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

   1 /*
   2  * Copyright (C) 2018 - Jérémie Galarneau <jeremie.galarneau@efficios.com>
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License, version 2 only, as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 51
  15  * Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  16  */
  17
  18 #include <urcu.h>
  19 #include <urcu/list.h>
  20 #include <urcu/rculfhash.h>
  21
  22 #include <sys/stat.h>
  23 #include <sys/types.h>
  24 #include <fcntl.h>
  25 #include <stdbool.h>
  26 #include <pthread.h>
  27 #include <inttypes.h>
  28
  29 #include "common/macros.h"
  30 #include "common/error.h"
  31 #include "common/defaults.h"
  32 #include "common/hashtable/utils.h"
  33 #include "common/hashtable/hashtable.h"
  34
  35 #include "fd-tracker.h"
  36 #include "inode.h"
  37
  38 /* Tracker lock must be taken by the user. */
  39 #define TRACKED_COUNT(tracker)                 \
  40         (tracker->count.suspendable.active +   \
  41         tracker->count.suspendable.suspended + \
  42         tracker->count.unsuspendable)
  43
  44 /* Tracker lock must be taken by the user. */
  45 #define ACTIVE_COUNT(tracker)                  \
  46         (tracker->count.suspendable.active +   \
  47         tracker->count.unsuspendable)
  48
  49 /* Tracker lock must be taken by the user. */
  50 #define SUSPENDED_COUNT(tracker)               \
  51         (tracker->count.suspendable.suspended)
  52
  53 /* Tracker lock must be taken by the user. */
  54 #define SUSPENDABLE_COUNT(tracker)             \
  55         (tracker->count.suspendable.active +   \
  56         tracker->count.suspendable.suspended)
  57
  58 /* Tracker lock must be taken by the user. */
  59 #define UNSUSPENDABLE_COUNT(tracker)           \
  60         (tracker->count.unsuspendable)
  61
  62 struct fd_tracker {
  63         pthread_mutex_t lock;
  64         struct {
  65                 struct {
  66                         unsigned int active;
  67                         unsigned int suspended;
  68                 } suspendable;
  69                 unsigned int unsuspendable;
  70         } count;
  71         unsigned int capacity;
  72         struct {
  73                 uint64_t uses;
  74                 uint64_t misses;
  75                 /* Failures to suspend or restore fs handles. */
  76                 uint64_t errors;
  77         } stats;
  78         /*
  79          * The head of the active_handles list is always the least recently
  80          * used active handle. When an handle is used, it is removed from the
  81          * list and added to the end. When a file has to be suspended, the
  82          * first element in the list is "popped", suspended, and added to the
  83          * list of suspended handles.
  84          */
  85         struct cds_list_head active_handles;
  86         struct cds_list_head suspended_handles;
  87         struct cds_lfht *unsuspendable_fds;
  88         struct lttng_inode_registry *inode_registry;
  89 };
  90
  91 struct open_properties {
  92         int flags;
  93         struct {
  94                 bool is_set;
  95                 mode_t value;
  96         } mode;
  97 };
  98
  99 /*
 100  * A fs_handle is not ref-counted. Therefore, it is assumed that a
 101  * handle is never in-use while it is being reclaimed. It can be
 102  * shared by multiple threads, but external synchronization is required
 103  * to ensure it is not still being used when it is reclaimed (close method).
 104  * In this respect, it is not different from a regular file descriptor.
 105  *
 106  * The fs_handle lock always nests _within_ the tracker's lock.
 107  */
 108 struct fs_handle {
 109         pthread_mutex_t lock;
 110         /*
 111          * Weak reference to the tracker. All fs_handles are assumed to have
 112          * been closed at the moment of the destruction of the fd_tracker.
 113          */
 114         struct fd_tracker *tracker;
 115         struct open_properties properties;
 116         struct lttng_inode *inode;
 117         int fd;
 118         /* inode number of the file at the time of the handle's creation. */
 119         uint64_t ino;
 120         bool in_use;
 121         /* Offset to which the file should be restored. */
 122         off_t offset;
 123         struct cds_list_head handles_list_node;
 124 };
 125
 126 struct unsuspendable_fd {
 127         /*
 128          * Accesses are only performed through the tracker, which is protected
 129          * by its own lock.
 130          */
 131         int fd;
 132         char *name;
 133         struct cds_lfht_node tracker_node;
 134         struct rcu_head rcu_head;
 135 };
 136
 137 static struct {
 138         pthread_mutex_t lock;
 139         bool initialized;
 140         unsigned long value;
 141 } seed = {
 142         .lock = PTHREAD_MUTEX_INITIALIZER,
 143 };
 144
 145 static int match_fd(struct cds_lfht_node *node, const void *key);
 146 static void unsuspendable_fd_destroy(struct unsuspendable_fd *entry);
 147 static struct unsuspendable_fd *unsuspendable_fd_create(const char *name,
 148                 int fd);
 149 static int open_from_properties(const char *path,
 150                 struct open_properties *properties);
 151
 152 static void fs_handle_log(struct fs_handle *handle);
 153 static int fs_handle_suspend(struct fs_handle *handle);
 154 static int fs_handle_restore(struct fs_handle *handle);
 155
 156 static void fd_tracker_track(struct fd_tracker *tracker,
 157                 struct fs_handle *handle);
 158 static void fd_tracker_untrack(struct fd_tracker *tracker,
 159                 struct fs_handle *handle);
 160 static int fd_tracker_suspend_handles(struct fd_tracker *tracker,
 161                 unsigned int count);
 162 static int fd_tracker_restore_handle(struct fd_tracker *tracker,
 163                 struct fs_handle *handle);
 164
 165 /* Match function of the tracker's unsuspendable_fds hash table. */
 166 static
 167 int match_fd(struct cds_lfht_node *node, const void *key)
 168 {
 169         struct unsuspendable_fd *entry =
 170                 caa_container_of(node, struct unsuspendable_fd, tracker_node);
 171
 172         return hash_match_key_ulong((void *) (unsigned long) entry->fd,
 173                         (void *) key);
 174 }
 175
 176 static
 177 void delete_unsuspendable_fd(struct rcu_head *head)
 178 {
 179         struct unsuspendable_fd *fd = caa_container_of(head,
 180                         struct unsuspendable_fd, rcu_head);
 181
 182         free(fd->name);
 183         free(fd);
 184 }
 185
 186 static
 187 void unsuspendable_fd_destroy(struct unsuspendable_fd *entry)
 188 {
 189         if (!entry) {
 190                 return;
 191         }
 192         call_rcu(&entry->rcu_head, delete_unsuspendable_fd);
 193 }
 194
 195 static
 196 struct unsuspendable_fd *unsuspendable_fd_create(const char *name, int fd)
 197 {
 198         struct unsuspendable_fd *entry =
 199                         zmalloc(sizeof(*entry));
 200
 201         if (!entry) {
 202                 goto error;
 203         }
 204         if (name) {
 205                 entry->name = strdup(name);
 206                 if (!entry->name) {
 207                         goto error;
 208                 }
 209         }
 210         cds_lfht_node_init(&entry->tracker_node);
 211         entry->fd = fd;
 212         return entry;
 213 error:
 214         unsuspendable_fd_destroy(entry);
 215         return NULL;
 216 }
 217
 218 static
 219 void fs_handle_log(struct fs_handle *handle)
 220 {
 221         const char *path;
 222
 223         pthread_mutex_lock(&handle->lock);
 224         path = lttng_inode_get_path(handle->inode);
 225
 226         if (handle->fd >= 0) {
 227                 DBG_NO_LOC("    %s [active, fd %d%s]",
 228                                 path,
 229                                 handle->fd,
 230                                 handle->in_use ? ", in use" : "");
 231         } else {
 232                 DBG_NO_LOC("    %s [suspended]", path);
 233         }
 234         pthread_mutex_unlock(&handle->lock);
 235 }
 236
 237 /* Tracker lock must be held by the caller. */
 238 static
 239 int fs_handle_suspend(struct fs_handle *handle)
 240 {
 241         int ret = 0;
 242         struct stat fs_stat;
 243         const char *path;
 244
 245         pthread_mutex_lock(&handle->lock);
 246         path = lttng_inode_get_path(handle->inode);
 247         assert(handle->fd >= 0);
 248         if (handle->in_use) {
 249                 /* This handle can't be suspended as it is currently in use. */
 250                 ret = -EAGAIN;
 251                 goto end;
 252         }
 253
 254         ret = stat(path, &fs_stat);
 255         if (ret) {
 256                 PERROR("Filesystem handle to %s cannot be suspended as stat() failed",
 257                                 path);
 258                 ret = -errno;
 259                 goto end;
 260         }
 261
 262         if (fs_stat.st_ino != handle->ino) {
 263                 /* Don't suspend as the handle would not be restorable. */
 264                 WARN("Filesystem handle to %s cannot be suspended as its inode changed",
 265                                 path);
 266                 ret = -ENOENT;
 267                 goto end;
 268         }
 269
 270         handle->offset = lseek(handle->fd, 0, SEEK_CUR);
 271         if (handle->offset == -1) {
 272                 WARN("Filesystem handle to %s cannot be suspended as lseek() failed to sample its current position",
 273                                 path);
 274                 ret = -errno;
 275                 goto end;
 276         }
 277
 278         ret = close(handle->fd);
 279         if (ret) {
 280                 PERROR("Filesystem handle to %s cannot be suspended as close() failed",
 281                                 path);
 282                 ret = -errno;
 283                 goto end;
 284         }
 285         DBG("Suspended filesystem handle to %s (fd %i) at position %" PRId64,
 286                         path, handle->fd, handle->offset);
 287         handle->fd = -1;
 288 end:
 289         if (ret) {
 290                 handle->tracker->stats.errors++;
 291         }
 292         pthread_mutex_unlock(&handle->lock);
 293         return ret;
 294 }
 295
 296 /* Caller must hold the tracker and handle's locks. */
 297 static
 298 int fs_handle_restore(struct fs_handle *handle)
 299 {
 300         int ret, fd = -1;
 301         const char *path = lttng_inode_get_path(handle->inode);
 302
 303         assert(handle->fd == -1);
 304         assert(path);
 305         ret = open_from_properties(path,
 306                         &handle->properties);
 307         if (ret < 0) {
 308                 errno = -ret;
 309                 PERROR("Failed to restore filesystem handle to %s, open() failed",
 310                                 path);
 311                 goto end;
 312         }
 313         fd = ret;
 314
 315         ret = lseek(fd, handle->offset, SEEK_SET);
 316         if (ret < 0) {
 317                 PERROR("Failed to restore filesystem handle to %s, lseek() failed",
 318                                 path);
 319                 ret = -errno;
 320                 goto end;
 321         }
 322         DBG("Restored filesystem handle to %s (fd %i) at position %" PRId64,
 323                         path, fd, handle->offset);
 324         ret = 0;
 325         handle->fd = fd;
 326         fd = -1;
 327 end:
 328         if (fd >= 0) {
 329                 (void) close(fd);
 330         }
 331         return ret;
 332 }
 333
 334 static
 335 int open_from_properties(const char *path, struct open_properties *properties)
 336 {
 337         int ret;
 338
 339         /*
 340          * open() ignores the 'flags' parameter unless the O_CREAT or O_TMPFILE
 341          * flags are set. O_TMPFILE would not make sense in the context of a
 342          * suspendable fs_handle as it would not be restorable (see OPEN(2)),
 343          * thus it is ignored here.
 344          */
 345         if ((properties->flags & O_CREAT) && properties->mode.is_set) {
 346                 ret = open(path, properties->flags,
 347                                 properties->mode.value);
 348         } else {
 349                 ret = open(path, properties->flags);
 350         }
 351         /*
 352          * Some flags should not be used beyond the initial open() of a
 353          * restorable file system handle. O_CREAT and O_TRUNC must
 354          * be cleared since it would be unexpected to re-use them
 355          * when the handle is retored:
 356          *  - O_CREAT should not be needed as the file has been created
 357          *    on the initial call to open(),
 358          *  - O_TRUNC would destroy the file's contents by truncating it
 359          *    to length 0.
 360          */
 361         properties->flags &= ~(O_CREAT | O_TRUNC);
 362         if (ret < 0) {
 363                 ret = -errno;
 364                 goto end;
 365         }
 366 end:
 367         return ret;
 368 }
 369
 370 struct fd_tracker *fd_tracker_create(unsigned int capacity)
 371 {
 372         struct fd_tracker *tracker = zmalloc(sizeof(struct fd_tracker));
 373
 374         if (!tracker) {
 375                 goto end;
 376         }
 377
 378         pthread_mutex_lock(&seed.lock);
 379         if (!seed.initialized) {
 380                 seed.value = (unsigned long) time(NULL);
 381                 seed.initialized = true;
 382         }
 383         pthread_mutex_unlock(&seed.lock);
 384
 385         CDS_INIT_LIST_HEAD(&tracker->active_handles);
 386         CDS_INIT_LIST_HEAD(&tracker->suspended_handles);
 387         tracker->capacity = capacity;
 388         tracker->unsuspendable_fds = cds_lfht_new(DEFAULT_HT_SIZE, 1, 0,
 389                         CDS_LFHT_AUTO_RESIZE | CDS_LFHT_ACCOUNTING, NULL);
 390         if (!tracker->unsuspendable_fds) {
 391                 ERR("Failed to create fd-tracker's unsuspendable_fds hash table");
 392                 goto error;
 393         }
 394         tracker->inode_registry = lttng_inode_registry_create();
 395         if (!tracker->inode_registry) {
 396                 ERR("Failed to create fd-tracker's inode registry");
 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 *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(handle, &tracker->active_handles,
 429                         handles_list_node) {
 430                 fs_handle_log(handle);
 431         }
 432         cds_list_for_each_entry(handle, &tracker->suspended_handles,
 433                         handles_list_node) {
 434                 fs_handle_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]", unsuspendable_fd->name ? : "Unnamed",
 445                                 unsuspendable_fd->fd);
 446         }
 447         rcu_read_unlock();
 448         if (!UNSUSPENDABLE_COUNT(tracker)) {
 449                 DBG_NO_LOC("    None");
 450         }
 451
 452         pthread_mutex_unlock(&tracker->lock);
 453 }
 454
 455 int fd_tracker_destroy(struct fd_tracker *tracker)
 456 {
 457         int ret = 0;
 458
 459         /*
 460          * Refuse to destroy the tracker as fs_handles may still old
 461          * weak references to the tracker.
 462          */
 463         pthread_mutex_lock(&tracker->lock);
 464         if (TRACKED_COUNT(tracker)) {
 465                 ERR("A file descriptor leak has been detected: %u tracked file descriptors are still being tracked",
 466                                 TRACKED_COUNT(tracker));
 467                 pthread_mutex_unlock(&tracker->lock);
 468                 fd_tracker_log(tracker);
 469                 ret = -1;
 470                 goto end;
 471         }
 472         pthread_mutex_unlock(&tracker->lock);
 473
 474         if (tracker->unsuspendable_fds) {
 475                 ret = cds_lfht_destroy(tracker->unsuspendable_fds, NULL);
 476                 assert(!ret);
 477         }
 478
 479         lttng_inode_registry_destroy(tracker->inode_registry);
 480         pthread_mutex_destroy(&tracker->lock);
 481         free(tracker);
 482 end:
 483         return ret;
 484 }
 485
 486 /*
 487  * If return NULL check errno for error.
 488  */
 489 struct fs_handle *fd_tracker_open_fs_handle(struct fd_tracker *tracker,
 490                 const char *path, int flags, mode_t *mode)
 491 {
 492         int ret = 0;
 493         struct fs_handle *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                                 ERR("Suspend handled failed");
 507                                 ret = EMFILE;
 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                         ret = EMFILE;
 519                         goto end;
 520                 }
 521         }
 522
 523         handle = zmalloc(sizeof(*handle));
 524         if (!handle) {
 525                 ret = ENOMEM;
 526                 goto end;
 527         }
 528         handle->tracker = tracker;
 529
 530         ret = pthread_mutex_init(&handle->lock, NULL);
 531         if (ret) {
 532                 PERROR("Failed to initialize handle mutex while creating fs handle");
 533                 ret = errno;
 534                 goto error_mutex_init;
 535         }
 536
 537         handle->fd = open_from_properties(path, &properties);
 538         if (handle->fd < 0) {
 539                 /* ret contains -errno on error. */
 540                 ret = -ret;
 541                 goto error;
 542         }
 543
 544         handle->properties = properties;
 545
 546         handle->inode = lttng_inode_registry_get_inode(tracker->inode_registry,
 547                         handle->fd, path);
 548         if (!handle->inode) {
 549                 ret = errno;
 550                 ERR("Failed to get lttng_inode corresponding to file %s",
 551                                 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                 ret = errno;
 558                 goto error;
 559         }
 560         handle->ino = fd_stat.st_ino;
 561
 562         fd_tracker_track(tracker, handle);
 563 end:
 564         pthread_mutex_unlock(&tracker->lock);
 565         errno = ret;
 566         return handle;
 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
 580 int fd_tracker_suspend_handles(struct fd_tracker *tracker,
 581                 unsigned int count)
 582 {
 583         unsigned int left_to_close = count;
 584         struct fs_handle *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_suspend(handle);
 592                 fd_tracker_track(tracker, handle);
 593                 if (!ret) {
 594                         left_to_close--;
 595                 }
 596
 597                 if (!left_to_close) {
 598                         break;
 599                 }
 600         }
 601         return left_to_close ? -EMFILE : 0;
 602 }
 603
 604 int fd_tracker_open_unsuspendable_fd(struct fd_tracker *tracker,
 605                 int *out_fds, const char **names, unsigned int fd_count,
 606                 fd_open_cb open, void *user_data)
 607 {
 608         int ret, user_ret, i, fds_to_suspend;
 609         unsigned int active_fds;
 610         struct unsuspendable_fd *entries[fd_count];
 611
 612         memset(entries, 0, sizeof(entries));
 613
 614         pthread_mutex_lock(&tracker->lock);
 615
 616         active_fds = ACTIVE_COUNT(tracker);
 617         fds_to_suspend = (int) active_fds + (int) fd_count - (int) tracker->capacity;
 618         if (fds_to_suspend > 0) {
 619                 if (fds_to_suspend <= tracker->count.suspendable.active) {
 620                         ret = fd_tracker_suspend_handles(tracker, fds_to_suspend);
 621                         if (ret) {
 622                                 goto end;
 623                         }
 624                 } else {
 625                         /*
 626                          * There are not enough active suspendable file
 627                          * descriptors to open a new fd and still accomodate the
 628                          * tracker's capacity.
 629                          */
 630                         WARN("Cannot open unsuspendable fd, too many unsuspendable file descriptors are opened (%u)",
 631                                         tracker->count.unsuspendable);
 632                         ret = -EMFILE;
 633                         goto end;
 634                 }
 635         }
 636
 637         user_ret = open(user_data, out_fds);
 638         if (user_ret) {
 639                 ret = user_ret;
 640                 goto end;
 641         }
 642
 643         /*
 644          * Add the fds returned by the user's callback to the hashtable
 645          * of unsuspendable fds.
 646          */
 647         for (i = 0; i < fd_count; i++) {
 648                 struct unsuspendable_fd *entry =
 649                                 unsuspendable_fd_create(names ? names[i] : NULL,
 650                                                 out_fds[i]);
 651
 652                 if (!entry) {
 653                         ret = -1;
 654                         goto end_free_entries;
 655                 }
 656                 entries[i] = entry;
 657         }
 658
 659         rcu_read_lock();
 660         for (i = 0; i < fd_count; i++) {
 661                 struct cds_lfht_node *node;
 662                 struct unsuspendable_fd *entry = entries[i];
 663
 664                 node = cds_lfht_add_unique(
 665                                 tracker->unsuspendable_fds,
 666                                 hash_key_ulong((void *) (unsigned long) out_fds[i],
 667                                                 seed.value),
 668                                 match_fd,
 669                                 (void *) (unsigned long) out_fds[i],
 670                                 &entry->tracker_node);
 671
 672                 if (node != &entry->tracker_node) {
 673                         ret = -EEXIST;
 674                         rcu_read_unlock();
 675                         goto end_free_entries;
 676                 }
 677                 entries[i] = NULL;
 678         }
 679         tracker->count.unsuspendable += fd_count;
 680         rcu_read_unlock();
 681         ret = user_ret;
 682 end:
 683         pthread_mutex_unlock(&tracker->lock);
 684         return ret;
 685 end_free_entries:
 686         for (i = 0; i < fd_count; i++) {
 687                 unsuspendable_fd_destroy(entries[i]);
 688         }
 689         goto end;
 690 }
 691
 692 int fd_tracker_close_unsuspendable_fd(struct fd_tracker *tracker,
 693                 int *fds_in, unsigned int fd_count, fd_close_cb close,
 694                 void *user_data)
 695 {
 696         int i, ret, user_ret;
 697         int fds[fd_count];
 698
 699         /*
 700          * Maintain a local copy of fds_in as the user's callback may modify its
 701          * contents (e.g. setting the fd(s) to -1 after close).
 702          */
 703         memcpy(fds, fds_in, sizeof(*fds) * fd_count);
 704
 705         pthread_mutex_lock(&tracker->lock);
 706         rcu_read_lock();
 707
 708         /* Let the user close the file descriptors. */
 709         user_ret = close(user_data, fds_in);
 710         if (user_ret) {
 711                 ret = user_ret;
 712                 goto end;
 713         }
 714
 715         /* Untrack the fds that were just closed by the user's callback. */
 716         for (i = 0; i < fd_count; i++) {
 717                 struct cds_lfht_node *node;
 718                 struct cds_lfht_iter iter;
 719                 struct unsuspendable_fd *entry;
 720
 721                 cds_lfht_lookup(tracker->unsuspendable_fds,
 722                                 hash_key_ulong((void *) (unsigned long) fds[i],
 723                                                 seed.value),
 724                                 match_fd,
 725                                 (void *) (unsigned long) fds[i],
 726                                 &iter);
 727                 node = cds_lfht_iter_get_node(&iter);
 728                 if (!node) {
 729                         /* Unknown file descriptor. */
 730                         WARN("Untracked file descriptor %d passed to fd_tracker_close_unsuspendable_fd()",
 731                                         fds[i]);
 732                         ret = -EINVAL;
 733                         goto end;
 734                 }
 735                 entry = caa_container_of(node,
 736                                 struct unsuspendable_fd,
 737                                 tracker_node);
 738
 739                 cds_lfht_del(tracker->unsuspendable_fds, node);
 740                 unsuspendable_fd_destroy(entry);
 741                 fds[i] = -1;
 742         }
 743
 744         tracker->count.unsuspendable -= fd_count;
 745         ret = 0;
 746 end:
 747         rcu_read_unlock();
 748         pthread_mutex_unlock(&tracker->lock);
 749         return ret;
 750 }
 751
 752 /* Caller must have taken the tracker's and handle's locks. */
 753 static
 754 void fd_tracker_track(struct fd_tracker *tracker, struct fs_handle *handle)
 755 {
 756         if (handle->fd >= 0) {
 757                 tracker->count.suspendable.active++;
 758                 cds_list_add_tail(&handle->handles_list_node,
 759                                 &tracker->active_handles);
 760         } else {
 761                 tracker->count.suspendable.suspended++;
 762                 cds_list_add_tail(&handle->handles_list_node,
 763                                 &tracker->suspended_handles);
 764         }
 765 }
 766
 767 /* Caller must have taken the tracker's and handle's locks. */
 768 static
 769 void fd_tracker_untrack(struct fd_tracker *tracker, struct fs_handle *handle)
 770 {
 771         if (handle->fd >= 0) {
 772                 tracker->count.suspendable.active--;
 773         } else {
 774                 tracker->count.suspendable.suspended--;
 775         }
 776         cds_list_del(&handle->handles_list_node);
 777 }
 778
 779 /* Caller must have taken the tracker's and handle's locks. */
 780 static
 781 int fd_tracker_restore_handle(struct fd_tracker *tracker,
 782                 struct fs_handle *handle)
 783 {
 784         int ret;
 785
 786         fd_tracker_untrack(tracker, handle);
 787         if (ACTIVE_COUNT(tracker) >= tracker->capacity) {
 788                 ret = fd_tracker_suspend_handles(tracker, 1);
 789                 if (ret) {
 790                         goto end;
 791                 }
 792         }
 793         ret = fs_handle_restore(handle);
 794 end:
 795         fd_tracker_track(tracker, handle);
 796         return ret ? ret : handle->fd;
 797 }
 798
 799 int fs_handle_get_fd(struct fs_handle *handle)
 800 {
 801         int ret;
 802
 803         /*
 804          * TODO This should be optimized as it is a fairly hot path.
 805          * The fd-tracker's lock should only be taken when a fs_handle is
 806          * restored (slow path). On the fast path (fs_handle is active),
 807          * the only effect on the fd_tracker is marking the handle as the
 808          * most recently used. Currently, it is done by a call to the
 809          * track/untrack helpers, but it should be done atomically.
 810          *
 811          * Note that the lock's nesting order must still be respected here.
 812          * The handle's lock nests inside the tracker's lock.
 813          */
 814         pthread_mutex_lock(&handle->tracker->lock);
 815         pthread_mutex_lock(&handle->lock);
 816         assert(!handle->in_use);
 817
 818         handle->tracker->stats.uses++;
 819         if (handle->fd >= 0) {
 820                 ret = handle->fd;
 821                 /* Mark as most recently used. */
 822                 fd_tracker_untrack(handle->tracker, handle);
 823                 fd_tracker_track(handle->tracker, handle);
 824         } else {
 825                 handle->tracker->stats.misses++;
 826                 ret = fd_tracker_restore_handle(handle->tracker, handle);
 827                 if (ret < 0) {
 828                         handle->tracker->stats.errors++;
 829                         goto end;
 830                 }
 831         }
 832         handle->in_use = true;
 833 end:
 834         pthread_mutex_unlock(&handle->lock);
 835         pthread_mutex_unlock(&handle->tracker->lock);
 836         return ret;
 837 }
 838
 839 void fs_handle_put_fd(struct fs_handle *handle)
 840 {
 841         pthread_mutex_lock(&handle->lock);
 842         handle->in_use = false;
 843         pthread_mutex_unlock(&handle->lock);
 844 }
 845
 846 int fs_handle_unlink(struct fs_handle *handle)
 847 {
 848         int ret;
 849
 850         pthread_mutex_lock(&handle->tracker->lock);
 851         pthread_mutex_lock(&handle->lock);
 852         ret = lttng_inode_defer_unlink(handle->inode);
 853         pthread_mutex_unlock(&handle->lock);
 854         pthread_mutex_unlock(&handle->tracker->lock);
 855         return ret;
 856 }
 857
 858 int fs_handle_close(struct fs_handle *handle)
 859 {
 860         int ret = 0;
 861         const char *path = NULL;
 862
 863         if (!handle) {
 864                 ret = -EINVAL;
 865                 goto end;
 866         }
 867
 868         pthread_mutex_lock(&handle->tracker->lock);
 869         pthread_mutex_lock(&handle->lock);
 870         if (handle->inode) {
 871                 path = lttng_inode_get_path(handle->inode);
 872         }
 873         fd_tracker_untrack(handle->tracker, handle);
 874         if (handle->fd >= 0) {
 875                 assert(!handle->in_use);
 876                 /*
 877                  * The return value of close() is not propagated as there
 878                  * isn't much the user can do about it.
 879                  */
 880                 if (close(handle->fd)) {
 881                         PERROR("Failed to close the file descritptor (%d) of fs handle to %s, close() returned",
 882                                         handle->fd, path ? path : "Unknown");
 883                 }
 884                 handle->fd = -1;
 885         }
 886         lttng_inode_put(handle->inode);
 887         pthread_mutex_unlock(&handle->lock);
 888         pthread_mutex_destroy(&handle->lock);
 889         pthread_mutex_unlock(&handle->tracker->lock);
 890         free(handle);
 891 end:
 892         return ret;
 893 }