2 * intel_pt.c: Intel Processor Trace support
3 * Copyright (c) 2013-2015, Intel Corporation.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
19 #include <linux/kernel.h>
20 #include <linux/types.h>
34 #include "thread-stack.h"
36 #include "callchain.h"
44 #include "intel-pt-decoder/intel-pt-log.h"
45 #include "intel-pt-decoder/intel-pt-decoder.h"
46 #include "intel-pt-decoder/intel-pt-insn-decoder.h"
47 #include "intel-pt-decoder/intel-pt-pkt-decoder.h"
49 #define MAX_TIMESTAMP (~0ULL)
52 struct auxtrace auxtrace;
53 struct auxtrace_queues queues;
54 struct auxtrace_heap heap;
56 struct perf_session *session;
57 struct machine *machine;
58 struct perf_evsel *switch_evsel;
59 struct thread *unknown_thread;
60 bool timeless_decoding;
69 int have_sched_switch;
75 struct perf_tsc_conversion tc;
76 bool cap_user_time_zero;
78 struct itrace_synth_opts synth_opts;
80 bool sample_instructions;
81 u64 instructions_sample_type;
82 u64 instructions_sample_period;
87 u64 branches_sample_type;
90 bool sample_transactions;
91 u64 transactions_sample_type;
94 bool synth_needs_swap;
103 unsigned max_non_turbo_ratio;
105 unsigned long num_events;
109 INTEL_PT_SS_NOT_TRACING,
112 INTEL_PT_SS_EXPECTING_SWITCH_EVENT,
113 INTEL_PT_SS_EXPECTING_SWITCH_IP,
116 struct intel_pt_queue {
118 unsigned int queue_nr;
119 struct auxtrace_buffer *buffer;
121 const struct intel_pt_state *state;
122 struct ip_callchain *chain;
123 struct branch_stack *last_branch;
124 struct branch_stack *last_branch_rb;
125 size_t last_branch_pos;
126 union perf_event *event_buf;
129 bool step_through_buffers;
130 bool use_buffer_pid_tid;
135 struct thread *thread;
145 static void intel_pt_dump(struct intel_pt *pt __maybe_unused,
146 unsigned char *buf, size_t len)
148 struct intel_pt_pkt packet;
151 char desc[INTEL_PT_PKT_DESC_MAX];
152 const char *color = PERF_COLOR_BLUE;
154 color_fprintf(stdout, color,
155 ". ... Intel Processor Trace data: size %zu bytes\n",
159 ret = intel_pt_get_packet(buf, len, &packet);
165 color_fprintf(stdout, color, " %08x: ", pos);
166 for (i = 0; i < pkt_len; i++)
167 color_fprintf(stdout, color, " %02x", buf[i]);
169 color_fprintf(stdout, color, " ");
171 ret = intel_pt_pkt_desc(&packet, desc,
172 INTEL_PT_PKT_DESC_MAX);
174 color_fprintf(stdout, color, " %s\n", desc);
176 color_fprintf(stdout, color, " Bad packet!\n");
184 static void intel_pt_dump_event(struct intel_pt *pt, unsigned char *buf,
188 intel_pt_dump(pt, buf, len);
191 static int intel_pt_do_fix_overlap(struct intel_pt *pt, struct auxtrace_buffer *a,
192 struct auxtrace_buffer *b)
196 start = intel_pt_find_overlap(a->data, a->size, b->data, b->size,
200 b->use_size = b->data + b->size - start;
205 static void intel_pt_use_buffer_pid_tid(struct intel_pt_queue *ptq,
206 struct auxtrace_queue *queue,
207 struct auxtrace_buffer *buffer)
209 if (queue->cpu == -1 && buffer->cpu != -1)
210 ptq->cpu = buffer->cpu;
212 ptq->pid = buffer->pid;
213 ptq->tid = buffer->tid;
215 intel_pt_log("queue %u cpu %d pid %d tid %d\n",
216 ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
218 thread__zput(ptq->thread);
220 if (ptq->tid != -1) {
222 ptq->thread = machine__findnew_thread(ptq->pt->machine,
226 ptq->thread = machine__find_thread(ptq->pt->machine, -1,
231 /* This function assumes data is processed sequentially only */
232 static int intel_pt_get_trace(struct intel_pt_buffer *b, void *data)
234 struct intel_pt_queue *ptq = data;
235 struct auxtrace_buffer *buffer = ptq->buffer, *old_buffer = buffer;
236 struct auxtrace_queue *queue;
243 queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
245 buffer = auxtrace_buffer__next(queue, buffer);
248 auxtrace_buffer__drop_data(old_buffer);
253 ptq->buffer = buffer;
256 int fd = perf_data_file__fd(ptq->pt->session->file);
258 buffer->data = auxtrace_buffer__get_data(buffer, fd);
263 if (ptq->pt->snapshot_mode && !buffer->consecutive && old_buffer &&
264 intel_pt_do_fix_overlap(ptq->pt, old_buffer, buffer))
268 auxtrace_buffer__drop_data(old_buffer);
270 if (buffer->use_data) {
271 b->len = buffer->use_size;
272 b->buf = buffer->use_data;
274 b->len = buffer->size;
275 b->buf = buffer->data;
277 b->ref_timestamp = buffer->reference;
279 if (!old_buffer || ptq->pt->sampling_mode || (ptq->pt->snapshot_mode &&
280 !buffer->consecutive)) {
281 b->consecutive = false;
282 b->trace_nr = buffer->buffer_nr + 1;
284 b->consecutive = true;
287 if (ptq->use_buffer_pid_tid && (ptq->pid != buffer->pid ||
288 ptq->tid != buffer->tid))
289 intel_pt_use_buffer_pid_tid(ptq, queue, buffer);
291 if (ptq->step_through_buffers)
295 return intel_pt_get_trace(b, data);
300 struct intel_pt_cache_entry {
301 struct auxtrace_cache_entry entry;
304 enum intel_pt_insn_op op;
305 enum intel_pt_insn_branch branch;
310 static int intel_pt_config_div(const char *var, const char *value, void *data)
315 if (!strcmp(var, "intel-pt.cache-divisor")) {
316 val = strtol(value, NULL, 0);
317 if (val > 0 && val <= INT_MAX)
324 static int intel_pt_cache_divisor(void)
331 perf_config(intel_pt_config_div, &d);
339 static unsigned int intel_pt_cache_size(struct dso *dso,
340 struct machine *machine)
344 size = dso__data_size(dso, machine);
345 size /= intel_pt_cache_divisor();
348 if (size > (1 << 21))
350 return 32 - __builtin_clz(size);
353 static struct auxtrace_cache *intel_pt_cache(struct dso *dso,
354 struct machine *machine)
356 struct auxtrace_cache *c;
359 if (dso->auxtrace_cache)
360 return dso->auxtrace_cache;
362 bits = intel_pt_cache_size(dso, machine);
364 /* Ignoring cache creation failure */
365 c = auxtrace_cache__new(bits, sizeof(struct intel_pt_cache_entry), 200);
367 dso->auxtrace_cache = c;
372 static int intel_pt_cache_add(struct dso *dso, struct machine *machine,
373 u64 offset, u64 insn_cnt, u64 byte_cnt,
374 struct intel_pt_insn *intel_pt_insn)
376 struct auxtrace_cache *c = intel_pt_cache(dso, machine);
377 struct intel_pt_cache_entry *e;
383 e = auxtrace_cache__alloc_entry(c);
387 e->insn_cnt = insn_cnt;
388 e->byte_cnt = byte_cnt;
389 e->op = intel_pt_insn->op;
390 e->branch = intel_pt_insn->branch;
391 e->length = intel_pt_insn->length;
392 e->rel = intel_pt_insn->rel;
394 err = auxtrace_cache__add(c, offset, &e->entry);
396 auxtrace_cache__free_entry(c, e);
401 static struct intel_pt_cache_entry *
402 intel_pt_cache_lookup(struct dso *dso, struct machine *machine, u64 offset)
404 struct auxtrace_cache *c = intel_pt_cache(dso, machine);
409 return auxtrace_cache__lookup(dso->auxtrace_cache, offset);
412 static int intel_pt_walk_next_insn(struct intel_pt_insn *intel_pt_insn,
413 uint64_t *insn_cnt_ptr, uint64_t *ip,
414 uint64_t to_ip, uint64_t max_insn_cnt,
417 struct intel_pt_queue *ptq = data;
418 struct machine *machine = ptq->pt->machine;
419 struct thread *thread;
420 struct addr_location al;
421 unsigned char buf[1024];
426 u64 offset, start_offset, start_ip;
430 if (to_ip && *ip == to_ip)
433 bufsz = intel_pt_insn_max_size();
435 if (*ip >= ptq->pt->kernel_start)
436 cpumode = PERF_RECORD_MISC_KERNEL;
438 cpumode = PERF_RECORD_MISC_USER;
440 thread = ptq->thread;
442 if (cpumode != PERF_RECORD_MISC_KERNEL)
444 thread = ptq->pt->unknown_thread;
448 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, *ip, &al);
449 if (!al.map || !al.map->dso)
452 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
453 dso__data_status_seen(al.map->dso,
454 DSO_DATA_STATUS_SEEN_ITRACE))
457 offset = al.map->map_ip(al.map, *ip);
459 if (!to_ip && one_map) {
460 struct intel_pt_cache_entry *e;
462 e = intel_pt_cache_lookup(al.map->dso, machine, offset);
464 (!max_insn_cnt || e->insn_cnt <= max_insn_cnt)) {
465 *insn_cnt_ptr = e->insn_cnt;
467 intel_pt_insn->op = e->op;
468 intel_pt_insn->branch = e->branch;
469 intel_pt_insn->length = e->length;
470 intel_pt_insn->rel = e->rel;
471 intel_pt_log_insn_no_data(intel_pt_insn, *ip);
476 start_offset = offset;
479 /* Load maps to ensure dso->is_64_bit has been updated */
480 map__load(al.map, machine->symbol_filter);
482 x86_64 = al.map->dso->is_64_bit;
485 len = dso__data_read_offset(al.map->dso, machine,
490 if (intel_pt_get_insn(buf, len, x86_64, intel_pt_insn))
493 intel_pt_log_insn(intel_pt_insn, *ip);
497 if (intel_pt_insn->branch != INTEL_PT_BR_NO_BRANCH)
500 if (max_insn_cnt && insn_cnt >= max_insn_cnt)
503 *ip += intel_pt_insn->length;
505 if (to_ip && *ip == to_ip)
508 if (*ip >= al.map->end)
511 offset += intel_pt_insn->length;
516 *insn_cnt_ptr = insn_cnt;
522 * Didn't lookup in the 'to_ip' case, so do it now to prevent duplicate
526 struct intel_pt_cache_entry *e;
528 e = intel_pt_cache_lookup(al.map->dso, machine, start_offset);
533 /* Ignore cache errors */
534 intel_pt_cache_add(al.map->dso, machine, start_offset, insn_cnt,
535 *ip - start_ip, intel_pt_insn);
540 *insn_cnt_ptr = insn_cnt;
544 static bool intel_pt_get_config(struct intel_pt *pt,
545 struct perf_event_attr *attr, u64 *config)
547 if (attr->type == pt->pmu_type) {
549 *config = attr->config;
556 static bool intel_pt_exclude_kernel(struct intel_pt *pt)
558 struct perf_evsel *evsel;
560 evlist__for_each_entry(pt->session->evlist, evsel) {
561 if (intel_pt_get_config(pt, &evsel->attr, NULL) &&
562 !evsel->attr.exclude_kernel)
568 static bool intel_pt_return_compression(struct intel_pt *pt)
570 struct perf_evsel *evsel;
573 if (!pt->noretcomp_bit)
576 evlist__for_each_entry(pt->session->evlist, evsel) {
577 if (intel_pt_get_config(pt, &evsel->attr, &config) &&
578 (config & pt->noretcomp_bit))
584 static unsigned int intel_pt_mtc_period(struct intel_pt *pt)
586 struct perf_evsel *evsel;
590 if (!pt->mtc_freq_bits)
593 for (shift = 0, config = pt->mtc_freq_bits; !(config & 1); shift++)
596 evlist__for_each_entry(pt->session->evlist, evsel) {
597 if (intel_pt_get_config(pt, &evsel->attr, &config))
598 return (config & pt->mtc_freq_bits) >> shift;
603 static bool intel_pt_timeless_decoding(struct intel_pt *pt)
605 struct perf_evsel *evsel;
606 bool timeless_decoding = true;
609 if (!pt->tsc_bit || !pt->cap_user_time_zero)
612 evlist__for_each_entry(pt->session->evlist, evsel) {
613 if (!(evsel->attr.sample_type & PERF_SAMPLE_TIME))
615 if (intel_pt_get_config(pt, &evsel->attr, &config)) {
616 if (config & pt->tsc_bit)
617 timeless_decoding = false;
622 return timeless_decoding;
625 static bool intel_pt_tracing_kernel(struct intel_pt *pt)
627 struct perf_evsel *evsel;
629 evlist__for_each_entry(pt->session->evlist, evsel) {
630 if (intel_pt_get_config(pt, &evsel->attr, NULL) &&
631 !evsel->attr.exclude_kernel)
637 static bool intel_pt_have_tsc(struct intel_pt *pt)
639 struct perf_evsel *evsel;
640 bool have_tsc = false;
646 evlist__for_each_entry(pt->session->evlist, evsel) {
647 if (intel_pt_get_config(pt, &evsel->attr, &config)) {
648 if (config & pt->tsc_bit)
657 static u64 intel_pt_ns_to_ticks(const struct intel_pt *pt, u64 ns)
661 quot = ns / pt->tc.time_mult;
662 rem = ns % pt->tc.time_mult;
663 return (quot << pt->tc.time_shift) + (rem << pt->tc.time_shift) /
667 static struct intel_pt_queue *intel_pt_alloc_queue(struct intel_pt *pt,
668 unsigned int queue_nr)
670 struct intel_pt_params params = { .get_trace = 0, };
671 struct intel_pt_queue *ptq;
673 ptq = zalloc(sizeof(struct intel_pt_queue));
677 if (pt->synth_opts.callchain) {
678 size_t sz = sizeof(struct ip_callchain);
680 sz += pt->synth_opts.callchain_sz * sizeof(u64);
681 ptq->chain = zalloc(sz);
686 if (pt->synth_opts.last_branch) {
687 size_t sz = sizeof(struct branch_stack);
689 sz += pt->synth_opts.last_branch_sz *
690 sizeof(struct branch_entry);
691 ptq->last_branch = zalloc(sz);
692 if (!ptq->last_branch)
694 ptq->last_branch_rb = zalloc(sz);
695 if (!ptq->last_branch_rb)
699 ptq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
704 ptq->queue_nr = queue_nr;
705 ptq->exclude_kernel = intel_pt_exclude_kernel(pt);
711 params.get_trace = intel_pt_get_trace;
712 params.walk_insn = intel_pt_walk_next_insn;
714 params.return_compression = intel_pt_return_compression(pt);
715 params.max_non_turbo_ratio = pt->max_non_turbo_ratio;
716 params.mtc_period = intel_pt_mtc_period(pt);
717 params.tsc_ctc_ratio_n = pt->tsc_ctc_ratio_n;
718 params.tsc_ctc_ratio_d = pt->tsc_ctc_ratio_d;
720 if (pt->synth_opts.instructions) {
721 if (pt->synth_opts.period) {
722 switch (pt->synth_opts.period_type) {
723 case PERF_ITRACE_PERIOD_INSTRUCTIONS:
725 INTEL_PT_PERIOD_INSTRUCTIONS;
726 params.period = pt->synth_opts.period;
728 case PERF_ITRACE_PERIOD_TICKS:
729 params.period_type = INTEL_PT_PERIOD_TICKS;
730 params.period = pt->synth_opts.period;
732 case PERF_ITRACE_PERIOD_NANOSECS:
733 params.period_type = INTEL_PT_PERIOD_TICKS;
734 params.period = intel_pt_ns_to_ticks(pt,
735 pt->synth_opts.period);
742 if (!params.period) {
743 params.period_type = INTEL_PT_PERIOD_INSTRUCTIONS;
748 ptq->decoder = intel_pt_decoder_new(¶ms);
755 zfree(&ptq->event_buf);
756 zfree(&ptq->last_branch);
757 zfree(&ptq->last_branch_rb);
763 static void intel_pt_free_queue(void *priv)
765 struct intel_pt_queue *ptq = priv;
769 thread__zput(ptq->thread);
770 intel_pt_decoder_free(ptq->decoder);
771 zfree(&ptq->event_buf);
772 zfree(&ptq->last_branch);
773 zfree(&ptq->last_branch_rb);
778 static void intel_pt_set_pid_tid_cpu(struct intel_pt *pt,
779 struct auxtrace_queue *queue)
781 struct intel_pt_queue *ptq = queue->priv;
783 if (queue->tid == -1 || pt->have_sched_switch) {
784 ptq->tid = machine__get_current_tid(pt->machine, ptq->cpu);
785 thread__zput(ptq->thread);
788 if (!ptq->thread && ptq->tid != -1)
789 ptq->thread = machine__find_thread(pt->machine, -1, ptq->tid);
792 ptq->pid = ptq->thread->pid_;
793 if (queue->cpu == -1)
794 ptq->cpu = ptq->thread->cpu;
798 static void intel_pt_sample_flags(struct intel_pt_queue *ptq)
800 if (ptq->state->flags & INTEL_PT_ABORT_TX) {
801 ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_TX_ABORT;
802 } else if (ptq->state->flags & INTEL_PT_ASYNC) {
803 if (ptq->state->to_ip)
804 ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_CALL |
806 PERF_IP_FLAG_INTERRUPT;
808 ptq->flags = PERF_IP_FLAG_BRANCH |
809 PERF_IP_FLAG_TRACE_END;
812 if (ptq->state->from_ip)
813 ptq->flags = intel_pt_insn_type(ptq->state->insn_op);
815 ptq->flags = PERF_IP_FLAG_BRANCH |
816 PERF_IP_FLAG_TRACE_BEGIN;
817 if (ptq->state->flags & INTEL_PT_IN_TX)
818 ptq->flags |= PERF_IP_FLAG_IN_TX;
819 ptq->insn_len = ptq->state->insn_len;
823 static int intel_pt_setup_queue(struct intel_pt *pt,
824 struct auxtrace_queue *queue,
825 unsigned int queue_nr)
827 struct intel_pt_queue *ptq = queue->priv;
829 if (list_empty(&queue->head))
833 ptq = intel_pt_alloc_queue(pt, queue_nr);
838 if (queue->cpu != -1)
839 ptq->cpu = queue->cpu;
840 ptq->tid = queue->tid;
842 if (pt->sampling_mode) {
843 if (pt->timeless_decoding)
844 ptq->step_through_buffers = true;
845 if (pt->timeless_decoding || !pt->have_sched_switch)
846 ptq->use_buffer_pid_tid = true;
852 ptq->switch_state != INTEL_PT_SS_EXPECTING_SWITCH_EVENT)) {
853 const struct intel_pt_state *state;
856 if (pt->timeless_decoding)
859 intel_pt_log("queue %u getting timestamp\n", queue_nr);
860 intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
861 queue_nr, ptq->cpu, ptq->pid, ptq->tid);
863 state = intel_pt_decode(ptq->decoder);
865 if (state->err == INTEL_PT_ERR_NODATA) {
866 intel_pt_log("queue %u has no timestamp\n",
872 if (state->timestamp)
876 ptq->timestamp = state->timestamp;
877 intel_pt_log("queue %u timestamp 0x%" PRIx64 "\n",
878 queue_nr, ptq->timestamp);
880 ptq->have_sample = true;
881 intel_pt_sample_flags(ptq);
882 ret = auxtrace_heap__add(&pt->heap, queue_nr, ptq->timestamp);
891 static int intel_pt_setup_queues(struct intel_pt *pt)
896 for (i = 0; i < pt->queues.nr_queues; i++) {
897 ret = intel_pt_setup_queue(pt, &pt->queues.queue_array[i], i);
904 static inline void intel_pt_copy_last_branch_rb(struct intel_pt_queue *ptq)
906 struct branch_stack *bs_src = ptq->last_branch_rb;
907 struct branch_stack *bs_dst = ptq->last_branch;
910 bs_dst->nr = bs_src->nr;
915 nr = ptq->pt->synth_opts.last_branch_sz - ptq->last_branch_pos;
916 memcpy(&bs_dst->entries[0],
917 &bs_src->entries[ptq->last_branch_pos],
918 sizeof(struct branch_entry) * nr);
920 if (bs_src->nr >= ptq->pt->synth_opts.last_branch_sz) {
921 memcpy(&bs_dst->entries[nr],
923 sizeof(struct branch_entry) * ptq->last_branch_pos);
927 static inline void intel_pt_reset_last_branch_rb(struct intel_pt_queue *ptq)
929 ptq->last_branch_pos = 0;
930 ptq->last_branch_rb->nr = 0;
933 static void intel_pt_update_last_branch_rb(struct intel_pt_queue *ptq)
935 const struct intel_pt_state *state = ptq->state;
936 struct branch_stack *bs = ptq->last_branch_rb;
937 struct branch_entry *be;
939 if (!ptq->last_branch_pos)
940 ptq->last_branch_pos = ptq->pt->synth_opts.last_branch_sz;
942 ptq->last_branch_pos -= 1;
944 be = &bs->entries[ptq->last_branch_pos];
945 be->from = state->from_ip;
946 be->to = state->to_ip;
947 be->flags.abort = !!(state->flags & INTEL_PT_ABORT_TX);
948 be->flags.in_tx = !!(state->flags & INTEL_PT_IN_TX);
949 /* No support for mispredict */
950 be->flags.mispred = ptq->pt->mispred_all;
952 if (bs->nr < ptq->pt->synth_opts.last_branch_sz)
956 static int intel_pt_inject_event(union perf_event *event,
957 struct perf_sample *sample, u64 type,
960 event->header.size = perf_event__sample_event_size(sample, type, 0);
961 return perf_event__synthesize_sample(event, type, 0, sample, swapped);
964 static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq)
967 struct intel_pt *pt = ptq->pt;
968 union perf_event *event = ptq->event_buf;
969 struct perf_sample sample = { .ip = 0, };
970 struct dummy_branch_stack {
972 struct branch_entry entries;
975 if (pt->branches_filter && !(pt->branches_filter & ptq->flags))
978 if (pt->synth_opts.initial_skip &&
979 pt->num_events++ < pt->synth_opts.initial_skip)
982 event->sample.header.type = PERF_RECORD_SAMPLE;
983 event->sample.header.misc = PERF_RECORD_MISC_USER;
984 event->sample.header.size = sizeof(struct perf_event_header);
986 if (!pt->timeless_decoding)
987 sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
989 sample.cpumode = PERF_RECORD_MISC_USER;
990 sample.ip = ptq->state->from_ip;
991 sample.pid = ptq->pid;
992 sample.tid = ptq->tid;
993 sample.addr = ptq->state->to_ip;
994 sample.id = ptq->pt->branches_id;
995 sample.stream_id = ptq->pt->branches_id;
997 sample.cpu = ptq->cpu;
998 sample.flags = ptq->flags;
999 sample.insn_len = ptq->insn_len;
1002 * perf report cannot handle events without a branch stack when using
1003 * SORT_MODE__BRANCH so make a dummy one.
1005 if (pt->synth_opts.last_branch && sort__mode == SORT_MODE__BRANCH) {
1006 dummy_bs = (struct dummy_branch_stack){
1013 sample.branch_stack = (struct branch_stack *)&dummy_bs;
1016 if (pt->synth_opts.inject) {
1017 ret = intel_pt_inject_event(event, &sample,
1018 pt->branches_sample_type,
1019 pt->synth_needs_swap);
1024 ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1026 pr_err("Intel Processor Trace: failed to deliver branch event, error %d\n",
1032 static int intel_pt_synth_instruction_sample(struct intel_pt_queue *ptq)
1035 struct intel_pt *pt = ptq->pt;
1036 union perf_event *event = ptq->event_buf;
1037 struct perf_sample sample = { .ip = 0, };
1039 if (pt->synth_opts.initial_skip &&
1040 pt->num_events++ < pt->synth_opts.initial_skip)
1043 event->sample.header.type = PERF_RECORD_SAMPLE;
1044 event->sample.header.misc = PERF_RECORD_MISC_USER;
1045 event->sample.header.size = sizeof(struct perf_event_header);
1047 if (!pt->timeless_decoding)
1048 sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1050 sample.cpumode = PERF_RECORD_MISC_USER;
1051 sample.ip = ptq->state->from_ip;
1052 sample.pid = ptq->pid;
1053 sample.tid = ptq->tid;
1054 sample.addr = ptq->state->to_ip;
1055 sample.id = ptq->pt->instructions_id;
1056 sample.stream_id = ptq->pt->instructions_id;
1057 sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt;
1058 sample.cpu = ptq->cpu;
1059 sample.flags = ptq->flags;
1060 sample.insn_len = ptq->insn_len;
1062 ptq->last_insn_cnt = ptq->state->tot_insn_cnt;
1064 if (pt->synth_opts.callchain) {
1065 thread_stack__sample(ptq->thread, ptq->chain,
1066 pt->synth_opts.callchain_sz, sample.ip);
1067 sample.callchain = ptq->chain;
1070 if (pt->synth_opts.last_branch) {
1071 intel_pt_copy_last_branch_rb(ptq);
1072 sample.branch_stack = ptq->last_branch;
1075 if (pt->synth_opts.inject) {
1076 ret = intel_pt_inject_event(event, &sample,
1077 pt->instructions_sample_type,
1078 pt->synth_needs_swap);
1083 ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1085 pr_err("Intel Processor Trace: failed to deliver instruction event, error %d\n",
1088 if (pt->synth_opts.last_branch)
1089 intel_pt_reset_last_branch_rb(ptq);
1094 static int intel_pt_synth_transaction_sample(struct intel_pt_queue *ptq)
1097 struct intel_pt *pt = ptq->pt;
1098 union perf_event *event = ptq->event_buf;
1099 struct perf_sample sample = { .ip = 0, };
1101 if (pt->synth_opts.initial_skip &&
1102 pt->num_events++ < pt->synth_opts.initial_skip)
1105 event->sample.header.type = PERF_RECORD_SAMPLE;
1106 event->sample.header.misc = PERF_RECORD_MISC_USER;
1107 event->sample.header.size = sizeof(struct perf_event_header);
1109 if (!pt->timeless_decoding)
1110 sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1112 sample.cpumode = PERF_RECORD_MISC_USER;
1113 sample.ip = ptq->state->from_ip;
1114 sample.pid = ptq->pid;
1115 sample.tid = ptq->tid;
1116 sample.addr = ptq->state->to_ip;
1117 sample.id = ptq->pt->transactions_id;
1118 sample.stream_id = ptq->pt->transactions_id;
1120 sample.cpu = ptq->cpu;
1121 sample.flags = ptq->flags;
1122 sample.insn_len = ptq->insn_len;
1124 if (pt->synth_opts.callchain) {
1125 thread_stack__sample(ptq->thread, ptq->chain,
1126 pt->synth_opts.callchain_sz, sample.ip);
1127 sample.callchain = ptq->chain;
1130 if (pt->synth_opts.last_branch) {
1131 intel_pt_copy_last_branch_rb(ptq);
1132 sample.branch_stack = ptq->last_branch;
1135 if (pt->synth_opts.inject) {
1136 ret = intel_pt_inject_event(event, &sample,
1137 pt->transactions_sample_type,
1138 pt->synth_needs_swap);
1143 ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1145 pr_err("Intel Processor Trace: failed to deliver transaction event, error %d\n",
1148 if (pt->synth_opts.last_branch)
1149 intel_pt_reset_last_branch_rb(ptq);
1154 static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,
1155 pid_t pid, pid_t tid, u64 ip)
1157 union perf_event event;
1158 char msg[MAX_AUXTRACE_ERROR_MSG];
1161 intel_pt__strerror(code, msg, MAX_AUXTRACE_ERROR_MSG);
1163 auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
1164 code, cpu, pid, tid, ip, msg);
1166 err = perf_session__deliver_synth_event(pt->session, &event, NULL);
1168 pr_err("Intel Processor Trace: failed to deliver error event, error %d\n",
1174 static int intel_pt_next_tid(struct intel_pt *pt, struct intel_pt_queue *ptq)
1176 struct auxtrace_queue *queue;
1177 pid_t tid = ptq->next_tid;
1183 intel_pt_log("switch: cpu %d tid %d\n", ptq->cpu, tid);
1185 err = machine__set_current_tid(pt->machine, ptq->cpu, -1, tid);
1187 queue = &pt->queues.queue_array[ptq->queue_nr];
1188 intel_pt_set_pid_tid_cpu(pt, queue);
1195 static inline bool intel_pt_is_switch_ip(struct intel_pt_queue *ptq, u64 ip)
1197 struct intel_pt *pt = ptq->pt;
1199 return ip == pt->switch_ip &&
1200 (ptq->flags & PERF_IP_FLAG_BRANCH) &&
1201 !(ptq->flags & (PERF_IP_FLAG_CONDITIONAL | PERF_IP_FLAG_ASYNC |
1202 PERF_IP_FLAG_INTERRUPT | PERF_IP_FLAG_TX_ABORT));
1205 static int intel_pt_sample(struct intel_pt_queue *ptq)
1207 const struct intel_pt_state *state = ptq->state;
1208 struct intel_pt *pt = ptq->pt;
1211 if (!ptq->have_sample)
1214 ptq->have_sample = false;
1216 if (pt->sample_instructions &&
1217 (state->type & INTEL_PT_INSTRUCTION) &&
1218 (!pt->synth_opts.initial_skip ||
1219 pt->num_events++ >= pt->synth_opts.initial_skip)) {
1220 err = intel_pt_synth_instruction_sample(ptq);
1225 if (pt->sample_transactions &&
1226 (state->type & INTEL_PT_TRANSACTION) &&
1227 (!pt->synth_opts.initial_skip ||
1228 pt->num_events++ >= pt->synth_opts.initial_skip)) {
1229 err = intel_pt_synth_transaction_sample(ptq);
1234 if (!(state->type & INTEL_PT_BRANCH))
1237 if (pt->synth_opts.callchain)
1238 thread_stack__event(ptq->thread, ptq->flags, state->from_ip,
1239 state->to_ip, ptq->insn_len,
1242 thread_stack__set_trace_nr(ptq->thread, state->trace_nr);
1244 if (pt->sample_branches) {
1245 err = intel_pt_synth_branch_sample(ptq);
1250 if (pt->synth_opts.last_branch)
1251 intel_pt_update_last_branch_rb(ptq);
1253 if (!pt->sync_switch)
1256 if (intel_pt_is_switch_ip(ptq, state->to_ip)) {
1257 switch (ptq->switch_state) {
1258 case INTEL_PT_SS_UNKNOWN:
1259 case INTEL_PT_SS_EXPECTING_SWITCH_IP:
1260 err = intel_pt_next_tid(pt, ptq);
1263 ptq->switch_state = INTEL_PT_SS_TRACING;
1266 ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_EVENT;
1269 } else if (!state->to_ip) {
1270 ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
1271 } else if (ptq->switch_state == INTEL_PT_SS_NOT_TRACING) {
1272 ptq->switch_state = INTEL_PT_SS_UNKNOWN;
1273 } else if (ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
1274 state->to_ip == pt->ptss_ip &&
1275 (ptq->flags & PERF_IP_FLAG_CALL)) {
1276 ptq->switch_state = INTEL_PT_SS_TRACING;
1282 static u64 intel_pt_switch_ip(struct intel_pt *pt, u64 *ptss_ip)
1284 struct machine *machine = pt->machine;
1286 struct symbol *sym, *start;
1287 u64 ip, switch_ip = 0;
1293 map = machine__kernel_map(machine);
1297 if (map__load(map, machine->symbol_filter))
1300 start = dso__first_symbol(map->dso, MAP__FUNCTION);
1302 for (sym = start; sym; sym = dso__next_symbol(sym)) {
1303 if (sym->binding == STB_GLOBAL &&
1304 !strcmp(sym->name, "__switch_to")) {
1305 ip = map->unmap_ip(map, sym->start);
1306 if (ip >= map->start && ip < map->end) {
1313 if (!switch_ip || !ptss_ip)
1316 if (pt->have_sched_switch == 1)
1317 ptss = "perf_trace_sched_switch";
1319 ptss = "__perf_event_task_sched_out";
1321 for (sym = start; sym; sym = dso__next_symbol(sym)) {
1322 if (!strcmp(sym->name, ptss)) {
1323 ip = map->unmap_ip(map, sym->start);
1324 if (ip >= map->start && ip < map->end) {
1334 static int intel_pt_run_decoder(struct intel_pt_queue *ptq, u64 *timestamp)
1336 const struct intel_pt_state *state = ptq->state;
1337 struct intel_pt *pt = ptq->pt;
1340 if (!pt->kernel_start) {
1341 pt->kernel_start = machine__kernel_start(pt->machine);
1342 if (pt->per_cpu_mmaps &&
1343 (pt->have_sched_switch == 1 || pt->have_sched_switch == 3) &&
1344 !pt->timeless_decoding && intel_pt_tracing_kernel(pt) &&
1345 !pt->sampling_mode) {
1346 pt->switch_ip = intel_pt_switch_ip(pt, &pt->ptss_ip);
1347 if (pt->switch_ip) {
1348 intel_pt_log("switch_ip: %"PRIx64" ptss_ip: %"PRIx64"\n",
1349 pt->switch_ip, pt->ptss_ip);
1350 pt->sync_switch = true;
1355 intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
1356 ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
1358 err = intel_pt_sample(ptq);
1362 state = intel_pt_decode(ptq->decoder);
1364 if (state->err == INTEL_PT_ERR_NODATA)
1366 if (pt->sync_switch &&
1367 state->from_ip >= pt->kernel_start) {
1368 pt->sync_switch = false;
1369 intel_pt_next_tid(pt, ptq);
1371 if (pt->synth_opts.errors) {
1372 err = intel_pt_synth_error(pt, state->err,
1383 ptq->have_sample = true;
1384 intel_pt_sample_flags(ptq);
1386 /* Use estimated TSC upon return to user space */
1388 (state->from_ip >= pt->kernel_start || !state->from_ip) &&
1389 state->to_ip && state->to_ip < pt->kernel_start) {
1390 intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
1391 state->timestamp, state->est_timestamp);
1392 ptq->timestamp = state->est_timestamp;
1393 /* Use estimated TSC in unknown switch state */
1394 } else if (pt->sync_switch &&
1395 ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
1396 intel_pt_is_switch_ip(ptq, state->to_ip) &&
1397 ptq->next_tid == -1) {
1398 intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
1399 state->timestamp, state->est_timestamp);
1400 ptq->timestamp = state->est_timestamp;
1401 } else if (state->timestamp > ptq->timestamp) {
1402 ptq->timestamp = state->timestamp;
1405 if (!pt->timeless_decoding && ptq->timestamp >= *timestamp) {
1406 *timestamp = ptq->timestamp;
1413 static inline int intel_pt_update_queues(struct intel_pt *pt)
1415 if (pt->queues.new_data) {
1416 pt->queues.new_data = false;
1417 return intel_pt_setup_queues(pt);
1422 static int intel_pt_process_queues(struct intel_pt *pt, u64 timestamp)
1424 unsigned int queue_nr;
1429 struct auxtrace_queue *queue;
1430 struct intel_pt_queue *ptq;
1432 if (!pt->heap.heap_cnt)
1435 if (pt->heap.heap_array[0].ordinal >= timestamp)
1438 queue_nr = pt->heap.heap_array[0].queue_nr;
1439 queue = &pt->queues.queue_array[queue_nr];
1442 intel_pt_log("queue %u processing 0x%" PRIx64 " to 0x%" PRIx64 "\n",
1443 queue_nr, pt->heap.heap_array[0].ordinal,
1446 auxtrace_heap__pop(&pt->heap);
1448 if (pt->heap.heap_cnt) {
1449 ts = pt->heap.heap_array[0].ordinal + 1;
1456 intel_pt_set_pid_tid_cpu(pt, queue);
1458 ret = intel_pt_run_decoder(ptq, &ts);
1461 auxtrace_heap__add(&pt->heap, queue_nr, ts);
1466 ret = auxtrace_heap__add(&pt->heap, queue_nr, ts);
1470 ptq->on_heap = false;
1477 static int intel_pt_process_timeless_queues(struct intel_pt *pt, pid_t tid,
1480 struct auxtrace_queues *queues = &pt->queues;
1484 for (i = 0; i < queues->nr_queues; i++) {
1485 struct auxtrace_queue *queue = &pt->queues.queue_array[i];
1486 struct intel_pt_queue *ptq = queue->priv;
1488 if (ptq && (tid == -1 || ptq->tid == tid)) {
1490 intel_pt_set_pid_tid_cpu(pt, queue);
1491 intel_pt_run_decoder(ptq, &ts);
1497 static int intel_pt_lost(struct intel_pt *pt, struct perf_sample *sample)
1499 return intel_pt_synth_error(pt, INTEL_PT_ERR_LOST, sample->cpu,
1500 sample->pid, sample->tid, 0);
1503 static struct intel_pt_queue *intel_pt_cpu_to_ptq(struct intel_pt *pt, int cpu)
1507 if (cpu < 0 || !pt->queues.nr_queues)
1510 if ((unsigned)cpu >= pt->queues.nr_queues)
1511 i = pt->queues.nr_queues - 1;
1515 if (pt->queues.queue_array[i].cpu == cpu)
1516 return pt->queues.queue_array[i].priv;
1518 for (j = 0; i > 0; j++) {
1519 if (pt->queues.queue_array[--i].cpu == cpu)
1520 return pt->queues.queue_array[i].priv;
1523 for (; j < pt->queues.nr_queues; j++) {
1524 if (pt->queues.queue_array[j].cpu == cpu)
1525 return pt->queues.queue_array[j].priv;
1531 static int intel_pt_sync_switch(struct intel_pt *pt, int cpu, pid_t tid,
1534 struct intel_pt_queue *ptq;
1537 if (!pt->sync_switch)
1540 ptq = intel_pt_cpu_to_ptq(pt, cpu);
1544 switch (ptq->switch_state) {
1545 case INTEL_PT_SS_NOT_TRACING:
1548 case INTEL_PT_SS_UNKNOWN:
1549 case INTEL_PT_SS_TRACING:
1550 ptq->next_tid = tid;
1551 ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_IP;
1553 case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
1554 if (!ptq->on_heap) {
1555 ptq->timestamp = perf_time_to_tsc(timestamp,
1557 err = auxtrace_heap__add(&pt->heap, ptq->queue_nr,
1561 ptq->on_heap = true;
1563 ptq->switch_state = INTEL_PT_SS_TRACING;
1565 case INTEL_PT_SS_EXPECTING_SWITCH_IP:
1566 ptq->next_tid = tid;
1567 intel_pt_log("ERROR: cpu %d expecting switch ip\n", cpu);
1576 static int intel_pt_process_switch(struct intel_pt *pt,
1577 struct perf_sample *sample)
1579 struct perf_evsel *evsel;
1583 evsel = perf_evlist__id2evsel(pt->session->evlist, sample->id);
1584 if (evsel != pt->switch_evsel)
1587 tid = perf_evsel__intval(evsel, sample, "next_pid");
1590 intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1591 cpu, tid, sample->time, perf_time_to_tsc(sample->time,
1594 ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
1598 return machine__set_current_tid(pt->machine, cpu, -1, tid);
1601 static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,
1602 struct perf_sample *sample)
1604 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1610 if (pt->have_sched_switch == 3) {
1613 if (event->header.type != PERF_RECORD_SWITCH_CPU_WIDE) {
1614 pr_err("Expecting CPU-wide context switch event\n");
1617 pid = event->context_switch.next_prev_pid;
1618 tid = event->context_switch.next_prev_tid;
1627 pr_err("context_switch event has no tid\n");
1631 intel_pt_log("context_switch: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1632 cpu, pid, tid, sample->time, perf_time_to_tsc(sample->time,
1635 ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
1639 return machine__set_current_tid(pt->machine, cpu, pid, tid);
1642 static int intel_pt_process_itrace_start(struct intel_pt *pt,
1643 union perf_event *event,
1644 struct perf_sample *sample)
1646 if (!pt->per_cpu_mmaps)
1649 intel_pt_log("itrace_start: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1650 sample->cpu, event->itrace_start.pid,
1651 event->itrace_start.tid, sample->time,
1652 perf_time_to_tsc(sample->time, &pt->tc));
1654 return machine__set_current_tid(pt->machine, sample->cpu,
1655 event->itrace_start.pid,
1656 event->itrace_start.tid);
1659 static int intel_pt_process_event(struct perf_session *session,
1660 union perf_event *event,
1661 struct perf_sample *sample,
1662 struct perf_tool *tool)
1664 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1672 if (!tool->ordered_events) {
1673 pr_err("Intel Processor Trace requires ordered events\n");
1677 if (sample->time && sample->time != (u64)-1)
1678 timestamp = perf_time_to_tsc(sample->time, &pt->tc);
1682 if (timestamp || pt->timeless_decoding) {
1683 err = intel_pt_update_queues(pt);
1688 if (pt->timeless_decoding) {
1689 if (event->header.type == PERF_RECORD_EXIT) {
1690 err = intel_pt_process_timeless_queues(pt,
1694 } else if (timestamp) {
1695 err = intel_pt_process_queues(pt, timestamp);
1700 if (event->header.type == PERF_RECORD_AUX &&
1701 (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
1702 pt->synth_opts.errors) {
1703 err = intel_pt_lost(pt, sample);
1708 if (pt->switch_evsel && event->header.type == PERF_RECORD_SAMPLE)
1709 err = intel_pt_process_switch(pt, sample);
1710 else if (event->header.type == PERF_RECORD_ITRACE_START)
1711 err = intel_pt_process_itrace_start(pt, event, sample);
1712 else if (event->header.type == PERF_RECORD_SWITCH ||
1713 event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
1714 err = intel_pt_context_switch(pt, event, sample);
1716 intel_pt_log("event %s (%u): cpu %d time %"PRIu64" tsc %#"PRIx64"\n",
1717 perf_event__name(event->header.type), event->header.type,
1718 sample->cpu, sample->time, timestamp);
1723 static int intel_pt_flush(struct perf_session *session, struct perf_tool *tool)
1725 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1732 if (!tool->ordered_events)
1735 ret = intel_pt_update_queues(pt);
1739 if (pt->timeless_decoding)
1740 return intel_pt_process_timeless_queues(pt, -1,
1743 return intel_pt_process_queues(pt, MAX_TIMESTAMP);
1746 static void intel_pt_free_events(struct perf_session *session)
1748 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1750 struct auxtrace_queues *queues = &pt->queues;
1753 for (i = 0; i < queues->nr_queues; i++) {
1754 intel_pt_free_queue(queues->queue_array[i].priv);
1755 queues->queue_array[i].priv = NULL;
1757 intel_pt_log_disable();
1758 auxtrace_queues__free(queues);
1761 static void intel_pt_free(struct perf_session *session)
1763 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1766 auxtrace_heap__free(&pt->heap);
1767 intel_pt_free_events(session);
1768 session->auxtrace = NULL;
1769 thread__put(pt->unknown_thread);
1773 static int intel_pt_process_auxtrace_event(struct perf_session *session,
1774 union perf_event *event,
1775 struct perf_tool *tool __maybe_unused)
1777 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1780 if (pt->sampling_mode)
1783 if (!pt->data_queued) {
1784 struct auxtrace_buffer *buffer;
1786 int fd = perf_data_file__fd(session->file);
1789 if (perf_data_file__is_pipe(session->file)) {
1792 data_offset = lseek(fd, 0, SEEK_CUR);
1793 if (data_offset == -1)
1797 err = auxtrace_queues__add_event(&pt->queues, session, event,
1798 data_offset, &buffer);
1802 /* Dump here now we have copied a piped trace out of the pipe */
1804 if (auxtrace_buffer__get_data(buffer, fd)) {
1805 intel_pt_dump_event(pt, buffer->data,
1807 auxtrace_buffer__put_data(buffer);
1815 struct intel_pt_synth {
1816 struct perf_tool dummy_tool;
1817 struct perf_session *session;
1820 static int intel_pt_event_synth(struct perf_tool *tool,
1821 union perf_event *event,
1822 struct perf_sample *sample __maybe_unused,
1823 struct machine *machine __maybe_unused)
1825 struct intel_pt_synth *intel_pt_synth =
1826 container_of(tool, struct intel_pt_synth, dummy_tool);
1828 return perf_session__deliver_synth_event(intel_pt_synth->session, event,
1832 static int intel_pt_synth_event(struct perf_session *session,
1833 struct perf_event_attr *attr, u64 id)
1835 struct intel_pt_synth intel_pt_synth;
1837 memset(&intel_pt_synth, 0, sizeof(struct intel_pt_synth));
1838 intel_pt_synth.session = session;
1840 return perf_event__synthesize_attr(&intel_pt_synth.dummy_tool, attr, 1,
1841 &id, intel_pt_event_synth);
1844 static int intel_pt_synth_events(struct intel_pt *pt,
1845 struct perf_session *session)
1847 struct perf_evlist *evlist = session->evlist;
1848 struct perf_evsel *evsel;
1849 struct perf_event_attr attr;
1854 evlist__for_each_entry(evlist, evsel) {
1855 if (evsel->attr.type == pt->pmu_type && evsel->ids) {
1862 pr_debug("There are no selected events with Intel Processor Trace data\n");
1866 memset(&attr, 0, sizeof(struct perf_event_attr));
1867 attr.size = sizeof(struct perf_event_attr);
1868 attr.type = PERF_TYPE_HARDWARE;
1869 attr.sample_type = evsel->attr.sample_type & PERF_SAMPLE_MASK;
1870 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
1872 if (pt->timeless_decoding)
1873 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
1875 attr.sample_type |= PERF_SAMPLE_TIME;
1876 if (!pt->per_cpu_mmaps)
1877 attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
1878 attr.exclude_user = evsel->attr.exclude_user;
1879 attr.exclude_kernel = evsel->attr.exclude_kernel;
1880 attr.exclude_hv = evsel->attr.exclude_hv;
1881 attr.exclude_host = evsel->attr.exclude_host;
1882 attr.exclude_guest = evsel->attr.exclude_guest;
1883 attr.sample_id_all = evsel->attr.sample_id_all;
1884 attr.read_format = evsel->attr.read_format;
1886 id = evsel->id[0] + 1000000000;
1890 if (pt->synth_opts.instructions) {
1891 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
1892 if (pt->synth_opts.period_type == PERF_ITRACE_PERIOD_NANOSECS)
1893 attr.sample_period =
1894 intel_pt_ns_to_ticks(pt, pt->synth_opts.period);
1896 attr.sample_period = pt->synth_opts.period;
1897 pt->instructions_sample_period = attr.sample_period;
1898 if (pt->synth_opts.callchain)
1899 attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
1900 if (pt->synth_opts.last_branch)
1901 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
1902 pr_debug("Synthesizing 'instructions' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
1903 id, (u64)attr.sample_type);
1904 err = intel_pt_synth_event(session, &attr, id);
1906 pr_err("%s: failed to synthesize 'instructions' event type\n",
1910 pt->sample_instructions = true;
1911 pt->instructions_sample_type = attr.sample_type;
1912 pt->instructions_id = id;
1916 if (pt->synth_opts.transactions) {
1917 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
1918 attr.sample_period = 1;
1919 if (pt->synth_opts.callchain)
1920 attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
1921 if (pt->synth_opts.last_branch)
1922 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
1923 pr_debug("Synthesizing 'transactions' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
1924 id, (u64)attr.sample_type);
1925 err = intel_pt_synth_event(session, &attr, id);
1927 pr_err("%s: failed to synthesize 'transactions' event type\n",
1931 pt->sample_transactions = true;
1932 pt->transactions_id = id;
1934 evlist__for_each_entry(evlist, evsel) {
1935 if (evsel->id && evsel->id[0] == pt->transactions_id) {
1937 zfree(&evsel->name);
1938 evsel->name = strdup("transactions");
1944 if (pt->synth_opts.branches) {
1945 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
1946 attr.sample_period = 1;
1947 attr.sample_type |= PERF_SAMPLE_ADDR;
1948 attr.sample_type &= ~(u64)PERF_SAMPLE_CALLCHAIN;
1949 attr.sample_type &= ~(u64)PERF_SAMPLE_BRANCH_STACK;
1950 pr_debug("Synthesizing 'branches' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
1951 id, (u64)attr.sample_type);
1952 err = intel_pt_synth_event(session, &attr, id);
1954 pr_err("%s: failed to synthesize 'branches' event type\n",
1958 pt->sample_branches = true;
1959 pt->branches_sample_type = attr.sample_type;
1960 pt->branches_id = id;
1963 pt->synth_needs_swap = evsel->needs_swap;
1968 static struct perf_evsel *intel_pt_find_sched_switch(struct perf_evlist *evlist)
1970 struct perf_evsel *evsel;
1972 evlist__for_each_entry_reverse(evlist, evsel) {
1973 const char *name = perf_evsel__name(evsel);
1975 if (!strcmp(name, "sched:sched_switch"))
1982 static bool intel_pt_find_switch(struct perf_evlist *evlist)
1984 struct perf_evsel *evsel;
1986 evlist__for_each_entry(evlist, evsel) {
1987 if (evsel->attr.context_switch)
1994 static int intel_pt_perf_config(const char *var, const char *value, void *data)
1996 struct intel_pt *pt = data;
1998 if (!strcmp(var, "intel-pt.mispred-all"))
1999 pt->mispred_all = perf_config_bool(var, value);
2004 static const char * const intel_pt_info_fmts[] = {
2005 [INTEL_PT_PMU_TYPE] = " PMU Type %"PRId64"\n",
2006 [INTEL_PT_TIME_SHIFT] = " Time Shift %"PRIu64"\n",
2007 [INTEL_PT_TIME_MULT] = " Time Muliplier %"PRIu64"\n",
2008 [INTEL_PT_TIME_ZERO] = " Time Zero %"PRIu64"\n",
2009 [INTEL_PT_CAP_USER_TIME_ZERO] = " Cap Time Zero %"PRId64"\n",
2010 [INTEL_PT_TSC_BIT] = " TSC bit %#"PRIx64"\n",
2011 [INTEL_PT_NORETCOMP_BIT] = " NoRETComp bit %#"PRIx64"\n",
2012 [INTEL_PT_HAVE_SCHED_SWITCH] = " Have sched_switch %"PRId64"\n",
2013 [INTEL_PT_SNAPSHOT_MODE] = " Snapshot mode %"PRId64"\n",
2014 [INTEL_PT_PER_CPU_MMAPS] = " Per-cpu maps %"PRId64"\n",
2015 [INTEL_PT_MTC_BIT] = " MTC bit %#"PRIx64"\n",
2016 [INTEL_PT_TSC_CTC_N] = " TSC:CTC numerator %"PRIu64"\n",
2017 [INTEL_PT_TSC_CTC_D] = " TSC:CTC denominator %"PRIu64"\n",
2018 [INTEL_PT_CYC_BIT] = " CYC bit %#"PRIx64"\n",
2021 static void intel_pt_print_info(u64 *arr, int start, int finish)
2028 for (i = start; i <= finish; i++)
2029 fprintf(stdout, intel_pt_info_fmts[i], arr[i]);
2032 int intel_pt_process_auxtrace_info(union perf_event *event,
2033 struct perf_session *session)
2035 struct auxtrace_info_event *auxtrace_info = &event->auxtrace_info;
2036 size_t min_sz = sizeof(u64) * INTEL_PT_PER_CPU_MMAPS;
2037 struct intel_pt *pt;
2040 if (auxtrace_info->header.size < sizeof(struct auxtrace_info_event) +
2044 pt = zalloc(sizeof(struct intel_pt));
2048 perf_config(intel_pt_perf_config, pt);
2050 err = auxtrace_queues__init(&pt->queues);
2054 intel_pt_log_set_name(INTEL_PT_PMU_NAME);
2056 pt->session = session;
2057 pt->machine = &session->machines.host; /* No kvm support */
2058 pt->auxtrace_type = auxtrace_info->type;
2059 pt->pmu_type = auxtrace_info->priv[INTEL_PT_PMU_TYPE];
2060 pt->tc.time_shift = auxtrace_info->priv[INTEL_PT_TIME_SHIFT];
2061 pt->tc.time_mult = auxtrace_info->priv[INTEL_PT_TIME_MULT];
2062 pt->tc.time_zero = auxtrace_info->priv[INTEL_PT_TIME_ZERO];
2063 pt->cap_user_time_zero = auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO];
2064 pt->tsc_bit = auxtrace_info->priv[INTEL_PT_TSC_BIT];
2065 pt->noretcomp_bit = auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT];
2066 pt->have_sched_switch = auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH];
2067 pt->snapshot_mode = auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE];
2068 pt->per_cpu_mmaps = auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS];
2069 intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_PMU_TYPE,
2070 INTEL_PT_PER_CPU_MMAPS);
2072 if (auxtrace_info->header.size >= sizeof(struct auxtrace_info_event) +
2073 (sizeof(u64) * INTEL_PT_CYC_BIT)) {
2074 pt->mtc_bit = auxtrace_info->priv[INTEL_PT_MTC_BIT];
2075 pt->mtc_freq_bits = auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS];
2076 pt->tsc_ctc_ratio_n = auxtrace_info->priv[INTEL_PT_TSC_CTC_N];
2077 pt->tsc_ctc_ratio_d = auxtrace_info->priv[INTEL_PT_TSC_CTC_D];
2078 pt->cyc_bit = auxtrace_info->priv[INTEL_PT_CYC_BIT];
2079 intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_MTC_BIT,
2083 pt->timeless_decoding = intel_pt_timeless_decoding(pt);
2084 pt->have_tsc = intel_pt_have_tsc(pt);
2085 pt->sampling_mode = false;
2086 pt->est_tsc = !pt->timeless_decoding;
2088 pt->unknown_thread = thread__new(999999999, 999999999);
2089 if (!pt->unknown_thread) {
2091 goto err_free_queues;
2095 * Since this thread will not be kept in any rbtree not in a
2096 * list, initialize its list node so that at thread__put() the
2097 * current thread lifetime assuption is kept and we don't segfault
2098 * at list_del_init().
2100 INIT_LIST_HEAD(&pt->unknown_thread->node);
2102 err = thread__set_comm(pt->unknown_thread, "unknown", 0);
2104 goto err_delete_thread;
2105 if (thread__init_map_groups(pt->unknown_thread, pt->machine)) {
2107 goto err_delete_thread;
2110 pt->auxtrace.process_event = intel_pt_process_event;
2111 pt->auxtrace.process_auxtrace_event = intel_pt_process_auxtrace_event;
2112 pt->auxtrace.flush_events = intel_pt_flush;
2113 pt->auxtrace.free_events = intel_pt_free_events;
2114 pt->auxtrace.free = intel_pt_free;
2115 session->auxtrace = &pt->auxtrace;
2120 if (pt->have_sched_switch == 1) {
2121 pt->switch_evsel = intel_pt_find_sched_switch(session->evlist);
2122 if (!pt->switch_evsel) {
2123 pr_err("%s: missing sched_switch event\n", __func__);
2124 goto err_delete_thread;
2126 } else if (pt->have_sched_switch == 2 &&
2127 !intel_pt_find_switch(session->evlist)) {
2128 pr_err("%s: missing context_switch attribute flag\n", __func__);
2129 goto err_delete_thread;
2132 if (session->itrace_synth_opts && session->itrace_synth_opts->set) {
2133 pt->synth_opts = *session->itrace_synth_opts;
2135 itrace_synth_opts__set_default(&pt->synth_opts);
2136 if (use_browser != -1) {
2137 pt->synth_opts.branches = false;
2138 pt->synth_opts.callchain = true;
2142 if (pt->synth_opts.log)
2143 intel_pt_log_enable();
2145 /* Maximum non-turbo ratio is TSC freq / 100 MHz */
2146 if (pt->tc.time_mult) {
2147 u64 tsc_freq = intel_pt_ns_to_ticks(pt, 1000000000);
2149 pt->max_non_turbo_ratio = (tsc_freq + 50000000) / 100000000;
2150 intel_pt_log("TSC frequency %"PRIu64"\n", tsc_freq);
2151 intel_pt_log("Maximum non-turbo ratio %u\n",
2152 pt->max_non_turbo_ratio);
2155 if (pt->synth_opts.calls)
2156 pt->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
2157 PERF_IP_FLAG_TRACE_END;
2158 if (pt->synth_opts.returns)
2159 pt->branches_filter |= PERF_IP_FLAG_RETURN |
2160 PERF_IP_FLAG_TRACE_BEGIN;
2162 if (pt->synth_opts.callchain && !symbol_conf.use_callchain) {
2163 symbol_conf.use_callchain = true;
2164 if (callchain_register_param(&callchain_param) < 0) {
2165 symbol_conf.use_callchain = false;
2166 pt->synth_opts.callchain = false;
2170 err = intel_pt_synth_events(pt, session);
2172 goto err_delete_thread;
2174 err = auxtrace_queues__process_index(&pt->queues, session);
2176 goto err_delete_thread;
2178 if (pt->queues.populated)
2179 pt->data_queued = true;
2181 if (pt->timeless_decoding)
2182 pr_debug2("Intel PT decoding without timestamps\n");
2187 thread__zput(pt->unknown_thread);
2189 intel_pt_log_disable();
2190 auxtrace_queues__free(&pt->queues);
2191 session->auxtrace = NULL;