tools/perf/util/evlist.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9 #include <poll.h>
  10 #include "cpumap.h"
  11 #include "thread_map.h"
  12 #include "evlist.h"
  13 #include "evsel.h"
  14 #include "util.h"
  15
  16 #include <sys/mman.h>
  17
  18 #include <linux/bitops.h>
  19 #include <linux/hash.h>
  20
  21 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  22 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  23
  24 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  25                        struct thread_map *threads)
  26 {
  27         int i;
  28
  29         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  30                 INIT_HLIST_HEAD(&evlist->heads[i]);
  31         INIT_LIST_HEAD(&evlist->entries);
  32         perf_evlist__set_maps(evlist, cpus, threads);
  33 }
  34
  35 struct perf_evlist *perf_evlist__new(struct cpu_map *cpus,
  36                                      struct thread_map *threads)
  37 {
  38         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  39
  40         if (evlist != NULL)
  41                 perf_evlist__init(evlist, cpus, threads);
  42
  43         return evlist;
  44 }
  45
  46 static void perf_evlist__purge(struct perf_evlist *evlist)
  47 {
  48         struct perf_evsel *pos, *n;
  49
  50         list_for_each_entry_safe(pos, n, &evlist->entries, node) {
  51                 list_del_init(&pos->node);
  52                 perf_evsel__delete(pos);
  53         }
  54
  55         evlist->nr_entries = 0;
  56 }
  57
  58 void perf_evlist__exit(struct perf_evlist *evlist)
  59 {
  60         free(evlist->mmap);
  61         free(evlist->pollfd);
  62         evlist->mmap = NULL;
  63         evlist->pollfd = NULL;
  64 }
  65
  66 void perf_evlist__delete(struct perf_evlist *evlist)
  67 {
  68         perf_evlist__purge(evlist);
  69         perf_evlist__exit(evlist);
  70         free(evlist);
  71 }
  72
  73 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
  74 {
  75         list_add_tail(&entry->node, &evlist->entries);
  76         ++evlist->nr_entries;
  77 }
  78
  79 int perf_evlist__add_default(struct perf_evlist *evlist)
  80 {
  81         struct perf_event_attr attr = {
  82                 .type = PERF_TYPE_HARDWARE,
  83                 .config = PERF_COUNT_HW_CPU_CYCLES,
  84         };
  85         struct perf_evsel *evsel = perf_evsel__new(&attr, 0);
  86
  87         if (evsel == NULL)
  88                 goto error;
  89
  90         /* use strdup() because free(evsel) assumes name is allocated */
  91         evsel->name = strdup("cycles");
  92         if (!evsel->name)
  93                 goto error_free;
  94
  95         perf_evlist__add(evlist, evsel);
  96         return 0;
  97 error_free:
  98         perf_evsel__delete(evsel);
  99 error:
 100         return -ENOMEM;
 101 }
 102
 103 void perf_evlist__disable(struct perf_evlist *evlist)
 104 {
 105         int cpu, thread;
 106         struct perf_evsel *pos;
 107
 108         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 109                 list_for_each_entry(pos, &evlist->entries, node) {
 110                         for (thread = 0; thread < evlist->threads->nr; thread++)
 111                                 ioctl(FD(pos, cpu, thread), PERF_EVENT_IOC_DISABLE);
 112                 }
 113         }
 114 }
 115
 116 void perf_evlist__enable(struct perf_evlist *evlist)
 117 {
 118         int cpu, thread;
 119         struct perf_evsel *pos;
 120
 121         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 122                 list_for_each_entry(pos, &evlist->entries, node) {
 123                         for (thread = 0; thread < evlist->threads->nr; thread++)
 124                                 ioctl(FD(pos, cpu, thread), PERF_EVENT_IOC_ENABLE);
 125                 }
 126         }
 127 }
 128
 129 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 130 {
 131         int nfds = evlist->cpus->nr * evlist->threads->nr * evlist->nr_entries;
 132         evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
 133         return evlist->pollfd != NULL ? 0 : -ENOMEM;
 134 }
 135
 136 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 137 {
 138         fcntl(fd, F_SETFL, O_NONBLOCK);
 139         evlist->pollfd[evlist->nr_fds].fd = fd;
 140         evlist->pollfd[evlist->nr_fds].events = POLLIN;
 141         evlist->nr_fds++;
 142 }
 143
 144 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 145                                  struct perf_evsel *evsel,
 146                                  int cpu, int thread, u64 id)
 147 {
 148         int hash;
 149         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 150
 151         sid->id = id;
 152         sid->evsel = evsel;
 153         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 154         hlist_add_head(&sid->node, &evlist->heads[hash]);
 155 }
 156
 157 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 158                          int cpu, int thread, u64 id)
 159 {
 160         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 161         evsel->id[evsel->ids++] = id;
 162 }
 163
 164 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 165                                   struct perf_evsel *evsel,
 166                                   int cpu, int thread, int fd)
 167 {
 168         u64 read_data[4] = { 0, };
 169         int id_idx = 1; /* The first entry is the counter value */
 170
 171         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 172             read(fd, &read_data, sizeof(read_data)) == -1)
 173                 return -1;
 174
 175         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 176                 ++id_idx;
 177         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 178                 ++id_idx;
 179
 180         perf_evlist__id_add(evlist, evsel, cpu, thread, read_data[id_idx]);
 181         return 0;
 182 }
 183
 184 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 185 {
 186         struct hlist_head *head;
 187         struct hlist_node *pos;
 188         struct perf_sample_id *sid;
 189         int hash;
 190
 191         if (evlist->nr_entries == 1)
 192                 return list_entry(evlist->entries.next, struct perf_evsel, node);
 193
 194         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 195         head = &evlist->heads[hash];
 196
 197         hlist_for_each_entry(sid, pos, head, node)
 198                 if (sid->id == id)
 199                         return sid->evsel;
 200         return NULL;
 201 }
 202
 203 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 204 {
 205         /* XXX Move this to perf.c, making it generally available */
 206         unsigned int page_size = sysconf(_SC_PAGE_SIZE);
 207         struct perf_mmap *md = &evlist->mmap[idx];
 208         unsigned int head = perf_mmap__read_head(md);
 209         unsigned int old = md->prev;
 210         unsigned char *data = md->base + page_size;
 211         union perf_event *event = NULL;
 212
 213         if (evlist->overwrite) {
 214                 /*
 215                  * If we're further behind than half the buffer, there's a chance
 216                  * the writer will bite our tail and mess up the samples under us.
 217                  *
 218                  * If we somehow ended up ahead of the head, we got messed up.
 219                  *
 220                  * In either case, truncate and restart at head.
 221                  */
 222                 int diff = head - old;
 223                 if (diff > md->mask / 2 || diff < 0) {
 224                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 225
 226                         /*
 227                          * head points to a known good entry, start there.
 228                          */
 229                         old = head;
 230                 }
 231         }
 232
 233         if (old != head) {
 234                 size_t size;
 235
 236                 event = (union perf_event *)&data[old & md->mask];
 237                 size = event->header.size;
 238
 239                 /*
 240                  * Event straddles the mmap boundary -- header should always
 241                  * be inside due to u64 alignment of output.
 242                  */
 243                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 244                         unsigned int offset = old;
 245                         unsigned int len = min(sizeof(*event), size), cpy;
 246                         void *dst = &evlist->event_copy;
 247
 248                         do {
 249                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 250                                 memcpy(dst, &data[offset & md->mask], cpy);
 251                                 offset += cpy;
 252                                 dst += cpy;
 253                                 len -= cpy;
 254                         } while (len);
 255
 256                         event = &evlist->event_copy;
 257                 }
 258
 259                 old += size;
 260         }
 261
 262         md->prev = old;
 263
 264         if (!evlist->overwrite)
 265                 perf_mmap__write_tail(md, old);
 266
 267         return event;
 268 }
 269
 270 void perf_evlist__munmap(struct perf_evlist *evlist)
 271 {
 272         int i;
 273
 274         for (i = 0; i < evlist->nr_mmaps; i++) {
 275                 if (evlist->mmap[i].base != NULL) {
 276                         munmap(evlist->mmap[i].base, evlist->mmap_len);
 277                         evlist->mmap[i].base = NULL;
 278                 }
 279         }
 280
 281         free(evlist->mmap);
 282         evlist->mmap = NULL;
 283 }
 284
 285 int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 286 {
 287         evlist->nr_mmaps = evlist->cpus->nr;
 288         if (evlist->cpus->map[0] == -1)
 289                 evlist->nr_mmaps = evlist->threads->nr;
 290         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 291         return evlist->mmap != NULL ? 0 : -ENOMEM;
 292 }
 293
 294 static int __perf_evlist__mmap(struct perf_evlist *evlist,
 295                                int idx, int prot, int mask, int fd)
 296 {
 297         evlist->mmap[idx].prev = 0;
 298         evlist->mmap[idx].mask = mask;
 299         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
 300                                       MAP_SHARED, fd, 0);
 301         if (evlist->mmap[idx].base == MAP_FAILED)
 302                 return -1;
 303
 304         perf_evlist__add_pollfd(evlist, fd);
 305         return 0;
 306 }
 307
 308 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot, int mask)
 309 {
 310         struct perf_evsel *evsel;
 311         int cpu, thread;
 312
 313         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 314                 int output = -1;
 315
 316                 for (thread = 0; thread < evlist->threads->nr; thread++) {
 317                         list_for_each_entry(evsel, &evlist->entries, node) {
 318                                 int fd = FD(evsel, cpu, thread);
 319
 320                                 if (output == -1) {
 321                                         output = fd;
 322                                         if (__perf_evlist__mmap(evlist, cpu,
 323                                                                 prot, mask, output) < 0)
 324                                                 goto out_unmap;
 325                                 } else {
 326                                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
 327                                                 goto out_unmap;
 328                                 }
 329
 330                                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 331                                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
 332                                         goto out_unmap;
 333                         }
 334                 }
 335         }
 336
 337         return 0;
 338
 339 out_unmap:
 340         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 341                 if (evlist->mmap[cpu].base != NULL) {
 342                         munmap(evlist->mmap[cpu].base, evlist->mmap_len);
 343                         evlist->mmap[cpu].base = NULL;
 344                 }
 345         }
 346         return -1;
 347 }
 348
 349 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot, int mask)
 350 {
 351         struct perf_evsel *evsel;
 352         int thread;
 353
 354         for (thread = 0; thread < evlist->threads->nr; thread++) {
 355                 int output = -1;
 356
 357                 list_for_each_entry(evsel, &evlist->entries, node) {
 358                         int fd = FD(evsel, 0, thread);
 359
 360                         if (output == -1) {
 361                                 output = fd;
 362                                 if (__perf_evlist__mmap(evlist, thread,
 363                                                         prot, mask, output) < 0)
 364                                         goto out_unmap;
 365                         } else {
 366                                 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
 367                                         goto out_unmap;
 368                         }
 369
 370                         if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 371                             perf_evlist__id_add_fd(evlist, evsel, 0, thread, fd) < 0)
 372                                 goto out_unmap;
 373                 }
 374         }
 375
 376         return 0;
 377
 378 out_unmap:
 379         for (thread = 0; thread < evlist->threads->nr; thread++) {
 380                 if (evlist->mmap[thread].base != NULL) {
 381                         munmap(evlist->mmap[thread].base, evlist->mmap_len);
 382                         evlist->mmap[thread].base = NULL;
 383                 }
 384         }
 385         return -1;
 386 }
 387
 388 /** perf_evlist__mmap - Create per cpu maps to receive events
 389  *
 390  * @evlist - list of events
 391  * @pages - map length in pages
 392  * @overwrite - overwrite older events?
 393  *
 394  * If overwrite is false the user needs to signal event consuption using:
 395  *
 396  *      struct perf_mmap *m = &evlist->mmap[cpu];
 397  *      unsigned int head = perf_mmap__read_head(m);
 398  *
 399  *      perf_mmap__write_tail(m, head)
 400  *
 401  * Using perf_evlist__read_on_cpu does this automatically.
 402  */
 403 int perf_evlist__mmap(struct perf_evlist *evlist, int pages, bool overwrite)
 404 {
 405         unsigned int page_size = sysconf(_SC_PAGE_SIZE);
 406         int mask = pages * page_size - 1;
 407         struct perf_evsel *evsel;
 408         const struct cpu_map *cpus = evlist->cpus;
 409         const struct thread_map *threads = evlist->threads;
 410         int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE);
 411
 412         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 413                 return -ENOMEM;
 414
 415         if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 416                 return -ENOMEM;
 417
 418         evlist->overwrite = overwrite;
 419         evlist->mmap_len = (pages + 1) * page_size;
 420
 421         list_for_each_entry(evsel, &evlist->entries, node) {
 422                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 423                     evsel->sample_id == NULL &&
 424                     perf_evsel__alloc_id(evsel, cpus->nr, threads->nr) < 0)
 425                         return -ENOMEM;
 426         }
 427
 428         if (evlist->cpus->map[0] == -1)
 429                 return perf_evlist__mmap_per_thread(evlist, prot, mask);
 430
 431         return perf_evlist__mmap_per_cpu(evlist, prot, mask);
 432 }
 433
 434 int perf_evlist__create_maps(struct perf_evlist *evlist, pid_t target_pid,
 435                              pid_t target_tid, const char *cpu_list)
 436 {
 437         evlist->threads = thread_map__new(target_pid, target_tid);
 438
 439         if (evlist->threads == NULL)
 440                 return -1;
 441
 442         if (cpu_list == NULL && target_tid != -1)
 443                 evlist->cpus = cpu_map__dummy_new();
 444         else
 445                 evlist->cpus = cpu_map__new(cpu_list);
 446
 447         if (evlist->cpus == NULL)
 448                 goto out_delete_threads;
 449
 450         return 0;
 451
 452 out_delete_threads:
 453         thread_map__delete(evlist->threads);
 454         return -1;
 455 }
 456
 457 void perf_evlist__delete_maps(struct perf_evlist *evlist)
 458 {
 459         cpu_map__delete(evlist->cpus);
 460         thread_map__delete(evlist->threads);
 461         evlist->cpus    = NULL;
 462         evlist->threads = NULL;
 463 }
 464
 465 int perf_evlist__set_filters(struct perf_evlist *evlist)
 466 {
 467         const struct thread_map *threads = evlist->threads;
 468         const struct cpu_map *cpus = evlist->cpus;
 469         struct perf_evsel *evsel;
 470         char *filter;
 471         int thread;
 472         int cpu;
 473         int err;
 474         int fd;
 475
 476         list_for_each_entry(evsel, &evlist->entries, node) {
 477                 filter = evsel->filter;
 478                 if (!filter)
 479                         continue;
 480                 for (cpu = 0; cpu < cpus->nr; cpu++) {
 481                         for (thread = 0; thread < threads->nr; thread++) {
 482                                 fd = FD(evsel, cpu, thread);
 483                                 err = ioctl(fd, PERF_EVENT_IOC_SET_FILTER, filter);
 484                                 if (err)
 485                                         return err;
 486                         }
 487                 }
 488         }
 489
 490         return 0;
 491 }
 492
 493 bool perf_evlist__valid_sample_type(const struct perf_evlist *evlist)
 494 {
 495         struct perf_evsel *pos, *first;
 496
 497         pos = first = list_entry(evlist->entries.next, struct perf_evsel, node);
 498
 499         list_for_each_entry_continue(pos, &evlist->entries, node) {
 500                 if (first->attr.sample_type != pos->attr.sample_type)
 501                         return false;
 502         }
 503
 504         return true;
 505 }
 506
 507 u64 perf_evlist__sample_type(const struct perf_evlist *evlist)
 508 {
 509         struct perf_evsel *first;
 510
 511         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 512         return first->attr.sample_type;
 513 }
 514
 515 bool perf_evlist__valid_sample_id_all(const struct perf_evlist *evlist)
 516 {
 517         struct perf_evsel *pos, *first;
 518
 519         pos = first = list_entry(evlist->entries.next, struct perf_evsel, node);
 520
 521         list_for_each_entry_continue(pos, &evlist->entries, node) {
 522                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
 523                         return false;
 524         }
 525
 526         return true;
 527 }
 528
 529 bool perf_evlist__sample_id_all(const struct perf_evlist *evlist)
 530 {
 531         struct perf_evsel *first;
 532
 533         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 534         return first->attr.sample_id_all;
 535 }