src/common/fd-tracker/fd-tracker.c

   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 = 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         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         assert(handle->fd == -1);
 296         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 = 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                 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.is_set = !!mode,
 503                 .mode.value = mode ? *mode : 0,
 504         };
 505
 506         pthread_mutex_lock(&tracker->lock);
 507         if (ACTIVE_COUNT(tracker) == tracker->capacity) {
 508                 if (tracker->count.suspendable.active > 0) {
 509                         ret = fd_tracker_suspend_handles(tracker, 1);
 510                         if (ret) {
 511                                 goto end;
 512                         }
 513                 } else {
 514                         /*
 515                          * There are not enough active suspendable file
 516                          * descriptors to open a new fd and still accommodate
 517                          * the tracker's capacity.
 518                          */
 519                         WARN("Cannot open file system handle, too many unsuspendable file descriptors are opened (%u)",
 520                                         tracker->count.unsuspendable);
 521                         goto end;
 522                 }
 523         }
 524
 525         handle = zmalloc(sizeof(*handle));
 526         if (!handle) {
 527                 goto end;
 528         }
 529         handle->parent = (typeof(handle->parent)) {
 530                 .get_fd = fs_handle_tracked_get_fd,
 531                 .put_fd = fs_handle_tracked_put_fd,
 532                 .unlink = fs_handle_tracked_unlink,
 533                 .close = fs_handle_tracked_close,
 534         };
 535
 536         handle->tracker = tracker;
 537
 538         ret = pthread_mutex_init(&handle->lock, NULL);
 539         if (ret) {
 540                 PERROR("Failed to initialize handle mutex while creating fs handle");
 541                 goto error_mutex_init;
 542         }
 543
 544         handle->fd = open_from_properties(directory, path, &properties);
 545         if (handle->fd < 0) {
 546                 PERROR("Failed to open fs handle to %s, open() returned", path);
 547                 goto error;
 548         }
 549
 550         handle->properties = properties;
 551
 552         handle->inode = lttng_inode_registry_get_inode(tracker->inode_registry,
 553                         directory, path, handle->fd,
 554                         tracker->unlinked_file_pool);
 555         if (!handle->inode) {
 556                 ERR("Failed to get lttng_inode corresponding to file %s", path);
 557                 goto error;
 558         }
 559
 560         if (fstat(handle->fd, &fd_stat)) {
 561                 PERROR("Failed to retrieve file descriptor inode while creating fs handle, fstat() returned");
 562                 goto error;
 563         }
 564         handle->ino = fd_stat.st_ino;
 565
 566         fd_tracker_track(tracker, handle);
 567 end:
 568         pthread_mutex_unlock(&tracker->lock);
 569         return handle ? &handle->parent : NULL;
 570 error:
 571         if (handle->inode) {
 572                 lttng_inode_put(handle->inode);
 573         }
 574         pthread_mutex_destroy(&handle->lock);
 575 error_mutex_init:
 576         free(handle);
 577         handle = NULL;
 578         goto end;
 579 }
 580
 581 /* Caller must hold the tracker's lock. */
 582 static int fd_tracker_suspend_handles(
 583                 struct fd_tracker *tracker, unsigned int count)
 584 {
 585         unsigned int left_to_close = count;
 586         unsigned int attempts_left = tracker->count.suspendable.active;
 587         struct fs_handle_tracked *handle, *tmp;
 588
 589         cds_list_for_each_entry_safe(handle, tmp, &tracker->active_handles,
 590                         handles_list_node) {
 591                 int ret;
 592
 593                 fd_tracker_untrack(tracker, handle);
 594                 ret = fs_handle_tracked_suspend(handle);
 595                 fd_tracker_track(tracker, handle);
 596                 if (!ret) {
 597                         left_to_close--;
 598                 }
 599                 attempts_left--;
 600
 601                 if (left_to_close == 0 || attempts_left == 0) {
 602                         break;
 603                 }
 604         }
 605         return left_to_close ? -EMFILE : 0;
 606 }
 607
 608 int fd_tracker_open_unsuspendable_fd(struct fd_tracker *tracker,
 609                 int *out_fds,
 610                 const char **names,
 611                 unsigned int fd_count,
 612                 fd_open_cb open,
 613                 void *user_data)
 614 {
 615         int ret, user_ret, i, fds_to_suspend;
 616         unsigned int active_fds;
 617         struct unsuspendable_fd **entries;
 618
 619         entries = zmalloc(fd_count * sizeof(*entries));
 620         if (!entries) {
 621                 ret = -1;
 622                 goto end;
 623         }
 624
 625         pthread_mutex_lock(&tracker->lock);
 626
 627         active_fds = ACTIVE_COUNT(tracker);
 628         fds_to_suspend = (int) active_fds + (int) fd_count -
 629                         (int) tracker->capacity;
 630         if (fds_to_suspend > 0) {
 631                 if (fds_to_suspend <= tracker->count.suspendable.active) {
 632                         ret = fd_tracker_suspend_handles(
 633                                         tracker, fds_to_suspend);
 634                         if (ret) {
 635                                 goto end_unlock;
 636                         }
 637                 } else {
 638                         /*
 639                          * There are not enough active suspendable file
 640                          * descriptors to open a new fd and still accomodate the
 641                          * tracker's capacity.
 642                          */
 643                         WARN("Cannot open unsuspendable fd, too many unsuspendable file descriptors are opened (%u)",
 644                                         tracker->count.unsuspendable);
 645                         ret = -EMFILE;
 646                         goto end_unlock;
 647                 }
 648         }
 649
 650         user_ret = open(user_data, out_fds);
 651         if (user_ret) {
 652                 ret = user_ret;
 653                 goto end_unlock;
 654         }
 655
 656         /*
 657          * Add the fds returned by the user's callback to the hashtable
 658          * of unsuspendable fds.
 659          */
 660         for (i = 0; i < fd_count; i++) {
 661                 struct unsuspendable_fd *entry = unsuspendable_fd_create(
 662                                 names ? names[i] : NULL, out_fds[i]);
 663
 664                 if (!entry) {
 665                         ret = -1;
 666                         goto end_free_entries;
 667                 }
 668                 entries[i] = entry;
 669         }
 670
 671         rcu_read_lock();
 672         for (i = 0; i < fd_count; i++) {
 673                 struct cds_lfht_node *node;
 674                 struct unsuspendable_fd *entry = entries[i];
 675
 676                 node = cds_lfht_add_unique(tracker->unsuspendable_fds,
 677                                 hash_key_ulong((void *) (unsigned long)
 678                                                                 out_fds[i],
 679                                                 seed.value),
 680                                 match_fd, (void *) (unsigned long) out_fds[i],
 681                                 &entry->tracker_node);
 682
 683                 if (node != &entry->tracker_node) {
 684                         ret = -EEXIST;
 685                         rcu_read_unlock();
 686                         goto end_free_entries;
 687                 }
 688                 entries[i] = NULL;
 689         }
 690         tracker->count.unsuspendable += fd_count;
 691         rcu_read_unlock();
 692         ret = user_ret;
 693 end_unlock:
 694         pthread_mutex_unlock(&tracker->lock);
 695 end:
 696         free(entries);
 697         return ret;
 698 end_free_entries:
 699         for (i = 0; i < fd_count; i++) {
 700                 unsuspendable_fd_destroy(entries[i]);
 701         }
 702         goto end_unlock;
 703 }
 704
 705 int fd_tracker_close_unsuspendable_fd(struct fd_tracker *tracker,
 706                 int *fds_in,
 707                 unsigned int fd_count,
 708                 fd_close_cb close,
 709                 void *user_data)
 710 {
 711         int i, ret, user_ret;
 712         int *fds = NULL;
 713
 714         /*
 715          * Maintain a local copy of fds_in as the user's callback may modify its
 716          * contents (e.g. setting the fd(s) to -1 after close).
 717          */
 718         fds = malloc(sizeof(*fds) * fd_count);
 719         if (!fds) {
 720                 ret = -1;
 721                 goto end;
 722         }
 723         memcpy(fds, fds_in, sizeof(*fds) * fd_count);
 724
 725         pthread_mutex_lock(&tracker->lock);
 726         rcu_read_lock();
 727
 728         /* Let the user close the file descriptors. */
 729         user_ret = close(user_data, fds_in);
 730         if (user_ret) {
 731                 ret = user_ret;
 732                 goto end_unlock;
 733         }
 734
 735         /* Untrack the fds that were just closed by the user's callback. */
 736         for (i = 0; i < fd_count; i++) {
 737                 struct cds_lfht_node *node;
 738                 struct cds_lfht_iter iter;
 739                 struct unsuspendable_fd *entry;
 740
 741                 cds_lfht_lookup(tracker->unsuspendable_fds,
 742                                 hash_key_ulong((void *) (unsigned long) fds[i],
 743                                                 seed.value),
 744                                 match_fd, (void *) (unsigned long) fds[i],
 745                                 &iter);
 746                 node = cds_lfht_iter_get_node(&iter);
 747                 if (!node) {
 748                         /* Unknown file descriptor. */
 749                         WARN("Untracked file descriptor %d passed to fd_tracker_close_unsuspendable_fd()",
 750                                         fds[i]);
 751                         ret = -EINVAL;
 752                         goto end_unlock;
 753                 }
 754                 entry = caa_container_of(
 755                                 node, struct unsuspendable_fd, tracker_node);
 756
 757                 cds_lfht_del(tracker->unsuspendable_fds, node);
 758                 unsuspendable_fd_destroy(entry);
 759                 fds[i] = -1;
 760         }
 761
 762         tracker->count.unsuspendable -= fd_count;
 763         ret = 0;
 764 end_unlock:
 765         rcu_read_unlock();
 766         pthread_mutex_unlock(&tracker->lock);
 767         free(fds);
 768 end:
 769         return ret;
 770 }
 771
 772 /* Caller must have taken the tracker's and handle's locks. */
 773 static void fd_tracker_track(
 774                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 775 {
 776         if (handle->fd >= 0) {
 777                 tracker->count.suspendable.active++;
 778                 cds_list_add_tail(&handle->handles_list_node,
 779                                 &tracker->active_handles);
 780         } else {
 781                 tracker->count.suspendable.suspended++;
 782                 cds_list_add_tail(&handle->handles_list_node,
 783                                 &tracker->suspended_handles);
 784         }
 785 }
 786
 787 /* Caller must have taken the tracker's and handle's locks. */
 788 static void fd_tracker_untrack(
 789                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 790 {
 791         if (handle->fd >= 0) {
 792                 tracker->count.suspendable.active--;
 793         } else {
 794                 tracker->count.suspendable.suspended--;
 795         }
 796         cds_list_del(&handle->handles_list_node);
 797 }
 798
 799 /* Caller must have taken the tracker's and handle's locks. */
 800 static int fd_tracker_restore_handle(
 801                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 802 {
 803         int ret;
 804
 805         fd_tracker_untrack(tracker, handle);
 806         if (ACTIVE_COUNT(tracker) >= tracker->capacity) {
 807                 ret = fd_tracker_suspend_handles(tracker, 1);
 808                 if (ret) {
 809                         goto end;
 810                 }
 811         }
 812         ret = fs_handle_tracked_restore(handle);
 813 end:
 814         fd_tracker_track(tracker, handle);
 815         return ret ? ret : handle->fd;
 816 }
 817
 818 static int fs_handle_tracked_get_fd(struct fs_handle *_handle)
 819 {
 820         int ret;
 821         struct fs_handle_tracked *handle =
 822                         container_of(_handle, struct fs_handle_tracked, parent);
 823
 824         /*
 825          * TODO This should be optimized as it is a fairly hot path.
 826          * The fd-tracker's lock should only be taken when a fs_handle is
 827          * restored (slow path). On the fast path (fs_handle is active),
 828          * the only effect on the fd_tracker is marking the handle as the
 829          * most recently used. Currently, it is done by a call to the
 830          * track/untrack helpers, but it should be done atomically.
 831          *
 832          * Note that the lock's nesting order must still be respected here.
 833          * The handle's lock nests inside the tracker's lock.
 834          */
 835         pthread_mutex_lock(&handle->tracker->lock);
 836         pthread_mutex_lock(&handle->lock);
 837         assert(!handle->in_use);
 838
 839         handle->tracker->stats.uses++;
 840         if (handle->fd >= 0) {
 841                 ret = handle->fd;
 842                 /* Mark as most recently used. */
 843                 fd_tracker_untrack(handle->tracker, handle);
 844                 fd_tracker_track(handle->tracker, handle);
 845         } else {
 846                 handle->tracker->stats.misses++;
 847                 ret = fd_tracker_restore_handle(handle->tracker, handle);
 848                 if (ret < 0) {
 849                         handle->tracker->stats.errors++;
 850                         goto end;
 851                 }
 852         }
 853         handle->in_use = true;
 854 end:
 855         pthread_mutex_unlock(&handle->lock);
 856         pthread_mutex_unlock(&handle->tracker->lock);
 857         return ret;
 858 }
 859
 860 static void fs_handle_tracked_put_fd(struct fs_handle *_handle)
 861 {
 862         struct fs_handle_tracked *handle =
 863                         container_of(_handle, struct fs_handle_tracked, parent);
 864
 865         pthread_mutex_lock(&handle->lock);
 866         handle->in_use = false;
 867         pthread_mutex_unlock(&handle->lock);
 868 }
 869
 870 static int fs_handle_tracked_unlink(struct fs_handle *_handle)
 871 {
 872         int ret;
 873         struct fs_handle_tracked *handle =
 874                         container_of(_handle, struct fs_handle_tracked, parent);
 875
 876         pthread_mutex_lock(&handle->tracker->lock);
 877         pthread_mutex_lock(&handle->lock);
 878         ret = lttng_inode_unlink(handle->inode);
 879         pthread_mutex_unlock(&handle->lock);
 880         pthread_mutex_unlock(&handle->tracker->lock);
 881         return ret;
 882 }
 883
 884 static int fs_handle_tracked_close(struct fs_handle *_handle)
 885 {
 886         int ret = 0;
 887         const char *path = NULL;
 888         struct fs_handle_tracked *handle =
 889                         container_of(_handle, struct fs_handle_tracked, parent);
 890         struct lttng_directory_handle *inode_directory_handle = NULL;
 891
 892         if (!handle) {
 893                 ret = -EINVAL;
 894                 goto end;
 895         }
 896
 897         pthread_mutex_lock(&handle->tracker->lock);
 898         pthread_mutex_lock(&handle->lock);
 899         if (handle->inode) {
 900                 lttng_inode_borrow_location(handle->inode, NULL, &path);
 901                 /*
 902                  * Here a reference to the inode's directory handle is acquired
 903                  * to prevent the last reference to it from being released while
 904                  * the tracker's lock is taken.
 905                  *
 906                  * If this wasn't done, the directory handle could attempt to
 907                  * close its underlying directory file descriptor, which would
 908                  * attempt to lock the tracker's lock, resulting in a deadlock.
 909                  *
 910                  * Since a new reference to the directory handle is taken within
 911                  * the scope of this function, it is not possible for the last
 912                  * reference to the inode's location directory handle to be
 913                  * released during the call to lttng_inode_put().
 914                  *
 915                  * We wait until the tracker's lock is released to release the
 916                  * reference. Hence, the call to the tracker is delayed just
 917                  * enough to not attempt to recursively acquire the tracker's
 918                  * lock twice.
 919                  */
 920                 inode_directory_handle =
 921                                 lttng_inode_get_location_directory_handle(
 922                                                 handle->inode);
 923         }
 924         fd_tracker_untrack(handle->tracker, handle);
 925         if (handle->fd >= 0) {
 926                 /*
 927                  * The return value of close() is not propagated as there
 928                  * isn't much the user can do about it.
 929                  */
 930                 if (close(handle->fd)) {
 931                         PERROR("Failed to close the file descritptor (%d) of fs handle to %s, close() returned",
 932                                         handle->fd, path ? path : "Unknown");
 933                 }
 934                 handle->fd = -1;
 935         }
 936         if (handle->inode) {
 937                 lttng_inode_put(handle->inode);
 938         }
 939         pthread_mutex_unlock(&handle->lock);
 940         pthread_mutex_destroy(&handle->lock);
 941         pthread_mutex_unlock(&handle->tracker->lock);
 942         free(handle);
 943         lttng_directory_handle_put(inode_directory_handle);
 944 end:
 945         return ret;
 946 }