fs/btrfs/inode.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/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include "compat.h"
  43 #include "ctree.h"
  44 #include "disk-io.h"
  45 #include "transaction.h"
  46 #include "btrfs_inode.h"
  47 #include "ioctl.h"
  48 #include "print-tree.h"
  49 #include "ordered-data.h"
  50 #include "xattr.h"
  51 #include "tree-log.h"
  52 #include "volumes.h"
  53 #include "compression.h"
  54 #include "locking.h"
  55 #include "free-space-cache.h"
  56 #include "inode-map.h"
  57
  58 struct btrfs_iget_args {
  59         u64 ino;
  60         struct btrfs_root *root;
  61 };
  62
  63 static const struct inode_operations btrfs_dir_inode_operations;
  64 static const struct inode_operations btrfs_symlink_inode_operations;
  65 static const struct inode_operations btrfs_dir_ro_inode_operations;
  66 static const struct inode_operations btrfs_special_inode_operations;
  67 static const struct inode_operations btrfs_file_inode_operations;
  68 static const struct address_space_operations btrfs_aops;
  69 static const struct address_space_operations btrfs_symlink_aops;
  70 static const struct file_operations btrfs_dir_file_operations;
  71 static struct extent_io_ops btrfs_extent_io_ops;
  72
  73 static struct kmem_cache *btrfs_inode_cachep;
  74 struct kmem_cache *btrfs_trans_handle_cachep;
  75 struct kmem_cache *btrfs_transaction_cachep;
  76 struct kmem_cache *btrfs_path_cachep;
  77 struct kmem_cache *btrfs_free_space_cachep;
  78
  79 #define S_SHIFT 12
  80 static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  81         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  82         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  83         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  84         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  85         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  86         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  87         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  88 };
  89
  90 static int btrfs_setsize(struct inode *inode, loff_t newsize);
  91 static int btrfs_truncate(struct inode *inode);
  92 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end);
  93 static noinline int cow_file_range(struct inode *inode,
  94                                    struct page *locked_page,
  95                                    u64 start, u64 end, int *page_started,
  96                                    unsigned long *nr_written, int unlock);
  97 static noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
  98                                 struct btrfs_root *root, struct inode *inode);
  99
 100 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 101                                      struct inode *inode,  struct inode *dir,
 102                                      const struct qstr *qstr)
 103 {
 104         int err;
 105
 106         err = btrfs_init_acl(trans, inode, dir);
 107         if (!err)
 108                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 109         return err;
 110 }
 111
 112 /*
 113  * this does all the hard work for inserting an inline extent into
 114  * the btree.  The caller should have done a btrfs_drop_extents so that
 115  * no overlapping inline items exist in the btree
 116  */
 117 static noinline int insert_inline_extent(struct btrfs_trans_handle *trans,
 118                                 struct btrfs_root *root, struct inode *inode,
 119                                 u64 start, size_t size, size_t compressed_size,
 120                                 int compress_type,
 121                                 struct page **compressed_pages)
 122 {
 123         struct btrfs_key key;
 124         struct btrfs_path *path;
 125         struct extent_buffer *leaf;
 126         struct page *page = NULL;
 127         char *kaddr;
 128         unsigned long ptr;
 129         struct btrfs_file_extent_item *ei;
 130         int err = 0;
 131         int ret;
 132         size_t cur_size = size;
 133         size_t datasize;
 134         unsigned long offset;
 135
 136         if (compressed_size && compressed_pages)
 137                 cur_size = compressed_size;
 138
 139         path = btrfs_alloc_path();
 140         if (!path)
 141                 return -ENOMEM;
 142
 143         path->leave_spinning = 1;
 144
 145         key.objectid = btrfs_ino(inode);
 146         key.offset = start;
 147         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
 148         datasize = btrfs_file_extent_calc_inline_size(cur_size);
 149
 150         inode_add_bytes(inode, size);
 151         ret = btrfs_insert_empty_item(trans, root, path, &key,
 152                                       datasize);
 153         BUG_ON(ret);
 154         if (ret) {
 155                 err = ret;
 156                 goto fail;
 157         }
 158         leaf = path->nodes[0];
 159         ei = btrfs_item_ptr(leaf, path->slots[0],
 160                             struct btrfs_file_extent_item);
 161         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 162         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 163         btrfs_set_file_extent_encryption(leaf, ei, 0);
 164         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 165         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 166         ptr = btrfs_file_extent_inline_start(ei);
 167
 168         if (compress_type != BTRFS_COMPRESS_NONE) {
 169                 struct page *cpage;
 170                 int i = 0;
 171                 while (compressed_size > 0) {
 172                         cpage = compressed_pages[i];
 173                         cur_size = min_t(unsigned long, compressed_size,
 174                                        PAGE_CACHE_SIZE);
 175
 176                         kaddr = kmap_atomic(cpage);
 177                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 178                         kunmap_atomic(kaddr);
 179
 180                         i++;
 181                         ptr += cur_size;
 182                         compressed_size -= cur_size;
 183                 }
 184                 btrfs_set_file_extent_compression(leaf, ei,
 185                                                   compress_type);
 186         } else {
 187                 page = find_get_page(inode->i_mapping,
 188                                      start >> PAGE_CACHE_SHIFT);
 189                 btrfs_set_file_extent_compression(leaf, ei, 0);
 190                 kaddr = kmap_atomic(page);
 191                 offset = start & (PAGE_CACHE_SIZE - 1);
 192                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 193                 kunmap_atomic(kaddr);
 194                 page_cache_release(page);
 195         }
 196         btrfs_mark_buffer_dirty(leaf);
 197         btrfs_free_path(path);
 198
 199         /*
 200          * we're an inline extent, so nobody can
 201          * extend the file past i_size without locking
 202          * a page we already have locked.
 203          *
 204          * We must do any isize and inode updates
 205          * before we unlock the pages.  Otherwise we
 206          * could end up racing with unlink.
 207          */
 208         BTRFS_I(inode)->disk_i_size = inode->i_size;
 209         btrfs_update_inode(trans, root, inode);
 210
 211         return 0;
 212 fail:
 213         btrfs_free_path(path);
 214         return err;
 215 }
 216
 217
 218 /*
 219  * conditionally insert an inline extent into the file.  This
 220  * does the checks required to make sure the data is small enough
 221  * to fit as an inline extent.
 222  */
 223 static noinline int cow_file_range_inline(struct btrfs_trans_handle *trans,
 224                                  struct btrfs_root *root,
 225                                  struct inode *inode, u64 start, u64 end,
 226                                  size_t compressed_size, int compress_type,
 227                                  struct page **compressed_pages)
 228 {
 229         u64 isize = i_size_read(inode);
 230         u64 actual_end = min(end + 1, isize);
 231         u64 inline_len = actual_end - start;
 232         u64 aligned_end = (end + root->sectorsize - 1) &
 233                         ~((u64)root->sectorsize - 1);
 234         u64 hint_byte;
 235         u64 data_len = inline_len;
 236         int ret;
 237
 238         if (compressed_size)
 239                 data_len = compressed_size;
 240
 241         if (start > 0 ||
 242             actual_end >= PAGE_CACHE_SIZE ||
 243             data_len >= BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 244             (!compressed_size &&
 245             (actual_end & (root->sectorsize - 1)) == 0) ||
 246             end + 1 < isize ||
 247             data_len > root->fs_info->max_inline) {
 248                 return 1;
 249         }
 250
 251         ret = btrfs_drop_extents(trans, inode, start, aligned_end,
 252                                  &hint_byte, 1);
 253         BUG_ON(ret);
 254
 255         if (isize > actual_end)
 256                 inline_len = min_t(u64, isize, actual_end);
 257         ret = insert_inline_extent(trans, root, inode, start,
 258                                    inline_len, compressed_size,
 259                                    compress_type, compressed_pages);
 260         BUG_ON(ret);
 261         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 262         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 263         return 0;
 264 }
 265
 266 struct async_extent {
 267         u64 start;
 268         u64 ram_size;
 269         u64 compressed_size;
 270         struct page **pages;
 271         unsigned long nr_pages;
 272         int compress_type;
 273         struct list_head list;
 274 };
 275
 276 struct async_cow {
 277         struct inode *inode;
 278         struct btrfs_root *root;
 279         struct page *locked_page;
 280         u64 start;
 281         u64 end;
 282         struct list_head extents;
 283         struct btrfs_work work;
 284 };
 285
 286 static noinline int add_async_extent(struct async_cow *cow,
 287                                      u64 start, u64 ram_size,
 288                                      u64 compressed_size,
 289                                      struct page **pages,
 290                                      unsigned long nr_pages,
 291                                      int compress_type)
 292 {
 293         struct async_extent *async_extent;
 294
 295         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 296         BUG_ON(!async_extent);
 297         async_extent->start = start;
 298         async_extent->ram_size = ram_size;
 299         async_extent->compressed_size = compressed_size;
 300         async_extent->pages = pages;
 301         async_extent->nr_pages = nr_pages;
 302         async_extent->compress_type = compress_type;
 303         list_add_tail(&async_extent->list, &cow->extents);
 304         return 0;
 305 }
 306
 307 /*
 308  * we create compressed extents in two phases.  The first
 309  * phase compresses a range of pages that have already been
 310  * locked (both pages and state bits are locked).
 311  *
 312  * This is done inside an ordered work queue, and the compression
 313  * is spread across many cpus.  The actual IO submission is step
 314  * two, and the ordered work queue takes care of making sure that
 315  * happens in the same order things were put onto the queue by
 316  * writepages and friends.
 317  *
 318  * If this code finds it can't get good compression, it puts an
 319  * entry onto the work queue to write the uncompressed bytes.  This
 320  * makes sure that both compressed inodes and uncompressed inodes
 321  * are written in the same order that pdflush sent them down.
 322  */
 323 static noinline int compress_file_range(struct inode *inode,
 324                                         struct page *locked_page,
 325                                         u64 start, u64 end,
 326                                         struct async_cow *async_cow,
 327                                         int *num_added)
 328 {
 329         struct btrfs_root *root = BTRFS_I(inode)->root;
 330         struct btrfs_trans_handle *trans;
 331         u64 num_bytes;
 332         u64 blocksize = root->sectorsize;
 333         u64 actual_end;
 334         u64 isize = i_size_read(inode);
 335         int ret = 0;
 336         struct page **pages = NULL;
 337         unsigned long nr_pages;
 338         unsigned long nr_pages_ret = 0;
 339         unsigned long total_compressed = 0;
 340         unsigned long total_in = 0;
 341         unsigned long max_compressed = 128 * 1024;
 342         unsigned long max_uncompressed = 128 * 1024;
 343         int i;
 344         int will_compress;
 345         int compress_type = root->fs_info->compress_type;
 346
 347         /* if this is a small write inside eof, kick off a defragbot */
 348         if (end <= BTRFS_I(inode)->disk_i_size && (end - start + 1) < 16 * 1024)
 349                 btrfs_add_inode_defrag(NULL, inode);
 350
 351         actual_end = min_t(u64, isize, end + 1);
 352 again:
 353         will_compress = 0;
 354         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 355         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
 356
 357         /*
 358          * we don't want to send crud past the end of i_size through
 359          * compression, that's just a waste of CPU time.  So, if the
 360          * end of the file is before the start of our current
 361          * requested range of bytes, we bail out to the uncompressed
 362          * cleanup code that can deal with all of this.
 363          *
 364          * It isn't really the fastest way to fix things, but this is a
 365          * very uncommon corner.
 366          */
 367         if (actual_end <= start)
 368                 goto cleanup_and_bail_uncompressed;
 369
 370         total_compressed = actual_end - start;
 371
 372         /* we want to make sure that amount of ram required to uncompress
 373          * an extent is reasonable, so we limit the total size in ram
 374          * of a compressed extent to 128k.  This is a crucial number
 375          * because it also controls how easily we can spread reads across
 376          * cpus for decompression.
 377          *
 378          * We also want to make sure the amount of IO required to do
 379          * a random read is reasonably small, so we limit the size of
 380          * a compressed extent to 128k.
 381          */
 382         total_compressed = min(total_compressed, max_uncompressed);
 383         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
 384         num_bytes = max(blocksize,  num_bytes);
 385         total_in = 0;
 386         ret = 0;
 387
 388         /*
 389          * we do compression for mount -o compress and when the
 390          * inode has not been flagged as nocompress.  This flag can
 391          * change at any time if we discover bad compression ratios.
 392          */
 393         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS) &&
 394             (btrfs_test_opt(root, COMPRESS) ||
 395              (BTRFS_I(inode)->force_compress) ||
 396              (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS))) {
 397                 WARN_ON(pages);
 398                 pages = kzalloc(sizeof(struct page *) * nr_pages, GFP_NOFS);
 399                 if (!pages) {
 400                         /* just bail out to the uncompressed code */
 401                         goto cont;
 402                 }
 403
 404                 if (BTRFS_I(inode)->force_compress)
 405                         compress_type = BTRFS_I(inode)->force_compress;
 406
 407                 ret = btrfs_compress_pages(compress_type,
 408                                            inode->i_mapping, start,
 409                                            total_compressed, pages,
 410                                            nr_pages, &nr_pages_ret,
 411                                            &total_in,
 412                                            &total_compressed,
 413                                            max_compressed);
 414
 415                 if (!ret) {
 416                         unsigned long offset = total_compressed &
 417                                 (PAGE_CACHE_SIZE - 1);
 418                         struct page *page = pages[nr_pages_ret - 1];
 419                         char *kaddr;
 420
 421                         /* zero the tail end of the last page, we might be
 422                          * sending it down to disk
 423                          */
 424                         if (offset) {
 425                                 kaddr = kmap_atomic(page);
 426                                 memset(kaddr + offset, 0,
 427                                        PAGE_CACHE_SIZE - offset);
 428                                 kunmap_atomic(kaddr);
 429                         }
 430                         will_compress = 1;
 431                 }
 432         }
 433 cont:
 434         if (start == 0) {
 435                 trans = btrfs_join_transaction(root);
 436                 BUG_ON(IS_ERR(trans));
 437                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 438
 439                 /* lets try to make an inline extent */
 440                 if (ret || total_in < (actual_end - start)) {
 441                         /* we didn't compress the entire range, try
 442                          * to make an uncompressed inline extent.
 443                          */
 444                         ret = cow_file_range_inline(trans, root, inode,
 445                                                     start, end, 0, 0, NULL);
 446                 } else {
 447                         /* try making a compressed inline extent */
 448                         ret = cow_file_range_inline(trans, root, inode,
 449                                                     start, end,
 450                                                     total_compressed,
 451                                                     compress_type, pages);
 452                 }
 453                 if (ret == 0) {
 454                         /*
 455                          * inline extent creation worked, we don't need
 456                          * to create any more async work items.  Unlock
 457                          * and free up our temp pages.
 458                          */
 459                         extent_clear_unlock_delalloc(inode,
 460                              &BTRFS_I(inode)->io_tree,
 461                              start, end, NULL,
 462                              EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
 463                              EXTENT_CLEAR_DELALLOC |
 464                              EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK);
 465
 466                         btrfs_end_transaction(trans, root);
 467                         goto free_pages_out;
 468                 }
 469                 btrfs_end_transaction(trans, root);
 470         }
 471
 472         if (will_compress) {
 473                 /*
 474                  * we aren't doing an inline extent round the compressed size
 475                  * up to a block size boundary so the allocator does sane
 476                  * things
 477                  */
 478                 total_compressed = (total_compressed + blocksize - 1) &
 479                         ~(blocksize - 1);
 480
 481                 /*
 482                  * one last check to make sure the compression is really a
 483                  * win, compare the page count read with the blocks on disk
 484                  */
 485                 total_in = (total_in + PAGE_CACHE_SIZE - 1) &
 486                         ~(PAGE_CACHE_SIZE - 1);
 487                 if (total_compressed >= total_in) {
 488                         will_compress = 0;
 489                 } else {
 490                         num_bytes = total_in;
 491                 }
 492         }
 493         if (!will_compress && pages) {
 494                 /*
 495                  * the compression code ran but failed to make things smaller,
 496                  * free any pages it allocated and our page pointer array
 497                  */
 498                 for (i = 0; i < nr_pages_ret; i++) {
 499                         WARN_ON(pages[i]->mapping);
 500                         page_cache_release(pages[i]);
 501                 }
 502                 kfree(pages);
 503                 pages = NULL;
 504                 total_compressed = 0;
 505                 nr_pages_ret = 0;
 506
 507                 /* flag the file so we don't compress in the future */
 508                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 509                     !(BTRFS_I(inode)->force_compress)) {
 510                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 511                 }
 512         }
 513         if (will_compress) {
 514                 *num_added += 1;
 515
 516                 /* the async work queues will take care of doing actual
 517                  * allocation on disk for these compressed pages,
 518                  * and will submit them to the elevator.
 519                  */
 520                 add_async_extent(async_cow, start, num_bytes,
 521                                  total_compressed, pages, nr_pages_ret,
 522                                  compress_type);
 523
 524                 if (start + num_bytes < end) {
 525                         start += num_bytes;
 526                         pages = NULL;
 527                         cond_resched();
 528                         goto again;
 529                 }
 530         } else {
 531 cleanup_and_bail_uncompressed:
 532                 /*
 533                  * No compression, but we still need to write the pages in
 534                  * the file we've been given so far.  redirty the locked
 535                  * page if it corresponds to our extent and set things up
 536                  * for the async work queue to run cow_file_range to do
 537                  * the normal delalloc dance
 538                  */
 539                 if (page_offset(locked_page) >= start &&
 540                     page_offset(locked_page) <= end) {
 541                         __set_page_dirty_nobuffers(locked_page);
 542                         /* unlocked later on in the async handlers */
 543                 }
 544                 add_async_extent(async_cow, start, end - start + 1,
 545                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 546                 *num_added += 1;
 547         }
 548
 549 out:
 550         return 0;
 551
 552 free_pages_out:
 553         for (i = 0; i < nr_pages_ret; i++) {
 554                 WARN_ON(pages[i]->mapping);
 555                 page_cache_release(pages[i]);
 556         }
 557         kfree(pages);
 558
 559         goto out;
 560 }
 561
 562 /*
 563  * phase two of compressed writeback.  This is the ordered portion
 564  * of the code, which only gets called in the order the work was
 565  * queued.  We walk all the async extents created by compress_file_range
 566  * and send them down to the disk.
 567  */
 568 static noinline int submit_compressed_extents(struct inode *inode,
 569                                               struct async_cow *async_cow)
 570 {
 571         struct async_extent *async_extent;
 572         u64 alloc_hint = 0;
 573         struct btrfs_trans_handle *trans;
 574         struct btrfs_key ins;
 575         struct extent_map *em;
 576         struct btrfs_root *root = BTRFS_I(inode)->root;
 577         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 578         struct extent_io_tree *io_tree;
 579         int ret = 0;
 580
 581         if (list_empty(&async_cow->extents))
 582                 return 0;
 583
 584
 585         while (!list_empty(&async_cow->extents)) {
 586                 async_extent = list_entry(async_cow->extents.next,
 587                                           struct async_extent, list);
 588                 list_del(&async_extent->list);
 589
 590                 io_tree = &BTRFS_I(inode)->io_tree;
 591
 592 retry:
 593                 /* did the compression code fall back to uncompressed IO? */
 594                 if (!async_extent->pages) {
 595                         int page_started = 0;
 596                         unsigned long nr_written = 0;
 597
 598                         lock_extent(io_tree, async_extent->start,
 599                                          async_extent->start +
 600                                          async_extent->ram_size - 1, GFP_NOFS);
 601
 602                         /* allocate blocks */
 603                         ret = cow_file_range(inode, async_cow->locked_page,
 604                                              async_extent->start,
 605                                              async_extent->start +
 606                                              async_extent->ram_size - 1,
 607                                              &page_started, &nr_written, 0);
 608
 609                         /*
 610                          * if page_started, cow_file_range inserted an
 611                          * inline extent and took care of all the unlocking
 612                          * and IO for us.  Otherwise, we need to submit
 613                          * all those pages down to the drive.
 614                          */
 615                         if (!page_started && !ret)
 616                                 extent_write_locked_range(io_tree,
 617                                                   inode, async_extent->start,
 618                                                   async_extent->start +
 619                                                   async_extent->ram_size - 1,
 620                                                   btrfs_get_extent,
 621                                                   WB_SYNC_ALL);
 622                         kfree(async_extent);
 623                         cond_resched();
 624                         continue;
 625                 }
 626
 627                 lock_extent(io_tree, async_extent->start,
 628                             async_extent->start + async_extent->ram_size - 1,
 629                             GFP_NOFS);
 630
 631                 trans = btrfs_join_transaction(root);
 632                 BUG_ON(IS_ERR(trans));
 633                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 634                 ret = btrfs_reserve_extent(trans, root,
 635                                            async_extent->compressed_size,
 636                                            async_extent->compressed_size,
 637                                            0, alloc_hint,
 638                                            (u64)-1, &ins, 1);
 639                 btrfs_end_transaction(trans, root);
 640
 641                 if (ret) {
 642                         int i;
 643                         for (i = 0; i < async_extent->nr_pages; i++) {
 644                                 WARN_ON(async_extent->pages[i]->mapping);
 645                                 page_cache_release(async_extent->pages[i]);
 646                         }
 647                         kfree(async_extent->pages);
 648                         async_extent->nr_pages = 0;
 649                         async_extent->pages = NULL;
 650                         unlock_extent(io_tree, async_extent->start,
 651                                       async_extent->start +
 652                                       async_extent->ram_size - 1, GFP_NOFS);
 653                         goto retry;
 654                 }
 655
 656                 /*
 657                  * here we're doing allocation and writeback of the
 658                  * compressed pages
 659                  */
 660                 btrfs_drop_extent_cache(inode, async_extent->start,
 661                                         async_extent->start +
 662                                         async_extent->ram_size - 1, 0);
 663
 664                 em = alloc_extent_map();
 665                 BUG_ON(!em);
 666                 em->start = async_extent->start;
 667                 em->len = async_extent->ram_size;
 668                 em->orig_start = em->start;
 669
 670                 em->block_start = ins.objectid;
 671                 em->block_len = ins.offset;
 672                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 673                 em->compress_type = async_extent->compress_type;
 674                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 675                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 676
 677                 while (1) {
 678                         write_lock(&em_tree->lock);
 679                         ret = add_extent_mapping(em_tree, em);
 680                         write_unlock(&em_tree->lock);
 681                         if (ret != -EEXIST) {
 682                                 free_extent_map(em);
 683                                 break;
 684                         }
 685                         btrfs_drop_extent_cache(inode, async_extent->start,
 686                                                 async_extent->start +
 687                                                 async_extent->ram_size - 1, 0);
 688                 }
 689
 690                 ret = btrfs_add_ordered_extent_compress(inode,
 691                                                 async_extent->start,
 692                                                 ins.objectid,
 693                                                 async_extent->ram_size,
 694                                                 ins.offset,
 695                                                 BTRFS_ORDERED_COMPRESSED,
 696                                                 async_extent->compress_type);
 697                 BUG_ON(ret);
 698
 699                 /*
 700                  * clear dirty, set writeback and unlock the pages.
 701                  */
 702                 extent_clear_unlock_delalloc(inode,
 703                                 &BTRFS_I(inode)->io_tree,
 704                                 async_extent->start,
 705                                 async_extent->start +
 706                                 async_extent->ram_size - 1,
 707                                 NULL, EXTENT_CLEAR_UNLOCK_PAGE |
 708                                 EXTENT_CLEAR_UNLOCK |
 709                                 EXTENT_CLEAR_DELALLOC |
 710                                 EXTENT_CLEAR_DIRTY | EXTENT_SET_WRITEBACK);
 711
 712                 ret = btrfs_submit_compressed_write(inode,
 713                                     async_extent->start,
 714                                     async_extent->ram_size,
 715                                     ins.objectid,
 716                                     ins.offset, async_extent->pages,
 717                                     async_extent->nr_pages);
 718
 719                 BUG_ON(ret);
 720                 alloc_hint = ins.objectid + ins.offset;
 721                 kfree(async_extent);
 722                 cond_resched();
 723         }
 724
 725         return 0;
 726 }
 727
 728 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 729                                       u64 num_bytes)
 730 {
 731         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 732         struct extent_map *em;
 733         u64 alloc_hint = 0;
 734
 735         read_lock(&em_tree->lock);
 736         em = search_extent_mapping(em_tree, start, num_bytes);
 737         if (em) {
 738                 /*
 739                  * if block start isn't an actual block number then find the
 740                  * first block in this inode and use that as a hint.  If that
 741                  * block is also bogus then just don't worry about it.
 742                  */
 743                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 744                         free_extent_map(em);
 745                         em = search_extent_mapping(em_tree, 0, 0);
 746                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 747                                 alloc_hint = em->block_start;
 748                         if (em)
 749                                 free_extent_map(em);
 750                 } else {
 751                         alloc_hint = em->block_start;
 752                         free_extent_map(em);
 753                 }
 754         }
 755         read_unlock(&em_tree->lock);
 756
 757         return alloc_hint;
 758 }
 759
 760 /*
 761  * when extent_io.c finds a delayed allocation range in the file,
 762  * the call backs end up in this code.  The basic idea is to
 763  * allocate extents on disk for the range, and create ordered data structs
 764  * in ram to track those extents.
 765  *
 766  * locked_page is the page that writepage had locked already.  We use
 767  * it to make sure we don't do extra locks or unlocks.
 768  *
 769  * *page_started is set to one if we unlock locked_page and do everything
 770  * required to start IO on it.  It may be clean and already done with
 771  * IO when we return.
 772  */
 773 static noinline int cow_file_range(struct inode *inode,
 774                                    struct page *locked_page,
 775                                    u64 start, u64 end, int *page_started,
 776                                    unsigned long *nr_written,
 777                                    int unlock)
 778 {
 779         struct btrfs_root *root = BTRFS_I(inode)->root;
 780         struct btrfs_trans_handle *trans;
 781         u64 alloc_hint = 0;
 782         u64 num_bytes;
 783         unsigned long ram_size;
 784         u64 disk_num_bytes;
 785         u64 cur_alloc_size;
 786         u64 blocksize = root->sectorsize;
 787         struct btrfs_key ins;
 788         struct extent_map *em;
 789         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 790         int ret = 0;
 791
 792         BUG_ON(btrfs_is_free_space_inode(root, inode));
 793         trans = btrfs_join_transaction(root);
 794         BUG_ON(IS_ERR(trans));
 795         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 796
 797         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
 798         num_bytes = max(blocksize,  num_bytes);
 799         disk_num_bytes = num_bytes;
 800         ret = 0;
 801
 802         /* if this is a small write inside eof, kick off defrag */
 803         if (end <= BTRFS_I(inode)->disk_i_size && num_bytes < 64 * 1024)
 804                 btrfs_add_inode_defrag(trans, inode);
 805
 806         if (start == 0) {
 807                 /* lets try to make an inline extent */
 808                 ret = cow_file_range_inline(trans, root, inode,
 809                                             start, end, 0, 0, NULL);
 810                 if (ret == 0) {
 811                         extent_clear_unlock_delalloc(inode,
 812                                      &BTRFS_I(inode)->io_tree,
 813                                      start, end, NULL,
 814                                      EXTENT_CLEAR_UNLOCK_PAGE |
 815                                      EXTENT_CLEAR_UNLOCK |
 816                                      EXTENT_CLEAR_DELALLOC |
 817                                      EXTENT_CLEAR_DIRTY |
 818                                      EXTENT_SET_WRITEBACK |
 819                                      EXTENT_END_WRITEBACK);
 820
 821                         *nr_written = *nr_written +
 822                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 823                         *page_started = 1;
 824                         ret = 0;
 825                         goto out;
 826                 }
 827         }
 828
 829         BUG_ON(disk_num_bytes >
 830                btrfs_super_total_bytes(root->fs_info->super_copy));
 831
 832         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 833         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 834
 835         while (disk_num_bytes > 0) {
 836                 unsigned long op;
 837
 838                 cur_alloc_size = disk_num_bytes;
 839                 ret = btrfs_reserve_extent(trans, root, cur_alloc_size,
 840                                            root->sectorsize, 0, alloc_hint,
 841                                            (u64)-1, &ins, 1);
 842                 BUG_ON(ret);
 843
 844                 em = alloc_extent_map();
 845                 BUG_ON(!em);
 846                 em->start = start;
 847                 em->orig_start = em->start;
 848                 ram_size = ins.offset;
 849                 em->len = ins.offset;
 850
 851                 em->block_start = ins.objectid;
 852                 em->block_len = ins.offset;
 853                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 854                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 855
 856                 while (1) {
 857                         write_lock(&em_tree->lock);
 858                         ret = add_extent_mapping(em_tree, em);
 859                         write_unlock(&em_tree->lock);
 860                         if (ret != -EEXIST) {
 861                                 free_extent_map(em);
 862                                 break;
 863                         }
 864                         btrfs_drop_extent_cache(inode, start,
 865                                                 start + ram_size - 1, 0);
 866                 }
 867
 868                 cur_alloc_size = ins.offset;
 869                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
 870                                                ram_size, cur_alloc_size, 0);
 871                 BUG_ON(ret);
 872
 873                 if (root->root_key.objectid ==
 874                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
 875                         ret = btrfs_reloc_clone_csums(inode, start,
 876                                                       cur_alloc_size);
 877                         BUG_ON(ret);
 878                 }
 879
 880                 if (disk_num_bytes < cur_alloc_size)
 881                         break;
 882
 883                 /* we're not doing compressed IO, don't unlock the first
 884                  * page (which the caller expects to stay locked), don't
 885                  * clear any dirty bits and don't set any writeback bits
 886                  *
 887                  * Do set the Private2 bit so we know this page was properly
 888                  * setup for writepage
 889                  */
 890                 op = unlock ? EXTENT_CLEAR_UNLOCK_PAGE : 0;
 891                 op |= EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
 892                         EXTENT_SET_PRIVATE2;
 893
 894                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
 895                                              start, start + ram_size - 1,
 896                                              locked_page, op);
 897                 disk_num_bytes -= cur_alloc_size;
 898                 num_bytes -= cur_alloc_size;
 899                 alloc_hint = ins.objectid + ins.offset;
 900                 start += cur_alloc_size;
 901         }
 902 out:
 903         ret = 0;
 904         btrfs_end_transaction(trans, root);
 905
 906         return ret;
 907 }
 908
 909 /*
 910  * work queue call back to started compression on a file and pages
 911  */
 912 static noinline void async_cow_start(struct btrfs_work *work)
 913 {
 914         struct async_cow *async_cow;
 915         int num_added = 0;
 916         async_cow = container_of(work, struct async_cow, work);
 917
 918         compress_file_range(async_cow->inode, async_cow->locked_page,
 919                             async_cow->start, async_cow->end, async_cow,
 920                             &num_added);
 921         if (num_added == 0)
 922                 async_cow->inode = NULL;
 923 }
 924
 925 /*
 926  * work queue call back to submit previously compressed pages
 927  */
 928 static noinline void async_cow_submit(struct btrfs_work *work)
 929 {
 930         struct async_cow *async_cow;
 931         struct btrfs_root *root;
 932         unsigned long nr_pages;
 933
 934         async_cow = container_of(work, struct async_cow, work);
 935
 936         root = async_cow->root;
 937         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
 938                 PAGE_CACHE_SHIFT;
 939
 940         atomic_sub(nr_pages, &root->fs_info->async_delalloc_pages);
 941
 942         if (atomic_read(&root->fs_info->async_delalloc_pages) <
 943             5 * 1042 * 1024 &&
 944             waitqueue_active(&root->fs_info->async_submit_wait))
 945                 wake_up(&root->fs_info->async_submit_wait);
 946
 947         if (async_cow->inode)
 948                 submit_compressed_extents(async_cow->inode, async_cow);
 949 }
 950
 951 static noinline void async_cow_free(struct btrfs_work *work)
 952 {
 953         struct async_cow *async_cow;
 954         async_cow = container_of(work, struct async_cow, work);
 955         kfree(async_cow);
 956 }
 957
 958 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
 959                                 u64 start, u64 end, int *page_started,
 960                                 unsigned long *nr_written)
 961 {
 962         struct async_cow *async_cow;
 963         struct btrfs_root *root = BTRFS_I(inode)->root;
 964         unsigned long nr_pages;
 965         u64 cur_end;
 966         int limit = 10 * 1024 * 1042;
 967
 968         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
 969                          1, 0, NULL, GFP_NOFS);
 970         while (start < end) {
 971                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
 972                 BUG_ON(!async_cow);
 973                 async_cow->inode = inode;
 974                 async_cow->root = root;
 975                 async_cow->locked_page = locked_page;
 976                 async_cow->start = start;
 977
 978                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 979                         cur_end = end;
 980                 else
 981                         cur_end = min(end, start + 512 * 1024 - 1);
 982
 983                 async_cow->end = cur_end;
 984                 INIT_LIST_HEAD(&async_cow->extents);
 985
 986                 async_cow->work.func = async_cow_start;
 987                 async_cow->work.ordered_func = async_cow_submit;
 988                 async_cow->work.ordered_free = async_cow_free;
 989                 async_cow->work.flags = 0;
 990
 991                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
 992                         PAGE_CACHE_SHIFT;
 993                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
 994
 995                 btrfs_queue_worker(&root->fs_info->delalloc_workers,
 996                                    &async_cow->work);
 997
 998                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
 999                         wait_event(root->fs_info->async_submit_wait,
1000                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1001                             limit));
1002                 }
1003
1004                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1005                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1006                         wait_event(root->fs_info->async_submit_wait,
1007                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1008                            0));
1009                 }
1010
1011                 *nr_written += nr_pages;
1012                 start = cur_end + 1;
1013         }
1014         *page_started = 1;
1015         return 0;
1016 }
1017
1018 static noinline int csum_exist_in_range(struct btrfs_root *root,
1019                                         u64 bytenr, u64 num_bytes)
1020 {
1021         int ret;
1022         struct btrfs_ordered_sum *sums;
1023         LIST_HEAD(list);
1024
1025         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1026                                        bytenr + num_bytes - 1, &list, 0);
1027         if (ret == 0 && list_empty(&list))
1028                 return 0;
1029
1030         while (!list_empty(&list)) {
1031                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1032                 list_del(&sums->list);
1033                 kfree(sums);
1034         }
1035         return 1;
1036 }
1037
1038 /*
1039  * when nowcow writeback call back.  This checks for snapshots or COW copies
1040  * of the extents that exist in the file, and COWs the file as required.
1041  *
1042  * If no cow copies or snapshots exist, we write directly to the existing
1043  * blocks on disk
1044  */
1045 static noinline int run_delalloc_nocow(struct inode *inode,
1046                                        struct page *locked_page,
1047                               u64 start, u64 end, int *page_started, int force,
1048                               unsigned long *nr_written)
1049 {
1050         struct btrfs_root *root = BTRFS_I(inode)->root;
1051         struct btrfs_trans_handle *trans;
1052         struct extent_buffer *leaf;
1053         struct btrfs_path *path;
1054         struct btrfs_file_extent_item *fi;
1055         struct btrfs_key found_key;
1056         u64 cow_start;
1057         u64 cur_offset;
1058         u64 extent_end;
1059         u64 extent_offset;
1060         u64 disk_bytenr;
1061         u64 num_bytes;
1062         int extent_type;
1063         int ret;
1064         int type;
1065         int nocow;
1066         int check_prev = 1;
1067         bool nolock;
1068         u64 ino = btrfs_ino(inode);
1069
1070         path = btrfs_alloc_path();
1071         if (!path)
1072                 return -ENOMEM;
1073
1074         nolock = btrfs_is_free_space_inode(root, inode);
1075
1076         if (nolock)
1077                 trans = btrfs_join_transaction_nolock(root);
1078         else
1079                 trans = btrfs_join_transaction(root);
1080
1081         BUG_ON(IS_ERR(trans));
1082         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1083
1084         cow_start = (u64)-1;
1085         cur_offset = start;
1086         while (1) {
1087                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1088                                                cur_offset, 0);
1089                 BUG_ON(ret < 0);
1090                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1091                         leaf = path->nodes[0];
1092                         btrfs_item_key_to_cpu(leaf, &found_key,
1093                                               path->slots[0] - 1);
1094                         if (found_key.objectid == ino &&
1095                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1096                                 path->slots[0]--;
1097                 }
1098                 check_prev = 0;
1099 next_slot:
1100                 leaf = path->nodes[0];
1101                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1102                         ret = btrfs_next_leaf(root, path);
1103                         if (ret < 0)
1104                                 BUG_ON(1);
1105                         if (ret > 0)
1106                                 break;
1107                         leaf = path->nodes[0];
1108                 }
1109
1110                 nocow = 0;
1111                 disk_bytenr = 0;
1112                 num_bytes = 0;
1113                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1114
1115                 if (found_key.objectid > ino ||
1116                     found_key.type > BTRFS_EXTENT_DATA_KEY ||
1117                     found_key.offset > end)
1118                         break;
1119
1120                 if (found_key.offset > cur_offset) {
1121                         extent_end = found_key.offset;
1122                         extent_type = 0;
1123                         goto out_check;
1124                 }
1125
1126                 fi = btrfs_item_ptr(leaf, path->slots[0],
1127                                     struct btrfs_file_extent_item);
1128                 extent_type = btrfs_file_extent_type(leaf, fi);
1129
1130                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1131                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1132                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1133                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1134                         extent_end = found_key.offset +
1135                                 btrfs_file_extent_num_bytes(leaf, fi);
1136                         if (extent_end <= start) {
1137                                 path->slots[0]++;
1138                                 goto next_slot;
1139                         }
1140                         if (disk_bytenr == 0)
1141                                 goto out_check;
1142                         if (btrfs_file_extent_compression(leaf, fi) ||
1143                             btrfs_file_extent_encryption(leaf, fi) ||
1144                             btrfs_file_extent_other_encoding(leaf, fi))
1145                                 goto out_check;
1146                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1147                                 goto out_check;
1148                         if (btrfs_extent_readonly(root, disk_bytenr))
1149                                 goto out_check;
1150                         if (btrfs_cross_ref_exist(trans, root, ino,
1151                                                   found_key.offset -
1152                                                   extent_offset, disk_bytenr))
1153                                 goto out_check;
1154                         disk_bytenr += extent_offset;
1155                         disk_bytenr += cur_offset - found_key.offset;
1156                         num_bytes = min(end + 1, extent_end) - cur_offset;
1157                         /*
1158                          * force cow if csum exists in the range.
1159                          * this ensure that csum for a given extent are
1160                          * either valid or do not exist.
1161                          */
1162                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1163                                 goto out_check;
1164                         nocow = 1;
1165                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1166                         extent_end = found_key.offset +
1167                                 btrfs_file_extent_inline_len(leaf, fi);
1168                         extent_end = ALIGN(extent_end, root->sectorsize);
1169                 } else {
1170                         BUG_ON(1);
1171                 }
1172 out_check:
1173                 if (extent_end <= start) {
1174                         path->slots[0]++;
1175                         goto next_slot;
1176                 }
1177                 if (!nocow) {
1178                         if (cow_start == (u64)-1)
1179                                 cow_start = cur_offset;
1180                         cur_offset = extent_end;
1181                         if (cur_offset > end)
1182                                 break;
1183                         path->slots[0]++;
1184                         goto next_slot;
1185                 }
1186
1187                 btrfs_release_path(path);
1188                 if (cow_start != (u64)-1) {
1189                         ret = cow_file_range(inode, locked_page, cow_start,
1190                                         found_key.offset - 1, page_started,
1191                                         nr_written, 1);
1192                         BUG_ON(ret);
1193                         cow_start = (u64)-1;
1194                 }
1195
1196                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1197                         struct extent_map *em;
1198                         struct extent_map_tree *em_tree;
1199                         em_tree = &BTRFS_I(inode)->extent_tree;
1200                         em = alloc_extent_map();
1201                         BUG_ON(!em);
1202                         em->start = cur_offset;
1203                         em->orig_start = em->start;
1204                         em->len = num_bytes;
1205                         em->block_len = num_bytes;
1206                         em->block_start = disk_bytenr;
1207                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1208                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1209                         while (1) {
1210                                 write_lock(&em_tree->lock);
1211                                 ret = add_extent_mapping(em_tree, em);
1212                                 write_unlock(&em_tree->lock);
1213                                 if (ret != -EEXIST) {
1214                                         free_extent_map(em);
1215                                         break;
1216                                 }
1217                                 btrfs_drop_extent_cache(inode, em->start,
1218                                                 em->start + em->len - 1, 0);
1219                         }
1220                         type = BTRFS_ORDERED_PREALLOC;
1221                 } else {
1222                         type = BTRFS_ORDERED_NOCOW;
1223                 }
1224
1225                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1226                                                num_bytes, num_bytes, type);
1227                 BUG_ON(ret);
1228
1229                 if (root->root_key.objectid ==
1230                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1231                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1232                                                       num_bytes);
1233                         BUG_ON(ret);
1234                 }
1235
1236                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
1237                                 cur_offset, cur_offset + num_bytes - 1,
1238                                 locked_page, EXTENT_CLEAR_UNLOCK_PAGE |
1239                                 EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
1240                                 EXTENT_SET_PRIVATE2);
1241                 cur_offset = extent_end;
1242                 if (cur_offset > end)
1243                         break;
1244         }
1245         btrfs_release_path(path);
1246
1247         if (cur_offset <= end && cow_start == (u64)-1)
1248                 cow_start = cur_offset;
1249         if (cow_start != (u64)-1) {
1250                 ret = cow_file_range(inode, locked_page, cow_start, end,
1251                                      page_started, nr_written, 1);
1252                 BUG_ON(ret);
1253         }
1254
1255         if (nolock) {
1256                 ret = btrfs_end_transaction_nolock(trans, root);
1257                 BUG_ON(ret);
1258         } else {
1259                 ret = btrfs_end_transaction(trans, root);
1260                 BUG_ON(ret);
1261         }
1262         btrfs_free_path(path);
1263         return 0;
1264 }
1265
1266 /*
1267  * extent_io.c call back to do delayed allocation processing
1268  */
1269 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1270                               u64 start, u64 end, int *page_started,
1271                               unsigned long *nr_written)
1272 {
1273         int ret;
1274         struct btrfs_root *root = BTRFS_I(inode)->root;
1275
1276         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)
1277                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1278                                          page_started, 1, nr_written);
1279         else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)
1280                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1281                                          page_started, 0, nr_written);
1282         else if (!btrfs_test_opt(root, COMPRESS) &&
1283                  !(BTRFS_I(inode)->force_compress) &&
1284                  !(BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS))
1285                 ret = cow_file_range(inode, locked_page, start, end,
1286                                       page_started, nr_written, 1);
1287         else
1288                 ret = cow_file_range_async(inode, locked_page, start, end,
1289                                            page_started, nr_written);
1290         return ret;
1291 }
1292
1293 static void btrfs_split_extent_hook(struct inode *inode,
1294                                     struct extent_state *orig, u64 split)
1295 {
1296         /* not delalloc, ignore it */
1297         if (!(orig->state & EXTENT_DELALLOC))
1298                 return;
1299
1300         spin_lock(&BTRFS_I(inode)->lock);
1301         BTRFS_I(inode)->outstanding_extents++;
1302         spin_unlock(&BTRFS_I(inode)->lock);
1303 }
1304
1305 /*
1306  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1307  * extents so we can keep track of new extents that are just merged onto old
1308  * extents, such as when we are doing sequential writes, so we can properly
1309  * account for the metadata space we'll need.
1310  */
1311 static void btrfs_merge_extent_hook(struct inode *inode,
1312                                     struct extent_state *new,
1313                                     struct extent_state *other)
1314 {
1315         /* not delalloc, ignore it */
1316         if (!(other->state & EXTENT_DELALLOC))
1317                 return;
1318
1319         spin_lock(&BTRFS_I(inode)->lock);
1320         BTRFS_I(inode)->outstanding_extents--;
1321         spin_unlock(&BTRFS_I(inode)->lock);
1322 }
1323
1324 /*
1325  * extent_io.c set_bit_hook, used to track delayed allocation
1326  * bytes in this file, and to maintain the list of inodes that
1327  * have pending delalloc work to be done.
1328  */
1329 static void btrfs_set_bit_hook(struct inode *inode,
1330                                struct extent_state *state, int *bits)
1331 {
1332
1333         /*
1334          * set_bit and clear bit hooks normally require _irqsave/restore
1335          * but in this case, we are only testing for the DELALLOC
1336          * bit, which is only set or cleared with irqs on
1337          */
1338         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1339                 struct btrfs_root *root = BTRFS_I(inode)->root;
1340                 u64 len = state->end + 1 - state->start;
1341                 bool do_list = !btrfs_is_free_space_inode(root, inode);
1342
1343                 if (*bits & EXTENT_FIRST_DELALLOC) {
1344                         *bits &= ~EXTENT_FIRST_DELALLOC;
1345                 } else {
1346                         spin_lock(&BTRFS_I(inode)->lock);
1347                         BTRFS_I(inode)->outstanding_extents++;
1348                         spin_unlock(&BTRFS_I(inode)->lock);
1349                 }
1350
1351                 spin_lock(&root->fs_info->delalloc_lock);
1352                 BTRFS_I(inode)->delalloc_bytes += len;
1353                 root->fs_info->delalloc_bytes += len;
1354                 if (do_list && list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1355                         list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1356                                       &root->fs_info->delalloc_inodes);
1357                 }
1358                 spin_unlock(&root->fs_info->delalloc_lock);
1359         }
1360 }
1361
1362 /*
1363  * extent_io.c clear_bit_hook, see set_bit_hook for why
1364  */
1365 static void btrfs_clear_bit_hook(struct inode *inode,
1366                                  struct extent_state *state, int *bits)
1367 {
1368         /*
1369          * set_bit and clear bit hooks normally require _irqsave/restore
1370          * but in this case, we are only testing for the DELALLOC
1371          * bit, which is only set or cleared with irqs on
1372          */
1373         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1374                 struct btrfs_root *root = BTRFS_I(inode)->root;
1375                 u64 len = state->end + 1 - state->start;
1376                 bool do_list = !btrfs_is_free_space_inode(root, inode);
1377
1378                 if (*bits & EXTENT_FIRST_DELALLOC) {
1379                         *bits &= ~EXTENT_FIRST_DELALLOC;
1380                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1381                         spin_lock(&BTRFS_I(inode)->lock);
1382                         BTRFS_I(inode)->outstanding_extents--;
1383                         spin_unlock(&BTRFS_I(inode)->lock);
1384                 }
1385
1386                 if (*bits & EXTENT_DO_ACCOUNTING)
1387                         btrfs_delalloc_release_metadata(inode, len);
1388
1389                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1390                     && do_list)
1391                         btrfs_free_reserved_data_space(inode, len);
1392
1393                 spin_lock(&root->fs_info->delalloc_lock);
1394                 root->fs_info->delalloc_bytes -= len;
1395                 BTRFS_I(inode)->delalloc_bytes -= len;
1396
1397                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1398                     !list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1399                         list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1400                 }
1401                 spin_unlock(&root->fs_info->delalloc_lock);
1402         }
1403 }
1404
1405 /*
1406  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1407  * we don't create bios that span stripes or chunks
1408  */
1409 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1410                          size_t size, struct bio *bio,
1411                          unsigned long bio_flags)
1412 {
1413         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1414         struct btrfs_mapping_tree *map_tree;
1415         u64 logical = (u64)bio->bi_sector << 9;
1416         u64 length = 0;
1417         u64 map_length;
1418         int ret;
1419
1420         if (bio_flags & EXTENT_BIO_COMPRESSED)
1421                 return 0;
1422
1423         length = bio->bi_size;
1424         map_tree = &root->fs_info->mapping_tree;
1425         map_length = length;
1426         ret = btrfs_map_block(map_tree, READ, logical,
1427                               &map_length, NULL, 0);
1428
1429         if (map_length < length + size)
1430                 return 1;
1431         return ret;
1432 }
1433
1434 /*
1435  * in order to insert checksums into the metadata in large chunks,
1436  * we wait until bio submission time.   All the pages in the bio are
1437  * checksummed and sums are attached onto the ordered extent record.
1438  *
1439  * At IO completion time the cums attached on the ordered extent record
1440  * are inserted into the btree
1441  */
1442 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1443                                     struct bio *bio, int mirror_num,
1444                                     unsigned long bio_flags,
1445                                     u64 bio_offset)
1446 {
1447         struct btrfs_root *root = BTRFS_I(inode)->root;
1448         int ret = 0;
1449
1450         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1451         BUG_ON(ret);
1452         return 0;
1453 }
1454
1455 /*
1456  * in order to insert checksums into the metadata in large chunks,
1457  * we wait until bio submission time.   All the pages in the bio are
1458  * checksummed and sums are attached onto the ordered extent record.
1459  *
1460  * At IO completion time the cums attached on the ordered extent record
1461  * are inserted into the btree
1462  */
1463 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1464                           int mirror_num, unsigned long bio_flags,
1465                           u64 bio_offset)
1466 {
1467         struct btrfs_root *root = BTRFS_I(inode)->root;
1468         return btrfs_map_bio(root, rw, bio, mirror_num, 1);
1469 }
1470
1471 /*
1472  * extent_io.c submission hook. This does the right thing for csum calculation
1473  * on write, or reading the csums from the tree before a read
1474  */
1475 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1476                           int mirror_num, unsigned long bio_flags,
1477                           u64 bio_offset)
1478 {
1479         struct btrfs_root *root = BTRFS_I(inode)->root;
1480         int ret = 0;
1481         int skip_sum;
1482
1483         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1484
1485         if (btrfs_is_free_space_inode(root, inode))
1486                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 2);
1487         else
1488                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
1489         BUG_ON(ret);
1490
1491         if (!(rw & REQ_WRITE)) {
1492                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1493                         return btrfs_submit_compressed_read(inode, bio,
1494                                                     mirror_num, bio_flags);
1495                 } else if (!skip_sum) {
1496                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1497                         if (ret)
1498                                 return ret;
1499                 }
1500                 goto mapit;
1501         } else if (!skip_sum) {
1502                 /* csum items have already been cloned */
1503                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1504                         goto mapit;
1505                 /* we're doing a write, do the async checksumming */
1506                 return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1507                                    inode, rw, bio, mirror_num,
1508                                    bio_flags, bio_offset,
1509                                    __btrfs_submit_bio_start,
1510                                    __btrfs_submit_bio_done);
1511         }
1512
1513 mapit:
1514         return btrfs_map_bio(root, rw, bio, mirror_num, 0);
1515 }
1516
1517 /*
1518  * given a list of ordered sums record them in the inode.  This happens
1519  * at IO completion time based on sums calculated at bio submission time.
1520  */
1521 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1522                              struct inode *inode, u64 file_offset,
1523                              struct list_head *list)
1524 {
1525         struct btrfs_ordered_sum *sum;
1526
1527         list_for_each_entry(sum, list, list) {
1528                 btrfs_csum_file_blocks(trans,
1529                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1530         }
1531         return 0;
1532 }
1533
1534 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1535                               struct extent_state **cached_state)
1536 {
1537         if ((end & (PAGE_CACHE_SIZE - 1)) == 0)
1538                 WARN_ON(1);
1539         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1540                                    cached_state, GFP_NOFS);
1541 }
1542
1543 /* see btrfs_writepage_start_hook for details on why this is required */
1544 struct btrfs_writepage_fixup {
1545         struct page *page;
1546         struct btrfs_work work;
1547 };
1548
1549 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1550 {
1551         struct btrfs_writepage_fixup *fixup;
1552         struct btrfs_ordered_extent *ordered;
1553         struct extent_state *cached_state = NULL;
1554         struct page *page;
1555         struct inode *inode;
1556         u64 page_start;
1557         u64 page_end;
1558         int ret;
1559
1560         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1561         page = fixup->page;
1562 again:
1563         lock_page(page);
1564         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1565                 ClearPageChecked(page);
1566                 goto out_page;
1567         }
1568
1569         inode = page->mapping->host;
1570         page_start = page_offset(page);
1571         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1572
1573         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end, 0,
1574                          &cached_state, GFP_NOFS);
1575
1576         /* already ordered? We're done */
1577         if (PagePrivate2(page))
1578                 goto out;
1579
1580         ordered = btrfs_lookup_ordered_extent(inode, page_start);
1581         if (ordered) {
1582                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
1583                                      page_end, &cached_state, GFP_NOFS);
1584                 unlock_page(page);
1585                 btrfs_start_ordered_extent(inode, ordered, 1);
1586                 btrfs_put_ordered_extent(ordered);
1587                 goto again;
1588         }
1589
1590         ret = btrfs_delalloc_reserve_space(inode, PAGE_CACHE_SIZE);
1591         if (ret) {
1592                 mapping_set_error(page->mapping, ret);
1593                 end_extent_writepage(page, ret, page_start, page_end);
1594                 ClearPageChecked(page);
1595                 goto out;
1596          }
1597
1598         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
1599         ClearPageChecked(page);
1600         set_page_dirty(page);
1601 out:
1602         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
1603                              &cached_state, GFP_NOFS);
1604 out_page:
1605         unlock_page(page);
1606         page_cache_release(page);
1607         kfree(fixup);
1608 }
1609
1610 /*
1611  * There are a few paths in the higher layers of the kernel that directly
1612  * set the page dirty bit without asking the filesystem if it is a
1613  * good idea.  This causes problems because we want to make sure COW
1614  * properly happens and the data=ordered rules are followed.
1615  *
1616  * In our case any range that doesn't have the ORDERED bit set
1617  * hasn't been properly setup for IO.  We kick off an async process
1618  * to fix it up.  The async helper will wait for ordered extents, set
1619  * the delalloc bit and make it safe to write the page.
1620  */
1621 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
1622 {
1623         struct inode *inode = page->mapping->host;
1624         struct btrfs_writepage_fixup *fixup;
1625         struct btrfs_root *root = BTRFS_I(inode)->root;
1626
1627         /* this page is properly in the ordered list */
1628         if (TestClearPagePrivate2(page))
1629                 return 0;
1630
1631         if (PageChecked(page))
1632                 return -EAGAIN;
1633
1634         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
1635         if (!fixup)
1636                 return -EAGAIN;
1637
1638         SetPageChecked(page);
1639         page_cache_get(page);
1640         fixup->work.func = btrfs_writepage_fixup_worker;
1641         fixup->page = page;
1642         btrfs_queue_worker(&root->fs_info->fixup_workers, &fixup->work);
1643         return -EBUSY;
1644 }
1645
1646 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
1647                                        struct inode *inode, u64 file_pos,
1648                                        u64 disk_bytenr, u64 disk_num_bytes,
1649                                        u64 num_bytes, u64 ram_bytes,
1650                                        u8 compression, u8 encryption,
1651                                        u16 other_encoding, int extent_type)
1652 {
1653         struct btrfs_root *root = BTRFS_I(inode)->root;
1654         struct btrfs_file_extent_item *fi;
1655         struct btrfs_path *path;
1656         struct extent_buffer *leaf;
1657         struct btrfs_key ins;
1658         u64 hint;
1659         int ret;
1660
1661         path = btrfs_alloc_path();
1662         if (!path)
1663                 return -ENOMEM;
1664
1665         path->leave_spinning = 1;
1666
1667         /*
1668          * we may be replacing one extent in the tree with another.
1669          * The new extent is pinned in the extent map, and we don't want
1670          * to drop it from the cache until it is completely in the btree.
1671          *
1672          * So, tell btrfs_drop_extents to leave this extent in the cache.
1673          * the caller is expected to unpin it and allow it to be merged
1674          * with the others.
1675          */
1676         ret = btrfs_drop_extents(trans, inode, file_pos, file_pos + num_bytes,
1677                                  &hint, 0);
1678         BUG_ON(ret);
1679
1680         ins.objectid = btrfs_ino(inode);
1681         ins.offset = file_pos;
1682         ins.type = BTRFS_EXTENT_DATA_KEY;
1683         ret = btrfs_insert_empty_item(trans, root, path, &ins, sizeof(*fi));
1684         BUG_ON(ret);
1685         leaf = path->nodes[0];
1686         fi = btrfs_item_ptr(leaf, path->slots[0],
1687                             struct btrfs_file_extent_item);
1688         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1689         btrfs_set_file_extent_type(leaf, fi, extent_type);
1690         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
1691         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
1692         btrfs_set_file_extent_offset(leaf, fi, 0);
1693         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1694         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
1695         btrfs_set_file_extent_compression(leaf, fi, compression);
1696         btrfs_set_file_extent_encryption(leaf, fi, encryption);
1697         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
1698
1699         btrfs_unlock_up_safe(path, 1);
1700         btrfs_set_lock_blocking(leaf);
1701
1702         btrfs_mark_buffer_dirty(leaf);
1703
1704         inode_add_bytes(inode, num_bytes);
1705
1706         ins.objectid = disk_bytenr;
1707         ins.offset = disk_num_bytes;
1708         ins.type = BTRFS_EXTENT_ITEM_KEY;
1709         ret = btrfs_alloc_reserved_file_extent(trans, root,
1710                                         root->root_key.objectid,
1711                                         btrfs_ino(inode), file_pos, &ins);
1712         BUG_ON(ret);
1713         btrfs_free_path(path);
1714
1715         return 0;
1716 }
1717
1718 /*
1719  * helper function for btrfs_finish_ordered_io, this
1720  * just reads in some of the csum leaves to prime them into ram
1721  * before we start the transaction.  It limits the amount of btree
1722  * reads required while inside the transaction.
1723  */
1724 /* as ordered data IO finishes, this gets called so we can finish
1725  * an ordered extent if the range of bytes in the file it covers are
1726  * fully written.
1727  */
1728 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end)
1729 {
1730         struct btrfs_root *root = BTRFS_I(inode)->root;
1731         struct btrfs_trans_handle *trans = NULL;
1732         struct btrfs_ordered_extent *ordered_extent = NULL;
1733         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1734         struct extent_state *cached_state = NULL;
1735         int compress_type = 0;
1736         int ret;
1737         bool nolock;
1738
1739         ret = btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
1740                                              end - start + 1);
1741         if (!ret)
1742                 return 0;
1743         BUG_ON(!ordered_extent);
1744
1745         nolock = btrfs_is_free_space_inode(root, inode);
1746
1747         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
1748                 BUG_ON(!list_empty(&ordered_extent->list));
1749                 ret = btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1750                 if (!ret) {
1751                         if (nolock)
1752                                 trans = btrfs_join_transaction_nolock(root);
1753                         else
1754                                 trans = btrfs_join_transaction(root);
1755                         BUG_ON(IS_ERR(trans));
1756                         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1757                         ret = btrfs_update_inode_fallback(trans, root, inode);
1758                         BUG_ON(ret);
1759                 }
1760                 goto out;
1761         }
1762
1763         lock_extent_bits(io_tree, ordered_extent->file_offset,
1764                          ordered_extent->file_offset + ordered_extent->len - 1,
1765                          0, &cached_state, GFP_NOFS);
1766
1767         if (nolock)
1768                 trans = btrfs_join_transaction_nolock(root);
1769         else
1770                 trans = btrfs_join_transaction(root);
1771         BUG_ON(IS_ERR(trans));
1772         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1773
1774         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
1775                 compress_type = ordered_extent->compress_type;
1776         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
1777                 BUG_ON(compress_type);
1778                 ret = btrfs_mark_extent_written(trans, inode,
1779                                                 ordered_extent->file_offset,
1780                                                 ordered_extent->file_offset +
1781                                                 ordered_extent->len);
1782                 BUG_ON(ret);
1783         } else {
1784                 BUG_ON(root == root->fs_info->tree_root);
1785                 ret = insert_reserved_file_extent(trans, inode,
1786                                                 ordered_extent->file_offset,
1787                                                 ordered_extent->start,
1788                                                 ordered_extent->disk_len,
1789                                                 ordered_extent->len,
1790                                                 ordered_extent->len,
1791                                                 compress_type, 0, 0,
1792                                                 BTRFS_FILE_EXTENT_REG);
1793                 unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
1794                                    ordered_extent->file_offset,
1795                                    ordered_extent->len);
1796                 BUG_ON(ret);
1797         }
1798         unlock_extent_cached(io_tree, ordered_extent->file_offset,
1799                              ordered_extent->file_offset +
1800                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
1801
1802         add_pending_csums(trans, inode, ordered_extent->file_offset,
1803                           &ordered_extent->list);
1804
1805         ret = btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1806         if (!ret || !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
1807                 ret = btrfs_update_inode_fallback(trans, root, inode);
1808                 BUG_ON(ret);
1809         }
1810         ret = 0;
1811 out:
1812         if (root != root->fs_info->tree_root)
1813                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
1814         if (trans) {
1815                 if (nolock)
1816                         btrfs_end_transaction_nolock(trans, root);
1817                 else
1818                         btrfs_end_transaction(trans, root);
1819         }
1820
1821         /* once for us */
1822         btrfs_put_ordered_extent(ordered_extent);
1823         /* once for the tree */
1824         btrfs_put_ordered_extent(ordered_extent);
1825
1826         return 0;
1827 }
1828
1829 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
1830                                 struct extent_state *state, int uptodate)
1831 {
1832         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
1833
1834         ClearPagePrivate2(page);
1835         return btrfs_finish_ordered_io(page->mapping->host, start, end);
1836 }
1837
1838 /*
1839  * when reads are done, we need to check csums to verify the data is correct
1840  * if there's a match, we allow the bio to finish.  If not, the code in
1841  * extent_io.c will try to find good copies for us.
1842  */
1843 static int btrfs_readpage_end_io_hook(struct page *page, u64 start, u64 end,
1844                                struct extent_state *state)
1845 {
1846         size_t offset = start - ((u64)page->index << PAGE_CACHE_SHIFT);
1847         struct inode *inode = page->mapping->host;
1848         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1849         char *kaddr;
1850         u64 private = ~(u32)0;
1851         int ret;
1852         struct btrfs_root *root = BTRFS_I(inode)->root;
1853         u32 csum = ~(u32)0;
1854
1855         if (PageChecked(page)) {
1856                 ClearPageChecked(page);
1857                 goto good;
1858         }
1859
1860         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
1861                 goto good;
1862
1863         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
1864             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
1865                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
1866                                   GFP_NOFS);
1867                 return 0;
1868         }
1869
1870         if (state && state->start == start) {
1871                 private = state->private;
1872                 ret = 0;
1873         } else {
1874                 ret = get_state_private(io_tree, start, &private);
1875         }
1876         kaddr = kmap_atomic(page);
1877         if (ret)
1878                 goto zeroit;
1879
1880         csum = btrfs_csum_data(root, kaddr + offset, csum,  end - start + 1);
1881         btrfs_csum_final(csum, (char *)&csum);
1882         if (csum != private)
1883                 goto zeroit;
1884
1885         kunmap_atomic(kaddr);
1886 good:
1887         return 0;
1888
1889 zeroit:
1890         printk_ratelimited(KERN_INFO "btrfs csum failed ino %llu off %llu csum %u "
1891                        "private %llu\n",
1892                        (unsigned long long)btrfs_ino(page->mapping->host),
1893                        (unsigned long long)start, csum,
1894                        (unsigned long long)private);
1895         memset(kaddr + offset, 1, end - start + 1);
1896         flush_dcache_page(page);
1897         kunmap_atomic(kaddr);
1898         if (private == 0)
1899                 return 0;
1900         return -EIO;
1901 }
1902
1903 struct delayed_iput {
1904         struct list_head list;
1905         struct inode *inode;
1906 };
1907
1908 void btrfs_add_delayed_iput(struct inode *inode)
1909 {
1910         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1911         struct delayed_iput *delayed;
1912
1913         if (atomic_add_unless(&inode->i_count, -1, 1))
1914                 return;
1915
1916         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
1917         delayed->inode = inode;
1918
1919         spin_lock(&fs_info->delayed_iput_lock);
1920         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
1921         spin_unlock(&fs_info->delayed_iput_lock);
1922 }
1923
1924 void btrfs_run_delayed_iputs(struct btrfs_root *root)
1925 {
1926         LIST_HEAD(list);
1927         struct btrfs_fs_info *fs_info = root->fs_info;
1928         struct delayed_iput *delayed;
1929         int empty;
1930
1931         spin_lock(&fs_info->delayed_iput_lock);
1932         empty = list_empty(&fs_info->delayed_iputs);
1933         spin_unlock(&fs_info->delayed_iput_lock);
1934         if (empty)
1935                 return;
1936
1937         down_read(&root->fs_info->cleanup_work_sem);
1938         spin_lock(&fs_info->delayed_iput_lock);
1939         list_splice_init(&fs_info->delayed_iputs, &list);
1940         spin_unlock(&fs_info->delayed_iput_lock);
1941
1942         while (!list_empty(&list)) {
1943                 delayed = list_entry(list.next, struct delayed_iput, list);
1944                 list_del(&delayed->list);
1945                 iput(delayed->inode);
1946                 kfree(delayed);
1947         }
1948         up_read(&root->fs_info->cleanup_work_sem);
1949 }
1950
1951 enum btrfs_orphan_cleanup_state {
1952         ORPHAN_CLEANUP_STARTED  = 1,
1953         ORPHAN_CLEANUP_DONE     = 2,
1954 };
1955
1956 /*
1957  * This is called in transaction commit time. If there are no orphan
1958  * files in the subvolume, it removes orphan item and frees block_rsv
1959  * structure.
1960  */
1961 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
1962                               struct btrfs_root *root)
1963 {
1964         struct btrfs_block_rsv *block_rsv;
1965         int ret;
1966
1967         if (!list_empty(&root->orphan_list) ||
1968             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
1969                 return;
1970
1971         spin_lock(&root->orphan_lock);
1972         if (!list_empty(&root->orphan_list)) {
1973                 spin_unlock(&root->orphan_lock);
1974                 return;
1975         }
1976
1977         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
1978                 spin_unlock(&root->orphan_lock);
1979                 return;
1980         }
1981
1982         block_rsv = root->orphan_block_rsv;
1983         root->orphan_block_rsv = NULL;
1984         spin_unlock(&root->orphan_lock);
1985
1986         if (root->orphan_item_inserted &&
1987             btrfs_root_refs(&root->root_item) > 0) {
1988                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
1989                                             root->root_key.objectid);
1990                 BUG_ON(ret);
1991                 root->orphan_item_inserted = 0;
1992         }
1993
1994         if (block_rsv) {
1995                 WARN_ON(block_rsv->size > 0);
1996                 btrfs_free_block_rsv(root, block_rsv);
1997         }
1998 }
1999
2000 /*
2001  * This creates an orphan entry for the given inode in case something goes
2002  * wrong in the middle of an unlink/truncate.
2003  *
2004  * NOTE: caller of this function should reserve 5 units of metadata for
2005  *       this function.
2006  */
2007 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
2008 {
2009         struct btrfs_root *root = BTRFS_I(inode)->root;
2010         struct btrfs_block_rsv *block_rsv = NULL;
2011         int reserve = 0;
2012         int insert = 0;
2013         int ret;
2014
2015         if (!root->orphan_block_rsv) {
2016                 block_rsv = btrfs_alloc_block_rsv(root);
2017                 if (!block_rsv)
2018                         return -ENOMEM;
2019         }
2020
2021         spin_lock(&root->orphan_lock);
2022         if (!root->orphan_block_rsv) {
2023                 root->orphan_block_rsv = block_rsv;
2024         } else if (block_rsv) {
2025                 btrfs_free_block_rsv(root, block_rsv);
2026                 block_rsv = NULL;
2027         }
2028
2029         if (list_empty(&BTRFS_I(inode)->i_orphan)) {
2030                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2031 #if 0
2032                 /*
2033                  * For proper ENOSPC handling, we should do orphan
2034                  * cleanup when mounting. But this introduces backward
2035                  * compatibility issue.
2036                  */
2037                 if (!xchg(&root->orphan_item_inserted, 1))
2038                         insert = 2;
2039                 else
2040                         insert = 1;
2041 #endif
2042                 insert = 1;
2043         }
2044
2045         if (!BTRFS_I(inode)->orphan_meta_reserved) {
2046                 BTRFS_I(inode)->orphan_meta_reserved = 1;
2047                 reserve = 1;
2048         }
2049         spin_unlock(&root->orphan_lock);
2050
2051         /* grab metadata reservation from transaction handle */
2052         if (reserve) {
2053                 ret = btrfs_orphan_reserve_metadata(trans, inode);
2054                 BUG_ON(ret);
2055         }
2056
2057         /* insert an orphan item to track this unlinked/truncated file */
2058         if (insert >= 1) {
2059                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
2060                 BUG_ON(ret && ret != -EEXIST);
2061         }
2062
2063         /* insert an orphan item to track subvolume contains orphan files */
2064         if (insert >= 2) {
2065                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2066                                                root->root_key.objectid);
2067                 BUG_ON(ret);
2068         }
2069         return 0;
2070 }
2071
2072 /*
2073  * We have done the truncate/delete so we can go ahead and remove the orphan
2074  * item for this particular inode.
2075  */
2076 int btrfs_orphan_del(struct btrfs_trans_handle *trans, struct inode *inode)
2077 {
2078         struct btrfs_root *root = BTRFS_I(inode)->root;
2079         int delete_item = 0;
2080         int release_rsv = 0;
2081         int ret = 0;
2082
2083         spin_lock(&root->orphan_lock);
2084         if (!list_empty(&BTRFS_I(inode)->i_orphan)) {
2085                 list_del_init(&BTRFS_I(inode)->i_orphan);
2086                 delete_item = 1;
2087         }
2088
2089         if (BTRFS_I(inode)->orphan_meta_reserved) {
2090                 BTRFS_I(inode)->orphan_meta_reserved = 0;
2091                 release_rsv = 1;
2092         }
2093         spin_unlock(&root->orphan_lock);
2094
2095         if (trans && delete_item) {
2096                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
2097                 BUG_ON(ret);
2098         }
2099
2100         if (release_rsv)
2101                 btrfs_orphan_release_metadata(inode);
2102
2103         return 0;
2104 }
2105
2106 /*
2107  * this cleans up any orphans that may be left on the list from the last use
2108  * of this root.
2109  */
2110 int btrfs_orphan_cleanup(struct btrfs_root *root)
2111 {
2112         struct btrfs_path *path;
2113         struct extent_buffer *leaf;
2114         struct btrfs_key key, found_key;
2115         struct btrfs_trans_handle *trans;
2116         struct inode *inode;
2117         u64 last_objectid = 0;
2118         int ret = 0, nr_unlink = 0, nr_truncate = 0;
2119
2120         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
2121                 return 0;
2122
2123         path = btrfs_alloc_path();
2124         if (!path) {
2125                 ret = -ENOMEM;
2126                 goto out;
2127         }
2128         path->reada = -1;
2129
2130         key.objectid = BTRFS_ORPHAN_OBJECTID;
2131         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
2132         key.offset = (u64)-1;
2133
2134         while (1) {
2135                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2136                 if (ret < 0)
2137                         goto out;
2138
2139                 /*
2140                  * if ret == 0 means we found what we were searching for, which
2141                  * is weird, but possible, so only screw with path if we didn't
2142                  * find the key and see if we have stuff that matches
2143                  */
2144                 if (ret > 0) {
2145                         ret = 0;
2146                         if (path->slots[0] == 0)
2147                                 break;
2148                         path->slots[0]--;
2149                 }
2150
2151                 /* pull out the item */
2152                 leaf = path->nodes[0];
2153                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2154
2155                 /* make sure the item matches what we want */
2156                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
2157                         break;
2158                 if (btrfs_key_type(&found_key) != BTRFS_ORPHAN_ITEM_KEY)
2159                         break;
2160
2161                 /* release the path since we're done with it */
2162                 btrfs_release_path(path);
2163
2164                 /*
2165                  * this is where we are basically btrfs_lookup, without the
2166                  * crossing root thing.  we store the inode number in the
2167                  * offset of the orphan item.
2168                  */
2169
2170                 if (found_key.offset == last_objectid) {
2171                         printk(KERN_ERR "btrfs: Error removing orphan entry, "
2172                                "stopping orphan cleanup\n");
2173                         ret = -EINVAL;
2174                         goto out;
2175                 }
2176
2177                 last_objectid = found_key.offset;
2178
2179                 found_key.objectid = found_key.offset;
2180                 found_key.type = BTRFS_INODE_ITEM_KEY;
2181                 found_key.offset = 0;
2182                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
2183                 ret = PTR_RET(inode);
2184                 if (ret && ret != -ESTALE)
2185                         goto out;
2186
2187                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
2188                         struct btrfs_root *dead_root;
2189                         struct btrfs_fs_info *fs_info = root->fs_info;
2190                         int is_dead_root = 0;
2191
2192                         /*
2193                          * this is an orphan in the tree root. Currently these
2194                          * could come from 2 sources:
2195                          *  a) a snapshot deletion in progress
2196                          *  b) a free space cache inode
2197                          * We need to distinguish those two, as the snapshot
2198                          * orphan must not get deleted.
2199                          * find_dead_roots already ran before us, so if this
2200                          * is a snapshot deletion, we should find the root
2201                          * in the dead_roots list
2202                          */
2203                         spin_lock(&fs_info->trans_lock);
2204                         list_for_each_entry(dead_root, &fs_info->dead_roots,
2205                                             root_list) {
2206                                 if (dead_root->root_key.objectid ==
2207                                     found_key.objectid) {
2208                                         is_dead_root = 1;
2209                                         break;
2210                                 }
2211                         }
2212                         spin_unlock(&fs_info->trans_lock);
2213                         if (is_dead_root) {
2214                                 /* prevent this orphan from being found again */
2215                                 key.offset = found_key.objectid - 1;
2216                                 continue;
2217                         }
2218                 }
2219                 /*
2220                  * Inode is already gone but the orphan item is still there,
2221                  * kill the orphan item.
2222                  */
2223                 if (ret == -ESTALE) {
2224                         trans = btrfs_start_transaction(root, 1);
2225                         if (IS_ERR(trans)) {
2226                                 ret = PTR_ERR(trans);
2227                                 goto out;
2228                         }
2229                         ret = btrfs_del_orphan_item(trans, root,
2230                                                     found_key.objectid);
2231                         BUG_ON(ret);
2232                         btrfs_end_transaction(trans, root);
2233                         continue;
2234                 }
2235
2236                 /*
2237                  * add this inode to the orphan list so btrfs_orphan_del does
2238                  * the proper thing when we hit it
2239                  */
2240                 spin_lock(&root->orphan_lock);
2241                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2242                 spin_unlock(&root->orphan_lock);
2243
2244                 /* if we have links, this was a truncate, lets do that */
2245                 if (inode->i_nlink) {
2246                         if (!S_ISREG(inode->i_mode)) {
2247                                 WARN_ON(1);
2248                                 iput(inode);
2249                                 continue;
2250                         }
2251                         nr_truncate++;
2252                         ret = btrfs_truncate(inode);
2253                 } else {
2254                         nr_unlink++;
2255                 }
2256
2257                 /* this will do delete_inode and everything for us */
2258                 iput(inode);
2259                 if (ret)
2260                         goto out;
2261         }
2262         /* release the path since we're done with it */
2263         btrfs_release_path(path);
2264
2265         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
2266
2267         if (root->orphan_block_rsv)
2268                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
2269                                         (u64)-1);
2270
2271         if (root->orphan_block_rsv || root->orphan_item_inserted) {
2272                 trans = btrfs_join_transaction(root);
2273                 if (!IS_ERR(trans))
2274                         btrfs_end_transaction(trans, root);
2275         }
2276
2277         if (nr_unlink)
2278                 printk(KERN_INFO "btrfs: unlinked %d orphans\n", nr_unlink);
2279         if (nr_truncate)
2280                 printk(KERN_INFO "btrfs: truncated %d orphans\n", nr_truncate);
2281
2282 out:
2283         if (ret)
2284                 printk(KERN_CRIT "btrfs: could not do orphan cleanup %d\n", ret);
2285         btrfs_free_path(path);
2286         return ret;
2287 }
2288
2289 /*
2290  * very simple check to peek ahead in the leaf looking for xattrs.  If we
2291  * don't find any xattrs, we know there can't be any acls.
2292  *
2293  * slot is the slot the inode is in, objectid is the objectid of the inode
2294  */
2295 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
2296                                           int slot, u64 objectid)
2297 {
2298         u32 nritems = btrfs_header_nritems(leaf);
2299         struct btrfs_key found_key;
2300         int scanned = 0;
2301
2302         slot++;
2303         while (slot < nritems) {
2304                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2305
2306                 /* we found a different objectid, there must not be acls */
2307                 if (found_key.objectid != objectid)
2308                         return 0;
2309
2310                 /* we found an xattr, assume we've got an acl */
2311                 if (found_key.type == BTRFS_XATTR_ITEM_KEY)
2312                         return 1;
2313
2314                 /*
2315                  * we found a key greater than an xattr key, there can't
2316                  * be any acls later on
2317                  */
2318                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
2319                         return 0;
2320
2321                 slot++;
2322                 scanned++;
2323
2324                 /*
2325                  * it goes inode, inode backrefs, xattrs, extents,
2326                  * so if there are a ton of hard links to an inode there can
2327                  * be a lot of backrefs.  Don't waste time searching too hard,
2328                  * this is just an optimization
2329                  */
2330                 if (scanned >= 8)
2331                         break;
2332         }
2333         /* we hit the end of the leaf before we found an xattr or
2334          * something larger than an xattr.  We have to assume the inode
2335          * has acls
2336          */
2337         return 1;
2338 }
2339
2340 /*
2341  * read an inode from the btree into the in-memory inode
2342  */
2343 static void btrfs_read_locked_inode(struct inode *inode)
2344 {
2345         struct btrfs_path *path;
2346         struct extent_buffer *leaf;
2347         struct btrfs_inode_item *inode_item;
2348         struct btrfs_timespec *tspec;
2349         struct btrfs_root *root = BTRFS_I(inode)->root;
2350         struct btrfs_key location;
2351         int maybe_acls;
2352         u32 rdev;
2353         int ret;
2354         bool filled = false;
2355
2356         ret = btrfs_fill_inode(inode, &rdev);
2357         if (!ret)
2358                 filled = true;
2359
2360         path = btrfs_alloc_path();
2361         if (!path)
2362                 goto make_bad;
2363
2364         path->leave_spinning = 1;
2365         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
2366
2367         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
2368         if (ret)
2369                 goto make_bad;
2370
2371         leaf = path->nodes[0];
2372
2373         if (filled)
2374                 goto cache_acl;
2375
2376         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2377                                     struct btrfs_inode_item);
2378         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
2379         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
2380         inode->i_uid = btrfs_inode_uid(leaf, inode_item);
2381         inode->i_gid = btrfs_inode_gid(leaf, inode_item);
2382         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
2383
2384         tspec = btrfs_inode_atime(inode_item);
2385         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2386         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2387
2388         tspec = btrfs_inode_mtime(inode_item);
2389         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2390         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2391
2392         tspec = btrfs_inode_ctime(inode_item);
2393         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2394         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2395
2396         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
2397         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
2398         BTRFS_I(inode)->sequence = btrfs_inode_sequence(leaf, inode_item);
2399         inode->i_generation = BTRFS_I(inode)->generation;
2400         inode->i_rdev = 0;
2401         rdev = btrfs_inode_rdev(leaf, inode_item);
2402
2403         BTRFS_I(inode)->index_cnt = (u64)-1;
2404         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
2405 cache_acl:
2406         /*
2407          * try to precache a NULL acl entry for files that don't have
2408          * any xattrs or acls
2409          */
2410         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
2411                                            btrfs_ino(inode));
2412         if (!maybe_acls)
2413                 cache_no_acl(inode);
2414
2415         btrfs_free_path(path);
2416
2417         switch (inode->i_mode & S_IFMT) {
2418         case S_IFREG:
2419                 inode->i_mapping->a_ops = &btrfs_aops;
2420                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2421                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
2422                 inode->i_fop = &btrfs_file_operations;
2423                 inode->i_op = &btrfs_file_inode_operations;
2424                 break;
2425         case S_IFDIR:
2426                 inode->i_fop = &btrfs_dir_file_operations;
2427                 if (root == root->fs_info->tree_root)
2428                         inode->i_op = &btrfs_dir_ro_inode_operations;
2429                 else
2430                         inode->i_op = &btrfs_dir_inode_operations;
2431                 break;
2432         case S_IFLNK:
2433                 inode->i_op = &btrfs_symlink_inode_operations;
2434                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
2435                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2436                 break;
2437         default:
2438                 inode->i_op = &btrfs_special_inode_operations;
2439                 init_special_inode(inode, inode->i_mode, rdev);
2440                 break;
2441         }
2442
2443         btrfs_update_iflags(inode);
2444         return;
2445
2446 make_bad:
2447         btrfs_free_path(path);
2448         make_bad_inode(inode);
2449 }
2450
2451 /*
2452  * given a leaf and an inode, copy the inode fields into the leaf
2453  */
2454 static void fill_inode_item(struct btrfs_trans_handle *trans,
2455                             struct extent_buffer *leaf,
2456                             struct btrfs_inode_item *item,
2457                             struct inode *inode)
2458 {
2459         btrfs_set_inode_uid(leaf, item, inode->i_uid);
2460         btrfs_set_inode_gid(leaf, item, inode->i_gid);
2461         btrfs_set_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size);
2462         btrfs_set_inode_mode(leaf, item, inode->i_mode);
2463         btrfs_set_inode_nlink(leaf, item, inode->i_nlink);
2464
2465         btrfs_set_timespec_sec(leaf, btrfs_inode_atime(item),
2466                                inode->i_atime.tv_sec);
2467         btrfs_set_timespec_nsec(leaf, btrfs_inode_atime(item),
2468                                 inode->i_atime.tv_nsec);
2469
2470         btrfs_set_timespec_sec(leaf, btrfs_inode_mtime(item),
2471                                inode->i_mtime.tv_sec);
2472         btrfs_set_timespec_nsec(leaf, btrfs_inode_mtime(item),
2473                                 inode->i_mtime.tv_nsec);
2474
2475         btrfs_set_timespec_sec(leaf, btrfs_inode_ctime(item),
2476                                inode->i_ctime.tv_sec);
2477         btrfs_set_timespec_nsec(leaf, btrfs_inode_ctime(item),
2478                                 inode->i_ctime.tv_nsec);
2479
2480         btrfs_set_inode_nbytes(leaf, item, inode_get_bytes(inode));
2481         btrfs_set_inode_generation(leaf, item, BTRFS_I(inode)->generation);
2482         btrfs_set_inode_sequence(leaf, item, BTRFS_I(inode)->sequence);
2483         btrfs_set_inode_transid(leaf, item, trans->transid);
2484         btrfs_set_inode_rdev(leaf, item, inode->i_rdev);
2485         btrfs_set_inode_flags(leaf, item, BTRFS_I(inode)->flags);
2486         btrfs_set_inode_block_group(leaf, item, 0);
2487 }
2488
2489 /*
2490  * copy everything in the in-memory inode into the btree.
2491  */
2492 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
2493                                 struct btrfs_root *root, struct inode *inode)
2494 {
2495         struct btrfs_inode_item *inode_item;
2496         struct btrfs_path *path;
2497         struct extent_buffer *leaf;
2498         int ret;
2499
2500         path = btrfs_alloc_path();
2501         if (!path)
2502                 return -ENOMEM;
2503
2504         path->leave_spinning = 1;
2505         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
2506                                  1);
2507         if (ret) {
2508                 if (ret > 0)
2509                         ret = -ENOENT;
2510                 goto failed;
2511         }
2512
2513         btrfs_unlock_up_safe(path, 1);
2514         leaf = path->nodes[0];
2515         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2516                                     struct btrfs_inode_item);
2517
2518         fill_inode_item(trans, leaf, inode_item, inode);
2519         btrfs_mark_buffer_dirty(leaf);
2520         btrfs_set_inode_last_trans(trans, inode);
2521         ret = 0;
2522 failed:
2523         btrfs_free_path(path);
2524         return ret;
2525 }
2526
2527 /*
2528  * copy everything in the in-memory inode into the btree.
2529  */
2530 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
2531                                 struct btrfs_root *root, struct inode *inode)
2532 {
2533         int ret;
2534
2535         /*
2536          * If the inode is a free space inode, we can deadlock during commit
2537          * if we put it into the delayed code.
2538          *
2539          * The data relocation inode should also be directly updated
2540          * without delay
2541          */
2542         if (!btrfs_is_free_space_inode(root, inode)
2543             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID) {
2544                 ret = btrfs_delayed_update_inode(trans, root, inode);
2545                 if (!ret)
2546                         btrfs_set_inode_last_trans(trans, inode);
2547                 return ret;
2548         }
2549
2550         return btrfs_update_inode_item(trans, root, inode);
2551 }
2552
2553 static noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
2554                                 struct btrfs_root *root, struct inode *inode)
2555 {
2556         int ret;
2557
2558         ret = btrfs_update_inode(trans, root, inode);
2559         if (ret == -ENOSPC)
2560                 return btrfs_update_inode_item(trans, root, inode);
2561         return ret;
2562 }
2563
2564 /*
2565  * unlink helper that gets used here in inode.c and in the tree logging
2566  * recovery code.  It remove a link in a directory with a given name, and
2567  * also drops the back refs in the inode to the directory
2568  */
2569 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
2570                                 struct btrfs_root *root,
2571                                 struct inode *dir, struct inode *inode,
2572                                 const char *name, int name_len)
2573 {
2574         struct btrfs_path *path;
2575         int ret = 0;
2576         struct extent_buffer *leaf;
2577         struct btrfs_dir_item *di;
2578         struct btrfs_key key;
2579         u64 index;
2580         u64 ino = btrfs_ino(inode);
2581         u64 dir_ino = btrfs_ino(dir);
2582
2583         path = btrfs_alloc_path();
2584         if (!path) {
2585                 ret = -ENOMEM;
2586                 goto out;
2587         }
2588
2589         path->leave_spinning = 1;
2590         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
2591                                     name, name_len, -1);
2592         if (IS_ERR(di)) {
2593                 ret = PTR_ERR(di);
2594                 goto err;
2595         }
2596         if (!di) {
2597                 ret = -ENOENT;
2598                 goto err;
2599         }
2600         leaf = path->nodes[0];
2601         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2602         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2603         if (ret)
2604                 goto err;
2605         btrfs_release_path(path);
2606
2607         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
2608                                   dir_ino, &index);
2609         if (ret) {
2610                 printk(KERN_INFO "btrfs failed to delete reference to %.*s, "
2611                        "inode %llu parent %llu\n", name_len, name,
2612                        (unsigned long long)ino, (unsigned long long)dir_ino);
2613                 goto err;
2614         }
2615
2616         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
2617         if (ret)
2618                 goto err;
2619
2620         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
2621                                          inode, dir_ino);
2622         BUG_ON(ret != 0 && ret != -ENOENT);
2623
2624         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
2625                                            dir, index);
2626         if (ret == -ENOENT)
2627                 ret = 0;
2628 err:
2629         btrfs_free_path(path);
2630         if (ret)
2631                 goto out;
2632
2633         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
2634         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
2635         btrfs_update_inode(trans, root, dir);
2636 out:
2637         return ret;
2638 }
2639
2640 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
2641                        struct btrfs_root *root,
2642                        struct inode *dir, struct inode *inode,
2643                        const char *name, int name_len)
2644 {
2645         int ret;
2646         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
2647         if (!ret) {
2648                 btrfs_drop_nlink(inode);
2649                 ret = btrfs_update_inode(trans, root, inode);
2650         }
2651         return ret;
2652 }
2653
2654
2655 /* helper to check if there is any shared block in the path */
2656 static int check_path_shared(struct btrfs_root *root,
2657                              struct btrfs_path *path)
2658 {
2659         struct extent_buffer *eb;
2660         int level;
2661         u64 refs = 1;
2662
2663         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2664                 int ret;
2665
2666                 if (!path->nodes[level])
2667                         break;
2668                 eb = path->nodes[level];
2669                 if (!btrfs_block_can_be_shared(root, eb))
2670                         continue;
2671                 ret = btrfs_lookup_extent_info(NULL, root, eb->start, eb->len,
2672                                                &refs, NULL);
2673                 if (refs > 1)
2674                         return 1;
2675         }
2676         return 0;
2677 }
2678
2679 /*
2680  * helper to start transaction for unlink and rmdir.
2681  *
2682  * unlink and rmdir are special in btrfs, they do not always free space.
2683  * so in enospc case, we should make sure they will free space before
2684  * allowing them to use the global metadata reservation.
2685  */
2686 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir,
2687                                                        struct dentry *dentry)
2688 {
2689         struct btrfs_trans_handle *trans;
2690         struct btrfs_root *root = BTRFS_I(dir)->root;
2691         struct btrfs_path *path;
2692         struct btrfs_inode_ref *ref;
2693         struct btrfs_dir_item *di;
2694         struct inode *inode = dentry->d_inode;
2695         u64 index;
2696         int check_link = 1;
2697         int err = -ENOSPC;
2698         int ret;
2699         u64 ino = btrfs_ino(inode);
2700         u64 dir_ino = btrfs_ino(dir);
2701
2702         /*
2703          * 1 for the possible orphan item
2704          * 1 for the dir item
2705          * 1 for the dir index
2706          * 1 for the inode ref
2707          * 1 for the inode ref in the tree log
2708          * 2 for the dir entries in the log
2709          * 1 for the inode
2710          */
2711         trans = btrfs_start_transaction(root, 8);
2712         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
2713                 return trans;
2714
2715         if (ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
2716                 return ERR_PTR(-ENOSPC);
2717
2718         /* check if there is someone else holds reference */
2719         if (S_ISDIR(inode->i_mode) && atomic_read(&inode->i_count) > 1)
2720                 return ERR_PTR(-ENOSPC);
2721
2722         if (atomic_read(&inode->i_count) > 2)
2723                 return ERR_PTR(-ENOSPC);
2724
2725         if (xchg(&root->fs_info->enospc_unlink, 1))
2726                 return ERR_PTR(-ENOSPC);
2727
2728         path = btrfs_alloc_path();
2729         if (!path) {
2730                 root->fs_info->enospc_unlink = 0;
2731                 return ERR_PTR(-ENOMEM);
2732         }
2733
2734         /* 1 for the orphan item */
2735         trans = btrfs_start_transaction(root, 1);
2736         if (IS_ERR(trans)) {
2737                 btrfs_free_path(path);
2738                 root->fs_info->enospc_unlink = 0;
2739                 return trans;
2740         }
2741
2742         path->skip_locking = 1;
2743         path->search_commit_root = 1;
2744
2745         ret = btrfs_lookup_inode(trans, root, path,
2746                                 &BTRFS_I(dir)->location, 0);
2747         if (ret < 0) {
2748                 err = ret;
2749                 goto out;
2750         }
2751         if (ret == 0) {
2752                 if (check_path_shared(root, path))
2753                         goto out;
2754         } else {
2755                 check_link = 0;
2756         }
2757         btrfs_release_path(path);
2758
2759         ret = btrfs_lookup_inode(trans, root, path,
2760                                 &BTRFS_I(inode)->location, 0);
2761         if (ret < 0) {
2762                 err = ret;
2763                 goto out;
2764         }
2765         if (ret == 0) {
2766                 if (check_path_shared(root, path))
2767                         goto out;
2768         } else {
2769                 check_link = 0;
2770         }
2771         btrfs_release_path(path);
2772
2773         if (ret == 0 && S_ISREG(inode->i_mode)) {
2774                 ret = btrfs_lookup_file_extent(trans, root, path,
2775                                                ino, (u64)-1, 0);
2776                 if (ret < 0) {
2777                         err = ret;
2778                         goto out;
2779                 }
2780                 BUG_ON(ret == 0);
2781                 if (check_path_shared(root, path))
2782                         goto out;
2783                 btrfs_release_path(path);
2784         }
2785
2786         if (!check_link) {
2787                 err = 0;
2788                 goto out;
2789         }
2790
2791         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
2792                                 dentry->d_name.name, dentry->d_name.len, 0);
2793         if (IS_ERR(di)) {
2794                 err = PTR_ERR(di);
2795                 goto out;
2796         }
2797         if (di) {
2798                 if (check_path_shared(root, path))
2799                         goto out;
2800         } else {
2801                 err = 0;
2802                 goto out;
2803         }
2804         btrfs_release_path(path);
2805
2806         ref = btrfs_lookup_inode_ref(trans, root, path,
2807                                 dentry->d_name.name, dentry->d_name.len,
2808                                 ino, dir_ino, 0);
2809         if (IS_ERR(ref)) {
2810                 err = PTR_ERR(ref);
2811                 goto out;
2812         }
2813         BUG_ON(!ref);
2814         if (check_path_shared(root, path))
2815                 goto out;
2816         index = btrfs_inode_ref_index(path->nodes[0], ref);
2817         btrfs_release_path(path);
2818
2819         /*
2820          * This is a commit root search, if we can lookup inode item and other
2821          * relative items in the commit root, it means the transaction of
2822          * dir/file creation has been committed, and the dir index item that we
2823          * delay to insert has also been inserted into the commit root. So
2824          * we needn't worry about the delayed insertion of the dir index item
2825          * here.
2826          */
2827         di = btrfs_lookup_dir_index_item(trans, root, path, dir_ino, index,
2828                                 dentry->d_name.name, dentry->d_name.len, 0);
2829         if (IS_ERR(di)) {
2830                 err = PTR_ERR(di);
2831                 goto out;
2832         }
2833         BUG_ON(ret == -ENOENT);
2834         if (check_path_shared(root, path))
2835                 goto out;
2836
2837         err = 0;
2838 out:
2839         btrfs_free_path(path);
2840         /* Migrate the orphan reservation over */
2841         if (!err)
2842                 err = btrfs_block_rsv_migrate(trans->block_rsv,
2843                                 &root->fs_info->global_block_rsv,
2844                                 trans->bytes_reserved);
2845
2846         if (err) {
2847                 btrfs_end_transaction(trans, root);
2848                 root->fs_info->enospc_unlink = 0;
2849                 return ERR_PTR(err);
2850         }
2851
2852         trans->block_rsv = &root->fs_info->global_block_rsv;
2853         return trans;
2854 }
2855
2856 static void __unlink_end_trans(struct btrfs_trans_handle *trans,
2857                                struct btrfs_root *root)
2858 {
2859         if (trans->block_rsv == &root->fs_info->global_block_rsv) {
2860                 btrfs_block_rsv_release(root, trans->block_rsv,
2861                                         trans->bytes_reserved);
2862                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2863                 BUG_ON(!root->fs_info->enospc_unlink);
2864                 root->fs_info->enospc_unlink = 0;
2865         }
2866         btrfs_end_transaction(trans, root);
2867 }
2868
2869 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
2870 {
2871         struct btrfs_root *root = BTRFS_I(dir)->root;
2872         struct btrfs_trans_handle *trans;
2873         struct inode *inode = dentry->d_inode;
2874         int ret;
2875         unsigned long nr = 0;
2876
2877         trans = __unlink_start_trans(dir, dentry);
2878         if (IS_ERR(trans))
2879                 return PTR_ERR(trans);
2880
2881         btrfs_record_unlink_dir(trans, dir, dentry->d_inode, 0);
2882
2883         ret = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
2884                                  dentry->d_name.name, dentry->d_name.len);
2885         if (ret)
2886                 goto out;
2887
2888         if (inode->i_nlink == 0) {
2889                 ret = btrfs_orphan_add(trans, inode);
2890                 if (ret)
2891                         goto out;
2892         }
2893
2894 out:
2895         nr = trans->blocks_used;
2896         __unlink_end_trans(trans, root);
2897         btrfs_btree_balance_dirty(root, nr);
2898         return ret;
2899 }
2900
2901 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
2902                         struct btrfs_root *root,
2903                         struct inode *dir, u64 objectid,
2904                         const char *name, int name_len)
2905 {
2906         struct btrfs_path *path;
2907         struct extent_buffer *leaf;
2908         struct btrfs_dir_item *di;
2909         struct btrfs_key key;
2910         u64 index;
2911         int ret;
2912         u64 dir_ino = btrfs_ino(dir);
2913
2914         path = btrfs_alloc_path();
2915         if (!path)
2916                 return -ENOMEM;
2917
2918         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
2919                                    name, name_len, -1);
2920         BUG_ON(IS_ERR_OR_NULL(di));
2921
2922         leaf = path->nodes[0];
2923         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2924         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
2925         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2926         BUG_ON(ret);
2927         btrfs_release_path(path);
2928
2929         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
2930                                  objectid, root->root_key.objectid,
2931                                  dir_ino, &index, name, name_len);
2932         if (ret < 0) {
2933                 BUG_ON(ret != -ENOENT);
2934                 di = btrfs_search_dir_index_item(root, path, dir_ino,
2935                                                  name, name_len);
2936                 BUG_ON(IS_ERR_OR_NULL(di));
2937
2938                 leaf = path->nodes[0];
2939                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2940                 btrfs_release_path(path);
2941                 index = key.offset;
2942         }
2943         btrfs_release_path(path);
2944
2945         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
2946         BUG_ON(ret);
2947
2948         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
2949         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
2950         ret = btrfs_update_inode(trans, root, dir);
2951         BUG_ON(ret);
2952
2953         btrfs_free_path(path);
2954         return 0;
2955 }
2956
2957 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
2958 {
2959         struct inode *inode = dentry->d_inode;
2960         int err = 0;
2961         struct btrfs_root *root = BTRFS_I(dir)->root;
2962         struct btrfs_trans_handle *trans;
2963         unsigned long nr = 0;
2964
2965         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE ||
2966             btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
2967                 return -ENOTEMPTY;
2968
2969         trans = __unlink_start_trans(dir, dentry);
2970         if (IS_ERR(trans))
2971                 return PTR_ERR(trans);
2972
2973         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
2974                 err = btrfs_unlink_subvol(trans, root, dir,
2975                                           BTRFS_I(inode)->location.objectid,
2976                                           dentry->d_name.name,
2977                                           dentry->d_name.len);
2978                 goto out;
2979         }
2980
2981         err = btrfs_orphan_add(trans, inode);
2982         if (err)
2983                 goto out;
2984
2985         /* now the directory is empty */
2986         err = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
2987                                  dentry->d_name.name, dentry->d_name.len);
2988         if (!err)
2989                 btrfs_i_size_write(inode, 0);
2990 out:
2991         nr = trans->blocks_used;
2992         __unlink_end_trans(trans, root);
2993         btrfs_btree_balance_dirty(root, nr);
2994
2995         return err;
2996 }
2997
2998 /*
2999  * this can truncate away extent items, csum items and directory items.
3000  * It starts at a high offset and removes keys until it can't find
3001  * any higher than new_size
3002  *
3003  * csum items that cross the new i_size are truncated to the new size
3004  * as well.
3005  *
3006  * min_type is the minimum key type to truncate down to.  If set to 0, this
3007  * will kill all the items on this inode, including the INODE_ITEM_KEY.
3008  */
3009 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
3010                                struct btrfs_root *root,
3011                                struct inode *inode,
3012                                u64 new_size, u32 min_type)
3013 {
3014         struct btrfs_path *path;
3015         struct extent_buffer *leaf;
3016         struct btrfs_file_extent_item *fi;
3017         struct btrfs_key key;
3018         struct btrfs_key found_key;
3019         u64 extent_start = 0;
3020         u64 extent_num_bytes = 0;
3021         u64 extent_offset = 0;
3022         u64 item_end = 0;
3023         u64 mask = root->sectorsize - 1;
3024         u32 found_type = (u8)-1;
3025         int found_extent;
3026         int del_item;
3027         int pending_del_nr = 0;
3028         int pending_del_slot = 0;
3029         int extent_type = -1;
3030         int ret;
3031         int err = 0;
3032         u64 ino = btrfs_ino(inode);
3033
3034         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
3035
3036         path = btrfs_alloc_path();
3037         if (!path)
3038                 return -ENOMEM;
3039         path->reada = -1;
3040
3041         if (root->ref_cows || root == root->fs_info->tree_root)
3042                 btrfs_drop_extent_cache(inode, new_size & (~mask), (u64)-1, 0);
3043
3044         /*
3045          * This function is also used to drop the items in the log tree before
3046          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
3047          * it is used to drop the loged items. So we shouldn't kill the delayed
3048          * items.
3049          */
3050         if (min_type == 0 && root == BTRFS_I(inode)->root)
3051                 btrfs_kill_delayed_inode_items(inode);
3052
3053         key.objectid = ino;
3054         key.offset = (u64)-1;
3055         key.type = (u8)-1;
3056
3057 search_again:
3058         path->leave_spinning = 1;
3059         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3060         if (ret < 0) {
3061                 err = ret;
3062                 goto out;
3063         }
3064
3065         if (ret > 0) {
3066                 /* there are no items in the tree for us to truncate, we're
3067                  * done
3068                  */
3069                 if (path->slots[0] == 0)
3070                         goto out;
3071                 path->slots[0]--;
3072         }
3073
3074         while (1) {
3075                 fi = NULL;
3076                 leaf = path->nodes[0];
3077                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3078                 found_type = btrfs_key_type(&found_key);
3079
3080                 if (found_key.objectid != ino)
3081                         break;
3082
3083                 if (found_type < min_type)
3084                         break;
3085
3086                 item_end = found_key.offset;
3087                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
3088                         fi = btrfs_item_ptr(leaf, path->slots[0],
3089                                             struct btrfs_file_extent_item);
3090                         extent_type = btrfs_file_extent_type(leaf, fi);
3091                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3092                                 item_end +=
3093                                     btrfs_file_extent_num_bytes(leaf, fi);
3094                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3095                                 item_end += btrfs_file_extent_inline_len(leaf,
3096                                                                          fi);
3097                         }
3098                         item_end--;
3099                 }
3100                 if (found_type > min_type) {
3101                         del_item = 1;
3102                 } else {
3103                         if (item_end < new_size)
3104                                 break;
3105                         if (found_key.offset >= new_size)
3106                                 del_item = 1;
3107                         else
3108                                 del_item = 0;
3109                 }
3110                 found_extent = 0;
3111                 /* FIXME, shrink the extent if the ref count is only 1 */
3112                 if (found_type != BTRFS_EXTENT_DATA_KEY)
3113                         goto delete;
3114
3115                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3116                         u64 num_dec;
3117                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
3118                         if (!del_item) {
3119                                 u64 orig_num_bytes =
3120                                         btrfs_file_extent_num_bytes(leaf, fi);
3121                                 extent_num_bytes = new_size -
3122                                         found_key.offset + root->sectorsize - 1;
3123                                 extent_num_bytes = extent_num_bytes &
3124                                         ~((u64)root->sectorsize - 1);
3125                                 btrfs_set_file_extent_num_bytes(leaf, fi,
3126                                                          extent_num_bytes);
3127                                 num_dec = (orig_num_bytes -
3128                                            extent_num_bytes);
3129                                 if (root->ref_cows && extent_start != 0)
3130                                         inode_sub_bytes(inode, num_dec);
3131                                 btrfs_mark_buffer_dirty(leaf);
3132                         } else {
3133                                 extent_num_bytes =
3134                                         btrfs_file_extent_disk_num_bytes(leaf,
3135                                                                          fi);
3136                                 extent_offset = found_key.offset -
3137                                         btrfs_file_extent_offset(leaf, fi);
3138
3139                                 /* FIXME blocksize != 4096 */
3140                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
3141                                 if (extent_start != 0) {
3142                                         found_extent = 1;
3143                                         if (root->ref_cows)
3144                                                 inode_sub_bytes(inode, num_dec);
3145                                 }
3146                         }
3147                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3148                         /*
3149                          * we can't truncate inline items that have had
3150                          * special encodings
3151                          */
3152                         if (!del_item &&
3153                             btrfs_file_extent_compression(leaf, fi) == 0 &&
3154                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
3155                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
3156                                 u32 size = new_size - found_key.offset;
3157
3158                                 if (root->ref_cows) {
3159                                         inode_sub_bytes(inode, item_end + 1 -
3160                                                         new_size);
3161                                 }
3162                                 size =
3163                                     btrfs_file_extent_calc_inline_size(size);
3164                                 ret = btrfs_truncate_item(trans, root, path,
3165                                                           size, 1);
3166                         } else if (root->ref_cows) {
3167                                 inode_sub_bytes(inode, item_end + 1 -
3168                                                 found_key.offset);
3169                         }
3170                 }
3171 delete:
3172                 if (del_item) {
3173                         if (!pending_del_nr) {
3174                                 /* no pending yet, add ourselves */
3175                                 pending_del_slot = path->slots[0];
3176                                 pending_del_nr = 1;
3177                         } else if (pending_del_nr &&
3178                                    path->slots[0] + 1 == pending_del_slot) {
3179                                 /* hop on the pending chunk */
3180                                 pending_del_nr++;
3181                                 pending_del_slot = path->slots[0];
3182                         } else {
3183                                 BUG();
3184                         }
3185                 } else {
3186                         break;
3187                 }
3188                 if (found_extent && (root->ref_cows ||
3189                                      root == root->fs_info->tree_root)) {
3190                         btrfs_set_path_blocking(path);
3191                         ret = btrfs_free_extent(trans, root, extent_start,
3192                                                 extent_num_bytes, 0,
3193                                                 btrfs_header_owner(leaf),
3194                                                 ino, extent_offset, 0);
3195                         BUG_ON(ret);
3196                 }
3197
3198                 if (found_type == BTRFS_INODE_ITEM_KEY)
3199                         break;
3200
3201                 if (path->slots[0] == 0 ||
3202                     path->slots[0] != pending_del_slot) {
3203                         if (root->ref_cows &&
3204                             BTRFS_I(inode)->location.objectid !=
3205                                                 BTRFS_FREE_INO_OBJECTID) {
3206                                 err = -EAGAIN;
3207                                 goto out;
3208                         }
3209                         if (pending_del_nr) {
3210                                 ret = btrfs_del_items(trans, root, path,
3211                                                 pending_del_slot,