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 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 117 {
 118         int nfds = evlist->cpus->nr * evlist->threads->nr * evlist->nr_entries;
 119         evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
 120         return evlist->pollfd != NULL ? 0 : -ENOMEM;
 121 }
 122
 123 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 124 {
 125         fcntl(fd, F_SETFL, O_NONBLOCK);
 126         evlist->pollfd[evlist->nr_fds].fd = fd;
 127         evlist->pollfd[evlist->nr_fds].events = POLLIN;
 128         evlist->nr_fds++;
 129 }
 130
 131 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 132                                  struct perf_evsel *evsel,
 133                                  int cpu, int thread, u64 id)
 134 {
 135         int hash;
 136         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 137
 138         sid->id = id;
 139         sid->evsel = evsel;
 140         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 141         hlist_add_head(&sid->node, &evlist->heads[hash]);
 142 }
 143
 144 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 145                          int cpu, int thread, u64 id)
 146 {
 147         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 148         evsel->id[evsel->ids++] = id;
 149 }
 150
 151 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 152                                   struct perf_evsel *evsel,
 153                                   int cpu, int thread, int fd)
 154 {
 155         u64 read_data[4] = { 0, };
 156         int id_idx = 1; /* The first entry is the counter value */
 157
 158         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 159             read(fd, &read_data, sizeof(read_data)) == -1)
 160                 return -1;
 161
 162         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 163                 ++id_idx;
 164         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 165                 ++id_idx;
 166
 167         perf_evlist__id_add(evlist, evsel, cpu, thread, read_data[id_idx]);
 168         return 0;
 169 }
 170
 171 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 172 {
 173         struct hlist_head *head;
 174         struct hlist_node *pos;
 175         struct perf_sample_id *sid;
 176         int hash;
 177
 178         if (evlist->nr_entries == 1)
 179                 return list_entry(evlist->entries.next, struct perf_evsel, node);
 180
 181         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 182         head = &evlist->heads[hash];
 183
 184         hlist_for_each_entry(sid, pos, head, node)
 185                 if (sid->id == id)
 186                         return sid->evsel;
 187         return NULL;
 188 }
 189
 190 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 191 {
 192         /* XXX Move this to perf.c, making it generally available */
 193         unsigned int page_size = sysconf(_SC_PAGE_SIZE);
 194         struct perf_mmap *md = &evlist->mmap[idx];
 195         unsigned int head = perf_mmap__read_head(md);
 196         unsigned int old = md->prev;
 197         unsigned char *data = md->base + page_size;
 198         union perf_event *event = NULL;
 199
 200         if (evlist->overwrite) {
 201                 /*
 202                  * If we're further behind than half the buffer, there's a chance
 203                  * the writer will bite our tail and mess up the samples under us.
 204                  *
 205                  * If we somehow ended up ahead of the head, we got messed up.
 206                  *
 207                  * In either case, truncate and restart at head.
 208                  */
 209                 int diff = head - old;
 210                 if (diff > md->mask / 2 || diff < 0) {
 211                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 212
 213                         /*
 214                          * head points to a known good entry, start there.
 215                          */
 216                         old = head;
 217                 }
 218         }
 219
 220         if (old != head) {
 221                 size_t size;
 222
 223                 event = (union perf_event *)&data[old & md->mask];
 224                 size = event->header.size;
 225
 226                 /*
 227                  * Event straddles the mmap boundary -- header should always
 228                  * be inside due to u64 alignment of output.
 229                  */
 230                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 231                         unsigned int offset = old;
 232                         unsigned int len = min(sizeof(*event), size), cpy;
 233                         void *dst = &evlist->event_copy;
 234
 235                         do {
 236                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 237                                 memcpy(dst, &data[offset & md->mask], cpy);
 238                                 offset += cpy;
 239                                 dst += cpy;
 240                                 len -= cpy;
 241                         } while (len);
 242
 243                         event = &evlist->event_copy;
 244                 }
 245
 246                 old += size;
 247         }
 248
 249         md->prev = old;
 250
 251         if (!evlist->overwrite)
 252                 perf_mmap__write_tail(md, old);
 253
 254         return event;
 255 }
 256
 257 void perf_evlist__munmap(struct perf_evlist *evlist)
 258 {
 259         int i;
 260
 261         for (i = 0; i < evlist->nr_mmaps; i++) {
 262                 if (evlist->mmap[i].base != NULL) {
 263                         munmap(evlist->mmap[i].base, evlist->mmap_len);
 264                         evlist->mmap[i].base = NULL;
 265                 }
 266         }
 267
 268         free(evlist->mmap);
 269         evlist->mmap = NULL;
 270 }
 271
 272 int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 273 {
 274         evlist->nr_mmaps = evlist->cpus->nr;
 275         if (evlist->cpus->map[0] == -1)
 276                 evlist->nr_mmaps = evlist->threads->nr;
 277         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 278         return evlist->mmap != NULL ? 0 : -ENOMEM;
 279 }
 280
 281 static int __perf_evlist__mmap(struct perf_evlist *evlist,
 282                                int idx, int prot, int mask, int fd)
 283 {
 284         evlist->mmap[idx].prev = 0;
 285         evlist->mmap[idx].mask = mask;
 286         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
 287                                       MAP_SHARED, fd, 0);
 288         if (evlist->mmap[idx].base == MAP_FAILED)
 289                 return -1;
 290
 291         perf_evlist__add_pollfd(evlist, fd);
 292         return 0;
 293 }
 294
 295 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot, int mask)
 296 {
 297         struct perf_evsel *evsel;
 298         int cpu, thread;
 299
 300         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 301                 int output = -1;
 302
 303                 for (thread = 0; thread < evlist->threads->nr; thread++) {
 304                         list_for_each_entry(evsel, &evlist->entries, node) {
 305                                 int fd = FD(evsel, cpu, thread);
 306
 307                                 if (output == -1) {
 308                                         output = fd;
 309                                         if (__perf_evlist__mmap(evlist, cpu,
 310                                                                 prot, mask, output) < 0)
 311                                                 goto out_unmap;
 312                                 } else {
 313                                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
 314                                                 goto out_unmap;
 315                                 }
 316
 317                                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 318                                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
 319                                         goto out_unmap;
 320                         }
 321                 }
 322         }
 323
 324         return 0;
 325
 326 out_unmap:
 327         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 328                 if (evlist->mmap[cpu].base != NULL) {
 329                         munmap(evlist->mmap[cpu].base, evlist->mmap_len);
 330                         evlist->mmap[cpu].base = NULL;
 331                 }
 332         }
 333         return -1;
 334 }
 335
 336 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot, int mask)
 337 {
 338         struct perf_evsel *evsel;
 339         int thread;
 340
 341         for (thread = 0; thread < evlist->threads->nr; thread++) {
 342                 int output = -1;
 343
 344                 list_for_each_entry(evsel, &evlist->entries, node) {
 345                         int fd = FD(evsel, 0, thread);
 346
 347                         if (output == -1) {
 348                                 output = fd;
 349                                 if (__perf_evlist__mmap(evlist, thread,
 350                                                         prot, mask, output) < 0)
 351                                         goto out_unmap;
 352                         } else {
 353                                 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
 354                                         goto out_unmap;
 355                         }
 356
 357                         if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 358                             perf_evlist__id_add_fd(evlist, evsel, 0, thread, fd) < 0)
 359                                 goto out_unmap;
 360                 }
 361         }
 362
 363         return 0;
 364
 365 out_unmap:
 366         for (thread = 0; thread < evlist->threads->nr; thread++) {
 367                 if (evlist->mmap[thread].base != NULL) {
 368                         munmap(evlist->mmap[thread].base, evlist->mmap_len);
 369                         evlist->mmap[thread].base = NULL;
 370                 }
 371         }
 372         return -1;
 373 }
 374
 375 /** perf_evlist__mmap - Create per cpu maps to receive events
 376  *
 377  * @evlist - list of events
 378  * @pages - map length in pages
 379  * @overwrite - overwrite older events?
 380  *
 381  * If overwrite is false the user needs to signal event consuption using:
 382  *
 383  *      struct perf_mmap *m = &evlist->mmap[cpu];
 384  *      unsigned int head = perf_mmap__read_head(m);
 385  *
 386  *      perf_mmap__write_tail(m, head)
 387  *
 388  * Using perf_evlist__read_on_cpu does this automatically.
 389  */
 390 int perf_evlist__mmap(struct perf_evlist *evlist, int pages, bool overwrite)
 391 {
 392         unsigned int page_size = sysconf(_SC_PAGE_SIZE);
 393         int mask = pages * page_size - 1;
 394         struct perf_evsel *evsel;
 395         const struct cpu_map *cpus = evlist->cpus;
 396         const struct thread_map *threads = evlist->threads;
 397         int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE);
 398
 399         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 400                 return -ENOMEM;
 401
 402         if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 403                 return -ENOMEM;
 404
 405         evlist->overwrite = overwrite;
 406         evlist->mmap_len = (pages + 1) * page_size;
 407
 408         list_for_each_entry(evsel, &evlist->entries, node) {
 409                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 410                     evsel->sample_id == NULL &&
 411                     perf_evsel__alloc_id(evsel, cpus->nr, threads->nr) < 0)
 412                         return -ENOMEM;
 413         }
 414
 415         if (evlist->cpus->map[0] == -1)
 416                 return perf_evlist__mmap_per_thread(evlist, prot, mask);
 417
 418         return perf_evlist__mmap_per_cpu(evlist, prot, mask);
 419 }
 420
 421 int perf_evlist__create_maps(struct perf_evlist *evlist, pid_t target_pid,
 422                              pid_t target_tid, const char *cpu_list)
 423 {
 424         evlist->threads = thread_map__new(target_pid, target_tid);
 425
 426         if (evlist->threads == NULL)
 427                 return -1;
 428
 429         if (cpu_list == NULL && target_tid != -1)
 430                 evlist->cpus = cpu_map__dummy_new();
 431         else
 432                 evlist->cpus = cpu_map__new(cpu_list);
 433
 434         if (evlist->cpus == NULL)
 435                 goto out_delete_threads;
 436
 437         return 0;
 438
 439 out_delete_threads:
 440         thread_map__delete(evlist->threads);
 441         return -1;
 442 }
 443
 444 void perf_evlist__delete_maps(struct perf_evlist *evlist)
 445 {
 446         cpu_map__delete(evlist->cpus);
 447         thread_map__delete(evlist->threads);
 448         evlist->cpus    = NULL;
 449         evlist->threads = NULL;
 450 }
 451
 452 int perf_evlist__set_filters(struct perf_evlist *evlist)
 453 {
 454         const struct thread_map *threads = evlist->threads;
 455         const struct cpu_map *cpus = evlist->cpus;
 456         struct perf_evsel *evsel;
 457         char *filter;
 458         int thread;
 459         int cpu;
 460         int err;
 461         int fd;
 462
 463         list_for_each_entry(evsel, &evlist->entries, node) {
 464                 filter = evsel->filter;
 465                 if (!filter)
 466                         continue;
 467                 for (cpu = 0; cpu < cpus->nr; cpu++) {
 468                         for (thread = 0; thread < threads->nr; thread++) {
 469                                 fd = FD(evsel, cpu, thread);
 470                                 err = ioctl(fd, PERF_EVENT_IOC_SET_FILTER, filter);
 471                                 if (err)
 472                                         return err;
 473                         }
 474                 }
 475         }
 476
 477         return 0;
 478 }
 479
 480 bool perf_evlist__valid_sample_type(const struct perf_evlist *evlist)
 481 {
 482         struct perf_evsel *pos, *first;
 483
 484         pos = first = list_entry(evlist->entries.next, struct perf_evsel, node);
 485
 486         list_for_each_entry_continue(pos, &evlist->entries, node) {
 487                 if (first->attr.sample_type != pos->attr.sample_type)
 488                         return false;
 489         }
 490
 491         return true;
 492 }
 493
 494 u64 perf_evlist__sample_type(const struct perf_evlist *evlist)
 495 {
 496         struct perf_evsel *first;
 497
 498         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 499         return first->attr.sample_type;
 500 }
 501
 502 bool perf_evlist__valid_sample_id_all(const struct perf_evlist *evlist)
 503 {
 504         struct perf_evsel *pos, *first;
 505
 506         pos = first = list_entry(evlist->entries.next, struct perf_evsel, node);
 507
 508         list_for_each_entry_continue(pos, &evlist->entries, node) {
 509                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
 510                         return false;
 511         }
 512
 513         return true;
 514 }
 515
 516 bool perf_evlist__sample_id_all(const struct perf_evlist *evlist)
 517 {
 518         struct perf_evsel *first;
 519
 520         first = list_entry(evlist->entries.next, struct perf_evsel, node);
 521         return first->attr.sample_id_all;
 522 }