linux/tools/perf/arch/x86/util/intel-pt.c
<<
>>
Prefs
   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
  42struct intel_pt_snapshot_ref {
  43        void *ref_buf;
  44        size_t ref_offset;
  45        bool wrapped;
  46};
  47
  48struct 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
  62static int intel_pt_parse_terms_with_default(const char *pmu_name,
  63                                             struct list_head *formats,
  64                                             const char *str,
  65                                             u64 *config)
  66{
  67        struct list_head *terms;
  68        struct perf_event_attr attr = { .size = 0, };
  69        int err;
  70
  71        terms = malloc(sizeof(struct list_head));
  72        if (!terms)
  73                return -ENOMEM;
  74
  75        INIT_LIST_HEAD(terms);
  76
  77        err = parse_events_terms(terms, str);
  78        if (err)
  79                goto out_free;
  80
  81        attr.config = *config;
  82        err = perf_pmu__config_terms(pmu_name, formats, &attr, terms, true,
  83                                     NULL);
  84        if (err)
  85                goto out_free;
  86
  87        *config = attr.config;
  88out_free:
  89        parse_events_terms__delete(terms);
  90        return err;
  91}
  92
  93static int intel_pt_parse_terms(const char *pmu_name, struct list_head *formats,
  94                                const char *str, u64 *config)
  95{
  96        *config = 0;
  97        return intel_pt_parse_terms_with_default(pmu_name, formats, str,
  98                                                 config);
  99}
 100
 101static u64 intel_pt_masked_bits(u64 mask, u64 bits)
 102{
 103        const u64 top_bit = 1ULL << 63;
 104        u64 res = 0;
 105        int i;
 106
 107        for (i = 0; i < 64; i++) {
 108                if (mask & top_bit) {
 109                        res <<= 1;
 110                        if (bits & top_bit)
 111                                res |= 1;
 112                }
 113                mask <<= 1;
 114                bits <<= 1;
 115        }
 116
 117        return res;
 118}
 119
 120static int intel_pt_read_config(struct perf_pmu *intel_pt_pmu, const char *str,
 121                                struct evlist *evlist, u64 *res)
 122{
 123        struct evsel *evsel;
 124        u64 mask;
 125
 126        *res = 0;
 127
 128        mask = perf_pmu__format_bits(&intel_pt_pmu->format, str);
 129        if (!mask)
 130                return -EINVAL;
 131
 132        evlist__for_each_entry(evlist, evsel) {
 133                if (evsel->core.attr.type == intel_pt_pmu->type) {
 134                        *res = intel_pt_masked_bits(mask, evsel->core.attr.config);
 135                        return 0;
 136                }
 137        }
 138
 139        return -EINVAL;
 140}
 141
 142static size_t intel_pt_psb_period(struct perf_pmu *intel_pt_pmu,
 143                                  struct evlist *evlist)
 144{
 145        u64 val;
 146        int err, topa_multiple_entries;
 147        size_t psb_period;
 148
 149        if (perf_pmu__scan_file(intel_pt_pmu, "caps/topa_multiple_entries",
 150                                "%d", &topa_multiple_entries) != 1)
 151                topa_multiple_entries = 0;
 152
 153        /*
 154         * Use caps/topa_multiple_entries to indicate early hardware that had
 155         * extra frequent PSBs.
 156         */
 157        if (!topa_multiple_entries) {
 158                psb_period = 256;
 159                goto out;
 160        }
 161
 162        err = intel_pt_read_config(intel_pt_pmu, "psb_period", evlist, &val);
 163        if (err)
 164                val = 0;
 165
 166        psb_period = 1 << (val + 11);
 167out:
 168        pr_debug2("%s psb_period %zu\n", intel_pt_pmu->name, psb_period);
 169        return psb_period;
 170}
 171
 172static int intel_pt_pick_bit(int bits, int target)
 173{
 174        int pos, pick = -1;
 175
 176        for (pos = 0; bits; bits >>= 1, pos++) {
 177                if (bits & 1) {
 178                        if (pos <= target || pick < 0)
 179                                pick = pos;
 180                        if (pos >= target)
 181                                break;
 182                }
 183        }
 184
 185        return pick;
 186}
 187
 188static u64 intel_pt_default_config(struct perf_pmu *intel_pt_pmu)
 189{
 190        char buf[256];
 191        int mtc, mtc_periods = 0, mtc_period;
 192        int psb_cyc, psb_periods, psb_period;
 193        int pos = 0;
 194        u64 config;
 195        char c;
 196
 197        pos += scnprintf(buf + pos, sizeof(buf) - pos, "tsc");
 198
 199        if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc", "%d",
 200                                &mtc) != 1)
 201                mtc = 1;
 202
 203        if (mtc) {
 204                if (perf_pmu__scan_file(intel_pt_pmu, "caps/mtc_periods", "%x",
 205                                        &mtc_periods) != 1)
 206                        mtc_periods = 0;
 207                if (mtc_periods) {
 208                        mtc_period = intel_pt_pick_bit(mtc_periods, 3);
 209                        pos += scnprintf(buf + pos, sizeof(buf) - pos,
 210                                         ",mtc,mtc_period=%d", mtc_period);
 211                }
 212        }
 213
 214        if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_cyc", "%d",
 215                                &psb_cyc) != 1)
 216                psb_cyc = 1;
 217
 218        if (psb_cyc && mtc_periods) {
 219                if (perf_pmu__scan_file(intel_pt_pmu, "caps/psb_periods", "%x",
 220                                        &psb_periods) != 1)
 221                        psb_periods = 0;
 222                if (psb_periods) {
 223                        psb_period = intel_pt_pick_bit(psb_periods, 3);
 224                        pos += scnprintf(buf + pos, sizeof(buf) - pos,
 225                                         ",psb_period=%d", psb_period);
 226                }
 227        }
 228
 229        if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
 230            perf_pmu__scan_file(intel_pt_pmu, "format/branch", "%c", &c) == 1)
 231                pos += scnprintf(buf + pos, sizeof(buf) - pos, ",pt,branch");
 232
 233        pr_debug2("%s default config: %s\n", intel_pt_pmu->name, buf);
 234
 235        intel_pt_parse_terms(intel_pt_pmu->name, &intel_pt_pmu->format, buf,
 236                             &config);
 237
 238        return config;
 239}
 240
 241static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
 242                                           struct record_opts *opts,
 243                                           const char *str)
 244{
 245        struct intel_pt_recording *ptr =
 246                        container_of(itr, struct intel_pt_recording, itr);
 247        unsigned long long snapshot_size = 0;
 248        char *endptr;
 249
 250        if (str) {
 251                snapshot_size = strtoull(str, &endptr, 0);
 252                if (*endptr || snapshot_size > SIZE_MAX)
 253                        return -1;
 254        }
 255
 256        opts->auxtrace_snapshot_mode = true;
 257        opts->auxtrace_snapshot_size = snapshot_size;
 258
 259        ptr->snapshot_size = snapshot_size;
 260
 261        return 0;
 262}
 263
 264struct perf_event_attr *
 265intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
 266{
 267        struct perf_event_attr *attr;
 268
 269        attr = zalloc(sizeof(struct perf_event_attr));
 270        if (!attr)
 271                return NULL;
 272
 273        attr->config = intel_pt_default_config(intel_pt_pmu);
 274
 275        intel_pt_pmu->selectable = true;
 276
 277        return attr;
 278}
 279
 280static const char *intel_pt_find_filter(struct evlist *evlist,
 281                                        struct perf_pmu *intel_pt_pmu)
 282{
 283        struct evsel *evsel;
 284
 285        evlist__for_each_entry(evlist, evsel) {
 286                if (evsel->core.attr.type == intel_pt_pmu->type)
 287                        return evsel->filter;
 288        }
 289
 290        return NULL;
 291}
 292
 293static size_t intel_pt_filter_bytes(const char *filter)
 294{
 295        size_t len = filter ? strlen(filter) : 0;
 296
 297        return len ? roundup(len + 1, 8) : 0;
 298}
 299
 300static size_t
 301intel_pt_info_priv_size(struct auxtrace_record *itr, struct evlist *evlist)
 302{
 303        struct intel_pt_recording *ptr =
 304                        container_of(itr, struct intel_pt_recording, itr);
 305        const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
 306
 307        ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
 308                         intel_pt_filter_bytes(filter);
 309        ptr->priv_size += sizeof(u64); /* Cap Event Trace */
 310
 311        return ptr->priv_size;
 312}
 313
 314static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
 315{
 316        unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
 317
 318        __get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
 319        *n = ebx;
 320        *d = eax;
 321}
 322
 323static int intel_pt_info_fill(struct auxtrace_record *itr,
 324                              struct perf_session *session,
 325                              struct perf_record_auxtrace_info *auxtrace_info,
 326                              size_t priv_size)
 327{
 328        struct intel_pt_recording *ptr =
 329                        container_of(itr, struct intel_pt_recording, itr);
 330        struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 331        struct perf_event_mmap_page *pc;
 332        struct perf_tsc_conversion tc = { .time_mult = 0, };
 333        bool cap_user_time_zero = false, per_cpu_mmaps;
 334        u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
 335        u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
 336        unsigned long max_non_turbo_ratio;
 337        size_t filter_str_len;
 338        const char *filter;
 339        int event_trace;
 340        __u64 *info;
 341        int err;
 342
 343        if (priv_size != ptr->priv_size)
 344                return -EINVAL;
 345
 346        intel_pt_parse_terms(intel_pt_pmu->name, &intel_pt_pmu->format,
 347                             "tsc", &tsc_bit);
 348        intel_pt_parse_terms(intel_pt_pmu->name, &intel_pt_pmu->format,
 349                             "noretcomp", &noretcomp_bit);
 350        intel_pt_parse_terms(intel_pt_pmu->name, &intel_pt_pmu->format,
 351                             "mtc", &mtc_bit);
 352        mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
 353                                              "mtc_period");
 354        intel_pt_parse_terms(intel_pt_pmu->name, &intel_pt_pmu->format,
 355                             "cyc", &cyc_bit);
 356
 357        intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
 358
 359        if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
 360                                "%lu", &max_non_turbo_ratio) != 1)
 361                max_non_turbo_ratio = 0;
 362        if (perf_pmu__scan_file(intel_pt_pmu, "caps/event_trace",
 363                                "%d", &event_trace) != 1)
 364                event_trace = 0;
 365
 366        filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
 367        filter_str_len = filter ? strlen(filter) : 0;
 368
 369        if (!session->evlist->core.nr_mmaps)
 370                return -EINVAL;
 371
 372        pc = session->evlist->mmap[0].core.base;
 373        if (pc) {
 374                err = perf_read_tsc_conversion(pc, &tc);
 375                if (err) {
 376                        if (err != -EOPNOTSUPP)
 377                                return err;
 378                } else {
 379                        cap_user_time_zero = tc.time_mult != 0;
 380                }
 381                if (!cap_user_time_zero)
 382                        ui__warning("Intel Processor Trace: TSC not available\n");
 383        }
 384
 385        per_cpu_mmaps = !perf_cpu_map__empty(session->evlist->core.user_requested_cpus);
 386
 387        auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
 388        auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
 389        auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
 390        auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
 391        auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
 392        auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
 393        auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
 394        auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
 395        auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
 396        auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
 397        auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
 398        auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
 399        auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
 400        auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
 401        auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
 402        auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
 403        auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
 404        auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
 405
 406        info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
 407
 408        if (filter_str_len) {
 409                size_t len = intel_pt_filter_bytes(filter);
 410
 411                strncpy((char *)info, filter, len);
 412                info += len >> 3;
 413        }
 414
 415        *info++ = event_trace;
 416
 417        return 0;
 418}
 419
 420static int intel_pt_track_switches(struct evlist *evlist)
 421{
 422        const char *sched_switch = "sched:sched_switch";
 423        struct evsel *evsel;
 424        int err;
 425
 426        if (!evlist__can_select_event(evlist, sched_switch))
 427                return -EPERM;
 428
 429        err = parse_events(evlist, sched_switch, NULL);
 430        if (err) {
 431                pr_debug2("%s: failed to parse %s, error %d\n",
 432                          __func__, sched_switch, err);
 433                return err;
 434        }
 435
 436        evsel = evlist__last(evlist);
 437
 438        evsel__set_sample_bit(evsel, CPU);
 439        evsel__set_sample_bit(evsel, TIME);
 440
 441        evsel->core.system_wide = true;
 442        evsel->no_aux_samples = true;
 443        evsel->immediate = true;
 444
 445        return 0;
 446}
 447
 448static void intel_pt_valid_str(char *str, size_t len, u64 valid)
 449{
 450        unsigned int val, last = 0, state = 1;
 451        int p = 0;
 452
 453        str[0] = '\0';
 454
 455        for (val = 0; val <= 64; val++, valid >>= 1) {
 456                if (valid & 1) {
 457                        last = val;
 458                        switch (state) {
 459                        case 0:
 460                                p += scnprintf(str + p, len - p, ",");
 461                                /* Fall through */
 462                        case 1:
 463                                p += scnprintf(str + p, len - p, "%u", val);
 464                                state = 2;
 465                                break;
 466                        case 2:
 467                                state = 3;
 468                                break;
 469                        case 3:
 470                                state = 4;
 471                                break;
 472                        default:
 473                                break;
 474                        }
 475                } else {
 476                        switch (state) {
 477                        case 3:
 478                                p += scnprintf(str + p, len - p, ",%u", last);
 479                                state = 0;
 480                                break;
 481                        case 4:
 482                                p += scnprintf(str + p, len - p, "-%u", last);
 483                                state = 0;
 484                                break;
 485                        default:
 486                                break;
 487                        }
 488                        if (state != 1)
 489                                state = 0;
 490                }
 491        }
 492}
 493
 494static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
 495                                    const char *caps, const char *name,
 496                                    const char *supported, u64 config)
 497{
 498        char valid_str[256];
 499        unsigned int shift;
 500        unsigned long long valid;
 501        u64 bits;
 502        int ok;
 503
 504        if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
 505                valid = 0;
 506
 507        if (supported &&
 508            perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
 509                valid = 0;
 510
 511        valid |= 1;
 512
 513        bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
 514
 515        config &= bits;
 516
 517        for (shift = 0; bits && !(bits & 1); shift++)
 518                bits >>= 1;
 519
 520        config >>= shift;
 521
 522        if (config > 63)
 523                goto out_err;
 524
 525        if (valid & (1 << config))
 526                return 0;
 527out_err:
 528        intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
 529        pr_err("Invalid %s for %s. Valid values are: %s\n",
 530               name, INTEL_PT_PMU_NAME, valid_str);
 531        return -EINVAL;
 532}
 533
 534static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
 535                                    struct evsel *evsel)
 536{
 537        int err;
 538        char c;
 539
 540        if (!evsel)
 541                return 0;
 542
 543        /*
 544         * If supported, force pass-through config term (pt=1) even if user
 545         * sets pt=0, which avoids senseless kernel errors.
 546         */
 547        if (perf_pmu__scan_file(intel_pt_pmu, "format/pt", "%c", &c) == 1 &&
 548            !(evsel->core.attr.config & 1)) {
 549                pr_warning("pt=0 doesn't make sense, forcing pt=1\n");
 550                evsel->core.attr.config |= 1;
 551        }
 552
 553        err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
 554                                       "cyc_thresh", "caps/psb_cyc",
 555                                       evsel->core.attr.config);
 556        if (err)
 557                return err;
 558
 559        err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
 560                                       "mtc_period", "caps/mtc",
 561                                       evsel->core.attr.config);
 562        if (err)
 563                return err;
 564
 565        return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
 566                                        "psb_period", "caps/psb_cyc",
 567                                        evsel->core.attr.config);
 568}
 569
 570static void intel_pt_config_sample_mode(struct perf_pmu *intel_pt_pmu,
 571                                        struct evsel *evsel)
 572{
 573        u64 user_bits = 0, bits;
 574        struct evsel_config_term *term = evsel__get_config_term(evsel, CFG_CHG);
 575
 576        if (term)
 577                user_bits = term->val.cfg_chg;
 578
 579        bits = perf_pmu__format_bits(&intel_pt_pmu->format, "psb_period");
 580
 581        /* Did user change psb_period */
 582        if (bits & user_bits)
 583                return;
 584
 585        /* Set psb_period to 0 */
 586        evsel->core.attr.config &= ~bits;
 587}
 588
 589static void intel_pt_min_max_sample_sz(struct evlist *evlist,
 590                                       size_t *min_sz, size_t *max_sz)
 591{
 592        struct evsel *evsel;
 593
 594        evlist__for_each_entry(evlist, evsel) {
 595                size_t sz = evsel->core.attr.aux_sample_size;
 596
 597                if (!sz)
 598                        continue;
 599                if (min_sz && (sz < *min_sz || !*min_sz))
 600                        *min_sz = sz;
 601                if (max_sz && sz > *max_sz)
 602                        *max_sz = sz;
 603        }
 604}
 605
 606/*
 607 * Currently, there is not enough information to disambiguate different PEBS
 608 * events, so only allow one.
 609 */
 610static bool intel_pt_too_many_aux_output(struct evlist *evlist)
 611{
 612        struct evsel *evsel;
 613        int aux_output_cnt = 0;
 614
 615        evlist__for_each_entry(evlist, evsel)
 616                aux_output_cnt += !!evsel->core.attr.aux_output;
 617
 618        if (aux_output_cnt > 1) {
 619                pr_err(INTEL_PT_PMU_NAME " supports at most one event with aux-output\n");
 620                return true;
 621        }
 622
 623        return false;
 624}
 625
 626static int intel_pt_recording_options(struct auxtrace_record *itr,
 627                                      struct evlist *evlist,
 628                                      struct record_opts *opts)
 629{
 630        struct intel_pt_recording *ptr =
 631                        container_of(itr, struct intel_pt_recording, itr);
 632        struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 633        bool have_timing_info, need_immediate = false;
 634        struct evsel *evsel, *intel_pt_evsel = NULL;
 635        const struct perf_cpu_map *cpus = evlist->core.user_requested_cpus;
 636        bool privileged = perf_event_paranoid_check(-1);
 637        u64 tsc_bit;
 638        int err;
 639
 640        ptr->evlist = evlist;
 641        ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
 642
 643        evlist__for_each_entry(evlist, evsel) {
 644                if (evsel->core.attr.type == intel_pt_pmu->type) {
 645                        if (intel_pt_evsel) {
 646                                pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
 647                                return -EINVAL;
 648                        }
 649                        evsel->core.attr.freq = 0;
 650                        evsel->core.attr.sample_period = 1;
 651                        evsel->no_aux_samples = true;
 652                        evsel->needs_auxtrace_mmap = true;
 653                        intel_pt_evsel = evsel;
 654                        opts->full_auxtrace = true;
 655                }
 656        }
 657
 658        if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
 659                pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
 660                return -EINVAL;
 661        }
 662
 663        if (opts->auxtrace_snapshot_mode && opts->auxtrace_sample_mode) {
 664                pr_err("Snapshot mode (" INTEL_PT_PMU_NAME " PMU) and sample trace cannot be used together\n");
 665                return -EINVAL;
 666        }
 667
 668        if (opts->use_clockid) {
 669                pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
 670                return -EINVAL;
 671        }
 672
 673        if (intel_pt_too_many_aux_output(evlist))
 674                return -EINVAL;
 675
 676        if (!opts->full_auxtrace)
 677                return 0;
 678
 679        if (opts->auxtrace_sample_mode)
 680                intel_pt_config_sample_mode(intel_pt_pmu, intel_pt_evsel);
 681
 682        err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
 683        if (err)
 684                return err;
 685
 686        /* Set default sizes for snapshot mode */
 687        if (opts->auxtrace_snapshot_mode) {
 688                size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 689
 690                if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
 691                        if (privileged) {
 692                                opts->auxtrace_mmap_pages = MiB(4) / page_size;
 693                        } else {
 694                                opts->auxtrace_mmap_pages = KiB(128) / page_size;
 695                                if (opts->mmap_pages == UINT_MAX)
 696                                        opts->mmap_pages = KiB(256) / page_size;
 697                        }
 698                } else if (!opts->auxtrace_mmap_pages && !privileged &&
 699                           opts->mmap_pages == UINT_MAX) {
 700                        opts->mmap_pages = KiB(256) / page_size;
 701                }
 702                if (!opts->auxtrace_snapshot_size)
 703                        opts->auxtrace_snapshot_size =
 704                                opts->auxtrace_mmap_pages * (size_t)page_size;
 705                if (!opts->auxtrace_mmap_pages) {
 706                        size_t sz = opts->auxtrace_snapshot_size;
 707
 708                        sz = round_up(sz, page_size) / page_size;
 709                        opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 710                }
 711                if (opts->auxtrace_snapshot_size >
 712                                opts->auxtrace_mmap_pages * (size_t)page_size) {
 713                        pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
 714                               opts->auxtrace_snapshot_size,
 715                               opts->auxtrace_mmap_pages * (size_t)page_size);
 716                        return -EINVAL;
 717                }
 718                if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
 719                        pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
 720                        return -EINVAL;
 721                }
 722                pr_debug2("Intel PT snapshot size: %zu\n",
 723                          opts->auxtrace_snapshot_size);
 724                if (psb_period &&
 725                    opts->auxtrace_snapshot_size <= psb_period +
 726                                                  INTEL_PT_PSB_PERIOD_NEAR)
 727                        ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
 728                                    opts->auxtrace_snapshot_size, psb_period);
 729        }
 730
 731        /* Set default sizes for sample mode */
 732        if (opts->auxtrace_sample_mode) {
 733                size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 734                size_t min_sz = 0, max_sz = 0;
 735
 736                intel_pt_min_max_sample_sz(evlist, &min_sz, &max_sz);
 737                if (!opts->auxtrace_mmap_pages && !privileged &&
 738                    opts->mmap_pages == UINT_MAX)
 739                        opts->mmap_pages = KiB(256) / page_size;
 740                if (!opts->auxtrace_mmap_pages) {
 741                        size_t sz = round_up(max_sz, page_size) / page_size;
 742
 743                        opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 744                }
 745                if (max_sz > opts->auxtrace_mmap_pages * (size_t)page_size) {
 746                        pr_err("Sample size %zu must not be greater than AUX area tracing mmap size %zu\n",
 747                               max_sz,
 748                               opts->auxtrace_mmap_pages * (size_t)page_size);
 749                        return -EINVAL;
 750                }
 751                pr_debug2("Intel PT min. sample size: %zu max. sample size: %zu\n",
 752                          min_sz, max_sz);
 753                if (psb_period &&
 754                    min_sz <= psb_period + INTEL_PT_PSB_PERIOD_NEAR)
 755                        ui__warning("Intel PT sample size (%zu) may be too small for PSB period (%zu)\n",
 756                                    min_sz, psb_period);
 757        }
 758
 759        /* Set default sizes for full trace mode */
 760        if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
 761                if (privileged) {
 762                        opts->auxtrace_mmap_pages = MiB(4) / page_size;
 763                } else {
 764                        opts->auxtrace_mmap_pages = KiB(128) / page_size;
 765                        if (opts->mmap_pages == UINT_MAX)
 766                                opts->mmap_pages = KiB(256) / page_size;
 767                }
 768        }
 769
 770        /* Validate auxtrace_mmap_pages */
 771        if (opts->auxtrace_mmap_pages) {
 772                size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 773                size_t min_sz;
 774
 775                if (opts->auxtrace_snapshot_mode || opts->auxtrace_sample_mode)
 776                        min_sz = KiB(4);
 777                else
 778                        min_sz = KiB(8);
 779
 780                if (sz < min_sz || !is_power_of_2(sz)) {
 781                        pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
 782                               min_sz / 1024);
 783                        return -EINVAL;
 784                }
 785        }
 786
 787        if (!opts->auxtrace_snapshot_mode && !opts->auxtrace_sample_mode) {
 788                u32 aux_watermark = opts->auxtrace_mmap_pages * page_size / 4;
 789
 790                intel_pt_evsel->core.attr.aux_watermark = aux_watermark;
 791        }
 792
 793        intel_pt_parse_terms(intel_pt_pmu->name, &intel_pt_pmu->format,
 794                             "tsc", &tsc_bit);
 795
 796        if (opts->full_auxtrace && (intel_pt_evsel->core.attr.config & tsc_bit))
 797                have_timing_info = true;
 798        else
 799                have_timing_info = false;
 800
 801        /*
 802         * Per-cpu recording needs sched_switch events to distinguish different
 803         * threads.
 804         */
 805        if (have_timing_info && !perf_cpu_map__empty(cpus) &&
 806            !record_opts__no_switch_events(opts)) {
 807                if (perf_can_record_switch_events()) {
 808                        bool cpu_wide = !target__none(&opts->target) &&
 809                                        !target__has_task(&opts->target);
 810
 811                        if (!cpu_wide && perf_can_record_cpu_wide()) {
 812                                struct evsel *switch_evsel;
 813
 814                                switch_evsel = evlist__add_dummy_on_all_cpus(evlist);
 815                                if (!switch_evsel)
 816                                        return -ENOMEM;
 817
 818                                switch_evsel->core.attr.context_switch = 1;
 819                                switch_evsel->immediate = true;
 820
 821                                evsel__set_sample_bit(switch_evsel, TID);
 822                                evsel__set_sample_bit(switch_evsel, TIME);
 823                                evsel__set_sample_bit(switch_evsel, CPU);
 824                                evsel__reset_sample_bit(switch_evsel, BRANCH_STACK);
 825
 826                                opts->record_switch_events = false;
 827                                ptr->have_sched_switch = 3;
 828                        } else {
 829                                opts->record_switch_events = true;
 830                                need_immediate = true;
 831                                if (cpu_wide)
 832                                        ptr->have_sched_switch = 3;
 833                                else
 834                                        ptr->have_sched_switch = 2;
 835                        }
 836                } else {
 837                        err = intel_pt_track_switches(evlist);
 838                        if (err == -EPERM)
 839                                pr_debug2("Unable to select sched:sched_switch\n");
 840                        else if (err)
 841                                return err;
 842                        else
 843                                ptr->have_sched_switch = 1;
 844                }
 845        }
 846
 847        if (have_timing_info && !intel_pt_evsel->core.attr.exclude_kernel &&
 848            perf_can_record_text_poke_events() && perf_can_record_cpu_wide())
 849                opts->text_poke = true;
 850
 851        if (intel_pt_evsel) {
 852                /*
 853                 * To obtain the auxtrace buffer file descriptor, the auxtrace
 854                 * event must come first.
 855                 */
 856                evlist__to_front(evlist, intel_pt_evsel);
 857                /*
 858                 * In the case of per-cpu mmaps, we need the CPU on the
 859                 * AUX event.
 860                 */
 861                if (!perf_cpu_map__empty(cpus))
 862                        evsel__set_sample_bit(intel_pt_evsel, CPU);
 863        }
 864
 865        /* Add dummy event to keep tracking */
 866        if (opts->full_auxtrace) {
 867                bool need_system_wide_tracking;
 868                struct evsel *tracking_evsel;
 869
 870                /*
 871                 * User space tasks can migrate between CPUs, so when tracing
 872                 * selected CPUs, sideband for all CPUs is still needed.
 873                 */
 874                need_system_wide_tracking = evlist->core.has_user_cpus &&
 875                                            !intel_pt_evsel->core.attr.exclude_user;
 876
 877                tracking_evsel = evlist__add_aux_dummy(evlist, need_system_wide_tracking);
 878                if (!tracking_evsel)
 879                        return -ENOMEM;
 880
 881                evlist__set_tracking_event(evlist, tracking_evsel);
 882
 883                if (need_immediate)
 884                        tracking_evsel->immediate = true;
 885
 886                /* In per-cpu case, always need the time of mmap events etc */
 887                if (!perf_cpu_map__empty(cpus)) {
 888                        evsel__set_sample_bit(tracking_evsel, TIME);
 889                        /* And the CPU for switch events */
 890                        evsel__set_sample_bit(tracking_evsel, CPU);
 891                }
 892                evsel__reset_sample_bit(tracking_evsel, BRANCH_STACK);
 893        }
 894
 895        /*
 896         * Warn the user when we do not have enough information to decode i.e.
 897         * per-cpu with no sched_switch (except workload-only).
 898         */
 899        if (!ptr->have_sched_switch && !perf_cpu_map__empty(cpus) &&
 900            !target__none(&opts->target) &&
 901            !intel_pt_evsel->core.attr.exclude_user)
 902                ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
 903
 904        return 0;
 905}
 906
 907static int intel_pt_snapshot_start(struct auxtrace_record *itr)
 908{
 909        struct intel_pt_recording *ptr =
 910                        container_of(itr, struct intel_pt_recording, itr);
 911        struct evsel *evsel;
 912
 913        evlist__for_each_entry(ptr->evlist, evsel) {
 914                if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
 915                        return evsel__disable(evsel);
 916        }
 917        return -EINVAL;
 918}
 919
 920static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
 921{
 922        struct intel_pt_recording *ptr =
 923                        container_of(itr, struct intel_pt_recording, itr);
 924        struct evsel *evsel;
 925
 926        evlist__for_each_entry(ptr->evlist, evsel) {
 927                if (evsel->core.attr.type == ptr->intel_pt_pmu->type)
 928                        return evsel__enable(evsel);
 929        }
 930        return -EINVAL;
 931}
 932
 933static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
 934{
 935        const size_t sz = sizeof(struct intel_pt_snapshot_ref);
 936        int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
 937        struct intel_pt_snapshot_ref *refs;
 938
 939        if (!new_cnt)
 940                new_cnt = 16;
 941
 942        while (new_cnt <= idx)
 943                new_cnt *= 2;
 944
 945        refs = calloc(new_cnt, sz);
 946        if (!refs)
 947                return -ENOMEM;
 948
 949        memcpy(refs, ptr->snapshot_refs, cnt * sz);
 950
 951        ptr->snapshot_refs = refs;
 952        ptr->snapshot_ref_cnt = new_cnt;
 953
 954        return 0;
 955}
 956
 957static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
 958{
 959        int i;
 960
 961        for (i = 0; i < ptr->snapshot_ref_cnt; i++)
 962                zfree(&ptr->snapshot_refs[i].ref_buf);
 963        zfree(&ptr->snapshot_refs);
 964}
 965
 966static void intel_pt_recording_free(struct auxtrace_record *itr)
 967{
 968        struct intel_pt_recording *ptr =
 969                        container_of(itr, struct intel_pt_recording, itr);
 970
 971        intel_pt_free_snapshot_refs(ptr);
 972        free(ptr);
 973}
 974
 975static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
 976                                       size_t snapshot_buf_size)
 977{
 978        size_t ref_buf_size = ptr->snapshot_ref_buf_size;
 979        void *ref_buf;
 980
 981        ref_buf = zalloc(ref_buf_size);
 982        if (!ref_buf)
 983                return -ENOMEM;
 984
 985        ptr->snapshot_refs[idx].ref_buf = ref_buf;
 986        ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
 987
 988        return 0;
 989}
 990
 991static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
 992                                             size_t snapshot_buf_size)
 993{
 994        const size_t max_size = 256 * 1024;
 995        size_t buf_size = 0, psb_period;
 996
 997        if (ptr->snapshot_size <= 64 * 1024)
 998                return 0;
 999
1000        psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
1001        if (psb_period)
1002                buf_size = psb_period * 2;
1003
1004        if (!buf_size || buf_size > max_size)
1005                buf_size = max_size;
1006
1007        if (buf_size >= snapshot_buf_size)
1008                return 0;
1009
1010        if (buf_size >= ptr->snapshot_size / 2)
1011                return 0;
1012
1013        return buf_size;
1014}
1015
1016static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
1017                                  size_t snapshot_buf_size)
1018{
1019        if (ptr->snapshot_init_done)
1020                return 0;
1021
1022        ptr->snapshot_init_done = true;
1023
1024        ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
1025                                                        snapshot_buf_size);
1026
1027        return 0;
1028}
1029
1030/**
1031 * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
1032 * @buf1: first buffer
1033 * @compare_size: number of bytes to compare
1034 * @buf2: second buffer (a circular buffer)
1035 * @offs2: offset in second buffer
1036 * @buf2_size: size of second buffer
1037 *
1038 * The comparison allows for the possibility that the bytes to compare in the
1039 * circular buffer are not contiguous.  It is assumed that @compare_size <=
1040 * @buf2_size.  This function returns %false if the bytes are identical, %true
1041 * otherwise.
1042 */
1043static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
1044                                     void *buf2, size_t offs2, size_t buf2_size)
1045{
1046        size_t end2 = offs2 + compare_size, part_size;
1047
1048        if (end2 <= buf2_size)
1049                return memcmp(buf1, buf2 + offs2, compare_size);
1050
1051        part_size = end2 - buf2_size;
1052        if (memcmp(buf1, buf2 + offs2, part_size))
1053                return true;
1054
1055        compare_size -= part_size;
1056
1057        return memcmp(buf1 + part_size, buf2, compare_size);
1058}
1059
1060static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
1061                                 size_t ref_size, size_t buf_size,
1062                                 void *data, size_t head)
1063{
1064        size_t ref_end = ref_offset + ref_size;
1065
1066        if (ref_end > buf_size) {
1067                if (head > ref_offset || head < ref_end - buf_size)
1068                        return true;
1069        } else if (head > ref_offset && head < ref_end) {
1070                return true;
1071        }
1072
1073        return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
1074                                        buf_size);
1075}
1076
1077static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
1078                              void *data, size_t head)
1079{
1080        if (head >= ref_size) {
1081                memcpy(ref_buf, data + head - ref_size, ref_size);
1082        } else {
1083                memcpy(ref_buf, data, head);
1084                ref_size -= head;
1085                memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
1086        }
1087}
1088
1089static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
1090                             struct auxtrace_mmap *mm, unsigned char *data,
1091                             u64 head)
1092{
1093        struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
1094        bool wrapped;
1095
1096        wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
1097                                       ptr->snapshot_ref_buf_size, mm->len,
1098                                       data, head);
1099
1100        intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
1101                          data, head);
1102
1103        return wrapped;
1104}
1105
1106static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
1107{
1108        int i, a, b;
1109
1110        b = buf_size >> 3;
1111        a = b - 512;
1112        if (a < 0)
1113                a = 0;
1114
1115        for (i = a; i < b; i++) {
1116                if (data[i])
1117                        return true;
1118        }
1119
1120        return false;
1121}
1122
1123static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
1124                                  struct auxtrace_mmap *mm, unsigned char *data,
1125                                  u64 *head, u64 *old)
1126{
1127        struct intel_pt_recording *ptr =
1128                        container_of(itr, struct intel_pt_recording, itr);
1129        bool wrapped;
1130        int err;
1131
1132        pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1133                  __func__, idx, (size_t)*old, (size_t)*head);
1134
1135        err = intel_pt_snapshot_init(ptr, mm->len);
1136        if (err)
1137                goto out_err;
1138
1139        if (idx >= ptr->snapshot_ref_cnt) {
1140                err = intel_pt_alloc_snapshot_refs(ptr, idx);
1141                if (err)
1142                        goto out_err;
1143        }
1144
1145        if (ptr->snapshot_ref_buf_size) {
1146                if (!ptr->snapshot_refs[idx].ref_buf) {
1147                        err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1148                        if (err)
1149                                goto out_err;
1150                }
1151                wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1152        } else {
1153                wrapped = ptr->snapshot_refs[idx].wrapped;
1154                if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1155                        ptr->snapshot_refs[idx].wrapped = true;
1156                        wrapped = true;
1157                }
1158        }
1159
1160        /*
1161         * In full trace mode 'head' continually increases.  However in snapshot
1162         * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1163         * are adjusted to match the full trace case which expects that 'old' is
1164         * always less than 'head'.
1165         */
1166        if (wrapped) {
1167                *old = *head;
1168                *head += mm->len;
1169        } else {
1170                if (mm->mask)
1171                        *old &= mm->mask;
1172                else
1173                        *old %= mm->len;
1174                if (*old > *head)
1175                        *head += mm->len;
1176        }
1177
1178        pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1179                  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1180
1181        return 0;
1182
1183out_err:
1184        pr_err("%s: failed, error %d\n", __func__, err);
1185        return err;
1186}
1187
1188static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1189{
1190        return rdtsc();
1191}
1192
1193struct auxtrace_record *intel_pt_recording_init(int *err)
1194{
1195        struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1196        struct intel_pt_recording *ptr;
1197
1198        if (!intel_pt_pmu)
1199                return NULL;
1200
1201        if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1202                *err = -errno;
1203                return NULL;
1204        }
1205
1206        ptr = zalloc(sizeof(struct intel_pt_recording));
1207        if (!ptr) {
1208                *err = -ENOMEM;
1209                return NULL;
1210        }
1211
1212        ptr->intel_pt_pmu = intel_pt_pmu;
1213        ptr->itr.pmu = intel_pt_pmu;
1214        ptr->itr.recording_options = intel_pt_recording_options;
1215        ptr->itr.info_priv_size = intel_pt_info_priv_size;
1216        ptr->itr.info_fill = intel_pt_info_fill;
1217        ptr->itr.free = intel_pt_recording_free;
1218        ptr->itr.snapshot_start = intel_pt_snapshot_start;
1219        ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1220        ptr->itr.find_snapshot = intel_pt_find_snapshot;
1221        ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1222        ptr->itr.reference = intel_pt_reference;
1223        ptr->itr.read_finish = auxtrace_record__read_finish;
1224        /*
1225         * Decoding starts at a PSB packet. Minimum PSB period is 2K so 4K
1226         * should give at least 1 PSB per sample.
1227         */
1228        ptr->itr.default_aux_sample_size = 4096;
1229        return &ptr->itr;
1230}
1231