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 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 void perf_evlist__id_hash(struct perf_evlist *evlist, 110 struct perf_evsel *evsel, 111 int cpu, int thread, u64 id) 112 { 113 int hash; 114 struct perf_sample_id *sid = SID(evsel, cpu, thread); 115 116 sid->id = id; 117 sid->evsel = evsel; 118 hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS); 119 hlist_add_head(&sid->node, &evlist->heads[hash]); 120 } 121 122 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel, 123 int cpu, int thread, u64 id) 124 { 125 perf_evlist__id_hash(evlist, evsel, cpu, thread, id); 126 evsel->id[evsel->ids++] = id; 127 } 128 129 static int perf_evlist__id_add_fd(struct perf_evlist *evlist, 130 struct perf_evsel *evsel, 131 int cpu, int thread, int fd) 132 { 133 u64 read_data[4] = { 0, }; 134 int id_idx = 1; /* The first entry is the counter value */ 135 136 if (!(evsel->attr.read_format & PERF_FORMAT_ID) || 137 read(fd, &read_data, sizeof(read_data)) == -1) 138 return -1; 139 140 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 141 ++id_idx; 142 if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 143 ++id_idx; 144 145 perf_evlist__id_add(evlist, evsel, cpu, thread, read_data[id_idx]); 146 return 0; 147 } 148 149 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id) 150 { 151 struct hlist_head *head; 152 struct hlist_node *pos; 153 struct perf_sample_id *sid; 154 int hash; 155 156 if (evlist->nr_entries == 1) 157 return list_entry(evlist->entries.next, struct perf_evsel, node); 158 159 hash = hash_64(id, PERF_EVLIST__HLIST_BITS); 160 head = &evlist->heads[hash]; 161 162 hlist_for_each_entry(sid, pos, head, node) 163 if (sid->id == id) 164 return sid->evsel; 165 return NULL; 166 } 167 168 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx) 169 { 170 /* XXX Move this to perf.c, making it generally available */ 171 unsigned int page_size = sysconf(_SC_PAGE_SIZE); 172 struct perf_mmap *md = &evlist->mmap[idx]; 173 unsigned int head = perf_mmap__read_head(md); 174 unsigned int old = md->prev; 175 unsigned char *data = md->base + page_size; 176 union perf_event *event = NULL; 177 178 if (evlist->overwrite) { 179 /* 180 * If we're further behind than half the buffer, there's a chance 181 * the writer will bite our tail and mess up the samples under us. 182 * 183 * If we somehow ended up ahead of the head, we got messed up. 184 * 185 * In either case, truncate and restart at head. 186 */ 187 int diff = head - old; 188 if (diff > md->mask / 2 || diff < 0) { 189 fprintf(stderr, "WARNING: failed to keep up with mmap data.\n"); 190 191 /* 192 * head points to a known good entry, start there. 193 */ 194 old = head; 195 } 196 } 197 198 if (old != head) { 199 size_t size; 200 201 event = (union perf_event *)&data[old & md->mask]; 202 size = event->header.size; 203 204 /* 205 * Event straddles the mmap boundary -- header should always 206 * be inside due to u64 alignment of output. 207 */ 208 if ((old & md->mask) + size != ((old + size) & md->mask)) { 209 unsigned int offset = old; 210 unsigned int len = min(sizeof(*event), size), cpy; 211 void *dst = &evlist->event_copy; 212 213 do { 214 cpy = min(md->mask + 1 - (offset & md->mask), len); 215 memcpy(dst, &data[offset & md->mask], cpy); 216 offset += cpy; 217 dst += cpy; 218 len -= cpy; 219 } while (len); 220 221 event = &evlist->event_copy; 222 } 223 224 old += size; 225 } 226 227 md->prev = old; 228 229 if (!evlist->overwrite) 230 perf_mmap__write_tail(md, old); 231 232 return event; 233 } 234 235 void perf_evlist__munmap(struct perf_evlist *evlist) 236 { 237 int i; 238 239 for (i = 0; i < evlist->nr_mmaps; i++) { 240 if (evlist->mmap[i].base != NULL) { 241 munmap(evlist->mmap[i].base, evlist->mmap_len); 242 evlist->mmap[i].base = NULL; 243 } 244 } 245 246 free(evlist->mmap); 247 evlist->mmap = NULL; 248 } 249 250 int perf_evlist__alloc_mmap(struct perf_evlist *evlist) 251 { 252 evlist->nr_mmaps = evlist->cpus->nr; 253 if (evlist->cpus->map[0] == -1) 254 evlist->nr_mmaps = evlist->threads->nr; 255 evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap)); 256 return evlist->mmap != NULL ? 0 : -ENOMEM; 257 } 258 259 static int __perf_evlist__mmap(struct perf_evlist *evlist, 260 int idx, int prot, int mask, int fd) 261 { 262 evlist->mmap[idx].prev = 0; 263 evlist->mmap[idx].mask = mask; 264 evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot, 265 MAP_SHARED, fd, 0); 266 if (evlist->mmap[idx].base == MAP_FAILED) 267 return -1; 268 269 perf_evlist__add_pollfd(evlist, fd); 270 return 0; 271 } 272 273 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot, int mask) 274 { 275 struct perf_evsel *evsel; 276 int cpu, thread; 277 278 for (cpu = 0; cpu < evlist->cpus->nr; cpu++) { 279 int output = -1; 280 281 for (thread = 0; thread < evlist->threads->nr; thread++) { 282 list_for_each_entry(evsel, &evlist->entries, node) { 283 int fd = FD(evsel, cpu, thread); 284 285 if (output == -1) { 286 output = fd; 287 if (__perf_evlist__mmap(evlist, cpu, 288 prot, mask, output) < 0) 289 goto out_unmap; 290 } else { 291 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0) 292 goto out_unmap; 293 } 294 295 if ((evsel->attr.read_format & PERF_FORMAT_ID) && 296 perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0) 297 goto out_unmap; 298 } 299 } 300 } 301 302 return 0; 303 304 out_unmap: 305 for (cpu = 0; cpu < evlist->cpus->nr; cpu++) { 306 if (evlist->mmap[cpu].base != NULL) { 307 munmap(evlist->mmap[cpu].base, evlist->mmap_len); 308 evlist->mmap[cpu].base = NULL; 309 } 310 } 311 return -1; 312 } 313 314 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot, int mask) 315 { 316 struct perf_evsel *evsel; 317 int thread; 318 319 for (thread = 0; thread < evlist->threads->nr; thread++) { 320 int output = -1; 321 322 list_for_each_entry(evsel, &evlist->entries, node) { 323 int fd = FD(evsel, 0, thread); 324 325 if (output == -1) { 326 output = fd; 327 if (__perf_evlist__mmap(evlist, thread, 328 prot, mask, output) < 0) 329 goto out_unmap; 330 } else { 331 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0) 332 goto out_unmap; 333 } 334 335 if ((evsel->attr.read_format & PERF_FORMAT_ID) && 336 perf_evlist__id_add_fd(evlist, evsel, 0, thread, fd) < 0) 337 goto out_unmap; 338 } 339 } 340 341 return 0; 342 343 out_unmap: 344 for (thread = 0; thread < evlist->threads->nr; thread++) { 345 if (evlist->mmap[thread].base != NULL) { 346 munmap(evlist->mmap[thread].base, evlist->mmap_len); 347 evlist->mmap[thread].base = NULL; 348 } 349 } 350 return -1; 351 } 352 353 /** perf_evlist__mmap - Create per cpu maps to receive events 354 * 355 * @evlist - list of events 356 * @pages - map length in pages 357 * @overwrite - overwrite older events? 358 * 359 * If overwrite is false the user needs to signal event consuption using: 360 * 361 * struct perf_mmap *m = &evlist->mmap[cpu]; 362 * unsigned int head = perf_mmap__read_head(m); 363 * 364 * perf_mmap__write_tail(m, head) 365 * 366 * Using perf_evlist__read_on_cpu does this automatically. 367 */ 368 int perf_evlist__mmap(struct perf_evlist *evlist, int pages, bool overwrite) 369 { 370 unsigned int page_size = sysconf(_SC_PAGE_SIZE); 371 int mask = pages * page_size - 1; 372 struct perf_evsel *evsel; 373 const struct cpu_map *cpus = evlist->cpus; 374 const struct thread_map *threads = evlist->threads; 375 int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE); 376 377 if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0) 378 return -ENOMEM; 379 380 if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0) 381 return -ENOMEM; 382 383 evlist->overwrite = overwrite; 384 evlist->mmap_len = (pages + 1) * page_size; 385 386 list_for_each_entry(evsel, &evlist->entries, node) { 387 if ((evsel->attr.read_format & PERF_FORMAT_ID) && 388 evsel->sample_id == NULL && 389 perf_evsel__alloc_id(evsel, cpus->nr, threads->nr) < 0) 390 return -ENOMEM; 391 } 392 393 if (evlist->cpus->map[0] == -1) 394 return perf_evlist__mmap_per_thread(evlist, prot, mask); 395 396 return perf_evlist__mmap_per_cpu(evlist, prot, mask); 397 } 398 399 int perf_evlist__create_maps(struct perf_evlist *evlist, pid_t target_pid, 400 pid_t target_tid, const char *cpu_list) 401 { 402 evlist->threads = thread_map__new(target_pid, target_tid); 403 404 if (evlist->threads == NULL) 405 return -1; 406 407 if (cpu_list == NULL && target_tid != -1) 408 evlist->cpus = cpu_map__dummy_new(); 409 else 410 evlist->cpus = cpu_map__new(cpu_list); 411 412 if (evlist->cpus == NULL) 413 goto out_delete_threads; 414 415 return 0; 416 417 out_delete_threads: 418 thread_map__delete(evlist->threads); 419 return -1; 420 } 421 422 void perf_evlist__delete_maps(struct perf_evlist *evlist) 423 { 424 cpu_map__delete(evlist->cpus); 425 thread_map__delete(evlist->threads); 426 evlist->cpus = NULL; 427 evlist->threads = NULL; 428 } 429 430 int perf_evlist__set_filters(struct perf_evlist *evlist) 431 { 432 const struct thread_map *threads = evlist->threads; 433 const struct cpu_map *cpus = evlist->cpus; 434 struct perf_evsel *evsel; 435 char *filter; 436 int thread; 437 int cpu; 438 int err; 439 int fd; 440 441 list_for_each_entry(evsel, &evlist->entries, node) { 442 filter = evsel->filter; 443 if (!filter) 444 continue; 445 for (cpu = 0; cpu < cpus->nr; cpu++) { 446 for (thread = 0; thread < threads->nr; thread++) { 447 fd = FD(evsel, cpu, thread); 448 err = ioctl(fd, PERF_EVENT_IOC_SET_FILTER, filter); 449 if (err) 450 return err; 451 } 452 } 453 } 454 455 return 0; 456 } 457 458 bool perf_evlist__valid_sample_type(const struct perf_evlist *evlist) 459 { 460 struct perf_evsel *pos, *first; 461 462 pos = first = list_entry(evlist->entries.next, struct perf_evsel, node); 463 464 list_for_each_entry_continue(pos, &evlist->entries, node) { 465 if (first->attr.sample_type != pos->attr.sample_type) 466 return false; 467 } 468 469 return true; 470 } 471 472 u64 perf_evlist__sample_type(const struct perf_evlist *evlist) 473 { 474 struct perf_evsel *first; 475 476 first = list_entry(evlist->entries.next, struct perf_evsel, node); 477 return first->attr.sample_type; 478 } 479 480 bool perf_evlist__valid_sample_id_all(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_id_all != pos->attr.sample_id_all) 488 return false; 489 } 490 491 return true; 492 } 493 494 bool perf_evlist__sample_id_all(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_id_all; 500 } 501