fs/btrfs/inode-map.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kthread.h>
   7 #include <linux/pagemap.h>
   8
   9 #include "ctree.h"
  10 #include "disk-io.h"
  11 #include "free-space-cache.h"
  12 #include "inode-map.h"
  13 #include "transaction.h"
  14 #include "delalloc-space.h"
  15
  16 static int caching_kthread(void *data)
  17 {
  18         struct btrfs_root *root = data;
  19         struct btrfs_fs_info *fs_info = root->fs_info;
  20         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
  21         struct btrfs_key key;
  22         struct btrfs_path *path;
  23         struct extent_buffer *leaf;
  24         u64 last = (u64)-1;
  25         int slot;
  26         int ret;
  27
  28         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
  29                 return 0;
  30
  31         path = btrfs_alloc_path();
  32         if (!path)
  33                 return -ENOMEM;
  34
  35         /* Since the commit root is read-only, we can safely skip locking. */
  36         path->skip_locking = 1;
  37         path->search_commit_root = 1;
  38         path->reada = READA_FORWARD;
  39
  40         key.objectid = BTRFS_FIRST_FREE_OBJECTID;
  41         key.offset = 0;
  42         key.type = BTRFS_INODE_ITEM_KEY;
  43 again:
  44         /* need to make sure the commit_root doesn't disappear */
  45         down_read(&fs_info->commit_root_sem);
  46
  47         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  48         if (ret < 0)
  49                 goto out;
  50
  51         while (1) {
  52                 if (btrfs_fs_closing(fs_info))
  53                         goto out;
  54
  55                 leaf = path->nodes[0];
  56                 slot = path->slots[0];
  57                 if (slot >= btrfs_header_nritems(leaf)) {
  58                         ret = btrfs_next_leaf(root, path);
  59                         if (ret < 0)
  60                                 goto out;
  61                         else if (ret > 0)
  62                                 break;
  63
  64                         if (need_resched() ||
  65                             btrfs_transaction_in_commit(fs_info)) {
  66                                 leaf = path->nodes[0];
  67
  68                                 if (WARN_ON(btrfs_header_nritems(leaf) == 0))
  69                                         break;
  70
  71                                 /*
  72                                  * Save the key so we can advances forward
  73                                  * in the next search.
  74                                  */
  75                                 btrfs_item_key_to_cpu(leaf, &key, 0);
  76                                 btrfs_release_path(path);
  77                                 root->ino_cache_progress = last;
  78                                 up_read(&fs_info->commit_root_sem);
  79                                 schedule_timeout(1);
  80                                 goto again;
  81                         } else
  82                                 continue;
  83                 }
  84
  85                 btrfs_item_key_to_cpu(leaf, &key, slot);
  86
  87                 if (key.type != BTRFS_INODE_ITEM_KEY)
  88                         goto next;
  89
  90                 if (key.objectid >= root->highest_objectid)
  91                         break;
  92
  93                 if (last != (u64)-1 && last + 1 != key.objectid) {
  94                         __btrfs_add_free_space(fs_info, ctl, last + 1,
  95                                                key.objectid - last - 1);
  96                         wake_up(&root->ino_cache_wait);
  97                 }
  98
  99                 last = key.objectid;
 100 next:
 101                 path->slots[0]++;
 102         }
 103
 104         if (last < root->highest_objectid - 1) {
 105                 __btrfs_add_free_space(fs_info, ctl, last + 1,
 106                                        root->highest_objectid - last - 1);
 107         }
 108
 109         spin_lock(&root->ino_cache_lock);
 110         root->ino_cache_state = BTRFS_CACHE_FINISHED;
 111         spin_unlock(&root->ino_cache_lock);
 112
 113         root->ino_cache_progress = (u64)-1;
 114         btrfs_unpin_free_ino(root);
 115 out:
 116         wake_up(&root->ino_cache_wait);
 117         up_read(&fs_info->commit_root_sem);
 118
 119         btrfs_free_path(path);
 120
 121         return ret;
 122 }
 123
 124 static void start_caching(struct btrfs_root *root)
 125 {
 126         struct btrfs_fs_info *fs_info = root->fs_info;
 127         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 128         struct task_struct *tsk;
 129         int ret;
 130         u64 objectid;
 131
 132         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 133                 return;
 134
 135         spin_lock(&root->ino_cache_lock);
 136         if (root->ino_cache_state != BTRFS_CACHE_NO) {
 137                 spin_unlock(&root->ino_cache_lock);
 138                 return;
 139         }
 140
 141         root->ino_cache_state = BTRFS_CACHE_STARTED;
 142         spin_unlock(&root->ino_cache_lock);
 143
 144         ret = load_free_ino_cache(fs_info, root);
 145         if (ret == 1) {
 146                 spin_lock(&root->ino_cache_lock);
 147                 root->ino_cache_state = BTRFS_CACHE_FINISHED;
 148                 spin_unlock(&root->ino_cache_lock);
 149                 return;
 150         }
 151
 152         /*
 153          * It can be quite time-consuming to fill the cache by searching
 154          * through the extent tree, and this can keep ino allocation path
 155          * waiting. Therefore at start we quickly find out the highest
 156          * inode number and we know we can use inode numbers which fall in
 157          * [highest_ino + 1, BTRFS_LAST_FREE_OBJECTID].
 158          */
 159         ret = btrfs_find_free_objectid(root, &objectid);
 160         if (!ret && objectid <= BTRFS_LAST_FREE_OBJECTID) {
 161                 __btrfs_add_free_space(fs_info, ctl, objectid,
 162                                        BTRFS_LAST_FREE_OBJECTID - objectid + 1);
 163         }
 164
 165         tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu",
 166                           root->root_key.objectid);
 167         if (IS_ERR(tsk)) {
 168                 btrfs_warn(fs_info, "failed to start inode caching task");
 169                 btrfs_clear_pending_and_info(fs_info, INODE_MAP_CACHE,
 170                                              "disabling inode map caching");
 171         }
 172 }
 173
 174 int btrfs_find_free_ino(struct btrfs_root *root, u64 *objectid)
 175 {
 176         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 177                 return btrfs_find_free_objectid(root, objectid);
 178
 179 again:
 180         *objectid = btrfs_find_ino_for_alloc(root);
 181
 182         if (*objectid != 0)
 183                 return 0;
 184
 185         start_caching(root);
 186
 187         wait_event(root->ino_cache_wait,
 188                    root->ino_cache_state == BTRFS_CACHE_FINISHED ||
 189                    root->free_ino_ctl->free_space > 0);
 190
 191         if (root->ino_cache_state == BTRFS_CACHE_FINISHED &&
 192             root->free_ino_ctl->free_space == 0)
 193                 return -ENOSPC;
 194         else
 195                 goto again;
 196 }
 197
 198 void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
 199 {
 200         struct btrfs_fs_info *fs_info = root->fs_info;
 201         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 202
 203         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 204                 return;
 205 again:
 206         if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 207                 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
 208         } else {
 209                 down_write(&fs_info->commit_root_sem);
 210                 spin_lock(&root->ino_cache_lock);
 211                 if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 212                         spin_unlock(&root->ino_cache_lock);
 213                         up_write(&fs_info->commit_root_sem);
 214                         goto again;
 215                 }
 216                 spin_unlock(&root->ino_cache_lock);
 217
 218                 start_caching(root);
 219
 220                 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
 221
 222                 up_write(&fs_info->commit_root_sem);
 223         }
 224 }
 225
 226 /*
 227  * When a transaction is committed, we'll move those inode numbers which are
 228  * smaller than root->ino_cache_progress from pinned tree to free_ino tree, and
 229  * others will just be dropped, because the commit root we were searching has
 230  * changed.
 231  *
 232  * Must be called with root->fs_info->commit_root_sem held
 233  */
 234 void btrfs_unpin_free_ino(struct btrfs_root *root)
 235 {
 236         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 237         struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
 238         spinlock_t *rbroot_lock = &root->free_ino_pinned->tree_lock;
 239         struct btrfs_free_space *info;
 240         struct rb_node *n;
 241         u64 count;
 242
 243         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 244                 return;
 245
 246         while (1) {
 247                 spin_lock(rbroot_lock);
 248                 n = rb_first(rbroot);
 249                 if (!n) {
 250                         spin_unlock(rbroot_lock);
 251                         break;
 252                 }
 253
 254                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 255                 BUG_ON(info->bitmap); /* Logic error */
 256
 257                 if (info->offset > root->ino_cache_progress)
 258                         count = 0;
 259                 else
 260                         count = min(root->ino_cache_progress - info->offset + 1,
 261                                     info->bytes);
 262
 263                 rb_erase(&info->offset_index, rbroot);
 264                 spin_unlock(rbroot_lock);
 265                 if (count)
 266                         __btrfs_add_free_space(root->fs_info, ctl,
 267                                                info->offset, count);
 268                 kmem_cache_free(btrfs_free_space_cachep, info);
 269         }
 270 }
 271
 272 #define INIT_THRESHOLD  ((SZ_32K / 2) / sizeof(struct btrfs_free_space))
 273 #define INODES_PER_BITMAP (PAGE_SIZE * 8)
 274
 275 /*
 276  * The goal is to keep the memory used by the free_ino tree won't
 277  * exceed the memory if we use bitmaps only.
 278  */
 279 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 280 {
 281         struct btrfs_free_space *info;
 282         struct rb_node *n;
 283         int max_ino;
 284         int max_bitmaps;
 285
 286         n = rb_last(&ctl->free_space_offset);
 287         if (!n) {
 288                 ctl->extents_thresh = INIT_THRESHOLD;
 289                 return;
 290         }
 291         info = rb_entry(n, struct btrfs_free_space, offset_index);
 292
 293         /*
 294          * Find the maximum inode number in the filesystem. Note we
 295          * ignore the fact that this can be a bitmap, because we are
 296          * not doing precise calculation.
 297          */
 298         max_ino = info->bytes - 1;
 299
 300         max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
 301         if (max_bitmaps <= ctl->total_bitmaps) {
 302                 ctl->extents_thresh = 0;
 303                 return;
 304         }
 305
 306         ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
 307                                 PAGE_SIZE / sizeof(*info);
 308 }
 309
 310 /*
 311  * We don't fall back to bitmap, if we are below the extents threshold
 312  * or this chunk of inode numbers is a big one.
 313  */
 314 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
 315                        struct btrfs_free_space *info)
 316 {
 317         if (ctl->free_extents < ctl->extents_thresh ||
 318             info->bytes > INODES_PER_BITMAP / 10)
 319                 return false;
 320
 321         return true;
 322 }
 323
 324 static const struct btrfs_free_space_op free_ino_op = {
 325         .recalc_thresholds      = recalculate_thresholds,
 326         .use_bitmap             = use_bitmap,
 327 };
 328
 329 static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
 330 {
 331 }
 332
 333 static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
 334                               struct btrfs_free_space *info)
 335 {
 336         /*
 337          * We always use extents for two reasons:
 338          *
 339          * - The pinned tree is only used during the process of caching
 340          *   work.
 341          * - Make code simpler. See btrfs_unpin_free_ino().
 342          */
 343         return false;
 344 }
 345
 346 static const struct btrfs_free_space_op pinned_free_ino_op = {
 347         .recalc_thresholds      = pinned_recalc_thresholds,
 348         .use_bitmap             = pinned_use_bitmap,
 349 };
 350
 351 void btrfs_init_free_ino_ctl(struct btrfs_root *root)
 352 {
 353         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 354         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 355
 356         spin_lock_init(&ctl->tree_lock);
 357         ctl->unit = 1;
 358         ctl->start = 0;
 359         ctl->private = NULL;
 360         ctl->op = &free_ino_op;
 361         INIT_LIST_HEAD(&ctl->trimming_ranges);
 362         mutex_init(&ctl->cache_writeout_mutex);
 363
 364         /*
 365          * Initially we allow to use 16K of ram to cache chunks of
 366          * inode numbers before we resort to bitmaps. This is somewhat
 367          * arbitrary, but it will be adjusted in runtime.
 368          */
 369         ctl->extents_thresh = INIT_THRESHOLD;
 370
 371         spin_lock_init(&pinned->tree_lock);
 372         pinned->unit = 1;
 373         pinned->start = 0;
 374         pinned->private = NULL;
 375         pinned->extents_thresh = 0;
 376         pinned->op = &pinned_free_ino_op;
 377 }
 378
 379 int btrfs_save_ino_cache(struct btrfs_root *root,
 380                          struct btrfs_trans_handle *trans)
 381 {
 382         struct btrfs_fs_info *fs_info = root->fs_info;
 383         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 384         struct btrfs_path *path;
 385         struct inode *inode;
 386         struct btrfs_block_rsv *rsv;
 387         struct extent_changeset *data_reserved = NULL;
 388         u64 num_bytes;
 389         u64 alloc_hint = 0;
 390         int ret;
 391         int prealloc;
 392         bool retry = false;
 393
 394         /* only fs tree and subvol/snap needs ino cache */
 395         if (root->root_key.objectid != BTRFS_FS_TREE_OBJECTID &&
 396             (root->root_key.objectid < BTRFS_FIRST_FREE_OBJECTID ||
 397              root->root_key.objectid > BTRFS_LAST_FREE_OBJECTID))
 398                 return 0;
 399
 400         /* Don't save inode cache if we are deleting this root */
 401         if (btrfs_root_refs(&root->root_item) == 0)
 402                 return 0;
 403
 404         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 405                 return 0;
 406
 407         path = btrfs_alloc_path();
 408         if (!path)
 409                 return -ENOMEM;
 410
 411         rsv = trans->block_rsv;
 412         trans->block_rsv = &fs_info->trans_block_rsv;
 413
 414         num_bytes = trans->bytes_reserved;
 415         /*
 416          * 1 item for inode item insertion if need
 417          * 4 items for inode item update (in the worst case)
 418          * 1 items for slack space if we need do truncation
 419          * 1 item for free space object
 420          * 3 items for pre-allocation
 421          */
 422         trans->bytes_reserved = btrfs_calc_trans_metadata_size(fs_info, 10);
 423         ret = btrfs_block_rsv_add(root, trans->block_rsv,
 424                                   trans->bytes_reserved,
 425                                   BTRFS_RESERVE_NO_FLUSH);
 426         if (ret)
 427                 goto out;
 428         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 429                                       trans->bytes_reserved, 1);
 430 again:
 431         inode = lookup_free_ino_inode(root, path);
 432         if (IS_ERR(inode) && (PTR_ERR(inode) != -ENOENT || retry)) {
 433                 ret = PTR_ERR(inode);
 434                 goto out_release;
 435         }
 436
 437         if (IS_ERR(inode)) {
 438                 BUG_ON(retry); /* Logic error */
 439                 retry = true;
 440
 441                 ret = create_free_ino_inode(root, trans, path);
 442                 if (ret)
 443                         goto out_release;
 444                 goto again;
 445         }
 446
 447         BTRFS_I(inode)->generation = 0;
 448         ret = btrfs_update_inode(trans, root, inode);
 449         if (ret) {
 450                 btrfs_abort_transaction(trans, ret);
 451                 goto out_put;
 452         }
 453
 454         if (i_size_read(inode) > 0) {
 455                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
 456                 if (ret) {
 457                         if (ret != -ENOSPC)
 458                                 btrfs_abort_transaction(trans, ret);
 459                         goto out_put;
 460                 }
 461         }
 462
 463         spin_lock(&root->ino_cache_lock);
 464         if (root->ino_cache_state != BTRFS_CACHE_FINISHED) {
 465                 ret = -1;
 466                 spin_unlock(&root->ino_cache_lock);
 467                 goto out_put;
 468         }
 469         spin_unlock(&root->ino_cache_lock);
 470
 471         spin_lock(&ctl->tree_lock);
 472         prealloc = sizeof(struct btrfs_free_space) * ctl->free_extents;
 473         prealloc = ALIGN(prealloc, PAGE_SIZE);
 474         prealloc += ctl->total_bitmaps * PAGE_SIZE;
 475         spin_unlock(&ctl->tree_lock);
 476
 477         /* Just to make sure we have enough space */
 478         prealloc += 8 * PAGE_SIZE;
 479
 480         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, 0, prealloc);
 481         if (ret)
 482                 goto out_put;
 483
 484         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, prealloc,
 485                                               prealloc, prealloc, &alloc_hint);
 486         if (ret) {
 487                 btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc);
 488                 btrfs_delalloc_release_metadata(BTRFS_I(inode), prealloc, true);
 489                 goto out_put;
 490         }
 491
 492         ret = btrfs_write_out_ino_cache(root, trans, path, inode);
 493         btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc);
 494 out_put:
 495         iput(inode);
 496 out_release:
 497         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 498                                       trans->bytes_reserved, 0);
 499         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 500                                 trans->bytes_reserved);
 501 out:
 502         trans->block_rsv = rsv;
 503         trans->bytes_reserved = num_bytes;
 504
 505         btrfs_free_path(path);
 506         extent_changeset_free(data_reserved);
 507         return ret;
 508 }
 509
 510 int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
 511 {
 512         struct btrfs_path *path;
 513         int ret;
 514         struct extent_buffer *l;
 515         struct btrfs_key search_key;
 516         struct btrfs_key found_key;
 517         int slot;
 518
 519         path = btrfs_alloc_path();
 520         if (!path)
 521                 return -ENOMEM;
 522
 523         search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
 524         search_key.type = -1;
 525         search_key.offset = (u64)-1;
 526         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 527         if (ret < 0)
 528                 goto error;
 529         BUG_ON(ret == 0); /* Corruption */
 530         if (path->slots[0] > 0) {
 531                 slot = path->slots[0] - 1;
 532                 l = path->nodes[0];
 533                 btrfs_item_key_to_cpu(l, &found_key, slot);
 534                 *objectid = max_t(u64, found_key.objectid,
 535                                   BTRFS_FIRST_FREE_OBJECTID - 1);
 536         } else {
 537                 *objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
 538         }
 539         ret = 0;
 540 error:
 541         btrfs_free_path(path);
 542         return ret;
 543 }
 544
 545 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
 546 {
 547         int ret;
 548         mutex_lock(&root->objectid_mutex);
 549
 550         if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
 551                 btrfs_warn(root->fs_info,
 552                            "the objectid of root %llu reaches its highest value",
 553                            root->root_key.objectid);
 554                 ret = -ENOSPC;
 555                 goto out;
 556         }
 557
 558         *objectid = ++root->highest_objectid;
 559         ret = 0;
 560 out:
 561         mutex_unlock(&root->objectid_mutex);
 562         return ret;
 563 }