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