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 "../../../util/session.h"
  17 #include "../../../util/event.h"
  18 #include "../../../util/evlist.h"
  19 #include "../../../util/evsel.h"
  20 #include "../../../util/evsel_config.h"
  21 #include "../../../util/cpumap.h"
  22 #include "../../../util/mmap.h"
  23 #include <subcmd/parse-options.h>
  24 #include "../../../util/parse-events.h"
  25 #include "../../../util/pmu.h"
  26 #include "../../../util/debug.h"
  27 #include "../../../util/auxtrace.h"
  28 #include "../../../util/perf_api_probe.h"
  29 #include "../../../util/record.h"
  30 #include "../../../util/target.h"
  31 #include "../../../util/tsc.h"
  32 #include <internal/lib.h> // page_size
  33 #include "../../../util/intel-pt.h"
  34
  35 #define KiB(x) ((x) * 1024)
  36 #define MiB(x) ((x) * 1024 * 1024)
  37 #define KiB_MASK(x) (KiB(x) - 1)
  38 #define MiB_MASK(x) (MiB(x) - 1)
  39
  40 #define INTEL_PT_PSB_PERIOD_NEAR        256
  41
  42 struct intel_pt_snapshot_ref {
  43         void *ref_buf;
  44         size_t ref_offset;
  45         bool wrapped;
  46 };
  47
  48 struct intel_pt_recording {
  49         struct auxtrace_record          itr;
  50         struct perf_pmu                 *intel_pt_pmu;
  51         int                             have_sched_switch;
  52         struct evlist           *evlist;
  53         bool                            snapshot_mode;
  54         bool                            snapshot_init_done;
  55         size_t                          snapshot_size;
  56         size_t                          snapshot_ref_buf_size;
  57         int                             snapshot_ref_cnt;
  58         struct intel_pt_snapshot_ref    *snapshot_refs;
  59         size_t                          priv_size;
  60 };
  61
  62 static int intel_pt_parse_terms_with_default(struct list_head *formats,
  63                                              const char *str,
  64                                              u64 *config)
  65 {
  66         struct list_head *terms;
  67         struct perf_event_attr attr = { .size = 0, };
  68         int err;
  69
  70         terms = malloc(sizeof(struct list_head));
  71         if (!terms)
  72                 return -ENOMEM;
  73
  74         INIT_LIST_HEAD(terms);
  75
  76         err = parse_events_terms(terms, str);
  77         if (err)
  78                 goto out_free;
  79
  80         attr.config = *config;
  81         err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
  82         if (err)
  83                 goto out_free;
  84
  85         *config = attr.config;
  86 out_free:
  87         parse_events_terms__delete(terms);
  88         return err;
  89 }
  90
  91 static int intel_pt_parse_terms(struct list_head *formats, const char *str,
  92                                 u64 *config)
  93 {
  94         *config = 0;
  95         return intel_pt_parse_terms_with_default(formats, str, config);
  96 }
  97
  98 static u64 intel_pt_masked_bits(u64 mask, u64 bits)
  99 {
 100         const u64 top_bit = 1ULL << 63;
 101         u64 res = 0;
 102         int i;
 103
 104         for (i = 0; i < 64; i++) {
 105                 if (mask & top_bit) {
 106                         res <<= 1;
 107                         if (bits & top_bit)
 108                                 res |= 1;
 109                 }
 110                 mask <<= 1;
 111                 bits <<= 1;
 112         }
 113
 114         return res;
 115 }
 116
 117 static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
 118                                 struct evlist *evlist, u64 *res)
 119 {
 120         struct evsel *evsel;
 121         u64 mask;
 122
 123         *res = 0;
 124
 125         mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
 126         if (!mask)
 127                 return -EINVAL;
 128
 129         evlist__for_each_entry(evlist, evsel) {
 130                 if (evsel->core.attr.type == intel_pt_pmu->type) {
 131                         *res = intel_pt_masked_bits(mask, evsel->core.attr.config);
 132                         return 0;
 133                 }
 134         }
 135
 136         return -EINVAL;
 137 }
 138
 139 static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
 140                                   struct evlist *evlist)
 141 {
 142         u64 val;
 143         int err, topa_multiple_entries;
 144         size_t psb_period;
 145
 146         if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
 147                                 "%d", &topa_multiple_entries) != 1)
 148                 topa_multiple_entries = 0;
 149
 150         /*
 151          * Use caps/topa_multiple_entries to indicate early hardware that had
 152          * extra frequent PSBs.
 153          */
 154         if (!topa_multiple_entries) {
 155                 psb_period = 256;
 156                 goto out;
 157         }
 158
 159         err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
 160         if (err)
 161                 val = 0;
 162
 163         psb_period = 1 << (val + 11);
 164 out:
 165         pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
 166         return psb_period;
 167 }
 168
 169 static int intel_pt_pick_bit(int bits, int target)
 170 {
 171         int pos, pick = -1;
 172
 173         for (pos = 0; bits; bits >>= 1, pos++) {
 174                 if (bits & 1) {
 175                         if (pos <= target || pick < 0)
 176                                 pick = pos;
 177                         if (pos >= target)
 178                                 break;
 179                 }
 180         }
 181
 182         return pick;
 183 }
 184
 185 static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
 186 {
 187         char buf[256];
 188         int mtc, mtc_periods = 0, mtc_period;
 189         int psb_cyc, psb_periods, psb_period;
 190         int pos = 0;
 191         u64 config;
 192         char c;
 193
 194         pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
 195
 196         if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
 197                                 &mtc) != 1)
 198                 mtc = 1;
 199
 200         if (mtc) {
 201                 if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
 202                                         &mtc_periods) != 1)
 203                         mtc_periods = 0;
 204                 if (mtc_periods) {
 205                         mtc_period = intel_pt_pick_bit(mtc_periods, 3);
 206                         pos += scnprintf(buf + pos, sizeof(buf) - pos,
 207                                          ",mtc,mtc_period=%d", mtc_period);
 208                 }
 209         }
 210
 211         if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
 212                                 &psb_cyc) != 1)
 213                 psb_cyc = 1;
 214
 215         if (psb_cyc && mtc_periods) {
 216                 if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
 217                                         &psb_periods) != 1)
 218                         psb_periods = 0;
 219                 if (psb_periods) {
 220                         psb_period = intel_pt_pick_bit(psb_periods, 3);
 221                         pos += scnprintf(buf + pos, sizeof(buf) - pos,
 222                                          ",psb_period=%d", psb_period);
 223                 }
 224         }
 225
 226         if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
 227             perf_pmu__scan_file(intel_pt_pmu, "format/branch", "%c", &c) == 1)
 228                 pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
 229
 230         pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
 231
 232         intel_pt_parse_terms(&intel_pt_pmu->format, buf, &config);
 233
 234         return config;
 235 }
 236
 237 static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
 238                                            struct record_opts *opts,
 239                                            const char *str)
 240 {
 241         struct intel_pt_recording *ptr =
 242                         container_of(itr, struct intel_pt_recording, itr);
 243         unsigned long long snapshot_size = 0;
 244         char *endptr;
 245
 246         if (str) {
 247                 snapshot_size = strtoull(str, &endptr, 0);
 248                 if (*endptr || snapshot_size > SIZE_MAX)
 249                         return -1;
 250         }
 251
 252         opts->auxtrace_snapshot_mode = true;
 253         opts->auxtrace_snapshot_size = snapshot_size;
 254
 255         ptr->snapshot_size = snapshot_size;
 256
 257         return 0;
 258 }
 259
 260 struct perf_event_attr *
 261 intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
 262 {
 263         struct perf_event_attr *attr;
 264
 265         attr = zalloc(sizeof(struct perf_event_attr));
 266         if (!attr)
 267                 return NULL;
 268
 269         attr->config = intel_pt_default_config(intel_pt_pmu);
 270
 271         intel_pt_pmu->selectable = true;
 272
 273         return attr;
 274 }
 275
 276 static const char *intel_pt_find_filter(struct evlist *evlist,
 277                                         struct perf_pmu *intel_pt_pmu)
 278 {
 279         struct evsel *evsel;
 280
 281         evlist__for_each_entry(evlist, evsel) {
 282                 if (evsel->core.attr.type == intel_pt_pmu->type)
 283                         return evsel->filter;
 284         }
 285
 286         return NULL;
 287 }
 288
 289 static size_t intel_pt_filter_bytes(const char *filter)
 290 {
 291         size_t len = filter ? strlen(filter) : 0;
 292
 293         return len ? roundup(len + 1, 8) : 0;
 294 }
 295
 296 static size_t
 297 intel_pt_info_priv_size(struct auxtrace_record *itr, struct evlist *evlist)
 298 {
 299         struct intel_pt_recording *ptr =
 300                         container_of(itr, struct intel_pt_recording, itr);
 301         const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
 302
 303         ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
 304                          intel_pt_filter_bytes(filter);
 305
 306         return ptr->priv_size;
 307 }
 308
 309 static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
 310 {
 311         unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
 312
 313         __get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
 314         *n = ebx;
 315         *d = eax;
 316 }
 317
 318 static int intel_pt_info_fill(struct auxtrace_record *itr,
 319                               struct perf_session *session,
 320                               struct perf_record_auxtrace_info *auxtrace_info,
 321                               size_t priv_size)
 322 {
 323         struct intel_pt_recording *ptr =
 324                         container_of(itr, struct intel_pt_recording, itr);
 325         struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 326         struct perf_event_mmap_page *pc;
 327         struct perf_tsc_conversion tc = { .time_mult = 0, };
 328         bool cap_user_time_zero = false, per_cpu_mmaps;
 329         u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
 330         u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
 331         unsigned long max_non_turbo_ratio;
 332         size_t filter_str_len;
 333         const char *filter;
 334         __u64 *info;
 335         int err;
 336
 337         if (priv_size != ptr->priv_size)
 338                 return -EINVAL;
 339
 340         intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
 341         intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
 342                              &noretcomp_bit);
 343         intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
 344         mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
 345                                               "mtc_period");
 346         intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
 347
 348         intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
 349
 350         if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
 351                                 "%lu", &max_non_turbo_ratio) != 1)
 352                 max_non_turbo_ratio = 0;
 353
 354         filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
 355         filter_str_len = filter ? strlen(filter) : 0;
 356
 357         if (!session->evlist->core.nr_mmaps)
 358                 return -EINVAL;
 359
 360         pc = session->evlist->mmap[0].core.base;
 361         if (pc) {
 362                 err = perf_read_tsc_conversion(pc, &tc);
 363                 if (err) {
 364                         if (err != -EOPNOTSUPP)
 365                                 return err;
 366                 } else {
 367                         cap_user_time_zero = tc.time_mult != 0;
 368                 }
 369                 if (!cap_user_time_zero)
 370                         ui__warning("Intel Processor Trace: TSC not available\n");
 371         }
 372
 373         per_cpu_mmaps = !perf_cpu_map__empty(session->evlist->core.cpus);
 374
 375         auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
 376         auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
 377         auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
 378         auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
 379         auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
 380         auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
 381         auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
 382         auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
 383         auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
 384         auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
 385         auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
 386         auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
 387         auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
 388         auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
 389         auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
 390         auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
 391         auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
 392         auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
 393
 394         info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
 395
 396         if (filter_str_len) {
 397                 size_t len = intel_pt_filter_bytes(filter);
 398
 399                 strncpy((char *)info, filter, len);
 400                 info += len >> 3;
 401         }
 402
 403         return 0;
 404 }
 405
 406 static int intel_pt_track_switches(struct evlist *evlist)
 407 {
 408         const char *sched_switch = "sched:sched_switch";
 409         struct evsel *evsel;
 410         int err;
 411
 412         if (!perf_evlist__can_select_event(evlist, sched_switch))
 413                 return -EPERM;
 414
 415         err = parse_events(evlist, sched_switch, NULL);
 416         if (err) {
 417                 pr_debug2("%s: failed to parse %s, error %d\n",
 418                           __func__, sched_switch, err);
 419                 return err;
 420         }
 421
 422         evsel = evlist__last(evlist);
 423
 424         evsel__set_sample_bit(evsel, CPU);
 425         evsel__set_sample_bit(evsel, TIME);
 426
 427         evsel->core.system_wide = true;
 428         evsel->no_aux_samples = true;
 429         evsel->immediate = true;
 430
 431         return 0;
 432 }
 433
 434 static void intel_pt_valid_str(char *str, size_t len, u64 valid)
 435 {
 436         unsigned int val, last = 0, state = 1;
 437         int p = 0;
 438
 439         str[0] = '\0';
 440
 441         for (val = 0; val <= 64; val++, valid >>= 1) {
 442                 if (valid & 1) {
 443                         last = val;
 444                         switch (state) {
 445                         case 0:
 446                                 p += scnprintf(str + p, len - p, ",");
 447                                 /* Fall through */
 448                         case 1:
 449                                 p += scnprintf(str + p, len - p, "%u", val);
 450                                 state = 2;
 451                                 break;
 452                         case 2:
 453                                 state = 3;
 454                                 break;
 455                         case 3:
 456                                 state = 4;
 457                                 break;
 458                         default:
 459                                 break;
 460                         }
 461                 } else {
 462                         switch (state) {
 463                         case 3:
 464                                 p += scnprintf(str + p, len - p, ",%u", last);
 465                                 state = 0;
 466                                 break;
 467                         case 4:
 468                                 p += scnprintf(str + p, len - p, "-%u", last);
 469                                 state = 0;
 470                                 break;
 471                         default:
 472                                 break;
 473                         }
 474                         if (state != 1)
 475                                 state = 0;
 476                 }
 477         }
 478 }
 479
 480 static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
 481                                     const char *caps, const char *name,
 482                                     const char *supported, u64 config)
 483 {
 484         char valid_str[256];
 485         unsigned int shift;
 486         unsigned long long valid;
 487         u64 bits;
 488         int ok;
 489
 490         if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
 491                 valid = 0;
 492
 493         if (supported &&
 494             perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
 495                 valid = 0;
 496
 497         valid |= 1;
 498
 499         bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
 500
 501         config &= bits;
 502
 503         for (shift = 0; bits && !(bits & 1); shift++)
 504                 bits >>= 1;
 505
 506         config >>= shift;
 507
 508         if (config > 63)
 509                 goto out_err;
 510
 511         if (valid & (1 << config))
 512                 return 0;
 513 out_err:
 514         intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
 515         pr_err("Invalid %s for %s. Valid values are: %s\n",
 516                name, INTEL_PT_PMU_NAME, valid_str);
 517         return -EINVAL;
 518 }
 519
 520 static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
 521                                     struct evsel *evsel)
 522 {
 523         int err;
 524         char c;
 525
 526         if (!evsel)
 527                 return 0;
 528
 529         /*
 530          * If supported, force pass-through config term (pt=1) even if user
 531          * sets pt=0, which avoids senseless kernel errors.
 532          */
 533         if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
 534             !(evsel->core.attr.config & 1)) {
 535                 pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
 536                 evsel->core.attr.config |= 1;
 537         }
 538
 539         err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
 540                                        "cyc_thresh", "caps/psb_cyc",
 541                                        evsel->core.attr.config);
 542         if (err)
 543                 return err;
 544
 545         err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
 546                                        "mtc_period", "caps/mtc",
 547                                        evsel->core.attr.config);
 548         if (err)
 549                 return err;
 550
 551         return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
 552                                         "psb_period", "caps/psb_cyc",
 553                                         evsel->core.attr.config);
 554 }
 555
 556 static void intel_pt_config_sample_mode(struct perf_pmu *intel_pt_pmu,
 557                                         struct evsel *evsel)
 558 {
 559         struct perf_evsel_config_term *term;
 560         u64 user_bits = 0, bits;
 561
 562         term = perf_evsel__get_config_term(evsel, CFG_CHG);
 563         if (term)
 564                 user_bits = term->val.cfg_chg;
 565
 566         bits = perf_pmu__format_bits(&intel_pt_pmu->format, "psb_period");
 567
 568         /* Did user change psb_period */
 569         if (bits & user_bits)
 570                 return;
 571
 572         /* Set psb_period to 0 */
 573         evsel->core.attr.config &= ~bits;
 574 }
 575
 576 static void intel_pt_min_max_sample_sz(struct evlist *evlist,
 577                                        size_t *min_sz, size_t *max_sz)
 578 {
 579         struct evsel *evsel;
 580
 581         evlist__for_each_entry(evlist, evsel) {
 582                 size_t sz = evsel->core.attr.aux_sample_size;
 583
 584                 if (!sz)
 585                         continue;
 586                 if (min_sz && (sz < *min_sz || !*min_sz))
 587                         *min_sz = sz;
 588                 if (max_sz && sz > *max_sz)
 589                         *max_sz = sz;
 590         }
 591 }
 592
 593 /*
 594  * Currently, there is not enough information to disambiguate different PEBS
 595  * events, so only allow one.
 596  */
 597 static bool intel_pt_too_many_aux_output(struct evlist *evlist)
 598 {
 599         struct evsel *evsel;
 600         int aux_output_cnt = 0;
 601
 602         evlist__for_each_entry(evlist, evsel)
 603                 aux_output_cnt += !!evsel->core.attr.aux_output;
 604
 605         if (aux_output_cnt > 1) {
 606                 pr_err(INTEL_PT_PMU_NAME " supports at most one event with aux-output\n");
 607                 return true;
 608         }
 609
 610         return false;
 611 }
 612
 613 static int intel_pt_recording_options(struct auxtrace_record *itr,
 614                                       struct evlist *evlist,
 615                                       struct record_opts *opts)
 616 {
 617         struct intel_pt_recording *ptr =
 618                         container_of(itr, struct intel_pt_recording, itr);
 619         struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 620         bool have_timing_info, need_immediate = false;
 621         struct evsel *evsel, *intel_pt_evsel = NULL;
 622         const struct perf_cpu_map *cpus = evlist->core.cpus;
 623         bool privileged = perf_event_paranoid_check(-1);
 624         u64 tsc_bit;
 625         int err;
 626
 627         ptr->evlist = evlist;
 628         ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
 629
 630         evlist__for_each_entry(evlist, evsel) {
 631                 if (evsel->core.attr.type == intel_pt_pmu->type) {
 632                         if (intel_pt_evsel) {
 633                                 pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
 634                                 return -EINVAL;
 635                         }
 636                         evsel->core.attr.freq = 0;
 637                         evsel->core.attr.sample_period = 1;
 638                         intel_pt_evsel = evsel;
 639                         opts->full_auxtrace = true;
 640                 }
 641         }
 642
 643         if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
 644                 pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
 645                 return -EINVAL;
 646         }
 647
 648         if (opts->auxtrace_snapshot_mode && opts->auxtrace_sample_mode) {
 649                 pr_err("Snapshot mode (" INTEL_PT_PMU_NAME " PMU) and sample trace cannot be used together\n");
 650                 return -EINVAL;
 651         }
 652
 653         if (opts->use_clockid) {
 654                 pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
 655                 return -EINVAL;
 656         }
 657
 658         if (intel_pt_too_many_aux_output(evlist))
 659                 return -EINVAL;
 660
 661         if (!opts->full_auxtrace)
 662                 return 0;
 663
 664         if (opts->auxtrace_sample_mode)
 665                 intel_pt_config_sample_mode(intel_pt_pmu, intel_pt_evsel);
 666
 667         err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
 668         if (err)
 669                 return err;
 670
 671         /* Set default sizes for snapshot mode */
 672         if (opts->auxtrace_snapshot_mode) {
 673                 size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 674
 675                 if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
 676                         if (privileged) {
 677                                 opts->auxtrace_mmap_pages = MiB(4) / page_size;
 678                         } else {
 679                                 opts->auxtrace_mmap_pages = KiB(128) / page_size;
 680                                 if (opts->mmap_pages == UINT_MAX)
 681                                         opts->mmap_pages = KiB(256) / page_size;
 682                         }
 683                 } else if (!opts->auxtrace_mmap_pages && !privileged &&
 684                            opts->mmap_pages == UINT_MAX) {
 685                         opts->mmap_pages = KiB(256) / page_size;
 686                 }
 687                 if (!opts->auxtrace_snapshot_size)
 688                         opts->auxtrace_snapshot_size =
 689                                 opts->auxtrace_mmap_pages * (size_t)page_size;
 690                 if (!opts->auxtrace_mmap_pages) {
 691                         size_t sz = opts->auxtrace_snapshot_size;
 692
 693                         sz = round_up(sz, page_size) / page_size;
 694                         opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 695                 }
 696                 if (opts->auxtrace_snapshot_size >
 697                                 opts->auxtrace_mmap_pages * (size_t)page_size) {
 698                         pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
 699                                opts->auxtrace_snapshot_size,
 700                                opts->auxtrace_mmap_pages * (size_t)page_size);
 701                         return -EINVAL;
 702                 }
 703                 if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
 704                         pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
 705                         return -EINVAL;
 706                 }
 707                 pr_debug2("Intel PT snapshot size: %zu\n",
 708                           opts->auxtrace_snapshot_size);
 709                 if (psb_period &&
 710                     opts->auxtrace_snapshot_size <= psb_period +
 711                                                   INTEL_PT_PSB_PERIOD_NEAR)
 712                         ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
 713                                     opts->auxtrace_snapshot_size, psb_period);
 714         }
 715
 716         /* Set default sizes for sample mode */
 717         if (opts->auxtrace_sample_mode) {
 718                 size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 719                 size_t min_sz = 0, max_sz = 0;
 720
 721                 intel_pt_min_max_sample_sz(evlist, &min_sz, &max_sz);
 722                 if (!opts->auxtrace_mmap_pages && !privileged &&
 723                     opts->mmap_pages == UINT_MAX)
 724                         opts->mmap_pages = KiB(256) / page_size;
 725                 if (!opts->auxtrace_mmap_pages) {
 726                         size_t sz = round_up(max_sz, page_size) / page_size;
 727
 728                         opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 729                 }
 730                 if (max_sz > opts->auxtrace_mmap_pages * (size_t)page_size) {
 731                         pr_err("Sample size %zu must not be greater than AUX area tracing mmap size %zu\n",
 732                                max_sz,
 733                                opts->auxtrace_mmap_pages * (size_t)page_size);
 734                         return -EINVAL;
 735                 }
 736                 pr_debug2("Intel PT min. sample size: %zu max. sample size: %zu\n",
 737                           min_sz, max_sz);
 738                 if (psb_period &&
 739                     min_sz <= psb_period + INTEL_PT_PSB_PERIOD_NEAR)
 740                         ui__warning("Intel PT sample size (%zu) may be too small for PSB period (%zu)\n",
 741                                     min_sz, psb_period);
 742         }
 743
 744         /* Set default sizes for full trace mode */
 745         if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
 746                 if (privileged) {
 747                         opts->auxtrace_mmap_pages = MiB(4) / page_size;
 748                 } else {
 749                         opts->auxtrace_mmap_pages = KiB(128) / page_size;
 750                         if (opts->mmap_pages == UINT_MAX)
 751                                 opts->mmap_pages = KiB(256) / page_size;
 752                 }
 753         }
 754
 755         /* Validate auxtrace_mmap_pages */
 756         if (opts->auxtrace_mmap_pages) {
 757                 size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 758                 size_t min_sz;
 759
 760                 if (opts->auxtrace_snapshot_mode || opts->auxtrace_sample_mode)
 761                         min_sz = KiB(4);
 762                 else
 763                         min_sz = KiB(8);
 764
 765                 if (sz < min_sz || !is_power_of_2(sz)) {
 766                         pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
 767                                min_sz / 1024);
 768                         return -EINVAL;
 769                 }
 770         }
 771
 772         intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
 773
 774         if (opts->full_auxtrace && (intel_pt_evsel->core.attr.config & tsc_bit))
 775                 have_timing_info = true;
 776         else
 777                 have_timing_info = false;
 778
 779         /*
 780          * Per-cpu recording needs sched_switch events to distinguish different
 781          * threads.
 782          */
 783         if (have_timing_info && !perf_cpu_map__empty(cpus)) {
 784                 if (perf_can_record_switch_events()) {
 785                         bool cpu_wide = !target__none(&opts->target) &&
 786                                         !target__has_task(&opts->target);
 787
 788                         if (!cpu_wide && perf_can_record_cpu_wide()) {
 789                                 struct evsel *switch_evsel;
 790
 791                                 err = parse_events(evlist, "dummy:u", NULL);
 792                                 if (err)
 793                                         return err;
 794
 795                                 switch_evsel = evlist__last(evlist);
 796
 797                                 switch_evsel->core.attr.freq = 0;
 798                                 switch_evsel->core.attr.sample_period = 1;
 799                                 switch_evsel->core.attr.context_switch = 1;
 800
 801                                 switch_evsel->core.system_wide = true;
 802                                 switch_evsel->no_aux_samples = true;
 803                                 switch_evsel->immediate = true;
 804
 805                                 evsel__set_sample_bit(switch_evsel, TID);
 806                                 evsel__set_sample_bit(switch_evsel, TIME);
 807                                 evsel__set_sample_bit(switch_evsel, CPU);
 808                                 evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
 809
 810                                 opts->record_switch_events = false;
 811                                 ptr->have_sched_switch = 3;
 812                         } else {
 813                                 opts->record_switch_events = true;
 814                                 need_immediate = true;
 815                                 if (cpu_wide)
 816                                         ptr->have_sched_switch = 3;
 817                                 else
 818                                         ptr->have_sched_switch = 2;
 819                         }
 820                 } else {
 821                         err = intel_pt_track_switches(evlist);
 822                         if (err == -EPERM)
 823                                 pr_debug2("Unable to select sched:sched_switch\n");
 824                         else if (err)
 825                                 return err;
 826                         else
 827                                 ptr->have_sched_switch = 1;
 828                 }
 829         }
 830
 831         if (intel_pt_evsel) {
 832                 /*
 833                  * To obtain the auxtrace buffer file descriptor, the auxtrace
 834                  * event must come first.
 835                  */
 836                 perf_evlist__to_front(evlist, intel_pt_evsel);
 837                 /*
 838                  * In the case of per-cpu mmaps, we need the CPU on the
 839                  * AUX event.
 840                  */
 841                 if (!perf_cpu_map__empty(cpus))
 842                         evsel__set_sample_bit(intel_pt_evsel, CPU);
 843         }
 844
 845         /* Add dummy event to keep tracking */
 846         if (opts->full_auxtrace) {
 847                 struct evsel *tracking_evsel;
 848
 849                 err = parse_events(evlist, "dummy:u", NULL);
 850                 if (err)
 851                         return err;
 852
 853                 tracking_evsel = evlist__last(evlist);
 854
 855                 perf_evlist__set_tracking_event(evlist, tracking_evsel);
 856
 857                 tracking_evsel->core.attr.freq = 0;
 858                 tracking_evsel->core.attr.sample_period = 1;
 859
 860                 tracking_evsel->no_aux_samples = true;
 861                 if (need_immediate)
 862                         tracking_evsel->immediate = true;
 863
 864                 /* In per-cpu case, always need the time of mmap events etc */
 865                 if (!perf_cpu_map__empty(cpus)) {
 866                         evsel__set_sample_bit(tracking_evsel, TIME);
 867                         /* And the CPU for switch events */
 868                         evsel__set_sample_bit(tracking_evsel, CPU);
 869                 }
 870                 evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
 871         }
 872
 873         /*
 874          * Warn the user when we do not have enough information to decode i.e.
 875          * per-cpu with no sched_switch (except workload-only).
 876          */
 877         if (!ptr->have_sched_switch && !perf_cpu_map__empty(cpus) &&
 878             !target__none(&opts->target))
 879                 ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
 880
 881         return 0;
 882 }
 883
 884 static int intel_pt_snapshot_start(struct auxtrace_record *itr)
 885 {
 886         struct intel_pt_recording *ptr =
 887                         container_of(itr, struct intel_pt_recording, itr);
 888         struct evsel *evsel;
 889
 890         evlist__for_each_entry(ptr->evlist, evsel) {
 891                 if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
 892                         return evsel__disable(evsel);
 893         }
 894         return -EINVAL;
 895 }
 896
 897 static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
 898 {
 899         struct intel_pt_recording *ptr =
 900                         container_of(itr, struct intel_pt_recording, itr);
 901         struct evsel *evsel;
 902
 903         evlist__for_each_entry(ptr->evlist, evsel) {
 904                 if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
 905                         return evsel__enable(evsel);
 906         }
 907         return -EINVAL;
 908 }
 909
 910 static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
 911 {
 912         const size_t sz = sizeof(struct intel_pt_snapshot_ref);
 913         int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
 914         struct intel_pt_snapshot_ref *refs;
 915
 916         if (!new_cnt)
 917                 new_cnt = 16;
 918
 919         while (new_cnt <= idx)
 920                 new_cnt *= 2;
 921
 922         refs = calloc(new_cnt, sz);
 923         if (!refs)
 924                 return -ENOMEM;
 925
 926         memcpy(refs, ptr->snapshot_refs, cnt * sz);
 927
 928         ptr->snapshot_refs = refs;
 929         ptr->snapshot_ref_cnt = new_cnt;
 930
 931         return 0;
 932 }
 933
 934 static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
 935 {
 936         int i;
 937
 938         for (i = 0; i < ptr->snapshot_ref_cnt; i++)
 939                 zfree(&ptr->snapshot_refs[i].ref_buf);
 940         zfree(&ptr->snapshot_refs);
 941 }
 942
 943 static void intel_pt_recording_free(struct auxtrace_record *itr)
 944 {
 945         struct intel_pt_recording *ptr =
 946                         container_of(itr, struct intel_pt_recording, itr);
 947
 948         intel_pt_free_snapshot_refs(ptr);
 949         free(ptr);
 950 }
 951
 952 static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
 953                                        size_t snapshot_buf_size)
 954 {
 955         size_t ref_buf_size = ptr->snapshot_ref_buf_size;
 956         void *ref_buf;
 957
 958         ref_buf = zalloc(ref_buf_size);
 959         if (!ref_buf)
 960                 return -ENOMEM;
 961
 962         ptr->snapshot_refs[idx].ref_buf = ref_buf;
 963         ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
 964
 965         return 0;
 966 }
 967
 968 static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
 969                                              size_t snapshot_buf_size)
 970 {
 971         const size_t max_size = 256 * 1024;
 972         size_t buf_size = 0, psb_period;
 973
 974         if (ptr->snapshot_size <= 64 * 1024)
 975                 return 0;
 976
 977         psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
 978         if (psb_period)
 979                 buf_size = psb_period * 2;
 980
 981         if (!buf_size || buf_size > max_size)
 982                 buf_size = max_size;
 983
 984         if (buf_size >= snapshot_buf_size)
 985                 return 0;
 986
 987         if (buf_size >= ptr->snapshot_size / 2)
 988                 return 0;
 989
 990         return buf_size;
 991 }
 992
 993 static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
 994                                   size_t snapshot_buf_size)
 995 {
 996         if (ptr->snapshot_init_done)
 997                 return 0;
 998
 999         ptr->snapshot_init_done = true;
1000
1001         ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
1002                                                         snapshot_buf_size);
1003
1004         return 0;
1005 }
1006
1007 /**
1008  * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
1009  * @buf1: first buffer
1010  * @compare_size: number of bytes to compare
1011  * @buf2: second buffer (a circular buffer)
1012  * @offs2: offset in second buffer
1013  * @buf2_size: size of second buffer
1014  *
1015  * The comparison allows for the possibility that the bytes to compare in the
1016  * circular buffer are not contiguous.  It is assumed that @compare_size <=
1017  * @buf2_size.  This function returns %false if the bytes are identical, %true
1018  * otherwise.
1019  */
1020 static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
1021                                      void *buf2, size_t offs2, size_t buf2_size)
1022 {
1023         size_t end2 = offs2 + compare_size, part_size;
1024
1025         if (end2 <= buf2_size)
1026                 return memcmp(buf1, buf2 + offs2, compare_size);
1027
1028         part_size = end2 - buf2_size;
1029         if (memcmp(buf1, buf2 + offs2, part_size))
1030                 return true;
1031
1032         compare_size -= part_size;
1033
1034         return memcmp(buf1 + part_size, buf2, compare_size);
1035 }
1036
1037 static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
1038                                  size_t ref_size, size_t buf_size,
1039                                  void *data, size_t head)
1040 {
1041         size_t ref_end = ref_offset + ref_size;
1042
1043         if (ref_end > buf_size) {
1044                 if (head > ref_offset || head < ref_end - buf_size)
1045                         return true;
1046         } else if (head > ref_offset && head < ref_end) {
1047                 return true;
1048         }
1049
1050         return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
1051                                         buf_size);
1052 }
1053
1054 static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
1055                               void *data, size_t head)
1056 {
1057         if (head >= ref_size) {
1058                 memcpy(ref_buf, data + head - ref_size, ref_size);
1059         } else {
1060                 memcpy(ref_buf, data, head);
1061                 ref_size -= head;
1062                 memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
1063         }
1064 }
1065
1066 static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
1067                              struct auxtrace_mmap *mm, unsigned char *data,
1068                              u64 head)
1069 {
1070         struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
1071         bool wrapped;
1072
1073         wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
1074                                        ptr->snapshot_ref_buf_size, mm->len,
1075                                        data, head);
1076
1077         intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
1078                           data, head);
1079
1080         return wrapped;
1081 }
1082
1083 static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
1084 {
1085         int i, a, b;
1086
1087         b = buf_size >> 3;
1088         a = b - 512;
1089         if (a < 0)
1090                 a = 0;
1091
1092         for (i = a; i < b; i++) {
1093                 if (data[i])
1094                         return true;
1095         }
1096
1097         return false;
1098 }
1099
1100 static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1101                                   struct auxtrace_mmap *mm, unsigned char *data,
1102                                   u64 *head, u64 *old)
1103 {
1104         struct intel_pt_recording *ptr =
1105                         container_of(itr, struct intel_pt_recording, itr);
1106         bool wrapped;
1107         int err;
1108
1109         pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1110                   __func__, idx, (size_t)*old, (size_t)*head);
1111
1112         err = intel_pt_snapshot_init(ptr, mm->len);
1113         if (err)
1114                 goto out_err;
1115
1116         if (idx >= ptr->snapshot_ref_cnt) {
1117                 err = intel_pt_alloc_snapshot_refs(ptr, idx);
1118                 if (err)
1119                         goto out_err;
1120         }
1121
1122         if (ptr->snapshot_ref_buf_size) {
1123                 if (!ptr->snapshot_refs[idx].ref_buf) {
1124                         err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1125                         if (err)
1126                                 goto out_err;
1127                 }
1128                 wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1129         } else {
1130                 wrapped = ptr->snapshot_refs[idx].wrapped;
1131                 if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1132                         ptr->snapshot_refs[idx].wrapped = true;
1133                         wrapped = true;
1134                 }
1135         }
1136
1137         /*
1138          * In full trace mode 'head' continually increases.  However in snapshot
1139          * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1140          * are adjusted to match the full trace case which expects that 'old' is
1141          * always less than 'head'.
1142          */
1143         if (wrapped) {
1144                 *old = *head;
1145                 *head += mm->len;
1146         } else {
1147                 if (mm->mask)
1148                         *old &= mm->mask;
1149                 else
1150                         *old %= mm->len;
1151                 if (*old > *head)
1152                         *head += mm->len;
1153         }
1154
1155         pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1156                   __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1157
1158         return 0;
1159
1160 out_err:
1161         pr_err("%s: failed, error %d\n", __func__, err);
1162         return err;
1163 }
1164
1165 static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1166 {
1167         return rdtsc();
1168 }
1169
1170 struct auxtrace_record *intel_pt_recording_init(int *err)
1171 {
1172         struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1173         struct intel_pt_recording *ptr;
1174
1175         if (!intel_pt_pmu)
1176                 return NULL;
1177
1178         if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1179                 *err = -errno;
1180                 return NULL;
1181         }
1182
1183         ptr = zalloc(sizeof(struct intel_pt_recording));
1184         if (!ptr) {
1185                 *err = -ENOMEM;
1186                 return NULL;
1187         }
1188
1189         ptr->intel_pt_pmu = intel_pt_pmu;
1190         ptr->itr.pmu = intel_pt_pmu;
1191         ptr->itr.recording_options = intel_pt_recording_options;
1192         ptr->itr.info_priv_size = intel_pt_info_priv_size;
1193         ptr->itr.info_fill = intel_pt_info_fill;
1194         ptr->itr.free = intel_pt_recording_free;
1195         ptr->itr.snapshot_start = intel_pt_snapshot_start;
1196         ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1197         ptr->itr.find_snapshot = intel_pt_find_snapshot;
1198         ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1199         ptr->itr.reference = intel_pt_reference;
1200         ptr->itr.read_finish = auxtrace_record__read_finish;
1201         /*
1202          * Decoding starts at a PSB packet. Minimum PSB period is 2K so 4K
1203          * should give at least 1 PSB per sample.
1204          */
1205         ptr->itr.default_aux_sample_size = 4096;
1206         return &ptr->itr;
1207 }