1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright(C) 2015-2018 Linaro Limited.
5 * Author: Tor Jeremiassen <tor@ti.com>
6 * Author: Mathieu Poirier <mathieu.poirier@linaro.org>
9 #include <linux/bitops.h>
10 #include <linux/err.h>
11 #include <linux/kernel.h>
12 #include <linux/log2.h>
13 #include <linux/types.h>
20 #include "cs-etm-decoder/cs-etm-decoder.h"
28 #include "thread_map.h"
29 #include "thread-stack.h"
32 #define MAX_TIMESTAMP (~0ULL)
35 * A64 instructions are always 4 bytes
37 * Only A64 is supported, so can use this constant for converting between
38 * addresses and instruction counts, calculting offsets etc
40 #define A64_INSTR_SIZE 4
42 struct cs_etm_auxtrace {
43 struct auxtrace auxtrace;
44 struct auxtrace_queues queues;
45 struct auxtrace_heap heap;
46 struct itrace_synth_opts synth_opts;
47 struct perf_session *session;
48 struct machine *machine;
49 struct thread *unknown_thread;
55 u8 sample_instructions;
59 u64 branches_sample_type;
61 u64 instructions_sample_type;
62 u64 instructions_sample_period;
66 unsigned int pmu_type;
70 struct cs_etm_auxtrace *etm;
71 struct thread *thread;
72 struct cs_etm_decoder *decoder;
73 struct auxtrace_buffer *buffer;
74 const struct cs_etm_state *state;
75 union perf_event *event_buf;
76 unsigned int queue_nr;
82 u64 period_instructions;
83 struct branch_stack *last_branch;
84 struct branch_stack *last_branch_rb;
85 size_t last_branch_pos;
86 struct cs_etm_packet *prev_packet;
87 struct cs_etm_packet *packet;
90 static int cs_etm__update_queues(struct cs_etm_auxtrace *etm);
91 static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm,
92 pid_t tid, u64 time_);
94 static void cs_etm__packet_dump(const char *pkt_string)
96 const char *color = PERF_COLOR_BLUE;
97 int len = strlen(pkt_string);
99 if (len && (pkt_string[len-1] == '\n'))
100 color_fprintf(stdout, color, " %s", pkt_string);
102 color_fprintf(stdout, color, " %s\n", pkt_string);
107 static void cs_etm__dump_event(struct cs_etm_auxtrace *etm,
108 struct auxtrace_buffer *buffer)
111 const char *color = PERF_COLOR_BLUE;
112 struct cs_etm_decoder_params d_params;
113 struct cs_etm_trace_params *t_params;
114 struct cs_etm_decoder *decoder;
115 size_t buffer_used = 0;
117 fprintf(stdout, "\n");
118 color_fprintf(stdout, color,
119 ". ... CoreSight ETM Trace data: size %zu bytes\n",
122 /* Use metadata to fill in trace parameters for trace decoder */
123 t_params = zalloc(sizeof(*t_params) * etm->num_cpu);
124 for (i = 0; i < etm->num_cpu; i++) {
125 t_params[i].protocol = CS_ETM_PROTO_ETMV4i;
126 t_params[i].etmv4.reg_idr0 = etm->metadata[i][CS_ETMV4_TRCIDR0];
127 t_params[i].etmv4.reg_idr1 = etm->metadata[i][CS_ETMV4_TRCIDR1];
128 t_params[i].etmv4.reg_idr2 = etm->metadata[i][CS_ETMV4_TRCIDR2];
129 t_params[i].etmv4.reg_idr8 = etm->metadata[i][CS_ETMV4_TRCIDR8];
130 t_params[i].etmv4.reg_configr =
131 etm->metadata[i][CS_ETMV4_TRCCONFIGR];
132 t_params[i].etmv4.reg_traceidr =
133 etm->metadata[i][CS_ETMV4_TRCTRACEIDR];
136 /* Set decoder parameters to simply print the trace packets */
137 d_params.packet_printer = cs_etm__packet_dump;
138 d_params.operation = CS_ETM_OPERATION_PRINT;
139 d_params.formatted = true;
140 d_params.fsyncs = false;
141 d_params.hsyncs = false;
142 d_params.frame_aligned = true;
144 decoder = cs_etm_decoder__new(etm->num_cpu, &d_params, t_params);
153 ret = cs_etm_decoder__process_data_block(
154 decoder, buffer->offset,
155 &((u8 *)buffer->data)[buffer_used],
156 buffer->size - buffer_used, &consumed);
160 buffer_used += consumed;
161 } while (buffer_used < buffer->size);
163 cs_etm_decoder__free(decoder);
166 static int cs_etm__flush_events(struct perf_session *session,
167 struct perf_tool *tool)
170 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
171 struct cs_etm_auxtrace,
176 if (!tool->ordered_events)
179 if (!etm->timeless_decoding)
182 ret = cs_etm__update_queues(etm);
187 return cs_etm__process_timeless_queues(etm, -1, MAX_TIMESTAMP - 1);
190 static void cs_etm__free_queue(void *priv)
192 struct cs_etm_queue *etmq = priv;
197 thread__zput(etmq->thread);
198 cs_etm_decoder__free(etmq->decoder);
199 zfree(&etmq->event_buf);
200 zfree(&etmq->last_branch);
201 zfree(&etmq->last_branch_rb);
202 zfree(&etmq->prev_packet);
203 zfree(&etmq->packet);
207 static void cs_etm__free_events(struct perf_session *session)
210 struct cs_etm_auxtrace *aux = container_of(session->auxtrace,
211 struct cs_etm_auxtrace,
213 struct auxtrace_queues *queues = &aux->queues;
215 for (i = 0; i < queues->nr_queues; i++) {
216 cs_etm__free_queue(queues->queue_array[i].priv);
217 queues->queue_array[i].priv = NULL;
220 auxtrace_queues__free(queues);
223 static void cs_etm__free(struct perf_session *session)
226 struct int_node *inode, *tmp;
227 struct cs_etm_auxtrace *aux = container_of(session->auxtrace,
228 struct cs_etm_auxtrace,
230 cs_etm__free_events(session);
231 session->auxtrace = NULL;
233 /* First remove all traceID/CPU# nodes for the RB tree */
234 intlist__for_each_entry_safe(inode, tmp, traceid_list)
235 intlist__remove(traceid_list, inode);
236 /* Then the RB tree itself */
237 intlist__delete(traceid_list);
239 for (i = 0; i < aux->num_cpu; i++)
240 zfree(&aux->metadata[i]);
242 thread__zput(aux->unknown_thread);
243 zfree(&aux->metadata);
247 static u32 cs_etm__mem_access(struct cs_etm_queue *etmq, u64 address,
248 size_t size, u8 *buffer)
253 struct thread *thread;
254 struct machine *machine;
255 struct addr_location al;
260 machine = etmq->etm->machine;
261 if (address >= etmq->etm->kernel_start)
262 cpumode = PERF_RECORD_MISC_KERNEL;
264 cpumode = PERF_RECORD_MISC_USER;
266 thread = etmq->thread;
268 if (cpumode != PERF_RECORD_MISC_KERNEL)
270 thread = etmq->etm->unknown_thread;
273 if (!thread__find_map(thread, cpumode, address, &al) || !al.map->dso)
276 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
277 dso__data_status_seen(al.map->dso, DSO_DATA_STATUS_SEEN_ITRACE))
280 offset = al.map->map_ip(al.map, address);
284 len = dso__data_read_offset(al.map->dso, machine, offset, buffer, size);
292 static struct cs_etm_queue *cs_etm__alloc_queue(struct cs_etm_auxtrace *etm,
293 unsigned int queue_nr)
296 struct cs_etm_decoder_params d_params;
297 struct cs_etm_trace_params *t_params;
298 struct cs_etm_queue *etmq;
299 size_t szp = sizeof(struct cs_etm_packet);
301 etmq = zalloc(sizeof(*etmq));
305 etmq->packet = zalloc(szp);
309 if (etm->synth_opts.last_branch || etm->sample_branches) {
310 etmq->prev_packet = zalloc(szp);
311 if (!etmq->prev_packet)
315 if (etm->synth_opts.last_branch) {
316 size_t sz = sizeof(struct branch_stack);
318 sz += etm->synth_opts.last_branch_sz *
319 sizeof(struct branch_entry);
320 etmq->last_branch = zalloc(sz);
321 if (!etmq->last_branch)
323 etmq->last_branch_rb = zalloc(sz);
324 if (!etmq->last_branch_rb)
328 etmq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
329 if (!etmq->event_buf)
333 etmq->queue_nr = queue_nr;
338 /* Use metadata to fill in trace parameters for trace decoder */
339 t_params = zalloc(sizeof(*t_params) * etm->num_cpu);
344 for (i = 0; i < etm->num_cpu; i++) {
345 t_params[i].protocol = CS_ETM_PROTO_ETMV4i;
346 t_params[i].etmv4.reg_idr0 = etm->metadata[i][CS_ETMV4_TRCIDR0];
347 t_params[i].etmv4.reg_idr1 = etm->metadata[i][CS_ETMV4_TRCIDR1];
348 t_params[i].etmv4.reg_idr2 = etm->metadata[i][CS_ETMV4_TRCIDR2];
349 t_params[i].etmv4.reg_idr8 = etm->metadata[i][CS_ETMV4_TRCIDR8];
350 t_params[i].etmv4.reg_configr =
351 etm->metadata[i][CS_ETMV4_TRCCONFIGR];
352 t_params[i].etmv4.reg_traceidr =
353 etm->metadata[i][CS_ETMV4_TRCTRACEIDR];
356 /* Set decoder parameters to simply print the trace packets */
357 d_params.packet_printer = cs_etm__packet_dump;
358 d_params.operation = CS_ETM_OPERATION_DECODE;
359 d_params.formatted = true;
360 d_params.fsyncs = false;
361 d_params.hsyncs = false;
362 d_params.frame_aligned = true;
363 d_params.data = etmq;
365 etmq->decoder = cs_etm_decoder__new(etm->num_cpu, &d_params, t_params);
373 * Register a function to handle all memory accesses required by
374 * the trace decoder library.
376 if (cs_etm_decoder__add_mem_access_cb(etmq->decoder,
379 goto out_free_decoder;
382 etmq->period_instructions = 0;
387 cs_etm_decoder__free(etmq->decoder);
389 zfree(&etmq->event_buf);
390 zfree(&etmq->last_branch);
391 zfree(&etmq->last_branch_rb);
392 zfree(&etmq->prev_packet);
393 zfree(&etmq->packet);
399 static int cs_etm__setup_queue(struct cs_etm_auxtrace *etm,
400 struct auxtrace_queue *queue,
401 unsigned int queue_nr)
403 struct cs_etm_queue *etmq = queue->priv;
405 if (list_empty(&queue->head) || etmq)
408 etmq = cs_etm__alloc_queue(etm, queue_nr);
415 if (queue->cpu != -1)
416 etmq->cpu = queue->cpu;
418 etmq->tid = queue->tid;
423 static int cs_etm__setup_queues(struct cs_etm_auxtrace *etm)
428 for (i = 0; i < etm->queues.nr_queues; i++) {
429 ret = cs_etm__setup_queue(etm, &etm->queues.queue_array[i], i);
437 static int cs_etm__update_queues(struct cs_etm_auxtrace *etm)
439 if (etm->queues.new_data) {
440 etm->queues.new_data = false;
441 return cs_etm__setup_queues(etm);
447 static inline void cs_etm__copy_last_branch_rb(struct cs_etm_queue *etmq)
449 struct branch_stack *bs_src = etmq->last_branch_rb;
450 struct branch_stack *bs_dst = etmq->last_branch;
454 * Set the number of records before early exit: ->nr is used to
455 * determine how many branches to copy from ->entries.
457 bs_dst->nr = bs_src->nr;
460 * Early exit when there is nothing to copy.
466 * As bs_src->entries is a circular buffer, we need to copy from it in
467 * two steps. First, copy the branches from the most recently inserted
468 * branch ->last_branch_pos until the end of bs_src->entries buffer.
470 nr = etmq->etm->synth_opts.last_branch_sz - etmq->last_branch_pos;
471 memcpy(&bs_dst->entries[0],
472 &bs_src->entries[etmq->last_branch_pos],
473 sizeof(struct branch_entry) * nr);
476 * If we wrapped around at least once, the branches from the beginning
477 * of the bs_src->entries buffer and until the ->last_branch_pos element
478 * are older valid branches: copy them over. The total number of
479 * branches copied over will be equal to the number of branches asked by
480 * the user in last_branch_sz.
482 if (bs_src->nr >= etmq->etm->synth_opts.last_branch_sz) {
483 memcpy(&bs_dst->entries[nr],
485 sizeof(struct branch_entry) * etmq->last_branch_pos);
489 static inline void cs_etm__reset_last_branch_rb(struct cs_etm_queue *etmq)
491 etmq->last_branch_pos = 0;
492 etmq->last_branch_rb->nr = 0;
495 static inline u64 cs_etm__last_executed_instr(struct cs_etm_packet *packet)
498 * The packet records the execution range with an exclusive end address
500 * A64 instructions are constant size, so the last executed
501 * instruction is A64_INSTR_SIZE before the end address
502 * Will need to do instruction level decode for T32 instructions as
503 * they can be variable size (not yet supported).
505 return packet->end_addr - A64_INSTR_SIZE;
508 static inline u64 cs_etm__instr_count(const struct cs_etm_packet *packet)
511 * Only A64 instructions are currently supported, so can get
512 * instruction count by dividing.
513 * Will need to do instruction level decode for T32 instructions as
514 * they can be variable size (not yet supported).
516 return (packet->end_addr - packet->start_addr) / A64_INSTR_SIZE;
519 static inline u64 cs_etm__instr_addr(const struct cs_etm_packet *packet,
523 * Only A64 instructions are currently supported, so can get
524 * instruction address by muliplying.
525 * Will need to do instruction level decode for T32 instructions as
526 * they can be variable size (not yet supported).
528 return packet->start_addr + offset * A64_INSTR_SIZE;
531 static void cs_etm__update_last_branch_rb(struct cs_etm_queue *etmq)
533 struct branch_stack *bs = etmq->last_branch_rb;
534 struct branch_entry *be;
537 * The branches are recorded in a circular buffer in reverse
538 * chronological order: we start recording from the last element of the
539 * buffer down. After writing the first element of the stack, move the
540 * insert position back to the end of the buffer.
542 if (!etmq->last_branch_pos)
543 etmq->last_branch_pos = etmq->etm->synth_opts.last_branch_sz;
545 etmq->last_branch_pos -= 1;
547 be = &bs->entries[etmq->last_branch_pos];
548 be->from = cs_etm__last_executed_instr(etmq->prev_packet);
549 be->to = etmq->packet->start_addr;
550 /* No support for mispredict */
551 be->flags.mispred = 0;
552 be->flags.predicted = 1;
555 * Increment bs->nr until reaching the number of last branches asked by
556 * the user on the command line.
558 if (bs->nr < etmq->etm->synth_opts.last_branch_sz)
562 static int cs_etm__inject_event(union perf_event *event,
563 struct perf_sample *sample, u64 type)
565 event->header.size = perf_event__sample_event_size(sample, type, 0);
566 return perf_event__synthesize_sample(event, type, 0, sample);
571 cs_etm__get_trace(struct cs_etm_buffer *buff, struct cs_etm_queue *etmq)
573 struct auxtrace_buffer *aux_buffer = etmq->buffer;
574 struct auxtrace_buffer *old_buffer = aux_buffer;
575 struct auxtrace_queue *queue;
577 queue = &etmq->etm->queues.queue_array[etmq->queue_nr];
579 aux_buffer = auxtrace_buffer__next(queue, aux_buffer);
581 /* If no more data, drop the previous auxtrace_buffer and return */
584 auxtrace_buffer__drop_data(old_buffer);
589 etmq->buffer = aux_buffer;
591 /* If the aux_buffer doesn't have data associated, try to load it */
592 if (!aux_buffer->data) {
593 /* get the file desc associated with the perf data file */
594 int fd = perf_data__fd(etmq->etm->session->data);
596 aux_buffer->data = auxtrace_buffer__get_data(aux_buffer, fd);
597 if (!aux_buffer->data)
601 /* If valid, drop the previous buffer */
603 auxtrace_buffer__drop_data(old_buffer);
605 buff->offset = aux_buffer->offset;
606 buff->len = aux_buffer->size;
607 buff->buf = aux_buffer->data;
609 buff->ref_timestamp = aux_buffer->reference;
614 static void cs_etm__set_pid_tid_cpu(struct cs_etm_auxtrace *etm,
615 struct auxtrace_queue *queue)
617 struct cs_etm_queue *etmq = queue->priv;
619 /* CPU-wide tracing isn't supported yet */
620 if (queue->tid == -1)
623 if ((!etmq->thread) && (etmq->tid != -1))
624 etmq->thread = machine__find_thread(etm->machine, -1,
628 etmq->pid = etmq->thread->pid_;
629 if (queue->cpu == -1)
630 etmq->cpu = etmq->thread->cpu;
634 static int cs_etm__synth_instruction_sample(struct cs_etm_queue *etmq,
635 u64 addr, u64 period)
638 struct cs_etm_auxtrace *etm = etmq->etm;
639 union perf_event *event = etmq->event_buf;
640 struct perf_sample sample = {.ip = 0,};
642 event->sample.header.type = PERF_RECORD_SAMPLE;
643 event->sample.header.misc = PERF_RECORD_MISC_USER;
644 event->sample.header.size = sizeof(struct perf_event_header);
647 sample.pid = etmq->pid;
648 sample.tid = etmq->tid;
649 sample.id = etmq->etm->instructions_id;
650 sample.stream_id = etmq->etm->instructions_id;
651 sample.period = period;
652 sample.cpu = etmq->packet->cpu;
655 sample.cpumode = event->header.misc;
657 if (etm->synth_opts.last_branch) {
658 cs_etm__copy_last_branch_rb(etmq);
659 sample.branch_stack = etmq->last_branch;
662 if (etm->synth_opts.inject) {
663 ret = cs_etm__inject_event(event, &sample,
664 etm->instructions_sample_type);
669 ret = perf_session__deliver_synth_event(etm->session, event, &sample);
673 "CS ETM Trace: failed to deliver instruction event, error %d\n",
676 if (etm->synth_opts.last_branch)
677 cs_etm__reset_last_branch_rb(etmq);
683 * The cs etm packet encodes an instruction range between a branch target
684 * and the next taken branch. Generate sample accordingly.
686 static int cs_etm__synth_branch_sample(struct cs_etm_queue *etmq)
689 struct cs_etm_auxtrace *etm = etmq->etm;
690 struct perf_sample sample = {.ip = 0,};
691 union perf_event *event = etmq->event_buf;
692 struct dummy_branch_stack {
694 struct branch_entry entries;
697 event->sample.header.type = PERF_RECORD_SAMPLE;
698 event->sample.header.misc = PERF_RECORD_MISC_USER;
699 event->sample.header.size = sizeof(struct perf_event_header);
701 sample.ip = cs_etm__last_executed_instr(etmq->prev_packet);
702 sample.pid = etmq->pid;
703 sample.tid = etmq->tid;
704 sample.addr = etmq->packet->start_addr;
705 sample.id = etmq->etm->branches_id;
706 sample.stream_id = etmq->etm->branches_id;
708 sample.cpu = etmq->packet->cpu;
710 sample.cpumode = PERF_RECORD_MISC_USER;
713 * perf report cannot handle events without a branch stack
715 if (etm->synth_opts.last_branch) {
716 dummy_bs = (struct dummy_branch_stack){
723 sample.branch_stack = (struct branch_stack *)&dummy_bs;
726 if (etm->synth_opts.inject) {
727 ret = cs_etm__inject_event(event, &sample,
728 etm->branches_sample_type);
733 ret = perf_session__deliver_synth_event(etm->session, event, &sample);
737 "CS ETM Trace: failed to deliver instruction event, error %d\n",
743 struct cs_etm_synth {
744 struct perf_tool dummy_tool;
745 struct perf_session *session;
748 static int cs_etm__event_synth(struct perf_tool *tool,
749 union perf_event *event,
750 struct perf_sample *sample __maybe_unused,
751 struct machine *machine __maybe_unused)
753 struct cs_etm_synth *cs_etm_synth =
754 container_of(tool, struct cs_etm_synth, dummy_tool);
756 return perf_session__deliver_synth_event(cs_etm_synth->session,
760 static int cs_etm__synth_event(struct perf_session *session,
761 struct perf_event_attr *attr, u64 id)
763 struct cs_etm_synth cs_etm_synth;
765 memset(&cs_etm_synth, 0, sizeof(struct cs_etm_synth));
766 cs_etm_synth.session = session;
768 return perf_event__synthesize_attr(&cs_etm_synth.dummy_tool, attr, 1,
769 &id, cs_etm__event_synth);
772 static int cs_etm__synth_events(struct cs_etm_auxtrace *etm,
773 struct perf_session *session)
775 struct perf_evlist *evlist = session->evlist;
776 struct perf_evsel *evsel;
777 struct perf_event_attr attr;
782 evlist__for_each_entry(evlist, evsel) {
783 if (evsel->attr.type == etm->pmu_type) {
790 pr_debug("No selected events with CoreSight Trace data\n");
794 memset(&attr, 0, sizeof(struct perf_event_attr));
795 attr.size = sizeof(struct perf_event_attr);
796 attr.type = PERF_TYPE_HARDWARE;
797 attr.sample_type = evsel->attr.sample_type & PERF_SAMPLE_MASK;
798 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
800 if (etm->timeless_decoding)
801 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
803 attr.sample_type |= PERF_SAMPLE_TIME;
805 attr.exclude_user = evsel->attr.exclude_user;
806 attr.exclude_kernel = evsel->attr.exclude_kernel;
807 attr.exclude_hv = evsel->attr.exclude_hv;
808 attr.exclude_host = evsel->attr.exclude_host;
809 attr.exclude_guest = evsel->attr.exclude_guest;
810 attr.sample_id_all = evsel->attr.sample_id_all;
811 attr.read_format = evsel->attr.read_format;
813 /* create new id val to be a fixed offset from evsel id */
814 id = evsel->id[0] + 1000000000;
819 if (etm->synth_opts.branches) {
820 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
821 attr.sample_period = 1;
822 attr.sample_type |= PERF_SAMPLE_ADDR;
823 err = cs_etm__synth_event(session, &attr, id);
826 etm->sample_branches = true;
827 etm->branches_sample_type = attr.sample_type;
828 etm->branches_id = id;
830 attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR;
833 if (etm->synth_opts.last_branch)
834 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
836 if (etm->synth_opts.instructions) {
837 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
838 attr.sample_period = etm->synth_opts.period;
839 etm->instructions_sample_period = attr.sample_period;
840 err = cs_etm__synth_event(session, &attr, id);
843 etm->sample_instructions = true;
844 etm->instructions_sample_type = attr.sample_type;
845 etm->instructions_id = id;
852 static int cs_etm__sample(struct cs_etm_queue *etmq)
854 struct cs_etm_auxtrace *etm = etmq->etm;
855 struct cs_etm_packet *tmp;
859 instrs_executed = cs_etm__instr_count(etmq->packet);
860 etmq->period_instructions += instrs_executed;
863 * Record a branch when the last instruction in
864 * PREV_PACKET is a branch.
866 if (etm->synth_opts.last_branch &&
868 etmq->prev_packet->sample_type == CS_ETM_RANGE &&
869 etmq->prev_packet->last_instr_taken_branch)
870 cs_etm__update_last_branch_rb(etmq);
872 if (etm->sample_instructions &&
873 etmq->period_instructions >= etm->instructions_sample_period) {
875 * Emit instruction sample periodically
876 * TODO: allow period to be defined in cycles and clock time
879 /* Get number of instructions executed after the sample point */
880 u64 instrs_over = etmq->period_instructions -
881 etm->instructions_sample_period;
884 * Calculate the address of the sampled instruction (-1 as
885 * sample is reported as though instruction has just been
886 * executed, but PC has not advanced to next instruction)
888 u64 offset = (instrs_executed - instrs_over - 1);
889 u64 addr = cs_etm__instr_addr(etmq->packet, offset);
891 ret = cs_etm__synth_instruction_sample(
892 etmq, addr, etm->instructions_sample_period);
896 /* Carry remaining instructions into next sample period */
897 etmq->period_instructions = instrs_over;
900 if (etm->sample_branches &&
902 etmq->prev_packet->sample_type == CS_ETM_RANGE &&
903 etmq->prev_packet->last_instr_taken_branch) {
904 ret = cs_etm__synth_branch_sample(etmq);
909 if (etm->sample_branches || etm->synth_opts.last_branch) {
911 * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for
912 * the next incoming packet.
915 etmq->packet = etmq->prev_packet;
916 etmq->prev_packet = tmp;
922 static int cs_etm__flush(struct cs_etm_queue *etmq)
925 struct cs_etm_packet *tmp;
927 if (etmq->etm->synth_opts.last_branch &&
929 etmq->prev_packet->sample_type == CS_ETM_RANGE) {
931 * Generate a last branch event for the branches left in the
932 * circular buffer at the end of the trace.
934 * Use the address of the end of the last reported execution
937 u64 addr = cs_etm__last_executed_instr(etmq->prev_packet);
939 err = cs_etm__synth_instruction_sample(
941 etmq->period_instructions);
942 etmq->period_instructions = 0;
945 * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for
946 * the next incoming packet.
949 etmq->packet = etmq->prev_packet;
950 etmq->prev_packet = tmp;
956 static int cs_etm__run_decoder(struct cs_etm_queue *etmq)
958 struct cs_etm_auxtrace *etm = etmq->etm;
959 struct cs_etm_buffer buffer;
960 size_t buffer_used, processed;
963 if (!etm->kernel_start)
964 etm->kernel_start = machine__kernel_start(etm->machine);
966 /* Go through each buffer in the queue and decode them one by one */
969 memset(&buffer, 0, sizeof(buffer));
970 err = cs_etm__get_trace(&buffer, etmq);
974 * We cannot assume consecutive blocks in the data file are
975 * contiguous, reset the decoder to force re-sync.
977 err = cs_etm_decoder__reset(etmq->decoder);
981 /* Run trace decoder until buffer consumed or end of trace */
984 err = cs_etm_decoder__process_data_block(
987 &buffer.buf[buffer_used],
988 buffer.len - buffer_used,
993 etmq->offset += processed;
994 buffer_used += processed;
996 /* Process each packet in this chunk */
998 err = cs_etm_decoder__get_packet(etmq->decoder,
1002 * Stop processing this chunk on
1003 * end of data or error
1007 switch (etmq->packet->sample_type) {
1010 * If the packet contains an instruction
1011 * range, generate instruction sequence
1014 cs_etm__sample(etmq);
1016 case CS_ETM_TRACE_ON:
1018 * Discontinuity in trace, flush
1019 * previous branch stack
1021 cs_etm__flush(etmq);
1027 } while (buffer.len > buffer_used);
1030 /* Flush any remaining branch stack entries */
1031 err = cs_etm__flush(etmq);
1037 static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm,
1038 pid_t tid, u64 time_)
1041 struct auxtrace_queues *queues = &etm->queues;
1043 for (i = 0; i < queues->nr_queues; i++) {
1044 struct auxtrace_queue *queue = &etm->queues.queue_array[i];
1045 struct cs_etm_queue *etmq = queue->priv;
1047 if (etmq && ((tid == -1) || (etmq->tid == tid))) {
1049 cs_etm__set_pid_tid_cpu(etm, queue);
1050 cs_etm__run_decoder(etmq);
1057 static int cs_etm__process_event(struct perf_session *session,
1058 union perf_event *event,
1059 struct perf_sample *sample,
1060 struct perf_tool *tool)
1064 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
1065 struct cs_etm_auxtrace,
1071 if (!tool->ordered_events) {
1072 pr_err("CoreSight ETM Trace requires ordered events\n");
1076 if (!etm->timeless_decoding)
1079 if (sample->time && (sample->time != (u64) -1))
1080 timestamp = sample->time;
1084 if (timestamp || etm->timeless_decoding) {
1085 err = cs_etm__update_queues(etm);
1090 if (event->header.type == PERF_RECORD_EXIT)
1091 return cs_etm__process_timeless_queues(etm,
1098 static int cs_etm__process_auxtrace_event(struct perf_session *session,
1099 union perf_event *event,
1100 struct perf_tool *tool __maybe_unused)
1102 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
1103 struct cs_etm_auxtrace,
1105 if (!etm->data_queued) {
1106 struct auxtrace_buffer *buffer;
1108 int fd = perf_data__fd(session->data);
1109 bool is_pipe = perf_data__is_pipe(session->data);
1115 data_offset = lseek(fd, 0, SEEK_CUR);
1116 if (data_offset == -1)
1120 err = auxtrace_queues__add_event(&etm->queues, session,
1121 event, data_offset, &buffer);
1126 if (auxtrace_buffer__get_data(buffer, fd)) {
1127 cs_etm__dump_event(etm, buffer);
1128 auxtrace_buffer__put_data(buffer);
1135 static bool cs_etm__is_timeless_decoding(struct cs_etm_auxtrace *etm)
1137 struct perf_evsel *evsel;
1138 struct perf_evlist *evlist = etm->session->evlist;
1139 bool timeless_decoding = true;
1142 * Circle through the list of event and complain if we find one
1143 * with the time bit set.
1145 evlist__for_each_entry(evlist, evsel) {
1146 if ((evsel->attr.sample_type & PERF_SAMPLE_TIME))
1147 timeless_decoding = false;
1150 return timeless_decoding;
1153 static const char * const cs_etm_global_header_fmts[] = {
1154 [CS_HEADER_VERSION_0] = " Header version %llx\n",
1155 [CS_PMU_TYPE_CPUS] = " PMU type/num cpus %llx\n",
1156 [CS_ETM_SNAPSHOT] = " Snapshot %llx\n",
1159 static const char * const cs_etm_priv_fmts[] = {
1160 [CS_ETM_MAGIC] = " Magic number %llx\n",
1161 [CS_ETM_CPU] = " CPU %lld\n",
1162 [CS_ETM_ETMCR] = " ETMCR %llx\n",
1163 [CS_ETM_ETMTRACEIDR] = " ETMTRACEIDR %llx\n",
1164 [CS_ETM_ETMCCER] = " ETMCCER %llx\n",
1165 [CS_ETM_ETMIDR] = " ETMIDR %llx\n",
1168 static const char * const cs_etmv4_priv_fmts[] = {
1169 [CS_ETM_MAGIC] = " Magic number %llx\n",
1170 [CS_ETM_CPU] = " CPU %lld\n",
1171 [CS_ETMV4_TRCCONFIGR] = " TRCCONFIGR %llx\n",
1172 [CS_ETMV4_TRCTRACEIDR] = " TRCTRACEIDR %llx\n",
1173 [CS_ETMV4_TRCIDR0] = " TRCIDR0 %llx\n",
1174 [CS_ETMV4_TRCIDR1] = " TRCIDR1 %llx\n",
1175 [CS_ETMV4_TRCIDR2] = " TRCIDR2 %llx\n",
1176 [CS_ETMV4_TRCIDR8] = " TRCIDR8 %llx\n",
1177 [CS_ETMV4_TRCAUTHSTATUS] = " TRCAUTHSTATUS %llx\n",
1180 static void cs_etm__print_auxtrace_info(u64 *val, int num)
1184 for (i = 0; i < CS_HEADER_VERSION_0_MAX; i++)
1185 fprintf(stdout, cs_etm_global_header_fmts[i], val[i]);
1187 for (i = CS_HEADER_VERSION_0_MAX; cpu < num; cpu++) {
1188 if (val[i] == __perf_cs_etmv3_magic)
1189 for (j = 0; j < CS_ETM_PRIV_MAX; j++, i++)
1190 fprintf(stdout, cs_etm_priv_fmts[j], val[i]);
1191 else if (val[i] == __perf_cs_etmv4_magic)
1192 for (j = 0; j < CS_ETMV4_PRIV_MAX; j++, i++)
1193 fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]);
1195 /* failure.. return */
1200 int cs_etm__process_auxtrace_info(union perf_event *event,
1201 struct perf_session *session)
1203 struct auxtrace_info_event *auxtrace_info = &event->auxtrace_info;
1204 struct cs_etm_auxtrace *etm = NULL;
1205 struct int_node *inode;
1206 unsigned int pmu_type;
1207 int event_header_size = sizeof(struct perf_event_header);
1208 int info_header_size;
1209 int total_size = auxtrace_info->header.size;
1212 int err = 0, idx = -1;
1214 u64 *ptr, *hdr = NULL;
1215 u64 **metadata = NULL;
1218 * sizeof(auxtrace_info_event::type) +
1219 * sizeof(auxtrace_info_event::reserved) == 8
1221 info_header_size = 8;
1223 if (total_size < (event_header_size + info_header_size))
1226 priv_size = total_size - event_header_size - info_header_size;
1228 /* First the global part */
1229 ptr = (u64 *) auxtrace_info->priv;
1231 /* Look for version '0' of the header */
1235 hdr = zalloc(sizeof(*hdr) * CS_HEADER_VERSION_0_MAX);
1239 /* Extract header information - see cs-etm.h for format */
1240 for (i = 0; i < CS_HEADER_VERSION_0_MAX; i++)
1242 num_cpu = hdr[CS_PMU_TYPE_CPUS] & 0xffffffff;
1243 pmu_type = (unsigned int) ((hdr[CS_PMU_TYPE_CPUS] >> 32) &
1247 * Create an RB tree for traceID-CPU# tuple. Since the conversion has
1248 * to be made for each packet that gets decoded, optimizing access in
1249 * anything other than a sequential array is worth doing.
1251 traceid_list = intlist__new(NULL);
1252 if (!traceid_list) {
1257 metadata = zalloc(sizeof(*metadata) * num_cpu);
1260 goto err_free_traceid_list;
1264 * The metadata is stored in the auxtrace_info section and encodes
1265 * the configuration of the ARM embedded trace macrocell which is
1266 * required by the trace decoder to properly decode the trace due
1267 * to its highly compressed nature.
1269 for (j = 0; j < num_cpu; j++) {
1270 if (ptr[i] == __perf_cs_etmv3_magic) {
1271 metadata[j] = zalloc(sizeof(*metadata[j]) *
1275 goto err_free_metadata;
1277 for (k = 0; k < CS_ETM_PRIV_MAX; k++)
1278 metadata[j][k] = ptr[i + k];
1280 /* The traceID is our handle */
1281 idx = metadata[j][CS_ETM_ETMTRACEIDR];
1282 i += CS_ETM_PRIV_MAX;
1283 } else if (ptr[i] == __perf_cs_etmv4_magic) {
1284 metadata[j] = zalloc(sizeof(*metadata[j]) *
1288 goto err_free_metadata;
1290 for (k = 0; k < CS_ETMV4_PRIV_MAX; k++)
1291 metadata[j][k] = ptr[i + k];
1293 /* The traceID is our handle */
1294 idx = metadata[j][CS_ETMV4_TRCTRACEIDR];
1295 i += CS_ETMV4_PRIV_MAX;
1298 /* Get an RB node for this CPU */
1299 inode = intlist__findnew(traceid_list, idx);
1301 /* Something went wrong, no need to continue */
1303 err = PTR_ERR(inode);
1304 goto err_free_metadata;
1308 * The node for that CPU should not be taken.
1309 * Back out if that's the case.
1313 goto err_free_metadata;
1315 /* All good, associate the traceID with the CPU# */
1316 inode->priv = &metadata[j][CS_ETM_CPU];
1320 * Each of CS_HEADER_VERSION_0_MAX, CS_ETM_PRIV_MAX and
1321 * CS_ETMV4_PRIV_MAX mark how many double words are in the
1322 * global metadata, and each cpu's metadata respectively.
1323 * The following tests if the correct number of double words was
1324 * present in the auxtrace info section.
1326 if (i * 8 != priv_size) {
1328 goto err_free_metadata;
1331 etm = zalloc(sizeof(*etm));
1335 goto err_free_metadata;
1338 err = auxtrace_queues__init(&etm->queues);
1342 etm->session = session;
1343 etm->machine = &session->machines.host;
1345 etm->num_cpu = num_cpu;
1346 etm->pmu_type = pmu_type;
1347 etm->snapshot_mode = (hdr[CS_ETM_SNAPSHOT] != 0);
1348 etm->metadata = metadata;
1349 etm->auxtrace_type = auxtrace_info->type;
1350 etm->timeless_decoding = cs_etm__is_timeless_decoding(etm);
1352 etm->auxtrace.process_event = cs_etm__process_event;
1353 etm->auxtrace.process_auxtrace_event = cs_etm__process_auxtrace_event;
1354 etm->auxtrace.flush_events = cs_etm__flush_events;
1355 etm->auxtrace.free_events = cs_etm__free_events;
1356 etm->auxtrace.free = cs_etm__free;
1357 session->auxtrace = &etm->auxtrace;
1359 etm->unknown_thread = thread__new(999999999, 999999999);
1360 if (!etm->unknown_thread)
1361 goto err_free_queues;
1364 * Initialize list node so that at thread__zput() we can avoid
1365 * segmentation fault at list_del_init().
1367 INIT_LIST_HEAD(&etm->unknown_thread->node);
1369 err = thread__set_comm(etm->unknown_thread, "unknown", 0);
1371 goto err_delete_thread;
1373 if (thread__init_map_groups(etm->unknown_thread, etm->machine))
1374 goto err_delete_thread;
1377 cs_etm__print_auxtrace_info(auxtrace_info->priv, num_cpu);
1381 if (session->itrace_synth_opts && session->itrace_synth_opts->set) {
1382 etm->synth_opts = *session->itrace_synth_opts;
1384 itrace_synth_opts__set_default(&etm->synth_opts);
1385 etm->synth_opts.callchain = false;
1388 err = cs_etm__synth_events(etm, session);
1390 goto err_delete_thread;
1392 err = auxtrace_queues__process_index(&etm->queues, session);
1394 goto err_delete_thread;
1396 etm->data_queued = etm->queues.populated;
1401 thread__zput(etm->unknown_thread);
1403 auxtrace_queues__free(&etm->queues);
1404 session->auxtrace = NULL;
1408 /* No need to check @metadata[j], free(NULL) is supported */
1409 for (j = 0; j < num_cpu; j++)
1412 err_free_traceid_list:
1413 intlist__delete(traceid_list);