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 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 } 123 124 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 } 130 131 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 } 147 148 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 } 165 166 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 } 205 206 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 } 219 220 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 } 228 229 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 } 247 248 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 } 264 265 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 } 271 272 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 /*read_format=*/0, /*branch_sample_type=*/0, 307 &sample); 308 if (ret) 309 return ret; 310 } 311 312 ret = perf_session__deliver_synth_event(bts->session, &event, &sample); 313 if (ret) 314 pr_err("Intel BTS: failed to deliver branch event, error %d\n", 315 ret); 316 317 perf_sample__exit(&sample); 318 return ret; 319 } 320 321 static int intel_bts_get_next_insn(struct intel_bts_queue *btsq, u64 ip) 322 { 323 struct machine *machine = btsq->bts->machine; 324 struct thread *thread; 325 unsigned char buf[INTEL_PT_INSN_BUF_SZ]; 326 ssize_t len; 327 bool x86_64; 328 int err = -1; 329 330 thread = machine__find_thread(machine, -1, btsq->tid); 331 if (!thread) 332 return -1; 333 334 len = thread__memcpy(thread, machine, buf, ip, INTEL_PT_INSN_BUF_SZ, &x86_64); 335 if (len <= 0) 336 goto out_put; 337 338 if (intel_pt_get_insn(buf, len, x86_64, &btsq->intel_pt_insn)) 339 goto out_put; 340 341 err = 0; 342 out_put: 343 thread__put(thread); 344 return err; 345 } 346 347 static int intel_bts_synth_error(struct intel_bts *bts, int cpu, pid_t pid, 348 pid_t tid, u64 ip) 349 { 350 union perf_event event; 351 int err; 352 353 auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE, 354 INTEL_BTS_ERR_NOINSN, cpu, pid, tid, ip, 355 "Failed to get instruction", 0); 356 357 err = perf_session__deliver_synth_event(bts->session, &event, NULL); 358 if (err) 359 pr_err("Intel BTS: failed to deliver error event, error %d\n", 360 err); 361 362 return err; 363 } 364 365 static int intel_bts_get_branch_type(struct intel_bts_queue *btsq, 366 struct branch *branch) 367 { 368 int err; 369 370 if (!branch->from) { 371 if (branch->to) 372 btsq->sample_flags = PERF_IP_FLAG_BRANCH | 373 PERF_IP_FLAG_TRACE_BEGIN; 374 else 375 btsq->sample_flags = 0; 376 btsq->intel_pt_insn.length = 0; 377 } else if (!branch->to) { 378 btsq->sample_flags = PERF_IP_FLAG_BRANCH | 379 PERF_IP_FLAG_TRACE_END; 380 btsq->intel_pt_insn.length = 0; 381 } else { 382 err = intel_bts_get_next_insn(btsq, branch->from); 383 if (err) { 384 btsq->sample_flags = 0; 385 btsq->intel_pt_insn.length = 0; 386 if (!btsq->bts->synth_opts.errors) 387 return 0; 388 err = intel_bts_synth_error(btsq->bts, btsq->cpu, 389 btsq->pid, btsq->tid, 390 branch->from); 391 return err; 392 } 393 btsq->sample_flags = intel_pt_insn_type(btsq->intel_pt_insn.op); 394 /* Check for an async branch into the kernel */ 395 if (!machine__kernel_ip(btsq->bts->machine, branch->from) && 396 machine__kernel_ip(btsq->bts->machine, branch->to) && 397 btsq->sample_flags != (PERF_IP_FLAG_BRANCH | 398 PERF_IP_FLAG_CALL | 399 PERF_IP_FLAG_SYSCALLRET)) 400 btsq->sample_flags = PERF_IP_FLAG_BRANCH | 401 PERF_IP_FLAG_CALL | 402 PERF_IP_FLAG_ASYNC | 403 PERF_IP_FLAG_INTERRUPT; 404 } 405 406 return 0; 407 } 408 409 static int intel_bts_process_buffer(struct intel_bts_queue *btsq, 410 struct auxtrace_buffer *buffer, 411 struct thread *thread) 412 { 413 struct branch *branch; 414 size_t sz, bsz = sizeof(struct branch); 415 u32 filter = btsq->bts->branches_filter; 416 int err = 0; 417 418 if (buffer->use_data) { 419 sz = buffer->use_size; 420 branch = buffer->use_data; 421 } else { 422 sz = buffer->size; 423 branch = buffer->data; 424 } 425 426 if (!btsq->bts->sample_branches) 427 return 0; 428 429 for (; sz > bsz; branch += 1, sz -= bsz) { 430 if (!branch->from && !branch->to) 431 continue; 432 intel_bts_get_branch_type(btsq, branch); 433 if (btsq->bts->synth_opts.thread_stack) 434 thread_stack__event(thread, btsq->cpu, btsq->sample_flags, 435 le64_to_cpu(branch->from), 436 le64_to_cpu(branch->to), 437 btsq->intel_pt_insn.length, 438 buffer->buffer_nr + 1, true, 0, 0); 439 if (filter && !(filter & btsq->sample_flags)) 440 continue; 441 err = intel_bts_synth_branch_sample(btsq, branch); 442 if (err) 443 break; 444 } 445 return err; 446 } 447 448 static int intel_bts_process_queue(struct intel_bts_queue *btsq, u64 *timestamp) 449 { 450 struct auxtrace_buffer *buffer = btsq->buffer, *old_buffer = buffer; 451 struct auxtrace_queue *queue; 452 struct thread *thread; 453 int err; 454 455 if (btsq->done) 456 return 1; 457 458 if (btsq->pid == -1) { 459 thread = machine__find_thread(btsq->bts->machine, -1, 460 btsq->tid); 461 if (thread) 462 btsq->pid = thread__pid(thread); 463 } else { 464 thread = machine__findnew_thread(btsq->bts->machine, btsq->pid, 465 btsq->tid); 466 } 467 468 queue = &btsq->bts->queues.queue_array[btsq->queue_nr]; 469 470 if (!buffer) 471 buffer = auxtrace_buffer__next(queue, NULL); 472 473 if (!buffer) { 474 if (!btsq->bts->sampling_mode) 475 btsq->done = 1; 476 err = 1; 477 goto out_put; 478 } 479 480 /* Currently there is no support for split buffers */ 481 if (buffer->consecutive) { 482 err = -EINVAL; 483 goto out_put; 484 } 485 486 if (!buffer->data) { 487 int fd = perf_data__fd(btsq->bts->session->data); 488 489 buffer->data = auxtrace_buffer__get_data(buffer, fd); 490 if (!buffer->data) { 491 err = -ENOMEM; 492 goto out_put; 493 } 494 } 495 496 if (btsq->bts->snapshot_mode && !buffer->consecutive && 497 intel_bts_do_fix_overlap(queue, buffer)) { 498 err = -ENOMEM; 499 goto out_put; 500 } 501 502 if (!btsq->bts->synth_opts.callchain && 503 !btsq->bts->synth_opts.thread_stack && thread && 504 (!old_buffer || btsq->bts->sampling_mode || 505 (btsq->bts->snapshot_mode && !buffer->consecutive))) 506 thread_stack__set_trace_nr(thread, btsq->cpu, buffer->buffer_nr + 1); 507 508 err = intel_bts_process_buffer(btsq, buffer, thread); 509 510 auxtrace_buffer__drop_data(buffer); 511 512 btsq->buffer = auxtrace_buffer__next(queue, buffer); 513 if (btsq->buffer) { 514 if (timestamp) 515 *timestamp = btsq->buffer->reference; 516 } else { 517 if (!btsq->bts->sampling_mode) 518 btsq->done = 1; 519 } 520 out_put: 521 thread__put(thread); 522 return err; 523 } 524 525 static int intel_bts_flush_queue(struct intel_bts_queue *btsq) 526 { 527 u64 ts = 0; 528 int ret; 529 530 while (1) { 531 ret = intel_bts_process_queue(btsq, &ts); 532 if (ret < 0) 533 return ret; 534 if (ret) 535 break; 536 } 537 return 0; 538 } 539 540 static int intel_bts_process_tid_exit(struct intel_bts *bts, pid_t tid) 541 { 542 struct auxtrace_queues *queues = &bts->queues; 543 unsigned int i; 544 545 for (i = 0; i < queues->nr_queues; i++) { 546 struct auxtrace_queue *queue = &bts->queues.queue_array[i]; 547 struct intel_bts_queue *btsq = queue->priv; 548 549 if (btsq && btsq->tid == tid) 550 return intel_bts_flush_queue(btsq); 551 } 552 return 0; 553 } 554 555 static int intel_bts_process_queues(struct intel_bts *bts, u64 timestamp) 556 { 557 while (1) { 558 unsigned int queue_nr; 559 struct auxtrace_queue *queue; 560 struct intel_bts_queue *btsq; 561 u64 ts = 0; 562 int ret; 563 564 if (!bts->heap.heap_cnt) 565 return 0; 566 567 if (bts->heap.heap_array[0].ordinal > timestamp) 568 return 0; 569 570 queue_nr = bts->heap.heap_array[0].queue_nr; 571 queue = &bts->queues.queue_array[queue_nr]; 572 btsq = queue->priv; 573 574 auxtrace_heap__pop(&bts->heap); 575 576 ret = intel_bts_process_queue(btsq, &ts); 577 if (ret < 0) { 578 auxtrace_heap__add(&bts->heap, queue_nr, ts); 579 return ret; 580 } 581 582 if (!ret) { 583 ret = auxtrace_heap__add(&bts->heap, queue_nr, ts); 584 if (ret < 0) 585 return ret; 586 } else { 587 btsq->on_heap = false; 588 } 589 } 590 591 return 0; 592 } 593 594 static int intel_bts_process_event(struct perf_session *session, 595 union perf_event *event, 596 struct perf_sample *sample, 597 const struct perf_tool *tool) 598 { 599 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, 600 auxtrace); 601 u64 timestamp; 602 int err; 603 604 if (dump_trace) 605 return 0; 606 607 if (!tool->ordered_events) { 608 pr_err("Intel BTS requires ordered events\n"); 609 return -EINVAL; 610 } 611 612 if (sample->time && sample->time != (u64)-1) 613 timestamp = perf_time_to_tsc(sample->time, &bts->tc); 614 else 615 timestamp = 0; 616 617 err = intel_bts_update_queues(bts); 618 if (err) 619 return err; 620 621 err = intel_bts_process_queues(bts, timestamp); 622 if (err) 623 return err; 624 if (event->header.type == PERF_RECORD_EXIT) { 625 err = intel_bts_process_tid_exit(bts, event->fork.tid); 626 if (err) 627 return err; 628 } 629 630 if (event->header.type == PERF_RECORD_AUX && 631 (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) && 632 bts->synth_opts.errors) 633 err = intel_bts_lost(bts, sample); 634 635 return err; 636 } 637 638 static int intel_bts_process_auxtrace_event(struct perf_session *session, 639 union perf_event *event, 640 const struct perf_tool *tool __maybe_unused) 641 { 642 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, 643 auxtrace); 644 645 if (bts->sampling_mode) 646 return 0; 647 648 if (!bts->data_queued) { 649 struct auxtrace_buffer *buffer; 650 off_t data_offset; 651 int fd = perf_data__fd(session->data); 652 int err; 653 654 if (perf_data__is_pipe(session->data)) { 655 data_offset = 0; 656 } else { 657 data_offset = lseek(fd, 0, SEEK_CUR); 658 if (data_offset == -1) 659 return -errno; 660 } 661 662 err = auxtrace_queues__add_event(&bts->queues, session, event, 663 data_offset, &buffer); 664 if (err) 665 return err; 666 667 /* Dump here now we have copied a piped trace out of the pipe */ 668 if (dump_trace) { 669 if (auxtrace_buffer__get_data(buffer, fd)) { 670 intel_bts_dump_event(bts, buffer->data, 671 buffer->size); 672 auxtrace_buffer__put_data(buffer); 673 } 674 } 675 } 676 677 return 0; 678 } 679 680 static int intel_bts_flush(struct perf_session *session, 681 const struct perf_tool *tool __maybe_unused) 682 { 683 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, 684 auxtrace); 685 int ret; 686 687 if (dump_trace || bts->sampling_mode) 688 return 0; 689 690 if (!tool->ordered_events) 691 return -EINVAL; 692 693 ret = intel_bts_update_queues(bts); 694 if (ret < 0) 695 return ret; 696 697 return intel_bts_process_queues(bts, MAX_TIMESTAMP); 698 } 699 700 static void intel_bts_free_queue(void *priv) 701 { 702 struct intel_bts_queue *btsq = priv; 703 704 if (!btsq) 705 return; 706 free(btsq); 707 } 708 709 static void intel_bts_free_events(struct perf_session *session) 710 { 711 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, 712 auxtrace); 713 struct auxtrace_queues *queues = &bts->queues; 714 unsigned int i; 715 716 for (i = 0; i < queues->nr_queues; i++) { 717 intel_bts_free_queue(queues->queue_array[i].priv); 718 queues->queue_array[i].priv = NULL; 719 } 720 auxtrace_queues__free(queues); 721 } 722 723 static void intel_bts_free(struct perf_session *session) 724 { 725 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, 726 auxtrace); 727 728 auxtrace_heap__free(&bts->heap); 729 intel_bts_free_events(session); 730 session->auxtrace = NULL; 731 free(bts); 732 } 733 734 static bool intel_bts_evsel_is_auxtrace(struct perf_session *session, 735 struct evsel *evsel) 736 { 737 struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts, 738 auxtrace); 739 740 return evsel->core.attr.type == bts->pmu_type; 741 } 742 743 static int intel_bts_synth_events(struct intel_bts *bts, 744 struct perf_session *session) 745 { 746 struct evlist *evlist = session->evlist; 747 struct evsel *evsel; 748 struct perf_event_attr attr; 749 bool found = false; 750 u64 id; 751 int err; 752 753 evlist__for_each_entry(evlist, evsel) { 754 if (evsel->core.attr.type == bts->pmu_type && evsel->core.ids) { 755 found = true; 756 break; 757 } 758 } 759 760 if (!found) { 761 pr_debug("There are no selected events with Intel BTS data\n"); 762 return 0; 763 } 764 765 memset(&attr, 0, sizeof(struct perf_event_attr)); 766 attr.size = sizeof(struct perf_event_attr); 767 attr.type = PERF_TYPE_HARDWARE; 768 attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK; 769 attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID | 770 PERF_SAMPLE_PERIOD; 771 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME; 772 attr.sample_type &= ~(u64)PERF_SAMPLE_CPU; 773 attr.exclude_user = evsel->core.attr.exclude_user; 774 attr.exclude_kernel = evsel->core.attr.exclude_kernel; 775 attr.exclude_hv = evsel->core.attr.exclude_hv; 776 attr.exclude_host = evsel->core.attr.exclude_host; 777 attr.exclude_guest = evsel->core.attr.exclude_guest; 778 attr.sample_id_all = evsel->core.attr.sample_id_all; 779 attr.read_format = evsel->core.attr.read_format; 780 781 id = auxtrace_synth_id_range_start(evsel); 782 783 if (bts->synth_opts.branches) { 784 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; 785 attr.sample_period = 1; 786 attr.sample_type |= PERF_SAMPLE_ADDR; 787 pr_debug("Synthesizing 'branches' event with id %" PRIu64 " sample type %#" PRIx64 "\n", 788 id, (u64)attr.sample_type); 789 err = perf_session__deliver_synth_attr_event(session, &attr, id); 790 if (err) { 791 pr_err("%s: failed to synthesize 'branches' event type\n", 792 __func__); 793 return err; 794 } 795 bts->sample_branches = true; 796 bts->branches_sample_type = attr.sample_type; 797 bts->branches_id = id; 798 /* 799 * We only use sample types from PERF_SAMPLE_MASK so we can use 800 * __evsel__sample_size() here. 801 */ 802 bts->branches_event_size = sizeof(struct perf_record_sample) + 803 __evsel__sample_size(attr.sample_type); 804 } 805 806 return 0; 807 } 808 809 static const char * const intel_bts_info_fmts[] = { 810 [INTEL_BTS_PMU_TYPE] = " PMU Type %"PRId64"\n", 811 [INTEL_BTS_TIME_SHIFT] = " Time Shift %"PRIu64"\n", 812 [INTEL_BTS_TIME_MULT] = " Time Multiplier %"PRIu64"\n", 813 [INTEL_BTS_TIME_ZERO] = " Time Zero %"PRIu64"\n", 814 [INTEL_BTS_CAP_USER_TIME_ZERO] = " Cap Time Zero %"PRId64"\n", 815 [INTEL_BTS_SNAPSHOT_MODE] = " Snapshot mode %"PRId64"\n", 816 }; 817 818 static void intel_bts_print_info(__u64 *arr, int start, int finish) 819 { 820 int i; 821 822 if (!dump_trace) 823 return; 824 825 for (i = start; i <= finish; i++) 826 fprintf(stdout, intel_bts_info_fmts[i], arr[i]); 827 } 828 829 int intel_bts_process_auxtrace_info(union perf_event *event, 830 struct perf_session *session) 831 { 832 struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info; 833 size_t min_sz = sizeof(u64) * INTEL_BTS_SNAPSHOT_MODE; 834 struct intel_bts *bts; 835 int err; 836 837 if (auxtrace_info->header.size < sizeof(struct perf_record_auxtrace_info) + 838 min_sz) 839 return -EINVAL; 840 841 bts = zalloc(sizeof(struct intel_bts)); 842 if (!bts) 843 return -ENOMEM; 844 845 err = auxtrace_queues__init(&bts->queues); 846 if (err) 847 goto err_free; 848 849 bts->session = session; 850 bts->machine = &session->machines.host; /* No kvm support */ 851 bts->auxtrace_type = auxtrace_info->type; 852 bts->pmu_type = auxtrace_info->priv[INTEL_BTS_PMU_TYPE]; 853 bts->tc.time_shift = auxtrace_info->priv[INTEL_BTS_TIME_SHIFT]; 854 bts->tc.time_mult = auxtrace_info->priv[INTEL_BTS_TIME_MULT]; 855 bts->tc.time_zero = auxtrace_info->priv[INTEL_BTS_TIME_ZERO]; 856 bts->cap_user_time_zero = 857 auxtrace_info->priv[INTEL_BTS_CAP_USER_TIME_ZERO]; 858 bts->snapshot_mode = auxtrace_info->priv[INTEL_BTS_SNAPSHOT_MODE]; 859 860 bts->sampling_mode = false; 861 862 bts->auxtrace.process_event = intel_bts_process_event; 863 bts->auxtrace.process_auxtrace_event = intel_bts_process_auxtrace_event; 864 bts->auxtrace.flush_events = intel_bts_flush; 865 bts->auxtrace.free_events = intel_bts_free_events; 866 bts->auxtrace.free = intel_bts_free; 867 bts->auxtrace.evsel_is_auxtrace = intel_bts_evsel_is_auxtrace; 868 session->auxtrace = &bts->auxtrace; 869 870 intel_bts_print_info(&auxtrace_info->priv[0], INTEL_BTS_PMU_TYPE, 871 INTEL_BTS_SNAPSHOT_MODE); 872 873 if (dump_trace) 874 return 0; 875 876 if (session->itrace_synth_opts->set) { 877 bts->synth_opts = *session->itrace_synth_opts; 878 } else { 879 itrace_synth_opts__set_default(&bts->synth_opts, 880 session->itrace_synth_opts->default_no_sample); 881 bts->synth_opts.thread_stack = 882 session->itrace_synth_opts->thread_stack; 883 } 884 885 if (bts->synth_opts.calls) 886 bts->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC | 887 PERF_IP_FLAG_TRACE_END; 888 if (bts->synth_opts.returns) 889 bts->branches_filter |= PERF_IP_FLAG_RETURN | 890 PERF_IP_FLAG_TRACE_BEGIN; 891 892 err = intel_bts_synth_events(bts, session); 893 if (err) 894 goto err_free_queues; 895 896 err = auxtrace_queues__process_index(&bts->queues, session); 897 if (err) 898 goto err_free_queues; 899 900 if (bts->queues.populated) 901 bts->data_queued = true; 902 903 return 0; 904 905 err_free_queues: 906 auxtrace_queues__free(&bts->queues); 907 session->auxtrace = NULL; 908 err_free: 909 free(bts); 910 return err; 911 } 912