tools/perf/arch/x86/util/intel-pt.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * intel_pt.c: Intel Processor Trace support
   4  * Copyright (c) 2013-2015, Intel Corporation.
   5  */
   6
   7 #include <errno.h>
   8 #include <stdbool.h>
   9 #include <linux/kernel.h>
  10 #include <linux/types.h>
  11 #include <linux/bitops.h>
  12 #include <linux/log2.h>
  13 #include <linux/zalloc.h>
  14 #include <cpuid.h>
  15
  16 #include "../../perf.h"
  17 #include "../../util/session.h"
  18 #include "../../util/event.h"
  19 #include "../../util/evlist.h"
  20 #include "../../util/evsel.h"
  21 #include "../../util/cpumap.h"
  22 #include <subcmd/parse-options.h>
  23 #include "../../util/parse-events.h"
  24 #include "../../util/pmu.h"
  25 #include "../../util/debug.h"
  26 #include "../../util/auxtrace.h"
  27 #include "../../util/tsc.h"
  28 #include "../../util/intel-pt.h"
  29
  30 #define KiB(x) ((x) * 1024)
  31 #define MiB(x) ((x) * 1024 * 1024)
  32 #define KiB_MASK(x) (KiB(x) - 1)
  33 #define MiB_MASK(x) (MiB(x) - 1)
  34
  35 #define INTEL_PT_PSB_PERIOD_NEAR        256
  36
  37 struct intel_pt_snapshot_ref {
  38         void *ref_buf;
  39         size_t ref_offset;
  40         bool wrapped;
  41 };
  42
  43 struct intel_pt_recording {
  44         struct auxtrace_record          itr;
  45         struct perf_pmu                 *intel_pt_pmu;
  46         int                             have_sched_switch;
  47         struct evlist           *evlist;
  48         bool                            snapshot_mode;
  49         bool                            snapshot_init_done;
  50         size_t                          snapshot_size;
  51         size_t                          snapshot_ref_buf_size;
  52         int                             snapshot_ref_cnt;
  53         struct intel_pt_snapshot_ref    *snapshot_refs;
  54         size_t                          priv_size;
  55 };
  56
  57 static int intel_pt_parse_terms_with_default(struct list_head *formats,
  58                                              const char *str,
  59                                              u64 *config)
  60 {
  61         struct list_head *terms;
  62         struct perf_event_attr attr = { .size = 0, };
  63         int err;
  64
  65         terms = malloc(sizeof(struct list_head));
  66         if (!terms)
  67                 return -ENOMEM;
  68
  69         INIT_LIST_HEAD(terms);
  70
  71         err = parse_events_terms(terms, str);
  72         if (err)
  73                 goto out_free;
  74
  75         attr.config = *config;
  76         err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
  77         if (err)
  78                 goto out_free;
  79
  80         *config = attr.config;
  81 out_free:
  82         parse_events_terms__delete(terms);
  83         return err;
  84 }
  85
  86 static int intel_pt_parse_terms(struct list_head *formats, const char *str,
  87                                 u64 *config)
  88 {
  89         *config = 0;
  90         return intel_pt_parse_terms_with_default(formats, str, config);
  91 }
  92
  93 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
  94 {
  95         const u64 top_bit = 1ULL << 63;
  96         u64 res = 0;
  97         int i;
  98
  99         for (i = 0; i < 64; i++) {
 100                 if (mask & top_bit) {
 101                         res <<= 1;
 102                         if (bits & top_bit)
 103                                 res |= 1;
 104                 }
 105                 mask <<= 1;
 106                 bits <<= 1;
 107         }
 108
 109         return res;
 110 }
 111
 112 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
 113                                 struct evlist *evlist, u64 *res)
 114 {
 115         struct evsel *evsel;
 116         u64 mask;
 117
 118         *res = 0;
 119
 120         mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
 121         if (!mask)
 122                 return -EINVAL;
 123
 124         evlist__for_each_entry(evlist, evsel) {
 125                 if (evsel->core.attr.type == intel_pt_pmu->type) {
 126                         *res = intel_pt_masked_bits(mask, evsel->core.attr.config);
 127                         return 0;
 128                 }
 129         }
 130
 131         return -EINVAL;
 132 }
 133
 134 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
 135                                   struct evlist *evlist)
 136 {
 137         u64 val;
 138         int err, topa_multiple_entries;
 139         size_t psb_period;
 140
 141         if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
 142                                 "%d", &topa_multiple_entries) != 1)
 143                 topa_multiple_entries = 0;
 144
 145         /*
 146          * Use caps/topa_multiple_entries to indicate early hardware that had
 147          * extra frequent PSBs.
 148          */
 149         if (!topa_multiple_entries) {
 150                 psb_period = 256;
 151                 goto out;
 152         }
 153
 154         err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
 155         if (err)
 156                 val = 0;
 157
 158         psb_period = 1 << (val + 11);
 159 out:
 160         pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
 161         return psb_period;
 162 }
 163
 164 static int intel_pt_pick_bit(int bits, int target)
 165 {
 166         int pos, pick = -1;
 167
 168         for (pos = 0; bits; bits >>= 1, pos++) {
 169                 if (bits & 1) {
 170                         if (pos <= target || pick < 0)
 171                                 pick = pos;
 172                         if (pos >= target)
 173                                 break;
 174                 }
 175         }
 176
 177         return pick;
 178 }
 179
 180 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
 181 {
 182         char buf[256];
 183         int mtc, mtc_periods = 0, mtc_period;
 184         int psb_cyc, psb_periods, psb_period;
 185         int pos = 0;
 186         u64 config;
 187         char c;
 188
 189         pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
 190
 191         if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
 192                                 &mtc) != 1)
 193                 mtc = 1;
 194
 195         if (mtc) {
 196                 if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
 197                                         &mtc_periods) != 1)
 198                         mtc_periods = 0;
 199                 if (mtc_periods) {
 200                         mtc_period = intel_pt_pick_bit(mtc_periods, 3);
 201                         pos += scnprintf(buf + pos, sizeof(buf) - pos,
 202                                          ",mtc,mtc_period=%d", mtc_period);
 203                 }
 204         }
 205
 206         if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
 207                                 &psb_cyc) != 1)
 208                 psb_cyc = 1;
 209
 210         if (psb_cyc && mtc_periods) {
 211                 if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
 212                                         &psb_periods) != 1)
 213                         psb_periods = 0;
 214                 if (psb_periods) {
 215                         psb_period = intel_pt_pick_bit(psb_periods, 3);
 216                         pos += scnprintf(buf + pos, sizeof(buf) - pos,
 217                                          ",psb_period=%d", psb_period);
 218                 }
 219         }
 220
 221         if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
 222             perf_pmu__scan_file(intel_pt_pmu, "format/branch", "%c", &c) == 1)
 223                 pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
 224
 225         pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
 226
 227         intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
 228
 229         return config;
 230 }
 231
 232 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
 233                                            struct record_opts *opts,
 234                                            const char *str)
 235 {
 236         struct intel_pt_recording *ptr =
 237                         container_of(itr, struct intel_pt_recording, itr);
 238         unsigned long long snapshot_size = 0;
 239         char *endptr;
 240
 241         if (str) {
 242                 snapshot_size = strtoull(str, &endptr, 0);
 243                 if (*endptr || snapshot_size > SIZE_MAX)
 244                         return -1;
 245         }
 246
 247         opts->auxtrace_snapshot_mode = true;
 248         opts->auxtrace_snapshot_size = snapshot_size;
 249
 250         ptr->snapshot_size = snapshot_size;
 251
 252         return 0;
 253 }
 254
 255 struct perf_event_attr *
 256 intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
 257 {
 258         struct perf_event_attr *attr;
 259
 260         attr = zalloc(sizeof(struct perf_event_attr));
 261         if (!attr)
 262                 return NULL;
 263
 264         attr->config = intel_pt_default_config(intel_pt_pmu);
 265
 266         intel_pt_pmu->selectable = true;
 267
 268         return attr;
 269 }
 270
 271 static const char *intel_pt_find_filter(struct evlist *evlist,
 272                                         struct perf_pmu *intel_pt_pmu)
 273 {
 274         struct evsel *evsel;
 275
 276         evlist__for_each_entry(evlist, evsel) {
 277                 if (evsel->core.attr.type == intel_pt_pmu->type)
 278                         return evsel->filter;
 279         }
 280
 281         return NULL;
 282 }
 283
 284 static size_t intel_pt_filter_bytes(const char *filter)
 285 {
 286         size_t len = filter ? strlen(filter) : 0;
 287
 288         return len ? roundup(len + 1, 8) : 0;
 289 }
 290
 291 static size_t
 292 intel_pt_info_priv_size(struct auxtrace_record *itr, struct evlist *evlist)
 293 {
 294         struct intel_pt_recording *ptr =
 295                         container_of(itr, struct intel_pt_recording, itr);
 296         const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
 297
 298         ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
 299                          intel_pt_filter_bytes(filter);
 300
 301         return ptr->priv_size;
 302 }
 303
 304 static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
 305 {
 306         unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
 307
 308         __get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
 309         *n = ebx;
 310         *d = eax;
 311 }
 312
 313 static int intel_pt_info_fill(struct auxtrace_record *itr,
 314                               struct perf_session *session,
 315                               struct auxtrace_info_event *auxtrace_info,
 316                               size_t priv_size)
 317 {
 318         struct intel_pt_recording *ptr =
 319                         container_of(itr, struct intel_pt_recording, itr);
 320         struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 321         struct perf_event_mmap_page *pc;
 322         struct perf_tsc_conversion tc = { .time_mult = 0, };
 323         bool cap_user_time_zero = false, per_cpu_mmaps;
 324         u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
 325         u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
 326         unsigned long max_non_turbo_ratio;
 327         size_t filter_str_len;
 328         const char *filter;
 329         u64 *info;
 330         int err;
 331
 332         if (priv_size != ptr->priv_size)
 333                 return -EINVAL;
 334
 335         intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
 336         intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
 337                              &noretcomp_bit);
 338         intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
 339         mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
 340                                               "mtc_period");
 341         intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
 342
 343         intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
 344
 345         if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
 346                                 "%lu", &max_non_turbo_ratio) != 1)
 347                 max_non_turbo_ratio = 0;
 348
 349         filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
 350         filter_str_len = filter ? strlen(filter) : 0;
 351
 352         if (!session->evlist->nr_mmaps)
 353                 return -EINVAL;
 354
 355         pc = session->evlist->mmap[0].base;
 356         if (pc) {
 357                 err = perf_read_tsc_conversion(pc, &tc);
 358                 if (err) {
 359                         if (err != -EOPNOTSUPP)
 360                                 return err;
 361                 } else {
 362                         cap_user_time_zero = tc.time_mult != 0;
 363                 }
 364                 if (!cap_user_time_zero)
 365                         ui__warning("Intel Processor Trace: TSC not available\n");
 366         }
 367
 368         per_cpu_mmaps = !cpu_map__empty(session->evlist->core.cpus);
 369
 370         auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
 371         auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
 372         auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
 373         auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
 374         auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
 375         auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
 376         auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
 377         auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
 378         auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
 379         auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
 380         auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
 381         auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
 382         auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
 383         auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
 384         auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
 385         auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
 386         auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
 387         auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
 388
 389         info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
 390
 391         if (filter_str_len) {
 392                 size_t len = intel_pt_filter_bytes(filter);
 393
 394                 strncpy((char *)info, filter, len);
 395                 info += len >> 3;
 396         }
 397
 398         return 0;
 399 }
 400
 401 static int intel_pt_track_switches(struct evlist *evlist)
 402 {
 403         const char *sched_switch = "sched:sched_switch";
 404         struct evsel *evsel;
 405         int err;
 406
 407         if (!perf_evlist__can_select_event(evlist, sched_switch))
 408                 return -EPERM;
 409
 410         err = parse_events(evlist, sched_switch, NULL);
 411         if (err) {
 412                 pr_debug2("%s: failed to parse %s, error %d\n",
 413                           __func__, sched_switch, err);
 414                 return err;
 415         }
 416
 417         evsel = perf_evlist__last(evlist);
 418
 419         perf_evsel__set_sample_bit(evsel, CPU);
 420         perf_evsel__set_sample_bit(evsel, TIME);
 421
 422         evsel->system_wide = true;
 423         evsel->no_aux_samples = true;
 424         evsel->immediate = true;
 425
 426         return 0;
 427 }
 428
 429 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
 430 {
 431         unsigned int val, last = 0, state = 1;
 432         int p = 0;
 433
 434         str[0] = '\0';
 435
 436         for (val = 0; val <= 64; val++, valid >>= 1) {
 437                 if (valid & 1) {
 438                         last = val;
 439                         switch (state) {
 440                         case 0:
 441                                 p += scnprintf(str + p, len - p, ",");
 442                                 /* Fall through */
 443                         case 1:
 444                                 p += scnprintf(str + p, len - p, "%u", val);
 445                                 state = 2;
 446                                 break;
 447                         case 2:
 448                                 state = 3;
 449                                 break;
 450                         case 3:
 451                                 state = 4;
 452                                 break;
 453                         default:
 454                                 break;
 455                         }
 456                 } else {
 457                         switch (state) {
 458                         case 3:
 459                                 p += scnprintf(str + p, len - p, ",%u", last);
 460                                 state = 0;
 461                                 break;
 462                         case 4:
 463                                 p += scnprintf(str + p, len - p, "-%u", last);
 464                                 state = 0;
 465                                 break;
 466                         default:
 467                                 break;
 468                         }
 469                         if (state != 1)
 470                                 state = 0;
 471                 }
 472         }
 473 }
 474
 475 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
 476                                     const char *caps, const char *name,
 477                                     const char *supported, u64 config)
 478 {
 479         char valid_str[256];
 480         unsigned int shift;
 481         unsigned long long valid;
 482         u64 bits;
 483         int ok;
 484
 485         if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
 486                 valid = 0;
 487
 488         if (supported &&
 489             perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
 490                 valid = 0;
 491
 492         valid |= 1;
 493
 494         bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
 495
 496         config &= bits;
 497
 498         for (shift = 0; bits && !(bits & 1); shift++)
 499                 bits >>= 1;
 500
 501         config >>= shift;
 502
 503         if (config > 63)
 504                 goto out_err;
 505
 506         if (valid & (1 << config))
 507                 return 0;
 508 out_err:
 509         intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
 510         pr_err("Invalid %s for %s. Valid values are: %s\n",
 511                name, INTEL_PT_PMU_NAME, valid_str);
 512         return -EINVAL;
 513 }
 514
 515 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
 516                                     struct evsel *evsel)
 517 {
 518         int err;
 519         char c;
 520
 521         if (!evsel)
 522                 return 0;
 523
 524         /*
 525          * If supported, force pass-through config term (pt=1) even if user
 526          * sets pt=0, which avoids senseless kernel errors.
 527          */
 528         if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
 529             !(evsel->core.attr.config & 1)) {
 530                 pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
 531                 evsel->core.attr.config |= 1;
 532         }
 533
 534         err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
 535                                        "cyc_thresh", "caps/psb_cyc",
 536                                        evsel->core.attr.config);
 537         if (err)
 538                 return err;
 539
 540         err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
 541                                        "mtc_period", "caps/mtc",
 542                                        evsel->core.attr.config);
 543         if (err)
 544                 return err;
 545
 546         return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
 547                                         "psb_period", "caps/psb_cyc",
 548                                         evsel->core.attr.config);
 549 }
 550
 551 /*
 552  * Currently, there is not enough information to disambiguate different PEBS
 553  * events, so only allow one.
 554  */
 555 static bool intel_pt_too_many_aux_output(struct evlist *evlist)
 556 {
 557         struct evsel *evsel;
 558         int aux_output_cnt = 0;
 559
 560         evlist__for_each_entry(evlist, evsel)
 561                 aux_output_cnt += !!evsel->core.attr.aux_output;
 562
 563         if (aux_output_cnt > 1) {
 564                 pr_err(INTEL_PT_PMU_NAME " supports at most one event with aux-output\n");
 565                 return true;
 566         }
 567
 568         return false;
 569 }
 570
 571 static int intel_pt_recording_options(struct auxtrace_record *itr,
 572                                       struct evlist *evlist,
 573                                       struct record_opts *opts)
 574 {
 575         struct intel_pt_recording *ptr =
 576                         container_of(itr, struct intel_pt_recording, itr);
 577         struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 578         bool have_timing_info, need_immediate = false;
 579         struct evsel *evsel, *intel_pt_evsel = NULL;
 580         const struct perf_cpu_map *cpus = evlist->core.cpus;
 581         bool privileged = geteuid() == 0 || perf_event_paranoid() < 0;
 582         u64 tsc_bit;
 583         int err;
 584
 585         ptr->evlist = evlist;
 586         ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
 587
 588         evlist__for_each_entry(evlist, evsel) {
 589                 if (evsel->core.attr.type == intel_pt_pmu->type) {
 590                         if (intel_pt_evsel) {
 591                                 pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
 592                                 return -EINVAL;
 593                         }
 594                         evsel->core.attr.freq = 0;
 595                         evsel->core.attr.sample_period = 1;
 596                         intel_pt_evsel = evsel;
 597                         opts->full_auxtrace = true;
 598                 }
 599         }
 600
 601         if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
 602                 pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
 603                 return -EINVAL;
 604         }
 605
 606         if (opts->use_clockid) {
 607                 pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
 608                 return -EINVAL;
 609         }
 610
 611         if (intel_pt_too_many_aux_output(evlist))
 612                 return -EINVAL;
 613
 614         if (!opts->full_auxtrace)
 615                 return 0;
 616
 617         err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
 618         if (err)
 619                 return err;
 620
 621         /* Set default sizes for snapshot mode */
 622         if (opts->auxtrace_snapshot_mode) {
 623                 size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 624
 625                 if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
 626                         if (privileged) {
 627                                 opts->auxtrace_mmap_pages = MiB(4) / page_size;
 628                         } else {
 629                                 opts->auxtrace_mmap_pages = KiB(128) / page_size;
 630                                 if (opts->mmap_pages == UINT_MAX)
 631                                         opts->mmap_pages = KiB(256) / page_size;
 632                         }
 633                 } else if (!opts->auxtrace_mmap_pages && !privileged &&
 634                            opts->mmap_pages == UINT_MAX) {
 635                         opts->mmap_pages = KiB(256) / page_size;
 636                 }
 637                 if (!opts->auxtrace_snapshot_size)
 638                         opts->auxtrace_snapshot_size =
 639                                 opts->auxtrace_mmap_pages * (size_t)page_size;
 640                 if (!opts->auxtrace_mmap_pages) {
 641                         size_t sz = opts->auxtrace_snapshot_size;
 642
 643                         sz = round_up(sz, page_size) / page_size;
 644                         opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 645                 }
 646                 if (opts->auxtrace_snapshot_size >
 647                                 opts->auxtrace_mmap_pages * (size_t)page_size) {
 648                         pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
 649                                opts->auxtrace_snapshot_size,
 650                                opts->auxtrace_mmap_pages * (size_t)page_size);
 651                         return -EINVAL;
 652                 }
 653                 if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
 654                         pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
 655                         return -EINVAL;
 656                 }
 657                 pr_debug2("Intel PT snapshot size: %zu\n",
 658                           opts->auxtrace_snapshot_size);
 659                 if (psb_period &&
 660                     opts->auxtrace_snapshot_size <= psb_period +
 661                                                   INTEL_PT_PSB_PERIOD_NEAR)
 662                         ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
 663                                     opts->auxtrace_snapshot_size, psb_period);
 664         }
 665
 666         /* Set default sizes for full trace mode */
 667         if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
 668                 if (privileged) {
 669                         opts->auxtrace_mmap_pages = MiB(4) / page_size;
 670                 } else {
 671                         opts->auxtrace_mmap_pages = KiB(128) / page_size;
 672                         if (opts->mmap_pages == UINT_MAX)
 673                                 opts->mmap_pages = KiB(256) / page_size;
 674                 }
 675         }
 676
 677         /* Validate auxtrace_mmap_pages */
 678         if (opts->auxtrace_mmap_pages) {
 679                 size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 680                 size_t min_sz;
 681
 682                 if (opts->auxtrace_snapshot_mode)
 683                         min_sz = KiB(4);
 684                 else
 685                         min_sz = KiB(8);
 686
 687                 if (sz < min_sz || !is_power_of_2(sz)) {
 688                         pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
 689                                min_sz / 1024);
 690                         return -EINVAL;
 691                 }
 692         }
 693
 694         intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
 695
 696         if (opts->full_auxtrace && (intel_pt_evsel->core.attr.config & tsc_bit))
 697                 have_timing_info = true;
 698         else
 699                 have_timing_info = false;
 700
 701         /*
 702          * Per-cpu recording needs sched_switch events to distinguish different
 703          * threads.
 704          */
 705         if (have_timing_info && !cpu_map__empty(cpus)) {
 706                 if (perf_can_record_switch_events()) {
 707                         bool cpu_wide = !target__none(&opts->target) &&
 708                                         !target__has_task(&opts->target);
 709
 710                         if (!cpu_wide && perf_can_record_cpu_wide()) {
 711                                 struct evsel *switch_evsel;
 712
 713                                 err = parse_events(evlist, "dummy:u", NULL);
 714                                 if (err)
 715                                         return err;
 716
 717                                 switch_evsel = perf_evlist__last(evlist);
 718
 719                                 switch_evsel->core.attr.freq = 0;
 720                                 switch_evsel->core.attr.sample_period = 1;
 721                                 switch_evsel->core.attr.context_switch = 1;
 722
 723                                 switch_evsel->system_wide = true;
 724                                 switch_evsel->no_aux_samples = true;
 725                                 switch_evsel->immediate = true;
 726
 727                                 perf_evsel__set_sample_bit(switch_evsel, TID);
 728                                 perf_evsel__set_sample_bit(switch_evsel, TIME);
 729                                 perf_evsel__set_sample_bit(switch_evsel, CPU);
 730                                 perf_evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
 731
 732                                 opts->record_switch_events = false;
 733                                 ptr->have_sched_switch = 3;
 734                         } else {
 735                                 opts->record_switch_events = true;
 736                                 need_immediate = true;
 737                                 if (cpu_wide)
 738                                         ptr->have_sched_switch = 3;
 739                                 else
 740                                         ptr->have_sched_switch = 2;
 741                         }
 742                 } else {
 743                         err = intel_pt_track_switches(evlist);
 744                         if (err == -EPERM)
 745                                 pr_debug2("Unable to select sched:sched_switch\n");
 746                         else if (err)
 747                                 return err;
 748                         else
 749                                 ptr->have_sched_switch = 1;
 750                 }
 751         }
 752
 753         if (intel_pt_evsel) {
 754                 /*
 755                  * To obtain the auxtrace buffer file descriptor, the auxtrace
 756                  * event must come first.
 757                  */
 758                 perf_evlist__to_front(evlist, intel_pt_evsel);
 759                 /*
 760                  * In the case of per-cpu mmaps, we need the CPU on the
 761                  * AUX event.
 762                  */
 763                 if (!cpu_map__empty(cpus))
 764                         perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
 765         }
 766
 767         /* Add dummy event to keep tracking */
 768         if (opts->full_auxtrace) {
 769                 struct evsel *tracking_evsel;
 770
 771                 err = parse_events(evlist, "dummy:u", NULL);
 772                 if (err)
 773                         return err;
 774
 775                 tracking_evsel = perf_evlist__last(evlist);
 776
 777                 perf_evlist__set_tracking_event(evlist, tracking_evsel);
 778
 779                 tracking_evsel->core.attr.freq = 0;
 780                 tracking_evsel->core.attr.sample_period = 1;
 781
 782                 tracking_evsel->no_aux_samples = true;
 783                 if (need_immediate)
 784                         tracking_evsel->immediate = true;
 785
 786                 /* In per-cpu case, always need the time of mmap events etc */
 787                 if (!cpu_map__empty(cpus)) {
 788                         perf_evsel__set_sample_bit(tracking_evsel, TIME);
 789                         /* And the CPU for switch events */
 790                         perf_evsel__set_sample_bit(tracking_evsel, CPU);
 791                 }
 792                 perf_evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
 793         }
 794
 795         /*
 796          * Warn the user when we do not have enough information to decode i.e.
 797          * per-cpu with no sched_switch (except workload-only).
 798          */
 799         if (!ptr->have_sched_switch && !cpu_map__empty(cpus) &&
 800             !target__none(&opts->target))
 801                 ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
 802
 803         return 0;
 804 }
 805
 806 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
 807 {
 808         struct intel_pt_recording *ptr =
 809                         container_of(itr, struct intel_pt_recording, itr);
 810         struct evsel *evsel;
 811
 812         evlist__for_each_entry(ptr->evlist, evsel) {
 813                 if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
 814                         return evsel__disable(evsel);
 815         }
 816         return -EINVAL;
 817 }
 818
 819 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
 820 {
 821         struct intel_pt_recording *ptr =
 822                         container_of(itr, struct intel_pt_recording, itr);
 823         struct evsel *evsel;
 824
 825         evlist__for_each_entry(ptr->evlist, evsel) {
 826                 if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
 827                         return evsel__enable(evsel);
 828         }
 829         return -EINVAL;
 830 }
 831
 832 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
 833 {
 834         const size_t sz = sizeof(struct intel_pt_snapshot_ref);
 835         int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
 836         struct intel_pt_snapshot_ref *refs;
 837
 838         if (!new_cnt)
 839                 new_cnt = 16;
 840
 841         while (new_cnt <= idx)
 842                 new_cnt *= 2;
 843
 844         refs = calloc(new_cnt, sz);
 845         if (!refs)
 846                 return -ENOMEM;
 847
 848         memcpy(refs, ptr->snapshot_refs, cnt * sz);
 849
 850         ptr->snapshot_refs = refs;
 851         ptr->snapshot_ref_cnt = new_cnt;
 852
 853         return 0;
 854 }
 855
 856 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
 857 {
 858         int i;
 859
 860         for (i = 0; i < ptr->snapshot_ref_cnt; i++)
 861                 zfree(&ptr->snapshot_refs[i].ref_buf);
 862         zfree(&ptr->snapshot_refs);
 863 }
 864
 865 static void intel_pt_recording_free(struct auxtrace_record *itr)
 866 {
 867         struct intel_pt_recording *ptr =
 868                         container_of(itr, struct intel_pt_recording, itr);
 869
 870         intel_pt_free_snapshot_refs(ptr);
 871         free(ptr);
 872 }
 873
 874 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
 875                                        size_t snapshot_buf_size)
 876 {
 877         size_t ref_buf_size = ptr->snapshot_ref_buf_size;
 878         void *ref_buf;
 879
 880         ref_buf = zalloc(ref_buf_size);
 881         if (!ref_buf)
 882                 return -ENOMEM;
 883
 884         ptr->snapshot_refs[idx].ref_buf = ref_buf;
 885         ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
 886
 887         return 0;
 888 }
 889
 890 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
 891                                              size_t snapshot_buf_size)
 892 {
 893         const size_t max_size = 256 * 1024;
 894         size_t buf_size = 0, psb_period;
 895
 896         if (ptr->snapshot_size <= 64 * 1024)
 897                 return 0;
 898
 899         psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
 900         if (psb_period)
 901                 buf_size = psb_period * 2;
 902
 903         if (!buf_size || buf_size > max_size)
 904                 buf_size = max_size;
 905
 906         if (buf_size >= snapshot_buf_size)
 907                 return 0;
 908
 909         if (buf_size >= ptr->snapshot_size / 2)
 910                 return 0;
 911
 912         return buf_size;
 913 }
 914
 915 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
 916                                   size_t snapshot_buf_size)
 917 {
 918         if (ptr->snapshot_init_done)
 919                 return 0;
 920
 921         ptr->snapshot_init_done = true;
 922
 923         ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
 924                                                         snapshot_buf_size);
 925
 926         return 0;
 927 }
 928
 929 /**
 930  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
 931  * @buf1: first buffer
 932  * @compare_size: number of bytes to compare
 933  * @buf2: second buffer (a circular buffer)
 934  * @offs2: offset in second buffer
 935  * @buf2_size: size of second buffer
 936  *
 937  * The comparison allows for the possibility that the bytes to compare in the
 938  * circular buffer are not contiguous.  It is assumed that @compare_size <=
 939  * @buf2_size.  This function returns %false if the bytes are identical, %true
 940  * otherwise.
 941  */
 942 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
 943                                      void *buf2, size_t offs2, size_t buf2_size)
 944 {
 945         size_t end2 = offs2 + compare_size, part_size;
 946
 947         if (end2 <= buf2_size)
 948                 return memcmp(buf1, buf2 + offs2, compare_size);
 949
 950         part_size = end2 - buf2_size;
 951         if (memcmp(buf1, buf2 + offs2, part_size))
 952                 return true;
 953
 954         compare_size -= part_size;
 955
 956         return memcmp(buf1 + part_size, buf2, compare_size);
 957 }
 958
 959 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
 960                                  size_t ref_size, size_t buf_size,
 961                                  void *data, size_t head)
 962 {
 963         size_t ref_end = ref_offset + ref_size;
 964
 965         if (ref_end > buf_size) {
 966                 if (head > ref_offset || head < ref_end - buf_size)
 967                         return true;
 968         } else if (head > ref_offset && head < ref_end) {
 969                 return true;
 970         }
 971
 972         return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
 973                                         buf_size);
 974 }
 975
 976 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
 977                               void *data, size_t head)
 978 {
 979         if (head >= ref_size) {
 980                 memcpy(ref_buf, data + head - ref_size, ref_size);
 981         } else {
 982                 memcpy(ref_buf, data, head);
 983                 ref_size -= head;
 984                 memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
 985         }
 986 }
 987
 988 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
 989                              struct auxtrace_mmap *mm, unsigned char *data,
 990                              u64 head)
 991 {
 992         struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
 993         bool wrapped;
 994
 995         wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
 996                                        ptr->snapshot_ref_buf_size, mm->len,
 997                                        data, head);
 998
 999         intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
1000                           data, head);
1001
1002         return wrapped;
1003 }
1004
1005 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
1006 {
1007         int i, a, b;
1008
1009         b = buf_size >> 3;
1010         a = b - 512;
1011         if (a < 0)
1012                 a = 0;
1013
1014         for (i = a; i < b; i++) {
1015                 if (data[i])
1016                         return true;
1017         }
1018
1019         return false;
1020 }
1021
1022 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1023                                   struct auxtrace_mmap *mm, unsigned char *data,
1024                                   u64 *head, u64 *old)
1025 {
1026         struct intel_pt_recording *ptr =
1027                         container_of(itr, struct intel_pt_recording, itr);
1028         bool wrapped;
1029         int err;
1030
1031         pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1032                   __func__, idx, (size_t)*old, (size_t)*head);
1033
1034         err = intel_pt_snapshot_init(ptr, mm->len);
1035         if (err)
1036                 goto out_err;
1037
1038         if (idx >= ptr->snapshot_ref_cnt) {
1039                 err = intel_pt_alloc_snapshot_refs(ptr, idx);
1040                 if (err)
1041                         goto out_err;
1042         }
1043
1044         if (ptr->snapshot_ref_buf_size) {
1045                 if (!ptr->snapshot_refs[idx].ref_buf) {
1046                         err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1047                         if (err)
1048                                 goto out_err;
1049                 }
1050                 wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1051         } else {
1052                 wrapped = ptr->snapshot_refs[idx].wrapped;
1053                 if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1054                         ptr->snapshot_refs[idx].wrapped = true;
1055                         wrapped = true;
1056                 }
1057         }
1058
1059         /*
1060          * In full trace mode 'head' continually increases.  However in snapshot
1061          * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1062          * are adjusted to match the full trace case which expects that 'old' is
1063          * always less than 'head'.
1064          */
1065         if (wrapped) {
1066                 *old = *head;
1067                 *head += mm->len;
1068         } else {
1069                 if (mm->mask)
1070                         *old &= mm->mask;
1071                 else
1072                         *old %= mm->len;
1073                 if (*old > *head)
1074                         *head += mm->len;
1075         }
1076
1077         pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1078                   __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1079
1080         return 0;
1081
1082 out_err:
1083         pr_err("%s: failed, error %d\n", __func__, err);
1084         return err;
1085 }
1086
1087 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1088 {
1089         return rdtsc();
1090 }
1091
1092 static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1093 {
1094         struct intel_pt_recording *ptr =
1095                         container_of(itr, struct intel_pt_recording, itr);
1096         struct evsel *evsel;
1097
1098         evlist__for_each_entry(ptr->evlist, evsel) {
1099                 if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
1100                         return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1101                                                              idx);
1102         }
1103         return -EINVAL;
1104 }
1105
1106 struct auxtrace_record *intel_pt_recording_init(int *err)
1107 {
1108         struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1109         struct intel_pt_recording *ptr;
1110
1111         if (!intel_pt_pmu)
1112                 return NULL;
1113
1114         if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1115                 *err = -errno;
1116                 return NULL;
1117         }
1118
1119         ptr = zalloc(sizeof(struct intel_pt_recording));
1120         if (!ptr) {
1121                 *err = -ENOMEM;
1122                 return NULL;
1123         }
1124
1125         ptr->intel_pt_pmu = intel_pt_pmu;
1126         ptr->itr.recording_options = intel_pt_recording_options;
1127         ptr->itr.info_priv_size = intel_pt_info_priv_size;
1128         ptr->itr.info_fill = intel_pt_info_fill;
1129         ptr->itr.free = intel_pt_recording_free;
1130         ptr->itr.snapshot_start = intel_pt_snapshot_start;
1131         ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1132         ptr->itr.find_snapshot = intel_pt_find_snapshot;
1133         ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1134         ptr->itr.reference = intel_pt_reference;
1135         ptr->itr.read_finish = intel_pt_read_finish;
1136         return &ptr->itr;
1137 }