tools/perf/util/intel-bts.c

   1 /*
   2  * intel-bts.c: Intel Processor Trace support
   3  * Copyright (c) 2013-2015, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  */
  15
  16 #include <endian.h>
  17 #include <byteswap.h>
  18 #include <linux/kernel.h>
  19 #include <linux/types.h>
  20 #include <linux/bitops.h>
  21 #include <linux/log2.h>
  22
  23 #include "cpumap.h"
  24 #include "color.h"
  25 #include "evsel.h"
  26 #include "evlist.h"
  27 #include "machine.h"
  28 #include "session.h"
  29 #include "util.h"
  30 #include "thread.h"
  31 #include "thread-stack.h"
  32 #include "debug.h"
  33 #include "tsc.h"
  34 #include "auxtrace.h"
  35 #include "intel-pt-decoder/intel-pt-insn-decoder.h"
  36 #include "intel-bts.h"
  37
  38 #define MAX_TIMESTAMP (~0ULL)
  39
  40 #define INTEL_BTS_ERR_NOINSN  5
  41 #define INTEL_BTS_ERR_LOST    9
  42
  43 #if __BYTE_ORDER == __BIG_ENDIAN
  44 #define le64_to_cpu bswap_64
  45 #else
  46 #define le64_to_cpu
  47 #endif
  48
  49 struct intel_bts {
  50         struct auxtrace                 auxtrace;
  51         struct auxtrace_queues          queues;
  52         struct auxtrace_heap            heap;
  53         u32                             auxtrace_type;
  54         struct perf_session             *session;
  55         struct machine                  *machine;
  56         bool                            sampling_mode;
  57         bool                            snapshot_mode;
  58         bool                            data_queued;
  59         u32                             pmu_type;
  60         struct perf_tsc_conversion      tc;
  61         bool                            cap_user_time_zero;
  62         struct itrace_synth_opts        synth_opts;
  63         bool                            sample_branches;
  64         u32                             branches_filter;
  65         u64                             branches_sample_type;
  66         u64                             branches_id;
  67         size_t                          branches_event_size;
  68         bool                            synth_needs_swap;
  69         unsigned long                   num_events;
  70 };
  71
  72 struct intel_bts_queue {
  73         struct intel_bts        *bts;
  74         unsigned int            queue_nr;
  75         struct auxtrace_buffer  *buffer;
  76         bool                    on_heap;
  77         bool                    done;
  78         pid_t                   pid;
  79         pid_t                   tid;
  80         int                     cpu;
  81         u64                     time;
  82         struct intel_pt_insn    intel_pt_insn;
  83         u32                     sample_flags;
  84 };
  85
  86 struct branch {
  87         u64 from;
  88         u64 to;
  89         u64 misc;
  90 };
  91
  92 static void intel_bts_dump(struct intel_bts *bts __maybe_unused,
  93                            unsigned char *buf, size_t len)
  94 {
  95         struct branch *branch;
  96         size_t i, pos = 0, br_sz = sizeof(struct branch), sz;
  97         const char *color = PERF_COLOR_BLUE;
  98
  99         color_fprintf(stdout, color,
 100                       ". ... Intel BTS data: size %zu bytes\n",
 101                       len);
 102
 103         while (len) {
 104                 if (len >= br_sz)
 105                         sz = br_sz;
 106                 else
 107                         sz = len;
 108                 printf(".");
 109                 color_fprintf(stdout, color, "  %08x: ", pos);
 110                 for (i = 0; i < sz; i++)
 111                         color_fprintf(stdout, color, " %02x", buf[i]);
 112                 for (; i < br_sz; i++)
 113                         color_fprintf(stdout, color, "   ");
 114                 if (len >= br_sz) {
 115                         branch = (struct branch *)buf;
 116                         color_fprintf(stdout, color, " %"PRIx64" -> %"PRIx64" %s\n",
 117                                       le64_to_cpu(branch->from),
 118                                       le64_to_cpu(branch->to),
 119                                       le64_to_cpu(branch->misc) & 0x10 ?
 120                                                         "pred" : "miss");
 121                 } else {
 122                         color_fprintf(stdout, color, " Bad record!\n");
 123                 }
 124                 pos += sz;
 125                 buf += sz;
 126                 len -= sz;
 127         }
 128 }
 129
 130 static void intel_bts_dump_event(struct intel_bts *bts, unsigned char *buf,
 131                                  size_t len)
 132 {
 133         printf(".\n");
 134         intel_bts_dump(bts, buf, len);
 135 }
 136
 137 static int intel_bts_lost(struct intel_bts *bts, struct perf_sample *sample)
 138 {
 139         union perf_event event;
 140         int err;
 141
 142         auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
 143                              INTEL_BTS_ERR_LOST, sample->cpu, sample->pid,
 144                              sample->tid, 0, "Lost trace data");
 145
 146         err = perf_session__deliver_synth_event(bts->session, &event, NULL);
 147         if (err)
 148                 pr_err("Intel BTS: failed to deliver error event, error %d\n",
 149                        err);
 150
 151         return err;
 152 }
 153
 154 static struct intel_bts_queue *intel_bts_alloc_queue(struct intel_bts *bts,
 155                                                      unsigned int queue_nr)
 156 {
 157         struct intel_bts_queue *btsq;
 158
 159         btsq = zalloc(sizeof(struct intel_bts_queue));
 160         if (!btsq)
 161                 return NULL;
 162
 163         btsq->bts = bts;
 164         btsq->queue_nr = queue_nr;
 165         btsq->pid = -1;
 166         btsq->tid = -1;
 167         btsq->cpu = -1;
 168
 169         return btsq;
 170 }
 171
 172 static int intel_bts_setup_queue(struct intel_bts *bts,
 173                                  struct auxtrace_queue *queue,
 174                                  unsigned int queue_nr)
 175 {
 176         struct intel_bts_queue *btsq = queue->priv;
 177
 178         if (list_empty(&queue->head))
 179                 return 0;
 180
 181         if (!btsq) {
 182                 btsq = intel_bts_alloc_queue(bts, queue_nr);
 183                 if (!btsq)
 184                         return -ENOMEM;
 185                 queue->priv = btsq;
 186
 187                 if (queue->cpu != -1)
 188                         btsq->cpu = queue->cpu;
 189                 btsq->tid = queue->tid;
 190         }
 191
 192         if (bts->sampling_mode)
 193                 return 0;
 194
 195         if (!btsq->on_heap && !btsq->buffer) {
 196                 int ret;
 197
 198                 btsq->buffer = auxtrace_buffer__next(queue, NULL);
 199                 if (!btsq->buffer)
 200                         return 0;
 201
 202                 ret = auxtrace_heap__add(&bts->heap, queue_nr,
 203                                          btsq->buffer->reference);
 204                 if (ret)
 205                         return ret;
 206                 btsq->on_heap = true;
 207         }
 208
 209         return 0;
 210 }
 211
 212 static int intel_bts_setup_queues(struct intel_bts *bts)
 213 {
 214         unsigned int i;
 215         int ret;
 216
 217         for (i = 0; i < bts->queues.nr_queues; i++) {
 218                 ret = intel_bts_setup_queue(bts, &bts->queues.queue_array[i],
 219                                             i);
 220                 if (ret)
 221                         return ret;
 222         }
 223         return 0;
 224 }
 225
 226 static inline int intel_bts_update_queues(struct intel_bts *bts)
 227 {
 228         if (bts->queues.new_data) {
 229                 bts->queues.new_data = false;
 230                 return intel_bts_setup_queues(bts);
 231         }
 232         return 0;
 233 }
 234
 235 static unsigned char *intel_bts_find_overlap(unsigned char *buf_a, size_t len_a,
 236                                              unsigned char *buf_b, size_t len_b)
 237 {
 238         size_t offs, len;
 239
 240         if (len_a > len_b)
 241                 offs = len_a - len_b;
 242         else
 243                 offs = 0;
 244
 245         for (; offs < len_a; offs += sizeof(struct branch)) {
 246                 len = len_a - offs;
 247                 if (!memcmp(buf_a + offs, buf_b, len))
 248                         return buf_b + len;
 249         }
 250
 251         return buf_b;
 252 }
 253
 254 static int intel_bts_do_fix_overlap(struct auxtrace_queue *queue,
 255                                     struct auxtrace_buffer *b)
 256 {
 257         struct auxtrace_buffer *a;
 258         void *start;
 259
 260         if (b->list.prev == &queue->head)
 261                 return 0;
 262         a = list_entry(b->list.prev, struct auxtrace_buffer, list);
 263         start = intel_bts_find_overlap(a->data, a->size, b->data, b->size);
 264         if (!start)
 265                 return -EINVAL;
 266         b->use_size = b->data + b->size - start;
 267         b->use_data = start;
 268         return 0;
 269 }
 270
 271 static int intel_bts_synth_branch_sample(struct intel_bts_queue *btsq,
 272                                          struct branch *branch)
 273 {
 274         int ret;
 275         struct intel_bts *bts = btsq->bts;
 276         union perf_event event;
 277         struct perf_sample sample = { .ip = 0, };
 278
 279         if (bts->synth_opts.initial_skip &&
 280             bts->num_events++ <= bts->synth_opts.initial_skip)
 281                 return 0;
 282
 283         event.sample.header.type = PERF_RECORD_SAMPLE;
 284         event.sample.header.misc = PERF_RECORD_MISC_USER;
 285         event.sample.header.size = sizeof(struct perf_event_header);
 286
 287         sample.cpumode = PERF_RECORD_MISC_USER;
 288         sample.ip = le64_to_cpu(branch->from);
 289         sample.pid = btsq->pid;
 290         sample.tid = btsq->tid;
 291         sample.addr = le64_to_cpu(branch->to);
 292         sample.id = btsq->bts->branches_id;
 293         sample.stream_id = btsq->bts->branches_id;
 294         sample.period = 1;
 295         sample.cpu = btsq->cpu;
 296         sample.flags = btsq->sample_flags;
 297         sample.insn_len = btsq->intel_pt_insn.length;
 298
 299         if (bts->synth_opts.inject) {
 300                 event.sample.header.size = bts->branches_event_size;
 301                 ret = perf_event__synthesize_sample(&event,
 302                                                     bts->branches_sample_type,
 303                                                     0, &sample,
 304                                                     bts->synth_needs_swap);
 305                 if (ret)
 306                         return ret;
 307         }
 308
 309         ret = perf_session__deliver_synth_event(bts->session, &event, &sample);
 310         if (ret)
 311                 pr_err("Intel BTS: failed to deliver branch event, error %d\n",
 312                        ret);
 313
 314         return ret;
 315 }
 316
 317 static int intel_bts_get_next_insn(struct intel_bts_queue *btsq, u64 ip)
 318 {
 319         struct machine *machine = btsq->bts->machine;
 320         struct thread *thread;
 321         struct addr_location al;
 322         unsigned char buf[INTEL_PT_INSN_BUF_SZ];
 323         ssize_t len;
 324         int x86_64;
 325         uint8_t cpumode;
 326         int err = -1;
 327
 328         if (machine__kernel_ip(machine, ip))
 329                 cpumode = PERF_RECORD_MISC_KERNEL;
 330         else
 331                 cpumode = PERF_RECORD_MISC_USER;
 332
 333         thread = machine__find_thread(machine, -1, btsq->tid);
 334         if (!thread)
 335                 return -1;
 336
 337         thread__find_addr_map(thread, cpumode, MAP__FUNCTION, ip, &al);
 338         if (!al.map || !al.map->dso)
 339                 goto out_put;
 340
 341         len = dso__data_read_addr(al.map->dso, al.map, machine, ip, buf,
 342                                   INTEL_PT_INSN_BUF_SZ);
 343         if (len <= 0)
 344                 goto out_put;
 345
 346         /* Load maps to ensure dso->is_64_bit has been updated */
 347         map__load(al.map);
 348
 349         x86_64 = al.map->dso->is_64_bit;
 350
 351         if (intel_pt_get_insn(buf, len, x86_64, &btsq->intel_pt_insn))
 352                 goto out_put;
 353
 354         err = 0;
 355 out_put:
 356         thread__put(thread);
 357         return err;
 358 }
 359
 360 static int intel_bts_synth_error(struct intel_bts *bts, int cpu, pid_t pid,
 361                                  pid_t tid, u64 ip)
 362 {
 363         union perf_event event;
 364         int err;
 365
 366         auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
 367                              INTEL_BTS_ERR_NOINSN, cpu, pid, tid, ip,
 368                              "Failed to get instruction");
 369
 370         err = perf_session__deliver_synth_event(bts->session, &event, NULL);
 371         if (err)
 372                 pr_err("Intel BTS: failed to deliver error event, error %d\n",
 373                        err);
 374
 375         return err;
 376 }
 377
 378 static int intel_bts_get_branch_type(struct intel_bts_queue *btsq,
 379                                      struct branch *branch)
 380 {
 381         int err;
 382
 383         if (!branch->from) {
 384                 if (branch->to)
 385                         btsq->sample_flags = PERF_IP_FLAG_BRANCH |
 386                                              PERF_IP_FLAG_TRACE_BEGIN;
 387                 else
 388                         btsq->sample_flags = 0;
 389                 btsq->intel_pt_insn.length = 0;
 390         } else if (!branch->to) {
 391                 btsq->sample_flags = PERF_IP_FLAG_BRANCH |
 392                                      PERF_IP_FLAG_TRACE_END;
 393                 btsq->intel_pt_insn.length = 0;
 394         } else {
 395                 err = intel_bts_get_next_insn(btsq, branch->from);
 396                 if (err) {
 397                         btsq->sample_flags = 0;
 398                         btsq->intel_pt_insn.length = 0;
 399                         if (!btsq->bts->synth_opts.errors)
 400                                 return 0;
 401                         err = intel_bts_synth_error(btsq->bts, btsq->cpu,
 402                                                     btsq->pid, btsq->tid,
 403                                                     branch->from);
 404                         return err;
 405                 }
 406                 btsq->sample_flags = intel_pt_insn_type(btsq->intel_pt_insn.op);
 407                 /* Check for an async branch into the kernel */
 408                 if (!machine__kernel_ip(btsq->bts->machine, branch->from) &&
 409                     machine__kernel_ip(btsq->bts->machine, branch->to) &&
 410                     btsq->sample_flags != (PERF_IP_FLAG_BRANCH |
 411                                            PERF_IP_FLAG_CALL |
 412                                            PERF_IP_FLAG_SYSCALLRET))
 413                         btsq->sample_flags = PERF_IP_FLAG_BRANCH |
 414                                              PERF_IP_FLAG_CALL |
 415                                              PERF_IP_FLAG_ASYNC |
 416                                              PERF_IP_FLAG_INTERRUPT;
 417         }
 418
 419         return 0;
 420 }
 421
 422 static int intel_bts_process_buffer(struct intel_bts_queue *btsq,
 423                                     struct auxtrace_buffer *buffer,
 424                                     struct thread *thread)
 425 {
 426         struct branch *branch;
 427         size_t sz, bsz = sizeof(struct branch);
 428         u32 filter = btsq->bts->branches_filter;
 429         int err = 0;
 430
 431         if (buffer->use_data) {
 432                 sz = buffer->use_size;
 433                 branch = buffer->use_data;
 434         } else {
 435                 sz = buffer->size;
 436                 branch = buffer->data;
 437         }
 438
 439         if (!btsq->bts->sample_branches)
 440                 return 0;
 441
 442         for (; sz > bsz; branch += 1, sz -= bsz) {
 443                 if (!branch->from && !branch->to)
 444                         continue;
 445                 intel_bts_get_branch_type(btsq, branch);
 446                 if (btsq->bts->synth_opts.thread_stack)
 447                         thread_stack__event(thread, btsq->sample_flags,
 448                                             le64_to_cpu(branch->from),
 449                                             le64_to_cpu(branch->to),
 450                                             btsq->intel_pt_insn.length,
 451                                             buffer->buffer_nr + 1);
 452                 if (filter && !(filter & btsq->sample_flags))
 453                         continue;
 454                 err = intel_bts_synth_branch_sample(btsq, branch);
 455                 if (err)
 456                         break;
 457         }
 458         return err;
 459 }
 460
 461 static int intel_bts_process_queue(struct intel_bts_queue *btsq, u64 *timestamp)
 462 {
 463         struct auxtrace_buffer *buffer = btsq->buffer, *old_buffer = buffer;
 464         struct auxtrace_queue *queue;
 465         struct thread *thread;
 466         int err;
 467
 468         if (btsq->done)
 469                 return 1;
 470
 471         if (btsq->pid == -1) {
 472                 thread = machine__find_thread(btsq->bts->machine, -1,
 473                                               btsq->tid);
 474                 if (thread)
 475                         btsq->pid = thread->pid_;
 476         } else {
 477                 thread = machine__findnew_thread(btsq->bts->machine, btsq->pid,
 478                                                  btsq->tid);
 479         }
 480
 481         queue = &btsq->bts->queues.queue_array[btsq->queue_nr];
 482
 483         if (!buffer)
 484                 buffer = auxtrace_buffer__next(queue, NULL);
 485
 486         if (!buffer) {
 487                 if (!btsq->bts->sampling_mode)
 488                         btsq->done = 1;
 489                 err = 1;
 490                 goto out_put;
 491         }
 492
 493         /* Currently there is no support for split buffers */
 494         if (buffer->consecutive) {
 495                 err = -EINVAL;
 496                 goto out_put;
 497         }
 498
 499         if (!buffer->data) {
 500                 int fd = perf_data_file__fd(btsq->bts->session->file);
 501
 502                 buffer->data = auxtrace_buffer__get_data(buffer, fd);
 503                 if (!buffer->data) {
 504                         err = -ENOMEM;
 505                         goto out_put;
 506                 }
 507         }
 508
 509         if (btsq->bts->snapshot_mode && !buffer->consecutive &&
 510             intel_bts_do_fix_overlap(queue, buffer)) {
 511                 err = -ENOMEM;
 512                 goto out_put;
 513         }
 514
 515         if (!btsq->bts->synth_opts.callchain &&
 516             !btsq->bts->synth_opts.thread_stack && thread &&
 517             (!old_buffer || btsq->bts->sampling_mode ||
 518              (btsq->bts->snapshot_mode && !buffer->consecutive)))
 519                 thread_stack__set_trace_nr(thread, buffer->buffer_nr + 1);
 520
 521         err = intel_bts_process_buffer(btsq, buffer, thread);
 522
 523         auxtrace_buffer__drop_data(buffer);
 524
 525         btsq->buffer = auxtrace_buffer__next(queue, buffer);
 526         if (btsq->buffer) {
 527                 if (timestamp)
 528                         *timestamp = btsq->buffer->reference;
 529         } else {
 530                 if (!btsq->bts->sampling_mode)
 531                         btsq->done = 1;
 532         }
 533 out_put:
 534         thread__put(thread);
 535         return err;
 536 }
 537
 538 static int intel_bts_flush_queue(struct intel_bts_queue *btsq)
 539 {
 540         u64 ts = 0;
 541         int ret;
 542
 543         while (1) {
 544                 ret = intel_bts_process_queue(btsq, &ts);
 545                 if (ret < 0)
 546                         return ret;
 547                 if (ret)
 548                         break;
 549         }
 550         return 0;
 551 }
 552
 553 static int intel_bts_process_tid_exit(struct intel_bts *bts, pid_t tid)
 554 {
 555         struct auxtrace_queues *queues = &bts->queues;
 556         unsigned int i;
 557
 558         for (i = 0; i < queues->nr_queues; i++) {
 559                 struct auxtrace_queue *queue = &bts->queues.queue_array[i];
 560                 struct intel_bts_queue *btsq = queue->priv;
 561
 562                 if (btsq && btsq->tid == tid)
 563                         return intel_bts_flush_queue(btsq);
 564         }
 565         return 0;
 566 }
 567
 568 static int intel_bts_process_queues(struct intel_bts *bts, u64 timestamp)
 569 {
 570         while (1) {
 571                 unsigned int queue_nr;
 572                 struct auxtrace_queue *queue;
 573                 struct intel_bts_queue *btsq;
 574                 u64 ts = 0;
 575                 int ret;
 576
 577                 if (!bts->heap.heap_cnt)
 578                         return 0;
 579
 580                 if (bts->heap.heap_array[0].ordinal > timestamp)
 581                         return 0;
 582
 583                 queue_nr = bts->heap.heap_array[0].queue_nr;
 584                 queue = &bts->queues.queue_array[queue_nr];
 585                 btsq = queue->priv;
 586
 587                 auxtrace_heap__pop(&bts->heap);
 588
 589                 ret = intel_bts_process_queue(btsq, &ts);
 590                 if (ret < 0) {
 591                         auxtrace_heap__add(&bts->heap, queue_nr, ts);
 592                         return ret;
 593                 }
 594
 595                 if (!ret) {
 596                         ret = auxtrace_heap__add(&bts->heap, queue_nr, ts);
 597                         if (ret < 0)
 598                                 return ret;
 599                 } else {
 600                         btsq->on_heap = false;
 601                 }
 602         }
 603
 604         return 0;
 605 }
 606
 607 static int intel_bts_process_event(struct perf_session *session,
 608                                    union perf_event *event,
 609                                    struct perf_sample *sample,
 610                                    struct perf_tool *tool)
 611 {
 612         struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
 613                                              auxtrace);
 614         u64 timestamp;
 615         int err;
 616
 617         if (dump_trace)
 618                 return 0;
 619
 620         if (!tool->ordered_events) {
 621                 pr_err("Intel BTS requires ordered events\n");
 622                 return -EINVAL;
 623         }
 624
 625         if (sample->time && sample->time != (u64)-1)
 626                 timestamp = perf_time_to_tsc(sample->time, &bts->tc);
 627         else
 628                 timestamp = 0;
 629
 630         err = intel_bts_update_queues(bts);
 631         if (err)
 632                 return err;
 633
 634         err = intel_bts_process_queues(bts, timestamp);
 635         if (err)
 636                 return err;
 637         if (event->header.type == PERF_RECORD_EXIT) {
 638                 err = intel_bts_process_tid_exit(bts, event->fork.tid);
 639                 if (err)
 640                         return err;
 641         }
 642
 643         if (event->header.type == PERF_RECORD_AUX &&
 644             (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
 645             bts->synth_opts.errors)
 646                 err = intel_bts_lost(bts, sample);
 647
 648         return err;
 649 }
 650
 651 static int intel_bts_process_auxtrace_event(struct perf_session *session,
 652                                             union perf_event *event,
 653                                             struct perf_tool *tool __maybe_unused)
 654 {
 655         struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
 656                                              auxtrace);
 657
 658         if (bts->sampling_mode)
 659                 return 0;
 660
 661         if (!bts->data_queued) {
 662                 struct auxtrace_buffer *buffer;
 663                 off_t data_offset;
 664                 int fd = perf_data_file__fd(session->file);
 665                 int err;
 666
 667                 if (perf_data_file__is_pipe(session->file)) {
 668                         data_offset = 0;
 669                 } else {
 670                         data_offset = lseek(fd, 0, SEEK_CUR);
 671                         if (data_offset == -1)
 672                                 return -errno;
 673                 }
 674
 675                 err = auxtrace_queues__add_event(&bts->queues, session, event,
 676                                                  data_offset, &buffer);
 677                 if (err)
 678                         return err;
 679
 680                 /* Dump here now we have copied a piped trace out of the pipe */
 681                 if (dump_trace) {
 682                         if (auxtrace_buffer__get_data(buffer, fd)) {
 683                                 intel_bts_dump_event(bts, buffer->data,
 684                                                      buffer->size);
 685                                 auxtrace_buffer__put_data(buffer);
 686                         }
 687                 }
 688         }
 689
 690         return 0;
 691 }
 692
 693 static int intel_bts_flush(struct perf_session *session,
 694                            struct perf_tool *tool __maybe_unused)
 695 {
 696         struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
 697                                              auxtrace);
 698         int ret;
 699
 700         if (dump_trace || bts->sampling_mode)
 701                 return 0;
 702
 703         if (!tool->ordered_events)
 704                 return -EINVAL;
 705
 706         ret = intel_bts_update_queues(bts);
 707         if (ret < 0)
 708                 return ret;
 709
 710         return intel_bts_process_queues(bts, MAX_TIMESTAMP);
 711 }
 712
 713 static void intel_bts_free_queue(void *priv)
 714 {
 715         struct intel_bts_queue *btsq = priv;
 716
 717         if (!btsq)
 718                 return;
 719         free(btsq);
 720 }
 721
 722 static void intel_bts_free_events(struct perf_session *session)
 723 {
 724         struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
 725                                              auxtrace);
 726         struct auxtrace_queues *queues = &bts->queues;
 727         unsigned int i;
 728
 729         for (i = 0; i < queues->nr_queues; i++) {
 730                 intel_bts_free_queue(queues->queue_array[i].priv);
 731                 queues->queue_array[i].priv = NULL;
 732         }
 733         auxtrace_queues__free(queues);
 734 }
 735
 736 static void intel_bts_free(struct perf_session *session)
 737 {
 738         struct intel_bts *bts = container_of(session->auxtrace, struct intel_bts,
 739                                              auxtrace);
 740
 741         auxtrace_heap__free(&bts->heap);
 742         intel_bts_free_events(session);
 743         session->auxtrace = NULL;
 744         free(bts);
 745 }
 746
 747 struct intel_bts_synth {
 748         struct perf_tool dummy_tool;
 749         struct perf_session *session;
 750 };
 751
 752 static int intel_bts_event_synth(struct perf_tool *tool,
 753                                  union perf_event *event,
 754                                  struct perf_sample *sample __maybe_unused,
 755                                  struct machine *machine __maybe_unused)
 756 {
 757         struct intel_bts_synth *intel_bts_synth =
 758                         container_of(tool, struct intel_bts_synth, dummy_tool);
 759
 760         return perf_session__deliver_synth_event(intel_bts_synth->session,
 761                                                  event, NULL);
 762 }
 763
 764 static int intel_bts_synth_event(struct perf_session *session,
 765                                  struct perf_event_attr *attr, u64 id)
 766 {
 767         struct intel_bts_synth intel_bts_synth;
 768
 769         memset(&intel_bts_synth, 0, sizeof(struct intel_bts_synth));
 770         intel_bts_synth.session = session;
 771
 772         return perf_event__synthesize_attr(&intel_bts_synth.dummy_tool, attr, 1,
 773                                            &id, intel_bts_event_synth);
 774 }
 775
 776 static int intel_bts_synth_events(struct intel_bts *bts,
 777                                   struct perf_session *session)
 778 {
 779         struct perf_evlist *evlist = session->evlist;
 780         struct perf_evsel *evsel;
 781         struct perf_event_attr attr;
 782         bool found = false;
 783         u64 id;
 784         int err;
 785
 786         evlist__for_each_entry(evlist, evsel) {
 787                 if (evsel->attr.type == bts->pmu_type && evsel->ids) {
 788                         found = true;
 789                         break;
 790                 }
 791         }
 792
 793         if (!found) {
 794                 pr_debug("There are no selected events with Intel BTS data\n");
 795                 return 0;
 796         }
 797
 798         memset(&attr, 0, sizeof(struct perf_event_attr));
 799         attr.size = sizeof(struct perf_event_attr);
 800         attr.type = PERF_TYPE_HARDWARE;
 801         attr.sample_type = evsel->attr.sample_type & PERF_SAMPLE_MASK;
 802         attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
 803                             PERF_SAMPLE_PERIOD;
 804         attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
 805         attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
 806         attr.exclude_user = evsel->attr.exclude_user;
 807         attr.exclude_kernel = evsel->attr.exclude_kernel;
 808         attr.exclude_hv = evsel->attr.exclude_hv;
 809         attr.exclude_host = evsel->attr.exclude_host;
 810         attr.exclude_guest = evsel->attr.exclude_guest;
 811         attr.sample_id_all = evsel->attr.sample_id_all;
 812         attr.read_format = evsel->attr.read_format;
 813
 814         id = evsel->id[0] + 1000000000;
 815         if (!id)
 816                 id = 1;
 817
 818         if (bts->synth_opts.branches) {
 819                 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
 820                 attr.sample_period = 1;
 821                 attr.sample_type |= PERF_SAMPLE_ADDR;
 822                 pr_debug("Synthesizing 'branches' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
 823                          id, (u64)attr.sample_type);
 824                 err = intel_bts_synth_event(session, &attr, id);
 825                 if (err) {
 826                         pr_err("%s: failed to synthesize 'branches' event type\n",
 827                                __func__);
 828                         return err;
 829                 }
 830                 bts->sample_branches = true;
 831                 bts->branches_sample_type = attr.sample_type;
 832                 bts->branches_id = id;
 833                 /*
 834                  * We only use sample types from PERF_SAMPLE_MASK so we can use
 835                  * __perf_evsel__sample_size() here.
 836                  */
 837                 bts->branches_event_size = sizeof(struct sample_event) +
 838                                 __perf_evsel__sample_size(attr.sample_type);
 839         }
 840
 841         bts->synth_needs_swap = evsel->needs_swap;
 842
 843         return 0;
 844 }
 845
 846 static const char * const intel_bts_info_fmts[] = {
 847         [INTEL_BTS_PMU_TYPE]            = "  PMU Type           %"PRId64"\n",
 848         [INTEL_BTS_TIME_SHIFT]          = "  Time Shift         %"PRIu64"\n",
 849         [INTEL_BTS_TIME_MULT]           = "  Time Muliplier     %"PRIu64"\n",
 850         [INTEL_BTS_TIME_ZERO]           = "  Time Zero          %"PRIu64"\n",
 851         [INTEL_BTS_CAP_USER_TIME_ZERO]  = "  Cap Time Zero      %"PRId64"\n",
 852         [INTEL_BTS_SNAPSHOT_MODE]       = "  Snapshot mode      %"PRId64"\n",
 853 };
 854
 855 static void intel_bts_print_info(u64 *arr, int start, int finish)
 856 {
 857         int i;
 858
 859         if (!dump_trace)
 860                 return;
 861
 862         for (i = start; i <= finish; i++)
 863                 fprintf(stdout, intel_bts_info_fmts[i], arr[i]);
 864 }
 865
 866 u64 intel_bts_auxtrace_info_priv[INTEL_BTS_AUXTRACE_PRIV_SIZE];
 867
 868 int intel_bts_process_auxtrace_info(union perf_event *event,
 869                                     struct perf_session *session)
 870 {
 871         struct auxtrace_info_event *auxtrace_info = &event->auxtrace_info;
 872         size_t min_sz = sizeof(u64) * INTEL_BTS_SNAPSHOT_MODE;
 873         struct intel_bts *bts;
 874         int err;
 875
 876         if (auxtrace_info->header.size < sizeof(struct auxtrace_info_event) +
 877                                         min_sz)
 878                 return -EINVAL;
 879
 880         bts = zalloc(sizeof(struct intel_bts));
 881         if (!bts)
 882                 return -ENOMEM;
 883
 884         err = auxtrace_queues__init(&bts->queues);
 885         if (err)
 886                 goto err_free;
 887
 888         bts->session = session;
 889         bts->machine = &session->machines.host; /* No kvm support */
 890         bts->auxtrace_type = auxtrace_info->type;
 891         bts->pmu_type = auxtrace_info->priv[INTEL_BTS_PMU_TYPE];
 892         bts->tc.time_shift = auxtrace_info->priv[INTEL_BTS_TIME_SHIFT];
 893         bts->tc.time_mult = auxtrace_info->priv[INTEL_BTS_TIME_MULT];
 894         bts->tc.time_zero = auxtrace_info->priv[INTEL_BTS_TIME_ZERO];
 895         bts->cap_user_time_zero =
 896                         auxtrace_info->priv[INTEL_BTS_CAP_USER_TIME_ZERO];
 897         bts->snapshot_mode = auxtrace_info->priv[INTEL_BTS_SNAPSHOT_MODE];
 898
 899         bts->sampling_mode = false;
 900
 901         bts->auxtrace.process_event = intel_bts_process_event;
 902         bts->auxtrace.process_auxtrace_event = intel_bts_process_auxtrace_event;
 903         bts->auxtrace.flush_events = intel_bts_flush;
 904         bts->auxtrace.free_events = intel_bts_free_events;
 905         bts->auxtrace.free = intel_bts_free;
 906         session->auxtrace = &bts->auxtrace;
 907
 908         intel_bts_print_info(&auxtrace_info->priv[0], INTEL_BTS_PMU_TYPE,
 909                              INTEL_BTS_SNAPSHOT_MODE);
 910
 911         if (dump_trace)
 912                 return 0;
 913
 914         if (session->itrace_synth_opts && session->itrace_synth_opts->set) {
 915                 bts->synth_opts = *session->itrace_synth_opts;
 916         } else {
 917                 itrace_synth_opts__set_default(&bts->synth_opts);
 918                 if (session->itrace_synth_opts)
 919                         bts->synth_opts.thread_stack =
 920                                 session->itrace_synth_opts->thread_stack;
 921         }
 922
 923         if (bts->synth_opts.calls)
 924                 bts->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
 925                                         PERF_IP_FLAG_TRACE_END;
 926         if (bts->synth_opts.returns)
 927                 bts->branches_filter |= PERF_IP_FLAG_RETURN |
 928                                         PERF_IP_FLAG_TRACE_BEGIN;
 929
 930         err = intel_bts_synth_events(bts, session);
 931         if (err)
 932                 goto err_free_queues;
 933
 934         err = auxtrace_queues__process_index(&bts->queues, session);
 935         if (err)
 936                 goto err_free_queues;
 937
 938         if (bts->queues.populated)
 939                 bts->data_queued = true;
 940
 941         return 0;
 942
 943 err_free_queues:
 944         auxtrace_queues__free(&bts->queues);
 945         session->auxtrace = NULL;
 946 err_free:
 947         free(bts);
 948         return err;
 949 }