linux/tools/perf/arch/x86/util/intel-pt.c
<<
>>
Prefs
   1/*
   2 * intel_pt.c: Intel Processor Trace support
   3 * Copyright (c) 2013-2015, Intel Corporation.
   4 *
   5 * This program is free software; you can redistribute it and/or modify it
   6 * under the terms and conditions of the GNU General Public License,
   7 * version 2, as published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope it will be useful, but WITHOUT
  10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12 * more details.
  13 *
  14 */
  15
  16#include <errno.h>
  17#include <stdbool.h>
  18#include <linux/kernel.h>
  19#include <linux/types.h>
  20#include <linux/bitops.h>
  21#include <linux/log2.h>
  22#include <cpuid.h>
  23
  24#include "../../perf.h"
  25#include "../../util/session.h"
  26#include "../../util/event.h"
  27#include "../../util/evlist.h"
  28#include "../../util/evsel.h"
  29#include "../../util/cpumap.h"
  30#include <subcmd/parse-options.h>
  31#include "../../util/parse-events.h"
  32#include "../../util/pmu.h"
  33#include "../../util/debug.h"
  34#include "../../util/auxtrace.h"
  35#include "../../util/tsc.h"
  36#include "../../util/intel-pt.h"
  37
  38#define KiB(x) ((x) * 1024)
  39#define MiB(x) ((x) * 1024 * 1024)
  40#define KiB_MASK(x) (KiB(x) - 1)
  41#define MiB_MASK(x) (MiB(x) - 1)
  42
  43#define INTEL_PT_PSB_PERIOD_NEAR        256
  44
  45struct intel_pt_snapshot_ref {
  46        void *ref_buf;
  47        size_t ref_offset;
  48        bool wrapped;
  49};
  50
  51struct intel_pt_recording {
  52        struct auxtrace_record          itr;
  53        struct perf_pmu                 *intel_pt_pmu;
  54        int                             have_sched_switch;
  55        struct perf_evlist              *evlist;
  56        bool                            snapshot_mode;
  57        bool                            snapshot_init_done;
  58        size_t                          snapshot_size;
  59        size_t                          snapshot_ref_buf_size;
  60        int                             snapshot_ref_cnt;
  61        struct intel_pt_snapshot_ref    *snapshot_refs;
  62        size_t                          priv_size;
  63};
  64
  65static int intel_pt_parse_terms_with_default(struct list_head *formats,
  66                                             const char *str,
  67                                             u64 *config)
  68{
  69        struct list_head *terms;
  70        struct perf_event_attr attr = { .size = 0, };
  71        int err;
  72
  73        terms = malloc(sizeof(struct list_head));
  74        if (!terms)
  75                return -ENOMEM;
  76
  77        INIT_LIST_HEAD(terms);
  78
  79        err = parse_events_terms(terms, str);
  80        if (err)
  81                goto out_free;
  82
  83        attr.config = *config;
  84        err = perf_pmu__config_terms(formats, &attr, terms, true, NULL);
  85        if (err)
  86                goto out_free;
  87
  88        *config = attr.config;
  89out_free:
  90        parse_events_terms__delete(terms);
  91        return err;
  92}
  93
  94static int intel_pt_parse_terms(struct list_head *formats, const char *str,
  95                                u64 *config)
  96{
  97        *config = 0;
  98        return intel_pt_parse_terms_with_default(formats, str, 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 perf_evlist *evlist, u64 *res)
 122{
 123        struct perf_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->attr.type == intel_pt_pmu->type) {
 134                        *res = intel_pt_masked_bits(mask, evsel->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 perf_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->format, buf, &config);
 236
 237        return config;
 238}
 239
 240static int intel_pt_parse_snapshot_options(struct auxtrace_record *itr,
 241                                           struct record_opts *opts,
 242                                           const char *str)
 243{
 244        struct intel_pt_recording *ptr =
 245                        container_of(itr, struct intel_pt_recording, itr);
 246        unsigned long long snapshot_size = 0;
 247        char *endptr;
 248
 249        if (str) {
 250                snapshot_size = strtoull(str, &endptr, 0);
 251                if (*endptr || snapshot_size > SIZE_MAX)
 252                        return -1;
 253        }
 254
 255        opts->auxtrace_snapshot_mode = true;
 256        opts->auxtrace_snapshot_size = snapshot_size;
 257
 258        ptr->snapshot_size = snapshot_size;
 259
 260        return 0;
 261}
 262
 263struct perf_event_attr *
 264intel_pt_pmu_default_config(struct perf_pmu *intel_pt_pmu)
 265{
 266        struct perf_event_attr *attr;
 267
 268        attr = zalloc(sizeof(struct perf_event_attr));
 269        if (!attr)
 270                return NULL;
 271
 272        attr->config = intel_pt_default_config(intel_pt_pmu);
 273
 274        intel_pt_pmu->selectable = true;
 275
 276        return attr;
 277}
 278
 279static const char *intel_pt_find_filter(struct perf_evlist *evlist,
 280                                        struct perf_pmu *intel_pt_pmu)
 281{
 282        struct perf_evsel *evsel;
 283
 284        evlist__for_each_entry(evlist, evsel) {
 285                if (evsel->attr.type == intel_pt_pmu->type)
 286                        return evsel->filter;
 287        }
 288
 289        return NULL;
 290}
 291
 292static size_t intel_pt_filter_bytes(const char *filter)
 293{
 294        size_t len = filter ? strlen(filter) : 0;
 295
 296        return len ? roundup(len + 1, 8) : 0;
 297}
 298
 299static size_t
 300intel_pt_info_priv_size(struct auxtrace_record *itr, struct perf_evlist *evlist)
 301{
 302        struct intel_pt_recording *ptr =
 303                        container_of(itr, struct intel_pt_recording, itr);
 304        const char *filter = intel_pt_find_filter(evlist, ptr->intel_pt_pmu);
 305
 306        ptr->priv_size = (INTEL_PT_AUXTRACE_PRIV_MAX * sizeof(u64)) +
 307                         intel_pt_filter_bytes(filter);
 308
 309        return ptr->priv_size;
 310}
 311
 312static void intel_pt_tsc_ctc_ratio(u32 *n, u32 *d)
 313{
 314        unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
 315
 316        __get_cpuid(0x15, &eax, &ebx, &ecx, &edx);
 317        *n = ebx;
 318        *d = eax;
 319}
 320
 321static int intel_pt_info_fill(struct auxtrace_record *itr,
 322                              struct perf_session *session,
 323                              struct auxtrace_info_event *auxtrace_info,
 324                              size_t priv_size)
 325{
 326        struct intel_pt_recording *ptr =
 327                        container_of(itr, struct intel_pt_recording, itr);
 328        struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 329        struct perf_event_mmap_page *pc;
 330        struct perf_tsc_conversion tc = { .time_mult = 0, };
 331        bool cap_user_time_zero = false, per_cpu_mmaps;
 332        u64 tsc_bit, mtc_bit, mtc_freq_bits, cyc_bit, noretcomp_bit;
 333        u32 tsc_ctc_ratio_n, tsc_ctc_ratio_d;
 334        unsigned long max_non_turbo_ratio;
 335        size_t filter_str_len;
 336        const char *filter;
 337        u64 *info;
 338        int err;
 339
 340        if (priv_size != ptr->priv_size)
 341                return -EINVAL;
 342
 343        intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
 344        intel_pt_parse_terms(&intel_pt_pmu->format, "noretcomp",
 345                             &noretcomp_bit);
 346        intel_pt_parse_terms(&intel_pt_pmu->format, "mtc", &mtc_bit);
 347        mtc_freq_bits = perf_pmu__format_bits(&intel_pt_pmu->format,
 348                                              "mtc_period");
 349        intel_pt_parse_terms(&intel_pt_pmu->format, "cyc", &cyc_bit);
 350
 351        intel_pt_tsc_ctc_ratio(&tsc_ctc_ratio_n, &tsc_ctc_ratio_d);
 352
 353        if (perf_pmu__scan_file(intel_pt_pmu, "max_nonturbo_ratio",
 354                                "%lu", &max_non_turbo_ratio) != 1)
 355                max_non_turbo_ratio = 0;
 356
 357        filter = intel_pt_find_filter(session->evlist, ptr->intel_pt_pmu);
 358        filter_str_len = filter ? strlen(filter) : 0;
 359
 360        if (!session->evlist->nr_mmaps)
 361                return -EINVAL;
 362
 363        pc = session->evlist->mmap[0].base;
 364        if (pc) {
 365                err = perf_read_tsc_conversion(pc, &tc);
 366                if (err) {
 367                        if (err != -EOPNOTSUPP)
 368                                return err;
 369                } else {
 370                        cap_user_time_zero = tc.time_mult != 0;
 371                }
 372                if (!cap_user_time_zero)
 373                        ui__warning("Intel Processor Trace: TSC not available\n");
 374        }
 375
 376        per_cpu_mmaps = !cpu_map__empty(session->evlist->cpus);
 377
 378        auxtrace_info->type = PERF_AUXTRACE_INTEL_PT;
 379        auxtrace_info->priv[INTEL_PT_PMU_TYPE] = intel_pt_pmu->type;
 380        auxtrace_info->priv[INTEL_PT_TIME_SHIFT] = tc.time_shift;
 381        auxtrace_info->priv[INTEL_PT_TIME_MULT] = tc.time_mult;
 382        auxtrace_info->priv[INTEL_PT_TIME_ZERO] = tc.time_zero;
 383        auxtrace_info->priv[INTEL_PT_CAP_USER_TIME_ZERO] = cap_user_time_zero;
 384        auxtrace_info->priv[INTEL_PT_TSC_BIT] = tsc_bit;
 385        auxtrace_info->priv[INTEL_PT_NORETCOMP_BIT] = noretcomp_bit;
 386        auxtrace_info->priv[INTEL_PT_HAVE_SCHED_SWITCH] = ptr->have_sched_switch;
 387        auxtrace_info->priv[INTEL_PT_SNAPSHOT_MODE] = ptr->snapshot_mode;
 388        auxtrace_info->priv[INTEL_PT_PER_CPU_MMAPS] = per_cpu_mmaps;
 389        auxtrace_info->priv[INTEL_PT_MTC_BIT] = mtc_bit;
 390        auxtrace_info->priv[INTEL_PT_MTC_FREQ_BITS] = mtc_freq_bits;
 391        auxtrace_info->priv[INTEL_PT_TSC_CTC_N] = tsc_ctc_ratio_n;
 392        auxtrace_info->priv[INTEL_PT_TSC_CTC_D] = tsc_ctc_ratio_d;
 393        auxtrace_info->priv[INTEL_PT_CYC_BIT] = cyc_bit;
 394        auxtrace_info->priv[INTEL_PT_MAX_NONTURBO_RATIO] = max_non_turbo_ratio;
 395        auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] = filter_str_len;
 396
 397        info = &auxtrace_info->priv[INTEL_PT_FILTER_STR_LEN] + 1;
 398
 399        if (filter_str_len) {
 400                size_t len = intel_pt_filter_bytes(filter);
 401
 402                strncpy((char *)info, filter, len);
 403                info += len >> 3;
 404        }
 405
 406        return 0;
 407}
 408
 409static int intel_pt_track_switches(struct perf_evlist *evlist)
 410{
 411        const char *sched_switch = "sched:sched_switch";
 412        struct perf_evsel *evsel;
 413        int err;
 414
 415        if (!perf_evlist__can_select_event(evlist, sched_switch))
 416                return -EPERM;
 417
 418        err = parse_events(evlist, sched_switch, NULL);
 419        if (err) {
 420                pr_debug2("%s: failed to parse %s, error %d\n",
 421                          __func__, sched_switch, err);
 422                return err;
 423        }
 424
 425        evsel = perf_evlist__last(evlist);
 426
 427        perf_evsel__set_sample_bit(evsel, CPU);
 428        perf_evsel__set_sample_bit(evsel, TIME);
 429
 430        evsel->system_wide = true;
 431        evsel->no_aux_samples = true;
 432        evsel->immediate = true;
 433
 434        return 0;
 435}
 436
 437static void intel_pt_valid_str(char *str, size_t len, u64 valid)
 438{
 439        unsigned int val, last = 0, state = 1;
 440        int p = 0;
 441
 442        str[0] = '\0';
 443
 444        for (val = 0; val <= 64; val++, valid >>= 1) {
 445                if (valid & 1) {
 446                        last = val;
 447                        switch (state) {
 448                        case 0:
 449                                p += scnprintf(str + p, len - p, ",");
 450                                /* Fall through */
 451                        case 1:
 452                                p += scnprintf(str + p, len - p, "%u", val);
 453                                state = 2;
 454                                break;
 455                        case 2:
 456                                state = 3;
 457                                break;
 458                        case 3:
 459                                state = 4;
 460                                break;
 461                        default:
 462                                break;
 463                        }
 464                } else {
 465                        switch (state) {
 466                        case 3:
 467                                p += scnprintf(str + p, len - p, ",%u", last);
 468                                state = 0;
 469                                break;
 470                        case 4:
 471                                p += scnprintf(str + p, len - p, "-%u", last);
 472                                state = 0;
 473                                break;
 474                        default:
 475                                break;
 476                        }
 477                        if (state != 1)
 478                                state = 0;
 479                }
 480        }
 481}
 482
 483static int intel_pt_val_config_term(struct perf_pmu *intel_pt_pmu,
 484                                    const char *caps, const char *name,
 485                                    const char *supported, u64 config)
 486{
 487        char valid_str[256];
 488        unsigned int shift;
 489        unsigned long long valid;
 490        u64 bits;
 491        int ok;
 492
 493        if (perf_pmu__scan_file(intel_pt_pmu, caps, "%llx", &valid) != 1)
 494                valid = 0;
 495
 496        if (supported &&
 497            perf_pmu__scan_file(intel_pt_pmu, supported, "%d", &ok) == 1 && !ok)
 498                valid = 0;
 499
 500        valid |= 1;
 501
 502        bits = perf_pmu__format_bits(&intel_pt_pmu->format, name);
 503
 504        config &= bits;
 505
 506        for (shift = 0; bits && !(bits & 1); shift++)
 507                bits >>= 1;
 508
 509        config >>= shift;
 510
 511        if (config > 63)
 512                goto out_err;
 513
 514        if (valid & (1 << config))
 515                return 0;
 516out_err:
 517        intel_pt_valid_str(valid_str, sizeof(valid_str), valid);
 518        pr_err("Invalid %s for %s. Valid values are: %s\n",
 519               name, INTEL_PT_PMU_NAME, valid_str);
 520        return -EINVAL;
 521}
 522
 523static int intel_pt_validate_config(struct perf_pmu *intel_pt_pmu,
 524                                    struct perf_evsel *evsel)
 525{
 526        int err;
 527
 528        if (!evsel)
 529                return 0;
 530
 531        err = intel_pt_val_config_term(intel_pt_pmu, "caps/cycle_thresholds",
 532                                       "cyc_thresh", "caps/psb_cyc",
 533                                       evsel->attr.config);
 534        if (err)
 535                return err;
 536
 537        err = intel_pt_val_config_term(intel_pt_pmu, "caps/mtc_periods",
 538                                       "mtc_period", "caps/mtc",
 539                                       evsel->attr.config);
 540        if (err)
 541                return err;
 542
 543        return intel_pt_val_config_term(intel_pt_pmu, "caps/psb_periods",
 544                                        "psb_period", "caps/psb_cyc",
 545                                        evsel->attr.config);
 546}
 547
 548static int intel_pt_recording_options(struct auxtrace_record *itr,
 549                                      struct perf_evlist *evlist,
 550                                      struct record_opts *opts)
 551{
 552        struct intel_pt_recording *ptr =
 553                        container_of(itr, struct intel_pt_recording, itr);
 554        struct perf_pmu *intel_pt_pmu = ptr->intel_pt_pmu;
 555        bool have_timing_info, need_immediate = false;
 556        struct perf_evsel *evsel, *intel_pt_evsel = NULL;
 557        const struct cpu_map *cpus = evlist->cpus;
 558        bool privileged = geteuid() == 0 || perf_event_paranoid() < 0;
 559        u64 tsc_bit;
 560        int err;
 561
 562        ptr->evlist = evlist;
 563        ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
 564
 565        evlist__for_each_entry(evlist, evsel) {
 566                if (evsel->attr.type == intel_pt_pmu->type) {
 567                        if (intel_pt_evsel) {
 568                                pr_err("There may be only one " INTEL_PT_PMU_NAME " event\n");
 569                                return -EINVAL;
 570                        }
 571                        evsel->attr.freq = 0;
 572                        evsel->attr.sample_period = 1;
 573                        intel_pt_evsel = evsel;
 574                        opts->full_auxtrace = true;
 575                }
 576        }
 577
 578        if (opts->auxtrace_snapshot_mode && !opts->full_auxtrace) {
 579                pr_err("Snapshot mode (-S option) requires " INTEL_PT_PMU_NAME " PMU event (-e " INTEL_PT_PMU_NAME ")\n");
 580                return -EINVAL;
 581        }
 582
 583        if (opts->use_clockid) {
 584                pr_err("Cannot use clockid (-k option) with " INTEL_PT_PMU_NAME "\n");
 585                return -EINVAL;
 586        }
 587
 588        if (!opts->full_auxtrace)
 589                return 0;
 590
 591        err = intel_pt_validate_config(intel_pt_pmu, intel_pt_evsel);
 592        if (err)
 593                return err;
 594
 595        /* Set default sizes for snapshot mode */
 596        if (opts->auxtrace_snapshot_mode) {
 597                size_t psb_period = intel_pt_psb_period(intel_pt_pmu, evlist);
 598
 599                if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
 600                        if (privileged) {
 601                                opts->auxtrace_mmap_pages = MiB(4) / page_size;
 602                        } else {
 603                                opts->auxtrace_mmap_pages = KiB(128) / page_size;
 604                                if (opts->mmap_pages == UINT_MAX)
 605                                        opts->mmap_pages = KiB(256) / page_size;
 606                        }
 607                } else if (!opts->auxtrace_mmap_pages && !privileged &&
 608                           opts->mmap_pages == UINT_MAX) {
 609                        opts->mmap_pages = KiB(256) / page_size;
 610                }
 611                if (!opts->auxtrace_snapshot_size)
 612                        opts->auxtrace_snapshot_size =
 613                                opts->auxtrace_mmap_pages * (size_t)page_size;
 614                if (!opts->auxtrace_mmap_pages) {
 615                        size_t sz = opts->auxtrace_snapshot_size;
 616
 617                        sz = round_up(sz, page_size) / page_size;
 618                        opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 619                }
 620                if (opts->auxtrace_snapshot_size >
 621                                opts->auxtrace_mmap_pages * (size_t)page_size) {
 622                        pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
 623                               opts->auxtrace_snapshot_size,
 624                               opts->auxtrace_mmap_pages * (size_t)page_size);
 625                        return -EINVAL;
 626                }
 627                if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
 628                        pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
 629                        return -EINVAL;
 630                }
 631                pr_debug2("Intel PT snapshot size: %zu\n",
 632                          opts->auxtrace_snapshot_size);
 633                if (psb_period &&
 634                    opts->auxtrace_snapshot_size <= psb_period +
 635                                                  INTEL_PT_PSB_PERIOD_NEAR)
 636                        ui__warning("Intel PT snapshot size (%zu) may be too small for PSB period (%zu)\n",
 637                                    opts->auxtrace_snapshot_size, psb_period);
 638        }
 639
 640        /* Set default sizes for full trace mode */
 641        if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
 642                if (privileged) {
 643                        opts->auxtrace_mmap_pages = MiB(4) / page_size;
 644                } else {
 645                        opts->auxtrace_mmap_pages = KiB(128) / page_size;
 646                        if (opts->mmap_pages == UINT_MAX)
 647                                opts->mmap_pages = KiB(256) / page_size;
 648                }
 649        }
 650
 651        /* Validate auxtrace_mmap_pages */
 652        if (opts->auxtrace_mmap_pages) {
 653                size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 654                size_t min_sz;
 655
 656                if (opts->auxtrace_snapshot_mode)
 657                        min_sz = KiB(4);
 658                else
 659                        min_sz = KiB(8);
 660
 661                if (sz < min_sz || !is_power_of_2(sz)) {
 662                        pr_err("Invalid mmap size for Intel Processor Trace: must be at least %zuKiB and a power of 2\n",
 663                               min_sz / 1024);
 664                        return -EINVAL;
 665                }
 666        }
 667
 668        intel_pt_parse_terms(&intel_pt_pmu->format, "tsc", &tsc_bit);
 669
 670        if (opts->full_auxtrace && (intel_pt_evsel->attr.config & tsc_bit))
 671                have_timing_info = true;
 672        else
 673                have_timing_info = false;
 674
 675        /*
 676         * Per-cpu recording needs sched_switch events to distinguish different
 677         * threads.
 678         */
 679        if (have_timing_info && !cpu_map__empty(cpus)) {
 680                if (perf_can_record_switch_events()) {
 681                        bool cpu_wide = !target__none(&opts->target) &&
 682                                        !target__has_task(&opts->target);
 683
 684                        if (!cpu_wide && perf_can_record_cpu_wide()) {
 685                                struct perf_evsel *switch_evsel;
 686
 687                                err = parse_events(evlist, "dummy:u", NULL);
 688                                if (err)
 689                                        return err;
 690
 691                                switch_evsel = perf_evlist__last(evlist);
 692
 693                                switch_evsel->attr.freq = 0;
 694                                switch_evsel->attr.sample_period = 1;
 695                                switch_evsel->attr.context_switch = 1;
 696
 697                                switch_evsel->system_wide = true;
 698                                switch_evsel->no_aux_samples = true;
 699                                switch_evsel->immediate = true;
 700
 701                                perf_evsel__set_sample_bit(switch_evsel, TID);
 702                                perf_evsel__set_sample_bit(switch_evsel, TIME);
 703                                perf_evsel__set_sample_bit(switch_evsel, CPU);
 704
 705                                opts->record_switch_events = false;
 706                                ptr->have_sched_switch = 3;
 707                        } else {
 708                                opts->record_switch_events = true;
 709                                need_immediate = true;
 710                                if (cpu_wide)
 711                                        ptr->have_sched_switch = 3;
 712                                else
 713                                        ptr->have_sched_switch = 2;
 714                        }
 715                } else {
 716                        err = intel_pt_track_switches(evlist);
 717                        if (err == -EPERM)
 718                                pr_debug2("Unable to select sched:sched_switch\n");
 719                        else if (err)
 720                                return err;
 721                        else
 722                                ptr->have_sched_switch = 1;
 723                }
 724        }
 725
 726        if (intel_pt_evsel) {
 727                /*
 728                 * To obtain the auxtrace buffer file descriptor, the auxtrace
 729                 * event must come first.
 730                 */
 731                perf_evlist__to_front(evlist, intel_pt_evsel);
 732                /*
 733                 * In the case of per-cpu mmaps, we need the CPU on the
 734                 * AUX event.
 735                 */
 736                if (!cpu_map__empty(cpus))
 737                        perf_evsel__set_sample_bit(intel_pt_evsel, CPU);
 738        }
 739
 740        /* Add dummy event to keep tracking */
 741        if (opts->full_auxtrace) {
 742                struct perf_evsel *tracking_evsel;
 743
 744                err = parse_events(evlist, "dummy:u", NULL);
 745                if (err)
 746                        return err;
 747
 748                tracking_evsel = perf_evlist__last(evlist);
 749
 750                perf_evlist__set_tracking_event(evlist, tracking_evsel);
 751
 752                tracking_evsel->attr.freq = 0;
 753                tracking_evsel->attr.sample_period = 1;
 754
 755                if (need_immediate)
 756                        tracking_evsel->immediate = true;
 757
 758                /* In per-cpu case, always need the time of mmap events etc */
 759                if (!cpu_map__empty(cpus)) {
 760                        perf_evsel__set_sample_bit(tracking_evsel, TIME);
 761                        /* And the CPU for switch events */
 762                        perf_evsel__set_sample_bit(tracking_evsel, CPU);
 763                }
 764        }
 765
 766        /*
 767         * Warn the user when we do not have enough information to decode i.e.
 768         * per-cpu with no sched_switch (except workload-only).
 769         */
 770        if (!ptr->have_sched_switch && !cpu_map__empty(cpus) &&
 771            !target__none(&opts->target))
 772                ui__warning("Intel Processor Trace decoding will not be possible except for kernel tracing!\n");
 773
 774        return 0;
 775}
 776
 777static int intel_pt_snapshot_start(struct auxtrace_record *itr)
 778{
 779        struct intel_pt_recording *ptr =
 780                        container_of(itr, struct intel_pt_recording, itr);
 781        struct perf_evsel *evsel;
 782
 783        evlist__for_each_entry(ptr->evlist, evsel) {
 784                if (evsel->attr.type == ptr->intel_pt_pmu->type)
 785                        return perf_evsel__disable(evsel);
 786        }
 787        return -EINVAL;
 788}
 789
 790static int intel_pt_snapshot_finish(struct auxtrace_record *itr)
 791{
 792        struct intel_pt_recording *ptr =
 793                        container_of(itr, struct intel_pt_recording, itr);
 794        struct perf_evsel *evsel;
 795
 796        evlist__for_each_entry(ptr->evlist, evsel) {
 797                if (evsel->attr.type == ptr->intel_pt_pmu->type)
 798                        return perf_evsel__enable(evsel);
 799        }
 800        return -EINVAL;
 801}
 802
 803static int intel_pt_alloc_snapshot_refs(struct intel_pt_recording *ptr, int idx)
 804{
 805        const size_t sz = sizeof(struct intel_pt_snapshot_ref);
 806        int cnt = ptr->snapshot_ref_cnt, new_cnt = cnt * 2;
 807        struct intel_pt_snapshot_ref *refs;
 808
 809        if (!new_cnt)
 810                new_cnt = 16;
 811
 812        while (new_cnt <= idx)
 813                new_cnt *= 2;
 814
 815        refs = calloc(new_cnt, sz);
 816        if (!refs)
 817                return -ENOMEM;
 818
 819        memcpy(refs, ptr->snapshot_refs, cnt * sz);
 820
 821        ptr->snapshot_refs = refs;
 822        ptr->snapshot_ref_cnt = new_cnt;
 823
 824        return 0;
 825}
 826
 827static void intel_pt_free_snapshot_refs(struct intel_pt_recording *ptr)
 828{
 829        int i;
 830
 831        for (i = 0; i < ptr->snapshot_ref_cnt; i++)
 832                zfree(&ptr->snapshot_refs[i].ref_buf);
 833        zfree(&ptr->snapshot_refs);
 834}
 835
 836static void intel_pt_recording_free(struct auxtrace_record *itr)
 837{
 838        struct intel_pt_recording *ptr =
 839                        container_of(itr, struct intel_pt_recording, itr);
 840
 841        intel_pt_free_snapshot_refs(ptr);
 842        free(ptr);
 843}
 844
 845static int intel_pt_alloc_snapshot_ref(struct intel_pt_recording *ptr, int idx,
 846                                       size_t snapshot_buf_size)
 847{
 848        size_t ref_buf_size = ptr->snapshot_ref_buf_size;
 849        void *ref_buf;
 850
 851        ref_buf = zalloc(ref_buf_size);
 852        if (!ref_buf)
 853                return -ENOMEM;
 854
 855        ptr->snapshot_refs[idx].ref_buf = ref_buf;
 856        ptr->snapshot_refs[idx].ref_offset = snapshot_buf_size - ref_buf_size;
 857
 858        return 0;
 859}
 860
 861static size_t intel_pt_snapshot_ref_buf_size(struct intel_pt_recording *ptr,
 862                                             size_t snapshot_buf_size)
 863{
 864        const size_t max_size = 256 * 1024;
 865        size_t buf_size = 0, psb_period;
 866
 867        if (ptr->snapshot_size <= 64 * 1024)
 868                return 0;
 869
 870        psb_period = intel_pt_psb_period(ptr->intel_pt_pmu, ptr->evlist);
 871        if (psb_period)
 872                buf_size = psb_period * 2;
 873
 874        if (!buf_size || buf_size > max_size)
 875                buf_size = max_size;
 876
 877        if (buf_size >= snapshot_buf_size)
 878                return 0;
 879
 880        if (buf_size >= ptr->snapshot_size / 2)
 881                return 0;
 882
 883        return buf_size;
 884}
 885
 886static int intel_pt_snapshot_init(struct intel_pt_recording *ptr,
 887                                  size_t snapshot_buf_size)
 888{
 889        if (ptr->snapshot_init_done)
 890                return 0;
 891
 892        ptr->snapshot_init_done = true;
 893
 894        ptr->snapshot_ref_buf_size = intel_pt_snapshot_ref_buf_size(ptr,
 895                                                        snapshot_buf_size);
 896
 897        return 0;
 898}
 899
 900/**
 901 * intel_pt_compare_buffers - compare bytes in a buffer to a circular buffer.
 902 * @buf1: first buffer
 903 * @compare_size: number of bytes to compare
 904 * @buf2: second buffer (a circular buffer)
 905 * @offs2: offset in second buffer
 906 * @buf2_size: size of second buffer
 907 *
 908 * The comparison allows for the possibility that the bytes to compare in the
 909 * circular buffer are not contiguous.  It is assumed that @compare_size <=
 910 * @buf2_size.  This function returns %false if the bytes are identical, %true
 911 * otherwise.
 912 */
 913static bool intel_pt_compare_buffers(void *buf1, size_t compare_size,
 914                                     void *buf2, size_t offs2, size_t buf2_size)
 915{
 916        size_t end2 = offs2 + compare_size, part_size;
 917
 918        if (end2 <= buf2_size)
 919                return memcmp(buf1, buf2 + offs2, compare_size);
 920
 921        part_size = end2 - buf2_size;
 922        if (memcmp(buf1, buf2 + offs2, part_size))
 923                return true;
 924
 925        compare_size -= part_size;
 926
 927        return memcmp(buf1 + part_size, buf2, compare_size);
 928}
 929
 930static bool intel_pt_compare_ref(void *ref_buf, size_t ref_offset,
 931                                 size_t ref_size, size_t buf_size,
 932                                 void *data, size_t head)
 933{
 934        size_t ref_end = ref_offset + ref_size;
 935
 936        if (ref_end > buf_size) {
 937                if (head > ref_offset || head < ref_end - buf_size)
 938                        return true;
 939        } else if (head > ref_offset && head < ref_end) {
 940                return true;
 941        }
 942
 943        return intel_pt_compare_buffers(ref_buf, ref_size, data, ref_offset,
 944                                        buf_size);
 945}
 946
 947static void intel_pt_copy_ref(void *ref_buf, size_t ref_size, size_t buf_size,
 948                              void *data, size_t head)
 949{
 950        if (head >= ref_size) {
 951                memcpy(ref_buf, data + head - ref_size, ref_size);
 952        } else {
 953                memcpy(ref_buf, data, head);
 954                ref_size -= head;
 955                memcpy(ref_buf + head, data + buf_size - ref_size, ref_size);
 956        }
 957}
 958
 959static bool intel_pt_wrapped(struct intel_pt_recording *ptr, int idx,
 960                             struct auxtrace_mmap *mm, unsigned char *data,
 961                             u64 head)
 962{
 963        struct intel_pt_snapshot_ref *ref = &ptr->snapshot_refs[idx];
 964        bool wrapped;
 965
 966        wrapped = intel_pt_compare_ref(ref->ref_buf, ref->ref_offset,
 967                                       ptr->snapshot_ref_buf_size, mm->len,
 968                                       data, head);
 969
 970        intel_pt_copy_ref(ref->ref_buf, ptr->snapshot_ref_buf_size, mm->len,
 971                          data, head);
 972
 973        return wrapped;
 974}
 975
 976static bool intel_pt_first_wrap(u64 *data, size_t buf_size)
 977{
 978        int i, a, b;
 979
 980        b = buf_size >> 3;
 981        a = b - 512;
 982        if (a < 0)
 983                a = 0;
 984
 985        for (i = a; i < b; i++) {
 986                if (data[i])
 987                        return true;
 988        }
 989
 990        return false;
 991}
 992
 993static int intel_pt_find_snapshot(struct auxtrace_record *itr, int idx,
 994                                  struct auxtrace_mmap *mm, unsigned char *data,
 995                                  u64 *head, u64 *old)
 996{
 997        struct intel_pt_recording *ptr =
 998                        container_of(itr, struct intel_pt_recording, itr);
 999        bool wrapped;
1000        int err;
1001
1002        pr_debug3("%s: mmap index %d old head %zu new head %zu\n",
1003                  __func__, idx, (size_t)*old, (size_t)*head);
1004
1005        err = intel_pt_snapshot_init(ptr, mm->len);
1006        if (err)
1007                goto out_err;
1008
1009        if (idx >= ptr->snapshot_ref_cnt) {
1010                err = intel_pt_alloc_snapshot_refs(ptr, idx);
1011                if (err)
1012                        goto out_err;
1013        }
1014
1015        if (ptr->snapshot_ref_buf_size) {
1016                if (!ptr->snapshot_refs[idx].ref_buf) {
1017                        err = intel_pt_alloc_snapshot_ref(ptr, idx, mm->len);
1018                        if (err)
1019                                goto out_err;
1020                }
1021                wrapped = intel_pt_wrapped(ptr, idx, mm, data, *head);
1022        } else {
1023                wrapped = ptr->snapshot_refs[idx].wrapped;
1024                if (!wrapped && intel_pt_first_wrap((u64 *)data, mm->len)) {
1025                        ptr->snapshot_refs[idx].wrapped = true;
1026                        wrapped = true;
1027                }
1028        }
1029
1030        /*
1031         * In full trace mode 'head' continually increases.  However in snapshot
1032         * mode 'head' is an offset within the buffer.  Here 'old' and 'head'
1033         * are adjusted to match the full trace case which expects that 'old' is
1034         * always less than 'head'.
1035         */
1036        if (wrapped) {
1037                *old = *head;
1038                *head += mm->len;
1039        } else {
1040                if (mm->mask)
1041                        *old &= mm->mask;
1042                else
1043                        *old %= mm->len;
1044                if (*old > *head)
1045                        *head += mm->len;
1046        }
1047
1048        pr_debug3("%s: wrap-around %sdetected, adjusted old head %zu adjusted new head %zu\n",
1049                  __func__, wrapped ? "" : "not ", (size_t)*old, (size_t)*head);
1050
1051        return 0;
1052
1053out_err:
1054        pr_err("%s: failed, error %d\n", __func__, err);
1055        return err;
1056}
1057
1058static u64 intel_pt_reference(struct auxtrace_record *itr __maybe_unused)
1059{
1060        return rdtsc();
1061}
1062
1063static int intel_pt_read_finish(struct auxtrace_record *itr, int idx)
1064{
1065        struct intel_pt_recording *ptr =
1066                        container_of(itr, struct intel_pt_recording, itr);
1067        struct perf_evsel *evsel;
1068
1069        evlist__for_each_entry(ptr->evlist, evsel) {
1070                if (evsel->attr.type == ptr->intel_pt_pmu->type)
1071                        return perf_evlist__enable_event_idx(ptr->evlist, evsel,
1072                                                             idx);
1073        }
1074        return -EINVAL;
1075}
1076
1077struct auxtrace_record *intel_pt_recording_init(int *err)
1078{
1079        struct perf_pmu *intel_pt_pmu = perf_pmu__find(INTEL_PT_PMU_NAME);
1080        struct intel_pt_recording *ptr;
1081
1082        if (!intel_pt_pmu)
1083                return NULL;
1084
1085        if (setenv("JITDUMP_USE_ARCH_TIMESTAMP", "1", 1)) {
1086                *err = -errno;
1087                return NULL;
1088        }
1089
1090        ptr = zalloc(sizeof(struct intel_pt_recording));
1091        if (!ptr) {
1092                *err = -ENOMEM;
1093                return NULL;
1094        }
1095
1096        ptr->intel_pt_pmu = intel_pt_pmu;
1097        ptr->itr.recording_options = intel_pt_recording_options;
1098        ptr->itr.info_priv_size = intel_pt_info_priv_size;
1099        ptr->itr.info_fill = intel_pt_info_fill;
1100        ptr->itr.free = intel_pt_recording_free;
1101        ptr->itr.snapshot_start = intel_pt_snapshot_start;
1102        ptr->itr.snapshot_finish = intel_pt_snapshot_finish;
1103        ptr->itr.find_snapshot = intel_pt_find_snapshot;
1104        ptr->itr.parse_snapshot_options = intel_pt_parse_snapshot_options;
1105        ptr->itr.reference = intel_pt_reference;
1106        ptr->itr.read_finish = intel_pt_read_finish;
1107        return &ptr->itr;
1108}
1109