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

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