1 // SPDX-License-Identifier: GPL-2.0-only
3 * intel_pt.c: Intel Processor Trace support
4 * Copyright (c) 2013-2015, Intel Corporation.
11 #include <linux/kernel.h>
12 #include <linux/string.h>
13 #include <linux/types.h>
14 #include <linux/zalloc.h>
27 #include "thread-stack.h"
29 #include "callchain.h"
36 #include "util/synthetic-events.h"
37 #include "time-utils.h"
39 #include "../arch/x86/include/uapi/asm/perf_regs.h"
41 #include "intel-pt-decoder/intel-pt-log.h"
42 #include "intel-pt-decoder/intel-pt-decoder.h"
43 #include "intel-pt-decoder/intel-pt-insn-decoder.h"
44 #include "intel-pt-decoder/intel-pt-pkt-decoder.h"
46 #define MAX_TIMESTAMP (~0ULL)
54 struct auxtrace auxtrace;
55 struct auxtrace_queues queues;
56 struct auxtrace_heap heap;
58 struct perf_session *session;
59 struct machine *machine;
60 struct evsel *switch_evsel;
61 struct thread *unknown_thread;
62 bool timeless_decoding;
71 int have_sched_switch;
77 struct perf_tsc_conversion tc;
78 bool cap_user_time_zero;
80 struct itrace_synth_opts synth_opts;
82 bool sample_instructions;
83 u64 instructions_sample_type;
88 u64 branches_sample_type;
91 bool sample_transactions;
92 u64 transactions_sample_type;
96 u64 ptwrites_sample_type;
99 bool sample_pwr_events;
100 u64 pwr_events_sample_type;
108 struct evsel *pebs_evsel;
117 unsigned max_non_turbo_ratio;
120 unsigned long num_events;
123 struct addr_filters filts;
125 struct range *time_ranges;
126 unsigned int range_cnt;
130 INTEL_PT_SS_NOT_TRACING,
133 INTEL_PT_SS_EXPECTING_SWITCH_EVENT,
134 INTEL_PT_SS_EXPECTING_SWITCH_IP,
137 struct intel_pt_queue {
139 unsigned int queue_nr;
140 struct auxtrace_buffer *buffer;
141 struct auxtrace_buffer *old_buffer;
143 const struct intel_pt_state *state;
144 struct ip_callchain *chain;
145 struct branch_stack *last_branch;
146 struct branch_stack *last_branch_rb;
147 size_t last_branch_pos;
148 union perf_event *event_buf;
151 bool step_through_buffers;
152 bool use_buffer_pid_tid;
158 struct thread *thread;
165 unsigned int sel_idx;
171 u64 last_in_insn_cnt;
173 u64 last_br_insn_cnt;
175 unsigned int cbr_seen;
176 char insn[INTEL_PT_INSN_BUF_SZ];
179 static void intel_pt_dump(struct intel_pt *pt __maybe_unused,
180 unsigned char *buf, size_t len)
182 struct intel_pt_pkt packet;
185 char desc[INTEL_PT_PKT_DESC_MAX];
186 const char *color = PERF_COLOR_BLUE;
187 enum intel_pt_pkt_ctx ctx = INTEL_PT_NO_CTX;
189 color_fprintf(stdout, color,
190 ". ... Intel Processor Trace data: size %zu bytes\n",
194 ret = intel_pt_get_packet(buf, len, &packet, &ctx);
200 color_fprintf(stdout, color, " %08x: ", pos);
201 for (i = 0; i < pkt_len; i++)
202 color_fprintf(stdout, color, " %02x", buf[i]);
204 color_fprintf(stdout, color, " ");
206 ret = intel_pt_pkt_desc(&packet, desc,
207 INTEL_PT_PKT_DESC_MAX);
209 color_fprintf(stdout, color, " %s\n", desc);
211 color_fprintf(stdout, color, " Bad packet!\n");
219 static void intel_pt_dump_event(struct intel_pt *pt, unsigned char *buf,
223 intel_pt_dump(pt, buf, len);
226 static void intel_pt_log_event(union perf_event *event)
228 FILE *f = intel_pt_log_fp();
230 if (!intel_pt_enable_logging || !f)
233 perf_event__fprintf(event, f);
236 static void intel_pt_dump_sample(struct perf_session *session,
237 struct perf_sample *sample)
239 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
243 intel_pt_dump(pt, sample->aux_sample.data, sample->aux_sample.size);
246 static int intel_pt_do_fix_overlap(struct intel_pt *pt, struct auxtrace_buffer *a,
247 struct auxtrace_buffer *b)
249 bool consecutive = false;
252 start = intel_pt_find_overlap(a->data, a->size, b->data, b->size,
253 pt->have_tsc, &consecutive);
256 b->use_size = b->data + b->size - start;
258 if (b->use_size && consecutive)
259 b->consecutive = true;
263 static int intel_pt_get_buffer(struct intel_pt_queue *ptq,
264 struct auxtrace_buffer *buffer,
265 struct auxtrace_buffer *old_buffer,
266 struct intel_pt_buffer *b)
271 int fd = perf_data__fd(ptq->pt->session->data);
273 buffer->data = auxtrace_buffer__get_data(buffer, fd);
278 might_overlap = ptq->pt->snapshot_mode || ptq->pt->sampling_mode;
279 if (might_overlap && !buffer->consecutive && old_buffer &&
280 intel_pt_do_fix_overlap(ptq->pt, old_buffer, buffer))
283 if (buffer->use_data) {
284 b->len = buffer->use_size;
285 b->buf = buffer->use_data;
287 b->len = buffer->size;
288 b->buf = buffer->data;
290 b->ref_timestamp = buffer->reference;
292 if (!old_buffer || (might_overlap && !buffer->consecutive)) {
293 b->consecutive = false;
294 b->trace_nr = buffer->buffer_nr + 1;
296 b->consecutive = true;
302 /* Do not drop buffers with references - refer intel_pt_get_trace() */
303 static void intel_pt_lookahead_drop_buffer(struct intel_pt_queue *ptq,
304 struct auxtrace_buffer *buffer)
306 if (!buffer || buffer == ptq->buffer || buffer == ptq->old_buffer)
309 auxtrace_buffer__drop_data(buffer);
312 /* Must be serialized with respect to intel_pt_get_trace() */
313 static int intel_pt_lookahead(void *data, intel_pt_lookahead_cb_t cb,
316 struct intel_pt_queue *ptq = data;
317 struct auxtrace_buffer *buffer = ptq->buffer;
318 struct auxtrace_buffer *old_buffer = ptq->old_buffer;
319 struct auxtrace_queue *queue;
322 queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
325 struct intel_pt_buffer b = { .len = 0 };
327 buffer = auxtrace_buffer__next(queue, buffer);
331 err = intel_pt_get_buffer(ptq, buffer, old_buffer, &b);
336 intel_pt_lookahead_drop_buffer(ptq, old_buffer);
339 intel_pt_lookahead_drop_buffer(ptq, buffer);
343 err = cb(&b, cb_data);
348 if (buffer != old_buffer)
349 intel_pt_lookahead_drop_buffer(ptq, buffer);
350 intel_pt_lookahead_drop_buffer(ptq, old_buffer);
356 * This function assumes data is processed sequentially only.
357 * Must be serialized with respect to intel_pt_lookahead()
359 static int intel_pt_get_trace(struct intel_pt_buffer *b, void *data)
361 struct intel_pt_queue *ptq = data;
362 struct auxtrace_buffer *buffer = ptq->buffer;
363 struct auxtrace_buffer *old_buffer = ptq->old_buffer;
364 struct auxtrace_queue *queue;
372 queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
374 buffer = auxtrace_buffer__next(queue, buffer);
377 auxtrace_buffer__drop_data(old_buffer);
382 ptq->buffer = buffer;
384 err = intel_pt_get_buffer(ptq, buffer, old_buffer, b);
388 if (ptq->step_through_buffers)
393 auxtrace_buffer__drop_data(old_buffer);
394 ptq->old_buffer = buffer;
396 auxtrace_buffer__drop_data(buffer);
397 return intel_pt_get_trace(b, data);
403 struct intel_pt_cache_entry {
404 struct auxtrace_cache_entry entry;
407 enum intel_pt_insn_op op;
408 enum intel_pt_insn_branch branch;
411 char insn[INTEL_PT_INSN_BUF_SZ];
414 static int intel_pt_config_div(const char *var, const char *value, void *data)
419 if (!strcmp(var, "intel-pt.cache-divisor")) {
420 val = strtol(value, NULL, 0);
421 if (val > 0 && val <= INT_MAX)
428 static int intel_pt_cache_divisor(void)
435 perf_config(intel_pt_config_div, &d);
443 static unsigned int intel_pt_cache_size(struct dso *dso,
444 struct machine *machine)
448 size = dso__data_size(dso, machine);
449 size /= intel_pt_cache_divisor();
452 if (size > (1 << 21))
454 return 32 - __builtin_clz(size);
457 static struct auxtrace_cache *intel_pt_cache(struct dso *dso,
458 struct machine *machine)
460 struct auxtrace_cache *c;
463 if (dso->auxtrace_cache)
464 return dso->auxtrace_cache;
466 bits = intel_pt_cache_size(dso, machine);
468 /* Ignoring cache creation failure */
469 c = auxtrace_cache__new(bits, sizeof(struct intel_pt_cache_entry), 200);
471 dso->auxtrace_cache = c;
476 static int intel_pt_cache_add(struct dso *dso, struct machine *machine,
477 u64 offset, u64 insn_cnt, u64 byte_cnt,
478 struct intel_pt_insn *intel_pt_insn)
480 struct auxtrace_cache *c = intel_pt_cache(dso, machine);
481 struct intel_pt_cache_entry *e;
487 e = auxtrace_cache__alloc_entry(c);
491 e->insn_cnt = insn_cnt;
492 e->byte_cnt = byte_cnt;
493 e->op = intel_pt_insn->op;
494 e->branch = intel_pt_insn->branch;
495 e->length = intel_pt_insn->length;
496 e->rel = intel_pt_insn->rel;
497 memcpy(e->insn, intel_pt_insn->buf, INTEL_PT_INSN_BUF_SZ);
499 err = auxtrace_cache__add(c, offset, &e->entry);
501 auxtrace_cache__free_entry(c, e);
506 static struct intel_pt_cache_entry *
507 intel_pt_cache_lookup(struct dso *dso, struct machine *machine, u64 offset)
509 struct auxtrace_cache *c = intel_pt_cache(dso, machine);
514 return auxtrace_cache__lookup(dso->auxtrace_cache, offset);
517 static inline u8 intel_pt_cpumode(struct intel_pt *pt, uint64_t ip)
519 return ip >= pt->kernel_start ?
520 PERF_RECORD_MISC_KERNEL :
521 PERF_RECORD_MISC_USER;
524 static int intel_pt_walk_next_insn(struct intel_pt_insn *intel_pt_insn,
525 uint64_t *insn_cnt_ptr, uint64_t *ip,
526 uint64_t to_ip, uint64_t max_insn_cnt,
529 struct intel_pt_queue *ptq = data;
530 struct machine *machine = ptq->pt->machine;
531 struct thread *thread;
532 struct addr_location al;
533 unsigned char buf[INTEL_PT_INSN_BUF_SZ];
537 u64 offset, start_offset, start_ip;
541 intel_pt_insn->length = 0;
543 if (to_ip && *ip == to_ip)
546 cpumode = intel_pt_cpumode(ptq->pt, *ip);
548 thread = ptq->thread;
550 if (cpumode != PERF_RECORD_MISC_KERNEL)
552 thread = ptq->pt->unknown_thread;
556 if (!thread__find_map(thread, cpumode, *ip, &al) || !al.map->dso)
559 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
560 dso__data_status_seen(al.map->dso,
561 DSO_DATA_STATUS_SEEN_ITRACE))
564 offset = al.map->map_ip(al.map, *ip);
566 if (!to_ip && one_map) {
567 struct intel_pt_cache_entry *e;
569 e = intel_pt_cache_lookup(al.map->dso, machine, offset);
571 (!max_insn_cnt || e->insn_cnt <= max_insn_cnt)) {
572 *insn_cnt_ptr = e->insn_cnt;
574 intel_pt_insn->op = e->op;
575 intel_pt_insn->branch = e->branch;
576 intel_pt_insn->length = e->length;
577 intel_pt_insn->rel = e->rel;
578 memcpy(intel_pt_insn->buf, e->insn,
579 INTEL_PT_INSN_BUF_SZ);
580 intel_pt_log_insn_no_data(intel_pt_insn, *ip);
585 start_offset = offset;
588 /* Load maps to ensure dso->is_64_bit has been updated */
591 x86_64 = al.map->dso->is_64_bit;
594 len = dso__data_read_offset(al.map->dso, machine,
596 INTEL_PT_INSN_BUF_SZ);
600 if (intel_pt_get_insn(buf, len, x86_64, intel_pt_insn))
603 intel_pt_log_insn(intel_pt_insn, *ip);
607 if (intel_pt_insn->branch != INTEL_PT_BR_NO_BRANCH)
610 if (max_insn_cnt && insn_cnt >= max_insn_cnt)
613 *ip += intel_pt_insn->length;
615 if (to_ip && *ip == to_ip)
618 if (*ip >= al.map->end)
621 offset += intel_pt_insn->length;
626 *insn_cnt_ptr = insn_cnt;
632 * Didn't lookup in the 'to_ip' case, so do it now to prevent duplicate
636 struct intel_pt_cache_entry *e;
638 e = intel_pt_cache_lookup(al.map->dso, machine, start_offset);
643 /* Ignore cache errors */
644 intel_pt_cache_add(al.map->dso, machine, start_offset, insn_cnt,
645 *ip - start_ip, intel_pt_insn);
650 *insn_cnt_ptr = insn_cnt;
654 static bool intel_pt_match_pgd_ip(struct intel_pt *pt, uint64_t ip,
655 uint64_t offset, const char *filename)
657 struct addr_filter *filt;
658 bool have_filter = false;
659 bool hit_tracestop = false;
660 bool hit_filter = false;
662 list_for_each_entry(filt, &pt->filts.head, list) {
666 if ((filename && !filt->filename) ||
667 (!filename && filt->filename) ||
668 (filename && strcmp(filename, filt->filename)))
671 if (!(offset >= filt->addr && offset < filt->addr + filt->size))
674 intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s hit filter: %s offset %#"PRIx64" size %#"PRIx64"\n",
675 ip, offset, filename ? filename : "[kernel]",
676 filt->start ? "filter" : "stop",
677 filt->addr, filt->size);
682 hit_tracestop = true;
685 if (!hit_tracestop && !hit_filter)
686 intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s is not in a filter region\n",
687 ip, offset, filename ? filename : "[kernel]");
689 return hit_tracestop || (have_filter && !hit_filter);
692 static int __intel_pt_pgd_ip(uint64_t ip, void *data)
694 struct intel_pt_queue *ptq = data;
695 struct thread *thread;
696 struct addr_location al;
700 if (ip >= ptq->pt->kernel_start)
701 return intel_pt_match_pgd_ip(ptq->pt, ip, ip, NULL);
703 cpumode = PERF_RECORD_MISC_USER;
705 thread = ptq->thread;
709 if (!thread__find_map(thread, cpumode, ip, &al) || !al.map->dso)
712 offset = al.map->map_ip(al.map, ip);
714 return intel_pt_match_pgd_ip(ptq->pt, ip, offset,
715 al.map->dso->long_name);
718 static bool intel_pt_pgd_ip(uint64_t ip, void *data)
720 return __intel_pt_pgd_ip(ip, data) > 0;
723 static bool intel_pt_get_config(struct intel_pt *pt,
724 struct perf_event_attr *attr, u64 *config)
726 if (attr->type == pt->pmu_type) {
728 *config = attr->config;
735 static bool intel_pt_exclude_kernel(struct intel_pt *pt)
739 evlist__for_each_entry(pt->session->evlist, evsel) {
740 if (intel_pt_get_config(pt, &evsel->core.attr, NULL) &&
741 !evsel->core.attr.exclude_kernel)
747 static bool intel_pt_return_compression(struct intel_pt *pt)
752 if (!pt->noretcomp_bit)
755 evlist__for_each_entry(pt->session->evlist, evsel) {
756 if (intel_pt_get_config(pt, &evsel->core.attr, &config) &&
757 (config & pt->noretcomp_bit))
763 static bool intel_pt_branch_enable(struct intel_pt *pt)
768 evlist__for_each_entry(pt->session->evlist, evsel) {
769 if (intel_pt_get_config(pt, &evsel->core.attr, &config) &&
770 (config & 1) && !(config & 0x2000))
776 static unsigned int intel_pt_mtc_period(struct intel_pt *pt)
782 if (!pt->mtc_freq_bits)
785 for (shift = 0, config = pt->mtc_freq_bits; !(config & 1); shift++)
788 evlist__for_each_entry(pt->session->evlist, evsel) {
789 if (intel_pt_get_config(pt, &evsel->core.attr, &config))
790 return (config & pt->mtc_freq_bits) >> shift;
795 static bool intel_pt_timeless_decoding(struct intel_pt *pt)
798 bool timeless_decoding = true;
801 if (!pt->tsc_bit || !pt->cap_user_time_zero)
804 evlist__for_each_entry(pt->session->evlist, evsel) {
805 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
807 if (intel_pt_get_config(pt, &evsel->core.attr, &config)) {
808 if (config & pt->tsc_bit)
809 timeless_decoding = false;
814 return timeless_decoding;
817 static bool intel_pt_tracing_kernel(struct intel_pt *pt)
821 evlist__for_each_entry(pt->session->evlist, evsel) {
822 if (intel_pt_get_config(pt, &evsel->core.attr, NULL) &&
823 !evsel->core.attr.exclude_kernel)
829 static bool intel_pt_have_tsc(struct intel_pt *pt)
832 bool have_tsc = false;
838 evlist__for_each_entry(pt->session->evlist, evsel) {
839 if (intel_pt_get_config(pt, &evsel->core.attr, &config)) {
840 if (config & pt->tsc_bit)
849 static bool intel_pt_sampling_mode(struct intel_pt *pt)
853 evlist__for_each_entry(pt->session->evlist, evsel) {
854 if ((evsel->core.attr.sample_type & PERF_SAMPLE_AUX) &&
855 evsel->core.attr.aux_sample_size)
861 static u64 intel_pt_ns_to_ticks(const struct intel_pt *pt, u64 ns)
865 quot = ns / pt->tc.time_mult;
866 rem = ns % pt->tc.time_mult;
867 return (quot << pt->tc.time_shift) + (rem << pt->tc.time_shift) /
871 static struct intel_pt_queue *intel_pt_alloc_queue(struct intel_pt *pt,
872 unsigned int queue_nr)
874 struct intel_pt_params params = { .get_trace = 0, };
875 struct perf_env *env = pt->machine->env;
876 struct intel_pt_queue *ptq;
878 ptq = zalloc(sizeof(struct intel_pt_queue));
882 if (pt->synth_opts.callchain) {
883 size_t sz = sizeof(struct ip_callchain);
885 /* Add 1 to callchain_sz for callchain context */
886 sz += (pt->synth_opts.callchain_sz + 1) * sizeof(u64);
887 ptq->chain = zalloc(sz);
892 if (pt->synth_opts.last_branch) {
893 size_t sz = sizeof(struct branch_stack);
895 sz += pt->synth_opts.last_branch_sz *
896 sizeof(struct branch_entry);
897 ptq->last_branch = zalloc(sz);
898 if (!ptq->last_branch)
900 ptq->last_branch_rb = zalloc(sz);
901 if (!ptq->last_branch_rb)
905 ptq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
910 ptq->queue_nr = queue_nr;
911 ptq->exclude_kernel = intel_pt_exclude_kernel(pt);
917 params.get_trace = intel_pt_get_trace;
918 params.walk_insn = intel_pt_walk_next_insn;
919 params.lookahead = intel_pt_lookahead;
921 params.return_compression = intel_pt_return_compression(pt);
922 params.branch_enable = intel_pt_branch_enable(pt);
923 params.max_non_turbo_ratio = pt->max_non_turbo_ratio;
924 params.mtc_period = intel_pt_mtc_period(pt);
925 params.tsc_ctc_ratio_n = pt->tsc_ctc_ratio_n;
926 params.tsc_ctc_ratio_d = pt->tsc_ctc_ratio_d;
928 if (pt->filts.cnt > 0)
929 params.pgd_ip = intel_pt_pgd_ip;
931 if (pt->synth_opts.instructions) {
932 if (pt->synth_opts.period) {
933 switch (pt->synth_opts.period_type) {
934 case PERF_ITRACE_PERIOD_INSTRUCTIONS:
936 INTEL_PT_PERIOD_INSTRUCTIONS;
937 params.period = pt->synth_opts.period;
939 case PERF_ITRACE_PERIOD_TICKS:
940 params.period_type = INTEL_PT_PERIOD_TICKS;
941 params.period = pt->synth_opts.period;
943 case PERF_ITRACE_PERIOD_NANOSECS:
944 params.period_type = INTEL_PT_PERIOD_TICKS;
945 params.period = intel_pt_ns_to_ticks(pt,
946 pt->synth_opts.period);
953 if (!params.period) {
954 params.period_type = INTEL_PT_PERIOD_INSTRUCTIONS;
959 if (env->cpuid && !strncmp(env->cpuid, "GenuineIntel,6,92,", 18))
960 params.flags |= INTEL_PT_FUP_WITH_NLIP;
962 ptq->decoder = intel_pt_decoder_new(¶ms);
969 zfree(&ptq->event_buf);
970 zfree(&ptq->last_branch);
971 zfree(&ptq->last_branch_rb);
977 static void intel_pt_free_queue(void *priv)
979 struct intel_pt_queue *ptq = priv;
983 thread__zput(ptq->thread);
984 intel_pt_decoder_free(ptq->decoder);
985 zfree(&ptq->event_buf);
986 zfree(&ptq->last_branch);
987 zfree(&ptq->last_branch_rb);
992 static void intel_pt_set_pid_tid_cpu(struct intel_pt *pt,
993 struct auxtrace_queue *queue)
995 struct intel_pt_queue *ptq = queue->priv;
997 if (queue->tid == -1 || pt->have_sched_switch) {
998 ptq->tid = machine__get_current_tid(pt->machine, ptq->cpu);
999 thread__zput(ptq->thread);
1002 if (!ptq->thread && ptq->tid != -1)
1003 ptq->thread = machine__find_thread(pt->machine, -1, ptq->tid);
1006 ptq->pid = ptq->thread->pid_;
1007 if (queue->cpu == -1)
1008 ptq->cpu = ptq->thread->cpu;
1012 static void intel_pt_sample_flags(struct intel_pt_queue *ptq)
1014 if (ptq->state->flags & INTEL_PT_ABORT_TX) {
1015 ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_TX_ABORT;
1016 } else if (ptq->state->flags & INTEL_PT_ASYNC) {
1017 if (ptq->state->to_ip)
1018 ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_CALL |
1019 PERF_IP_FLAG_ASYNC |
1020 PERF_IP_FLAG_INTERRUPT;
1022 ptq->flags = PERF_IP_FLAG_BRANCH |
1023 PERF_IP_FLAG_TRACE_END;
1026 if (ptq->state->from_ip)
1027 ptq->flags = intel_pt_insn_type(ptq->state->insn_op);
1029 ptq->flags = PERF_IP_FLAG_BRANCH |
1030 PERF_IP_FLAG_TRACE_BEGIN;
1031 if (ptq->state->flags & INTEL_PT_IN_TX)
1032 ptq->flags |= PERF_IP_FLAG_IN_TX;
1033 ptq->insn_len = ptq->state->insn_len;
1034 memcpy(ptq->insn, ptq->state->insn, INTEL_PT_INSN_BUF_SZ);
1037 if (ptq->state->type & INTEL_PT_TRACE_BEGIN)
1038 ptq->flags |= PERF_IP_FLAG_TRACE_BEGIN;
1039 if (ptq->state->type & INTEL_PT_TRACE_END)
1040 ptq->flags |= PERF_IP_FLAG_TRACE_END;
1043 static void intel_pt_setup_time_range(struct intel_pt *pt,
1044 struct intel_pt_queue *ptq)
1049 ptq->sel_timestamp = pt->time_ranges[0].start;
1052 if (ptq->sel_timestamp) {
1053 ptq->sel_start = true;
1055 ptq->sel_timestamp = pt->time_ranges[0].end;
1056 ptq->sel_start = false;
1060 static int intel_pt_setup_queue(struct intel_pt *pt,
1061 struct auxtrace_queue *queue,
1062 unsigned int queue_nr)
1064 struct intel_pt_queue *ptq = queue->priv;
1066 if (list_empty(&queue->head))
1070 ptq = intel_pt_alloc_queue(pt, queue_nr);
1075 if (queue->cpu != -1)
1076 ptq->cpu = queue->cpu;
1077 ptq->tid = queue->tid;
1079 ptq->cbr_seen = UINT_MAX;
1081 if (pt->sampling_mode && !pt->snapshot_mode &&
1082 pt->timeless_decoding)
1083 ptq->step_through_buffers = true;
1085 ptq->sync_switch = pt->sync_switch;
1087 intel_pt_setup_time_range(pt, ptq);
1090 if (!ptq->on_heap &&
1091 (!ptq->sync_switch ||
1092 ptq->switch_state != INTEL_PT_SS_EXPECTING_SWITCH_EVENT)) {
1093 const struct intel_pt_state *state;
1096 if (pt->timeless_decoding)
1099 intel_pt_log("queue %u getting timestamp\n", queue_nr);
1100 intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
1101 queue_nr, ptq->cpu, ptq->pid, ptq->tid);
1103 if (ptq->sel_start && ptq->sel_timestamp) {
1104 ret = intel_pt_fast_forward(ptq->decoder,
1105 ptq->sel_timestamp);
1111 state = intel_pt_decode(ptq->decoder);
1113 if (state->err == INTEL_PT_ERR_NODATA) {
1114 intel_pt_log("queue %u has no timestamp\n",
1120 if (state->timestamp)
1124 ptq->timestamp = state->timestamp;
1125 intel_pt_log("queue %u timestamp 0x%" PRIx64 "\n",
1126 queue_nr, ptq->timestamp);
1128 ptq->have_sample = true;
1129 if (ptq->sel_start && ptq->sel_timestamp &&
1130 ptq->timestamp < ptq->sel_timestamp)
1131 ptq->have_sample = false;
1132 intel_pt_sample_flags(ptq);
1133 ret = auxtrace_heap__add(&pt->heap, queue_nr, ptq->timestamp);
1136 ptq->on_heap = true;
1142 static int intel_pt_setup_queues(struct intel_pt *pt)
1147 for (i = 0; i < pt->queues.nr_queues; i++) {
1148 ret = intel_pt_setup_queue(pt, &pt->queues.queue_array[i], i);
1155 static inline void intel_pt_copy_last_branch_rb(struct intel_pt_queue *ptq)
1157 struct branch_stack *bs_src = ptq->last_branch_rb;
1158 struct branch_stack *bs_dst = ptq->last_branch;
1161 bs_dst->nr = bs_src->nr;
1166 nr = ptq->pt->synth_opts.last_branch_sz - ptq->last_branch_pos;
1167 memcpy(&bs_dst->entries[0],
1168 &bs_src->entries[ptq->last_branch_pos],
1169 sizeof(struct branch_entry) * nr);
1171 if (bs_src->nr >= ptq->pt->synth_opts.last_branch_sz) {
1172 memcpy(&bs_dst->entries[nr],
1173 &bs_src->entries[0],
1174 sizeof(struct branch_entry) * ptq->last_branch_pos);
1178 static inline void intel_pt_reset_last_branch_rb(struct intel_pt_queue *ptq)
1180 ptq->last_branch_pos = 0;
1181 ptq->last_branch_rb->nr = 0;
1184 static void intel_pt_update_last_branch_rb(struct intel_pt_queue *ptq)
1186 const struct intel_pt_state *state = ptq->state;
1187 struct branch_stack *bs = ptq->last_branch_rb;
1188 struct branch_entry *be;
1190 if (!ptq->last_branch_pos)
1191 ptq->last_branch_pos = ptq->pt->synth_opts.last_branch_sz;
1193 ptq->last_branch_pos -= 1;
1195 be = &bs->entries[ptq->last_branch_pos];
1196 be->from = state->from_ip;
1197 be->to = state->to_ip;
1198 be->flags.abort = !!(state->flags & INTEL_PT_ABORT_TX);
1199 be->flags.in_tx = !!(state->flags & INTEL_PT_IN_TX);
1200 /* No support for mispredict */
1201 be->flags.mispred = ptq->pt->mispred_all;
1203 if (bs->nr < ptq->pt->synth_opts.last_branch_sz)
1207 static inline bool intel_pt_skip_event(struct intel_pt *pt)
1209 return pt->synth_opts.initial_skip &&
1210 pt->num_events++ < pt->synth_opts.initial_skip;
1214 * Cannot count CBR as skipped because it won't go away until cbr == cbr_seen.
1215 * Also ensure CBR is first non-skipped event by allowing for 4 more samples
1216 * from this decoder state.
1218 static inline bool intel_pt_skip_cbr_event(struct intel_pt *pt)
1220 return pt->synth_opts.initial_skip &&
1221 pt->num_events + 4 < pt->synth_opts.initial_skip;
1224 static void intel_pt_prep_a_sample(struct intel_pt_queue *ptq,
1225 union perf_event *event,
1226 struct perf_sample *sample)
1228 event->sample.header.type = PERF_RECORD_SAMPLE;
1229 event->sample.header.size = sizeof(struct perf_event_header);
1231 sample->pid = ptq->pid;
1232 sample->tid = ptq->tid;
1233 sample->cpu = ptq->cpu;
1234 sample->insn_len = ptq->insn_len;
1235 memcpy(sample->insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1238 static void intel_pt_prep_b_sample(struct intel_pt *pt,
1239 struct intel_pt_queue *ptq,
1240 union perf_event *event,
1241 struct perf_sample *sample)
1243 intel_pt_prep_a_sample(ptq, event, sample);
1245 if (!pt->timeless_decoding)
1246 sample->time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1248 sample->ip = ptq->state->from_ip;
1249 sample->cpumode = intel_pt_cpumode(pt, sample->ip);
1250 sample->addr = ptq->state->to_ip;
1252 sample->flags = ptq->flags;
1254 event->sample.header.misc = sample->cpumode;
1257 static int intel_pt_inject_event(union perf_event *event,
1258 struct perf_sample *sample, u64 type)
1260 event->header.size = perf_event__sample_event_size(sample, type, 0);
1261 return perf_event__synthesize_sample(event, type, 0, sample);
1264 static inline int intel_pt_opt_inject(struct intel_pt *pt,
1265 union perf_event *event,
1266 struct perf_sample *sample, u64 type)
1268 if (!pt->synth_opts.inject)
1271 return intel_pt_inject_event(event, sample, type);
1274 static int intel_pt_deliver_synth_b_event(struct intel_pt *pt,
1275 union perf_event *event,
1276 struct perf_sample *sample, u64 type)
1280 ret = intel_pt_opt_inject(pt, event, sample, type);
1284 ret = perf_session__deliver_synth_event(pt->session, event, sample);
1286 pr_err("Intel PT: failed to deliver event, error %d\n", ret);
1291 static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq)
1293 struct intel_pt *pt = ptq->pt;
1294 union perf_event *event = ptq->event_buf;
1295 struct perf_sample sample = { .ip = 0, };
1296 struct dummy_branch_stack {
1298 struct branch_entry entries;
1301 if (pt->branches_filter && !(pt->branches_filter & ptq->flags))
1304 if (intel_pt_skip_event(pt))
1307 intel_pt_prep_b_sample(pt, ptq, event, &sample);
1309 sample.id = ptq->pt->branches_id;
1310 sample.stream_id = ptq->pt->branches_id;
1313 * perf report cannot handle events without a branch stack when using
1314 * SORT_MODE__BRANCH so make a dummy one.
1316 if (pt->synth_opts.last_branch && sort__mode == SORT_MODE__BRANCH) {
1317 dummy_bs = (struct dummy_branch_stack){
1324 sample.branch_stack = (struct branch_stack *)&dummy_bs;
1327 sample.cyc_cnt = ptq->ipc_cyc_cnt - ptq->last_br_cyc_cnt;
1328 if (sample.cyc_cnt) {
1329 sample.insn_cnt = ptq->ipc_insn_cnt - ptq->last_br_insn_cnt;
1330 ptq->last_br_insn_cnt = ptq->ipc_insn_cnt;
1331 ptq->last_br_cyc_cnt = ptq->ipc_cyc_cnt;
1334 return intel_pt_deliver_synth_b_event(pt, event, &sample,
1335 pt->branches_sample_type);
1338 static void intel_pt_prep_sample(struct intel_pt *pt,
1339 struct intel_pt_queue *ptq,
1340 union perf_event *event,
1341 struct perf_sample *sample)
1343 intel_pt_prep_b_sample(pt, ptq, event, sample);
1345 if (pt->synth_opts.callchain) {
1346 thread_stack__sample(ptq->thread, ptq->cpu, ptq->chain,
1347 pt->synth_opts.callchain_sz + 1,
1348 sample->ip, pt->kernel_start);
1349 sample->callchain = ptq->chain;
1352 if (pt->synth_opts.last_branch) {
1353 intel_pt_copy_last_branch_rb(ptq);
1354 sample->branch_stack = ptq->last_branch;
1358 static inline int intel_pt_deliver_synth_event(struct intel_pt *pt,
1359 struct intel_pt_queue *ptq,
1360 union perf_event *event,
1361 struct perf_sample *sample,
1366 ret = intel_pt_deliver_synth_b_event(pt, event, sample, type);
1368 if (pt->synth_opts.last_branch)
1369 intel_pt_reset_last_branch_rb(ptq);
1374 static int intel_pt_synth_instruction_sample(struct intel_pt_queue *ptq)
1376 struct intel_pt *pt = ptq->pt;
1377 union perf_event *event = ptq->event_buf;
1378 struct perf_sample sample = { .ip = 0, };
1380 if (intel_pt_skip_event(pt))
1383 intel_pt_prep_sample(pt, ptq, event, &sample);
1385 sample.id = ptq->pt->instructions_id;
1386 sample.stream_id = ptq->pt->instructions_id;
1387 sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt;
1389 sample.cyc_cnt = ptq->ipc_cyc_cnt - ptq->last_in_cyc_cnt;
1390 if (sample.cyc_cnt) {
1391 sample.insn_cnt = ptq->ipc_insn_cnt - ptq->last_in_insn_cnt;
1392 ptq->last_in_insn_cnt = ptq->ipc_insn_cnt;
1393 ptq->last_in_cyc_cnt = ptq->ipc_cyc_cnt;
1396 ptq->last_insn_cnt = ptq->state->tot_insn_cnt;
1398 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1399 pt->instructions_sample_type);
1402 static int intel_pt_synth_transaction_sample(struct intel_pt_queue *ptq)
1404 struct intel_pt *pt = ptq->pt;
1405 union perf_event *event = ptq->event_buf;
1406 struct perf_sample sample = { .ip = 0, };
1408 if (intel_pt_skip_event(pt))
1411 intel_pt_prep_sample(pt, ptq, event, &sample);
1413 sample.id = ptq->pt->transactions_id;
1414 sample.stream_id = ptq->pt->transactions_id;
1416 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1417 pt->transactions_sample_type);
1420 static void intel_pt_prep_p_sample(struct intel_pt *pt,
1421 struct intel_pt_queue *ptq,
1422 union perf_event *event,
1423 struct perf_sample *sample)
1425 intel_pt_prep_sample(pt, ptq, event, sample);
1428 * Zero IP is used to mean "trace start" but that is not the case for
1429 * power or PTWRITE events with no IP, so clear the flags.
1435 static int intel_pt_synth_ptwrite_sample(struct intel_pt_queue *ptq)
1437 struct intel_pt *pt = ptq->pt;
1438 union perf_event *event = ptq->event_buf;
1439 struct perf_sample sample = { .ip = 0, };
1440 struct perf_synth_intel_ptwrite raw;
1442 if (intel_pt_skip_event(pt))
1445 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1447 sample.id = ptq->pt->ptwrites_id;
1448 sample.stream_id = ptq->pt->ptwrites_id;
1451 raw.ip = !!(ptq->state->flags & INTEL_PT_FUP_IP);
1452 raw.payload = cpu_to_le64(ptq->state->ptw_payload);
1454 sample.raw_size = perf_synth__raw_size(raw);
1455 sample.raw_data = perf_synth__raw_data(&raw);
1457 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1458 pt->ptwrites_sample_type);
1461 static int intel_pt_synth_cbr_sample(struct intel_pt_queue *ptq)
1463 struct intel_pt *pt = ptq->pt;
1464 union perf_event *event = ptq->event_buf;
1465 struct perf_sample sample = { .ip = 0, };
1466 struct perf_synth_intel_cbr raw;
1469 if (intel_pt_skip_cbr_event(pt))
1472 ptq->cbr_seen = ptq->state->cbr;
1474 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1476 sample.id = ptq->pt->cbr_id;
1477 sample.stream_id = ptq->pt->cbr_id;
1479 flags = (u16)ptq->state->cbr_payload | (pt->max_non_turbo_ratio << 16);
1480 raw.flags = cpu_to_le32(flags);
1481 raw.freq = cpu_to_le32(raw.cbr * pt->cbr2khz);
1484 sample.raw_size = perf_synth__raw_size(raw);
1485 sample.raw_data = perf_synth__raw_data(&raw);
1487 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1488 pt->pwr_events_sample_type);
1491 static int intel_pt_synth_mwait_sample(struct intel_pt_queue *ptq)
1493 struct intel_pt *pt = ptq->pt;
1494 union perf_event *event = ptq->event_buf;
1495 struct perf_sample sample = { .ip = 0, };
1496 struct perf_synth_intel_mwait raw;
1498 if (intel_pt_skip_event(pt))
1501 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1503 sample.id = ptq->pt->mwait_id;
1504 sample.stream_id = ptq->pt->mwait_id;
1507 raw.payload = cpu_to_le64(ptq->state->mwait_payload);
1509 sample.raw_size = perf_synth__raw_size(raw);
1510 sample.raw_data = perf_synth__raw_data(&raw);
1512 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1513 pt->pwr_events_sample_type);
1516 static int intel_pt_synth_pwre_sample(struct intel_pt_queue *ptq)
1518 struct intel_pt *pt = ptq->pt;
1519 union perf_event *event = ptq->event_buf;
1520 struct perf_sample sample = { .ip = 0, };
1521 struct perf_synth_intel_pwre raw;
1523 if (intel_pt_skip_event(pt))
1526 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1528 sample.id = ptq->pt->pwre_id;
1529 sample.stream_id = ptq->pt->pwre_id;
1532 raw.payload = cpu_to_le64(ptq->state->pwre_payload);
1534 sample.raw_size = perf_synth__raw_size(raw);
1535 sample.raw_data = perf_synth__raw_data(&raw);
1537 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1538 pt->pwr_events_sample_type);
1541 static int intel_pt_synth_exstop_sample(struct intel_pt_queue *ptq)
1543 struct intel_pt *pt = ptq->pt;
1544 union perf_event *event = ptq->event_buf;
1545 struct perf_sample sample = { .ip = 0, };
1546 struct perf_synth_intel_exstop raw;
1548 if (intel_pt_skip_event(pt))
1551 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1553 sample.id = ptq->pt->exstop_id;
1554 sample.stream_id = ptq->pt->exstop_id;
1557 raw.ip = !!(ptq->state->flags & INTEL_PT_FUP_IP);
1559 sample.raw_size = perf_synth__raw_size(raw);
1560 sample.raw_data = perf_synth__raw_data(&raw);
1562 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1563 pt->pwr_events_sample_type);
1566 static int intel_pt_synth_pwrx_sample(struct intel_pt_queue *ptq)
1568 struct intel_pt *pt = ptq->pt;
1569 union perf_event *event = ptq->event_buf;
1570 struct perf_sample sample = { .ip = 0, };
1571 struct perf_synth_intel_pwrx raw;
1573 if (intel_pt_skip_event(pt))
1576 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1578 sample.id = ptq->pt->pwrx_id;
1579 sample.stream_id = ptq->pt->pwrx_id;
1582 raw.payload = cpu_to_le64(ptq->state->pwrx_payload);
1584 sample.raw_size = perf_synth__raw_size(raw);
1585 sample.raw_data = perf_synth__raw_data(&raw);
1587 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1588 pt->pwr_events_sample_type);
1592 * PEBS gp_regs array indexes plus 1 so that 0 means not present. Refer
1593 * intel_pt_add_gp_regs().
1595 static const int pebs_gp_regs[] = {
1596 [PERF_REG_X86_FLAGS] = 1,
1597 [PERF_REG_X86_IP] = 2,
1598 [PERF_REG_X86_AX] = 3,
1599 [PERF_REG_X86_CX] = 4,
1600 [PERF_REG_X86_DX] = 5,
1601 [PERF_REG_X86_BX] = 6,
1602 [PERF_REG_X86_SP] = 7,
1603 [PERF_REG_X86_BP] = 8,
1604 [PERF_REG_X86_SI] = 9,
1605 [PERF_REG_X86_DI] = 10,
1606 [PERF_REG_X86_R8] = 11,
1607 [PERF_REG_X86_R9] = 12,
1608 [PERF_REG_X86_R10] = 13,
1609 [PERF_REG_X86_R11] = 14,
1610 [PERF_REG_X86_R12] = 15,
1611 [PERF_REG_X86_R13] = 16,
1612 [PERF_REG_X86_R14] = 17,
1613 [PERF_REG_X86_R15] = 18,
1616 static u64 *intel_pt_add_gp_regs(struct regs_dump *intr_regs, u64 *pos,
1617 const struct intel_pt_blk_items *items,
1620 const u64 *gp_regs = items->val[INTEL_PT_GP_REGS_POS];
1621 u32 mask = items->mask[INTEL_PT_GP_REGS_POS];
1625 for (i = 0, bit = 1; i < PERF_REG_X86_64_MAX; i++, bit <<= 1) {
1626 /* Get the PEBS gp_regs array index */
1627 int n = pebs_gp_regs[i] - 1;
1632 * Add only registers that were requested (i.e. 'regs_mask') and
1633 * that were provided (i.e. 'mask'), and update the resulting
1634 * mask (i.e. 'intr_regs->mask') accordingly.
1636 if (mask & 1 << n && regs_mask & bit) {
1637 intr_regs->mask |= bit;
1638 *pos++ = gp_regs[n];
1645 #ifndef PERF_REG_X86_XMM0
1646 #define PERF_REG_X86_XMM0 32
1649 static void intel_pt_add_xmm(struct regs_dump *intr_regs, u64 *pos,
1650 const struct intel_pt_blk_items *items,
1653 u32 mask = items->has_xmm & (regs_mask >> PERF_REG_X86_XMM0);
1654 const u64 *xmm = items->xmm;
1657 * If there are any XMM registers, then there should be all of them.
1658 * Nevertheless, follow the logic to add only registers that were
1659 * requested (i.e. 'regs_mask') and that were provided (i.e. 'mask'),
1660 * and update the resulting mask (i.e. 'intr_regs->mask') accordingly.
1662 intr_regs->mask |= (u64)mask << PERF_REG_X86_XMM0;
1664 for (; mask; mask >>= 1, xmm++) {
1670 #define LBR_INFO_MISPRED (1ULL << 63)
1671 #define LBR_INFO_IN_TX (1ULL << 62)
1672 #define LBR_INFO_ABORT (1ULL << 61)
1673 #define LBR_INFO_CYCLES 0xffff
1675 /* Refer kernel's intel_pmu_store_pebs_lbrs() */
1676 static u64 intel_pt_lbr_flags(u64 info)
1679 struct branch_flags flags;
1683 .mispred = !!(info & LBR_INFO_MISPRED),
1684 .predicted = !(info & LBR_INFO_MISPRED),
1685 .in_tx = !!(info & LBR_INFO_IN_TX),
1686 .abort = !!(info & LBR_INFO_ABORT),
1687 .cycles = info & LBR_INFO_CYCLES,
1694 static void intel_pt_add_lbrs(struct branch_stack *br_stack,
1695 const struct intel_pt_blk_items *items)
1702 to = &br_stack->entries[0].from;
1704 for (i = INTEL_PT_LBR_0_POS; i <= INTEL_PT_LBR_2_POS; i++) {
1705 u32 mask = items->mask[i];
1706 const u64 *from = items->val[i];
1708 for (; mask; mask >>= 3, from += 3) {
1709 if ((mask & 7) == 7) {
1712 *to++ = intel_pt_lbr_flags(from[2]);
1719 /* INTEL_PT_LBR_0, INTEL_PT_LBR_1 and INTEL_PT_LBR_2 */
1720 #define LBRS_MAX (INTEL_PT_BLK_ITEM_ID_CNT * 3)
1722 static int intel_pt_synth_pebs_sample(struct intel_pt_queue *ptq)
1724 const struct intel_pt_blk_items *items = &ptq->state->items;
1725 struct perf_sample sample = { .ip = 0, };
1726 union perf_event *event = ptq->event_buf;
1727 struct intel_pt *pt = ptq->pt;
1728 struct evsel *evsel = pt->pebs_evsel;
1729 u64 sample_type = evsel->core.attr.sample_type;
1730 u64 id = evsel->core.id[0];
1733 if (intel_pt_skip_event(pt))
1736 intel_pt_prep_a_sample(ptq, event, &sample);
1739 sample.stream_id = id;
1741 if (!evsel->core.attr.freq)
1742 sample.period = evsel->core.attr.sample_period;
1744 /* No support for non-zero CS base */
1746 sample.ip = items->ip;
1747 else if (items->has_rip)
1748 sample.ip = items->rip;
1750 sample.ip = ptq->state->from_ip;
1752 /* No support for guest mode at this time */
1753 cpumode = sample.ip < ptq->pt->kernel_start ?
1754 PERF_RECORD_MISC_USER :
1755 PERF_RECORD_MISC_KERNEL;
1757 event->sample.header.misc = cpumode | PERF_RECORD_MISC_EXACT_IP;
1759 sample.cpumode = cpumode;
1761 if (sample_type & PERF_SAMPLE_TIME) {
1764 if (items->has_timestamp)
1765 timestamp = items->timestamp;
1766 else if (!pt->timeless_decoding)
1767 timestamp = ptq->timestamp;
1769 sample.time = tsc_to_perf_time(timestamp, &pt->tc);
1772 if (sample_type & PERF_SAMPLE_CALLCHAIN &&
1773 pt->synth_opts.callchain) {
1774 thread_stack__sample(ptq->thread, ptq->cpu, ptq->chain,
1775 pt->synth_opts.callchain_sz, sample.ip,
1777 sample.callchain = ptq->chain;
1780 if (sample_type & PERF_SAMPLE_REGS_INTR &&
1781 items->mask[INTEL_PT_GP_REGS_POS]) {
1782 u64 regs[sizeof(sample.intr_regs.mask)];
1783 u64 regs_mask = evsel->core.attr.sample_regs_intr;
1786 sample.intr_regs.abi = items->is_32_bit ?
1787 PERF_SAMPLE_REGS_ABI_32 :
1788 PERF_SAMPLE_REGS_ABI_64;
1789 sample.intr_regs.regs = regs;
1791 pos = intel_pt_add_gp_regs(&sample.intr_regs, regs, items, regs_mask);
1793 intel_pt_add_xmm(&sample.intr_regs, pos, items, regs_mask);
1796 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1798 struct branch_stack br_stack;
1799 struct branch_entry entries[LBRS_MAX];
1802 if (items->mask[INTEL_PT_LBR_0_POS] ||
1803 items->mask[INTEL_PT_LBR_1_POS] ||
1804 items->mask[INTEL_PT_LBR_2_POS]) {
1805 intel_pt_add_lbrs(&br.br_stack, items);
1806 sample.branch_stack = &br.br_stack;
1807 } else if (pt->synth_opts.last_branch) {
1808 intel_pt_copy_last_branch_rb(ptq);
1809 sample.branch_stack = ptq->last_branch;
1812 sample.branch_stack = &br.br_stack;
1816 if (sample_type & PERF_SAMPLE_ADDR && items->has_mem_access_address)
1817 sample.addr = items->mem_access_address;
1819 if (sample_type & PERF_SAMPLE_WEIGHT) {
1821 * Refer kernel's setup_pebs_adaptive_sample_data() and
1822 * intel_hsw_weight().
1824 if (items->has_mem_access_latency)
1825 sample.weight = items->mem_access_latency;
1826 if (!sample.weight && items->has_tsx_aux_info) {
1827 /* Cycles last block */
1828 sample.weight = (u32)items->tsx_aux_info;
1832 if (sample_type & PERF_SAMPLE_TRANSACTION && items->has_tsx_aux_info) {
1833 u64 ax = items->has_rax ? items->rax : 0;
1834 /* Refer kernel's intel_hsw_transaction() */
1835 u64 txn = (u8)(items->tsx_aux_info >> 32);
1837 /* For RTM XABORTs also log the abort code from AX */
1838 if (txn & PERF_TXN_TRANSACTION && ax & 1)
1839 txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1840 sample.transaction = txn;
1843 return intel_pt_deliver_synth_event(pt, ptq, event, &sample, sample_type);
1846 static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,
1847 pid_t pid, pid_t tid, u64 ip, u64 timestamp)
1849 union perf_event event;
1850 char msg[MAX_AUXTRACE_ERROR_MSG];
1853 intel_pt__strerror(code, msg, MAX_AUXTRACE_ERROR_MSG);
1855 auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
1856 code, cpu, pid, tid, ip, msg, timestamp);
1858 err = perf_session__deliver_synth_event(pt->session, &event, NULL);
1860 pr_err("Intel Processor Trace: failed to deliver error event, error %d\n",
1866 static int intel_ptq_synth_error(struct intel_pt_queue *ptq,
1867 const struct intel_pt_state *state)
1869 struct intel_pt *pt = ptq->pt;
1870 u64 tm = ptq->timestamp;
1872 tm = pt->timeless_decoding ? 0 : tsc_to_perf_time(tm, &pt->tc);
1874 return intel_pt_synth_error(pt, state->err, ptq->cpu, ptq->pid,
1875 ptq->tid, state->from_ip, tm);
1878 static int intel_pt_next_tid(struct intel_pt *pt, struct intel_pt_queue *ptq)
1880 struct auxtrace_queue *queue;
1881 pid_t tid = ptq->next_tid;
1887 intel_pt_log("switch: cpu %d tid %d\n", ptq->cpu, tid);
1889 err = machine__set_current_tid(pt->machine, ptq->cpu, -1, tid);
1891 queue = &pt->queues.queue_array[ptq->queue_nr];
1892 intel_pt_set_pid_tid_cpu(pt, queue);
1899 static inline bool intel_pt_is_switch_ip(struct intel_pt_queue *ptq, u64 ip)
1901 struct intel_pt *pt = ptq->pt;
1903 return ip == pt->switch_ip &&
1904 (ptq->flags & PERF_IP_FLAG_BRANCH) &&
1905 !(ptq->flags & (PERF_IP_FLAG_CONDITIONAL | PERF_IP_FLAG_ASYNC |
1906 PERF_IP_FLAG_INTERRUPT | PERF_IP_FLAG_TX_ABORT));
1909 #define INTEL_PT_PWR_EVT (INTEL_PT_MWAIT_OP | INTEL_PT_PWR_ENTRY | \
1910 INTEL_PT_EX_STOP | INTEL_PT_PWR_EXIT)
1912 static int intel_pt_sample(struct intel_pt_queue *ptq)
1914 const struct intel_pt_state *state = ptq->state;
1915 struct intel_pt *pt = ptq->pt;
1918 if (!ptq->have_sample)
1921 ptq->have_sample = false;
1923 if (ptq->state->tot_cyc_cnt > ptq->ipc_cyc_cnt) {
1925 * Cycle count and instruction count only go together to create
1926 * a valid IPC ratio when the cycle count changes.
1928 ptq->ipc_insn_cnt = ptq->state->tot_insn_cnt;
1929 ptq->ipc_cyc_cnt = ptq->state->tot_cyc_cnt;
1933 * Do PEBS first to allow for the possibility that the PEBS timestamp
1934 * precedes the current timestamp.
1936 if (pt->sample_pebs && state->type & INTEL_PT_BLK_ITEMS) {
1937 err = intel_pt_synth_pebs_sample(ptq);
1942 if (pt->sample_pwr_events) {
1943 if (ptq->state->cbr != ptq->cbr_seen) {
1944 err = intel_pt_synth_cbr_sample(ptq);
1948 if (state->type & INTEL_PT_PWR_EVT) {
1949 if (state->type & INTEL_PT_MWAIT_OP) {
1950 err = intel_pt_synth_mwait_sample(ptq);
1954 if (state->type & INTEL_PT_PWR_ENTRY) {
1955 err = intel_pt_synth_pwre_sample(ptq);
1959 if (state->type & INTEL_PT_EX_STOP) {
1960 err = intel_pt_synth_exstop_sample(ptq);
1964 if (state->type & INTEL_PT_PWR_EXIT) {
1965 err = intel_pt_synth_pwrx_sample(ptq);
1972 if (pt->sample_instructions && (state->type & INTEL_PT_INSTRUCTION)) {
1973 err = intel_pt_synth_instruction_sample(ptq);
1978 if (pt->sample_transactions && (state->type & INTEL_PT_TRANSACTION)) {
1979 err = intel_pt_synth_transaction_sample(ptq);
1984 if (pt->sample_ptwrites && (state->type & INTEL_PT_PTW)) {
1985 err = intel_pt_synth_ptwrite_sample(ptq);
1990 if (!(state->type & INTEL_PT_BRANCH))
1993 if (pt->synth_opts.callchain || pt->synth_opts.thread_stack)
1994 thread_stack__event(ptq->thread, ptq->cpu, ptq->flags, state->from_ip,
1995 state->to_ip, ptq->insn_len,
1998 thread_stack__set_trace_nr(ptq->thread, ptq->cpu, state->trace_nr);
2000 if (pt->sample_branches) {
2001 err = intel_pt_synth_branch_sample(ptq);
2006 if (pt->synth_opts.last_branch)
2007 intel_pt_update_last_branch_rb(ptq);
2009 if (!ptq->sync_switch)
2012 if (intel_pt_is_switch_ip(ptq, state->to_ip)) {
2013 switch (ptq->switch_state) {
2014 case INTEL_PT_SS_NOT_TRACING:
2015 case INTEL_PT_SS_UNKNOWN:
2016 case INTEL_PT_SS_EXPECTING_SWITCH_IP:
2017 err = intel_pt_next_tid(pt, ptq);
2020 ptq->switch_state = INTEL_PT_SS_TRACING;
2023 ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_EVENT;
2026 } else if (!state->to_ip) {
2027 ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
2028 } else if (ptq->switch_state == INTEL_PT_SS_NOT_TRACING) {
2029 ptq->switch_state = INTEL_PT_SS_UNKNOWN;
2030 } else if (ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
2031 state->to_ip == pt->ptss_ip &&
2032 (ptq->flags & PERF_IP_FLAG_CALL)) {
2033 ptq->switch_state = INTEL_PT_SS_TRACING;
2039 static u64 intel_pt_switch_ip(struct intel_pt *pt, u64 *ptss_ip)
2041 struct machine *machine = pt->machine;
2043 struct symbol *sym, *start;
2044 u64 ip, switch_ip = 0;
2050 map = machine__kernel_map(machine);
2057 start = dso__first_symbol(map->dso);
2059 for (sym = start; sym; sym = dso__next_symbol(sym)) {
2060 if (sym->binding == STB_GLOBAL &&
2061 !strcmp(sym->name, "__switch_to")) {
2062 ip = map->unmap_ip(map, sym->start);
2063 if (ip >= map->start && ip < map->end) {
2070 if (!switch_ip || !ptss_ip)
2073 if (pt->have_sched_switch == 1)
2074 ptss = "perf_trace_sched_switch";
2076 ptss = "__perf_event_task_sched_out";
2078 for (sym = start; sym; sym = dso__next_symbol(sym)) {
2079 if (!strcmp(sym->name, ptss)) {
2080 ip = map->unmap_ip(map, sym->start);
2081 if (ip >= map->start && ip < map->end) {
2091 static void intel_pt_enable_sync_switch(struct intel_pt *pt)
2095 pt->sync_switch = true;
2097 for (i = 0; i < pt->queues.nr_queues; i++) {
2098 struct auxtrace_queue *queue = &pt->queues.queue_array[i];
2099 struct intel_pt_queue *ptq = queue->priv;
2102 ptq->sync_switch = true;
2107 * To filter against time ranges, it is only necessary to look at the next start
2110 static bool intel_pt_next_time(struct intel_pt_queue *ptq)
2112 struct intel_pt *pt = ptq->pt;
2114 if (ptq->sel_start) {
2115 /* Next time is an end time */
2116 ptq->sel_start = false;
2117 ptq->sel_timestamp = pt->time_ranges[ptq->sel_idx].end;
2119 } else if (ptq->sel_idx + 1 < pt->range_cnt) {
2120 /* Next time is a start time */
2121 ptq->sel_start = true;
2123 ptq->sel_timestamp = pt->time_ranges[ptq->sel_idx].start;
2131 static int intel_pt_time_filter(struct intel_pt_queue *ptq, u64 *ff_timestamp)
2136 if (ptq->sel_start) {
2137 if (ptq->timestamp >= ptq->sel_timestamp) {
2138 /* After start time, so consider next time */
2139 intel_pt_next_time(ptq);
2140 if (!ptq->sel_timestamp) {
2144 /* Check against end time */
2147 /* Before start time, so fast forward */
2148 ptq->have_sample = false;
2149 if (ptq->sel_timestamp > *ff_timestamp) {
2150 if (ptq->sync_switch) {
2151 intel_pt_next_tid(ptq->pt, ptq);
2152 ptq->switch_state = INTEL_PT_SS_UNKNOWN;
2154 *ff_timestamp = ptq->sel_timestamp;
2155 err = intel_pt_fast_forward(ptq->decoder,
2156 ptq->sel_timestamp);
2161 } else if (ptq->timestamp > ptq->sel_timestamp) {
2162 /* After end time, so consider next time */
2163 if (!intel_pt_next_time(ptq)) {
2164 /* No next time range, so stop decoding */
2165 ptq->have_sample = false;
2166 ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
2169 /* Check against next start time */
2172 /* Before end time */
2178 static int intel_pt_run_decoder(struct intel_pt_queue *ptq, u64 *timestamp)
2180 const struct intel_pt_state *state = ptq->state;
2181 struct intel_pt *pt = ptq->pt;
2182 u64 ff_timestamp = 0;
2185 if (!pt->kernel_start) {
2186 pt->kernel_start = machine__kernel_start(pt->machine);
2187 if (pt->per_cpu_mmaps &&
2188 (pt->have_sched_switch == 1 || pt->have_sched_switch == 3) &&
2189 !pt->timeless_decoding && intel_pt_tracing_kernel(pt) &&
2190 !pt->sampling_mode) {
2191 pt->switch_ip = intel_pt_switch_ip(pt, &pt->ptss_ip);
2192 if (pt->switch_ip) {
2193 intel_pt_log("switch_ip: %"PRIx64" ptss_ip: %"PRIx64"\n",
2194 pt->switch_ip, pt->ptss_ip);
2195 intel_pt_enable_sync_switch(pt);
2200 intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
2201 ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
2203 err = intel_pt_sample(ptq);
2207 state = intel_pt_decode(ptq->decoder);
2209 if (state->err == INTEL_PT_ERR_NODATA)
2211 if (ptq->sync_switch &&
2212 state->from_ip >= pt->kernel_start) {
2213 ptq->sync_switch = false;
2214 intel_pt_next_tid(pt, ptq);
2216 if (pt->synth_opts.errors) {
2217 err = intel_ptq_synth_error(ptq, state);
2225 ptq->have_sample = true;
2226 intel_pt_sample_flags(ptq);
2228 /* Use estimated TSC upon return to user space */
2230 (state->from_ip >= pt->kernel_start || !state->from_ip) &&
2231 state->to_ip && state->to_ip < pt->kernel_start) {
2232 intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
2233 state->timestamp, state->est_timestamp);
2234 ptq->timestamp = state->est_timestamp;
2235 /* Use estimated TSC in unknown switch state */
2236 } else if (ptq->sync_switch &&
2237 ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
2238 intel_pt_is_switch_ip(ptq, state->to_ip) &&
2239 ptq->next_tid == -1) {
2240 intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
2241 state->timestamp, state->est_timestamp);
2242 ptq->timestamp = state->est_timestamp;
2243 } else if (state->timestamp > ptq->timestamp) {
2244 ptq->timestamp = state->timestamp;
2247 if (ptq->sel_timestamp) {
2248 err = intel_pt_time_filter(ptq, &ff_timestamp);
2253 if (!pt->timeless_decoding && ptq->timestamp >= *timestamp) {
2254 *timestamp = ptq->timestamp;
2261 static inline int intel_pt_update_queues(struct intel_pt *pt)
2263 if (pt->queues.new_data) {
2264 pt->queues.new_data = false;
2265 return intel_pt_setup_queues(pt);
2270 static int intel_pt_process_queues(struct intel_pt *pt, u64 timestamp)
2272 unsigned int queue_nr;
2277 struct auxtrace_queue *queue;
2278 struct intel_pt_queue *ptq;
2280 if (!pt->heap.heap_cnt)
2283 if (pt->heap.heap_array[0].ordinal >= timestamp)
2286 queue_nr = pt->heap.heap_array[0].queue_nr;
2287 queue = &pt->queues.queue_array[queue_nr];
2290 intel_pt_log("queue %u processing 0x%" PRIx64 " to 0x%" PRIx64 "\n",
2291 queue_nr, pt->heap.heap_array[0].ordinal,
2294 auxtrace_heap__pop(&pt->heap);
2296 if (pt->heap.heap_cnt) {
2297 ts = pt->heap.heap_array[0].ordinal + 1;
2304 intel_pt_set_pid_tid_cpu(pt, queue);
2306 ret = intel_pt_run_decoder(ptq, &ts);
2309 auxtrace_heap__add(&pt->heap, queue_nr, ts);
2314 ret = auxtrace_heap__add(&pt->heap, queue_nr, ts);
2318 ptq->on_heap = false;
2325 static int intel_pt_process_timeless_queues(struct intel_pt *pt, pid_t tid,
2328 struct auxtrace_queues *queues = &pt->queues;
2332 for (i = 0; i < queues->nr_queues; i++) {
2333 struct auxtrace_queue *queue = &pt->queues.queue_array[i];
2334 struct intel_pt_queue *ptq = queue->priv;
2336 if (ptq && (tid == -1 || ptq->tid == tid)) {
2338 intel_pt_set_pid_tid_cpu(pt, queue);
2339 intel_pt_run_decoder(ptq, &ts);
2345 static void intel_pt_sample_set_pid_tid_cpu(struct intel_pt_queue *ptq,
2346 struct auxtrace_queue *queue,
2347 struct perf_sample *sample)
2349 struct machine *m = ptq->pt->machine;
2351 ptq->pid = sample->pid;
2352 ptq->tid = sample->tid;
2353 ptq->cpu = queue->cpu;
2355 intel_pt_log("queue %u cpu %d pid %d tid %d\n",
2356 ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
2358 thread__zput(ptq->thread);
2363 if (ptq->pid == -1) {
2364 ptq->thread = machine__find_thread(m, -1, ptq->tid);
2366 ptq->pid = ptq->thread->pid_;
2370 ptq->thread = machine__findnew_thread(m, ptq->pid, ptq->tid);
2373 static int intel_pt_process_timeless_sample(struct intel_pt *pt,
2374 struct perf_sample *sample)
2376 struct auxtrace_queue *queue;
2377 struct intel_pt_queue *ptq;
2380 queue = auxtrace_queues__sample_queue(&pt->queues, sample, pt->session);
2389 ptq->time = sample->time;
2390 intel_pt_sample_set_pid_tid_cpu(ptq, queue, sample);
2391 intel_pt_run_decoder(ptq, &ts);
2395 static int intel_pt_lost(struct intel_pt *pt, struct perf_sample *sample)
2397 return intel_pt_synth_error(pt, INTEL_PT_ERR_LOST, sample->cpu,
2398 sample->pid, sample->tid, 0, sample->time);
2401 static struct intel_pt_queue *intel_pt_cpu_to_ptq(struct intel_pt *pt, int cpu)
2405 if (cpu < 0 || !pt->queues.nr_queues)
2408 if ((unsigned)cpu >= pt->queues.nr_queues)
2409 i = pt->queues.nr_queues - 1;
2413 if (pt->queues.queue_array[i].cpu == cpu)
2414 return pt->queues.queue_array[i].priv;
2416 for (j = 0; i > 0; j++) {
2417 if (pt->queues.queue_array[--i].cpu == cpu)
2418 return pt->queues.queue_array[i].priv;
2421 for (; j < pt->queues.nr_queues; j++) {
2422 if (pt->queues.queue_array[j].cpu == cpu)
2423 return pt->queues.queue_array[j].priv;
2429 static int intel_pt_sync_switch(struct intel_pt *pt, int cpu, pid_t tid,
2432 struct intel_pt_queue *ptq;
2435 if (!pt->sync_switch)
2438 ptq = intel_pt_cpu_to_ptq(pt, cpu);
2439 if (!ptq || !ptq->sync_switch)
2442 switch (ptq->switch_state) {
2443 case INTEL_PT_SS_NOT_TRACING:
2445 case INTEL_PT_SS_UNKNOWN:
2446 case INTEL_PT_SS_TRACING:
2447 ptq->next_tid = tid;
2448 ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_IP;
2450 case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
2451 if (!ptq->on_heap) {
2452 ptq->timestamp = perf_time_to_tsc(timestamp,
2454 err = auxtrace_heap__add(&pt->heap, ptq->queue_nr,
2458 ptq->on_heap = true;
2460 ptq->switch_state = INTEL_PT_SS_TRACING;
2462 case INTEL_PT_SS_EXPECTING_SWITCH_IP:
2463 intel_pt_log("ERROR: cpu %d expecting switch ip\n", cpu);
2474 static int intel_pt_process_switch(struct intel_pt *pt,
2475 struct perf_sample *sample)
2477 struct evsel *evsel;
2481 evsel = perf_evlist__id2evsel(pt->session->evlist, sample->id);
2482 if (evsel != pt->switch_evsel)
2485 tid = perf_evsel__intval(evsel, sample, "next_pid");
2488 intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
2489 cpu, tid, sample->time, perf_time_to_tsc(sample->time,
2492 ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
2496 return machine__set_current_tid(pt->machine, cpu, -1, tid);
2499 static int intel_pt_context_switch_in(struct intel_pt *pt,
2500 struct perf_sample *sample)
2502 pid_t pid = sample->pid;
2503 pid_t tid = sample->tid;
2504 int cpu = sample->cpu;
2506 if (pt->sync_switch) {
2507 struct intel_pt_queue *ptq;
2509 ptq = intel_pt_cpu_to_ptq(pt, cpu);
2510 if (ptq && ptq->sync_switch) {
2512 switch (ptq->switch_state) {
2513 case INTEL_PT_SS_NOT_TRACING:
2514 case INTEL_PT_SS_UNKNOWN:
2515 case INTEL_PT_SS_TRACING:
2517 case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
2518 case INTEL_PT_SS_EXPECTING_SWITCH_IP:
2519 ptq->switch_state = INTEL_PT_SS_TRACING;
2528 * If the current tid has not been updated yet, ensure it is now that
2529 * a "switch in" event has occurred.
2531 if (machine__get_current_tid(pt->machine, cpu) == tid)
2534 return machine__set_current_tid(pt->machine, cpu, pid, tid);
2537 static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,
2538 struct perf_sample *sample)
2540 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
2546 if (pt->have_sched_switch == 3) {
2548 return intel_pt_context_switch_in(pt, sample);
2549 if (event->header.type != PERF_RECORD_SWITCH_CPU_WIDE) {
2550 pr_err("Expecting CPU-wide context switch event\n");
2553 pid = event->context_switch.next_prev_pid;
2554 tid = event->context_switch.next_prev_tid;
2563 pr_err("context_switch event has no tid\n");
2567 intel_pt_log("context_switch: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
2568 cpu, pid, tid, sample->time, perf_time_to_tsc(sample->time,
2571 ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
2575 return machine__set_current_tid(pt->machine, cpu, pid, tid);
2578 static int intel_pt_process_itrace_start(struct intel_pt *pt,
2579 union perf_event *event,
2580 struct perf_sample *sample)
2582 if (!pt->per_cpu_mmaps)
2585 intel_pt_log("itrace_start: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
2586 sample->cpu, event->itrace_start.pid,
2587 event->itrace_start.tid, sample->time,
2588 perf_time_to_tsc(sample->time, &pt->tc));
2590 return machine__set_current_tid(pt->machine, sample->cpu,
2591 event->itrace_start.pid,
2592 event->itrace_start.tid);
2595 static int intel_pt_process_event(struct perf_session *session,
2596 union perf_event *event,
2597 struct perf_sample *sample,
2598 struct perf_tool *tool)
2600 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2608 if (!tool->ordered_events) {
2609 pr_err("Intel Processor Trace requires ordered events\n");
2613 if (sample->time && sample->time != (u64)-1)
2614 timestamp = perf_time_to_tsc(sample->time, &pt->tc);
2618 if (timestamp || pt->timeless_decoding) {
2619 err = intel_pt_update_queues(pt);
2624 if (pt->timeless_decoding) {
2625 if (pt->sampling_mode) {
2626 if (sample->aux_sample.size)
2627 err = intel_pt_process_timeless_sample(pt,
2629 } else if (event->header.type == PERF_RECORD_EXIT) {
2630 err = intel_pt_process_timeless_queues(pt,
2634 } else if (timestamp) {
2635 err = intel_pt_process_queues(pt, timestamp);
2640 if (event->header.type == PERF_RECORD_AUX &&
2641 (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
2642 pt->synth_opts.errors) {
2643 err = intel_pt_lost(pt, sample);
2648 if (pt->switch_evsel && event->header.type == PERF_RECORD_SAMPLE)
2649 err = intel_pt_process_switch(pt, sample);
2650 else if (event->header.type == PERF_RECORD_ITRACE_START)
2651 err = intel_pt_process_itrace_start(pt, event, sample);
2652 else if (event->header.type == PERF_RECORD_SWITCH ||
2653 event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
2654 err = intel_pt_context_switch(pt, event, sample);
2656 intel_pt_log("event %u: cpu %d time %"PRIu64" tsc %#"PRIx64" ",
2657 event->header.type, sample->cpu, sample->time, timestamp);
2658 intel_pt_log_event(event);
2663 static int intel_pt_flush(struct perf_session *session, struct perf_tool *tool)
2665 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2672 if (!tool->ordered_events)
2675 ret = intel_pt_update_queues(pt);
2679 if (pt->timeless_decoding)
2680 return intel_pt_process_timeless_queues(pt, -1,
2683 return intel_pt_process_queues(pt, MAX_TIMESTAMP);
2686 static void intel_pt_free_events(struct perf_session *session)
2688 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2690 struct auxtrace_queues *queues = &pt->queues;
2693 for (i = 0; i < queues->nr_queues; i++) {
2694 intel_pt_free_queue(queues->queue_array[i].priv);
2695 queues->queue_array[i].priv = NULL;
2697 intel_pt_log_disable();
2698 auxtrace_queues__free(queues);
2701 static void intel_pt_free(struct perf_session *session)
2703 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2706 auxtrace_heap__free(&pt->heap);
2707 intel_pt_free_events(session);
2708 session->auxtrace = NULL;
2709 thread__put(pt->unknown_thread);
2710 addr_filters__exit(&pt->filts);
2712 zfree(&pt->time_ranges);
2716 static int intel_pt_process_auxtrace_event(struct perf_session *session,
2717 union perf_event *event,
2718 struct perf_tool *tool __maybe_unused)
2720 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2723 if (!pt->data_queued) {
2724 struct auxtrace_buffer *buffer;
2726 int fd = perf_data__fd(session->data);
2729 if (perf_data__is_pipe(session->data)) {
2732 data_offset = lseek(fd, 0, SEEK_CUR);
2733 if (data_offset == -1)
2737 err = auxtrace_queues__add_event(&pt->queues, session, event,
2738 data_offset, &buffer);
2742 /* Dump here now we have copied a piped trace out of the pipe */
2744 if (auxtrace_buffer__get_data(buffer, fd)) {
2745 intel_pt_dump_event(pt, buffer->data,
2747 auxtrace_buffer__put_data(buffer);
2755 static int intel_pt_queue_data(struct perf_session *session,
2756 struct perf_sample *sample,
2757 union perf_event *event, u64 data_offset)
2759 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2764 return auxtrace_queues__add_event(&pt->queues, session, event,
2768 if (sample->time && sample->time != (u64)-1)
2769 timestamp = perf_time_to_tsc(sample->time, &pt->tc);
2773 return auxtrace_queues__add_sample(&pt->queues, session, sample,
2774 data_offset, timestamp);
2777 struct intel_pt_synth {
2778 struct perf_tool dummy_tool;
2779 struct perf_session *session;
2782 static int intel_pt_event_synth(struct perf_tool *tool,
2783 union perf_event *event,
2784 struct perf_sample *sample __maybe_unused,
2785 struct machine *machine __maybe_unused)
2787 struct intel_pt_synth *intel_pt_synth =
2788 container_of(tool, struct intel_pt_synth, dummy_tool);
2790 return perf_session__deliver_synth_event(intel_pt_synth->session, event,
2794 static int intel_pt_synth_event(struct perf_session *session, const char *name,
2795 struct perf_event_attr *attr, u64 id)
2797 struct intel_pt_synth intel_pt_synth;
2800 pr_debug("Synthesizing '%s' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
2801 name, id, (u64)attr->sample_type);
2803 memset(&intel_pt_synth, 0, sizeof(struct intel_pt_synth));
2804 intel_pt_synth.session = session;
2806 err = perf_event__synthesize_attr(&intel_pt_synth.dummy_tool, attr, 1,
2807 &id, intel_pt_event_synth);
2809 pr_err("%s: failed to synthesize '%s' event type\n",
2815 static void intel_pt_set_event_name(struct evlist *evlist, u64 id,
2818 struct evsel *evsel;
2820 evlist__for_each_entry(evlist, evsel) {
2821 if (evsel->core.id && evsel->core.id[0] == id) {
2823 zfree(&evsel->name);
2824 evsel->name = strdup(name);
2830 static struct evsel *intel_pt_evsel(struct intel_pt *pt,
2831 struct evlist *evlist)
2833 struct evsel *evsel;
2835 evlist__for_each_entry(evlist, evsel) {
2836 if (evsel->core.attr.type == pt->pmu_type && evsel->core.ids)
2843 static int intel_pt_synth_events(struct intel_pt *pt,
2844 struct perf_session *session)
2846 struct evlist *evlist = session->evlist;
2847 struct evsel *evsel = intel_pt_evsel(pt, evlist);
2848 struct perf_event_attr attr;
2853 pr_debug("There are no selected events with Intel Processor Trace data\n");
2857 memset(&attr, 0, sizeof(struct perf_event_attr));
2858 attr.size = sizeof(struct perf_event_attr);
2859 attr.type = PERF_TYPE_HARDWARE;
2860 attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK;
2861 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
2863 if (pt->timeless_decoding)
2864 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
2866 attr.sample_type |= PERF_SAMPLE_TIME;
2867 if (!pt->per_cpu_mmaps)
2868 attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
2869 attr.exclude_user = evsel->core.attr.exclude_user;
2870 attr.exclude_kernel = evsel->core.attr.exclude_kernel;
2871 attr.exclude_hv = evsel->core.attr.exclude_hv;
2872 attr.exclude_host = evsel->core.attr.exclude_host;
2873 attr.exclude_guest = evsel->core.attr.exclude_guest;
2874 attr.sample_id_all = evsel->core.attr.sample_id_all;
2875 attr.read_format = evsel->core.attr.read_format;
2877 id = evsel->core.id[0] + 1000000000;
2881 if (pt->synth_opts.branches) {
2882 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
2883 attr.sample_period = 1;
2884 attr.sample_type |= PERF_SAMPLE_ADDR;
2885 err = intel_pt_synth_event(session, "branches", &attr, id);
2888 pt->sample_branches = true;
2889 pt->branches_sample_type = attr.sample_type;
2890 pt->branches_id = id;
2892 attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR;
2895 if (pt->synth_opts.callchain)
2896 attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
2897 if (pt->synth_opts.last_branch)
2898 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
2900 if (pt->synth_opts.instructions) {
2901 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
2902 if (pt->synth_opts.period_type == PERF_ITRACE_PERIOD_NANOSECS)
2903 attr.sample_period =
2904 intel_pt_ns_to_ticks(pt, pt->synth_opts.period);
2906 attr.sample_period = pt->synth_opts.period;
2907 err = intel_pt_synth_event(session, "instructions", &attr, id);
2910 pt->sample_instructions = true;
2911 pt->instructions_sample_type = attr.sample_type;
2912 pt->instructions_id = id;
2916 attr.sample_type &= ~(u64)PERF_SAMPLE_PERIOD;
2917 attr.sample_period = 1;
2919 if (pt->synth_opts.transactions) {
2920 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
2921 err = intel_pt_synth_event(session, "transactions", &attr, id);
2924 pt->sample_transactions = true;
2925 pt->transactions_sample_type = attr.sample_type;
2926 pt->transactions_id = id;
2927 intel_pt_set_event_name(evlist, id, "transactions");
2931 attr.type = PERF_TYPE_SYNTH;
2932 attr.sample_type |= PERF_SAMPLE_RAW;
2934 if (pt->synth_opts.ptwrites) {
2935 attr.config = PERF_SYNTH_INTEL_PTWRITE;
2936 err = intel_pt_synth_event(session, "ptwrite", &attr, id);
2939 pt->sample_ptwrites = true;
2940 pt->ptwrites_sample_type = attr.sample_type;
2941 pt->ptwrites_id = id;
2942 intel_pt_set_event_name(evlist, id, "ptwrite");
2946 if (pt->synth_opts.pwr_events) {
2947 pt->sample_pwr_events = true;
2948 pt->pwr_events_sample_type = attr.sample_type;
2950 attr.config = PERF_SYNTH_INTEL_CBR;
2951 err = intel_pt_synth_event(session, "cbr", &attr, id);
2955 intel_pt_set_event_name(evlist, id, "cbr");
2959 if (pt->synth_opts.pwr_events && (evsel->core.attr.config & 0x10)) {
2960 attr.config = PERF_SYNTH_INTEL_MWAIT;
2961 err = intel_pt_synth_event(session, "mwait", &attr, id);
2965 intel_pt_set_event_name(evlist, id, "mwait");
2968 attr.config = PERF_SYNTH_INTEL_PWRE;
2969 err = intel_pt_synth_event(session, "pwre", &attr, id);
2973 intel_pt_set_event_name(evlist, id, "pwre");
2976 attr.config = PERF_SYNTH_INTEL_EXSTOP;
2977 err = intel_pt_synth_event(session, "exstop", &attr, id);
2981 intel_pt_set_event_name(evlist, id, "exstop");
2984 attr.config = PERF_SYNTH_INTEL_PWRX;
2985 err = intel_pt_synth_event(session, "pwrx", &attr, id);
2989 intel_pt_set_event_name(evlist, id, "pwrx");
2996 static void intel_pt_setup_pebs_events(struct intel_pt *pt)
2998 struct evsel *evsel;
3000 if (!pt->synth_opts.other_events)
3003 evlist__for_each_entry(pt->session->evlist, evsel) {
3004 if (evsel->core.attr.aux_output && evsel->core.id) {
3005 pt->sample_pebs = true;
3006 pt->pebs_evsel = evsel;
3012 static struct evsel *intel_pt_find_sched_switch(struct evlist *evlist)
3014 struct evsel *evsel;
3016 evlist__for_each_entry_reverse(evlist, evsel) {
3017 const char *name = perf_evsel__name(evsel);
3019 if (!strcmp(name, "sched:sched_switch"))
3026 static bool intel_pt_find_switch(struct evlist *evlist)
3028 struct evsel *evsel;
3030 evlist__for_each_entry(evlist, evsel) {
3031 if (evsel->core.attr.context_switch)
3038 static int intel_pt_perf_config(const char *var, const char *value, void *data)
3040 struct intel_pt *pt = data;
3042 if (!strcmp(var, "intel-pt.mispred-all"))
3043 pt->mispred_all = perf_config_bool(var, value);
3048 /* Find least TSC which converts to ns or later */
3049 static u64 intel_pt_tsc_start(u64 ns, struct intel_pt *pt)
3053 tsc = perf_time_to_tsc(ns, &pt->tc);
3056 tm = tsc_to_perf_time(tsc, &pt->tc);
3063 tm = tsc_to_perf_time(++tsc, &pt->tc);
3068 /* Find greatest TSC which converts to ns or earlier */
3069 static u64 intel_pt_tsc_end(u64 ns, struct intel_pt *pt)
3073 tsc = perf_time_to_tsc(ns, &pt->tc);
3076 tm = tsc_to_perf_time(tsc, &pt->tc);
3083 tm = tsc_to_perf_time(--tsc, &pt->tc);
3088 static int intel_pt_setup_time_ranges(struct intel_pt *pt,
3089 struct itrace_synth_opts *opts)
3091 struct perf_time_interval *p = opts->ptime_range;
3092 int n = opts->range_num;
3095 if (!n || !p || pt->timeless_decoding)
3098 pt->time_ranges = calloc(n, sizeof(struct range));
3099 if (!pt->time_ranges)
3104 intel_pt_log("%s: %u range(s)\n", __func__, n);
3106 for (i = 0; i < n; i++) {
3107 struct range *r = &pt->time_ranges[i];
3108 u64 ts = p[i].start;
3112 * Take care to ensure the TSC range matches the perf-time range
3113 * when converted back to perf-time.
3115 r->start = ts ? intel_pt_tsc_start(ts, pt) : 0;
3116 r->end = te ? intel_pt_tsc_end(te, pt) : 0;
3118 intel_pt_log("range %d: perf time interval: %"PRIu64" to %"PRIu64"\n",
3120 intel_pt_log("range %d: TSC time interval: %#"PRIx64" to %#"PRIx64"\n",
3121 i, r->start, r->end);
3127 static const char * const intel_pt_info_fmts[] = {
3128 [INTEL_PT_PMU_TYPE] = " PMU Type %"PRId64"\n",
3129 [INTEL_PT_TIME_SHIFT] = " Time Shift %"PRIu64"\n",
3130 [INTEL_PT_TIME_MULT] = " Time Muliplier %"PRIu64"\n",
3131 [INTEL_PT_TIME_ZERO] = " Time Zero %"PRIu64"\n",
3132 [INTEL_PT_CAP_USER_TIME_ZERO] = " Cap Time Zero %"PRId64"\n",
3133 [INTEL_PT_TSC_BIT] = " TSC bit %#"PRIx64"\n",
3134 [INTEL_PT_NORETCOMP_BIT] = " NoRETComp bit %#"PRIx64"\n",
3135 [INTEL_PT_HAVE_SCHED_SWITCH] = " Have sched_switch %"PRId64"\n",
3136 [INTEL_PT_SNAPSHOT_MODE] = " Snapshot mode %"PRId64"\n",
3137 [INTEL_PT_PER_CPU_MMAPS] = " Per-cpu maps %"PRId64"\n",
3138 [INTEL_PT_MTC_BIT] = " MTC bit %#"PRIx64"\n",
3139 [INTEL_PT_TSC_CTC_N] = " TSC:CTC numerator %"PRIu64"\n",
3140 [INTEL_PT_TSC_CTC_D] = " TSC:CTC denominator %"PRIu64"\n",
3141 [INTEL_PT_CYC_BIT] = " CYC bit %#"PRIx64"\n",
3142 [INTEL_PT_MAX_NONTURBO_RATIO] = " Max non-turbo ratio %"PRIu64"\n",
3143 [INTEL_PT_FILTER_STR_LEN] = " Filter string len. %"PRIu64"\n",
3146 static void intel_pt_print_info(__u64 *arr, int start, int finish)
3153 for (i = start; i <= finish; i++)
3154 fprintf(stdout, intel_pt_info_fmts[i], arr[i]);
3157 static void intel_pt_print_info_str(const char *name, const char *str)
3162 fprintf(stdout, " %-20s%s\n", name, str ? str : "");
3165 static bool intel_pt_has(struct perf_record_auxtrace_info *auxtrace_info, int pos)
3167 return auxtrace_info->header.size >=
3168 sizeof(struct perf_record_auxtrace_info) + (sizeof(u64) * (pos + 1));
3171 int intel_pt_process_auxtrace_info(union perf_event *event,
3172 struct perf_session *session)
3174 struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info;
3175 size_t min_sz = sizeof(u64) * INTEL_PT_PER_CPU_MMAPS;
3176 struct intel_pt *pt;
3181 if (auxtrace_info->header.size < sizeof(struct perf_record_auxtrace_info) +
3185 pt = zalloc(sizeof(struct intel_pt));
3189 addr_filters__init(&pt->filts);
3191 err = perf_config(intel_pt_perf_config, pt);
3195 err = auxtrace_queues__init(&pt->queues);
3199 intel_pt_log_set_name(INTEL_PT_PMU_NAME);
3201 pt->session = session;
3202 pt->machine = &session->machines.host; /* No kvm support */
3203 pt->auxtrace_type = auxtrace_info->type;
3204 pt->pmu_type = auxtrace_info->priv[INTEL_PT_PMU_TYPE];
3205 pt->tc.time_shift = auxtrace_info->priv[INTEL_PT_TIME_SHIFT];
3206 pt->tc.time_mult = auxtrace_info->priv[INTEL_PT_TIME_MULT];
3207 pt->tc.time_zero = auxtrace_info->priv[INTEL_PT_TIME_ZERO];
3208 pt->cap_user_time_zero = auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO];
3209 pt->tsc_bit = auxtrace_info->priv[INTEL_PT_TSC_BIT];
3210 pt->noretcomp_bit = auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT];
3211 pt->have_sched_switch = auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH];
3212 pt->snapshot_mode = auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE];
3213 pt->per_cpu_mmaps = auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS];
3214 intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_PMU_TYPE,
3215 INTEL_PT_PER_CPU_MMAPS);
3217 if (intel_pt_has(auxtrace_info, INTEL_PT_CYC_BIT)) {
3218 pt->mtc_bit = auxtrace_info->priv[INTEL_PT_MTC_BIT];
3219 pt->mtc_freq_bits = auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS];
3220 pt->tsc_ctc_ratio_n = auxtrace_info->priv[INTEL_PT_TSC_CTC_N];
3221 pt->tsc_ctc_ratio_d = auxtrace_info->priv[INTEL_PT_TSC_CTC_D];
3222 pt->cyc_bit = auxtrace_info->priv[INTEL_PT_CYC_BIT];
3223 intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_MTC_BIT,
3227 if (intel_pt_has(auxtrace_info, INTEL_PT_MAX_NONTURBO_RATIO)) {
3228 pt->max_non_turbo_ratio =
3229 auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO];
3230 intel_pt_print_info(&auxtrace_info->priv[0],
3231 INTEL_PT_MAX_NONTURBO_RATIO,
3232 INTEL_PT_MAX_NONTURBO_RATIO);
3235 info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
3236 info_end = (void *)info + auxtrace_info->header.size;
3238 if (intel_pt_has(auxtrace_info, INTEL_PT_FILTER_STR_LEN)) {
3241 len = auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN];
3242 intel_pt_print_info(&auxtrace_info->priv[0],
3243 INTEL_PT_FILTER_STR_LEN,
3244 INTEL_PT_FILTER_STR_LEN);
3246 const char *filter = (const char *)info;
3248 len = roundup(len + 1, 8);
3250 if ((void *)info > info_end) {
3251 pr_err("%s: bad filter string length\n", __func__);
3253 goto err_free_queues;
3255 pt->filter = memdup(filter, len);
3258 goto err_free_queues;
3260 if (session->header.needs_swap)
3261 mem_bswap_64(pt->filter, len);
3262 if (pt->filter[len - 1]) {
3263 pr_err("%s: filter string not null terminated\n", __func__);
3265 goto err_free_queues;
3267 err = addr_filters__parse_bare_filter(&pt->filts,
3270 goto err_free_queues;
3272 intel_pt_print_info_str("Filter string", pt->filter);
3275 pt->timeless_decoding = intel_pt_timeless_decoding(pt);
3276 if (pt->timeless_decoding && !pt->tc.time_mult)
3277 pt->tc.time_mult = 1;
3278 pt->have_tsc = intel_pt_have_tsc(pt);
3279 pt->sampling_mode = intel_pt_sampling_mode(pt);
3280 pt->est_tsc = !pt->timeless_decoding;
3282 pt->unknown_thread = thread__new(999999999, 999999999);
3283 if (!pt->unknown_thread) {
3285 goto err_free_queues;
3289 * Since this thread will not be kept in any rbtree not in a
3290 * list, initialize its list node so that at thread__put() the
3291 * current thread lifetime assuption is kept and we don't segfault
3292 * at list_del_init().
3294 INIT_LIST_HEAD(&pt->unknown_thread->node);
3296 err = thread__set_comm(pt->unknown_thread, "unknown", 0);
3298 goto err_delete_thread;
3299 if (thread__init_maps(pt->unknown_thread, pt->machine)) {
3301 goto err_delete_thread;
3304 pt->auxtrace.process_event = intel_pt_process_event;
3305 pt->auxtrace.process_auxtrace_event = intel_pt_process_auxtrace_event;
3306 pt->auxtrace.queue_data = intel_pt_queue_data;
3307 pt->auxtrace.dump_auxtrace_sample = intel_pt_dump_sample;
3308 pt->auxtrace.flush_events = intel_pt_flush;
3309 pt->auxtrace.free_events = intel_pt_free_events;
3310 pt->auxtrace.free = intel_pt_free;
3311 session->auxtrace = &pt->auxtrace;
3316 if (pt->have_sched_switch == 1) {
3317 pt->switch_evsel = intel_pt_find_sched_switch(session->evlist);
3318 if (!pt->switch_evsel) {
3319 pr_err("%s: missing sched_switch event\n", __func__);
3321 goto err_delete_thread;
3323 } else if (pt->have_sched_switch == 2 &&
3324 !intel_pt_find_switch(session->evlist)) {
3325 pr_err("%s: missing context_switch attribute flag\n", __func__);
3327 goto err_delete_thread;
3330 if (session->itrace_synth_opts->set) {
3331 pt->synth_opts = *session->itrace_synth_opts;
3333 itrace_synth_opts__set_default(&pt->synth_opts,
3334 session->itrace_synth_opts->default_no_sample);
3335 if (!session->itrace_synth_opts->default_no_sample &&
3336 !session->itrace_synth_opts->inject) {
3337 pt->synth_opts.branches = false;
3338 pt->synth_opts.callchain = true;
3340 pt->synth_opts.thread_stack =
3341 session->itrace_synth_opts->thread_stack;
3344 if (pt->synth_opts.log)
3345 intel_pt_log_enable();
3347 /* Maximum non-turbo ratio is TSC freq / 100 MHz */
3348 if (pt->tc.time_mult) {
3349 u64 tsc_freq = intel_pt_ns_to_ticks(pt, 1000000000);
3351 if (!pt->max_non_turbo_ratio)
3352 pt->max_non_turbo_ratio =
3353 (tsc_freq + 50000000) / 100000000;
3354 intel_pt_log("TSC frequency %"PRIu64"\n", tsc_freq);
3355 intel_pt_log("Maximum non-turbo ratio %u\n",
3356 pt->max_non_turbo_ratio);
3357 pt->cbr2khz = tsc_freq / pt->max_non_turbo_ratio / 1000;
3360 err = intel_pt_setup_time_ranges(pt, session->itrace_synth_opts);
3362 goto err_delete_thread;
3364 if (pt->synth_opts.calls)
3365 pt->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
3366 PERF_IP_FLAG_TRACE_END;
3367 if (pt->synth_opts.returns)
3368 pt->branches_filter |= PERF_IP_FLAG_RETURN |
3369 PERF_IP_FLAG_TRACE_BEGIN;
3371 if (pt->synth_opts.callchain && !symbol_conf.use_callchain) {
3372 symbol_conf.use_callchain = true;
3373 if (callchain_register_param(&callchain_param) < 0) {
3374 symbol_conf.use_callchain = false;
3375 pt->synth_opts.callchain = false;
3379 err = intel_pt_synth_events(pt, session);
3381 goto err_delete_thread;
3383 intel_pt_setup_pebs_events(pt);
3385 if (pt->sampling_mode || list_empty(&session->auxtrace_index))
3386 err = auxtrace_queue_data(session, true, true);
3388 err = auxtrace_queues__process_index(&pt->queues, session);
3390 goto err_delete_thread;
3392 if (pt->queues.populated)
3393 pt->data_queued = true;
3395 if (pt->timeless_decoding)
3396 pr_debug2("Intel PT decoding without timestamps\n");
3401 thread__zput(pt->unknown_thread);
3403 intel_pt_log_disable();
3404 auxtrace_queues__free(&pt->queues);
3405 session->auxtrace = NULL;
3407 addr_filters__exit(&pt->filts);
3409 zfree(&pt->time_ranges);