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

   1 /*
   2  * Copyright (C) 2018-2020 Jérémie Galarneau <jeremie.galarneau@efficios.com>
   3  *
   4  * SPDX-License-Identifier: GPL-2.0-only
   5  *
   6  */
   7
   8 #include <urcu.h>
   9 #include <urcu/list.h>
  10 #include <urcu/rculfhash.h>
  11
  12 #include <fcntl.h>
  13 #include <inttypes.h>
  14 #include <pthread.h>
  15 #include <stdbool.h>
  16 #include <sys/stat.h>
  17 #include <sys/types.h>
  18
  19 #include <common/defaults.h>
  20 #include <common/error.h>
  21 #include <common/fs-handle-internal.h>
  22 #include <common/hashtable/hashtable.h>
  23 #include <common/hashtable/utils.h>
  24 #include <common/macros.h>
  25 #include <common/optional.h>
  26
  27 #include "fd-tracker.h"
  28 #include "inode.h"
  29
  30 /* Tracker lock must be taken by the user. */
  31 #define TRACKED_COUNT(tracker)                                 \
  32         (tracker->count.suspendable.active +                   \
  33                         tracker->count.suspendable.suspended + \
  34                         tracker->count.unsuspendable)
  35
  36 /* Tracker lock must be taken by the user. */
  37 #define ACTIVE_COUNT(tracker) \
  38         (tracker->count.suspendable.active + tracker->count.unsuspendable)
  39
  40 /* Tracker lock must be taken by the user. */
  41 #define SUSPENDED_COUNT(tracker) (tracker->count.suspendable.suspended)
  42
  43 /* Tracker lock must be taken by the user. */
  44 #define SUSPENDABLE_COUNT(tracker)           \
  45         (tracker->count.suspendable.active + \
  46                         tracker->count.suspendable.suspended)
  47
  48 /* Tracker lock must be taken by the user. */
  49 #define UNSUSPENDABLE_COUNT(tracker) (tracker->count.unsuspendable)
  50
  51 struct fd_tracker {
  52         pthread_mutex_t lock;
  53         struct {
  54                 struct {
  55                         unsigned int active;
  56                         unsigned int suspended;
  57                 } suspendable;
  58                 unsigned int unsuspendable;
  59         } count;
  60         unsigned int capacity;
  61         struct {
  62                 uint64_t uses;
  63                 uint64_t misses;
  64                 /* Failures to suspend or restore fs handles. */
  65                 uint64_t errors;
  66         } stats;
  67         /*
  68          * The head of the active_handles list is always the least recently
  69          * used active handle. When an handle is used, it is removed from the
  70          * list and added to the end. When a file has to be suspended, the
  71          * first element in the list is "popped", suspended, and added to the
  72          * list of suspended handles.
  73          */
  74         struct cds_list_head active_handles;
  75         struct cds_list_head suspended_handles;
  76         struct cds_lfht *unsuspendable_fds;
  77         struct lttng_inode_registry *inode_registry;
  78         /* Unlinked files are moved in this directory under a unique name. */
  79         struct lttng_directory_handle *unlink_directory_handle;
  80         struct lttng_unlinked_file_pool *unlinked_file_pool;
  81 };
  82
  83 struct open_properties {
  84         int flags;
  85         LTTNG_OPTIONAL(mode_t) mode;
  86 };
  87
  88 /*
  89  * A fs_handle_tracked is not ref-counted. Therefore, it is assumed that a
  90  * handle is never in-use while it is being reclaimed. It can be
  91  * shared by multiple threads, but external synchronization is required
  92  * to ensure it is not still being used when it is reclaimed (close method).
  93  * In this respect, it is not different from a regular file descriptor.
  94  *
  95  * The fs_handle lock always nests _within_ the tracker's lock.
  96  */
  97 struct fs_handle_tracked {
  98         struct fs_handle parent;
  99         pthread_mutex_t lock;
 100         /*
 101          * Weak reference to the tracker. All fs_handles are assumed to have
 102          * been closed at the moment of the destruction of the fd_tracker.
 103          */
 104         struct fd_tracker *tracker;
 105         struct open_properties properties;
 106         struct lttng_inode *inode;
 107         int fd;
 108         /* inode number of the file at the time of the handle's creation. */
 109         uint64_t ino;
 110         bool in_use;
 111         /* Offset to which the file should be restored. */
 112         off_t offset;
 113         struct cds_list_head handles_list_node;
 114 };
 115
 116 struct unsuspendable_fd {
 117         /*
 118          * Accesses are only performed through the tracker, which is protected
 119          * by its own lock.
 120          */
 121         int fd;
 122         char *name;
 123         struct cds_lfht_node tracker_node;
 124         struct rcu_head rcu_head;
 125 };
 126
 127 static struct {
 128         pthread_mutex_t lock;
 129         bool initialized;
 130         unsigned long value;
 131 } seed = {
 132         .lock = PTHREAD_MUTEX_INITIALIZER,
 133 };
 134
 135 static int match_fd(struct cds_lfht_node *node, const void *key);
 136 static void unsuspendable_fd_destroy(struct unsuspendable_fd *entry);
 137 static struct unsuspendable_fd *unsuspendable_fd_create(
 138                 const char *name, int fd);
 139 static int open_from_properties(const struct lttng_directory_handle *dir_handle,
 140                 const char *path, struct open_properties *properties);
 141
 142 static void fs_handle_tracked_log(struct fs_handle_tracked *handle);
 143 static int fs_handle_tracked_suspend(struct fs_handle_tracked *handle);
 144 static int fs_handle_tracked_restore(struct fs_handle_tracked *handle);
 145 static int fs_handle_tracked_get_fd(struct fs_handle *_handle);
 146 static void fs_handle_tracked_put_fd(struct fs_handle *_handle);
 147 static int fs_handle_tracked_unlink(struct fs_handle *_handle);
 148 static int fs_handle_tracked_close(struct fs_handle *_handle);
 149
 150 static void fd_tracker_track(
 151                 struct fd_tracker *tracker, struct fs_handle_tracked *handle);
 152 static void fd_tracker_untrack(
 153                 struct fd_tracker *tracker, struct fs_handle_tracked *handle);
 154 static int fd_tracker_suspend_handles(
 155                 struct fd_tracker *tracker, unsigned int count);
 156 static int fd_tracker_restore_handle(
 157                 struct fd_tracker *tracker, struct fs_handle_tracked *handle);
 158
 159 /* Match function of the tracker's unsuspendable_fds hash table. */
 160 static int match_fd(struct cds_lfht_node *node, const void *key)
 161 {
 162         struct unsuspendable_fd *entry = caa_container_of(
 163                         node, struct unsuspendable_fd, tracker_node);
 164
 165         return hash_match_key_ulong(
 166                         (void *) (unsigned long) entry->fd, (void *) key);
 167 }
 168
 169 static void delete_unsuspendable_fd(struct rcu_head *head)
 170 {
 171         struct unsuspendable_fd *fd = caa_container_of(
 172                         head, struct unsuspendable_fd, rcu_head);
 173
 174         free(fd->name);
 175         free(fd);
 176 }
 177
 178 static void unsuspendable_fd_destroy(struct unsuspendable_fd *entry)
 179 {
 180         if (!entry) {
 181                 return;
 182         }
 183         call_rcu(&entry->rcu_head, delete_unsuspendable_fd);
 184 }
 185
 186 static struct unsuspendable_fd *unsuspendable_fd_create(
 187                 const char *name, int fd)
 188 {
 189         struct unsuspendable_fd *entry = zmalloc(sizeof(*entry));
 190
 191         if (!entry) {
 192                 goto error;
 193         }
 194         if (name) {
 195                 entry->name = strdup(name);
 196                 if (!entry->name) {
 197                         goto error;
 198                 }
 199         }
 200         cds_lfht_node_init(&entry->tracker_node);
 201         entry->fd = fd;
 202         return entry;
 203 error:
 204         unsuspendable_fd_destroy(entry);
 205         return NULL;
 206 }
 207
 208 static void fs_handle_tracked_log(struct fs_handle_tracked *handle)
 209 {
 210         const char *path;
 211
 212         pthread_mutex_lock(&handle->lock);
 213         lttng_inode_borrow_location(handle->inode, NULL, &path);
 214
 215         if (handle->fd >= 0) {
 216                 DBG_NO_LOC("    %s [active, fd %d%s]", path, handle->fd,
 217                                 handle->in_use ? ", in use" : "");
 218         } else {
 219                 DBG_NO_LOC("    %s [suspended]", path);
 220         }
 221         pthread_mutex_unlock(&handle->lock);
 222 }
 223
 224 /* Tracker lock must be held by the caller. */
 225 static int fs_handle_tracked_suspend(struct fs_handle_tracked *handle)
 226 {
 227         int ret = 0;
 228         struct stat fs_stat;
 229         const char *path;
 230         const struct lttng_directory_handle *node_directory_handle;
 231
 232         pthread_mutex_lock(&handle->lock);
 233         lttng_inode_borrow_location(
 234                         handle->inode, &node_directory_handle, &path);
 235         assert(handle->fd >= 0);
 236         if (handle->in_use) {
 237                 /* This handle can't be suspended as it is currently in use. */
 238                 ret = -EAGAIN;
 239                 goto end;
 240         }
 241
 242         ret = lttng_directory_handle_stat(
 243                         node_directory_handle, path, &fs_stat);
 244         if (ret) {
 245                 PERROR("Filesystem handle to %s cannot be suspended as stat() failed",
 246                                 path);
 247                 ret = -errno;
 248                 goto end;
 249         }
 250
 251         if (fs_stat.st_ino != handle->ino) {
 252                 /* Don't suspend as the handle would not be restorable. */
 253                 WARN("Filesystem handle to %s cannot be suspended as its inode changed",
 254                                 path);
 255                 ret = -ENOENT;
 256                 goto end;
 257         }
 258
 259         handle->offset = lseek(handle->fd, 0, SEEK_CUR);
 260         if (handle->offset == -1) {
 261                 WARN("Filesystem handle to %s cannot be suspended as lseek() failed to sample its current position",
 262                                 path);
 263                 ret = -errno;
 264                 goto end;
 265         }
 266
 267         ret = close(handle->fd);
 268         if (ret) {
 269                 PERROR("Filesystem handle to %s cannot be suspended as close() failed",
 270                                 path);
 271                 ret = -errno;
 272                 goto end;
 273         }
 274         DBG("Suspended filesystem handle to %s (fd %i) at position %" PRId64,
 275                         path, handle->fd, handle->offset);
 276         handle->fd = -1;
 277 end:
 278         if (ret) {
 279                 handle->tracker->stats.errors++;
 280         }
 281         pthread_mutex_unlock(&handle->lock);
 282         return ret;
 283 }
 284
 285 /* Caller must hold the tracker and handle's locks. */
 286 static int fs_handle_tracked_restore(struct fs_handle_tracked *handle)
 287 {
 288         int ret, fd = -1;
 289         const char *path;
 290         const struct lttng_directory_handle *node_directory_handle;
 291
 292         lttng_inode_borrow_location(
 293                         handle->inode, &node_directory_handle, &path);
 294
 295         assert(handle->fd == -1);
 296         assert(path);
 297         ret = open_from_properties(
 298                         node_directory_handle, path, &handle->properties);
 299         if (ret < 0) {
 300                 PERROR("Failed to restore filesystem handle to %s, open() failed",
 301                                 path);
 302                 ret = -errno;
 303                 goto end;
 304         }
 305         fd = ret;
 306
 307         ret = lseek(fd, handle->offset, SEEK_SET);
 308         if (ret < 0) {
 309                 PERROR("Failed to restore filesystem handle to %s, lseek() failed",
 310                                 path);
 311                 ret = -errno;
 312                 goto end;
 313         }
 314         DBG("Restored filesystem handle to %s (fd %i) at position %" PRId64,
 315                         path, fd, handle->offset);
 316         ret = 0;
 317         handle->fd = fd;
 318         fd = -1;
 319 end:
 320         if (fd >= 0) {
 321                 (void) close(fd);
 322         }
 323         return ret;
 324 }
 325
 326 static int open_from_properties(const struct lttng_directory_handle *dir_handle,
 327                 const char *path, struct open_properties *properties)
 328 {
 329         int ret;
 330
 331         /*
 332          * open() ignores the 'flags' parameter unless the O_CREAT or O_TMPFILE
 333          * flags are set. O_TMPFILE would not make sense in the context of a
 334          * suspendable fs_handle as it would not be restorable (see OPEN(2)),
 335          * thus it is ignored here.
 336          */
 337         if ((properties->flags & O_CREAT) && properties->mode.is_set) {
 338                 ret = lttng_directory_handle_open_file(dir_handle, path,
 339                                 properties->flags, properties->mode.value);
 340         } else {
 341                 ret = lttng_directory_handle_open_file(dir_handle, path,
 342                                 properties->flags, 0);
 343         }
 344         /*
 345          * Some flags should not be used beyond the initial open() of a
 346          * restorable file system handle. O_CREAT and O_TRUNC must
 347          * be cleared since it would be unexpected to re-use them
 348          * when the handle is retored:
 349          *  - O_CREAT should not be needed as the file has been created
 350          *    on the initial call to open(),
 351          *  - O_TRUNC would destroy the file's contents by truncating it
 352          *    to length 0.
 353          */
 354         properties->flags &= ~(O_CREAT | O_TRUNC);
 355         if (ret < 0) {
 356                 ret = -errno;
 357                 goto end;
 358         }
 359 end:
 360         return ret;
 361 }
 362
 363 LTTNG_HIDDEN
 364 struct fd_tracker *fd_tracker_create(const char *unlinked_file_path,
 365                 unsigned int capacity)
 366 {
 367         struct fd_tracker *tracker = zmalloc(sizeof(struct fd_tracker));
 368
 369         if (!tracker) {
 370                 goto end;
 371         }
 372
 373         pthread_mutex_lock(&seed.lock);
 374         if (!seed.initialized) {
 375                 seed.value = (unsigned long) time(NULL);
 376                 seed.initialized = true;
 377         }
 378         pthread_mutex_unlock(&seed.lock);
 379
 380         CDS_INIT_LIST_HEAD(&tracker->active_handles);
 381         CDS_INIT_LIST_HEAD(&tracker->suspended_handles);
 382         tracker->capacity = capacity;
 383         tracker->unsuspendable_fds = cds_lfht_new(DEFAULT_HT_SIZE, 1, 0,
 384                         CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING, NULL);
 385         if (!tracker->unsuspendable_fds) {
 386                 ERR("Failed to create fd-tracker's unsuspendable_fds hash table");
 387                 goto error;
 388         }
 389         tracker->inode_registry = lttng_inode_registry_create();
 390         if (!tracker->inode_registry) {
 391                 ERR("Failed to create fd-tracker's inode registry");
 392                 goto error;
 393         }
 394
 395         tracker->unlinked_file_pool =
 396                         lttng_unlinked_file_pool_create(unlinked_file_path);
 397         if (!tracker->unlinked_file_pool) {
 398                 goto error;
 399         }
 400         DBG("File descriptor tracker created with a limit of %u simultaneously-opened FDs",
 401                         capacity);
 402 end:
 403         return tracker;
 404 error:
 405         fd_tracker_destroy(tracker);
 406         return NULL;
 407 }
 408
 409 LTTNG_HIDDEN
 410 void fd_tracker_log(struct fd_tracker *tracker)
 411 {
 412         struct fs_handle_tracked *handle;
 413         struct unsuspendable_fd *unsuspendable_fd;
 414         struct cds_lfht_iter iter;
 415
 416         pthread_mutex_lock(&tracker->lock);
 417         DBG_NO_LOC("File descriptor tracker");
 418         DBG_NO_LOC("  Stats:");
 419         DBG_NO_LOC("    uses:            %" PRIu64, tracker->stats.uses);
 420         DBG_NO_LOC("    misses:          %" PRIu64, tracker->stats.misses);
 421         DBG_NO_LOC("    errors:          %" PRIu64, tracker->stats.errors);
 422         DBG_NO_LOC("  Tracked:           %u", TRACKED_COUNT(tracker));
 423         DBG_NO_LOC("    active:          %u", ACTIVE_COUNT(tracker));
 424         DBG_NO_LOC("      suspendable:   %u", SUSPENDABLE_COUNT(tracker));
 425         DBG_NO_LOC("      unsuspendable: %u", UNSUSPENDABLE_COUNT(tracker));
 426         DBG_NO_LOC("    suspended:       %u", SUSPENDED_COUNT(tracker));
 427         DBG_NO_LOC("    capacity:        %u", tracker->capacity);
 428
 429         DBG_NO_LOC("  Tracked suspendable file descriptors");
 430         cds_list_for_each_entry(
 431                         handle, &tracker->active_handles, handles_list_node) {
 432                 fs_handle_tracked_log(handle);
 433         }
 434         cds_list_for_each_entry(handle, &tracker->suspended_handles,
 435                         handles_list_node) {
 436                 fs_handle_tracked_log(handle);
 437         }
 438         if (!SUSPENDABLE_COUNT(tracker)) {
 439                 DBG_NO_LOC("    None");
 440         }
 441
 442         DBG_NO_LOC("  Tracked unsuspendable file descriptors");
 443         rcu_read_lock();
 444         cds_lfht_for_each_entry(tracker->unsuspendable_fds, &iter,
 445                         unsuspendable_fd, tracker_node) {
 446                 DBG_NO_LOC("    %s [active, fd %d]",
 447                                 unsuspendable_fd->name ?: "Unnamed",
 448                                 unsuspendable_fd->fd);
 449         }
 450         rcu_read_unlock();
 451         if (!UNSUSPENDABLE_COUNT(tracker)) {
 452                 DBG_NO_LOC("    None");
 453         }
 454
 455         pthread_mutex_unlock(&tracker->lock);
 456 }
 457
 458 LTTNG_HIDDEN
 459 int fd_tracker_destroy(struct fd_tracker *tracker)
 460 {
 461         int ret = 0;
 462
 463         if (!tracker) {
 464                 goto end;
 465         }
 466         /*
 467          * Refuse to destroy the tracker as fs_handles may still old
 468          * weak references to the tracker.
 469          */
 470         pthread_mutex_lock(&tracker->lock);
 471         if (TRACKED_COUNT(tracker)) {
 472                 ERR("A file descriptor leak has been detected: %u tracked file descriptors are still being tracked",
 473                                 TRACKED_COUNT(tracker));
 474                 pthread_mutex_unlock(&tracker->lock);
 475                 fd_tracker_log(tracker);
 476                 ret = -1;
 477                 goto end;
 478         }
 479         pthread_mutex_unlock(&tracker->lock);
 480
 481         if (tracker->unsuspendable_fds) {
 482                 ret = cds_lfht_destroy(tracker->unsuspendable_fds, NULL);
 483                 assert(!ret);
 484         }
 485
 486         lttng_inode_registry_destroy(tracker->inode_registry);
 487         lttng_unlinked_file_pool_destroy(tracker->unlinked_file_pool);
 488         pthread_mutex_destroy(&tracker->lock);
 489         free(tracker);
 490 end:
 491         return ret;
 492 }
 493
 494 LTTNG_HIDDEN
 495 struct fs_handle *fd_tracker_open_fs_handle(struct fd_tracker *tracker,
 496                 struct lttng_directory_handle *directory,
 497                 const char *path,
 498                 int flags,
 499                 mode_t *mode)
 500 {
 501         int ret;
 502         struct fs_handle_tracked *handle = NULL;
 503         struct stat fd_stat;
 504         struct open_properties properties = {
 505                 .flags = flags,
 506                 .mode.is_set = !!mode,
 507                 .mode.value = mode ? *mode : 0,
 508         };
 509
 510         pthread_mutex_lock(&tracker->lock);
 511         if (ACTIVE_COUNT(tracker) == tracker->capacity) {
 512                 if (tracker->count.suspendable.active > 0) {
 513                         ret = fd_tracker_suspend_handles(tracker, 1);
 514                         if (ret) {
 515                                 goto end;
 516                         }
 517                 } else {
 518                         /*
 519                          * There are not enough active suspendable file
 520                          * descriptors to open a new fd and still accommodate
 521                          * the tracker's capacity.
 522                          */
 523                         WARN("Cannot open file system handle, too many unsuspendable file descriptors are opened (%u)",
 524                                         tracker->count.unsuspendable);
 525                         goto end;
 526                 }
 527         }
 528
 529         handle = zmalloc(sizeof(*handle));
 530         if (!handle) {
 531                 goto end;
 532         }
 533         handle->parent = (typeof(handle->parent)) {
 534                 .get_fd = fs_handle_tracked_get_fd,
 535                 .put_fd = fs_handle_tracked_put_fd,
 536                 .unlink = fs_handle_tracked_unlink,
 537                 .close = fs_handle_tracked_close,
 538         };
 539
 540         handle->tracker = tracker;
 541
 542         ret = pthread_mutex_init(&handle->lock, NULL);
 543         if (ret) {
 544                 PERROR("Failed to initialize handle mutex while creating fs handle");
 545                 goto error_mutex_init;
 546         }
 547
 548         handle->fd = open_from_properties(directory, path, &properties);
 549         if (handle->fd < 0) {
 550                 PERROR("Failed to open fs handle to %s, open() returned", path);
 551                 goto error;
 552         }
 553
 554         handle->properties = properties;
 555
 556         handle->inode = lttng_inode_registry_get_inode(tracker->inode_registry,
 557                         directory, path, handle->fd,
 558                         tracker->unlinked_file_pool);
 559         if (!handle->inode) {
 560                 ERR("Failed to get lttng_inode corresponding to file %s", path);
 561                 goto error;
 562         }
 563
 564         if (fstat(handle->fd, &fd_stat)) {
 565                 PERROR("Failed to retrieve file descriptor inode while creating fs handle, fstat() returned");
 566                 goto error;
 567         }
 568         handle->ino = fd_stat.st_ino;
 569
 570         fd_tracker_track(tracker, handle);
 571 end:
 572         pthread_mutex_unlock(&tracker->lock);
 573         return handle ? &handle->parent : NULL;
 574 error:
 575         if (handle->inode) {
 576                 lttng_inode_put(handle->inode);
 577         }
 578         pthread_mutex_destroy(&handle->lock);
 579 error_mutex_init:
 580         free(handle);
 581         handle = NULL;
 582         goto end;
 583 }
 584
 585 /* Caller must hold the tracker's lock. */
 586 static int fd_tracker_suspend_handles(
 587                 struct fd_tracker *tracker, unsigned int count)
 588 {
 589         unsigned int left_to_close = count;
 590         unsigned int attempts_left = tracker->count.suspendable.active;
 591         struct fs_handle_tracked *handle, *tmp;
 592
 593         cds_list_for_each_entry_safe(handle, tmp, &tracker->active_handles,
 594                         handles_list_node) {
 595                 int ret;
 596
 597                 fd_tracker_untrack(tracker, handle);
 598                 ret = fs_handle_tracked_suspend(handle);
 599                 fd_tracker_track(tracker, handle);
 600                 if (!ret) {
 601                         left_to_close--;
 602                 }
 603                 attempts_left--;
 604
 605                 if (left_to_close == 0 || attempts_left == 0) {
 606                         break;
 607                 }
 608         }
 609         return left_to_close ? -EMFILE : 0;
 610 }
 611
 612 LTTNG_HIDDEN
 613 int fd_tracker_open_unsuspendable_fd(struct fd_tracker *tracker,
 614                 int *out_fds,
 615                 const char **names,
 616                 unsigned int fd_count,
 617                 fd_open_cb open,
 618                 void *user_data)
 619 {
 620         int ret, user_ret, i, fds_to_suspend;
 621         unsigned int active_fds;
 622         struct unsuspendable_fd **entries;
 623
 624         entries = zmalloc(fd_count * sizeof(*entries));
 625         if (!entries) {
 626                 ret = -1;
 627                 goto end;
 628         }
 629
 630         pthread_mutex_lock(&tracker->lock);
 631
 632         active_fds = ACTIVE_COUNT(tracker);
 633         fds_to_suspend = (int) active_fds + (int) fd_count -
 634                         (int) tracker->capacity;
 635         if (fds_to_suspend > 0) {
 636                 if (fds_to_suspend <= tracker->count.suspendable.active) {
 637                         ret = fd_tracker_suspend_handles(
 638                                         tracker, fds_to_suspend);
 639                         if (ret) {
 640                                 goto end_unlock;
 641                         }
 642                 } else {
 643                         /*
 644                          * There are not enough active suspendable file
 645                          * descriptors to open a new fd and still accommodates the
 646                          * tracker's capacity.
 647                          */
 648                         WARN("Cannot open unsuspendable fd, too many unsuspendable file descriptors are opened (%u)",
 649                                         tracker->count.unsuspendable);
 650                         ret = -EMFILE;
 651                         goto end_unlock;
 652                 }
 653         }
 654
 655         user_ret = open(user_data, out_fds);
 656         if (user_ret) {
 657                 ret = user_ret;
 658                 goto end_unlock;
 659         }
 660
 661         /*
 662          * Add the fds returned by the user's callback to the hashtable
 663          * of unsuspendable fds.
 664          */
 665         for (i = 0; i < fd_count; i++) {
 666                 struct unsuspendable_fd *entry = unsuspendable_fd_create(
 667                                 names ? names[i] : NULL, out_fds[i]);
 668
 669                 if (!entry) {
 670                         ret = -1;
 671                         goto end_free_entries;
 672                 }
 673                 entries[i] = entry;
 674         }
 675
 676         rcu_read_lock();
 677         for (i = 0; i < fd_count; i++) {
 678                 struct cds_lfht_node *node;
 679                 struct unsuspendable_fd *entry = entries[i];
 680
 681                 node = cds_lfht_add_unique(tracker->unsuspendable_fds,
 682                                 hash_key_ulong((void *) (unsigned long)
 683                                                                 out_fds[i],
 684                                                 seed.value),
 685                                 match_fd, (void *) (unsigned long) out_fds[i],
 686                                 &entry->tracker_node);
 687
 688                 if (node != &entry->tracker_node) {
 689                         ret = -EEXIST;
 690                         rcu_read_unlock();
 691                         goto end_free_entries;
 692                 }
 693                 entries[i] = NULL;
 694         }
 695         tracker->count.unsuspendable += fd_count;
 696         rcu_read_unlock();
 697         ret = user_ret;
 698 end_unlock:
 699         pthread_mutex_unlock(&tracker->lock);
 700 end:
 701         free(entries);
 702         return ret;
 703 end_free_entries:
 704         for (i = 0; i < fd_count; i++) {
 705                 unsuspendable_fd_destroy(entries[i]);
 706         }
 707         goto end_unlock;
 708 }
 709
 710 LTTNG_HIDDEN
 711 int fd_tracker_close_unsuspendable_fd(struct fd_tracker *tracker,
 712                 int *fds_in,
 713                 unsigned int fd_count,
 714                 fd_close_cb close,
 715                 void *user_data)
 716 {
 717         int i, ret, user_ret;
 718         int *fds = NULL;
 719
 720         /*
 721          * Maintain a local copy of fds_in as the user's callback may modify its
 722          * contents (e.g. setting the fd(s) to -1 after close).
 723          */
 724         fds = malloc(sizeof(*fds) * fd_count);
 725         if (!fds) {
 726                 ret = -1;
 727                 goto end;
 728         }
 729         memcpy(fds, fds_in, sizeof(*fds) * fd_count);
 730
 731         pthread_mutex_lock(&tracker->lock);
 732         rcu_read_lock();
 733
 734         /* Let the user close the file descriptors. */
 735         user_ret = close(user_data, fds_in);
 736         if (user_ret) {
 737                 ret = user_ret;
 738                 goto end_unlock;
 739         }
 740
 741         /* Untrack the fds that were just closed by the user's callback. */
 742         for (i = 0; i < fd_count; i++) {
 743                 struct cds_lfht_node *node;
 744                 struct cds_lfht_iter iter;
 745                 struct unsuspendable_fd *entry;
 746
 747                 cds_lfht_lookup(tracker->unsuspendable_fds,
 748                                 hash_key_ulong((void *) (unsigned long) fds[i],
 749                                                 seed.value),
 750                                 match_fd, (void *) (unsigned long) fds[i],
 751                                 &iter);
 752                 node = cds_lfht_iter_get_node(&iter);
 753                 if (!node) {
 754                         /* Unknown file descriptor. */
 755                         WARN("Untracked file descriptor %d passed to fd_tracker_close_unsuspendable_fd()",
 756                                         fds[i]);
 757                         ret = -EINVAL;
 758                         goto end_unlock;
 759                 }
 760                 entry = caa_container_of(
 761                                 node, struct unsuspendable_fd, tracker_node);
 762
 763                 cds_lfht_del(tracker->unsuspendable_fds, node);
 764                 unsuspendable_fd_destroy(entry);
 765                 fds[i] = -1;
 766         }
 767
 768         tracker->count.unsuspendable -= fd_count;
 769         ret = 0;
 770 end_unlock:
 771         rcu_read_unlock();
 772         pthread_mutex_unlock(&tracker->lock);
 773         free(fds);
 774 end:
 775         return ret;
 776 }
 777
 778 /* Caller must have taken the tracker's and handle's locks. */
 779 static void fd_tracker_track(
 780                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 781 {
 782         if (handle->fd >= 0) {
 783                 tracker->count.suspendable.active++;
 784                 cds_list_add_tail(&handle->handles_list_node,
 785                                 &tracker->active_handles);
 786         } else {
 787                 tracker->count.suspendable.suspended++;
 788                 cds_list_add_tail(&handle->handles_list_node,
 789                                 &tracker->suspended_handles);
 790         }
 791 }
 792
 793 /* Caller must have taken the tracker's and handle's locks. */
 794 static void fd_tracker_untrack(
 795                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 796 {
 797         if (handle->fd >= 0) {
 798                 tracker->count.suspendable.active--;
 799         } else {
 800                 tracker->count.suspendable.suspended--;
 801         }
 802         cds_list_del(&handle->handles_list_node);
 803 }
 804
 805 /* Caller must have taken the tracker's and handle's locks. */
 806 static int fd_tracker_restore_handle(
 807                 struct fd_tracker *tracker, struct fs_handle_tracked *handle)
 808 {
 809         int ret;
 810
 811         fd_tracker_untrack(tracker, handle);
 812         if (ACTIVE_COUNT(tracker) >= tracker->capacity) {
 813                 ret = fd_tracker_suspend_handles(tracker, 1);
 814                 if (ret) {
 815                         goto end;
 816                 }
 817         }
 818         ret = fs_handle_tracked_restore(handle);
 819 end:
 820         fd_tracker_track(tracker, handle);
 821         return ret ? ret : handle->fd;
 822 }
 823
 824 static int fs_handle_tracked_get_fd(struct fs_handle *_handle)
 825 {
 826         int ret;
 827         struct fs_handle_tracked *handle =
 828                         container_of(_handle, struct fs_handle_tracked, parent);
 829
 830         /*
 831          * TODO This should be optimized as it is a fairly hot path.
 832          * The fd-tracker's lock should only be taken when a fs_handle is
 833          * restored (slow path). On the fast path (fs_handle is active),
 834          * the only effect on the fd_tracker is marking the handle as the
 835          * most recently used. Currently, it is done by a call to the
 836          * track/untrack helpers, but it should be done atomically.
 837          *
 838          * Note that the lock's nesting order must still be respected here.
 839          * The handle's lock nests inside the tracker's lock.
 840          */
 841         pthread_mutex_lock(&handle->tracker->lock);
 842         pthread_mutex_lock(&handle->lock);
 843         assert(!handle->in_use);
 844
 845         handle->tracker->stats.uses++;
 846         if (handle->fd >= 0) {
 847                 ret = handle->fd;
 848                 /* Mark as most recently used. */
 849                 fd_tracker_untrack(handle->tracker, handle);
 850                 fd_tracker_track(handle->tracker, handle);
 851         } else {
 852                 handle->tracker->stats.misses++;
 853                 ret = fd_tracker_restore_handle(handle->tracker, handle);
 854                 if (ret < 0) {
 855                         handle->tracker->stats.errors++;
 856                         goto end;
 857                 }
 858         }
 859         handle->in_use = true;
 860 end:
 861         pthread_mutex_unlock(&handle->lock);
 862         pthread_mutex_unlock(&handle->tracker->lock);
 863         return ret;
 864 }
 865
 866 static void fs_handle_tracked_put_fd(struct fs_handle *_handle)
 867 {
 868         struct fs_handle_tracked *handle =
 869                         container_of(_handle, struct fs_handle_tracked, parent);
 870
 871         pthread_mutex_lock(&handle->lock);
 872         handle->in_use = false;
 873         pthread_mutex_unlock(&handle->lock);
 874 }
 875
 876 static int fs_handle_tracked_unlink(struct fs_handle *_handle)
 877 {
 878         int ret;
 879         struct fs_handle_tracked *handle =
 880                         container_of(_handle, struct fs_handle_tracked, parent);
 881
 882         pthread_mutex_lock(&handle->tracker->lock);
 883         pthread_mutex_lock(&handle->lock);
 884         ret = lttng_inode_unlink(handle->inode);
 885         pthread_mutex_unlock(&handle->lock);
 886         pthread_mutex_unlock(&handle->tracker->lock);
 887         return ret;
 888 }
 889
 890 static int fs_handle_tracked_close(struct fs_handle *_handle)
 891 {
 892         int ret = 0;
 893         const char *path = NULL;
 894         struct fs_handle_tracked *handle =
 895                         container_of(_handle, struct fs_handle_tracked, parent);
 896         struct lttng_directory_handle *inode_directory_handle = NULL;
 897
 898         if (!handle) {
 899                 ret = -EINVAL;
 900                 goto end;
 901         }
 902
 903         pthread_mutex_lock(&handle->tracker->lock);
 904         pthread_mutex_lock(&handle->lock);
 905         if (handle->inode) {
 906                 lttng_inode_borrow_location(handle->inode, NULL, &path);
 907                 /*
 908                  * Here a reference to the inode's directory handle is acquired
 909                  * to prevent the last reference to it from being released while
 910                  * the tracker's lock is taken.
 911                  *
 912                  * If this wasn't done, the directory handle could attempt to
 913                  * close its underlying directory file descriptor, which would
 914                  * attempt to lock the tracker's lock, resulting in a deadlock.
 915                  *
 916                  * Since a new reference to the directory handle is taken within
 917                  * the scope of this function, it is not possible for the last
 918                  * reference to the inode's location directory handle to be
 919                  * released during the call to lttng_inode_put().
 920                  *
 921                  * We wait until the tracker's lock is released to release the
 922                  * reference. Hence, the call to the tracker is delayed just
 923                  * enough to not attempt to recursively acquire the tracker's
 924                  * lock twice.
 925                  */
 926                 inode_directory_handle =
 927                                 lttng_inode_get_location_directory_handle(
 928                                                 handle->inode);
 929         }
 930         fd_tracker_untrack(handle->tracker, handle);
 931         if (handle->fd >= 0) {
 932                 /*
 933                  * The return value of close() is not propagated as there
 934                  * isn't much the user can do about it.
 935                  */
 936                 if (close(handle->fd)) {
 937                         PERROR("Failed to close the file descriptor (%d) of fs handle to %s, close() returned",
 938                                         handle->fd, path ? path : "Unknown");
 939                 }
 940                 handle->fd = -1;
 941         }
 942         if (handle->inode) {
 943                 lttng_inode_put(handle->inode);
 944         }
 945         pthread_mutex_unlock(&handle->lock);
 946         pthread_mutex_destroy(&handle->lock);
 947         pthread_mutex_unlock(&handle->tracker->lock);
 948         free(handle);
 949         lttng_directory_handle_put(inode_directory_handle);
 950 end:
 951         return ret;
 952 }