| 1 | /* |
| 2 | * Copyright (C) 2018-2020 Jérémie Galarneau <jeremie.galarneau@efficios.com> |
| 3 | * |
| 4 | * SPDX-License-Identifier: GPL-2.0-only |
| 5 | * |
| 6 | */ |
| 7 | |
| 8 | #include <urcu.h> |
| 9 | #include <urcu/list.h> |
| 10 | #include <urcu/rculfhash.h> |
| 11 | |
| 12 | #include <fcntl.h> |
| 13 | #include <inttypes.h> |
| 14 | #include <pthread.h> |
| 15 | #include <stdbool.h> |
| 16 | #include <sys/stat.h> |
| 17 | #include <sys/types.h> |
| 18 | |
| 19 | #include <common/defaults.h> |
| 20 | #include <common/error.h> |
| 21 | #include <common/fs-handle-internal.h> |
| 22 | #include <common/hashtable/hashtable.h> |
| 23 | #include <common/hashtable/utils.h> |
| 24 | #include <common/macros.h> |
| 25 | #include <common/optional.h> |
| 26 | |
| 27 | #include "fd-tracker.h" |
| 28 | #include "inode.h" |
| 29 | |
| 30 | /* Tracker lock must be taken by the user. */ |
| 31 | #define TRACKED_COUNT(tracker) \ |
| 32 | (tracker->count.suspendable.active + \ |
| 33 | tracker->count.suspendable.suspended + \ |
| 34 | tracker->count.unsuspendable) |
| 35 | |
| 36 | /* Tracker lock must be taken by the user. */ |
| 37 | #define ACTIVE_COUNT(tracker) \ |
| 38 | (tracker->count.suspendable.active + tracker->count.unsuspendable) |
| 39 | |
| 40 | /* Tracker lock must be taken by the user. */ |
| 41 | #define SUSPENDED_COUNT(tracker) (tracker->count.suspendable.suspended) |
| 42 | |
| 43 | /* Tracker lock must be taken by the user. */ |
| 44 | #define SUSPENDABLE_COUNT(tracker) \ |
| 45 | (tracker->count.suspendable.active + \ |
| 46 | tracker->count.suspendable.suspended) |
| 47 | |
| 48 | /* Tracker lock must be taken by the user. */ |
| 49 | #define UNSUSPENDABLE_COUNT(tracker) (tracker->count.unsuspendable) |
| 50 | |
| 51 | struct fd_tracker { |
| 52 | pthread_mutex_t lock; |
| 53 | struct { |
| 54 | struct { |
| 55 | unsigned int active; |
| 56 | unsigned int suspended; |
| 57 | } suspendable; |
| 58 | unsigned int unsuspendable; |
| 59 | } count; |
| 60 | unsigned int capacity; |
| 61 | struct { |
| 62 | uint64_t uses; |
| 63 | uint64_t misses; |
| 64 | /* Failures to suspend or restore fs handles. */ |
| 65 | uint64_t errors; |
| 66 | } stats; |
| 67 | /* |
| 68 | * The head of the active_handles list is always the least recently |
| 69 | * used active handle. When an handle is used, it is removed from the |
| 70 | * list and added to the end. When a file has to be suspended, the |
| 71 | * first element in the list is "popped", suspended, and added to the |
| 72 | * list of suspended handles. |
| 73 | */ |
| 74 | struct cds_list_head active_handles; |
| 75 | struct cds_list_head suspended_handles; |
| 76 | struct cds_lfht *unsuspendable_fds; |
| 77 | struct lttng_inode_registry *inode_registry; |
| 78 | /* Unlinked files are moved in this directory under a unique name. */ |
| 79 | struct lttng_directory_handle *unlink_directory_handle; |
| 80 | struct lttng_unlinked_file_pool *unlinked_file_pool; |
| 81 | }; |
| 82 | |
| 83 | struct open_properties { |
| 84 | int flags; |
| 85 | LTTNG_OPTIONAL(mode_t) mode; |
| 86 | }; |
| 87 | |
| 88 | /* |
| 89 | * A fs_handle_tracked is not ref-counted. Therefore, it is assumed that a |
| 90 | * handle is never in-use while it is being reclaimed. It can be |
| 91 | * shared by multiple threads, but external synchronization is required |
| 92 | * to ensure it is not still being used when it is reclaimed (close method). |
| 93 | * In this respect, it is not different from a regular file descriptor. |
| 94 | * |
| 95 | * The fs_handle lock always nests _within_ the tracker's lock. |
| 96 | */ |
| 97 | struct fs_handle_tracked { |
| 98 | struct fs_handle parent; |
| 99 | pthread_mutex_t lock; |
| 100 | /* |
| 101 | * Weak reference to the tracker. All fs_handles are assumed to have |
| 102 | * been closed at the moment of the destruction of the fd_tracker. |
| 103 | */ |
| 104 | struct fd_tracker *tracker; |
| 105 | struct open_properties properties; |
| 106 | struct lttng_inode *inode; |
| 107 | int fd; |
| 108 | /* inode number of the file at the time of the handle's creation. */ |
| 109 | uint64_t ino; |
| 110 | bool in_use; |
| 111 | /* Offset to which the file should be restored. */ |
| 112 | off_t offset; |
| 113 | struct cds_list_head handles_list_node; |
| 114 | }; |
| 115 | |
| 116 | struct unsuspendable_fd { |
| 117 | /* |
| 118 | * Accesses are only performed through the tracker, which is protected |
| 119 | * by its own lock. |
| 120 | */ |
| 121 | int fd; |
| 122 | char *name; |
| 123 | struct cds_lfht_node tracker_node; |
| 124 | struct rcu_head rcu_head; |
| 125 | }; |
| 126 | |
| 127 | static struct { |
| 128 | pthread_mutex_t lock; |
| 129 | bool initialized; |
| 130 | unsigned long value; |
| 131 | } seed = { |
| 132 | .lock = PTHREAD_MUTEX_INITIALIZER, |
| 133 | }; |
| 134 | |
| 135 | static int match_fd(struct cds_lfht_node *node, const void *key); |
| 136 | static void unsuspendable_fd_destroy(struct unsuspendable_fd *entry); |
| 137 | static struct unsuspendable_fd *unsuspendable_fd_create( |
| 138 | const char *name, int fd); |
| 139 | static int open_from_properties(const struct lttng_directory_handle *dir_handle, |
| 140 | const char *path, struct open_properties *properties); |
| 141 | |
| 142 | static void fs_handle_tracked_log(struct fs_handle_tracked *handle); |
| 143 | static int fs_handle_tracked_suspend(struct fs_handle_tracked *handle); |
| 144 | static int fs_handle_tracked_restore(struct fs_handle_tracked *handle); |
| 145 | static int fs_handle_tracked_get_fd(struct fs_handle *_handle); |
| 146 | static void fs_handle_tracked_put_fd(struct fs_handle *_handle); |
| 147 | static int fs_handle_tracked_unlink(struct fs_handle *_handle); |
| 148 | static int fs_handle_tracked_close(struct fs_handle *_handle); |
| 149 | |
| 150 | static void fd_tracker_track( |
| 151 | struct fd_tracker *tracker, struct fs_handle_tracked *handle); |
| 152 | static void fd_tracker_untrack( |
| 153 | struct fd_tracker *tracker, struct fs_handle_tracked *handle); |
| 154 | static int fd_tracker_suspend_handles( |
| 155 | struct fd_tracker *tracker, unsigned int count); |
| 156 | static int fd_tracker_restore_handle( |
| 157 | struct fd_tracker *tracker, struct fs_handle_tracked *handle); |
| 158 | |
| 159 | /* Match function of the tracker's unsuspendable_fds hash table. */ |
| 160 | static int match_fd(struct cds_lfht_node *node, const void *key) |
| 161 | { |
| 162 | struct unsuspendable_fd *entry = caa_container_of( |
| 163 | node, struct unsuspendable_fd, tracker_node); |
| 164 | |
| 165 | return hash_match_key_ulong( |
| 166 | (void *) (unsigned long) entry->fd, (void *) key); |
| 167 | } |
| 168 | |
| 169 | static void delete_unsuspendable_fd(struct rcu_head *head) |
| 170 | { |
| 171 | struct unsuspendable_fd *fd = caa_container_of( |
| 172 | head, struct unsuspendable_fd, rcu_head); |
| 173 | |
| 174 | free(fd->name); |
| 175 | free(fd); |
| 176 | } |
| 177 | |
| 178 | static void unsuspendable_fd_destroy(struct unsuspendable_fd *entry) |
| 179 | { |
| 180 | if (!entry) { |
| 181 | return; |
| 182 | } |
| 183 | call_rcu(&entry->rcu_head, delete_unsuspendable_fd); |
| 184 | } |
| 185 | |
| 186 | static struct unsuspendable_fd *unsuspendable_fd_create( |
| 187 | const char *name, int fd) |
| 188 | { |
| 189 | struct unsuspendable_fd *entry = (unsuspendable_fd *) zmalloc(sizeof(*entry)); |
| 190 | |
| 191 | if (!entry) { |
| 192 | goto error; |
| 193 | } |
| 194 | if (name) { |
| 195 | entry->name = strdup(name); |
| 196 | if (!entry->name) { |
| 197 | goto error; |
| 198 | } |
| 199 | } |
| 200 | cds_lfht_node_init(&entry->tracker_node); |
| 201 | entry->fd = fd; |
| 202 | return entry; |
| 203 | error: |
| 204 | unsuspendable_fd_destroy(entry); |
| 205 | return NULL; |
| 206 | } |
| 207 | |
| 208 | static void fs_handle_tracked_log(struct fs_handle_tracked *handle) |
| 209 | { |
| 210 | const char *path; |
| 211 | |
| 212 | pthread_mutex_lock(&handle->lock); |
| 213 | lttng_inode_borrow_location(handle->inode, NULL, &path); |
| 214 | |
| 215 | if (handle->fd >= 0) { |
| 216 | DBG_NO_LOC(" %s [active, fd %d%s]", path, handle->fd, |
| 217 | handle->in_use ? ", in use" : ""); |
| 218 | } else { |
| 219 | DBG_NO_LOC(" %s [suspended]", path); |
| 220 | } |
| 221 | pthread_mutex_unlock(&handle->lock); |
| 222 | } |
| 223 | |
| 224 | /* Tracker lock must be held by the caller. */ |
| 225 | static int fs_handle_tracked_suspend(struct fs_handle_tracked *handle) |
| 226 | { |
| 227 | int ret = 0; |
| 228 | struct stat fs_stat; |
| 229 | const char *path; |
| 230 | const struct lttng_directory_handle *node_directory_handle; |
| 231 | |
| 232 | pthread_mutex_lock(&handle->lock); |
| 233 | lttng_inode_borrow_location( |
| 234 | handle->inode, &node_directory_handle, &path); |
| 235 | LTTNG_ASSERT(handle->fd >= 0); |
| 236 | if (handle->in_use) { |
| 237 | /* This handle can't be suspended as it is currently in use. */ |
| 238 | ret = -EAGAIN; |
| 239 | goto end; |
| 240 | } |
| 241 | |
| 242 | ret = lttng_directory_handle_stat( |
| 243 | node_directory_handle, path, &fs_stat); |
| 244 | if (ret) { |
| 245 | PERROR("Filesystem handle to %s cannot be suspended as stat() failed", |
| 246 | path); |
| 247 | ret = -errno; |
| 248 | goto end; |
| 249 | } |
| 250 | |
| 251 | if (fs_stat.st_ino != handle->ino) { |
| 252 | /* Don't suspend as the handle would not be restorable. */ |
| 253 | WARN("Filesystem handle to %s cannot be suspended as its inode changed", |
| 254 | path); |
| 255 | ret = -ENOENT; |
| 256 | goto end; |
| 257 | } |
| 258 | |
| 259 | handle->offset = lseek(handle->fd, 0, SEEK_CUR); |
| 260 | if (handle->offset == -1) { |
| 261 | WARN("Filesystem handle to %s cannot be suspended as lseek() failed to sample its current position", |
| 262 | path); |
| 263 | ret = -errno; |
| 264 | goto end; |
| 265 | } |
| 266 | |
| 267 | ret = close(handle->fd); |
| 268 | if (ret) { |
| 269 | PERROR("Filesystem handle to %s cannot be suspended as close() failed", |
| 270 | path); |
| 271 | ret = -errno; |
| 272 | goto end; |
| 273 | } |
| 274 | DBG("Suspended filesystem handle to %s (fd %i) at position %" PRId64, |
| 275 | path, handle->fd, handle->offset); |
| 276 | handle->fd = -1; |
| 277 | end: |
| 278 | if (ret) { |
| 279 | handle->tracker->stats.errors++; |
| 280 | } |
| 281 | pthread_mutex_unlock(&handle->lock); |
| 282 | return ret; |
| 283 | } |
| 284 | |
| 285 | /* Caller must hold the tracker and handle's locks. */ |
| 286 | static int fs_handle_tracked_restore(struct fs_handle_tracked *handle) |
| 287 | { |
| 288 | int ret, fd = -1; |
| 289 | const char *path; |
| 290 | const struct lttng_directory_handle *node_directory_handle; |
| 291 | |
| 292 | lttng_inode_borrow_location( |
| 293 | handle->inode, &node_directory_handle, &path); |
| 294 | |
| 295 | LTTNG_ASSERT(handle->fd == -1); |
| 296 | LTTNG_ASSERT(path); |
| 297 | ret = open_from_properties( |
| 298 | node_directory_handle, path, &handle->properties); |
| 299 | if (ret < 0) { |
| 300 | PERROR("Failed to restore filesystem handle to %s, open() failed", |
| 301 | path); |
| 302 | ret = -errno; |
| 303 | goto end; |
| 304 | } |
| 305 | fd = ret; |
| 306 | |
| 307 | ret = lseek(fd, handle->offset, SEEK_SET); |
| 308 | if (ret < 0) { |
| 309 | PERROR("Failed to restore filesystem handle to %s, lseek() failed", |
| 310 | path); |
| 311 | ret = -errno; |
| 312 | goto end; |
| 313 | } |
| 314 | DBG("Restored filesystem handle to %s (fd %i) at position %" PRId64, |
| 315 | path, fd, handle->offset); |
| 316 | ret = 0; |
| 317 | handle->fd = fd; |
| 318 | fd = -1; |
| 319 | end: |
| 320 | if (fd >= 0) { |
| 321 | (void) close(fd); |
| 322 | } |
| 323 | return ret; |
| 324 | } |
| 325 | |
| 326 | static int open_from_properties(const struct lttng_directory_handle *dir_handle, |
| 327 | const char *path, struct open_properties *properties) |
| 328 | { |
| 329 | int ret; |
| 330 | |
| 331 | /* |
| 332 | * open() ignores the 'flags' parameter unless the O_CREAT or O_TMPFILE |
| 333 | * flags are set. O_TMPFILE would not make sense in the context of a |
| 334 | * suspendable fs_handle as it would not be restorable (see OPEN(2)), |
| 335 | * thus it is ignored here. |
| 336 | */ |
| 337 | if ((properties->flags & O_CREAT) && properties->mode.is_set) { |
| 338 | ret = lttng_directory_handle_open_file(dir_handle, path, |
| 339 | properties->flags, properties->mode.value); |
| 340 | } else { |
| 341 | ret = lttng_directory_handle_open_file(dir_handle, path, |
| 342 | properties->flags, 0); |
| 343 | } |
| 344 | /* |
| 345 | * Some flags should not be used beyond the initial open() of a |
| 346 | * restorable file system handle. O_CREAT and O_TRUNC must |
| 347 | * be cleared since it would be unexpected to re-use them |
| 348 | * when the handle is retored: |
| 349 | * - O_CREAT should not be needed as the file has been created |
| 350 | * on the initial call to open(), |
| 351 | * - O_TRUNC would destroy the file's contents by truncating it |
| 352 | * to length 0. |
| 353 | */ |
| 354 | properties->flags &= ~(O_CREAT | O_TRUNC); |
| 355 | if (ret < 0) { |
| 356 | ret = -errno; |
| 357 | goto end; |
| 358 | } |
| 359 | end: |
| 360 | return ret; |
| 361 | } |
| 362 | |
| 363 | struct fd_tracker *fd_tracker_create(const char *unlinked_file_path, |
| 364 | unsigned int capacity) |
| 365 | { |
| 366 | struct fd_tracker *tracker = (fd_tracker *) zmalloc(sizeof(struct fd_tracker)); |
| 367 | |
| 368 | if (!tracker) { |
| 369 | goto end; |
| 370 | } |
| 371 | |
| 372 | pthread_mutex_lock(&seed.lock); |
| 373 | if (!seed.initialized) { |
| 374 | seed.value = (unsigned long) time(NULL); |
| 375 | seed.initialized = true; |
| 376 | } |
| 377 | pthread_mutex_unlock(&seed.lock); |
| 378 | |
| 379 | CDS_INIT_LIST_HEAD(&tracker->active_handles); |
| 380 | CDS_INIT_LIST_HEAD(&tracker->suspended_handles); |
| 381 | tracker->capacity = capacity; |
| 382 | tracker->unsuspendable_fds = cds_lfht_new(DEFAULT_HT_SIZE, 1, 0, |
| 383 | CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING, NULL); |
| 384 | if (!tracker->unsuspendable_fds) { |
| 385 | ERR("Failed to create fd-tracker's unsuspendable_fds hash table"); |
| 386 | goto error; |
| 387 | } |
| 388 | tracker->inode_registry = lttng_inode_registry_create(); |
| 389 | if (!tracker->inode_registry) { |
| 390 | ERR("Failed to create fd-tracker's inode registry"); |
| 391 | goto error; |
| 392 | } |
| 393 | |
| 394 | tracker->unlinked_file_pool = |
| 395 | lttng_unlinked_file_pool_create(unlinked_file_path); |
| 396 | if (!tracker->unlinked_file_pool) { |
| 397 | goto error; |
| 398 | } |
| 399 | DBG("File descriptor tracker created with a limit of %u simultaneously-opened FDs", |
| 400 | capacity); |
| 401 | end: |
| 402 | return tracker; |
| 403 | error: |
| 404 | fd_tracker_destroy(tracker); |
| 405 | return NULL; |
| 406 | } |
| 407 | |
| 408 | void fd_tracker_log(struct fd_tracker *tracker) |
| 409 | { |
| 410 | struct fs_handle_tracked *handle; |
| 411 | struct unsuspendable_fd *unsuspendable_fd; |
| 412 | struct cds_lfht_iter iter; |
| 413 | |
| 414 | pthread_mutex_lock(&tracker->lock); |
| 415 | DBG_NO_LOC("File descriptor tracker"); |
| 416 | DBG_NO_LOC(" Stats:"); |
| 417 | DBG_NO_LOC(" uses: %" PRIu64, tracker->stats.uses); |
| 418 | DBG_NO_LOC(" misses: %" PRIu64, tracker->stats.misses); |
| 419 | DBG_NO_LOC(" errors: %" PRIu64, tracker->stats.errors); |
| 420 | DBG_NO_LOC(" Tracked: %u", TRACKED_COUNT(tracker)); |
| 421 | DBG_NO_LOC(" active: %u", ACTIVE_COUNT(tracker)); |
| 422 | DBG_NO_LOC(" suspendable: %u", SUSPENDABLE_COUNT(tracker)); |
| 423 | DBG_NO_LOC(" unsuspendable: %u", UNSUSPENDABLE_COUNT(tracker)); |
| 424 | DBG_NO_LOC(" suspended: %u", SUSPENDED_COUNT(tracker)); |
| 425 | DBG_NO_LOC(" capacity: %u", tracker->capacity); |
| 426 | |
| 427 | DBG_NO_LOC(" Tracked suspendable file descriptors"); |
| 428 | cds_list_for_each_entry( |
| 429 | handle, &tracker->active_handles, handles_list_node) { |
| 430 | fs_handle_tracked_log(handle); |
| 431 | } |
| 432 | cds_list_for_each_entry(handle, &tracker->suspended_handles, |
| 433 | handles_list_node) { |
| 434 | fs_handle_tracked_log(handle); |
| 435 | } |
| 436 | if (!SUSPENDABLE_COUNT(tracker)) { |
| 437 | DBG_NO_LOC(" None"); |
| 438 | } |
| 439 | |
| 440 | DBG_NO_LOC(" Tracked unsuspendable file descriptors"); |
| 441 | rcu_read_lock(); |
| 442 | cds_lfht_for_each_entry(tracker->unsuspendable_fds, &iter, |
| 443 | unsuspendable_fd, tracker_node) { |
| 444 | DBG_NO_LOC(" %s [active, fd %d]", |
| 445 | unsuspendable_fd->name ?: "Unnamed", |
| 446 | unsuspendable_fd->fd); |
| 447 | } |
| 448 | rcu_read_unlock(); |
| 449 | if (!UNSUSPENDABLE_COUNT(tracker)) { |
| 450 | DBG_NO_LOC(" None"); |
| 451 | } |
| 452 | |
| 453 | pthread_mutex_unlock(&tracker->lock); |
| 454 | } |
| 455 | |
| 456 | int fd_tracker_destroy(struct fd_tracker *tracker) |
| 457 | { |
| 458 | int ret = 0; |
| 459 | |
| 460 | if (!tracker) { |
| 461 | goto end; |
| 462 | } |
| 463 | /* |
| 464 | * Refuse to destroy the tracker as fs_handles may still old |
| 465 | * weak references to the tracker. |
| 466 | */ |
| 467 | pthread_mutex_lock(&tracker->lock); |
| 468 | if (TRACKED_COUNT(tracker)) { |
| 469 | ERR("A file descriptor leak has been detected: %u tracked file descriptors are still being tracked", |
| 470 | TRACKED_COUNT(tracker)); |
| 471 | pthread_mutex_unlock(&tracker->lock); |
| 472 | fd_tracker_log(tracker); |
| 473 | ret = -1; |
| 474 | goto end; |
| 475 | } |
| 476 | pthread_mutex_unlock(&tracker->lock); |
| 477 | |
| 478 | if (tracker->unsuspendable_fds) { |
| 479 | ret = cds_lfht_destroy(tracker->unsuspendable_fds, NULL); |
| 480 | LTTNG_ASSERT(!ret); |
| 481 | } |
| 482 | |
| 483 | lttng_inode_registry_destroy(tracker->inode_registry); |
| 484 | lttng_unlinked_file_pool_destroy(tracker->unlinked_file_pool); |
| 485 | pthread_mutex_destroy(&tracker->lock); |
| 486 | free(tracker); |
| 487 | end: |
| 488 | return ret; |
| 489 | } |
| 490 | |
| 491 | struct fs_handle *fd_tracker_open_fs_handle(struct fd_tracker *tracker, |
| 492 | struct lttng_directory_handle *directory, |
| 493 | const char *path, |
| 494 | int flags, |
| 495 | mode_t *mode) |
| 496 | { |
| 497 | int ret; |
| 498 | struct fs_handle_tracked *handle = NULL; |
| 499 | struct stat fd_stat; |
| 500 | struct open_properties properties = { |
| 501 | .flags = flags, |
| 502 | .mode = { |
| 503 | .is_set = !!mode, |
| 504 | .value = static_cast<mode_t>(mode ? *mode : 0), |
| 505 | } |
| 506 | }; |
| 507 | |
| 508 | pthread_mutex_lock(&tracker->lock); |
| 509 | if (ACTIVE_COUNT(tracker) == tracker->capacity) { |
| 510 | if (tracker->count.suspendable.active > 0) { |
| 511 | ret = fd_tracker_suspend_handles(tracker, 1); |
| 512 | if (ret) { |
| 513 | goto end; |
| 514 | } |
| 515 | } else { |
| 516 | /* |
| 517 | * There are not enough active suspendable file |
| 518 | * descriptors to open a new fd and still accommodate |
| 519 | * the tracker's capacity. |
| 520 | */ |
| 521 | WARN("Cannot open file system handle, too many unsuspendable file descriptors are opened (%u)", |
| 522 | tracker->count.unsuspendable); |
| 523 | goto end; |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | handle = (fs_handle_tracked *) zmalloc(sizeof(*handle)); |
| 528 | if (!handle) { |
| 529 | goto end; |
| 530 | } |
| 531 | handle->parent = (typeof(handle->parent)) { |
| 532 | .get_fd = fs_handle_tracked_get_fd, |
| 533 | .put_fd = fs_handle_tracked_put_fd, |
| 534 | .unlink = fs_handle_tracked_unlink, |
| 535 | .close = fs_handle_tracked_close, |
| 536 | }; |
| 537 | |
| 538 | handle->tracker = tracker; |
| 539 | |
| 540 | ret = pthread_mutex_init(&handle->lock, NULL); |
| 541 | if (ret) { |
| 542 | PERROR("Failed to initialize handle mutex while creating fs handle"); |
| 543 | goto error_mutex_init; |
| 544 | } |
| 545 | |
| 546 | handle->fd = open_from_properties(directory, path, &properties); |
| 547 | if (handle->fd < 0) { |
| 548 | PERROR("Failed to open fs handle to %s, open() returned", path); |
| 549 | goto error; |
| 550 | } |
| 551 | |
| 552 | handle->properties = properties; |
| 553 | |
| 554 | handle->inode = lttng_inode_registry_get_inode(tracker->inode_registry, |
| 555 | directory, path, handle->fd, |
| 556 | tracker->unlinked_file_pool); |
| 557 | if (!handle->inode) { |
| 558 | ERR("Failed to get lttng_inode corresponding to file %s", path); |
| 559 | goto error; |
| 560 | } |
| 561 | |
| 562 | if (fstat(handle->fd, &fd_stat)) { |
| 563 | PERROR("Failed to retrieve file descriptor inode while creating fs handle, fstat() returned"); |
| 564 | goto error; |
| 565 | } |
| 566 | handle->ino = fd_stat.st_ino; |
| 567 | |
| 568 | fd_tracker_track(tracker, handle); |
| 569 | end: |
| 570 | pthread_mutex_unlock(&tracker->lock); |
| 571 | return handle ? &handle->parent : NULL; |
| 572 | error: |
| 573 | if (handle->inode) { |
| 574 | lttng_inode_put(handle->inode); |
| 575 | } |
| 576 | pthread_mutex_destroy(&handle->lock); |
| 577 | error_mutex_init: |
| 578 | free(handle); |
| 579 | handle = NULL; |
| 580 | goto end; |
| 581 | } |
| 582 | |
| 583 | /* Caller must hold the tracker's lock. */ |
| 584 | static int fd_tracker_suspend_handles( |
| 585 | struct fd_tracker *tracker, unsigned int count) |
| 586 | { |
| 587 | unsigned int left_to_close = count; |
| 588 | unsigned int attempts_left = tracker->count.suspendable.active; |
| 589 | struct fs_handle_tracked *handle, *tmp; |
| 590 | |
| 591 | cds_list_for_each_entry_safe(handle, tmp, &tracker->active_handles, |
| 592 | handles_list_node) { |
| 593 | int ret; |
| 594 | |
| 595 | fd_tracker_untrack(tracker, handle); |
| 596 | ret = fs_handle_tracked_suspend(handle); |
| 597 | fd_tracker_track(tracker, handle); |
| 598 | if (!ret) { |
| 599 | left_to_close--; |
| 600 | } |
| 601 | attempts_left--; |
| 602 | |
| 603 | if (left_to_close == 0 || attempts_left == 0) { |
| 604 | break; |
| 605 | } |
| 606 | } |
| 607 | return left_to_close ? -EMFILE : 0; |
| 608 | } |
| 609 | |
| 610 | int fd_tracker_open_unsuspendable_fd(struct fd_tracker *tracker, |
| 611 | int *out_fds, |
| 612 | const char **names, |
| 613 | unsigned int fd_count, |
| 614 | fd_open_cb open, |
| 615 | void *user_data) |
| 616 | { |
| 617 | int ret, user_ret, i, fds_to_suspend; |
| 618 | unsigned int active_fds; |
| 619 | struct unsuspendable_fd **entries; |
| 620 | |
| 621 | entries = (unsuspendable_fd **) zmalloc(fd_count * sizeof(*entries)); |
| 622 | if (!entries) { |
| 623 | ret = -1; |
| 624 | goto end; |
| 625 | } |
| 626 | |
| 627 | pthread_mutex_lock(&tracker->lock); |
| 628 | |
| 629 | active_fds = ACTIVE_COUNT(tracker); |
| 630 | fds_to_suspend = (int) active_fds + (int) fd_count - |
| 631 | (int) tracker->capacity; |
| 632 | if (fds_to_suspend > 0) { |
| 633 | if (fds_to_suspend <= tracker->count.suspendable.active) { |
| 634 | ret = fd_tracker_suspend_handles( |
| 635 | tracker, fds_to_suspend); |
| 636 | if (ret) { |
| 637 | goto end_unlock; |
| 638 | } |
| 639 | } else { |
| 640 | /* |
| 641 | * There are not enough active suspendable file |
| 642 | * descriptors to open a new fd and still accommodates the |
| 643 | * tracker's capacity. |
| 644 | */ |
| 645 | WARN("Cannot open unsuspendable fd, too many unsuspendable file descriptors are opened (%u)", |
| 646 | tracker->count.unsuspendable); |
| 647 | ret = -EMFILE; |
| 648 | goto end_unlock; |
| 649 | } |
| 650 | } |
| 651 | |
| 652 | user_ret = open(user_data, out_fds); |
| 653 | if (user_ret) { |
| 654 | ret = user_ret; |
| 655 | goto end_unlock; |
| 656 | } |
| 657 | |
| 658 | /* |
| 659 | * Add the fds returned by the user's callback to the hashtable |
| 660 | * of unsuspendable fds. |
| 661 | */ |
| 662 | for (i = 0; i < fd_count; i++) { |
| 663 | struct unsuspendable_fd *entry = unsuspendable_fd_create( |
| 664 | names ? names[i] : NULL, out_fds[i]); |
| 665 | |
| 666 | if (!entry) { |
| 667 | ret = -1; |
| 668 | goto end_free_entries; |
| 669 | } |
| 670 | entries[i] = entry; |
| 671 | } |
| 672 | |
| 673 | rcu_read_lock(); |
| 674 | for (i = 0; i < fd_count; i++) { |
| 675 | struct cds_lfht_node *node; |
| 676 | struct unsuspendable_fd *entry = entries[i]; |
| 677 | |
| 678 | node = cds_lfht_add_unique(tracker->unsuspendable_fds, |
| 679 | hash_key_ulong((void *) (unsigned long) |
| 680 | out_fds[i], |
| 681 | seed.value), |
| 682 | match_fd, (void *) (unsigned long) out_fds[i], |
| 683 | &entry->tracker_node); |
| 684 | |
| 685 | if (node != &entry->tracker_node) { |
| 686 | ret = -EEXIST; |
| 687 | rcu_read_unlock(); |
| 688 | goto end_free_entries; |
| 689 | } |
| 690 | entries[i] = NULL; |
| 691 | } |
| 692 | tracker->count.unsuspendable += fd_count; |
| 693 | rcu_read_unlock(); |
| 694 | ret = user_ret; |
| 695 | end_unlock: |
| 696 | pthread_mutex_unlock(&tracker->lock); |
| 697 | end: |
| 698 | free(entries); |
| 699 | return ret; |
| 700 | end_free_entries: |
| 701 | for (i = 0; i < fd_count; i++) { |
| 702 | unsuspendable_fd_destroy(entries[i]); |
| 703 | } |
| 704 | goto end_unlock; |
| 705 | } |
| 706 | |
| 707 | int fd_tracker_close_unsuspendable_fd(struct fd_tracker *tracker, |
| 708 | int *fds_in, |
| 709 | unsigned int fd_count, |
| 710 | fd_close_cb close, |
| 711 | void *user_data) |
| 712 | { |
| 713 | int i, ret, user_ret; |
| 714 | int *fds = NULL; |
| 715 | |
| 716 | /* |
| 717 | * Maintain a local copy of fds_in as the user's callback may modify its |
| 718 | * contents (e.g. setting the fd(s) to -1 after close). |
| 719 | */ |
| 720 | fds = (int *) malloc(sizeof(*fds) * fd_count); |
| 721 | if (!fds) { |
| 722 | ret = -1; |
| 723 | goto end; |
| 724 | } |
| 725 | memcpy(fds, fds_in, sizeof(*fds) * fd_count); |
| 726 | |
| 727 | pthread_mutex_lock(&tracker->lock); |
| 728 | rcu_read_lock(); |
| 729 | |
| 730 | /* Let the user close the file descriptors. */ |
| 731 | user_ret = close(user_data, fds_in); |
| 732 | if (user_ret) { |
| 733 | ret = user_ret; |
| 734 | goto end_unlock; |
| 735 | } |
| 736 | |
| 737 | /* Untrack the fds that were just closed by the user's callback. */ |
| 738 | for (i = 0; i < fd_count; i++) { |
| 739 | struct cds_lfht_node *node; |
| 740 | struct cds_lfht_iter iter; |
| 741 | struct unsuspendable_fd *entry; |
| 742 | |
| 743 | cds_lfht_lookup(tracker->unsuspendable_fds, |
| 744 | hash_key_ulong((void *) (unsigned long) fds[i], |
| 745 | seed.value), |
| 746 | match_fd, (void *) (unsigned long) fds[i], |
| 747 | &iter); |
| 748 | node = cds_lfht_iter_get_node(&iter); |
| 749 | if (!node) { |
| 750 | /* Unknown file descriptor. */ |
| 751 | WARN("Untracked file descriptor %d passed to fd_tracker_close_unsuspendable_fd()", |
| 752 | fds[i]); |
| 753 | ret = -EINVAL; |
| 754 | goto end_unlock; |
| 755 | } |
| 756 | entry = caa_container_of( |
| 757 | node, struct unsuspendable_fd, tracker_node); |
| 758 | |
| 759 | cds_lfht_del(tracker->unsuspendable_fds, node); |
| 760 | unsuspendable_fd_destroy(entry); |
| 761 | fds[i] = -1; |
| 762 | } |
| 763 | |
| 764 | tracker->count.unsuspendable -= fd_count; |
| 765 | ret = 0; |
| 766 | end_unlock: |
| 767 | rcu_read_unlock(); |
| 768 | pthread_mutex_unlock(&tracker->lock); |
| 769 | free(fds); |
| 770 | end: |
| 771 | return ret; |
| 772 | } |
| 773 | |
| 774 | /* Caller must have taken the tracker's and handle's locks. */ |
| 775 | static void fd_tracker_track( |
| 776 | struct fd_tracker *tracker, struct fs_handle_tracked *handle) |
| 777 | { |
| 778 | if (handle->fd >= 0) { |
| 779 | tracker->count.suspendable.active++; |
| 780 | cds_list_add_tail(&handle->handles_list_node, |
| 781 | &tracker->active_handles); |
| 782 | } else { |
| 783 | tracker->count.suspendable.suspended++; |
| 784 | cds_list_add_tail(&handle->handles_list_node, |
| 785 | &tracker->suspended_handles); |
| 786 | } |
| 787 | } |
| 788 | |
| 789 | /* Caller must have taken the tracker's and handle's locks. */ |
| 790 | static void fd_tracker_untrack( |
| 791 | struct fd_tracker *tracker, struct fs_handle_tracked *handle) |
| 792 | { |
| 793 | if (handle->fd >= 0) { |
| 794 | tracker->count.suspendable.active--; |
| 795 | } else { |
| 796 | tracker->count.suspendable.suspended--; |
| 797 | } |
| 798 | cds_list_del(&handle->handles_list_node); |
| 799 | } |
| 800 | |
| 801 | /* Caller must have taken the tracker's and handle's locks. */ |
| 802 | static int fd_tracker_restore_handle( |
| 803 | struct fd_tracker *tracker, struct fs_handle_tracked *handle) |
| 804 | { |
| 805 | int ret; |
| 806 | |
| 807 | fd_tracker_untrack(tracker, handle); |
| 808 | if (ACTIVE_COUNT(tracker) >= tracker->capacity) { |
| 809 | ret = fd_tracker_suspend_handles(tracker, 1); |
| 810 | if (ret) { |
| 811 | goto end; |
| 812 | } |
| 813 | } |
| 814 | ret = fs_handle_tracked_restore(handle); |
| 815 | end: |
| 816 | fd_tracker_track(tracker, handle); |
| 817 | return ret ? ret : handle->fd; |
| 818 | } |
| 819 | |
| 820 | static int fs_handle_tracked_get_fd(struct fs_handle *_handle) |
| 821 | { |
| 822 | int ret; |
| 823 | struct fs_handle_tracked *handle = |
| 824 | container_of(_handle, struct fs_handle_tracked, parent); |
| 825 | |
| 826 | /* |
| 827 | * TODO This should be optimized as it is a fairly hot path. |
| 828 | * The fd-tracker's lock should only be taken when a fs_handle is |
| 829 | * restored (slow path). On the fast path (fs_handle is active), |
| 830 | * the only effect on the fd_tracker is marking the handle as the |
| 831 | * most recently used. Currently, it is done by a call to the |
| 832 | * track/untrack helpers, but it should be done atomically. |
| 833 | * |
| 834 | * Note that the lock's nesting order must still be respected here. |
| 835 | * The handle's lock nests inside the tracker's lock. |
| 836 | */ |
| 837 | pthread_mutex_lock(&handle->tracker->lock); |
| 838 | pthread_mutex_lock(&handle->lock); |
| 839 | LTTNG_ASSERT(!handle->in_use); |
| 840 | |
| 841 | handle->tracker->stats.uses++; |
| 842 | if (handle->fd >= 0) { |
| 843 | ret = handle->fd; |
| 844 | /* Mark as most recently used. */ |
| 845 | fd_tracker_untrack(handle->tracker, handle); |
| 846 | fd_tracker_track(handle->tracker, handle); |
| 847 | } else { |
| 848 | handle->tracker->stats.misses++; |
| 849 | ret = fd_tracker_restore_handle(handle->tracker, handle); |
| 850 | if (ret < 0) { |
| 851 | handle->tracker->stats.errors++; |
| 852 | goto end; |
| 853 | } |
| 854 | } |
| 855 | handle->in_use = true; |
| 856 | end: |
| 857 | pthread_mutex_unlock(&handle->lock); |
| 858 | pthread_mutex_unlock(&handle->tracker->lock); |
| 859 | return ret; |
| 860 | } |
| 861 | |
| 862 | static void fs_handle_tracked_put_fd(struct fs_handle *_handle) |
| 863 | { |
| 864 | struct fs_handle_tracked *handle = |
| 865 | container_of(_handle, struct fs_handle_tracked, parent); |
| 866 | |
| 867 | pthread_mutex_lock(&handle->lock); |
| 868 | handle->in_use = false; |
| 869 | pthread_mutex_unlock(&handle->lock); |
| 870 | } |
| 871 | |
| 872 | static int fs_handle_tracked_unlink(struct fs_handle *_handle) |
| 873 | { |
| 874 | int ret; |
| 875 | struct fs_handle_tracked *handle = |
| 876 | container_of(_handle, struct fs_handle_tracked, parent); |
| 877 | |
| 878 | pthread_mutex_lock(&handle->tracker->lock); |
| 879 | pthread_mutex_lock(&handle->lock); |
| 880 | ret = lttng_inode_unlink(handle->inode); |
| 881 | pthread_mutex_unlock(&handle->lock); |
| 882 | pthread_mutex_unlock(&handle->tracker->lock); |
| 883 | return ret; |
| 884 | } |
| 885 | |
| 886 | static int fs_handle_tracked_close(struct fs_handle *_handle) |
| 887 | { |
| 888 | int ret = 0; |
| 889 | const char *path = NULL; |
| 890 | struct fs_handle_tracked *handle = |
| 891 | container_of(_handle, struct fs_handle_tracked, parent); |
| 892 | struct lttng_directory_handle *inode_directory_handle = NULL; |
| 893 | |
| 894 | if (!handle) { |
| 895 | ret = -EINVAL; |
| 896 | goto end; |
| 897 | } |
| 898 | |
| 899 | pthread_mutex_lock(&handle->tracker->lock); |
| 900 | pthread_mutex_lock(&handle->lock); |
| 901 | if (handle->inode) { |
| 902 | lttng_inode_borrow_location(handle->inode, NULL, &path); |
| 903 | /* |
| 904 | * Here a reference to the inode's directory handle is acquired |
| 905 | * to prevent the last reference to it from being released while |
| 906 | * the tracker's lock is taken. |
| 907 | * |
| 908 | * If this wasn't done, the directory handle could attempt to |
| 909 | * close its underlying directory file descriptor, which would |
| 910 | * attempt to lock the tracker's lock, resulting in a deadlock. |
| 911 | * |
| 912 | * Since a new reference to the directory handle is taken within |
| 913 | * the scope of this function, it is not possible for the last |
| 914 | * reference to the inode's location directory handle to be |
| 915 | * released during the call to lttng_inode_put(). |
| 916 | * |
| 917 | * We wait until the tracker's lock is released to release the |
| 918 | * reference. Hence, the call to the tracker is delayed just |
| 919 | * enough to not attempt to recursively acquire the tracker's |
| 920 | * lock twice. |
| 921 | */ |
| 922 | inode_directory_handle = |
| 923 | lttng_inode_get_location_directory_handle( |
| 924 | handle->inode); |
| 925 | } |
| 926 | fd_tracker_untrack(handle->tracker, handle); |
| 927 | if (handle->fd >= 0) { |
| 928 | /* |
| 929 | * The return value of close() is not propagated as there |
| 930 | * isn't much the user can do about it. |
| 931 | */ |
| 932 | if (close(handle->fd)) { |
| 933 | PERROR("Failed to close the file descriptor (%d) of fs handle to %s, close() returned", |
| 934 | handle->fd, path ? path : "Unknown"); |
| 935 | } |
| 936 | handle->fd = -1; |
| 937 | } |
| 938 | if (handle->inode) { |
| 939 | lttng_inode_put(handle->inode); |
| 940 | } |
| 941 | pthread_mutex_unlock(&handle->lock); |
| 942 | pthread_mutex_destroy(&handle->lock); |
| 943 | pthread_mutex_unlock(&handle->tracker->lock); |
| 944 | free(handle); |
| 945 | lttng_directory_handle_put(inode_directory_handle); |
| 946 | end: |
| 947 | return ret; |
| 948 | } |