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