| 1 | /* |
| 2 | * libringbuffer/shm.c |
| 3 | * |
| 4 | * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 5 | * |
| 6 | * This library is free software; you can redistribute it and/or |
| 7 | * modify it under the terms of the GNU Lesser General Public |
| 8 | * License as published by the Free Software Foundation; only |
| 9 | * version 2.1 of the License. |
| 10 | * |
| 11 | * This library is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 14 | * Lesser General Public License for more details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU Lesser General Public |
| 17 | * License along with this library; if not, write to the Free Software |
| 18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 19 | */ |
| 20 | |
| 21 | #define _LGPL_SOURCE |
| 22 | #include <config.h> |
| 23 | #include "shm.h" |
| 24 | #include <unistd.h> |
| 25 | #include <fcntl.h> |
| 26 | #include <sys/mman.h> |
| 27 | #include <sys/types.h> |
| 28 | #include <sys/stat.h> /* For mode constants */ |
| 29 | #include <fcntl.h> /* For O_* constants */ |
| 30 | #include <assert.h> |
| 31 | #include <stdio.h> |
| 32 | #include <signal.h> |
| 33 | #include <dirent.h> |
| 34 | #include <lttng/align.h> |
| 35 | #include <limits.h> |
| 36 | #include <stdbool.h> |
| 37 | #ifdef HAVE_LIBNUMA |
| 38 | #include <numa.h> |
| 39 | #include <numaif.h> |
| 40 | #endif |
| 41 | #include <helper.h> |
| 42 | #include <ust-fd.h> |
| 43 | #include "mmap.h" |
| 44 | |
| 45 | /* |
| 46 | * Ensure we have the required amount of space available by writing 0 |
| 47 | * into the entire buffer. Not doing so can trigger SIGBUS when going |
| 48 | * beyond the available shm space. |
| 49 | */ |
| 50 | static |
| 51 | int zero_file(int fd, size_t len) |
| 52 | { |
| 53 | ssize_t retlen; |
| 54 | size_t written = 0; |
| 55 | char *zeropage; |
| 56 | long pagelen; |
| 57 | int ret; |
| 58 | |
| 59 | pagelen = sysconf(_SC_PAGESIZE); |
| 60 | if (pagelen < 0) |
| 61 | return (int) pagelen; |
| 62 | zeropage = calloc(pagelen, 1); |
| 63 | if (!zeropage) |
| 64 | return -ENOMEM; |
| 65 | |
| 66 | while (len > written) { |
| 67 | do { |
| 68 | retlen = write(fd, zeropage, |
| 69 | min_t(size_t, pagelen, len - written)); |
| 70 | } while (retlen == -1UL && errno == EINTR); |
| 71 | if (retlen < 0) { |
| 72 | ret = (int) retlen; |
| 73 | goto error; |
| 74 | } |
| 75 | written += retlen; |
| 76 | } |
| 77 | ret = 0; |
| 78 | error: |
| 79 | free(zeropage); |
| 80 | return ret; |
| 81 | } |
| 82 | |
| 83 | struct shm_object_table *shm_object_table_create(size_t max_nb_obj) |
| 84 | { |
| 85 | struct shm_object_table *table; |
| 86 | |
| 87 | table = zmalloc(sizeof(struct shm_object_table) + |
| 88 | max_nb_obj * sizeof(table->objects[0])); |
| 89 | if (!table) |
| 90 | return NULL; |
| 91 | table->size = max_nb_obj; |
| 92 | return table; |
| 93 | } |
| 94 | |
| 95 | static |
| 96 | struct shm_object *_shm_object_table_alloc_shm(struct shm_object_table *table, |
| 97 | size_t memory_map_size, |
| 98 | int stream_fd) |
| 99 | { |
| 100 | int shmfd, waitfd[2], ret, i; |
| 101 | struct shm_object *obj; |
| 102 | char *memory_map; |
| 103 | |
| 104 | if (stream_fd < 0) |
| 105 | return NULL; |
| 106 | if (table->allocated_len >= table->size) |
| 107 | return NULL; |
| 108 | obj = &table->objects[table->allocated_len]; |
| 109 | |
| 110 | /* wait_fd: create pipe */ |
| 111 | ret = pipe(waitfd); |
| 112 | if (ret < 0) { |
| 113 | PERROR("pipe"); |
| 114 | goto error_pipe; |
| 115 | } |
| 116 | for (i = 0; i < 2; i++) { |
| 117 | ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC); |
| 118 | if (ret < 0) { |
| 119 | PERROR("fcntl"); |
| 120 | goto error_fcntl; |
| 121 | } |
| 122 | } |
| 123 | /* The write end of the pipe needs to be non-blocking */ |
| 124 | ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK); |
| 125 | if (ret < 0) { |
| 126 | PERROR("fcntl"); |
| 127 | goto error_fcntl; |
| 128 | } |
| 129 | memcpy(obj->wait_fd, waitfd, sizeof(waitfd)); |
| 130 | |
| 131 | /* create shm */ |
| 132 | |
| 133 | shmfd = stream_fd; |
| 134 | ret = zero_file(shmfd, memory_map_size); |
| 135 | if (ret) { |
| 136 | PERROR("zero_file"); |
| 137 | goto error_zero_file; |
| 138 | } |
| 139 | ret = ftruncate(shmfd, memory_map_size); |
| 140 | if (ret) { |
| 141 | PERROR("ftruncate"); |
| 142 | goto error_ftruncate; |
| 143 | } |
| 144 | /* |
| 145 | * Also ensure the file metadata is synced with the storage by using |
| 146 | * fsync(2). |
| 147 | */ |
| 148 | ret = fsync(shmfd); |
| 149 | if (ret) { |
| 150 | PERROR("fsync"); |
| 151 | goto error_fsync; |
| 152 | } |
| 153 | obj->shm_fd_ownership = 0; |
| 154 | obj->shm_fd = shmfd; |
| 155 | |
| 156 | /* memory_map: mmap */ |
| 157 | memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE, |
| 158 | MAP_SHARED | LTTNG_MAP_POPULATE, shmfd, 0); |
| 159 | if (memory_map == MAP_FAILED) { |
| 160 | PERROR("mmap"); |
| 161 | goto error_mmap; |
| 162 | } |
| 163 | obj->type = SHM_OBJECT_SHM; |
| 164 | obj->memory_map = memory_map; |
| 165 | obj->memory_map_size = memory_map_size; |
| 166 | obj->allocated_len = 0; |
| 167 | obj->index = table->allocated_len++; |
| 168 | |
| 169 | return obj; |
| 170 | |
| 171 | error_mmap: |
| 172 | error_fsync: |
| 173 | error_ftruncate: |
| 174 | error_zero_file: |
| 175 | error_fcntl: |
| 176 | for (i = 0; i < 2; i++) { |
| 177 | ret = close(waitfd[i]); |
| 178 | if (ret) { |
| 179 | PERROR("close"); |
| 180 | assert(0); |
| 181 | } |
| 182 | } |
| 183 | error_pipe: |
| 184 | return NULL; |
| 185 | } |
| 186 | |
| 187 | static |
| 188 | struct shm_object *_shm_object_table_alloc_mem(struct shm_object_table *table, |
| 189 | size_t memory_map_size) |
| 190 | { |
| 191 | struct shm_object *obj; |
| 192 | void *memory_map; |
| 193 | int waitfd[2], i, ret; |
| 194 | |
| 195 | if (table->allocated_len >= table->size) |
| 196 | return NULL; |
| 197 | obj = &table->objects[table->allocated_len]; |
| 198 | |
| 199 | memory_map = zmalloc(memory_map_size); |
| 200 | if (!memory_map) |
| 201 | goto alloc_error; |
| 202 | |
| 203 | /* wait_fd: create pipe */ |
| 204 | ret = pipe(waitfd); |
| 205 | if (ret < 0) { |
| 206 | PERROR("pipe"); |
| 207 | goto error_pipe; |
| 208 | } |
| 209 | for (i = 0; i < 2; i++) { |
| 210 | ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC); |
| 211 | if (ret < 0) { |
| 212 | PERROR("fcntl"); |
| 213 | goto error_fcntl; |
| 214 | } |
| 215 | } |
| 216 | /* The write end of the pipe needs to be non-blocking */ |
| 217 | ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK); |
| 218 | if (ret < 0) { |
| 219 | PERROR("fcntl"); |
| 220 | goto error_fcntl; |
| 221 | } |
| 222 | memcpy(obj->wait_fd, waitfd, sizeof(waitfd)); |
| 223 | |
| 224 | /* no shm_fd */ |
| 225 | obj->shm_fd = -1; |
| 226 | obj->shm_fd_ownership = 0; |
| 227 | |
| 228 | obj->type = SHM_OBJECT_MEM; |
| 229 | obj->memory_map = memory_map; |
| 230 | obj->memory_map_size = memory_map_size; |
| 231 | obj->allocated_len = 0; |
| 232 | obj->index = table->allocated_len++; |
| 233 | |
| 234 | return obj; |
| 235 | |
| 236 | error_fcntl: |
| 237 | for (i = 0; i < 2; i++) { |
| 238 | ret = close(waitfd[i]); |
| 239 | if (ret) { |
| 240 | PERROR("close"); |
| 241 | assert(0); |
| 242 | } |
| 243 | } |
| 244 | error_pipe: |
| 245 | free(memory_map); |
| 246 | alloc_error: |
| 247 | return NULL; |
| 248 | } |
| 249 | |
| 250 | /* |
| 251 | * libnuma prints errors on the console even for numa_available(). |
| 252 | * Work-around this limitation by using get_mempolicy() directly to |
| 253 | * check whether the kernel supports mempolicy. |
| 254 | */ |
| 255 | #ifdef HAVE_LIBNUMA |
| 256 | static bool lttng_is_numa_available(void) |
| 257 | { |
| 258 | int ret; |
| 259 | |
| 260 | ret = get_mempolicy(NULL, NULL, 0, NULL, 0); |
| 261 | if (ret && errno == ENOSYS) { |
| 262 | return false; |
| 263 | } |
| 264 | return numa_available() > 0; |
| 265 | } |
| 266 | #endif |
| 267 | |
| 268 | struct shm_object *shm_object_table_alloc(struct shm_object_table *table, |
| 269 | size_t memory_map_size, |
| 270 | enum shm_object_type type, |
| 271 | int stream_fd, |
| 272 | int cpu) |
| 273 | { |
| 274 | struct shm_object *shm_object; |
| 275 | #ifdef HAVE_LIBNUMA |
| 276 | int oldnode = 0, node; |
| 277 | bool numa_avail; |
| 278 | |
| 279 | numa_avail = lttng_is_numa_available(); |
| 280 | if (numa_avail) { |
| 281 | oldnode = numa_preferred(); |
| 282 | if (cpu >= 0) { |
| 283 | node = numa_node_of_cpu(cpu); |
| 284 | if (node >= 0) |
| 285 | numa_set_preferred(node); |
| 286 | } |
| 287 | if (cpu < 0 || node < 0) |
| 288 | numa_set_localalloc(); |
| 289 | } |
| 290 | #endif /* HAVE_LIBNUMA */ |
| 291 | switch (type) { |
| 292 | case SHM_OBJECT_SHM: |
| 293 | shm_object = _shm_object_table_alloc_shm(table, memory_map_size, |
| 294 | stream_fd); |
| 295 | break; |
| 296 | case SHM_OBJECT_MEM: |
| 297 | shm_object = _shm_object_table_alloc_mem(table, memory_map_size); |
| 298 | break; |
| 299 | default: |
| 300 | assert(0); |
| 301 | } |
| 302 | #ifdef HAVE_LIBNUMA |
| 303 | if (numa_avail) |
| 304 | numa_set_preferred(oldnode); |
| 305 | #endif /* HAVE_LIBNUMA */ |
| 306 | return shm_object; |
| 307 | } |
| 308 | |
| 309 | struct shm_object *shm_object_table_append_shm(struct shm_object_table *table, |
| 310 | int shm_fd, int wakeup_fd, uint32_t stream_nr, |
| 311 | size_t memory_map_size) |
| 312 | { |
| 313 | struct shm_object *obj; |
| 314 | char *memory_map; |
| 315 | int ret; |
| 316 | |
| 317 | if (table->allocated_len >= table->size) |
| 318 | return NULL; |
| 319 | /* streams _must_ be received in sequential order, else fail. */ |
| 320 | if (stream_nr + 1 != table->allocated_len) |
| 321 | return NULL; |
| 322 | |
| 323 | obj = &table->objects[table->allocated_len]; |
| 324 | |
| 325 | /* wait_fd: set write end of the pipe. */ |
| 326 | obj->wait_fd[0] = -1; /* read end is unset */ |
| 327 | obj->wait_fd[1] = wakeup_fd; |
| 328 | obj->shm_fd = shm_fd; |
| 329 | obj->shm_fd_ownership = 1; |
| 330 | |
| 331 | ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC); |
| 332 | if (ret < 0) { |
| 333 | PERROR("fcntl"); |
| 334 | goto error_fcntl; |
| 335 | } |
| 336 | /* The write end of the pipe needs to be non-blocking */ |
| 337 | ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK); |
| 338 | if (ret < 0) { |
| 339 | PERROR("fcntl"); |
| 340 | goto error_fcntl; |
| 341 | } |
| 342 | |
| 343 | /* memory_map: mmap */ |
| 344 | memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE, |
| 345 | MAP_SHARED | LTTNG_MAP_POPULATE, shm_fd, 0); |
| 346 | if (memory_map == MAP_FAILED) { |
| 347 | PERROR("mmap"); |
| 348 | goto error_mmap; |
| 349 | } |
| 350 | obj->type = SHM_OBJECT_SHM; |
| 351 | obj->memory_map = memory_map; |
| 352 | obj->memory_map_size = memory_map_size; |
| 353 | obj->allocated_len = memory_map_size; |
| 354 | obj->index = table->allocated_len++; |
| 355 | |
| 356 | return obj; |
| 357 | |
| 358 | error_fcntl: |
| 359 | error_mmap: |
| 360 | return NULL; |
| 361 | } |
| 362 | |
| 363 | /* |
| 364 | * Passing ownership of mem to object. |
| 365 | */ |
| 366 | struct shm_object *shm_object_table_append_mem(struct shm_object_table *table, |
| 367 | void *mem, size_t memory_map_size, int wakeup_fd) |
| 368 | { |
| 369 | struct shm_object *obj; |
| 370 | int ret; |
| 371 | |
| 372 | if (table->allocated_len >= table->size) |
| 373 | return NULL; |
| 374 | obj = &table->objects[table->allocated_len]; |
| 375 | |
| 376 | obj->wait_fd[0] = -1; /* read end is unset */ |
| 377 | obj->wait_fd[1] = wakeup_fd; |
| 378 | obj->shm_fd = -1; |
| 379 | obj->shm_fd_ownership = 0; |
| 380 | |
| 381 | ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC); |
| 382 | if (ret < 0) { |
| 383 | PERROR("fcntl"); |
| 384 | goto error_fcntl; |
| 385 | } |
| 386 | /* The write end of the pipe needs to be non-blocking */ |
| 387 | ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK); |
| 388 | if (ret < 0) { |
| 389 | PERROR("fcntl"); |
| 390 | goto error_fcntl; |
| 391 | } |
| 392 | |
| 393 | obj->type = SHM_OBJECT_MEM; |
| 394 | obj->memory_map = mem; |
| 395 | obj->memory_map_size = memory_map_size; |
| 396 | obj->allocated_len = memory_map_size; |
| 397 | obj->index = table->allocated_len++; |
| 398 | |
| 399 | return obj; |
| 400 | |
| 401 | error_fcntl: |
| 402 | return NULL; |
| 403 | } |
| 404 | |
| 405 | static |
| 406 | void shmp_object_destroy(struct shm_object *obj, int consumer) |
| 407 | { |
| 408 | switch (obj->type) { |
| 409 | case SHM_OBJECT_SHM: |
| 410 | { |
| 411 | int ret, i; |
| 412 | |
| 413 | ret = munmap(obj->memory_map, obj->memory_map_size); |
| 414 | if (ret) { |
| 415 | PERROR("umnmap"); |
| 416 | assert(0); |
| 417 | } |
| 418 | |
| 419 | if (obj->shm_fd_ownership) { |
| 420 | /* Delete FDs only if called from app (not consumer). */ |
| 421 | if (!consumer) { |
| 422 | lttng_ust_lock_fd_tracker(); |
| 423 | ret = close(obj->shm_fd); |
| 424 | if (!ret) { |
| 425 | lttng_ust_delete_fd_from_tracker(obj->shm_fd); |
| 426 | } else { |
| 427 | PERROR("close"); |
| 428 | assert(0); |
| 429 | } |
| 430 | lttng_ust_unlock_fd_tracker(); |
| 431 | } else { |
| 432 | ret = close(obj->shm_fd); |
| 433 | if (ret) { |
| 434 | PERROR("close"); |
| 435 | assert(0); |
| 436 | } |
| 437 | } |
| 438 | } |
| 439 | for (i = 0; i < 2; i++) { |
| 440 | if (obj->wait_fd[i] < 0) |
| 441 | continue; |
| 442 | if (!consumer) { |
| 443 | lttng_ust_lock_fd_tracker(); |
| 444 | ret = close(obj->wait_fd[i]); |
| 445 | if (!ret) { |
| 446 | lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]); |
| 447 | } else { |
| 448 | PERROR("close"); |
| 449 | assert(0); |
| 450 | } |
| 451 | lttng_ust_unlock_fd_tracker(); |
| 452 | } else { |
| 453 | ret = close(obj->wait_fd[i]); |
| 454 | if (ret) { |
| 455 | PERROR("close"); |
| 456 | assert(0); |
| 457 | } |
| 458 | } |
| 459 | } |
| 460 | break; |
| 461 | } |
| 462 | case SHM_OBJECT_MEM: |
| 463 | { |
| 464 | int ret, i; |
| 465 | |
| 466 | for (i = 0; i < 2; i++) { |
| 467 | if (obj->wait_fd[i] < 0) |
| 468 | continue; |
| 469 | if (!consumer) { |
| 470 | lttng_ust_lock_fd_tracker(); |
| 471 | ret = close(obj->wait_fd[i]); |
| 472 | if (!ret) { |
| 473 | lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]); |
| 474 | } else { |
| 475 | PERROR("close"); |
| 476 | assert(0); |
| 477 | } |
| 478 | lttng_ust_unlock_fd_tracker(); |
| 479 | } else { |
| 480 | ret = close(obj->wait_fd[i]); |
| 481 | if (ret) { |
| 482 | PERROR("close"); |
| 483 | assert(0); |
| 484 | } |
| 485 | } |
| 486 | } |
| 487 | free(obj->memory_map); |
| 488 | break; |
| 489 | } |
| 490 | default: |
| 491 | assert(0); |
| 492 | } |
| 493 | } |
| 494 | |
| 495 | void shm_object_table_destroy(struct shm_object_table *table, int consumer) |
| 496 | { |
| 497 | int i; |
| 498 | |
| 499 | for (i = 0; i < table->allocated_len; i++) |
| 500 | shmp_object_destroy(&table->objects[i], consumer); |
| 501 | free(table); |
| 502 | } |
| 503 | |
| 504 | /* |
| 505 | * zalloc_shm - allocate memory within a shm object. |
| 506 | * |
| 507 | * Shared memory is already zeroed by shmget. |
| 508 | * *NOT* multithread-safe (should be protected by mutex). |
| 509 | * Returns a -1, -1 tuple on error. |
| 510 | */ |
| 511 | struct shm_ref zalloc_shm(struct shm_object *obj, size_t len) |
| 512 | { |
| 513 | struct shm_ref ref; |
| 514 | struct shm_ref shm_ref_error = { -1, -1 }; |
| 515 | |
| 516 | if (obj->memory_map_size - obj->allocated_len < len) |
| 517 | return shm_ref_error; |
| 518 | ref.index = obj->index; |
| 519 | ref.offset = obj->allocated_len; |
| 520 | obj->allocated_len += len; |
| 521 | return ref; |
| 522 | } |
| 523 | |
| 524 | void align_shm(struct shm_object *obj, size_t align) |
| 525 | { |
| 526 | size_t offset_len = offset_align(obj->allocated_len, align); |
| 527 | obj->allocated_len += offset_len; |
| 528 | } |