tools/perf/util/intel-pt.c

   1 /*
   2  * intel_pt.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 <inttypes.h>
  17 #include <stdio.h>
  18 #include <stdbool.h>
  19 #include <errno.h>
  20 #include <linux/kernel.h>
  21 #include <linux/types.h>
  22
  23 #include "../perf.h"
  24 #include "session.h"
  25 #include "machine.h"
  26 #include "memswap.h"
  27 #include "sort.h"
  28 #include "tool.h"
  29 #include "event.h"
  30 #include "evlist.h"
  31 #include "evsel.h"
  32 #include "map.h"
  33 #include "color.h"
  34 #include "util.h"
  35 #include "thread.h"
  36 #include "thread-stack.h"
  37 #include "symbol.h"
  38 #include "callchain.h"
  39 #include "dso.h"
  40 #include "debug.h"
  41 #include "auxtrace.h"
  42 #include "tsc.h"
  43 #include "intel-pt.h"
  44 #include "config.h"
  45
  46 #include "intel-pt-decoder/intel-pt-log.h"
  47 #include "intel-pt-decoder/intel-pt-decoder.h"
  48 #include "intel-pt-decoder/intel-pt-insn-decoder.h"
  49 #include "intel-pt-decoder/intel-pt-pkt-decoder.h"
  50
  51 #define MAX_TIMESTAMP (~0ULL)
  52
  53 struct intel_pt {
  54         struct auxtrace auxtrace;
  55         struct auxtrace_queues queues;
  56         struct auxtrace_heap heap;
  57         u32 auxtrace_type;
  58         struct perf_session *session;
  59         struct machine *machine;
  60         struct perf_evsel *switch_evsel;
  61         struct thread *unknown_thread;
  62         bool timeless_decoding;
  63         bool sampling_mode;
  64         bool snapshot_mode;
  65         bool per_cpu_mmaps;
  66         bool have_tsc;
  67         bool data_queued;
  68         bool est_tsc;
  69         bool sync_switch;
  70         bool mispred_all;
  71         int have_sched_switch;
  72         u32 pmu_type;
  73         u64 kernel_start;
  74         u64 switch_ip;
  75         u64 ptss_ip;
  76
  77         struct perf_tsc_conversion tc;
  78         bool cap_user_time_zero;
  79
  80         struct itrace_synth_opts synth_opts;
  81
  82         bool sample_instructions;
  83         u64 instructions_sample_type;
  84         u64 instructions_sample_period;
  85         u64 instructions_id;
  86
  87         bool sample_branches;
  88         u32 branches_filter;
  89         u64 branches_sample_type;
  90         u64 branches_id;
  91
  92         bool sample_transactions;
  93         u64 transactions_sample_type;
  94         u64 transactions_id;
  95
  96         bool synth_needs_swap;
  97
  98         u64 tsc_bit;
  99         u64 mtc_bit;
 100         u64 mtc_freq_bits;
 101         u32 tsc_ctc_ratio_n;
 102         u32 tsc_ctc_ratio_d;
 103         u64 cyc_bit;
 104         u64 noretcomp_bit;
 105         unsigned max_non_turbo_ratio;
 106
 107         unsigned long num_events;
 108
 109         char *filter;
 110         struct addr_filters filts;
 111 };
 112
 113 enum switch_state {
 114         INTEL_PT_SS_NOT_TRACING,
 115         INTEL_PT_SS_UNKNOWN,
 116         INTEL_PT_SS_TRACING,
 117         INTEL_PT_SS_EXPECTING_SWITCH_EVENT,
 118         INTEL_PT_SS_EXPECTING_SWITCH_IP,
 119 };
 120
 121 struct intel_pt_queue {
 122         struct intel_pt *pt;
 123         unsigned int queue_nr;
 124         struct auxtrace_buffer *buffer;
 125         void *decoder;
 126         const struct intel_pt_state *state;
 127         struct ip_callchain *chain;
 128         struct branch_stack *last_branch;
 129         struct branch_stack *last_branch_rb;
 130         size_t last_branch_pos;
 131         union perf_event *event_buf;
 132         bool on_heap;
 133         bool stop;
 134         bool step_through_buffers;
 135         bool use_buffer_pid_tid;
 136         pid_t pid, tid;
 137         int cpu;
 138         int switch_state;
 139         pid_t next_tid;
 140         struct thread *thread;
 141         bool exclude_kernel;
 142         bool have_sample;
 143         u64 time;
 144         u64 timestamp;
 145         u32 flags;
 146         u16 insn_len;
 147         u64 last_insn_cnt;
 148         char insn[INTEL_PT_INSN_BUF_SZ];
 149 };
 150
 151 static void intel_pt_dump(struct intel_pt *pt __maybe_unused,
 152                           unsigned char *buf, size_t len)
 153 {
 154         struct intel_pt_pkt packet;
 155         size_t pos = 0;
 156         int ret, pkt_len, i;
 157         char desc[INTEL_PT_PKT_DESC_MAX];
 158         const char *color = PERF_COLOR_BLUE;
 159
 160         color_fprintf(stdout, color,
 161                       ". ... Intel Processor Trace data: size %zu bytes\n",
 162                       len);
 163
 164         while (len) {
 165                 ret = intel_pt_get_packet(buf, len, &packet);
 166                 if (ret > 0)
 167                         pkt_len = ret;
 168                 else
 169                         pkt_len = 1;
 170                 printf(".");
 171                 color_fprintf(stdout, color, "  %08x: ", pos);
 172                 for (i = 0; i < pkt_len; i++)
 173                         color_fprintf(stdout, color, " %02x", buf[i]);
 174                 for (; i < 16; i++)
 175                         color_fprintf(stdout, color, "   ");
 176                 if (ret > 0) {
 177                         ret = intel_pt_pkt_desc(&packet, desc,
 178                                                 INTEL_PT_PKT_DESC_MAX);
 179                         if (ret > 0)
 180                                 color_fprintf(stdout, color, " %s\n", desc);
 181                 } else {
 182                         color_fprintf(stdout, color, " Bad packet!\n");
 183                 }
 184                 pos += pkt_len;
 185                 buf += pkt_len;
 186                 len -= pkt_len;
 187         }
 188 }
 189
 190 static void intel_pt_dump_event(struct intel_pt *pt, unsigned char *buf,
 191                                 size_t len)
 192 {
 193         printf(".\n");
 194         intel_pt_dump(pt, buf, len);
 195 }
 196
 197 static int intel_pt_do_fix_overlap(struct intel_pt *pt, struct auxtrace_buffer *a,
 198                                    struct auxtrace_buffer *b)
 199 {
 200         void *start;
 201
 202         start = intel_pt_find_overlap(a->data, a->size, b->data, b->size,
 203                                       pt->have_tsc);
 204         if (!start)
 205                 return -EINVAL;
 206         b->use_size = b->data + b->size - start;
 207         b->use_data = start;
 208         return 0;
 209 }
 210
 211 static void intel_pt_use_buffer_pid_tid(struct intel_pt_queue *ptq,
 212                                         struct auxtrace_queue *queue,
 213                                         struct auxtrace_buffer *buffer)
 214 {
 215         if (queue->cpu == -1 && buffer->cpu != -1)
 216                 ptq->cpu = buffer->cpu;
 217
 218         ptq->pid = buffer->pid;
 219         ptq->tid = buffer->tid;
 220
 221         intel_pt_log("queue %u cpu %d pid %d tid %d\n",
 222                      ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
 223
 224         thread__zput(ptq->thread);
 225
 226         if (ptq->tid != -1) {
 227                 if (ptq->pid != -1)
 228                         ptq->thread = machine__findnew_thread(ptq->pt->machine,
 229                                                               ptq->pid,
 230                                                               ptq->tid);
 231                 else
 232                         ptq->thread = machine__find_thread(ptq->pt->machine, -1,
 233                                                            ptq->tid);
 234         }
 235 }
 236
 237 /* This function assumes data is processed sequentially only */
 238 static int intel_pt_get_trace(struct intel_pt_buffer *b, void *data)
 239 {
 240         struct intel_pt_queue *ptq = data;
 241         struct auxtrace_buffer *buffer = ptq->buffer, *old_buffer = buffer;
 242         struct auxtrace_queue *queue;
 243
 244         if (ptq->stop) {
 245                 b->len = 0;
 246                 return 0;
 247         }
 248
 249         queue = &ptq->pt->queues.queue_array[ptq->queue_nr];
 250 next:
 251         buffer = auxtrace_buffer__next(queue, buffer);
 252         if (!buffer) {
 253                 if (old_buffer)
 254                         auxtrace_buffer__drop_data(old_buffer);
 255                 b->len = 0;
 256                 return 0;
 257         }
 258
 259         ptq->buffer = buffer;
 260
 261         if (!buffer->data) {
 262                 int fd = perf_data_file__fd(ptq->pt->session->file);
 263
 264                 buffer->data = auxtrace_buffer__get_data(buffer, fd);
 265                 if (!buffer->data)
 266                         return -ENOMEM;
 267         }
 268
 269         if (ptq->pt->snapshot_mode && !buffer->consecutive && old_buffer &&
 270             intel_pt_do_fix_overlap(ptq->pt, old_buffer, buffer))
 271                 return -ENOMEM;
 272
 273         if (buffer->use_data) {
 274                 b->len = buffer->use_size;
 275                 b->buf = buffer->use_data;
 276         } else {
 277                 b->len = buffer->size;
 278                 b->buf = buffer->data;
 279         }
 280         b->ref_timestamp = buffer->reference;
 281
 282         /*
 283          * If in snapshot mode and the buffer has no usable data, get next
 284          * buffer and again check overlap against old_buffer.
 285          */
 286         if (ptq->pt->snapshot_mode && !b->len)
 287                 goto next;
 288
 289         if (old_buffer)
 290                 auxtrace_buffer__drop_data(old_buffer);
 291
 292         if (!old_buffer || ptq->pt->sampling_mode || (ptq->pt->snapshot_mode &&
 293                                                       !buffer->consecutive)) {
 294                 b->consecutive = false;
 295                 b->trace_nr = buffer->buffer_nr + 1;
 296         } else {
 297                 b->consecutive = true;
 298         }
 299
 300         if (ptq->use_buffer_pid_tid && (ptq->pid != buffer->pid ||
 301                                         ptq->tid != buffer->tid))
 302                 intel_pt_use_buffer_pid_tid(ptq, queue, buffer);
 303
 304         if (ptq->step_through_buffers)
 305                 ptq->stop = true;
 306
 307         if (!b->len)
 308                 return intel_pt_get_trace(b, data);
 309
 310         return 0;
 311 }
 312
 313 struct intel_pt_cache_entry {
 314         struct auxtrace_cache_entry     entry;
 315         u64                             insn_cnt;
 316         u64                             byte_cnt;
 317         enum intel_pt_insn_op           op;
 318         enum intel_pt_insn_branch       branch;
 319         int                             length;
 320         int32_t                         rel;
 321         char                            insn[INTEL_PT_INSN_BUF_SZ];
 322 };
 323
 324 static int intel_pt_config_div(const char *var, const char *value, void *data)
 325 {
 326         int *d = data;
 327         long val;
 328
 329         if (!strcmp(var, "intel-pt.cache-divisor")) {
 330                 val = strtol(value, NULL, 0);
 331                 if (val > 0 && val <= INT_MAX)
 332                         *d = val;
 333         }
 334
 335         return 0;
 336 }
 337
 338 static int intel_pt_cache_divisor(void)
 339 {
 340         static int d;
 341
 342         if (d)
 343                 return d;
 344
 345         perf_config(intel_pt_config_div, &d);
 346
 347         if (!d)
 348                 d = 64;
 349
 350         return d;
 351 }
 352
 353 static unsigned int intel_pt_cache_size(struct dso *dso,
 354                                         struct machine *machine)
 355 {
 356         off_t size;
 357
 358         size = dso__data_size(dso, machine);
 359         size /= intel_pt_cache_divisor();
 360         if (size < 1000)
 361                 return 10;
 362         if (size > (1 << 21))
 363                 return 21;
 364         return 32 - __builtin_clz(size);
 365 }
 366
 367 static struct auxtrace_cache *intel_pt_cache(struct dso *dso,
 368                                              struct machine *machine)
 369 {
 370         struct auxtrace_cache *c;
 371         unsigned int bits;
 372
 373         if (dso->auxtrace_cache)
 374                 return dso->auxtrace_cache;
 375
 376         bits = intel_pt_cache_size(dso, machine);
 377
 378         /* Ignoring cache creation failure */
 379         c = auxtrace_cache__new(bits, sizeof(struct intel_pt_cache_entry), 200);
 380
 381         dso->auxtrace_cache = c;
 382
 383         return c;
 384 }
 385
 386 static int intel_pt_cache_add(struct dso *dso, struct machine *machine,
 387                               u64 offset, u64 insn_cnt, u64 byte_cnt,
 388                               struct intel_pt_insn *intel_pt_insn)
 389 {
 390         struct auxtrace_cache *c = intel_pt_cache(dso, machine);
 391         struct intel_pt_cache_entry *e;
 392         int err;
 393
 394         if (!c)
 395                 return -ENOMEM;
 396
 397         e = auxtrace_cache__alloc_entry(c);
 398         if (!e)
 399                 return -ENOMEM;
 400
 401         e->insn_cnt = insn_cnt;
 402         e->byte_cnt = byte_cnt;
 403         e->op = intel_pt_insn->op;
 404         e->branch = intel_pt_insn->branch;
 405         e->length = intel_pt_insn->length;
 406         e->rel = intel_pt_insn->rel;
 407         memcpy(e->insn, intel_pt_insn->buf, INTEL_PT_INSN_BUF_SZ);
 408
 409         err = auxtrace_cache__add(c, offset, &e->entry);
 410         if (err)
 411                 auxtrace_cache__free_entry(c, e);
 412
 413         return err;
 414 }
 415
 416 static struct intel_pt_cache_entry *
 417 intel_pt_cache_lookup(struct dso *dso, struct machine *machine, u64 offset)
 418 {
 419         struct auxtrace_cache *c = intel_pt_cache(dso, machine);
 420
 421         if (!c)
 422                 return NULL;
 423
 424         return auxtrace_cache__lookup(dso->auxtrace_cache, offset);
 425 }
 426
 427 static int intel_pt_walk_next_insn(struct intel_pt_insn *intel_pt_insn,
 428                                    uint64_t *insn_cnt_ptr, uint64_t *ip,
 429                                    uint64_t to_ip, uint64_t max_insn_cnt,
 430                                    void *data)
 431 {
 432         struct intel_pt_queue *ptq = data;
 433         struct machine *machine = ptq->pt->machine;
 434         struct thread *thread;
 435         struct addr_location al;
 436         unsigned char buf[INTEL_PT_INSN_BUF_SZ];
 437         ssize_t len;
 438         int x86_64;
 439         u8 cpumode;
 440         u64 offset, start_offset, start_ip;
 441         u64 insn_cnt = 0;
 442         bool one_map = true;
 443
 444         intel_pt_insn->length = 0;
 445
 446         if (to_ip && *ip == to_ip)
 447                 goto out_no_cache;
 448
 449         if (*ip >= ptq->pt->kernel_start)
 450                 cpumode = PERF_RECORD_MISC_KERNEL;
 451         else
 452                 cpumode = PERF_RECORD_MISC_USER;
 453
 454         thread = ptq->thread;
 455         if (!thread) {
 456                 if (cpumode != PERF_RECORD_MISC_KERNEL)
 457                         return -EINVAL;
 458                 thread = ptq->pt->unknown_thread;
 459         }
 460
 461         while (1) {
 462                 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, *ip, &al);
 463                 if (!al.map || !al.map->dso)
 464                         return -EINVAL;
 465
 466                 if (al.map->dso->data.status == DSO_DATA_STATUS_ERROR &&
 467                     dso__data_status_seen(al.map->dso,
 468                                           DSO_DATA_STATUS_SEEN_ITRACE))
 469                         return -ENOENT;
 470
 471                 offset = al.map->map_ip(al.map, *ip);
 472
 473                 if (!to_ip && one_map) {
 474                         struct intel_pt_cache_entry *e;
 475
 476                         e = intel_pt_cache_lookup(al.map->dso, machine, offset);
 477                         if (e &&
 478                             (!max_insn_cnt || e->insn_cnt <= max_insn_cnt)) {
 479                                 *insn_cnt_ptr = e->insn_cnt;
 480                                 *ip += e->byte_cnt;
 481                                 intel_pt_insn->op = e->op;
 482                                 intel_pt_insn->branch = e->branch;
 483                                 intel_pt_insn->length = e->length;
 484                                 intel_pt_insn->rel = e->rel;
 485                                 memcpy(intel_pt_insn->buf, e->insn,
 486                                        INTEL_PT_INSN_BUF_SZ);
 487                                 intel_pt_log_insn_no_data(intel_pt_insn, *ip);
 488                                 return 0;
 489                         }
 490                 }
 491
 492                 start_offset = offset;
 493                 start_ip = *ip;
 494
 495                 /* Load maps to ensure dso->is_64_bit has been updated */
 496                 map__load(al.map);
 497
 498                 x86_64 = al.map->dso->is_64_bit;
 499
 500                 while (1) {
 501                         len = dso__data_read_offset(al.map->dso, machine,
 502                                                     offset, buf,
 503                                                     INTEL_PT_INSN_BUF_SZ);
 504                         if (len <= 0)
 505                                 return -EINVAL;
 506
 507                         if (intel_pt_get_insn(buf, len, x86_64, intel_pt_insn))
 508                                 return -EINVAL;
 509
 510                         intel_pt_log_insn(intel_pt_insn, *ip);
 511
 512                         insn_cnt += 1;
 513
 514                         if (intel_pt_insn->branch != INTEL_PT_BR_NO_BRANCH)
 515                                 goto out;
 516
 517                         if (max_insn_cnt && insn_cnt >= max_insn_cnt)
 518                                 goto out_no_cache;
 519
 520                         *ip += intel_pt_insn->length;
 521
 522                         if (to_ip && *ip == to_ip)
 523                                 goto out_no_cache;
 524
 525                         if (*ip >= al.map->end)
 526                                 break;
 527
 528                         offset += intel_pt_insn->length;
 529                 }
 530                 one_map = false;
 531         }
 532 out:
 533         *insn_cnt_ptr = insn_cnt;
 534
 535         if (!one_map)
 536                 goto out_no_cache;
 537
 538         /*
 539          * Didn't lookup in the 'to_ip' case, so do it now to prevent duplicate
 540          * entries.
 541          */
 542         if (to_ip) {
 543                 struct intel_pt_cache_entry *e;
 544
 545                 e = intel_pt_cache_lookup(al.map->dso, machine, start_offset);
 546                 if (e)
 547                         return 0;
 548         }
 549
 550         /* Ignore cache errors */
 551         intel_pt_cache_add(al.map->dso, machine, start_offset, insn_cnt,
 552                            *ip - start_ip, intel_pt_insn);
 553
 554         return 0;
 555
 556 out_no_cache:
 557         *insn_cnt_ptr = insn_cnt;
 558         return 0;
 559 }
 560
 561 static bool intel_pt_match_pgd_ip(struct intel_pt *pt, uint64_t ip,
 562                                   uint64_t offset, const char *filename)
 563 {
 564         struct addr_filter *filt;
 565         bool have_filter   = false;
 566         bool hit_tracestop = false;
 567         bool hit_filter    = false;
 568
 569         list_for_each_entry(filt, &pt->filts.head, list) {
 570                 if (filt->start)
 571                         have_filter = true;
 572
 573                 if ((filename && !filt->filename) ||
 574                     (!filename && filt->filename) ||
 575                     (filename && strcmp(filename, filt->filename)))
 576                         continue;
 577
 578                 if (!(offset >= filt->addr && offset < filt->addr + filt->size))
 579                         continue;
 580
 581                 intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s hit filter: %s offset %#"PRIx64" size %#"PRIx64"\n",
 582                              ip, offset, filename ? filename : "[kernel]",
 583                              filt->start ? "filter" : "stop",
 584                              filt->addr, filt->size);
 585
 586                 if (filt->start)
 587                         hit_filter = true;
 588                 else
 589                         hit_tracestop = true;
 590         }
 591
 592         if (!hit_tracestop && !hit_filter)
 593                 intel_pt_log("TIP.PGD ip %#"PRIx64" offset %#"PRIx64" in %s is not in a filter region\n",
 594                              ip, offset, filename ? filename : "[kernel]");
 595
 596         return hit_tracestop || (have_filter && !hit_filter);
 597 }
 598
 599 static int __intel_pt_pgd_ip(uint64_t ip, void *data)
 600 {
 601         struct intel_pt_queue *ptq = data;
 602         struct thread *thread;
 603         struct addr_location al;
 604         u8 cpumode;
 605         u64 offset;
 606
 607         if (ip >= ptq->pt->kernel_start)
 608                 return intel_pt_match_pgd_ip(ptq->pt, ip, ip, NULL);
 609
 610         cpumode = PERF_RECORD_MISC_USER;
 611
 612         thread = ptq->thread;
 613         if (!thread)
 614                 return -EINVAL;
 615
 616         thread__find_addr_map(thread, cpumode, MAP__FUNCTION, ip, &al);
 617         if (!al.map || !al.map->dso)
 618                 return -EINVAL;
 619
 620         offset = al.map->map_ip(al.map, ip);
 621
 622         return intel_pt_match_pgd_ip(ptq->pt, ip, offset,
 623                                      al.map->dso->long_name);
 624 }
 625
 626 static bool intel_pt_pgd_ip(uint64_t ip, void *data)
 627 {
 628         return __intel_pt_pgd_ip(ip, data) > 0;
 629 }
 630
 631 static bool intel_pt_get_config(struct intel_pt *pt,
 632                                 struct perf_event_attr *attr, u64 *config)
 633 {
 634         if (attr->type == pt->pmu_type) {
 635                 if (config)
 636                         *config = attr->config;
 637                 return true;
 638         }
 639
 640         return false;
 641 }
 642
 643 static bool intel_pt_exclude_kernel(struct intel_pt *pt)
 644 {
 645         struct perf_evsel *evsel;
 646
 647         evlist__for_each_entry(pt->session->evlist, evsel) {
 648                 if (intel_pt_get_config(pt, &evsel->attr, NULL) &&
 649                     !evsel->attr.exclude_kernel)
 650                         return false;
 651         }
 652         return true;
 653 }
 654
 655 static bool intel_pt_return_compression(struct intel_pt *pt)
 656 {
 657         struct perf_evsel *evsel;
 658         u64 config;
 659
 660         if (!pt->noretcomp_bit)
 661                 return true;
 662
 663         evlist__for_each_entry(pt->session->evlist, evsel) {
 664                 if (intel_pt_get_config(pt, &evsel->attr, &config) &&
 665                     (config & pt->noretcomp_bit))
 666                         return false;
 667         }
 668         return true;
 669 }
 670
 671 static unsigned int intel_pt_mtc_period(struct intel_pt *pt)
 672 {
 673         struct perf_evsel *evsel;
 674         unsigned int shift;
 675         u64 config;
 676
 677         if (!pt->mtc_freq_bits)
 678                 return 0;
 679
 680         for (shift = 0, config = pt->mtc_freq_bits; !(config & 1); shift++)
 681                 config >>= 1;
 682
 683         evlist__for_each_entry(pt->session->evlist, evsel) {
 684                 if (intel_pt_get_config(pt, &evsel->attr, &config))
 685                         return (config & pt->mtc_freq_bits) >> shift;
 686         }
 687         return 0;
 688 }
 689
 690 static bool intel_pt_timeless_decoding(struct intel_pt *pt)
 691 {
 692         struct perf_evsel *evsel;
 693         bool timeless_decoding = true;
 694         u64 config;
 695
 696         if (!pt->tsc_bit || !pt->cap_user_time_zero)
 697                 return true;
 698
 699         evlist__for_each_entry(pt->session->evlist, evsel) {
 700                 if (!(evsel->attr.sample_type & PERF_SAMPLE_TIME))
 701                         return true;
 702                 if (intel_pt_get_config(pt, &evsel->attr, &config)) {
 703                         if (config & pt->tsc_bit)
 704                                 timeless_decoding = false;
 705                         else
 706                                 return true;
 707                 }
 708         }
 709         return timeless_decoding;
 710 }
 711
 712 static bool intel_pt_tracing_kernel(struct intel_pt *pt)
 713 {
 714         struct perf_evsel *evsel;
 715
 716         evlist__for_each_entry(pt->session->evlist, evsel) {
 717                 if (intel_pt_get_config(pt, &evsel->attr, NULL) &&
 718                     !evsel->attr.exclude_kernel)
 719                         return true;
 720         }
 721         return false;
 722 }
 723
 724 static bool intel_pt_have_tsc(struct intel_pt *pt)
 725 {
 726         struct perf_evsel *evsel;
 727         bool have_tsc = false;
 728         u64 config;
 729
 730         if (!pt->tsc_bit)
 731                 return false;
 732
 733         evlist__for_each_entry(pt->session->evlist, evsel) {
 734                 if (intel_pt_get_config(pt, &evsel->attr, &config)) {
 735                         if (config & pt->tsc_bit)
 736                                 have_tsc = true;
 737                         else
 738                                 return false;
 739                 }
 740         }
 741         return have_tsc;
 742 }
 743
 744 static u64 intel_pt_ns_to_ticks(const struct intel_pt *pt, u64 ns)
 745 {
 746         u64 quot, rem;
 747
 748         quot = ns / pt->tc.time_mult;
 749         rem  = ns % pt->tc.time_mult;
 750         return (quot << pt->tc.time_shift) + (rem << pt->tc.time_shift) /
 751                 pt->tc.time_mult;
 752 }
 753
 754 static struct intel_pt_queue *intel_pt_alloc_queue(struct intel_pt *pt,
 755                                                    unsigned int queue_nr)
 756 {
 757         struct intel_pt_params params = { .get_trace = 0, };
 758         struct intel_pt_queue *ptq;
 759
 760         ptq = zalloc(sizeof(struct intel_pt_queue));
 761         if (!ptq)
 762                 return NULL;
 763
 764         if (pt->synth_opts.callchain) {
 765                 size_t sz = sizeof(struct ip_callchain);
 766
 767                 sz += pt->synth_opts.callchain_sz * sizeof(u64);
 768                 ptq->chain = zalloc(sz);
 769                 if (!ptq->chain)
 770                         goto out_free;
 771         }
 772
 773         if (pt->synth_opts.last_branch) {
 774                 size_t sz = sizeof(struct branch_stack);
 775
 776                 sz += pt->synth_opts.last_branch_sz *
 777                       sizeof(struct branch_entry);
 778                 ptq->last_branch = zalloc(sz);
 779                 if (!ptq->last_branch)
 780                         goto out_free;
 781                 ptq->last_branch_rb = zalloc(sz);
 782                 if (!ptq->last_branch_rb)
 783                         goto out_free;
 784         }
 785
 786         ptq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
 787         if (!ptq->event_buf)
 788                 goto out_free;
 789
 790         ptq->pt = pt;
 791         ptq->queue_nr = queue_nr;
 792         ptq->exclude_kernel = intel_pt_exclude_kernel(pt);
 793         ptq->pid = -1;
 794         ptq->tid = -1;
 795         ptq->cpu = -1;
 796         ptq->next_tid = -1;
 797
 798         params.get_trace = intel_pt_get_trace;
 799         params.walk_insn = intel_pt_walk_next_insn;
 800         params.data = ptq;
 801         params.return_compression = intel_pt_return_compression(pt);
 802         params.max_non_turbo_ratio = pt->max_non_turbo_ratio;
 803         params.mtc_period = intel_pt_mtc_period(pt);
 804         params.tsc_ctc_ratio_n = pt->tsc_ctc_ratio_n;
 805         params.tsc_ctc_ratio_d = pt->tsc_ctc_ratio_d;
 806
 807         if (pt->filts.cnt > 0)
 808                 params.pgd_ip = intel_pt_pgd_ip;
 809
 810         if (pt->synth_opts.instructions) {
 811                 if (pt->synth_opts.period) {
 812                         switch (pt->synth_opts.period_type) {
 813                         case PERF_ITRACE_PERIOD_INSTRUCTIONS:
 814                                 params.period_type =
 815                                                 INTEL_PT_PERIOD_INSTRUCTIONS;
 816                                 params.period = pt->synth_opts.period;
 817                                 break;
 818                         case PERF_ITRACE_PERIOD_TICKS:
 819                                 params.period_type = INTEL_PT_PERIOD_TICKS;
 820                                 params.period = pt->synth_opts.period;
 821                                 break;
 822                         case PERF_ITRACE_PERIOD_NANOSECS:
 823                                 params.period_type = INTEL_PT_PERIOD_TICKS;
 824                                 params.period = intel_pt_ns_to_ticks(pt,
 825                                                         pt->synth_opts.period);
 826                                 break;
 827                         default:
 828                                 break;
 829                         }
 830                 }
 831
 832                 if (!params.period) {
 833                         params.period_type = INTEL_PT_PERIOD_INSTRUCTIONS;
 834                         params.period = 1;
 835                 }
 836         }
 837
 838         ptq->decoder = intel_pt_decoder_new(&params);
 839         if (!ptq->decoder)
 840                 goto out_free;
 841
 842         return ptq;
 843
 844 out_free:
 845         zfree(&ptq->event_buf);
 846         zfree(&ptq->last_branch);
 847         zfree(&ptq->last_branch_rb);
 848         zfree(&ptq->chain);
 849         free(ptq);
 850         return NULL;
 851 }
 852
 853 static void intel_pt_free_queue(void *priv)
 854 {
 855         struct intel_pt_queue *ptq = priv;
 856
 857         if (!ptq)
 858                 return;
 859         thread__zput(ptq->thread);
 860         intel_pt_decoder_free(ptq->decoder);
 861         zfree(&ptq->event_buf);
 862         zfree(&ptq->last_branch);
 863         zfree(&ptq->last_branch_rb);
 864         zfree(&ptq->chain);
 865         free(ptq);
 866 }
 867
 868 static void intel_pt_set_pid_tid_cpu(struct intel_pt *pt,
 869                                      struct auxtrace_queue *queue)
 870 {
 871         struct intel_pt_queue *ptq = queue->priv;
 872
 873         if (queue->tid == -1 || pt->have_sched_switch) {
 874                 ptq->tid = machine__get_current_tid(pt->machine, ptq->cpu);
 875                 thread__zput(ptq->thread);
 876         }
 877
 878         if (!ptq->thread && ptq->tid != -1)
 879                 ptq->thread = machine__find_thread(pt->machine, -1, ptq->tid);
 880
 881         if (ptq->thread) {
 882                 ptq->pid = ptq->thread->pid_;
 883                 if (queue->cpu == -1)
 884                         ptq->cpu = ptq->thread->cpu;
 885         }
 886 }
 887
 888 static void intel_pt_sample_flags(struct intel_pt_queue *ptq)
 889 {
 890         if (ptq->state->flags & INTEL_PT_ABORT_TX) {
 891                 ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_TX_ABORT;
 892         } else if (ptq->state->flags & INTEL_PT_ASYNC) {
 893                 if (ptq->state->to_ip)
 894                         ptq->flags = PERF_IP_FLAG_BRANCH | PERF_IP_FLAG_CALL |
 895                                      PERF_IP_FLAG_ASYNC |
 896                                      PERF_IP_FLAG_INTERRUPT;
 897                 else
 898                         ptq->flags = PERF_IP_FLAG_BRANCH |
 899                                      PERF_IP_FLAG_TRACE_END;
 900                 ptq->insn_len = 0;
 901         } else {
 902                 if (ptq->state->from_ip)
 903                         ptq->flags = intel_pt_insn_type(ptq->state->insn_op);
 904                 else
 905                         ptq->flags = PERF_IP_FLAG_BRANCH |
 906                                      PERF_IP_FLAG_TRACE_BEGIN;
 907                 if (ptq->state->flags & INTEL_PT_IN_TX)
 908                         ptq->flags |= PERF_IP_FLAG_IN_TX;
 909                 ptq->insn_len = ptq->state->insn_len;
 910                 memcpy(ptq->insn, ptq->state->insn, INTEL_PT_INSN_BUF_SZ);
 911         }
 912 }
 913
 914 static int intel_pt_setup_queue(struct intel_pt *pt,
 915                                 struct auxtrace_queue *queue,
 916                                 unsigned int queue_nr)
 917 {
 918         struct intel_pt_queue *ptq = queue->priv;
 919
 920         if (list_empty(&queue->head))
 921                 return 0;
 922
 923         if (!ptq) {
 924                 ptq = intel_pt_alloc_queue(pt, queue_nr);
 925                 if (!ptq)
 926                         return -ENOMEM;
 927                 queue->priv = ptq;
 928
 929                 if (queue->cpu != -1)
 930                         ptq->cpu = queue->cpu;
 931                 ptq->tid = queue->tid;
 932
 933                 if (pt->sampling_mode) {
 934                         if (pt->timeless_decoding)
 935                                 ptq->step_through_buffers = true;
 936                         if (pt->timeless_decoding || !pt->have_sched_switch)
 937                                 ptq->use_buffer_pid_tid = true;
 938                 }
 939         }
 940
 941         if (!ptq->on_heap &&
 942             (!pt->sync_switch ||
 943              ptq->switch_state != INTEL_PT_SS_EXPECTING_SWITCH_EVENT)) {
 944                 const struct intel_pt_state *state;
 945                 int ret;
 946
 947                 if (pt->timeless_decoding)
 948                         return 0;
 949
 950                 intel_pt_log("queue %u getting timestamp\n", queue_nr);
 951                 intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
 952                              queue_nr, ptq->cpu, ptq->pid, ptq->tid);
 953                 while (1) {
 954                         state = intel_pt_decode(ptq->decoder);
 955                         if (state->err) {
 956                                 if (state->err == INTEL_PT_ERR_NODATA) {
 957                                         intel_pt_log("queue %u has no timestamp\n",
 958                                                      queue_nr);
 959                                         return 0;
 960                                 }
 961                                 continue;
 962                         }
 963                         if (state->timestamp)
 964                                 break;
 965                 }
 966
 967                 ptq->timestamp = state->timestamp;
 968                 intel_pt_log("queue %u timestamp 0x%" PRIx64 "\n",
 969                              queue_nr, ptq->timestamp);
 970                 ptq->state = state;
 971                 ptq->have_sample = true;
 972                 intel_pt_sample_flags(ptq);
 973                 ret = auxtrace_heap__add(&pt->heap, queue_nr, ptq->timestamp);
 974                 if (ret)
 975                         return ret;
 976                 ptq->on_heap = true;
 977         }
 978
 979         return 0;
 980 }
 981
 982 static int intel_pt_setup_queues(struct intel_pt *pt)
 983 {
 984         unsigned int i;
 985         int ret;
 986
 987         for (i = 0; i < pt->queues.nr_queues; i++) {
 988                 ret = intel_pt_setup_queue(pt, &pt->queues.queue_array[i], i);
 989                 if (ret)
 990                         return ret;
 991         }
 992         return 0;
 993 }
 994
 995 static inline void intel_pt_copy_last_branch_rb(struct intel_pt_queue *ptq)
 996 {
 997         struct branch_stack *bs_src = ptq->last_branch_rb;
 998         struct branch_stack *bs_dst = ptq->last_branch;
 999         size_t nr = 0;
1000
1001         bs_dst->nr = bs_src->nr;
1002
1003         if (!bs_src->nr)
1004                 return;
1005
1006         nr = ptq->pt->synth_opts.last_branch_sz - ptq->last_branch_pos;
1007         memcpy(&bs_dst->entries[0],
1008                &bs_src->entries[ptq->last_branch_pos],
1009                sizeof(struct branch_entry) * nr);
1010
1011         if (bs_src->nr >= ptq->pt->synth_opts.last_branch_sz) {
1012                 memcpy(&bs_dst->entries[nr],
1013                        &bs_src->entries[0],
1014                        sizeof(struct branch_entry) * ptq->last_branch_pos);
1015         }
1016 }
1017
1018 static inline void intel_pt_reset_last_branch_rb(struct intel_pt_queue *ptq)
1019 {
1020         ptq->last_branch_pos = 0;
1021         ptq->last_branch_rb->nr = 0;
1022 }
1023
1024 static void intel_pt_update_last_branch_rb(struct intel_pt_queue *ptq)
1025 {
1026         const struct intel_pt_state *state = ptq->state;
1027         struct branch_stack *bs = ptq->last_branch_rb;
1028         struct branch_entry *be;
1029
1030         if (!ptq->last_branch_pos)
1031                 ptq->last_branch_pos = ptq->pt->synth_opts.last_branch_sz;
1032
1033         ptq->last_branch_pos -= 1;
1034
1035         be              = &bs->entries[ptq->last_branch_pos];
1036         be->from        = state->from_ip;
1037         be->to          = state->to_ip;
1038         be->flags.abort = !!(state->flags & INTEL_PT_ABORT_TX);
1039         be->flags.in_tx = !!(state->flags & INTEL_PT_IN_TX);
1040         /* No support for mispredict */
1041         be->flags.mispred = ptq->pt->mispred_all;
1042
1043         if (bs->nr < ptq->pt->synth_opts.last_branch_sz)
1044                 bs->nr += 1;
1045 }
1046
1047 static int intel_pt_inject_event(union perf_event *event,
1048                                  struct perf_sample *sample, u64 type,
1049                                  bool swapped)
1050 {
1051         event->header.size = perf_event__sample_event_size(sample, type, 0);
1052         return perf_event__synthesize_sample(event, type, 0, sample, swapped);
1053 }
1054
1055 static int intel_pt_synth_branch_sample(struct intel_pt_queue *ptq)
1056 {
1057         int ret;
1058         struct intel_pt *pt = ptq->pt;
1059         union perf_event *event = ptq->event_buf;
1060         struct perf_sample sample = { .ip = 0, };
1061         struct dummy_branch_stack {
1062                 u64                     nr;
1063                 struct branch_entry     entries;
1064         } dummy_bs;
1065
1066         if (pt->branches_filter && !(pt->branches_filter & ptq->flags))
1067                 return 0;
1068
1069         if (pt->synth_opts.initial_skip &&
1070             pt->num_events++ < pt->synth_opts.initial_skip)
1071                 return 0;
1072
1073         event->sample.header.type = PERF_RECORD_SAMPLE;
1074         event->sample.header.misc = PERF_RECORD_MISC_USER;
1075         event->sample.header.size = sizeof(struct perf_event_header);
1076
1077         if (!pt->timeless_decoding)
1078                 sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1079
1080         sample.cpumode = PERF_RECORD_MISC_USER;
1081         sample.ip = ptq->state->from_ip;
1082         sample.pid = ptq->pid;
1083         sample.tid = ptq->tid;
1084         sample.addr = ptq->state->to_ip;
1085         sample.id = ptq->pt->branches_id;
1086         sample.stream_id = ptq->pt->branches_id;
1087         sample.period = 1;
1088         sample.cpu = ptq->cpu;
1089         sample.flags = ptq->flags;
1090         sample.insn_len = ptq->insn_len;
1091         memcpy(sample.insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1092
1093         /*
1094          * perf report cannot handle events without a branch stack when using
1095          * SORT_MODE__BRANCH so make a dummy one.
1096          */
1097         if (pt->synth_opts.last_branch && sort__mode == SORT_MODE__BRANCH) {
1098                 dummy_bs = (struct dummy_branch_stack){
1099                         .nr = 1,
1100                         .entries = {
1101                                 .from = sample.ip,
1102                                 .to = sample.addr,
1103                         },
1104                 };
1105                 sample.branch_stack = (struct branch_stack *)&dummy_bs;
1106         }
1107
1108         if (pt->synth_opts.inject) {
1109                 ret = intel_pt_inject_event(event, &sample,
1110                                             pt->branches_sample_type,
1111                                             pt->synth_needs_swap);
1112                 if (ret)
1113                         return ret;
1114         }
1115
1116         ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1117         if (ret)
1118                 pr_err("Intel Processor Trace: failed to deliver branch event, error %d\n",
1119                        ret);
1120
1121         return ret;
1122 }
1123
1124 static int intel_pt_synth_instruction_sample(struct intel_pt_queue *ptq)
1125 {
1126         int ret;
1127         struct intel_pt *pt = ptq->pt;
1128         union perf_event *event = ptq->event_buf;
1129         struct perf_sample sample = { .ip = 0, };
1130
1131         if (pt->synth_opts.initial_skip &&
1132             pt->num_events++ < pt->synth_opts.initial_skip)
1133                 return 0;
1134
1135         event->sample.header.type = PERF_RECORD_SAMPLE;
1136         event->sample.header.misc = PERF_RECORD_MISC_USER;
1137         event->sample.header.size = sizeof(struct perf_event_header);
1138
1139         if (!pt->timeless_decoding)
1140                 sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1141
1142         sample.cpumode = PERF_RECORD_MISC_USER;
1143         sample.ip = ptq->state->from_ip;
1144         sample.pid = ptq->pid;
1145         sample.tid = ptq->tid;
1146         sample.addr = ptq->state->to_ip;
1147         sample.id = ptq->pt->instructions_id;
1148         sample.stream_id = ptq->pt->instructions_id;
1149         sample.period = ptq->state->tot_insn_cnt - ptq->last_insn_cnt;
1150         sample.cpu = ptq->cpu;
1151         sample.flags = ptq->flags;
1152         sample.insn_len = ptq->insn_len;
1153         memcpy(sample.insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1154
1155         ptq->last_insn_cnt = ptq->state->tot_insn_cnt;
1156
1157         if (pt->synth_opts.callchain) {
1158                 thread_stack__sample(ptq->thread, ptq->chain,
1159                                      pt->synth_opts.callchain_sz, sample.ip);
1160                 sample.callchain = ptq->chain;
1161         }
1162
1163         if (pt->synth_opts.last_branch) {
1164                 intel_pt_copy_last_branch_rb(ptq);
1165                 sample.branch_stack = ptq->last_branch;
1166         }
1167
1168         if (pt->synth_opts.inject) {
1169                 ret = intel_pt_inject_event(event, &sample,
1170                                             pt->instructions_sample_type,
1171                                             pt->synth_needs_swap);
1172                 if (ret)
1173                         return ret;
1174         }
1175
1176         ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1177         if (ret)
1178                 pr_err("Intel Processor Trace: failed to deliver instruction event, error %d\n",
1179                        ret);
1180
1181         if (pt->synth_opts.last_branch)
1182                 intel_pt_reset_last_branch_rb(ptq);
1183
1184         return ret;
1185 }
1186
1187 static int intel_pt_synth_transaction_sample(struct intel_pt_queue *ptq)
1188 {
1189         int ret;
1190         struct intel_pt *pt = ptq->pt;
1191         union perf_event *event = ptq->event_buf;
1192         struct perf_sample sample = { .ip = 0, };
1193
1194         if (pt->synth_opts.initial_skip &&
1195             pt->num_events++ < pt->synth_opts.initial_skip)
1196                 return 0;
1197
1198         event->sample.header.type = PERF_RECORD_SAMPLE;
1199         event->sample.header.misc = PERF_RECORD_MISC_USER;
1200         event->sample.header.size = sizeof(struct perf_event_header);
1201
1202         if (!pt->timeless_decoding)
1203                 sample.time = tsc_to_perf_time(ptq->timestamp, &pt->tc);
1204
1205         sample.cpumode = PERF_RECORD_MISC_USER;
1206         sample.ip = ptq->state->from_ip;
1207         sample.pid = ptq->pid;
1208         sample.tid = ptq->tid;
1209         sample.addr = ptq->state->to_ip;
1210         sample.id = ptq->pt->transactions_id;
1211         sample.stream_id = ptq->pt->transactions_id;
1212         sample.period = 1;
1213         sample.cpu = ptq->cpu;
1214         sample.flags = ptq->flags;
1215         sample.insn_len = ptq->insn_len;
1216         memcpy(sample.insn, ptq->insn, INTEL_PT_INSN_BUF_SZ);
1217
1218         if (pt->synth_opts.callchain) {
1219                 thread_stack__sample(ptq->thread, ptq->chain,
1220                                      pt->synth_opts.callchain_sz, sample.ip);
1221                 sample.callchain = ptq->chain;
1222         }
1223
1224         if (pt->synth_opts.last_branch) {
1225                 intel_pt_copy_last_branch_rb(ptq);
1226                 sample.branch_stack = ptq->last_branch;
1227         }
1228
1229         if (pt->synth_opts.inject) {
1230                 ret = intel_pt_inject_event(event, &sample,
1231                                             pt->transactions_sample_type,
1232                                             pt->synth_needs_swap);
1233                 if (ret)
1234                         return ret;
1235         }
1236
1237         ret = perf_session__deliver_synth_event(pt->session, event, &sample);
1238         if (ret)
1239                 pr_err("Intel Processor Trace: failed to deliver transaction event, error %d\n",
1240                        ret);
1241
1242         if (pt->synth_opts.last_branch)
1243                 intel_pt_reset_last_branch_rb(ptq);
1244
1245         return ret;
1246 }
1247
1248 static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,
1249                                 pid_t pid, pid_t tid, u64 ip)
1250 {
1251         union perf_event event;
1252         char msg[MAX_AUXTRACE_ERROR_MSG];
1253         int err;
1254
1255         intel_pt__strerror(code, msg, MAX_AUXTRACE_ERROR_MSG);
1256
1257         auxtrace_synth_error(&event.auxtrace_error, PERF_AUXTRACE_ERROR_ITRACE,
1258                              code, cpu, pid, tid, ip, msg);
1259
1260         err = perf_session__deliver_synth_event(pt->session, &event, NULL);
1261         if (err)
1262                 pr_err("Intel Processor Trace: failed to deliver error event, error %d\n",
1263                        err);
1264
1265         return err;
1266 }
1267
1268 static int intel_pt_next_tid(struct intel_pt *pt, struct intel_pt_queue *ptq)
1269 {
1270         struct auxtrace_queue *queue;
1271         pid_t tid = ptq->next_tid;
1272         int err;
1273
1274         if (tid == -1)
1275                 return 0;
1276
1277         intel_pt_log("switch: cpu %d tid %d\n", ptq->cpu, tid);
1278
1279         err = machine__set_current_tid(pt->machine, ptq->cpu, -1, tid);
1280
1281         queue = &pt->queues.queue_array[ptq->queue_nr];
1282         intel_pt_set_pid_tid_cpu(pt, queue);
1283
1284         ptq->next_tid = -1;
1285
1286         return err;
1287 }
1288
1289 static inline bool intel_pt_is_switch_ip(struct intel_pt_queue *ptq, u64 ip)
1290 {
1291         struct intel_pt *pt = ptq->pt;
1292
1293         return ip == pt->switch_ip &&
1294                (ptq->flags & PERF_IP_FLAG_BRANCH) &&
1295                !(ptq->flags & (PERF_IP_FLAG_CONDITIONAL | PERF_IP_FLAG_ASYNC |
1296                                PERF_IP_FLAG_INTERRUPT | PERF_IP_FLAG_TX_ABORT));
1297 }
1298
1299 static int intel_pt_sample(struct intel_pt_queue *ptq)
1300 {
1301         const struct intel_pt_state *state = ptq->state;
1302         struct intel_pt *pt = ptq->pt;
1303         int err;
1304
1305         if (!ptq->have_sample)
1306                 return 0;
1307
1308         ptq->have_sample = false;
1309
1310         if (pt->sample_instructions &&
1311             (state->type & INTEL_PT_INSTRUCTION) &&
1312             (!pt->synth_opts.initial_skip ||
1313              pt->num_events++ >= pt->synth_opts.initial_skip)) {
1314                 err = intel_pt_synth_instruction_sample(ptq);
1315                 if (err)
1316                         return err;
1317         }
1318
1319         if (pt->sample_transactions &&
1320             (state->type & INTEL_PT_TRANSACTION) &&
1321             (!pt->synth_opts.initial_skip ||
1322              pt->num_events++ >= pt->synth_opts.initial_skip)) {
1323                 err = intel_pt_synth_transaction_sample(ptq);
1324                 if (err)
1325                         return err;
1326         }
1327
1328         if (!(state->type & INTEL_PT_BRANCH))
1329                 return 0;
1330
1331         if (pt->synth_opts.callchain || pt->synth_opts.thread_stack)
1332                 thread_stack__event(ptq->thread, ptq->flags, state->from_ip,
1333                                     state->to_ip, ptq->insn_len,
1334                                     state->trace_nr);
1335         else
1336                 thread_stack__set_trace_nr(ptq->thread, state->trace_nr);
1337
1338         if (pt->sample_branches) {
1339                 err = intel_pt_synth_branch_sample(ptq);
1340                 if (err)
1341                         return err;
1342         }
1343
1344         if (pt->synth_opts.last_branch)
1345                 intel_pt_update_last_branch_rb(ptq);
1346
1347         if (!pt->sync_switch)
1348                 return 0;
1349
1350         if (intel_pt_is_switch_ip(ptq, state->to_ip)) {
1351                 switch (ptq->switch_state) {
1352                 case INTEL_PT_SS_UNKNOWN:
1353                 case INTEL_PT_SS_EXPECTING_SWITCH_IP:
1354                         err = intel_pt_next_tid(pt, ptq);
1355                         if (err)
1356                                 return err;
1357                         ptq->switch_state = INTEL_PT_SS_TRACING;
1358                         break;
1359                 default:
1360                         ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_EVENT;
1361                         return 1;
1362                 }
1363         } else if (!state->to_ip) {
1364                 ptq->switch_state = INTEL_PT_SS_NOT_TRACING;
1365         } else if (ptq->switch_state == INTEL_PT_SS_NOT_TRACING) {
1366                 ptq->switch_state = INTEL_PT_SS_UNKNOWN;
1367         } else if (ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
1368                    state->to_ip == pt->ptss_ip &&
1369                    (ptq->flags & PERF_IP_FLAG_CALL)) {
1370                 ptq->switch_state = INTEL_PT_SS_TRACING;
1371         }
1372
1373         return 0;
1374 }
1375
1376 static u64 intel_pt_switch_ip(struct intel_pt *pt, u64 *ptss_ip)
1377 {
1378         struct machine *machine = pt->machine;
1379         struct map *map;
1380         struct symbol *sym, *start;
1381         u64 ip, switch_ip = 0;
1382         const char *ptss;
1383
1384         if (ptss_ip)
1385                 *ptss_ip = 0;
1386
1387         map = machine__kernel_map(machine);
1388         if (!map)
1389                 return 0;
1390
1391         if (map__load(map))
1392                 return 0;
1393
1394         start = dso__first_symbol(map->dso, MAP__FUNCTION);
1395
1396         for (sym = start; sym; sym = dso__next_symbol(sym)) {
1397                 if (sym->binding == STB_GLOBAL &&
1398                     !strcmp(sym->name, "__switch_to")) {
1399                         ip = map->unmap_ip(map, sym->start);
1400                         if (ip >= map->start && ip < map->end) {
1401                                 switch_ip = ip;
1402                                 break;
1403                         }
1404                 }
1405         }
1406
1407         if (!switch_ip || !ptss_ip)
1408                 return 0;
1409
1410         if (pt->have_sched_switch == 1)
1411                 ptss = "perf_trace_sched_switch";
1412         else
1413                 ptss = "__perf_event_task_sched_out";
1414
1415         for (sym = start; sym; sym = dso__next_symbol(sym)) {
1416                 if (!strcmp(sym->name, ptss)) {
1417                         ip = map->unmap_ip(map, sym->start);
1418                         if (ip >= map->start && ip < map->end) {
1419                                 *ptss_ip = ip;
1420                                 break;
1421                         }
1422                 }
1423         }
1424
1425         return switch_ip;
1426 }
1427
1428 static int intel_pt_run_decoder(struct intel_pt_queue *ptq, u64 *timestamp)
1429 {
1430         const struct intel_pt_state *state = ptq->state;
1431         struct intel_pt *pt = ptq->pt;
1432         int err;
1433
1434         if (!pt->kernel_start) {
1435                 pt->kernel_start = machine__kernel_start(pt->machine);
1436                 if (pt->per_cpu_mmaps &&
1437                     (pt->have_sched_switch == 1 || pt->have_sched_switch == 3) &&
1438                     !pt->timeless_decoding && intel_pt_tracing_kernel(pt) &&
1439                     !pt->sampling_mode) {
1440                         pt->switch_ip = intel_pt_switch_ip(pt, &pt->ptss_ip);
1441                         if (pt->switch_ip) {
1442                                 intel_pt_log("switch_ip: %"PRIx64" ptss_ip: %"PRIx64"\n",
1443                                              pt->switch_ip, pt->ptss_ip);
1444                                 pt->sync_switch = true;
1445                         }
1446                 }
1447         }
1448
1449         intel_pt_log("queue %u decoding cpu %d pid %d tid %d\n",
1450                      ptq->queue_nr, ptq->cpu, ptq->pid, ptq->tid);
1451         while (1) {
1452                 err = intel_pt_sample(ptq);
1453                 if (err)
1454                         return err;
1455
1456                 state = intel_pt_decode(ptq->decoder);
1457                 if (state->err) {
1458                         if (state->err == INTEL_PT_ERR_NODATA)
1459                                 return 1;
1460                         if (pt->sync_switch &&
1461                             state->from_ip >= pt->kernel_start) {
1462                                 pt->sync_switch = false;
1463                                 intel_pt_next_tid(pt, ptq);
1464                         }
1465                         if (pt->synth_opts.errors) {
1466                                 err = intel_pt_synth_error(pt, state->err,
1467                                                            ptq->cpu, ptq->pid,
1468                                                            ptq->tid,
1469                                                            state->from_ip);
1470                                 if (err)
1471                                         return err;
1472                         }
1473                         continue;
1474                 }
1475
1476                 ptq->state = state;
1477                 ptq->have_sample = true;
1478                 intel_pt_sample_flags(ptq);
1479
1480                 /* Use estimated TSC upon return to user space */
1481                 if (pt->est_tsc &&
1482                     (state->from_ip >= pt->kernel_start || !state->from_ip) &&
1483                     state->to_ip && state->to_ip < pt->kernel_start) {
1484                         intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
1485                                      state->timestamp, state->est_timestamp);
1486                         ptq->timestamp = state->est_timestamp;
1487                 /* Use estimated TSC in unknown switch state */
1488                 } else if (pt->sync_switch &&
1489                            ptq->switch_state == INTEL_PT_SS_UNKNOWN &&
1490                            intel_pt_is_switch_ip(ptq, state->to_ip) &&
1491                            ptq->next_tid == -1) {
1492                         intel_pt_log("TSC %"PRIx64" est. TSC %"PRIx64"\n",
1493                                      state->timestamp, state->est_timestamp);
1494                         ptq->timestamp = state->est_timestamp;
1495                 } else if (state->timestamp > ptq->timestamp) {
1496                         ptq->timestamp = state->timestamp;
1497                 }
1498
1499                 if (!pt->timeless_decoding && ptq->timestamp >= *timestamp) {
1500                         *timestamp = ptq->timestamp;
1501                         return 0;
1502                 }
1503         }
1504         return 0;
1505 }
1506
1507 static inline int intel_pt_update_queues(struct intel_pt *pt)
1508 {
1509         if (pt->queues.new_data) {
1510                 pt->queues.new_data = false;
1511                 return intel_pt_setup_queues(pt);
1512         }
1513         return 0;
1514 }
1515
1516 static int intel_pt_process_queues(struct intel_pt *pt, u64 timestamp)
1517 {
1518         unsigned int queue_nr;
1519         u64 ts;
1520         int ret;
1521
1522         while (1) {
1523                 struct auxtrace_queue *queue;
1524                 struct intel_pt_queue *ptq;
1525
1526                 if (!pt->heap.heap_cnt)
1527                         return 0;
1528
1529                 if (pt->heap.heap_array[0].ordinal >= timestamp)
1530                         return 0;
1531
1532                 queue_nr = pt->heap.heap_array[0].queue_nr;
1533                 queue = &pt->queues.queue_array[queue_nr];
1534                 ptq = queue->priv;
1535
1536                 intel_pt_log("queue %u processing 0x%" PRIx64 " to 0x%" PRIx64 "\n",
1537                              queue_nr, pt->heap.heap_array[0].ordinal,
1538                              timestamp);
1539
1540                 auxtrace_heap__pop(&pt->heap);
1541
1542                 if (pt->heap.heap_cnt) {
1543                         ts = pt->heap.heap_array[0].ordinal + 1;
1544                         if (ts > timestamp)
1545                                 ts = timestamp;
1546                 } else {
1547                         ts = timestamp;
1548                 }
1549
1550                 intel_pt_set_pid_tid_cpu(pt, queue);
1551
1552                 ret = intel_pt_run_decoder(ptq, &ts);
1553
1554                 if (ret < 0) {
1555                         auxtrace_heap__add(&pt->heap, queue_nr, ts);
1556                         return ret;
1557                 }
1558
1559                 if (!ret) {
1560                         ret = auxtrace_heap__add(&pt->heap, queue_nr, ts);
1561                         if (ret < 0)
1562                                 return ret;
1563                 } else {
1564                         ptq->on_heap = false;
1565                 }
1566         }
1567
1568         return 0;
1569 }
1570
1571 static int intel_pt_process_timeless_queues(struct intel_pt *pt, pid_t tid,
1572                                             u64 time_)
1573 {
1574         struct auxtrace_queues *queues = &pt->queues;
1575         unsigned int i;
1576         u64 ts = 0;
1577
1578         for (i = 0; i < queues->nr_queues; i++) {
1579                 struct auxtrace_queue *queue = &pt->queues.queue_array[i];
1580                 struct intel_pt_queue *ptq = queue->priv;
1581
1582                 if (ptq && (tid == -1 || ptq->tid == tid)) {
1583                         ptq->time = time_;
1584                         intel_pt_set_pid_tid_cpu(pt, queue);
1585                         intel_pt_run_decoder(ptq, &ts);
1586                 }
1587         }
1588         return 0;
1589 }
1590
1591 static int intel_pt_lost(struct intel_pt *pt, struct perf_sample *sample)
1592 {
1593         return intel_pt_synth_error(pt, INTEL_PT_ERR_LOST, sample->cpu,
1594                                     sample->pid, sample->tid, 0);
1595 }
1596
1597 static struct intel_pt_queue *intel_pt_cpu_to_ptq(struct intel_pt *pt, int cpu)
1598 {
1599         unsigned i, j;
1600
1601         if (cpu < 0 || !pt->queues.nr_queues)
1602                 return NULL;
1603
1604         if ((unsigned)cpu >= pt->queues.nr_queues)
1605                 i = pt->queues.nr_queues - 1;
1606         else
1607                 i = cpu;
1608
1609         if (pt->queues.queue_array[i].cpu == cpu)
1610                 return pt->queues.queue_array[i].priv;
1611
1612         for (j = 0; i > 0; j++) {
1613                 if (pt->queues.queue_array[--i].cpu == cpu)
1614                         return pt->queues.queue_array[i].priv;
1615         }
1616
1617         for (; j < pt->queues.nr_queues; j++) {
1618                 if (pt->queues.queue_array[j].cpu == cpu)
1619                         return pt->queues.queue_array[j].priv;
1620         }
1621
1622         return NULL;
1623 }
1624
1625 static int intel_pt_sync_switch(struct intel_pt *pt, int cpu, pid_t tid,
1626                                 u64 timestamp)
1627 {
1628         struct intel_pt_queue *ptq;
1629         int err;
1630
1631         if (!pt->sync_switch)
1632                 return 1;
1633
1634         ptq = intel_pt_cpu_to_ptq(pt, cpu);
1635         if (!ptq)
1636                 return 1;
1637
1638         switch (ptq->switch_state) {
1639         case INTEL_PT_SS_NOT_TRACING:
1640                 ptq->next_tid = -1;
1641                 break;
1642         case INTEL_PT_SS_UNKNOWN:
1643         case INTEL_PT_SS_TRACING:
1644                 ptq->next_tid = tid;
1645                 ptq->switch_state = INTEL_PT_SS_EXPECTING_SWITCH_IP;
1646                 return 0;
1647         case INTEL_PT_SS_EXPECTING_SWITCH_EVENT:
1648                 if (!ptq->on_heap) {
1649                         ptq->timestamp = perf_time_to_tsc(timestamp,
1650                                                           &pt->tc);
1651                         err = auxtrace_heap__add(&pt->heap, ptq->queue_nr,
1652                                                  ptq->timestamp);
1653                         if (err)
1654                                 return err;
1655                         ptq->on_heap = true;
1656                 }
1657                 ptq->switch_state = INTEL_PT_SS_TRACING;
1658                 break;
1659         case INTEL_PT_SS_EXPECTING_SWITCH_IP:
1660                 ptq->next_tid = tid;
1661                 intel_pt_log("ERROR: cpu %d expecting switch ip\n", cpu);
1662                 break;
1663         default:
1664                 break;
1665         }
1666
1667         return 1;
1668 }
1669
1670 static int intel_pt_process_switch(struct intel_pt *pt,
1671                                    struct perf_sample *sample)
1672 {
1673         struct perf_evsel *evsel;
1674         pid_t tid;
1675         int cpu, ret;
1676
1677         evsel = perf_evlist__id2evsel(pt->session->evlist, sample->id);
1678         if (evsel != pt->switch_evsel)
1679                 return 0;
1680
1681         tid = perf_evsel__intval(evsel, sample, "next_pid");
1682         cpu = sample->cpu;
1683
1684         intel_pt_log("sched_switch: cpu %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1685                      cpu, tid, sample->time, perf_time_to_tsc(sample->time,
1686                      &pt->tc));
1687
1688         ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
1689         if (ret <= 0)
1690                 return ret;
1691
1692         return machine__set_current_tid(pt->machine, cpu, -1, tid);
1693 }
1694
1695 static int intel_pt_context_switch(struct intel_pt *pt, union perf_event *event,
1696                                    struct perf_sample *sample)
1697 {
1698         bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
1699         pid_t pid, tid;
1700         int cpu, ret;
1701
1702         cpu = sample->cpu;
1703
1704         if (pt->have_sched_switch == 3) {
1705                 if (!out)
1706                         return 0;
1707                 if (event->header.type != PERF_RECORD_SWITCH_CPU_WIDE) {
1708                         pr_err("Expecting CPU-wide context switch event\n");
1709                         return -EINVAL;
1710                 }
1711                 pid = event->context_switch.next_prev_pid;
1712                 tid = event->context_switch.next_prev_tid;
1713         } else {
1714                 if (out)
1715                         return 0;
1716                 pid = sample->pid;
1717                 tid = sample->tid;
1718         }
1719
1720         if (tid == -1) {
1721                 pr_err("context_switch event has no tid\n");
1722                 return -EINVAL;
1723         }
1724
1725         intel_pt_log("context_switch: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1726                      cpu, pid, tid, sample->time, perf_time_to_tsc(sample->time,
1727                      &pt->tc));
1728
1729         ret = intel_pt_sync_switch(pt, cpu, tid, sample->time);
1730         if (ret <= 0)
1731                 return ret;
1732
1733         return machine__set_current_tid(pt->machine, cpu, pid, tid);
1734 }
1735
1736 static int intel_pt_process_itrace_start(struct intel_pt *pt,
1737                                          union perf_event *event,
1738                                          struct perf_sample *sample)
1739 {
1740         if (!pt->per_cpu_mmaps)
1741                 return 0;
1742
1743         intel_pt_log("itrace_start: cpu %d pid %d tid %d time %"PRIu64" tsc %#"PRIx64"\n",
1744                      sample->cpu, event->itrace_start.pid,
1745                      event->itrace_start.tid, sample->time,
1746                      perf_time_to_tsc(sample->time, &pt->tc));
1747
1748         return machine__set_current_tid(pt->machine, sample->cpu,
1749                                         event->itrace_start.pid,
1750                                         event->itrace_start.tid);
1751 }
1752
1753 static int intel_pt_process_event(struct perf_session *session,
1754                                   union perf_event *event,
1755                                   struct perf_sample *sample,
1756                                   struct perf_tool *tool)
1757 {
1758         struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1759                                            auxtrace);
1760         u64 timestamp;
1761         int err = 0;
1762
1763         if (dump_trace)
1764                 return 0;
1765
1766         if (!tool->ordered_events) {
1767                 pr_err("Intel Processor Trace requires ordered events\n");
1768                 return -EINVAL;
1769         }
1770
1771         if (sample->time && sample->time != (u64)-1)
1772                 timestamp = perf_time_to_tsc(sample->time, &pt->tc);
1773         else
1774                 timestamp = 0;
1775
1776         if (timestamp || pt->timeless_decoding) {
1777                 err = intel_pt_update_queues(pt);
1778                 if (err)
1779                         return err;
1780         }
1781
1782         if (pt->timeless_decoding) {
1783                 if (event->header.type == PERF_RECORD_EXIT) {
1784                         err = intel_pt_process_timeless_queues(pt,
1785                                                                event->fork.tid,
1786                                                                sample->time);
1787                 }
1788         } else if (timestamp) {
1789                 err = intel_pt_process_queues(pt, timestamp);
1790         }
1791         if (err)
1792                 return err;
1793
1794         if (event->header.type == PERF_RECORD_AUX &&
1795             (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) &&
1796             pt->synth_opts.errors) {
1797                 err = intel_pt_lost(pt, sample);
1798                 if (err)
1799                         return err;
1800         }
1801
1802         if (pt->switch_evsel && event->header.type == PERF_RECORD_SAMPLE)
1803                 err = intel_pt_process_switch(pt, sample);
1804         else if (event->header.type == PERF_RECORD_ITRACE_START)
1805                 err = intel_pt_process_itrace_start(pt, event, sample);
1806         else if (event->header.type == PERF_RECORD_SWITCH ||
1807                  event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
1808                 err = intel_pt_context_switch(pt, event, sample);
1809
1810         intel_pt_log("event %s (%u): cpu %d time %"PRIu64" tsc %#"PRIx64"\n",
1811                      perf_event__name(event->header.type), event->header.type,
1812                      sample->cpu, sample->time, timestamp);
1813
1814         return err;
1815 }
1816
1817 static int intel_pt_flush(struct perf_session *session, struct perf_tool *tool)
1818 {
1819         struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1820                                            auxtrace);
1821         int ret;
1822
1823         if (dump_trace)
1824                 return 0;
1825
1826         if (!tool->ordered_events)
1827                 return -EINVAL;
1828
1829         ret = intel_pt_update_queues(pt);
1830         if (ret < 0)
1831                 return ret;
1832
1833         if (pt->timeless_decoding)
1834                 return intel_pt_process_timeless_queues(pt, -1,
1835                                                         MAX_TIMESTAMP - 1);
1836
1837         return intel_pt_process_queues(pt, MAX_TIMESTAMP);
1838 }
1839
1840 static void intel_pt_free_events(struct perf_session *session)
1841 {
1842         struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1843                                            auxtrace);
1844         struct auxtrace_queues *queues = &pt->queues;
1845         unsigned int i;
1846
1847         for (i = 0; i < queues->nr_queues; i++) {
1848                 intel_pt_free_queue(queues->queue_array[i].priv);
1849                 queues->queue_array[i].priv = NULL;
1850         }
1851         intel_pt_log_disable();
1852         auxtrace_queues__free(queues);
1853 }
1854
1855 static void intel_pt_free(struct perf_session *session)
1856 {
1857         struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1858                                            auxtrace);
1859
1860         auxtrace_heap__free(&pt->heap);
1861         intel_pt_free_events(session);
1862         session->auxtrace = NULL;
1863         thread__put(pt->unknown_thread);
1864         addr_filters__exit(&pt->filts);
1865         zfree(&pt->filter);
1866         free(pt);
1867 }
1868
1869 static int intel_pt_process_auxtrace_event(struct perf_session *session,
1870                                            union perf_event *event,
1871                                            struct perf_tool *tool __maybe_unused)
1872 {
1873         struct intel_pt *pt = container_of(session->auxtrace, struct intel_pt,
1874                                            auxtrace);
1875
1876         if (pt->sampling_mode)
1877                 return 0;
1878
1879         if (!pt->data_queued) {
1880                 struct auxtrace_buffer *buffer;
1881                 off_t data_offset;
1882                 int fd = perf_data_file__fd(session->file);
1883                 int err;
1884
1885                 if (perf_data_file__is_pipe(session->file)) {
1886                         data_offset = 0;
1887                 } else {
1888                         data_offset = lseek(fd, 0, SEEK_CUR);
1889                         if (data_offset == -1)
1890                                 return -errno;
1891                 }
1892
1893                 err = auxtrace_queues__add_event(&pt->queues, session, event,
1894                                                  data_offset, &buffer);
1895                 if (err)
1896                         return err;
1897
1898                 /* Dump here now we have copied a piped trace out of the pipe */
1899                 if (dump_trace) {
1900                         if (auxtrace_buffer__get_data(buffer, fd)) {
1901                                 intel_pt_dump_event(pt, buffer->data,
1902                                                     buffer->size);
1903                                 auxtrace_buffer__put_data(buffer);
1904                         }
1905                 }
1906         }
1907
1908         return 0;
1909 }
1910
1911 struct intel_pt_synth {
1912         struct perf_tool dummy_tool;
1913         struct perf_session *session;
1914 };
1915
1916 static int intel_pt_event_synth(struct perf_tool *tool,
1917                                 union perf_event *event,
1918                                 struct perf_sample *sample __maybe_unused,
1919                                 struct machine *machine __maybe_unused)
1920 {
1921         struct intel_pt_synth *intel_pt_synth =
1922                         container_of(tool, struct intel_pt_synth, dummy_tool);
1923
1924         return perf_session__deliver_synth_event(intel_pt_synth->session, event,
1925                                                  NULL);
1926 }
1927
1928 static int intel_pt_synth_event(struct perf_session *session,
1929                                 struct perf_event_attr *attr, u64 id)
1930 {
1931         struct intel_pt_synth intel_pt_synth;
1932
1933         memset(&intel_pt_synth, 0, sizeof(struct intel_pt_synth));
1934         intel_pt_synth.session = session;
1935
1936         return perf_event__synthesize_attr(&intel_pt_synth.dummy_tool, attr, 1,
1937                                            &id, intel_pt_event_synth);
1938 }
1939
1940 static int intel_pt_synth_events(struct intel_pt *pt,
1941                                  struct perf_session *session)
1942 {
1943         struct perf_evlist *evlist = session->evlist;
1944         struct perf_evsel *evsel;
1945         struct perf_event_attr attr;
1946         bool found = false;
1947         u64 id;
1948         int err;
1949
1950         evlist__for_each_entry(evlist, evsel) {
1951                 if (evsel->attr.type == pt->pmu_type && evsel->ids) {
1952                         found = true;
1953                         break;
1954                 }
1955         }
1956
1957         if (!found) {
1958                 pr_debug("There are no selected events with Intel Processor Trace data\n");
1959                 return 0;
1960         }
1961
1962         memset(&attr, 0, sizeof(struct perf_event_attr));
1963         attr.size = sizeof(struct perf_event_attr);
1964         attr.type = PERF_TYPE_HARDWARE;
1965         attr.sample_type = evsel->attr.sample_type & PERF_SAMPLE_MASK;
1966         attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
1967                             PERF_SAMPLE_PERIOD;
1968         if (pt->timeless_decoding)
1969                 attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
1970         else
1971                 attr.sample_type |= PERF_SAMPLE_TIME;
1972         if (!pt->per_cpu_mmaps)
1973                 attr.sample_type &= ~(u64)PERF_SAMPLE_CPU;
1974         attr.exclude_user = evsel->attr.exclude_user;
1975         attr.exclude_kernel = evsel->attr.exclude_kernel;
1976         attr.exclude_hv = evsel->attr.exclude_hv;
1977         attr.exclude_host = evsel->attr.exclude_host;
1978         attr.exclude_guest = evsel->attr.exclude_guest;
1979         attr.sample_id_all = evsel->attr.sample_id_all;
1980         attr.read_format = evsel->attr.read_format;
1981
1982         id = evsel->id[0] + 1000000000;
1983         if (!id)
1984                 id = 1;
1985
1986         if (pt->synth_opts.instructions) {
1987                 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
1988                 if (pt->synth_opts.period_type == PERF_ITRACE_PERIOD_NANOSECS)
1989                         attr.sample_period =
1990                                 intel_pt_ns_to_ticks(pt, pt->synth_opts.period);
1991                 else
1992                         attr.sample_period = pt->synth_opts.period;
1993                 pt->instructions_sample_period = attr.sample_period;
1994                 if (pt->synth_opts.callchain)
1995                         attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
1996                 if (pt->synth_opts.last_branch)
1997                         attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
1998                 pr_debug("Synthesizing 'instructions' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
1999                          id, (u64)attr.sample_type);
2000                 err = intel_pt_synth_event(session, &attr, id);
2001                 if (err) {
2002                         pr_err("%s: failed to synthesize 'instructions' event type\n",
2003                                __func__);
2004                         return err;
2005                 }
2006                 pt->sample_instructions = true;
2007                 pt->instructions_sample_type = attr.sample_type;
2008                 pt->instructions_id = id;
2009                 id += 1;
2010         }
2011
2012         if (pt->synth_opts.transactions) {
2013                 attr.config = PERF_COUNT_HW_INSTRUCTIONS;
2014                 attr.sample_period = 1;
2015                 if (pt->synth_opts.callchain)
2016                         attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
2017                 if (pt->synth_opts.last_branch)
2018                         attr.sample_type |= PERF_SAMPLE_BRANCH_STACK;
2019                 pr_debug("Synthesizing 'transactions' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
2020                          id, (u64)attr.sample_type);
2021                 err = intel_pt_synth_event(session, &attr, id);
2022                 if (err) {
2023                         pr_err("%s: failed to synthesize 'transactions' event type\n",
2024                                __func__);
2025                         return err;
2026                 }
2027                 pt->sample_transactions = true;
2028                 pt->transactions_id = id;
2029                 id += 1;
2030                 evlist__for_each_entry(evlist, evsel) {
2031                         if (evsel->id && evsel->id[0] == pt->transactions_id) {
2032                                 if (evsel->name)
2033                                         zfree(&evsel->name);
2034                                 evsel->name = strdup("transactions");
2035                                 break;
2036                         }
2037                 }
2038         }
2039
2040         if (pt->synth_opts.branches) {
2041                 attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
2042                 attr.sample_period = 1;
2043                 attr.sample_type |= PERF_SAMPLE_ADDR;
2044                 attr.sample_type &= ~(u64)PERF_SAMPLE_CALLCHAIN;
2045                 attr.sample_type &= ~(u64)PERF_SAMPLE_BRANCH_STACK;
2046                 pr_debug("Synthesizing 'branches' event with id %" PRIu64 " sample type %#" PRIx64 "\n",
2047                          id, (u64)attr.sample_type);
2048                 err = intel_pt_synth_event(session, &attr, id);
2049                 if (err) {
2050                         pr_err("%s: failed to synthesize 'branches' event type\n",
2051                                __func__);
2052                         return err;
2053                 }
2054                 pt->sample_branches = true;
2055                 pt->branches_sample_type = attr.sample_type;
2056                 pt->branches_id = id;
2057         }
2058
2059         pt->synth_needs_swap = evsel->needs_swap;
2060
2061         return 0;
2062 }
2063
2064 static struct perf_evsel *intel_pt_find_sched_switch(struct perf_evlist *evlist)
2065 {
2066         struct perf_evsel *evsel;
2067
2068         evlist__for_each_entry_reverse(evlist, evsel) {
2069                 const char *name = perf_evsel__name(evsel);
2070
2071                 if (!strcmp(name, "sched:sched_switch"))
2072                         return evsel;
2073         }
2074
2075         return NULL;
2076 }
2077
2078 static bool intel_pt_find_switch(struct perf_evlist *evlist)
2079 {
2080         struct perf_evsel *evsel;
2081
2082         evlist__for_each_entry(evlist, evsel) {
2083                 if (evsel->attr.context_switch)
2084                         return true;
2085         }
2086
2087         return false;
2088 }
2089
2090 static int intel_pt_perf_config(const char *var, const char *value, void *data)
2091 {
2092         struct intel_pt *pt = data;
2093
2094         if (!strcmp(var, "intel-pt.mispred-all"))
2095                 pt->mispred_all = perf_config_bool(var, value);
2096
2097         return 0;
2098 }
2099
2100 static const char * const intel_pt_info_fmts[] = {
2101         [INTEL_PT_PMU_TYPE]             = "  PMU Type            %"PRId64"\n",
2102         [INTEL_PT_TIME_SHIFT]           = "  Time Shift          %"PRIu64"\n",
2103         [INTEL_PT_TIME_MULT]            = "  Time Muliplier      %"PRIu64"\n",
2104         [INTEL_PT_TIME_ZERO]            = "  Time Zero           %"PRIu64"\n",
2105         [INTEL_PT_CAP_USER_TIME_ZERO]   = "  Cap Time Zero       %"PRId64"\n",
2106         [INTEL_PT_TSC_BIT]              = "  TSC bit             %#"PRIx64"\n",
2107         [INTEL_PT_NORETCOMP_BIT]        = "  NoRETComp bit       %#"PRIx64"\n",
2108         [INTEL_PT_HAVE_SCHED_SWITCH]    = "  Have sched_switch   %"PRId64"\n",
2109         [INTEL_PT_SNAPSHOT_MODE]        = "  Snapshot mode       %"PRId64"\n",
2110         [INTEL_PT_PER_CPU_MMAPS]        = "  Per-cpu maps        %"PRId64"\n",
2111         [INTEL_PT_MTC_BIT]              = "  MTC bit             %#"PRIx64"\n",
2112         [INTEL_PT_TSC_CTC_N]            = "  TSC:CTC numerator   %"PRIu64"\n",
2113         [INTEL_PT_TSC_CTC_D]            = "  TSC:CTC denominator %"PRIu64"\n",
2114         [INTEL_PT_CYC_BIT]              = "  CYC bit             %#"PRIx64"\n",
2115         [INTEL_PT_MAX_NONTURBO_RATIO]   = "  Max non-turbo ratio %"PRIu64"\n",
2116         [INTEL_PT_FILTER_STR_LEN]       = "  Filter string len.  %"PRIu64"\n",
2117 };
2118
2119 static void intel_pt_print_info(u64 *arr, int start, int finish)
2120 {
2121         int i;
2122
2123         if (!dump_trace)
2124                 return;
2125
2126         for (i = start; i <= finish; i++)
2127                 fprintf(stdout, intel_pt_info_fmts[i], arr[i]);
2128 }
2129
2130 static void intel_pt_print_info_str(const char *name, const char *str)
2131 {
2132         if (!dump_trace)
2133                 return;
2134
2135         fprintf(stdout, "  %-20s%s\n", name, str ? str : "");
2136 }
2137
2138 static bool intel_pt_has(struct auxtrace_info_event *auxtrace_info, int pos)
2139 {
2140         return auxtrace_info->header.size >=
2141                 sizeof(struct auxtrace_info_event) + (sizeof(u64) * (pos + 1));
2142 }
2143
2144 int intel_pt_process_auxtrace_info(union perf_event *event,
2145                                    struct perf_session *session)
2146 {
2147         struct auxtrace_info_event *auxtrace_info = &event->auxtrace_info;
2148         size_t min_sz = sizeof(u64) * INTEL_PT_PER_CPU_MMAPS;
2149         struct intel_pt *pt;
2150         void *info_end;
2151         u64 *info;
2152         int err;
2153
2154         if (auxtrace_info->header.size < sizeof(struct auxtrace_info_event) +
2155                                         min_sz)
2156                 return -EINVAL;
2157
2158         pt = zalloc(sizeof(struct intel_pt));
2159         if (!pt)
2160                 return -ENOMEM;
2161
2162         addr_filters__init(&pt->filts);
2163
2164         err = perf_config(intel_pt_perf_config, pt);
2165         if (err)
2166                 goto err_free;
2167
2168         err = auxtrace_queues__init(&pt->queues);
2169         if (err)
2170                 goto err_free;
2171
2172         intel_pt_log_set_name(INTEL_PT_PMU_NAME);
2173
2174         pt->session = session;
2175         pt->machine = &session->machines.host; /* No kvm support */
2176         pt->auxtrace_type = auxtrace_info->type;
2177         pt->pmu_type = auxtrace_info->priv[INTEL_PT_PMU_TYPE];
2178         pt->tc.time_shift = auxtrace_info->priv[INTEL_PT_TIME_SHIFT];
2179         pt->tc.time_mult = auxtrace_info->priv[INTEL_PT_TIME_MULT];
2180         pt->tc.time_zero = auxtrace_info->priv[INTEL_PT_TIME_ZERO];
2181         pt->cap_user_time_zero = auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO];
2182         pt->tsc_bit = auxtrace_info->priv[INTEL_PT_TSC_BIT];
2183         pt->noretcomp_bit = auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT];
2184         pt->have_sched_switch = auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH];
2185         pt->snapshot_mode = auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE];
2186         pt->per_cpu_mmaps = auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS];
2187         intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_PMU_TYPE,
2188                             INTEL_PT_PER_CPU_MMAPS);
2189
2190         if (intel_pt_has(auxtrace_info, INTEL_PT_CYC_BIT)) {
2191                 pt->mtc_bit = auxtrace_info->priv[INTEL_PT_MTC_BIT];
2192                 pt->mtc_freq_bits = auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS];
2193                 pt->tsc_ctc_ratio_n = auxtrace_info->priv[INTEL_PT_TSC_CTC_N];
2194                 pt->tsc_ctc_ratio_d = auxtrace_info->priv[INTEL_PT_TSC_CTC_D];
2195                 pt->cyc_bit = auxtrace_info->priv[INTEL_PT_CYC_BIT];
2196                 intel_pt_print_info(&auxtrace_info->priv[0], INTEL_PT_MTC_BIT,
2197                                     INTEL_PT_CYC_BIT);
2198         }
2199
2200         if (intel_pt_has(auxtrace_info, INTEL_PT_MAX_NONTURBO_RATIO)) {
2201                 pt->max_non_turbo_ratio =
2202                         auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO];
2203                 intel_pt_print_info(&auxtrace_info->priv[0],
2204                                     INTEL_PT_MAX_NONTURBO_RATIO,
2205                                     INTEL_PT_MAX_NONTURBO_RATIO);
2206         }
2207
2208         info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
2209         info_end = (void *)info + auxtrace_info->header.size;
2210
2211         if (intel_pt_has(auxtrace_info, INTEL_PT_FILTER_STR_LEN)) {
2212                 size_t len;
2213
2214                 len = auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN];
2215                 intel_pt_print_info(&auxtrace_info->priv[0],
2216                                     INTEL_PT_FILTER_STR_LEN,
2217                                     INTEL_PT_FILTER_STR_LEN);
2218                 if (len) {
2219                         const char *filter = (const char *)info;
2220
2221                         len = roundup(len + 1, 8);
2222                         info += len >> 3;
2223                         if ((void *)info > info_end) {
2224                                 pr_err("%s: bad filter string length\n", __func__);
2225                                 err = -EINVAL;
2226                                 goto err_free_queues;
2227                         }
2228                         pt->filter = memdup(filter, len);
2229                         if (!pt->filter) {
2230                                 err = -ENOMEM;
2231                                 goto err_free_queues;
2232                         }
2233                         if (session->header.needs_swap)
2234                                 mem_bswap_64(pt->filter, len);
2235                         if (pt->filter[len - 1]) {
2236                                 pr_err("%s: filter string not null terminated\n", __func__);
2237                                 err = -EINVAL;
2238                                 goto err_free_queues;
2239                         }
2240                         err = addr_filters__parse_bare_filter(&pt->filts,
2241                                                               filter);
2242                         if (err)
2243                                 goto err_free_queues;
2244                 }
2245                 intel_pt_print_info_str("Filter string", pt->filter);
2246         }
2247
2248         pt->timeless_decoding = intel_pt_timeless_decoding(pt);
2249         pt->have_tsc = intel_pt_have_tsc(pt);
2250         pt->sampling_mode = false;
2251         pt->est_tsc = !pt->timeless_decoding;
2252
2253         pt->unknown_thread = thread__new(999999999, 999999999);
2254         if (!pt->unknown_thread) {
2255                 err = -ENOMEM;
2256                 goto err_free_queues;
2257         }
2258
2259         /*
2260          * Since this thread will not be kept in any rbtree not in a
2261          * list, initialize its list node so that at thread__put() the
2262          * current thread lifetime assuption is kept and we don't segfault
2263          * at list_del_init().
2264          */
2265         INIT_LIST_HEAD(&pt->unknown_thread->node);
2266
2267         err = thread__set_comm(pt->unknown_thread, "unknown", 0);
2268         if (err)
2269                 goto err_delete_thread;
2270         if (thread__init_map_groups(pt->unknown_thread, pt->machine)) {
2271                 err = -ENOMEM;
2272                 goto err_delete_thread;
2273         }
2274
2275         pt->auxtrace.process_event = intel_pt_process_event;
2276         pt->auxtrace.process_auxtrace_event = intel_pt_process_auxtrace_event;
2277         pt->auxtrace.flush_events = intel_pt_flush;
2278         pt->auxtrace.free_events = intel_pt_free_events;
2279         pt->auxtrace.free = intel_pt_free;
2280         session->auxtrace = &pt->auxtrace;
2281
2282         if (dump_trace)
2283                 return 0;
2284
2285         if (pt->have_sched_switch == 1) {
2286                 pt->switch_evsel = intel_pt_find_sched_switch(session->evlist);
2287                 if (!pt->switch_evsel) {
2288                         pr_err("%s: missing sched_switch event\n", __func__);
2289                         err = -EINVAL;
2290                         goto err_delete_thread;
2291                 }
2292         } else if (pt->have_sched_switch == 2 &&
2293                    !intel_pt_find_switch(session->evlist)) {
2294                 pr_err("%s: missing context_switch attribute flag\n", __func__);
2295                 err = -EINVAL;
2296                 goto err_delete_thread;
2297         }
2298
2299         if (session->itrace_synth_opts && session->itrace_synth_opts->set) {
2300                 pt->synth_opts = *session->itrace_synth_opts;
2301         } else {
2302                 itrace_synth_opts__set_default(&pt->synth_opts);
2303                 if (use_browser != -1) {
2304                         pt->synth_opts.branches = false;
2305                         pt->synth_opts.callchain = true;
2306                 }
2307                 if (session->itrace_synth_opts)
2308                         pt->synth_opts.thread_stack =
2309                                 session->itrace_synth_opts->thread_stack;
2310         }
2311
2312         if (pt->synth_opts.log)
2313                 intel_pt_log_enable();
2314
2315         /* Maximum non-turbo ratio is TSC freq / 100 MHz */
2316         if (pt->tc.time_mult) {
2317                 u64 tsc_freq = intel_pt_ns_to_ticks(pt, 1000000000);
2318
2319                 if (!pt->max_non_turbo_ratio)
2320                         pt->max_non_turbo_ratio =
2321                                         (tsc_freq + 50000000) / 100000000;
2322                 intel_pt_log("TSC frequency %"PRIu64"\n", tsc_freq);
2323                 intel_pt_log("Maximum non-turbo ratio %u\n",
2324                              pt->max_non_turbo_ratio);
2325         }
2326
2327         if (pt->synth_opts.calls)
2328                 pt->branches_filter |= PERF_IP_FLAG_CALL | PERF_IP_FLAG_ASYNC |
2329                                        PERF_IP_FLAG_TRACE_END;
2330         if (pt->synth_opts.returns)
2331                 pt->branches_filter |= PERF_IP_FLAG_RETURN |
2332                                        PERF_IP_FLAG_TRACE_BEGIN;
2333
2334         if (pt->synth_opts.callchain && !symbol_conf.use_callchain) {
2335                 symbol_conf.use_callchain = true;
2336                 if (callchain_register_param(&callchain_param) < 0) {
2337                         symbol_conf.use_callchain = false;
2338                         pt->synth_opts.callchain = false;
2339                 }
2340         }
2341
2342         err = intel_pt_synth_events(pt, session);
2343         if (err)
2344                 goto err_delete_thread;
2345
2346         err = auxtrace_queues__process_index(&pt->queues, session);
2347         if (err)
2348                 goto err_delete_thread;
2349
2350         if (pt->queues.populated)
2351                 pt->data_queued = true;
2352
2353         if (pt->timeless_decoding)
2354                 pr_debug2("Intel PT decoding without timestamps\n");
2355
2356         return 0;
2357
2358 err_delete_thread:
2359         thread__zput(pt->unknown_thread);
2360 err_free_queues:
2361         intel_pt_log_disable();
2362         auxtrace_queues__free(&pt->queues);
2363         session->auxtrace = NULL;
2364 err_free:
2365         addr_filters__exit(&pt->filts);
2366         zfree(&pt->filter);
2367         free(pt);
2368         return err;
2369 }