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/xyarray.h>
  15 #include <internal/lib.h>
  16 #include <linux/zalloc.h>
  17 #include <sys/ioctl.h>
  18 #include <stdlib.h>
  19 #include <errno.h>
  20 #include <unistd.h>
  21 #include <fcntl.h>
  22 #include <signal.h>
  23 #include <poll.h>
  24 #include <sys/mman.h>
  25 #include <perf/cpumap.h>
  26 #include <perf/threadmap.h>
  27 #include <api/fd/array.h>
  28
  29 void perf_evlist__init(struct perf_evlist *evlist)
  30 {
  31         int i;
  32
  33         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  34                 INIT_HLIST_HEAD(&evlist->heads[i]);
  35         INIT_LIST_HEAD(&evlist->entries);
  36         evlist->nr_entries = 0;
  37         fdarray__init(&evlist->pollfd, 64);
  38 }
  39
  40 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
  41                                           struct perf_evsel *evsel)
  42 {
  43         /*
  44          * We already have cpus for evsel (via PMU sysfs) so
  45          * keep it, if there's no target cpu list defined.
  46          */
  47         if (!evsel->own_cpus || evlist->has_user_cpus) {
  48                 perf_cpu_map__put(evsel->cpus);
  49                 evsel->cpus = perf_cpu_map__get(evlist->cpus);
  50         } else if (evsel->cpus != evsel->own_cpus) {
  51                 perf_cpu_map__put(evsel->cpus);
  52                 evsel->cpus = perf_cpu_map__get(evsel->own_cpus);
  53         }
  54
  55         perf_thread_map__put(evsel->threads);
  56         evsel->threads = perf_thread_map__get(evlist->threads);
  57         evlist->all_cpus = perf_cpu_map__merge(evlist->all_cpus, evsel->cpus);
  58 }
  59
  60 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
  61 {
  62         struct perf_evsel *evsel;
  63
  64         perf_evlist__for_each_evsel(evlist, evsel)
  65                 __perf_evlist__propagate_maps(evlist, evsel);
  66 }
  67
  68 void perf_evlist__add(struct perf_evlist *evlist,
  69                       struct perf_evsel *evsel)
  70 {
  71         list_add_tail(&evsel->node, &evlist->entries);
  72         evlist->nr_entries += 1;
  73         __perf_evlist__propagate_maps(evlist, evsel);
  74 }
  75
  76 void perf_evlist__remove(struct perf_evlist *evlist,
  77                          struct perf_evsel *evsel)
  78 {
  79         list_del_init(&evsel->node);
  80         evlist->nr_entries -= 1;
  81 }
  82
  83 struct perf_evlist *perf_evlist__new(void)
  84 {
  85         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  86
  87         if (evlist != NULL)
  88                 perf_evlist__init(evlist);
  89
  90         return evlist;
  91 }
  92
  93 struct perf_evsel *
  94 perf_evlist__next(struct perf_evlist *evlist, struct perf_evsel *prev)
  95 {
  96         struct perf_evsel *next;
  97
  98         if (!prev) {
  99                 next = list_first_entry(&evlist->entries,
 100                                         struct perf_evsel,
 101                                         node);
 102         } else {
 103                 next = list_next_entry(prev, node);
 104         }
 105
 106         /* Empty list is noticed here so don't need checking on entry. */
 107         if (&next->node == &evlist->entries)
 108                 return NULL;
 109
 110         return next;
 111 }
 112
 113 static void perf_evlist__purge(struct perf_evlist *evlist)
 114 {
 115         struct perf_evsel *pos, *n;
 116
 117         perf_evlist__for_each_entry_safe(evlist, n, pos) {
 118                 list_del_init(&pos->node);
 119                 perf_evsel__delete(pos);
 120         }
 121
 122         evlist->nr_entries = 0;
 123 }
 124
 125 void perf_evlist__exit(struct perf_evlist *evlist)
 126 {
 127         perf_cpu_map__put(evlist->cpus);
 128         perf_cpu_map__put(evlist->all_cpus);
 129         perf_thread_map__put(evlist->threads);
 130         evlist->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__id_add(struct perf_evlist *evlist,
 239                          struct perf_evsel *evsel,
 240                          int cpu, int thread, u64 id)
 241 {
 242         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 243         evsel->id[evsel->ids++] = id;
 244 }
 245
 246 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 247                            struct perf_evsel *evsel,
 248                            int cpu, int thread, int fd)
 249 {
 250         u64 read_data[4] = { 0, };
 251         int id_idx = 1; /* The first entry is the counter value */
 252         u64 id;
 253         int ret;
 254
 255         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 256         if (!ret)
 257                 goto add;
 258
 259         if (errno != ENOTTY)
 260                 return -1;
 261
 262         /* Legacy way to get event id.. All hail to old kernels! */
 263
 264         /*
 265          * This way does not work with group format read, so bail
 266          * out in that case.
 267          */
 268         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 269                 return -1;
 270
 271         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 272             read(fd, &read_data, sizeof(read_data)) == -1)
 273                 return -1;
 274
 275         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 276                 ++id_idx;
 277         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 278                 ++id_idx;
 279
 280         id = read_data[id_idx];
 281
 282 add:
 283         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 284         return 0;
 285 }
 286
 287 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 288 {
 289         int nr_cpus = perf_cpu_map__nr(evlist->cpus);
 290         int nr_threads = perf_thread_map__nr(evlist->threads);
 291         int nfds = 0;
 292         struct perf_evsel *evsel;
 293
 294         perf_evlist__for_each_entry(evlist, evsel) {
 295                 if (evsel->system_wide)
 296                         nfds += nr_cpus;
 297                 else
 298                         nfds += nr_cpus * nr_threads;
 299         }
 300
 301         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 302             fdarray__grow(&evlist->pollfd, nfds) < 0)
 303                 return -ENOMEM;
 304
 305         return 0;
 306 }
 307
 308 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 309                             void *ptr, short revent)
 310 {
 311         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 312
 313         if (pos >= 0) {
 314                 evlist->pollfd.priv[pos].ptr = ptr;
 315                 fcntl(fd, F_SETFL, O_NONBLOCK);
 316         }
 317
 318         return pos;
 319 }
 320
 321 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 322                                          void *arg __maybe_unused)
 323 {
 324         struct perf_mmap *map = fda->priv[fd].ptr;
 325
 326         if (map)
 327                 perf_mmap__put(map);
 328 }
 329
 330 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 331 {
 332         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 333                                perf_evlist__munmap_filtered, NULL);
 334 }
 335
 336 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 337 {
 338         return fdarray__poll(&evlist->pollfd, timeout);
 339 }
 340
 341 static struct perf_mmap* perf_evlist__alloc_mmap(struct perf_evlist *evlist, bool overwrite)
 342 {
 343         int i;
 344         struct perf_mmap *map;
 345
 346         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 347         if (!map)
 348                 return NULL;
 349
 350         for (i = 0; i < evlist->nr_mmaps; i++) {
 351                 struct perf_mmap *prev = i ? &map[i - 1] : NULL;
 352
 353                 /*
 354                  * When the perf_mmap() call is made we grab one refcount, plus
 355                  * one extra to let perf_mmap__consume() get the last
 356                  * events after all real references (perf_mmap__get()) are
 357                  * dropped.
 358                  *
 359                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 360                  * thus does perf_mmap__get() on it.
 361                  */
 362                 perf_mmap__init(&map[i], prev, overwrite, NULL);
 363         }
 364
 365         return map;
 366 }
 367
 368 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 369                                      struct perf_evsel *evsel, int idx, int cpu,
 370                                      int thread)
 371 {
 372         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 373
 374         sid->idx = idx;
 375         if (evlist->cpus && cpu >= 0)
 376                 sid->cpu = evlist->cpus->map[cpu];
 377         else
 378                 sid->cpu = -1;
 379         if (!evsel->system_wide && evlist->threads && thread >= 0)
 380                 sid->tid = perf_thread_map__pid(evlist->threads, thread);
 381         else
 382                 sid->tid = -1;
 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) < 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_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 502                                                  thread);
 503                 }
 504         }
 505
 506         return 0;
 507 }
 508
 509 static int
 510 mmap_per_thread(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
 511                 struct perf_mmap_param *mp)
 512 {
 513         int thread;
 514         int nr_threads = perf_thread_map__nr(evlist->threads);
 515
 516         for (thread = 0; thread < nr_threads; thread++) {
 517                 int output = -1;
 518                 int output_overwrite = -1;
 519
 520                 if (ops->idx)
 521                         ops->idx(evlist, mp, thread, false);
 522
 523                 if (mmap_per_evsel(evlist, ops, thread, mp, 0, thread,
 524                                    &output, &output_overwrite))
 525                         goto out_unmap;
 526         }
 527
 528         return 0;
 529
 530 out_unmap:
 531         perf_evlist__munmap(evlist);
 532         return -1;
 533 }
 534
 535 static int
 536 mmap_per_cpu(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
 537              struct perf_mmap_param *mp)
 538 {
 539         int nr_threads = perf_thread_map__nr(evlist->threads);
 540         int nr_cpus    = perf_cpu_map__nr(evlist->cpus);
 541         int cpu, thread;
 542
 543         for (cpu = 0; cpu < nr_cpus; cpu++) {
 544                 int output = -1;
 545                 int output_overwrite = -1;
 546
 547                 if (ops->idx)
 548                         ops->idx(evlist, mp, cpu, true);
 549
 550                 for (thread = 0; thread < nr_threads; thread++) {
 551                         if (mmap_per_evsel(evlist, ops, cpu, mp, cpu,
 552                                            thread, &output, &output_overwrite))
 553                                 goto out_unmap;
 554                 }
 555         }
 556
 557         return 0;
 558
 559 out_unmap:
 560         perf_evlist__munmap(evlist);
 561         return -1;
 562 }
 563
 564 static int perf_evlist__nr_mmaps(struct perf_evlist *evlist)
 565 {
 566         int nr_mmaps;
 567
 568         nr_mmaps = perf_cpu_map__nr(evlist->cpus);
 569         if (perf_cpu_map__empty(evlist->cpus))
 570                 nr_mmaps = perf_thread_map__nr(evlist->threads);
 571
 572         return nr_mmaps;
 573 }
 574
 575 int perf_evlist__mmap_ops(struct perf_evlist *evlist,
 576                           struct perf_evlist_mmap_ops *ops,
 577                           struct perf_mmap_param *mp)
 578 {
 579         struct perf_evsel *evsel;
 580         const struct perf_cpu_map *cpus = evlist->cpus;
 581         const struct perf_thread_map *threads = evlist->threads;
 582
 583         if (!ops || !ops->get || !ops->mmap)
 584                 return -EINVAL;
 585
 586         mp->mask = evlist->mmap_len - page_size - 1;
 587
 588         evlist->nr_mmaps = perf_evlist__nr_mmaps(evlist);
 589
 590         perf_evlist__for_each_entry(evlist, evsel) {
 591                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 592                     evsel->sample_id == NULL &&
 593                     perf_evsel__alloc_id(evsel, perf_cpu_map__nr(cpus), threads->nr) < 0)
 594                         return -ENOMEM;
 595         }
 596
 597         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 598                 return -ENOMEM;
 599
 600         if (perf_cpu_map__empty(cpus))
 601                 return mmap_per_thread(evlist, ops, mp);
 602
 603         return mmap_per_cpu(evlist, ops, mp);
 604 }
 605
 606 int perf_evlist__mmap(struct perf_evlist *evlist, int pages)
 607 {
 608         struct perf_mmap_param mp;
 609         struct perf_evlist_mmap_ops ops = {
 610                 .get  = perf_evlist__mmap_cb_get,
 611                 .mmap = perf_evlist__mmap_cb_mmap,
 612         };
 613
 614         evlist->mmap_len = (pages + 1) * page_size;
 615
 616         return perf_evlist__mmap_ops(evlist, &ops, &mp);
 617 }
 618
 619 void perf_evlist__munmap(struct perf_evlist *evlist)
 620 {
 621         int i;
 622
 623         if (evlist->mmap) {
 624                 for (i = 0; i < evlist->nr_mmaps; i++)
 625                         perf_mmap__munmap(&evlist->mmap[i]);
 626         }
 627
 628         if (evlist->mmap_ovw) {
 629                 for (i = 0; i < evlist->nr_mmaps; i++)
 630                         perf_mmap__munmap(&evlist->mmap_ovw[i]);
 631         }
 632
 633         zfree(&evlist->mmap);
 634         zfree(&evlist->mmap_ovw);
 635 }
 636
 637 struct perf_mmap*
 638 perf_evlist__next_mmap(struct perf_evlist *evlist, struct perf_mmap *map,
 639                        bool overwrite)
 640 {
 641         if (map)
 642                 return map->next;
 643
 644         return overwrite ? evlist->mmap_ovw_first : evlist->mmap_first;
 645 }