tools/perf/arch/arm64/util/arm-spe.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Arm Statistical Profiling Extensions (SPE) support
   4  * Copyright (c) 2017-2018, Arm Ltd.
   5  */
   6
   7 #include <linux/kernel.h>
   8 #include <linux/types.h>
   9 #include <linux/bitops.h>
  10 #include <linux/log2.h>
  11 #include <linux/zalloc.h>
  12 #include <time.h>
  13
  14 #include "../../../util/cpumap.h"
  15 #include "../../../util/event.h"
  16 #include "../../../util/evsel.h"
  17 #include "../../../util/evsel_config.h"
  18 #include "../../../util/evlist.h"
  19 #include "../../../util/session.h"
  20 #include <internal/lib.h> // page_size
  21 #include "../../../util/pmu.h"
  22 #include "../../../util/debug.h"
  23 #include "../../../util/auxtrace.h"
  24 #include "../../../util/record.h"
  25 #include "../../../util/arm-spe.h"
  26 #include <tools/libc_compat.h> // reallocarray
  27
  28 #define KiB(x) ((x) * 1024)
  29 #define MiB(x) ((x) * 1024 * 1024)
  30
  31 struct arm_spe_recording {
  32         struct auxtrace_record          itr;
  33         struct perf_pmu                 *arm_spe_pmu;
  34         struct evlist           *evlist;
  35         int                     wrapped_cnt;
  36         bool                    *wrapped;
  37 };
  38
  39 static void arm_spe_set_timestamp(struct auxtrace_record *itr,
  40                                   struct evsel *evsel)
  41 {
  42         struct arm_spe_recording *ptr;
  43         struct perf_pmu *arm_spe_pmu;
  44         struct evsel_config_term *term = evsel__get_config_term(evsel, CFG_CHG);
  45         u64 user_bits = 0, bit;
  46
  47         ptr = container_of(itr, struct arm_spe_recording, itr);
  48         arm_spe_pmu = ptr->arm_spe_pmu;
  49
  50         if (term)
  51                 user_bits = term->val.cfg_chg;
  52
  53         bit = perf_pmu__format_bits(&arm_spe_pmu->format, "ts_enable");
  54
  55         /* Skip if user has set it */
  56         if (bit & user_bits)
  57                 return;
  58
  59         evsel->core.attr.config |= bit;
  60 }
  61
  62 static size_t
  63 arm_spe_info_priv_size(struct auxtrace_record *itr __maybe_unused,
  64                        struct evlist *evlist __maybe_unused)
  65 {
  66         return ARM_SPE_AUXTRACE_PRIV_SIZE;
  67 }
  68
  69 static int arm_spe_info_fill(struct auxtrace_record *itr,
  70                              struct perf_session *session,
  71                              struct perf_record_auxtrace_info *auxtrace_info,
  72                              size_t priv_size)
  73 {
  74         struct arm_spe_recording *sper =
  75                         container_of(itr, struct arm_spe_recording, itr);
  76         struct perf_pmu *arm_spe_pmu = sper->arm_spe_pmu;
  77
  78         if (priv_size != ARM_SPE_AUXTRACE_PRIV_SIZE)
  79                 return -EINVAL;
  80
  81         if (!session->evlist->core.nr_mmaps)
  82                 return -EINVAL;
  83
  84         auxtrace_info->type = PERF_AUXTRACE_ARM_SPE;
  85         auxtrace_info->priv[ARM_SPE_PMU_TYPE] = arm_spe_pmu->type;
  86
  87         return 0;
  88 }
  89
  90 static void
  91 arm_spe_snapshot_resolve_auxtrace_defaults(struct record_opts *opts,
  92                                            bool privileged)
  93 {
  94         /*
  95          * The default snapshot size is the auxtrace mmap size. If neither auxtrace mmap size nor
  96          * snapshot size is specified, then the default is 4MiB for privileged users, 128KiB for
  97          * unprivileged users.
  98          *
  99          * The default auxtrace mmap size is 4MiB/page_size for privileged users, 128KiB for
 100          * unprivileged users. If an unprivileged user does not specify mmap pages, the mmap pages
 101          * will be reduced from the default 512KiB/page_size to 256KiB/page_size, otherwise the
 102          * user is likely to get an error as they exceed their mlock limmit.
 103          */
 104
 105         /*
 106          * No size were given to '-S' or '-m,', so go with the default
 107          */
 108         if (!opts->auxtrace_snapshot_size && !opts->auxtrace_mmap_pages) {
 109                 if (privileged) {
 110                         opts->auxtrace_mmap_pages = MiB(4) / page_size;
 111                 } else {
 112                         opts->auxtrace_mmap_pages = KiB(128) / page_size;
 113                         if (opts->mmap_pages == UINT_MAX)
 114                                 opts->mmap_pages = KiB(256) / page_size;
 115                 }
 116         } else if (!opts->auxtrace_mmap_pages && !privileged && opts->mmap_pages == UINT_MAX) {
 117                 opts->mmap_pages = KiB(256) / page_size;
 118         }
 119
 120         /*
 121          * '-m,xyz' was specified but no snapshot size, so make the snapshot size as big as the
 122          * auxtrace mmap area.
 123          */
 124         if (!opts->auxtrace_snapshot_size)
 125                 opts->auxtrace_snapshot_size = opts->auxtrace_mmap_pages * (size_t)page_size;
 126
 127         /*
 128          * '-Sxyz' was specified but no auxtrace mmap area, so make the auxtrace mmap area big
 129          * enough to fit the requested snapshot size.
 130          */
 131         if (!opts->auxtrace_mmap_pages) {
 132                 size_t sz = opts->auxtrace_snapshot_size;
 133
 134                 sz = round_up(sz, page_size) / page_size;
 135                 opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 136         }
 137 }
 138
 139 static int arm_spe_recording_options(struct auxtrace_record *itr,
 140                                      struct evlist *evlist,
 141                                      struct record_opts *opts)
 142 {
 143         struct arm_spe_recording *sper =
 144                         container_of(itr, struct arm_spe_recording, itr);
 145         struct perf_pmu *arm_spe_pmu = sper->arm_spe_pmu;
 146         struct evsel *evsel, *arm_spe_evsel = NULL;
 147         struct perf_cpu_map *cpus = evlist->core.user_requested_cpus;
 148         bool privileged = perf_event_paranoid_check(-1);
 149         struct evsel *tracking_evsel;
 150         int err;
 151
 152         sper->evlist = evlist;
 153
 154         evlist__for_each_entry(evlist, evsel) {
 155                 if (evsel->core.attr.type == arm_spe_pmu->type) {
 156                         if (arm_spe_evsel) {
 157                                 pr_err("There may be only one " ARM_SPE_PMU_NAME "x event\n");
 158                                 return -EINVAL;
 159                         }
 160                         evsel->core.attr.freq = 0;
 161                         evsel->core.attr.sample_period = arm_spe_pmu->default_config->sample_period;
 162                         arm_spe_evsel = evsel;
 163                         opts->full_auxtrace = true;
 164                 }
 165         }
 166
 167         if (!opts->full_auxtrace)
 168                 return 0;
 169
 170         /*
 171          * we are in snapshot mode.
 172          */
 173         if (opts->auxtrace_snapshot_mode) {
 174                 /*
 175                  * Command arguments '-Sxyz' and/or '-m,xyz' are missing, so fill those in with
 176                  * default values.
 177                  */
 178                 if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages)
 179                         arm_spe_snapshot_resolve_auxtrace_defaults(opts, privileged);
 180
 181                 /*
 182                  * Snapshot size can't be bigger than the auxtrace area.
 183                  */
 184                 if (opts->auxtrace_snapshot_size > opts->auxtrace_mmap_pages * (size_t)page_size) {
 185                         pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
 186                                opts->auxtrace_snapshot_size,
 187                                opts->auxtrace_mmap_pages * (size_t)page_size);
 188                         return -EINVAL;
 189                 }
 190
 191                 /*
 192                  * Something went wrong somewhere - this shouldn't happen.
 193                  */
 194                 if (!opts->auxtrace_snapshot_size || !opts->auxtrace_mmap_pages) {
 195                         pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
 196                         return -EINVAL;
 197                 }
 198         }
 199
 200         /* We are in full trace mode but '-m,xyz' wasn't specified */
 201         if (!opts->auxtrace_mmap_pages) {
 202                 if (privileged) {
 203                         opts->auxtrace_mmap_pages = MiB(4) / page_size;
 204                 } else {
 205                         opts->auxtrace_mmap_pages = KiB(128) / page_size;
 206                         if (opts->mmap_pages == UINT_MAX)
 207                                 opts->mmap_pages = KiB(256) / page_size;
 208                 }
 209         }
 210
 211         /* Validate auxtrace_mmap_pages */
 212         if (opts->auxtrace_mmap_pages) {
 213                 size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 214                 size_t min_sz = KiB(8);
 215
 216                 if (sz < min_sz || !is_power_of_2(sz)) {
 217                         pr_err("Invalid mmap size for ARM SPE: must be at least %zuKiB and a power of 2\n",
 218                                min_sz / 1024);
 219                         return -EINVAL;
 220                 }
 221         }
 222
 223         if (opts->auxtrace_snapshot_mode)
 224                 pr_debug2("%sx snapshot size: %zu\n", ARM_SPE_PMU_NAME,
 225                           opts->auxtrace_snapshot_size);
 226
 227         /*
 228          * To obtain the auxtrace buffer file descriptor, the auxtrace event
 229          * must come first.
 230          */
 231         evlist__to_front(evlist, arm_spe_evsel);
 232
 233         /*
 234          * In the case of per-cpu mmaps, sample CPU for AUX event;
 235          * also enable the timestamp tracing for samples correlation.
 236          */
 237         if (!perf_cpu_map__empty(cpus)) {
 238                 evsel__set_sample_bit(arm_spe_evsel, CPU);
 239                 arm_spe_set_timestamp(itr, arm_spe_evsel);
 240         }
 241
 242         /*
 243          * Set this only so that perf report knows that SPE generates memory info. It has no effect
 244          * on the opening of the event or the SPE data produced.
 245          */
 246         evsel__set_sample_bit(arm_spe_evsel, DATA_SRC);
 247
 248         /* Add dummy event to keep tracking */
 249         err = parse_events(evlist, "dummy:u", NULL);
 250         if (err)
 251                 return err;
 252
 253         tracking_evsel = evlist__last(evlist);
 254         evlist__set_tracking_event(evlist, tracking_evsel);
 255
 256         tracking_evsel->core.attr.freq = 0;
 257         tracking_evsel->core.attr.sample_period = 1;
 258
 259         /* In per-cpu case, always need the time of mmap events etc */
 260         if (!perf_cpu_map__empty(cpus)) {
 261                 evsel__set_sample_bit(tracking_evsel, TIME);
 262                 evsel__set_sample_bit(tracking_evsel, CPU);
 263
 264                 /* also track task context switch */
 265                 if (!record_opts__no_switch_events(opts))
 266                         tracking_evsel->core.attr.context_switch = 1;
 267         }
 268
 269         return 0;
 270 }
 271
 272 static int arm_spe_parse_snapshot_options(struct auxtrace_record *itr __maybe_unused,
 273                                          struct record_opts *opts,
 274                                          const char *str)
 275 {
 276         unsigned long long snapshot_size = 0;
 277         char *endptr;
 278
 279         if (str) {
 280                 snapshot_size = strtoull(str, &endptr, 0);
 281                 if (*endptr || snapshot_size > SIZE_MAX)
 282                         return -1;
 283         }
 284
 285         opts->auxtrace_snapshot_mode = true;
 286         opts->auxtrace_snapshot_size = snapshot_size;
 287
 288         return 0;
 289 }
 290
 291 static int arm_spe_snapshot_start(struct auxtrace_record *itr)
 292 {
 293         struct arm_spe_recording *ptr =
 294                         container_of(itr, struct arm_spe_recording, itr);
 295         struct evsel *evsel;
 296
 297         evlist__for_each_entry(ptr->evlist, evsel) {
 298                 if (evsel->core.attr.type == ptr->arm_spe_pmu->type)
 299                         return evsel__disable(evsel);
 300         }
 301         return -EINVAL;
 302 }
 303
 304 static int arm_spe_snapshot_finish(struct auxtrace_record *itr)
 305 {
 306         struct arm_spe_recording *ptr =
 307                         container_of(itr, struct arm_spe_recording, itr);
 308         struct evsel *evsel;
 309
 310         evlist__for_each_entry(ptr->evlist, evsel) {
 311                 if (evsel->core.attr.type == ptr->arm_spe_pmu->type)
 312                         return evsel__enable(evsel);
 313         }
 314         return -EINVAL;
 315 }
 316
 317 static int arm_spe_alloc_wrapped_array(struct arm_spe_recording *ptr, int idx)
 318 {
 319         bool *wrapped;
 320         int cnt = ptr->wrapped_cnt, new_cnt, i;
 321
 322         /*
 323          * No need to allocate, so return early.
 324          */
 325         if (idx < cnt)
 326                 return 0;
 327
 328         /*
 329          * Make ptr->wrapped as big as idx.
 330          */
 331         new_cnt = idx + 1;
 332
 333         /*
 334          * Free'ed in arm_spe_recording_free().
 335          */
 336         wrapped = reallocarray(ptr->wrapped, new_cnt, sizeof(bool));
 337         if (!wrapped)
 338                 return -ENOMEM;
 339
 340         /*
 341          * init new allocated values.
 342          */
 343         for (i = cnt; i < new_cnt; i++)
 344                 wrapped[i] = false;
 345
 346         ptr->wrapped_cnt = new_cnt;
 347         ptr->wrapped = wrapped;
 348
 349         return 0;
 350 }
 351
 352 static bool arm_spe_buffer_has_wrapped(unsigned char *buffer,
 353                                       size_t buffer_size, u64 head)
 354 {
 355         u64 i, watermark;
 356         u64 *buf = (u64 *)buffer;
 357         size_t buf_size = buffer_size;
 358
 359         /*
 360          * Defensively handle the case where head might be continually increasing - if its value is
 361          * equal or greater than the size of the ring buffer, then we can safely determine it has
 362          * wrapped around. Otherwise, continue to detect if head might have wrapped.
 363          */
 364         if (head >= buffer_size)
 365                 return true;
 366
 367         /*
 368          * We want to look the very last 512 byte (chosen arbitrarily) in the ring buffer.
 369          */
 370         watermark = buf_size - 512;
 371
 372         /*
 373          * The value of head is somewhere within the size of the ring buffer. This can be that there
 374          * hasn't been enough data to fill the ring buffer yet or the trace time was so long that
 375          * head has numerically wrapped around.  To find we need to check if we have data at the
 376          * very end of the ring buffer.  We can reliably do this because mmap'ed pages are zeroed
 377          * out and there is a fresh mapping with every new session.
 378          */
 379
 380         /*
 381          * head is less than 512 byte from the end of the ring buffer.
 382          */
 383         if (head > watermark)
 384                 watermark = head;
 385
 386         /*
 387          * Speed things up by using 64 bit transactions (see "u64 *buf" above)
 388          */
 389         watermark /= sizeof(u64);
 390         buf_size /= sizeof(u64);
 391
 392         /*
 393          * If we find trace data at the end of the ring buffer, head has been there and has
 394          * numerically wrapped around at least once.
 395          */
 396         for (i = watermark; i < buf_size; i++)
 397                 if (buf[i])
 398                         return true;
 399
 400         return false;
 401 }
 402
 403 static int arm_spe_find_snapshot(struct auxtrace_record *itr, int idx,
 404                                   struct auxtrace_mmap *mm, unsigned char *data,
 405                                   u64 *head, u64 *old)
 406 {
 407         int err;
 408         bool wrapped;
 409         struct arm_spe_recording *ptr =
 410                         container_of(itr, struct arm_spe_recording, itr);
 411
 412         /*
 413          * Allocate memory to keep track of wrapping if this is the first
 414          * time we deal with this *mm.
 415          */
 416         if (idx >= ptr->wrapped_cnt) {
 417                 err = arm_spe_alloc_wrapped_array(ptr, idx);
 418                 if (err)
 419                         return err;
 420         }
 421
 422         /*
 423          * Check to see if *head has wrapped around.  If it hasn't only the
 424          * amount of data between *head and *old is snapshot'ed to avoid
 425          * bloating the perf.data file with zeros.  But as soon as *head has
 426          * wrapped around the entire size of the AUX ring buffer it taken.
 427          */
 428         wrapped = ptr->wrapped[idx];
 429         if (!wrapped && arm_spe_buffer_has_wrapped(data, mm->len, *head)) {
 430                 wrapped = true;
 431                 ptr->wrapped[idx] = true;
 432         }
 433
 434         pr_debug3("%s: mmap index %d old head %zu new head %zu size %zu\n",
 435                   __func__, idx, (size_t)*old, (size_t)*head, mm->len);
 436
 437         /*
 438          * No wrap has occurred, we can just use *head and *old.
 439          */
 440         if (!wrapped)
 441                 return 0;
 442
 443         /*
 444          * *head has wrapped around - adjust *head and *old to pickup the
 445          * entire content of the AUX buffer.
 446          */
 447         if (*head >= mm->len) {
 448                 *old = *head - mm->len;
 449         } else {
 450                 *head += mm->len;
 451                 *old = *head - mm->len;
 452         }
 453
 454         return 0;
 455 }
 456
 457 static u64 arm_spe_reference(struct auxtrace_record *itr __maybe_unused)
 458 {
 459         struct timespec ts;
 460
 461         clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
 462
 463         return ts.tv_sec ^ ts.tv_nsec;
 464 }
 465
 466 static void arm_spe_recording_free(struct auxtrace_record *itr)
 467 {
 468         struct arm_spe_recording *sper =
 469                         container_of(itr, struct arm_spe_recording, itr);
 470
 471         free(sper->wrapped);
 472         free(sper);
 473 }
 474
 475 struct auxtrace_record *arm_spe_recording_init(int *err,
 476                                                struct perf_pmu *arm_spe_pmu)
 477 {
 478         struct arm_spe_recording *sper;
 479
 480         if (!arm_spe_pmu) {
 481                 *err = -ENODEV;
 482                 return NULL;
 483         }
 484
 485         sper = zalloc(sizeof(struct arm_spe_recording));
 486         if (!sper) {
 487                 *err = -ENOMEM;
 488                 return NULL;
 489         }
 490
 491         sper->arm_spe_pmu = arm_spe_pmu;
 492         sper->itr.pmu = arm_spe_pmu;
 493         sper->itr.snapshot_start = arm_spe_snapshot_start;
 494         sper->itr.snapshot_finish = arm_spe_snapshot_finish;
 495         sper->itr.find_snapshot = arm_spe_find_snapshot;
 496         sper->itr.parse_snapshot_options = arm_spe_parse_snapshot_options;
 497         sper->itr.recording_options = arm_spe_recording_options;
 498         sper->itr.info_priv_size = arm_spe_info_priv_size;
 499         sper->itr.info_fill = arm_spe_info_fill;
 500         sper->itr.free = arm_spe_recording_free;
 501         sper->itr.reference = arm_spe_reference;
 502         sper->itr.read_finish = auxtrace_record__read_finish;
 503         sper->itr.alignment = 0;
 504
 505         *err = 0;
 506         return &sper->itr;
 507 }
 508
 509 struct perf_event_attr
 510 *arm_spe_pmu_default_config(struct perf_pmu *arm_spe_pmu)
 511 {
 512         struct perf_event_attr *attr;
 513
 514         attr = zalloc(sizeof(struct perf_event_attr));
 515         if (!attr) {
 516                 pr_err("arm_spe default config cannot allocate a perf_event_attr\n");
 517                 return NULL;
 518         }
 519
 520         /*
 521          * If kernel driver doesn't advertise a minimum,
 522          * use max allowable by PMSIDR_EL1.INTERVAL
 523          */
 524         if (perf_pmu__scan_file(arm_spe_pmu, "caps/min_interval", "%llu",
 525                                   &attr->sample_period) != 1) {
 526                 pr_debug("arm_spe driver doesn't advertise a min. interval. Using 4096\n");
 527                 attr->sample_period = 4096;
 528         }
 529
 530         arm_spe_pmu->selectable = true;
 531         arm_spe_pmu->is_uncore = false;
 532
 533         return attr;
 534 }