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_get_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_get_location(handle->inode, &node_directory_handle, &path);
 234         assert(handle->fd >= 0);
 235         if (handle->in_use) {
 236                 /* This handle can't be suspended as it is currently in use. */
 237                 ret = -EAGAIN;
 238                 goto end;
 239         }
 240
 241         ret = lttng_directory_handle_stat(
 242                         node_directory_handle, path, &fs_stat);
 243         if (ret) {
 244                 PERROR("Filesystem handle to %s cannot be suspended as stat() failed",
 245                                 path);
 246                 ret = -errno;
 247                 goto end;
 248         }
 249
 250         if (fs_stat.st_ino != handle->ino) {
 251                 /* Don't suspend as the handle would not be restorable. */
 252                 WARN("Filesystem handle to %s cannot be suspended as its inode changed",
 253                                 path);
 254                 ret = -ENOENT;
 255                 goto end;
 256         }
 257
 258         handle->offset = lseek(handle->fd, 0, SEEK_CUR);
 259         if (handle->offset == -1) {
 260                 WARN("Filesystem handle to %s cannot be suspended as lseek() failed to sample its current position",
 261                                 path);
 262                 ret = -errno;
 263                 goto end;
 264         }
 265
 266         ret = close(handle->fd);
 267         if (ret) {
 268                 PERROR("Filesystem handle to %s cannot be suspended as close() failed",
 269                                 path);
 270                 ret = -errno;
 271                 goto end;
 272         }
 273         DBG("Suspended filesystem handle to %s (fd %i) at position %" PRId64,
 274                         path, handle->fd, handle->offset);
 275         handle->fd = -1;
 276 end:
 277         if (ret) {
 278                 handle->tracker->stats.errors++;
 279         }
 280         pthread_mutex_unlock(&handle->lock);
 281         return ret;
 282 }
 283
 284 /* Caller must hold the tracker and handle's locks. */
 285 static int fs_handle_tracked_restore(struct fs_handle_tracked *handle)
 286 {
 287         int ret, fd = -1;
 288         const char *path;
 289         const struct lttng_directory_handle *node_directory_handle;
 290
 291         lttng_inode_get_location(handle->inode, &node_directory_handle, &path);
 292
 293         assert(handle->fd == -1);
 294         assert(path);
 295         ret = open_from_properties(
 296                         node_directory_handle, path, &handle->properties);
 297         if (ret < 0) {
 298                 PERROR("Failed to restore filesystem handle to %s, open() failed",
 299                                 path);
 300                 ret = -errno;
 301                 goto end;
 302         }
 303         fd = ret;
 304
 305         ret = lseek(fd, handle->offset, SEEK_SET);
 306         if (ret < 0) {
 307                 PERROR("Failed to restore filesystem handle to %s, lseek() failed",
 308                                 path);
 309                 ret = -errno;
 310                 goto end;
 311         }
 312         DBG("Restored filesystem handle to %s (fd %i) at position %" PRId64,
 313                         path, fd, handle->offset);
 314         ret = 0;
 315         handle->fd = fd;
 316         fd = -1;
 317 end:
 318         if (fd >= 0) {
 319                 (void) close(fd);
 320         }
 321         return ret;
 322 }
 323
 324 static int open_from_properties(const struct lttng_directory_handle *dir_handle,
 325                 const char *path, struct open_properties *properties)
 326 {
 327         int ret;
 328
 329         /*
 330          * open() ignores the 'flags' parameter unless the O_CREAT or O_TMPFILE
 331          * flags are set. O_TMPFILE would not make sense in the context of a
 332          * suspendable fs_handle as it would not be restorable (see OPEN(2)),
 333          * thus it is ignored here.
 334          */
 335         if ((properties->flags & O_CREAT) && properties->mode.is_set) {
 336                 ret = lttng_directory_handle_open_file(dir_handle, path,
 337                                 properties->flags, properties->mode.value);
 338         } else {
 339                 ret = lttng_directory_handle_open_file(dir_handle, path,
 340                                 properties->flags, 0);
 341         }
 342         /*
 343          * Some flags should not be used beyond the initial open() of a
 344          * restorable file system handle. O_CREAT and O_TRUNC must
 345          * be cleared since it would be unexpected to re-use them
 346          * when the handle is retored:
 347          *  - O_CREAT should not be needed as the file has been created
 348          *    on the initial call to open(),
 349          *  - O_TRUNC would destroy the file's contents by truncating it
 350          *    to length 0.
 351          */
 352         properties->flags &= ~(O_CREAT | O_TRUNC);
 353         if (ret < 0) {
 354                 ret = -errno;
 355                 goto end;
 356         }
 357 end:
 358         return ret;
 359 }
 360
 361 struct fd_tracker *fd_tracker_create(const char *unlinked_file_path,
 362                 unsigned int capacity)
 363 {
 364         struct fd_tracker *tracker = zmalloc(sizeof(struct fd_tracker));
 365
 366         if (!tracker) {
 367                 goto end;
 368         }
 369
 370         pthread_mutex_lock(&seed.lock);
 371         if (!seed.initialized) {
 372                 seed.value = (unsigned long) time(NULL);
 373                 seed.initialized = true;
 374         }
 375         pthread_mutex_unlock(&seed.lock);
 376
 377         CDS_INIT_LIST_HEAD(&tracker->active_handles);
 378         CDS_INIT_LIST_HEAD(&tracker->suspended_handles);
 379         tracker->capacity = capacity;
 380         tracker->unsuspendable_fds = cds_lfht_new(DEFAULT_HT_SIZE, 1, 0,
 381                         CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING, NULL);
 382         if (!tracker->unsuspendable_fds) {
 383                 ERR("Failed to create fd-tracker's unsuspendable_fds hash table");
 384                 goto error;
 385         }
 386         tracker->inode_registry = lttng_inode_registry_create();
 387         if (!tracker->inode_registry) {
 388                 ERR("Failed to create fd-tracker's inode registry");
 389                 goto error;
 390         }
 391
 392         tracker->unlinked_file_pool =
 393                         lttng_unlinked_file_pool_create(unlinked_file_path);
 394         if (!tracker->unlinked_file_pool) {
 395                 goto error;
 396         }
 397         DBG("File descriptor tracker created with a limit of %u simultaneously-opened FDs",
 398                         capacity);
 399 end:
 400         return tracker;
 401 error:
 402         fd_tracker_destroy(tracker);
 403         return NULL;
 404 }
 405
 406 void fd_tracker_log(struct fd_tracker *tracker)
 407 {
 408         struct fs_handle_tracked *handle;
 409         struct unsuspendable_fd *unsuspendable_fd;
 410         struct cds_lfht_iter iter;
 411
 412         pthread_mutex_lock(&tracker->lock);
 413         DBG_NO_LOC("File descriptor tracker");
 414         DBG_NO_LOC("  Stats:");
 415         DBG_NO_LOC("    uses:            %" PRIu64, tracker->stats.uses);
 416         DBG_NO_LOC("    misses:          %" PRIu64, tracker->stats.misses);
 417         DBG_NO_LOC("    errors:          %" PRIu64, tracker->stats.errors);
 418         DBG_NO_LOC("  Tracked:           %u", TRACKED_COUNT(tracker));
 419         DBG_NO_LOC("    active:          %u", ACTIVE_COUNT(tracker));
 420         DBG_NO_LOC("      suspendable:   %u", SUSPENDABLE_COUNT(tracker));
 421         DBG_NO_LOC("      unsuspendable: %u", UNSUSPENDABLE_COUNT(tracker));
 422         DBG_NO_LOC("    suspended:       %u", SUSPENDED_COUNT(tracker));
 423         DBG_NO_LOC("    capacity:        %u", tracker->capacity);
 424
 425         DBG_NO_LOC("  Tracked suspendable file descriptors");
 426         cds_list_for_each_entry(
 427                         handle, &tracker->active_handles, handles_list_node) {
 428                 fs_handle_tracked_log(handle);
 429         }
 430         cds_list_for_each_entry(handle, &tracker->suspended_handles,
 431                         handles_list_node) {
 432                 fs_handle_tracked_log(handle);
 433         }
 434         if (!SUSPENDABLE_COUNT(tracker)) {
 435                 DBG_NO_LOC("    None");
 436         }
 437
 438         DBG_NO_LOC("  Tracked unsuspendable file descriptors");
 439         rcu_read_lock();
 440         cds_lfht_for_each_entry(tracker->unsuspendable_fds, &iter,
 441                         unsuspendable_fd, tracker_node) {
 442                 DBG_NO_LOC("    %s [active, fd %d]",
 443                                 unsuspendable_fd->name ?: "Unnamed",
 444                                 unsuspendable_fd->fd);
 445         }
 446         rcu_read_unlock();
 447         if (!UNSUSPENDABLE_COUNT(tracker)) {
 448                 DBG_NO_LOC("    None");
 449         }
 450
 451         pthread_mutex_unlock(&tracker->lock);
 452 }
 453
 454 int fd_tracker_destroy(struct fd_tracker *tracker)
 455 {
 456         int ret = 0;
 457
 458         /*
 459          * Refuse to destroy the tracker as fs_handles may still old
 460          * weak references to the tracker.
 461          */
 462         pthread_mutex_lock(&tracker->lock);
 463         if (TRACKED_COUNT(tracker)) {
 464                 ERR("A file descriptor leak has been detected: %u tracked file descriptors are still being tracked",
 465                                 TRACKED_COUNT(tracker));
 466                 pthread_mutex_unlock(&tracker->lock);
 467                 fd_tracker_log(tracker);
 468                 ret = -1;
 469                 goto end;
 470         }
 471         pthread_mutex_unlock(&tracker->lock);
 472
 473         if (tracker->unsuspendable_fds) {
 474                 ret = cds_lfht_destroy(tracker->unsuspendable_fds, NULL);
 475                 assert(!ret);
 476         }
 477
 478         lttng_inode_registry_destroy(tracker->inode_registry);
 479         lttng_unlinked_file_pool_destroy(tracker->unlinked_file_pool);
 480         pthread_mutex_destroy(&tracker->lock);
 481         free(tracker);
 482 end:
 483         return ret;
 484 }
 485
 486 struct fs_handle *fd_tracker_open_fs_handle(struct fd_tracker *tracker,
 487                 struct lttng_directory_handle *directory,
 488                 const char *path,
 489                 int flags,
 490                 mode_t *mode)
 491 {
 492         int ret;
 493         struct fs_handle_tracked *handle = NULL;
 494         struct stat fd_stat;
 495         struct open_properties properties = {
 496                 .flags = flags,
 497                 .mode.is_set = !!mode,
 498                 .mode.value = mode ? *mode : 0,
 499         };
 500
 501         pthread_mutex_lock(&tracker->lock);
 502         if (ACTIVE_COUNT(tracker) == tracker->capacity) {
 503                 if (tracker->count.suspendable.active > 0) {
 504                         ret = fd_tracker_suspend_handles(tracker, 1);
 505                         if (ret) {
 506                                 goto end;
 507                         }
 508                 } else {
 509                         /*
 510                          * There are not enough active suspendable file
 511                          * descriptors to open a new fd and still accommodate
 512                          * the tracker's capacity.
 513                          */
 514                         WARN("Cannot open file system handle, too many unsuspendable file descriptors are opened (%u)",
 515                                         tracker->count.unsuspendable);
 516                         goto end;
 517                 }
 518         }
 519
 520         handle = zmalloc(sizeof(*handle));
 521         if (!handle) {
 522                 goto end;
 523         }
 524         handle->parent = (typeof(handle->parent)) {
 525                 .get_fd = fs_handle_tracked_get_fd,
 526                 .put_fd = fs_handle_tracked_put_fd,
 527                 .unlink = fs_handle_tracked_unlink,
 528                 .close = fs_handle_tracked_close,
 529         };
 530
 531         handle->tracker = tracker;
 532
 533         ret = pthread_mutex_init(&handle->lock, NULL);
 534         if (ret) {
 535                 PERROR("Failed to initialize handle mutex while creating fs handle");
 536                 goto error_mutex_init;
 537         }
 538
 539         handle->fd = open_from_properties(directory, path, &properties);
 540         if (handle->fd < 0) {
 541                 PERROR("Failed to open fs handle to %s, open() returned", path);
 542                 goto error;
 543         }
 544
 545         handle->properties = properties;
 546
 547         handle->inode = lttng_inode_registry_get_inode(tracker->inode_registry,
 548                         directory, path, handle->fd,
 549                         tracker->unlinked_file_pool);
 550         if (!handle->inode) {
 551                 ERR("Failed to get lttng_inode corresponding to file %s", path);
 552                 goto error;
 553         }
 554
 555         if (fstat(handle->fd, &fd_stat)) {
 556                 PERROR("Failed to retrieve file descriptor inode while creating fs handle, fstat() returned");
 557                 goto error;
 558         }
 559         handle->ino = fd_stat.st_ino;
 560
 561         fd_tracker_track(tracker, handle);
 562 end:
 563         pthread_mutex_unlock(&tracker->lock);
 564         return handle ? &handle->parent : NULL;
 565 error:
 566         if (handle->inode) {
 567                 lttng_inode_put(handle->inode);
 568         }
 569         pthread_mutex_destroy(&handle->lock);
 570 error_mutex_init:
 571         free(handle);
 572         handle = NULL;
 573         goto end;
 574 }
 575
 576 /* Caller must hold the tracker's lock. */
 577 static int fd_tracker_suspend_handles(
 578                 struct fd_tracker *tracker, unsigned int count)
 579 {
 580         unsigned int left_to_close = count;
 581         unsigned int attempts_left = tracker->count.suspendable.active;
 582         struct fs_handle_tracked *handle, *tmp;
 583
 584         cds_list_for_each_entry_safe(handle, tmp, &tracker->active_handles,
 585                         handles_list_node) {
 586                 int ret;
 587
 588                 fd_tracker_untrack(tracker, handle);
 589                 ret = fs_handle_tracked_suspend(handle);
 590                 fd_tracker_track(tracker, handle);
 591                 if (!ret) {
 592                         left_to_close--;
 593                 }
 594                 attempts_left--;
 595
 596                 if (left_to_close == 0 || attempts_left == 0) {
 597                         break;
 598                 }
 599         }
 600         return left_to_close ? -EMFILE : 0;
 601 }
 602
 603 int fd_tracker_open_unsuspendable_fd(struct fd_tracker *tracker,
 604                 int *out_fds,
 605                 const char **names,
 606                 unsigned int fd_count,
 607                 fd_open_cb open,
 608                 void *user_data)
 609 {
 610         int ret, user_ret, i, fds_to_suspend;
 611         unsigned int active_fds;
 612         struct unsuspendable_fd **entries;
 613
 614         entries = zmalloc(fd_count * sizeof(*entries));
 615         if (!entries) {
 616                 ret = -1;
 617                 goto end;
 618         }
 619
 620         pthread_mutex_lock(&tracker->lock);
 621
 622         active_fds = ACTIVE_COUNT(tracker);
 623         fds_to_suspend = (int) active_fds + (int) fd_count -
 624                         (int) tracker->capacity;
 625         if (fds_to_suspend > 0) {
 626                 if (fds_to_suspend <= tracker->count.suspendable.active) {
 627                         ret = fd_tracker_suspend_handles(
 628                                         tracker, fds_to_suspend);
 629                         if (ret) {
 630                                 goto end_unlock;
 631                         }
 632                 } else {
 633                         /*
 634                          * There are not enough active suspendable file
 635                          * descriptors to open a new fd and still accomodate the
 636                          * tracker's capacity.
 637                          */
 638                         WARN("Cannot open unsuspendable fd, too many unsuspendable file descriptors are opened (%u)",
 639                                         tracker->count.unsuspendable);
 640                         ret = -EMFILE;
 641                         goto end_unlock;
 642                 }
 643         }
 644
 645         user_ret = open(user_data, out_fds);
 646         if (user_ret) {
 647                 ret = user_ret;
 648                 goto end_unlock;
 649         }
 650
 651         /*
 652          * Add the fds returned by the user's callback to the hashtable
 653          * of unsuspendable fds.
 654          */
 655         for (i = 0; i < fd_count; i++) {
 656                 struct unsuspendable_fd *entry = unsuspendable_fd_create(
 657                                 names ? names[i] : NULL, out_fds[i]);
 658
 659                 if (!entry) {
 660                         ret = -1;
 661                         goto end_free_entries;
 662                 }
 663                 entries[i] = entry;
 664         }
 665
 666         rcu_read_lock();
 667         for (i = 0; i < fd_count; i++) {
 668                 struct cds_lfht_node *node;
 669                 struct unsuspendable_fd *entry = entries[i];
 670
 671                 node = cds_lfht_add_unique(tracker->unsuspendable_fds,
 672                                 hash_key_ulong((void *) (unsigned long)
 673                                                                 out_fds[i],
 674                                                 seed.value),
 675                                 match_fd, (void *) (unsigned long) out_fds[i],
 676                                 &entry->tracker_node);
 677
 678                 if (node != &entry->tracker_node) {
 679                         ret = -EEXIST;
 680                         rcu_read_unlock();
 681                         goto end_free_entries;
 682                 }
 683                 entries[i] = NULL;
 684         }
 685         tracker->count.unsuspendable += fd_count;
 686         rcu_read_unlock();
 687         ret = user_ret;
 688 end_unlock:
 689         pthread_mutex_unlock(&tracker->lock);
 690 end:
 691         free(entries);
 692         return ret;
 693 end_free_entries:
 694         for (i = 0; i < fd_count; i++) {
 695                 unsuspendable_fd_destroy(entries[i]);
 696         }
 697         goto end_unlock;
 698 }
 699
 700 int fd_tracker_close_unsuspendable_fd(struct fd_tracker *tracker,
 701                 int *fds_in,
 702                 unsigned int fd_count,
 703                 fd_close_cb close,
 704                 void *user_data)
 705 {
 706         int i, ret, user_ret;
 707         int *fds = NULL;
 708
 709         /*
 710          * Maintain a local copy of fds_in as the user's callback may modify its
 711          * contents (e.g. setting the fd(s) to -1 after close).
 712          */
 713         fds = malloc(sizeof(*fds) * fd_count);
 714         if (!fds) {
 715                 ret = -1;
 716                 goto end;
 717         }
 718         memcpy(fds, fds_in, sizeof(*fds) * fd_count);
 719
 720         pthread_mutex_lock(&tracker->lock);
 721         rcu_read_lock();
 722
 723         /* Let the user close the file descriptors. */
 724         user_ret = close(user_data, fds_in);
 725         if (user_ret) {
 726                 ret = user_ret;
 727                 goto end_unlock;
 728         }
 729
 730         /* Untrack the fds that were just closed by the user's callback. */
 731         for (i = 0; i < fd_count; i++) {
 732                 struct cds_lfht_node *node;
 733                 struct cds_lfht_iter iter;
 734                 struct unsuspendable_fd *entry;
 735
 736                 cds_lfht_lookup(tracker->unsuspendable_fds,
 737                                 hash_key_ulong((void *) (unsigned long) fds[i],
 738                                                 seed.value),
 739                                 match_fd, (void *) (unsigned long) fds[i],
 740                                 &iter);
 741                 node = cds_lfht_iter_get_node(&iter);
 742                 if (!node) {
 743                         /* Unknown file descriptor. */
 744                         WARN("Untracked file descriptor %d passed to fd_tracker_close_unsuspendable_fd()",
 745                                         fds[i]);
 746                         ret = -EINVAL;
 747                         goto end_unlock;
 748                 }
 749                 entry = caa_container_of(
 750                                 node, struct unsuspendable_fd, tracker_node);
 751
 752                 cds_lfht_del(tracker->unsuspendable_fds, node);
 753                 unsuspendable_fd_destroy(entry);
 754                 fds[i] = -1;
 755         }
 756
 757         tracker->count.unsuspendable -= fd_count;
 758         ret = 0;
 759 end_unlock:
 760         rcu_read_unlock();
 761         pthread_mutex_unlock(&tracker->lock);
 762         free(fds);
 763 end:
 764         return ret;
 765 }
 766
 767 /* Caller must have taken the tracker's and handle's locks. */
 768 static void fd_tracker_track(
 769                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 770 {
 771         if (handle->fd >= 0) {
 772                 tracker->count.suspendable.active++;
 773                 cds_list_add_tail(&handle->handles_list_node,
 774                                 &tracker->active_handles);
 775         } else {
 776                 tracker->count.suspendable.suspended++;
 777                 cds_list_add_tail(&handle->handles_list_node,
 778                                 &tracker->suspended_handles);
 779         }
 780 }
 781
 782 /* Caller must have taken the tracker's and handle's locks. */
 783 static void fd_tracker_untrack(
 784                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 785 {
 786         if (handle->fd >= 0) {
 787                 tracker->count.suspendable.active--;
 788         } else {
 789                 tracker->count.suspendable.suspended--;
 790         }
 791         cds_list_del(&handle->handles_list_node);
 792 }
 793
 794 /* Caller must have taken the tracker's and handle's locks. */
 795 static int fd_tracker_restore_handle(
 796                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 797 {
 798         int ret;
 799
 800         fd_tracker_untrack(tracker, handle);
 801         if (ACTIVE_COUNT(tracker) >= tracker->capacity) {
 802                 ret = fd_tracker_suspend_handles(tracker, 1);
 803                 if (ret) {
 804                         goto end;
 805                 }
 806         }
 807         ret = fs_handle_tracked_restore(handle);
 808 end:
 809         fd_tracker_track(tracker, handle);
 810         return ret ? ret : handle->fd;
 811 }
 812
 813 static int fs_handle_tracked_get_fd(struct fs_handle *_handle)
 814 {
 815         int ret;
 816         struct fs_handle_tracked *handle =
 817                         container_of(_handle, struct fs_handle_tracked, parent);
 818
 819         /*
 820          * TODO This should be optimized as it is a fairly hot path.
 821          * The fd-tracker's lock should only be taken when a fs_handle is
 822          * restored (slow path). On the fast path (fs_handle is active),
 823          * the only effect on the fd_tracker is marking the handle as the
 824          * most recently used. Currently, it is done by a call to the
 825          * track/untrack helpers, but it should be done atomically.
 826          *
 827          * Note that the lock's nesting order must still be respected here.
 828          * The handle's lock nests inside the tracker's lock.
 829          */
 830         pthread_mutex_lock(&handle->tracker->lock);
 831         pthread_mutex_lock(&handle->lock);
 832         assert(!handle->in_use);
 833
 834         handle->tracker->stats.uses++;
 835         if (handle->fd >= 0) {
 836                 ret = handle->fd;
 837                 /* Mark as most recently used. */
 838                 fd_tracker_untrack(handle->tracker, handle);
 839                 fd_tracker_track(handle->tracker, handle);
 840         } else {
 841                 handle->tracker->stats.misses++;
 842                 ret = fd_tracker_restore_handle(handle->tracker, handle);
 843                 if (ret < 0) {
 844                         handle->tracker->stats.errors++;
 845                         goto end;
 846                 }
 847         }
 848         handle->in_use = true;
 849 end:
 850         pthread_mutex_unlock(&handle->lock);
 851         pthread_mutex_unlock(&handle->tracker->lock);
 852         return ret;
 853 }
 854
 855 static void fs_handle_tracked_put_fd(struct fs_handle *_handle)
 856 {
 857         struct fs_handle_tracked *handle =
 858                         container_of(_handle, struct fs_handle_tracked, parent);
 859
 860         pthread_mutex_lock(&handle->lock);
 861         handle->in_use = false;
 862         pthread_mutex_unlock(&handle->lock);
 863 }
 864
 865 static int fs_handle_tracked_unlink(struct fs_handle *_handle)
 866 {
 867         int ret;
 868         struct fs_handle_tracked *handle =
 869                         container_of(_handle, struct fs_handle_tracked, parent);
 870
 871         pthread_mutex_lock(&handle->tracker->lock);
 872         pthread_mutex_lock(&handle->lock);
 873         ret = lttng_inode_unlink(handle->inode);
 874         pthread_mutex_unlock(&handle->lock);
 875         pthread_mutex_unlock(&handle->tracker->lock);
 876         return ret;
 877 }
 878
 879 static int fs_handle_tracked_close(struct fs_handle *_handle)
 880 {
 881         int ret = 0;
 882         const char *path = NULL;
 883         struct fs_handle_tracked *handle =
 884                         container_of(_handle, struct fs_handle_tracked, parent);
 885
 886         if (!handle) {
 887                 ret = -EINVAL;
 888                 goto end;
 889         }
 890
 891         pthread_mutex_lock(&handle->tracker->lock);
 892         pthread_mutex_lock(&handle->lock);
 893         if (handle->inode) {
 894                 lttng_inode_get_location(handle->inode, NULL, &path);
 895         }
 896         fd_tracker_untrack(handle->tracker, handle);
 897         if (handle->fd >= 0) {
 898                 /*
 899                  * The return value of close() is not propagated as there
 900                  * isn't much the user can do about it.
 901                  */
 902                 if (close(handle->fd)) {
 903                         PERROR("Failed to close the file descritptor (%d) of fs handle to %s, close() returned",
 904                                         handle->fd, path ? path : "Unknown");
 905                 }
 906                 handle->fd = -1;
 907         }
 908         if (handle->inode) {
 909                 lttng_inode_put(handle->inode);
 910         }
 911         pthread_mutex_unlock(&handle->lock);
 912         pthread_mutex_destroy(&handle->lock);
 913         pthread_mutex_unlock(&handle->tracker->lock);
 914         free(handle);
 915 end:
 916         return ret;
 917 }