src/common/fd-tracker/fd-tracker.cpp

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