fs/btrfs/inode-map.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/delay.h>
  20 #include <linux/kthread.h>
  21 #include <linux/pagemap.h>
  22
  23 #include "ctree.h"
  24 #include "disk-io.h"
  25 #include "free-space-cache.h"
  26 #include "inode-map.h"
  27 #include "transaction.h"
  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(root->fs_info, INODE_MAP_CACHE))
  42                 return 0;
  43
  44         path = btrfs_alloc_path();
  45         if (!path)
  46                 return -ENOMEM;
  47
  48         /* Since the commit root is read-only, we can safely skip locking. */
  49         path->skip_locking = 1;
  50         path->search_commit_root = 1;
  51         path->reada = READA_FORWARD;
  52
  53         key.objectid = BTRFS_FIRST_FREE_OBJECTID;
  54         key.offset = 0;
  55         key.type = BTRFS_INODE_ITEM_KEY;
  56 again:
  57         /* need to make sure the commit_root doesn't disappear */
  58         down_read(&fs_info->commit_root_sem);
  59
  60         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  61         if (ret < 0)
  62                 goto out;
  63
  64         while (1) {
  65                 if (btrfs_fs_closing(fs_info))
  66                         goto out;
  67
  68                 leaf = path->nodes[0];
  69                 slot = path->slots[0];
  70                 if (slot >= btrfs_header_nritems(leaf)) {
  71                         ret = btrfs_next_leaf(root, path);
  72                         if (ret < 0)
  73                                 goto out;
  74                         else if (ret > 0)
  75                                 break;
  76
  77                         if (need_resched() ||
  78                             btrfs_transaction_in_commit(fs_info)) {
  79                                 leaf = path->nodes[0];
  80
  81                                 if (WARN_ON(btrfs_header_nritems(leaf) == 0))
  82                                         break;
  83
  84                                 /*
  85                                  * Save the key so we can advances forward
  86                                  * in the next search.
  87                                  */
  88                                 btrfs_item_key_to_cpu(leaf, &key, 0);
  89                                 btrfs_release_path(path);
  90                                 root->ino_cache_progress = last;
  91                                 up_read(&fs_info->commit_root_sem);
  92                                 schedule_timeout(1);
  93                                 goto again;
  94                         } else
  95                                 continue;
  96                 }
  97
  98                 btrfs_item_key_to_cpu(leaf, &key, slot);
  99
 100                 if (key.type != BTRFS_INODE_ITEM_KEY)
 101                         goto next;
 102
 103                 if (key.objectid >= root->highest_objectid)
 104                         break;
 105
 106                 if (last != (u64)-1 && last + 1 != key.objectid) {
 107                         __btrfs_add_free_space(fs_info, ctl, last + 1,
 108                                                key.objectid - last - 1);
 109                         wake_up(&root->ino_cache_wait);
 110                 }
 111
 112                 last = key.objectid;
 113 next:
 114                 path->slots[0]++;
 115         }
 116
 117         if (last < root->highest_objectid - 1) {
 118                 __btrfs_add_free_space(fs_info, ctl, last + 1,
 119                                        root->highest_objectid - last - 1);
 120         }
 121
 122         spin_lock(&root->ino_cache_lock);
 123         root->ino_cache_state = BTRFS_CACHE_FINISHED;
 124         spin_unlock(&root->ino_cache_lock);
 125
 126         root->ino_cache_progress = (u64)-1;
 127         btrfs_unpin_free_ino(root);
 128 out:
 129         wake_up(&root->ino_cache_wait);
 130         up_read(&fs_info->commit_root_sem);
 131
 132         btrfs_free_path(path);
 133
 134         return ret;
 135 }
 136
 137 static void start_caching(struct btrfs_root *root)
 138 {
 139         struct btrfs_fs_info *fs_info = root->fs_info;
 140         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 141         struct task_struct *tsk;
 142         int ret;
 143         u64 objectid;
 144
 145         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 146                 return;
 147
 148         spin_lock(&root->ino_cache_lock);
 149         if (root->ino_cache_state != BTRFS_CACHE_NO) {
 150                 spin_unlock(&root->ino_cache_lock);
 151                 return;
 152         }
 153
 154         root->ino_cache_state = BTRFS_CACHE_STARTED;
 155         spin_unlock(&root->ino_cache_lock);
 156
 157         ret = load_free_ino_cache(fs_info, root);
 158         if (ret == 1) {
 159                 spin_lock(&root->ino_cache_lock);
 160                 root->ino_cache_state = BTRFS_CACHE_FINISHED;
 161                 spin_unlock(&root->ino_cache_lock);
 162                 wake_up(&root->ino_cache_wait);
 163                 return;
 164         }
 165
 166         /*
 167          * It can be quite time-consuming to fill the cache by searching
 168          * through the extent tree, and this can keep ino allocation path
 169          * waiting. Therefore at start we quickly find out the highest
 170          * inode number and we know we can use inode numbers which fall in
 171          * [highest_ino + 1, BTRFS_LAST_FREE_OBJECTID].
 172          */
 173         ret = btrfs_find_free_objectid(root, &objectid);
 174         if (!ret && objectid <= BTRFS_LAST_FREE_OBJECTID) {
 175                 __btrfs_add_free_space(fs_info, ctl, objectid,
 176                                        BTRFS_LAST_FREE_OBJECTID - objectid + 1);
 177         }
 178
 179         tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu",
 180                           root->root_key.objectid);
 181         if (IS_ERR(tsk)) {
 182                 btrfs_warn(fs_info, "failed to start inode caching task");
 183                 btrfs_clear_pending_and_info(fs_info, INODE_MAP_CACHE,
 184                                 "disabling inode map caching");
 185         }
 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->free_ino_ctl->free_space > 0);
 204
 205         if (root->ino_cache_state == BTRFS_CACHE_FINISHED &&
 206             root->free_ino_ctl->free_space == 0)
 207                 return -ENOSPC;
 208         else
 209                 goto again;
 210 }
 211
 212 void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
 213 {
 214         struct btrfs_fs_info *fs_info = root->fs_info;
 215         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 216
 217         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
 218                 return;
 219 again:
 220         if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 221                 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
 222         } else {
 223                 down_write(&fs_info->commit_root_sem);
 224                 spin_lock(&root->ino_cache_lock);
 225                 if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
 226                         spin_unlock(&root->ino_cache_lock);
 227                         up_write(&fs_info->commit_root_sem);
 228                         goto again;
 229                 }
 230                 spin_unlock(&root->ino_cache_lock);
 231
 232                 start_caching(root);
 233
 234                 __btrfs_add_free_space(fs_info, pinned, objectid, 1);
 235
 236                 up_write(&fs_info->commit_root_sem);
 237         }
 238 }
 239
 240 /*
 241  * When a transaction is committed, we'll move those inode numbers which are
 242  * smaller than root->ino_cache_progress from pinned tree to free_ino tree, and
 243  * others will just be dropped, because the commit root we were searching has
 244  * changed.
 245  *
 246  * Must be called with root->fs_info->commit_root_sem held
 247  */
 248 void btrfs_unpin_free_ino(struct btrfs_root *root)
 249 {
 250         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 251         struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
 252         spinlock_t *rbroot_lock = &root->free_ino_pinned->tree_lock;
 253         struct btrfs_free_space *info;
 254         struct rb_node *n;
 255         u64 count;
 256
 257         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
 258                 return;
 259
 260         while (1) {
 261                 bool add_to_ctl = true;
 262
 263                 spin_lock(rbroot_lock);
 264                 n = rb_first(rbroot);
 265                 if (!n) {
 266                         spin_unlock(rbroot_lock);
 267                         break;
 268                 }
 269
 270                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 271                 BUG_ON(info->bitmap); /* Logic error */
 272
 273                 if (info->offset > root->ino_cache_progress)
 274                         add_to_ctl = false;
 275                 else if (info->offset + info->bytes > root->ino_cache_progress)
 276                         count = root->ino_cache_progress - info->offset + 1;
 277                 else
 278                         count = info->bytes;
 279
 280                 rb_erase(&info->offset_index, rbroot);
 281                 spin_unlock(rbroot_lock);
 282                 if (add_to_ctl)
 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_free_space_ctl *ctl = root->free_ino_ctl;
 400         struct btrfs_path *path;
 401         struct inode *inode;
 402         struct btrfs_block_rsv *rsv;
 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(root->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 = &root->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(root, 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(root->fs_info, "ino_cache",
 444                                       trans->transid, 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(root, 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, 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_metadata(inode, prealloc);
 503                 goto out_put;
 504         }
 505
 506         ret = btrfs_write_out_ino_cache(root, trans, path, inode);
 507 out_put:
 508         iput(inode);
 509 out_release:
 510         trace_btrfs_space_reservation(root->fs_info, "ino_cache",
 511                                       trans->transid, trans->bytes_reserved, 0);
 512         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
 513 out:
 514         trans->block_rsv = rsv;
 515         trans->bytes_reserved = num_bytes;
 516
 517         btrfs_free_path(path);
 518         return ret;
 519 }
 520
 521 int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
 522 {
 523         struct btrfs_path *path;
 524         int ret;
 525         struct extent_buffer *l;
 526         struct btrfs_key search_key;
 527         struct btrfs_key found_key;
 528         int slot;
 529
 530         path = btrfs_alloc_path();
 531         if (!path)
 532                 return -ENOMEM;
 533
 534         search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
 535         search_key.type = -1;
 536         search_key.offset = (u64)-1;
 537         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 538         if (ret < 0)
 539                 goto error;
 540         BUG_ON(ret == 0); /* Corruption */
 541         if (path->slots[0] > 0) {
 542                 slot = path->slots[0] - 1;
 543                 l = path->nodes[0];
 544                 btrfs_item_key_to_cpu(l, &found_key, slot);
 545                 *objectid = max_t(u64, found_key.objectid,
 546                                   BTRFS_FIRST_FREE_OBJECTID - 1);
 547         } else {
 548                 *objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
 549         }
 550         ret = 0;
 551 error:
 552         btrfs_free_path(path);
 553         return ret;
 554 }
 555
 556 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
 557 {
 558         int ret;
 559         mutex_lock(&root->objectid_mutex);
 560
 561         if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
 562                 btrfs_warn(root->fs_info,
 563                            "the objectid of root %llu reaches its highest value",
 564                            root->root_key.objectid);
 565                 ret = -ENOSPC;
 566                 goto out;
 567         }
 568
 569         *objectid = ++root->highest_objectid;
 570         ret = 0;
 571 out:
 572         mutex_unlock(&root->objectid_mutex);
 573         return ret;
 574 }