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 void fail_caching_thread(struct btrfs_root *root)
  17 {
  18         struct btrfs_fs_info *fs_info = root->fs_info;
  19
  20         btrfs_warn(fs_info, "failed to start inode caching task");
  21         btrfs_clear_pending_and_info(fs_info, INODE_MAP_CACHE,
  22                                      "disabling inode map caching");
  23         spin_lock(&root->ino_cache_lock);
  24         root->ino_cache_state = BTRFS_CACHE_ERROR;
  25         spin_unlock(&root->ino_cache_lock);
  26         wake_up(&root->ino_cache_wait);
  27 }
  28
  29 static int caching_kthread(void *data)
  30 {
  31         struct btrfs_root *root = data;
  32         struct btrfs_fs_info *fs_info = root->fs_info;
  33         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
  34         struct btrfs_key key;
  35         struct btrfs_path *path;
  36         struct extent_buffer *leaf;
  37         u64 last = (u64)-1;
  38         int slot;
  39         int ret;
  40
  41         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
  42                 return 0;
  43
  44         path = btrfs_alloc_path();
  45         if (!path) {
  46                 fail_caching_thread(root);
  47                 return -ENOMEM;
  48         }
  49
  50         /* Since the commit root is read-only, we can safely skip locking. */
  51         path->skip_locking = 1;
  52         path->search_commit_root = 1;
  53         path->reada = READA_FORWARD;
  54
  55         key.objectid = BTRFS_FIRST_FREE_OBJECTID;
  56         key.offset = 0;
  57         key.type = BTRFS_INODE_ITEM_KEY;
  58 again:
  59         /* need to make sure the commit_root doesn't disappear */
  60         down_read(&fs_info->commit_root_sem);
  61
  62         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  63         if (ret < 0)
  64                 goto out;
  65
  66         while (1) {
  67                 if (btrfs_fs_closing(fs_info))
  68                         goto out;
  69
  70                 leaf = path->nodes[0];
  71                 slot = path->slots[0];
  72                 if (slot >= btrfs_header_nritems(leaf)) {
  73                         ret = btrfs_next_leaf(root, path);
  74                         if (ret < 0)
  75                                 goto out;
  76                         else if (ret > 0)
  77                                 break;
  78
  79                         if (need_resched() ||
  80                             btrfs_transaction_in_commit(fs_info)) {
  81                                 leaf = path->nodes[0];
  82
  83                                 if (WARN_ON(btrfs_header_nritems(leaf) == 0))
  84                                         break;
  85
  86                                 /*
  87                                  * Save the key so we can advances forward
  88                                  * in the next search.
  89                                  */
  90                                 btrfs_item_key_to_cpu(leaf, &key, 0);
  91                                 btrfs_release_path(path);
  92                                 root->ino_cache_progress = last;
  93                                 up_read(&fs_info->commit_root_sem);
  94                                 schedule_timeout(1);
  95                                 goto again;
  96                         } else
  97                                 continue;
  98                 }
  99
 100                 btrfs_item_key_to_cpu(leaf, &key, slot);
 101
 102                 if (key.type != BTRFS_INODE_ITEM_KEY)
 103                         goto next;
 104
 105                 if (key.objectid >= root->highest_objectid)
 106                         break;
 107
 108                 if (last != (u64)-1 && last + 1 != key.objectid) {
 109                         __btrfs_add_free_space(fs_info, ctl, last + 1,
 110                                                key.objectid - last - 1);
 111                         wake_up(&root->ino_cache_wait);
 112                 }
 113
 114                 last = key.objectid;
 115 next:
 116                 path->slots[0]++;
 117         }
 118
 119         if (last < root->highest_objectid - 1) {
 120                 __btrfs_add_free_space(fs_info, ctl, last + 1,
 121                                        root->highest_objectid - last - 1);
 122         }
 123
 124         spin_lock(&root->ino_cache_lock);
 125         root->ino_cache_state = BTRFS_CACHE_FINISHED;
 126         spin_unlock(&root->ino_cache_lock);
 127
 128         root->ino_cache_progress = (u64)-1;
 129         btrfs_unpin_free_ino(root);
 130 out:
 131         wake_up(&root->ino_cache_wait);
 132         up_read(&fs_info->commit_root_sem);
 133
 134         btrfs_free_path(path);
 135
 136         return ret;
 137 }
 138
 139 static void start_caching(struct btrfs_root *root)
 140 {
 141         struct btrfs_fs_info *fs_info = root->fs_info;
 142         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 143         struct task_struct *tsk;
 144         int ret;
 145         u64 objectid;
 146
 147         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 148                 return;
 149
 150         spin_lock(&root->ino_cache_lock);
 151         if (root->ino_cache_state != BTRFS_CACHE_NO) {
 152                 spin_unlock(&root->ino_cache_lock);
 153                 return;
 154         }
 155
 156         root->ino_cache_state = BTRFS_CACHE_STARTED;
 157         spin_unlock(&root->ino_cache_lock);
 158
 159         ret = load_free_ino_cache(fs_info, root);
 160         if (ret == 1) {
 161                 spin_lock(&root->ino_cache_lock);
 162                 root->ino_cache_state = BTRFS_CACHE_FINISHED;
 163                 spin_unlock(&root->ino_cache_lock);
 164                 wake_up(&root->ino_cache_wait);
 165                 return;
 166         }
 167
 168         /*
 169          * It can be quite time-consuming to fill the cache by searching
 170          * through the extent tree, and this can keep ino allocation path
 171          * waiting. Therefore at start we quickly find out the highest
 172          * inode number and we know we can use inode numbers which fall in
 173          * [highest_ino + 1, BTRFS_LAST_FREE_OBJECTID].
 174          */
 175         ret = btrfs_find_free_objectid(root, &objectid);
 176         if (!ret && objectid <= BTRFS_LAST_FREE_OBJECTID) {
 177                 __btrfs_add_free_space(fs_info, ctl, objectid,
 178                                        BTRFS_LAST_FREE_OBJECTID - objectid + 1);
 179                 wake_up(&root->ino_cache_wait);
 180         }
 181
 182         tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu",
 183                           root->root_key.objectid);
 184         if (IS_ERR(tsk))
 185                 fail_caching_thread(root);
 186 }
 187
 188 int btrfs_find_free_ino(struct btrfs_root *root, u64 *objectid)
 189 {
 190         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 191                 return btrfs_find_free_objectid(root, objectid);
 192
 193 again:
 194         *objectid = btrfs_find_ino_for_alloc(root);
 195
 196         if (*objectid != 0)
 197                 return 0;
 198
 199         start_caching(root);
 200
 201         wait_event(root->ino_cache_wait,
 202                    root->ino_cache_state == BTRFS_CACHE_FINISHED ||
 203                    root->ino_cache_state == BTRFS_CACHE_ERROR ||
 204                    root->free_ino_ctl->free_space > 0);
 205
 206         if (root->ino_cache_state == BTRFS_CACHE_FINISHED &&
 207             root->free_ino_ctl->free_space == 0)
 208                 return -ENOSPC;
 209         else if (root->ino_cache_state == BTRFS_CACHE_ERROR)
 210                 return btrfs_find_free_objectid(root, objectid);
 211         else
 212                 goto again;
 213 }
 214
 215 void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
 216 {
 217         struct btrfs_fs_info *fs_info = root->fs_info;
 218         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 219
 220         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 221                 return;
 222 again:
 223         if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 224                 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
 225         } else {
 226                 down_write(&fs_info->commit_root_sem);
 227                 spin_lock(&root->ino_cache_lock);
 228                 if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 229                         spin_unlock(&root->ino_cache_lock);
 230                         up_write(&fs_info->commit_root_sem);
 231                         goto again;
 232                 }
 233                 spin_unlock(&root->ino_cache_lock);
 234
 235                 start_caching(root);
 236
 237                 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
 238
 239                 up_write(&fs_info->commit_root_sem);
 240         }
 241 }
 242
 243 /*
 244  * When a transaction is committed, we'll move those inode numbers which are
 245  * smaller than root->ino_cache_progress from pinned tree to free_ino tree, and
 246  * others will just be dropped, because the commit root we were searching has
 247  * changed.
 248  *
 249  * Must be called with root->fs_info->commit_root_sem held
 250  */
 251 void btrfs_unpin_free_ino(struct btrfs_root *root)
 252 {
 253         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 254         struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
 255         spinlock_t *rbroot_lock = &root->free_ino_pinned->tree_lock;
 256         struct btrfs_free_space *info;
 257         struct rb_node *n;
 258         u64 count;
 259
 260         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 261                 return;
 262
 263         while (1) {
 264                 spin_lock(rbroot_lock);
 265                 n = rb_first(rbroot);
 266                 if (!n) {
 267                         spin_unlock(rbroot_lock);
 268                         break;
 269                 }
 270
 271                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 272                 BUG_ON(info->bitmap); /* Logic error */
 273
 274                 if (info->offset > root->ino_cache_progress)
 275                         count = 0;
 276                 else
 277                         count = min(root->ino_cache_progress - info->offset + 1,
 278                                     info->bytes);
 279
 280                 rb_erase(&info->offset_index, rbroot);
 281                 spin_unlock(rbroot_lock);
 282                 if (count)
 283                         __btrfs_add_free_space(root->fs_info, ctl,
 284                                                info->offset, count);
 285                 kmem_cache_free(btrfs_free_space_cachep, info);
 286         }
 287 }
 288
 289 #define INIT_THRESHOLD  ((SZ_32K / 2) / sizeof(struct btrfs_free_space))
 290 #define INODES_PER_BITMAP (PAGE_SIZE * 8)
 291
 292 /*
 293  * The goal is to keep the memory used by the free_ino tree won't
 294  * exceed the memory if we use bitmaps only.
 295  */
 296 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 297 {
 298         struct btrfs_free_space *info;
 299         struct rb_node *n;
 300         int max_ino;
 301         int max_bitmaps;
 302
 303         n = rb_last(&ctl->free_space_offset);
 304         if (!n) {
 305                 ctl->extents_thresh = INIT_THRESHOLD;
 306                 return;
 307         }
 308         info = rb_entry(n, struct btrfs_free_space, offset_index);
 309
 310         /*
 311          * Find the maximum inode number in the filesystem. Note we
 312          * ignore the fact that this can be a bitmap, because we are
 313          * not doing precise calculation.
 314          */
 315         max_ino = info->bytes - 1;
 316
 317         max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
 318         if (max_bitmaps <= ctl->total_bitmaps) {
 319                 ctl->extents_thresh = 0;
 320                 return;
 321         }
 322
 323         ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
 324                                 PAGE_SIZE / sizeof(*info);
 325 }
 326
 327 /*
 328  * We don't fall back to bitmap, if we are below the extents threshold
 329  * or this chunk of inode numbers is a big one.
 330  */
 331 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
 332                        struct btrfs_free_space *info)
 333 {
 334         if (ctl->free_extents < ctl->extents_thresh ||
 335             info->bytes > INODES_PER_BITMAP / 10)
 336                 return false;
 337
 338         return true;
 339 }
 340
 341 static const struct btrfs_free_space_op free_ino_op = {
 342         .recalc_thresholds      = recalculate_thresholds,
 343         .use_bitmap             = use_bitmap,
 344 };
 345
 346 static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
 347 {
 348 }
 349
 350 static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
 351                               struct btrfs_free_space *info)
 352 {
 353         /*
 354          * We always use extents for two reasons:
 355          *
 356          * - The pinned tree is only used during the process of caching
 357          *   work.
 358          * - Make code simpler. See btrfs_unpin_free_ino().
 359          */
 360         return false;
 361 }
 362
 363 static const struct btrfs_free_space_op pinned_free_ino_op = {
 364         .recalc_thresholds      = pinned_recalc_thresholds,
 365         .use_bitmap             = pinned_use_bitmap,
 366 };
 367
 368 void btrfs_init_free_ino_ctl(struct btrfs_root *root)
 369 {
 370         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 371         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 372
 373         spin_lock_init(&ctl->tree_lock);
 374         ctl->unit = 1;
 375         ctl->start = 0;
 376         ctl->private = NULL;
 377         ctl->op = &free_ino_op;
 378         INIT_LIST_HEAD(&ctl->trimming_ranges);
 379         mutex_init(&ctl->cache_writeout_mutex);
 380
 381         /*
 382          * Initially we allow to use 16K of ram to cache chunks of
 383          * inode numbers before we resort to bitmaps. This is somewhat
 384          * arbitrary, but it will be adjusted in runtime.
 385          */
 386         ctl->extents_thresh = INIT_THRESHOLD;
 387
 388         spin_lock_init(&pinned->tree_lock);
 389         pinned->unit = 1;
 390         pinned->start = 0;
 391         pinned->private = NULL;
 392         pinned->extents_thresh = 0;
 393         pinned->op = &pinned_free_ino_op;
 394 }
 395
 396 int btrfs_save_ino_cache(struct btrfs_root *root,
 397                          struct btrfs_trans_handle *trans)
 398 {
 399         struct btrfs_fs_info *fs_info = root->fs_info;
 400         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 401         struct btrfs_path *path;
 402         struct inode *inode;
 403         struct btrfs_block_rsv *rsv;
 404         struct extent_changeset *data_reserved = NULL;
 405         u64 num_bytes;
 406         u64 alloc_hint = 0;
 407         int ret;
 408         int prealloc;
 409         bool retry = false;
 410
 411         /* only fs tree and subvol/snap needs ino cache */
 412         if (root->root_key.objectid != BTRFS_FS_TREE_OBJECTID &&
 413             (root->root_key.objectid < BTRFS_FIRST_FREE_OBJECTID ||
 414              root->root_key.objectid > BTRFS_LAST_FREE_OBJECTID))
 415                 return 0;
 416
 417         /* Don't save inode cache if we are deleting this root */
 418         if (btrfs_root_refs(&root->root_item) == 0)
 419                 return 0;
 420
 421         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 422                 return 0;
 423
 424         path = btrfs_alloc_path();
 425         if (!path)
 426                 return -ENOMEM;
 427
 428         rsv = trans->block_rsv;
 429         trans->block_rsv = &fs_info->trans_block_rsv;
 430
 431         num_bytes = trans->bytes_reserved;
 432         /*
 433          * 1 item for inode item insertion if need
 434          * 4 items for inode item update (in the worst case)
 435          * 1 items for slack space if we need do truncation
 436          * 1 item for free space object
 437          * 3 items for pre-allocation
 438          */
 439         trans->bytes_reserved = btrfs_calc_insert_metadata_size(fs_info, 10);
 440         ret = btrfs_block_rsv_add(root, trans->block_rsv,
 441                                   trans->bytes_reserved,
 442                                   BTRFS_RESERVE_NO_FLUSH);
 443         if (ret)
 444                 goto out;
 445         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 446                                       trans->bytes_reserved, 1);
 447 again:
 448         inode = lookup_free_ino_inode(root, path);
 449         if (IS_ERR(inode) && (PTR_ERR(inode) != -ENOENT || retry)) {
 450                 ret = PTR_ERR(inode);
 451                 goto out_release;
 452         }
 453
 454         if (IS_ERR(inode)) {
 455                 BUG_ON(retry); /* Logic error */
 456                 retry = true;
 457
 458                 ret = create_free_ino_inode(root, trans, path);
 459                 if (ret)
 460                         goto out_release;
 461                 goto again;
 462         }
 463
 464         BTRFS_I(inode)->generation = 0;
 465         ret = btrfs_update_inode(trans, root, inode);
 466         if (ret) {
 467                 btrfs_abort_transaction(trans, ret);
 468                 goto out_put;
 469         }
 470
 471         if (i_size_read(inode) > 0) {
 472                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
 473                 if (ret) {
 474                         if (ret != -ENOSPC)
 475                                 btrfs_abort_transaction(trans, ret);
 476                         goto out_put;
 477                 }
 478         }
 479
 480         spin_lock(&root->ino_cache_lock);
 481         if (root->ino_cache_state != BTRFS_CACHE_FINISHED) {
 482                 ret = -1;
 483                 spin_unlock(&root->ino_cache_lock);
 484                 goto out_put;
 485         }
 486         spin_unlock(&root->ino_cache_lock);
 487
 488         spin_lock(&ctl->tree_lock);
 489         prealloc = sizeof(struct btrfs_free_space) * ctl->free_extents;
 490         prealloc = ALIGN(prealloc, PAGE_SIZE);
 491         prealloc += ctl->total_bitmaps * PAGE_SIZE;
 492         spin_unlock(&ctl->tree_lock);
 493
 494         /* Just to make sure we have enough space */
 495         prealloc += 8 * PAGE_SIZE;
 496
 497         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, 0, prealloc);
 498         if (ret)
 499                 goto out_put;
 500
 501         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, prealloc,
 502                                               prealloc, prealloc, &alloc_hint);
 503         if (ret) {
 504                 btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc);
 505                 btrfs_delalloc_release_metadata(BTRFS_I(inode), prealloc, true);
 506                 goto out_put;
 507         }
 508
 509         ret = btrfs_write_out_ino_cache(root, trans, path, inode);
 510         btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc);
 511 out_put:
 512         iput(inode);
 513 out_release:
 514         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 515                                       trans->bytes_reserved, 0);
 516         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 517                                 trans->bytes_reserved);
 518 out:
 519         trans->block_rsv = rsv;
 520         trans->bytes_reserved = num_bytes;
 521
 522         btrfs_free_path(path);
 523         extent_changeset_free(data_reserved);
 524         return ret;
 525 }
 526
 527 int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
 528 {
 529         struct btrfs_path *path;
 530         int ret;
 531         struct extent_buffer *l;
 532         struct btrfs_key search_key;
 533         struct btrfs_key found_key;
 534         int slot;
 535
 536         path = btrfs_alloc_path();
 537         if (!path)
 538                 return -ENOMEM;
 539
 540         search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
 541         search_key.type = -1;
 542         search_key.offset = (u64)-1;
 543         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 544         if (ret < 0)
 545                 goto error;
 546         BUG_ON(ret == 0); /* Corruption */
 547         if (path->slots[0] > 0) {
 548                 slot = path->slots[0] - 1;
 549                 l = path->nodes[0];
 550                 btrfs_item_key_to_cpu(l, &found_key, slot);
 551                 *objectid = max_t(u64, found_key.objectid,
 552                                   BTRFS_FIRST_FREE_OBJECTID - 1);
 553         } else {
 554                 *objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
 555         }
 556         ret = 0;
 557 error:
 558         btrfs_free_path(path);
 559         return ret;
 560 }
 561
 562 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
 563 {
 564         int ret;
 565         mutex_lock(&root->objectid_mutex);
 566
 567         if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
 568                 btrfs_warn(root->fs_info,
 569                            "the objectid of root %llu reaches its highest value",
 570                            root->root_key.objectid);
 571                 ret = -ENOSPC;
 572                 goto out;
 573         }
 574
 575         *objectid = ++root->highest_objectid;
 576         ret = 0;
 577 out:
 578         mutex_unlock(&root->objectid_mutex);
 579         return ret;
 580 }