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

   1 /*
   2  * Copyright (C) 2018 - Jérémie Galarneau <jeremie.galarneau@efficios.com>
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License, version 2 only, as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 51
  15  * Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  16  */
  17
  18 #include <urcu.h>
  19 #include <urcu/list.h>
  20 #include <urcu/rculfhash.h>
  21
  22 #include <sys/stat.h>
  23 #include <sys/types.h>
  24 #include <fcntl.h>
  25 #include <stdbool.h>
  26 #include <pthread.h>
  27 #include <inttypes.h>
  28
  29 #include "common/macros.h"
  30 #include "common/error.h"
  31 #include "common/defaults.h"
  32 #include "common/hashtable/utils.h"
  33 #include "common/hashtable/hashtable.h"
  34
  35 #include "fd-tracker.h"
  36
  37 /* Tracker lock must be taken by the user. */
  38 #define TRACKED_COUNT(tracker)                 \
  39         (tracker->count.suspendable.active +   \
  40         tracker->count.suspendable.suspended + \
  41         tracker->count.unsuspendable)
  42
  43 /* Tracker lock must be taken by the user. */
  44 #define ACTIVE_COUNT(tracker)                  \
  45         (tracker->count.suspendable.active +   \
  46         tracker->count.unsuspendable)
  47
  48 /* Tracker lock must be taken by the user. */
  49 #define SUSPENDED_COUNT(tracker)               \
  50         (tracker->count.suspendable.suspended)
  51
  52 /* Tracker lock must be taken by the user. */
  53 #define SUSPENDABLE_COUNT(tracker)             \
  54         (tracker->count.suspendable.active +   \
  55         tracker->count.suspendable.suspended)
  56
  57 /* Tracker lock must be taken by the user. */
  58 #define UNSUSPENDABLE_COUNT(tracker)           \
  59         (tracker->count.unsuspendable)
  60
  61 struct fd_tracker {
  62         pthread_mutex_t lock;
  63         struct {
  64                 struct {
  65                         unsigned int active;
  66                         unsigned int suspended;
  67                 } suspendable;
  68                 unsigned int unsuspendable;
  69         } count;
  70         unsigned int capacity;
  71         struct {
  72                 uint64_t uses;
  73                 uint64_t misses;
  74                 /* Failures to suspend or restore fs handles. */
  75                 uint64_t errors;
  76         } stats;
  77         /*
  78          * The head of the active_handles list is always the least recently
  79          * used active handle. When an handle is used, it is removed from the
  80          * list and added to the end. When a file has to be suspended, the
  81          * first element in the list is "popped", suspended, and added to the
  82          * list of suspended handles.
  83          */
  84         struct cds_list_head active_handles;
  85         struct cds_list_head suspended_handles;
  86         struct cds_lfht *unsuspendable_fds;
  87 };
  88
  89 struct open_properties {
  90         char *path;
  91         int flags;
  92         struct {
  93                 bool is_set;
  94                 mode_t value;
  95         } mode;
  96 };
  97
  98 /*
  99  * A fs_handle is not ref-counted. Therefore, it is assumed that a
 100  * handle is never in-use while it is being reclaimed. It can be
 101  * shared by multiple threads, but external synchronization is required
 102  * to ensure it is not still being used when it is reclaimed (close method).
 103  * In this respect, it is not different from a regular file descriptor.
 104  *
 105  * The fs_handle lock always nests _within_ the tracker's lock.
 106  */
 107 struct fs_handle {
 108         pthread_mutex_t lock;
 109         /*
 110          * Weak reference to the tracker. All fs_handles are assumed to have
 111          * been closed at the moment of the destruction of the fd_tracker.
 112          */
 113         struct fd_tracker *tracker;
 114         struct open_properties properties;
 115         int fd;
 116         /* inode number of the file at the time of the handle's creation. */
 117         uint64_t ino;
 118         bool in_use;
 119         /* Offset to which the file should be restored. */
 120         off_t offset;
 121         struct cds_list_head handles_list_node;
 122 };
 123
 124 struct unsuspendable_fd {
 125         /*
 126          * Accesses are only performed through the tracker, which is protected
 127          * by its own lock.
 128          */
 129         int fd;
 130         char *name;
 131         struct cds_lfht_node tracker_node;
 132         struct rcu_head rcu_head;
 133 };
 134
 135 static struct {
 136         pthread_mutex_t lock;
 137         bool initialized;
 138         unsigned long value;
 139 } seed = {
 140         .lock = PTHREAD_MUTEX_INITIALIZER,
 141 };
 142
 143 static int match_fd(struct cds_lfht_node *node, const void *key);
 144 static void unsuspendable_fd_destroy(struct unsuspendable_fd *entry);
 145 static struct unsuspendable_fd *unsuspendable_fd_create(const char *name,
 146                 int fd);
 147 static int open_from_properties(struct open_properties *properties);
 148
 149 static void fs_handle_log(struct fs_handle *handle);
 150 static int fs_handle_suspend(struct fs_handle *handle);
 151 static int fs_handle_restore(struct fs_handle *handle);
 152
 153 static void fd_tracker_track(struct fd_tracker *tracker,
 154                 struct fs_handle *handle);
 155 static void fd_tracker_untrack(struct fd_tracker *tracker,
 156                 struct fs_handle *handle);
 157 static int fd_tracker_suspend_handles(struct fd_tracker *tracker,
 158                 unsigned int count);
 159 static int fd_tracker_restore_handle(struct fd_tracker *tracker,
 160                 struct fs_handle *handle);
 161
 162 /* Match function of the tracker's unsuspendable_fds hash table. */
 163 static
 164 int match_fd(struct cds_lfht_node *node, const void *key)
 165 {
 166         struct unsuspendable_fd *entry =
 167                 caa_container_of(node, struct unsuspendable_fd, tracker_node);
 168
 169         return hash_match_key_ulong((void *) (unsigned long) entry->fd,
 170                         (void *) key);
 171 }
 172
 173 static
 174 void delete_unsuspendable_fd(struct rcu_head *head)
 175 {
 176         struct unsuspendable_fd *fd = caa_container_of(head,
 177                         struct unsuspendable_fd, rcu_head);
 178
 179         free(fd->name);
 180         free(fd);
 181 }
 182
 183 static
 184 void unsuspendable_fd_destroy(struct unsuspendable_fd *entry)
 185 {
 186         if (!entry) {
 187                 return;
 188         }
 189         call_rcu(&entry->rcu_head, delete_unsuspendable_fd);
 190 }
 191
 192 static
 193 struct unsuspendable_fd *unsuspendable_fd_create(const char *name, int fd)
 194 {
 195         struct unsuspendable_fd *entry =
 196                         zmalloc(sizeof(*entry));
 197
 198         if (!entry) {
 199                 goto error;
 200         }
 201         if (name) {
 202                 entry->name = strdup(name);
 203                 if (!entry->name) {
 204                         goto error;
 205                 }
 206         }
 207         cds_lfht_node_init(&entry->tracker_node);
 208         entry->fd = fd;
 209         return entry;
 210 error:
 211         unsuspendable_fd_destroy(entry);
 212         return NULL;
 213 }
 214
 215 static
 216 void fs_handle_log(struct fs_handle *handle)
 217 {
 218         pthread_mutex_lock(&handle->lock);
 219         if (handle->fd >= 0) {
 220                 DBG_NO_LOC("    %s [active, fd %d%s]",
 221                                 handle->properties.path,
 222                                 handle->fd,
 223                                 handle->in_use ? ", in use" : "");
 224         } else {
 225                 DBG_NO_LOC("    %s [suspended]", handle->properties.path);
 226         }
 227         pthread_mutex_unlock(&handle->lock);
 228 }
 229
 230 static
 231 int fs_handle_suspend(struct fs_handle *handle)
 232 {
 233         int ret = 0;
 234         struct stat fs_stat;
 235
 236         pthread_mutex_lock(&handle->lock);
 237         assert(handle->fd >= 0);
 238         if (handle->in_use) {
 239                 /* This handle can't be suspended as it is currently in use. */
 240                 ret = -EAGAIN;
 241                 goto end;
 242         }
 243
 244         ret = stat(handle->properties.path, &fs_stat);
 245         if (ret) {
 246                 PERROR("Filesystem handle to %s cannot be suspended as stat() failed",
 247                                 handle->properties.path);
 248                 ret = -errno;
 249                 goto end;
 250         }
 251
 252         if (fs_stat.st_ino != handle->ino) {
 253                 /* Don't suspend as the handle would not be restorable. */
 254                 WARN("Filesystem handle to %s cannot be suspended as its inode changed",
 255                                 handle->properties.path);
 256                 ret = -ENOENT;
 257                 goto end;
 258         }
 259
 260         handle->offset = lseek(handle->fd, 0, SEEK_CUR);
 261         if (handle->offset == -1) {
 262                 WARN("Filesystem handle to %s cannot be suspended as lseek() failed to sample its current position",
 263                                 handle->properties.path);
 264                 ret = -errno;
 265                 goto end;
 266         }
 267
 268         ret = close(handle->fd);
 269         if (ret) {
 270                 PERROR("Filesystem handle to %s cannot be suspended as close() failed",
 271                                 handle->properties.path);
 272                 ret = -errno;
 273                 goto end;
 274         }
 275         DBG("Suspended filesystem handle to %s (fd %i) at position %" PRId64,
 276                         handle->properties.path, handle->fd, handle->offset);
 277         handle->fd = -1;
 278 end:
 279         if (ret) {
 280                 handle->tracker->stats.errors++;
 281         }
 282         pthread_mutex_unlock(&handle->lock);
 283         return ret;
 284 }
 285
 286 /* Caller must hold the tracker and handle's locks. */
 287 static
 288 int fs_handle_restore(struct fs_handle *handle)
 289 {
 290         int ret, fd = -1;
 291
 292         assert(handle->fd == -1);
 293         ret = open_from_properties(&handle->properties);
 294         if (ret < 0) {
 295                 PERROR("Failed to restore filesystem handle to %s, open() failed",
 296                                 handle->properties.path);
 297                 ret = -errno;
 298                 goto end;
 299         }
 300         fd = ret;
 301
 302         ret = lseek(fd, handle->offset, SEEK_SET);
 303         if (ret < 0) {
 304                 PERROR("Failed to restore filesystem handle to %s, lseek() failed",
 305                                 handle->properties.path);
 306                 ret = -errno;
 307                 goto end;
 308         }
 309         DBG("Restored filesystem handle to %s (fd %i) at position %" PRId64,
 310                         handle->properties.path, fd, handle->offset);
 311         ret = 0;
 312         handle->fd = fd;
 313         fd = -1;
 314 end:
 315         if (fd >= 0) {
 316                 (void) close(fd);
 317         }
 318         return ret;
 319 }
 320
 321 static
 322 int open_from_properties(struct open_properties *properties)
 323 {
 324         int ret;
 325
 326         /*
 327          * open() ignores the 'flags' parameter unless the O_CREAT or O_TMPFILE
 328          * flags are set. O_TMPFILE would not make sense in the context of a
 329          * suspendable fs_handle as it would not be restorable (see OPEN(2)),
 330          * thus it is ignored here.
 331          */
 332         if ((properties->flags & O_CREAT) && properties->mode.is_set) {
 333                 ret = open(properties->path, properties->flags,
 334                                 properties->mode.value);
 335         } else {
 336                 ret = open(properties->path, properties->flags);
 337         }
 338         /*
 339          * Some flags should not be used beyond the initial open() of a
 340          * restorable file system handle. O_CREAT and O_TRUNC must
 341          * be cleared since it would be unexpected to re-use them
 342          * when the handle is retored:
 343          *  - O_CREAT should not be needed as the file has been created
 344          *    on the initial call to open(),
 345          *  - O_TRUNC would destroy the file's contents by truncating it
 346          *    to length 0.
 347          */
 348         properties->flags &= ~(O_CREAT | O_TRUNC);
 349         if (ret < 0) {
 350                 ret = -errno;
 351                 goto end;
 352         }
 353 end:
 354         return ret;
 355 }
 356
 357 struct fd_tracker *fd_tracker_create(unsigned int capacity)
 358 {
 359         struct fd_tracker *tracker = zmalloc(sizeof(struct fd_tracker));
 360
 361         if (!tracker) {
 362                 goto end;
 363         }
 364
 365         pthread_mutex_lock(&seed.lock);
 366         if (!seed.initialized) {
 367                 seed.value = (unsigned long) time(NULL);
 368                 seed.initialized = true;
 369         }
 370         pthread_mutex_unlock(&seed.lock);
 371
 372         CDS_INIT_LIST_HEAD(&tracker->active_handles);
 373         CDS_INIT_LIST_HEAD(&tracker->suspended_handles);
 374         tracker->capacity = capacity;
 375         tracker->unsuspendable_fds = cds_lfht_new(DEFAULT_HT_SIZE, 1, 0,
 376                         CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING, NULL);
 377         DBG("File descriptor tracker created with a limit of %u simultaneously-opened FDs",
 378                         capacity);
 379 end:
 380         return tracker;
 381 }
 382
 383 void fd_tracker_log(struct fd_tracker *tracker)
 384 {
 385         struct fs_handle *handle;
 386         struct unsuspendable_fd *unsuspendable_fd;
 387         struct cds_lfht_iter iter;
 388
 389         pthread_mutex_lock(&tracker->lock);
 390         DBG_NO_LOC("File descriptor tracker");
 391         DBG_NO_LOC("  Stats:");
 392         DBG_NO_LOC("    uses:            %" PRIu64, tracker->stats.uses);
 393         DBG_NO_LOC("    misses:          %" PRIu64, tracker->stats.misses);
 394         DBG_NO_LOC("    errors:          %" PRIu64, tracker->stats.errors);
 395         DBG_NO_LOC("  Tracked:           %u", TRACKED_COUNT(tracker));
 396         DBG_NO_LOC("    active:          %u", ACTIVE_COUNT(tracker));
 397         DBG_NO_LOC("      suspendable:   %u", SUSPENDABLE_COUNT(tracker));
 398         DBG_NO_LOC("      unsuspendable: %u", UNSUSPENDABLE_COUNT(tracker));
 399         DBG_NO_LOC("    suspended:       %u", SUSPENDED_COUNT(tracker));
 400         DBG_NO_LOC("    capacity:        %u", tracker->capacity);
 401
 402         DBG_NO_LOC("  Tracked suspendable file descriptors");
 403         cds_list_for_each_entry(handle, &tracker->active_handles,
 404                         handles_list_node) {
 405                 fs_handle_log(handle);
 406         }
 407         cds_list_for_each_entry(handle, &tracker->suspended_handles,
 408                         handles_list_node) {
 409                 fs_handle_log(handle);
 410         }
 411         if (!SUSPENDABLE_COUNT(tracker)) {
 412                 DBG_NO_LOC("    None");
 413         }
 414
 415         DBG_NO_LOC("  Tracked unsuspendable file descriptors");
 416         rcu_read_lock();
 417         cds_lfht_for_each_entry(tracker->unsuspendable_fds, &iter,
 418                         unsuspendable_fd, tracker_node) {
 419                 DBG_NO_LOC("    %s [active, fd %d]", unsuspendable_fd->name ? : "Unnamed",
 420                                 unsuspendable_fd->fd);
 421         }
 422         rcu_read_unlock();
 423         if (!UNSUSPENDABLE_COUNT(tracker)) {
 424                 DBG_NO_LOC("    None");
 425         }
 426
 427         pthread_mutex_unlock(&tracker->lock);
 428 }
 429
 430 int fd_tracker_destroy(struct fd_tracker *tracker)
 431 {
 432         int ret = 0;
 433
 434         /*
 435          * Refuse to destroy the tracker as fs_handles may still old
 436          * weak references to the tracker.
 437          */
 438         pthread_mutex_lock(&tracker->lock);
 439         if (TRACKED_COUNT(tracker)) {
 440                 ERR("A file descriptor leak has been detected: %u tracked file descriptors are still being tracked",
 441                                 TRACKED_COUNT(tracker));
 442                 pthread_mutex_unlock(&tracker->lock);
 443                 fd_tracker_log(tracker);
 444                 ret = -1;
 445                 goto end;
 446         }
 447         pthread_mutex_unlock(&tracker->lock);
 448
 449         ret = cds_lfht_destroy(tracker->unsuspendable_fds, NULL);
 450         assert(!ret);
 451         pthread_mutex_destroy(&tracker->lock);
 452         free(tracker);
 453 end:
 454         return ret;
 455 }
 456
 457 struct fs_handle *fd_tracker_open_fs_handle(struct fd_tracker *tracker,
 458                 const char *path, int flags, mode_t *mode)
 459 {
 460         int ret;
 461         struct fs_handle *handle = NULL;
 462         struct stat fd_stat;
 463         struct open_properties properties = {
 464                 .path = strdup(path),
 465                 .flags = flags,
 466                 .mode.is_set = !!mode,
 467                 .mode.value = mode ? *mode : 0,
 468         };
 469
 470         if (!properties.path) {
 471                 goto end;
 472         }
 473
 474         pthread_mutex_lock(&tracker->lock);
 475         if (ACTIVE_COUNT(tracker) == tracker->capacity) {
 476                 if (tracker->count.suspendable.active > 0) {
 477                         ret = fd_tracker_suspend_handles(tracker, 1);
 478                         if (ret) {
 479                                 goto error_destroy;
 480                         }
 481                 } else {
 482                         /*
 483                          * There are not enough active suspendable file
 484                          * descriptors to open a new fd and still accomodate the
 485                          * tracker's capacity.
 486                          */
 487                         WARN("Cannot open file system handle, too many unsuspendable file descriptors are opened (%u)",
 488                                         tracker->count.unsuspendable);
 489                         ret = -EMFILE;
 490                         goto error_destroy;
 491                 }
 492         }
 493
 494         handle = zmalloc(sizeof(*handle));
 495         if (!handle) {
 496                 goto end;
 497         }
 498
 499         ret = pthread_mutex_init(&handle->lock, NULL);
 500         if (ret) {
 501                 PERROR("Failed to initialize handle mutex while creating fs handle");
 502                 free(handle);
 503                 goto end;
 504         }
 505
 506         handle->fd = open_from_properties(&properties);
 507         if (handle->fd < 0) {
 508                 PERROR("Failed to open fs handle to %s, open() returned", path);
 509                 ret = -errno;
 510                 goto error_destroy;
 511         }
 512
 513         /*
 514          * Clear the create flag from the open flags as it would make no sense
 515          * to use it when restoring a fs handle.
 516          */
 517         properties.flags &= ~O_CREAT;
 518         handle->properties = properties;
 519         properties.path = NULL;
 520
 521         if (fstat(handle->fd, &fd_stat)) {
 522                 PERROR("Failed to retrieve file descriptor inode while creating fs handle, fstat() returned");
 523                 ret = -errno;
 524                 goto error_destroy;
 525         }
 526         handle->ino = fd_stat.st_ino;
 527
 528         fd_tracker_track(tracker, handle);
 529         handle->tracker = tracker;
 530         pthread_mutex_unlock(&tracker->lock);
 531 end:
 532         free(properties.path);
 533         return handle;
 534 error_destroy:
 535         pthread_mutex_unlock(&tracker->lock);
 536         (void) fs_handle_close(handle);
 537         handle = NULL;
 538         goto end;
 539 }
 540
 541 /* Caller must hold the tracker's lock. */
 542 static
 543 int fd_tracker_suspend_handles(struct fd_tracker *tracker,
 544                 unsigned int count)
 545 {
 546         unsigned int left_to_close = count;
 547         struct fs_handle *handle, *tmp;
 548
 549         cds_list_for_each_entry_safe(handle, tmp, &tracker->active_handles,
 550                         handles_list_node) {
 551                 int ret;
 552
 553                 fd_tracker_untrack(tracker, handle);
 554                 ret = fs_handle_suspend(handle);
 555                 fd_tracker_track(tracker, handle);
 556                 if (!ret) {
 557                         left_to_close--;
 558                 }
 559
 560                 if (!left_to_close) {
 561                         break;
 562                 }
 563         }
 564         return left_to_close ? -EMFILE : 0;
 565 }
 566
 567 int fd_tracker_open_unsuspendable_fd(struct fd_tracker *tracker,
 568                 int *out_fds, const char **names, unsigned int fd_count,
 569                 fd_open_cb open, void *user_data)
 570 {
 571         int ret, user_ret, i, fds_to_suspend;
 572         unsigned int active_fds;
 573         struct unsuspendable_fd *entries[fd_count];
 574
 575         memset(entries, 0, sizeof(entries));
 576
 577         pthread_mutex_lock(&tracker->lock);
 578
 579         active_fds = ACTIVE_COUNT(tracker);
 580         fds_to_suspend = (int) active_fds + (int) fd_count - (int) tracker->capacity;
 581         if (fds_to_suspend > 0) {
 582                 if (fds_to_suspend <= tracker->count.suspendable.active) {
 583                         ret = fd_tracker_suspend_handles(tracker, fds_to_suspend);
 584                         if (ret) {
 585                                 goto end;
 586                         }
 587                 } else {
 588                         /*
 589                          * There are not enough active suspendable file
 590                          * descriptors to open a new fd and still accomodate the
 591                          * tracker's capacity.
 592                          */
 593                         WARN("Cannot open unsuspendable fd, too many unsuspendable file descriptors are opened (%u)",
 594                                         tracker->count.unsuspendable);
 595                         ret = -EMFILE;
 596                         goto end;
 597                 }
 598         }
 599
 600         user_ret = open(user_data, out_fds);
 601         if (user_ret) {
 602                 ret = user_ret;
 603                 goto end;
 604         }
 605
 606         /*
 607          * Add the fds returned by the user's callback to the hashtable
 608          * of unsuspendable fds.
 609          */
 610         for (i = 0; i < fd_count; i++) {
 611                 struct unsuspendable_fd *entry =
 612                                 unsuspendable_fd_create(names ? names[i] : NULL,
 613                                                 out_fds[i]);
 614
 615                 if (!entry) {
 616                         ret = -1;
 617                         goto end_free_entries;
 618                 }
 619                 entries[i] = entry;
 620         }
 621
 622         rcu_read_lock();
 623         for (i = 0; i < fd_count; i++) {
 624                 struct cds_lfht_node *node;
 625                 struct unsuspendable_fd *entry = entries[i];
 626
 627                 node = cds_lfht_add_unique(
 628                                 tracker->unsuspendable_fds,
 629                                 hash_key_ulong((void *) (unsigned long) out_fds[i],
 630                                                 seed.value),
 631                                 match_fd,
 632                                 (void *) (unsigned long) out_fds[i],
 633                                 &entry->tracker_node);
 634
 635                 if (node != &entry->tracker_node) {
 636                         ret = -EEXIST;
 637                         rcu_read_unlock();
 638                         goto end_free_entries;
 639                 }
 640                 entries[i] = NULL;
 641         }
 642         tracker->count.unsuspendable += fd_count;
 643         rcu_read_unlock();
 644         ret = user_ret;
 645 end:
 646         pthread_mutex_unlock(&tracker->lock);
 647         return ret;
 648 end_free_entries:
 649         for (i = 0; i < fd_count; i++) {
 650                 unsuspendable_fd_destroy(entries[i]);
 651         }
 652         goto end;
 653 }
 654
 655 int fd_tracker_close_unsuspendable_fd(struct fd_tracker *tracker,
 656                 int *fds_in, unsigned int fd_count, fd_close_cb close,
 657                 void *user_data)
 658 {
 659         int i, ret, user_ret;
 660         int fds[fd_count];
 661
 662         /*
 663          * Maintain a local copy of fds_in as the user's callback may modify its
 664          * contents (e.g. setting the fd(s) to -1 after close).
 665          */
 666         memcpy(fds, fds_in, sizeof(*fds) * fd_count);
 667
 668         pthread_mutex_lock(&tracker->lock);
 669         rcu_read_lock();
 670
 671         /* Let the user close the file descriptors. */
 672         user_ret = close(user_data, fds_in);
 673         if (user_ret) {
 674                 ret = user_ret;
 675                 goto end;
 676         }
 677
 678         /* Untrack the fds that were just closed by the user's callback. */
 679         for (i = 0; i < fd_count; i++) {
 680                 struct cds_lfht_node *node;
 681                 struct cds_lfht_iter iter;
 682                 struct unsuspendable_fd *entry;
 683
 684                 cds_lfht_lookup(tracker->unsuspendable_fds,
 685                                 hash_key_ulong((void *) (unsigned long) fds[i],
 686                                                 seed.value),
 687                                 match_fd,
 688                                 (void *) (unsigned long) fds[i],
 689                                 &iter);
 690                 node = cds_lfht_iter_get_node(&iter);
 691                 if (!node) {
 692                         /* Unknown file descriptor. */
 693                         WARN("Untracked file descriptor %d passed to fd_tracker_close_unsuspendable_fd()",
 694                                         fds[i]);
 695                         ret = -EINVAL;
 696                         goto end;
 697                 }
 698                 entry = caa_container_of(node,
 699                                 struct unsuspendable_fd,
 700                                 tracker_node);
 701
 702                 cds_lfht_del(tracker->unsuspendable_fds, node);
 703                 unsuspendable_fd_destroy(entry);
 704                 fds[i] = -1;
 705         }
 706
 707         tracker->count.unsuspendable -= fd_count;
 708         ret = 0;
 709 end:
 710         rcu_read_unlock();
 711         pthread_mutex_unlock(&tracker->lock);
 712         return ret;
 713 }
 714
 715 /* Caller must have taken the tracker's and handle's locks. */
 716 static
 717 void fd_tracker_track(struct fd_tracker *tracker, struct fs_handle *handle)
 718 {
 719         if (handle->fd >= 0) {
 720                 tracker->count.suspendable.active++;
 721                 cds_list_add_tail(&handle->handles_list_node,
 722                                 &tracker->active_handles);
 723         } else {
 724                 tracker->count.suspendable.suspended++;
 725                 cds_list_add_tail(&handle->handles_list_node,
 726                                 &tracker->suspended_handles);
 727         }
 728 }
 729
 730 /* Caller must have taken the tracker's and handle's locks. */
 731 static
 732 void fd_tracker_untrack(struct fd_tracker *tracker, struct fs_handle *handle)
 733 {
 734         if (handle->fd >= 0) {
 735                 tracker->count.suspendable.active--;
 736         } else {
 737                 tracker->count.suspendable.suspended--;
 738         }
 739         cds_list_del(&handle->handles_list_node);
 740 }
 741
 742 /* Caller must have taken the tracker's and handle's locks. */
 743 static
 744 int fd_tracker_restore_handle(struct fd_tracker *tracker,
 745                 struct fs_handle *handle)
 746 {
 747         int ret;
 748
 749         fd_tracker_untrack(tracker, handle);
 750         if (ACTIVE_COUNT(tracker) >= tracker->capacity) {
 751                 ret = fd_tracker_suspend_handles(tracker, 1);
 752                 if (ret) {
 753                         goto end;
 754                 }
 755         }
 756         ret = fs_handle_restore(handle);
 757 end:
 758         fd_tracker_track(tracker, handle);
 759         return ret ? ret : handle->fd;
 760 }
 761
 762 int fs_handle_get_fd(struct fs_handle *handle)
 763 {
 764         int ret;
 765
 766         /*
 767          * TODO This should be optimized as it is a fairly hot path.
 768          * The fd-tracker's lock should only be taken when a fs_handle is
 769          * restored (slow path). On the fast path (fs_handle is active),
 770          * the only effect on the fd_tracker is marking the handle as the
 771          * most recently used. Currently, it is done by a call to the
 772          * track/untrack helpers, but it should be done atomically.
 773          *
 774          * Note that the lock's nesting order must still be respected here.
 775          * The handle's lock nests inside the tracker's lock.
 776          */
 777         pthread_mutex_lock(&handle->tracker->lock);
 778         pthread_mutex_lock(&handle->lock);
 779         assert(!handle->in_use);
 780
 781         handle->tracker->stats.uses++;
 782         if (handle->fd >= 0) {
 783                 ret = handle->fd;
 784                 /* Mark as most recently used. */
 785                 fd_tracker_untrack(handle->tracker, handle);
 786                 fd_tracker_track(handle->tracker, handle);
 787         } else {
 788                 handle->tracker->stats.misses++;
 789                 ret = fd_tracker_restore_handle(handle->tracker, handle);
 790                 if (ret < 0) {
 791                         handle->tracker->stats.errors++;
 792                         goto end;
 793                 }
 794         }
 795         handle->in_use = true;
 796 end:
 797         pthread_mutex_unlock(&handle->lock);
 798         pthread_mutex_unlock(&handle->tracker->lock);
 799         return ret;
 800 }
 801
 802 void fs_handle_put_fd(struct fs_handle *handle)
 803 {
 804         pthread_mutex_lock(&handle->lock);
 805         handle->in_use = false;
 806         pthread_mutex_unlock(&handle->lock);
 807 }
 808
 809 int fs_handle_unlink(struct fs_handle *handle)
 810 {
 811         int ret;
 812
 813         pthread_mutex_lock(&handle->tracker->lock);
 814         pthread_mutex_lock(&handle->lock);
 815         ret = lttng_inode_defer_unlink(handle->inode);
 816         pthread_mutex_unlock(&handle->lock);
 817         pthread_mutex_unlock(&handle->tracker->lock);
 818         return ret;
 819 }
 820
 821 int fs_handle_close(struct fs_handle *handle)
 822 {
 823         int ret = 0;
 824
 825         if (!handle) {
 826                 ret = -EINVAL;
 827                 goto end;
 828         }
 829
 830         pthread_mutex_lock(&handle->tracker->lock);
 831         pthread_mutex_lock(&handle->lock);
 832         fd_tracker_untrack(handle->tracker, handle);
 833         if (handle->fd >= 0) {
 834                 assert(!handle->in_use);
 835                 /*
 836                  * The return value of close() is not propagated as there
 837                  * isn't much the user can do about it.
 838                  */
 839                 if (close(handle->fd)) {
 840                         PERROR("Failed to close the file descritptor (%d) of fs handle to %s, close() returned",
 841                                         handle->fd, handle->properties.path);
 842                 }
 843                 handle->fd = -1;
 844         }
 845         pthread_mutex_unlock(&handle->lock);
 846         pthread_mutex_destroy(&handle->lock);
 847         pthread_mutex_unlock(&handle->tracker->lock);
 848         free(handle->properties.path);
 849         free(handle);
 850 end:
 851         return ret;
 852 }