Loading...
Note: File does not exist in v4.6.
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}