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

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