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