tools/perf/arch/x86/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 <stdbool.h>
  17 #include <linux/kernel.h>
  18 #include <linux/types.h>
  19 #include <linux/bitops.h>
  20 #include <linux/log2.h>
  21 #include <cpuid.h>
  22
  23 #include "../../perf.h"
  24 #include "../../util/session.h"
  25 #include "../../util/event.h"
  26 #include "../../util/evlist.h"
  27 #include "../../util/evsel.h"
  28 #include "../../util/cpumap.h"
  29 #include <subcmd/parse-options.h>
  30 #include "../../util/parse-events.h"
  31 #include "../../util/pmu.h"
  32 #include "../../util/debug.h"
  33 #include "../../util/auxtrace.h"
  34 #include "../../util/tsc.h"
  35 #include "../../util/intel-pt.h"
  36
  37 #define KiB(x) ((x) * 1024)
  38 #define MiB(x) ((x) * 1024 * 1024)
  39 #define KiB_MASK(x) (KiB(x) - 1)
  40 #define MiB_MASK(x) (MiB(x) - 1)
  41
  42 #define INTEL_PT_DEFAULT_SAMPLE_SIZE    KiB(4)
  43
  44 #define INTEL_PT_MAX_SAMPLE_SIZE        KiB(60)
  45
  46 #define INTEL_PT_PSB_PERIOD_NEAR        256
  47
  48 struct intel_pt_snapshot_ref {
  49         void *ref_buf;
  50         size_t ref_offset;
  51         bool wrapped;
  52 };
  53
  54 struct intel_pt_recording {
  55         struct auxtrace_record          itr;
  56         struct perf_pmu                 *intel_pt_pmu;
  57         int                             have_sched_switch;
  58         struct perf_evlist              *evlist;
  59         bool                            snapshot_mode;
  60         bool                            snapshot_init_done;
  61         size_t                          snapshot_size;
  62         size_t                          snapshot_ref_buf_size;
  63         int                             snapshot_ref_cnt;
  64         struct intel_pt_snapshot_ref    *snapshot_refs;
  65         size_t                          priv_size;
  66 };
  67
  68 static int intel_pt_parse_terms_with_default(struct list_head *formats,
  69                                              const char *str,
  70                                              u64 *config)
  71 {
  72         struct list_head *terms;
  73         struct perf_event_attr attr = { .size = 0, };
  74         int err;
  75
  76         terms = malloc(sizeof(struct list_head));
  77         if (!terms)
  78                 return -ENOMEM;
  79
  80         INIT_LIST_HEAD(terms);
  81
  82         err = parse_events_terms(terms, str);
  83         if (err)
  84                 goto out_free;
  85
  86         attr.config = *config;
  87         err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
  88         if (err)
  89                 goto out_free;
  90
  91         *config = attr.config;
  92 out_free:
  93         parse_events_terms__delete(terms);
  94         return err;
  95 }
  96
  97 static int intel_pt_parse_terms(struct list_head *formats, const char *str,
  98                                 u64 *config)
  99 {
 100         *config = 0;
 101         return intel_pt_parse_terms_with_default(formats, str, config);
 102 }
 103
 104 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
 105 {
 106         const u64 top_bit = 1ULL << 63;
 107         u64 res = 0;
 108         int i;
 109
 110         for (i = 0; i < 64; i++) {
 111                 if (mask & top_bit) {
 112                         res <<= 1;
 113                         if (bits & top_bit)
 114                                 res |= 1;
 115                 }
 116                 mask <<= 1;
 117                 bits <<= 1;
 118         }
 119
 120         return res;
 121 }
 122
 123 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
 124                                 struct perf_evlist *evlist, u64 *res)
 125 {
 126         struct perf_evsel *evsel;
 127         u64 mask;
 128
 129         *res = 0;
 130
 131         mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
 132         if (!mask)
 133                 return -EINVAL;
 134
 135         evlist__for_each_entry(evlist, evsel) {
 136                 if (evsel->attr.type == intel_pt_pmu->type) {
 137                         *res = intel_pt_masked_bits(mask, evsel->attr.config);
 138                         return 0;
 139                 }
 140         }
 141
 142         return -EINVAL;
 143 }
 144
 145 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
 146                                   struct perf_evlist *evlist)
 147 {
 148         u64 val;
 149         int err, topa_multiple_entries;
 150         size_t psb_period;
 151
 152         if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
 153                                 "%d", &topa_multiple_entries) != 1)
 154                 topa_multiple_entries = 0;
 155
 156         /*
 157          * Use caps/topa_multiple_entries to indicate early hardware that had
 158          * extra frequent PSBs.
 159          */
 160         if (!topa_multiple_entries) {
 161                 psb_period = 256;
 162                 goto out;
 163         }
 164
 165         err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
 166         if (err)
 167                 val = 0;
 168
 169         psb_period = 1 << (val + 11);
 170 out:
 171         pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
 172         return psb_period;
 173 }
 174
 175 static int intel_pt_pick_bit(int bits, int target)
 176 {
 177         int pos, pick = -1;
 178
 179         for (pos = 0; bits; bits >>= 1, pos++) {
 180                 if (bits & 1) {
 181                         if (pos <= target || pick < 0)
 182                                 pick = pos;
 183                         if (pos >= target)
 184                                 break;
 185                 }
 186         }
 187
 188         return pick;
 189 }
 190
 191 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
 192 {
 193         char buf[256];
 194         int mtc, mtc_periods = 0, mtc_period;
 195         int psb_cyc, psb_periods, psb_period;
 196         int pos = 0;
 197         u64 config;
 198
 199         pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
 200
 201         if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
 202                                 &mtc) != 1)
 203                 mtc = 1;
 204
 205         if (mtc) {
 206                 if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
 207                                         &mtc_periods) != 1)
 208                         mtc_periods = 0;
 209                 if (mtc_periods) {
 210                         mtc_period = intel_pt_pick_bit(mtc_periods, 3);
 211                         pos += scnprintf(buf + pos, sizeof(buf) - pos,
 212                                          ",mtc,mtc_period=%d", mtc_period);
 213                 }
 214         }
 215
 216         if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
 217                                 &psb_cyc) != 1)
 218                 psb_cyc = 1;
 219
 220         if (psb_cyc && mtc_periods) {
 221                 if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
 222                                         &psb_periods) != 1)
 223                         psb_periods = 0;
 224                 if (psb_periods) {
 225                         psb_period = intel_pt_pick_bit(psb_periods, 3);
 226                         pos += scnprintf(buf + pos, sizeof(buf) - pos,
 227                                          ",psb_period=%d", psb_period);
 228                 }
 229         }
 230
 231         pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
 232
 233         intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
 234
 235         return config;
 236 }
 237
 238 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
 239                                            struct record_opts *opts,
 240                                            const char *str)
 241 {
 242         struct intel_pt_recording *ptr =
 243                         container_of(itr, struct intel_pt_recording, itr);
 244         unsigned long long snapshot_size = 0;
 245         char *endptr;
 246
 247         if (str) {
 248                 snapshot_size = strtoull(str, &endptr, 0);
 249                 if (*endptr || snapshot_size > SIZE_MAX)
 250                         return -1;
 251         }
 252
 253         opts->auxtrace_snapshot_mode = true;
 254         opts->auxtrace_snapshot_size = snapshot_size;
 255
 256         ptr->snapshot_size = snapshot_size;
 257
 258         return 0;
 259 }
 260
 261 struct perf_event_attr *
 262 intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
 263 {
 264         struct perf_event_attr *attr;
 265
 266         attr = zalloc(sizeof(struct perf_event_attr));
 267         if (!attr)
 268                 return NULL;
 269
 270         attr->config = intel_pt_default_config(intel_pt_pmu);
 271
 272         intel_pt_pmu->selectable = true;
 273
 274         return attr;
 275 }
 276
 277 static const char *intel_pt_find_filter(struct perf_evlist *evlist,
 278                                         struct perf_pmu *intel_pt_pmu)
 279 {
 280         struct perf_evsel *evsel;
 281
 282         evlist__for_each_entry(evlist, evsel) {
 283                 if (evsel->attr.type == intel_pt_pmu->type)
 284                         return evsel->filter;
 285         }
 286
 287         return NULL;
 288 }
 289
 290 static size_t intel_pt_filter_bytes(const char *filter)
 291 {
 292         size_t len = filter ? strlen(filter) : 0;
 293
 294         return len ? roundup(len + 1, 8) : 0;
 295 }
 296
 297 static size_t
 298 intel_pt_info_priv_size(struct auxtrace_record *itr, struct perf_evlist *evlist)
 299 {
 300         struct intel_pt_recording *ptr =
 301                         container_of(itr, struct intel_pt_recording, itr);
 302         const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
 303
 304         ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
 305                          intel_pt_filter_bytes(filter);
 306
 307         return ptr->priv_size;
 308 }
 309
 310 static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
 311 {
 312         unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
 313
 314         __get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
 315         *n = ebx;
 316         *d = eax;
 317 }
 318
 319 static int intel_pt_info_fill(struct auxtrace_record *itr,
 320                               struct perf_session *session,
 321                               struct auxtrace_info_event *auxtrace_info,
 322                               size_t priv_size)
 323 {
 324         struct intel_pt_recording *ptr =
 325                         container_of(itr, struct intel_pt_recording, itr);
 326         struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 327         struct perf_event_mmap_page *pc;
 328         struct perf_tsc_conversion tc = { .time_mult = 0, };
 329         bool cap_user_time_zero = false, per_cpu_mmaps;
 330         u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
 331         u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
 332         unsigned long max_non_turbo_ratio;
 333         size_t filter_str_len;
 334         const char *filter;
 335         u64 *info;
 336         int err;
 337
 338         if (priv_size != ptr->priv_size)
 339                 return -EINVAL;
 340
 341         intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
 342         intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
 343                              &noretcomp_bit);
 344         intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
 345         mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
 346                                               "mtc_period");
 347         intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
 348
 349         intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
 350
 351         if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
 352                                 "%lu", &max_non_turbo_ratio) != 1)
 353                 max_non_turbo_ratio = 0;
 354
 355         filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
 356         filter_str_len = filter ? strlen(filter) : 0;
 357
 358         if (!session->evlist->nr_mmaps)
 359                 return -EINVAL;
 360
 361         pc = session->evlist->mmap[0].base;
 362         if (pc) {
 363                 err = perf_read_tsc_conversion(pc, &tc);
 364                 if (err) {
 365                         if (err != -EOPNOTSUPP)
 366                                 return err;
 367                 } else {
 368                         cap_user_time_zero = tc.time_mult != 0;
 369                 }
 370                 if (!cap_user_time_zero)
 371                         ui__warning("Intel Processor Trace: TSC not available\n");
 372         }
 373
 374         per_cpu_mmaps = !cpu_map__empty(session->evlist->cpus);
 375
 376         auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
 377         auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
 378         auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
 379         auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
 380         auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
 381         auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
 382         auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
 383         auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
 384         auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
 385         auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
 386         auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
 387         auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
 388         auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
 389         auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
 390         auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
 391         auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
 392         auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
 393         auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
 394
 395         info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
 396
 397         if (filter_str_len) {
 398                 size_t len = intel_pt_filter_bytes(filter);
 399
 400                 strncpy((char *)info, filter, len);
 401                 info += len >> 3;
 402         }
 403
 404         return 0;
 405 }
 406
 407 static int intel_pt_track_switches(struct perf_evlist *evlist)
 408 {
 409         const char *sched_switch = "sched:sched_switch";
 410         struct perf_evsel *evsel;
 411         int err;
 412
 413         if (!perf_evlist__can_select_event(evlist, sched_switch))
 414                 return -EPERM;
 415
 416         err = parse_events(evlist, sched_switch, NULL);
 417         if (err) {
 418                 pr_debug2("%s: failed to parse %s, error %d\n",
 419                           __func__, sched_switch, err);
 420                 return err;
 421         }
 422
 423         evsel = perf_evlist__last(evlist);
 424
 425         perf_evsel__set_sample_bit(evsel, CPU);
 426         perf_evsel__set_sample_bit(evsel, TIME);
 427
 428         evsel->system_wide = true;
 429         evsel->no_aux_samples = true;
 430         evsel->immediate = true;
 431
 432         return 0;
 433 }
 434
 435 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
 436 {
 437         unsigned int val, last = 0, state = 1;
 438         int p = 0;
 439
 440         str[0] = '\0';
 441
 442         for (val = 0; val <= 64; val++, valid >>= 1) {
 443                 if (valid & 1) {
 444                         last = val;
 445                         switch (state) {
 446                         case 0:
 447                                 p += scnprintf(str + p, len - p, ",");
 448                                 /* Fall through */
 449                         case 1:
 450                                 p += scnprintf(str + p, len - p, "%u", val);
 451                                 state = 2;
 452                                 break;
 453                         case 2:
 454                                 state = 3;
 455                                 break;
 456                         case 3:
 457                                 state = 4;
 458                                 break;
 459                         default:
 460                                 break;
 461                         }
 462                 } else {
 463                         switch (state) {
 464                         case 3:
 465                                 p += scnprintf(str + p, len - p, ",%u", last);
 466                                 state = 0;
 467                                 break;
 468                         case 4:
 469                                 p += scnprintf(str + p, len - p, "-%u", last);
 470                                 state = 0;
 471                                 break;
 472                         default:
 473                                 break;
 474                         }
 475                         if (state != 1)
 476                                 state = 0;
 477                 }
 478         }
 479 }
 480
 481 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
 482                                     const char *caps, const char *name,
 483                                     const char *supported, u64 config)
 484 {
 485         char valid_str[256];
 486         unsigned int shift;
 487         unsigned long long valid;
 488         u64 bits;
 489         int ok;
 490
 491         if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
 492                 valid = 0;
 493
 494         if (supported &&
 495             perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
 496                 valid = 0;
 497
 498         valid |= 1;
 499
 500         bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
 501
 502         config &= bits;
 503
 504         for (shift = 0; bits && !(bits & 1); shift++)
 505                 bits >>= 1;
 506
 507         config >>= shift;
 508
 509         if (config > 63)
 510                 goto out_err;
 511
 512         if (valid & (1 << config))
 513                 return 0;
 514 out_err:
 515         intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
 516         pr_err("Invalid %s for %s. Valid values are: %s\n",
 517                name, INTEL_PT_PMU_NAME, valid_str);
 518         return -EINVAL;
 519 }
 520
 521 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
 522                                     struct perf_evsel *evsel)
 523 {
 524         int err;
 525         char c;
 526
 527         if (!evsel)
 528                 return 0;
 529
 530         /*
 531          * If supported, force pass-through config term (pt=1) even if user
 532          * sets pt=0, which avoids senseless kernel errors.
 533          */
 534         if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
 535             !(evsel->attr.config & 1)) {
 536                 pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
 537                 evsel->attr.config |= 1;
 538         }
 539
 540         err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
 541                                        "cyc_thresh", "caps/psb_cyc",
 542                                        evsel->attr.config);
 543         if (err)
 544                 return err;
 545
 546         err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
 547                                        "mtc_period", "caps/mtc",
 548                                        evsel->attr.config);
 549         if (err)
 550                 return err;
 551
 552         return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
 553                                         "psb_period", "caps/psb_cyc",
 554                                         evsel->attr.config);
 555 }
 556
 557 static int intel_pt_recording_options(struct auxtrace_record *itr,
 558                                       struct perf_evlist *evlist,
 559                                       struct record_opts *opts)
 560 {
 561         struct intel_pt_recording *ptr =
 562                         container_of(itr, struct intel_pt_recording, itr);
 563         struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 564         bool have_timing_info, need_immediate = false;
 565         struct perf_evsel *evsel, *intel_pt_evsel = NULL;
 566         const struct cpu_map *cpus = evlist->cpus;
 567         bool privileged = geteuid() == 0 || perf_event_paranoid() < 0;
 568         u64 tsc_bit;
 569         int err;
 570
 571         ptr->evlist = evlist;
 572         ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
 573
 574         evlist__for_each_entry(evlist, evsel) {
 575                 if (evsel->attr.type == intel_pt_pmu->type) {
 576                         if (intel_pt_evsel) {
 577                                 pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
 578                                 return -EINVAL;
 579                         }
 580                         evsel->attr.freq = 0;
 581                         evsel->attr.sample_period = 1;
 582                         intel_pt_evsel = evsel;
 583                         opts->full_auxtrace = true;
 584                 }
 585         }
 586
 587         if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
 588                 pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
 589                 return -EINVAL;
 590         }
 591
 592         if (opts->use_clockid) {
 593                 pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
 594                 return -EINVAL;
 595         }
 596
 597         if (!opts->full_auxtrace)
 598                 return 0;
 599
 600         err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
 601         if (err)
 602                 return err;
 603
 604         /* Set default sizes for snapshot mode */
 605         if (opts->auxtrace_snapshot_mode) {
 606                 size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 607
 608                 if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
 609                         if (privileged) {
 610                                 opts->auxtrace_mmap_pages = MiB(4) / page_size;
 611                         } else {
 612                                 opts->auxtrace_mmap_pages = KiB(128) / page_size;
 613                                 if (opts->mmap_pages == UINT_MAX)
 614                                         opts->mmap_pages = KiB(256) / page_size;
 615                         }
 616                 } else if (!opts->auxtrace_mmap_pages && !privileged &&
 617                            opts->mmap_pages == UINT_MAX) {
 618                         opts->mmap_pages = KiB(256) / page_size;
 619                 }
 620                 if (!opts->auxtrace_snapshot_size)
 621                         opts->auxtrace_snapshot_size =
 622                                 opts->auxtrace_mmap_pages * (size_t)page_size;
 623                 if (!opts->auxtrace_mmap_pages) {
 624                         size_t sz = opts->auxtrace_snapshot_size;
 625
 626                         sz = round_up(sz, page_size) / page_size;
 627                         opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 628                 }
 629                 if (opts->auxtrace_snapshot_size >
 630                                 opts->auxtrace_mmap_pages * (size_t)page_size) {
 631                         pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
 632                                opts->auxtrace_snapshot_size,
 633                                opts->auxtrace_mmap_pages * (size_t)page_size);
 634                         return -EINVAL;
 635                 }
 636                 if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
 637                         pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
 638                         return -EINVAL;
 639                 }
 640                 pr_debug2("Intel PT snapshot size: %zu\n",
 641                           opts->auxtrace_snapshot_size);
 642                 if (psb_period &&
 643                     opts->auxtrace_snapshot_size <= psb_period +
 644                                                   INTEL_PT_PSB_PERIOD_NEAR)
 645                         ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
 646                                     opts->auxtrace_snapshot_size, psb_period);
 647         }
 648
 649         /* Set default sizes for full trace mode */
 650         if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
 651                 if (privileged) {
 652                         opts->auxtrace_mmap_pages = MiB(4) / page_size;
 653                 } else {
 654                         opts->auxtrace_mmap_pages = KiB(128) / page_size;
 655                         if (opts->mmap_pages == UINT_MAX)
 656                                 opts->mmap_pages = KiB(256) / page_size;
 657                 }
 658         }
 659
 660         /* Validate auxtrace_mmap_pages */
 661         if (opts->auxtrace_mmap_pages) {
 662                 size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 663                 size_t min_sz;
 664
 665                 if (opts->auxtrace_snapshot_mode)
 666                         min_sz = KiB(4);
 667                 else
 668                         min_sz = KiB(8);
 669
 670                 if (sz < min_sz || !is_power_of_2(sz)) {
 671                         pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
 672                                min_sz / 1024);
 673                         return -EINVAL;
 674                 }
 675         }
 676
 677         intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
 678
 679         if (opts->full_auxtrace && (intel_pt_evsel->attr.config & tsc_bit))
 680                 have_timing_info = true;
 681         else
 682                 have_timing_info = false;
 683
 684         /*
 685          * Per-cpu recording needs sched_switch events to distinguish different
 686          * threads.
 687          */
 688         if (have_timing_info && !cpu_map__empty(cpus)) {
 689                 if (perf_can_record_switch_events()) {
 690                         bool cpu_wide = !target__none(&opts->target) &&
 691                                         !target__has_task(&opts->target);
 692
 693                         if (!cpu_wide && perf_can_record_cpu_wide()) {
 694                                 struct perf_evsel *switch_evsel;
 695
 696                                 err = parse_events(evlist, "dummy:u", NULL);
 697                                 if (err)
 698                                         return err;
 699
 700                                 switch_evsel = perf_evlist__last(evlist);
 701
 702                                 switch_evsel->attr.freq = 0;
 703                                 switch_evsel->attr.sample_period = 1;
 704                                 switch_evsel->attr.context_switch = 1;
 705
 706                                 switch_evsel->system_wide = true;
 707                                 switch_evsel->no_aux_samples = true;
 708                                 switch_evsel->immediate = true;
 709
 710                                 perf_evsel__set_sample_bit(switch_evsel, TID);
 711                                 perf_evsel__set_sample_bit(switch_evsel, TIME);
 712                                 perf_evsel__set_sample_bit(switch_evsel, CPU);
 713
 714                                 opts->record_switch_events = false;
 715                                 ptr->have_sched_switch = 3;
 716                         } else {
 717                                 opts->record_switch_events = true;
 718                                 need_immediate = true;
 719                                 if (cpu_wide)
 720                                         ptr->have_sched_switch = 3;
 721                                 else
 722                                         ptr->have_sched_switch = 2;
 723                         }
 724                 } else {
 725                         err = intel_pt_track_switches(evlist);
 726                         if (err == -EPERM)
 727                                 pr_debug2("Unable to select sched:sched_switch\n");
 728                         else if (err)
 729                                 return err;
 730                         else
 731                                 ptr->have_sched_switch = 1;
 732                 }
 733         }
 734
 735         if (intel_pt_evsel) {
 736                 /*
 737                  * To obtain the auxtrace buffer file descriptor, the auxtrace
 738                  * event must come first.
 739                  */
 740                 perf_evlist__to_front(evlist, intel_pt_evsel);
 741                 /*
 742                  * In the case of per-cpu mmaps, we need the CPU on the
 743                  * AUX event.
 744                  */
 745                 if (!cpu_map__empty(cpus))
 746                         perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
 747         }
 748
 749         /* Add dummy event to keep tracking */
 750         if (opts->full_auxtrace) {
 751                 struct perf_evsel *tracking_evsel;
 752
 753                 err = parse_events(evlist, "dummy:u", NULL);
 754                 if (err)
 755                         return err;
 756
 757                 tracking_evsel = perf_evlist__last(evlist);
 758
 759                 perf_evlist__set_tracking_event(evlist, tracking_evsel);
 760
 761                 tracking_evsel->attr.freq = 0;
 762                 tracking_evsel->attr.sample_period = 1;
 763
 764                 if (need_immediate)
 765                         tracking_evsel->immediate = true;
 766
 767                 /* In per-cpu case, always need the time of mmap events etc */
 768                 if (!cpu_map__empty(cpus)) {
 769                         perf_evsel__set_sample_bit(tracking_evsel, TIME);
 770                         /* And the CPU for switch events */
 771                         perf_evsel__set_sample_bit(tracking_evsel, CPU);
 772                 }
 773         }
 774
 775         /*
 776          * Warn the user when we do not have enough information to decode i.e.
 777          * per-cpu with no sched_switch (except workload-only).
 778          */
 779         if (!ptr->have_sched_switch && !cpu_map__empty(cpus) &&
 780             !target__none(&opts->target))
 781                 ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
 782
 783         return 0;
 784 }
 785
 786 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
 787 {
 788         struct intel_pt_recording *ptr =
 789                         container_of(itr, struct intel_pt_recording, itr);
 790         struct perf_evsel *evsel;
 791
 792         evlist__for_each_entry(ptr->evlist, evsel) {
 793                 if (evsel->attr.type == ptr->intel_pt_pmu->type)
 794                         return perf_evsel__disable(evsel);
 795         }
 796         return -EINVAL;
 797 }
 798
 799 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
 800 {
 801         struct intel_pt_recording *ptr =
 802                         container_of(itr, struct intel_pt_recording, itr);
 803         struct perf_evsel *evsel;
 804
 805         evlist__for_each_entry(ptr->evlist, evsel) {
 806                 if (evsel->attr.type == ptr->intel_pt_pmu->type)
 807                         return perf_evsel__enable(evsel);
 808         }
 809         return -EINVAL;
 810 }
 811
 812 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
 813 {
 814         const size_t sz = sizeof(struct intel_pt_snapshot_ref);
 815         int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
 816         struct intel_pt_snapshot_ref *refs;
 817
 818         if (!new_cnt)
 819                 new_cnt = 16;
 820
 821         while (new_cnt <= idx)
 822                 new_cnt *= 2;
 823
 824         refs = calloc(new_cnt, sz);
 825         if (!refs)
 826                 return -ENOMEM;
 827
 828         memcpy(refs, ptr->snapshot_refs, cnt * sz);
 829
 830         ptr->snapshot_refs = refs;
 831         ptr->snapshot_ref_cnt = new_cnt;
 832
 833         return 0;
 834 }
 835
 836 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
 837 {
 838         int i;
 839
 840         for (i = 0; i < ptr->snapshot_ref_cnt; i++)
 841                 zfree(&ptr->snapshot_refs[i].ref_buf);
 842         zfree(&ptr->snapshot_refs);
 843 }
 844
 845 static void intel_pt_recording_free(struct auxtrace_record *itr)
 846 {
 847         struct intel_pt_recording *ptr =
 848                         container_of(itr, struct intel_pt_recording, itr);
 849
 850         intel_pt_free_snapshot_refs(ptr);
 851         free(ptr);
 852 }
 853
 854 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
 855                                        size_t snapshot_buf_size)
 856 {
 857         size_t ref_buf_size = ptr->snapshot_ref_buf_size;
 858         void *ref_buf;
 859
 860         ref_buf = zalloc(ref_buf_size);
 861         if (!ref_buf)
 862                 return -ENOMEM;
 863
 864         ptr->snapshot_refs[idx].ref_buf = ref_buf;
 865         ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
 866
 867         return 0;
 868 }
 869
 870 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
 871                                              size_t snapshot_buf_size)
 872 {
 873         const size_t max_size = 256 * 1024;
 874         size_t buf_size = 0, psb_period;
 875
 876         if (ptr->snapshot_size <= 64 * 1024)
 877                 return 0;
 878
 879         psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
 880         if (psb_period)
 881                 buf_size = psb_period * 2;
 882
 883         if (!buf_size || buf_size > max_size)
 884                 buf_size = max_size;
 885
 886         if (buf_size >= snapshot_buf_size)
 887                 return 0;
 888
 889         if (buf_size >= ptr->snapshot_size / 2)
 890                 return 0;
 891
 892         return buf_size;
 893 }
 894
 895 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
 896                                   size_t snapshot_buf_size)
 897 {
 898         if (ptr->snapshot_init_done)
 899                 return 0;
 900
 901         ptr->snapshot_init_done = true;
 902
 903         ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
 904                                                         snapshot_buf_size);
 905
 906         return 0;
 907 }
 908
 909 /**
 910  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
 911  * @buf1: first buffer
 912  * @compare_size: number of bytes to compare
 913  * @buf2: second buffer (a circular buffer)
 914  * @offs2: offset in second buffer
 915  * @buf2_size: size of second buffer
 916  *
 917  * The comparison allows for the possibility that the bytes to compare in the
 918  * circular buffer are not contiguous.  It is assumed that @compare_size <=
 919  * @buf2_size.  This function returns %false if the bytes are identical, %true
 920  * otherwise.
 921  */
 922 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
 923                                      void *buf2, size_t offs2, size_t buf2_size)
 924 {
 925         size_t end2 = offs2 + compare_size, part_size;
 926
 927         if (end2 <= buf2_size)
 928                 return memcmp(buf1, buf2 + offs2, compare_size);
 929
 930         part_size = end2 - buf2_size;
 931         if (memcmp(buf1, buf2 + offs2, part_size))
 932                 return true;
 933
 934         compare_size -= part_size;
 935
 936         return memcmp(buf1 + part_size, buf2, compare_size);
 937 }
 938
 939 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
 940                                  size_t ref_size, size_t buf_size,
 941                                  void *data, size_t head)
 942 {
 943         size_t ref_end = ref_offset + ref_size;
 944
 945         if (ref_end > buf_size) {
 946                 if (head > ref_offset || head < ref_end - buf_size)
 947                         return true;
 948         } else if (head > ref_offset && head < ref_end) {
 949                 return true;
 950         }
 951
 952         return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
 953                                         buf_size);
 954 }
 955
 956 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
 957                               void *data, size_t head)
 958 {
 959         if (head >= ref_size) {
 960                 memcpy(ref_buf, data + head - ref_size, ref_size);
 961         } else {
 962                 memcpy(ref_buf, data, head);
 963                 ref_size -= head;
 964                 memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
 965         }
 966 }
 967
 968 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
 969                              struct auxtrace_mmap *mm, unsigned char *data,
 970                              u64 head)
 971 {
 972         struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
 973         bool wrapped;
 974
 975         wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
 976                                        ptr->snapshot_ref_buf_size, mm->len,
 977                                        data, head);
 978
 979         intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
 980                           data, head);
 981
 982         return wrapped;
 983 }
 984
 985 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
 986 {
 987         int i, a, b;
 988
 989         b = buf_size >> 3;
 990         a = b - 512;
 991         if (a < 0)
 992                 a = 0;
 993
 994         for (i = a; i < b; i++) {
 995                 if (data[i])
 996                         return true;
 997         }
 998
 999         return false;
1000 }
1001
1002 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1003                                   struct auxtrace_mmap *mm, unsigned char *data,
1004                                   u64 *head, u64 *old)
1005 {
1006         struct intel_pt_recording *ptr =
1007                         container_of(itr, struct intel_pt_recording, itr);
1008         bool wrapped;
1009         int err;
1010
1011         pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1012                   __func__, idx, (size_t)*old, (size_t)*head);
1013
1014         err = intel_pt_snapshot_init(ptr, mm->len);
1015         if (err)
1016                 goto out_err;
1017
1018         if (idx >= ptr->snapshot_ref_cnt) {
1019                 err = intel_pt_alloc_snapshot_refs(ptr, idx);
1020                 if (err)
1021                         goto out_err;
1022         }
1023
1024         if (ptr->snapshot_ref_buf_size) {
1025                 if (!ptr->snapshot_refs[idx].ref_buf) {
1026                         err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1027                         if (err)
1028                                 goto out_err;
1029                 }
1030                 wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1031         } else {
1032                 wrapped = ptr->snapshot_refs[idx].wrapped;
1033                 if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1034                         ptr->snapshot_refs[idx].wrapped = true;
1035                         wrapped = true;
1036                 }
1037         }
1038
1039         /*
1040          * In full trace mode 'head' continually increases.  However in snapshot
1041          * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1042          * are adjusted to match the full trace case which expects that 'old' is
1043          * always less than 'head'.
1044          */
1045         if (wrapped) {
1046                 *old = *head;
1047                 *head += mm->len;
1048         } else {
1049                 if (mm->mask)
1050                         *old &= mm->mask;
1051                 else
1052                         *old %= mm->len;
1053                 if (*old > *head)
1054                         *head += mm->len;
1055         }
1056
1057         pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1058                   __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1059
1060         return 0;
1061
1062 out_err:
1063         pr_err("%s: failed, error %d\n", __func__, err);
1064         return err;
1065 }
1066
1067 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1068 {
1069         return rdtsc();
1070 }
1071
1072 static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1073 {
1074         struct intel_pt_recording *ptr =
1075                         container_of(itr, struct intel_pt_recording, itr);
1076         struct perf_evsel *evsel;
1077
1078         evlist__for_each_entry(ptr->evlist, evsel) {
1079                 if (evsel->attr.type == ptr->intel_pt_pmu->type)
1080                         return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1081                                                              idx);
1082         }
1083         return -EINVAL;
1084 }
1085
1086 struct auxtrace_record *intel_pt_recording_init(int *err)
1087 {
1088         struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1089         struct intel_pt_recording *ptr;
1090
1091         if (!intel_pt_pmu)
1092                 return NULL;
1093
1094         if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1095                 *err = -errno;
1096                 return NULL;
1097         }
1098
1099         ptr = zalloc(sizeof(struct intel_pt_recording));
1100         if (!ptr) {
1101                 *err = -ENOMEM;
1102                 return NULL;
1103         }
1104
1105         ptr->intel_pt_pmu = intel_pt_pmu;
1106         ptr->itr.recording_options = intel_pt_recording_options;
1107         ptr->itr.info_priv_size = intel_pt_info_priv_size;
1108         ptr->itr.info_fill = intel_pt_info_fill;
1109         ptr->itr.free = intel_pt_recording_free;
1110         ptr->itr.snapshot_start = intel_pt_snapshot_start;
1111         ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1112         ptr->itr.find_snapshot = intel_pt_find_snapshot;
1113         ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1114         ptr->itr.reference = intel_pt_reference;
1115         ptr->itr.read_finish = intel_pt_read_finish;
1116         return &ptr->itr;
1117 }