fs/btrfs/disk-io.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/fs.h>
  20 #include <linux/blkdev.h>
  21 #include <linux/crc32c.h>
  22 #include <linux/scatterlist.h>
  23 #include <linux/swap.h>
  24 #include <linux/radix-tree.h>
  25 #include <linux/writeback.h>
  26 #include <linux/buffer_head.h> // for block_sync_page
  27 #include "ctree.h"
  28 #include "disk-io.h"
  29 #include "transaction.h"
  30 #include "btrfs_inode.h"
  31 #include "print-tree.h"
  32
  33 #if 0
  34 static int check_tree_block(struct btrfs_root *root, struct extent_buffer *buf)
  35 {
  36         if (extent_buffer_blocknr(buf) != btrfs_header_blocknr(buf)) {
  37                 printk(KERN_CRIT "buf blocknr(buf) is %llu, header is %llu\n",
  38                        (unsigned long long)extent_buffer_blocknr(buf),
  39                        (unsigned long long)btrfs_header_blocknr(buf));
  40                 return 1;
  41         }
  42         return 0;
  43 }
  44 #endif
  45
  46 static struct extent_map_ops btree_extent_map_ops;
  47
  48 struct extent_buffer *btrfs_find_tree_block(struct btrfs_root *root,
  49                                             u64 bytenr, u32 blocksize)
  50 {
  51         struct inode *btree_inode = root->fs_info->btree_inode;
  52         struct extent_buffer *eb;
  53         eb = find_extent_buffer(&BTRFS_I(btree_inode)->extent_tree,
  54                                 bytenr, blocksize, GFP_NOFS);
  55         return eb;
  56 }
  57
  58 struct extent_buffer *btrfs_find_create_tree_block(struct btrfs_root *root,
  59                                                  u64 bytenr, u32 blocksize)
  60 {
  61         struct inode *btree_inode = root->fs_info->btree_inode;
  62         struct extent_buffer *eb;
  63
  64         eb = alloc_extent_buffer(&BTRFS_I(btree_inode)->extent_tree,
  65                                  bytenr, blocksize, NULL, GFP_NOFS);
  66         return eb;
  67 }
  68
  69 struct extent_map *btree_get_extent(struct inode *inode, struct page *page,
  70                                     size_t page_offset, u64 start, u64 end,
  71                                     int create)
  72 {
  73         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
  74         struct extent_map *em;
  75         int ret;
  76
  77 again:
  78         em = lookup_extent_mapping(em_tree, start, end);
  79         if (em) {
  80                 goto out;
  81         }
  82         em = alloc_extent_map(GFP_NOFS);
  83         if (!em) {
  84                 em = ERR_PTR(-ENOMEM);
  85                 goto out;
  86         }
  87         em->start = 0;
  88         em->end = (i_size_read(inode) & ~((u64)PAGE_CACHE_SIZE -1)) - 1;
  89         em->block_start = 0;
  90         em->block_end = em->end;
  91         em->bdev = inode->i_sb->s_bdev;
  92         ret = add_extent_mapping(em_tree, em);
  93         if (ret == -EEXIST) {
  94                 free_extent_map(em);
  95                 em = NULL;
  96                 goto again;
  97         } else if (ret) {
  98                 em = ERR_PTR(ret);
  99         }
 100 out:
 101         return em;
 102 }
 103
 104 u32 btrfs_csum_data(struct btrfs_root *root, char *data, u32 seed, size_t len)
 105 {
 106         return crc32c(seed, data, len);
 107 }
 108
 109 void btrfs_csum_final(u32 crc, char *result)
 110 {
 111         *(__le32 *)result = ~cpu_to_le32(crc);
 112 }
 113
 114 static int csum_tree_block(struct btrfs_root *root, struct extent_buffer *buf,
 115                            int verify)
 116 {
 117         char result[BTRFS_CRC32_SIZE];
 118         unsigned long len;
 119         unsigned long cur_len;
 120         unsigned long offset = BTRFS_CSUM_SIZE;
 121         char *map_token = NULL;
 122         char *kaddr;
 123         unsigned long map_start;
 124         unsigned long map_len;
 125         int err;
 126         u32 crc = ~(u32)0;
 127
 128         len = buf->len - offset;
 129         while(len > 0) {
 130                 err = map_private_extent_buffer(buf, offset, 32,
 131                                         &map_token, &kaddr,
 132                                         &map_start, &map_len, KM_USER0);
 133                 if (err) {
 134                         printk("failed to map extent buffer! %lu\n",
 135                                offset);
 136                         return 1;
 137                 }
 138                 cur_len = min(len, map_len - (offset - map_start));
 139                 crc = btrfs_csum_data(root, kaddr + offset - map_start,
 140                                       crc, cur_len);
 141                 len -= cur_len;
 142                 offset += cur_len;
 143                 unmap_extent_buffer(buf, map_token, KM_USER0);
 144         }
 145         btrfs_csum_final(crc, result);
 146
 147         if (verify) {
 148                 int from_this_trans = 0;
 149
 150                 if (root->fs_info->running_transaction &&
 151                     btrfs_header_generation(buf) ==
 152                     root->fs_info->running_transaction->transid)
 153                         from_this_trans = 1;
 154
 155                 /* FIXME, this is not good */
 156                 if (from_this_trans == 0 &&
 157                     memcmp_extent_buffer(buf, result, 0, BTRFS_CRC32_SIZE)) {
 158                         u32 val;
 159                         u32 found = 0;
 160                         memcpy(&found, result, BTRFS_CRC32_SIZE);
 161
 162                         read_extent_buffer(buf, &val, 0, BTRFS_CRC32_SIZE);
 163                         printk("btrfs: %s checksum verify failed on %llu "
 164                                "wanted %X found %X from_this_trans %d\n",
 165                                root->fs_info->sb->s_id,
 166                                buf->start, val, found, from_this_trans);
 167                         return 1;
 168                 }
 169         } else {
 170                 write_extent_buffer(buf, result, 0, BTRFS_CRC32_SIZE);
 171         }
 172         return 0;
 173 }
 174
 175
 176 int csum_dirty_buffer(struct btrfs_root *root, struct page *page)
 177 {
 178         struct extent_map_tree *tree;
 179         u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
 180         u64 found_start;
 181         int found_level;
 182         unsigned long len;
 183         struct extent_buffer *eb;
 184         tree = &BTRFS_I(page->mapping->host)->extent_tree;
 185
 186         if (page->private == EXTENT_PAGE_PRIVATE)
 187                 goto out;
 188         if (!page->private)
 189                 goto out;
 190         len = page->private >> 2;
 191         if (len == 0) {
 192                 WARN_ON(1);
 193         }
 194         eb = alloc_extent_buffer(tree, start, len, page, GFP_NOFS);
 195         read_extent_buffer_pages(tree, eb, start + PAGE_CACHE_SIZE, 1);
 196         btrfs_clear_buffer_defrag(eb);
 197         found_start = btrfs_header_bytenr(eb);
 198         if (found_start != start) {
 199                 printk("warning: eb start incorrect %Lu buffer %Lu len %lu\n",
 200                        start, found_start, len);
 201                 WARN_ON(1);
 202                 goto err;
 203         }
 204         if (eb->first_page != page) {
 205                 printk("bad first page %lu %lu\n", eb->first_page->index,
 206                        page->index);
 207                 WARN_ON(1);
 208                 goto err;
 209         }
 210         if (!PageUptodate(page)) {
 211                 printk("csum not up to date page %lu\n", page->index);
 212                 WARN_ON(1);
 213                 goto err;
 214         }
 215         found_level = btrfs_header_level(eb);
 216         csum_tree_block(root, eb, 0);
 217 err:
 218         free_extent_buffer(eb);
 219 out:
 220         return 0;
 221 }
 222
 223 static int btree_writepage_io_hook(struct page *page, u64 start, u64 end)
 224 {
 225         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
 226
 227         csum_dirty_buffer(root, page);
 228         return 0;
 229 }
 230
 231 static int btree_writepage(struct page *page, struct writeback_control *wbc)
 232 {
 233         struct extent_map_tree *tree;
 234         tree = &BTRFS_I(page->mapping->host)->extent_tree;
 235         return extent_write_full_page(tree, page, btree_get_extent, wbc);
 236 }
 237
 238 static int btree_writepages(struct address_space *mapping,
 239                             struct writeback_control *wbc)
 240 {
 241         struct extent_map_tree *tree;
 242         tree = &BTRFS_I(mapping->host)->extent_tree;
 243         if (wbc->sync_mode == WB_SYNC_NONE) {
 244                 u64 num_dirty;
 245                 u64 start = 0;
 246                 unsigned long thresh = 96 * 1024 * 1024;
 247
 248                 if (wbc->for_kupdate)
 249                         return 0;
 250
 251                 if (current_is_pdflush()) {
 252                         thresh = 96 * 1024 * 1024;
 253                 } else {
 254                         thresh = 8 * 1024 * 1024;
 255                 }
 256                 num_dirty = count_range_bits(tree, &start, (u64)-1,
 257                                              thresh, EXTENT_DIRTY);
 258                 if (num_dirty < thresh) {
 259                         return 0;
 260                 }
 261         }
 262         return extent_writepages(tree, mapping, btree_get_extent, wbc);
 263 }
 264
 265 int btree_readpage(struct file *file, struct page *page)
 266 {
 267         struct extent_map_tree *tree;
 268         tree = &BTRFS_I(page->mapping->host)->extent_tree;
 269         return extent_read_full_page(tree, page, btree_get_extent);
 270 }
 271
 272 static int btree_releasepage(struct page *page, gfp_t unused_gfp_flags)
 273 {
 274         struct extent_map_tree *tree;
 275         int ret;
 276
 277         tree = &BTRFS_I(page->mapping->host)->extent_tree;
 278         ret = try_release_extent_mapping(tree, page);
 279         if (ret == 1) {
 280                 ClearPagePrivate(page);
 281                 set_page_private(page, 0);
 282                 page_cache_release(page);
 283         }
 284         return ret;
 285 }
 286
 287 static void btree_invalidatepage(struct page *page, unsigned long offset)
 288 {
 289         struct extent_map_tree *tree;
 290         tree = &BTRFS_I(page->mapping->host)->extent_tree;
 291         extent_invalidatepage(tree, page, offset);
 292         btree_releasepage(page, GFP_NOFS);
 293 }
 294
 295 #if 0
 296 static int btree_writepage(struct page *page, struct writeback_control *wbc)
 297 {
 298         struct buffer_head *bh;
 299         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
 300         struct buffer_head *head;
 301         if (!page_has_buffers(page)) {
 302                 create_empty_buffers(page, root->fs_info->sb->s_blocksize,
 303                                         (1 << BH_Dirty)|(1 << BH_Uptodate));
 304         }
 305         head = page_buffers(page);
 306         bh = head;
 307         do {
 308                 if (buffer_dirty(bh))
 309                         csum_tree_block(root, bh, 0);
 310                 bh = bh->b_this_page;
 311         } while (bh != head);
 312         return block_write_full_page(page, btree_get_block, wbc);
 313 }
 314 #endif
 315
 316 static struct address_space_operations btree_aops = {
 317         .readpage       = btree_readpage,
 318         .writepage      = btree_writepage,
 319         .writepages     = btree_writepages,
 320         .releasepage    = btree_releasepage,
 321         .invalidatepage = btree_invalidatepage,
 322         .sync_page      = block_sync_page,
 323 };
 324
 325 int readahead_tree_block(struct btrfs_root *root, u64 bytenr, u32 blocksize)
 326 {
 327         struct extent_buffer *buf = NULL;
 328         struct inode *btree_inode = root->fs_info->btree_inode;
 329         int ret = 0;
 330
 331         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
 332         if (!buf)
 333                 return 0;
 334         read_extent_buffer_pages(&BTRFS_I(btree_inode)->extent_tree,
 335                                  buf, 0, 0);
 336         free_extent_buffer(buf);
 337         return ret;
 338 }
 339
 340 struct extent_buffer *read_tree_block(struct btrfs_root *root, u64 bytenr,
 341                                       u32 blocksize)
 342 {
 343         struct extent_buffer *buf = NULL;
 344         struct inode *btree_inode = root->fs_info->btree_inode;
 345         struct extent_map_tree *extent_tree;
 346         u64 end;
 347         int ret;
 348
 349         extent_tree = &BTRFS_I(btree_inode)->extent_tree;
 350
 351         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
 352         if (!buf)
 353                 return NULL;
 354         read_extent_buffer_pages(&BTRFS_I(btree_inode)->extent_tree,
 355                                  buf, 0, 1);
 356
 357         if (buf->flags & EXTENT_CSUM)
 358                 return buf;
 359
 360         end = buf->start + PAGE_CACHE_SIZE - 1;
 361         if (test_range_bit(extent_tree, buf->start, end, EXTENT_CSUM, 1)) {
 362                 buf->flags |= EXTENT_CSUM;
 363                 return buf;
 364         }
 365
 366         lock_extent(extent_tree, buf->start, end, GFP_NOFS);
 367
 368         if (test_range_bit(extent_tree, buf->start, end, EXTENT_CSUM, 1)) {
 369                 buf->flags |= EXTENT_CSUM;
 370                 goto out_unlock;
 371         }
 372
 373         ret = csum_tree_block(root, buf, 1);
 374         set_extent_bits(extent_tree, buf->start, end, EXTENT_CSUM, GFP_NOFS);
 375         buf->flags |= EXTENT_CSUM;
 376
 377 out_unlock:
 378         unlock_extent(extent_tree, buf->start, end, GFP_NOFS);
 379         return buf;
 380 }
 381
 382 int clean_tree_block(struct btrfs_trans_handle *trans, struct btrfs_root *root,
 383                      struct extent_buffer *buf)
 384 {
 385         struct inode *btree_inode = root->fs_info->btree_inode;
 386         if (btrfs_header_generation(buf) ==
 387             root->fs_info->running_transaction->transid)
 388                 clear_extent_buffer_dirty(&BTRFS_I(btree_inode)->extent_tree,
 389                                           buf);
 390         return 0;
 391 }
 392
 393 int wait_on_tree_block_writeback(struct btrfs_root *root,
 394                                  struct extent_buffer *buf)
 395 {
 396         struct inode *btree_inode = root->fs_info->btree_inode;
 397         wait_on_extent_buffer_writeback(&BTRFS_I(btree_inode)->extent_tree,
 398                                         buf);
 399         return 0;
 400 }
 401
 402 static int __setup_root(u32 nodesize, u32 leafsize, u32 sectorsize,
 403                         u32 stripesize, struct btrfs_root *root,
 404                         struct btrfs_fs_info *fs_info,
 405                         u64 objectid)
 406 {
 407         root->node = NULL;
 408         root->inode = NULL;
 409         root->commit_root = NULL;
 410         root->sectorsize = sectorsize;
 411         root->nodesize = nodesize;
 412         root->leafsize = leafsize;
 413         root->stripesize = stripesize;
 414         root->ref_cows = 0;
 415         root->fs_info = fs_info;
 416         root->objectid = objectid;
 417         root->last_trans = 0;
 418         root->highest_inode = 0;
 419         root->last_inode_alloc = 0;
 420         root->name = NULL;
 421         root->in_sysfs = 0;
 422         memset(&root->root_key, 0, sizeof(root->root_key));
 423         memset(&root->root_item, 0, sizeof(root->root_item));
 424         memset(&root->defrag_progress, 0, sizeof(root->defrag_progress));
 425         memset(&root->root_kobj, 0, sizeof(root->root_kobj));
 426         init_completion(&root->kobj_unregister);
 427         root->defrag_running = 0;
 428         root->defrag_level = 0;
 429         root->root_key.objectid = objectid;
 430         return 0;
 431 }
 432
 433 static int find_and_setup_root(struct btrfs_root *tree_root,
 434                                struct btrfs_fs_info *fs_info,
 435                                u64 objectid,
 436                                struct btrfs_root *root)
 437 {
 438         int ret;
 439         u32 blocksize;
 440
 441         __setup_root(tree_root->nodesize, tree_root->leafsize,
 442                      tree_root->sectorsize, tree_root->stripesize,
 443                      root, fs_info, objectid);
 444         ret = btrfs_find_last_root(tree_root, objectid,
 445                                    &root->root_item, &root->root_key);
 446         BUG_ON(ret);
 447
 448         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 449         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 450                                      blocksize);
 451         BUG_ON(!root->node);
 452         return 0;
 453 }
 454
 455 struct btrfs_root *btrfs_read_fs_root_no_radix(struct btrfs_fs_info *fs_info,
 456                                                struct btrfs_key *location)
 457 {
 458         struct btrfs_root *root;
 459         struct btrfs_root *tree_root = fs_info->tree_root;
 460         struct btrfs_path *path;
 461         struct extent_buffer *l;
 462         u64 highest_inode;
 463         u32 blocksize;
 464         int ret = 0;
 465
 466         root = kzalloc(sizeof(*root), GFP_NOFS);
 467         if (!root)
 468                 return ERR_PTR(-ENOMEM);
 469         if (location->offset == (u64)-1) {
 470                 ret = find_and_setup_root(tree_root, fs_info,
 471                                           location->objectid, root);
 472                 if (ret) {
 473                         kfree(root);
 474                         return ERR_PTR(ret);
 475                 }
 476                 goto insert;
 477         }
 478
 479         __setup_root(tree_root->nodesize, tree_root->leafsize,
 480                      tree_root->sectorsize, tree_root->stripesize,
 481                      root, fs_info, location->objectid);
 482
 483         path = btrfs_alloc_path();
 484         BUG_ON(!path);
 485         ret = btrfs_search_slot(NULL, tree_root, location, path, 0, 0);
 486         if (ret != 0) {
 487                 if (ret > 0)
 488                         ret = -ENOENT;
 489                 goto out;
 490         }
 491         l = path->nodes[0];
 492         read_extent_buffer(l, &root->root_item,
 493                btrfs_item_ptr_offset(l, path->slots[0]),
 494                sizeof(root->root_item));
 495         memcpy(&root->root_key, location, sizeof(*location));
 496         ret = 0;
 497 out:
 498         btrfs_release_path(root, path);
 499         btrfs_free_path(path);
 500         if (ret) {
 501                 kfree(root);
 502                 return ERR_PTR(ret);
 503         }
 504         blocksize = btrfs_level_size(root, btrfs_root_level(&root->root_item));
 505         root->node = read_tree_block(root, btrfs_root_bytenr(&root->root_item),
 506                                      blocksize);
 507         BUG_ON(!root->node);
 508 insert:
 509         root->ref_cows = 1;
 510         ret = btrfs_find_highest_inode(root, &highest_inode);
 511         if (ret == 0) {
 512                 root->highest_inode = highest_inode;
 513                 root->last_inode_alloc = highest_inode;
 514         }
 515         return root;
 516 }
 517
 518 struct btrfs_root *btrfs_lookup_fs_root(struct btrfs_fs_info *fs_info,
 519                                         u64 root_objectid)
 520 {
 521         struct btrfs_root *root;
 522
 523         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID)
 524                 return fs_info->tree_root;
 525         if (root_objectid == BTRFS_EXTENT_TREE_OBJECTID)
 526                 return fs_info->extent_root;
 527
 528         root = radix_tree_lookup(&fs_info->fs_roots_radix,
 529                                  (unsigned long)root_objectid);
 530         return root;
 531 }
 532
 533 struct btrfs_root *btrfs_read_fs_root_no_name(struct btrfs_fs_info *fs_info,
 534                                               struct btrfs_key *location)
 535 {
 536         struct btrfs_root *root;
 537         int ret;
 538
 539         if (location->objectid == BTRFS_ROOT_TREE_OBJECTID)
 540                 return fs_info->tree_root;
 541         if (location->objectid == BTRFS_EXTENT_TREE_OBJECTID)
 542                 return fs_info->extent_root;
 543
 544         root = radix_tree_lookup(&fs_info->fs_roots_radix,
 545                                  (unsigned long)location->objectid);
 546         if (root)
 547                 return root;
 548
 549         root = btrfs_read_fs_root_no_radix(fs_info, location);
 550         if (IS_ERR(root))
 551                 return root;
 552         ret = radix_tree_insert(&fs_info->fs_roots_radix,
 553                                 (unsigned long)root->root_key.objectid,
 554                                 root);
 555         if (ret) {
 556                 free_extent_buffer(root->node);
 557                 kfree(root);
 558                 return ERR_PTR(ret);
 559         }
 560         ret = btrfs_find_dead_roots(fs_info->tree_root,
 561                                     root->root_key.objectid, root);
 562         BUG_ON(ret);
 563
 564         return root;
 565 }
 566
 567 struct btrfs_root *btrfs_read_fs_root(struct btrfs_fs_info *fs_info,
 568                                       struct btrfs_key *location,
 569                                       const char *name, int namelen)
 570 {
 571         struct btrfs_root *root;
 572         int ret;
 573
 574         root = btrfs_read_fs_root_no_name(fs_info, location);
 575         if (!root)
 576                 return NULL;
 577
 578         if (root->in_sysfs)
 579                 return root;
 580
 581         ret = btrfs_set_root_name(root, name, namelen);
 582         if (ret) {
 583                 free_extent_buffer(root->node);
 584                 kfree(root);
 585                 return ERR_PTR(ret);
 586         }
 587
 588         ret = btrfs_sysfs_add_root(root);
 589         if (ret) {
 590                 free_extent_buffer(root->node);
 591                 kfree(root->name);
 592                 kfree(root);
 593                 return ERR_PTR(ret);
 594         }
 595         root->in_sysfs = 1;
 596         return root;
 597 }
 598 #if 0
 599 static int add_hasher(struct btrfs_fs_info *info, char *type) {
 600         struct btrfs_hasher *hasher;
 601
 602         hasher = kmalloc(sizeof(*hasher), GFP_NOFS);
 603         if (!hasher)
 604                 return -ENOMEM;
 605         hasher->hash_tfm = crypto_alloc_hash(type, 0, CRYPTO_ALG_ASYNC);
 606         if (!hasher->hash_tfm) {
 607                 kfree(hasher);
 608                 return -EINVAL;
 609         }
 610         spin_lock(&info->hash_lock);
 611         list_add(&hasher->list, &info->hashers);
 612         spin_unlock(&info->hash_lock);
 613         return 0;
 614 }
 615 #endif
 616 struct btrfs_root *open_ctree(struct super_block *sb)
 617 {
 618         u32 sectorsize;
 619         u32 nodesize;
 620         u32 leafsize;
 621         u32 blocksize;
 622         u32 stripesize;
 623         struct btrfs_root *extent_root = kmalloc(sizeof(struct btrfs_root),
 624                                                  GFP_NOFS);
 625         struct btrfs_root *tree_root = kmalloc(sizeof(struct btrfs_root),
 626                                                GFP_NOFS);
 627         struct btrfs_fs_info *fs_info = kmalloc(sizeof(*fs_info),
 628                                                 GFP_NOFS);
 629         int ret;
 630         int err = -EIO;
 631         struct btrfs_super_block *disk_super;
 632
 633         if (!extent_root || !tree_root || !fs_info) {
 634                 err = -ENOMEM;
 635                 goto fail;
 636         }
 637         INIT_RADIX_TREE(&fs_info->fs_roots_radix, GFP_NOFS);
 638         INIT_LIST_HEAD(&fs_info->trans_list);
 639         INIT_LIST_HEAD(&fs_info->dead_roots);
 640         INIT_LIST_HEAD(&fs_info->hashers);
 641         spin_lock_init(&fs_info->hash_lock);
 642         spin_lock_init(&fs_info->delalloc_lock);
 643         spin_lock_init(&fs_info->new_trans_lock);
 644
 645         memset(&fs_info->super_kobj, 0, sizeof(fs_info->super_kobj));
 646         init_completion(&fs_info->kobj_unregister);
 647         sb_set_blocksize(sb, 4096);
 648         fs_info->running_transaction = NULL;
 649         fs_info->last_trans_committed = 0;
 650         fs_info->tree_root = tree_root;
 651         fs_info->extent_root = extent_root;
 652         fs_info->sb = sb;
 653         fs_info->throttles = 0;
 654         fs_info->mount_opt = 0;
 655         fs_info->max_extent = (u64)-1;
 656         fs_info->delalloc_bytes = 0;
 657         fs_info->btree_inode = new_inode(sb);
 658         fs_info->btree_inode->i_ino = 1;
 659         fs_info->btree_inode->i_nlink = 1;
 660         fs_info->btree_inode->i_size = sb->s_bdev->bd_inode->i_size;
 661         fs_info->btree_inode->i_mapping->a_ops = &btree_aops;
 662         extent_map_tree_init(&BTRFS_I(fs_info->btree_inode)->extent_tree,
 663                              fs_info->btree_inode->i_mapping,
 664                              GFP_NOFS);
 665         BTRFS_I(fs_info->btree_inode)->extent_tree.ops = &btree_extent_map_ops;
 666
 667         extent_map_tree_init(&fs_info->free_space_cache,
 668                              fs_info->btree_inode->i_mapping, GFP_NOFS);
 669         extent_map_tree_init(&fs_info->block_group_cache,
 670                              fs_info->btree_inode->i_mapping, GFP_NOFS);
 671         extent_map_tree_init(&fs_info->pinned_extents,
 672                              fs_info->btree_inode->i_mapping, GFP_NOFS);
 673         extent_map_tree_init(&fs_info->pending_del,
 674                              fs_info->btree_inode->i_mapping, GFP_NOFS);
 675         extent_map_tree_init(&fs_info->extent_ins,
 676                              fs_info->btree_inode->i_mapping, GFP_NOFS);
 677         fs_info->do_barriers = 1;
 678         fs_info->closing = 0;
 679         fs_info->total_pinned = 0;
 680         fs_info->last_alloc = 0;
 681
 682 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,18)
 683         INIT_WORK(&fs_info->trans_work, btrfs_transaction_cleaner, fs_info);
 684 #else
 685         INIT_DELAYED_WORK(&fs_info->trans_work, btrfs_transaction_cleaner);
 686 #endif
 687         BTRFS_I(fs_info->btree_inode)->root = tree_root;
 688         memset(&BTRFS_I(fs_info->btree_inode)->location, 0,
 689                sizeof(struct btrfs_key));
 690         insert_inode_hash(fs_info->btree_inode);
 691         mapping_set_gfp_mask(fs_info->btree_inode->i_mapping, GFP_NOFS);
 692
 693         mutex_init(&fs_info->trans_mutex);
 694         mutex_init(&fs_info->fs_mutex);
 695
 696 #if 0
 697         ret = add_hasher(fs_info, "crc32c");
 698         if (ret) {
 699                 printk("btrfs: failed hash setup, modprobe cryptomgr?\n");
 700                 err = -ENOMEM;
 701                 goto fail_iput;
 702         }
 703 #endif
 704         __setup_root(512, 512, 512, 512, tree_root,
 705                      fs_info, BTRFS_ROOT_TREE_OBJECTID);
 706
 707         fs_info->sb_buffer = read_tree_block(tree_root,
 708                                              BTRFS_SUPER_INFO_OFFSET,
 709                                              512);
 710
 711         if (!fs_info->sb_buffer)
 712                 goto fail_iput;
 713
 714         read_extent_buffer(fs_info->sb_buffer, &fs_info->super_copy, 0,
 715                            sizeof(fs_info->super_copy));
 716
 717         read_extent_buffer(fs_info->sb_buffer, fs_info->fsid,
 718                            (unsigned long)btrfs_super_fsid(fs_info->sb_buffer),
 719                            BTRFS_FSID_SIZE);
 720         disk_super = &fs_info->super_copy;
 721         if (!btrfs_super_root(disk_super))
 722                 goto fail_sb_buffer;
 723
 724         nodesize = btrfs_super_nodesize(disk_super);
 725         leafsize = btrfs_super_leafsize(disk_super);
 726         sectorsize = btrfs_super_sectorsize(disk_super);
 727         stripesize = btrfs_super_stripesize(disk_super);
 728         tree_root->nodesize = nodesize;
 729         tree_root->leafsize = leafsize;
 730         tree_root->sectorsize = sectorsize;
 731         tree_root->stripesize = stripesize;
 732         sb_set_blocksize(sb, sectorsize);
 733
 734         i_size_write(fs_info->btree_inode,
 735                      btrfs_super_total_bytes(disk_super));
 736
 737         if (strncmp((char *)(&disk_super->magic), BTRFS_MAGIC,
 738                     sizeof(disk_super->magic))) {
 739                 printk("btrfs: valid FS not found on %s\n", sb->s_id);
 740                 goto fail_sb_buffer;
 741         }
 742
 743         blocksize = btrfs_level_size(tree_root,
 744                                      btrfs_super_root_level(disk_super));
 745
 746         tree_root->node = read_tree_block(tree_root,
 747                                           btrfs_super_root(disk_super),
 748                                           blocksize);
 749         if (!tree_root->node)
 750                 goto fail_sb_buffer;
 751
 752         mutex_lock(&fs_info->fs_mutex);
 753
 754         ret = find_and_setup_root(tree_root, fs_info,
 755                                   BTRFS_EXTENT_TREE_OBJECTID, extent_root);
 756         if (ret) {
 757                 mutex_unlock(&fs_info->fs_mutex);
 758                 goto fail_tree_root;
 759         }
 760
 761         btrfs_read_block_groups(extent_root);
 762
 763         fs_info->generation = btrfs_super_generation(disk_super) + 1;
 764         mutex_unlock(&fs_info->fs_mutex);
 765         return tree_root;
 766
 767 fail_tree_root:
 768         free_extent_buffer(tree_root->node);
 769 fail_sb_buffer:
 770         free_extent_buffer(fs_info->sb_buffer);
 771 fail_iput:
 772         iput(fs_info->btree_inode);
 773 fail:
 774         kfree(extent_root);
 775         kfree(tree_root);
 776         kfree(fs_info);
 777         return ERR_PTR(err);
 778 }
 779
 780 int write_ctree_super(struct btrfs_trans_handle *trans, struct btrfs_root
 781                       *root)
 782 {
 783         int ret;
 784         struct extent_buffer *super = root->fs_info->sb_buffer;
 785         struct inode *btree_inode = root->fs_info->btree_inode;
 786         struct super_block *sb = root->fs_info->sb;
 787
 788         if (!btrfs_test_opt(root, NOBARRIER))
 789                 blkdev_issue_flush(sb->s_bdev, NULL);
 790         set_extent_buffer_dirty(&BTRFS_I(btree_inode)->extent_tree, super);
 791         ret = sync_page_range_nolock(btree_inode, btree_inode->i_mapping,
 792                                      super->start, super->len);
 793         if (!btrfs_test_opt(root, NOBARRIER))
 794                 blkdev_issue_flush(sb->s_bdev, NULL);
 795         return ret;
 796 }
 797
 798 int btrfs_free_fs_root(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
 799 {
 800         radix_tree_delete(&fs_info->fs_roots_radix,
 801                           (unsigned long)root->root_key.objectid);
 802         if (root->in_sysfs)
 803                 btrfs_sysfs_del_root(root);
 804         if (root->inode)
 805                 iput(root->inode);
 806         if (root->node)
 807                 free_extent_buffer(root->node);
 808         if (root->commit_root)
 809                 free_extent_buffer(root->commit_root);
 810         if (root->name)
 811                 kfree(root->name);
 812         kfree(root);
 813         return 0;
 814 }
 815
 816 static int del_fs_roots(struct btrfs_fs_info *fs_info)
 817 {
 818         int ret;
 819         struct btrfs_root *gang[8];
 820         int i;
 821
 822         while(1) {
 823                 ret = radix_tree_gang_lookup(&fs_info->fs_roots_radix,
 824                                              (void **)gang, 0,
 825                                              ARRAY_SIZE(gang));
 826                 if (!ret)
 827                         break;
 828                 for (i = 0; i < ret; i++)
 829                         btrfs_free_fs_root(fs_info, gang[i]);
 830         }
 831         return 0;
 832 }
 833
 834 int close_ctree(struct btrfs_root *root)
 835 {
 836         int ret;
 837         struct btrfs_trans_handle *trans;
 838         struct btrfs_fs_info *fs_info = root->fs_info;
 839
 840         fs_info->closing = 1;
 841         btrfs_transaction_flush_work(root);
 842         mutex_lock(&fs_info->fs_mutex);
 843         btrfs_defrag_dirty_roots(root->fs_info);
 844         trans = btrfs_start_transaction(root, 1);
 845         ret = btrfs_commit_transaction(trans, root);
 846         /* run commit again to  drop the original snapshot */
 847         trans = btrfs_start_transaction(root, 1);
 848         btrfs_commit_transaction(trans, root);
 849         ret = btrfs_write_and_wait_transaction(NULL, root);
 850         BUG_ON(ret);
 851         write_ctree_super(NULL, root);
 852         mutex_unlock(&fs_info->fs_mutex);
 853
 854         if (fs_info->extent_root->node)
 855                 free_extent_buffer(fs_info->extent_root->node);
 856
 857         if (fs_info->tree_root->node)
 858                 free_extent_buffer(fs_info->tree_root->node);
 859
 860         free_extent_buffer(fs_info->sb_buffer);
 861
 862         btrfs_free_block_groups(root->fs_info);
 863         del_fs_roots(fs_info);
 864
 865         filemap_write_and_wait(fs_info->btree_inode->i_mapping);
 866
 867         extent_map_tree_empty_lru(&fs_info->free_space_cache);
 868         extent_map_tree_empty_lru(&fs_info->block_group_cache);
 869         extent_map_tree_empty_lru(&fs_info->pinned_extents);
 870         extent_map_tree_empty_lru(&fs_info->pending_del);
 871         extent_map_tree_empty_lru(&fs_info->extent_ins);
 872         extent_map_tree_empty_lru(&BTRFS_I(fs_info->btree_inode)->extent_tree);
 873
 874         truncate_inode_pages(fs_info->btree_inode->i_mapping, 0);
 875
 876         iput(fs_info->btree_inode);
 877 #if 0
 878         while(!list_empty(&fs_info->hashers)) {
 879                 struct btrfs_hasher *hasher;
 880                 hasher = list_entry(fs_info->hashers.next, struct btrfs_hasher,
 881                                     hashers);
 882                 list_del(&hasher->hashers);
 883                 crypto_free_hash(&fs_info->hash_tfm);
 884                 kfree(hasher);
 885         }
 886 #endif
 887         kfree(fs_info->extent_root);
 888         kfree(fs_info->tree_root);
 889         return 0;
 890 }
 891
 892 int btrfs_buffer_uptodate(struct extent_buffer *buf)
 893 {
 894         struct inode *btree_inode = buf->first_page->mapping->host;
 895         return extent_buffer_uptodate(&BTRFS_I(btree_inode)->extent_tree, buf);
 896 }
 897
 898 int btrfs_set_buffer_uptodate(struct extent_buffer *buf)
 899 {
 900         struct inode *btree_inode = buf->first_page->mapping->host;
 901         return set_extent_buffer_uptodate(&BTRFS_I(btree_inode)->extent_tree,
 902                                           buf);
 903 }
 904
 905 void btrfs_mark_buffer_dirty(struct extent_buffer *buf)
 906 {
 907         struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
 908         u64 transid = btrfs_header_generation(buf);
 909         struct inode *btree_inode = root->fs_info->btree_inode;
 910
 911         if (transid != root->fs_info->generation) {
 912                 printk(KERN_CRIT "transid mismatch buffer %llu, found %Lu running %Lu\n",
 913                         (unsigned long long)buf->start,
 914                         transid, root->fs_info->generation);
 915                 WARN_ON(1);
 916         }
 917         set_extent_buffer_dirty(&BTRFS_I(btree_inode)->extent_tree, buf);
 918 }
 919
 920 void btrfs_throttle(struct btrfs_root *root)
 921 {
 922         struct backing_dev_info *bdi;
 923
 924         bdi = root->fs_info->sb->s_bdev->bd_inode->i_mapping->backing_dev_info;
 925         if (root->fs_info->throttles && bdi_write_congested(bdi)) {
 926 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,18)
 927                 congestion_wait(WRITE, HZ/20);
 928 #else
 929                 blk_congestion_wait(WRITE, HZ/20);
 930 #endif
 931         }
 932 }
 933
 934 void btrfs_btree_balance_dirty(struct btrfs_root *root, unsigned long nr)
 935 {
 936         balance_dirty_pages_ratelimited_nr(
 937                         root->fs_info->btree_inode->i_mapping, 1);
 938 }
 939
 940 void btrfs_set_buffer_defrag(struct extent_buffer *buf)
 941 {
 942         struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
 943         struct inode *btree_inode = root->fs_info->btree_inode;
 944         set_extent_bits(&BTRFS_I(btree_inode)->extent_tree, buf->start,
 945                         buf->start + buf->len - 1, EXTENT_DEFRAG, GFP_NOFS);
 946 }
 947
 948 void btrfs_set_buffer_defrag_done(struct extent_buffer *buf)
 949 {
 950         struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
 951         struct inode *btree_inode = root->fs_info->btree_inode;
 952         set_extent_bits(&BTRFS_I(btree_inode)->extent_tree, buf->start,
 953                         buf->start + buf->len - 1, EXTENT_DEFRAG_DONE,
 954                         GFP_NOFS);
 955 }
 956
 957 int btrfs_buffer_defrag(struct extent_buffer *buf)
 958 {
 959         struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
 960         struct inode *btree_inode = root->fs_info->btree_inode;
 961         return test_range_bit(&BTRFS_I(btree_inode)->extent_tree,
 962                      buf->start, buf->start + buf->len - 1, EXTENT_DEFRAG, 0);
 963 }
 964
 965 int btrfs_buffer_defrag_done(struct extent_buffer *buf)
 966 {
 967         struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
 968         struct inode *btree_inode = root->fs_info->btree_inode;
 969         return test_range_bit(&BTRFS_I(btree_inode)->extent_tree,
 970                      buf->start, buf->start + buf->len - 1,
 971                      EXTENT_DEFRAG_DONE, 0);
 972 }
 973
 974 int btrfs_clear_buffer_defrag_done(struct extent_buffer *buf)
 975 {
 976         struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
 977         struct inode *btree_inode = root->fs_info->btree_inode;
 978         return clear_extent_bits(&BTRFS_I(btree_inode)->extent_tree,
 979                      buf->start, buf->start + buf->len - 1,
 980                      EXTENT_DEFRAG_DONE, GFP_NOFS);
 981 }
 982
 983 int btrfs_clear_buffer_defrag(struct extent_buffer *buf)
 984 {
 985         struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
 986         struct inode *btree_inode = root->fs_info->btree_inode;
 987         return clear_extent_bits(&BTRFS_I(btree_inode)->extent_tree,
 988                      buf->start, buf->start + buf->len - 1,
 989                      EXTENT_DEFRAG, GFP_NOFS);
 990 }
 991
 992 int btrfs_read_buffer(struct extent_buffer *buf)
 993 {
 994         struct btrfs_root *root = BTRFS_I(buf->first_page->mapping->host)->root;
 995         struct inode *btree_inode = root->fs_info->btree_inode;
 996         return read_extent_buffer_pages(&BTRFS_I(btree_inode)->extent_tree,
 997                                         buf, 0, 1);
 998 }
 999
1000 static struct extent_map_ops btree_extent_map_ops = {
1001         .writepage_io_hook = btree_writepage_io_hook,
1002 };