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/types.h>
27 #include "thread-stack.h"
29 #include "callchain.h"
37 #include "intel-pt-decoder/intel-pt-log.h"
38 #include "intel-pt-decoder/intel-pt-decoder.h"
39 #include "intel-pt-decoder/intel-pt-insn-decoder.h"
40 #include "intel-pt-decoder/intel-pt-pkt-decoder.h"
42 #define MAX_TIMESTAMP (~0ULL)
45 struct auxtrace auxtrace;
46 struct auxtrace_queues queues;
47 struct auxtrace_heap heap;
49 struct perf_session *session;
50 struct machine *machine;
51 struct perf_evsel *switch_evsel;
52 struct thread *unknown_thread;
53 bool timeless_decoding;
62 int have_sched_switch;
68 struct perf_tsc_conversion tc;
69 bool cap_user_time_zero;
71 struct itrace_synth_opts synth_opts;
73 bool sample_instructions;
74 u64 instructions_sample_type;
79 u64 branches_sample_type;
82 bool sample_transactions;
83 u64 transactions_sample_type;
87 u64 ptwrites_sample_type;
90 bool sample_pwr_events;
91 u64 pwr_events_sample_type;
105 unsigned max_non_turbo_ratio;
108 unsigned long num_events;
111 struct addr_filters filts;
115 INTEL_PT_SS_NOT_TRACING,
118 INTEL_PT_SS_EXPECTING_SWITCH_EVENT,
119 INTEL_PT_SS_EXPECTING_SWITCH_IP,
122 struct intel_pt_queue {
124 unsigned int queue_nr;
125 struct auxtrace_buffer *buffer;
126 struct auxtrace_buffer *old_buffer;
128 const struct intel_pt_state *state;
129 struct ip_callchain *chain;
130 struct branch_stack *last_branch;
131 struct branch_stack *last_branch_rb;
132 size_t last_branch_pos;
133 union perf_event *event_buf;
136 bool step_through_buffers;
137 bool use_buffer_pid_tid;
143 struct thread *thread;
151 char insn[INTEL_PT_INSN_BUF_SZ];
154 static void intel_pt_dump(struct intel_pt *pt __maybe_unused,
155 unsigned char *buf, size_t len)
157 struct intel_pt_pkt packet;
160 char desc[INTEL_PT_PKT_DESC_MAX];
161 const char *color = PERF_COLOR_BLUE;
163 color_fprintf(stdout, color,
164 ". ... Intel Processor Trace data: size %zu bytes\n",
168 ret = intel_pt_get_packet(buf, len, &packet);
174 color_fprintf(stdout, color, " %08x: ", pos);
175 for (i = 0; i < pkt_len; i++)
176 color_fprintf(stdout, color, " %02x", buf[i]);
178 color_fprintf(stdout, color, " ");
180 ret = intel_pt_pkt_desc(&packet, desc,
181 INTEL_PT_PKT_DESC_MAX);
183 color_fprintf(stdout, color, " %s\n", desc);
185 color_fprintf(stdout, color, " Bad packet!\n");
193 static void intel_pt_dump_event(struct intel_pt *pt, unsigned char *buf,
197 intel_pt_dump(pt, buf, len);
200 static void intel_pt_log_event(union perf_event *event)
202 FILE *f = intel_pt_log_fp();
204 if (!intel_pt_enable_logging || !f)
207 perf_event__fprintf(event, f);
210 static int intel_pt_do_fix_overlap(struct intel_pt *pt, struct auxtrace_buffer *a,
211 struct auxtrace_buffer *b)
213 bool consecutive = false;
216 start = intel_pt_find_overlap(a->data, a->size, b->data, b->size,
217 pt->have_tsc, &consecutive);
220 b->use_size = b->data + b->size - start;
222 if (b->use_size && consecutive)
223 b->consecutive = true;
227 /* This function assumes data is processed sequentially only */
228 static int intel_pt_get_trace(struct intel_pt_buffer *b, void *data)
230 struct intel_pt_queue *ptq = data;
231 struct auxtrace_buffer *buffer = ptq->buffer;
232 struct auxtrace_buffer *old_buffer = ptq->old_buffer;
233 struct auxtrace_queue *queue;
241 queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
243 buffer = auxtrace_buffer__next(queue, buffer);
246 auxtrace_buffer__drop_data(old_buffer);
251 ptq->buffer = buffer;
254 int fd = perf_data__fd(ptq->pt->session->data);
256 buffer->data = auxtrace_buffer__get_data(buffer, fd);
261 might_overlap = ptq->pt->snapshot_mode || ptq->pt->sampling_mode;
262 if (might_overlap && !buffer->consecutive && old_buffer &&
263 intel_pt_do_fix_overlap(ptq->pt, old_buffer, buffer))
266 if (buffer->use_data) {
267 b->len = buffer->use_size;
268 b->buf = buffer->use_data;
270 b->len = buffer->size;
271 b->buf = buffer->data;
273 b->ref_timestamp = buffer->reference;
275 if (!old_buffer || (might_overlap && !buffer->consecutive)) {
276 b->consecutive = false;
277 b->trace_nr = buffer->buffer_nr + 1;
279 b->consecutive = true;
282 if (ptq->step_through_buffers)
287 auxtrace_buffer__drop_data(old_buffer);
288 ptq->old_buffer = buffer;
290 auxtrace_buffer__drop_data(buffer);
291 return intel_pt_get_trace(b, data);
297 struct intel_pt_cache_entry {
298 struct auxtrace_cache_entry entry;
301 enum intel_pt_insn_op op;
302 enum intel_pt_insn_branch branch;
305 char insn[INTEL_PT_INSN_BUF_SZ];
308 static int intel_pt_config_div(const char *var, const char *value, void *data)
313 if (!strcmp(var, "intel-pt.cache-divisor")) {
314 val = strtol(value, NULL, 0);
315 if (val > 0 && val <= INT_MAX)
322 static int intel_pt_cache_divisor(void)
329 perf_config(intel_pt_config_div, &d);
337 static unsigned int intel_pt_cache_size(struct dso *dso,
338 struct machine *machine)
342 size = dso__data_size(dso, machine);
343 size /= intel_pt_cache_divisor();
346 if (size > (1 << 21))
348 return 32 - __builtin_clz(size);
351 static struct auxtrace_cache *intel_pt_cache(struct dso *dso,
352 struct machine *machine)
354 struct auxtrace_cache *c;
357 if (dso->auxtrace_cache)
358 return dso->auxtrace_cache;
360 bits = intel_pt_cache_size(dso, machine);
362 /* Ignoring cache creation failure */
363 c = auxtrace_cache__new(bits, sizeof(struct intel_pt_cache_entry), 200);
365 dso->auxtrace_cache = c;
370 static int intel_pt_cache_add(struct dso *dso, struct machine *machine,
371 u64 offset, u64 insn_cnt, u64 byte_cnt,
372 struct intel_pt_insn *intel_pt_insn)
374 struct auxtrace_cache *c = intel_pt_cache(dso, machine);
375 struct intel_pt_cache_entry *e;
381 e = auxtrace_cache__alloc_entry(c);
385 e->insn_cnt = insn_cnt;
386 e->byte_cnt = byte_cnt;
387 e->op = intel_pt_insn->op;
388 e->branch = intel_pt_insn->branch;
389 e->length = intel_pt_insn->length;
390 e->rel = intel_pt_insn->rel;
391 memcpy(e->insn, intel_pt_insn->buf, INTEL_PT_INSN_BUF_SZ);
393 err = auxtrace_cache__add(c, offset, &e->entry);
395 auxtrace_cache__free_entry(c, e);
400 static struct intel_pt_cache_entry *
401 intel_pt_cache_lookup(struct dso *dso, struct machine *machine, u64 offset)
403 struct auxtrace_cache *c = intel_pt_cache(dso, machine);
408 return auxtrace_cache__lookup(dso->auxtrace_cache, offset);
411 static inline u8 intel_pt_cpumode(struct intel_pt *pt, uint64_t ip)
413 return ip >= pt->kernel_start ?
414 PERF_RECORD_MISC_KERNEL :
415 PERF_RECORD_MISC_USER;
418 static int intel_pt_walk_next_insn(struct intel_pt_insn *intel_pt_insn,
419 uint64_t *insn_cnt_ptr, uint64_t *ip,
420 uint64_t to_ip, uint64_t max_insn_cnt,
423 struct intel_pt_queue *ptq = data;
424 struct machine *machine = ptq->pt->machine;
425 struct thread *thread;
426 struct addr_location al;
427 unsigned char buf[INTEL_PT_INSN_BUF_SZ];
431 u64 offset, start_offset, start_ip;
435 intel_pt_insn->length = 0;
437 if (to_ip && *ip == to_ip)
440 cpumode = intel_pt_cpumode(ptq->pt, *ip);
442 thread = ptq->thread;
444 if (cpumode != PERF_RECORD_MISC_KERNEL)
446 thread = ptq->pt->unknown_thread;
450 if (!thread__find_map(thread, cpumode, *ip, &al) || !al.map->dso)
453 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
454 dso__data_status_seen(al.map->dso,
455 DSO_DATA_STATUS_SEEN_ITRACE))
458 offset = al.map->map_ip(al.map, *ip);
460 if (!to_ip && one_map) {
461 struct intel_pt_cache_entry *e;
463 e = intel_pt_cache_lookup(al.map->dso, machine, offset);
465 (!max_insn_cnt || e->insn_cnt <= max_insn_cnt)) {
466 *insn_cnt_ptr = e->insn_cnt;
468 intel_pt_insn->op = e->op;
469 intel_pt_insn->branch = e->branch;
470 intel_pt_insn->length = e->length;
471 intel_pt_insn->rel = e->rel;
472 memcpy(intel_pt_insn->buf, e->insn,
473 INTEL_PT_INSN_BUF_SZ);
474 intel_pt_log_insn_no_data(intel_pt_insn, *ip);
479 start_offset = offset;
482 /* Load maps to ensure dso->is_64_bit has been updated */
485 x86_64 = al.map->dso->is_64_bit;
488 len = dso__data_read_offset(al.map->dso, machine,
490 INTEL_PT_INSN_BUF_SZ);
494 if (intel_pt_get_insn(buf, len, x86_64, intel_pt_insn))
497 intel_pt_log_insn(intel_pt_insn, *ip);
501 if (intel_pt_insn->branch != INTEL_PT_BR_NO_BRANCH)
504 if (max_insn_cnt && insn_cnt >= max_insn_cnt)
507 *ip += intel_pt_insn->length;
509 if (to_ip && *ip == to_ip)
512 if (*ip >= al.map->end)
515 offset += intel_pt_insn->length;
520 *insn_cnt_ptr = insn_cnt;
526 * Didn't lookup in the 'to_ip' case, so do it now to prevent duplicate
530 struct intel_pt_cache_entry *e;
532 e = intel_pt_cache_lookup(al.map->dso, machine, start_offset);
537 /* Ignore cache errors */
538 intel_pt_cache_add(al.map->dso, machine, start_offset, insn_cnt,
539 *ip - start_ip, intel_pt_insn);
544 *insn_cnt_ptr = insn_cnt;
548 static bool intel_pt_match_pgd_ip(struct intel_pt *pt, uint64_t ip,
549 uint64_t offset, const char *filename)
551 struct addr_filter *filt;
552 bool have_filter = false;
553 bool hit_tracestop = false;
554 bool hit_filter = false;
556 list_for_each_entry(filt, &pt->filts.head, list) {
560 if ((filename && !filt->filename) ||
561 (!filename && filt->filename) ||
562 (filename && strcmp(filename, filt->filename)))
565 if (!(offset >= filt->addr && offset < filt->addr + filt->size))
568 intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s hit filter: %s offset %#"PRIx64" size %#"PRIx64"\n",
569 ip, offset, filename ? filename : "[kernel]",
570 filt->start ? "filter" : "stop",
571 filt->addr, filt->size);
576 hit_tracestop = true;
579 if (!hit_tracestop && !hit_filter)
580 intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s is not in a filter region\n",
581 ip, offset, filename ? filename : "[kernel]");
583 return hit_tracestop || (have_filter && !hit_filter);
586 static int __intel_pt_pgd_ip(uint64_t ip, void *data)
588 struct intel_pt_queue *ptq = data;
589 struct thread *thread;
590 struct addr_location al;
594 if (ip >= ptq->pt->kernel_start)
595 return intel_pt_match_pgd_ip(ptq->pt, ip, ip, NULL);
597 cpumode = PERF_RECORD_MISC_USER;
599 thread = ptq->thread;
603 if (!thread__find_map(thread, cpumode, ip, &al) || !al.map->dso)
606 offset = al.map->map_ip(al.map, ip);
608 return intel_pt_match_pgd_ip(ptq->pt, ip, offset,
609 al.map->dso->long_name);
612 static bool intel_pt_pgd_ip(uint64_t ip, void *data)
614 return __intel_pt_pgd_ip(ip, data) > 0;
617 static bool intel_pt_get_config(struct intel_pt *pt,
618 struct perf_event_attr *attr, u64 *config)
620 if (attr->type == pt->pmu_type) {
622 *config = attr->config;
629 static bool intel_pt_exclude_kernel(struct intel_pt *pt)
631 struct perf_evsel *evsel;
633 evlist__for_each_entry(pt->session->evlist, evsel) {
634 if (intel_pt_get_config(pt, &evsel->attr, NULL) &&
635 !evsel->attr.exclude_kernel)
641 static bool intel_pt_return_compression(struct intel_pt *pt)
643 struct perf_evsel *evsel;
646 if (!pt->noretcomp_bit)
649 evlist__for_each_entry(pt->session->evlist, evsel) {
650 if (intel_pt_get_config(pt, &evsel->attr, &config) &&
651 (config & pt->noretcomp_bit))
657 static bool intel_pt_branch_enable(struct intel_pt *pt)
659 struct perf_evsel *evsel;
662 evlist__for_each_entry(pt->session->evlist, evsel) {
663 if (intel_pt_get_config(pt, &evsel->attr, &config) &&
664 (config & 1) && !(config & 0x2000))
670 static unsigned int intel_pt_mtc_period(struct intel_pt *pt)
672 struct perf_evsel *evsel;
676 if (!pt->mtc_freq_bits)
679 for (shift = 0, config = pt->mtc_freq_bits; !(config & 1); shift++)
682 evlist__for_each_entry(pt->session->evlist, evsel) {
683 if (intel_pt_get_config(pt, &evsel->attr, &config))
684 return (config & pt->mtc_freq_bits) >> shift;
689 static bool intel_pt_timeless_decoding(struct intel_pt *pt)
691 struct perf_evsel *evsel;
692 bool timeless_decoding = true;
695 if (!pt->tsc_bit || !pt->cap_user_time_zero)
698 evlist__for_each_entry(pt->session->evlist, evsel) {
699 if (!(evsel->attr.sample_type & PERF_SAMPLE_TIME))
701 if (intel_pt_get_config(pt, &evsel->attr, &config)) {
702 if (config & pt->tsc_bit)
703 timeless_decoding = false;
708 return timeless_decoding;
711 static bool intel_pt_tracing_kernel(struct intel_pt *pt)
713 struct perf_evsel *evsel;
715 evlist__for_each_entry(pt->session->evlist, evsel) {
716 if (intel_pt_get_config(pt, &evsel->attr, NULL) &&
717 !evsel->attr.exclude_kernel)
723 static bool intel_pt_have_tsc(struct intel_pt *pt)
725 struct perf_evsel *evsel;
726 bool have_tsc = false;
732 evlist__for_each_entry(pt->session->evlist, evsel) {
733 if (intel_pt_get_config(pt, &evsel->attr, &config)) {
734 if (config & pt->tsc_bit)
743 static u64 intel_pt_ns_to_ticks(const struct intel_pt *pt, u64 ns)
747 quot = ns / pt->tc.time_mult;
748 rem = ns % pt->tc.time_mult;
749 return (quot << pt->tc.time_shift) + (rem << pt->tc.time_shift) /
753 static struct intel_pt_queue *intel_pt_alloc_queue(struct intel_pt *pt,
754 unsigned int queue_nr)
756 struct intel_pt_params params = { .get_trace = 0, };
757 struct perf_env *env = pt->machine->env;
758 struct intel_pt_queue *ptq;
760 ptq = zalloc(sizeof(struct intel_pt_queue));
764 if (pt->synth_opts.callchain) {
765 size_t sz = sizeof(struct ip_callchain);
767 /* Add 1 to callchain_sz for callchain context */
768 sz += (pt->synth_opts.callchain_sz + 1) * sizeof(u64);
769 ptq->chain = zalloc(sz);
774 if (pt->synth_opts.last_branch) {
775 size_t sz = sizeof(struct branch_stack);
777 sz += pt->synth_opts.last_branch_sz *
778 sizeof(struct branch_entry);
779 ptq->last_branch = zalloc(sz);
780 if (!ptq->last_branch)
782 ptq->last_branch_rb = zalloc(sz);
783 if (!ptq->last_branch_rb)
787 ptq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
792 ptq->queue_nr = queue_nr;
793 ptq->exclude_kernel = intel_pt_exclude_kernel(pt);
799 params.get_trace = intel_pt_get_trace;
800 params.walk_insn = intel_pt_walk_next_insn;
802 params.return_compression = intel_pt_return_compression(pt);
803 params.branch_enable = intel_pt_branch_enable(pt);
804 params.max_non_turbo_ratio = pt->max_non_turbo_ratio;
805 params.mtc_period = intel_pt_mtc_period(pt);
806 params.tsc_ctc_ratio_n = pt->tsc_ctc_ratio_n;
807 params.tsc_ctc_ratio_d = pt->tsc_ctc_ratio_d;
809 if (pt->filts.cnt > 0)
810 params.pgd_ip = intel_pt_pgd_ip;
812 if (pt->synth_opts.instructions) {
813 if (pt->synth_opts.period) {
814 switch (pt->synth_opts.period_type) {
815 case PERF_ITRACE_PERIOD_INSTRUCTIONS:
817 INTEL_PT_PERIOD_INSTRUCTIONS;
818 params.period = pt->synth_opts.period;
820 case PERF_ITRACE_PERIOD_TICKS:
821 params.period_type = INTEL_PT_PERIOD_TICKS;
822 params.period = pt->synth_opts.period;
824 case PERF_ITRACE_PERIOD_NANOSECS:
825 params.period_type = INTEL_PT_PERIOD_TICKS;
826 params.period = intel_pt_ns_to_ticks(pt,
827 pt->synth_opts.period);
834 if (!params.period) {
835 params.period_type = INTEL_PT_PERIOD_INSTRUCTIONS;
840 if (env->cpuid && !strncmp(env->cpuid, "GenuineIntel,6,92,", 18))
841 params.flags |= INTEL_PT_FUP_WITH_NLIP;
843 ptq->decoder = intel_pt_decoder_new(¶ms);
850 zfree(&ptq->event_buf);
851 zfree(&ptq->last_branch);
852 zfree(&ptq->last_branch_rb);
858 static void intel_pt_free_queue(void *priv)
860 struct intel_pt_queue *ptq = priv;
864 thread__zput(ptq->thread);
865 intel_pt_decoder_free(ptq->decoder);
866 zfree(&ptq->event_buf);
867 zfree(&ptq->last_branch);
868 zfree(&ptq->last_branch_rb);
873 static void intel_pt_set_pid_tid_cpu(struct intel_pt *pt,
874 struct auxtrace_queue *queue)
876 struct intel_pt_queue *ptq = queue->priv;
878 if (queue->tid == -1 || pt->have_sched_switch) {
879 ptq->tid = machine__get_current_tid(pt->machine, ptq->cpu);
880 thread__zput(ptq->thread);
883 if (!ptq->thread && ptq->tid != -1)
884 ptq->thread = machine__find_thread(pt->machine, -1, ptq->tid);
887 ptq->pid = ptq->thread->pid_;
888 if (queue->cpu == -1)
889 ptq->cpu = ptq->thread->cpu;
893 static void intel_pt_sample_flags(struct intel_pt_queue *ptq)
895 if (ptq->state->flags & INTEL_PT_ABORT_TX) {
896 ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_TX_ABORT;
897 } else if (ptq->state->flags & INTEL_PT_ASYNC) {
898 if (ptq->state->to_ip)
899 ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_CALL |
901 PERF_IP_FLAG_INTERRUPT;
903 ptq->flags = PERF_IP_FLAG_BRANCH |
904 PERF_IP_FLAG_TRACE_END;
907 if (ptq->state->from_ip)
908 ptq->flags = intel_pt_insn_type(ptq->state->insn_op);
910 ptq->flags = PERF_IP_FLAG_BRANCH |
911 PERF_IP_FLAG_TRACE_BEGIN;
912 if (ptq->state->flags & INTEL_PT_IN_TX)
913 ptq->flags |= PERF_IP_FLAG_IN_TX;
914 ptq->insn_len = ptq->state->insn_len;
915 memcpy(ptq->insn, ptq->state->insn, INTEL_PT_INSN_BUF_SZ);
918 if (ptq->state->type & INTEL_PT_TRACE_BEGIN)
919 ptq->flags |= PERF_IP_FLAG_TRACE_BEGIN;
920 if (ptq->state->type & INTEL_PT_TRACE_END)
921 ptq->flags |= PERF_IP_FLAG_TRACE_END;
924 static int intel_pt_setup_queue(struct intel_pt *pt,
925 struct auxtrace_queue *queue,
926 unsigned int queue_nr)
928 struct intel_pt_queue *ptq = queue->priv;
930 if (list_empty(&queue->head))
934 ptq = intel_pt_alloc_queue(pt, queue_nr);
939 if (queue->cpu != -1)
940 ptq->cpu = queue->cpu;
941 ptq->tid = queue->tid;
943 if (pt->sampling_mode && !pt->snapshot_mode &&
944 pt->timeless_decoding)
945 ptq->step_through_buffers = true;
947 ptq->sync_switch = pt->sync_switch;
951 (!ptq->sync_switch ||
952 ptq->switch_state != INTEL_PT_SS_EXPECTING_SWITCH_EVENT)) {
953 const struct intel_pt_state *state;
956 if (pt->timeless_decoding)
959 intel_pt_log("queue %u getting timestamp\n", queue_nr);
960 intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
961 queue_nr, ptq->cpu, ptq->pid, ptq->tid);
963 state = intel_pt_decode(ptq->decoder);
965 if (state->err == INTEL_PT_ERR_NODATA) {
966 intel_pt_log("queue %u has no timestamp\n",
972 if (state->timestamp)
976 ptq->timestamp = state->timestamp;
977 intel_pt_log("queue %u timestamp 0x%" PRIx64 "\n",
978 queue_nr, ptq->timestamp);
980 ptq->have_sample = true;
981 intel_pt_sample_flags(ptq);
982 ret = auxtrace_heap__add(&pt->heap, queue_nr, ptq->timestamp);
991 static int intel_pt_setup_queues(struct intel_pt *pt)
996 for (i = 0; i < pt->queues.nr_queues; i++) {
997 ret = intel_pt_setup_queue(pt, &pt->queues.queue_array[i], i);
1004 static inline void intel_pt_copy_last_branch_rb(struct intel_pt_queue *ptq)
1006 struct branch_stack *bs_src = ptq->last_branch_rb;
1007 struct branch_stack *bs_dst = ptq->last_branch;
1010 bs_dst->nr = bs_src->nr;
1015 nr = ptq->pt->synth_opts.last_branch_sz - ptq->last_branch_pos;
1016 memcpy(&bs_dst->entries[0],
1017 &bs_src->entries[ptq->last_branch_pos],
1018 sizeof(struct branch_entry) * nr);
1020 if (bs_src->nr >= ptq->pt->synth_opts.last_branch_sz) {
1021 memcpy(&bs_dst->entries[nr],
1022 &bs_src->entries[0],
1023 sizeof(struct branch_entry) * ptq->last_branch_pos);
1027 static inline void intel_pt_reset_last_branch_rb(struct intel_pt_queue *ptq)
1029 ptq->last_branch_pos = 0;
1030 ptq->last_branch_rb->nr = 0;
1033 static void intel_pt_update_last_branch_rb(struct intel_pt_queue *ptq)
1035 const struct intel_pt_state *state = ptq->state;
1036 struct branch_stack *bs = ptq->last_branch_rb;
1037 struct branch_entry *be;
1039 if (!ptq->last_branch_pos)
1040 ptq->last_branch_pos = ptq->pt->synth_opts.last_branch_sz;
1042 ptq->last_branch_pos -= 1;
1044 be = &bs->entries[ptq->last_branch_pos];
1045 be->from = state->from_ip;
1046 be->to = state->to_ip;
1047 be->flags.abort = !!(state->flags & INTEL_PT_ABORT_TX);
1048 be->flags.in_tx = !!(state->flags & INTEL_PT_IN_TX);
1049 /* No support for mispredict */
1050 be->flags.mispred = ptq->pt->mispred_all;
1052 if (bs->nr < ptq->pt->synth_opts.last_branch_sz)
1056 static inline bool intel_pt_skip_event(struct intel_pt *pt)
1058 return pt->synth_opts.initial_skip &&
1059 pt->num_events++ < pt->synth_opts.initial_skip;
1062 static void intel_pt_prep_b_sample(struct intel_pt *pt,
1063 struct intel_pt_queue *ptq,
1064 union perf_event *event,
1065 struct perf_sample *sample)
1067 if (!pt->timeless_decoding)
1068 sample->time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1070 sample->ip = ptq->state->from_ip;
1071 sample->cpumode = intel_pt_cpumode(pt, sample->ip);
1072 sample->pid = ptq->pid;
1073 sample->tid = ptq->tid;
1074 sample->addr = ptq->state->to_ip;
1076 sample->cpu = ptq->cpu;
1077 sample->flags = ptq->flags;
1078 sample->insn_len = ptq->insn_len;
1079 memcpy(sample->insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1081 event->sample.header.type = PERF_RECORD_SAMPLE;
1082 event->sample.header.misc = sample->cpumode;
1083 event->sample.header.size = sizeof(struct perf_event_header);
1086 static int intel_pt_inject_event(union perf_event *event,
1087 struct perf_sample *sample, u64 type)
1089 event->header.size = perf_event__sample_event_size(sample, type, 0);
1090 return perf_event__synthesize_sample(event, type, 0, sample);
1093 static inline int intel_pt_opt_inject(struct intel_pt *pt,
1094 union perf_event *event,
1095 struct perf_sample *sample, u64 type)
1097 if (!pt->synth_opts.inject)
1100 return intel_pt_inject_event(event, sample, type);
1103 static int intel_pt_deliver_synth_b_event(struct intel_pt *pt,
1104 union perf_event *event,
1105 struct perf_sample *sample, u64 type)
1109 ret = intel_pt_opt_inject(pt, event, sample, type);
1113 ret = perf_session__deliver_synth_event(pt->session, event, sample);
1115 pr_err("Intel PT: failed to deliver event, error %d\n", ret);
1120 static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq)
1122 struct intel_pt *pt = ptq->pt;
1123 union perf_event *event = ptq->event_buf;
1124 struct perf_sample sample = { .ip = 0, };
1125 struct dummy_branch_stack {
1127 struct branch_entry entries;
1130 if (pt->branches_filter && !(pt->branches_filter & ptq->flags))
1133 if (intel_pt_skip_event(pt))
1136 intel_pt_prep_b_sample(pt, ptq, event, &sample);
1138 sample.id = ptq->pt->branches_id;
1139 sample.stream_id = ptq->pt->branches_id;
1142 * perf report cannot handle events without a branch stack when using
1143 * SORT_MODE__BRANCH so make a dummy one.
1145 if (pt->synth_opts.last_branch && sort__mode == SORT_MODE__BRANCH) {
1146 dummy_bs = (struct dummy_branch_stack){
1153 sample.branch_stack = (struct branch_stack *)&dummy_bs;
1156 return intel_pt_deliver_synth_b_event(pt, event, &sample,
1157 pt->branches_sample_type);
1160 static void intel_pt_prep_sample(struct intel_pt *pt,
1161 struct intel_pt_queue *ptq,
1162 union perf_event *event,
1163 struct perf_sample *sample)
1165 intel_pt_prep_b_sample(pt, ptq, event, sample);
1167 if (pt->synth_opts.callchain) {
1168 thread_stack__sample(ptq->thread, ptq->cpu, ptq->chain,
1169 pt->synth_opts.callchain_sz + 1,
1170 sample->ip, pt->kernel_start);
1171 sample->callchain = ptq->chain;
1174 if (pt->synth_opts.last_branch) {
1175 intel_pt_copy_last_branch_rb(ptq);
1176 sample->branch_stack = ptq->last_branch;
1180 static inline int intel_pt_deliver_synth_event(struct intel_pt *pt,
1181 struct intel_pt_queue *ptq,
1182 union perf_event *event,
1183 struct perf_sample *sample,
1188 ret = intel_pt_deliver_synth_b_event(pt, event, sample, type);
1190 if (pt->synth_opts.last_branch)
1191 intel_pt_reset_last_branch_rb(ptq);
1196 static int intel_pt_synth_instruction_sample(struct intel_pt_queue *ptq)
1198 struct intel_pt *pt = ptq->pt;
1199 union perf_event *event = ptq->event_buf;
1200 struct perf_sample sample = { .ip = 0, };
1202 if (intel_pt_skip_event(pt))
1205 intel_pt_prep_sample(pt, ptq, event, &sample);
1207 sample.id = ptq->pt->instructions_id;
1208 sample.stream_id = ptq->pt->instructions_id;
1209 sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt;
1211 ptq->last_insn_cnt = ptq->state->tot_insn_cnt;
1213 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1214 pt->instructions_sample_type);
1217 static int intel_pt_synth_transaction_sample(struct intel_pt_queue *ptq)
1219 struct intel_pt *pt = ptq->pt;
1220 union perf_event *event = ptq->event_buf;
1221 struct perf_sample sample = { .ip = 0, };
1223 if (intel_pt_skip_event(pt))
1226 intel_pt_prep_sample(pt, ptq, event, &sample);
1228 sample.id = ptq->pt->transactions_id;
1229 sample.stream_id = ptq->pt->transactions_id;
1231 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1232 pt->transactions_sample_type);
1235 static void intel_pt_prep_p_sample(struct intel_pt *pt,
1236 struct intel_pt_queue *ptq,
1237 union perf_event *event,
1238 struct perf_sample *sample)
1240 intel_pt_prep_sample(pt, ptq, event, sample);
1243 * Zero IP is used to mean "trace start" but that is not the case for
1244 * power or PTWRITE events with no IP, so clear the flags.
1250 static int intel_pt_synth_ptwrite_sample(struct intel_pt_queue *ptq)
1252 struct intel_pt *pt = ptq->pt;
1253 union perf_event *event = ptq->event_buf;
1254 struct perf_sample sample = { .ip = 0, };
1255 struct perf_synth_intel_ptwrite raw;
1257 if (intel_pt_skip_event(pt))
1260 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1262 sample.id = ptq->pt->ptwrites_id;
1263 sample.stream_id = ptq->pt->ptwrites_id;
1266 raw.ip = !!(ptq->state->flags & INTEL_PT_FUP_IP);
1267 raw.payload = cpu_to_le64(ptq->state->ptw_payload);
1269 sample.raw_size = perf_synth__raw_size(raw);
1270 sample.raw_data = perf_synth__raw_data(&raw);
1272 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1273 pt->ptwrites_sample_type);
1276 static int intel_pt_synth_cbr_sample(struct intel_pt_queue *ptq)
1278 struct intel_pt *pt = ptq->pt;
1279 union perf_event *event = ptq->event_buf;
1280 struct perf_sample sample = { .ip = 0, };
1281 struct perf_synth_intel_cbr raw;
1284 if (intel_pt_skip_event(pt))
1287 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1289 sample.id = ptq->pt->cbr_id;
1290 sample.stream_id = ptq->pt->cbr_id;
1292 flags = (u16)ptq->state->cbr_payload | (pt->max_non_turbo_ratio << 16);
1293 raw.flags = cpu_to_le32(flags);
1294 raw.freq = cpu_to_le32(raw.cbr * pt->cbr2khz);
1297 sample.raw_size = perf_synth__raw_size(raw);
1298 sample.raw_data = perf_synth__raw_data(&raw);
1300 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1301 pt->pwr_events_sample_type);
1304 static int intel_pt_synth_mwait_sample(struct intel_pt_queue *ptq)
1306 struct intel_pt *pt = ptq->pt;
1307 union perf_event *event = ptq->event_buf;
1308 struct perf_sample sample = { .ip = 0, };
1309 struct perf_synth_intel_mwait raw;
1311 if (intel_pt_skip_event(pt))
1314 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1316 sample.id = ptq->pt->mwait_id;
1317 sample.stream_id = ptq->pt->mwait_id;
1320 raw.payload = cpu_to_le64(ptq->state->mwait_payload);
1322 sample.raw_size = perf_synth__raw_size(raw);
1323 sample.raw_data = perf_synth__raw_data(&raw);
1325 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1326 pt->pwr_events_sample_type);
1329 static int intel_pt_synth_pwre_sample(struct intel_pt_queue *ptq)
1331 struct intel_pt *pt = ptq->pt;
1332 union perf_event *event = ptq->event_buf;
1333 struct perf_sample sample = { .ip = 0, };
1334 struct perf_synth_intel_pwre raw;
1336 if (intel_pt_skip_event(pt))
1339 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1341 sample.id = ptq->pt->pwre_id;
1342 sample.stream_id = ptq->pt->pwre_id;
1345 raw.payload = cpu_to_le64(ptq->state->pwre_payload);
1347 sample.raw_size = perf_synth__raw_size(raw);
1348 sample.raw_data = perf_synth__raw_data(&raw);
1350 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1351 pt->pwr_events_sample_type);
1354 static int intel_pt_synth_exstop_sample(struct intel_pt_queue *ptq)
1356 struct intel_pt *pt = ptq->pt;
1357 union perf_event *event = ptq->event_buf;
1358 struct perf_sample sample = { .ip = 0, };
1359 struct perf_synth_intel_exstop raw;
1361 if (intel_pt_skip_event(pt))
1364 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1366 sample.id = ptq->pt->exstop_id;
1367 sample.stream_id = ptq->pt->exstop_id;
1370 raw.ip = !!(ptq->state->flags & INTEL_PT_FUP_IP);
1372 sample.raw_size = perf_synth__raw_size(raw);
1373 sample.raw_data = perf_synth__raw_data(&raw);
1375 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1376 pt->pwr_events_sample_type);
1379 static int intel_pt_synth_pwrx_sample(struct intel_pt_queue *ptq)
1381 struct intel_pt *pt = ptq->pt;
1382 union perf_event *event = ptq->event_buf;
1383 struct perf_sample sample = { .ip = 0, };
1384 struct perf_synth_intel_pwrx raw;
1386 if (intel_pt_skip_event(pt))
1389 intel_pt_prep_p_sample(pt, ptq, event, &sample);
1391 sample.id = ptq->pt->pwrx_id;
1392 sample.stream_id = ptq->pt->pwrx_id;
1395 raw.payload = cpu_to_le64(ptq->state->pwrx_payload);
1397 sample.raw_size = perf_synth__raw_size(raw);
1398 sample.raw_data = perf_synth__raw_data(&raw);
1400 return intel_pt_deliver_synth_event(pt, ptq, event, &sample,
1401 pt->pwr_events_sample_type);
1404 static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,
1405 pid_t pid, pid_t tid, u64 ip, u64 timestamp)
1407 union perf_event event;
1408 char msg[MAX_AUXTRACE_ERROR_MSG];
1411 intel_pt__strerror(code, msg, MAX_AUXTRACE_ERROR_MSG);
1413 auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
1414 code, cpu, pid, tid, ip, msg, timestamp);
1416 err = perf_session__deliver_synth_event(pt->session, &event, NULL);
1418 pr_err("Intel Processor Trace: failed to deliver error event, error %d\n",
1424 static int intel_ptq_synth_error(struct intel_pt_queue *ptq,
1425 const struct intel_pt_state *state)
1427 struct intel_pt *pt = ptq->pt;
1428 u64 tm = ptq->timestamp;
1430 tm = pt->timeless_decoding ? 0 : tsc_to_perf_time(tm, &pt->tc);
1432 return intel_pt_synth_error(pt, state->err, ptq->cpu, ptq->pid,
1433 ptq->tid, state->from_ip, tm);
1436 static int intel_pt_next_tid(struct intel_pt *pt, struct intel_pt_queue *ptq)
1438 struct auxtrace_queue *queue;
1439 pid_t tid = ptq->next_tid;
1445 intel_pt_log("switch: cpu %d tid %d\n", ptq->cpu, tid);
1447 err = machine__set_current_tid(pt->machine, ptq->cpu, -1, tid);
1449 queue = &pt->queues.queue_array[ptq->queue_nr];
1450 intel_pt_set_pid_tid_cpu(pt, queue);
1457 static inline bool intel_pt_is_switch_ip(struct intel_pt_queue *ptq, u64 ip)
1459 struct intel_pt *pt = ptq->pt;
1461 return ip == pt->switch_ip &&
1462 (ptq->flags & PERF_IP_FLAG_BRANCH) &&
1463 !(ptq->flags & (PERF_IP_FLAG_CONDITIONAL | PERF_IP_FLAG_ASYNC |
1464 PERF_IP_FLAG_INTERRUPT | PERF_IP_FLAG_TX_ABORT));
1467 #define INTEL_PT_PWR_EVT (INTEL_PT_MWAIT_OP | INTEL_PT_PWR_ENTRY | \
1468 INTEL_PT_EX_STOP | INTEL_PT_PWR_EXIT | \
1471 static int intel_pt_sample(struct intel_pt_queue *ptq)
1473 const struct intel_pt_state *state = ptq->state;
1474 struct intel_pt *pt = ptq->pt;
1477 if (!ptq->have_sample)
1480 ptq->have_sample = false;
1482 if (pt->sample_pwr_events && (state->type & INTEL_PT_PWR_EVT)) {
1483 if (state->type & INTEL_PT_CBR_CHG) {
1484 err = intel_pt_synth_cbr_sample(ptq);
1488 if (state->type & INTEL_PT_MWAIT_OP) {
1489 err = intel_pt_synth_mwait_sample(ptq);
1493 if (state->type & INTEL_PT_PWR_ENTRY) {
1494 err = intel_pt_synth_pwre_sample(ptq);
1498 if (state->type & INTEL_PT_EX_STOP) {
1499 err = intel_pt_synth_exstop_sample(ptq);
1503 if (state->type & INTEL_PT_PWR_EXIT) {
1504 err = intel_pt_synth_pwrx_sample(ptq);
1510 if (pt->sample_instructions && (state->type & INTEL_PT_INSTRUCTION)) {
1511 err = intel_pt_synth_instruction_sample(ptq);
1516 if (pt->sample_transactions && (state->type & INTEL_PT_TRANSACTION)) {
1517 err = intel_pt_synth_transaction_sample(ptq);
1522 if (pt->sample_ptwrites && (state->type & INTEL_PT_PTW)) {
1523 err = intel_pt_synth_ptwrite_sample(ptq);
1528 if (!(state->type & INTEL_PT_BRANCH))
1531 if (pt->synth_opts.callchain || pt->synth_opts.thread_stack)
1532 thread_stack__event(ptq->thread, ptq->cpu, ptq->flags, state->from_ip,
1533 state->to_ip, ptq->insn_len,
1536 thread_stack__set_trace_nr(ptq->thread, ptq->cpu, state->trace_nr);
1538 if (pt->sample_branches) {
1539 err = intel_pt_synth_branch_sample(ptq);
1544 if (pt->synth_opts.last_branch)
1545 intel_pt_update_last_branch_rb(ptq);
1547 if (!ptq->sync_switch)
1550 if (intel_pt_is_switch_ip(ptq, state->to_ip)) {
1551 switch (ptq->switch_state) {
1552 case INTEL_PT_SS_NOT_TRACING:
1553 case INTEL_PT_SS_UNKNOWN:
1554 case INTEL_PT_SS_EXPECTING_SWITCH_IP:
1555 err = intel_pt_next_tid(pt, ptq);
1558 ptq->switch_state = INTEL_PT_SS_TRACING;
1561 ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_EVENT;
1564 } else if (!state->to_ip) {
1565 ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
1566 } else if (ptq->switch_state == INTEL_PT_SS_NOT_TRACING) {
1567 ptq->switch_state = INTEL_PT_SS_UNKNOWN;
1568 } else if (ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
1569 state->to_ip == pt->ptss_ip &&
1570 (ptq->flags & PERF_IP_FLAG_CALL)) {
1571 ptq->switch_state = INTEL_PT_SS_TRACING;
1577 static u64 intel_pt_switch_ip(struct intel_pt *pt, u64 *ptss_ip)
1579 struct machine *machine = pt->machine;
1581 struct symbol *sym, *start;
1582 u64 ip, switch_ip = 0;
1588 map = machine__kernel_map(machine);
1595 start = dso__first_symbol(map->dso);
1597 for (sym = start; sym; sym = dso__next_symbol(sym)) {
1598 if (sym->binding == STB_GLOBAL &&
1599 !strcmp(sym->name, "__switch_to")) {
1600 ip = map->unmap_ip(map, sym->start);
1601 if (ip >= map->start && ip < map->end) {
1608 if (!switch_ip || !ptss_ip)
1611 if (pt->have_sched_switch == 1)
1612 ptss = "perf_trace_sched_switch";
1614 ptss = "__perf_event_task_sched_out";
1616 for (sym = start; sym; sym = dso__next_symbol(sym)) {
1617 if (!strcmp(sym->name, ptss)) {
1618 ip = map->unmap_ip(map, sym->start);
1619 if (ip >= map->start && ip < map->end) {
1629 static void intel_pt_enable_sync_switch(struct intel_pt *pt)
1633 pt->sync_switch = true;
1635 for (i = 0; i < pt->queues.nr_queues; i++) {
1636 struct auxtrace_queue *queue = &pt->queues.queue_array[i];
1637 struct intel_pt_queue *ptq = queue->priv;
1640 ptq->sync_switch = true;
1644 static int intel_pt_run_decoder(struct intel_pt_queue *ptq, u64 *timestamp)
1646 const struct intel_pt_state *state = ptq->state;
1647 struct intel_pt *pt = ptq->pt;
1650 if (!pt->kernel_start) {
1651 pt->kernel_start = machine__kernel_start(pt->machine);
1652 if (pt->per_cpu_mmaps &&
1653 (pt->have_sched_switch == 1 || pt->have_sched_switch == 3) &&
1654 !pt->timeless_decoding && intel_pt_tracing_kernel(pt) &&
1655 !pt->sampling_mode) {
1656 pt->switch_ip = intel_pt_switch_ip(pt, &pt->ptss_ip);
1657 if (pt->switch_ip) {
1658 intel_pt_log("switch_ip: %"PRIx64" ptss_ip: %"PRIx64"\n",
1659 pt->switch_ip, pt->ptss_ip);
1660 intel_pt_enable_sync_switch(pt);
1665 intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
1666 ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
1668 err = intel_pt_sample(ptq);
1672 state = intel_pt_decode(ptq->decoder);
1674 if (state->err == INTEL_PT_ERR_NODATA)
1676 if (ptq->sync_switch &&
1677 state->from_ip >= pt->kernel_start) {
1678 ptq->sync_switch = false;
1679 intel_pt_next_tid(pt, ptq);
1681 if (pt->synth_opts.errors) {
1682 err = intel_ptq_synth_error(ptq, state);
1690 ptq->have_sample = true;
1691 intel_pt_sample_flags(ptq);
1693 /* Use estimated TSC upon return to user space */
1695 (state->from_ip >= pt->kernel_start || !state->from_ip) &&
1696 state->to_ip && state->to_ip < pt->kernel_start) {
1697 intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
1698 state->timestamp, state->est_timestamp);
1699 ptq->timestamp = state->est_timestamp;
1700 /* Use estimated TSC in unknown switch state */
1701 } else if (ptq->sync_switch &&
1702 ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
1703 intel_pt_is_switch_ip(ptq, state->to_ip) &&
1704 ptq->next_tid == -1) {
1705 intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
1706 state->timestamp, state->est_timestamp);
1707 ptq->timestamp = state->est_timestamp;
1708 } else if (state->timestamp > ptq->timestamp) {
1709 ptq->timestamp = state->timestamp;
1712 if (!pt->timeless_decoding && ptq->timestamp >= *timestamp) {
1713 *timestamp = ptq->timestamp;
1720 static inline int intel_pt_update_queues(struct intel_pt *pt)
1722 if (pt->queues.new_data) {
1723 pt->queues.new_data = false;
1724 return intel_pt_setup_queues(pt);
1729 static int intel_pt_process_queues(struct intel_pt *pt, u64 timestamp)
1731 unsigned int queue_nr;
1736 struct auxtrace_queue *queue;
1737 struct intel_pt_queue *ptq;
1739 if (!pt->heap.heap_cnt)
1742 if (pt->heap.heap_array[0].ordinal >= timestamp)
1745 queue_nr = pt->heap.heap_array[0].queue_nr;
1746 queue = &pt->queues.queue_array[queue_nr];
1749 intel_pt_log("queue %u processing 0x%" PRIx64 " to 0x%" PRIx64 "\n",
1750 queue_nr, pt->heap.heap_array[0].ordinal,
1753 auxtrace_heap__pop(&pt->heap);
1755 if (pt->heap.heap_cnt) {
1756 ts = pt->heap.heap_array[0].ordinal + 1;
1763 intel_pt_set_pid_tid_cpu(pt, queue);
1765 ret = intel_pt_run_decoder(ptq, &ts);
1768 auxtrace_heap__add(&pt->heap, queue_nr, ts);
1773 ret = auxtrace_heap__add(&pt->heap, queue_nr, ts);
1777 ptq->on_heap = false;
1784 static int intel_pt_process_timeless_queues(struct intel_pt *pt, pid_t tid,
1787 struct auxtrace_queues *queues = &pt->queues;
1791 for (i = 0; i < queues->nr_queues; i++) {
1792 struct auxtrace_queue *queue = &pt->queues.queue_array[i];
1793 struct intel_pt_queue *ptq = queue->priv;
1795 if (ptq && (tid == -1 || ptq->tid == tid)) {
1797 intel_pt_set_pid_tid_cpu(pt, queue);
1798 intel_pt_run_decoder(ptq, &ts);
1804 static int intel_pt_lost(struct intel_pt *pt, struct perf_sample *sample)
1806 return intel_pt_synth_error(pt, INTEL_PT_ERR_LOST, sample->cpu,
1807 sample->pid, sample->tid, 0, sample->time);
1810 static struct intel_pt_queue *intel_pt_cpu_to_ptq(struct intel_pt *pt, int cpu)
1814 if (cpu < 0 || !pt->queues.nr_queues)
1817 if ((unsigned)cpu >= pt->queues.nr_queues)
1818 i = pt->queues.nr_queues - 1;
1822 if (pt->queues.queue_array[i].cpu == cpu)
1823 return pt->queues.queue_array[i].priv;
1825 for (j = 0; i > 0; j++) {
1826 if (pt->queues.queue_array[--i].cpu == cpu)
1827 return pt->queues.queue_array[i].priv;
1830 for (; j < pt->queues.nr_queues; j++) {
1831 if (pt->queues.queue_array[j].cpu == cpu)
1832 return pt->queues.queue_array[j].priv;
1838 static int intel_pt_sync_switch(struct intel_pt *pt, int cpu, pid_t tid,
1841 struct intel_pt_queue *ptq;
1844 if (!pt->sync_switch)
1847 ptq = intel_pt_cpu_to_ptq(pt, cpu);
1848 if (!ptq || !ptq->sync_switch)
1851 switch (ptq->switch_state) {
1852 case INTEL_PT_SS_NOT_TRACING:
1855 case INTEL_PT_SS_UNKNOWN:
1856 case INTEL_PT_SS_TRACING:
1857 ptq->next_tid = tid;
1858 ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_IP;
1860 case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
1861 if (!ptq->on_heap) {
1862 ptq->timestamp = perf_time_to_tsc(timestamp,
1864 err = auxtrace_heap__add(&pt->heap, ptq->queue_nr,
1868 ptq->on_heap = true;
1870 ptq->switch_state = INTEL_PT_SS_TRACING;
1872 case INTEL_PT_SS_EXPECTING_SWITCH_IP:
1873 ptq->next_tid = tid;
1874 intel_pt_log("ERROR: cpu %d expecting switch ip\n", cpu);
1883 static int intel_pt_process_switch(struct intel_pt *pt,
1884 struct perf_sample *sample)
1886 struct perf_evsel *evsel;
1890 evsel = perf_evlist__id2evsel(pt->session->evlist, sample->id);
1891 if (evsel != pt->switch_evsel)
1894 tid = perf_evsel__intval(evsel, sample, "next_pid");
1897 intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1898 cpu, tid, sample->time, perf_time_to_tsc(sample->time,
1901 ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
1905 return machine__set_current_tid(pt->machine, cpu, -1, tid);
1908 static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,
1909 struct perf_sample *sample)
1911 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1917 if (pt->have_sched_switch == 3) {
1920 if (event->header.type != PERF_RECORD_SWITCH_CPU_WIDE) {
1921 pr_err("Expecting CPU-wide context switch event\n");
1924 pid = event->context_switch.next_prev_pid;
1925 tid = event->context_switch.next_prev_tid;
1934 pr_err("context_switch event has no tid\n");
1938 intel_pt_log("context_switch: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1939 cpu, pid, tid, sample->time, perf_time_to_tsc(sample->time,
1942 ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
1946 return machine__set_current_tid(pt->machine, cpu, pid, tid);
1949 static int intel_pt_process_itrace_start(struct intel_pt *pt,
1950 union perf_event *event,
1951 struct perf_sample *sample)
1953 if (!pt->per_cpu_mmaps)
1956 intel_pt_log("itrace_start: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1957 sample->cpu, event->itrace_start.pid,
1958 event->itrace_start.tid, sample->time,
1959 perf_time_to_tsc(sample->time, &pt->tc));
1961 return machine__set_current_tid(pt->machine, sample->cpu,
1962 event->itrace_start.pid,
1963 event->itrace_start.tid);
1966 static int intel_pt_process_event(struct perf_session *session,
1967 union perf_event *event,
1968 struct perf_sample *sample,
1969 struct perf_tool *tool)
1971 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1979 if (!tool->ordered_events) {
1980 pr_err("Intel Processor Trace requires ordered events\n");
1984 if (sample->time && sample->time != (u64)-1)
1985 timestamp = perf_time_to_tsc(sample->time, &pt->tc);
1989 if (timestamp || pt->timeless_decoding) {
1990 err = intel_pt_update_queues(pt);
1995 if (pt->timeless_decoding) {
1996 if (event->header.type == PERF_RECORD_EXIT) {
1997 err = intel_pt_process_timeless_queues(pt,
2001 } else if (timestamp) {
2002 err = intel_pt_process_queues(pt, timestamp);
2007 if (event->header.type == PERF_RECORD_AUX &&
2008 (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
2009 pt->synth_opts.errors) {
2010 err = intel_pt_lost(pt, sample);
2015 if (pt->switch_evsel && event->header.type == PERF_RECORD_SAMPLE)
2016 err = intel_pt_process_switch(pt, sample);
2017 else if (event->header.type == PERF_RECORD_ITRACE_START)
2018 err = intel_pt_process_itrace_start(pt, event, sample);
2019 else if (event->header.type == PERF_RECORD_SWITCH ||
2020 event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
2021 err = intel_pt_context_switch(pt, event, sample);
2023 intel_pt_log("event %u: cpu %d time %"PRIu64" tsc %#"PRIx64" ",
2024 event->header.type, sample->cpu, sample->time, timestamp);
2025 intel_pt_log_event(event);
2030 static int intel_pt_flush(struct perf_session *session, struct perf_tool *tool)
2032 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2039 if (!tool->ordered_events)
2042 ret = intel_pt_update_queues(pt);
2046 if (pt->timeless_decoding)
2047 return intel_pt_process_timeless_queues(pt, -1,
2050 return intel_pt_process_queues(pt, MAX_TIMESTAMP);
2053 static void intel_pt_free_events(struct perf_session *session)
2055 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2057 struct auxtrace_queues *queues = &pt->queues;
2060 for (i = 0; i < queues->nr_queues; i++) {
2061 intel_pt_free_queue(queues->queue_array[i].priv);
2062 queues->queue_array[i].priv = NULL;
2064 intel_pt_log_disable();
2065 auxtrace_queues__free(queues);
2068 static void intel_pt_free(struct perf_session *session)
2070 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2073 auxtrace_heap__free(&pt->heap);
2074 intel_pt_free_events(session);
2075 session->auxtrace = NULL;
2076 thread__put(pt->unknown_thread);
2077 addr_filters__exit(&pt->filts);
2082 static int intel_pt_process_auxtrace_event(struct perf_session *session,
2083 union perf_event *event,
2084 struct perf_tool *tool __maybe_unused)
2086 struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
2089 if (!pt->data_queued) {
2090 struct auxtrace_buffer *buffer;
2092 int fd = perf_data__fd(session->data);
2095 if (perf_data__is_pipe(session->data)) {
2098 data_offset = lseek(fd, 0, SEEK_CUR);
2099 if (data_offset == -1)
2103 err = auxtrace_queues__add_event(&pt->queues, session, event,
2104 data_offset, &buffer);
2108 /* Dump here now we have copied a piped trace out of the pipe */
2110 if (auxtrace_buffer__get_data(buffer, fd)) {
2111 intel_pt_dump_event(pt, buffer->data,
2113 auxtrace_buffer__put_data(buffer);
2121 struct intel_pt_synth {
2122 struct perf_tool dummy_tool;
2123 struct perf_session *session;
2126 static int intel_pt_event_synth(struct perf_tool *tool,
2127 union perf_event *event,
2128 struct perf_sample *sample __maybe_unused,
2129 struct machine *machine __maybe_unused)
2131 struct intel_pt_synth *intel_pt_synth =
2132 container_of(tool, struct intel_pt_synth, dummy_tool);
2134 return perf_session__deliver_synth_event(intel_pt_synth->session, event,
2138 static int intel_pt_synth_event(struct perf_session *session, const char *name,
2139 struct perf_event_attr *attr, u64 id)
2141 struct intel_pt_synth intel_pt_synth;
2144 pr_debug("Synthesizing '%s' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
2145 name, id, (u64)attr->sample_type);
2147 memset(&intel_pt_synth, 0, sizeof(struct intel_pt_synth));
2148 intel_pt_synth.session = session;
2150 err = perf_event__synthesize_attr(&intel_pt_synth.dummy_tool, attr, 1,
2151 &id, intel_pt_event_synth);
2153 pr_err("%s: failed to synthesize '%s' event type\n",
2159 static void intel_pt_set_event_name(struct perf_evlist *evlist, u64 id,
2162 struct perf_evsel *evsel;
2164 evlist__for_each_entry(evlist, evsel) {
2165 if (evsel->id && evsel->id[0] == id) {
2167 zfree(&evsel->name);
2168 evsel->name = strdup(name);
2174 static struct perf_evsel *intel_pt_evsel(struct intel_pt *pt,
2175 struct perf_evlist *evlist)
2177 struct perf_evsel *evsel;
2179 evlist__for_each_entry(evlist, evsel) {
2180 if (evsel->attr.type == pt->pmu_type && evsel->ids)
2187 static int intel_pt_synth_events(struct intel_pt *pt,
2188 struct perf_session *session)
2190 struct perf_evlist *evlist = session->evlist;
2191 struct perf_evsel *evsel = intel_pt_evsel(pt, evlist);
2192 struct perf_event_attr attr;
2197 pr_debug("There are no selected events with Intel Processor Trace data\n");
2201 memset(&attr, 0, sizeof(struct perf_event_attr));
2202 attr.size = sizeof(struct perf_event_attr);
2203 attr.type = PERF_TYPE_HARDWARE;
2204 attr.sample_type = evsel->attr.sample_type & PERF_SAMPLE_MASK;
2205 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
2207 if (pt->timeless_decoding)
2208 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
2210 attr.sample_type |= PERF_SAMPLE_TIME;
2211 if (!pt->per_cpu_mmaps)
2212 attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
2213 attr.exclude_user = evsel->attr.exclude_user;
2214 attr.exclude_kernel = evsel->attr.exclude_kernel;
2215 attr.exclude_hv = evsel->attr.exclude_hv;
2216 attr.exclude_host = evsel->attr.exclude_host;
2217 attr.exclude_guest = evsel->attr.exclude_guest;
2218 attr.sample_id_all = evsel->attr.sample_id_all;
2219 attr.read_format = evsel->attr.read_format;
2221 id = evsel->id[0] + 1000000000;
2225 if (pt->synth_opts.branches) {
2226 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
2227 attr.sample_period = 1;
2228 attr.sample_type |= PERF_SAMPLE_ADDR;
2229 err = intel_pt_synth_event(session, "branches", &attr, id);
2232 pt->sample_branches = true;
2233 pt->branches_sample_type = attr.sample_type;
2234 pt->branches_id = id;
2236 attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR;
2239 if (pt->synth_opts.callchain)
2240 attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
2241 if (pt->synth_opts.last_branch)
2242 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
2244 if (pt->synth_opts.instructions) {
2245 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
2246 if (pt->synth_opts.period_type == PERF_ITRACE_PERIOD_NANOSECS)
2247 attr.sample_period =
2248 intel_pt_ns_to_ticks(pt, pt->synth_opts.period);
2250 attr.sample_period = pt->synth_opts.period;
2251 err = intel_pt_synth_event(session, "instructions", &attr, id);
2254 pt->sample_instructions = true;
2255 pt->instructions_sample_type = attr.sample_type;
2256 pt->instructions_id = id;
2260 attr.sample_type &= ~(u64)PERF_SAMPLE_PERIOD;
2261 attr.sample_period = 1;
2263 if (pt->synth_opts.transactions) {
2264 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
2265 err = intel_pt_synth_event(session, "transactions", &attr, id);
2268 pt->sample_transactions = true;
2269 pt->transactions_sample_type = attr.sample_type;
2270 pt->transactions_id = id;
2271 intel_pt_set_event_name(evlist, id, "transactions");
2275 attr.type = PERF_TYPE_SYNTH;
2276 attr.sample_type |= PERF_SAMPLE_RAW;
2278 if (pt->synth_opts.ptwrites) {
2279 attr.config = PERF_SYNTH_INTEL_PTWRITE;
2280 err = intel_pt_synth_event(session, "ptwrite", &attr, id);
2283 pt->sample_ptwrites = true;
2284 pt->ptwrites_sample_type = attr.sample_type;
2285 pt->ptwrites_id = id;
2286 intel_pt_set_event_name(evlist, id, "ptwrite");
2290 if (pt->synth_opts.pwr_events) {
2291 pt->sample_pwr_events = true;
2292 pt->pwr_events_sample_type = attr.sample_type;
2294 attr.config = PERF_SYNTH_INTEL_CBR;
2295 err = intel_pt_synth_event(session, "cbr", &attr, id);
2299 intel_pt_set_event_name(evlist, id, "cbr");
2303 if (pt->synth_opts.pwr_events && (evsel->attr.config & 0x10)) {
2304 attr.config = PERF_SYNTH_INTEL_MWAIT;
2305 err = intel_pt_synth_event(session, "mwait", &attr, id);
2309 intel_pt_set_event_name(evlist, id, "mwait");
2312 attr.config = PERF_SYNTH_INTEL_PWRE;
2313 err = intel_pt_synth_event(session, "pwre", &attr, id);
2317 intel_pt_set_event_name(evlist, id, "pwre");
2320 attr.config = PERF_SYNTH_INTEL_EXSTOP;
2321 err = intel_pt_synth_event(session, "exstop", &attr, id);
2325 intel_pt_set_event_name(evlist, id, "exstop");
2328 attr.config = PERF_SYNTH_INTEL_PWRX;
2329 err = intel_pt_synth_event(session, "pwrx", &attr, id);
2333 intel_pt_set_event_name(evlist, id, "pwrx");
2340 static struct perf_evsel *intel_pt_find_sched_switch(struct perf_evlist *evlist)
2342 struct perf_evsel *evsel;
2344 evlist__for_each_entry_reverse(evlist, evsel) {
2345 const char *name = perf_evsel__name(evsel);
2347 if (!strcmp(name, "sched:sched_switch"))
2354 static bool intel_pt_find_switch(struct perf_evlist *evlist)
2356 struct perf_evsel *evsel;
2358 evlist__for_each_entry(evlist, evsel) {
2359 if (evsel->attr.context_switch)
2366 static int intel_pt_perf_config(const char *var, const char *value, void *data)
2368 struct intel_pt *pt = data;
2370 if (!strcmp(var, "intel-pt.mispred-all"))
2371 pt->mispred_all = perf_config_bool(var, value);
2376 static const char * const intel_pt_info_fmts[] = {
2377 [INTEL_PT_PMU_TYPE] = " PMU Type %"PRId64"\n",
2378 [INTEL_PT_TIME_SHIFT] = " Time Shift %"PRIu64"\n",
2379 [INTEL_PT_TIME_MULT] = " Time Muliplier %"PRIu64"\n",
2380 [INTEL_PT_TIME_ZERO] = " Time Zero %"PRIu64"\n",
2381 [INTEL_PT_CAP_USER_TIME_ZERO] = " Cap Time Zero %"PRId64"\n",
2382 [INTEL_PT_TSC_BIT] = " TSC bit %#"PRIx64"\n",
2383 [INTEL_PT_NORETCOMP_BIT] = " NoRETComp bit %#"PRIx64"\n",
2384 [INTEL_PT_HAVE_SCHED_SWITCH] = " Have sched_switch %"PRId64"\n",
2385 [INTEL_PT_SNAPSHOT_MODE] = " Snapshot mode %"PRId64"\n",
2386 [INTEL_PT_PER_CPU_MMAPS] = " Per-cpu maps %"PRId64"\n",
2387 [INTEL_PT_MTC_BIT] = " MTC bit %#"PRIx64"\n",
2388 [INTEL_PT_TSC_CTC_N] = " TSC:CTC numerator %"PRIu64"\n",
2389 [INTEL_PT_TSC_CTC_D] = " TSC:CTC denominator %"PRIu64"\n",
2390 [INTEL_PT_CYC_BIT] = " CYC bit %#"PRIx64"\n",
2391 [INTEL_PT_MAX_NONTURBO_RATIO] = " Max non-turbo ratio %"PRIu64"\n",
2392 [INTEL_PT_FILTER_STR_LEN] = " Filter string len. %"PRIu64"\n",
2395 static void intel_pt_print_info(u64 *arr, int start, int finish)
2402 for (i = start; i <= finish; i++)
2403 fprintf(stdout, intel_pt_info_fmts[i], arr[i]);
2406 static void intel_pt_print_info_str(const char *name, const char *str)
2411 fprintf(stdout, " %-20s%s\n", name, str ? str : "");
2414 static bool intel_pt_has(struct auxtrace_info_event *auxtrace_info, int pos)
2416 return auxtrace_info->header.size >=
2417 sizeof(struct auxtrace_info_event) + (sizeof(u64) * (pos + 1));
2420 int intel_pt_process_auxtrace_info(union perf_event *event,
2421 struct perf_session *session)
2423 struct auxtrace_info_event *auxtrace_info = &event->auxtrace_info;
2424 size_t min_sz = sizeof(u64) * INTEL_PT_PER_CPU_MMAPS;
2425 struct intel_pt *pt;
2430 if (auxtrace_info->header.size < sizeof(struct auxtrace_info_event) +
2434 pt = zalloc(sizeof(struct intel_pt));
2438 addr_filters__init(&pt->filts);
2440 err = perf_config(intel_pt_perf_config, pt);
2444 err = auxtrace_queues__init(&pt->queues);
2448 intel_pt_log_set_name(INTEL_PT_PMU_NAME);
2450 pt->session = session;
2451 pt->machine = &session->machines.host; /* No kvm support */
2452 pt->auxtrace_type = auxtrace_info->type;
2453 pt->pmu_type = auxtrace_info->priv[INTEL_PT_PMU_TYPE];
2454 pt->tc.time_shift = auxtrace_info->priv[INTEL_PT_TIME_SHIFT];
2455 pt->tc.time_mult = auxtrace_info->priv[INTEL_PT_TIME_MULT];
2456 pt->tc.time_zero = auxtrace_info->priv[INTEL_PT_TIME_ZERO];
2457 pt->cap_user_time_zero = auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO];
2458 pt->tsc_bit = auxtrace_info->priv[INTEL_PT_TSC_BIT];
2459 pt->noretcomp_bit = auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT];
2460 pt->have_sched_switch = auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH];
2461 pt->snapshot_mode = auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE];
2462 pt->per_cpu_mmaps = auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS];
2463 intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_PMU_TYPE,
2464 INTEL_PT_PER_CPU_MMAPS);
2466 if (intel_pt_has(auxtrace_info, INTEL_PT_CYC_BIT)) {
2467 pt->mtc_bit = auxtrace_info->priv[INTEL_PT_MTC_BIT];
2468 pt->mtc_freq_bits = auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS];
2469 pt->tsc_ctc_ratio_n = auxtrace_info->priv[INTEL_PT_TSC_CTC_N];
2470 pt->tsc_ctc_ratio_d = auxtrace_info->priv[INTEL_PT_TSC_CTC_D];
2471 pt->cyc_bit = auxtrace_info->priv[INTEL_PT_CYC_BIT];
2472 intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_MTC_BIT,
2476 if (intel_pt_has(auxtrace_info, INTEL_PT_MAX_NONTURBO_RATIO)) {
2477 pt->max_non_turbo_ratio =
2478 auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO];
2479 intel_pt_print_info(&auxtrace_info->priv[0],
2480 INTEL_PT_MAX_NONTURBO_RATIO,
2481 INTEL_PT_MAX_NONTURBO_RATIO);
2484 info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
2485 info_end = (void *)info + auxtrace_info->header.size;
2487 if (intel_pt_has(auxtrace_info, INTEL_PT_FILTER_STR_LEN)) {
2490 len = auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN];
2491 intel_pt_print_info(&auxtrace_info->priv[0],
2492 INTEL_PT_FILTER_STR_LEN,
2493 INTEL_PT_FILTER_STR_LEN);
2495 const char *filter = (const char *)info;
2497 len = roundup(len + 1, 8);
2499 if ((void *)info > info_end) {
2500 pr_err("%s: bad filter string length\n", __func__);
2502 goto err_free_queues;
2504 pt->filter = memdup(filter, len);
2507 goto err_free_queues;
2509 if (session->header.needs_swap)
2510 mem_bswap_64(pt->filter, len);
2511 if (pt->filter[len - 1]) {
2512 pr_err("%s: filter string not null terminated\n", __func__);
2514 goto err_free_queues;
2516 err = addr_filters__parse_bare_filter(&pt->filts,
2519 goto err_free_queues;
2521 intel_pt_print_info_str("Filter string", pt->filter);
2524 pt->timeless_decoding = intel_pt_timeless_decoding(pt);
2525 if (pt->timeless_decoding && !pt->tc.time_mult)
2526 pt->tc.time_mult = 1;
2527 pt->have_tsc = intel_pt_have_tsc(pt);
2528 pt->sampling_mode = false;
2529 pt->est_tsc = !pt->timeless_decoding;
2531 pt->unknown_thread = thread__new(999999999, 999999999);
2532 if (!pt->unknown_thread) {
2534 goto err_free_queues;
2538 * Since this thread will not be kept in any rbtree not in a
2539 * list, initialize its list node so that at thread__put() the
2540 * current thread lifetime assuption is kept and we don't segfault
2541 * at list_del_init().
2543 INIT_LIST_HEAD(&pt->unknown_thread->node);
2545 err = thread__set_comm(pt->unknown_thread, "unknown", 0);
2547 goto err_delete_thread;
2548 if (thread__init_map_groups(pt->unknown_thread, pt->machine)) {
2550 goto err_delete_thread;
2553 pt->auxtrace.process_event = intel_pt_process_event;
2554 pt->auxtrace.process_auxtrace_event = intel_pt_process_auxtrace_event;
2555 pt->auxtrace.flush_events = intel_pt_flush;
2556 pt->auxtrace.free_events = intel_pt_free_events;
2557 pt->auxtrace.free = intel_pt_free;
2558 session->auxtrace = &pt->auxtrace;
2563 if (pt->have_sched_switch == 1) {
2564 pt->switch_evsel = intel_pt_find_sched_switch(session->evlist);
2565 if (!pt->switch_evsel) {
2566 pr_err("%s: missing sched_switch event\n", __func__);
2568 goto err_delete_thread;
2570 } else if (pt->have_sched_switch == 2 &&
2571 !intel_pt_find_switch(session->evlist)) {
2572 pr_err("%s: missing context_switch attribute flag\n", __func__);
2574 goto err_delete_thread;
2577 if (session->itrace_synth_opts && session->itrace_synth_opts->set) {
2578 pt->synth_opts = *session->itrace_synth_opts;
2580 itrace_synth_opts__set_default(&pt->synth_opts,
2581 session->itrace_synth_opts->default_no_sample);
2582 if (use_browser != -1) {
2583 pt->synth_opts.branches = false;
2584 pt->synth_opts.callchain = true;
2586 if (session->itrace_synth_opts)
2587 pt->synth_opts.thread_stack =
2588 session->itrace_synth_opts->thread_stack;
2591 if (pt->synth_opts.log)
2592 intel_pt_log_enable();
2594 /* Maximum non-turbo ratio is TSC freq / 100 MHz */
2595 if (pt->tc.time_mult) {
2596 u64 tsc_freq = intel_pt_ns_to_ticks(pt, 1000000000);
2598 if (!pt->max_non_turbo_ratio)
2599 pt->max_non_turbo_ratio =
2600 (tsc_freq + 50000000) / 100000000;
2601 intel_pt_log("TSC frequency %"PRIu64"\n", tsc_freq);
2602 intel_pt_log("Maximum non-turbo ratio %u\n",
2603 pt->max_non_turbo_ratio);
2604 pt->cbr2khz = tsc_freq / pt->max_non_turbo_ratio / 1000;
2607 if (pt->synth_opts.calls)
2608 pt->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
2609 PERF_IP_FLAG_TRACE_END;
2610 if (pt->synth_opts.returns)
2611 pt->branches_filter |= PERF_IP_FLAG_RETURN |
2612 PERF_IP_FLAG_TRACE_BEGIN;
2614 if (pt->synth_opts.callchain && !symbol_conf.use_callchain) {
2615 symbol_conf.use_callchain = true;
2616 if (callchain_register_param(&callchain_param) < 0) {
2617 symbol_conf.use_callchain = false;
2618 pt->synth_opts.callchain = false;
2622 err = intel_pt_synth_events(pt, session);
2624 goto err_delete_thread;
2626 err = auxtrace_queues__process_index(&pt->queues, session);
2628 goto err_delete_thread;
2630 if (pt->queues.populated)
2631 pt->data_queued = true;
2633 if (pt->timeless_decoding)
2634 pr_debug2("Intel PT decoding without timestamps\n");
2639 thread__zput(pt->unknown_thread);
2641 intel_pt_log_disable();
2642 auxtrace_queues__free(&pt->queues);
2643 session->auxtrace = NULL;
2645 addr_filters__exit(&pt->filts);