tools/lib/perf/evlist.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <perf/evlist.h>
   3 #include <perf/evsel.h>
   4 #include <linux/bitops.h>
   5 #include <linux/list.h>
   6 #include <linux/hash.h>
   7 #include <sys/ioctl.h>
   8 #include <internal/evlist.h>
   9 #include <internal/evsel.h>
  10 #include <internal/xyarray.h>
  11 #include <internal/mmap.h>
  12 #include <internal/cpumap.h>
  13 #include <internal/threadmap.h>
  14 #include <internal/lib.h>
  15 #include <linux/zalloc.h>
  16 #include <stdlib.h>
  17 #include <errno.h>
  18 #include <unistd.h>
  19 #include <fcntl.h>
  20 #include <signal.h>
  21 #include <poll.h>
  22 #include <sys/mman.h>
  23 #include <perf/cpumap.h>
  24 #include <perf/threadmap.h>
  25 #include <api/fd/array.h>
  26
  27 void perf_evlist__init(struct perf_evlist *evlist)
  28 {
  29         INIT_LIST_HEAD(&evlist->entries);
  30         evlist->nr_entries = 0;
  31         fdarray__init(&evlist->pollfd, 64);
  32         perf_evlist__reset_id_hash(evlist);
  33 }
  34
  35 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
  36                                           struct perf_evsel *evsel)
  37 {
  38         /*
  39          * We already have cpus for evsel (via PMU sysfs) so
  40          * keep it, if there's no target cpu list defined.
  41          */
  42         if (!evsel->own_cpus || evlist->has_user_cpus) {
  43                 perf_cpu_map__put(evsel->cpus);
  44                 evsel->cpus = perf_cpu_map__get(evlist->cpus);
  45         } else if (!evsel->system_wide && perf_cpu_map__empty(evlist->cpus)) {
  46                 perf_cpu_map__put(evsel->cpus);
  47                 evsel->cpus = perf_cpu_map__get(evlist->cpus);
  48         } else if (evsel->cpus != evsel->own_cpus) {
  49                 perf_cpu_map__put(evsel->cpus);
  50                 evsel->cpus = perf_cpu_map__get(evsel->own_cpus);
  51         }
  52
  53         perf_thread_map__put(evsel->threads);
  54         evsel->threads = perf_thread_map__get(evlist->threads);
  55         evlist->all_cpus = perf_cpu_map__merge(evlist->all_cpus, evsel->cpus);
  56 }
  57
  58 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
  59 {
  60         struct perf_evsel *evsel;
  61
  62         perf_evlist__for_each_evsel(evlist, evsel)
  63                 __perf_evlist__propagate_maps(evlist, evsel);
  64 }
  65
  66 void perf_evlist__add(struct perf_evlist *evlist,
  67                       struct perf_evsel *evsel)
  68 {
  69         evsel->idx = evlist->nr_entries;
  70         list_add_tail(&evsel->node, &evlist->entries);
  71         evlist->nr_entries += 1;
  72         __perf_evlist__propagate_maps(evlist, evsel);
  73 }
  74
  75 void perf_evlist__remove(struct perf_evlist *evlist,
  76                          struct perf_evsel *evsel)
  77 {
  78         list_del_init(&evsel->node);
  79         evlist->nr_entries -= 1;
  80 }
  81
  82 struct perf_evlist *perf_evlist__new(void)
  83 {
  84         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  85
  86         if (evlist != NULL)
  87                 perf_evlist__init(evlist);
  88
  89         return evlist;
  90 }
  91
  92 struct perf_evsel *
  93 perf_evlist__next(struct perf_evlist *evlist, struct perf_evsel *prev)
  94 {
  95         struct perf_evsel *next;
  96
  97         if (!prev) {
  98                 next = list_first_entry(&evlist->entries,
  99                                         struct perf_evsel,
 100                                         node);
 101         } else {
 102                 next = list_next_entry(prev, node);
 103         }
 104
 105         /* Empty list is noticed here so don't need checking on entry. */
 106         if (&next->node == &evlist->entries)
 107                 return NULL;
 108
 109         return next;
 110 }
 111
 112 static void perf_evlist__purge(struct perf_evlist *evlist)
 113 {
 114         struct perf_evsel *pos, *n;
 115
 116         perf_evlist__for_each_entry_safe(evlist, n, pos) {
 117                 list_del_init(&pos->node);
 118                 perf_evsel__delete(pos);
 119         }
 120
 121         evlist->nr_entries = 0;
 122 }
 123
 124 void perf_evlist__exit(struct perf_evlist *evlist)
 125 {
 126         perf_cpu_map__put(evlist->cpus);
 127         perf_cpu_map__put(evlist->all_cpus);
 128         perf_thread_map__put(evlist->threads);
 129         evlist->cpus = NULL;
 130         evlist->all_cpus = NULL;
 131         evlist->threads = NULL;
 132         fdarray__exit(&evlist->pollfd);
 133 }
 134
 135 void perf_evlist__delete(struct perf_evlist *evlist)
 136 {
 137         if (evlist == NULL)
 138                 return;
 139
 140         perf_evlist__munmap(evlist);
 141         perf_evlist__close(evlist);
 142         perf_evlist__purge(evlist);
 143         perf_evlist__exit(evlist);
 144         free(evlist);
 145 }
 146
 147 void perf_evlist__set_maps(struct perf_evlist *evlist,
 148                            struct perf_cpu_map *cpus,
 149                            struct perf_thread_map *threads)
 150 {
 151         /*
 152          * Allow for the possibility that one or another of the maps isn't being
 153          * changed i.e. don't put it.  Note we are assuming the maps that are
 154          * being applied are brand new and evlist is taking ownership of the
 155          * original reference count of 1.  If that is not the case it is up to
 156          * the caller to increase the reference count.
 157          */
 158         if (cpus != evlist->cpus) {
 159                 perf_cpu_map__put(evlist->cpus);
 160                 evlist->cpus = perf_cpu_map__get(cpus);
 161         }
 162
 163         if (threads != evlist->threads) {
 164                 perf_thread_map__put(evlist->threads);
 165                 evlist->threads = perf_thread_map__get(threads);
 166         }
 167
 168         if (!evlist->all_cpus && cpus)
 169                 evlist->all_cpus = perf_cpu_map__get(cpus);
 170
 171         perf_evlist__propagate_maps(evlist);
 172 }
 173
 174 int perf_evlist__open(struct perf_evlist *evlist)
 175 {
 176         struct perf_evsel *evsel;
 177         int err;
 178
 179         perf_evlist__for_each_entry(evlist, evsel) {
 180                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
 181                 if (err < 0)
 182                         goto out_err;
 183         }
 184
 185         return 0;
 186
 187 out_err:
 188         perf_evlist__close(evlist);
 189         return err;
 190 }
 191
 192 void perf_evlist__close(struct perf_evlist *evlist)
 193 {
 194         struct perf_evsel *evsel;
 195
 196         perf_evlist__for_each_entry_reverse(evlist, evsel)
 197                 perf_evsel__close(evsel);
 198 }
 199
 200 void perf_evlist__enable(struct perf_evlist *evlist)
 201 {
 202         struct perf_evsel *evsel;
 203
 204         perf_evlist__for_each_entry(evlist, evsel)
 205                 perf_evsel__enable(evsel);
 206 }
 207
 208 void perf_evlist__disable(struct perf_evlist *evlist)
 209 {
 210         struct perf_evsel *evsel;
 211
 212         perf_evlist__for_each_entry(evlist, evsel)
 213                 perf_evsel__disable(evsel);
 214 }
 215
 216 u64 perf_evlist__read_format(struct perf_evlist *evlist)
 217 {
 218         struct perf_evsel *first = perf_evlist__first(evlist);
 219
 220         return first->attr.read_format;
 221 }
 222
 223 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
 224
 225 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 226                                  struct perf_evsel *evsel,
 227                                  int cpu, int thread, u64 id)
 228 {
 229         int hash;
 230         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 231
 232         sid->id = id;
 233         sid->evsel = evsel;
 234         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 235         hlist_add_head(&sid->node, &evlist->heads[hash]);
 236 }
 237
 238 void perf_evlist__reset_id_hash(struct perf_evlist *evlist)
 239 {
 240         int i;
 241
 242         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
 243                 INIT_HLIST_HEAD(&evlist->heads[i]);
 244 }
 245
 246 void perf_evlist__id_add(struct perf_evlist *evlist,
 247                          struct perf_evsel *evsel,
 248                          int cpu, int thread, u64 id)
 249 {
 250         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 251         evsel->id[evsel->ids++] = id;
 252 }
 253
 254 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 255                            struct perf_evsel *evsel,
 256                            int cpu, int thread, int fd)
 257 {
 258         u64 read_data[4] = { 0, };
 259         int id_idx = 1; /* The first entry is the counter value */
 260         u64 id;
 261         int ret;
 262
 263         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 264         if (!ret)
 265                 goto add;
 266
 267         if (errno != ENOTTY)
 268                 return -1;
 269
 270         /* Legacy way to get event id.. All hail to old kernels! */
 271
 272         /*
 273          * This way does not work with group format read, so bail
 274          * out in that case.
 275          */
 276         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 277                 return -1;
 278
 279         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 280             read(fd, &read_data, sizeof(read_data)) == -1)
 281                 return -1;
 282
 283         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 284                 ++id_idx;
 285         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 286                 ++id_idx;
 287
 288         id = read_data[id_idx];
 289
 290 add:
 291         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 292         return 0;
 293 }
 294
 295 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 296 {
 297         int nr_cpus = perf_cpu_map__nr(evlist->cpus);
 298         int nr_threads = perf_thread_map__nr(evlist->threads);
 299         int nfds = 0;
 300         struct perf_evsel *evsel;
 301
 302         perf_evlist__for_each_entry(evlist, evsel) {
 303                 if (evsel->system_wide)
 304                         nfds += nr_cpus;
 305                 else
 306                         nfds += nr_cpus * nr_threads;
 307         }
 308
 309         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 310             fdarray__grow(&evlist->pollfd, nfds) < 0)
 311                 return -ENOMEM;
 312
 313         return 0;
 314 }
 315
 316 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 317                             void *ptr, short revent, enum fdarray_flags flags)
 318 {
 319         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP, flags);
 320
 321         if (pos >= 0) {
 322                 evlist->pollfd.priv[pos].ptr = ptr;
 323                 fcntl(fd, F_SETFL, O_NONBLOCK);
 324         }
 325
 326         return pos;
 327 }
 328
 329 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 330                                          void *arg __maybe_unused)
 331 {
 332         struct perf_mmap *map = fda->priv[fd].ptr;
 333
 334         if (map)
 335                 perf_mmap__put(map);
 336 }
 337
 338 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 339 {
 340         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 341                                perf_evlist__munmap_filtered, NULL);
 342 }
 343
 344 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 345 {
 346         return fdarray__poll(&evlist->pollfd, timeout);
 347 }
 348
 349 static struct perf_mmap* perf_evlist__alloc_mmap(struct perf_evlist *evlist, bool overwrite)
 350 {
 351         int i;
 352         struct perf_mmap *map;
 353
 354         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 355         if (!map)
 356                 return NULL;
 357
 358         for (i = 0; i < evlist->nr_mmaps; i++) {
 359                 struct perf_mmap *prev = i ? &map[i - 1] : NULL;
 360
 361                 /*
 362                  * When the perf_mmap() call is made we grab one refcount, plus
 363                  * one extra to let perf_mmap__consume() get the last
 364                  * events after all real references (perf_mmap__get()) are
 365                  * dropped.
 366                  *
 367                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 368                  * thus does perf_mmap__get() on it.
 369                  */
 370                 perf_mmap__init(&map[i], prev, overwrite, NULL);
 371         }
 372
 373         return map;
 374 }
 375
 376 static void perf_evsel__set_sid_idx(struct perf_evsel *evsel, int idx, int cpu, int thread)
 377 {
 378         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 379
 380         sid->idx = idx;
 381         sid->cpu = perf_cpu_map__cpu(evsel->cpus, cpu);
 382         sid->tid = perf_thread_map__pid(evsel->threads, thread);
 383 }
 384
 385 static struct perf_mmap*
 386 perf_evlist__mmap_cb_get(struct perf_evlist *evlist, bool overwrite, int idx)
 387 {
 388         struct perf_mmap *maps;
 389
 390         maps = overwrite ? evlist->mmap_ovw : evlist->mmap;
 391
 392         if (!maps) {
 393                 maps = perf_evlist__alloc_mmap(evlist, overwrite);
 394                 if (!maps)
 395                         return NULL;
 396
 397                 if (overwrite)
 398                         evlist->mmap_ovw = maps;
 399                 else
 400                         evlist->mmap = maps;
 401         }
 402
 403         return &maps[idx];
 404 }
 405
 406 #define FD(e, x, y) (*(int *) xyarray__entry(e->fd, x, y))
 407
 408 static int
 409 perf_evlist__mmap_cb_mmap(struct perf_mmap *map, struct perf_mmap_param *mp,
 410                           int output, int cpu)
 411 {
 412         return perf_mmap__mmap(map, mp, output, cpu);
 413 }
 414
 415 static void perf_evlist__set_mmap_first(struct perf_evlist *evlist, struct perf_mmap *map,
 416                                         bool overwrite)
 417 {
 418         if (overwrite)
 419                 evlist->mmap_ovw_first = map;
 420         else
 421                 evlist->mmap_first = map;
 422 }
 423
 424 static int
 425 mmap_per_evsel(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
 426                int idx, struct perf_mmap_param *mp, int cpu_idx,
 427                int thread, int *_output, int *_output_overwrite)
 428 {
 429         int evlist_cpu = perf_cpu_map__cpu(evlist->cpus, cpu_idx);
 430         struct perf_evsel *evsel;
 431         int revent;
 432
 433         perf_evlist__for_each_entry(evlist, evsel) {
 434                 bool overwrite = evsel->attr.write_backward;
 435                 struct perf_mmap *map;
 436                 int *output, fd, cpu;
 437
 438                 if (evsel->system_wide && thread)
 439                         continue;
 440
 441                 cpu = perf_cpu_map__idx(evsel->cpus, evlist_cpu);
 442                 if (cpu == -1)
 443                         continue;
 444
 445                 map = ops->get(evlist, overwrite, idx);
 446                 if (map == NULL)
 447                         return -ENOMEM;
 448
 449                 if (overwrite) {
 450                         mp->prot = PROT_READ;
 451                         output   = _output_overwrite;
 452                 } else {
 453                         mp->prot = PROT_READ | PROT_WRITE;
 454                         output   = _output;
 455                 }
 456
 457                 fd = FD(evsel, cpu, thread);
 458
 459                 if (*output == -1) {
 460                         *output = fd;
 461
 462                         /*
 463                          * The last one will be done at perf_mmap__consume(), so that we
 464                          * make sure we don't prevent tools from consuming every last event in
 465                          * the ring buffer.
 466                          *
 467                          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 468                          * anymore, but the last events for it are still in the ring buffer,
 469                          * waiting to be consumed.
 470                          *
 471                          * Tools can chose to ignore this at their own discretion, but the
 472                          * evlist layer can't just drop it when filtering events in
 473                          * perf_evlist__filter_pollfd().
 474                          */
 475                         refcount_set(&map->refcnt, 2);
 476
 477                         if (ops->mmap(map, mp, *output, evlist_cpu) < 0)
 478                                 return -1;
 479
 480                         if (!idx)
 481                                 perf_evlist__set_mmap_first(evlist, map, overwrite);
 482                 } else {
 483                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 484                                 return -1;
 485
 486                         perf_mmap__get(map);
 487                 }
 488
 489                 revent = !overwrite ? POLLIN : 0;
 490
 491                 if (!evsel->system_wide &&
 492                     perf_evlist__add_pollfd(evlist, fd, map, revent, fdarray_flag__default) < 0) {
 493                         perf_mmap__put(map);
 494                         return -1;
 495                 }
 496
 497                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 498                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 499                                                    fd) < 0)
 500                                 return -1;
 501                         perf_evsel__set_sid_idx(evsel, idx, cpu, thread);
 502                 }
 503         }
 504
 505         return 0;
 506 }
 507
 508 static int
 509 mmap_per_thread(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
 510                 struct perf_mmap_param *mp)
 511 {
 512         int thread;
 513         int nr_threads = perf_thread_map__nr(evlist->threads);
 514
 515         for (thread = 0; thread < nr_threads; thread++) {
 516                 int output = -1;
 517                 int output_overwrite = -1;
 518
 519                 if (ops->idx)
 520                         ops->idx(evlist, mp, thread, false);
 521
 522                 if (mmap_per_evsel(evlist, ops, thread, mp, 0, thread,
 523                                    &output, &output_overwrite))
 524                         goto out_unmap;
 525         }
 526
 527         return 0;
 528
 529 out_unmap:
 530         perf_evlist__munmap(evlist);
 531         return -1;
 532 }
 533
 534 static int
 535 mmap_per_cpu(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
 536              struct perf_mmap_param *mp)
 537 {
 538         int nr_threads = perf_thread_map__nr(evlist->threads);
 539         int nr_cpus    = perf_cpu_map__nr(evlist->cpus);
 540         int cpu, thread;
 541
 542         for (cpu = 0; cpu < nr_cpus; cpu++) {
 543                 int output = -1;
 544                 int output_overwrite = -1;
 545
 546                 if (ops->idx)
 547                         ops->idx(evlist, mp, cpu, true);
 548
 549                 for (thread = 0; thread < nr_threads; thread++) {
 550                         if (mmap_per_evsel(evlist, ops, cpu, mp, cpu,
 551                                            thread, &output, &output_overwrite))
 552                                 goto out_unmap;
 553                 }
 554         }
 555
 556         return 0;
 557
 558 out_unmap:
 559         perf_evlist__munmap(evlist);
 560         return -1;
 561 }
 562
 563 static int perf_evlist__nr_mmaps(struct perf_evlist *evlist)
 564 {
 565         int nr_mmaps;
 566
 567         nr_mmaps = perf_cpu_map__nr(evlist->cpus);
 568         if (perf_cpu_map__empty(evlist->cpus))
 569                 nr_mmaps = perf_thread_map__nr(evlist->threads);
 570
 571         return nr_mmaps;
 572 }
 573
 574 int perf_evlist__mmap_ops(struct perf_evlist *evlist,
 575                           struct perf_evlist_mmap_ops *ops,
 576                           struct perf_mmap_param *mp)
 577 {
 578         struct perf_evsel *evsel;
 579         const struct perf_cpu_map *cpus = evlist->cpus;
 580         const struct perf_thread_map *threads = evlist->threads;
 581
 582         if (!ops || !ops->get || !ops->mmap)
 583                 return -EINVAL;
 584
 585         mp->mask = evlist->mmap_len - page_size - 1;
 586
 587         evlist->nr_mmaps = perf_evlist__nr_mmaps(evlist);
 588
 589         perf_evlist__for_each_entry(evlist, evsel) {
 590                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 591                     evsel->sample_id == NULL &&
 592                     perf_evsel__alloc_id(evsel, perf_cpu_map__nr(cpus), threads->nr) < 0)
 593                         return -ENOMEM;
 594         }
 595
 596         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 597                 return -ENOMEM;
 598
 599         if (perf_cpu_map__empty(cpus))
 600                 return mmap_per_thread(evlist, ops, mp);
 601
 602         return mmap_per_cpu(evlist, ops, mp);
 603 }
 604
 605 int perf_evlist__mmap(struct perf_evlist *evlist, int pages)
 606 {
 607         struct perf_mmap_param mp;
 608         struct perf_evlist_mmap_ops ops = {
 609                 .get  = perf_evlist__mmap_cb_get,
 610                 .mmap = perf_evlist__mmap_cb_mmap,
 611         };
 612
 613         evlist->mmap_len = (pages + 1) * page_size;
 614
 615         return perf_evlist__mmap_ops(evlist, &ops, &mp);
 616 }
 617
 618 void perf_evlist__munmap(struct perf_evlist *evlist)
 619 {
 620         int i;
 621
 622         if (evlist->mmap) {
 623                 for (i = 0; i < evlist->nr_mmaps; i++)
 624                         perf_mmap__munmap(&evlist->mmap[i]);
 625         }
 626
 627         if (evlist->mmap_ovw) {
 628                 for (i = 0; i < evlist->nr_mmaps; i++)
 629                         perf_mmap__munmap(&evlist->mmap_ovw[i]);
 630         }
 631
 632         zfree(&evlist->mmap);
 633         zfree(&evlist->mmap_ovw);
 634 }
 635
 636 struct perf_mmap*
 637 perf_evlist__next_mmap(struct perf_evlist *evlist, struct perf_mmap *map,
 638                        bool overwrite)
 639 {
 640         if (map)
 641                 return map->next;
 642
 643         return overwrite ? evlist->mmap_ovw_first : evlist->mmap_first;
 644 }
 645
 646 void __perf_evlist__set_leader(struct list_head *list)
 647 {
 648         struct perf_evsel *evsel, *leader;
 649
 650         leader = list_entry(list->next, struct perf_evsel, node);
 651         evsel = list_entry(list->prev, struct perf_evsel, node);
 652
 653         leader->nr_members = evsel->idx - leader->idx + 1;
 654
 655         __perf_evlist__for_each_entry(list, evsel)
 656                 evsel->leader = leader;
 657 }
 658
 659 void perf_evlist__set_leader(struct perf_evlist *evlist)
 660 {
 661         if (evlist->nr_entries) {
 662                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 663                 __perf_evlist__set_leader(&evlist->entries);
 664         }
 665 }