 perf_event__swap_op swap;
1571 swap = perf_event__swap_ops[event->header.type];
1573 swap(event, sample_id_all);
1576 int perf_session__peek_event(struct perf_session *session, off_t file_offset,
1577 void *buf, size_t buf_sz,
1578 union perf_event **event_ptr,
1579 struct perf_sample *sample)
1581 union perf_event *event;
1582 size_t hdr_sz, rest;
1585 if (session->one_mmap && !session->header.needs_swap) {
1586 event = file_offset - session->one_mmap_offset +
1587 session->one_mmap_addr;
1588 goto out_parse_sample;
1591 if (perf_data__is_pipe(session->data))
1594 fd = perf_data__fd(session->data);
1595 hdr_sz = sizeof(struct perf_event_header);
1597 if (buf_sz < hdr_sz)
1600 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 ||
1601 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz)
1604 event = (union perf_event *)buf;
1606 if (session->header.needs_swap)
1607 perf_event_header__bswap(&event->header);
1609 if (event->header.size < hdr_sz || event->header.size > buf_sz)
1612 rest = event->header.size - hdr_sz;
1614 if (readn(fd, buf, rest) != (ssize_t)rest)
1617 if (session->header.needs_swap)
1618 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1622 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START &&
1623 perf_evlist__parse_sample(session->evlist, event, sample))
1631 static s64 perf_session__process_event(struct perf_session *session,
1632 union perf_event *event, u64 file_offset)
1634 struct evlist *evlist = session->evlist;
1635 struct perf_tool *tool = session->tool;
1638 if (session->header.needs_swap)
1639 event_swap(event, perf_evlist__sample_id_all(evlist));
1641 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1644 events_stats__inc(&evlist->stats, event->header.type);
1646 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1647 return perf_session__process_user_event(session, event, file_offset);
1649 if (tool->ordered_events) {
1650 u64 timestamp = -1ULL;
1652 ret = perf_evlist__parse_sample_timestamp(evlist, event, ×tamp);
1653 if (ret && ret != -1)
1656 ret = perf_session__queue_event(session, event, timestamp, file_offset);
1661 return perf_session__deliver_event(session, event, tool, file_offset);
1664 void perf_event_header__bswap(struct perf_event_header *hdr)
1666 hdr->type = bswap_32(hdr->type);
1667 hdr->misc = bswap_16(hdr->misc);
1668 hdr->size = bswap_16(hdr->size);
1671 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1673 return machine__findnew_thread(&session->machines.host, -1, pid);
1677 * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
1678 * So here a single thread is created for that, but actually there is a separate
1679 * idle task per cpu, so there should be one 'struct thread' per cpu, but there
1680 * is only 1. That causes problems for some tools, requiring workarounds. For
1681 * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
1683 int perf_session__register_idle_thread(struct perf_session *session)
1685 struct thread *thread;
1688 thread = machine__findnew_thread(&session->machines.host, 0, 0);
1689 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1690 pr_err("problem inserting idle task.\n");
1694 if (thread == NULL || thread__set_namespaces(thread, 0, NULL)) {
1695 pr_err("problem inserting idle task.\n");
1699 /* machine__findnew_thread() got the thread, so put it */
1700 thread__put(thread);
1705 perf_session__warn_order(const struct perf_session *session)
1707 const struct ordered_events *oe = &session->ordered_events;
1708 struct evsel *evsel;
1709 bool should_warn = true;
1711 evlist__for_each_entry(session->evlist, evsel) {
1712 if (evsel->attr.write_backward)
1713 should_warn = false;
1718 if (oe->nr_unordered_events != 0)
1719 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events);
1722 static void perf_session__warn_about_errors(const struct perf_session *session)
1724 const struct events_stats *stats = &session->evlist->stats;
1726 if (session->tool->lost == perf_event__process_lost &&
1727 stats->nr_events[PERF_RECORD_LOST] != 0) {
1728 ui__warning("Processed %d events and lost %d chunks!\n\n"
1729 "Check IO/CPU overload!\n\n",
1730 stats->nr_events[0],
1731 stats->nr_events[PERF_RECORD_LOST]);
1734 if (session->tool->lost_samples == perf_event__process_lost_samples) {
1737 drop_rate = (double)stats->total_lost_samples /
1738 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples);
1739 if (drop_rate > 0.05) {
1740 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n",
1741 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples,
1746 if (session->tool->aux == perf_event__process_aux &&
1747 stats->total_aux_lost != 0) {
1748 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n",
1749 stats->total_aux_lost,
1750 stats->nr_events[PERF_RECORD_AUX]);
1753 if (session->tool->aux == perf_event__process_aux &&
1754 stats->total_aux_partial != 0) {
1755 bool vmm_exclusive = false;
1757 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive",
1760 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n"
1761 "Are you running a KVM guest in the background?%s\n\n",
1762 stats->total_aux_partial,
1763 stats->nr_events[PERF_RECORD_AUX],
1765 "\nReloading kvm_intel module with vmm_exclusive=0\n"
1766 "will reduce the gaps to only guest's timeslices." :
1770 if (stats->nr_unknown_events != 0) {
1771 ui__warning("Found %u unknown events!\n\n"
1772 "Is this an older tool processing a perf.data "
1773 "file generated by a more recent tool?\n\n"
1774 "If that is not the case, consider "
1775 "reporting to linux-kernel@vger.kernel.org.\n\n",
1776 stats->nr_unknown_events);
1779 if (stats->nr_unknown_id != 0) {
1780 ui__warning("%u samples with id not present in the header\n",
1781 stats->nr_unknown_id);
1784 if (stats->nr_invalid_chains != 0) {
1785 ui__warning("Found invalid callchains!\n\n"
1786 "%u out of %u events were discarded for this reason.\n\n"
1787 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1788 stats->nr_invalid_chains,
1789 stats->nr_events[PERF_RECORD_SAMPLE]);
1792 if (stats->nr_unprocessable_samples != 0) {
1793 ui__warning("%u unprocessable samples recorded.\n"
1794 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1795 stats->nr_unprocessable_samples);
1798 perf_session__warn_order(session);
1800 events_stats__auxtrace_error_warn(stats);
1802 if (stats->nr_proc_map_timeout != 0) {
1803 ui__warning("%d map information files for pre-existing threads were\n"
1804 "not processed, if there are samples for addresses they\n"
1805 "will not be resolved, you may find out which are these\n"
1806 "threads by running with -v and redirecting the output\n"
1808 "The time limit to process proc map is too short?\n"
1809 "Increase it by --proc-map-timeout\n",
1810 stats->nr_proc_map_timeout);
1814 static int perf_session__flush_thread_stack(struct thread *thread,
1815 void *p __maybe_unused)
1817 return thread_stack__flush(thread);
1820 static int perf_session__flush_thread_stacks(struct perf_session *session)
1822 return machines__for_each_thread(&session->machines,
1823 perf_session__flush_thread_stack,
1827 volatile int session_done;
1829 static int __perf_session__process_decomp_events(struct perf_session *session);
1831 static int __perf_session__process_pipe_events(struct perf_session *session)
1833 struct ordered_events *oe = &session->ordered_events;
1834 struct perf_tool *tool = session->tool;
1835 int fd = perf_data__fd(session->data);
1836 union perf_event *event;
1837 uint32_t size, cur_size = 0;
1844 perf_tool__fill_defaults(tool);
1847 cur_size = sizeof(union perf_event);
1849 buf = malloc(cur_size);
1852 ordered_events__set_copy_on_queue(oe, true);
1855 err = readn(fd, event, sizeof(struct perf_event_header));
1860 pr_err("failed to read event header\n");
1864 if (session->header.needs_swap)
1865 perf_event_header__bswap(&event->header);
1867 size = event->header.size;
1868 if (size < sizeof(struct perf_event_header)) {
1869 pr_err("bad event header size\n");
1873 if (size > cur_size) {
1874 void *new = realloc(buf, size);
1876 pr_err("failed to allocate memory to read event\n");
1884 p += sizeof(struct perf_event_header);
1886 if (size - sizeof(struct perf_event_header)) {
1887 err = readn(fd, p, size - sizeof(struct perf_event_header));
1890 pr_err("unexpected end of event stream\n");
1894 pr_err("failed to read event data\n");
1899 if ((skip = perf_session__process_event(session, event, head)) < 0) {
1900 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1901 head, event->header.size, event->header.type);
1911 err = __perf_session__process_decomp_events(session);
1915 if (!session_done())
1918 /* do the final flush for ordered samples */
1919 err = ordered_events__flush(oe, OE_FLUSH__FINAL);
1922 err = auxtrace__flush_events(session, tool);
1925 err = perf_session__flush_thread_stacks(session);
1929 perf_session__warn_about_errors(session);
1930 ordered_events__free(&session->ordered_events);
1931 auxtrace__free_events(session);
1935 static union perf_event *
1936 fetch_mmaped_event(struct perf_session *session,
1937 u64 head, size_t mmap_size, char *buf)
1939 union perf_event *event;
1942 * Ensure we have enough space remaining to read
1943 * the size of the event in the headers.
1945 if (head + sizeof(event->header) > mmap_size)
1948 event = (union perf_event *)(buf + head);
1950 if (session->header.needs_swap)
1951 perf_event_header__bswap(&event->header);
1953 if (head + event->header.size > mmap_size) {
1954 /* We're not fetching the event so swap back again */
1955 if (session->header.needs_swap)
1956 perf_event_header__bswap(&event->header);
1963 static int __perf_session__process_decomp_events(struct perf_session *session)
1966 u64 size, file_pos = 0;
1967 struct decomp *decomp = session->decomp_last;
1972 while (decomp->head < decomp->size && !session_done()) {
1973 union perf_event *event = fetch_mmaped_event(session, decomp->head, decomp->size, decomp->data);
1978 size = event->header.size;
1980 if (size < sizeof(struct perf_event_header) ||
1981 (skip = perf_session__process_event(session, event, file_pos)) < 0) {
1982 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1983 decomp->file_pos + decomp->head, event->header.size, event->header.type);
1990 decomp->head += size;
1997 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1998 * slices. On 32bit we use 32MB.
2000 #if BITS_PER_LONG == 64
2001 #define MMAP_SIZE ULLONG_MAX
2004 #define MMAP_SIZE (32 * 1024 * 1024ULL)
2005 #define NUM_MMAPS 128
2010 typedef s64 (*reader_cb_t)(struct perf_session *session,
2011 union perf_event *event,
2018 reader_cb_t process;
2022 reader__process_events(struct reader *rd, struct perf_session *session,
2023 struct ui_progress *prog)
2025 u64 data_size = rd->data_size;
2026 u64 head, page_offset, file_offset, file_pos, size;
2027 int err = 0, mmap_prot, mmap_flags, map_idx = 0;
2029 char *buf, *mmaps[NUM_MMAPS];
2030 union perf_event *event;
2033 page_offset = page_size * (rd->data_offset / page_size);
2034 file_offset = page_offset;
2035 head = rd->data_offset - page_offset;
2037 ui_progress__init_size(prog, data_size, "Processing events...");
2039 data_size += rd->data_offset;
2041 mmap_size = MMAP_SIZE;
2042 if (mmap_size > data_size) {
2043 mmap_size = data_size;
2044 session->one_mmap = true;
2047 memset(mmaps, 0, sizeof(mmaps));
2049 mmap_prot = PROT_READ;
2050 mmap_flags = MAP_SHARED;
2052 if (session->header.needs_swap) {
2053 mmap_prot |= PROT_WRITE;
2054 mmap_flags = MAP_PRIVATE;
2057 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, rd->fd,
2059 if (buf == MAP_FAILED) {
2060 pr_err("failed to mmap file\n");
2064 mmaps[map_idx] = buf;
2065 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
2066 file_pos = file_offset + head;
2067 if (session->one_mmap) {
2068 session->one_mmap_addr = buf;
2069 session->one_mmap_offset = file_offset;
2073 event = fetch_mmaped_event(session, head, mmap_size, buf);
2075 if (mmaps[map_idx]) {
2076 munmap(mmaps[map_idx], mmap_size);
2077 mmaps[map_idx] = NULL;
2080 page_offset = page_size * (head / page_size);
2081 file_offset += page_offset;
2082 head -= page_offset;
2086 size = event->header.size;
2090 if (size < sizeof(struct perf_event_header) ||
2091 (skip = rd->process(session, event, file_pos)) < 0) {
2092 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n",
2093 file_offset + head, event->header.size,
2094 event->header.type, strerror(-skip));
2105 err = __perf_session__process_decomp_events(session);
2109 ui_progress__update(prog, size);
2114 if (file_pos < data_size)
2121 static s64 process_simple(struct perf_session *session,
2122 union perf_event *event,
2125 return perf_session__process_event(session, event, file_offset);
2128 static int __perf_session__process_events(struct perf_session *session)
2130 struct reader rd = {
2131 .fd = perf_data__fd(session->data),
2132 .data_size = session->header.data_size,
2133 .data_offset = session->header.data_offset,
2134 .process = process_simple,
2136 struct ordered_events *oe = &session->ordered_events;
2137 struct perf_tool *tool = session->tool;
2138 struct ui_progress prog;
2141 perf_tool__fill_defaults(tool);
2143 if (rd.data_size == 0)
2146 ui_progress__init_size(&prog, rd.data_size, "Processing events...");
2148 err = reader__process_events(&rd, session, &prog);
2151 /* do the final flush for ordered samples */
2152 err = ordered_events__flush(oe, OE_FLUSH__FINAL);
2155 err = auxtrace__flush_events(session, tool);
2158 err = perf_session__flush_thread_stacks(session);
2160 ui_progress__finish();
2162 perf_session__warn_about_errors(session);
2164 * We may switching perf.data output, make ordered_events
2167 ordered_events__reinit(&session->ordered_events);
2168 auxtrace__free_events(session);
2169 session->one_mmap = false;
2173 int perf_session__process_events(struct perf_session *session)
2175 if (perf_session__register_idle_thread(session) < 0)
2178 if (perf_data__is_pipe(session->data))
2179 return __perf_session__process_pipe_events(session);
2181 return __perf_session__process_events(session);
2184 bool perf_session__has_traces(struct perf_session *session, const char *msg)
2186 struct evsel *evsel;
2188 evlist__for_each_entry(session->evlist, evsel) {
2189 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
2193 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
2197 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr)
2200 struct ref_reloc_sym *ref;
2203 ref = zalloc(sizeof(struct ref_reloc_sym));
2207 ref->name = strdup(symbol_name);
2208 if (ref->name == NULL) {
2213 bracket = strchr(ref->name, ']');
2219 kmap = map__kmap(map);
2221 kmap->ref_reloc_sym = ref;
2226 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
2228 return machines__fprintf_dsos(&session->machines, fp);
2231 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
2232 bool (skip)(struct dso *dso, int parm), int parm)
2234 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
2237 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
2240 const char *msg = "";
2242 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE))
2243 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)";
2245 ret = fprintf(fp, "\nAggregated stats:%s\n", msg);
2247 ret += events_stats__fprintf(&session->evlist->stats, fp);
2251 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
2254 * FIXME: Here we have to actually print all the machines in this
2255 * session, not just the host...
2257 return machine__fprintf(&session->machines.host, fp);
2260 struct evsel *perf_session__find_first_evtype(struct perf_session *session,
2265 evlist__for_each_entry(session->evlist, pos) {
2266 if (pos->attr.type == type)
2272 int perf_session__cpu_bitmap(struct perf_session *session,
2273 const char *cpu_list, unsigned long *cpu_bitmap)
2276 struct perf_cpu_map *map;
2278 for (i = 0; i < PERF_TYPE_MAX; ++i) {
2279 struct evsel *evsel;
2281 evsel = perf_session__find_first_evtype(session, i);
2285 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
2286 pr_err("File does not contain CPU events. "
2287 "Remove -C option to proceed.\n");
2292 map = cpu_map__new(cpu_list);
2294 pr_err("Invalid cpu_list\n");
2298 for (i = 0; i < map->nr; i++) {
2299 int cpu = map->map[i];
2301 if (cpu >= MAX_NR_CPUS) {
2302 pr_err("Requested CPU %d too large. "
2303 "Consider raising MAX_NR_CPUS\n", cpu);
2304 goto out_delete_map;
2307 set_bit(cpu, cpu_bitmap);
2317 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
2320 if (session == NULL || fp == NULL)
2323 fprintf(fp, "# ========\n");
2324 perf_header__fprintf_info(session, fp, full);
2325 fprintf(fp, "# ========\n#\n");
2329 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
2330 const struct evsel_str_handler *assocs,
2333 struct evsel *evsel;
2337 for (i = 0; i < nr_assocs; i++) {
2339 * Adding a handler for an event not in the session,
2342 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
2347 if (evsel->handler != NULL)
2349 evsel->handler = assocs[i].handler;
2357 int perf_event__process_id_index(struct perf_session *session,
2358 union perf_event *event)
2360 struct evlist *evlist = session->evlist;
2361 struct id_index_event *ie = &event->id_index;
2362 size_t i, nr, max_nr;
2364 max_nr = (ie->header.size - sizeof(struct id_index_event)) /
2365 sizeof(struct id_index_entry);
2371 fprintf(stdout, " nr: %zu\n", nr);
2373 for (i = 0; i < nr; i++) {
2374 struct id_index_entry *e = &ie->entries[i];
2375 struct perf_sample_id *sid;
2378 fprintf(stdout, " ... id: %"PRIu64, e->id);
2379 fprintf(stdout, " idx: %"PRIu64, e->idx);
2380 fprintf(stdout, " cpu: %"PRId64, e->cpu);
2381 fprintf(stdout, " tid: %"PRId64"\n", e->tid);
2384 sid = perf_evlist__id2sid(evlist, e->id);
2394 int perf_event__synthesize_id_index(struct perf_tool *tool,
2395 perf_event__handler_t process,
2396 struct evlist *evlist,
2397 struct machine *machine)
2399 union perf_event *ev;
2400 struct evsel *evsel;
2401 size_t nr = 0, i = 0, sz, max_nr, n;
2404 pr_debug2("Synthesizing id index\n");
2406 max_nr = (UINT16_MAX - sizeof(struct id_index_event)) /
2407 sizeof(struct id_index_entry);
2409 evlist__for_each_entry(evlist, evsel)
2412 n = nr > max_nr ? max_nr : nr;
2413 sz = sizeof(struct id_index_event) + n * sizeof(struct id_index_entry);
2418 ev->id_index.header.type = PERF_RECORD_ID_INDEX;
2419 ev->id_index.header.size = sz;
2420 ev->id_index.nr = n;
2422 evlist__for_each_entry(evlist, evsel) {
2425 for (j = 0; j < evsel->ids; j++) {
2426 struct id_index_entry *e;
2427 struct perf_sample_id *sid;
2430 err = process(tool, ev, NULL, machine);
2437 e = &ev->id_index.entries[i++];
2439 e->id = evsel->id[j];
2441 sid = perf_evlist__id2sid(evlist, e->id);
2453 sz = sizeof(struct id_index_event) + nr * sizeof(struct id_index_entry);
2454 ev->id_index.header.size = sz;
2455 ev->id_index.nr = nr;
2457 err = process(tool, ev, NULL, machine);