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, 0);
 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, 0);
 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                                        0);
 180                 wake_up(&root->ino_cache_wait);
 181         }
 182
 183         tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu",
 184                           root->root_key.objectid);
 185         if (IS_ERR(tsk))
 186                 fail_caching_thread(root);
 187 }
 188
 189 int btrfs_find_free_ino(struct btrfs_root *root, u64 *objectid)
 190 {
 191         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 192                 return btrfs_find_free_objectid(root, objectid);
 193
 194 again:
 195         *objectid = btrfs_find_ino_for_alloc(root);
 196
 197         if (*objectid != 0)
 198                 return 0;
 199
 200         start_caching(root);
 201
 202         wait_event(root->ino_cache_wait,
 203                    root->ino_cache_state == BTRFS_CACHE_FINISHED ||
 204                    root->ino_cache_state == BTRFS_CACHE_ERROR ||
 205                    root->free_ino_ctl->free_space > 0);
 206
 207         if (root->ino_cache_state == BTRFS_CACHE_FINISHED &&
 208             root->free_ino_ctl->free_space == 0)
 209                 return -ENOSPC;
 210         else if (root->ino_cache_state == BTRFS_CACHE_ERROR)
 211                 return btrfs_find_free_objectid(root, objectid);
 212         else
 213                 goto again;
 214 }
 215
 216 void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
 217 {
 218         struct btrfs_fs_info *fs_info = root->fs_info;
 219         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 220
 221         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 222                 return;
 223 again:
 224         if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 225                 __btrfs_add_free_space(fs_info, pinned, objectid, 1, 0);
 226         } else {
 227                 down_write(&fs_info->commit_root_sem);
 228                 spin_lock(&root->ino_cache_lock);
 229                 if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 230                         spin_unlock(&root->ino_cache_lock);
 231                         up_write(&fs_info->commit_root_sem);
 232                         goto again;
 233                 }
 234                 spin_unlock(&root->ino_cache_lock);
 235
 236                 start_caching(root);
 237
 238                 __btrfs_add_free_space(fs_info, pinned, objectid, 1, 0);
 239
 240                 up_write(&fs_info->commit_root_sem);
 241         }
 242 }
 243
 244 /*
 245  * When a transaction is committed, we'll move those inode numbers which are
 246  * smaller than root->ino_cache_progress from pinned tree to free_ino tree, and
 247  * others will just be dropped, because the commit root we were searching has
 248  * changed.
 249  *
 250  * Must be called with root->fs_info->commit_root_sem held
 251  */
 252 void btrfs_unpin_free_ino(struct btrfs_root *root)
 253 {
 254         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 255         struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
 256         spinlock_t *rbroot_lock = &root->free_ino_pinned->tree_lock;
 257         struct btrfs_free_space *info;
 258         struct rb_node *n;
 259         u64 count;
 260
 261         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 262                 return;
 263
 264         while (1) {
 265                 spin_lock(rbroot_lock);
 266                 n = rb_first(rbroot);
 267                 if (!n) {
 268                         spin_unlock(rbroot_lock);
 269                         break;
 270                 }
 271
 272                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 273                 BUG_ON(info->bitmap); /* Logic error */
 274
 275                 if (info->offset > root->ino_cache_progress)
 276                         count = 0;
 277                 else
 278                         count = min(root->ino_cache_progress - info->offset + 1,
 279                                     info->bytes);
 280
 281                 rb_erase(&info->offset_index, rbroot);
 282                 spin_unlock(rbroot_lock);
 283                 if (count)
 284                         __btrfs_add_free_space(root->fs_info, ctl,
 285                                                info->offset, count, 0);
 286                 kmem_cache_free(btrfs_free_space_cachep, info);
 287         }
 288 }
 289
 290 #define INIT_THRESHOLD  ((SZ_32K / 2) / sizeof(struct btrfs_free_space))
 291 #define INODES_PER_BITMAP (PAGE_SIZE * 8)
 292
 293 /*
 294  * The goal is to keep the memory used by the free_ino tree won't
 295  * exceed the memory if we use bitmaps only.
 296  */
 297 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 298 {
 299         struct btrfs_free_space *info;
 300         struct rb_node *n;
 301         int max_ino;
 302         int max_bitmaps;
 303
 304         n = rb_last(&ctl->free_space_offset);
 305         if (!n) {
 306                 ctl->extents_thresh = INIT_THRESHOLD;
 307                 return;
 308         }
 309         info = rb_entry(n, struct btrfs_free_space, offset_index);
 310
 311         /*
 312          * Find the maximum inode number in the filesystem. Note we
 313          * ignore the fact that this can be a bitmap, because we are
 314          * not doing precise calculation.
 315          */
 316         max_ino = info->bytes - 1;
 317
 318         max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
 319         if (max_bitmaps <= ctl->total_bitmaps) {
 320                 ctl->extents_thresh = 0;
 321                 return;
 322         }
 323
 324         ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
 325                                 PAGE_SIZE / sizeof(*info);
 326 }
 327
 328 /*
 329  * We don't fall back to bitmap, if we are below the extents threshold
 330  * or this chunk of inode numbers is a big one.
 331  */
 332 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
 333                        struct btrfs_free_space *info)
 334 {
 335         if (ctl->free_extents < ctl->extents_thresh ||
 336             info->bytes > INODES_PER_BITMAP / 10)
 337                 return false;
 338
 339         return true;
 340 }
 341
 342 static const struct btrfs_free_space_op free_ino_op = {
 343         .recalc_thresholds      = recalculate_thresholds,
 344         .use_bitmap             = use_bitmap,
 345 };
 346
 347 static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
 348 {
 349 }
 350
 351 static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
 352                               struct btrfs_free_space *info)
 353 {
 354         /*
 355          * We always use extents for two reasons:
 356          *
 357          * - The pinned tree is only used during the process of caching
 358          *   work.
 359          * - Make code simpler. See btrfs_unpin_free_ino().
 360          */
 361         return false;
 362 }
 363
 364 static const struct btrfs_free_space_op pinned_free_ino_op = {
 365         .recalc_thresholds      = pinned_recalc_thresholds,
 366         .use_bitmap             = pinned_use_bitmap,
 367 };
 368
 369 void btrfs_init_free_ino_ctl(struct btrfs_root *root)
 370 {
 371         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 372         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 373
 374         spin_lock_init(&ctl->tree_lock);
 375         ctl->unit = 1;
 376         ctl->start = 0;
 377         ctl->private = NULL;
 378         ctl->op = &free_ino_op;
 379         INIT_LIST_HEAD(&ctl->trimming_ranges);
 380         mutex_init(&ctl->cache_writeout_mutex);
 381
 382         /*
 383          * Initially we allow to use 16K of ram to cache chunks of
 384          * inode numbers before we resort to bitmaps. This is somewhat
 385          * arbitrary, but it will be adjusted in runtime.
 386          */
 387         ctl->extents_thresh = INIT_THRESHOLD;
 388
 389         spin_lock_init(&pinned->tree_lock);
 390         pinned->unit = 1;
 391         pinned->start = 0;
 392         pinned->private = NULL;
 393         pinned->extents_thresh = 0;
 394         pinned->op = &pinned_free_ino_op;
 395 }
 396
 397 int btrfs_save_ino_cache(struct btrfs_root *root,
 398                          struct btrfs_trans_handle *trans)
 399 {
 400         struct btrfs_fs_info *fs_info = root->fs_info;
 401         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 402         struct btrfs_path *path;
 403         struct inode *inode;
 404         struct btrfs_block_rsv *rsv;
 405         struct extent_changeset *data_reserved = NULL;
 406         u64 num_bytes;
 407         u64 alloc_hint = 0;
 408         int ret;
 409         int prealloc;
 410         bool retry = false;
 411
 412         /* only fs tree and subvol/snap needs ino cache */
 413         if (root->root_key.objectid != BTRFS_FS_TREE_OBJECTID &&
 414             (root->root_key.objectid < BTRFS_FIRST_FREE_OBJECTID ||
 415              root->root_key.objectid > BTRFS_LAST_FREE_OBJECTID))
 416                 return 0;
 417
 418         /* Don't save inode cache if we are deleting this root */
 419         if (btrfs_root_refs(&root->root_item) == 0)
 420                 return 0;
 421
 422         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 423                 return 0;
 424
 425         path = btrfs_alloc_path();
 426         if (!path)
 427                 return -ENOMEM;
 428
 429         rsv = trans->block_rsv;
 430         trans->block_rsv = &fs_info->trans_block_rsv;
 431
 432         num_bytes = trans->bytes_reserved;
 433         /*
 434          * 1 item for inode item insertion if need
 435          * 4 items for inode item update (in the worst case)
 436          * 1 items for slack space if we need do truncation
 437          * 1 item for free space object
 438          * 3 items for pre-allocation
 439          */
 440         trans->bytes_reserved = btrfs_calc_insert_metadata_size(fs_info, 10);
 441         ret = btrfs_block_rsv_add(root, trans->block_rsv,
 442                                   trans->bytes_reserved,
 443                                   BTRFS_RESERVE_NO_FLUSH);
 444         if (ret)
 445                 goto out;
 446         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 447                                       trans->bytes_reserved, 1);
 448 again:
 449         inode = lookup_free_ino_inode(root, path);
 450         if (IS_ERR(inode) && (PTR_ERR(inode) != -ENOENT || retry)) {
 451                 ret = PTR_ERR(inode);
 452                 goto out_release;
 453         }
 454
 455         if (IS_ERR(inode)) {
 456                 BUG_ON(retry); /* Logic error */
 457                 retry = true;
 458
 459                 ret = create_free_ino_inode(root, trans, path);
 460                 if (ret)
 461                         goto out_release;
 462                 goto again;
 463         }
 464
 465         BTRFS_I(inode)->generation = 0;
 466         ret = btrfs_update_inode(trans, root, inode);
 467         if (ret) {
 468                 btrfs_abort_transaction(trans, ret);
 469                 goto out_put;
 470         }
 471
 472         if (i_size_read(inode) > 0) {
 473                 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
 474                 if (ret) {
 475                         if (ret != -ENOSPC)
 476                                 btrfs_abort_transaction(trans, ret);
 477                         goto out_put;
 478                 }
 479         }
 480
 481         spin_lock(&root->ino_cache_lock);
 482         if (root->ino_cache_state != BTRFS_CACHE_FINISHED) {
 483                 ret = -1;
 484                 spin_unlock(&root->ino_cache_lock);
 485                 goto out_put;
 486         }
 487         spin_unlock(&root->ino_cache_lock);
 488
 489         spin_lock(&ctl->tree_lock);
 490         prealloc = sizeof(struct btrfs_free_space) * ctl->free_extents;
 491         prealloc = ALIGN(prealloc, PAGE_SIZE);
 492         prealloc += ctl->total_bitmaps * PAGE_SIZE;
 493         spin_unlock(&ctl->tree_lock);
 494
 495         /* Just to make sure we have enough space */
 496         prealloc += 8 * PAGE_SIZE;
 497
 498         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, 0, prealloc);
 499         if (ret)
 500                 goto out_put;
 501
 502         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, prealloc,
 503                                               prealloc, prealloc, &alloc_hint);
 504         if (ret) {
 505                 btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc);
 506                 btrfs_delalloc_release_metadata(BTRFS_I(inode), prealloc, true);
 507                 goto out_put;
 508         }
 509
 510         ret = btrfs_write_out_ino_cache(root, trans, path, inode);
 511         btrfs_delalloc_release_extents(BTRFS_I(inode), prealloc);
 512 out_put:
 513         iput(inode);
 514 out_release:
 515         trace_btrfs_space_reservation(fs_info, "ino_cache", trans->transid,
 516                                       trans->bytes_reserved, 0);
 517         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 518                                 trans->bytes_reserved, NULL);
 519 out:
 520         trans->block_rsv = rsv;
 521         trans->bytes_reserved = num_bytes;
 522
 523         btrfs_free_path(path);
 524         extent_changeset_free(data_reserved);
 525         return ret;
 526 }
 527
 528 int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
 529 {
 530         struct btrfs_path *path;
 531         int ret;
 532         struct extent_buffer *l;
 533         struct btrfs_key search_key;
 534         struct btrfs_key found_key;
 535         int slot;
 536
 537         path = btrfs_alloc_path();
 538         if (!path)
 539                 return -ENOMEM;
 540
 541         search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
 542         search_key.type = -1;
 543         search_key.offset = (u64)-1;
 544         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 545         if (ret < 0)
 546                 goto error;
 547         BUG_ON(ret == 0); /* Corruption */
 548         if (path->slots[0] > 0) {
 549                 slot = path->slots[0] - 1;
 550                 l = path->nodes[0];
 551                 btrfs_item_key_to_cpu(l, &found_key, slot);
 552                 *objectid = max_t(u64, found_key.objectid,
 553                                   BTRFS_FIRST_FREE_OBJECTID - 1);
 554         } else {
 555                 *objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
 556         }
 557         ret = 0;
 558 error:
 559         btrfs_free_path(path);
 560         return ret;
 561 }
 562
 563 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
 564 {
 565         int ret;
 566         mutex_lock(&root->objectid_mutex);
 567
 568         if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
 569                 btrfs_warn(root->fs_info,
 570                            "the objectid of root %llu reaches its highest value",
 571                            root->root_key.objectid);
 572                 ret = -ENOSPC;
 573                 goto out;
 574         }
 575
 576         *objectid = ++root->highest_objectid;
 577         ret = 0;
 578 out:
 579         mutex_unlock(&root->objectid_mutex);
 580         return ret;
 581 }