1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * intel-bts.c: Intel Processor Trace support
4 * Copyright (c) 2013-2015, Intel Corporation.
5 */
6
7 #include <endian.h>
8 #include <errno.h>
9 #include <byteswap.h>
10 #include <inttypes.h>
11 #include <linux/kernel.h>
12 #include <linux/types.h>
13 #include <linux/bitops.h>
14 #include <linux/log2.h>
15 #include <linux/zalloc.h>
16
17 #include "color.h"
18 #include "evsel.h"
19 #include "evlist.h"
20 #include "machine.h"
21 #include "symbol.h"
22 #include "session.h"
23 #include "tool.h"
24 #include "thread.h"
25 #include "thread-stack.h"
26 #include "debug.h"
27 #include "tsc.h"
28 #include "auxtrace.h"
29 #include "intel-pt-decoder/intel-pt-insn-decoder.h"
30 #include "intel-bts.h"
31 #include "util/synthetic-events.h"
32
33 #define MAX_TIMESTAMP (~0ULL)
34
35 #define INTEL_BTS_ERR_NOINSN 5
36 #define INTEL_BTS_ERR_LOST 9
37
38 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
39 #define le64_to_cpu bswap_64
40 #else
41 #define le64_to_cpu
42 #endif
43
44 struct intel_bts {
45 struct auxtrace auxtrace;
46 struct auxtrace_queues queues;
47 struct auxtrace_heap heap;
48 u32 auxtrace_type;
49 struct perf_session *session;
50 struct machine *machine;
51 bool sampling_mode;
52 bool snapshot_mode;
53 bool data_queued;
54 u32 pmu_type;
55 struct perf_tsc_conversion tc;
56 bool cap_user_time_zero;
57 struct itrace_synth_opts synth_opts;
58 bool sample_branches;
59 u32 branches_filter;
60 u64 branches_sample_type;
61 u64 branches_id;
62 size_t branches_event_size;
63 unsigned long num_events;
64 };
65
66 struct intel_bts_queue {
67 struct intel_bts *bts;
68 unsigned int queue_nr;
69 struct auxtrace_buffer *buffer;
70 bool on_heap;
71 bool done;
72 pid_t pid;
73 pid_t tid;
74 int cpu;
75 u64 time;
76 struct intel_pt_insn intel_pt_insn;
77 u32 sample_flags;
78 };
79
80 struct branch {
81 u64 from;
82 u64 to;
83 u64 misc;
84 };
85
intel_bts_dump(struct intel_bts * bts __maybe_unused,unsigned char * buf,size_t len)86 static void intel_bts_dump(struct intel_bts *bts __maybe_unused,
87 unsigned char *buf, size_t len)
88 {
89 struct branch *branch;
90 size_t i, pos = 0, br_sz = sizeof(struct branch), sz;
91 const char *color = PERF_COLOR_BLUE;
92
93 color_fprintf(stdout, color,
94 ". ... Intel BTS data: size %zu bytes\n",
95 len);
96
97 while (len) {
98 if (len >= br_sz)
99 sz = br_sz;
100 else
101 sz = len;
102 printf(".");
103 color_fprintf(stdout, color, " %08zx: ", pos);
104 for (i = 0; i < sz; i++)
105 color_fprintf(stdout, color, " %02x", buf[i]);
106 for (; i < br_sz; i++)
107 color_fprintf(stdout, color, " ");
108 if (len >= br_sz) {
109 branch = (struct branch *)buf;
110 color_fprintf(stdout, color, " %"PRIx64" -> %"PRIx64" %s\n",
111 le64_to_cpu(branch->from),
112 le64_to_cpu(branch->to),
113 le64_to_cpu(branch->misc) & 0x10 ?
114 "pred" : "miss");
115 } else {
116 color_fprintf(stdout, color, " Bad record!\n");
117 }
118 pos += sz;
119 buf += sz;
120 len -= sz;
121 }
122 }
123
intel_bts_dump_event(struct intel_bts * bts,unsigned char * buf,size_t len)124 static void intel_bts_dump_event(struct intel_bts *bts, unsigned char *buf,
125 size_t len)
126 {
127 printf(".\n");
128 intel_bts_dump(bts, buf, len);
129 }
130
intel_bts_lost(struct intel_bts * bts,struct perf_sample * sample)131 static int intel_bts_lost(struct intel_bts *bts, struct perf_sample *sample)
132 {
133 union perf_event event;
134 int err;
135
136 auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
137 INTEL_BTS_ERR_LOST, sample->cpu, sample->pid,
138 sample->tid, 0, "Lost trace data", sample->time);
139
140 err = perf_session__deliver_synth_event(bts->session, &event, NULL);
141 if (err)
142 pr_err("Intel BTS: failed to deliver error event, error %d\n",
143 err);
144
145 return err;
146 }
147
intel_bts_alloc_queue(struct intel_bts * bts,unsigned int queue_nr)148 static struct intel_bts_queue *intel_bts_alloc_queue(struct intel_bts *bts,
149 unsigned int queue_nr)
150 {
151 struct intel_bts_queue *btsq;
152
153 btsq = zalloc(sizeof(struct intel_bts_queue));
154 if (!btsq)
155 return NULL;
156
157 btsq->bts = bts;
158 btsq->queue_nr = queue_nr;
159 btsq->pid = -1;
160 btsq->tid = -1;
161 btsq->cpu = -1;
162
163 return btsq;
164 }
165
intel_bts_setup_queue(struct intel_bts * bts,struct auxtrace_queue * queue,unsigned int queue_nr)166 static int intel_bts_setup_queue(struct intel_bts *bts,
167 struct auxtrace_queue *queue,
168 unsigned int queue_nr)
169 {
170 struct intel_bts_queue *btsq = queue->priv;
171
172 if (list_empty(&queue->head))
173 return 0;
174
175 if (!btsq) {
176 btsq = intel_bts_alloc_queue(bts, queue_nr);
177 if (!btsq)
178 return -ENOMEM;
179 queue->priv = btsq;
180
181 if (queue->cpu != -1)
182 btsq->cpu = queue->cpu;
183 btsq->tid = queue->tid;
184 }
185
186 if (bts->sampling_mode)
187 return 0;
188
189 if (!btsq->on_heap && !btsq->buffer) {
190 int ret;
191
192 btsq->buffer = auxtrace_buffer__next(queue, NULL);
193 if (!btsq->buffer)
194 return 0;
195
196 ret = auxtrace_heap__add(&bts->heap, queue_nr,
197 btsq->buffer->reference);
198 if (ret)
199 return ret;
200 btsq->on_heap = true;
201 }
202
203 return 0;
204 }
205
intel_bts_setup_queues(struct intel_bts * bts)206 static int intel_bts_setup_queues(struct intel_bts *bts)
207 {
208 unsigned int i;
209 int ret;
210
211 for (i = 0; i < bts->queues.nr_queues; i++) {
212 ret = intel_bts_setup_queue(bts, &bts->queues.queue_array[i],
213 i);
214 if (ret)
215 return ret;
216 }
217 return 0;
218 }
219
intel_bts_update_queues(struct intel_bts * bts)220 static inline int intel_bts_update_queues(struct intel_bts *bts)
221 {
222 if (bts->queues.new_data) {
223 bts->queues.new_data = false;
224 return intel_bts_setup_queues(bts);
225 }
226 return 0;
227 }
228
intel_bts_find_overlap(unsigned char * buf_a,size_t len_a,unsigned char * buf_b,size_t len_b)229 static unsigned char *intel_bts_find_overlap(unsigned char *buf_a, size_t len_a,
230 unsigned char *buf_b, size_t len_b)
231 {
232 size_t offs, len;
233
234 if (len_a > len_b)
235 offs = len_a - len_b;
236 else
237 offs = 0;
238
239 for (; offs < len_a; offs += sizeof(struct branch)) {
240 len = len_a - offs;
241 if (!memcmp(buf_a + offs, buf_b, len))
242 return buf_b + len;
243 }
244
245 return buf_b;
246 }
247
intel_bts_do_fix_overlap(struct auxtrace_queue * queue,struct auxtrace_buffer * b)248 static int intel_bts_do_fix_overlap(struct auxtrace_queue *queue,
249 struct auxtrace_buffer *b)
250 {
251 struct auxtrace_buffer *a;
252 void *start;
253
254 if (b->list.prev == &queue->head)
255 return 0;
256 a = list_entry(b->list.prev, struct auxtrace_buffer, list);
257 start = intel_bts_find_overlap(a->data, a->size, b->data, b->size);
258 if (!start)
259 return -EINVAL;
260 b->use_size = b->data + b->size - start;
261 b->use_data = start;
262 return 0;
263 }
264
intel_bts_cpumode(struct intel_bts * bts,uint64_t ip)265 static inline u8 intel_bts_cpumode(struct intel_bts *bts, uint64_t ip)
266 {
267 return machine__kernel_ip(bts->machine, ip) ?
268 PERF_RECORD_MISC_KERNEL :
269 PERF_RECORD_MISC_USER;
270 }
271
intel_bts_synth_branch_sample(struct intel_bts_queue * btsq,struct branch * branch)272 static int intel_bts_synth_branch_sample(struct intel_bts_queue *btsq,
273 struct branch *branch)
274 {
275 int ret;
276 struct intel_bts *bts = btsq->bts;
277 union perf_event event;
278 struct perf_sample sample;
279
280 if (bts->synth_opts.initial_skip &&
281 bts->num_events++ <= bts->synth_opts.initial_skip)
282 return 0;
283
284 perf_sample__init(&sample, /*all=*/true);
285 sample.ip = le64_to_cpu(branch->from);
286 sample.cpumode = intel_bts_cpumode(bts, sample.ip);
287 sample.pid = btsq->pid;
288 sample.tid = btsq->tid;
289 sample.addr = le64_to_cpu(branch->to);
290 sample.id = btsq->bts->branches_id;
291 sample.stream_id = btsq->bts->branches_id;
292 sample.period = 1;
293 sample.cpu = btsq->cpu;
294 sample.flags = btsq->sample_flags;
295 sample.insn_len = btsq->intel_pt_insn.length;
296 memcpy(sample.insn, btsq->intel_pt_insn.buf, INTEL_PT_INSN_BUF_SZ);
297
298 event.sample.header.type = PERF_RECORD_SAMPLE;
299 event.sample.header.misc = sample.cpumode;
300 event.sample.header.size = sizeof(struct perf_event_header);
301
302 if (bts->synth_opts.inject) {
303 event.sample.header.size = bts->branches_event_size;
304 ret = perf_event__synthesize_sample(&event,
305 bts->branches_sample_type,
306 0, &sample);
307 if (ret)
308 return ret;
309 }
310
311 ret = perf_session__deliver_synth_event(bts->session, &event, &sample);
312 if (ret)
313 pr_err("Intel BTS: failed to deliver branch event, error %d\n",
314 ret);
315
316 perf_sample__exit(&sample);
317 return ret;
318 }
319
intel_bts_get_next_insn(struct intel_bts_queue * btsq,u64 ip)320 static int intel_bts_get_next_insn(struct intel_bts_queue *btsq, u64 ip)
321 {
322 struct machine *machine = btsq->bts->machine;
323 struct thread *thread;
324 unsigned char buf[INTEL_PT_INSN_BUF_SZ];
325 ssize_t len;
326 bool x86_64;
327 int err = -1;
328
329 thread = machine__find_thread(machine, -1, btsq->tid);
330 if (!thread)
331 return -1;
332
333 len = thread__memcpy(thread, machine, buf, ip, INTEL_PT_INSN_BUF_SZ, &x86_64);
334 if (len <= 0)
335 goto out_put;
336
337 if (intel_pt_get_insn(buf, len, x86_64, &btsq->intel_pt_insn))
338 goto out_put;
339
340 err = 0;
341 out_put:
342 thread__put(thread);
343 return err;
344 }
345
intel_bts_synth_error(struct intel_bts * bts,int cpu,pid_t pid,pid_t tid,u64 ip)346 static int intel_bts_synth_error(struct intel_bts *bts, int cpu, pid_t pid,
347 pid_t tid, u64 ip)
348 {
349 union perf_event event;
350 int err;
351
352 auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
353 INTEL_BTS_ERR_NOINSN, cpu, pid, tid, ip,
354 "Failed to get instruction", 0);
355
356 err = perf_session__deliver_synth_event(bts->session, &event, NULL);
357 if (err)
358 pr_err("Intel BTS: failed to deliver error event, error %d\n",
359 err);
360
361 return err;
362 }
363
intel_bts_get_branch_type(struct intel_bts_queue * btsq,struct branch * branch)364 static int intel_bts_get_branch_type(struct intel_bts_queue *btsq,
365 struct branch *branch)
366 {
367 int err;
368
369 if (!branch->from) {
370 if (branch->to)
371 btsq->sample_flags = PERF_IP_FLAG_BRANCH |
372 PERF_IP_FLAG_TRACE_BEGIN;
373 else
374 btsq->sample_flags = 0;
375 btsq->intel_pt_insn.length = 0;
376 } else if (!branch->to) {
377 btsq->sample_flags = PERF_IP_FLAG_BRANCH |
378 PERF_IP_FLAG_TRACE_END;
379 btsq->intel_pt_insn.length = 0;
380 } else {
381 err = intel_bts_get_next_insn(btsq, branch->from);
382 if (err) {
383 btsq->sample_flags = 0;
384 btsq->intel_pt_insn.length = 0;
385 if (!btsq->bts->synth_opts.errors)
386 return 0;
387 err = intel_bts_synth_error(btsq->bts, btsq->cpu,
388 btsq->pid, btsq->tid,
389 branch->from);
390 return err;
391 }
392 btsq->sample_flags = intel_pt_insn_type(btsq->intel_pt_insn.op);
393 /* Check for an async branch into the kernel */
394 if (!machine__kernel_ip(btsq->bts->machine, branch->from) &&
395 machine__kernel_ip(btsq->bts->machine, branch->to) &&
396 btsq->sample_flags != (PERF_IP_FLAG_BRANCH |
397 PERF_IP_FLAG_CALL |
398 PERF_IP_FLAG_SYSCALLRET))
399 btsq->sample_flags = PERF_IP_FLAG_BRANCH |
400 PERF_IP_FLAG_CALL |
401 PERF_IP_FLAG_ASYNC |
402 PERF_IP_FLAG_INTERRUPT;
403 }
404
405 return 0;
406 }
407
intel_bts_process_buffer(struct intel_bts_queue * btsq,struct auxtrace_buffer * buffer,struct thread * thread)408 static int intel_bts_process_buffer(struct intel_bts_queue *btsq,
409 struct auxtrace_buffer *buffer,
410 struct thread *thread)
411 {
412 struct branch *branch;
413 size_t sz, bsz = sizeof(struct branch);
414 u32 filter = btsq->bts->branches_filter;
415 int err = 0;
416
417 if (buffer->use_data) {
418 sz = buffer->use_size;
419 branch = buffer->use_data;
420 } else {
421 sz = buffer->size;
422 branch = buffer->data;
423 }
424
425 if (!btsq->bts->sample_branches)
426 return 0;
427
428 for (; sz > bsz; branch += 1, sz -= bsz) {
429 if (!branch->from && !branch->to)
430 continue;
431 intel_bts_get_branch_type(btsq, branch);
432 if (btsq->bts->synth_opts.thread_stack)
433 thread_stack__event(thread, btsq->cpu, btsq->sample_flags,
434 le64_to_cpu(branch->from),
435 le64_to_cpu(branch->to),
436 btsq->intel_pt_insn.length,
437 buffer->buffer_nr + 1, true, 0, 0);
438 if (filter && !(filter & btsq->sample_flags))
439 continue;
440 err = intel_bts_synth_branch_sample(btsq, branch);
441 if (err)
442 break;
443 }
444 return err;
445 }
446
intel_bts_process_queue(struct intel_bts_queue * btsq,u64 * timestamp)447 static int intel_bts_process_queue(struct intel_bts_queue *btsq, u64 *timestamp)
448 {
449 struct auxtrace_buffer *buffer = btsq->buffer, *old_buffer = buffer;
450 struct auxtrace_queue *queue;
451 struct thread *thread;
452 int err;
453
454 if (btsq->done)
455 return 1;
456
457 if (btsq->pid == -1) {
458 thread = machine__find_thread(btsq->bts->machine, -1,
459 btsq->tid);
460 if (thread)
461 btsq->pid = thread__pid(thread);
462 } else {
463 thread = machine__findnew_thread(btsq->bts->machine, btsq->pid,
464 btsq->tid);
465 }
466
467 queue = &btsq->bts->queues.queue_array[btsq->queue_nr];
468
469 if (!buffer)
470 buffer = auxtrace_buffer__next(queue, NULL);
471
472 if (!buffer) {
473 if (!btsq->bts->sampling_mode)
474 btsq->done = 1;
475 err = 1;
476 goto out_put;
477 }
478
479 /* Currently there is no support for split buffers */
480 if (buffer->consecutive) {
481 err = -EINVAL;
482 goto out_put;
483 }
484
485 if (!buffer->data) {
486 int fd = perf_data__fd(btsq->bts->session->data);
487
488 buffer->data = auxtrace_buffer__get_data(buffer, fd);
489 if (!buffer->data) {
490 err = -ENOMEM;
491 goto out_put;
492 }
493 }
494
495 if (btsq->bts->snapshot_mode && !buffer->consecutive &&
496 intel_bts_do_fix_overlap(queue, buffer)) {
497 err = -ENOMEM;
498 goto out_put;
499 }
500
501 if (!btsq->bts->synth_opts.callchain &&
502 !btsq->bts->synth_opts.thread_stack && thread &&
503 (!old_buffer || btsq->bts->sampling_mode ||
504 (btsq->bts->snapshot_mode && !buffer->consecutive)))
505 thread_stack__set_trace_nr(thread, btsq->cpu, buffer->buffer_nr + 1);
506
507 err = intel_bts_process_buffer(btsq, buffer, thread);
508
509 auxtrace_buffer__drop_data(buffer);
510
511 btsq->buffer = auxtrace_buffer__next(queue, buffer);
512 if (btsq->buffer) {
513 if (timestamp)
514 *timestamp = btsq->buffer->reference;
515 } else {
516 if (!btsq->bts->sampling_mode)
517 btsq->done = 1;
518 }
519 out_put:
520 thread__put(thread);
521 return err;
522 }
523
intel_bts_flush_queue(struct intel_bts_queue * btsq)524 static int intel_bts_flush_queue(struct intel_bts_queue *btsq)
525 {
526 u64 ts = 0;
527 int ret;
528
529 while (1) {
530 ret = intel_bts_process_queue(btsq, &ts);
531 if (ret < 0)
532 return ret;
533 if (ret)
534 break;
535 }
536 return 0;
537 }
538
intel_bts_process_tid_exit(struct intel_bts * bts,pid_t tid)539 static int intel_bts_process_tid_exit(struct intel_bts *bts, pid_t tid)
540 {
541 struct auxtrace_queues *queues = &bts->queues;
542 unsigned int i;
543
544 for (i = 0; i < queues->nr_queues; i++) {
545 struct auxtrace_queue *queue = &bts->queues.queue_array[i];
546 struct intel_bts_queue *btsq = queue->priv;
547
548 if (btsq && btsq->tid == tid)
549 return intel_bts_flush_queue(btsq);
550 }
551 return 0;
552 }
553
intel_bts_process_queues(struct intel_bts * bts,u64 timestamp)554 static int intel_bts_process_queues(struct intel_bts *bts, u64 timestamp)
555 {
556 while (1) {
557 unsigned int queue_nr;
558 struct auxtrace_queue *queue;
559 struct intel_bts_queue *btsq;
560 u64 ts = 0;
561 int ret;
562
563 if (!bts->heap.heap_cnt)
564 return 0;
565
566 if (bts->heap.heap_array[0].ordinal > timestamp)
567 return 0;
568
569 queue_nr = bts->heap.heap_array[0].queue_nr;
570 queue = &bts->queues.queue_array[queue_nr];
571 btsq = queue->priv;
572
573 auxtrace_heap__pop(&bts->heap);
574
575 ret = intel_bts_process_queue(btsq, &ts);
576 if (ret < 0) {
577 auxtrace_heap__add(&bts->heap, queue_nr, ts);
578 return ret;
579 }
580
581 if (!ret) {
582 ret = auxtrace_heap__add(&bts->heap, queue_nr, ts);
583 if (ret < 0)
584 return ret;
585 } else {
586 btsq->on_heap = false;
587 }
588 }
589
590 return 0;
591 }
592
intel_bts_process_event(struct perf_session * session,union perf_event * event,struct perf_sample * sample,const struct perf_tool * tool)593 static int intel_bts_process_event(struct perf_session *session,
594 union perf_event *event,
595 struct perf_sample *sample,
596 const struct perf_tool *tool)
597 {
598 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
599 auxtrace);
600 u64 timestamp;
601 int err;
602
603 if (dump_trace)
604 return 0;
605
606 if (!tool->ordered_events) {
607 pr_err("Intel BTS requires ordered events\n");
608 return -EINVAL;
609 }
610
611 if (sample->time && sample->time != (u64)-1)
612 timestamp = perf_time_to_tsc(sample->time, &bts->tc);
613 else
614 timestamp = 0;
615
616 err = intel_bts_update_queues(bts);
617 if (err)
618 return err;
619
620 err = intel_bts_process_queues(bts, timestamp);
621 if (err)
622 return err;
623 if (event->header.type == PERF_RECORD_EXIT) {
624 err = intel_bts_process_tid_exit(bts, event->fork.tid);
625 if (err)
626 return err;
627 }
628
629 if (event->header.type == PERF_RECORD_AUX &&
630 (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
631 bts->synth_opts.errors)
632 err = intel_bts_lost(bts, sample);
633
634 return err;
635 }
636
intel_bts_process_auxtrace_event(struct perf_session * session,union perf_event * event,const struct perf_tool * tool __maybe_unused)637 static int intel_bts_process_auxtrace_event(struct perf_session *session,
638 union perf_event *event,
639 const struct perf_tool *tool __maybe_unused)
640 {
641 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
642 auxtrace);
643
644 if (bts->sampling_mode)
645 return 0;
646
647 if (!bts->data_queued) {
648 struct auxtrace_buffer *buffer;
649 off_t data_offset;
650 int fd = perf_data__fd(session->data);
651 int err;
652
653 if (perf_data__is_pipe(session->data)) {
654 data_offset = 0;
655 } else {
656 data_offset = lseek(fd, 0, SEEK_CUR);
657 if (data_offset == -1)
658 return -errno;
659 }
660
661 err = auxtrace_queues__add_event(&bts->queues, session, event,
662 data_offset, &buffer);
663 if (err)
664 return err;
665
666 /* Dump here now we have copied a piped trace out of the pipe */
667 if (dump_trace) {
668 if (auxtrace_buffer__get_data(buffer, fd)) {
669 intel_bts_dump_event(bts, buffer->data,
670 buffer->size);
671 auxtrace_buffer__put_data(buffer);
672 }
673 }
674 }
675
676 return 0;
677 }
678
intel_bts_flush(struct perf_session * session,const struct perf_tool * tool __maybe_unused)679 static int intel_bts_flush(struct perf_session *session,
680 const struct perf_tool *tool __maybe_unused)
681 {
682 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
683 auxtrace);
684 int ret;
685
686 if (dump_trace || bts->sampling_mode)
687 return 0;
688
689 if (!tool->ordered_events)
690 return -EINVAL;
691
692 ret = intel_bts_update_queues(bts);
693 if (ret < 0)
694 return ret;
695
696 return intel_bts_process_queues(bts, MAX_TIMESTAMP);
697 }
698
intel_bts_free_queue(void * priv)699 static void intel_bts_free_queue(void *priv)
700 {
701 struct intel_bts_queue *btsq = priv;
702
703 if (!btsq)
704 return;
705 free(btsq);
706 }
707
intel_bts_free_events(struct perf_session * session)708 static void intel_bts_free_events(struct perf_session *session)
709 {
710 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
711 auxtrace);
712 struct auxtrace_queues *queues = &bts->queues;
713 unsigned int i;
714
715 for (i = 0; i < queues->nr_queues; i++) {
716 intel_bts_free_queue(queues->queue_array[i].priv);
717 queues->queue_array[i].priv = NULL;
718 }
719 auxtrace_queues__free(queues);
720 }
721
intel_bts_free(struct perf_session * session)722 static void intel_bts_free(struct perf_session *session)
723 {
724 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
725 auxtrace);
726
727 auxtrace_heap__free(&bts->heap);
728 intel_bts_free_events(session);
729 session->auxtrace = NULL;
730 free(bts);
731 }
732
intel_bts_evsel_is_auxtrace(struct perf_session * session,struct evsel * evsel)733 static bool intel_bts_evsel_is_auxtrace(struct perf_session *session,
734 struct evsel *evsel)
735 {
736 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
737 auxtrace);
738
739 return evsel->core.attr.type == bts->pmu_type;
740 }
741
intel_bts_synth_events(struct intel_bts * bts,struct perf_session * session)742 static int intel_bts_synth_events(struct intel_bts *bts,
743 struct perf_session *session)
744 {
745 struct evlist *evlist = session->evlist;
746 struct evsel *evsel;
747 struct perf_event_attr attr;
748 bool found = false;
749 u64 id;
750 int err;
751
752 evlist__for_each_entry(evlist, evsel) {
753 if (evsel->core.attr.type == bts->pmu_type && evsel->core.ids) {
754 found = true;
755 break;
756 }
757 }
758
759 if (!found) {
760 pr_debug("There are no selected events with Intel BTS data\n");
761 return 0;
762 }
763
764 memset(&attr, 0, sizeof(struct perf_event_attr));
765 attr.size = sizeof(struct perf_event_attr);
766 attr.type = PERF_TYPE_HARDWARE;
767 attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK;
768 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
769 PERF_SAMPLE_PERIOD;
770 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
771 attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
772 attr.exclude_user = evsel->core.attr.exclude_user;
773 attr.exclude_kernel = evsel->core.attr.exclude_kernel;
774 attr.exclude_hv = evsel->core.attr.exclude_hv;
775 attr.exclude_host = evsel->core.attr.exclude_host;
776 attr.exclude_guest = evsel->core.attr.exclude_guest;
777 attr.sample_id_all = evsel->core.attr.sample_id_all;
778 attr.read_format = evsel->core.attr.read_format;
779
780 id = evsel->core.id[0] + 1000000000;
781 if (!id)
782 id = 1;
783
784 if (bts->synth_opts.branches) {
785 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
786 attr.sample_period = 1;
787 attr.sample_type |= PERF_SAMPLE_ADDR;
788 pr_debug("Synthesizing 'branches' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
789 id, (u64)attr.sample_type);
790 err = perf_session__deliver_synth_attr_event(session, &attr, id);
791 if (err) {
792 pr_err("%s: failed to synthesize 'branches' event type\n",
793 __func__);
794 return err;
795 }
796 bts->sample_branches = true;
797 bts->branches_sample_type = attr.sample_type;
798 bts->branches_id = id;
799 /*
800 * We only use sample types from PERF_SAMPLE_MASK so we can use
801 * __evsel__sample_size() here.
802 */
803 bts->branches_event_size = sizeof(struct perf_record_sample) +
804 __evsel__sample_size(attr.sample_type);
805 }
806
807 return 0;
808 }
809
810 static const char * const intel_bts_info_fmts[] = {
811 [INTEL_BTS_PMU_TYPE] = " PMU Type %"PRId64"\n",
812 [INTEL_BTS_TIME_SHIFT] = " Time Shift %"PRIu64"\n",
813 [INTEL_BTS_TIME_MULT] = " Time Multiplier %"PRIu64"\n",
814 [INTEL_BTS_TIME_ZERO] = " Time Zero %"PRIu64"\n",
815 [INTEL_BTS_CAP_USER_TIME_ZERO] = " Cap Time Zero %"PRId64"\n",
816 [INTEL_BTS_SNAPSHOT_MODE] = " Snapshot mode %"PRId64"\n",
817 };
818
intel_bts_print_info(__u64 * arr,int start,int finish)819 static void intel_bts_print_info(__u64 *arr, int start, int finish)
820 {
821 int i;
822
823 if (!dump_trace)
824 return;
825
826 for (i = start; i <= finish; i++)
827 fprintf(stdout, intel_bts_info_fmts[i], arr[i]);
828 }
829
intel_bts_process_auxtrace_info(union perf_event * event,struct perf_session * session)830 int intel_bts_process_auxtrace_info(union perf_event *event,
831 struct perf_session *session)
832 {
833 struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info;
834 size_t min_sz = sizeof(u64) * INTEL_BTS_SNAPSHOT_MODE;
835 struct intel_bts *bts;
836 int err;
837
838 if (auxtrace_info->header.size < sizeof(struct perf_record_auxtrace_info) +
839 min_sz)
840 return -EINVAL;
841
842 bts = zalloc(sizeof(struct intel_bts));
843 if (!bts)
844 return -ENOMEM;
845
846 err = auxtrace_queues__init(&bts->queues);
847 if (err)
848 goto err_free;
849
850 bts->session = session;
851 bts->machine = &session->machines.host; /* No kvm support */
852 bts->auxtrace_type = auxtrace_info->type;
853 bts->pmu_type = auxtrace_info->priv[INTEL_BTS_PMU_TYPE];
854 bts->tc.time_shift = auxtrace_info->priv[INTEL_BTS_TIME_SHIFT];
855 bts->tc.time_mult = auxtrace_info->priv[INTEL_BTS_TIME_MULT];
856 bts->tc.time_zero = auxtrace_info->priv[INTEL_BTS_TIME_ZERO];
857 bts->cap_user_time_zero =
858 auxtrace_info->priv[INTEL_BTS_CAP_USER_TIME_ZERO];
859 bts->snapshot_mode = auxtrace_info->priv[INTEL_BTS_SNAPSHOT_MODE];
860
861 bts->sampling_mode = false;
862
863 bts->auxtrace.process_event = intel_bts_process_event;
864 bts->auxtrace.process_auxtrace_event = intel_bts_process_auxtrace_event;
865 bts->auxtrace.flush_events = intel_bts_flush;
866 bts->auxtrace.free_events = intel_bts_free_events;
867 bts->auxtrace.free = intel_bts_free;
868 bts->auxtrace.evsel_is_auxtrace = intel_bts_evsel_is_auxtrace;
869 session->auxtrace = &bts->auxtrace;
870
871 intel_bts_print_info(&auxtrace_info->priv[0], INTEL_BTS_PMU_TYPE,
872 INTEL_BTS_SNAPSHOT_MODE);
873
874 if (dump_trace)
875 return 0;
876
877 if (session->itrace_synth_opts->set) {
878 bts->synth_opts = *session->itrace_synth_opts;
879 } else {
880 itrace_synth_opts__set_default(&bts->synth_opts,
881 session->itrace_synth_opts->default_no_sample);
882 bts->synth_opts.thread_stack =
883 session->itrace_synth_opts->thread_stack;
884 }
885
886 if (bts->synth_opts.calls)
887 bts->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
888 PERF_IP_FLAG_TRACE_END;
889 if (bts->synth_opts.returns)
890 bts->branches_filter |= PERF_IP_FLAG_RETURN |
891 PERF_IP_FLAG_TRACE_BEGIN;
892
893 err = intel_bts_synth_events(bts, session);
894 if (err)
895 goto err_free_queues;
896
897 err = auxtrace_queues__process_index(&bts->queues, session);
898 if (err)
899 goto err_free_queues;
900
901 if (bts->queues.populated)
902 bts->data_queued = true;
903
904 return 0;
905
906 err_free_queues:
907 auxtrace_queues__free(&bts->queues);
908 session->auxtrace = NULL;
909 err_free:
910 free(bts);
911 return err;
912 }
913