Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright(C) 2015-2018 Linaro Limited.
4 *
5 * Author: Tor Jeremiassen <tor@ti.com>
6 * Author: Mathieu Poirier <mathieu.poirier@linaro.org>
7 */
8
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>
14
15#include <stdlib.h>
16
17#include "auxtrace.h"
18#include "color.h"
19#include "cs-etm.h"
20#include "cs-etm-decoder/cs-etm-decoder.h"
21#include "debug.h"
22#include "evlist.h"
23#include "intlist.h"
24#include "machine.h"
25#include "map.h"
26#include "perf.h"
27#include "thread.h"
28#include "thread_map.h"
29#include "thread-stack.h"
30#include "util.h"
31
32#define MAX_TIMESTAMP (~0ULL)
33
34/*
35 * A64 instructions are always 4 bytes
36 *
37 * Only A64 is supported, so can use this constant for converting between
38 * addresses and instruction counts, calculting offsets etc
39 */
40#define A64_INSTR_SIZE 4
41
42struct 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;
50
51 u8 timeless_decoding;
52 u8 snapshot_mode;
53 u8 data_queued;
54 u8 sample_branches;
55 u8 sample_instructions;
56
57 int num_cpu;
58 u32 auxtrace_type;
59 u64 branches_sample_type;
60 u64 branches_id;
61 u64 instructions_sample_type;
62 u64 instructions_sample_period;
63 u64 instructions_id;
64 u64 **metadata;
65 u64 kernel_start;
66 unsigned int pmu_type;
67};
68
69struct cs_etm_queue {
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;
77 pid_t pid, tid;
78 int cpu;
79 u64 time;
80 u64 timestamp;
81 u64 offset;
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;
88};
89
90static int cs_etm__update_queues(struct cs_etm_auxtrace *etm);
91static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm,
92 pid_t tid, u64 time_);
93
94static void cs_etm__packet_dump(const char *pkt_string)
95{
96 const char *color = PERF_COLOR_BLUE;
97 int len = strlen(pkt_string);
98
99 if (len && (pkt_string[len-1] == '\n'))
100 color_fprintf(stdout, color, " %s", pkt_string);
101 else
102 color_fprintf(stdout, color, " %s\n", pkt_string);
103
104 fflush(stdout);
105}
106
107static void cs_etm__dump_event(struct cs_etm_auxtrace *etm,
108 struct auxtrace_buffer *buffer)
109{
110 int i, ret;
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;
116
117 fprintf(stdout, "\n");
118 color_fprintf(stdout, color,
119 ". ... CoreSight ETM Trace data: size %zu bytes\n",
120 buffer->size);
121
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];
134 }
135
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;
143
144 decoder = cs_etm_decoder__new(etm->num_cpu, &d_params, t_params);
145
146 zfree(&t_params);
147
148 if (!decoder)
149 return;
150 do {
151 size_t consumed;
152
153 ret = cs_etm_decoder__process_data_block(
154 decoder, buffer->offset,
155 &((u8 *)buffer->data)[buffer_used],
156 buffer->size - buffer_used, &consumed);
157 if (ret)
158 break;
159
160 buffer_used += consumed;
161 } while (buffer_used < buffer->size);
162
163 cs_etm_decoder__free(decoder);
164}
165
166static int cs_etm__flush_events(struct perf_session *session,
167 struct perf_tool *tool)
168{
169 int ret;
170 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
171 struct cs_etm_auxtrace,
172 auxtrace);
173 if (dump_trace)
174 return 0;
175
176 if (!tool->ordered_events)
177 return -EINVAL;
178
179 if (!etm->timeless_decoding)
180 return -EINVAL;
181
182 ret = cs_etm__update_queues(etm);
183
184 if (ret < 0)
185 return ret;
186
187 return cs_etm__process_timeless_queues(etm, -1, MAX_TIMESTAMP - 1);
188}
189
190static void cs_etm__free_queue(void *priv)
191{
192 struct cs_etm_queue *etmq = priv;
193
194 if (!etmq)
195 return;
196
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);
204 free(etmq);
205}
206
207static void cs_etm__free_events(struct perf_session *session)
208{
209 unsigned int i;
210 struct cs_etm_auxtrace *aux = container_of(session->auxtrace,
211 struct cs_etm_auxtrace,
212 auxtrace);
213 struct auxtrace_queues *queues = &aux->queues;
214
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;
218 }
219
220 auxtrace_queues__free(queues);
221}
222
223static void cs_etm__free(struct perf_session *session)
224{
225 int i;
226 struct int_node *inode, *tmp;
227 struct cs_etm_auxtrace *aux = container_of(session->auxtrace,
228 struct cs_etm_auxtrace,
229 auxtrace);
230 cs_etm__free_events(session);
231 session->auxtrace = NULL;
232
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);
238
239 for (i = 0; i < aux->num_cpu; i++)
240 zfree(&aux->metadata[i]);
241
242 thread__zput(aux->unknown_thread);
243 zfree(&aux->metadata);
244 zfree(&aux);
245}
246
247static u32 cs_etm__mem_access(struct cs_etm_queue *etmq, u64 address,
248 size_t size, u8 *buffer)
249{
250 u8 cpumode;
251 u64 offset;
252 int len;
253 struct thread *thread;
254 struct machine *machine;
255 struct addr_location al;
256
257 if (!etmq)
258 return -1;
259
260 machine = etmq->etm->machine;
261 if (address >= etmq->etm->kernel_start)
262 cpumode = PERF_RECORD_MISC_KERNEL;
263 else
264 cpumode = PERF_RECORD_MISC_USER;
265
266 thread = etmq->thread;
267 if (!thread) {
268 if (cpumode != PERF_RECORD_MISC_KERNEL)
269 return -EINVAL;
270 thread = etmq->etm->unknown_thread;
271 }
272
273 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, address, &al);
274
275 if (!al.map || !al.map->dso)
276 return 0;
277
278 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
279 dso__data_status_seen(al.map->dso, DSO_DATA_STATUS_SEEN_ITRACE))
280 return 0;
281
282 offset = al.map->map_ip(al.map, address);
283
284 map__load(al.map);
285
286 len = dso__data_read_offset(al.map->dso, machine, offset, buffer, size);
287
288 if (len <= 0)
289 return 0;
290
291 return len;
292}
293
294static struct cs_etm_queue *cs_etm__alloc_queue(struct cs_etm_auxtrace *etm,
295 unsigned int queue_nr)
296{
297 int i;
298 struct cs_etm_decoder_params d_params;
299 struct cs_etm_trace_params *t_params;
300 struct cs_etm_queue *etmq;
301 size_t szp = sizeof(struct cs_etm_packet);
302
303 etmq = zalloc(sizeof(*etmq));
304 if (!etmq)
305 return NULL;
306
307 etmq->packet = zalloc(szp);
308 if (!etmq->packet)
309 goto out_free;
310
311 if (etm->synth_opts.last_branch || etm->sample_branches) {
312 etmq->prev_packet = zalloc(szp);
313 if (!etmq->prev_packet)
314 goto out_free;
315 }
316
317 if (etm->synth_opts.last_branch) {
318 size_t sz = sizeof(struct branch_stack);
319
320 sz += etm->synth_opts.last_branch_sz *
321 sizeof(struct branch_entry);
322 etmq->last_branch = zalloc(sz);
323 if (!etmq->last_branch)
324 goto out_free;
325 etmq->last_branch_rb = zalloc(sz);
326 if (!etmq->last_branch_rb)
327 goto out_free;
328 }
329
330 etmq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
331 if (!etmq->event_buf)
332 goto out_free;
333
334 etmq->etm = etm;
335 etmq->queue_nr = queue_nr;
336 etmq->pid = -1;
337 etmq->tid = -1;
338 etmq->cpu = -1;
339
340 /* Use metadata to fill in trace parameters for trace decoder */
341 t_params = zalloc(sizeof(*t_params) * etm->num_cpu);
342
343 if (!t_params)
344 goto out_free;
345
346 for (i = 0; i < etm->num_cpu; i++) {
347 t_params[i].protocol = CS_ETM_PROTO_ETMV4i;
348 t_params[i].etmv4.reg_idr0 = etm->metadata[i][CS_ETMV4_TRCIDR0];
349 t_params[i].etmv4.reg_idr1 = etm->metadata[i][CS_ETMV4_TRCIDR1];
350 t_params[i].etmv4.reg_idr2 = etm->metadata[i][CS_ETMV4_TRCIDR2];
351 t_params[i].etmv4.reg_idr8 = etm->metadata[i][CS_ETMV4_TRCIDR8];
352 t_params[i].etmv4.reg_configr =
353 etm->metadata[i][CS_ETMV4_TRCCONFIGR];
354 t_params[i].etmv4.reg_traceidr =
355 etm->metadata[i][CS_ETMV4_TRCTRACEIDR];
356 }
357
358 /* Set decoder parameters to simply print the trace packets */
359 d_params.packet_printer = cs_etm__packet_dump;
360 d_params.operation = CS_ETM_OPERATION_DECODE;
361 d_params.formatted = true;
362 d_params.fsyncs = false;
363 d_params.hsyncs = false;
364 d_params.frame_aligned = true;
365 d_params.data = etmq;
366
367 etmq->decoder = cs_etm_decoder__new(etm->num_cpu, &d_params, t_params);
368
369 zfree(&t_params);
370
371 if (!etmq->decoder)
372 goto out_free;
373
374 /*
375 * Register a function to handle all memory accesses required by
376 * the trace decoder library.
377 */
378 if (cs_etm_decoder__add_mem_access_cb(etmq->decoder,
379 0x0L, ((u64) -1L),
380 cs_etm__mem_access))
381 goto out_free_decoder;
382
383 etmq->offset = 0;
384 etmq->period_instructions = 0;
385
386 return etmq;
387
388out_free_decoder:
389 cs_etm_decoder__free(etmq->decoder);
390out_free:
391 zfree(&etmq->event_buf);
392 zfree(&etmq->last_branch);
393 zfree(&etmq->last_branch_rb);
394 zfree(&etmq->prev_packet);
395 zfree(&etmq->packet);
396 free(etmq);
397
398 return NULL;
399}
400
401static int cs_etm__setup_queue(struct cs_etm_auxtrace *etm,
402 struct auxtrace_queue *queue,
403 unsigned int queue_nr)
404{
405 struct cs_etm_queue *etmq = queue->priv;
406
407 if (list_empty(&queue->head) || etmq)
408 return 0;
409
410 etmq = cs_etm__alloc_queue(etm, queue_nr);
411
412 if (!etmq)
413 return -ENOMEM;
414
415 queue->priv = etmq;
416
417 if (queue->cpu != -1)
418 etmq->cpu = queue->cpu;
419
420 etmq->tid = queue->tid;
421
422 return 0;
423}
424
425static int cs_etm__setup_queues(struct cs_etm_auxtrace *etm)
426{
427 unsigned int i;
428 int ret;
429
430 for (i = 0; i < etm->queues.nr_queues; i++) {
431 ret = cs_etm__setup_queue(etm, &etm->queues.queue_array[i], i);
432 if (ret)
433 return ret;
434 }
435
436 return 0;
437}
438
439static int cs_etm__update_queues(struct cs_etm_auxtrace *etm)
440{
441 if (etm->queues.new_data) {
442 etm->queues.new_data = false;
443 return cs_etm__setup_queues(etm);
444 }
445
446 return 0;
447}
448
449static inline void cs_etm__copy_last_branch_rb(struct cs_etm_queue *etmq)
450{
451 struct branch_stack *bs_src = etmq->last_branch_rb;
452 struct branch_stack *bs_dst = etmq->last_branch;
453 size_t nr = 0;
454
455 /*
456 * Set the number of records before early exit: ->nr is used to
457 * determine how many branches to copy from ->entries.
458 */
459 bs_dst->nr = bs_src->nr;
460
461 /*
462 * Early exit when there is nothing to copy.
463 */
464 if (!bs_src->nr)
465 return;
466
467 /*
468 * As bs_src->entries is a circular buffer, we need to copy from it in
469 * two steps. First, copy the branches from the most recently inserted
470 * branch ->last_branch_pos until the end of bs_src->entries buffer.
471 */
472 nr = etmq->etm->synth_opts.last_branch_sz - etmq->last_branch_pos;
473 memcpy(&bs_dst->entries[0],
474 &bs_src->entries[etmq->last_branch_pos],
475 sizeof(struct branch_entry) * nr);
476
477 /*
478 * If we wrapped around at least once, the branches from the beginning
479 * of the bs_src->entries buffer and until the ->last_branch_pos element
480 * are older valid branches: copy them over. The total number of
481 * branches copied over will be equal to the number of branches asked by
482 * the user in last_branch_sz.
483 */
484 if (bs_src->nr >= etmq->etm->synth_opts.last_branch_sz) {
485 memcpy(&bs_dst->entries[nr],
486 &bs_src->entries[0],
487 sizeof(struct branch_entry) * etmq->last_branch_pos);
488 }
489}
490
491static inline void cs_etm__reset_last_branch_rb(struct cs_etm_queue *etmq)
492{
493 etmq->last_branch_pos = 0;
494 etmq->last_branch_rb->nr = 0;
495}
496
497static inline u64 cs_etm__last_executed_instr(struct cs_etm_packet *packet)
498{
499 /*
500 * The packet records the execution range with an exclusive end address
501 *
502 * A64 instructions are constant size, so the last executed
503 * instruction is A64_INSTR_SIZE before the end address
504 * Will need to do instruction level decode for T32 instructions as
505 * they can be variable size (not yet supported).
506 */
507 return packet->end_addr - A64_INSTR_SIZE;
508}
509
510static inline u64 cs_etm__instr_count(const struct cs_etm_packet *packet)
511{
512 /*
513 * Only A64 instructions are currently supported, so can get
514 * instruction count by dividing.
515 * Will need to do instruction level decode for T32 instructions as
516 * they can be variable size (not yet supported).
517 */
518 return (packet->end_addr - packet->start_addr) / A64_INSTR_SIZE;
519}
520
521static inline u64 cs_etm__instr_addr(const struct cs_etm_packet *packet,
522 u64 offset)
523{
524 /*
525 * Only A64 instructions are currently supported, so can get
526 * instruction address by muliplying.
527 * Will need to do instruction level decode for T32 instructions as
528 * they can be variable size (not yet supported).
529 */
530 return packet->start_addr + offset * A64_INSTR_SIZE;
531}
532
533static void cs_etm__update_last_branch_rb(struct cs_etm_queue *etmq)
534{
535 struct branch_stack *bs = etmq->last_branch_rb;
536 struct branch_entry *be;
537
538 /*
539 * The branches are recorded in a circular buffer in reverse
540 * chronological order: we start recording from the last element of the
541 * buffer down. After writing the first element of the stack, move the
542 * insert position back to the end of the buffer.
543 */
544 if (!etmq->last_branch_pos)
545 etmq->last_branch_pos = etmq->etm->synth_opts.last_branch_sz;
546
547 etmq->last_branch_pos -= 1;
548
549 be = &bs->entries[etmq->last_branch_pos];
550 be->from = cs_etm__last_executed_instr(etmq->prev_packet);
551 be->to = etmq->packet->start_addr;
552 /* No support for mispredict */
553 be->flags.mispred = 0;
554 be->flags.predicted = 1;
555
556 /*
557 * Increment bs->nr until reaching the number of last branches asked by
558 * the user on the command line.
559 */
560 if (bs->nr < etmq->etm->synth_opts.last_branch_sz)
561 bs->nr += 1;
562}
563
564static int cs_etm__inject_event(union perf_event *event,
565 struct perf_sample *sample, u64 type)
566{
567 event->header.size = perf_event__sample_event_size(sample, type, 0);
568 return perf_event__synthesize_sample(event, type, 0, sample);
569}
570
571
572static int
573cs_etm__get_trace(struct cs_etm_buffer *buff, struct cs_etm_queue *etmq)
574{
575 struct auxtrace_buffer *aux_buffer = etmq->buffer;
576 struct auxtrace_buffer *old_buffer = aux_buffer;
577 struct auxtrace_queue *queue;
578
579 queue = &etmq->etm->queues.queue_array[etmq->queue_nr];
580
581 aux_buffer = auxtrace_buffer__next(queue, aux_buffer);
582
583 /* If no more data, drop the previous auxtrace_buffer and return */
584 if (!aux_buffer) {
585 if (old_buffer)
586 auxtrace_buffer__drop_data(old_buffer);
587 buff->len = 0;
588 return 0;
589 }
590
591 etmq->buffer = aux_buffer;
592
593 /* If the aux_buffer doesn't have data associated, try to load it */
594 if (!aux_buffer->data) {
595 /* get the file desc associated with the perf data file */
596 int fd = perf_data__fd(etmq->etm->session->data);
597
598 aux_buffer->data = auxtrace_buffer__get_data(aux_buffer, fd);
599 if (!aux_buffer->data)
600 return -ENOMEM;
601 }
602
603 /* If valid, drop the previous buffer */
604 if (old_buffer)
605 auxtrace_buffer__drop_data(old_buffer);
606
607 buff->offset = aux_buffer->offset;
608 buff->len = aux_buffer->size;
609 buff->buf = aux_buffer->data;
610
611 buff->ref_timestamp = aux_buffer->reference;
612
613 return buff->len;
614}
615
616static void cs_etm__set_pid_tid_cpu(struct cs_etm_auxtrace *etm,
617 struct auxtrace_queue *queue)
618{
619 struct cs_etm_queue *etmq = queue->priv;
620
621 /* CPU-wide tracing isn't supported yet */
622 if (queue->tid == -1)
623 return;
624
625 if ((!etmq->thread) && (etmq->tid != -1))
626 etmq->thread = machine__find_thread(etm->machine, -1,
627 etmq->tid);
628
629 if (etmq->thread) {
630 etmq->pid = etmq->thread->pid_;
631 if (queue->cpu == -1)
632 etmq->cpu = etmq->thread->cpu;
633 }
634}
635
636static int cs_etm__synth_instruction_sample(struct cs_etm_queue *etmq,
637 u64 addr, u64 period)
638{
639 int ret = 0;
640 struct cs_etm_auxtrace *etm = etmq->etm;
641 union perf_event *event = etmq->event_buf;
642 struct perf_sample sample = {.ip = 0,};
643
644 event->sample.header.type = PERF_RECORD_SAMPLE;
645 event->sample.header.misc = PERF_RECORD_MISC_USER;
646 event->sample.header.size = sizeof(struct perf_event_header);
647
648 sample.ip = addr;
649 sample.pid = etmq->pid;
650 sample.tid = etmq->tid;
651 sample.id = etmq->etm->instructions_id;
652 sample.stream_id = etmq->etm->instructions_id;
653 sample.period = period;
654 sample.cpu = etmq->packet->cpu;
655 sample.flags = 0;
656 sample.insn_len = 1;
657 sample.cpumode = event->header.misc;
658
659 if (etm->synth_opts.last_branch) {
660 cs_etm__copy_last_branch_rb(etmq);
661 sample.branch_stack = etmq->last_branch;
662 }
663
664 if (etm->synth_opts.inject) {
665 ret = cs_etm__inject_event(event, &sample,
666 etm->instructions_sample_type);
667 if (ret)
668 return ret;
669 }
670
671 ret = perf_session__deliver_synth_event(etm->session, event, &sample);
672
673 if (ret)
674 pr_err(
675 "CS ETM Trace: failed to deliver instruction event, error %d\n",
676 ret);
677
678 if (etm->synth_opts.last_branch)
679 cs_etm__reset_last_branch_rb(etmq);
680
681 return ret;
682}
683
684/*
685 * The cs etm packet encodes an instruction range between a branch target
686 * and the next taken branch. Generate sample accordingly.
687 */
688static int cs_etm__synth_branch_sample(struct cs_etm_queue *etmq)
689{
690 int ret = 0;
691 struct cs_etm_auxtrace *etm = etmq->etm;
692 struct perf_sample sample = {.ip = 0,};
693 union perf_event *event = etmq->event_buf;
694 struct dummy_branch_stack {
695 u64 nr;
696 struct branch_entry entries;
697 } dummy_bs;
698
699 event->sample.header.type = PERF_RECORD_SAMPLE;
700 event->sample.header.misc = PERF_RECORD_MISC_USER;
701 event->sample.header.size = sizeof(struct perf_event_header);
702
703 sample.ip = cs_etm__last_executed_instr(etmq->prev_packet);
704 sample.pid = etmq->pid;
705 sample.tid = etmq->tid;
706 sample.addr = etmq->packet->start_addr;
707 sample.id = etmq->etm->branches_id;
708 sample.stream_id = etmq->etm->branches_id;
709 sample.period = 1;
710 sample.cpu = etmq->packet->cpu;
711 sample.flags = 0;
712 sample.cpumode = PERF_RECORD_MISC_USER;
713
714 /*
715 * perf report cannot handle events without a branch stack
716 */
717 if (etm->synth_opts.last_branch) {
718 dummy_bs = (struct dummy_branch_stack){
719 .nr = 1,
720 .entries = {
721 .from = sample.ip,
722 .to = sample.addr,
723 },
724 };
725 sample.branch_stack = (struct branch_stack *)&dummy_bs;
726 }
727
728 if (etm->synth_opts.inject) {
729 ret = cs_etm__inject_event(event, &sample,
730 etm->branches_sample_type);
731 if (ret)
732 return ret;
733 }
734
735 ret = perf_session__deliver_synth_event(etm->session, event, &sample);
736
737 if (ret)
738 pr_err(
739 "CS ETM Trace: failed to deliver instruction event, error %d\n",
740 ret);
741
742 return ret;
743}
744
745struct cs_etm_synth {
746 struct perf_tool dummy_tool;
747 struct perf_session *session;
748};
749
750static int cs_etm__event_synth(struct perf_tool *tool,
751 union perf_event *event,
752 struct perf_sample *sample __maybe_unused,
753 struct machine *machine __maybe_unused)
754{
755 struct cs_etm_synth *cs_etm_synth =
756 container_of(tool, struct cs_etm_synth, dummy_tool);
757
758 return perf_session__deliver_synth_event(cs_etm_synth->session,
759 event, NULL);
760}
761
762static int cs_etm__synth_event(struct perf_session *session,
763 struct perf_event_attr *attr, u64 id)
764{
765 struct cs_etm_synth cs_etm_synth;
766
767 memset(&cs_etm_synth, 0, sizeof(struct cs_etm_synth));
768 cs_etm_synth.session = session;
769
770 return perf_event__synthesize_attr(&cs_etm_synth.dummy_tool, attr, 1,
771 &id, cs_etm__event_synth);
772}
773
774static int cs_etm__synth_events(struct cs_etm_auxtrace *etm,
775 struct perf_session *session)
776{
777 struct perf_evlist *evlist = session->evlist;
778 struct perf_evsel *evsel;
779 struct perf_event_attr attr;
780 bool found = false;
781 u64 id;
782 int err;
783
784 evlist__for_each_entry(evlist, evsel) {
785 if (evsel->attr.type == etm->pmu_type) {
786 found = true;
787 break;
788 }
789 }
790
791 if (!found) {
792 pr_debug("No selected events with CoreSight Trace data\n");
793 return 0;
794 }
795
796 memset(&attr, 0, sizeof(struct perf_event_attr));
797 attr.size = sizeof(struct perf_event_attr);
798 attr.type = PERF_TYPE_HARDWARE;
799 attr.sample_type = evsel->attr.sample_type & PERF_SAMPLE_MASK;
800 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
801 PERF_SAMPLE_PERIOD;
802 if (etm->timeless_decoding)
803 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
804 else
805 attr.sample_type |= PERF_SAMPLE_TIME;
806
807 attr.exclude_user = evsel->attr.exclude_user;
808 attr.exclude_kernel = evsel->attr.exclude_kernel;
809 attr.exclude_hv = evsel->attr.exclude_hv;
810 attr.exclude_host = evsel->attr.exclude_host;
811 attr.exclude_guest = evsel->attr.exclude_guest;
812 attr.sample_id_all = evsel->attr.sample_id_all;
813 attr.read_format = evsel->attr.read_format;
814
815 /* create new id val to be a fixed offset from evsel id */
816 id = evsel->id[0] + 1000000000;
817
818 if (!id)
819 id = 1;
820
821 if (etm->synth_opts.branches) {
822 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
823 attr.sample_period = 1;
824 attr.sample_type |= PERF_SAMPLE_ADDR;
825 err = cs_etm__synth_event(session, &attr, id);
826 if (err)
827 return err;
828 etm->sample_branches = true;
829 etm->branches_sample_type = attr.sample_type;
830 etm->branches_id = id;
831 id += 1;
832 attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR;
833 }
834
835 if (etm->synth_opts.last_branch)
836 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
837
838 if (etm->synth_opts.instructions) {
839 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
840 attr.sample_period = etm->synth_opts.period;
841 etm->instructions_sample_period = attr.sample_period;
842 err = cs_etm__synth_event(session, &attr, id);
843 if (err)
844 return err;
845 etm->sample_instructions = true;
846 etm->instructions_sample_type = attr.sample_type;
847 etm->instructions_id = id;
848 id += 1;
849 }
850
851 return 0;
852}
853
854static int cs_etm__sample(struct cs_etm_queue *etmq)
855{
856 struct cs_etm_auxtrace *etm = etmq->etm;
857 struct cs_etm_packet *tmp;
858 int ret;
859 u64 instrs_executed;
860
861 instrs_executed = cs_etm__instr_count(etmq->packet);
862 etmq->period_instructions += instrs_executed;
863
864 /*
865 * Record a branch when the last instruction in
866 * PREV_PACKET is a branch.
867 */
868 if (etm->synth_opts.last_branch &&
869 etmq->prev_packet &&
870 etmq->prev_packet->sample_type == CS_ETM_RANGE &&
871 etmq->prev_packet->last_instr_taken_branch)
872 cs_etm__update_last_branch_rb(etmq);
873
874 if (etm->sample_instructions &&
875 etmq->period_instructions >= etm->instructions_sample_period) {
876 /*
877 * Emit instruction sample periodically
878 * TODO: allow period to be defined in cycles and clock time
879 */
880
881 /* Get number of instructions executed after the sample point */
882 u64 instrs_over = etmq->period_instructions -
883 etm->instructions_sample_period;
884
885 /*
886 * Calculate the address of the sampled instruction (-1 as
887 * sample is reported as though instruction has just been
888 * executed, but PC has not advanced to next instruction)
889 */
890 u64 offset = (instrs_executed - instrs_over - 1);
891 u64 addr = cs_etm__instr_addr(etmq->packet, offset);
892
893 ret = cs_etm__synth_instruction_sample(
894 etmq, addr, etm->instructions_sample_period);
895 if (ret)
896 return ret;
897
898 /* Carry remaining instructions into next sample period */
899 etmq->period_instructions = instrs_over;
900 }
901
902 if (etm->sample_branches &&
903 etmq->prev_packet &&
904 etmq->prev_packet->sample_type == CS_ETM_RANGE &&
905 etmq->prev_packet->last_instr_taken_branch) {
906 ret = cs_etm__synth_branch_sample(etmq);
907 if (ret)
908 return ret;
909 }
910
911 if (etm->sample_branches || etm->synth_opts.last_branch) {
912 /*
913 * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for
914 * the next incoming packet.
915 */
916 tmp = etmq->packet;
917 etmq->packet = etmq->prev_packet;
918 etmq->prev_packet = tmp;
919 }
920
921 return 0;
922}
923
924static int cs_etm__flush(struct cs_etm_queue *etmq)
925{
926 int err = 0;
927 struct cs_etm_packet *tmp;
928
929 if (etmq->etm->synth_opts.last_branch &&
930 etmq->prev_packet &&
931 etmq->prev_packet->sample_type == CS_ETM_RANGE) {
932 /*
933 * Generate a last branch event for the branches left in the
934 * circular buffer at the end of the trace.
935 *
936 * Use the address of the end of the last reported execution
937 * range
938 */
939 u64 addr = cs_etm__last_executed_instr(etmq->prev_packet);
940
941 err = cs_etm__synth_instruction_sample(
942 etmq, addr,
943 etmq->period_instructions);
944 etmq->period_instructions = 0;
945
946 /*
947 * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for
948 * the next incoming packet.
949 */
950 tmp = etmq->packet;
951 etmq->packet = etmq->prev_packet;
952 etmq->prev_packet = tmp;
953 }
954
955 return err;
956}
957
958static int cs_etm__run_decoder(struct cs_etm_queue *etmq)
959{
960 struct cs_etm_auxtrace *etm = etmq->etm;
961 struct cs_etm_buffer buffer;
962 size_t buffer_used, processed;
963 int err = 0;
964
965 if (!etm->kernel_start)
966 etm->kernel_start = machine__kernel_start(etm->machine);
967
968 /* Go through each buffer in the queue and decode them one by one */
969 while (1) {
970 buffer_used = 0;
971 memset(&buffer, 0, sizeof(buffer));
972 err = cs_etm__get_trace(&buffer, etmq);
973 if (err <= 0)
974 return err;
975 /*
976 * We cannot assume consecutive blocks in the data file are
977 * contiguous, reset the decoder to force re-sync.
978 */
979 err = cs_etm_decoder__reset(etmq->decoder);
980 if (err != 0)
981 return err;
982
983 /* Run trace decoder until buffer consumed or end of trace */
984 do {
985 processed = 0;
986 err = cs_etm_decoder__process_data_block(
987 etmq->decoder,
988 etmq->offset,
989 &buffer.buf[buffer_used],
990 buffer.len - buffer_used,
991 &processed);
992 if (err)
993 return err;
994
995 etmq->offset += processed;
996 buffer_used += processed;
997
998 /* Process each packet in this chunk */
999 while (1) {
1000 err = cs_etm_decoder__get_packet(etmq->decoder,
1001 etmq->packet);
1002 if (err <= 0)
1003 /*
1004 * Stop processing this chunk on
1005 * end of data or error
1006 */
1007 break;
1008
1009 switch (etmq->packet->sample_type) {
1010 case CS_ETM_RANGE:
1011 /*
1012 * If the packet contains an instruction
1013 * range, generate instruction sequence
1014 * events.
1015 */
1016 cs_etm__sample(etmq);
1017 break;
1018 case CS_ETM_TRACE_ON:
1019 /*
1020 * Discontinuity in trace, flush
1021 * previous branch stack
1022 */
1023 cs_etm__flush(etmq);
1024 break;
1025 default:
1026 break;
1027 }
1028 }
1029 } while (buffer.len > buffer_used);
1030
1031 if (err == 0)
1032 /* Flush any remaining branch stack entries */
1033 err = cs_etm__flush(etmq);
1034 }
1035
1036 return err;
1037}
1038
1039static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm,
1040 pid_t tid, u64 time_)
1041{
1042 unsigned int i;
1043 struct auxtrace_queues *queues = &etm->queues;
1044
1045 for (i = 0; i < queues->nr_queues; i++) {
1046 struct auxtrace_queue *queue = &etm->queues.queue_array[i];
1047 struct cs_etm_queue *etmq = queue->priv;
1048
1049 if (etmq && ((tid == -1) || (etmq->tid == tid))) {
1050 etmq->time = time_;
1051 cs_etm__set_pid_tid_cpu(etm, queue);
1052 cs_etm__run_decoder(etmq);
1053 }
1054 }
1055
1056 return 0;
1057}
1058
1059static int cs_etm__process_event(struct perf_session *session,
1060 union perf_event *event,
1061 struct perf_sample *sample,
1062 struct perf_tool *tool)
1063{
1064 int err = 0;
1065 u64 timestamp;
1066 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
1067 struct cs_etm_auxtrace,
1068 auxtrace);
1069
1070 if (dump_trace)
1071 return 0;
1072
1073 if (!tool->ordered_events) {
1074 pr_err("CoreSight ETM Trace requires ordered events\n");
1075 return -EINVAL;
1076 }
1077
1078 if (!etm->timeless_decoding)
1079 return -EINVAL;
1080
1081 if (sample->time && (sample->time != (u64) -1))
1082 timestamp = sample->time;
1083 else
1084 timestamp = 0;
1085
1086 if (timestamp || etm->timeless_decoding) {
1087 err = cs_etm__update_queues(etm);
1088 if (err)
1089 return err;
1090 }
1091
1092 if (event->header.type == PERF_RECORD_EXIT)
1093 return cs_etm__process_timeless_queues(etm,
1094 event->fork.tid,
1095 sample->time);
1096
1097 return 0;
1098}
1099
1100static int cs_etm__process_auxtrace_event(struct perf_session *session,
1101 union perf_event *event,
1102 struct perf_tool *tool __maybe_unused)
1103{
1104 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
1105 struct cs_etm_auxtrace,
1106 auxtrace);
1107 if (!etm->data_queued) {
1108 struct auxtrace_buffer *buffer;
1109 off_t data_offset;
1110 int fd = perf_data__fd(session->data);
1111 bool is_pipe = perf_data__is_pipe(session->data);
1112 int err;
1113
1114 if (is_pipe)
1115 data_offset = 0;
1116 else {
1117 data_offset = lseek(fd, 0, SEEK_CUR);
1118 if (data_offset == -1)
1119 return -errno;
1120 }
1121
1122 err = auxtrace_queues__add_event(&etm->queues, session,
1123 event, data_offset, &buffer);
1124 if (err)
1125 return err;
1126
1127 if (dump_trace)
1128 if (auxtrace_buffer__get_data(buffer, fd)) {
1129 cs_etm__dump_event(etm, buffer);
1130 auxtrace_buffer__put_data(buffer);
1131 }
1132 }
1133
1134 return 0;
1135}
1136
1137static bool cs_etm__is_timeless_decoding(struct cs_etm_auxtrace *etm)
1138{
1139 struct perf_evsel *evsel;
1140 struct perf_evlist *evlist = etm->session->evlist;
1141 bool timeless_decoding = true;
1142
1143 /*
1144 * Circle through the list of event and complain if we find one
1145 * with the time bit set.
1146 */
1147 evlist__for_each_entry(evlist, evsel) {
1148 if ((evsel->attr.sample_type & PERF_SAMPLE_TIME))
1149 timeless_decoding = false;
1150 }
1151
1152 return timeless_decoding;
1153}
1154
1155static const char * const cs_etm_global_header_fmts[] = {
1156 [CS_HEADER_VERSION_0] = " Header version %llx\n",
1157 [CS_PMU_TYPE_CPUS] = " PMU type/num cpus %llx\n",
1158 [CS_ETM_SNAPSHOT] = " Snapshot %llx\n",
1159};
1160
1161static const char * const cs_etm_priv_fmts[] = {
1162 [CS_ETM_MAGIC] = " Magic number %llx\n",
1163 [CS_ETM_CPU] = " CPU %lld\n",
1164 [CS_ETM_ETMCR] = " ETMCR %llx\n",
1165 [CS_ETM_ETMTRACEIDR] = " ETMTRACEIDR %llx\n",
1166 [CS_ETM_ETMCCER] = " ETMCCER %llx\n",
1167 [CS_ETM_ETMIDR] = " ETMIDR %llx\n",
1168};
1169
1170static const char * const cs_etmv4_priv_fmts[] = {
1171 [CS_ETM_MAGIC] = " Magic number %llx\n",
1172 [CS_ETM_CPU] = " CPU %lld\n",
1173 [CS_ETMV4_TRCCONFIGR] = " TRCCONFIGR %llx\n",
1174 [CS_ETMV4_TRCTRACEIDR] = " TRCTRACEIDR %llx\n",
1175 [CS_ETMV4_TRCIDR0] = " TRCIDR0 %llx\n",
1176 [CS_ETMV4_TRCIDR1] = " TRCIDR1 %llx\n",
1177 [CS_ETMV4_TRCIDR2] = " TRCIDR2 %llx\n",
1178 [CS_ETMV4_TRCIDR8] = " TRCIDR8 %llx\n",
1179 [CS_ETMV4_TRCAUTHSTATUS] = " TRCAUTHSTATUS %llx\n",
1180};
1181
1182static void cs_etm__print_auxtrace_info(u64 *val, int num)
1183{
1184 int i, j, cpu = 0;
1185
1186 for (i = 0; i < CS_HEADER_VERSION_0_MAX; i++)
1187 fprintf(stdout, cs_etm_global_header_fmts[i], val[i]);
1188
1189 for (i = CS_HEADER_VERSION_0_MAX; cpu < num; cpu++) {
1190 if (val[i] == __perf_cs_etmv3_magic)
1191 for (j = 0; j < CS_ETM_PRIV_MAX; j++, i++)
1192 fprintf(stdout, cs_etm_priv_fmts[j], val[i]);
1193 else if (val[i] == __perf_cs_etmv4_magic)
1194 for (j = 0; j < CS_ETMV4_PRIV_MAX; j++, i++)
1195 fprintf(stdout, cs_etmv4_priv_fmts[j], val[i]);
1196 else
1197 /* failure.. return */
1198 return;
1199 }
1200}
1201
1202int cs_etm__process_auxtrace_info(union perf_event *event,
1203 struct perf_session *session)
1204{
1205 struct auxtrace_info_event *auxtrace_info = &event->auxtrace_info;
1206 struct cs_etm_auxtrace *etm = NULL;
1207 struct int_node *inode;
1208 unsigned int pmu_type;
1209 int event_header_size = sizeof(struct perf_event_header);
1210 int info_header_size;
1211 int total_size = auxtrace_info->header.size;
1212 int priv_size = 0;
1213 int num_cpu;
1214 int err = 0, idx = -1;
1215 int i, j, k;
1216 u64 *ptr, *hdr = NULL;
1217 u64 **metadata = NULL;
1218
1219 /*
1220 * sizeof(auxtrace_info_event::type) +
1221 * sizeof(auxtrace_info_event::reserved) == 8
1222 */
1223 info_header_size = 8;
1224
1225 if (total_size < (event_header_size + info_header_size))
1226 return -EINVAL;
1227
1228 priv_size = total_size - event_header_size - info_header_size;
1229
1230 /* First the global part */
1231 ptr = (u64 *) auxtrace_info->priv;
1232
1233 /* Look for version '0' of the header */
1234 if (ptr[0] != 0)
1235 return -EINVAL;
1236
1237 hdr = zalloc(sizeof(*hdr) * CS_HEADER_VERSION_0_MAX);
1238 if (!hdr)
1239 return -ENOMEM;
1240
1241 /* Extract header information - see cs-etm.h for format */
1242 for (i = 0; i < CS_HEADER_VERSION_0_MAX; i++)
1243 hdr[i] = ptr[i];
1244 num_cpu = hdr[CS_PMU_TYPE_CPUS] & 0xffffffff;
1245 pmu_type = (unsigned int) ((hdr[CS_PMU_TYPE_CPUS] >> 32) &
1246 0xffffffff);
1247
1248 /*
1249 * Create an RB tree for traceID-CPU# tuple. Since the conversion has
1250 * to be made for each packet that gets decoded, optimizing access in
1251 * anything other than a sequential array is worth doing.
1252 */
1253 traceid_list = intlist__new(NULL);
1254 if (!traceid_list) {
1255 err = -ENOMEM;
1256 goto err_free_hdr;
1257 }
1258
1259 metadata = zalloc(sizeof(*metadata) * num_cpu);
1260 if (!metadata) {
1261 err = -ENOMEM;
1262 goto err_free_traceid_list;
1263 }
1264
1265 /*
1266 * The metadata is stored in the auxtrace_info section and encodes
1267 * the configuration of the ARM embedded trace macrocell which is
1268 * required by the trace decoder to properly decode the trace due
1269 * to its highly compressed nature.
1270 */
1271 for (j = 0; j < num_cpu; j++) {
1272 if (ptr[i] == __perf_cs_etmv3_magic) {
1273 metadata[j] = zalloc(sizeof(*metadata[j]) *
1274 CS_ETM_PRIV_MAX);
1275 if (!metadata[j]) {
1276 err = -ENOMEM;
1277 goto err_free_metadata;
1278 }
1279 for (k = 0; k < CS_ETM_PRIV_MAX; k++)
1280 metadata[j][k] = ptr[i + k];
1281
1282 /* The traceID is our handle */
1283 idx = metadata[j][CS_ETM_ETMTRACEIDR];
1284 i += CS_ETM_PRIV_MAX;
1285 } else if (ptr[i] == __perf_cs_etmv4_magic) {
1286 metadata[j] = zalloc(sizeof(*metadata[j]) *
1287 CS_ETMV4_PRIV_MAX);
1288 if (!metadata[j]) {
1289 err = -ENOMEM;
1290 goto err_free_metadata;
1291 }
1292 for (k = 0; k < CS_ETMV4_PRIV_MAX; k++)
1293 metadata[j][k] = ptr[i + k];
1294
1295 /* The traceID is our handle */
1296 idx = metadata[j][CS_ETMV4_TRCTRACEIDR];
1297 i += CS_ETMV4_PRIV_MAX;
1298 }
1299
1300 /* Get an RB node for this CPU */
1301 inode = intlist__findnew(traceid_list, idx);
1302
1303 /* Something went wrong, no need to continue */
1304 if (!inode) {
1305 err = PTR_ERR(inode);
1306 goto err_free_metadata;
1307 }
1308
1309 /*
1310 * The node for that CPU should not be taken.
1311 * Back out if that's the case.
1312 */
1313 if (inode->priv) {
1314 err = -EINVAL;
1315 goto err_free_metadata;
1316 }
1317 /* All good, associate the traceID with the CPU# */
1318 inode->priv = &metadata[j][CS_ETM_CPU];
1319 }
1320
1321 /*
1322 * Each of CS_HEADER_VERSION_0_MAX, CS_ETM_PRIV_MAX and
1323 * CS_ETMV4_PRIV_MAX mark how many double words are in the
1324 * global metadata, and each cpu's metadata respectively.
1325 * The following tests if the correct number of double words was
1326 * present in the auxtrace info section.
1327 */
1328 if (i * 8 != priv_size) {
1329 err = -EINVAL;
1330 goto err_free_metadata;
1331 }
1332
1333 etm = zalloc(sizeof(*etm));
1334
1335 if (!etm) {
1336 err = -ENOMEM;
1337 goto err_free_metadata;
1338 }
1339
1340 err = auxtrace_queues__init(&etm->queues);
1341 if (err)
1342 goto err_free_etm;
1343
1344 etm->session = session;
1345 etm->machine = &session->machines.host;
1346
1347 etm->num_cpu = num_cpu;
1348 etm->pmu_type = pmu_type;
1349 etm->snapshot_mode = (hdr[CS_ETM_SNAPSHOT] != 0);
1350 etm->metadata = metadata;
1351 etm->auxtrace_type = auxtrace_info->type;
1352 etm->timeless_decoding = cs_etm__is_timeless_decoding(etm);
1353
1354 etm->auxtrace.process_event = cs_etm__process_event;
1355 etm->auxtrace.process_auxtrace_event = cs_etm__process_auxtrace_event;
1356 etm->auxtrace.flush_events = cs_etm__flush_events;
1357 etm->auxtrace.free_events = cs_etm__free_events;
1358 etm->auxtrace.free = cs_etm__free;
1359 session->auxtrace = &etm->auxtrace;
1360
1361 etm->unknown_thread = thread__new(999999999, 999999999);
1362 if (!etm->unknown_thread)
1363 goto err_free_queues;
1364
1365 /*
1366 * Initialize list node so that at thread__zput() we can avoid
1367 * segmentation fault at list_del_init().
1368 */
1369 INIT_LIST_HEAD(&etm->unknown_thread->node);
1370
1371 err = thread__set_comm(etm->unknown_thread, "unknown", 0);
1372 if (err)
1373 goto err_delete_thread;
1374
1375 if (thread__init_map_groups(etm->unknown_thread, etm->machine))
1376 goto err_delete_thread;
1377
1378 if (dump_trace) {
1379 cs_etm__print_auxtrace_info(auxtrace_info->priv, num_cpu);
1380 return 0;
1381 }
1382
1383 if (session->itrace_synth_opts && session->itrace_synth_opts->set) {
1384 etm->synth_opts = *session->itrace_synth_opts;
1385 } else {
1386 itrace_synth_opts__set_default(&etm->synth_opts);
1387 etm->synth_opts.callchain = false;
1388 }
1389
1390 err = cs_etm__synth_events(etm, session);
1391 if (err)
1392 goto err_delete_thread;
1393
1394 err = auxtrace_queues__process_index(&etm->queues, session);
1395 if (err)
1396 goto err_delete_thread;
1397
1398 etm->data_queued = etm->queues.populated;
1399
1400 return 0;
1401
1402err_delete_thread:
1403 thread__zput(etm->unknown_thread);
1404err_free_queues:
1405 auxtrace_queues__free(&etm->queues);
1406 session->auxtrace = NULL;
1407err_free_etm:
1408 zfree(&etm);
1409err_free_metadata:
1410 /* No need to check @metadata[j], free(NULL) is supported */
1411 for (j = 0; j < num_cpu; j++)
1412 free(metadata[j]);
1413 zfree(&metadata);
1414err_free_traceid_list:
1415 intlist__delete(traceid_list);
1416err_free_hdr:
1417 zfree(&hdr);
1418
1419 return -EINVAL;
1420}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright(C) 2015-2018 Linaro Limited.
4 *
5 * Author: Tor Jeremiassen <tor@ti.com>
6 * Author: Mathieu Poirier <mathieu.poirier@linaro.org>
7 */
8
9#include <linux/bitops.h>
10#include <linux/coresight-pmu.h>
11#include <linux/err.h>
12#include <linux/kernel.h>
13#include <linux/log2.h>
14#include <linux/types.h>
15#include <linux/zalloc.h>
16
17#include <opencsd/ocsd_if_types.h>
18#include <stdlib.h>
19
20#include "auxtrace.h"
21#include "color.h"
22#include "cs-etm.h"
23#include "cs-etm-decoder/cs-etm-decoder.h"
24#include "debug.h"
25#include "dso.h"
26#include "evlist.h"
27#include "intlist.h"
28#include "machine.h"
29#include "map.h"
30#include "perf.h"
31#include "session.h"
32#include "map_symbol.h"
33#include "branch.h"
34#include "symbol.h"
35#include "tool.h"
36#include "thread.h"
37#include "thread-stack.h"
38#include <tools/libc_compat.h>
39#include "util/synthetic-events.h"
40
41struct cs_etm_auxtrace {
42 struct auxtrace auxtrace;
43 struct auxtrace_queues queues;
44 struct auxtrace_heap heap;
45 struct itrace_synth_opts synth_opts;
46 struct perf_session *session;
47 struct machine *machine;
48 struct thread *unknown_thread;
49
50 u8 timeless_decoding;
51 u8 snapshot_mode;
52 u8 data_queued;
53
54 int num_cpu;
55 u64 latest_kernel_timestamp;
56 u32 auxtrace_type;
57 u64 branches_sample_type;
58 u64 branches_id;
59 u64 instructions_sample_type;
60 u64 instructions_sample_period;
61 u64 instructions_id;
62 u64 **metadata;
63 unsigned int pmu_type;
64};
65
66struct cs_etm_traceid_queue {
67 u8 trace_chan_id;
68 pid_t pid, tid;
69 u64 period_instructions;
70 size_t last_branch_pos;
71 union perf_event *event_buf;
72 struct thread *thread;
73 struct branch_stack *last_branch;
74 struct branch_stack *last_branch_rb;
75 struct cs_etm_packet *prev_packet;
76 struct cs_etm_packet *packet;
77 struct cs_etm_packet_queue packet_queue;
78};
79
80struct cs_etm_queue {
81 struct cs_etm_auxtrace *etm;
82 struct cs_etm_decoder *decoder;
83 struct auxtrace_buffer *buffer;
84 unsigned int queue_nr;
85 u8 pending_timestamp_chan_id;
86 u64 offset;
87 const unsigned char *buf;
88 size_t buf_len, buf_used;
89 /* Conversion between traceID and index in traceid_queues array */
90 struct intlist *traceid_queues_list;
91 struct cs_etm_traceid_queue **traceid_queues;
92};
93
94/* RB tree for quick conversion between traceID and metadata pointers */
95static struct intlist *traceid_list;
96
97static int cs_etm__process_queues(struct cs_etm_auxtrace *etm);
98static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm,
99 pid_t tid);
100static int cs_etm__get_data_block(struct cs_etm_queue *etmq);
101static int cs_etm__decode_data_block(struct cs_etm_queue *etmq);
102
103/* PTMs ETMIDR [11:8] set to b0011 */
104#define ETMIDR_PTM_VERSION 0x00000300
105
106/*
107 * A struct auxtrace_heap_item only has a queue_nr and a timestamp to
108 * work with. One option is to modify to auxtrace_heap_XYZ() API or simply
109 * encode the etm queue number as the upper 16 bit and the channel as
110 * the lower 16 bit.
111 */
112#define TO_CS_QUEUE_NR(queue_nr, trace_chan_id) \
113 (queue_nr << 16 | trace_chan_id)
114#define TO_QUEUE_NR(cs_queue_nr) (cs_queue_nr >> 16)
115#define TO_TRACE_CHAN_ID(cs_queue_nr) (cs_queue_nr & 0x0000ffff)
116
117static u32 cs_etm__get_v7_protocol_version(u32 etmidr)
118{
119 etmidr &= ETMIDR_PTM_VERSION;
120
121 if (etmidr == ETMIDR_PTM_VERSION)
122 return CS_ETM_PROTO_PTM;
123
124 return CS_ETM_PROTO_ETMV3;
125}
126
127static int cs_etm__get_magic(u8 trace_chan_id, u64 *magic)
128{
129 struct int_node *inode;
130 u64 *metadata;
131
132 inode = intlist__find(traceid_list, trace_chan_id);
133 if (!inode)
134 return -EINVAL;
135
136 metadata = inode->priv;
137 *magic = metadata[CS_ETM_MAGIC];
138 return 0;
139}
140
141int cs_etm__get_cpu(u8 trace_chan_id, int *cpu)
142{
143 struct int_node *inode;
144 u64 *metadata;
145
146 inode = intlist__find(traceid_list, trace_chan_id);
147 if (!inode)
148 return -EINVAL;
149
150 metadata = inode->priv;
151 *cpu = (int)metadata[CS_ETM_CPU];
152 return 0;
153}
154
155/*
156 * The returned PID format is presented by two bits:
157 *
158 * Bit ETM_OPT_CTXTID: CONTEXTIDR or CONTEXTIDR_EL1 is traced;
159 * Bit ETM_OPT_CTXTID2: CONTEXTIDR_EL2 is traced.
160 *
161 * It's possible that the two bits ETM_OPT_CTXTID and ETM_OPT_CTXTID2
162 * are enabled at the same time when the session runs on an EL2 kernel.
163 * This means the CONTEXTIDR_EL1 and CONTEXTIDR_EL2 both will be
164 * recorded in the trace data, the tool will selectively use
165 * CONTEXTIDR_EL2 as PID.
166 */
167int cs_etm__get_pid_fmt(u8 trace_chan_id, u64 *pid_fmt)
168{
169 struct int_node *inode;
170 u64 *metadata, val;
171
172 inode = intlist__find(traceid_list, trace_chan_id);
173 if (!inode)
174 return -EINVAL;
175
176 metadata = inode->priv;
177
178 if (metadata[CS_ETM_MAGIC] == __perf_cs_etmv3_magic) {
179 val = metadata[CS_ETM_ETMCR];
180 /* CONTEXTIDR is traced */
181 if (val & BIT(ETM_OPT_CTXTID))
182 *pid_fmt = BIT(ETM_OPT_CTXTID);
183 } else {
184 val = metadata[CS_ETMV4_TRCCONFIGR];
185 /* CONTEXTIDR_EL2 is traced */
186 if (val & (BIT(ETM4_CFG_BIT_VMID) | BIT(ETM4_CFG_BIT_VMID_OPT)))
187 *pid_fmt = BIT(ETM_OPT_CTXTID2);
188 /* CONTEXTIDR_EL1 is traced */
189 else if (val & BIT(ETM4_CFG_BIT_CTXTID))
190 *pid_fmt = BIT(ETM_OPT_CTXTID);
191 }
192
193 return 0;
194}
195
196void cs_etm__etmq_set_traceid_queue_timestamp(struct cs_etm_queue *etmq,
197 u8 trace_chan_id)
198{
199 /*
200 * When a timestamp packet is encountered the backend code
201 * is stopped so that the front end has time to process packets
202 * that were accumulated in the traceID queue. Since there can
203 * be more than one channel per cs_etm_queue, we need to specify
204 * what traceID queue needs servicing.
205 */
206 etmq->pending_timestamp_chan_id = trace_chan_id;
207}
208
209static u64 cs_etm__etmq_get_timestamp(struct cs_etm_queue *etmq,
210 u8 *trace_chan_id)
211{
212 struct cs_etm_packet_queue *packet_queue;
213
214 if (!etmq->pending_timestamp_chan_id)
215 return 0;
216
217 if (trace_chan_id)
218 *trace_chan_id = etmq->pending_timestamp_chan_id;
219
220 packet_queue = cs_etm__etmq_get_packet_queue(etmq,
221 etmq->pending_timestamp_chan_id);
222 if (!packet_queue)
223 return 0;
224
225 /* Acknowledge pending status */
226 etmq->pending_timestamp_chan_id = 0;
227
228 /* See function cs_etm_decoder__do_{hard|soft}_timestamp() */
229 return packet_queue->cs_timestamp;
230}
231
232static void cs_etm__clear_packet_queue(struct cs_etm_packet_queue *queue)
233{
234 int i;
235
236 queue->head = 0;
237 queue->tail = 0;
238 queue->packet_count = 0;
239 for (i = 0; i < CS_ETM_PACKET_MAX_BUFFER; i++) {
240 queue->packet_buffer[i].isa = CS_ETM_ISA_UNKNOWN;
241 queue->packet_buffer[i].start_addr = CS_ETM_INVAL_ADDR;
242 queue->packet_buffer[i].end_addr = CS_ETM_INVAL_ADDR;
243 queue->packet_buffer[i].instr_count = 0;
244 queue->packet_buffer[i].last_instr_taken_branch = false;
245 queue->packet_buffer[i].last_instr_size = 0;
246 queue->packet_buffer[i].last_instr_type = 0;
247 queue->packet_buffer[i].last_instr_subtype = 0;
248 queue->packet_buffer[i].last_instr_cond = 0;
249 queue->packet_buffer[i].flags = 0;
250 queue->packet_buffer[i].exception_number = UINT32_MAX;
251 queue->packet_buffer[i].trace_chan_id = UINT8_MAX;
252 queue->packet_buffer[i].cpu = INT_MIN;
253 }
254}
255
256static void cs_etm__clear_all_packet_queues(struct cs_etm_queue *etmq)
257{
258 int idx;
259 struct int_node *inode;
260 struct cs_etm_traceid_queue *tidq;
261 struct intlist *traceid_queues_list = etmq->traceid_queues_list;
262
263 intlist__for_each_entry(inode, traceid_queues_list) {
264 idx = (int)(intptr_t)inode->priv;
265 tidq = etmq->traceid_queues[idx];
266 cs_etm__clear_packet_queue(&tidq->packet_queue);
267 }
268}
269
270static int cs_etm__init_traceid_queue(struct cs_etm_queue *etmq,
271 struct cs_etm_traceid_queue *tidq,
272 u8 trace_chan_id)
273{
274 int rc = -ENOMEM;
275 struct auxtrace_queue *queue;
276 struct cs_etm_auxtrace *etm = etmq->etm;
277
278 cs_etm__clear_packet_queue(&tidq->packet_queue);
279
280 queue = &etmq->etm->queues.queue_array[etmq->queue_nr];
281 tidq->tid = queue->tid;
282 tidq->pid = -1;
283 tidq->trace_chan_id = trace_chan_id;
284
285 tidq->packet = zalloc(sizeof(struct cs_etm_packet));
286 if (!tidq->packet)
287 goto out;
288
289 tidq->prev_packet = zalloc(sizeof(struct cs_etm_packet));
290 if (!tidq->prev_packet)
291 goto out_free;
292
293 if (etm->synth_opts.last_branch) {
294 size_t sz = sizeof(struct branch_stack);
295
296 sz += etm->synth_opts.last_branch_sz *
297 sizeof(struct branch_entry);
298 tidq->last_branch = zalloc(sz);
299 if (!tidq->last_branch)
300 goto out_free;
301 tidq->last_branch_rb = zalloc(sz);
302 if (!tidq->last_branch_rb)
303 goto out_free;
304 }
305
306 tidq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
307 if (!tidq->event_buf)
308 goto out_free;
309
310 return 0;
311
312out_free:
313 zfree(&tidq->last_branch_rb);
314 zfree(&tidq->last_branch);
315 zfree(&tidq->prev_packet);
316 zfree(&tidq->packet);
317out:
318 return rc;
319}
320
321static struct cs_etm_traceid_queue
322*cs_etm__etmq_get_traceid_queue(struct cs_etm_queue *etmq, u8 trace_chan_id)
323{
324 int idx;
325 struct int_node *inode;
326 struct intlist *traceid_queues_list;
327 struct cs_etm_traceid_queue *tidq, **traceid_queues;
328 struct cs_etm_auxtrace *etm = etmq->etm;
329
330 if (etm->timeless_decoding)
331 trace_chan_id = CS_ETM_PER_THREAD_TRACEID;
332
333 traceid_queues_list = etmq->traceid_queues_list;
334
335 /*
336 * Check if the traceid_queue exist for this traceID by looking
337 * in the queue list.
338 */
339 inode = intlist__find(traceid_queues_list, trace_chan_id);
340 if (inode) {
341 idx = (int)(intptr_t)inode->priv;
342 return etmq->traceid_queues[idx];
343 }
344
345 /* We couldn't find a traceid_queue for this traceID, allocate one */
346 tidq = malloc(sizeof(*tidq));
347 if (!tidq)
348 return NULL;
349
350 memset(tidq, 0, sizeof(*tidq));
351
352 /* Get a valid index for the new traceid_queue */
353 idx = intlist__nr_entries(traceid_queues_list);
354 /* Memory for the inode is free'ed in cs_etm_free_traceid_queues () */
355 inode = intlist__findnew(traceid_queues_list, trace_chan_id);
356 if (!inode)
357 goto out_free;
358
359 /* Associate this traceID with this index */
360 inode->priv = (void *)(intptr_t)idx;
361
362 if (cs_etm__init_traceid_queue(etmq, tidq, trace_chan_id))
363 goto out_free;
364
365 /* Grow the traceid_queues array by one unit */
366 traceid_queues = etmq->traceid_queues;
367 traceid_queues = reallocarray(traceid_queues,
368 idx + 1,
369 sizeof(*traceid_queues));
370
371 /*
372 * On failure reallocarray() returns NULL and the original block of
373 * memory is left untouched.
374 */
375 if (!traceid_queues)
376 goto out_free;
377
378 traceid_queues[idx] = tidq;
379 etmq->traceid_queues = traceid_queues;
380
381 return etmq->traceid_queues[idx];
382
383out_free:
384 /*
385 * Function intlist__remove() removes the inode from the list
386 * and delete the memory associated to it.
387 */
388 intlist__remove(traceid_queues_list, inode);
389 free(tidq);
390
391 return NULL;
392}
393
394struct cs_etm_packet_queue
395*cs_etm__etmq_get_packet_queue(struct cs_etm_queue *etmq, u8 trace_chan_id)
396{
397 struct cs_etm_traceid_queue *tidq;
398
399 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id);
400 if (tidq)
401 return &tidq->packet_queue;
402
403 return NULL;
404}
405
406static void cs_etm__packet_swap(struct cs_etm_auxtrace *etm,
407 struct cs_etm_traceid_queue *tidq)
408{
409 struct cs_etm_packet *tmp;
410
411 if (etm->synth_opts.branches || etm->synth_opts.last_branch ||
412 etm->synth_opts.instructions) {
413 /*
414 * Swap PACKET with PREV_PACKET: PACKET becomes PREV_PACKET for
415 * the next incoming packet.
416 */
417 tmp = tidq->packet;
418 tidq->packet = tidq->prev_packet;
419 tidq->prev_packet = tmp;
420 }
421}
422
423static void cs_etm__packet_dump(const char *pkt_string)
424{
425 const char *color = PERF_COLOR_BLUE;
426 int len = strlen(pkt_string);
427
428 if (len && (pkt_string[len-1] == '\n'))
429 color_fprintf(stdout, color, " %s", pkt_string);
430 else
431 color_fprintf(stdout, color, " %s\n", pkt_string);
432
433 fflush(stdout);
434}
435
436static void cs_etm__set_trace_param_etmv3(struct cs_etm_trace_params *t_params,
437 struct cs_etm_auxtrace *etm, int idx,
438 u32 etmidr)
439{
440 u64 **metadata = etm->metadata;
441
442 t_params[idx].protocol = cs_etm__get_v7_protocol_version(etmidr);
443 t_params[idx].etmv3.reg_ctrl = metadata[idx][CS_ETM_ETMCR];
444 t_params[idx].etmv3.reg_trc_id = metadata[idx][CS_ETM_ETMTRACEIDR];
445}
446
447static void cs_etm__set_trace_param_etmv4(struct cs_etm_trace_params *t_params,
448 struct cs_etm_auxtrace *etm, int idx)
449{
450 u64 **metadata = etm->metadata;
451
452 t_params[idx].protocol = CS_ETM_PROTO_ETMV4i;
453 t_params[idx].etmv4.reg_idr0 = metadata[idx][CS_ETMV4_TRCIDR0];
454 t_params[idx].etmv4.reg_idr1 = metadata[idx][CS_ETMV4_TRCIDR1];
455 t_params[idx].etmv4.reg_idr2 = metadata[idx][CS_ETMV4_TRCIDR2];
456 t_params[idx].etmv4.reg_idr8 = metadata[idx][CS_ETMV4_TRCIDR8];
457 t_params[idx].etmv4.reg_configr = metadata[idx][CS_ETMV4_TRCCONFIGR];
458 t_params[idx].etmv4.reg_traceidr = metadata[idx][CS_ETMV4_TRCTRACEIDR];
459}
460
461static void cs_etm__set_trace_param_ete(struct cs_etm_trace_params *t_params,
462 struct cs_etm_auxtrace *etm, int idx)
463{
464 u64 **metadata = etm->metadata;
465
466 t_params[idx].protocol = CS_ETM_PROTO_ETE;
467 t_params[idx].ete.reg_idr0 = metadata[idx][CS_ETMV4_TRCIDR0];
468 t_params[idx].ete.reg_idr1 = metadata[idx][CS_ETMV4_TRCIDR1];
469 t_params[idx].ete.reg_idr2 = metadata[idx][CS_ETMV4_TRCIDR2];
470 t_params[idx].ete.reg_idr8 = metadata[idx][CS_ETMV4_TRCIDR8];
471 t_params[idx].ete.reg_configr = metadata[idx][CS_ETMV4_TRCCONFIGR];
472 t_params[idx].ete.reg_traceidr = metadata[idx][CS_ETMV4_TRCTRACEIDR];
473 t_params[idx].ete.reg_devarch = metadata[idx][CS_ETE_TRCDEVARCH];
474}
475
476static int cs_etm__init_trace_params(struct cs_etm_trace_params *t_params,
477 struct cs_etm_auxtrace *etm,
478 int decoders)
479{
480 int i;
481 u32 etmidr;
482 u64 architecture;
483
484 for (i = 0; i < decoders; i++) {
485 architecture = etm->metadata[i][CS_ETM_MAGIC];
486
487 switch (architecture) {
488 case __perf_cs_etmv3_magic:
489 etmidr = etm->metadata[i][CS_ETM_ETMIDR];
490 cs_etm__set_trace_param_etmv3(t_params, etm, i, etmidr);
491 break;
492 case __perf_cs_etmv4_magic:
493 cs_etm__set_trace_param_etmv4(t_params, etm, i);
494 break;
495 case __perf_cs_ete_magic:
496 cs_etm__set_trace_param_ete(t_params, etm, i);
497 break;
498 default:
499 return -EINVAL;
500 }
501 }
502
503 return 0;
504}
505
506static int cs_etm__init_decoder_params(struct cs_etm_decoder_params *d_params,
507 struct cs_etm_queue *etmq,
508 enum cs_etm_decoder_operation mode,
509 bool formatted)
510{
511 int ret = -EINVAL;
512
513 if (!(mode < CS_ETM_OPERATION_MAX))
514 goto out;
515
516 d_params->packet_printer = cs_etm__packet_dump;
517 d_params->operation = mode;
518 d_params->data = etmq;
519 d_params->formatted = formatted;
520 d_params->fsyncs = false;
521 d_params->hsyncs = false;
522 d_params->frame_aligned = true;
523
524 ret = 0;
525out:
526 return ret;
527}
528
529static void cs_etm__dump_event(struct cs_etm_queue *etmq,
530 struct auxtrace_buffer *buffer)
531{
532 int ret;
533 const char *color = PERF_COLOR_BLUE;
534 size_t buffer_used = 0;
535
536 fprintf(stdout, "\n");
537 color_fprintf(stdout, color,
538 ". ... CoreSight %s Trace data: size %#zx bytes\n",
539 cs_etm_decoder__get_name(etmq->decoder), buffer->size);
540
541 do {
542 size_t consumed;
543
544 ret = cs_etm_decoder__process_data_block(
545 etmq->decoder, buffer->offset,
546 &((u8 *)buffer->data)[buffer_used],
547 buffer->size - buffer_used, &consumed);
548 if (ret)
549 break;
550
551 buffer_used += consumed;
552 } while (buffer_used < buffer->size);
553
554 cs_etm_decoder__reset(etmq->decoder);
555}
556
557static int cs_etm__flush_events(struct perf_session *session,
558 struct perf_tool *tool)
559{
560 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
561 struct cs_etm_auxtrace,
562 auxtrace);
563 if (dump_trace)
564 return 0;
565
566 if (!tool->ordered_events)
567 return -EINVAL;
568
569 if (etm->timeless_decoding)
570 return cs_etm__process_timeless_queues(etm, -1);
571
572 return cs_etm__process_queues(etm);
573}
574
575static void cs_etm__free_traceid_queues(struct cs_etm_queue *etmq)
576{
577 int idx;
578 uintptr_t priv;
579 struct int_node *inode, *tmp;
580 struct cs_etm_traceid_queue *tidq;
581 struct intlist *traceid_queues_list = etmq->traceid_queues_list;
582
583 intlist__for_each_entry_safe(inode, tmp, traceid_queues_list) {
584 priv = (uintptr_t)inode->priv;
585 idx = priv;
586
587 /* Free this traceid_queue from the array */
588 tidq = etmq->traceid_queues[idx];
589 thread__zput(tidq->thread);
590 zfree(&tidq->event_buf);
591 zfree(&tidq->last_branch);
592 zfree(&tidq->last_branch_rb);
593 zfree(&tidq->prev_packet);
594 zfree(&tidq->packet);
595 zfree(&tidq);
596
597 /*
598 * Function intlist__remove() removes the inode from the list
599 * and delete the memory associated to it.
600 */
601 intlist__remove(traceid_queues_list, inode);
602 }
603
604 /* Then the RB tree itself */
605 intlist__delete(traceid_queues_list);
606 etmq->traceid_queues_list = NULL;
607
608 /* finally free the traceid_queues array */
609 zfree(&etmq->traceid_queues);
610}
611
612static void cs_etm__free_queue(void *priv)
613{
614 struct cs_etm_queue *etmq = priv;
615
616 if (!etmq)
617 return;
618
619 cs_etm_decoder__free(etmq->decoder);
620 cs_etm__free_traceid_queues(etmq);
621 free(etmq);
622}
623
624static void cs_etm__free_events(struct perf_session *session)
625{
626 unsigned int i;
627 struct cs_etm_auxtrace *aux = container_of(session->auxtrace,
628 struct cs_etm_auxtrace,
629 auxtrace);
630 struct auxtrace_queues *queues = &aux->queues;
631
632 for (i = 0; i < queues->nr_queues; i++) {
633 cs_etm__free_queue(queues->queue_array[i].priv);
634 queues->queue_array[i].priv = NULL;
635 }
636
637 auxtrace_queues__free(queues);
638}
639
640static void cs_etm__free(struct perf_session *session)
641{
642 int i;
643 struct int_node *inode, *tmp;
644 struct cs_etm_auxtrace *aux = container_of(session->auxtrace,
645 struct cs_etm_auxtrace,
646 auxtrace);
647 cs_etm__free_events(session);
648 session->auxtrace = NULL;
649
650 /* First remove all traceID/metadata nodes for the RB tree */
651 intlist__for_each_entry_safe(inode, tmp, traceid_list)
652 intlist__remove(traceid_list, inode);
653 /* Then the RB tree itself */
654 intlist__delete(traceid_list);
655
656 for (i = 0; i < aux->num_cpu; i++)
657 zfree(&aux->metadata[i]);
658
659 thread__zput(aux->unknown_thread);
660 zfree(&aux->metadata);
661 zfree(&aux);
662}
663
664static bool cs_etm__evsel_is_auxtrace(struct perf_session *session,
665 struct evsel *evsel)
666{
667 struct cs_etm_auxtrace *aux = container_of(session->auxtrace,
668 struct cs_etm_auxtrace,
669 auxtrace);
670
671 return evsel->core.attr.type == aux->pmu_type;
672}
673
674static u8 cs_etm__cpu_mode(struct cs_etm_queue *etmq, u64 address)
675{
676 struct machine *machine;
677
678 machine = etmq->etm->machine;
679
680 if (address >= machine__kernel_start(machine)) {
681 if (machine__is_host(machine))
682 return PERF_RECORD_MISC_KERNEL;
683 else
684 return PERF_RECORD_MISC_GUEST_KERNEL;
685 } else {
686 if (machine__is_host(machine))
687 return PERF_RECORD_MISC_USER;
688 else if (perf_guest)
689 return PERF_RECORD_MISC_GUEST_USER;
690 else
691 return PERF_RECORD_MISC_HYPERVISOR;
692 }
693}
694
695static u32 cs_etm__mem_access(struct cs_etm_queue *etmq, u8 trace_chan_id,
696 u64 address, size_t size, u8 *buffer)
697{
698 u8 cpumode;
699 u64 offset;
700 int len;
701 struct thread *thread;
702 struct machine *machine;
703 struct addr_location al;
704 struct cs_etm_traceid_queue *tidq;
705
706 if (!etmq)
707 return 0;
708
709 machine = etmq->etm->machine;
710 cpumode = cs_etm__cpu_mode(etmq, address);
711 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id);
712 if (!tidq)
713 return 0;
714
715 thread = tidq->thread;
716 if (!thread) {
717 if (cpumode != PERF_RECORD_MISC_KERNEL)
718 return 0;
719 thread = etmq->etm->unknown_thread;
720 }
721
722 if (!thread__find_map(thread, cpumode, address, &al) || !al.map->dso)
723 return 0;
724
725 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
726 dso__data_status_seen(al.map->dso, DSO_DATA_STATUS_SEEN_ITRACE))
727 return 0;
728
729 offset = al.map->map_ip(al.map, address);
730
731 map__load(al.map);
732
733 len = dso__data_read_offset(al.map->dso, machine, offset, buffer, size);
734
735 if (len <= 0) {
736 ui__warning_once("CS ETM Trace: Missing DSO. Use 'perf archive' or debuginfod to export data from the traced system.\n"
737 " Enable CONFIG_PROC_KCORE or use option '-k /path/to/vmlinux' for kernel symbols.\n");
738 if (!al.map->dso->auxtrace_warned) {
739 pr_err("CS ETM Trace: Debug data not found for address %#"PRIx64" in %s\n",
740 address,
741 al.map->dso->long_name ? al.map->dso->long_name : "Unknown");
742 al.map->dso->auxtrace_warned = true;
743 }
744 return 0;
745 }
746
747 return len;
748}
749
750static struct cs_etm_queue *cs_etm__alloc_queue(struct cs_etm_auxtrace *etm,
751 bool formatted)
752{
753 struct cs_etm_decoder_params d_params;
754 struct cs_etm_trace_params *t_params = NULL;
755 struct cs_etm_queue *etmq;
756 /*
757 * Each queue can only contain data from one CPU when unformatted, so only one decoder is
758 * needed.
759 */
760 int decoders = formatted ? etm->num_cpu : 1;
761
762 etmq = zalloc(sizeof(*etmq));
763 if (!etmq)
764 return NULL;
765
766 etmq->traceid_queues_list = intlist__new(NULL);
767 if (!etmq->traceid_queues_list)
768 goto out_free;
769
770 /* Use metadata to fill in trace parameters for trace decoder */
771 t_params = zalloc(sizeof(*t_params) * decoders);
772
773 if (!t_params)
774 goto out_free;
775
776 if (cs_etm__init_trace_params(t_params, etm, decoders))
777 goto out_free;
778
779 /* Set decoder parameters to decode trace packets */
780 if (cs_etm__init_decoder_params(&d_params, etmq,
781 dump_trace ? CS_ETM_OPERATION_PRINT :
782 CS_ETM_OPERATION_DECODE,
783 formatted))
784 goto out_free;
785
786 etmq->decoder = cs_etm_decoder__new(decoders, &d_params,
787 t_params);
788
789 if (!etmq->decoder)
790 goto out_free;
791
792 /*
793 * Register a function to handle all memory accesses required by
794 * the trace decoder library.
795 */
796 if (cs_etm_decoder__add_mem_access_cb(etmq->decoder,
797 0x0L, ((u64) -1L),
798 cs_etm__mem_access))
799 goto out_free_decoder;
800
801 zfree(&t_params);
802 return etmq;
803
804out_free_decoder:
805 cs_etm_decoder__free(etmq->decoder);
806out_free:
807 intlist__delete(etmq->traceid_queues_list);
808 free(etmq);
809
810 return NULL;
811}
812
813static int cs_etm__setup_queue(struct cs_etm_auxtrace *etm,
814 struct auxtrace_queue *queue,
815 unsigned int queue_nr,
816 bool formatted)
817{
818 struct cs_etm_queue *etmq = queue->priv;
819
820 if (list_empty(&queue->head) || etmq)
821 return 0;
822
823 etmq = cs_etm__alloc_queue(etm, formatted);
824
825 if (!etmq)
826 return -ENOMEM;
827
828 queue->priv = etmq;
829 etmq->etm = etm;
830 etmq->queue_nr = queue_nr;
831 etmq->offset = 0;
832
833 return 0;
834}
835
836static int cs_etm__queue_first_cs_timestamp(struct cs_etm_auxtrace *etm,
837 struct cs_etm_queue *etmq,
838 unsigned int queue_nr)
839{
840 int ret = 0;
841 unsigned int cs_queue_nr;
842 u8 trace_chan_id;
843 u64 cs_timestamp;
844
845 /*
846 * We are under a CPU-wide trace scenario. As such we need to know
847 * when the code that generated the traces started to execute so that
848 * it can be correlated with execution on other CPUs. So we get a
849 * handle on the beginning of traces and decode until we find a
850 * timestamp. The timestamp is then added to the auxtrace min heap
851 * in order to know what nibble (of all the etmqs) to decode first.
852 */
853 while (1) {
854 /*
855 * Fetch an aux_buffer from this etmq. Bail if no more
856 * blocks or an error has been encountered.
857 */
858 ret = cs_etm__get_data_block(etmq);
859 if (ret <= 0)
860 goto out;
861
862 /*
863 * Run decoder on the trace block. The decoder will stop when
864 * encountering a CS timestamp, a full packet queue or the end of
865 * trace for that block.
866 */
867 ret = cs_etm__decode_data_block(etmq);
868 if (ret)
869 goto out;
870
871 /*
872 * Function cs_etm_decoder__do_{hard|soft}_timestamp() does all
873 * the timestamp calculation for us.
874 */
875 cs_timestamp = cs_etm__etmq_get_timestamp(etmq, &trace_chan_id);
876
877 /* We found a timestamp, no need to continue. */
878 if (cs_timestamp)
879 break;
880
881 /*
882 * We didn't find a timestamp so empty all the traceid packet
883 * queues before looking for another timestamp packet, either
884 * in the current data block or a new one. Packets that were
885 * just decoded are useless since no timestamp has been
886 * associated with them. As such simply discard them.
887 */
888 cs_etm__clear_all_packet_queues(etmq);
889 }
890
891 /*
892 * We have a timestamp. Add it to the min heap to reflect when
893 * instructions conveyed by the range packets of this traceID queue
894 * started to execute. Once the same has been done for all the traceID
895 * queues of each etmq, redenring and decoding can start in
896 * chronological order.
897 *
898 * Note that packets decoded above are still in the traceID's packet
899 * queue and will be processed in cs_etm__process_queues().
900 */
901 cs_queue_nr = TO_CS_QUEUE_NR(queue_nr, trace_chan_id);
902 ret = auxtrace_heap__add(&etm->heap, cs_queue_nr, cs_timestamp);
903out:
904 return ret;
905}
906
907static inline
908void cs_etm__copy_last_branch_rb(struct cs_etm_queue *etmq,
909 struct cs_etm_traceid_queue *tidq)
910{
911 struct branch_stack *bs_src = tidq->last_branch_rb;
912 struct branch_stack *bs_dst = tidq->last_branch;
913 size_t nr = 0;
914
915 /*
916 * Set the number of records before early exit: ->nr is used to
917 * determine how many branches to copy from ->entries.
918 */
919 bs_dst->nr = bs_src->nr;
920
921 /*
922 * Early exit when there is nothing to copy.
923 */
924 if (!bs_src->nr)
925 return;
926
927 /*
928 * As bs_src->entries is a circular buffer, we need to copy from it in
929 * two steps. First, copy the branches from the most recently inserted
930 * branch ->last_branch_pos until the end of bs_src->entries buffer.
931 */
932 nr = etmq->etm->synth_opts.last_branch_sz - tidq->last_branch_pos;
933 memcpy(&bs_dst->entries[0],
934 &bs_src->entries[tidq->last_branch_pos],
935 sizeof(struct branch_entry) * nr);
936
937 /*
938 * If we wrapped around at least once, the branches from the beginning
939 * of the bs_src->entries buffer and until the ->last_branch_pos element
940 * are older valid branches: copy them over. The total number of
941 * branches copied over will be equal to the number of branches asked by
942 * the user in last_branch_sz.
943 */
944 if (bs_src->nr >= etmq->etm->synth_opts.last_branch_sz) {
945 memcpy(&bs_dst->entries[nr],
946 &bs_src->entries[0],
947 sizeof(struct branch_entry) * tidq->last_branch_pos);
948 }
949}
950
951static inline
952void cs_etm__reset_last_branch_rb(struct cs_etm_traceid_queue *tidq)
953{
954 tidq->last_branch_pos = 0;
955 tidq->last_branch_rb->nr = 0;
956}
957
958static inline int cs_etm__t32_instr_size(struct cs_etm_queue *etmq,
959 u8 trace_chan_id, u64 addr)
960{
961 u8 instrBytes[2];
962
963 cs_etm__mem_access(etmq, trace_chan_id, addr,
964 ARRAY_SIZE(instrBytes), instrBytes);
965 /*
966 * T32 instruction size is indicated by bits[15:11] of the first
967 * 16-bit word of the instruction: 0b11101, 0b11110 and 0b11111
968 * denote a 32-bit instruction.
969 */
970 return ((instrBytes[1] & 0xF8) >= 0xE8) ? 4 : 2;
971}
972
973static inline u64 cs_etm__first_executed_instr(struct cs_etm_packet *packet)
974{
975 /* Returns 0 for the CS_ETM_DISCONTINUITY packet */
976 if (packet->sample_type == CS_ETM_DISCONTINUITY)
977 return 0;
978
979 return packet->start_addr;
980}
981
982static inline
983u64 cs_etm__last_executed_instr(const struct cs_etm_packet *packet)
984{
985 /* Returns 0 for the CS_ETM_DISCONTINUITY packet */
986 if (packet->sample_type == CS_ETM_DISCONTINUITY)
987 return 0;
988
989 return packet->end_addr - packet->last_instr_size;
990}
991
992static inline u64 cs_etm__instr_addr(struct cs_etm_queue *etmq,
993 u64 trace_chan_id,
994 const struct cs_etm_packet *packet,
995 u64 offset)
996{
997 if (packet->isa == CS_ETM_ISA_T32) {
998 u64 addr = packet->start_addr;
999
1000 while (offset) {
1001 addr += cs_etm__t32_instr_size(etmq,
1002 trace_chan_id, addr);
1003 offset--;
1004 }
1005 return addr;
1006 }
1007
1008 /* Assume a 4 byte instruction size (A32/A64) */
1009 return packet->start_addr + offset * 4;
1010}
1011
1012static void cs_etm__update_last_branch_rb(struct cs_etm_queue *etmq,
1013 struct cs_etm_traceid_queue *tidq)
1014{
1015 struct branch_stack *bs = tidq->last_branch_rb;
1016 struct branch_entry *be;
1017
1018 /*
1019 * The branches are recorded in a circular buffer in reverse
1020 * chronological order: we start recording from the last element of the
1021 * buffer down. After writing the first element of the stack, move the
1022 * insert position back to the end of the buffer.
1023 */
1024 if (!tidq->last_branch_pos)
1025 tidq->last_branch_pos = etmq->etm->synth_opts.last_branch_sz;
1026
1027 tidq->last_branch_pos -= 1;
1028
1029 be = &bs->entries[tidq->last_branch_pos];
1030 be->from = cs_etm__last_executed_instr(tidq->prev_packet);
1031 be->to = cs_etm__first_executed_instr(tidq->packet);
1032 /* No support for mispredict */
1033 be->flags.mispred = 0;
1034 be->flags.predicted = 1;
1035
1036 /*
1037 * Increment bs->nr until reaching the number of last branches asked by
1038 * the user on the command line.
1039 */
1040 if (bs->nr < etmq->etm->synth_opts.last_branch_sz)
1041 bs->nr += 1;
1042}
1043
1044static int cs_etm__inject_event(union perf_event *event,
1045 struct perf_sample *sample, u64 type)
1046{
1047 event->header.size = perf_event__sample_event_size(sample, type, 0);
1048 return perf_event__synthesize_sample(event, type, 0, sample);
1049}
1050
1051
1052static int
1053cs_etm__get_trace(struct cs_etm_queue *etmq)
1054{
1055 struct auxtrace_buffer *aux_buffer = etmq->buffer;
1056 struct auxtrace_buffer *old_buffer = aux_buffer;
1057 struct auxtrace_queue *queue;
1058
1059 queue = &etmq->etm->queues.queue_array[etmq->queue_nr];
1060
1061 aux_buffer = auxtrace_buffer__next(queue, aux_buffer);
1062
1063 /* If no more data, drop the previous auxtrace_buffer and return */
1064 if (!aux_buffer) {
1065 if (old_buffer)
1066 auxtrace_buffer__drop_data(old_buffer);
1067 etmq->buf_len = 0;
1068 return 0;
1069 }
1070
1071 etmq->buffer = aux_buffer;
1072
1073 /* If the aux_buffer doesn't have data associated, try to load it */
1074 if (!aux_buffer->data) {
1075 /* get the file desc associated with the perf data file */
1076 int fd = perf_data__fd(etmq->etm->session->data);
1077
1078 aux_buffer->data = auxtrace_buffer__get_data(aux_buffer, fd);
1079 if (!aux_buffer->data)
1080 return -ENOMEM;
1081 }
1082
1083 /* If valid, drop the previous buffer */
1084 if (old_buffer)
1085 auxtrace_buffer__drop_data(old_buffer);
1086
1087 etmq->buf_used = 0;
1088 etmq->buf_len = aux_buffer->size;
1089 etmq->buf = aux_buffer->data;
1090
1091 return etmq->buf_len;
1092}
1093
1094static void cs_etm__set_pid_tid_cpu(struct cs_etm_auxtrace *etm,
1095 struct cs_etm_traceid_queue *tidq)
1096{
1097 if ((!tidq->thread) && (tidq->tid != -1))
1098 tidq->thread = machine__find_thread(etm->machine, -1,
1099 tidq->tid);
1100
1101 if (tidq->thread)
1102 tidq->pid = tidq->thread->pid_;
1103}
1104
1105int cs_etm__etmq_set_tid(struct cs_etm_queue *etmq,
1106 pid_t tid, u8 trace_chan_id)
1107{
1108 int cpu, err = -EINVAL;
1109 struct cs_etm_auxtrace *etm = etmq->etm;
1110 struct cs_etm_traceid_queue *tidq;
1111
1112 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id);
1113 if (!tidq)
1114 return err;
1115
1116 if (cs_etm__get_cpu(trace_chan_id, &cpu) < 0)
1117 return err;
1118
1119 err = machine__set_current_tid(etm->machine, cpu, tid, tid);
1120 if (err)
1121 return err;
1122
1123 tidq->tid = tid;
1124 thread__zput(tidq->thread);
1125
1126 cs_etm__set_pid_tid_cpu(etm, tidq);
1127 return 0;
1128}
1129
1130bool cs_etm__etmq_is_timeless(struct cs_etm_queue *etmq)
1131{
1132 return !!etmq->etm->timeless_decoding;
1133}
1134
1135static void cs_etm__copy_insn(struct cs_etm_queue *etmq,
1136 u64 trace_chan_id,
1137 const struct cs_etm_packet *packet,
1138 struct perf_sample *sample)
1139{
1140 /*
1141 * It's pointless to read instructions for the CS_ETM_DISCONTINUITY
1142 * packet, so directly bail out with 'insn_len' = 0.
1143 */
1144 if (packet->sample_type == CS_ETM_DISCONTINUITY) {
1145 sample->insn_len = 0;
1146 return;
1147 }
1148
1149 /*
1150 * T32 instruction size might be 32-bit or 16-bit, decide by calling
1151 * cs_etm__t32_instr_size().
1152 */
1153 if (packet->isa == CS_ETM_ISA_T32)
1154 sample->insn_len = cs_etm__t32_instr_size(etmq, trace_chan_id,
1155 sample->ip);
1156 /* Otherwise, A64 and A32 instruction size are always 32-bit. */
1157 else
1158 sample->insn_len = 4;
1159
1160 cs_etm__mem_access(etmq, trace_chan_id, sample->ip,
1161 sample->insn_len, (void *)sample->insn);
1162}
1163
1164static int cs_etm__synth_instruction_sample(struct cs_etm_queue *etmq,
1165 struct cs_etm_traceid_queue *tidq,
1166 u64 addr, u64 period)
1167{
1168 int ret = 0;
1169 struct cs_etm_auxtrace *etm = etmq->etm;
1170 union perf_event *event = tidq->event_buf;
1171 struct perf_sample sample = {.ip = 0,};
1172
1173 event->sample.header.type = PERF_RECORD_SAMPLE;
1174 event->sample.header.misc = cs_etm__cpu_mode(etmq, addr);
1175 event->sample.header.size = sizeof(struct perf_event_header);
1176
1177 if (!etm->timeless_decoding)
1178 sample.time = etm->latest_kernel_timestamp;
1179 sample.ip = addr;
1180 sample.pid = tidq->pid;
1181 sample.tid = tidq->tid;
1182 sample.id = etmq->etm->instructions_id;
1183 sample.stream_id = etmq->etm->instructions_id;
1184 sample.period = period;
1185 sample.cpu = tidq->packet->cpu;
1186 sample.flags = tidq->prev_packet->flags;
1187 sample.cpumode = event->sample.header.misc;
1188
1189 cs_etm__copy_insn(etmq, tidq->trace_chan_id, tidq->packet, &sample);
1190
1191 if (etm->synth_opts.last_branch)
1192 sample.branch_stack = tidq->last_branch;
1193
1194 if (etm->synth_opts.inject) {
1195 ret = cs_etm__inject_event(event, &sample,
1196 etm->instructions_sample_type);
1197 if (ret)
1198 return ret;
1199 }
1200
1201 ret = perf_session__deliver_synth_event(etm->session, event, &sample);
1202
1203 if (ret)
1204 pr_err(
1205 "CS ETM Trace: failed to deliver instruction event, error %d\n",
1206 ret);
1207
1208 return ret;
1209}
1210
1211/*
1212 * The cs etm packet encodes an instruction range between a branch target
1213 * and the next taken branch. Generate sample accordingly.
1214 */
1215static int cs_etm__synth_branch_sample(struct cs_etm_queue *etmq,
1216 struct cs_etm_traceid_queue *tidq)
1217{
1218 int ret = 0;
1219 struct cs_etm_auxtrace *etm = etmq->etm;
1220 struct perf_sample sample = {.ip = 0,};
1221 union perf_event *event = tidq->event_buf;
1222 struct dummy_branch_stack {
1223 u64 nr;
1224 u64 hw_idx;
1225 struct branch_entry entries;
1226 } dummy_bs;
1227 u64 ip;
1228
1229 ip = cs_etm__last_executed_instr(tidq->prev_packet);
1230
1231 event->sample.header.type = PERF_RECORD_SAMPLE;
1232 event->sample.header.misc = cs_etm__cpu_mode(etmq, ip);
1233 event->sample.header.size = sizeof(struct perf_event_header);
1234
1235 if (!etm->timeless_decoding)
1236 sample.time = etm->latest_kernel_timestamp;
1237 sample.ip = ip;
1238 sample.pid = tidq->pid;
1239 sample.tid = tidq->tid;
1240 sample.addr = cs_etm__first_executed_instr(tidq->packet);
1241 sample.id = etmq->etm->branches_id;
1242 sample.stream_id = etmq->etm->branches_id;
1243 sample.period = 1;
1244 sample.cpu = tidq->packet->cpu;
1245 sample.flags = tidq->prev_packet->flags;
1246 sample.cpumode = event->sample.header.misc;
1247
1248 cs_etm__copy_insn(etmq, tidq->trace_chan_id, tidq->prev_packet,
1249 &sample);
1250
1251 /*
1252 * perf report cannot handle events without a branch stack
1253 */
1254 if (etm->synth_opts.last_branch) {
1255 dummy_bs = (struct dummy_branch_stack){
1256 .nr = 1,
1257 .hw_idx = -1ULL,
1258 .entries = {
1259 .from = sample.ip,
1260 .to = sample.addr,
1261 },
1262 };
1263 sample.branch_stack = (struct branch_stack *)&dummy_bs;
1264 }
1265
1266 if (etm->synth_opts.inject) {
1267 ret = cs_etm__inject_event(event, &sample,
1268 etm->branches_sample_type);
1269 if (ret)
1270 return ret;
1271 }
1272
1273 ret = perf_session__deliver_synth_event(etm->session, event, &sample);
1274
1275 if (ret)
1276 pr_err(
1277 "CS ETM Trace: failed to deliver instruction event, error %d\n",
1278 ret);
1279
1280 return ret;
1281}
1282
1283struct cs_etm_synth {
1284 struct perf_tool dummy_tool;
1285 struct perf_session *session;
1286};
1287
1288static int cs_etm__event_synth(struct perf_tool *tool,
1289 union perf_event *event,
1290 struct perf_sample *sample __maybe_unused,
1291 struct machine *machine __maybe_unused)
1292{
1293 struct cs_etm_synth *cs_etm_synth =
1294 container_of(tool, struct cs_etm_synth, dummy_tool);
1295
1296 return perf_session__deliver_synth_event(cs_etm_synth->session,
1297 event, NULL);
1298}
1299
1300static int cs_etm__synth_event(struct perf_session *session,
1301 struct perf_event_attr *attr, u64 id)
1302{
1303 struct cs_etm_synth cs_etm_synth;
1304
1305 memset(&cs_etm_synth, 0, sizeof(struct cs_etm_synth));
1306 cs_etm_synth.session = session;
1307
1308 return perf_event__synthesize_attr(&cs_etm_synth.dummy_tool, attr, 1,
1309 &id, cs_etm__event_synth);
1310}
1311
1312static int cs_etm__synth_events(struct cs_etm_auxtrace *etm,
1313 struct perf_session *session)
1314{
1315 struct evlist *evlist = session->evlist;
1316 struct evsel *evsel;
1317 struct perf_event_attr attr;
1318 bool found = false;
1319 u64 id;
1320 int err;
1321
1322 evlist__for_each_entry(evlist, evsel) {
1323 if (evsel->core.attr.type == etm->pmu_type) {
1324 found = true;
1325 break;
1326 }
1327 }
1328
1329 if (!found) {
1330 pr_debug("No selected events with CoreSight Trace data\n");
1331 return 0;
1332 }
1333
1334 memset(&attr, 0, sizeof(struct perf_event_attr));
1335 attr.size = sizeof(struct perf_event_attr);
1336 attr.type = PERF_TYPE_HARDWARE;
1337 attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK;
1338 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
1339 PERF_SAMPLE_PERIOD;
1340 if (etm->timeless_decoding)
1341 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
1342 else
1343 attr.sample_type |= PERF_SAMPLE_TIME;
1344
1345 attr.exclude_user = evsel->core.attr.exclude_user;
1346 attr.exclude_kernel = evsel->core.attr.exclude_kernel;
1347 attr.exclude_hv = evsel->core.attr.exclude_hv;
1348 attr.exclude_host = evsel->core.attr.exclude_host;
1349 attr.exclude_guest = evsel->core.attr.exclude_guest;
1350 attr.sample_id_all = evsel->core.attr.sample_id_all;
1351 attr.read_format = evsel->core.attr.read_format;
1352
1353 /* create new id val to be a fixed offset from evsel id */
1354 id = evsel->core.id[0] + 1000000000;
1355
1356 if (!id)
1357 id = 1;
1358
1359 if (etm->synth_opts.branches) {
1360 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
1361 attr.sample_period = 1;
1362 attr.sample_type |= PERF_SAMPLE_ADDR;
1363 err = cs_etm__synth_event(session, &attr, id);
1364 if (err)
1365 return err;
1366 etm->branches_sample_type = attr.sample_type;
1367 etm->branches_id = id;
1368 id += 1;
1369 attr.sample_type &= ~(u64)PERF_SAMPLE_ADDR;
1370 }
1371
1372 if (etm->synth_opts.last_branch) {
1373 attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
1374 /*
1375 * We don't use the hardware index, but the sample generation
1376 * code uses the new format branch_stack with this field,
1377 * so the event attributes must indicate that it's present.
1378 */
1379 attr.branch_sample_type |= PERF_SAMPLE_BRANCH_HW_INDEX;
1380 }
1381
1382 if (etm->synth_opts.instructions) {
1383 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
1384 attr.sample_period = etm->synth_opts.period;
1385 etm->instructions_sample_period = attr.sample_period;
1386 err = cs_etm__synth_event(session, &attr, id);
1387 if (err)
1388 return err;
1389 etm->instructions_sample_type = attr.sample_type;
1390 etm->instructions_id = id;
1391 id += 1;
1392 }
1393
1394 return 0;
1395}
1396
1397static int cs_etm__sample(struct cs_etm_queue *etmq,
1398 struct cs_etm_traceid_queue *tidq)
1399{
1400 struct cs_etm_auxtrace *etm = etmq->etm;
1401 int ret;
1402 u8 trace_chan_id = tidq->trace_chan_id;
1403 u64 instrs_prev;
1404
1405 /* Get instructions remainder from previous packet */
1406 instrs_prev = tidq->period_instructions;
1407
1408 tidq->period_instructions += tidq->packet->instr_count;
1409
1410 /*
1411 * Record a branch when the last instruction in
1412 * PREV_PACKET is a branch.
1413 */
1414 if (etm->synth_opts.last_branch &&
1415 tidq->prev_packet->sample_type == CS_ETM_RANGE &&
1416 tidq->prev_packet->last_instr_taken_branch)
1417 cs_etm__update_last_branch_rb(etmq, tidq);
1418
1419 if (etm->synth_opts.instructions &&
1420 tidq->period_instructions >= etm->instructions_sample_period) {
1421 /*
1422 * Emit instruction sample periodically
1423 * TODO: allow period to be defined in cycles and clock time
1424 */
1425
1426 /*
1427 * Below diagram demonstrates the instruction samples
1428 * generation flows:
1429 *
1430 * Instrs Instrs Instrs Instrs
1431 * Sample(n) Sample(n+1) Sample(n+2) Sample(n+3)
1432 * | | | |
1433 * V V V V
1434 * --------------------------------------------------
1435 * ^ ^
1436 * | |
1437 * Period Period
1438 * instructions(Pi) instructions(Pi')
1439 *
1440 * | |
1441 * \---------------- -----------------/
1442 * V
1443 * tidq->packet->instr_count
1444 *
1445 * Instrs Sample(n...) are the synthesised samples occurring
1446 * every etm->instructions_sample_period instructions - as
1447 * defined on the perf command line. Sample(n) is being the
1448 * last sample before the current etm packet, n+1 to n+3
1449 * samples are generated from the current etm packet.
1450 *
1451 * tidq->packet->instr_count represents the number of
1452 * instructions in the current etm packet.
1453 *
1454 * Period instructions (Pi) contains the number of
1455 * instructions executed after the sample point(n) from the
1456 * previous etm packet. This will always be less than
1457 * etm->instructions_sample_period.
1458 *
1459 * When generate new samples, it combines with two parts
1460 * instructions, one is the tail of the old packet and another
1461 * is the head of the new coming packet, to generate
1462 * sample(n+1); sample(n+2) and sample(n+3) consume the
1463 * instructions with sample period. After sample(n+3), the rest
1464 * instructions will be used by later packet and it is assigned
1465 * to tidq->period_instructions for next round calculation.
1466 */
1467
1468 /*
1469 * Get the initial offset into the current packet instructions;
1470 * entry conditions ensure that instrs_prev is less than
1471 * etm->instructions_sample_period.
1472 */
1473 u64 offset = etm->instructions_sample_period - instrs_prev;
1474 u64 addr;
1475
1476 /* Prepare last branches for instruction sample */
1477 if (etm->synth_opts.last_branch)
1478 cs_etm__copy_last_branch_rb(etmq, tidq);
1479
1480 while (tidq->period_instructions >=
1481 etm->instructions_sample_period) {
1482 /*
1483 * Calculate the address of the sampled instruction (-1
1484 * as sample is reported as though instruction has just
1485 * been executed, but PC has not advanced to next
1486 * instruction)
1487 */
1488 addr = cs_etm__instr_addr(etmq, trace_chan_id,
1489 tidq->packet, offset - 1);
1490 ret = cs_etm__synth_instruction_sample(
1491 etmq, tidq, addr,
1492 etm->instructions_sample_period);
1493 if (ret)
1494 return ret;
1495
1496 offset += etm->instructions_sample_period;
1497 tidq->period_instructions -=
1498 etm->instructions_sample_period;
1499 }
1500 }
1501
1502 if (etm->synth_opts.branches) {
1503 bool generate_sample = false;
1504
1505 /* Generate sample for tracing on packet */
1506 if (tidq->prev_packet->sample_type == CS_ETM_DISCONTINUITY)
1507 generate_sample = true;
1508
1509 /* Generate sample for branch taken packet */
1510 if (tidq->prev_packet->sample_type == CS_ETM_RANGE &&
1511 tidq->prev_packet->last_instr_taken_branch)
1512 generate_sample = true;
1513
1514 if (generate_sample) {
1515 ret = cs_etm__synth_branch_sample(etmq, tidq);
1516 if (ret)
1517 return ret;
1518 }
1519 }
1520
1521 cs_etm__packet_swap(etm, tidq);
1522
1523 return 0;
1524}
1525
1526static int cs_etm__exception(struct cs_etm_traceid_queue *tidq)
1527{
1528 /*
1529 * When the exception packet is inserted, whether the last instruction
1530 * in previous range packet is taken branch or not, we need to force
1531 * to set 'prev_packet->last_instr_taken_branch' to true. This ensures
1532 * to generate branch sample for the instruction range before the
1533 * exception is trapped to kernel or before the exception returning.
1534 *
1535 * The exception packet includes the dummy address values, so don't
1536 * swap PACKET with PREV_PACKET. This keeps PREV_PACKET to be useful
1537 * for generating instruction and branch samples.
1538 */
1539 if (tidq->prev_packet->sample_type == CS_ETM_RANGE)
1540 tidq->prev_packet->last_instr_taken_branch = true;
1541
1542 return 0;
1543}
1544
1545static int cs_etm__flush(struct cs_etm_queue *etmq,
1546 struct cs_etm_traceid_queue *tidq)
1547{
1548 int err = 0;
1549 struct cs_etm_auxtrace *etm = etmq->etm;
1550
1551 /* Handle start tracing packet */
1552 if (tidq->prev_packet->sample_type == CS_ETM_EMPTY)
1553 goto swap_packet;
1554
1555 if (etmq->etm->synth_opts.last_branch &&
1556 etmq->etm->synth_opts.instructions &&
1557 tidq->prev_packet->sample_type == CS_ETM_RANGE) {
1558 u64 addr;
1559
1560 /* Prepare last branches for instruction sample */
1561 cs_etm__copy_last_branch_rb(etmq, tidq);
1562
1563 /*
1564 * Generate a last branch event for the branches left in the
1565 * circular buffer at the end of the trace.
1566 *
1567 * Use the address of the end of the last reported execution
1568 * range
1569 */
1570 addr = cs_etm__last_executed_instr(tidq->prev_packet);
1571
1572 err = cs_etm__synth_instruction_sample(
1573 etmq, tidq, addr,
1574 tidq->period_instructions);
1575 if (err)
1576 return err;
1577
1578 tidq->period_instructions = 0;
1579
1580 }
1581
1582 if (etm->synth_opts.branches &&
1583 tidq->prev_packet->sample_type == CS_ETM_RANGE) {
1584 err = cs_etm__synth_branch_sample(etmq, tidq);
1585 if (err)
1586 return err;
1587 }
1588
1589swap_packet:
1590 cs_etm__packet_swap(etm, tidq);
1591
1592 /* Reset last branches after flush the trace */
1593 if (etm->synth_opts.last_branch)
1594 cs_etm__reset_last_branch_rb(tidq);
1595
1596 return err;
1597}
1598
1599static int cs_etm__end_block(struct cs_etm_queue *etmq,
1600 struct cs_etm_traceid_queue *tidq)
1601{
1602 int err;
1603
1604 /*
1605 * It has no new packet coming and 'etmq->packet' contains the stale
1606 * packet which was set at the previous time with packets swapping;
1607 * so skip to generate branch sample to avoid stale packet.
1608 *
1609 * For this case only flush branch stack and generate a last branch
1610 * event for the branches left in the circular buffer at the end of
1611 * the trace.
1612 */
1613 if (etmq->etm->synth_opts.last_branch &&
1614 etmq->etm->synth_opts.instructions &&
1615 tidq->prev_packet->sample_type == CS_ETM_RANGE) {
1616 u64 addr;
1617
1618 /* Prepare last branches for instruction sample */
1619 cs_etm__copy_last_branch_rb(etmq, tidq);
1620
1621 /*
1622 * Use the address of the end of the last reported execution
1623 * range.
1624 */
1625 addr = cs_etm__last_executed_instr(tidq->prev_packet);
1626
1627 err = cs_etm__synth_instruction_sample(
1628 etmq, tidq, addr,
1629 tidq->period_instructions);
1630 if (err)
1631 return err;
1632
1633 tidq->period_instructions = 0;
1634 }
1635
1636 return 0;
1637}
1638/*
1639 * cs_etm__get_data_block: Fetch a block from the auxtrace_buffer queue
1640 * if need be.
1641 * Returns: < 0 if error
1642 * = 0 if no more auxtrace_buffer to read
1643 * > 0 if the current buffer isn't empty yet
1644 */
1645static int cs_etm__get_data_block(struct cs_etm_queue *etmq)
1646{
1647 int ret;
1648
1649 if (!etmq->buf_len) {
1650 ret = cs_etm__get_trace(etmq);
1651 if (ret <= 0)
1652 return ret;
1653 /*
1654 * We cannot assume consecutive blocks in the data file
1655 * are contiguous, reset the decoder to force re-sync.
1656 */
1657 ret = cs_etm_decoder__reset(etmq->decoder);
1658 if (ret)
1659 return ret;
1660 }
1661
1662 return etmq->buf_len;
1663}
1664
1665static bool cs_etm__is_svc_instr(struct cs_etm_queue *etmq, u8 trace_chan_id,
1666 struct cs_etm_packet *packet,
1667 u64 end_addr)
1668{
1669 /* Initialise to keep compiler happy */
1670 u16 instr16 = 0;
1671 u32 instr32 = 0;
1672 u64 addr;
1673
1674 switch (packet->isa) {
1675 case CS_ETM_ISA_T32:
1676 /*
1677 * The SVC of T32 is defined in ARM DDI 0487D.a, F5.1.247:
1678 *
1679 * b'15 b'8
1680 * +-----------------+--------+
1681 * | 1 1 0 1 1 1 1 1 | imm8 |
1682 * +-----------------+--------+
1683 *
1684 * According to the specification, it only defines SVC for T32
1685 * with 16 bits instruction and has no definition for 32bits;
1686 * so below only read 2 bytes as instruction size for T32.
1687 */
1688 addr = end_addr - 2;
1689 cs_etm__mem_access(etmq, trace_chan_id, addr,
1690 sizeof(instr16), (u8 *)&instr16);
1691 if ((instr16 & 0xFF00) == 0xDF00)
1692 return true;
1693
1694 break;
1695 case CS_ETM_ISA_A32:
1696 /*
1697 * The SVC of A32 is defined in ARM DDI 0487D.a, F5.1.247:
1698 *
1699 * b'31 b'28 b'27 b'24
1700 * +---------+---------+-------------------------+
1701 * | !1111 | 1 1 1 1 | imm24 |
1702 * +---------+---------+-------------------------+
1703 */
1704 addr = end_addr - 4;
1705 cs_etm__mem_access(etmq, trace_chan_id, addr,
1706 sizeof(instr32), (u8 *)&instr32);
1707 if ((instr32 & 0x0F000000) == 0x0F000000 &&
1708 (instr32 & 0xF0000000) != 0xF0000000)
1709 return true;
1710
1711 break;
1712 case CS_ETM_ISA_A64:
1713 /*
1714 * The SVC of A64 is defined in ARM DDI 0487D.a, C6.2.294:
1715 *
1716 * b'31 b'21 b'4 b'0
1717 * +-----------------------+---------+-----------+
1718 * | 1 1 0 1 0 1 0 0 0 0 0 | imm16 | 0 0 0 0 1 |
1719 * +-----------------------+---------+-----------+
1720 */
1721 addr = end_addr - 4;
1722 cs_etm__mem_access(etmq, trace_chan_id, addr,
1723 sizeof(instr32), (u8 *)&instr32);
1724 if ((instr32 & 0xFFE0001F) == 0xd4000001)
1725 return true;
1726
1727 break;
1728 case CS_ETM_ISA_UNKNOWN:
1729 default:
1730 break;
1731 }
1732
1733 return false;
1734}
1735
1736static bool cs_etm__is_syscall(struct cs_etm_queue *etmq,
1737 struct cs_etm_traceid_queue *tidq, u64 magic)
1738{
1739 u8 trace_chan_id = tidq->trace_chan_id;
1740 struct cs_etm_packet *packet = tidq->packet;
1741 struct cs_etm_packet *prev_packet = tidq->prev_packet;
1742
1743 if (magic == __perf_cs_etmv3_magic)
1744 if (packet->exception_number == CS_ETMV3_EXC_SVC)
1745 return true;
1746
1747 /*
1748 * ETMv4 exception type CS_ETMV4_EXC_CALL covers SVC, SMC and
1749 * HVC cases; need to check if it's SVC instruction based on
1750 * packet address.
1751 */
1752 if (magic == __perf_cs_etmv4_magic) {
1753 if (packet->exception_number == CS_ETMV4_EXC_CALL &&
1754 cs_etm__is_svc_instr(etmq, trace_chan_id, prev_packet,
1755 prev_packet->end_addr))
1756 return true;
1757 }
1758
1759 return false;
1760}
1761
1762static bool cs_etm__is_async_exception(struct cs_etm_traceid_queue *tidq,
1763 u64 magic)
1764{
1765 struct cs_etm_packet *packet = tidq->packet;
1766
1767 if (magic == __perf_cs_etmv3_magic)
1768 if (packet->exception_number == CS_ETMV3_EXC_DEBUG_HALT ||
1769 packet->exception_number == CS_ETMV3_EXC_ASYNC_DATA_ABORT ||
1770 packet->exception_number == CS_ETMV3_EXC_PE_RESET ||
1771 packet->exception_number == CS_ETMV3_EXC_IRQ ||
1772 packet->exception_number == CS_ETMV3_EXC_FIQ)
1773 return true;
1774
1775 if (magic == __perf_cs_etmv4_magic)
1776 if (packet->exception_number == CS_ETMV4_EXC_RESET ||
1777 packet->exception_number == CS_ETMV4_EXC_DEBUG_HALT ||
1778 packet->exception_number == CS_ETMV4_EXC_SYSTEM_ERROR ||
1779 packet->exception_number == CS_ETMV4_EXC_INST_DEBUG ||
1780 packet->exception_number == CS_ETMV4_EXC_DATA_DEBUG ||
1781 packet->exception_number == CS_ETMV4_EXC_IRQ ||
1782 packet->exception_number == CS_ETMV4_EXC_FIQ)
1783 return true;
1784
1785 return false;
1786}
1787
1788static bool cs_etm__is_sync_exception(struct cs_etm_queue *etmq,
1789 struct cs_etm_traceid_queue *tidq,
1790 u64 magic)
1791{
1792 u8 trace_chan_id = tidq->trace_chan_id;
1793 struct cs_etm_packet *packet = tidq->packet;
1794 struct cs_etm_packet *prev_packet = tidq->prev_packet;
1795
1796 if (magic == __perf_cs_etmv3_magic)
1797 if (packet->exception_number == CS_ETMV3_EXC_SMC ||
1798 packet->exception_number == CS_ETMV3_EXC_HYP ||
1799 packet->exception_number == CS_ETMV3_EXC_JAZELLE_THUMBEE ||
1800 packet->exception_number == CS_ETMV3_EXC_UNDEFINED_INSTR ||
1801 packet->exception_number == CS_ETMV3_EXC_PREFETCH_ABORT ||
1802 packet->exception_number == CS_ETMV3_EXC_DATA_FAULT ||
1803 packet->exception_number == CS_ETMV3_EXC_GENERIC)
1804 return true;
1805
1806 if (magic == __perf_cs_etmv4_magic) {
1807 if (packet->exception_number == CS_ETMV4_EXC_TRAP ||
1808 packet->exception_number == CS_ETMV4_EXC_ALIGNMENT ||
1809 packet->exception_number == CS_ETMV4_EXC_INST_FAULT ||
1810 packet->exception_number == CS_ETMV4_EXC_DATA_FAULT)
1811 return true;
1812
1813 /*
1814 * For CS_ETMV4_EXC_CALL, except SVC other instructions
1815 * (SMC, HVC) are taken as sync exceptions.
1816 */
1817 if (packet->exception_number == CS_ETMV4_EXC_CALL &&
1818 !cs_etm__is_svc_instr(etmq, trace_chan_id, prev_packet,
1819 prev_packet->end_addr))
1820 return true;
1821
1822 /*
1823 * ETMv4 has 5 bits for exception number; if the numbers
1824 * are in the range ( CS_ETMV4_EXC_FIQ, CS_ETMV4_EXC_END ]
1825 * they are implementation defined exceptions.
1826 *
1827 * For this case, simply take it as sync exception.
1828 */
1829 if (packet->exception_number > CS_ETMV4_EXC_FIQ &&
1830 packet->exception_number <= CS_ETMV4_EXC_END)
1831 return true;
1832 }
1833
1834 return false;
1835}
1836
1837static int cs_etm__set_sample_flags(struct cs_etm_queue *etmq,
1838 struct cs_etm_traceid_queue *tidq)
1839{
1840 struct cs_etm_packet *packet = tidq->packet;
1841 struct cs_etm_packet *prev_packet = tidq->prev_packet;
1842 u8 trace_chan_id = tidq->trace_chan_id;
1843 u64 magic;
1844 int ret;
1845
1846 switch (packet->sample_type) {
1847 case CS_ETM_RANGE:
1848 /*
1849 * Immediate branch instruction without neither link nor
1850 * return flag, it's normal branch instruction within
1851 * the function.
1852 */
1853 if (packet->last_instr_type == OCSD_INSTR_BR &&
1854 packet->last_instr_subtype == OCSD_S_INSTR_NONE) {
1855 packet->flags = PERF_IP_FLAG_BRANCH;
1856
1857 if (packet->last_instr_cond)
1858 packet->flags |= PERF_IP_FLAG_CONDITIONAL;
1859 }
1860
1861 /*
1862 * Immediate branch instruction with link (e.g. BL), this is
1863 * branch instruction for function call.
1864 */
1865 if (packet->last_instr_type == OCSD_INSTR_BR &&
1866 packet->last_instr_subtype == OCSD_S_INSTR_BR_LINK)
1867 packet->flags = PERF_IP_FLAG_BRANCH |
1868 PERF_IP_FLAG_CALL;
1869
1870 /*
1871 * Indirect branch instruction with link (e.g. BLR), this is
1872 * branch instruction for function call.
1873 */
1874 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT &&
1875 packet->last_instr_subtype == OCSD_S_INSTR_BR_LINK)
1876 packet->flags = PERF_IP_FLAG_BRANCH |
1877 PERF_IP_FLAG_CALL;
1878
1879 /*
1880 * Indirect branch instruction with subtype of
1881 * OCSD_S_INSTR_V7_IMPLIED_RET, this is explicit hint for
1882 * function return for A32/T32.
1883 */
1884 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT &&
1885 packet->last_instr_subtype == OCSD_S_INSTR_V7_IMPLIED_RET)
1886 packet->flags = PERF_IP_FLAG_BRANCH |
1887 PERF_IP_FLAG_RETURN;
1888
1889 /*
1890 * Indirect branch instruction without link (e.g. BR), usually
1891 * this is used for function return, especially for functions
1892 * within dynamic link lib.
1893 */
1894 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT &&
1895 packet->last_instr_subtype == OCSD_S_INSTR_NONE)
1896 packet->flags = PERF_IP_FLAG_BRANCH |
1897 PERF_IP_FLAG_RETURN;
1898
1899 /* Return instruction for function return. */
1900 if (packet->last_instr_type == OCSD_INSTR_BR_INDIRECT &&
1901 packet->last_instr_subtype == OCSD_S_INSTR_V8_RET)
1902 packet->flags = PERF_IP_FLAG_BRANCH |
1903 PERF_IP_FLAG_RETURN;
1904
1905 /*
1906 * Decoder might insert a discontinuity in the middle of
1907 * instruction packets, fixup prev_packet with flag
1908 * PERF_IP_FLAG_TRACE_BEGIN to indicate restarting trace.
1909 */
1910 if (prev_packet->sample_type == CS_ETM_DISCONTINUITY)
1911 prev_packet->flags |= PERF_IP_FLAG_BRANCH |
1912 PERF_IP_FLAG_TRACE_BEGIN;
1913
1914 /*
1915 * If the previous packet is an exception return packet
1916 * and the return address just follows SVC instruction,
1917 * it needs to calibrate the previous packet sample flags
1918 * as PERF_IP_FLAG_SYSCALLRET.
1919 */
1920 if (prev_packet->flags == (PERF_IP_FLAG_BRANCH |
1921 PERF_IP_FLAG_RETURN |
1922 PERF_IP_FLAG_INTERRUPT) &&
1923 cs_etm__is_svc_instr(etmq, trace_chan_id,
1924 packet, packet->start_addr))
1925 prev_packet->flags = PERF_IP_FLAG_BRANCH |
1926 PERF_IP_FLAG_RETURN |
1927 PERF_IP_FLAG_SYSCALLRET;
1928 break;
1929 case CS_ETM_DISCONTINUITY:
1930 /*
1931 * The trace is discontinuous, if the previous packet is
1932 * instruction packet, set flag PERF_IP_FLAG_TRACE_END
1933 * for previous packet.
1934 */
1935 if (prev_packet->sample_type == CS_ETM_RANGE)
1936 prev_packet->flags |= PERF_IP_FLAG_BRANCH |
1937 PERF_IP_FLAG_TRACE_END;
1938 break;
1939 case CS_ETM_EXCEPTION:
1940 ret = cs_etm__get_magic(packet->trace_chan_id, &magic);
1941 if (ret)
1942 return ret;
1943
1944 /* The exception is for system call. */
1945 if (cs_etm__is_syscall(etmq, tidq, magic))
1946 packet->flags = PERF_IP_FLAG_BRANCH |
1947 PERF_IP_FLAG_CALL |
1948 PERF_IP_FLAG_SYSCALLRET;
1949 /*
1950 * The exceptions are triggered by external signals from bus,
1951 * interrupt controller, debug module, PE reset or halt.
1952 */
1953 else if (cs_etm__is_async_exception(tidq, magic))
1954 packet->flags = PERF_IP_FLAG_BRANCH |
1955 PERF_IP_FLAG_CALL |
1956 PERF_IP_FLAG_ASYNC |
1957 PERF_IP_FLAG_INTERRUPT;
1958 /*
1959 * Otherwise, exception is caused by trap, instruction &
1960 * data fault, or alignment errors.
1961 */
1962 else if (cs_etm__is_sync_exception(etmq, tidq, magic))
1963 packet->flags = PERF_IP_FLAG_BRANCH |
1964 PERF_IP_FLAG_CALL |
1965 PERF_IP_FLAG_INTERRUPT;
1966
1967 /*
1968 * When the exception packet is inserted, since exception
1969 * packet is not used standalone for generating samples
1970 * and it's affiliation to the previous instruction range
1971 * packet; so set previous range packet flags to tell perf
1972 * it is an exception taken branch.
1973 */
1974 if (prev_packet->sample_type == CS_ETM_RANGE)
1975 prev_packet->flags = packet->flags;
1976 break;
1977 case CS_ETM_EXCEPTION_RET:
1978 /*
1979 * When the exception return packet is inserted, since
1980 * exception return packet is not used standalone for
1981 * generating samples and it's affiliation to the previous
1982 * instruction range packet; so set previous range packet
1983 * flags to tell perf it is an exception return branch.
1984 *
1985 * The exception return can be for either system call or
1986 * other exception types; unfortunately the packet doesn't
1987 * contain exception type related info so we cannot decide
1988 * the exception type purely based on exception return packet.
1989 * If we record the exception number from exception packet and
1990 * reuse it for exception return packet, this is not reliable
1991 * due the trace can be discontinuity or the interrupt can
1992 * be nested, thus the recorded exception number cannot be
1993 * used for exception return packet for these two cases.
1994 *
1995 * For exception return packet, we only need to distinguish the
1996 * packet is for system call or for other types. Thus the
1997 * decision can be deferred when receive the next packet which
1998 * contains the return address, based on the return address we
1999 * can read out the previous instruction and check if it's a
2000 * system call instruction and then calibrate the sample flag
2001 * as needed.
2002 */
2003 if (prev_packet->sample_type == CS_ETM_RANGE)
2004 prev_packet->flags = PERF_IP_FLAG_BRANCH |
2005 PERF_IP_FLAG_RETURN |
2006 PERF_IP_FLAG_INTERRUPT;
2007 break;
2008 case CS_ETM_EMPTY:
2009 default:
2010 break;
2011 }
2012
2013 return 0;
2014}
2015
2016static int cs_etm__decode_data_block(struct cs_etm_queue *etmq)
2017{
2018 int ret = 0;
2019 size_t processed = 0;
2020
2021 /*
2022 * Packets are decoded and added to the decoder's packet queue
2023 * until the decoder packet processing callback has requested that
2024 * processing stops or there is nothing left in the buffer. Normal
2025 * operations that stop processing are a timestamp packet or a full
2026 * decoder buffer queue.
2027 */
2028 ret = cs_etm_decoder__process_data_block(etmq->decoder,
2029 etmq->offset,
2030 &etmq->buf[etmq->buf_used],
2031 etmq->buf_len,
2032 &processed);
2033 if (ret)
2034 goto out;
2035
2036 etmq->offset += processed;
2037 etmq->buf_used += processed;
2038 etmq->buf_len -= processed;
2039
2040out:
2041 return ret;
2042}
2043
2044static int cs_etm__process_traceid_queue(struct cs_etm_queue *etmq,
2045 struct cs_etm_traceid_queue *tidq)
2046{
2047 int ret;
2048 struct cs_etm_packet_queue *packet_queue;
2049
2050 packet_queue = &tidq->packet_queue;
2051
2052 /* Process each packet in this chunk */
2053 while (1) {
2054 ret = cs_etm_decoder__get_packet(packet_queue,
2055 tidq->packet);
2056 if (ret <= 0)
2057 /*
2058 * Stop processing this chunk on
2059 * end of data or error
2060 */
2061 break;
2062
2063 /*
2064 * Since packet addresses are swapped in packet
2065 * handling within below switch() statements,
2066 * thus setting sample flags must be called
2067 * prior to switch() statement to use address
2068 * information before packets swapping.
2069 */
2070 ret = cs_etm__set_sample_flags(etmq, tidq);
2071 if (ret < 0)
2072 break;
2073
2074 switch (tidq->packet->sample_type) {
2075 case CS_ETM_RANGE:
2076 /*
2077 * If the packet contains an instruction
2078 * range, generate instruction sequence
2079 * events.
2080 */
2081 cs_etm__sample(etmq, tidq);
2082 break;
2083 case CS_ETM_EXCEPTION:
2084 case CS_ETM_EXCEPTION_RET:
2085 /*
2086 * If the exception packet is coming,
2087 * make sure the previous instruction
2088 * range packet to be handled properly.
2089 */
2090 cs_etm__exception(tidq);
2091 break;
2092 case CS_ETM_DISCONTINUITY:
2093 /*
2094 * Discontinuity in trace, flush
2095 * previous branch stack
2096 */
2097 cs_etm__flush(etmq, tidq);
2098 break;
2099 case CS_ETM_EMPTY:
2100 /*
2101 * Should not receive empty packet,
2102 * report error.
2103 */
2104 pr_err("CS ETM Trace: empty packet\n");
2105 return -EINVAL;
2106 default:
2107 break;
2108 }
2109 }
2110
2111 return ret;
2112}
2113
2114static void cs_etm__clear_all_traceid_queues(struct cs_etm_queue *etmq)
2115{
2116 int idx;
2117 struct int_node *inode;
2118 struct cs_etm_traceid_queue *tidq;
2119 struct intlist *traceid_queues_list = etmq->traceid_queues_list;
2120
2121 intlist__for_each_entry(inode, traceid_queues_list) {
2122 idx = (int)(intptr_t)inode->priv;
2123 tidq = etmq->traceid_queues[idx];
2124
2125 /* Ignore return value */
2126 cs_etm__process_traceid_queue(etmq, tidq);
2127
2128 /*
2129 * Generate an instruction sample with the remaining
2130 * branchstack entries.
2131 */
2132 cs_etm__flush(etmq, tidq);
2133 }
2134}
2135
2136static int cs_etm__run_decoder(struct cs_etm_queue *etmq)
2137{
2138 int err = 0;
2139 struct cs_etm_traceid_queue *tidq;
2140
2141 tidq = cs_etm__etmq_get_traceid_queue(etmq, CS_ETM_PER_THREAD_TRACEID);
2142 if (!tidq)
2143 return -EINVAL;
2144
2145 /* Go through each buffer in the queue and decode them one by one */
2146 while (1) {
2147 err = cs_etm__get_data_block(etmq);
2148 if (err <= 0)
2149 return err;
2150
2151 /* Run trace decoder until buffer consumed or end of trace */
2152 do {
2153 err = cs_etm__decode_data_block(etmq);
2154 if (err)
2155 return err;
2156
2157 /*
2158 * Process each packet in this chunk, nothing to do if
2159 * an error occurs other than hoping the next one will
2160 * be better.
2161 */
2162 err = cs_etm__process_traceid_queue(etmq, tidq);
2163
2164 } while (etmq->buf_len);
2165
2166 if (err == 0)
2167 /* Flush any remaining branch stack entries */
2168 err = cs_etm__end_block(etmq, tidq);
2169 }
2170
2171 return err;
2172}
2173
2174static int cs_etm__process_timeless_queues(struct cs_etm_auxtrace *etm,
2175 pid_t tid)
2176{
2177 unsigned int i;
2178 struct auxtrace_queues *queues = &etm->queues;
2179
2180 for (i = 0; i < queues->nr_queues; i++) {
2181 struct auxtrace_queue *queue = &etm->queues.queue_array[i];
2182 struct cs_etm_queue *etmq = queue->priv;
2183 struct cs_etm_traceid_queue *tidq;
2184
2185 if (!etmq)
2186 continue;
2187
2188 tidq = cs_etm__etmq_get_traceid_queue(etmq,
2189 CS_ETM_PER_THREAD_TRACEID);
2190
2191 if (!tidq)
2192 continue;
2193
2194 if ((tid == -1) || (tidq->tid == tid)) {
2195 cs_etm__set_pid_tid_cpu(etm, tidq);
2196 cs_etm__run_decoder(etmq);
2197 }
2198 }
2199
2200 return 0;
2201}
2202
2203static int cs_etm__process_queues(struct cs_etm_auxtrace *etm)
2204{
2205 int ret = 0;
2206 unsigned int cs_queue_nr, queue_nr, i;
2207 u8 trace_chan_id;
2208 u64 cs_timestamp;
2209 struct auxtrace_queue *queue;
2210 struct cs_etm_queue *etmq;
2211 struct cs_etm_traceid_queue *tidq;
2212
2213 /*
2214 * Pre-populate the heap with one entry from each queue so that we can
2215 * start processing in time order across all queues.
2216 */
2217 for (i = 0; i < etm->queues.nr_queues; i++) {
2218 etmq = etm->queues.queue_array[i].priv;
2219 if (!etmq)
2220 continue;
2221
2222 ret = cs_etm__queue_first_cs_timestamp(etm, etmq, i);
2223 if (ret)
2224 return ret;
2225 }
2226
2227 while (1) {
2228 if (!etm->heap.heap_cnt)
2229 goto out;
2230
2231 /* Take the entry at the top of the min heap */
2232 cs_queue_nr = etm->heap.heap_array[0].queue_nr;
2233 queue_nr = TO_QUEUE_NR(cs_queue_nr);
2234 trace_chan_id = TO_TRACE_CHAN_ID(cs_queue_nr);
2235 queue = &etm->queues.queue_array[queue_nr];
2236 etmq = queue->priv;
2237
2238 /*
2239 * Remove the top entry from the heap since we are about
2240 * to process it.
2241 */
2242 auxtrace_heap__pop(&etm->heap);
2243
2244 tidq = cs_etm__etmq_get_traceid_queue(etmq, trace_chan_id);
2245 if (!tidq) {
2246 /*
2247 * No traceID queue has been allocated for this traceID,
2248 * which means something somewhere went very wrong. No
2249 * other choice than simply exit.
2250 */
2251 ret = -EINVAL;
2252 goto out;
2253 }
2254
2255 /*
2256 * Packets associated with this timestamp are already in
2257 * the etmq's traceID queue, so process them.
2258 */
2259 ret = cs_etm__process_traceid_queue(etmq, tidq);
2260 if (ret < 0)
2261 goto out;
2262
2263 /*
2264 * Packets for this timestamp have been processed, time to
2265 * move on to the next timestamp, fetching a new auxtrace_buffer
2266 * if need be.
2267 */
2268refetch:
2269 ret = cs_etm__get_data_block(etmq);
2270 if (ret < 0)
2271 goto out;
2272
2273 /*
2274 * No more auxtrace_buffers to process in this etmq, simply
2275 * move on to another entry in the auxtrace_heap.
2276 */
2277 if (!ret)
2278 continue;
2279
2280 ret = cs_etm__decode_data_block(etmq);
2281 if (ret)
2282 goto out;
2283
2284 cs_timestamp = cs_etm__etmq_get_timestamp(etmq, &trace_chan_id);
2285
2286 if (!cs_timestamp) {
2287 /*
2288 * Function cs_etm__decode_data_block() returns when
2289 * there is no more traces to decode in the current
2290 * auxtrace_buffer OR when a timestamp has been
2291 * encountered on any of the traceID queues. Since we
2292 * did not get a timestamp, there is no more traces to
2293 * process in this auxtrace_buffer. As such empty and
2294 * flush all traceID queues.
2295 */
2296 cs_etm__clear_all_traceid_queues(etmq);
2297
2298 /* Fetch another auxtrace_buffer for this etmq */
2299 goto refetch;
2300 }
2301
2302 /*
2303 * Add to the min heap the timestamp for packets that have
2304 * just been decoded. They will be processed and synthesized
2305 * during the next call to cs_etm__process_traceid_queue() for
2306 * this queue/traceID.
2307 */
2308 cs_queue_nr = TO_CS_QUEUE_NR(queue_nr, trace_chan_id);
2309 ret = auxtrace_heap__add(&etm->heap, cs_queue_nr, cs_timestamp);
2310 }
2311
2312out:
2313 return ret;
2314}
2315
2316static int cs_etm__process_itrace_start(struct cs_etm_auxtrace *etm,
2317 union perf_event *event)
2318{
2319 struct thread *th;
2320
2321 if (etm->timeless_decoding)
2322 return 0;
2323
2324 /*
2325 * Add the tid/pid to the log so that we can get a match when
2326 * we get a contextID from the decoder.
2327 */
2328 th = machine__findnew_thread(etm->machine,
2329 event->itrace_start.pid,
2330 event->itrace_start.tid);
2331 if (!th)
2332 return -ENOMEM;
2333
2334 thread__put(th);
2335
2336 return 0;
2337}
2338
2339static int cs_etm__process_switch_cpu_wide(struct cs_etm_auxtrace *etm,
2340 union perf_event *event)
2341{
2342 struct thread *th;
2343 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
2344
2345 /*
2346 * Context switch in per-thread mode are irrelevant since perf
2347 * will start/stop tracing as the process is scheduled.
2348 */
2349 if (etm->timeless_decoding)
2350 return 0;
2351
2352 /*
2353 * SWITCH_IN events carry the next process to be switched out while
2354 * SWITCH_OUT events carry the process to be switched in. As such
2355 * we don't care about IN events.
2356 */
2357 if (!out)
2358 return 0;
2359
2360 /*
2361 * Add the tid/pid to the log so that we can get a match when
2362 * we get a contextID from the decoder.
2363 */
2364 th = machine__findnew_thread(etm->machine,
2365 event->context_switch.next_prev_pid,
2366 event->context_switch.next_prev_tid);
2367 if (!th)
2368 return -ENOMEM;
2369
2370 thread__put(th);
2371
2372 return 0;
2373}
2374
2375static int cs_etm__process_event(struct perf_session *session,
2376 union perf_event *event,
2377 struct perf_sample *sample,
2378 struct perf_tool *tool)
2379{
2380 u64 sample_kernel_timestamp;
2381 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
2382 struct cs_etm_auxtrace,
2383 auxtrace);
2384
2385 if (dump_trace)
2386 return 0;
2387
2388 if (!tool->ordered_events) {
2389 pr_err("CoreSight ETM Trace requires ordered events\n");
2390 return -EINVAL;
2391 }
2392
2393 if (sample->time && (sample->time != (u64) -1))
2394 sample_kernel_timestamp = sample->time;
2395 else
2396 sample_kernel_timestamp = 0;
2397
2398 /*
2399 * Don't wait for cs_etm__flush_events() in per-thread/timeless mode to start the decode. We
2400 * need the tid of the PERF_RECORD_EXIT event to assign to the synthesised samples because
2401 * ETM_OPT_CTXTID is not enabled.
2402 */
2403 if (etm->timeless_decoding &&
2404 event->header.type == PERF_RECORD_EXIT)
2405 return cs_etm__process_timeless_queues(etm,
2406 event->fork.tid);
2407
2408 if (event->header.type == PERF_RECORD_ITRACE_START)
2409 return cs_etm__process_itrace_start(etm, event);
2410 else if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
2411 return cs_etm__process_switch_cpu_wide(etm, event);
2412
2413 if (!etm->timeless_decoding && event->header.type == PERF_RECORD_AUX) {
2414 /*
2415 * Record the latest kernel timestamp available in the header
2416 * for samples so that synthesised samples occur from this point
2417 * onwards.
2418 */
2419 etm->latest_kernel_timestamp = sample_kernel_timestamp;
2420 }
2421
2422 return 0;
2423}
2424
2425static void dump_queued_data(struct cs_etm_auxtrace *etm,
2426 struct perf_record_auxtrace *event)
2427{
2428 struct auxtrace_buffer *buf;
2429 unsigned int i;
2430 /*
2431 * Find all buffers with same reference in the queues and dump them.
2432 * This is because the queues can contain multiple entries of the same
2433 * buffer that were split on aux records.
2434 */
2435 for (i = 0; i < etm->queues.nr_queues; ++i)
2436 list_for_each_entry(buf, &etm->queues.queue_array[i].head, list)
2437 if (buf->reference == event->reference)
2438 cs_etm__dump_event(etm->queues.queue_array[i].priv, buf);
2439}
2440
2441static int cs_etm__process_auxtrace_event(struct perf_session *session,
2442 union perf_event *event,
2443 struct perf_tool *tool __maybe_unused)
2444{
2445 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
2446 struct cs_etm_auxtrace,
2447 auxtrace);
2448 if (!etm->data_queued) {
2449 struct auxtrace_buffer *buffer;
2450 off_t data_offset;
2451 int fd = perf_data__fd(session->data);
2452 bool is_pipe = perf_data__is_pipe(session->data);
2453 int err;
2454 int idx = event->auxtrace.idx;
2455
2456 if (is_pipe)
2457 data_offset = 0;
2458 else {
2459 data_offset = lseek(fd, 0, SEEK_CUR);
2460 if (data_offset == -1)
2461 return -errno;
2462 }
2463
2464 err = auxtrace_queues__add_event(&etm->queues, session,
2465 event, data_offset, &buffer);
2466 if (err)
2467 return err;
2468
2469 /*
2470 * Knowing if the trace is formatted or not requires a lookup of
2471 * the aux record so only works in non-piped mode where data is
2472 * queued in cs_etm__queue_aux_records(). Always assume
2473 * formatted in piped mode (true).
2474 */
2475 err = cs_etm__setup_queue(etm, &etm->queues.queue_array[idx],
2476 idx, true);
2477 if (err)
2478 return err;
2479
2480 if (dump_trace)
2481 if (auxtrace_buffer__get_data(buffer, fd)) {
2482 cs_etm__dump_event(etm->queues.queue_array[idx].priv, buffer);
2483 auxtrace_buffer__put_data(buffer);
2484 }
2485 } else if (dump_trace)
2486 dump_queued_data(etm, &event->auxtrace);
2487
2488 return 0;
2489}
2490
2491static bool cs_etm__is_timeless_decoding(struct cs_etm_auxtrace *etm)
2492{
2493 struct evsel *evsel;
2494 struct evlist *evlist = etm->session->evlist;
2495 bool timeless_decoding = true;
2496
2497 /* Override timeless mode with user input from --itrace=Z */
2498 if (etm->synth_opts.timeless_decoding)
2499 return true;
2500
2501 /*
2502 * Circle through the list of event and complain if we find one
2503 * with the time bit set.
2504 */
2505 evlist__for_each_entry(evlist, evsel) {
2506 if ((evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2507 timeless_decoding = false;
2508 }
2509
2510 return timeless_decoding;
2511}
2512
2513/*
2514 * Read a single cpu parameter block from the auxtrace_info priv block.
2515 *
2516 * For version 1 there is a per cpu nr_params entry. If we are handling
2517 * version 1 file, then there may be less, the same, or more params
2518 * indicated by this value than the compile time number we understand.
2519 *
2520 * For a version 0 info block, there are a fixed number, and we need to
2521 * fill out the nr_param value in the metadata we create.
2522 */
2523static u64 *cs_etm__create_meta_blk(u64 *buff_in, int *buff_in_offset,
2524 int out_blk_size, int nr_params_v0)
2525{
2526 u64 *metadata = NULL;
2527 int hdr_version;
2528 int nr_in_params, nr_out_params, nr_cmn_params;
2529 int i, k;
2530
2531 metadata = zalloc(sizeof(*metadata) * out_blk_size);
2532 if (!metadata)
2533 return NULL;
2534
2535 /* read block current index & version */
2536 i = *buff_in_offset;
2537 hdr_version = buff_in[CS_HEADER_VERSION];
2538
2539 if (!hdr_version) {
2540 /* read version 0 info block into a version 1 metadata block */
2541 nr_in_params = nr_params_v0;
2542 metadata[CS_ETM_MAGIC] = buff_in[i + CS_ETM_MAGIC];
2543 metadata[CS_ETM_CPU] = buff_in[i + CS_ETM_CPU];
2544 metadata[CS_ETM_NR_TRC_PARAMS] = nr_in_params;
2545 /* remaining block params at offset +1 from source */
2546 for (k = CS_ETM_COMMON_BLK_MAX_V1 - 1; k < nr_in_params; k++)
2547 metadata[k + 1] = buff_in[i + k];
2548 /* version 0 has 2 common params */
2549 nr_cmn_params = 2;
2550 } else {
2551 /* read version 1 info block - input and output nr_params may differ */
2552 /* version 1 has 3 common params */
2553 nr_cmn_params = 3;
2554 nr_in_params = buff_in[i + CS_ETM_NR_TRC_PARAMS];
2555
2556 /* if input has more params than output - skip excess */
2557 nr_out_params = nr_in_params + nr_cmn_params;
2558 if (nr_out_params > out_blk_size)
2559 nr_out_params = out_blk_size;
2560
2561 for (k = CS_ETM_MAGIC; k < nr_out_params; k++)
2562 metadata[k] = buff_in[i + k];
2563
2564 /* record the actual nr params we copied */
2565 metadata[CS_ETM_NR_TRC_PARAMS] = nr_out_params - nr_cmn_params;
2566 }
2567
2568 /* adjust in offset by number of in params used */
2569 i += nr_in_params + nr_cmn_params;
2570 *buff_in_offset = i;
2571 return metadata;
2572}
2573
2574/**
2575 * Puts a fragment of an auxtrace buffer into the auxtrace queues based
2576 * on the bounds of aux_event, if it matches with the buffer that's at
2577 * file_offset.
2578 *
2579 * Normally, whole auxtrace buffers would be added to the queue. But we
2580 * want to reset the decoder for every PERF_RECORD_AUX event, and the decoder
2581 * is reset across each buffer, so splitting the buffers up in advance has
2582 * the same effect.
2583 */
2584static int cs_etm__queue_aux_fragment(struct perf_session *session, off_t file_offset, size_t sz,
2585 struct perf_record_aux *aux_event, struct perf_sample *sample)
2586{
2587 int err;
2588 char buf[PERF_SAMPLE_MAX_SIZE];
2589 union perf_event *auxtrace_event_union;
2590 struct perf_record_auxtrace *auxtrace_event;
2591 union perf_event auxtrace_fragment;
2592 __u64 aux_offset, aux_size;
2593 __u32 idx;
2594 bool formatted;
2595
2596 struct cs_etm_auxtrace *etm = container_of(session->auxtrace,
2597 struct cs_etm_auxtrace,
2598 auxtrace);
2599
2600 /*
2601 * There should be a PERF_RECORD_AUXTRACE event at the file_offset that we got
2602 * from looping through the auxtrace index.
2603 */
2604 err = perf_session__peek_event(session, file_offset, buf,
2605 PERF_SAMPLE_MAX_SIZE, &auxtrace_event_union, NULL);
2606 if (err)
2607 return err;
2608 auxtrace_event = &auxtrace_event_union->auxtrace;
2609 if (auxtrace_event->header.type != PERF_RECORD_AUXTRACE)
2610 return -EINVAL;
2611
2612 if (auxtrace_event->header.size < sizeof(struct perf_record_auxtrace) ||
2613 auxtrace_event->header.size != sz) {
2614 return -EINVAL;
2615 }
2616
2617 /*
2618 * In per-thread mode, CPU is set to -1, but TID will be set instead. See
2619 * auxtrace_mmap_params__set_idx(). Return 'not found' if neither CPU nor TID match.
2620 */
2621 if ((auxtrace_event->cpu == (__u32) -1 && auxtrace_event->tid != sample->tid) ||
2622 auxtrace_event->cpu != sample->cpu)
2623 return 1;
2624
2625 if (aux_event->flags & PERF_AUX_FLAG_OVERWRITE) {
2626 /*
2627 * Clamp size in snapshot mode. The buffer size is clamped in
2628 * __auxtrace_mmap__read() for snapshots, so the aux record size doesn't reflect
2629 * the buffer size.
2630 */
2631 aux_size = min(aux_event->aux_size, auxtrace_event->size);
2632
2633 /*
2634 * In this mode, the head also points to the end of the buffer so aux_offset
2635 * needs to have the size subtracted so it points to the beginning as in normal mode
2636 */
2637 aux_offset = aux_event->aux_offset - aux_size;
2638 } else {
2639 aux_size = aux_event->aux_size;
2640 aux_offset = aux_event->aux_offset;
2641 }
2642
2643 if (aux_offset >= auxtrace_event->offset &&
2644 aux_offset + aux_size <= auxtrace_event->offset + auxtrace_event->size) {
2645 /*
2646 * If this AUX event was inside this buffer somewhere, create a new auxtrace event
2647 * based on the sizes of the aux event, and queue that fragment.
2648 */
2649 auxtrace_fragment.auxtrace = *auxtrace_event;
2650 auxtrace_fragment.auxtrace.size = aux_size;
2651 auxtrace_fragment.auxtrace.offset = aux_offset;
2652 file_offset += aux_offset - auxtrace_event->offset + auxtrace_event->header.size;
2653
2654 pr_debug3("CS ETM: Queue buffer size: %#"PRI_lx64" offset: %#"PRI_lx64
2655 " tid: %d cpu: %d\n", aux_size, aux_offset, sample->tid, sample->cpu);
2656 err = auxtrace_queues__add_event(&etm->queues, session, &auxtrace_fragment,
2657 file_offset, NULL);
2658 if (err)
2659 return err;
2660
2661 idx = auxtrace_event->idx;
2662 formatted = !(aux_event->flags & PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW);
2663 return cs_etm__setup_queue(etm, &etm->queues.queue_array[idx],
2664 idx, formatted);
2665 }
2666
2667 /* Wasn't inside this buffer, but there were no parse errors. 1 == 'not found' */
2668 return 1;
2669}
2670
2671static int cs_etm__queue_aux_records_cb(struct perf_session *session, union perf_event *event,
2672 u64 offset __maybe_unused, void *data __maybe_unused)
2673{
2674 struct perf_sample sample;
2675 int ret;
2676 struct auxtrace_index_entry *ent;
2677 struct auxtrace_index *auxtrace_index;
2678 struct evsel *evsel;
2679 size_t i;
2680
2681 /* Don't care about any other events, we're only queuing buffers for AUX events */
2682 if (event->header.type != PERF_RECORD_AUX)
2683 return 0;
2684
2685 if (event->header.size < sizeof(struct perf_record_aux))
2686 return -EINVAL;
2687
2688 /* Truncated Aux records can have 0 size and shouldn't result in anything being queued. */
2689 if (!event->aux.aux_size)
2690 return 0;
2691
2692 /*
2693 * Parse the sample, we need the sample_id_all data that comes after the event so that the
2694 * CPU or PID can be matched to an AUXTRACE buffer's CPU or PID.
2695 */
2696 evsel = evlist__event2evsel(session->evlist, event);
2697 if (!evsel)
2698 return -EINVAL;
2699 ret = evsel__parse_sample(evsel, event, &sample);
2700 if (ret)
2701 return ret;
2702
2703 /*
2704 * Loop through the auxtrace index to find the buffer that matches up with this aux event.
2705 */
2706 list_for_each_entry(auxtrace_index, &session->auxtrace_index, list) {
2707 for (i = 0; i < auxtrace_index->nr; i++) {
2708 ent = &auxtrace_index->entries[i];
2709 ret = cs_etm__queue_aux_fragment(session, ent->file_offset,
2710 ent->sz, &event->aux, &sample);
2711 /*
2712 * Stop search on error or successful values. Continue search on
2713 * 1 ('not found')
2714 */
2715 if (ret != 1)
2716 return ret;
2717 }
2718 }
2719
2720 /*
2721 * Couldn't find the buffer corresponding to this aux record, something went wrong. Warn but
2722 * don't exit with an error because it will still be possible to decode other aux records.
2723 */
2724 pr_err("CS ETM: Couldn't find auxtrace buffer for aux_offset: %#"PRI_lx64
2725 " tid: %d cpu: %d\n", event->aux.aux_offset, sample.tid, sample.cpu);
2726 return 0;
2727}
2728
2729static int cs_etm__queue_aux_records(struct perf_session *session)
2730{
2731 struct auxtrace_index *index = list_first_entry_or_null(&session->auxtrace_index,
2732 struct auxtrace_index, list);
2733 if (index && index->nr > 0)
2734 return perf_session__peek_events(session, session->header.data_offset,
2735 session->header.data_size,
2736 cs_etm__queue_aux_records_cb, NULL);
2737
2738 /*
2739 * We would get here if there are no entries in the index (either no auxtrace
2740 * buffers or no index at all). Fail silently as there is the possibility of
2741 * queueing them in cs_etm__process_auxtrace_event() if etm->data_queued is still
2742 * false.
2743 *
2744 * In that scenario, buffers will not be split by AUX records.
2745 */
2746 return 0;
2747}
2748
2749int cs_etm__process_auxtrace_info_full(union perf_event *event,
2750 struct perf_session *session)
2751{
2752 struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info;
2753 struct cs_etm_auxtrace *etm = NULL;
2754 struct int_node *inode;
2755 int event_header_size = sizeof(struct perf_event_header);
2756 int total_size = auxtrace_info->header.size;
2757 int priv_size = 0;
2758 int num_cpu, trcidr_idx;
2759 int err = 0;
2760 int i, j;
2761 u64 *ptr = NULL;
2762 u64 **metadata = NULL;
2763
2764 /*
2765 * Create an RB tree for traceID-metadata tuple. Since the conversion
2766 * has to be made for each packet that gets decoded, optimizing access
2767 * in anything other than a sequential array is worth doing.
2768 */
2769 traceid_list = intlist__new(NULL);
2770 if (!traceid_list)
2771 return -ENOMEM;
2772
2773 /* First the global part */
2774 ptr = (u64 *) auxtrace_info->priv;
2775 num_cpu = ptr[CS_PMU_TYPE_CPUS] & 0xffffffff;
2776 metadata = zalloc(sizeof(*metadata) * num_cpu);
2777 if (!metadata) {
2778 err = -ENOMEM;
2779 goto err_free_traceid_list;
2780 }
2781
2782 /* Start parsing after the common part of the header */
2783 i = CS_HEADER_VERSION_MAX;
2784
2785 /*
2786 * The metadata is stored in the auxtrace_info section and encodes
2787 * the configuration of the ARM embedded trace macrocell which is
2788 * required by the trace decoder to properly decode the trace due
2789 * to its highly compressed nature.
2790 */
2791 for (j = 0; j < num_cpu; j++) {
2792 if (ptr[i] == __perf_cs_etmv3_magic) {
2793 metadata[j] =
2794 cs_etm__create_meta_blk(ptr, &i,
2795 CS_ETM_PRIV_MAX,
2796 CS_ETM_NR_TRC_PARAMS_V0);
2797
2798 /* The traceID is our handle */
2799 trcidr_idx = CS_ETM_ETMTRACEIDR;
2800
2801 } else if (ptr[i] == __perf_cs_etmv4_magic) {
2802 metadata[j] =
2803 cs_etm__create_meta_blk(ptr, &i,
2804 CS_ETMV4_PRIV_MAX,
2805 CS_ETMV4_NR_TRC_PARAMS_V0);
2806
2807 /* The traceID is our handle */
2808 trcidr_idx = CS_ETMV4_TRCTRACEIDR;
2809 } else if (ptr[i] == __perf_cs_ete_magic) {
2810 metadata[j] = cs_etm__create_meta_blk(ptr, &i, CS_ETE_PRIV_MAX, -1);
2811
2812 /* ETE shares first part of metadata with ETMv4 */
2813 trcidr_idx = CS_ETMV4_TRCTRACEIDR;
2814 } else {
2815 ui__error("CS ETM Trace: Unrecognised magic number %#"PRIx64". File could be from a newer version of perf.\n",
2816 ptr[i]);
2817 err = -EINVAL;
2818 goto err_free_metadata;
2819 }
2820
2821 if (!metadata[j]) {
2822 err = -ENOMEM;
2823 goto err_free_metadata;
2824 }
2825
2826 /* Get an RB node for this CPU */
2827 inode = intlist__findnew(traceid_list, metadata[j][trcidr_idx]);
2828
2829 /* Something went wrong, no need to continue */
2830 if (!inode) {
2831 err = -ENOMEM;
2832 goto err_free_metadata;
2833 }
2834
2835 /*
2836 * The node for that CPU should not be taken.
2837 * Back out if that's the case.
2838 */
2839 if (inode->priv) {
2840 err = -EINVAL;
2841 goto err_free_metadata;
2842 }
2843 /* All good, associate the traceID with the metadata pointer */
2844 inode->priv = metadata[j];
2845 }
2846
2847 /*
2848 * Each of CS_HEADER_VERSION_MAX, CS_ETM_PRIV_MAX and
2849 * CS_ETMV4_PRIV_MAX mark how many double words are in the
2850 * global metadata, and each cpu's metadata respectively.
2851 * The following tests if the correct number of double words was
2852 * present in the auxtrace info section.
2853 */
2854 priv_size = total_size - event_header_size - INFO_HEADER_SIZE;
2855 if (i * 8 != priv_size) {
2856 err = -EINVAL;
2857 goto err_free_metadata;
2858 }
2859
2860 etm = zalloc(sizeof(*etm));
2861
2862 if (!etm) {
2863 err = -ENOMEM;
2864 goto err_free_metadata;
2865 }
2866
2867 err = auxtrace_queues__init(&etm->queues);
2868 if (err)
2869 goto err_free_etm;
2870
2871 if (session->itrace_synth_opts->set) {
2872 etm->synth_opts = *session->itrace_synth_opts;
2873 } else {
2874 itrace_synth_opts__set_default(&etm->synth_opts,
2875 session->itrace_synth_opts->default_no_sample);
2876 etm->synth_opts.callchain = false;
2877 }
2878
2879 etm->session = session;
2880 etm->machine = &session->machines.host;
2881
2882 etm->num_cpu = num_cpu;
2883 etm->pmu_type = (unsigned int) ((ptr[CS_PMU_TYPE_CPUS] >> 32) & 0xffffffff);
2884 etm->snapshot_mode = (ptr[CS_ETM_SNAPSHOT] != 0);
2885 etm->metadata = metadata;
2886 etm->auxtrace_type = auxtrace_info->type;
2887 etm->timeless_decoding = cs_etm__is_timeless_decoding(etm);
2888
2889 etm->auxtrace.process_event = cs_etm__process_event;
2890 etm->auxtrace.process_auxtrace_event = cs_etm__process_auxtrace_event;
2891 etm->auxtrace.flush_events = cs_etm__flush_events;
2892 etm->auxtrace.free_events = cs_etm__free_events;
2893 etm->auxtrace.free = cs_etm__free;
2894 etm->auxtrace.evsel_is_auxtrace = cs_etm__evsel_is_auxtrace;
2895 session->auxtrace = &etm->auxtrace;
2896
2897 etm->unknown_thread = thread__new(999999999, 999999999);
2898 if (!etm->unknown_thread) {
2899 err = -ENOMEM;
2900 goto err_free_queues;
2901 }
2902
2903 /*
2904 * Initialize list node so that at thread__zput() we can avoid
2905 * segmentation fault at list_del_init().
2906 */
2907 INIT_LIST_HEAD(&etm->unknown_thread->node);
2908
2909 err = thread__set_comm(etm->unknown_thread, "unknown", 0);
2910 if (err)
2911 goto err_delete_thread;
2912
2913 if (thread__init_maps(etm->unknown_thread, etm->machine)) {
2914 err = -ENOMEM;
2915 goto err_delete_thread;
2916 }
2917
2918 err = cs_etm__synth_events(etm, session);
2919 if (err)
2920 goto err_delete_thread;
2921
2922 err = cs_etm__queue_aux_records(session);
2923 if (err)
2924 goto err_delete_thread;
2925
2926 etm->data_queued = etm->queues.populated;
2927 /*
2928 * Print warning in pipe mode, see cs_etm__process_auxtrace_event() and
2929 * cs_etm__queue_aux_fragment() for details relating to limitations.
2930 */
2931 if (!etm->data_queued)
2932 pr_warning("CS ETM warning: Coresight decode and TRBE support requires random file access.\n"
2933 "Continuing with best effort decoding in piped mode.\n\n");
2934
2935 return 0;
2936
2937err_delete_thread:
2938 thread__zput(etm->unknown_thread);
2939err_free_queues:
2940 auxtrace_queues__free(&etm->queues);
2941 session->auxtrace = NULL;
2942err_free_etm:
2943 zfree(&etm);
2944err_free_metadata:
2945 /* No need to check @metadata[j], free(NULL) is supported */
2946 for (j = 0; j < num_cpu; j++)
2947 zfree(&metadata[j]);
2948 zfree(&metadata);
2949err_free_traceid_list:
2950 intlist__delete(traceid_list);
2951 return err;
2952}