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