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 void fail_caching_thread(struct btrfs_root *root)
  16 {
  17         struct btrfs_fs_info *fs_info = root->fs_info;
  18
  19         btrfs_warn(fs_info, "failed to start inode caching task");
  20         btrfs_clear_pending_and_info(fs_info, INODE_MAP_CACHE,
  21                                      "disabling inode map caching");
  22         spin_lock(&root->ino_cache_lock);
  23         root->ino_cache_state = BTRFS_CACHE_ERROR;
  24         spin_unlock(&root->ino_cache_lock);
  25         wake_up(&root->ino_cache_wait);
  26 }
  27
  28 static int caching_kthread(void *data)
  29 {
  30         struct btrfs_root *root = data;
  31         struct btrfs_fs_info *fs_info = root->fs_info;
  32         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
  33         struct btrfs_key key;
  34         struct btrfs_path *path;
  35         struct extent_buffer *leaf;
  36         u64 last = (u64)-1;
  37         int slot;
  38         int ret;
  39
  40         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
  41                 return 0;
  42
  43         path = btrfs_alloc_path();
  44         if (!path) {
  45                 fail_caching_thread(root);
  46                 return -ENOMEM;
  47         }
  48
  49         /* Since the commit root is read-only, we can safely skip locking. */
  50         path->skip_locking = 1;
  51         path->search_commit_root = 1;
  52         path->reada = READA_FORWARD;
  53
  54         key.objectid = BTRFS_FIRST_FREE_OBJECTID;
  55         key.offset = 0;
  56         key.type = BTRFS_INODE_ITEM_KEY;
  57 again:
  58         /* need to make sure the commit_root doesn't disappear */
  59         down_read(&fs_info->commit_root_sem);
  60
  61         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  62         if (ret < 0)
  63                 goto out;
  64
  65         while (1) {
  66                 if (btrfs_fs_closing(fs_info))
  67                         goto out;
  68
  69                 leaf = path->nodes[0];
  70                 slot = path->slots[0];
  71                 if (slot >= btrfs_header_nritems(leaf)) {
  72                         ret = btrfs_next_leaf(root, path);
  73                         if (ret < 0)
  74                                 goto out;
  75                         else if (ret > 0)
  76                                 break;
  77
  78                         if (need_resched() ||
  79                             btrfs_transaction_in_commit(fs_info)) {
  80                                 leaf = path->nodes[0];
  81
  82                                 if (WARN_ON(btrfs_header_nritems(leaf) == 0))
  83                                         break;
  84
  85                                 /*
  86                                  * Save the key so we can advances forward
  87                                  * in the next search.
  88                                  */
  89                                 btrfs_item_key_to_cpu(leaf, &key, 0);
  90                                 btrfs_release_path(path);
  91                                 root->ino_cache_progress = last;
  92                                 up_read(&fs_info->commit_root_sem);
  93                                 schedule_timeout(1);
  94                                 goto again;
  95                         } else
  96                                 continue;
  97                 }
  98
  99                 btrfs_item_key_to_cpu(leaf, &key, slot);
 100
 101                 if (key.type != BTRFS_INODE_ITEM_KEY)
 102                         goto next;
 103
 104                 if (key.objectid >= root->highest_objectid)
 105                         break;
 106
 107                 if (last != (u64)-1 && last + 1 != key.objectid) {
 108                         __btrfs_add_free_space(fs_info, ctl, last + 1,
 109                                                key.objectid - last - 1);
 110                         wake_up(&root->ino_cache_wait);
 111                 }
 112
 113                 last = key.objectid;
 114 next:
 115                 path->slots[0]++;
 116         }
 117
 118         if (last < root->highest_objectid - 1) {
 119                 __btrfs_add_free_space(fs_info, ctl, last + 1,
 120                                        root->highest_objectid - last - 1);
 121         }
 122
 123         spin_lock(&root->ino_cache_lock);
 124         root->ino_cache_state = BTRFS_CACHE_FINISHED;
 125         spin_unlock(&root->ino_cache_lock);
 126
 127         root->ino_cache_progress = (u64)-1;
 128         btrfs_unpin_free_ino(root);
 129 out:
 130         wake_up(&root->ino_cache_wait);
 131         up_read(&fs_info->commit_root_sem);
 132
 133         btrfs_free_path(path);
 134
 135         return ret;
 136 }
 137
 138 static void start_caching(struct btrfs_root *root)
 139 {
 140         struct btrfs_fs_info *fs_info = root->fs_info;
 141         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 142         struct task_struct *tsk;
 143         int ret;
 144         u64 objectid;
 145
 146         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 147                 return;
 148
 149         spin_lock(&root->ino_cache_lock);
 150         if (root->ino_cache_state != BTRFS_CACHE_NO) {
 151                 spin_unlock(&root->ino_cache_lock);
 152                 return;
 153         }
 154
 155         root->ino_cache_state = BTRFS_CACHE_STARTED;
 156         spin_unlock(&root->ino_cache_lock);
 157
 158         ret = load_free_ino_cache(fs_info, root);
 159         if (ret == 1) {
 160                 spin_lock(&root->ino_cache_lock);
 161                 root->ino_cache_state = BTRFS_CACHE_FINISHED;
 162                 spin_unlock(&root->ino_cache_lock);
 163                 wake_up(&root->ino_cache_wait);
 164                 return;
 165         }
 166
 167         /*
 168          * It can be quite time-consuming to fill the cache by searching
 169          * through the extent tree, and this can keep ino allocation path
 170          * waiting. Therefore at start we quickly find out the highest
 171          * inode number and we know we can use inode numbers which fall in
 172          * [highest_ino + 1, BTRFS_LAST_FREE_OBJECTID].
 173          */
 174         ret = btrfs_find_free_objectid(root, &objectid);
 175         if (!ret && objectid <= BTRFS_LAST_FREE_OBJECTID) {
 176                 __btrfs_add_free_space(fs_info, ctl, objectid,
 177                                        BTRFS_LAST_FREE_OBJECTID - objectid + 1);
 178         }
 179
 180         tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu",
 181                           root->root_key.objectid);
 182         if (IS_ERR(tsk))
 183                 fail_caching_thread(root);
 184 }
 185
 186 int btrfs_find_free_ino(struct btrfs_root *root, u64 *objectid)
 187 {
 188         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 189                 return btrfs_find_free_objectid(root, objectid);
 190
 191 again:
 192         *objectid = btrfs_find_ino_for_alloc(root);
 193
 194         if (*objectid != 0)
 195                 return 0;
 196
 197         start_caching(root);
 198
 199         wait_event(root->ino_cache_wait,
 200                    root->ino_cache_state == BTRFS_CACHE_FINISHED ||
 201                    root->ino_cache_state == BTRFS_CACHE_ERROR ||
 202                    root->free_ino_ctl->free_space > 0);
 203
 204         if (root->ino_cache_state == BTRFS_CACHE_FINISHED &&
 205             root->free_ino_ctl->free_space == 0)
 206                 return -ENOSPC;
 207         else if (root->ino_cache_state == BTRFS_CACHE_ERROR)
 208                 return btrfs_find_free_objectid(root, objectid);
 209         else
 210                 goto again;
 211 }
 212
 213 void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
 214 {
 215         struct btrfs_fs_info *fs_info = root->fs_info;
 216         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 217
 218         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 219                 return;
 220 again:
 221         if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 222                 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
 223         } else {
 224                 down_write(&fs_info->commit_root_sem);
 225                 spin_lock(&root->ino_cache_lock);
 226                 if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 227                         spin_unlock(&root->ino_cache_lock);
 228                         up_write(&fs_info->commit_root_sem);
 229                         goto again;
 230                 }
 231                 spin_unlock(&root->ino_cache_lock);
 232
 233                 start_caching(root);
 234
 235                 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
 236
 237                 up_write(&fs_info->commit_root_sem);
 238         }
 239 }
 240
 241 /*
 242  * When a transaction is committed, we'll move those inode numbers which are
 243  * smaller than root->ino_cache_progress from pinned tree to free_ino tree, and
 244  * others will just be dropped, because the commit root we were searching has
 245  * changed.
 246  *
 247  * Must be called with root->fs_info->commit_root_sem held
 248  */
 249 void btrfs_unpin_free_ino(struct btrfs_root *root)
 250 {
 251         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 252         struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
 253         spinlock_t *rbroot_lock = &root->free_ino_pinned->tree_lock;
 254         struct btrfs_free_space *info;
 255         struct rb_node *n;
 256         u64 count;
 257
 258         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 259                 return;
 260
 261         while (1) {
 262                 spin_lock(rbroot_lock);
 263                 n = rb_first(rbroot);
 264                 if (!n) {
 265                         spin_unlock(rbroot_lock);
 266                         break;
 267                 }
 268
 269                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 270                 BUG_ON(info->bitmap); /* Logic error */
 271
 272                 if (info->offset > root->ino_cache_progress)
 273                         count = 0;
 274                 else
 275                         count = min(root->ino_cache_progress - info->offset + 1,
 276                                     info->bytes);
 277
 278                 rb_erase(&info->offset_index, rbroot);
 279                 spin_unlock(rbroot_lock);
 280                 if (count)
 281                         __btrfs_add_free_space(root->fs_info, ctl,
 282                                                info->offset, count);
 283                 kmem_cache_free(btrfs_free_space_cachep, info);
 284         }
 285 }
 286
 287 #define INIT_THRESHOLD  ((SZ_32K / 2) / sizeof(struct btrfs_free_space))
 288 #define INODES_PER_BITMAP (PAGE_SIZE * 8)
 289
 290 /*
 291  * The goal is to keep the memory used by the free_ino tree won't
 292  * exceed the memory if we use bitmaps only.
 293  */
 294 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 295 {
 296         struct btrfs_free_space *info;
 297         struct rb_node *n;
 298         int max_ino;
 299         int max_bitmaps;
 300
 301         n = rb_last(&ctl->free_space_offset);
 302         if (!n) {
 303                 ctl->extents_thresh = INIT_THRESHOLD;
 304                 return;
 305         }
 306         info = rb_entry(n, struct btrfs_free_space, offset_index);
 307
 308         /*
 309          * Find the maximum inode number in the filesystem. Note we
 310          * ignore the fact that this can be a bitmap, because we are
 311          * not doing precise calculation.
 312          */
 313         max_ino = info->bytes - 1;
 314
 315         max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
 316         if (max_bitmaps <= ctl->total_bitmaps) {
 317                 ctl->extents_thresh = 0;
 318                 return;
 319         }
 320
 321         ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
 322                                 PAGE_SIZE / sizeof(*info);
 323 }
 324
 325 /*
 326  * We don't fall back to bitmap, if we are below the extents threshold
 327  * or this chunk of inode numbers is a big one.
 328  */
 329 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
 330                        struct btrfs_free_space *info)
 331 {
 332         if (ctl->free_extents < ctl->extents_thresh ||
 333             info->bytes > INODES_PER_BITMAP / 10)
 334                 return false;
 335
 336         return true;
 337 }
 338
 339 static const struct btrfs_free_space_op free_ino_op = {
 340         .recalc_thresholds      = recalculate_thresholds,
 341         .use_bitmap             = use_bitmap,
 342 };
 343
 344 static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
 345 {
 346 }
 347
 348 static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
 349                               struct btrfs_free_space *info)
 350 {
 351         /*
 352          * We always use extents for two reasons:
 353          *
 354          * - The pinned tree is only used during the process of caching
 355          *   work.
 356          * - Make code simpler. See btrfs_unpin_free_ino().
 357          */
 358         return false;
 359 }
 360
 361 static const struct btrfs_free_space_op pinned_free_ino_op = {
 362         .recalc_thresholds      = pinned_recalc_thresholds,
 363         .use_bitmap             = pinned_use_bitmap,
 364 };
 365
 366 void btrfs_init_free_ino_ctl(struct btrfs_root *root)
 367 {
 368         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 369         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 370
 371         spin_lock_init(&ctl->tree_lock);
 372         ctl->unit = 1;
 373         ctl->start = 0;
 374         ctl->private = NULL;
 375         ctl->op = &free_ino_op;
 376         INIT_LIST_HEAD(&ctl->trimming_ranges);
 377         mutex_init(&ctl->cache_writeout_mutex);
 378
 379         /*
 380          * Initially we allow to use 16K of ram to cache chunks of
 381          * inode numbers before we resort to bitmaps. This is somewhat
 382          * arbitrary, but it will be adjusted in runtime.
 383          */
 384         ctl->extents_thresh = INIT_THRESHOLD;
 385
 386         spin_lock_init(&pinned->tree_lock);
 387         pinned->unit = 1;
 388         pinned->start = 0;
 389         pinned->private = NULL;
 390         pinned->extents_thresh = 0;
 391         pinned->op = &pinned_free_ino_op;
 392 }
 393
 394 int btrfs_save_ino_cache(struct btrfs_root *root,
 395                          struct btrfs_trans_handle *trans)
 396 {
 397         struct btrfs_fs_info *fs_info = root->fs_info;
 398         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 399         struct btrfs_path *path;
 400         struct inode *inode;
 401         struct btrfs_block_rsv *rsv;
 402         struct extent_changeset *data_reserved = NULL;
 403         u64 num_bytes;
 404         u64 alloc_hint = 0;
 405         int ret;
 406         int prealloc;
 407         bool retry = false;
 408
 409         /* only fs tree and subvol/snap needs ino cache */
 410         if (root->root_key.objectid != BTRFS_FS_TREE_OBJECTID &&
 411             (root->root_key.objectid < BTRFS_FIRST_FREE_OBJECTID ||
 412              root->root_key.objectid > BTRFS_LAST_FREE_OBJECTID))
 413                 return 0;
 414
 415         /* Don't save inode cache if we are deleting this root */
 416         if (btrfs_root_refs(&root->root_item) == 0)
 417                 return 0;
 418
 419         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 420                 return 0;
 421
 422         path = btrfs_alloc_path();
 423         if (!path)
 424                 return -ENOMEM;
 425
 426         rsv = trans->block_rsv;
 427         trans->block_rsv = &fs_info->trans_block_rsv;
 428
 429         num_bytes = trans->bytes_reserved;
 430         /*
 431          * 1 item for inode item insertion if need
 432          * 4 items for inode item update (in the worst case)
 433          * 1 items for slack space if we need do truncation
 434          * 1 item for free space object
 435          * 3 items for pre-allocation
 436          */
 437         trans->bytes_reserved = btrfs_calc_trans_metadata_size(fs_info, 10);
 438         ret = btrfs_block_rsv_add(root, trans->block_rsv,
 439                                   trans->bytes_reserved,
 440                                   BTRFS_RESERVE_NO_FLUSH);
 441         if (ret)
 442                 goto out;
 443         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 444                                       trans->bytes_reserved, 1);
 445 again:
 446         inode = lookup_free_ino_inode(root, path);
 447         if (IS_ERR(inode) && (PTR_ERR(inode) != -ENOENT || retry)) {
 448                 ret = PTR_ERR(inode);
 449                 goto out_release;
 450         }
 451
 452         if (IS_ERR(inode)) {
 453                 BUG_ON(retry); /* Logic error */
 454                 retry = true;
 455
 456                 ret = create_free_ino_inode(root, trans, path);
 457                 if (ret)
 458                         goto out_release;
 459                 goto again;
 460         }
 461
 462         BTRFS_I(inode)->generation = 0;
 463         ret = btrfs_update_inode(trans, root, inode);
 464         if (ret) {
 465                 btrfs_abort_transaction(trans, ret);
 466                 goto out_put;
 467         }
 468
 469         if (i_size_read(inode) > 0) {
 470                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
 471                 if (ret) {
 472                         if (ret != -ENOSPC)
 473                                 btrfs_abort_transaction(trans, ret);
 474                         goto out_put;
 475                 }
 476         }
 477
 478         spin_lock(&root->ino_cache_lock);
 479         if (root->ino_cache_state != BTRFS_CACHE_FINISHED) {
 480                 ret = -1;
 481                 spin_unlock(&root->ino_cache_lock);
 482                 goto out_put;
 483         }
 484         spin_unlock(&root->ino_cache_lock);
 485
 486         spin_lock(&ctl->tree_lock);
 487         prealloc = sizeof(struct btrfs_free_space) * ctl->free_extents;
 488         prealloc = ALIGN(prealloc, PAGE_SIZE);
 489         prealloc += ctl->total_bitmaps * PAGE_SIZE;
 490         spin_unlock(&ctl->tree_lock);
 491
 492         /* Just to make sure we have enough space */
 493         prealloc += 8 * PAGE_SIZE;
 494
 495         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, 0, prealloc);
 496         if (ret)
 497                 goto out_put;
 498
 499         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, prealloc,
 500                                               prealloc, prealloc, &alloc_hint);
 501         if (ret) {
 502                 btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc);
 503                 btrfs_delalloc_release_metadata(BTRFS_I(inode), prealloc, true);
 504                 goto out_put;
 505         }
 506
 507         ret = btrfs_write_out_ino_cache(root, trans, path, inode);
 508         btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc);
 509 out_put:
 510         iput(inode);
 511 out_release:
 512         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 513                                       trans->bytes_reserved, 0);
 514         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 515                                 trans->bytes_reserved);
 516 out:
 517         trans->block_rsv = rsv;
 518         trans->bytes_reserved = num_bytes;
 519
 520         btrfs_free_path(path);
 521         extent_changeset_free(data_reserved);
 522         return ret;
 523 }
 524
 525 int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
 526 {
 527         struct btrfs_path *path;
 528         int ret;
 529         struct extent_buffer *l;
 530         struct btrfs_key search_key;
 531         struct btrfs_key found_key;
 532         int slot;
 533
 534         path = btrfs_alloc_path();
 535         if (!path)
 536                 return -ENOMEM;
 537
 538         search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
 539         search_key.type = -1;
 540         search_key.offset = (u64)-1;
 541         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 542         if (ret < 0)
 543                 goto error;
 544         BUG_ON(ret == 0); /* Corruption */
 545         if (path->slots[0] > 0) {
 546                 slot = path->slots[0] - 1;
 547                 l = path->nodes[0];
 548                 btrfs_item_key_to_cpu(l, &found_key, slot);
 549                 *objectid = max_t(u64, found_key.objectid,
 550                                   BTRFS_FIRST_FREE_OBJECTID - 1);
 551         } else {
 552                 *objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
 553         }
 554         ret = 0;
 555 error:
 556         btrfs_free_path(path);
 557         return ret;
 558 }
 559
 560 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
 561 {
 562         int ret;
 563         mutex_lock(&root->objectid_mutex);
 564
 565         if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
 566                 btrfs_warn(root->fs_info,
 567                            "the objectid of root %llu reaches its highest value",
 568                            root->root_key.objectid);
 569                 ret = -ENOSPC;
 570                 goto out;
 571         }
 572
 573         *objectid = ++root->highest_objectid;
 574         ret = 0;
 575 out:
 576         mutex_unlock(&root->objectid_mutex);
 577         return ret;
 578 }