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                 return -ENOMEM;
  89
  90         perf_evlist__add(evlist, evsel);
  91         return 0;
  92 }
  93
  94 void perf_evlist__disable(struct perf_evlist *evlist)
  95 {
  96         int cpu, thread;
  97         struct perf_evsel *pos;
  98
  99         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 100                 list_for_each_entry(pos, &evlist->entries, node) {
 101                         for (thread = 0; thread < evlist->threads->nr; thread++)
 102                                 ioctl(FD(pos, cpu, thread), PERF_EVENT_IOC_DISABLE);
 103                 }
 104         }
 105 }
 106
 107 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 108 {
 109         int nfds = evlist->cpus->nr * evlist->threads->nr * evlist->nr_entries;
 110         evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
 111         return evlist->pollfd != NULL ? 0 : -ENOMEM;
 112 }
 113
 114 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 115 {
 116         fcntl(fd, F_SETFL, O_NONBLOCK);
 117         evlist->pollfd[evlist->nr_fds].fd = fd;
 118         evlist->pollfd[evlist->nr_fds].events = POLLIN;
 119         evlist->nr_fds++;
 120 }
 121
 122 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 123                                  struct perf_evsel *evsel,
 124                                  int cpu, int thread, u64 id)
 125 {
 126         int hash;
 127         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 128
 129         sid->id = id;
 130         sid->evsel = evsel;
 131         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 132         hlist_add_head(&sid->node, &evlist->heads[hash]);
 133 }
 134
 135 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 136                          int cpu, int thread, u64 id)
 137 {
 138         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 139         evsel->id[evsel->ids++] = id;
 140 }
 141
 142 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 143                                   struct perf_evsel *evsel,
 144                                   int cpu, int thread, int fd)
 145 {
 146         u64 read_data[4] = { 0, };
 147         int id_idx = 1; /* The first entry is the counter value */
 148
 149         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 150             read(fd, &read_data, sizeof(read_data)) == -1)
 151                 return -1;
 152
 153         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 154                 ++id_idx;
 155         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 156                 ++id_idx;
 157
 158         perf_evlist__id_add(evlist, evsel, cpu, thread, read_data[id_idx]);
 159         return 0;
 160 }
 161
 162 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 163 {
 164         struct hlist_head *head;
 165         struct hlist_node *pos;
 166         struct perf_sample_id *sid;
 167         int hash;
 168
 169         if (evlist->nr_entries == 1)
 170                 return list_entry(evlist->entries.next, struct perf_evsel, node);
 171
 172         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 173         head = &evlist->heads[hash];
 174
 175         hlist_for_each_entry(sid, pos, head, node)
 176                 if (sid->id == id)
 177                         return sid->evsel;
 178         return NULL;
 179 }
 180
 181 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 182 {
 183         /* XXX Move this to perf.c, making it generally available */
 184         unsigned int page_size = sysconf(_SC_PAGE_SIZE);
 185         struct perf_mmap *md = &evlist->mmap[idx];
 186         unsigned int head = perf_mmap__read_head(md);
 187         unsigned int old = md->prev;
 188         unsigned char *data = md->base + page_size;
 189         union perf_event *event = NULL;
 190
 191         if (evlist->overwrite) {
 192                 /*
 193                  * If we're further behind than half the buffer, there's a chance
 194                  * the writer will bite our tail and mess up the samples under us.
 195                  *
 196                  * If we somehow ended up ahead of the head, we got messed up.
 197                  *
 198                  * In either case, truncate and restart at head.
 199                  */
 200                 int diff = head - old;
 201                 if (diff > md->mask / 2 || diff < 0) {
 202                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 203
 204                         /*
 205                          * head points to a known good entry, start there.
 206                          */
 207                         old = head;
 208                 }
 209         }
 210
 211         if (old != head) {
 212                 size_t size;
 213
 214                 event = (union perf_event *)&data[old & md->mask];
 215                 size = event->header.size;
 216
 217                 /*
 218                  * Event straddles the mmap boundary -- header should always
 219                  * be inside due to u64 alignment of output.
 220                  */
 221                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 222                         unsigned int offset = old;
 223                         unsigned int len = min(sizeof(*event), size), cpy;
 224                         void *dst = &evlist->event_copy;
 225
 226                         do {
 227                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 228                                 memcpy(dst, &data[offset & md->mask], cpy);
 229                                 offset += cpy;
 230                                 dst += cpy;
 231                                 len -= cpy;
 232                         } while (len);
 233
 234                         event = &evlist->event_copy;
 235                 }
 236
 237                 old += size;
 238         }
 239
 240         md->prev = old;
 241
 242         if (!evlist->overwrite)
 243                 perf_mmap__write_tail(md, old);
 244
 245         return event;
 246 }
 247
 248 void perf_evlist__munmap(struct perf_evlist *evlist)
 249 {
 250         int i;
 251
 252         for (i = 0; i < evlist->nr_mmaps; i++) {
 253                 if (evlist->mmap[i].base != NULL) {
 254                         munmap(evlist->mmap[i].base, evlist->mmap_len);
 255                         evlist->mmap[i].base = NULL;
 256                 }
 257         }
 258
 259         free(evlist->mmap);
 260         evlist->mmap = NULL;
 261 }
 262
 263 int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 264 {
 265         evlist->nr_mmaps = evlist->cpus->nr;
 266         if (evlist->cpus->map[0] == -1)
 267                 evlist->nr_mmaps = evlist->threads->nr;
 268         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 269         return evlist->mmap != NULL ? 0 : -ENOMEM;
 270 }
 271
 272 static int __perf_evlist__mmap(struct perf_evlist *evlist,
 273                                int idx, int prot, int mask, int fd)
 274 {
 275         evlist->mmap[idx].prev = 0;
 276         evlist->mmap[idx].mask = mask;
 277         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
 278                                       MAP_SHARED, fd, 0);
 279         if (evlist->mmap[idx].base == MAP_FAILED)
 280                 return -1;
 281
 282         perf_evlist__add_pollfd(evlist, fd);
 283         return 0;
 284 }
 285
 286 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot, int mask)
 287 {
 288         struct perf_evsel *evsel;
 289         int cpu, thread;
 290
 291         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 292                 int output = -1;
 293
 294                 for (thread = 0; thread < evlist->threads->nr; thread++) {
 295                         list_for_each_entry(evsel, &evlist->entries, node) {
 296                                 int fd = FD(evsel, cpu, thread);
 297
 298                                 if (output == -1) {
 299                                         output = fd;
 300                                         if (__perf_evlist__mmap(evlist, cpu,
 301                                                                 prot, mask, output) < 0)
 302                                                 goto out_unmap;
 303                                 } else {
 304                                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
 305                                                 goto out_unmap;
 306                                 }
 307
 308                                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 309                                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
 310                                         goto out_unmap;
 311                         }
 312                 }
 313         }
 314
 315         return 0;
 316
 317 out_unmap:
 318         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 319                 if (evlist->mmap[cpu].base != NULL) {
 320                         munmap(evlist->mmap[cpu].base, evlist->mmap_len);
 321                         evlist->mmap[cpu].base = NULL;
 322                 }
 323         }
 324         return -1;
 325 }
 326
 327 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot, int mask)
 328 {
 329         struct perf_evsel *evsel;
 330         int thread;
 331
 332         for (thread = 0; thread < evlist->threads->nr; thread++) {
 333                 int output = -1;
 334
 335                 list_for_each_entry(evsel, &evlist->entries, node) {
 336                         int fd = FD(evsel, 0, thread);
 337
 338                         if (output == -1) {
 339                                 output = fd;
 340                                 if (__perf_evlist__mmap(evlist, thread,
 341                                                         prot, mask, output) < 0)
 342                                         goto out_unmap;
 343                         } else {
 344                                 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
 345                                         goto out_unmap;
 346                         }
 347
 348                         if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 349                             perf_evlist__id_add_fd(evlist, evsel, 0, thread, fd) < 0)
 350                                 goto out_unmap;
 351                 }
 352         }
 353
 354         return 0;
 355
 356 out_unmap:
 357         for (thread = 0; thread < evlist->threads->nr; thread++) {
 358                 if (evlist->mmap[thread].base != NULL) {
 359                         munmap(evlist->mmap[thread].base, evlist->mmap_len);
 360                         evlist->mmap[thread].base = NULL;
 361                 }
 362         }
 363         return -1;
 364 }
 365
 366 /** perf_evlist__mmap - Create per cpu maps to receive events
 367  *
 368  * @evlist - list of events
 369  * @pages - map length in pages
 370  * @overwrite - overwrite older events?
 371  *
 372  * If overwrite is false the user needs to signal event consuption using:
 373  *
 374  *      struct perf_mmap *m = &evlist->mmap[cpu];
 375  *      unsigned int head = perf_mmap__read_head(m);
 376  *
 377  *      perf_mmap__write_tail(m, head)
 378  *
 379  * Using perf_evlist__read_on_cpu does this automatically.
 380  */
 381 int perf_evlist__mmap(struct perf_evlist *evlist, int pages, bool overwrite)
 382 {
 383         unsigned int page_size = sysconf(_SC_PAGE_SIZE);
 384         int mask = pages * page_size - 1;
 385         struct perf_evsel *evsel;
 386         const struct cpu_map *cpus = evlist->cpus;
 387         const struct thread_map *threads = evlist->threads;
 388         int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE);
 389
 390         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 391                 return -ENOMEM;
 392
 393         if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 394                 return -ENOMEM;
 395
 396         evlist->overwrite = overwrite;
 397         evlist->mmap_len = (pages + 1) * page_size;
 398
 399         list_for_each_entry(evsel, &evlist->entries, node) {
 400                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 401                     evsel->sample_id == NULL &&
 402                     perf_evsel__alloc_id(evsel, cpus->nr, threads->nr) < 0)
 403                         return -ENOMEM;
 404         }
 405
 406         if (evlist->cpus->map[0] == -1)
 407                 return perf_evlist__mmap_per_thread(evlist, prot, mask);
 408
 409         return perf_evlist__mmap_per_cpu(evlist, prot, mask);
 410 }
 411
 412 int perf_evlist__create_maps(struct perf_evlist *evlist, pid_t target_pid,
 413                              pid_t target_tid, const char *cpu_list)
 414 {
 415         evlist->threads = thread_map__new(target_pid, target_tid);
 416
 417         if (evlist->threads == NULL)
 418                 return -1;
 419
 420         if (cpu_list == NULL && target_tid != -1)
 421                 evlist->cpus = cpu_map__dummy_new();
 422         else
 423                 evlist->cpus = cpu_map__new(cpu_list);
 424
 425         if (evlist->cpus == NULL)
 426                 goto out_delete_threads;
 427
 428         return 0;
 429
 430 out_delete_threads:
 431         thread_map__delete(evlist->threads);
 432         return -1;
 433 }
 434
 435 void perf_evlist__delete_maps(struct perf_evlist *evlist)
 436 {
 437         cpu_map__delete(evlist->cpus);
 438         thread_map__delete(evlist->threads);
 439         evlist->cpus    = NULL;
 440         evlist->threads = NULL;
 441 }
 442
 443 int perf_evlist__set_filters(struct perf_evlist *evlist)
 444 {
 445         const struct thread_map *threads = evlist->threads;
 446         const struct cpu_map *cpus = evlist->cpus;
 447         struct perf_evsel *evsel;
 448         char *filter;
 449         int thread;
 450         int cpu;
 451         int err;
 452         int fd;
 453
 454         list_for_each_entry(evsel, &evlist->entries, node) {
 455                 filter = evsel->filter;
 456                 if (!filter)
 457                         continue;
 458                 for (cpu = 0; cpu < cpus->nr; cpu++) {
 459                         for (thread = 0; thread < threads->nr; thread++) {
 460                                 fd = FD(evsel, cpu, thread);
 461                                 err = ioctl(fd, PERF_EVENT_IOC_SET_FILTER, filter);
 462                                 if (err)
 463                                         return err;
 464                         }
 465                 }
 466         }
 467
 468         return 0;
 469 }
 470
 471 bool perf_evlist__valid_sample_type(const struct perf_evlist *evlist)
 472 {
 473         struct perf_evsel *pos, *first;
 474
 475         pos = first = list_entry(evlist->entries.next, struct perf_evsel, node);
 476
 477         list_for_each_entry_continue(pos, &evlist->entries, node) {
 478                 if (first->attr.sample_type != pos->attr.sample_type)
 479                         return false;
 480         }
 481
 482         return true;
 483 }
 484
 485 u64 perf_evlist__sample_type(const struct perf_evlist *evlist)
 486 {
 487         struct perf_evsel *first;
 488
 489         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 490         return first->attr.sample_type;
 491 }
 492
 493 bool perf_evlist__valid_sample_id_all(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_id_all != pos->attr.sample_id_all)
 501                         return false;
 502         }
 503
 504         return true;
 505 }
 506
 507 bool perf_evlist__sample_id_all(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_id_all;
 513 }