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/delay.h>
   7 #include <linux/kthread.h>
   8 #include <linux/pagemap.h>
   9
  10 #include "ctree.h"
  11 #include "disk-io.h"
  12 #include "free-space-cache.h"
  13 #include "inode-map.h"
  14 #include "transaction.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                 bool add_to_ctl = true;
 248
 249                 spin_lock(rbroot_lock);
 250                 n = rb_first(rbroot);
 251                 if (!n) {
 252                         spin_unlock(rbroot_lock);
 253                         break;
 254                 }
 255
 256                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 257                 BUG_ON(info->bitmap); /* Logic error */
 258
 259                 if (info->offset > root->ino_cache_progress)
 260                         add_to_ctl = false;
 261                 else if (info->offset + info->bytes > root->ino_cache_progress)
 262                         count = root->ino_cache_progress - info->offset + 1;
 263                 else
 264                         count = info->bytes;
 265
 266                 rb_erase(&info->offset_index, rbroot);
 267                 spin_unlock(rbroot_lock);
 268                 if (add_to_ctl)
 269                         __btrfs_add_free_space(root->fs_info, ctl,
 270                                                info->offset, count);
 271                 kmem_cache_free(btrfs_free_space_cachep, info);
 272         }
 273 }
 274
 275 #define INIT_THRESHOLD  ((SZ_32K / 2) / sizeof(struct btrfs_free_space))
 276 #define INODES_PER_BITMAP (PAGE_SIZE * 8)
 277
 278 /*
 279  * The goal is to keep the memory used by the free_ino tree won't
 280  * exceed the memory if we use bitmaps only.
 281  */
 282 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 283 {
 284         struct btrfs_free_space *info;
 285         struct rb_node *n;
 286         int max_ino;
 287         int max_bitmaps;
 288
 289         n = rb_last(&ctl->free_space_offset);
 290         if (!n) {
 291                 ctl->extents_thresh = INIT_THRESHOLD;
 292                 return;
 293         }
 294         info = rb_entry(n, struct btrfs_free_space, offset_index);
 295
 296         /*
 297          * Find the maximum inode number in the filesystem. Note we
 298          * ignore the fact that this can be a bitmap, because we are
 299          * not doing precise calculation.
 300          */
 301         max_ino = info->bytes - 1;
 302
 303         max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
 304         if (max_bitmaps <= ctl->total_bitmaps) {
 305                 ctl->extents_thresh = 0;
 306                 return;
 307         }
 308
 309         ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
 310                                 PAGE_SIZE / sizeof(*info);
 311 }
 312
 313 /*
 314  * We don't fall back to bitmap, if we are below the extents threshold
 315  * or this chunk of inode numbers is a big one.
 316  */
 317 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
 318                        struct btrfs_free_space *info)
 319 {
 320         if (ctl->free_extents < ctl->extents_thresh ||
 321             info->bytes > INODES_PER_BITMAP / 10)
 322                 return false;
 323
 324         return true;
 325 }
 326
 327 static const struct btrfs_free_space_op free_ino_op = {
 328         .recalc_thresholds      = recalculate_thresholds,
 329         .use_bitmap             = use_bitmap,
 330 };
 331
 332 static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
 333 {
 334 }
 335
 336 static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
 337                               struct btrfs_free_space *info)
 338 {
 339         /*
 340          * We always use extents for two reasons:
 341          *
 342          * - The pinned tree is only used during the process of caching
 343          *   work.
 344          * - Make code simpler. See btrfs_unpin_free_ino().
 345          */
 346         return false;
 347 }
 348
 349 static const struct btrfs_free_space_op pinned_free_ino_op = {
 350         .recalc_thresholds      = pinned_recalc_thresholds,
 351         .use_bitmap             = pinned_use_bitmap,
 352 };
 353
 354 void btrfs_init_free_ino_ctl(struct btrfs_root *root)
 355 {
 356         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 357         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 358
 359         spin_lock_init(&ctl->tree_lock);
 360         ctl->unit = 1;
 361         ctl->start = 0;
 362         ctl->private = NULL;
 363         ctl->op = &free_ino_op;
 364         INIT_LIST_HEAD(&ctl->trimming_ranges);
 365         mutex_init(&ctl->cache_writeout_mutex);
 366
 367         /*
 368          * Initially we allow to use 16K of ram to cache chunks of
 369          * inode numbers before we resort to bitmaps. This is somewhat
 370          * arbitrary, but it will be adjusted in runtime.
 371          */
 372         ctl->extents_thresh = INIT_THRESHOLD;
 373
 374         spin_lock_init(&pinned->tree_lock);
 375         pinned->unit = 1;
 376         pinned->start = 0;
 377         pinned->private = NULL;
 378         pinned->extents_thresh = 0;
 379         pinned->op = &pinned_free_ino_op;
 380 }
 381
 382 int btrfs_save_ino_cache(struct btrfs_root *root,
 383                          struct btrfs_trans_handle *trans)
 384 {
 385         struct btrfs_fs_info *fs_info = root->fs_info;
 386         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 387         struct btrfs_path *path;
 388         struct inode *inode;
 389         struct btrfs_block_rsv *rsv;
 390         struct extent_changeset *data_reserved = NULL;
 391         u64 num_bytes;
 392         u64 alloc_hint = 0;
 393         int ret;
 394         int prealloc;
 395         bool retry = false;
 396
 397         /* only fs tree and subvol/snap needs ino cache */
 398         if (root->root_key.objectid != BTRFS_FS_TREE_OBJECTID &&
 399             (root->root_key.objectid < BTRFS_FIRST_FREE_OBJECTID ||
 400              root->root_key.objectid > BTRFS_LAST_FREE_OBJECTID))
 401                 return 0;
 402
 403         /* Don't save inode cache if we are deleting this root */
 404         if (btrfs_root_refs(&root->root_item) == 0)
 405                 return 0;
 406
 407         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 408                 return 0;
 409
 410         path = btrfs_alloc_path();
 411         if (!path)
 412                 return -ENOMEM;
 413
 414         rsv = trans->block_rsv;
 415         trans->block_rsv = &fs_info->trans_block_rsv;
 416
 417         num_bytes = trans->bytes_reserved;
 418         /*
 419          * 1 item for inode item insertion if need
 420          * 4 items for inode item update (in the worst case)
 421          * 1 items for slack space if we need do truncation
 422          * 1 item for free space object
 423          * 3 items for pre-allocation
 424          */
 425         trans->bytes_reserved = btrfs_calc_trans_metadata_size(fs_info, 10);
 426         ret = btrfs_block_rsv_add(root, trans->block_rsv,
 427                                   trans->bytes_reserved,
 428                                   BTRFS_RESERVE_NO_FLUSH);
 429         if (ret)
 430                 goto out;
 431         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 432                                       trans->bytes_reserved, 1);
 433 again:
 434         inode = lookup_free_ino_inode(root, path);
 435         if (IS_ERR(inode) && (PTR_ERR(inode) != -ENOENT || retry)) {
 436                 ret = PTR_ERR(inode);
 437                 goto out_release;
 438         }
 439
 440         if (IS_ERR(inode)) {
 441                 BUG_ON(retry); /* Logic error */
 442                 retry = true;
 443
 444                 ret = create_free_ino_inode(root, trans, path);
 445                 if (ret)
 446                         goto out_release;
 447                 goto again;
 448         }
 449
 450         BTRFS_I(inode)->generation = 0;
 451         ret = btrfs_update_inode(trans, root, inode);
 452         if (ret) {
 453                 btrfs_abort_transaction(trans, ret);
 454                 goto out_put;
 455         }
 456
 457         if (i_size_read(inode) > 0) {
 458                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
 459                 if (ret) {
 460                         if (ret != -ENOSPC)
 461                                 btrfs_abort_transaction(trans, ret);
 462                         goto out_put;
 463                 }
 464         }
 465
 466         spin_lock(&root->ino_cache_lock);
 467         if (root->ino_cache_state != BTRFS_CACHE_FINISHED) {
 468                 ret = -1;
 469                 spin_unlock(&root->ino_cache_lock);
 470                 goto out_put;
 471         }
 472         spin_unlock(&root->ino_cache_lock);
 473
 474         spin_lock(&ctl->tree_lock);
 475         prealloc = sizeof(struct btrfs_free_space) * ctl->free_extents;
 476         prealloc = ALIGN(prealloc, PAGE_SIZE);
 477         prealloc += ctl->total_bitmaps * PAGE_SIZE;
 478         spin_unlock(&ctl->tree_lock);
 479
 480         /* Just to make sure we have enough space */
 481         prealloc += 8 * PAGE_SIZE;
 482
 483         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, 0, prealloc);
 484         if (ret)
 485                 goto out_put;
 486
 487         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, prealloc,
 488                                               prealloc, prealloc, &alloc_hint);
 489         if (ret) {
 490                 btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc, true);
 491                 goto out_put;
 492         }
 493
 494         ret = btrfs_write_out_ino_cache(root, trans, path, inode);
 495         btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc, false);
 496 out_put:
 497         iput(inode);
 498 out_release:
 499         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 500                                       trans->bytes_reserved, 0);
 501         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 502                                 trans->bytes_reserved);
 503 out:
 504         trans->block_rsv = rsv;
 505         trans->bytes_reserved = num_bytes;
 506
 507         btrfs_free_path(path);
 508         extent_changeset_free(data_reserved);
 509         return ret;
 510 }
 511
 512 int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
 513 {
 514         struct btrfs_path *path;
 515         int ret;
 516         struct extent_buffer *l;
 517         struct btrfs_key search_key;
 518         struct btrfs_key found_key;
 519         int slot;
 520
 521         path = btrfs_alloc_path();
 522         if (!path)
 523                 return -ENOMEM;
 524
 525         search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
 526         search_key.type = -1;
 527         search_key.offset = (u64)-1;
 528         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 529         if (ret < 0)
 530                 goto error;
 531         BUG_ON(ret == 0); /* Corruption */
 532         if (path->slots[0] > 0) {
 533                 slot = path->slots[0] - 1;
 534                 l = path->nodes[0];
 535                 btrfs_item_key_to_cpu(l, &found_key, slot);
 536                 *objectid = max_t(u64, found_key.objectid,
 537                                   BTRFS_FIRST_FREE_OBJECTID - 1);
 538         } else {
 539                 *objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
 540         }
 541         ret = 0;
 542 error:
 543         btrfs_free_path(path);
 544         return ret;
 545 }
 546
 547 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
 548 {
 549         int ret;
 550         mutex_lock(&root->objectid_mutex);
 551
 552         if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
 553                 btrfs_warn(root->fs_info,
 554                            "the objectid of root %llu reaches its highest value",
 555                            root->root_key.objectid);
 556                 ret = -ENOSPC;
 557                 goto out;
 558         }
 559
 560         *objectid = ++root->highest_objectid;
 561         ret = 0;
 562 out:
 563         mutex_unlock(&root->objectid_mutex);
 564         return ret;
 565 }