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