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->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 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
  95 {
  96         int nfds = evlist->cpus->nr * evlist->threads->nr * evlist->nr_entries;
  97         evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
  98         return evlist->pollfd != NULL ? 0 : -ENOMEM;
  99 }
 100
 101 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 102 {
 103         fcntl(fd, F_SETFL, O_NONBLOCK);
 104         evlist->pollfd[evlist->nr_fds].fd = fd;
 105         evlist->pollfd[evlist->nr_fds].events = POLLIN;
 106         evlist->nr_fds++;
 107 }
 108
 109 static int perf_evlist__id_hash(struct perf_evlist *evlist, struct perf_evsel *evsel,
 110                                int cpu, int thread, int fd)
 111 {
 112         struct perf_sample_id *sid;
 113         u64 read_data[4] = { 0, };
 114         int hash, id_idx = 1; /* The first entry is the counter value */
 115
 116         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 117             read(fd, &read_data, sizeof(read_data)) == -1)
 118                 return -1;
 119
 120         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 121                 ++id_idx;
 122         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 123                 ++id_idx;
 124
 125         sid = SID(evsel, cpu, thread);
 126         sid->id = read_data[id_idx];
 127         sid->evsel = evsel;
 128         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 129         hlist_add_head(&sid->node, &evlist->heads[hash]);
 130         return 0;
 131 }
 132
 133 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 134 {
 135         struct hlist_head *head;
 136         struct hlist_node *pos;
 137         struct perf_sample_id *sid;
 138         int hash;
 139
 140         if (evlist->nr_entries == 1)
 141                 return list_entry(evlist->entries.next, struct perf_evsel, node);
 142
 143         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 144         head = &evlist->heads[hash];
 145
 146         hlist_for_each_entry(sid, pos, head, node)
 147                 if (sid->id == id)
 148                         return sid->evsel;
 149         return NULL;
 150 }
 151
 152 union perf_event *perf_evlist__read_on_cpu(struct perf_evlist *evlist, int cpu)
 153 {
 154         /* XXX Move this to perf.c, making it generally available */
 155         unsigned int page_size = sysconf(_SC_PAGE_SIZE);
 156         struct perf_mmap *md = &evlist->mmap[cpu];
 157         unsigned int head = perf_mmap__read_head(md);
 158         unsigned int old = md->prev;
 159         unsigned char *data = md->base + page_size;
 160         union perf_event *event = NULL;
 161
 162         if (evlist->overwrite) {
 163                 /*
 164                  * If we're further behind than half the buffer, there's a chance
 165                  * the writer will bite our tail and mess up the samples under us.
 166                  *
 167                  * If we somehow ended up ahead of the head, we got messed up.
 168                  *
 169                  * In either case, truncate and restart at head.
 170                  */
 171                 int diff = head - old;
 172                 if (diff > md->mask / 2 || diff < 0) {
 173                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 174
 175                         /*
 176                          * head points to a known good entry, start there.
 177                          */
 178                         old = head;
 179                 }
 180         }
 181
 182         if (old != head) {
 183                 size_t size;
 184
 185                 event = (union perf_event *)&data[old & md->mask];
 186                 size = event->header.size;
 187
 188                 /*
 189                  * Event straddles the mmap boundary -- header should always
 190                  * be inside due to u64 alignment of output.
 191                  */
 192                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 193                         unsigned int offset = old;
 194                         unsigned int len = min(sizeof(*event), size), cpy;
 195                         void *dst = &evlist->event_copy;
 196
 197                         do {
 198                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 199                                 memcpy(dst, &data[offset & md->mask], cpy);
 200                                 offset += cpy;
 201                                 dst += cpy;
 202                                 len -= cpy;
 203                         } while (len);
 204
 205                         event = &evlist->event_copy;
 206                 }
 207
 208                 old += size;
 209         }
 210
 211         md->prev = old;
 212
 213         if (!evlist->overwrite)
 214                 perf_mmap__write_tail(md, old);
 215
 216         return event;
 217 }
 218
 219 void perf_evlist__munmap(struct perf_evlist *evlist)
 220 {
 221         int cpu;
 222
 223         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 224                 if (evlist->mmap[cpu].base != NULL) {
 225                         munmap(evlist->mmap[cpu].base, evlist->mmap_len);
 226                         evlist->mmap[cpu].base = NULL;
 227                 }
 228         }
 229 }
 230
 231 int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 232 {
 233         evlist->mmap = zalloc(evlist->cpus->nr * sizeof(struct perf_mmap));
 234         return evlist->mmap != NULL ? 0 : -ENOMEM;
 235 }
 236
 237 static int __perf_evlist__mmap(struct perf_evlist *evlist, int cpu, int prot,
 238                                int mask, int fd)
 239 {
 240         evlist->mmap[cpu].prev = 0;
 241         evlist->mmap[cpu].mask = mask;
 242         evlist->mmap[cpu].base = mmap(NULL, evlist->mmap_len, prot,
 243                                       MAP_SHARED, fd, 0);
 244         if (evlist->mmap[cpu].base == MAP_FAILED)
 245                 return -1;
 246
 247         perf_evlist__add_pollfd(evlist, fd);
 248         return 0;
 249 }
 250
 251 /** perf_evlist__mmap - Create per cpu maps to receive events
 252  *
 253  * @evlist - list of events
 254  * @pages - map length in pages
 255  * @overwrite - overwrite older events?
 256  *
 257  * If overwrite is false the user needs to signal event consuption using:
 258  *
 259  *      struct perf_mmap *m = &evlist->mmap[cpu];
 260  *      unsigned int head = perf_mmap__read_head(m);
 261  *
 262  *      perf_mmap__write_tail(m, head)
 263  *
 264  * Using perf_evlist__read_on_cpu does this automatically.
 265  */
 266 int perf_evlist__mmap(struct perf_evlist *evlist, int pages, bool overwrite)
 267 {
 268         unsigned int page_size = sysconf(_SC_PAGE_SIZE);
 269         int mask = pages * page_size - 1, cpu;
 270         struct perf_evsel *first_evsel, *evsel;
 271         const struct cpu_map *cpus = evlist->cpus;
 272         const struct thread_map *threads = evlist->threads;
 273         int thread, prot = PROT_READ | (overwrite ? 0 : PROT_WRITE);
 274
 275         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 276                 return -ENOMEM;
 277
 278         if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 279                 return -ENOMEM;
 280
 281         evlist->overwrite = overwrite;
 282         evlist->mmap_len = (pages + 1) * page_size;
 283         first_evsel = list_entry(evlist->entries.next, struct perf_evsel, node);
 284
 285         list_for_each_entry(evsel, &evlist->entries, node) {
 286                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 287                     evsel->id == NULL &&
 288                     perf_evsel__alloc_id(evsel, cpus->nr, threads->nr) < 0)
 289                         return -ENOMEM;
 290
 291                 for (cpu = 0; cpu < cpus->nr; cpu++) {
 292                         for (thread = 0; thread < threads->nr; thread++) {
 293                                 int fd = FD(evsel, cpu, thread);
 294
 295                                 if (evsel->idx || thread) {
 296                                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT,
 297                                                   FD(first_evsel, cpu, 0)) != 0)
 298                                                 goto out_unmap;
 299                                 } else if (__perf_evlist__mmap(evlist, cpu, prot, mask, fd) < 0)
 300                                         goto out_unmap;
 301
 302                                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 303                                     perf_evlist__id_hash(evlist, evsel, cpu, thread, fd) < 0)
 304                                         goto out_unmap;
 305                         }
 306                 }
 307         }
 308
 309         return 0;
 310
 311 out_unmap:
 312         for (cpu = 0; cpu < cpus->nr; cpu++) {
 313                 if (evlist->mmap[cpu].base != NULL) {
 314                         munmap(evlist->mmap[cpu].base, evlist->mmap_len);
 315                         evlist->mmap[cpu].base = NULL;
 316                 }
 317         }
 318         return -1;
 319 }
 320
 321 int perf_evlist__create_maps(struct perf_evlist *evlist, pid_t target_pid,
 322                              pid_t target_tid, const char *cpu_list)
 323 {
 324         evlist->threads = thread_map__new(target_pid, target_tid);
 325
 326         if (evlist->threads == NULL)
 327                 return -1;
 328
 329         if (target_tid != -1)
 330                 evlist->cpus = cpu_map__dummy_new();
 331         else
 332                 evlist->cpus = cpu_map__new(cpu_list);
 333
 334         if (evlist->cpus == NULL)
 335                 goto out_delete_threads;
 336
 337         return 0;
 338
 339 out_delete_threads:
 340         thread_map__delete(evlist->threads);
 341         return -1;
 342 }
 343
 344 void perf_evlist__delete_maps(struct perf_evlist *evlist)
 345 {
 346         cpu_map__delete(evlist->cpus);
 347         thread_map__delete(evlist->threads);
 348         evlist->cpus    = NULL;
 349         evlist->threads = NULL;
 350 }
 351
 352 int perf_evlist__set_filters(struct perf_evlist *evlist)
 353 {
 354         const struct thread_map *threads = evlist->threads;
 355         const struct cpu_map *cpus = evlist->cpus;
 356         struct perf_evsel *evsel;
 357         char *filter;
 358         int thread;
 359         int cpu;
 360         int err;
 361         int fd;
 362
 363         list_for_each_entry(evsel, &evlist->entries, node) {
 364                 filter = evsel->filter;
 365                 if (!filter)
 366                         continue;
 367                 for (cpu = 0; cpu < cpus->nr; cpu++) {
 368                         for (thread = 0; thread < threads->nr; thread++) {
 369                                 fd = FD(evsel, cpu, thread);
 370                                 err = ioctl(fd, PERF_EVENT_IOC_SET_FILTER, filter);
 371                                 if (err)
 372                                         return err;
 373                         }
 374                 }
 375         }
 376
 377         return 0;
 378 }