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/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include "ctree.h"
  46 #include "disk-io.h"
  47 #include "transaction.h"
  48 #include "btrfs_inode.h"
  49 #include "print-tree.h"
  50 #include "ordered-data.h"
  51 #include "xattr.h"
  52 #include "tree-log.h"
  53 #include "volumes.h"
  54 #include "compression.h"
  55 #include "locking.h"
  56 #include "free-space-cache.h"
  57 #include "inode-map.h"
  58 #include "backref.h"
  59 #include "hash.h"
  60 #include "props.h"
  61 #include "qgroup.h"
  62 #include "dedupe.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74         int overwrite;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_transaction_cachep;
  90 struct kmem_cache *btrfs_path_cachep;
  91 struct kmem_cache *btrfs_free_space_cachep;
  92
  93 #define S_SHIFT 12
  94 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  95         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  96         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  97         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  98         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  99         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 100         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 101         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 102 };
 103
 104 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 105 static int btrfs_truncate(struct inode *inode);
 106 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 107 static noinline int cow_file_range(struct inode *inode,
 108                                    struct page *locked_page,
 109                                    u64 start, u64 end, u64 delalloc_end,
 110                                    int *page_started, unsigned long *nr_written,
 111                                    int unlock, struct btrfs_dedupe_hash *hash);
 112 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 113                                        u64 orig_start, u64 block_start,
 114                                        u64 block_len, u64 orig_block_len,
 115                                        u64 ram_bytes, int compress_type,
 116                                        int type);
 117
 118 static void __endio_write_update_ordered(struct inode *inode,
 119                                          const u64 offset, const u64 bytes,
 120                                          const bool uptodate);
 121
 122 /*
 123  * Cleanup all submitted ordered extents in specified range to handle errors
 124  * from the fill_dellaloc() callback.
 125  *
 126  * NOTE: caller must ensure that when an error happens, it can not call
 127  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 128  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 129  * to be released, which we want to happen only when finishing the ordered
 130  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 131  * fill_delalloc() callback already does proper cleanup for the first page of
 132  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 133  * range of the first page.
 134  */
 135 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 136                                                  const u64 offset,
 137                                                  const u64 bytes)
 138 {
 139         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 140                                             bytes - PAGE_SIZE, false);
 141 }
 142
 143 static int btrfs_dirty_inode(struct inode *inode);
 144
 145 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 146 void btrfs_test_inode_set_ops(struct inode *inode)
 147 {
 148         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 149 }
 150 #endif
 151
 152 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 153                                      struct inode *inode,  struct inode *dir,
 154                                      const struct qstr *qstr)
 155 {
 156         int err;
 157
 158         err = btrfs_init_acl(trans, inode, dir);
 159         if (!err)
 160                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 161         return err;
 162 }
 163
 164 /*
 165  * this does all the hard work for inserting an inline extent into
 166  * the btree.  The caller should have done a btrfs_drop_extents so that
 167  * no overlapping inline items exist in the btree
 168  */
 169 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 170                                 struct btrfs_path *path, int extent_inserted,
 171                                 struct btrfs_root *root, struct inode *inode,
 172                                 u64 start, size_t size, size_t compressed_size,
 173                                 int compress_type,
 174                                 struct page **compressed_pages)
 175 {
 176         struct extent_buffer *leaf;
 177         struct page *page = NULL;
 178         char *kaddr;
 179         unsigned long ptr;
 180         struct btrfs_file_extent_item *ei;
 181         int err = 0;
 182         int ret;
 183         size_t cur_size = size;
 184         unsigned long offset;
 185
 186         if (compressed_size && compressed_pages)
 187                 cur_size = compressed_size;
 188
 189         inode_add_bytes(inode, size);
 190
 191         if (!extent_inserted) {
 192                 struct btrfs_key key;
 193                 size_t datasize;
 194
 195                 key.objectid = btrfs_ino(BTRFS_I(inode));
 196                 key.offset = start;
 197                 key.type = BTRFS_EXTENT_DATA_KEY;
 198
 199                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 200                 path->leave_spinning = 1;
 201                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 202                                               datasize);
 203                 if (ret) {
 204                         err = ret;
 205                         goto fail;
 206                 }
 207         }
 208         leaf = path->nodes[0];
 209         ei = btrfs_item_ptr(leaf, path->slots[0],
 210                             struct btrfs_file_extent_item);
 211         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 212         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 213         btrfs_set_file_extent_encryption(leaf, ei, 0);
 214         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 215         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 216         ptr = btrfs_file_extent_inline_start(ei);
 217
 218         if (compress_type != BTRFS_COMPRESS_NONE) {
 219                 struct page *cpage;
 220                 int i = 0;
 221                 while (compressed_size > 0) {
 222                         cpage = compressed_pages[i];
 223                         cur_size = min_t(unsigned long, compressed_size,
 224                                        PAGE_SIZE);
 225
 226                         kaddr = kmap_atomic(cpage);
 227                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 228                         kunmap_atomic(kaddr);
 229
 230                         i++;
 231                         ptr += cur_size;
 232                         compressed_size -= cur_size;
 233                 }
 234                 btrfs_set_file_extent_compression(leaf, ei,
 235                                                   compress_type);
 236         } else {
 237                 page = find_get_page(inode->i_mapping,
 238                                      start >> PAGE_SHIFT);
 239                 btrfs_set_file_extent_compression(leaf, ei, 0);
 240                 kaddr = kmap_atomic(page);
 241                 offset = start & (PAGE_SIZE - 1);
 242                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 243                 kunmap_atomic(kaddr);
 244                 put_page(page);
 245         }
 246         btrfs_mark_buffer_dirty(leaf);
 247         btrfs_release_path(path);
 248
 249         /*
 250          * we're an inline extent, so nobody can
 251          * extend the file past i_size without locking
 252          * a page we already have locked.
 253          *
 254          * We must do any isize and inode updates
 255          * before we unlock the pages.  Otherwise we
 256          * could end up racing with unlink.
 257          */
 258         BTRFS_I(inode)->disk_i_size = inode->i_size;
 259         ret = btrfs_update_inode(trans, root, inode);
 260
 261         return ret;
 262 fail:
 263         return err;
 264 }
 265
 266
 267 /*
 268  * conditionally insert an inline extent into the file.  This
 269  * does the checks required to make sure the data is small enough
 270  * to fit as an inline extent.
 271  */
 272 static noinline int cow_file_range_inline(struct btrfs_root *root,
 273                                           struct inode *inode, u64 start,
 274                                           u64 end, size_t compressed_size,
 275                                           int compress_type,
 276                                           struct page **compressed_pages)
 277 {
 278         struct btrfs_fs_info *fs_info = root->fs_info;
 279         struct btrfs_trans_handle *trans;
 280         u64 isize = i_size_read(inode);
 281         u64 actual_end = min(end + 1, isize);
 282         u64 inline_len = actual_end - start;
 283         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 284         u64 data_len = inline_len;
 285         int ret;
 286         struct btrfs_path *path;
 287         int extent_inserted = 0;
 288         u32 extent_item_size;
 289
 290         if (compressed_size)
 291                 data_len = compressed_size;
 292
 293         if (start > 0 ||
 294             actual_end > fs_info->sectorsize ||
 295             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 296             (!compressed_size &&
 297             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 298             end + 1 < isize ||
 299             data_len > fs_info->max_inline) {
 300                 return 1;
 301         }
 302
 303         path = btrfs_alloc_path();
 304         if (!path)
 305                 return -ENOMEM;
 306
 307         trans = btrfs_join_transaction(root);
 308         if (IS_ERR(trans)) {
 309                 btrfs_free_path(path);
 310                 return PTR_ERR(trans);
 311         }
 312         trans->block_rsv = &fs_info->delalloc_block_rsv;
 313
 314         if (compressed_size && compressed_pages)
 315                 extent_item_size = btrfs_file_extent_calc_inline_size(
 316                    compressed_size);
 317         else
 318                 extent_item_size = btrfs_file_extent_calc_inline_size(
 319                     inline_len);
 320
 321         ret = __btrfs_drop_extents(trans, root, inode, path,
 322                                    start, aligned_end, NULL,
 323                                    1, 1, extent_item_size, &extent_inserted);
 324         if (ret) {
 325                 btrfs_abort_transaction(trans, ret);
 326                 goto out;
 327         }
 328
 329         if (isize > actual_end)
 330                 inline_len = min_t(u64, isize, actual_end);
 331         ret = insert_inline_extent(trans, path, extent_inserted,
 332                                    root, inode, start,
 333                                    inline_len, compressed_size,
 334                                    compress_type, compressed_pages);
 335         if (ret && ret != -ENOSPC) {
 336                 btrfs_abort_transaction(trans, ret);
 337                 goto out;
 338         } else if (ret == -ENOSPC) {
 339                 ret = 1;
 340                 goto out;
 341         }
 342
 343         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 344         btrfs_delalloc_release_metadata(BTRFS_I(inode), end + 1 - start);
 345         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 346 out:
 347         /*
 348          * Don't forget to free the reserved space, as for inlined extent
 349          * it won't count as data extent, free them directly here.
 350          * And at reserve time, it's always aligned to page size, so
 351          * just free one page here.
 352          */
 353         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 354         btrfs_free_path(path);
 355         btrfs_end_transaction(trans);
 356         return ret;
 357 }
 358
 359 struct async_extent {
 360         u64 start;
 361         u64 ram_size;
 362         u64 compressed_size;
 363         struct page **pages;
 364         unsigned long nr_pages;
 365         int compress_type;
 366         struct list_head list;
 367 };
 368
 369 struct async_cow {
 370         struct inode *inode;
 371         struct btrfs_root *root;
 372         struct page *locked_page;
 373         u64 start;
 374         u64 end;
 375         struct list_head extents;
 376         struct btrfs_work work;
 377 };
 378
 379 static noinline int add_async_extent(struct async_cow *cow,
 380                                      u64 start, u64 ram_size,
 381                                      u64 compressed_size,
 382                                      struct page **pages,
 383                                      unsigned long nr_pages,
 384                                      int compress_type)
 385 {
 386         struct async_extent *async_extent;
 387
 388         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 389         BUG_ON(!async_extent); /* -ENOMEM */
 390         async_extent->start = start;
 391         async_extent->ram_size = ram_size;
 392         async_extent->compressed_size = compressed_size;
 393         async_extent->pages = pages;
 394         async_extent->nr_pages = nr_pages;
 395         async_extent->compress_type = compress_type;
 396         list_add_tail(&async_extent->list, &cow->extents);
 397         return 0;
 398 }
 399
 400 static inline int inode_need_compress(struct inode *inode)
 401 {
 402         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 403
 404         /* force compress */
 405         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 406                 return 1;
 407         /* bad compression ratios */
 408         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 409                 return 0;
 410         if (btrfs_test_opt(fs_info, COMPRESS) ||
 411             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 412             BTRFS_I(inode)->force_compress)
 413                 return 1;
 414         return 0;
 415 }
 416
 417 static inline void inode_should_defrag(struct btrfs_inode *inode,
 418                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 419 {
 420         /* If this is a small write inside eof, kick off a defrag */
 421         if (num_bytes < small_write &&
 422             (start > 0 || end + 1 < inode->disk_i_size))
 423                 btrfs_add_inode_defrag(NULL, inode);
 424 }
 425
 426 /*
 427  * we create compressed extents in two phases.  The first
 428  * phase compresses a range of pages that have already been
 429  * locked (both pages and state bits are locked).
 430  *
 431  * This is done inside an ordered work queue, and the compression
 432  * is spread across many cpus.  The actual IO submission is step
 433  * two, and the ordered work queue takes care of making sure that
 434  * happens in the same order things were put onto the queue by
 435  * writepages and friends.
 436  *
 437  * If this code finds it can't get good compression, it puts an
 438  * entry onto the work queue to write the uncompressed bytes.  This
 439  * makes sure that both compressed inodes and uncompressed inodes
 440  * are written in the same order that the flusher thread sent them
 441  * down.
 442  */
 443 static noinline void compress_file_range(struct inode *inode,
 444                                         struct page *locked_page,
 445                                         u64 start, u64 end,
 446                                         struct async_cow *async_cow,
 447                                         int *num_added)
 448 {
 449         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 450         struct btrfs_root *root = BTRFS_I(inode)->root;
 451         u64 num_bytes;
 452         u64 blocksize = fs_info->sectorsize;
 453         u64 actual_end;
 454         u64 isize = i_size_read(inode);
 455         int ret = 0;
 456         struct page **pages = NULL;
 457         unsigned long nr_pages;
 458         unsigned long total_compressed = 0;
 459         unsigned long total_in = 0;
 460         int i;
 461         int will_compress;
 462         int compress_type = fs_info->compress_type;
 463         int redirty = 0;
 464
 465         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 466                         SZ_16K);
 467
 468         actual_end = min_t(u64, isize, end + 1);
 469 again:
 470         will_compress = 0;
 471         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 472         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 473         nr_pages = min_t(unsigned long, nr_pages,
 474                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 475
 476         /*
 477          * we don't want to send crud past the end of i_size through
 478          * compression, that's just a waste of CPU time.  So, if the
 479          * end of the file is before the start of our current
 480          * requested range of bytes, we bail out to the uncompressed
 481          * cleanup code that can deal with all of this.
 482          *
 483          * It isn't really the fastest way to fix things, but this is a
 484          * very uncommon corner.
 485          */
 486         if (actual_end <= start)
 487                 goto cleanup_and_bail_uncompressed;
 488
 489         total_compressed = actual_end - start;
 490
 491         /*
 492          * skip compression for a small file range(<=blocksize) that
 493          * isn't an inline extent, since it doesn't save disk space at all.
 494          */
 495         if (total_compressed <= blocksize &&
 496            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 497                 goto cleanup_and_bail_uncompressed;
 498
 499         total_compressed = min_t(unsigned long, total_compressed,
 500                         BTRFS_MAX_UNCOMPRESSED);
 501         num_bytes = ALIGN(end - start + 1, blocksize);
 502         num_bytes = max(blocksize,  num_bytes);
 503         total_in = 0;
 504         ret = 0;
 505
 506         /*
 507          * we do compression for mount -o compress and when the
 508          * inode has not been flagged as nocompress.  This flag can
 509          * change at any time if we discover bad compression ratios.
 510          */
 511         if (inode_need_compress(inode)) {
 512                 WARN_ON(pages);
 513                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 514                 if (!pages) {
 515                         /* just bail out to the uncompressed code */
 516                         goto cont;
 517                 }
 518
 519                 if (BTRFS_I(inode)->force_compress)
 520                         compress_type = BTRFS_I(inode)->force_compress;
 521
 522                 /*
 523                  * we need to call clear_page_dirty_for_io on each
 524                  * page in the range.  Otherwise applications with the file
 525                  * mmap'd can wander in and change the page contents while
 526                  * we are compressing them.
 527                  *
 528                  * If the compression fails for any reason, we set the pages
 529                  * dirty again later on.
 530                  */
 531                 extent_range_clear_dirty_for_io(inode, start, end);
 532                 redirty = 1;
 533                 ret = btrfs_compress_pages(compress_type,
 534                                            inode->i_mapping, start,
 535                                            pages,
 536                                            &nr_pages,
 537                                            &total_in,
 538                                            &total_compressed);
 539
 540                 if (!ret) {
 541                         unsigned long offset = total_compressed &
 542                                 (PAGE_SIZE - 1);
 543                         struct page *page = pages[nr_pages - 1];
 544                         char *kaddr;
 545
 546                         /* zero the tail end of the last page, we might be
 547                          * sending it down to disk
 548                          */
 549                         if (offset) {
 550                                 kaddr = kmap_atomic(page);
 551                                 memset(kaddr + offset, 0,
 552                                        PAGE_SIZE - offset);
 553                                 kunmap_atomic(kaddr);
 554                         }
 555                         will_compress = 1;
 556                 }
 557         }
 558 cont:
 559         if (start == 0) {
 560                 /* lets try to make an inline extent */
 561                 if (ret || total_in < (actual_end - start)) {
 562                         /* we didn't compress the entire range, try
 563                          * to make an uncompressed inline extent.
 564                          */
 565                         ret = cow_file_range_inline(root, inode, start, end,
 566                                             0, BTRFS_COMPRESS_NONE, NULL);
 567                 } else {
 568                         /* try making a compressed inline extent */
 569                         ret = cow_file_range_inline(root, inode, start, end,
 570                                                     total_compressed,
 571                                                     compress_type, pages);
 572                 }
 573                 if (ret <= 0) {
 574                         unsigned long clear_flags = EXTENT_DELALLOC |
 575                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG;
 576                         unsigned long page_error_op;
 577
 578                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 579                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 580
 581                         /*
 582                          * inline extent creation worked or returned error,
 583                          * we don't need to create any more async work items.
 584                          * Unlock and free up our temp pages.
 585                          */
 586                         extent_clear_unlock_delalloc(inode, start, end, end,
 587                                                      NULL, clear_flags,
 588                                                      PAGE_UNLOCK |
 589                                                      PAGE_CLEAR_DIRTY |
 590                                                      PAGE_SET_WRITEBACK |
 591                                                      page_error_op |
 592                                                      PAGE_END_WRITEBACK);
 593                         if (ret == 0)
 594                                 btrfs_free_reserved_data_space_noquota(inode,
 595                                                                start,
 596                                                                end - start + 1);
 597                         goto free_pages_out;
 598                 }
 599         }
 600
 601         if (will_compress) {
 602                 /*
 603                  * we aren't doing an inline extent round the compressed size
 604                  * up to a block size boundary so the allocator does sane
 605                  * things
 606                  */
 607                 total_compressed = ALIGN(total_compressed, blocksize);
 608
 609                 /*
 610                  * one last check to make sure the compression is really a
 611                  * win, compare the page count read with the blocks on disk
 612                  */
 613                 total_in = ALIGN(total_in, PAGE_SIZE);
 614                 if (total_compressed >= total_in) {
 615                         will_compress = 0;
 616                 } else {
 617                         num_bytes = total_in;
 618                         *num_added += 1;
 619
 620                         /*
 621                          * The async work queues will take care of doing actual
 622                          * allocation on disk for these compressed pages, and
 623                          * will submit them to the elevator.
 624                          */
 625                         add_async_extent(async_cow, start, num_bytes,
 626                                         total_compressed, pages, nr_pages,
 627                                         compress_type);
 628
 629                         if (start + num_bytes < end) {
 630                                 start += num_bytes;
 631                                 pages = NULL;
 632                                 cond_resched();
 633                                 goto again;
 634                         }
 635                         return;
 636                 }
 637         }
 638         if (pages) {
 639                 /*
 640                  * the compression code ran but failed to make things smaller,
 641                  * free any pages it allocated and our page pointer array
 642                  */
 643                 for (i = 0; i < nr_pages; i++) {
 644                         WARN_ON(pages[i]->mapping);
 645                         put_page(pages[i]);
 646                 }
 647                 kfree(pages);
 648                 pages = NULL;
 649                 total_compressed = 0;
 650                 nr_pages = 0;
 651
 652                 /* flag the file so we don't compress in the future */
 653                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 654                     !(BTRFS_I(inode)->force_compress)) {
 655                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 656                 }
 657         }
 658 cleanup_and_bail_uncompressed:
 659         /*
 660          * No compression, but we still need to write the pages in the file
 661          * we've been given so far.  redirty the locked page if it corresponds
 662          * to our extent and set things up for the async work queue to run
 663          * cow_file_range to do the normal delalloc dance.
 664          */
 665         if (page_offset(locked_page) >= start &&
 666             page_offset(locked_page) <= end)
 667                 __set_page_dirty_nobuffers(locked_page);
 668                 /* unlocked later on in the async handlers */
 669
 670         if (redirty)
 671                 extent_range_redirty_for_io(inode, start, end);
 672         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 673                          BTRFS_COMPRESS_NONE);
 674         *num_added += 1;
 675
 676         return;
 677
 678 free_pages_out:
 679         for (i = 0; i < nr_pages; i++) {
 680                 WARN_ON(pages[i]->mapping);
 681                 put_page(pages[i]);
 682         }
 683         kfree(pages);
 684 }
 685
 686 static void free_async_extent_pages(struct async_extent *async_extent)
 687 {
 688         int i;
 689
 690         if (!async_extent->pages)
 691                 return;
 692
 693         for (i = 0; i < async_extent->nr_pages; i++) {
 694                 WARN_ON(async_extent->pages[i]->mapping);
 695                 put_page(async_extent->pages[i]);
 696         }
 697         kfree(async_extent->pages);
 698         async_extent->nr_pages = 0;
 699         async_extent->pages = NULL;
 700 }
 701
 702 /*
 703  * phase two of compressed writeback.  This is the ordered portion
 704  * of the code, which only gets called in the order the work was
 705  * queued.  We walk all the async extents created by compress_file_range
 706  * and send them down to the disk.
 707  */
 708 static noinline void submit_compressed_extents(struct inode *inode,
 709                                               struct async_cow *async_cow)
 710 {
 711         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 712         struct async_extent *async_extent;
 713         u64 alloc_hint = 0;
 714         struct btrfs_key ins;
 715         struct extent_map *em;
 716         struct btrfs_root *root = BTRFS_I(inode)->root;
 717         struct extent_io_tree *io_tree;
 718         int ret = 0;
 719
 720 again:
 721         while (!list_empty(&async_cow->extents)) {
 722                 async_extent = list_entry(async_cow->extents.next,
 723                                           struct async_extent, list);
 724                 list_del(&async_extent->list);
 725
 726                 io_tree = &BTRFS_I(inode)->io_tree;
 727
 728 retry:
 729                 /* did the compression code fall back to uncompressed IO? */
 730                 if (!async_extent->pages) {
 731                         int page_started = 0;
 732                         unsigned long nr_written = 0;
 733
 734                         lock_extent(io_tree, async_extent->start,
 735                                          async_extent->start +
 736                                          async_extent->ram_size - 1);
 737
 738                         /* allocate blocks */
 739                         ret = cow_file_range(inode, async_cow->locked_page,
 740                                              async_extent->start,
 741                                              async_extent->start +
 742                                              async_extent->ram_size - 1,
 743                                              async_extent->start +
 744                                              async_extent->ram_size - 1,
 745                                              &page_started, &nr_written, 0,
 746                                              NULL);
 747
 748                         /* JDM XXX */
 749
 750                         /*
 751                          * if page_started, cow_file_range inserted an
 752                          * inline extent and took care of all the unlocking
 753                          * and IO for us.  Otherwise, we need to submit
 754                          * all those pages down to the drive.
 755                          */
 756                         if (!page_started && !ret)
 757                                 extent_write_locked_range(io_tree,
 758                                                   inode, async_extent->start,
 759                                                   async_extent->start +
 760                                                   async_extent->ram_size - 1,
 761                                                   btrfs_get_extent,
 762                                                   WB_SYNC_ALL);
 763                         else if (ret)
 764                                 unlock_page(async_cow->locked_page);
 765                         kfree(async_extent);
 766                         cond_resched();
 767                         continue;
 768                 }
 769
 770                 lock_extent(io_tree, async_extent->start,
 771                             async_extent->start + async_extent->ram_size - 1);
 772
 773                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 774                                            async_extent->compressed_size,
 775                                            async_extent->compressed_size,
 776                                            0, alloc_hint, &ins, 1, 1);
 777                 if (ret) {
 778                         free_async_extent_pages(async_extent);
 779
 780                         if (ret == -ENOSPC) {
 781                                 unlock_extent(io_tree, async_extent->start,
 782                                               async_extent->start +
 783                                               async_extent->ram_size - 1);
 784
 785                                 /*
 786                                  * we need to redirty the pages if we decide to
 787                                  * fallback to uncompressed IO, otherwise we
 788                                  * will not submit these pages down to lower
 789                                  * layers.
 790                                  */
 791                                 extent_range_redirty_for_io(inode,
 792                                                 async_extent->start,
 793                                                 async_extent->start +
 794                                                 async_extent->ram_size - 1);
 795
 796                                 goto retry;
 797                         }
 798                         goto out_free;
 799                 }
 800                 /*
 801                  * here we're doing allocation and writeback of the
 802                  * compressed pages
 803                  */
 804                 em = create_io_em(inode, async_extent->start,
 805                                   async_extent->ram_size, /* len */
 806                                   async_extent->start, /* orig_start */
 807                                   ins.objectid, /* block_start */
 808                                   ins.offset, /* block_len */
 809                                   ins.offset, /* orig_block_len */
 810                                   async_extent->ram_size, /* ram_bytes */
 811                                   async_extent->compress_type,
 812                                   BTRFS_ORDERED_COMPRESSED);
 813                 if (IS_ERR(em))
 814                         /* ret value is not necessary due to void function */
 815                         goto out_free_reserve;
 816                 free_extent_map(em);
 817
 818                 ret = btrfs_add_ordered_extent_compress(inode,
 819                                                 async_extent->start,
 820                                                 ins.objectid,
 821                                                 async_extent->ram_size,
 822                                                 ins.offset,
 823                                                 BTRFS_ORDERED_COMPRESSED,
 824                                                 async_extent->compress_type);
 825                 if (ret) {
 826                         btrfs_drop_extent_cache(BTRFS_I(inode),
 827                                                 async_extent->start,
 828                                                 async_extent->start +
 829                                                 async_extent->ram_size - 1, 0);
 830                         goto out_free_reserve;
 831                 }
 832                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 833
 834                 /*
 835                  * clear dirty, set writeback and unlock the pages.
 836                  */
 837                 extent_clear_unlock_delalloc(inode, async_extent->start,
 838                                 async_extent->start +
 839                                 async_extent->ram_size - 1,
 840                                 async_extent->start +
 841                                 async_extent->ram_size - 1,
 842                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 843                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 844                                 PAGE_SET_WRITEBACK);
 845                 if (btrfs_submit_compressed_write(inode,
 846                                     async_extent->start,
 847                                     async_extent->ram_size,
 848                                     ins.objectid,
 849                                     ins.offset, async_extent->pages,
 850                                     async_extent->nr_pages)) {
 851                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 852                         struct page *p = async_extent->pages[0];
 853                         const u64 start = async_extent->start;
 854                         const u64 end = start + async_extent->ram_size - 1;
 855
 856                         p->mapping = inode->i_mapping;
 857                         tree->ops->writepage_end_io_hook(p, start, end,
 858                                                          NULL, 0);
 859                         p->mapping = NULL;
 860                         extent_clear_unlock_delalloc(inode, start, end, end,
 861                                                      NULL, 0,
 862                                                      PAGE_END_WRITEBACK |
 863                                                      PAGE_SET_ERROR);
 864                         free_async_extent_pages(async_extent);
 865                 }
 866                 alloc_hint = ins.objectid + ins.offset;
 867                 kfree(async_extent);
 868                 cond_resched();
 869         }
 870         return;
 871 out_free_reserve:
 872         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 873         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 874 out_free:
 875         extent_clear_unlock_delalloc(inode, async_extent->start,
 876                                      async_extent->start +
 877                                      async_extent->ram_size - 1,
 878                                      async_extent->start +
 879                                      async_extent->ram_size - 1,
 880                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 881                                      EXTENT_DELALLOC_NEW |
 882                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 883                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 884                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 885                                      PAGE_SET_ERROR);
 886         free_async_extent_pages(async_extent);
 887         kfree(async_extent);
 888         goto again;
 889 }
 890
 891 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 892                                       u64 num_bytes)
 893 {
 894         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 895         struct extent_map *em;
 896         u64 alloc_hint = 0;
 897
 898         read_lock(&em_tree->lock);
 899         em = search_extent_mapping(em_tree, start, num_bytes);
 900         if (em) {
 901                 /*
 902                  * if block start isn't an actual block number then find the
 903                  * first block in this inode and use that as a hint.  If that
 904                  * block is also bogus then just don't worry about it.
 905                  */
 906                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 907                         free_extent_map(em);
 908                         em = search_extent_mapping(em_tree, 0, 0);
 909                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 910                                 alloc_hint = em->block_start;
 911                         if (em)
 912                                 free_extent_map(em);
 913                 } else {
 914                         alloc_hint = em->block_start;
 915                         free_extent_map(em);
 916                 }
 917         }
 918         read_unlock(&em_tree->lock);
 919
 920         return alloc_hint;
 921 }
 922
 923 /*
 924  * when extent_io.c finds a delayed allocation range in the file,
 925  * the call backs end up in this code.  The basic idea is to
 926  * allocate extents on disk for the range, and create ordered data structs
 927  * in ram to track those extents.
 928  *
 929  * locked_page is the page that writepage had locked already.  We use
 930  * it to make sure we don't do extra locks or unlocks.
 931  *
 932  * *page_started is set to one if we unlock locked_page and do everything
 933  * required to start IO on it.  It may be clean and already done with
 934  * IO when we return.
 935  */
 936 static noinline int cow_file_range(struct inode *inode,
 937                                    struct page *locked_page,
 938                                    u64 start, u64 end, u64 delalloc_end,
 939                                    int *page_started, unsigned long *nr_written,
 940                                    int unlock, struct btrfs_dedupe_hash *hash)
 941 {
 942         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 943         struct btrfs_root *root = BTRFS_I(inode)->root;
 944         u64 alloc_hint = 0;
 945         u64 num_bytes;
 946         unsigned long ram_size;
 947         u64 disk_num_bytes;
 948         u64 cur_alloc_size = 0;
 949         u64 blocksize = fs_info->sectorsize;
 950         struct btrfs_key ins;
 951         struct extent_map *em;
 952         unsigned clear_bits;
 953         unsigned long page_ops;
 954         bool extent_reserved = false;
 955         int ret = 0;
 956
 957         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 958                 WARN_ON_ONCE(1);
 959                 ret = -EINVAL;
 960                 goto out_unlock;
 961         }
 962
 963         num_bytes = ALIGN(end - start + 1, blocksize);
 964         num_bytes = max(blocksize,  num_bytes);
 965         disk_num_bytes = num_bytes;
 966
 967         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 968
 969         if (start == 0) {
 970                 /* lets try to make an inline extent */
 971                 ret = cow_file_range_inline(root, inode, start, end, 0,
 972                                         BTRFS_COMPRESS_NONE, NULL);
 973                 if (ret == 0) {
 974                         extent_clear_unlock_delalloc(inode, start, end,
 975                                      delalloc_end, NULL,
 976                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 977                                      EXTENT_DELALLOC_NEW |
 978                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 979                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 980                                      PAGE_END_WRITEBACK);
 981                         btrfs_free_reserved_data_space_noquota(inode, start,
 982                                                 end - start + 1);
 983                         *nr_written = *nr_written +
 984                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 985                         *page_started = 1;
 986                         goto out;
 987                 } else if (ret < 0) {
 988                         goto out_unlock;
 989                 }
 990         }
 991
 992         BUG_ON(disk_num_bytes >
 993                btrfs_super_total_bytes(fs_info->super_copy));
 994
 995         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 996         btrfs_drop_extent_cache(BTRFS_I(inode), start,
 997                         start + num_bytes - 1, 0);
 998
 999         while (disk_num_bytes > 0) {
1000                 cur_alloc_size = disk_num_bytes;
1001                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1002                                            fs_info->sectorsize, 0, alloc_hint,
1003                                            &ins, 1, 1);
1004                 if (ret < 0)
1005                         goto out_unlock;
1006                 cur_alloc_size = ins.offset;
1007                 extent_reserved = true;
1008
1009                 ram_size = ins.offset;
1010                 em = create_io_em(inode, start, ins.offset, /* len */
1011                                   start, /* orig_start */
1012                                   ins.objectid, /* block_start */
1013                                   ins.offset, /* block_len */
1014                                   ins.offset, /* orig_block_len */
1015                                   ram_size, /* ram_bytes */
1016                                   BTRFS_COMPRESS_NONE, /* compress_type */
1017                                   BTRFS_ORDERED_REGULAR /* type */);
1018                 if (IS_ERR(em))
1019                         goto out_reserve;
1020                 free_extent_map(em);
1021
1022                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1023                                                ram_size, cur_alloc_size, 0);
1024                 if (ret)
1025                         goto out_drop_extent_cache;
1026
1027                 if (root->root_key.objectid ==
1028                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1029                         ret = btrfs_reloc_clone_csums(inode, start,
1030                                                       cur_alloc_size);
1031                         /*
1032                          * Only drop cache here, and process as normal.
1033                          *
1034                          * We must not allow extent_clear_unlock_delalloc()
1035                          * at out_unlock label to free meta of this ordered
1036                          * extent, as its meta should be freed by
1037                          * btrfs_finish_ordered_io().
1038                          *
1039                          * So we must continue until @start is increased to
1040                          * skip current ordered extent.
1041                          */
1042                         if (ret)
1043                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1044                                                 start + ram_size - 1, 0);
1045                 }
1046
1047                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1048
1049                 /* we're not doing compressed IO, don't unlock the first
1050                  * page (which the caller expects to stay locked), don't
1051                  * clear any dirty bits and don't set any writeback bits
1052                  *
1053                  * Do set the Private2 bit so we know this page was properly
1054                  * setup for writepage
1055                  */
1056                 page_ops = unlock ? PAGE_UNLOCK : 0;
1057                 page_ops |= PAGE_SET_PRIVATE2;
1058
1059                 extent_clear_unlock_delalloc(inode, start,
1060                                              start + ram_size - 1,
1061                                              delalloc_end, locked_page,
1062                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1063                                              page_ops);
1064                 if (disk_num_bytes < cur_alloc_size)
1065                         disk_num_bytes = 0;
1066                 else
1067                         disk_num_bytes -= cur_alloc_size;
1068                 num_bytes -= cur_alloc_size;
1069                 alloc_hint = ins.objectid + ins.offset;
1070                 start += cur_alloc_size;
1071                 extent_reserved = false;
1072
1073                 /*
1074                  * btrfs_reloc_clone_csums() error, since start is increased
1075                  * extent_clear_unlock_delalloc() at out_unlock label won't
1076                  * free metadata of current ordered extent, we're OK to exit.
1077                  */
1078                 if (ret)
1079                         goto out_unlock;
1080         }
1081 out:
1082         return ret;
1083
1084 out_drop_extent_cache:
1085         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1086 out_reserve:
1087         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1088         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1089 out_unlock:
1090         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1091                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1092         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1093                 PAGE_END_WRITEBACK;
1094         /*
1095          * If we reserved an extent for our delalloc range (or a subrange) and
1096          * failed to create the respective ordered extent, then it means that
1097          * when we reserved the extent we decremented the extent's size from
1098          * the data space_info's bytes_may_use counter and incremented the
1099          * space_info's bytes_reserved counter by the same amount. We must make
1100          * sure extent_clear_unlock_delalloc() does not try to decrement again
1101          * the data space_info's bytes_may_use counter, therefore we do not pass
1102          * it the flag EXTENT_CLEAR_DATA_RESV.
1103          */
1104         if (extent_reserved) {
1105                 extent_clear_unlock_delalloc(inode, start,
1106                                              start + cur_alloc_size,
1107                                              start + cur_alloc_size,
1108                                              locked_page,
1109                                              clear_bits,
1110                                              page_ops);
1111                 start += cur_alloc_size;
1112                 if (start >= end)
1113                         goto out;
1114         }
1115         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1116                                      locked_page,
1117                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1118                                      page_ops);
1119         goto out;
1120 }
1121
1122 /*
1123  * work queue call back to started compression on a file and pages
1124  */
1125 static noinline void async_cow_start(struct btrfs_work *work)
1126 {
1127         struct async_cow *async_cow;
1128         int num_added = 0;
1129         async_cow = container_of(work, struct async_cow, work);
1130
1131         compress_file_range(async_cow->inode, async_cow->locked_page,
1132                             async_cow->start, async_cow->end, async_cow,
1133                             &num_added);
1134         if (num_added == 0) {
1135                 btrfs_add_delayed_iput(async_cow->inode);
1136                 async_cow->inode = NULL;
1137         }
1138 }
1139
1140 /*
1141  * work queue call back to submit previously compressed pages
1142  */
1143 static noinline void async_cow_submit(struct btrfs_work *work)
1144 {
1145         struct btrfs_fs_info *fs_info;
1146         struct async_cow *async_cow;
1147         struct btrfs_root *root;
1148         unsigned long nr_pages;
1149
1150         async_cow = container_of(work, struct async_cow, work);
1151
1152         root = async_cow->root;
1153         fs_info = root->fs_info;
1154         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1155                 PAGE_SHIFT;
1156
1157         /*
1158          * atomic_sub_return implies a barrier for waitqueue_active
1159          */
1160         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1161             5 * SZ_1M &&
1162             waitqueue_active(&fs_info->async_submit_wait))
1163                 wake_up(&fs_info->async_submit_wait);
1164
1165         if (async_cow->inode)
1166                 submit_compressed_extents(async_cow->inode, async_cow);
1167 }
1168
1169 static noinline void async_cow_free(struct btrfs_work *work)
1170 {
1171         struct async_cow *async_cow;
1172         async_cow = container_of(work, struct async_cow, work);
1173         if (async_cow->inode)
1174                 btrfs_add_delayed_iput(async_cow->inode);
1175         kfree(async_cow);
1176 }
1177
1178 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1179                                 u64 start, u64 end, int *page_started,
1180                                 unsigned long *nr_written)
1181 {
1182         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1183         struct async_cow *async_cow;
1184         struct btrfs_root *root = BTRFS_I(inode)->root;
1185         unsigned long nr_pages;
1186         u64 cur_end;
1187
1188         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1189                          1, 0, NULL, GFP_NOFS);
1190         while (start < end) {
1191                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1192                 BUG_ON(!async_cow); /* -ENOMEM */
1193                 async_cow->inode = igrab(inode);
1194                 async_cow->root = root;
1195                 async_cow->locked_page = locked_page;
1196                 async_cow->start = start;
1197
1198                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1199                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1200                         cur_end = end;
1201                 else
1202                         cur_end = min(end, start + SZ_512K - 1);
1203
1204                 async_cow->end = cur_end;
1205                 INIT_LIST_HEAD(&async_cow->extents);
1206
1207                 btrfs_init_work(&async_cow->work,
1208                                 btrfs_delalloc_helper,
1209                                 async_cow_start, async_cow_submit,
1210                                 async_cow_free);
1211
1212                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1213                         PAGE_SHIFT;
1214                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1215
1216                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1217
1218                 while (atomic_read(&fs_info->async_submit_draining) &&
1219                        atomic_read(&fs_info->async_delalloc_pages)) {
1220                         wait_event(fs_info->async_submit_wait,
1221                                    (atomic_read(&fs_info->async_delalloc_pages) ==
1222                                     0));
1223                 }
1224
1225                 *nr_written += nr_pages;
1226                 start = cur_end + 1;
1227         }
1228         *page_started = 1;
1229         return 0;
1230 }
1231
1232 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1233                                         u64 bytenr, u64 num_bytes)
1234 {
1235         int ret;
1236         struct btrfs_ordered_sum *sums;
1237         LIST_HEAD(list);
1238
1239         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1240                                        bytenr + num_bytes - 1, &list, 0);
1241         if (ret == 0 && list_empty(&list))
1242                 return 0;
1243
1244         while (!list_empty(&list)) {
1245                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1246                 list_del(&sums->list);
1247                 kfree(sums);
1248         }
1249         return 1;
1250 }
1251
1252 /*
1253  * when nowcow writeback call back.  This checks for snapshots or COW copies
1254  * of the extents that exist in the file, and COWs the file as required.
1255  *
1256  * If no cow copies or snapshots exist, we write directly to the existing
1257  * blocks on disk
1258  */
1259 static noinline int run_delalloc_nocow(struct inode *inode,
1260                                        struct page *locked_page,
1261                               u64 start, u64 end, int *page_started, int force,
1262                               unsigned long *nr_written)
1263 {
1264         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1265         struct btrfs_root *root = BTRFS_I(inode)->root;
1266         struct extent_buffer *leaf;
1267         struct btrfs_path *path;
1268         struct btrfs_file_extent_item *fi;
1269         struct btrfs_key found_key;
1270         struct extent_map *em;
1271         u64 cow_start;
1272         u64 cur_offset;
1273         u64 extent_end;
1274         u64 extent_offset;
1275         u64 disk_bytenr;
1276         u64 num_bytes;
1277         u64 disk_num_bytes;
1278         u64 ram_bytes;
1279         int extent_type;
1280         int ret, err;
1281         int type;
1282         int nocow;
1283         int check_prev = 1;
1284         bool nolock;
1285         u64 ino = btrfs_ino(BTRFS_I(inode));
1286
1287         path = btrfs_alloc_path();
1288         if (!path) {
1289                 extent_clear_unlock_delalloc(inode, start, end, end,
1290                                              locked_page,
1291                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1292                                              EXTENT_DO_ACCOUNTING |
1293                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1294                                              PAGE_CLEAR_DIRTY |
1295                                              PAGE_SET_WRITEBACK |
1296                                              PAGE_END_WRITEBACK);
1297                 return -ENOMEM;
1298         }
1299
1300         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1301
1302         cow_start = (u64)-1;
1303         cur_offset = start;
1304         while (1) {
1305                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1306                                                cur_offset, 0);
1307                 if (ret < 0)
1308                         goto error;
1309                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1310                         leaf = path->nodes[0];
1311                         btrfs_item_key_to_cpu(leaf, &found_key,
1312                                               path->slots[0] - 1);
1313                         if (found_key.objectid == ino &&
1314                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1315                                 path->slots[0]--;
1316                 }
1317                 check_prev = 0;
1318 next_slot:
1319                 leaf = path->nodes[0];
1320                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1321                         ret = btrfs_next_leaf(root, path);
1322                         if (ret < 0)
1323                                 goto error;
1324                         if (ret > 0)
1325                                 break;
1326                         leaf = path->nodes[0];
1327                 }
1328
1329                 nocow = 0;
1330                 disk_bytenr = 0;
1331                 num_bytes = 0;
1332                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1333
1334                 if (found_key.objectid > ino)
1335                         break;
1336                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1337                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1338                         path->slots[0]++;
1339                         goto next_slot;
1340                 }
1341                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1342                     found_key.offset > end)
1343                         break;
1344
1345                 if (found_key.offset > cur_offset) {
1346                         extent_end = found_key.offset;
1347                         extent_type = 0;
1348                         goto out_check;
1349                 }
1350
1351                 fi = btrfs_item_ptr(leaf, path->slots[0],
1352                                     struct btrfs_file_extent_item);
1353                 extent_type = btrfs_file_extent_type(leaf, fi);
1354
1355                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1356                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1357                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1358                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1359                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1360                         extent_end = found_key.offset +
1361                                 btrfs_file_extent_num_bytes(leaf, fi);
1362                         disk_num_bytes =
1363                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1364                         if (extent_end <= start) {
1365                                 path->slots[0]++;
1366                                 goto next_slot;
1367                         }
1368                         if (disk_bytenr == 0)
1369                                 goto out_check;
1370                         if (btrfs_file_extent_compression(leaf, fi) ||
1371                             btrfs_file_extent_encryption(leaf, fi) ||
1372                             btrfs_file_extent_other_encoding(leaf, fi))
1373                                 goto out_check;
1374                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1375                                 goto out_check;
1376                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1377                                 goto out_check;
1378                         if (btrfs_cross_ref_exist(root, ino,
1379                                                   found_key.offset -
1380                                                   extent_offset, disk_bytenr))
1381                                 goto out_check;
1382                         disk_bytenr += extent_offset;
1383                         disk_bytenr += cur_offset - found_key.offset;
1384                         num_bytes = min(end + 1, extent_end) - cur_offset;
1385                         /*
1386                          * if there are pending snapshots for this root,
1387                          * we fall into common COW way.
1388                          */
1389                         if (!nolock) {
1390                                 err = btrfs_start_write_no_snapshoting(root);
1391                                 if (!err)
1392                                         goto out_check;
1393                         }
1394                         /*
1395                          * force cow if csum exists in the range.
1396                          * this ensure that csum for a given extent are
1397                          * either valid or do not exist.
1398                          */
1399                         if (csum_exist_in_range(fs_info, disk_bytenr,
1400                                                 num_bytes)) {
1401                                 if (!nolock)
1402                                         btrfs_end_write_no_snapshoting(root);
1403                                 goto out_check;
1404                         }
1405                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1406                                 if (!nolock)
1407                                         btrfs_end_write_no_snapshoting(root);
1408                                 goto out_check;
1409                         }
1410                         nocow = 1;
1411                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1412                         extent_end = found_key.offset +
1413                                 btrfs_file_extent_inline_len(leaf,
1414                                                      path->slots[0], fi);
1415                         extent_end = ALIGN(extent_end,
1416                                            fs_info->sectorsize);
1417                 } else {
1418                         BUG_ON(1);
1419                 }
1420 out_check:
1421                 if (extent_end <= start) {
1422                         path->slots[0]++;
1423                         if (!nolock && nocow)
1424                                 btrfs_end_write_no_snapshoting(root);
1425                         if (nocow)
1426                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1427                         goto next_slot;
1428                 }
1429                 if (!nocow) {
1430                         if (cow_start == (u64)-1)
1431                                 cow_start = cur_offset;
1432                         cur_offset = extent_end;
1433                         if (cur_offset > end)
1434                                 break;
1435                         path->slots[0]++;
1436                         goto next_slot;
1437                 }
1438
1439                 btrfs_release_path(path);
1440                 if (cow_start != (u64)-1) {
1441                         ret = cow_file_range(inode, locked_page,
1442                                              cow_start, found_key.offset - 1,
1443                                              end, page_started, nr_written, 1,
1444                                              NULL);
1445                         if (ret) {
1446                                 if (!nolock && nocow)
1447                                         btrfs_end_write_no_snapshoting(root);
1448                                 if (nocow)
1449                                         btrfs_dec_nocow_writers(fs_info,
1450                                                                 disk_bytenr);
1451                                 goto error;
1452                         }
1453                         cow_start = (u64)-1;
1454                 }
1455
1456                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1457                         u64 orig_start = found_key.offset - extent_offset;
1458
1459                         em = create_io_em(inode, cur_offset, num_bytes,
1460                                           orig_start,
1461                                           disk_bytenr, /* block_start */
1462                                           num_bytes, /* block_len */
1463                                           disk_num_bytes, /* orig_block_len */
1464                                           ram_bytes, BTRFS_COMPRESS_NONE,
1465                                           BTRFS_ORDERED_PREALLOC);
1466                         if (IS_ERR(em)) {
1467                                 if (!nolock && nocow)
1468                                         btrfs_end_write_no_snapshoting(root);
1469                                 if (nocow)
1470                                         btrfs_dec_nocow_writers(fs_info,
1471                                                                 disk_bytenr);
1472                                 ret = PTR_ERR(em);
1473                                 goto error;
1474                         }
1475                         free_extent_map(em);
1476                 }
1477
1478                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1479                         type = BTRFS_ORDERED_PREALLOC;
1480                 } else {
1481                         type = BTRFS_ORDERED_NOCOW;
1482                 }
1483
1484                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1485                                                num_bytes, num_bytes, type);
1486                 if (nocow)
1487                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1488                 BUG_ON(ret); /* -ENOMEM */
1489
1490                 if (root->root_key.objectid ==
1491                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1492                         /*
1493                          * Error handled later, as we must prevent
1494                          * extent_clear_unlock_delalloc() in error handler
1495                          * from freeing metadata of created ordered extent.
1496                          */
1497                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1498                                                       num_bytes);
1499
1500                 extent_clear_unlock_delalloc(inode, cur_offset,
1501                                              cur_offset + num_bytes - 1, end,
1502                                              locked_page, EXTENT_LOCKED |
1503                                              EXTENT_DELALLOC |
1504                                              EXTENT_CLEAR_DATA_RESV,
1505                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1506
1507                 if (!nolock && nocow)
1508                         btrfs_end_write_no_snapshoting(root);
1509                 cur_offset = extent_end;
1510
1511                 /*
1512                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1513                  * handler, as metadata for created ordered extent will only
1514                  * be freed by btrfs_finish_ordered_io().
1515                  */
1516                 if (ret)
1517                         goto error;
1518                 if (cur_offset > end)
1519                         break;
1520         }
1521         btrfs_release_path(path);
1522
1523         if (cur_offset <= end && cow_start == (u64)-1) {
1524                 cow_start = cur_offset;
1525                 cur_offset = end;
1526         }
1527
1528         if (cow_start != (u64)-1) {
1529                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1530                                      page_started, nr_written, 1, NULL);
1531                 if (ret)
1532                         goto error;
1533         }
1534
1535 error:
1536         if (ret && cur_offset < end)
1537                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1538                                              locked_page, EXTENT_LOCKED |
1539                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1540                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1541                                              PAGE_CLEAR_DIRTY |
1542                                              PAGE_SET_WRITEBACK |
1543                                              PAGE_END_WRITEBACK);
1544         btrfs_free_path(path);
1545         return ret;
1546 }
1547
1548 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1549 {
1550
1551         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1552             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1553                 return 0;
1554
1555         /*
1556          * @defrag_bytes is a hint value, no spinlock held here,
1557          * if is not zero, it means the file is defragging.
1558          * Force cow if given extent needs to be defragged.
1559          */
1560         if (BTRFS_I(inode)->defrag_bytes &&
1561             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1562                            EXTENT_DEFRAG, 0, NULL))
1563                 return 1;
1564
1565         return 0;
1566 }
1567
1568 /*
1569  * extent_io.c call back to do delayed allocation processing
1570  */
1571 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1572                               u64 start, u64 end, int *page_started,
1573                               unsigned long *nr_written)
1574 {
1575         int ret;
1576         int force_cow = need_force_cow(inode, start, end);
1577
1578         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1579                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1580                                          page_started, 1, nr_written);
1581         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1582                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1583                                          page_started, 0, nr_written);
1584         } else if (!inode_need_compress(inode)) {
1585                 ret = cow_file_range(inode, locked_page, start, end, end,
1586                                       page_started, nr_written, 1, NULL);
1587         } else {
1588                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1589                         &BTRFS_I(inode)->runtime_flags);
1590                 ret = cow_file_range_async(inode, locked_page, start, end,
1591                                            page_started, nr_written);
1592         }
1593         if (ret)
1594                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1595         return ret;
1596 }
1597
1598 static void btrfs_split_extent_hook(struct inode *inode,
1599                                     struct extent_state *orig, u64 split)
1600 {
1601         u64 size;
1602
1603         /* not delalloc, ignore it */
1604         if (!(orig->state & EXTENT_DELALLOC))
1605                 return;
1606
1607         size = orig->end - orig->start + 1;
1608         if (size > BTRFS_MAX_EXTENT_SIZE) {
1609                 u32 num_extents;
1610                 u64 new_size;
1611
1612                 /*
1613                  * See the explanation in btrfs_merge_extent_hook, the same
1614                  * applies here, just in reverse.
1615                  */
1616                 new_size = orig->end - split + 1;
1617                 num_extents = count_max_extents(new_size);
1618                 new_size = split - orig->start;
1619                 num_extents += count_max_extents(new_size);
1620                 if (count_max_extents(size) >= num_extents)
1621                         return;
1622         }
1623
1624         spin_lock(&BTRFS_I(inode)->lock);
1625         BTRFS_I(inode)->outstanding_extents++;
1626         spin_unlock(&BTRFS_I(inode)->lock);
1627 }
1628
1629 /*
1630  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1631  * extents so we can keep track of new extents that are just merged onto old
1632  * extents, such as when we are doing sequential writes, so we can properly
1633  * account for the metadata space we'll need.
1634  */
1635 static void btrfs_merge_extent_hook(struct inode *inode,
1636                                     struct extent_state *new,
1637                                     struct extent_state *other)
1638 {
1639         u64 new_size, old_size;
1640         u32 num_extents;
1641
1642         /* not delalloc, ignore it */
1643         if (!(other->state & EXTENT_DELALLOC))
1644                 return;
1645
1646         if (new->start > other->start)
1647                 new_size = new->end - other->start + 1;
1648         else
1649                 new_size = other->end - new->start + 1;
1650
1651         /* we're not bigger than the max, unreserve the space and go */
1652         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1653                 spin_lock(&BTRFS_I(inode)->lock);
1654                 BTRFS_I(inode)->outstanding_extents--;
1655                 spin_unlock(&BTRFS_I(inode)->lock);
1656                 return;
1657         }
1658
1659         /*
1660          * We have to add up either side to figure out how many extents were
1661          * accounted for before we merged into one big extent.  If the number of
1662          * extents we accounted for is <= the amount we need for the new range
1663          * then we can return, otherwise drop.  Think of it like this
1664          *
1665          * [ 4k][MAX_SIZE]
1666          *
1667          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1668          * need 2 outstanding extents, on one side we have 1 and the other side
1669          * we have 1 so they are == and we can return.  But in this case
1670          *
1671          * [MAX_SIZE+4k][MAX_SIZE+4k]
1672          *
1673          * Each range on their own accounts for 2 extents, but merged together
1674          * they are only 3 extents worth of accounting, so we need to drop in
1675          * this case.
1676          */
1677         old_size = other->end - other->start + 1;
1678         num_extents = count_max_extents(old_size);
1679         old_size = new->end - new->start + 1;
1680         num_extents += count_max_extents(old_size);
1681         if (count_max_extents(new_size) >= num_extents)
1682                 return;
1683
1684         spin_lock(&BTRFS_I(inode)->lock);
1685         BTRFS_I(inode)->outstanding_extents--;
1686         spin_unlock(&BTRFS_I(inode)->lock);
1687 }
1688
1689 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1690                                       struct inode *inode)
1691 {
1692         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1693
1694         spin_lock(&root->delalloc_lock);
1695         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1696                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1697                               &root->delalloc_inodes);
1698                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1699                         &BTRFS_I(inode)->runtime_flags);
1700                 root->nr_delalloc_inodes++;
1701                 if (root->nr_delalloc_inodes == 1) {
1702                         spin_lock(&fs_info->delalloc_root_lock);
1703                         BUG_ON(!list_empty(&root->delalloc_root));
1704                         list_add_tail(&root->delalloc_root,
1705                                       &fs_info->delalloc_roots);
1706                         spin_unlock(&fs_info->delalloc_root_lock);
1707                 }
1708         }
1709         spin_unlock(&root->delalloc_lock);
1710 }
1711
1712 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1713                                      struct btrfs_inode *inode)
1714 {
1715         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1716
1717         spin_lock(&root->delalloc_lock);
1718         if (!list_empty(&inode->delalloc_inodes)) {
1719                 list_del_init(&inode->delalloc_inodes);
1720                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1721                           &inode->runtime_flags);
1722                 root->nr_delalloc_inodes--;
1723                 if (!root->nr_delalloc_inodes) {
1724                         spin_lock(&fs_info->delalloc_root_lock);
1725                         BUG_ON(list_empty(&root->delalloc_root));
1726                         list_del_init(&root->delalloc_root);
1727                         spin_unlock(&fs_info->delalloc_root_lock);
1728                 }
1729         }
1730         spin_unlock(&root->delalloc_lock);
1731 }
1732
1733 /*
1734  * extent_io.c set_bit_hook, used to track delayed allocation
1735  * bytes in this file, and to maintain the list of inodes that
1736  * have pending delalloc work to be done.
1737  */
1738 static void btrfs_set_bit_hook(struct inode *inode,
1739                                struct extent_state *state, unsigned *bits)
1740 {
1741
1742         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1743
1744         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1745                 WARN_ON(1);
1746         /*
1747          * set_bit and clear bit hooks normally require _irqsave/restore
1748          * but in this case, we are only testing for the DELALLOC
1749          * bit, which is only set or cleared with irqs on
1750          */
1751         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1752                 struct btrfs_root *root = BTRFS_I(inode)->root;
1753                 u64 len = state->end + 1 - state->start;
1754                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1755
1756                 if (*bits & EXTENT_FIRST_DELALLOC) {
1757                         *bits &= ~EXTENT_FIRST_DELALLOC;
1758                 } else {
1759                         spin_lock(&BTRFS_I(inode)->lock);
1760                         BTRFS_I(inode)->outstanding_extents++;
1761                         spin_unlock(&BTRFS_I(inode)->lock);
1762                 }
1763
1764                 /* For sanity tests */
1765                 if (btrfs_is_testing(fs_info))
1766                         return;
1767
1768                 __percpu_counter_add(&fs_info->delalloc_bytes, len,
1769                                      fs_info->delalloc_batch);
1770                 spin_lock(&BTRFS_I(inode)->lock);
1771                 BTRFS_I(inode)->delalloc_bytes += len;
1772                 if (*bits & EXTENT_DEFRAG)
1773                         BTRFS_I(inode)->defrag_bytes += len;
1774                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1775                                          &BTRFS_I(inode)->runtime_flags))
1776                         btrfs_add_delalloc_inodes(root, inode);
1777                 spin_unlock(&BTRFS_I(inode)->lock);
1778         }
1779
1780         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1781             (*bits & EXTENT_DELALLOC_NEW)) {
1782                 spin_lock(&BTRFS_I(inode)->lock);
1783                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1784                         state->start;
1785                 spin_unlock(&BTRFS_I(inode)->lock);
1786         }
1787 }
1788
1789 /*
1790  * extent_io.c clear_bit_hook, see set_bit_hook for why
1791  */
1792 static void btrfs_clear_bit_hook(struct btrfs_inode *inode,
1793                                  struct extent_state *state,
1794                                  unsigned *bits)
1795 {
1796         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1797         u64 len = state->end + 1 - state->start;
1798         u32 num_extents = count_max_extents(len);
1799
1800         spin_lock(&inode->lock);
1801         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1802                 inode->defrag_bytes -= len;
1803         spin_unlock(&inode->lock);
1804
1805         /*
1806          * set_bit and clear bit hooks normally require _irqsave/restore
1807          * but in this case, we are only testing for the DELALLOC
1808          * bit, which is only set or cleared with irqs on
1809          */
1810         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1811                 struct btrfs_root *root = inode->root;
1812                 bool do_list = !btrfs_is_free_space_inode(inode);
1813
1814                 if (*bits & EXTENT_FIRST_DELALLOC) {
1815                         *bits &= ~EXTENT_FIRST_DELALLOC;
1816                 } else if (!(*bits & EXTENT_CLEAR_META_RESV)) {
1817                         spin_lock(&inode->lock);
1818                         inode->outstanding_extents -= num_extents;
1819                         spin_unlock(&inode->lock);
1820                 }
1821
1822                 /*
1823                  * We don't reserve metadata space for space cache inodes so we
1824                  * don't need to call dellalloc_release_metadata if there is an
1825                  * error.
1826                  */
1827                 if (*bits & EXTENT_CLEAR_META_RESV &&
1828                     root != fs_info->tree_root)
1829                         btrfs_delalloc_release_metadata(inode, len);
1830
1831                 /* For sanity tests. */
1832                 if (btrfs_is_testing(fs_info))
1833                         return;
1834
1835                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1836                     do_list && !(state->state & EXTENT_NORESERVE) &&
1837                     (*bits & EXTENT_CLEAR_DATA_RESV))
1838                         btrfs_free_reserved_data_space_noquota(
1839                                         &inode->vfs_inode,
1840                                         state->start, len);
1841
1842                 __percpu_counter_add(&fs_info->delalloc_bytes, -len,
1843                                      fs_info->delalloc_batch);
1844                 spin_lock(&inode->lock);
1845                 inode->delalloc_bytes -= len;
1846                 if (do_list && inode->delalloc_bytes == 0 &&
1847                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1848                                         &inode->runtime_flags))
1849                         btrfs_del_delalloc_inode(root, inode);
1850                 spin_unlock(&inode->lock);
1851         }
1852
1853         if ((state->state & EXTENT_DELALLOC_NEW) &&
1854             (*bits & EXTENT_DELALLOC_NEW)) {
1855                 spin_lock(&inode->lock);
1856                 ASSERT(inode->new_delalloc_bytes >= len);
1857                 inode->new_delalloc_bytes -= len;
1858                 spin_unlock(&inode->lock);
1859         }
1860 }
1861
1862 /*
1863  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1864  * we don't create bios that span stripes or chunks
1865  *
1866  * return 1 if page cannot be merged to bio
1867  * return 0 if page can be merged to bio
1868  * return error otherwise
1869  */
1870 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1871                          size_t size, struct bio *bio,
1872                          unsigned long bio_flags)
1873 {
1874         struct inode *inode = page->mapping->host;
1875         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1876         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1877         u64 length = 0;
1878         u64 map_length;
1879         int ret;
1880
1881         if (bio_flags & EXTENT_BIO_COMPRESSED)
1882                 return 0;
1883
1884         length = bio->bi_iter.bi_size;
1885         map_length = length;
1886         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1887                               NULL, 0);
1888         if (ret < 0)
1889                 return ret;
1890         if (map_length < length + size)
1891                 return 1;
1892         return 0;
1893 }
1894
1895 /*
1896  * in order to insert checksums into the metadata in large chunks,
1897  * we wait until bio submission time.   All the pages in the bio are
1898  * checksummed and sums are attached onto the ordered extent record.
1899  *
1900  * At IO completion time the cums attached on the ordered extent record
1901  * are inserted into the btree
1902  */
1903 static blk_status_t __btrfs_submit_bio_start(struct inode *inode,
1904                 struct bio *bio, int mirror_num, unsigned long bio_flags,
1905                 u64 bio_offset)
1906 {
1907         blk_status_t ret = 0;
1908
1909         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1910         BUG_ON(ret); /* -ENOMEM */
1911         return 0;
1912 }
1913
1914 /*
1915  * in order to insert checksums into the metadata in large chunks,
1916  * we wait until bio submission time.   All the pages in the bio are
1917  * checksummed and sums are attached onto the ordered extent record.
1918  *
1919  * At IO completion time the cums attached on the ordered extent record
1920  * are inserted into the btree
1921  */
1922 static blk_status_t __btrfs_submit_bio_done(struct inode *inode,
1923                 struct bio *bio, int mirror_num, unsigned long bio_flags,
1924                 u64 bio_offset)
1925 {
1926         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1927         blk_status_t ret;
1928
1929         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1930         if (ret) {
1931                 bio->bi_status = ret;
1932                 bio_endio(bio);
1933         }
1934         return ret;
1935 }
1936
1937 /*
1938  * extent_io.c submission hook. This does the right thing for csum calculation
1939  * on write, or reading the csums from the tree before a read
1940  */
1941 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
1942                           int mirror_num, unsigned long bio_flags,
1943                           u64 bio_offset)
1944 {
1945         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1946         struct btrfs_root *root = BTRFS_I(inode)->root;
1947         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1948         blk_status_t ret = 0;
1949         int skip_sum;
1950         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1951
1952         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1953
1954         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1955                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1956
1957         if (bio_op(bio) != REQ_OP_WRITE) {
1958                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1959                 if (ret)
1960                         goto out;
1961
1962                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1963                         ret = btrfs_submit_compressed_read(inode, bio,
1964                                                            mirror_num,
1965                                                            bio_flags);
1966                         goto out;
1967                 } else if (!skip_sum) {
1968                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1969                         if (ret)
1970                                 goto out;
1971                 }
1972                 goto mapit;
1973         } else if (async && !skip_sum) {
1974                 /* csum items have already been cloned */
1975                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1976                         goto mapit;
1977                 /* we're doing a write, do the async checksumming */
1978                 ret = btrfs_wq_submit_bio(fs_info, inode, bio, mirror_num,
1979                                           bio_flags, bio_offset,
1980                                           __btrfs_submit_bio_start,
1981                                           __btrfs_submit_bio_done);
1982                 goto out;
1983         } else if (!skip_sum) {
1984                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1985                 if (ret)
1986                         goto out;
1987         }
1988
1989 mapit:
1990         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
1991
1992 out:
1993         if (ret) {
1994                 bio->bi_status = ret;
1995                 bio_endio(bio);
1996         }
1997         return ret;
1998 }
1999
2000 /*
2001  * given a list of ordered sums record them in the inode.  This happens
2002  * at IO completion time based on sums calculated at bio submission time.
2003  */
2004 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2005                              struct inode *inode, struct list_head *list)
2006 {
2007         struct btrfs_ordered_sum *sum;
2008
2009         list_for_each_entry(sum, list, list) {
2010                 trans->adding_csums = 1;
2011                 btrfs_csum_file_blocks(trans,
2012                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2013                 trans->adding_csums = 0;
2014         }
2015         return 0;
2016 }
2017
2018 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2019                               struct extent_state **cached_state, int dedupe)
2020 {
2021         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2022         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2023                                    cached_state);
2024 }
2025
2026 /* see btrfs_writepage_start_hook for details on why this is required */
2027 struct btrfs_writepage_fixup {
2028         struct page *page;
2029         struct btrfs_work work;
2030 };
2031
2032 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2033 {
2034         struct btrfs_writepage_fixup *fixup;
2035         struct btrfs_ordered_extent *ordered;
2036         struct extent_state *cached_state = NULL;
2037         struct page *page;
2038         struct inode *inode;
2039         u64 page_start;
2040         u64 page_end;
2041         int ret;
2042
2043         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2044         page = fixup->page;
2045 again:
2046         lock_page(page);
2047         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2048                 ClearPageChecked(page);
2049                 goto out_page;
2050         }
2051
2052         inode = page->mapping->host;
2053         page_start = page_offset(page);
2054         page_end = page_offset(page) + PAGE_SIZE - 1;
2055
2056         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2057                          &cached_state);
2058
2059         /* already ordered? We're done */
2060         if (PagePrivate2(page))
2061                 goto out;
2062
2063         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2064                                         PAGE_SIZE);
2065         if (ordered) {
2066                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2067                                      page_end, &cached_state, GFP_NOFS);
2068                 unlock_page(page);
2069                 btrfs_start_ordered_extent(inode, ordered, 1);
2070                 btrfs_put_ordered_extent(ordered);
2071                 goto again;
2072         }
2073
2074         ret = btrfs_delalloc_reserve_space(inode, page_start,
2075                                            PAGE_SIZE);
2076         if (ret) {
2077                 mapping_set_error(page->mapping, ret);
2078                 end_extent_writepage(page, ret, page_start, page_end);
2079                 ClearPageChecked(page);
2080                 goto out;
2081          }
2082
2083         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state,
2084                                   0);
2085         ClearPageChecked(page);
2086         set_page_dirty(page);
2087 out:
2088         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2089                              &cached_state, GFP_NOFS);
2090 out_page:
2091         unlock_page(page);
2092         put_page(page);
2093         kfree(fixup);
2094 }
2095
2096 /*
2097  * There are a few paths in the higher layers of the kernel that directly
2098  * set the page dirty bit without asking the filesystem if it is a
2099  * good idea.  This causes problems because we want to make sure COW
2100  * properly happens and the data=ordered rules are followed.
2101  *
2102  * In our case any range that doesn't have the ORDERED bit set
2103  * hasn't been properly setup for IO.  We kick off an async process
2104  * to fix it up.  The async helper will wait for ordered extents, set
2105  * the delalloc bit and make it safe to write the page.
2106  */
2107 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2108 {
2109         struct inode *inode = page->mapping->host;
2110         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2111         struct btrfs_writepage_fixup *fixup;
2112
2113         /* this page is properly in the ordered list */
2114         if (TestClearPagePrivate2(page))
2115                 return 0;
2116
2117         if (PageChecked(page))
2118                 return -EAGAIN;
2119
2120         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2121         if (!fixup)
2122                 return -EAGAIN;
2123
2124         SetPageChecked(page);
2125         get_page(page);
2126         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2127                         btrfs_writepage_fixup_worker, NULL, NULL);
2128         fixup->page = page;
2129         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2130         return -EBUSY;
2131 }
2132
2133 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2134                                        struct inode *inode, u64 file_pos,
2135                                        u64 disk_bytenr, u64 disk_num_bytes,
2136                                        u64 num_bytes, u64 ram_bytes,
2137                                        u8 compression, u8 encryption,
2138                                        u16 other_encoding, int extent_type)
2139 {
2140         struct btrfs_root *root = BTRFS_I(inode)->root;
2141         struct btrfs_file_extent_item *fi;
2142         struct btrfs_path *path;
2143         struct extent_buffer *leaf;
2144         struct btrfs_key ins;
2145         int extent_inserted = 0;
2146         int ret;
2147
2148         path = btrfs_alloc_path();
2149         if (!path)
2150                 return -ENOMEM;
2151
2152         /*
2153          * we may be replacing one extent in the tree with another.
2154          * The new extent is pinned in the extent map, and we don't want
2155          * to drop it from the cache until it is completely in the btree.
2156          *
2157          * So, tell btrfs_drop_extents to leave this extent in the cache.
2158          * the caller is expected to unpin it and allow it to be merged
2159          * with the others.
2160          */
2161         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2162                                    file_pos + num_bytes, NULL, 0,
2163                                    1, sizeof(*fi), &extent_inserted);
2164         if (ret)
2165                 goto out;
2166
2167         if (!extent_inserted) {
2168                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2169                 ins.offset = file_pos;
2170                 ins.type = BTRFS_EXTENT_DATA_KEY;
2171
2172                 path->leave_spinning = 1;
2173                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2174                                               sizeof(*fi));
2175                 if (ret)
2176                         goto out;
2177         }
2178         leaf = path->nodes[0];
2179         fi = btrfs_item_ptr(leaf, path->slots[0],
2180                             struct btrfs_file_extent_item);
2181         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2182         btrfs_set_file_extent_type(leaf, fi, extent_type);
2183         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2184         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2185         btrfs_set_file_extent_offset(leaf, fi, 0);
2186         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2187         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2188         btrfs_set_file_extent_compression(leaf, fi, compression);
2189         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2190         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2191
2192         btrfs_mark_buffer_dirty(leaf);
2193         btrfs_release_path(path);
2194
2195         inode_add_bytes(inode, num_bytes);
2196
2197         ins.objectid = disk_bytenr;
2198         ins.offset = disk_num_bytes;
2199         ins.type = BTRFS_EXTENT_ITEM_KEY;
2200         ret = btrfs_alloc_reserved_file_extent(trans, root->root_key.objectid,
2201                         btrfs_ino(BTRFS_I(inode)), file_pos, ram_bytes, &ins);
2202         /*
2203          * Release the reserved range from inode dirty range map, as it is
2204          * already moved into delayed_ref_head
2205          */
2206         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2207 out:
2208         btrfs_free_path(path);
2209
2210         return ret;
2211 }
2212
2213 /* snapshot-aware defrag */
2214 struct sa_defrag_extent_backref {
2215         struct rb_node node;
2216         struct old_sa_defrag_extent *old;
2217         u64 root_id;
2218         u64 inum;
2219         u64 file_pos;
2220         u64 extent_offset;
2221         u64 num_bytes;
2222         u64 generation;
2223 };
2224
2225 struct old_sa_defrag_extent {
2226         struct list_head list;
2227         struct new_sa_defrag_extent *new;
2228
2229         u64 extent_offset;
2230         u64 bytenr;
2231         u64 offset;
2232         u64 len;
2233         int count;
2234 };
2235
2236 struct new_sa_defrag_extent {
2237         struct rb_root root;
2238         struct list_head head;
2239         struct btrfs_path *path;
2240         struct inode *inode;
2241         u64 file_pos;
2242         u64 len;
2243         u64 bytenr;
2244         u64 disk_len;
2245         u8 compress_type;
2246 };
2247
2248 static int backref_comp(struct sa_defrag_extent_backref *b1,
2249                         struct sa_defrag_extent_backref *b2)
2250 {
2251         if (b1->root_id < b2->root_id)
2252                 return -1;
2253         else if (b1->root_id > b2->root_id)
2254                 return 1;
2255
2256         if (b1->inum < b2->inum)
2257                 return -1;
2258         else if (b1->inum > b2->inum)
2259                 return 1;
2260
2261         if (b1->file_pos < b2->file_pos)
2262                 return -1;
2263         else if (b1->file_pos > b2->file_pos)
2264                 return 1;
2265
2266         /*
2267          * [------------------------------] ===> (a range of space)
2268          *     |<--->|   |<---->| =============> (fs/file tree A)
2269          * |<---------------------------->| ===> (fs/file tree B)
2270          *
2271          * A range of space can refer to two file extents in one tree while
2272          * refer to only one file extent in another tree.
2273          *
2274          * So we may process a disk offset more than one time(two extents in A)
2275          * and locate at the same extent(one extent in B), then insert two same
2276          * backrefs(both refer to the extent in B).
2277          */
2278         return 0;
2279 }
2280
2281 static void backref_insert(struct rb_root *root,
2282                            struct sa_defrag_extent_backref *backref)
2283 {
2284         struct rb_node **p = &root->rb_node;
2285         struct rb_node *parent = NULL;
2286         struct sa_defrag_extent_backref *entry;
2287         int ret;
2288
2289         while (*p) {
2290                 parent = *p;
2291                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2292
2293                 ret = backref_comp(backref, entry);
2294                 if (ret < 0)
2295                         p = &(*p)->rb_left;
2296                 else
2297                         p = &(*p)->rb_right;
2298         }
2299
2300         rb_link_node(&backref->node, parent, p);
2301         rb_insert_color(&backref->node, root);
2302 }
2303
2304 /*
2305  * Note the backref might has changed, and in this case we just return 0.
2306  */
2307 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2308                                        void *ctx)
2309 {
2310         struct btrfs_file_extent_item *extent;
2311         struct old_sa_defrag_extent *old = ctx;
2312         struct new_sa_defrag_extent *new = old->new;
2313         struct btrfs_path *path = new->path;
2314         struct btrfs_key key;
2315         struct btrfs_root *root;
2316         struct sa_defrag_extent_backref *backref;
2317         struct extent_buffer *leaf;
2318         struct inode *inode = new->inode;
2319         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2320         int slot;
2321         int ret;
2322         u64 extent_offset;
2323         u64 num_bytes;
2324
2325         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2326             inum == btrfs_ino(BTRFS_I(inode)))
2327                 return 0;
2328
2329         key.objectid = root_id;
2330         key.type = BTRFS_ROOT_ITEM_KEY;
2331         key.offset = (u64)-1;
2332
2333         root = btrfs_read_fs_root_no_name(fs_info, &key);
2334         if (IS_ERR(root)) {
2335                 if (PTR_ERR(root) == -ENOENT)
2336                         return 0;
2337                 WARN_ON(1);
2338                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2339                          inum, offset, root_id);
2340                 return PTR_ERR(root);
2341         }
2342
2343         key.objectid = inum;
2344         key.type = BTRFS_EXTENT_DATA_KEY;
2345         if (offset > (u64)-1 << 32)
2346                 key.offset = 0;
2347         else
2348                 key.offset = offset;
2349
2350         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2351         if (WARN_ON(ret < 0))
2352                 return ret;
2353         ret = 0;
2354
2355         while (1) {
2356                 cond_resched();
2357
2358                 leaf = path->nodes[0];
2359                 slot = path->slots[0];
2360
2361                 if (slot >= btrfs_header_nritems(leaf)) {
2362                         ret = btrfs_next_leaf(root, path);
2363                         if (ret < 0) {
2364                                 goto out;
2365                         } else if (ret > 0) {
2366                                 ret = 0;
2367                                 goto out;
2368                         }
2369                         continue;
2370                 }
2371
2372                 path->slots[0]++;
2373
2374                 btrfs_item_key_to_cpu(leaf, &key, slot);
2375
2376                 if (key.objectid > inum)
2377                         goto out;
2378
2379                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2380                         continue;
2381
2382                 extent = btrfs_item_ptr(leaf, slot,
2383                                         struct btrfs_file_extent_item);
2384
2385                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2386                         continue;
2387
2388                 /*
2389                  * 'offset' refers to the exact key.offset,
2390                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2391                  * (key.offset - extent_offset).
2392                  */
2393                 if (key.offset != offset)
2394                         continue;
2395
2396                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2397                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2398
2399                 if (extent_offset >= old->extent_offset + old->offset +
2400                     old->len || extent_offset + num_bytes <=
2401                     old->extent_offset + old->offset)
2402                         continue;
2403                 break;
2404         }
2405
2406         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2407         if (!backref) {
2408                 ret = -ENOENT;
2409                 goto out;
2410         }
2411
2412         backref->root_id = root_id;
2413         backref->inum = inum;
2414         backref->file_pos = offset;
2415         backref->num_bytes = num_bytes;
2416         backref->extent_offset = extent_offset;
2417         backref->generation = btrfs_file_extent_generation(leaf, extent);
2418         backref->old = old;
2419         backref_insert(&new->root, backref);
2420         old->count++;
2421 out:
2422         btrfs_release_path(path);
2423         WARN_ON(ret);
2424         return ret;
2425 }
2426
2427 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2428                                    struct new_sa_defrag_extent *new)
2429 {
2430         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2431         struct old_sa_defrag_extent *old, *tmp;
2432         int ret;
2433
2434         new->path = path;
2435
2436         list_for_each_entry_safe(old, tmp, &new->head, list) {
2437                 ret = iterate_inodes_from_logical(old->bytenr +
2438                                                   old->extent_offset, fs_info,
2439                                                   path, record_one_backref,
2440                                                   old);
2441                 if (ret < 0 && ret != -ENOENT)
2442                         return false;
2443
2444                 /* no backref to be processed for this extent */
2445                 if (!old->count) {
2446                         list_del(&old->list);
2447                         kfree(old);
2448                 }
2449         }
2450
2451         if (list_empty(&new->head))
2452                 return false;
2453
2454         return true;
2455 }
2456
2457 static int relink_is_mergable(struct extent_buffer *leaf,
2458                               struct btrfs_file_extent_item *fi,
2459                               struct new_sa_defrag_extent *new)
2460 {
2461         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2462                 return 0;
2463
2464         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2465                 return 0;
2466
2467         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2468                 return 0;
2469
2470         if (btrfs_file_extent_encryption(leaf, fi) ||
2471             btrfs_file_extent_other_encoding(leaf, fi))
2472                 return 0;
2473
2474         return 1;
2475 }
2476
2477 /*
2478  * Note the backref might has changed, and in this case we just return 0.
2479  */
2480 static noinline int relink_extent_backref(struct btrfs_path *path,
2481                                  struct sa_defrag_extent_backref *prev,
2482                                  struct sa_defrag_extent_backref *backref)
2483 {
2484         struct btrfs_file_extent_item *extent;
2485         struct btrfs_file_extent_item *item;
2486         struct btrfs_ordered_extent *ordered;
2487         struct btrfs_trans_handle *trans;
2488         struct btrfs_root *root;
2489         struct btrfs_key key;
2490         struct extent_buffer *leaf;
2491         struct old_sa_defrag_extent *old = backref->old;
2492         struct new_sa_defrag_extent *new = old->new;
2493         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2494         struct inode *inode;
2495         struct extent_state *cached = NULL;
2496         int ret = 0;
2497         u64 start;
2498         u64 len;
2499         u64 lock_start;
2500         u64 lock_end;
2501         bool merge = false;
2502         int index;
2503
2504         if (prev && prev->root_id == backref->root_id &&
2505             prev->inum == backref->inum &&
2506             prev->file_pos + prev->num_bytes == backref->file_pos)
2507                 merge = true;
2508
2509         /* step 1: get root */
2510         key.objectid = backref->root_id;
2511         key.type = BTRFS_ROOT_ITEM_KEY;
2512         key.offset = (u64)-1;
2513
2514         index = srcu_read_lock(&fs_info->subvol_srcu);
2515
2516         root = btrfs_read_fs_root_no_name(fs_info, &key);
2517         if (IS_ERR(root)) {
2518                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2519                 if (PTR_ERR(root) == -ENOENT)
2520                         return 0;
2521                 return PTR_ERR(root);
2522         }
2523
2524         if (btrfs_root_readonly(root)) {
2525                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2526                 return 0;
2527         }
2528
2529         /* step 2: get inode */
2530         key.objectid = backref->inum;
2531         key.type = BTRFS_INODE_ITEM_KEY;
2532         key.offset = 0;
2533
2534         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2535         if (IS_ERR(inode)) {
2536                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2537                 return 0;
2538         }
2539
2540         srcu_read_unlock(&fs_info->subvol_srcu, index);
2541
2542         /* step 3: relink backref */
2543         lock_start = backref->file_pos;
2544         lock_end = backref->file_pos + backref->num_bytes - 1;
2545         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2546                          &cached);
2547
2548         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2549         if (ordered) {
2550                 btrfs_put_ordered_extent(ordered);
2551                 goto out_unlock;
2552         }
2553
2554         trans = btrfs_join_transaction(root);
2555         if (IS_ERR(trans)) {
2556                 ret = PTR_ERR(trans);
2557                 goto out_unlock;
2558         }
2559
2560         key.objectid = backref->inum;
2561         key.type = BTRFS_EXTENT_DATA_KEY;
2562         key.offset = backref->file_pos;
2563
2564         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2565         if (ret < 0) {
2566                 goto out_free_path;
2567         } else if (ret > 0) {
2568                 ret = 0;
2569                 goto out_free_path;
2570         }
2571
2572         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2573                                 struct btrfs_file_extent_item);
2574
2575         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2576             backref->generation)
2577                 goto out_free_path;
2578
2579         btrfs_release_path(path);
2580
2581         start = backref->file_pos;
2582         if (backref->extent_offset < old->extent_offset + old->offset)
2583                 start += old->extent_offset + old->offset -
2584                          backref->extent_offset;
2585
2586         len = min(backref->extent_offset + backref->num_bytes,
2587                   old->extent_offset + old->offset + old->len);
2588         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2589
2590         ret = btrfs_drop_extents(trans, root, inode, start,
2591                                  start + len, 1);
2592         if (ret)
2593                 goto out_free_path;
2594 again:
2595         key.objectid = btrfs_ino(BTRFS_I(inode));
2596         key.type = BTRFS_EXTENT_DATA_KEY;
2597         key.offset = start;
2598
2599         path->leave_spinning = 1;
2600         if (merge) {
2601                 struct btrfs_file_extent_item *fi;
2602                 u64 extent_len;
2603                 struct btrfs_key found_key;
2604
2605                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2606                 if (ret < 0)
2607                         goto out_free_path;
2608
2609                 path->slots[0]--;
2610                 leaf = path->nodes[0];
2611                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2612
2613                 fi = btrfs_item_ptr(leaf, path->slots[0],
2614                                     struct btrfs_file_extent_item);
2615                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2616
2617                 if (extent_len + found_key.offset == start &&
2618                     relink_is_mergable(leaf, fi, new)) {
2619                         btrfs_set_file_extent_num_bytes(leaf, fi,
2620                                                         extent_len + len);
2621                         btrfs_mark_buffer_dirty(leaf);
2622                         inode_add_bytes(inode, len);
2623
2624                         ret = 1;
2625                         goto out_free_path;
2626                 } else {
2627                         merge = false;
2628                         btrfs_release_path(path);
2629                         goto again;
2630                 }
2631         }
2632
2633         ret = btrfs_insert_empty_item(trans, root, path, &key,
2634                                         sizeof(*extent));
2635         if (ret) {
2636                 btrfs_abort_transaction(trans, ret);
2637                 goto out_free_path;
2638         }
2639
2640         leaf = path->nodes[0];
2641         item = btrfs_item_ptr(leaf, path->slots[0],
2642                                 struct btrfs_file_extent_item);
2643         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2644         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2645         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2646         btrfs_set_file_extent_num_bytes(leaf, item, len);
2647         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2648         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2649         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2650         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2651         btrfs_set_file_extent_encryption(leaf, item, 0);
2652         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2653
2654         btrfs_mark_buffer_dirty(leaf);
2655         inode_add_bytes(inode, len);
2656         btrfs_release_path(path);
2657
2658         ret = btrfs_inc_extent_ref(trans, fs_info, new->bytenr,
2659                         new->disk_len, 0,
2660                         backref->root_id, backref->inum,
2661                         new->file_pos); /* start - extent_offset */
2662         if (ret) {
2663                 btrfs_abort_transaction(trans, ret);
2664                 goto out_free_path;
2665         }
2666
2667         ret = 1;
2668 out_free_path:
2669         btrfs_release_path(path);
2670         path->leave_spinning = 0;
2671         btrfs_end_transaction(trans);
2672 out_unlock:
2673         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2674                              &cached, GFP_NOFS);
2675         iput(inode);
2676         return ret;
2677 }
2678
2679 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2680 {
2681         struct old_sa_defrag_extent *old, *tmp;
2682
2683         if (!new)
2684                 return;
2685
2686         list_for_each_entry_safe(old, tmp, &new->head, list) {
2687                 kfree(old);
2688         }
2689         kfree(new);
2690 }
2691
2692 static void relink_file_extents(struct new_sa_defrag_extent *new)
2693 {
2694         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2695         struct btrfs_path *path;
2696         struct sa_defrag_extent_backref *backref;
2697         struct sa_defrag_extent_backref *prev = NULL;
2698         struct inode *inode;
2699         struct btrfs_root *root;
2700         struct rb_node *node;
2701         int ret;
2702
2703         inode = new->inode;
2704         root = BTRFS_I(inode)->root;
2705
2706         path = btrfs_alloc_path();
2707         if (!path)
2708                 return;
2709
2710         if (!record_extent_backrefs(path, new)) {
2711                 btrfs_free_path(path);
2712                 goto out;
2713         }
2714         btrfs_release_path(path);
2715
2716         while (1) {
2717                 node = rb_first(&new->root);
2718                 if (!node)
2719                         break;
2720                 rb_erase(node, &new->root);
2721
2722                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2723
2724                 ret = relink_extent_backref(path, prev, backref);
2725                 WARN_ON(ret < 0);
2726
2727                 kfree(prev);
2728
2729                 if (ret == 1)
2730                         prev = backref;
2731                 else
2732                         prev = NULL;
2733                 cond_resched();
2734         }
2735         kfree(prev);
2736
2737         btrfs_free_path(path);
2738 out:
2739         free_sa_defrag_extent(new);
2740
2741         atomic_dec(&fs_info->defrag_running);
2742         wake_up(&fs_info->transaction_wait);
2743 }
2744
2745 static struct new_sa_defrag_extent *
2746 record_old_file_extents(struct inode *inode,
2747                         struct btrfs_ordered_extent *ordered)
2748 {
2749         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2750         struct btrfs_root *root = BTRFS_I(inode)->root;
2751         struct btrfs_path *path;
2752         struct btrfs_key key;
2753         struct old_sa_defrag_extent *old;
2754         struct new_sa_defrag_extent *new;
2755         int ret;
2756
2757         new = kmalloc(sizeof(*new), GFP_NOFS);
2758         if (!new)
2759                 return NULL;
2760
2761         new->inode = inode;
2762         new->file_pos = ordered->file_offset;
2763         new->len = ordered->len;
2764         new->bytenr = ordered->start;
2765         new->disk_len = ordered->disk_len;
2766         new->compress_type = ordered->compress_type;
2767         new->root = RB_ROOT;
2768         INIT_LIST_HEAD(&new->head);
2769
2770         path = btrfs_alloc_path();
2771         if (!path)
2772                 goto out_kfree;
2773
2774         key.objectid = btrfs_ino(BTRFS_I(inode));
2775         key.type = BTRFS_EXTENT_DATA_KEY;
2776         key.offset = new->file_pos;
2777
2778         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2779         if (ret < 0)
2780                 goto out_free_path;
2781         if (ret > 0 && path->slots[0] > 0)
2782                 path->slots[0]--;
2783
2784         /* find out all the old extents for the file range */
2785         while (1) {
2786                 struct btrfs_file_extent_item *extent;
2787                 struct extent_buffer *l;
2788                 int slot;
2789                 u64 num_bytes;
2790                 u64 offset;
2791                 u64 end;
2792                 u64 disk_bytenr;
2793                 u64 extent_offset;
2794
2795                 l = path->nodes[0];
2796                 slot = path->slots[0];
2797
2798                 if (slot >= btrfs_header_nritems(l)) {
2799                         ret = btrfs_next_leaf(root, path);
2800                         if (ret < 0)
2801                                 goto out_free_path;
2802                         else if (ret > 0)
2803                                 break;
2804                         continue;
2805                 }
2806
2807                 btrfs_item_key_to_cpu(l, &key, slot);
2808
2809                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2810                         break;
2811                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2812                         break;
2813                 if (key.offset >= new->file_pos + new->len)
2814                         break;
2815
2816                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2817
2818                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2819                 if (key.offset + num_bytes < new->file_pos)
2820                         goto next;
2821
2822                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2823                 if (!disk_bytenr)
2824                         goto next;
2825
2826                 extent_offset = btrfs_file_extent_offset(l, extent);
2827
2828                 old = kmalloc(sizeof(*old), GFP_NOFS);
2829                 if (!old)
2830                         goto out_free_path;
2831
2832                 offset = max(new->file_pos, key.offset);
2833                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2834
2835                 old->bytenr = disk_bytenr;
2836                 old->extent_offset = extent_offset;
2837                 old->offset = offset - key.offset;
2838                 old->len = end - offset;
2839                 old->new = new;
2840                 old->count = 0;
2841                 list_add_tail(&old->list, &new->head);
2842 next:
2843                 path->slots[0]++;
2844                 cond_resched();
2845         }
2846
2847         btrfs_free_path(path);
2848         atomic_inc(&fs_info->defrag_running);
2849
2850         return new;
2851
2852 out_free_path:
2853         btrfs_free_path(path);
2854 out_kfree:
2855         free_sa_defrag_extent(new);
2856         return NULL;
2857 }
2858
2859 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2860                                          u64 start, u64 len)
2861 {
2862         struct btrfs_block_group_cache *cache;
2863
2864         cache = btrfs_lookup_block_group(fs_info, start);
2865         ASSERT(cache);
2866
2867         spin_lock(&cache->lock);
2868         cache->delalloc_bytes -= len;
2869         spin_unlock(&cache->lock);
2870
2871         btrfs_put_block_group(cache);
2872 }
2873
2874 /* as ordered data IO finishes, this gets called so we can finish
2875  * an ordered extent if the range of bytes in the file it covers are
2876  * fully written.
2877  */
2878 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2879 {
2880         struct inode *inode = ordered_extent->inode;
2881         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2882         struct btrfs_root *root = BTRFS_I(inode)->root;
2883         struct btrfs_trans_handle *trans = NULL;
2884         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2885         struct extent_state *cached_state = NULL;
2886         struct new_sa_defrag_extent *new = NULL;
2887         int compress_type = 0;
2888         int ret = 0;
2889         u64 logical_len = ordered_extent->len;
2890         bool nolock;
2891         bool truncated = false;
2892         bool range_locked = false;
2893         bool clear_new_delalloc_bytes = false;
2894
2895         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2896             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2897             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2898                 clear_new_delalloc_bytes = true;
2899
2900         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2901
2902         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2903                 ret = -EIO;
2904                 goto out;
2905         }
2906
2907         btrfs_free_io_failure_record(BTRFS_I(inode),
2908                         ordered_extent->file_offset,
2909                         ordered_extent->file_offset +
2910                         ordered_extent->len - 1);
2911
2912         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2913                 truncated = true;
2914                 logical_len = ordered_extent->truncated_len;
2915                 /* Truncated the entire extent, don't bother adding */
2916                 if (!logical_len)
2917                         goto out;
2918         }
2919
2920         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2921                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2922
2923                 /*
2924                  * For mwrite(mmap + memset to write) case, we still reserve
2925                  * space for NOCOW range.
2926                  * As NOCOW won't cause a new delayed ref, just free the space
2927                  */
2928                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2929                                        ordered_extent->len);
2930                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2931                 if (nolock)
2932                         trans = btrfs_join_transaction_nolock(root);
2933                 else
2934                         trans = btrfs_join_transaction(root);
2935                 if (IS_ERR(trans)) {
2936                         ret = PTR_ERR(trans);
2937                         trans = NULL;
2938                         goto out;
2939                 }
2940                 trans->block_rsv = &fs_info->delalloc_block_rsv;
2941                 ret = btrfs_update_inode_fallback(trans, root, inode);
2942                 if (ret) /* -ENOMEM or corruption */
2943                         btrfs_abort_transaction(trans, ret);
2944                 goto out;
2945         }
2946
2947         range_locked = true;
2948         lock_extent_bits(io_tree, ordered_extent->file_offset,
2949                          ordered_extent->file_offset + ordered_extent->len - 1,
2950                          &cached_state);
2951
2952         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2953                         ordered_extent->file_offset + ordered_extent->len - 1,
2954                         EXTENT_DEFRAG, 0, cached_state);
2955         if (ret) {
2956                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2957                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2958                         /* the inode is shared */
2959                         new = record_old_file_extents(inode, ordered_extent);
2960
2961                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2962                         ordered_extent->file_offset + ordered_extent->len - 1,
2963                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2964         }
2965
2966         if (nolock)
2967                 trans = btrfs_join_transaction_nolock(root);
2968         else
2969                 trans = btrfs_join_transaction(root);
2970         if (IS_ERR(trans)) {
2971                 ret = PTR_ERR(trans);
2972                 trans = NULL;
2973                 goto out;
2974         }
2975
2976         trans->block_rsv = &fs_info->delalloc_block_rsv;
2977
2978         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2979                 compress_type = ordered_extent->compress_type;
2980         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2981                 BUG_ON(compress_type);
2982                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
2983                                                 ordered_extent->file_offset,
2984                                                 ordered_extent->file_offset +
2985                                                 logical_len);
2986         } else {
2987                 BUG_ON(root == fs_info->tree_root);
2988                 ret = insert_reserved_file_extent(trans, inode,
2989                                                 ordered_extent->file_offset,
2990                                                 ordered_extent->start,
2991                                                 ordered_extent->disk_len,
2992                                                 logical_len, logical_len,
2993                                                 compress_type, 0, 0,
2994                                                 BTRFS_FILE_EXTENT_REG);
2995                 if (!ret)
2996                         btrfs_release_delalloc_bytes(fs_info,
2997                                                      ordered_extent->start,
2998                                                      ordered_extent->disk_len);
2999         }
3000         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3001                            ordered_extent->file_offset, ordered_extent->len,
3002                            trans->transid);
3003         if (ret < 0) {
3004                 btrfs_abort_transaction(trans, ret);
3005                 goto out;
3006         }
3007
3008         add_pending_csums(trans, inode, &ordered_extent->list);
3009
3010         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3011         ret = btrfs_update_inode_fallback(trans, root, inode);
3012         if (ret) { /* -ENOMEM or corruption */
3013                 btrfs_abort_transaction(trans, ret);
3014                 goto out;
3015         }
3016         ret = 0;
3017 out:
3018         if (range_locked || clear_new_delalloc_bytes) {
3019                 unsigned int clear_bits = 0;
3020
3021                 if (range_locked)
3022                         clear_bits |= EXTENT_LOCKED;
3023                 if (clear_new_delalloc_bytes)
3024                         clear_bits |= EXTENT_DELALLOC_NEW;
3025                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3026                                  ordered_extent->file_offset,
3027                                  ordered_extent->file_offset +
3028                                  ordered_extent->len - 1,
3029                                  clear_bits,
3030                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3031                                  0, &cached_state, GFP_NOFS);
3032         }
3033
3034         if (root != fs_info->tree_root)
3035                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3036                                 ordered_extent->len);
3037         if (trans)
3038                 btrfs_end_transaction(trans);
3039
3040         if (ret || truncated) {
3041                 u64 start, end;
3042
3043                 if (truncated)
3044                         start = ordered_extent->file_offset + logical_len;
3045                 else
3046                         start = ordered_extent->file_offset;
3047                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3048                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3049
3050                 /* Drop the cache for the part of the extent we didn't write. */
3051                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3052
3053                 /*
3054                  * If the ordered extent had an IOERR or something else went
3055                  * wrong we need to return the space for this ordered extent
3056                  * back to the allocator.  We only free the extent in the
3057                  * truncated case if we didn't write out the extent at all.
3058                  */
3059                 if ((ret || !logical_len) &&
3060                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3061                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3062                         btrfs_free_reserved_extent(fs_info,
3063                                                    ordered_extent->start,
3064                                                    ordered_extent->disk_len, 1);
3065         }
3066
3067
3068         /*
3069          * This needs to be done to make sure anybody waiting knows we are done
3070          * updating everything for this ordered extent.
3071          */
3072         btrfs_remove_ordered_extent(inode, ordered_extent);
3073
3074         /* for snapshot-aware defrag */
3075         if (new) {
3076                 if (ret) {
3077                         free_sa_defrag_extent(new);
3078                         atomic_dec(&fs_info->defrag_running);
3079                 } else {
3080                         relink_file_extents(new);
3081                 }
3082         }
3083
3084         /* once for us */
3085         btrfs_put_ordered_extent(ordered_extent);
3086         /* once for the tree */
3087         btrfs_put_ordered_extent(ordered_extent);
3088
3089         return ret;
3090 }
3091
3092 static void finish_ordered_fn(struct btrfs_work *work)
3093 {
3094         struct btrfs_ordered_extent *ordered_extent;
3095         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3096         btrfs_finish_ordered_io(ordered_extent);
3097 }
3098
3099 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3100                                 struct extent_state *state, int uptodate)
3101 {
3102         struct inode *inode = page->mapping->host;
3103         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3104         struct btrfs_ordered_extent *ordered_extent = NULL;
3105         struct btrfs_workqueue *wq;
3106         btrfs_work_func_t func;
3107
3108         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3109
3110         ClearPagePrivate2(page);
3111         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3112                                             end - start + 1, uptodate))
3113                 return;
3114
3115         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3116                 wq = fs_info->endio_freespace_worker;
3117                 func = btrfs_freespace_write_helper;
3118         } else {
3119                 wq = fs_info->endio_write_workers;
3120                 func = btrfs_endio_write_helper;
3121         }
3122
3123         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3124                         NULL);
3125         btrfs_queue_work(wq, &ordered_extent->work);
3126 }
3127
3128 static int __readpage_endio_check(struct inode *inode,
3129                                   struct btrfs_io_bio *io_bio,
3130                                   int icsum, struct page *page,
3131                                   int pgoff, u64 start, size_t len)
3132 {
3133         char *kaddr;
3134         u32 csum_expected;
3135         u32 csum = ~(u32)0;
3136
3137         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3138
3139         kaddr = kmap_atomic(page);
3140         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3141         btrfs_csum_final(csum, (u8 *)&csum);
3142         if (csum != csum_expected)
3143                 goto zeroit;
3144
3145         kunmap_atomic(kaddr);
3146         return 0;
3147 zeroit:
3148         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3149                                     io_bio->mirror_num);
3150         memset(kaddr + pgoff, 1, len);
3151         flush_dcache_page(page);
3152         kunmap_atomic(kaddr);
3153         if (csum_expected == 0)
3154                 return 0;
3155         return -EIO;
3156 }
3157
3158 /*
3159  * when reads are done, we need to check csums to verify the data is correct
3160  * if there's a match, we allow the bio to finish.  If not, the code in
3161  * extent_io.c will try to find good copies for us.
3162  */
3163 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3164                                       u64 phy_offset, struct page *page,
3165                                       u64 start, u64 end, int mirror)
3166 {
3167         size_t offset = start - page_offset(page);
3168         struct inode *inode = page->mapping->host;
3169         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3170         struct btrfs_root *root = BTRFS_I(inode)->root;
3171
3172         if (PageChecked(page)) {
3173                 ClearPageChecked(page);
3174                 return 0;
3175         }
3176
3177         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3178                 return 0;
3179
3180         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3181             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3182                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3183                 return 0;
3184         }
3185
3186         phy_offset >>= inode->i_sb->s_blocksize_bits;
3187         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3188                                       start, (size_t)(end - start + 1));
3189 }
3190
3191 void btrfs_add_delayed_iput(struct inode *inode)
3192 {
3193         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3194         struct btrfs_inode *binode = BTRFS_I(inode);
3195
3196         if (atomic_add_unless(&inode->i_count, -1, 1))
3197                 return;
3198
3199         spin_lock(&fs_info->delayed_iput_lock);
3200         if (binode->delayed_iput_count == 0) {
3201                 ASSERT(list_empty(&binode->delayed_iput));
3202                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3203         } else {
3204                 binode->delayed_iput_count++;
3205         }
3206         spin_unlock(&fs_info->delayed_iput_lock);
3207 }
3208
3209 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3210 {
3211
3212         spin_lock(&fs_info->delayed_iput_lock);
3213         while (!list_empty(&fs_info->delayed_iputs)) {
3214                 struct btrfs_inode *inode;
3215
3216                 inode = list_first_entry(&fs_info->delayed_iputs,
3217                                 struct btrfs_inode, delayed_iput);
3218                 if (inode->delayed_iput_count) {
3219                         inode->delayed_iput_count--;
3220                         list_move_tail(&inode->delayed_iput,
3221                                         &fs_info->delayed_iputs);
3222                 } else {
3223                         list_del_init(&inode->delayed_iput);
3224                 }
3225                 spin_unlock(&fs_info->delayed_iput_lock);
3226                 iput(&inode->vfs_inode);
3227                 spin_lock(&fs_info->delayed_iput_lock);
3228         }
3229         spin_unlock(&fs_info->delayed_iput_lock);
3230 }
3231
3232 /*
3233  * This is called in transaction commit time. If there are no orphan
3234  * files in the subvolume, it removes orphan item and frees block_rsv
3235  * structure.
3236  */
3237 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3238                               struct btrfs_root *root)
3239 {
3240         struct btrfs_fs_info *fs_info = root->fs_info;
3241         struct btrfs_block_rsv *block_rsv;
3242         int ret;
3243
3244         if (atomic_read(&root->orphan_inodes) ||
3245             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3246                 return;
3247
3248         spin_lock(&root->orphan_lock);
3249         if (atomic_read(&root->orphan_inodes)) {
3250                 spin_unlock(&root->orphan_lock);
3251                 return;
3252         }
3253
3254         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3255                 spin_unlock(&root->orphan_lock);
3256                 return;
3257         }
3258
3259         block_rsv = root->orphan_block_rsv;
3260         root->orphan_block_rsv = NULL;
3261         spin_unlock(&root->orphan_lock);
3262
3263         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3264             btrfs_root_refs(&root->root_item) > 0) {
3265                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3266                                             root->root_key.objectid);
3267                 if (ret)
3268                         btrfs_abort_transaction(trans, ret);
3269                 else
3270                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3271                                   &root->state);
3272         }
3273
3274         if (block_rsv) {
3275                 WARN_ON(block_rsv->size > 0);
3276                 btrfs_free_block_rsv(fs_info, block_rsv);
3277         }
3278 }
3279
3280 /*
3281  * This creates an orphan entry for the given inode in case something goes
3282  * wrong in the middle of an unlink/truncate.
3283  *
3284  * NOTE: caller of this function should reserve 5 units of metadata for
3285  *       this function.
3286  */
3287 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3288                 struct btrfs_inode *inode)
3289 {
3290         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3291         struct btrfs_root *root = inode->root;
3292         struct btrfs_block_rsv *block_rsv = NULL;
3293         int reserve = 0;
3294         int insert = 0;
3295         int ret;
3296
3297         if (!root->orphan_block_rsv) {
3298                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3299                                                   BTRFS_BLOCK_RSV_TEMP);
3300                 if (!block_rsv)
3301                         return -ENOMEM;
3302         }
3303
3304         spin_lock(&root->orphan_lock);
3305         if (!root->orphan_block_rsv) {
3306                 root->orphan_block_rsv = block_rsv;
3307         } else if (block_rsv) {
3308                 btrfs_free_block_rsv(fs_info, block_rsv);
3309                 block_rsv = NULL;
3310         }
3311
3312         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3313                               &inode->runtime_flags)) {
3314 #if 0
3315                 /*
3316                  * For proper ENOSPC handling, we should do orphan
3317                  * cleanup when mounting. But this introduces backward
3318                  * compatibility issue.
3319                  */
3320                 if (!xchg(&root->orphan_item_inserted, 1))
3321                         insert = 2;
3322                 else
3323                         insert = 1;
3324 #endif
3325                 insert = 1;
3326                 atomic_inc(&root->orphan_inodes);
3327         }
3328
3329         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3330                               &inode->runtime_flags))
3331                 reserve = 1;
3332         spin_unlock(&root->orphan_lock);
3333
3334         /* grab metadata reservation from transaction handle */
3335         if (reserve) {
3336                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3337                 ASSERT(!ret);
3338                 if (ret) {
3339                         atomic_dec(&root->orphan_inodes);
3340                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3341                                   &inode->runtime_flags);
3342                         if (insert)
3343                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3344                                           &inode->runtime_flags);
3345                         return ret;
3346                 }
3347         }
3348
3349         /* insert an orphan item to track this unlinked/truncated file */
3350         if (insert >= 1) {
3351                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3352                 if (ret) {
3353                         atomic_dec(&root->orphan_inodes);
3354                         if (reserve) {
3355                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3356                                           &inode->runtime_flags);
3357                                 btrfs_orphan_release_metadata(inode);
3358                         }
3359                         if (ret != -EEXIST) {
3360                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3361                                           &inode->runtime_flags);
3362                                 btrfs_abort_transaction(trans, ret);
3363                                 return ret;
3364                         }
3365                 }
3366                 ret = 0;
3367         }
3368
3369         /* insert an orphan item to track subvolume contains orphan files */
3370         if (insert >= 2) {
3371                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3372                                                root->root_key.objectid);
3373                 if (ret && ret != -EEXIST) {
3374                         btrfs_abort_transaction(trans, ret);
3375                         return ret;
3376                 }
3377         }
3378         return 0;
3379 }
3380
3381 /*
3382  * We have done the truncate/delete so we can go ahead and remove the orphan
3383  * item for this particular inode.
3384  */
3385 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3386                             struct btrfs_inode *inode)
3387 {
3388         struct btrfs_root *root = inode->root;
3389         int delete_item = 0;
3390         int release_rsv = 0;
3391         int ret = 0;
3392
3393         spin_lock(&root->orphan_lock);
3394         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3395                                &inode->runtime_flags))
3396                 delete_item = 1;
3397
3398         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3399                                &inode->runtime_flags))
3400                 release_rsv = 1;
3401         spin_unlock(&root->orphan_lock);
3402
3403         if (delete_item) {
3404                 atomic_dec(&root->orphan_inodes);
3405                 if (trans)
3406                         ret = btrfs_del_orphan_item(trans, root,
3407                                                     btrfs_ino(inode));
3408         }
3409
3410         if (release_rsv)
3411                 btrfs_orphan_release_metadata(inode);
3412
3413         return ret;
3414 }
3415
3416 /*
3417  * this cleans up any orphans that may be left on the list from the last use
3418  * of this root.
3419  */
3420 int btrfs_orphan_cleanup(struct btrfs_root *root)
3421 {
3422         struct btrfs_fs_info *fs_info = root->fs_info;
3423         struct btrfs_path *path;
3424         struct extent_buffer *leaf;
3425         struct btrfs_key key, found_key;
3426         struct btrfs_trans_handle *trans;
3427         struct inode *inode;
3428         u64 last_objectid = 0;
3429         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3430
3431         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3432                 return 0;
3433
3434         path = btrfs_alloc_path();
3435         if (!path) {
3436                 ret = -ENOMEM;
3437                 goto out;
3438         }
3439         path->reada = READA_BACK;
3440
3441         key.objectid = BTRFS_ORPHAN_OBJECTID;
3442         key.type = BTRFS_ORPHAN_ITEM_KEY;
3443         key.offset = (u64)-1;
3444
3445         while (1) {
3446                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3447                 if (ret < 0)
3448                         goto out;
3449
3450                 /*
3451                  * if ret == 0 means we found what we were searching for, which
3452                  * is weird, but possible, so only screw with path if we didn't
3453                  * find the key and see if we have stuff that matches
3454                  */
3455                 if (ret > 0) {
3456                         ret = 0;
3457                         if (path->slots[0] == 0)
3458                                 break;
3459                         path->slots[0]--;
3460                 }
3461
3462                 /* pull out the item */
3463                 leaf = path->nodes[0];
3464                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3465
3466                 /* make sure the item matches what we want */
3467                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3468                         break;
3469                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3470                         break;
3471
3472                 /* release the path since we're done with it */
3473                 btrfs_release_path(path);
3474
3475                 /*
3476                  * this is where we are basically btrfs_lookup, without the
3477                  * crossing root thing.  we store the inode number in the
3478                  * offset of the orphan item.
3479                  */
3480
3481                 if (found_key.offset == last_objectid) {
3482                         btrfs_err(fs_info,
3483                                   "Error removing orphan entry, stopping orphan cleanup");
3484                         ret = -EINVAL;
3485                         goto out;
3486                 }
3487
3488                 last_objectid = found_key.offset;
3489
3490                 found_key.objectid = found_key.offset;
3491                 found_key.type = BTRFS_INODE_ITEM_KEY;
3492                 found_key.offset = 0;
3493                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3494                 ret = PTR_ERR_OR_ZERO(inode);
3495                 if (ret && ret != -ENOENT)
3496                         goto out;
3497
3498                 if (ret == -ENOENT && root == fs_info->tree_root) {
3499                         struct btrfs_root *dead_root;
3500                         struct btrfs_fs_info *fs_info = root->fs_info;
3501                         int is_dead_root = 0;
3502
3503                         /*
3504                          * this is an orphan in the tree root. Currently these
3505                          * could come from 2 sources:
3506                          *  a) a snapshot deletion in progress
3507                          *  b) a free space cache inode
3508                          * We need to distinguish those two, as the snapshot
3509                          * orphan must not get deleted.
3510                          * find_dead_roots already ran before us, so if this
3511                          * is a snapshot deletion, we should find the root
3512                          * in the dead_roots list
3513                          */
3514                         spin_lock(&fs_info->trans_lock);
3515                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3516                                             root_list) {
3517                                 if (dead_root->root_key.objectid ==
3518                                     found_key.objectid) {
3519                                         is_dead_root = 1;
3520                                         break;
3521                                 }
3522                         }
3523                         spin_unlock(&fs_info->trans_lock);
3524                         if (is_dead_root) {
3525                                 /* prevent this orphan from being found again */
3526                                 key.offset = found_key.objectid - 1;
3527                                 continue;
3528                         }
3529                 }
3530                 /*
3531                  * Inode is already gone but the orphan item is still there,
3532                  * kill the orphan item.
3533                  */
3534                 if (ret == -ENOENT) {
3535                         trans = btrfs_start_transaction(root, 1);
3536                         if (IS_ERR(trans)) {
3537                                 ret = PTR_ERR(trans);
3538                                 goto out;
3539                         }
3540                         btrfs_debug(fs_info, "auto deleting %Lu",
3541                                     found_key.objectid);
3542                         ret = btrfs_del_orphan_item(trans, root,
3543                                                     found_key.objectid);
3544                         btrfs_end_transaction(trans);
3545                         if (ret)
3546                                 goto out;
3547                         continue;
3548                 }
3549
3550                 /*
3551                  * add this inode to the orphan list so btrfs_orphan_del does
3552                  * the proper thing when we hit it
3553                  */
3554                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3555                         &BTRFS_I(inode)->runtime_flags);
3556                 atomic_inc(&root->orphan_inodes);
3557
3558                 /* if we have links, this was a truncate, lets do that */
3559                 if (inode->i_nlink) {
3560                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3561                                 iput(inode);
3562                                 continue;
3563                         }
3564                         nr_truncate++;
3565
3566                         /* 1 for the orphan item deletion. */
3567                         trans = btrfs_start_transaction(root, 1);
3568                         if (IS_ERR(trans)) {
3569                                 iput(inode);
3570                                 ret = PTR_ERR(trans);
3571                                 goto out;
3572                         }
3573                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3574                         btrfs_end_transaction(trans);
3575                         if (ret) {
3576                                 iput(inode);
3577                                 goto out;
3578                         }
3579
3580                         ret = btrfs_truncate(inode);
3581                         if (ret)
3582                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3583                 } else {
3584                         nr_unlink++;
3585                 }
3586
3587                 /* this will do delete_inode and everything for us */
3588                 iput(inode);
3589                 if (ret)
3590                         goto out;
3591         }
3592         /* release the path since we're done with it */
3593         btrfs_release_path(path);
3594
3595         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3596
3597         if (root->orphan_block_rsv)
3598                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3599                                         (u64)-1);
3600
3601         if (root->orphan_block_rsv ||
3602             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3603                 trans = btrfs_join_transaction(root);
3604                 if (!IS_ERR(trans))
3605                         btrfs_end_transaction(trans);
3606         }
3607
3608         if (nr_unlink)
3609                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3610         if (nr_truncate)
3611                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3612
3613 out:
3614         if (ret)
3615                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3616         btrfs_free_path(path);
3617         return ret;
3618 }
3619
3620 /*
3621  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3622  * don't find any xattrs, we know there can't be any acls.
3623  *
3624  * slot is the slot the inode is in, objectid is the objectid of the inode
3625  */
3626 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3627                                           int slot, u64 objectid,
3628                                           int *first_xattr_slot)
3629 {
3630         u32 nritems = btrfs_header_nritems(leaf);
3631         struct btrfs_key found_key;
3632         static u64 xattr_access = 0;
3633         static u64 xattr_default = 0;
3634         int scanned = 0;
3635
3636         if (!xattr_access) {
3637                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3638                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3639                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3640                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3641         }
3642
3643         slot++;
3644         *first_xattr_slot = -1;
3645         while (slot < nritems) {
3646                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3647
3648                 /* we found a different objectid, there must not be acls */
3649                 if (found_key.objectid != objectid)
3650                         return 0;
3651
3652                 /* we found an xattr, assume we've got an acl */
3653                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3654                         if (*first_xattr_slot == -1)
3655                                 *first_xattr_slot = slot;
3656                         if (found_key.offset == xattr_access ||
3657                             found_key.offset == xattr_default)
3658                                 return 1;
3659                 }
3660
3661                 /*
3662                  * we found a key greater than an xattr key, there can't
3663                  * be any acls later on
3664                  */
3665                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3666                         return 0;
3667
3668                 slot++;
3669                 scanned++;
3670
3671                 /*
3672                  * it goes inode, inode backrefs, xattrs, extents,
3673                  * so if there are a ton of hard links to an inode there can
3674                  * be a lot of backrefs.  Don't waste time searching too hard,
3675                  * this is just an optimization
3676                  */
3677                 if (scanned >= 8)
3678                         break;
3679         }
3680         /* we hit the end of the leaf before we found an xattr or
3681          * something larger than an xattr.  We have to assume the inode
3682          * has acls
3683          */
3684         if (*first_xattr_slot == -1)
3685                 *first_xattr_slot = slot;
3686         return 1;
3687 }
3688
3689 /*
3690  * read an inode from the btree into the in-memory inode
3691  */
3692 static int btrfs_read_locked_inode(struct inode *inode)
3693 {
3694         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3695         struct btrfs_path *path;
3696         struct extent_buffer *leaf;
3697         struct btrfs_inode_item *inode_item;
3698         struct btrfs_root *root = BTRFS_I(inode)->root;
3699         struct btrfs_key location;
3700         unsigned long ptr;
3701         int maybe_acls;
3702         u32 rdev;
3703         int ret;
3704         bool filled = false;
3705         int first_xattr_slot;
3706
3707         ret = btrfs_fill_inode(inode, &rdev);
3708         if (!ret)
3709                 filled = true;
3710
3711         path = btrfs_alloc_path();
3712         if (!path) {
3713                 ret = -ENOMEM;
3714                 goto make_bad;
3715         }
3716
3717         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3718
3719         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3720         if (ret) {
3721                 if (ret > 0)
3722                         ret = -ENOENT;
3723                 goto make_bad;
3724         }
3725
3726         leaf = path->nodes[0];
3727
3728         if (filled)
3729                 goto cache_index;
3730
3731         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3732                                     struct btrfs_inode_item);
3733         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3734         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3735         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3736         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3737         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3738
3739         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3740         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3741
3742         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3743         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3744
3745         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3746         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3747
3748         BTRFS_I(inode)->i_otime.tv_sec =
3749                 btrfs_timespec_sec(leaf, &inode_item->otime);
3750         BTRFS_I(inode)->i_otime.tv_nsec =
3751                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3752
3753         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3754         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3755         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3756
3757         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3758         inode->i_generation = BTRFS_I(inode)->generation;
3759         inode->i_rdev = 0;
3760         rdev = btrfs_inode_rdev(leaf, inode_item);
3761
3762         BTRFS_I(inode)->index_cnt = (u64)-1;
3763         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3764
3765 cache_index:
3766         /*
3767          * If we were modified in the current generation and evicted from memory
3768          * and then re-read we need to do a full sync since we don't have any
3769          * idea about which extents were modified before we were evicted from
3770          * cache.
3771          *
3772          * This is required for both inode re-read from disk and delayed inode
3773          * in delayed_nodes_tree.
3774          */
3775         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3776                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3777                         &BTRFS_I(inode)->runtime_flags);
3778
3779         /*
3780          * We don't persist the id of the transaction where an unlink operation
3781          * against the inode was last made. So here we assume the inode might
3782          * have been evicted, and therefore the exact value of last_unlink_trans
3783          * lost, and set it to last_trans to avoid metadata inconsistencies
3784          * between the inode and its parent if the inode is fsync'ed and the log
3785          * replayed. For example, in the scenario:
3786          *
3787          * touch mydir/foo
3788          * ln mydir/foo mydir/bar
3789          * sync
3790          * unlink mydir/bar
3791          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3792          * xfs_io -c fsync mydir/foo
3793          * <power failure>
3794          * mount fs, triggers fsync log replay
3795          *
3796          * We must make sure that when we fsync our inode foo we also log its
3797          * parent inode, otherwise after log replay the parent still has the
3798          * dentry with the "bar" name but our inode foo has a link count of 1
3799          * and doesn't have an inode ref with the name "bar" anymore.
3800          *
3801          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3802          * but it guarantees correctness at the expense of occasional full
3803          * transaction commits on fsync if our inode is a directory, or if our
3804          * inode is not a directory, logging its parent unnecessarily.
3805          */
3806         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3807
3808         path->slots[0]++;
3809         if (inode->i_nlink != 1 ||
3810             path->slots[0] >= btrfs_header_nritems(leaf))
3811                 goto cache_acl;
3812
3813         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3814         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3815                 goto cache_acl;
3816
3817         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3818         if (location.type == BTRFS_INODE_REF_KEY) {
3819                 struct btrfs_inode_ref *ref;
3820
3821                 ref = (struct btrfs_inode_ref *)ptr;
3822                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3823         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3824                 struct btrfs_inode_extref *extref;
3825
3826                 extref = (struct btrfs_inode_extref *)ptr;
3827                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3828                                                                      extref);
3829         }
3830 cache_acl:
3831         /*
3832          * try to precache a NULL acl entry for files that don't have
3833          * any xattrs or acls
3834          */
3835         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3836                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3837         if (first_xattr_slot != -1) {
3838                 path->slots[0] = first_xattr_slot;
3839                 ret = btrfs_load_inode_props(inode, path);
3840                 if (ret)
3841                         btrfs_err(fs_info,
3842                                   "error loading props for ino %llu (root %llu): %d",
3843                                   btrfs_ino(BTRFS_I(inode)),
3844                                   root->root_key.objectid, ret);
3845         }
3846         btrfs_free_path(path);
3847
3848         if (!maybe_acls)
3849                 cache_no_acl(inode);
3850
3851         switch (inode->i_mode & S_IFMT) {
3852         case S_IFREG:
3853                 inode->i_mapping->a_ops = &btrfs_aops;
3854                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3855                 inode->i_fop = &btrfs_file_operations;
3856                 inode->i_op = &btrfs_file_inode_operations;
3857                 break;
3858         case S_IFDIR:
3859                 inode->i_fop = &btrfs_dir_file_operations;
3860                 inode->i_op = &btrfs_dir_inode_operations;
3861                 break;
3862         case S_IFLNK:
3863                 inode->i_op = &btrfs_symlink_inode_operations;
3864                 inode_nohighmem(inode);
3865                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3866                 break;
3867         default:
3868                 inode->i_op = &btrfs_special_inode_operations;
3869                 init_special_inode(inode, inode->i_mode, rdev);
3870                 break;
3871         }
3872
3873         btrfs_update_iflags(inode);
3874         return 0;
3875
3876 make_bad:
3877         btrfs_free_path(path);
3878         make_bad_inode(inode);
3879         return ret;
3880 }
3881
3882 /*
3883  * given a leaf and an inode, copy the inode fields into the leaf
3884  */
3885 static void fill_inode_item(struct btrfs_trans_handle *trans,
3886                             struct extent_buffer *leaf,
3887                             struct btrfs_inode_item *item,
3888                             struct inode *inode)
3889 {
3890         struct btrfs_map_token token;
3891
3892         btrfs_init_map_token(&token);
3893
3894         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3895         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3896         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3897                                    &token);
3898         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3899         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3900
3901         btrfs_set_token_timespec_sec(leaf, &item->atime,
3902                                      inode->i_atime.tv_sec, &token);
3903         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3904                                       inode->i_atime.tv_nsec, &token);
3905
3906         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3907                                      inode->i_mtime.tv_sec, &token);
3908         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3909                                       inode->i_mtime.tv_nsec, &token);
3910
3911         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3912                                      inode->i_ctime.tv_sec, &token);
3913         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3914                                       inode->i_ctime.tv_nsec, &token);
3915
3916         btrfs_set_token_timespec_sec(leaf, &item->otime,
3917                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3918         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3919                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3920
3921         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3922                                      &token);
3923         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3924                                          &token);
3925         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3926         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3927         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3928         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3929         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3930 }
3931
3932 /*
3933  * copy everything in the in-memory inode into the btree.
3934  */
3935 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3936                                 struct btrfs_root *root, struct inode *inode)
3937 {
3938         struct btrfs_inode_item *inode_item;
3939         struct btrfs_path *path;
3940         struct extent_buffer *leaf;
3941         int ret;
3942
3943         path = btrfs_alloc_path();
3944         if (!path)
3945                 return -ENOMEM;
3946
3947         path->leave_spinning = 1;
3948         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3949                                  1);
3950         if (ret) {
3951                 if (ret > 0)
3952                         ret = -ENOENT;
3953                 goto failed;
3954         }
3955
3956         leaf = path->nodes[0];
3957         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3958                                     struct btrfs_inode_item);
3959
3960         fill_inode_item(trans, leaf, inode_item, inode);
3961         btrfs_mark_buffer_dirty(leaf);
3962         btrfs_set_inode_last_trans(trans, inode);
3963         ret = 0;
3964 failed:
3965         btrfs_free_path(path);
3966         return ret;
3967 }
3968
3969 /*
3970  * copy everything in the in-memory inode into the btree.
3971  */
3972 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3973                                 struct btrfs_root *root, struct inode *inode)
3974 {
3975         struct btrfs_fs_info *fs_info = root->fs_info;
3976         int ret;
3977
3978         /*
3979          * If the inode is a free space inode, we can deadlock during commit
3980          * if we put it into the delayed code.
3981          *
3982          * The data relocation inode should also be directly updated
3983          * without delay
3984          */
3985         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3986             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3987             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3988                 btrfs_update_root_times(trans, root);
3989
3990                 ret = btrfs_delayed_update_inode(trans, root, inode);
3991                 if (!ret)
3992                         btrfs_set_inode_last_trans(trans, inode);
3993                 return ret;
3994         }
3995
3996         return btrfs_update_inode_item(trans, root, inode);
3997 }
3998
3999 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4000                                          struct btrfs_root *root,
4001                                          struct inode *inode)
4002 {
4003         int ret;
4004
4005         ret = btrfs_update_inode(trans, root, inode);
4006         if (ret == -ENOSPC)
4007                 return btrfs_update_inode_item(trans, root, inode);
4008         return ret;
4009 }
4010
4011 /*
4012  * unlink helper that gets used here in inode.c and in the tree logging
4013  * recovery code.  It remove a link in a directory with a given name, and
4014  * also drops the back refs in the inode to the directory
4015  */
4016 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4017                                 struct btrfs_root *root,
4018                                 struct btrfs_inode *dir,
4019                                 struct btrfs_inode *inode,
4020                                 const char *name, int name_len)
4021 {
4022         struct btrfs_fs_info *fs_info = root->fs_info;
4023         struct btrfs_path *path;
4024         int ret = 0;
4025         struct extent_buffer *leaf;
4026         struct btrfs_dir_item *di;
4027         struct btrfs_key key;
4028         u64 index;
4029         u64 ino = btrfs_ino(inode);
4030         u64 dir_ino = btrfs_ino(dir);
4031
4032         path = btrfs_alloc_path();
4033         if (!path) {
4034                 ret = -ENOMEM;
4035                 goto out;
4036         }
4037
4038         path->leave_spinning = 1;
4039         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4040                                     name, name_len, -1);
4041         if (IS_ERR(di)) {
4042                 ret = PTR_ERR(di);
4043                 goto err;
4044         }
4045         if (!di) {
4046                 ret = -ENOENT;
4047                 goto err;
4048         }
4049         leaf = path->nodes[0];
4050         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4051         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4052         if (ret)
4053                 goto err;
4054         btrfs_release_path(path);
4055
4056         /*
4057          * If we don't have dir index, we have to get it by looking up
4058          * the inode ref, since we get the inode ref, remove it directly,
4059          * it is unnecessary to do delayed deletion.
4060          *
4061          * But if we have dir index, needn't search inode ref to get it.
4062          * Since the inode ref is close to the inode item, it is better
4063          * that we delay to delete it, and just do this deletion when
4064          * we update the inode item.
4065          */
4066         if (inode->dir_index) {
4067                 ret = btrfs_delayed_delete_inode_ref(inode);
4068                 if (!ret) {
4069                         index = inode->dir_index;
4070                         goto skip_backref;
4071                 }
4072         }
4073
4074         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4075                                   dir_ino, &index);
4076         if (ret) {
4077                 btrfs_info(fs_info,
4078                         "failed to delete reference to %.*s, inode %llu parent %llu",
4079                         name_len, name, ino, dir_ino);
4080                 btrfs_abort_transaction(trans, ret);
4081                 goto err;
4082         }
4083 skip_backref:
4084         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4085         if (ret) {
4086                 btrfs_abort_transaction(trans, ret);
4087                 goto err;
4088         }
4089
4090         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4091                         dir_ino);
4092         if (ret != 0 && ret != -ENOENT) {
4093                 btrfs_abort_transaction(trans, ret);
4094                 goto err;
4095         }
4096
4097         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4098                         index);
4099         if (ret == -ENOENT)
4100                 ret = 0;
4101         else if (ret)
4102                 btrfs_abort_transaction(trans, ret);
4103 err:
4104         btrfs_free_path(path);
4105         if (ret)
4106                 goto out;
4107
4108         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4109         inode_inc_iversion(&inode->vfs_inode);
4110         inode_inc_iversion(&dir->vfs_inode);
4111         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4112                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4113         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4114 out:
4115         return ret;
4116 }
4117
4118 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4119                        struct btrfs_root *root,
4120                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4121                        const char *name, int name_len)
4122 {
4123         int ret;
4124         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4125         if (!ret) {
4126                 drop_nlink(&inode->vfs_inode);
4127                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4128         }
4129         return ret;
4130 }
4131
4132 /*
4133  * helper to start transaction for unlink and rmdir.
4134  *
4135  * unlink and rmdir are special in btrfs, they do not always free space, so
4136  * if we cannot make our reservations the normal way try and see if there is
4137  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4138  * allow the unlink to occur.
4139  */
4140 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4141 {
4142         struct btrfs_root *root = BTRFS_I(dir)->root;
4143
4144         /*
4145          * 1 for the possible orphan item
4146          * 1 for the dir item
4147          * 1 for the dir index
4148          * 1 for the inode ref
4149          * 1 for the inode
4150          */
4151         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4152 }
4153
4154 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4155 {
4156         struct btrfs_root *root = BTRFS_I(dir)->root;
4157         struct btrfs_trans_handle *trans;
4158         struct inode *inode = d_inode(dentry);
4159         int ret;
4160
4161         trans = __unlink_start_trans(dir);
4162         if (IS_ERR(trans))
4163                 return PTR_ERR(trans);
4164
4165         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4166                         0);
4167
4168         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4169                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4170                         dentry->d_name.len);
4171         if (ret)
4172                 goto out;
4173
4174         if (inode->i_nlink == 0) {
4175                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4176                 if (ret)
4177                         goto out;
4178         }
4179
4180 out:
4181         btrfs_end_transaction(trans);
4182         btrfs_btree_balance_dirty(root->fs_info);
4183         return ret;
4184 }
4185
4186 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4187                         struct btrfs_root *root,
4188                         struct inode *dir, u64 objectid,
4189                         const char *name, int name_len)
4190 {
4191         struct btrfs_fs_info *fs_info = root->fs_info;
4192         struct btrfs_path *path;
4193         struct extent_buffer *leaf;
4194         struct btrfs_dir_item *di;
4195         struct btrfs_key key;
4196         u64 index;
4197         int ret;
4198         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4199
4200         path = btrfs_alloc_path();
4201         if (!path)
4202                 return -ENOMEM;
4203
4204         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4205                                    name, name_len, -1);
4206         if (IS_ERR_OR_NULL(di)) {
4207                 if (!di)
4208                         ret = -ENOENT;
4209                 else
4210                         ret = PTR_ERR(di);
4211                 goto out;
4212         }
4213
4214         leaf = path->nodes[0];
4215         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4216         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4217         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4218         if (ret) {
4219                 btrfs_abort_transaction(trans, ret);
4220                 goto out;
4221         }
4222         btrfs_release_path(path);
4223
4224         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4225                                  root->root_key.objectid, dir_ino,
4226                                  &index, name, name_len);
4227         if (ret < 0) {
4228                 if (ret != -ENOENT) {
4229                         btrfs_abort_transaction(trans, ret);
4230                         goto out;
4231                 }
4232                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4233                                                  name, name_len);
4234                 if (IS_ERR_OR_NULL(di)) {
4235                         if (!di)
4236                                 ret = -ENOENT;
4237                         else
4238                                 ret = PTR_ERR(di);
4239                         btrfs_abort_transaction(trans, ret);
4240                         goto out;
4241                 }
4242
4243                 leaf = path->nodes[0];
4244                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4245                 btrfs_release_path(path);
4246                 index = key.offset;
4247         }
4248         btrfs_release_path(path);
4249
4250         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4251         if (ret) {
4252                 btrfs_abort_transaction(trans, ret);
4253                 goto out;
4254         }
4255
4256         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4257         inode_inc_iversion(dir);
4258         dir->i_mtime = dir->i_ctime = current_time(dir);
4259         ret = btrfs_update_inode_fallback(trans, root, dir);
4260         if (ret)
4261                 btrfs_abort_transaction(trans, ret);
4262 out:
4263         btrfs_free_path(path);
4264         return ret;
4265 }
4266
4267 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4268 {
4269         struct inode *inode = d_inode(dentry);
4270         int err = 0;
4271         struct btrfs_root *root = BTRFS_I(dir)->root;
4272         struct btrfs_trans_handle *trans;
4273         u64 last_unlink_trans;
4274
4275         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4276                 return -ENOTEMPTY;
4277         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4278                 return -EPERM;
4279
4280         trans = __unlink_start_trans(dir);
4281         if (IS_ERR(trans))
4282                 return PTR_ERR(trans);
4283
4284         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4285                 err = btrfs_unlink_subvol(trans, root, dir,
4286                                           BTRFS_I(inode)->location.objectid,
4287                                           dentry->d_name.name,
4288                                           dentry->d_name.len);
4289                 goto out;
4290         }
4291
4292         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4293         if (err)
4294                 goto out;
4295
4296         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4297
4298         /* now the directory is empty */
4299         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4300                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4301                         dentry->d_name.len);
4302         if (!err) {
4303                 btrfs_i_size_write(BTRFS_I(inode), 0);
4304                 /*
4305                  * Propagate the last_unlink_trans value of the deleted dir to
4306                  * its parent directory. This is to prevent an unrecoverable
4307                  * log tree in the case we do something like this:
4308                  * 1) create dir foo
4309                  * 2) create snapshot under dir foo
4310                  * 3) delete the snapshot
4311                  * 4) rmdir foo
4312                  * 5) mkdir foo
4313                  * 6) fsync foo or some file inside foo
4314                  */
4315                 if (last_unlink_trans >= trans->transid)
4316                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4317         }
4318 out:
4319         btrfs_end_transaction(trans);
4320         btrfs_btree_balance_dirty(root->fs_info);
4321
4322         return err;
4323 }
4324
4325 static int truncate_space_check(struct btrfs_trans_handle *trans,
4326                                 struct btrfs_root *root,
4327                                 u64 bytes_deleted)
4328 {
4329         struct btrfs_fs_info *fs_info = root->fs_info;
4330         int ret;
4331
4332         /*
4333          * This is only used to apply pressure to the enospc system, we don't
4334          * intend to use this reservation at all.
4335          */
4336         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4337         bytes_deleted *= fs_info->nodesize;
4338         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4339                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4340         if (!ret) {
4341                 trace_btrfs_space_reservation(fs_info, "transaction",
4342                                               trans->transid,
4343                                               bytes_deleted, 1);
4344                 trans->bytes_reserved += bytes_deleted;
4345         }
4346         return ret;
4347
4348 }
4349
4350 static int truncate_inline_extent(struct inode *inode,
4351                                   struct btrfs_path *path,
4352                                   struct btrfs_key *found_key,
4353                                   const u64 item_end,
4354                                   const u64 new_size)
4355 {
4356         struct extent_buffer *leaf = path->nodes[0];
4357         int slot = path->slots[0];
4358         struct btrfs_file_extent_item *fi;
4359         u32 size = (u32)(new_size - found_key->offset);
4360         struct btrfs_root *root = BTRFS_I(inode)->root;
4361
4362         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4363
4364         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4365                 loff_t offset = new_size;
4366                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4367
4368                 /*
4369                  * Zero out the remaining of the last page of our inline extent,
4370                  * instead of directly truncating our inline extent here - that
4371                  * would be much more complex (decompressing all the data, then
4372                  * compressing the truncated data, which might be bigger than
4373                  * the size of the inline extent, resize the extent, etc).
4374                  * We release the path because to get the page we might need to
4375                  * read the extent item from disk (data not in the page cache).
4376                  */
4377                 btrfs_release_path(path);
4378                 return btrfs_truncate_block(inode, offset, page_end - offset,
4379                                         0);
4380         }
4381
4382         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4383         size = btrfs_file_extent_calc_inline_size(size);
4384         btrfs_truncate_item(root->fs_info, path, size, 1);
4385
4386         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4387                 inode_sub_bytes(inode, item_end + 1 - new_size);
4388
4389         return 0;
4390 }
4391
4392 /*
4393  * this can truncate away extent items, csum items and directory items.
4394  * It starts at a high offset and removes keys until it can't find
4395  * any higher than new_size
4396  *
4397  * csum items that cross the new i_size are truncated to the new size
4398  * as well.
4399  *
4400  * min_type is the minimum key type to truncate down to.  If set to 0, this
4401  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4402  */
4403 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4404                                struct btrfs_root *root,
4405                                struct inode *inode,
4406                                u64 new_size, u32 min_type)
4407 {
4408         struct btrfs_fs_info *fs_info = root->fs_info;
4409         struct btrfs_path *path;
4410         struct extent_buffer *leaf;
4411         struct btrfs_file_extent_item *fi;
4412         struct btrfs_key key;
4413         struct btrfs_key found_key;
4414         u64 extent_start = 0;
4415         u64 extent_num_bytes = 0;
4416         u64 extent_offset = 0;
4417         u64 item_end = 0;
4418         u64 last_size = new_size;
4419         u32 found_type = (u8)-1;
4420         int found_extent;
4421         int del_item;
4422         int pending_del_nr = 0;
4423         int pending_del_slot = 0;
4424         int extent_type = -1;
4425         int ret;
4426         int err = 0;
4427         u64 ino = btrfs_ino(BTRFS_I(inode));
4428         u64 bytes_deleted = 0;
4429         bool be_nice = 0;
4430         bool should_throttle = 0;
4431         bool should_end = 0;
4432
4433         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4434
4435         /*
4436          * for non-free space inodes and ref cows, we want to back off from
4437          * time to time
4438          */
4439         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4440             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4441                 be_nice = 1;
4442
4443         path = btrfs_alloc_path();
4444         if (!path)
4445                 return -ENOMEM;
4446         path->reada = READA_BACK;
4447
4448         /*
4449          * We want to drop from the next block forward in case this new size is
4450          * not block aligned since we will be keeping the last block of the
4451          * extent just the way it is.
4452          */
4453         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4454             root == fs_info->tree_root)
4455                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4456                                         fs_info->sectorsize),
4457                                         (u64)-1, 0);
4458
4459         /*
4460          * This function is also used to drop the items in the log tree before
4461          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4462          * it is used to drop the loged items. So we shouldn't kill the delayed
4463          * items.
4464          */
4465         if (min_type == 0 && root == BTRFS_I(inode)->root)
4466                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4467
4468         key.objectid = ino;
4469         key.offset = (u64)-1;
4470         key.type = (u8)-1;
4471
4472 search_again:
4473         /*
4474          * with a 16K leaf size and 128MB extents, you can actually queue
4475          * up a huge file in a single leaf.  Most of the time that
4476          * bytes_deleted is > 0, it will be huge by the time we get here
4477          */
4478         if (be_nice && bytes_deleted > SZ_32M) {
4479                 if (btrfs_should_end_transaction(trans)) {
4480                         err = -EAGAIN;
4481                         goto error;
4482                 }
4483         }
4484
4485
4486         path->leave_spinning = 1;
4487         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4488         if (ret < 0) {
4489                 err = ret;
4490                 goto out;
4491         }
4492
4493         if (ret > 0) {
4494                 /* there are no items in the tree for us to truncate, we're
4495                  * done
4496                  */
4497                 if (path->slots[0] == 0)
4498                         goto out;
4499                 path->slots[0]--;
4500         }
4501
4502         while (1) {
4503                 fi = NULL;
4504                 leaf = path->nodes[0];
4505                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4506                 found_type = found_key.type;
4507
4508                 if (found_key.objectid != ino)
4509                         break;
4510
4511                 if (found_type < min_type)
4512                         break;
4513
4514                 item_end = found_key.offset;
4515                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4516                         fi = btrfs_item_ptr(leaf, path->slots[0],
4517                                             struct btrfs_file_extent_item);
4518                         extent_type = btrfs_file_extent_type(leaf, fi);
4519                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4520                                 item_end +=
4521                                     btrfs_file_extent_num_bytes(leaf, fi);
4522
4523                                 trace_btrfs_truncate_show_fi_regular(
4524                                         BTRFS_I(inode), leaf, fi,
4525                                         found_key.offset);
4526                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4527                                 item_end += btrfs_file_extent_inline_len(leaf,
4528                                                          path->slots[0], fi);
4529
4530                                 trace_btrfs_truncate_show_fi_inline(
4531                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4532                                         found_key.offset);
4533                         }
4534                         item_end--;
4535                 }
4536                 if (found_type > min_type) {
4537                         del_item = 1;
4538                 } else {
4539                         if (item_end < new_size)
4540                                 break;
4541                         if (found_key.offset >= new_size)
4542                                 del_item = 1;
4543                         else
4544                                 del_item = 0;
4545                 }
4546                 found_extent = 0;
4547                 /* FIXME, shrink the extent if the ref count is only 1 */
4548                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4549                         goto delete;
4550
4551                 if (del_item)
4552                         last_size = found_key.offset;
4553                 else
4554                         last_size = new_size;
4555
4556                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4557                         u64 num_dec;
4558                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4559                         if (!del_item) {
4560                                 u64 orig_num_bytes =
4561                                         btrfs_file_extent_num_bytes(leaf, fi);
4562                                 extent_num_bytes = ALIGN(new_size -
4563                                                 found_key.offset,
4564                                                 fs_info->sectorsize);
4565                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4566                                                          extent_num_bytes);
4567                                 num_dec = (orig_num_bytes -
4568                                            extent_num_bytes);
4569                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4570                                              &root->state) &&
4571                                     extent_start != 0)
4572                                         inode_sub_bytes(inode, num_dec);
4573                                 btrfs_mark_buffer_dirty(leaf);
4574                         } else {
4575                                 extent_num_bytes =
4576                                         btrfs_file_extent_disk_num_bytes(leaf,
4577                                                                          fi);
4578                                 extent_offset = found_key.offset -
4579                                         btrfs_file_extent_offset(leaf, fi);
4580
4581                                 /* FIXME blocksize != 4096 */
4582                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4583                                 if (extent_start != 0) {
4584                                         found_extent = 1;
4585                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4586                                                      &root->state))
4587                                                 inode_sub_bytes(inode, num_dec);
4588                                 }
4589                         }
4590                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4591                         /*
4592                          * we can't truncate inline items that have had
4593                          * special encodings
4594                          */
4595                         if (!del_item &&
4596                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4597                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4598
4599                                 /*
4600                                  * Need to release path in order to truncate a
4601                                  * compressed extent. So delete any accumulated
4602                                  * extent items so far.
4603                                  */
4604                                 if (btrfs_file_extent_compression(leaf, fi) !=
4605                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4606                                         err = btrfs_del_items(trans, root, path,
4607                                                               pending_del_slot,
4608                                                               pending_del_nr);
4609                                         if (err) {
4610                                                 btrfs_abort_transaction(trans,
4611                                                                         err);
4612                                                 goto error;
4613                                         }
4614                                         pending_del_nr = 0;
4615                                 }
4616
4617                                 err = truncate_inline_extent(inode, path,
4618                                                              &found_key,
4619                                                              item_end,
4620                                                              new_size);
4621                                 if (err) {
4622                                         btrfs_abort_transaction(trans, err);
4623                                         goto error;
4624                                 }
4625                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4626                                             &root->state)) {
4627                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4628                         }
4629                 }
4630 delete:
4631                 if (del_item) {
4632                         if (!pending_del_nr) {
4633                                 /* no pending yet, add ourselves */
4634                                 pending_del_slot = path->slots[0];
4635                                 pending_del_nr = 1;
4636                         } else if (pending_del_nr &&
4637                                    path->slots[0] + 1 == pending_del_slot) {
4638                                 /* hop on the pending chunk */
4639                                 pending_del_nr++;
4640                                 pending_del_slot = path->slots[0];
4641                         } else {
4642                                 BUG();
4643                         }
4644                 } else {
4645                         break;
4646                 }
4647                 should_throttle = 0;
4648
4649                 if (found_extent &&
4650                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4651                      root == fs_info->tree_root)) {
4652                         btrfs_set_path_blocking(path);
4653                         bytes_deleted += extent_num_bytes;
4654                         ret = btrfs_free_extent(trans, fs_info, extent_start,
4655                                                 extent_num_bytes, 0,
4656                                                 btrfs_header_owner(leaf),
4657                                                 ino, extent_offset);
4658                         BUG_ON(ret);
4659                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4660                                 btrfs_async_run_delayed_refs(fs_info,
4661                                         trans->delayed_ref_updates * 2,
4662                                         trans->transid, 0);
4663                         if (be_nice) {
4664                                 if (truncate_space_check(trans, root,
4665                                                          extent_num_bytes)) {
4666                                         should_end = 1;
4667                                 }
4668                                 if (btrfs_should_throttle_delayed_refs(trans,
4669                                                                        fs_info))
4670                                         should_throttle = 1;
4671                         }
4672                 }
4673
4674                 if (found_type == BTRFS_INODE_ITEM_KEY)
4675                         break;
4676
4677                 if (path->slots[0] == 0 ||
4678                     path->slots[0] != pending_del_slot ||
4679                     should_throttle || should_end) {
4680                         if (pending_del_nr) {
4681                                 ret = btrfs_del_items(trans, root, path,
4682                                                 pending_del_slot,
4683                                                 pending_del_nr);
4684                                 if (ret) {
4685                                         btrfs_abort_transaction(trans, ret);
4686                                         goto error;
4687                                 }
4688                                 pending_del_nr = 0;
4689                         }
4690                         btrfs_release_path(path);
4691                         if (should_throttle) {
4692                                 unsigned long updates = trans->delayed_ref_updates;
4693                                 if (updates) {
4694                                         trans->delayed_ref_updates = 0;
4695                                         ret = btrfs_run_delayed_refs(trans,
4696                                                                    fs_info,
4697                                                                    updates * 2);
4698                                         if (ret && !err)
4699                                                 err = ret;
4700                                 }
4701                         }
4702                         /*
4703                          * if we failed to refill our space rsv, bail out
4704                          * and let the transaction restart
4705                          */
4706                         if (should_end) {
4707                                 err = -EAGAIN;
4708                                 goto error;
4709                         }
4710                         goto search_again;
4711                 } else {
4712                         path->slots[0]--;
4713                 }
4714         }
4715 out:
4716         if (pending_del_nr) {
4717                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4718                                       pending_del_nr);
4719                 if (ret)
4720                         btrfs_abort_transaction(trans, ret);
4721         }
4722 error:
4723         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4724                 ASSERT(last_size >= new_size);
4725                 if (!err && last_size > new_size)
4726                         last_size = new_size;
4727                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4728         }
4729
4730         btrfs_free_path(path);
4731
4732         if (be_nice && bytes_deleted > SZ_32M) {
4733                 unsigned long updates = trans->delayed_ref_updates;
4734                 if (updates) {
4735                         trans->delayed_ref_updates = 0;
4736                         ret = btrfs_run_delayed_refs(trans, fs_info,
4737                                                      updates * 2);
4738                         if (ret && !err)
4739                                 err = ret;
4740                 }
4741         }
4742         return err;
4743 }
4744
4745 /*
4746  * btrfs_truncate_block - read, zero a chunk and write a block
4747  * @inode - inode that we're zeroing
4748  * @from - the offset to start zeroing
4749  * @len - the length to zero, 0 to zero the entire range respective to the
4750  *      offset
4751  * @front - zero up to the offset instead of from the offset on
4752  *
4753  * This will find the block for the "from" offset and cow the block and zero the
4754  * part we want to zero.  This is used with truncate and hole punching.
4755  */
4756 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4757                         int front)
4758 {
4759         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4760         struct address_space *mapping = inode->i_mapping;
4761         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4762         struct btrfs_ordered_extent *ordered;
4763         struct extent_state *cached_state = NULL;
4764         char *kaddr;
4765         u32 blocksize = fs_info->sectorsize;
4766         pgoff_t index = from >> PAGE_SHIFT;
4767         unsigned offset = from & (blocksize - 1);
4768         struct page *page;
4769         gfp_t mask = btrfs_alloc_write_mask(mapping);
4770         int ret = 0;
4771         u64 block_start;
4772         u64 block_end;
4773
4774         if ((offset & (blocksize - 1)) == 0 &&
4775             (!len || ((len & (blocksize - 1)) == 0)))
4776                 goto out;
4777
4778         ret = btrfs_delalloc_reserve_space(inode,
4779                         round_down(from, blocksize), blocksize);
4780         if (ret)
4781                 goto out;
4782
4783 again:
4784         page = find_or_create_page(mapping, index, mask);
4785         if (!page) {
4786                 btrfs_delalloc_release_space(inode,
4787                                 round_down(from, blocksize),
4788                                 blocksize);
4789                 ret = -ENOMEM;
4790                 goto out;
4791         }
4792
4793         block_start = round_down(from, blocksize);
4794         block_end = block_start + blocksize - 1;
4795
4796         if (!PageUptodate(page)) {
4797                 ret = btrfs_readpage(NULL, page);
4798                 lock_page(page);
4799                 if (page->mapping != mapping) {
4800                         unlock_page(page);
4801                         put_page(page);
4802                         goto again;
4803                 }
4804                 if (!PageUptodate(page)) {
4805                         ret = -EIO;
4806                         goto out_unlock;
4807                 }
4808         }
4809         wait_on_page_writeback(page);
4810
4811         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4812         set_page_extent_mapped(page);
4813
4814         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4815         if (ordered) {
4816                 unlock_extent_cached(io_tree, block_start, block_end,
4817                                      &cached_state, GFP_NOFS);
4818                 unlock_page(page);
4819                 put_page(page);
4820                 btrfs_start_ordered_extent(inode, ordered, 1);
4821                 btrfs_put_ordered_extent(ordered);
4822                 goto again;
4823         }
4824
4825         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4826                           EXTENT_DIRTY | EXTENT_DELALLOC |
4827                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4828                           0, 0, &cached_state, GFP_NOFS);
4829
4830         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4831                                         &cached_state, 0);
4832         if (ret) {
4833                 unlock_extent_cached(io_tree, block_start, block_end,
4834                                      &cached_state, GFP_NOFS);
4835                 goto out_unlock;
4836         }
4837
4838         if (offset != blocksize) {
4839                 if (!len)
4840                         len = blocksize - offset;
4841                 kaddr = kmap(page);
4842                 if (front)
4843                         memset(kaddr + (block_start - page_offset(page)),
4844                                 0, offset);
4845                 else
4846                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4847                                 0, len);
4848                 flush_dcache_page(page);
4849                 kunmap(page);
4850         }
4851         ClearPageChecked(page);
4852         set_page_dirty(page);
4853         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4854                              GFP_NOFS);
4855
4856 out_unlock:
4857         if (ret)
4858                 btrfs_delalloc_release_space(inode, block_start,
4859                                              blocksize);
4860         unlock_page(page);
4861         put_page(page);
4862 out:
4863         return ret;
4864 }
4865
4866 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4867                              u64 offset, u64 len)
4868 {
4869         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4870         struct btrfs_trans_handle *trans;
4871         int ret;
4872
4873         /*
4874          * Still need to make sure the inode looks like it's been updated so
4875          * that any holes get logged if we fsync.
4876          */
4877         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4878                 BTRFS_I(inode)->last_trans = fs_info->generation;
4879                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4880                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4881                 return 0;
4882         }
4883
4884         /*
4885          * 1 - for the one we're dropping
4886          * 1 - for the one we're adding
4887          * 1 - for updating the inode.
4888          */
4889         trans = btrfs_start_transaction(root, 3);
4890         if (IS_ERR(trans))
4891                 return PTR_ERR(trans);
4892
4893         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4894         if (ret) {
4895                 btrfs_abort_transaction(trans, ret);
4896                 btrfs_end_transaction(trans);
4897                 return ret;
4898         }
4899
4900         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4901                         offset, 0, 0, len, 0, len, 0, 0, 0);
4902         if (ret)
4903                 btrfs_abort_transaction(trans, ret);
4904         else
4905                 btrfs_update_inode(trans, root, inode);
4906         btrfs_end_transaction(trans);
4907         return ret;
4908 }
4909
4910 /*
4911  * This function puts in dummy file extents for the area we're creating a hole
4912  * for.  So if we are truncating this file to a larger size we need to insert
4913  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4914  * the range between oldsize and size
4915  */
4916 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4917 {
4918         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4919         struct btrfs_root *root = BTRFS_I(inode)->root;
4920         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4921         struct extent_map *em = NULL;
4922         struct extent_state *cached_state = NULL;
4923         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4924         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4925         u64 block_end = ALIGN(size, fs_info->sectorsize);
4926         u64 last_byte;
4927         u64 cur_offset;
4928         u64 hole_size;
4929         int err = 0;
4930
4931         /*
4932          * If our size started in the middle of a block we need to zero out the
4933          * rest of the block before we expand the i_size, otherwise we could
4934          * expose stale data.
4935          */
4936         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4937         if (err)
4938                 return err;
4939
4940         if (size <= hole_start)
4941                 return 0;
4942
4943         while (1) {
4944                 struct btrfs_ordered_extent *ordered;
4945
4946                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4947                                  &cached_state);
4948                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4949                                                      block_end - hole_start);
4950                 if (!ordered)
4951                         break;
4952                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4953                                      &cached_state, GFP_NOFS);
4954                 btrfs_start_ordered_extent(inode, ordered, 1);
4955                 btrfs_put_ordered_extent(ordered);
4956         }
4957
4958         cur_offset = hole_start;
4959         while (1) {
4960                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4961                                 block_end - cur_offset, 0);
4962                 if (IS_ERR(em)) {
4963                         err = PTR_ERR(em);
4964                         em = NULL;
4965                         break;
4966                 }
4967                 last_byte = min(extent_map_end(em), block_end);
4968                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4969                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4970                         struct extent_map *hole_em;
4971                         hole_size = last_byte - cur_offset;
4972
4973                         err = maybe_insert_hole(root, inode, cur_offset,
4974                                                 hole_size);
4975                         if (err)
4976                                 break;
4977                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4978                                                 cur_offset + hole_size - 1, 0);
4979                         hole_em = alloc_extent_map();
4980                         if (!hole_em) {
4981                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4982                                         &BTRFS_I(inode)->runtime_flags);
4983                                 goto next;
4984                         }
4985                         hole_em->start = cur_offset;
4986                         hole_em->len = hole_size;
4987                         hole_em->orig_start = cur_offset;
4988
4989                         hole_em->block_start = EXTENT_MAP_HOLE;
4990                         hole_em->block_len = 0;
4991                         hole_em->orig_block_len = 0;
4992                         hole_em->ram_bytes = hole_size;
4993                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
4994                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4995                         hole_em->generation = fs_info->generation;
4996
4997                         while (1) {
4998                                 write_lock(&em_tree->lock);
4999                                 err = add_extent_mapping(em_tree, hole_em, 1);
5000                                 write_unlock(&em_tree->lock);
5001                                 if (err != -EEXIST)
5002                                         break;
5003                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5004                                                         cur_offset,
5005                                                         cur_offset +
5006                                                         hole_size - 1, 0);
5007                         }
5008                         free_extent_map(hole_em);
5009                 }
5010 next:
5011                 free_extent_map(em);
5012                 em = NULL;
5013                 cur_offset = last_byte;
5014                 if (cur_offset >= block_end)
5015                         break;
5016         }
5017         free_extent_map(em);
5018         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
5019                              GFP_NOFS);
5020         return err;
5021 }
5022
5023 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5024 {
5025         struct btrfs_root *root = BTRFS_I(inode)->root;
5026         struct btrfs_trans_handle *trans;
5027         loff_t oldsize = i_size_read(inode);
5028         loff_t newsize = attr->ia_size;
5029         int mask = attr->ia_valid;
5030         int ret;
5031
5032         /*
5033          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5034          * special case where we need to update the times despite not having
5035          * these flags set.  For all other operations the VFS set these flags
5036          * explicitly if it wants a timestamp update.
5037          */
5038         if (newsize != oldsize) {
5039                 inode_inc_iversion(inode);
5040                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5041                         inode->i_ctime = inode->i_mtime =
5042                                 current_time(inode);
5043         }
5044
5045         if (newsize > oldsize) {
5046                 /*
5047                  * Don't do an expanding truncate while snapshoting is ongoing.
5048                  * This is to ensure the snapshot captures a fully consistent
5049                  * state of this file - if the snapshot captures this expanding
5050                  * truncation, it must capture all writes that happened before
5051                  * this truncation.
5052                  */
5053                 btrfs_wait_for_snapshot_creation(root);
5054                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5055                 if (ret) {
5056                         btrfs_end_write_no_snapshoting(root);
5057                         return ret;
5058                 }
5059
5060                 trans = btrfs_start_transaction(root, 1);
5061                 if (IS_ERR(trans)) {
5062                         btrfs_end_write_no_snapshoting(root);
5063                         return PTR_ERR(trans);
5064                 }
5065
5066                 i_size_write(inode, newsize);
5067                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5068                 pagecache_isize_extended(inode, oldsize, newsize);
5069                 ret = btrfs_update_inode(trans, root, inode);
5070                 btrfs_end_write_no_snapshoting(root);
5071                 btrfs_end_transaction(trans);
5072         } else {
5073
5074                 /*
5075                  * We're truncating a file that used to have good data down to
5076                  * zero. Make sure it gets into the ordered flush list so that
5077                  * any new writes get down to disk quickly.
5078                  */
5079                 if (newsize == 0)
5080                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5081                                 &BTRFS_I(inode)->runtime_flags);
5082
5083                 /*
5084                  * 1 for the orphan item we're going to add
5085                  * 1 for the orphan item deletion.
5086                  */
5087                 trans = btrfs_start_transaction(root, 2);
5088                 if (IS_ERR(trans))
5089                         return PTR_ERR(trans);
5090
5091                 /*
5092                  * We need to do this in case we fail at _any_ point during the
5093                  * actual truncate.  Once we do the truncate_setsize we could
5094                  * invalidate pages which forces any outstanding ordered io to
5095                  * be instantly completed which will give us extents that need
5096                  * to be truncated.  If we fail to get an orphan inode down we
5097                  * could have left over extents that were never meant to live,
5098                  * so we need to guarantee from this point on that everything
5099                  * will be consistent.
5100                  */
5101                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5102                 btrfs_end_transaction(trans);
5103                 if (ret)
5104                         return ret;
5105
5106                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5107                 truncate_setsize(inode, newsize);
5108
5109                 /* Disable nonlocked read DIO to avoid the end less truncate */
5110                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5111                 inode_dio_wait(inode);
5112                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5113
5114                 ret = btrfs_truncate(inode);
5115                 if (ret && inode->i_nlink) {
5116                         int err;
5117
5118                         /* To get a stable disk_i_size */
5119                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5120                         if (err) {
5121                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5122                                 return err;
5123                         }
5124
5125                         /*
5126                          * failed to truncate, disk_i_size is only adjusted down
5127                          * as we remove extents, so it should represent the true
5128                          * size of the inode, so reset the in memory size and
5129                          * delete our orphan entry.
5130                          */
5131                         trans = btrfs_join_transaction(root);
5132                         if (IS_ERR(trans)) {
5133                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5134                                 return ret;
5135                         }
5136                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5137                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5138                         if (err)
5139                                 btrfs_abort_transaction(trans, err);
5140                         btrfs_end_transaction(trans);
5141                 }
5142         }
5143
5144         return ret;
5145 }
5146
5147 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5148 {
5149         struct inode *inode = d_inode(dentry);
5150         struct btrfs_root *root = BTRFS_I(inode)->root;
5151         int err;
5152
5153         if (btrfs_root_readonly(root))
5154                 return -EROFS;
5155
5156         err = setattr_prepare(dentry, attr);
5157         if (err)
5158                 return err;
5159
5160         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5161                 err = btrfs_setsize(inode, attr);
5162                 if (err)
5163                         return err;
5164         }
5165
5166         if (attr->ia_valid) {
5167                 setattr_copy(inode, attr);
5168                 inode_inc_iversion(inode);
5169                 err = btrfs_dirty_inode(inode);
5170
5171                 if (!err && attr->ia_valid & ATTR_MODE)
5172                         err = posix_acl_chmod(inode, inode->i_mode);
5173         }
5174
5175         return err;
5176 }
5177
5178 /*
5179  * While truncating the inode pages during eviction, we get the VFS calling
5180  * btrfs_invalidatepage() against each page of the inode. This is slow because
5181  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5182  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5183  * extent_state structures over and over, wasting lots of time.
5184  *
5185  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5186  * those expensive operations on a per page basis and do only the ordered io
5187  * finishing, while we release here the extent_map and extent_state structures,
5188  * without the excessive merging and splitting.
5189  */
5190 static void evict_inode_truncate_pages(struct inode *inode)
5191 {
5192         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5193         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5194         struct rb_node *node;
5195
5196         ASSERT(inode->i_state & I_FREEING);
5197         truncate_inode_pages_final(&inode->i_data);
5198
5199         write_lock(&map_tree->lock);
5200         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5201                 struct extent_map *em;
5202
5203                 node = rb_first(&map_tree->map);
5204                 em = rb_entry(node, struct extent_map, rb_node);
5205                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5206                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5207                 remove_extent_mapping(map_tree, em);
5208                 free_extent_map(em);
5209                 if (need_resched()) {
5210                         write_unlock(&map_tree->lock);
5211                         cond_resched();
5212                         write_lock(&map_tree->lock);
5213                 }
5214         }
5215         write_unlock(&map_tree->lock);
5216
5217         /*
5218          * Keep looping until we have no more ranges in the io tree.
5219          * We can have ongoing bios started by readpages (called from readahead)
5220          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5221          * still in progress (unlocked the pages in the bio but did not yet
5222          * unlocked the ranges in the io tree). Therefore this means some
5223          * ranges can still be locked and eviction started because before
5224          * submitting those bios, which are executed by a separate task (work
5225          * queue kthread), inode references (inode->i_count) were not taken
5226          * (which would be dropped in the end io callback of each bio).
5227          * Therefore here we effectively end up waiting for those bios and
5228          * anyone else holding locked ranges without having bumped the inode's
5229          * reference count - if we don't do it, when they access the inode's
5230          * io_tree to unlock a range it may be too late, leading to an
5231          * use-after-free issue.
5232          */
5233         spin_lock(&io_tree->lock);
5234         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5235                 struct extent_state *state;
5236                 struct extent_state *cached_state = NULL;
5237                 u64 start;
5238                 u64 end;
5239
5240                 node = rb_first(&io_tree->state);
5241                 state = rb_entry(node, struct extent_state, rb_node);
5242                 start = state->start;
5243                 end = state->end;
5244                 spin_unlock(&io_tree->lock);
5245
5246                 lock_extent_bits(io_tree, start, end, &cached_state);
5247
5248                 /*
5249                  * If still has DELALLOC flag, the extent didn't reach disk,
5250                  * and its reserved space won't be freed by delayed_ref.
5251                  * So we need to free its reserved space here.
5252                  * (Refer to comment in btrfs_invalidatepage, case 2)
5253                  *
5254                  * Note, end is the bytenr of last byte, so we need + 1 here.
5255                  */
5256                 if (state->state & EXTENT_DELALLOC)
5257                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5258
5259                 clear_extent_bit(io_tree, start, end,
5260                                  EXTENT_LOCKED | EXTENT_DIRTY |
5261                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5262                                  EXTENT_DEFRAG, 1, 1,
5263                                  &cached_state, GFP_NOFS);
5264
5265                 cond_resched();
5266                 spin_lock(&io_tree->lock);
5267         }
5268         spin_unlock(&io_tree->lock);
5269 }
5270
5271 void btrfs_evict_inode(struct inode *inode)
5272 {
5273         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5274         struct btrfs_trans_handle *trans;
5275         struct btrfs_root *root = BTRFS_I(inode)->root;
5276         struct btrfs_block_rsv *rsv, *global_rsv;
5277         int steal_from_global = 0;
5278         u64 min_size;
5279         int ret;
5280
5281         trace_btrfs_inode_evict(inode);
5282
5283         if (!root) {
5284                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5285                 return;
5286         }
5287
5288         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5289
5290         evict_inode_truncate_pages(inode);
5291
5292         if (inode->i_nlink &&
5293             ((btrfs_root_refs(&root->root_item) != 0 &&
5294               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5295              btrfs_is_free_space_inode(BTRFS_I(inode))))
5296                 goto no_delete;
5297
5298         if (is_bad_inode(inode)) {
5299                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5300                 goto no_delete;
5301         }
5302         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5303         if (!special_file(inode->i_mode))
5304                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5305
5306         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5307
5308         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5309                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5310                                  &BTRFS_I(inode)->runtime_flags));
5311                 goto no_delete;
5312         }
5313
5314         if (inode->i_nlink > 0) {
5315                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5316                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5317                 goto no_delete;
5318         }
5319
5320         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5321         if (ret) {
5322                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5323                 goto no_delete;
5324         }
5325
5326         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5327         if (!rsv) {
5328                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5329                 goto no_delete;
5330         }
5331         rsv->size = min_size;
5332         rsv->failfast = 1;
5333         global_rsv = &fs_info->global_block_rsv;
5334
5335         btrfs_i_size_write(BTRFS_I(inode), 0);
5336
5337         /*
5338          * This is a bit simpler than btrfs_truncate since we've already
5339          * reserved our space for our orphan item in the unlink, so we just
5340          * need to reserve some slack space in case we add bytes and update
5341          * inode item when doing the truncate.
5342          */
5343         while (1) {
5344                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5345                                              BTRFS_RESERVE_FLUSH_LIMIT);
5346
5347                 /*
5348                  * Try and steal from the global reserve since we will
5349                  * likely not use this space anyway, we want to try as
5350                  * hard as possible to get this to work.
5351                  */
5352                 if (ret)
5353                         steal_from_global++;
5354                 else
5355                         steal_from_global = 0;
5356                 ret = 0;
5357
5358                 /*
5359                  * steal_from_global == 0: we reserved stuff, hooray!
5360                  * steal_from_global == 1: we didn't reserve stuff, boo!
5361                  * steal_from_global == 2: we've committed, still not a lot of
5362                  * room but maybe we'll have room in the global reserve this
5363                  * time.
5364                  * steal_from_global == 3: abandon all hope!
5365                  */
5366                 if (steal_from_global > 2) {
5367                         btrfs_warn(fs_info,
5368                                    "Could not get space for a delete, will truncate on mount %d",
5369                                    ret);
5370                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5371                         btrfs_free_block_rsv(fs_info, rsv);
5372                         goto no_delete;
5373                 }
5374
5375                 trans = btrfs_join_transaction(root);
5376                 if (IS_ERR(trans)) {
5377                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5378                         btrfs_free_block_rsv(fs_info, rsv);
5379                         goto no_delete;
5380                 }
5381
5382                 /*
5383                  * We can't just steal from the global reserve, we need to make
5384                  * sure there is room to do it, if not we need to commit and try
5385                  * again.
5386                  */
5387                 if (steal_from_global) {
5388                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5389                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5390                                                               min_size, 0);
5391                         else
5392                                 ret = -ENOSPC;
5393                 }
5394
5395                 /*
5396                  * Couldn't steal from the global reserve, we have too much
5397                  * pending stuff built up, commit the transaction and try it
5398                  * again.
5399                  */
5400                 if (ret) {
5401                         ret = btrfs_commit_transaction(trans);
5402                         if (ret) {
5403                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5404                                 btrfs_free_block_rsv(fs_info, rsv);
5405                                 goto no_delete;
5406                         }
5407                         continue;
5408                 } else {
5409                         steal_from_global = 0;
5410                 }
5411
5412                 trans->block_rsv = rsv;
5413
5414                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5415                 if (ret != -ENOSPC && ret != -EAGAIN)
5416                         break;
5417
5418                 trans->block_rsv = &fs_info->trans_block_rsv;
5419                 btrfs_end_transaction(trans);
5420                 trans = NULL;
5421                 btrfs_btree_balance_dirty(fs_info);
5422         }
5423
5424         btrfs_free_block_rsv(fs_info, rsv);
5425
5426         /*
5427          * Errors here aren't a big deal, it just means we leave orphan items
5428          * in the tree.  They will be cleaned up on the next mount.
5429          */
5430         if (ret == 0) {
5431                 trans->block_rsv = root->orphan_block_rsv;
5432                 btrfs_orphan_del(trans, BTRFS_I(inode));
5433         } else {
5434                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5435         }
5436
5437         trans->block_rsv = &fs_info->trans_block_rsv;
5438         if (!(root == fs_info->tree_root ||
5439               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5440                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5441
5442         btrfs_end_transaction(trans);
5443         btrfs_btree_balance_dirty(fs_info);
5444 no_delete:
5445         btrfs_remove_delayed_node(BTRFS_I(inode));
5446         clear_inode(inode);
5447 }
5448
5449 /*
5450  * this returns the key found in the dir entry in the location pointer.
5451  * If no dir entries were found, location->objectid is 0.
5452  */
5453 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5454                                struct btrfs_key *location)
5455 {
5456         const char *name = dentry->d_name.name;
5457         int namelen = dentry->d_name.len;
5458         struct btrfs_dir_item *di;
5459         struct btrfs_path *path;
5460         struct btrfs_root *root = BTRFS_I(dir)->root;
5461         int ret = 0;
5462
5463         path = btrfs_alloc_path();
5464         if (!path)
5465                 return -ENOMEM;
5466
5467         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5468                         name, namelen, 0);
5469         if (IS_ERR(di))
5470                 ret = PTR_ERR(di);
5471
5472         if (IS_ERR_OR_NULL(di))
5473                 goto out_err;
5474
5475         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5476 out:
5477         btrfs_free_path(path);
5478         return ret;
5479 out_err:
5480         location->objectid = 0;
5481         goto out;
5482 }
5483
5484 /*
5485  * when we hit a tree root in a directory, the btrfs part of the inode
5486  * needs to be changed to reflect the root directory of the tree root.  This
5487  * is kind of like crossing a mount point.
5488  */
5489 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5490                                     struct inode *dir,
5491                                     struct dentry *dentry,
5492                                     struct btrfs_key *location,
5493                                     struct btrfs_root **sub_root)
5494 {
5495         struct btrfs_path *path;
5496         struct btrfs_root *new_root;
5497         struct btrfs_root_ref *ref;
5498         struct extent_buffer *leaf;
5499         struct btrfs_key key;
5500         int ret;
5501         int err = 0;
5502
5503         path = btrfs_alloc_path();
5504         if (!path) {
5505                 err = -ENOMEM;
5506                 goto out;
5507         }
5508
5509         err = -ENOENT;
5510         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5511         key.type = BTRFS_ROOT_REF_KEY;
5512         key.offset = location->objectid;
5513
5514         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5515         if (ret) {
5516                 if (ret < 0)
5517                         err = ret;
5518                 goto out;
5519         }
5520
5521         leaf = path->nodes[0];
5522         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5523         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5524             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5525                 goto out;
5526
5527         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5528                                    (unsigned long)(ref + 1),
5529                                    dentry->d_name.len);
5530         if (ret)
5531                 goto out;
5532
5533         btrfs_release_path(path);
5534
5535         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5536         if (IS_ERR(new_root)) {
5537                 err = PTR_ERR(new_root);
5538                 goto out;
5539         }
5540
5541         *sub_root = new_root;
5542         location->objectid = btrfs_root_dirid(&new_root->root_item);
5543         location->type = BTRFS_INODE_ITEM_KEY;
5544         location->offset = 0;
5545         err = 0;
5546 out:
5547         btrfs_free_path(path);
5548         return err;
5549 }
5550
5551 static void inode_tree_add(struct inode *inode)
5552 {
5553         struct btrfs_root *root = BTRFS_I(inode)->root;
5554         struct btrfs_inode *entry;
5555         struct rb_node **p;
5556         struct rb_node *parent;
5557         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5558         u64 ino = btrfs_ino(BTRFS_I(inode));
5559
5560         if (inode_unhashed(inode))
5561                 return;
5562         parent = NULL;
5563         spin_lock(&root->inode_lock);
5564         p = &root->inode_tree.rb_node;
5565         while (*p) {
5566                 parent = *p;
5567                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5568
5569                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5570                         p = &parent->rb_left;
5571                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5572                         p = &parent->rb_right;
5573                 else {
5574                         WARN_ON(!(entry->vfs_inode.i_state &
5575                                   (I_WILL_FREE | I_FREEING)));
5576                         rb_replace_node(parent, new, &root->inode_tree);
5577                         RB_CLEAR_NODE(parent);
5578                         spin_unlock(&root->inode_lock);
5579                         return;
5580                 }
5581         }
5582         rb_link_node(new, parent, p);
5583         rb_insert_color(new, &root->inode_tree);
5584         spin_unlock(&root->inode_lock);
5585 }
5586
5587 static void inode_tree_del(struct inode *inode)
5588 {
5589         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5590         struct btrfs_root *root = BTRFS_I(inode)->root;
5591         int empty = 0;
5592
5593         spin_lock(&root->inode_lock);
5594         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5595                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5596                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5597                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5598         }
5599         spin_unlock(&root->inode_lock);
5600
5601         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5602                 synchronize_srcu(&fs_info->subvol_srcu);
5603                 spin_lock(&root->inode_lock);
5604                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5605                 spin_unlock(&root->inode_lock);
5606                 if (empty)
5607                         btrfs_add_dead_root(root);
5608         }
5609 }
5610
5611 void btrfs_invalidate_inodes(struct btrfs_root *root)
5612 {
5613         struct btrfs_fs_info *fs_info = root->fs_info;
5614         struct rb_node *node;
5615         struct rb_node *prev;
5616         struct btrfs_inode *entry;
5617         struct inode *inode;
5618         u64 objectid = 0;
5619
5620         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5621                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5622
5623         spin_lock(&root->inode_lock);
5624 again:
5625         node = root->inode_tree.rb_node;
5626         prev = NULL;
5627         while (node) {
5628                 prev = node;
5629                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5630
5631                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5632                         node = node->rb_left;
5633                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5634                         node = node->rb_right;
5635                 else
5636                         break;
5637         }
5638         if (!node) {
5639                 while (prev) {
5640                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5641                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5642                                 node = prev;
5643                                 break;
5644                         }
5645                         prev = rb_next(prev);
5646                 }
5647         }
5648         while (node) {
5649                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5650                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5651                 inode = igrab(&entry->vfs_inode);
5652                 if (inode) {
5653                         spin_unlock(&root->inode_lock);
5654                         if (atomic_read(&inode->i_count) > 1)
5655                                 d_prune_aliases(inode);
5656                         /*
5657                          * btrfs_drop_inode will have it removed from
5658                          * the inode cache when its usage count
5659                          * hits zero.
5660                          */
5661                         iput(inode);
5662                         cond_resched();
5663                         spin_lock(&root->inode_lock);
5664                         goto again;
5665                 }
5666
5667                 if (cond_resched_lock(&root->inode_lock))
5668                         goto again;
5669
5670                 node = rb_next(node);
5671         }
5672         spin_unlock(&root->inode_lock);
5673 }
5674
5675 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5676 {
5677         struct btrfs_iget_args *args = p;
5678         inode->i_ino = args->location->objectid;
5679         memcpy(&BTRFS_I(inode)->location, args->location,
5680                sizeof(*args->location));
5681         BTRFS_I(inode)->root = args->root;
5682         return 0;
5683 }
5684
5685 static int btrfs_find_actor(struct inode *inode, void *opaque)
5686 {
5687         struct btrfs_iget_args *args = opaque;
5688         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5689                 args->root == BTRFS_I(inode)->root;
5690 }
5691
5692 static struct inode *btrfs_iget_locked(struct super_block *s,
5693                                        struct btrfs_key *location,
5694                                        struct btrfs_root *root)
5695 {
5696         struct inode *inode;
5697         struct btrfs_iget_args args;
5698         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5699
5700         args.location = location;
5701         args.root = root;
5702
5703         inode = iget5_locked(s, hashval, btrfs_find_actor,
5704                              btrfs_init_locked_inode,
5705                              (void *)&args);
5706         return inode;
5707 }
5708
5709 /* Get an inode object given its location and corresponding root.
5710  * Returns in *is_new if the inode was read from disk
5711  */
5712 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5713                          struct btrfs_root *root, int *new)
5714 {
5715         struct inode *inode;
5716
5717         inode = btrfs_iget_locked(s, location, root);
5718         if (!inode)
5719                 return ERR_PTR(-ENOMEM);
5720
5721         if (inode->i_state & I_NEW) {
5722                 int ret;
5723
5724                 ret = btrfs_read_locked_inode(inode);
5725                 if (!is_bad_inode(inode)) {
5726                         inode_tree_add(inode);
5727                         unlock_new_inode(inode);
5728                         if (new)
5729                                 *new = 1;
5730                 } else {
5731                         unlock_new_inode(inode);
5732                         iput(inode);
5733                         ASSERT(ret < 0);
5734                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5735                 }
5736         }
5737
5738         return inode;
5739 }
5740
5741 static struct inode *new_simple_dir(struct super_block *s,
5742                                     struct btrfs_key *key,
5743                                     struct btrfs_root *root)
5744 {
5745         struct inode *inode = new_inode(s);
5746
5747         if (!inode)
5748                 return ERR_PTR(-ENOMEM);
5749
5750         BTRFS_I(inode)->root = root;
5751         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5752         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5753
5754         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5755         inode->i_op = &btrfs_dir_ro_inode_operations;
5756         inode->i_opflags &= ~IOP_XATTR;
5757         inode->i_fop = &simple_dir_operations;
5758         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5759         inode->i_mtime = current_time(inode);
5760         inode->i_atime = inode->i_mtime;
5761         inode->i_ctime = inode->i_mtime;
5762         BTRFS_I(inode)->i_otime = inode->i_mtime;
5763
5764         return inode;
5765 }
5766
5767 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5768 {
5769         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5770         struct inode *inode;
5771         struct btrfs_root *root = BTRFS_I(dir)->root;
5772         struct btrfs_root *sub_root = root;
5773         struct btrfs_key location;
5774         int index;
5775         int ret = 0;
5776
5777         if (dentry->d_name.len > BTRFS_NAME_LEN)
5778                 return ERR_PTR(-ENAMETOOLONG);
5779
5780         ret = btrfs_inode_by_name(dir, dentry, &location);
5781         if (ret < 0)
5782                 return ERR_PTR(ret);
5783
5784         if (location.objectid == 0)
5785                 return ERR_PTR(-ENOENT);
5786
5787         if (location.type == BTRFS_INODE_ITEM_KEY) {
5788                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5789                 return inode;
5790         }
5791
5792         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5793
5794         index = srcu_read_lock(&fs_info->subvol_srcu);
5795         ret = fixup_tree_root_location(fs_info, dir, dentry,
5796                                        &location, &sub_root);
5797         if (ret < 0) {
5798                 if (ret != -ENOENT)
5799                         inode = ERR_PTR(ret);
5800                 else
5801                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5802         } else {
5803                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5804         }
5805         srcu_read_unlock(&fs_info->subvol_srcu, index);
5806
5807         if (!IS_ERR(inode) && root != sub_root) {
5808                 down_read(&fs_info->cleanup_work_sem);
5809                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5810                         ret = btrfs_orphan_cleanup(sub_root);
5811                 up_read(&fs_info->cleanup_work_sem);
5812                 if (ret) {
5813                         iput(inode);
5814                         inode = ERR_PTR(ret);
5815                 }
5816         }
5817
5818         return inode;
5819 }
5820
5821 static int btrfs_dentry_delete(const struct dentry *dentry)
5822 {
5823         struct btrfs_root *root;
5824         struct inode *inode = d_inode(dentry);
5825
5826         if (!inode && !IS_ROOT(dentry))
5827                 inode = d_inode(dentry->d_parent);
5828
5829         if (inode) {
5830                 root = BTRFS_I(inode)->root;
5831                 if (btrfs_root_refs(&root->root_item) == 0)
5832                         return 1;
5833
5834                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5835                         return 1;
5836         }
5837         return 0;
5838 }
5839
5840 static void btrfs_dentry_release(struct dentry *dentry)
5841 {
5842         kfree(dentry->d_fsdata);
5843 }
5844
5845 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5846                                    unsigned int flags)
5847 {
5848         struct inode *inode;
5849
5850         inode = btrfs_lookup_dentry(dir, dentry);
5851         if (IS_ERR(inode)) {
5852                 if (PTR_ERR(inode) == -ENOENT)
5853                         inode = NULL;
5854                 else
5855                         return ERR_CAST(inode);
5856         }
5857
5858         return d_splice_alias(inode, dentry);
5859 }
5860
5861 unsigned char btrfs_filetype_table[] = {
5862         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5863 };
5864
5865 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5866 {
5867         struct inode *inode = file_inode(file);
5868         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5869         struct btrfs_root *root = BTRFS_I(inode)->root;
5870         struct btrfs_item *item;
5871         struct btrfs_dir_item *di;
5872         struct btrfs_key key;
5873         struct btrfs_key found_key;
5874         struct btrfs_path *path;
5875         struct list_head ins_list;
5876         struct list_head del_list;
5877         int ret;
5878         struct extent_buffer *leaf;
5879         int slot;
5880         unsigned char d_type;
5881         int over = 0;
5882         char tmp_name[32];
5883         char *name_ptr;
5884         int name_len;
5885         bool put = false;
5886         struct btrfs_key location;
5887
5888         if (!dir_emit_dots(file, ctx))
5889                 return 0;
5890
5891         path = btrfs_alloc_path();
5892         if (!path)
5893                 return -ENOMEM;
5894
5895         path->reada = READA_FORWARD;
5896
5897         INIT_LIST_HEAD(&ins_list);
5898         INIT_LIST_HEAD(&del_list);
5899         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5900
5901         key.type = BTRFS_DIR_INDEX_KEY;
5902         key.offset = ctx->pos;
5903         key.objectid = btrfs_ino(BTRFS_I(inode));
5904
5905         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5906         if (ret < 0)
5907                 goto err;
5908
5909         while (1) {
5910                 leaf = path->nodes[0];
5911                 slot = path->slots[0];
5912                 if (slot >= btrfs_header_nritems(leaf)) {
5913                         ret = btrfs_next_leaf(root, path);
5914                         if (ret < 0)
5915                                 goto err;
5916                         else if (ret > 0)
5917                                 break;
5918                         continue;
5919                 }
5920
5921                 item = btrfs_item_nr(slot);
5922                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5923
5924                 if (found_key.objectid != key.objectid)
5925                         break;
5926                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5927                         break;
5928                 if (found_key.offset < ctx->pos)
5929                         goto next;
5930                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5931                         goto next;
5932
5933                 ctx->pos = found_key.offset;
5934
5935                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5936                 if (verify_dir_item(fs_info, leaf, di))
5937                         goto next;
5938
5939                 name_len = btrfs_dir_name_len(leaf, di);
5940                 if (name_len <= sizeof(tmp_name)) {
5941                         name_ptr = tmp_name;
5942                 } else {
5943                         name_ptr = kmalloc(name_len, GFP_KERNEL);
5944                         if (!name_ptr) {
5945                                 ret = -ENOMEM;
5946                                 goto err;
5947                         }
5948                 }
5949                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5950                                    name_len);
5951
5952                 d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5953                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5954
5955                 over = !dir_emit(ctx, name_ptr, name_len, location.objectid,
5956                                  d_type);
5957
5958                 if (name_ptr != tmp_name)
5959                         kfree(name_ptr);
5960
5961                 if (over)
5962                         goto nopos;
5963                 ctx->pos++;
5964 next:
5965                 path->slots[0]++;
5966         }
5967
5968         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5969         if (ret)
5970                 goto nopos;
5971
5972         /*
5973          * Stop new entries from being returned after we return the last
5974          * entry.
5975          *
5976          * New directory entries are assigned a strictly increasing
5977          * offset.  This means that new entries created during readdir
5978          * are *guaranteed* to be seen in the future by that readdir.
5979          * This has broken buggy programs which operate on names as
5980          * they're returned by readdir.  Until we re-use freed offsets
5981          * we have this hack to stop new entries from being returned
5982          * under the assumption that they'll never reach this huge
5983          * offset.
5984          *
5985          * This is being careful not to overflow 32bit loff_t unless the
5986          * last entry requires it because doing so has broken 32bit apps
5987          * in the past.
5988          */
5989         if (ctx->pos >= INT_MAX)
5990                 ctx->pos = LLONG_MAX;
5991         else
5992                 ctx->pos = INT_MAX;
5993 nopos:
5994         ret = 0;
5995 err:
5996         if (put)
5997                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5998         btrfs_free_path(path);
5999         return ret;
6000 }
6001
6002 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6003 {
6004         struct btrfs_root *root = BTRFS_I(inode)->root;
6005         struct btrfs_trans_handle *trans;
6006         int ret = 0;
6007         bool nolock = false;
6008
6009         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6010                 return 0;
6011
6012         if (btrfs_fs_closing(root->fs_info) &&
6013                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6014                 nolock = true;
6015
6016         if (wbc->sync_mode == WB_SYNC_ALL) {
6017                 if (nolock)
6018                         trans = btrfs_join_transaction_nolock(root);
6019                 else
6020                         trans = btrfs_join_transaction(root);
6021                 if (IS_ERR(trans))
6022                         return PTR_ERR(trans);
6023                 ret = btrfs_commit_transaction(trans);
6024         }
6025         return ret;
6026 }
6027
6028 /*
6029  * This is somewhat expensive, updating the tree every time the
6030  * inode changes.  But, it is most likely to find the inode in cache.
6031  * FIXME, needs more benchmarking...there are no reasons other than performance
6032  * to keep or drop this code.
6033  */
6034 static int btrfs_dirty_inode(struct inode *inode)
6035 {
6036         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6037         struct btrfs_root *root = BTRFS_I(inode)->root;
6038         struct btrfs_trans_handle *trans;
6039         int ret;
6040
6041         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6042                 return 0;
6043
6044         trans = btrfs_join_transaction(root);
6045         if (IS_ERR(trans))
6046                 return PTR_ERR(trans);
6047
6048         ret = btrfs_update_inode(trans, root, inode);
6049         if (ret && ret == -ENOSPC) {
6050                 /* whoops, lets try again with the full transaction */
6051                 btrfs_end_transaction(trans);
6052                 trans = btrfs_start_transaction(root, 1);
6053                 if (IS_ERR(trans))
6054                         return PTR_ERR(trans);
6055
6056                 ret = btrfs_update_inode(trans, root, inode);
6057         }
6058         btrfs_end_transaction(trans);
6059         if (BTRFS_I(inode)->delayed_node)
6060                 btrfs_balance_delayed_items(fs_info);
6061
6062         return ret;
6063 }
6064
6065 /*
6066  * This is a copy of file_update_time.  We need this so we can return error on
6067  * ENOSPC for updating the inode in the case of file write and mmap writes.
6068  */
6069 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6070                              int flags)
6071 {
6072         struct btrfs_root *root = BTRFS_I(inode)->root;
6073
6074         if (btrfs_root_readonly(root))
6075                 return -EROFS;
6076
6077         if (flags & S_VERSION)
6078                 inode_inc_iversion(inode);
6079         if (flags & S_CTIME)
6080                 inode->i_ctime = *now;
6081         if (flags & S_MTIME)
6082                 inode->i_mtime = *now;
6083         if (flags & S_ATIME)
6084                 inode->i_atime = *now;
6085         return btrfs_dirty_inode(inode);
6086 }
6087
6088 /*
6089  * find the highest existing sequence number in a directory
6090  * and then set the in-memory index_cnt variable to reflect
6091  * free sequence numbers
6092  */
6093 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6094 {
6095         struct btrfs_root *root = inode->root;
6096         struct btrfs_key key, found_key;
6097         struct btrfs_path *path;
6098         struct extent_buffer *leaf;
6099         int ret;
6100
6101         key.objectid = btrfs_ino(inode);
6102         key.type = BTRFS_DIR_INDEX_KEY;
6103         key.offset = (u64)-1;
6104
6105         path = btrfs_alloc_path();
6106         if (!path)
6107                 return -ENOMEM;
6108
6109         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6110         if (ret < 0)
6111                 goto out;
6112         /* FIXME: we should be able to handle this */
6113         if (ret == 0)
6114                 goto out;
6115         ret = 0;
6116
6117         /*
6118          * MAGIC NUMBER EXPLANATION:
6119          * since we search a directory based on f_pos we have to start at 2
6120          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6121          * else has to start at 2
6122          */
6123         if (path->slots[0] == 0) {
6124                 inode->index_cnt = 2;
6125                 goto out;
6126         }
6127
6128         path->slots[0]--;
6129
6130         leaf = path->nodes[0];
6131         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6132
6133         if (found_key.objectid != btrfs_ino(inode) ||
6134             found_key.type != BTRFS_DIR_INDEX_KEY) {
6135                 inode->index_cnt = 2;
6136                 goto out;
6137         }
6138
6139         inode->index_cnt = found_key.offset + 1;
6140 out:
6141         btrfs_free_path(path);
6142         return ret;
6143 }
6144
6145 /*
6146  * helper to find a free sequence number in a given directory.  This current
6147  * code is very simple, later versions will do smarter things in the btree
6148  */
6149 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6150 {
6151         int ret = 0;
6152
6153         if (dir->index_cnt == (u64)-1) {
6154                 ret = btrfs_inode_delayed_dir_index_count(dir);
6155                 if (ret) {
6156                         ret = btrfs_set_inode_index_count(dir);
6157                         if (ret)
6158                                 return ret;
6159                 }
6160         }
6161
6162         *index = dir->index_cnt;
6163         dir->index_cnt++;
6164
6165         return ret;
6166 }
6167
6168 static int btrfs_insert_inode_locked(struct inode *inode)
6169 {
6170         struct btrfs_iget_args args;
6171         args.location = &BTRFS_I(inode)->location;
6172         args.root = BTRFS_I(inode)->root;
6173
6174         return insert_inode_locked4(inode,
6175                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6176                    btrfs_find_actor, &args);
6177 }
6178
6179 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6180                                      struct btrfs_root *root,
6181                                      struct inode *dir,
6182                                      const char *name, int name_len,
6183                                      u64 ref_objectid, u64 objectid,
6184                                      umode_t mode, u64 *index)
6185 {
6186         struct btrfs_fs_info *fs_info = root->fs_info;
6187         struct inode *inode;
6188         struct btrfs_inode_item *inode_item;
6189         struct btrfs_key *location;
6190         struct btrfs_path *path;
6191         struct btrfs_inode_ref *ref;
6192         struct btrfs_key key[2];
6193         u32 sizes[2];
6194         int nitems = name ? 2 : 1;
6195         unsigned long ptr;
6196         int ret;
6197
6198         path = btrfs_alloc_path();
6199         if (!path)
6200                 return ERR_PTR(-ENOMEM);
6201
6202         inode = new_inode(fs_info->sb);
6203         if (!inode) {
6204                 btrfs_free_path(path);
6205                 return ERR_PTR(-ENOMEM);
6206         }
6207
6208         /*
6209          * O_TMPFILE, set link count to 0, so that after this point,
6210          * we fill in an inode item with the correct link count.
6211          */
6212         if (!name)
6213                 set_nlink(inode, 0);
6214
6215         /*
6216          * we have to initialize this early, so we can reclaim the inode
6217          * number if we fail afterwards in this function.
6218          */
6219         inode->i_ino = objectid;
6220
6221         if (dir && name) {
6222                 trace_btrfs_inode_request(dir);
6223
6224                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6225                 if (ret) {
6226                         btrfs_free_path(path);
6227                         iput(inode);
6228                         return ERR_PTR(ret);
6229                 }
6230         } else if (dir) {
6231                 *index = 0;
6232         }
6233         /*
6234          * index_cnt is ignored for everything but a dir,
6235          * btrfs_get_inode_index_count has an explanation for the magic
6236          * number
6237          */
6238         BTRFS_I(inode)->index_cnt = 2;
6239         BTRFS_I(inode)->dir_index = *index;
6240         BTRFS_I(inode)->root = root;
6241         BTRFS_I(inode)->generation = trans->transid;
6242         inode->i_generation = BTRFS_I(inode)->generation;
6243
6244         /*
6245          * We could have gotten an inode number from somebody who was fsynced
6246          * and then removed in this same transaction, so let's just set full
6247          * sync since it will be a full sync anyway and this will blow away the
6248          * old info in the log.
6249          */
6250         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6251
6252         key[0].objectid = objectid;
6253         key[0].type = BTRFS_INODE_ITEM_KEY;
6254         key[0].offset = 0;
6255
6256         sizes[0] = sizeof(struct btrfs_inode_item);
6257
6258         if (name) {
6259                 /*
6260                  * Start new inodes with an inode_ref. This is slightly more
6261                  * efficient for small numbers of hard links since they will
6262                  * be packed into one item. Extended refs will kick in if we
6263                  * add more hard links than can fit in the ref item.
6264                  */
6265                 key[1].objectid = objectid;
6266                 key[1].type = BTRFS_INODE_REF_KEY;
6267                 key[1].offset = ref_objectid;
6268
6269                 sizes[1] = name_len + sizeof(*ref);
6270         }
6271
6272         location = &BTRFS_I(inode)->location;
6273         location->objectid = objectid;
6274         location->offset = 0;
6275         location->type = BTRFS_INODE_ITEM_KEY;
6276
6277         ret = btrfs_insert_inode_locked(inode);
6278         if (ret < 0)
6279                 goto fail;
6280
6281         path->leave_spinning = 1;
6282         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6283         if (ret != 0)
6284                 goto fail_unlock;
6285
6286         inode_init_owner(inode, dir, mode);
6287         inode_set_bytes(inode, 0);
6288
6289         inode->i_mtime = current_time(inode);
6290         inode->i_atime = inode->i_mtime;
6291         inode->i_ctime = inode->i_mtime;
6292         BTRFS_I(inode)->i_otime = inode->i_mtime;
6293
6294         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6295                                   struct btrfs_inode_item);
6296         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6297                              sizeof(*inode_item));
6298         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6299
6300         if (name) {
6301                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6302                                      struct btrfs_inode_ref);
6303                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6304                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6305                 ptr = (unsigned long)(ref + 1);
6306                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6307         }
6308
6309         btrfs_mark_buffer_dirty(path->nodes[0]);
6310         btrfs_free_path(path);
6311
6312         btrfs_inherit_iflags(inode, dir);
6313
6314         if (S_ISREG(mode)) {
6315                 if (btrfs_test_opt(fs_info, NODATASUM))
6316                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6317                 if (btrfs_test_opt(fs_info, NODATACOW))
6318                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6319                                 BTRFS_INODE_NODATASUM;
6320         }
6321
6322         inode_tree_add(inode);
6323
6324         trace_btrfs_inode_new(inode);
6325         btrfs_set_inode_last_trans(trans, inode);
6326
6327         btrfs_update_root_times(trans, root);
6328
6329         ret = btrfs_inode_inherit_props(trans, inode, dir);
6330         if (ret)
6331                 btrfs_err(fs_info,
6332                           "error inheriting props for ino %llu (root %llu): %d",
6333                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6334
6335         return inode;
6336
6337 fail_unlock:
6338         unlock_new_inode(inode);
6339 fail:
6340         if (dir && name)
6341                 BTRFS_I(dir)->index_cnt--;
6342         btrfs_free_path(path);
6343         iput(inode);
6344         return ERR_PTR(ret);
6345 }
6346
6347 static inline u8 btrfs_inode_type(struct inode *inode)
6348 {
6349         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6350 }
6351
6352 /*
6353  * utility function to add 'inode' into 'parent_inode' with
6354  * a give name and a given sequence number.
6355  * if 'add_backref' is true, also insert a backref from the
6356  * inode to the parent directory.
6357  */
6358 int btrfs_add_link(struct btrfs_trans_handle *trans,
6359                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6360                    const char *name, int name_len, int add_backref, u64 index)
6361 {
6362         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6363         int ret = 0;
6364         struct btrfs_key key;
6365         struct btrfs_root *root = parent_inode->root;
6366         u64 ino = btrfs_ino(inode);
6367         u64 parent_ino = btrfs_ino(parent_inode);
6368
6369         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6370                 memcpy(&key, &inode->root->root_key, sizeof(key));
6371         } else {
6372                 key.objectid = ino;
6373                 key.type = BTRFS_INODE_ITEM_KEY;
6374                 key.offset = 0;
6375         }
6376
6377         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6378                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6379                                          root->root_key.objectid, parent_ino,
6380                                          index, name, name_len);
6381         } else if (add_backref) {
6382                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6383                                              parent_ino, index);
6384         }
6385
6386         /* Nothing to clean up yet */
6387         if (ret)
6388                 return ret;
6389
6390         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6391                                     parent_inode, &key,
6392                                     btrfs_inode_type(&inode->vfs_inode), index);
6393         if (ret == -EEXIST || ret == -EOVERFLOW)
6394                 goto fail_dir_item;
6395         else if (ret) {
6396                 btrfs_abort_transaction(trans, ret);
6397                 return ret;
6398         }
6399
6400         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6401                            name_len * 2);
6402         inode_inc_iversion(&parent_inode->vfs_inode);
6403         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6404                 current_time(&parent_inode->vfs_inode);
6405         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6406         if (ret)
6407                 btrfs_abort_transaction(trans, ret);
6408         return ret;
6409
6410 fail_dir_item:
6411         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6412                 u64 local_index;
6413                 int err;
6414                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6415                                          root->root_key.objectid, parent_ino,
6416                                          &local_index, name, name_len);
6417
6418         } else if (add_backref) {
6419                 u64 local_index;
6420                 int err;
6421
6422                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6423                                           ino, parent_ino, &local_index);
6424         }
6425         return ret;
6426 }
6427
6428 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6429                             struct btrfs_inode *dir, struct dentry *dentry,
6430                             struct btrfs_inode *inode, int backref, u64 index)
6431 {
6432         int err = btrfs_add_link(trans, dir, inode,
6433                                  dentry->d_name.name, dentry->d_name.len,
6434                                  backref, index);
6435         if (err > 0)
6436                 err = -EEXIST;
6437         return err;
6438 }
6439
6440 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6441                         umode_t mode, dev_t rdev)
6442 {
6443         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6444         struct btrfs_trans_handle *trans;
6445         struct btrfs_root *root = BTRFS_I(dir)->root;
6446         struct inode *inode = NULL;
6447         int err;
6448         int drop_inode = 0;
6449         u64 objectid;
6450         u64 index = 0;
6451
6452         /*
6453          * 2 for inode item and ref
6454          * 2 for dir items
6455          * 1 for xattr if selinux is on
6456          */
6457         trans = btrfs_start_transaction(root, 5);
6458         if (IS_ERR(trans))
6459                 return PTR_ERR(trans);
6460
6461         err = btrfs_find_free_ino(root, &objectid);
6462         if (err)
6463                 goto out_unlock;
6464
6465         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6466                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6467                         mode, &index);
6468         if (IS_ERR(inode)) {
6469                 err = PTR_ERR(inode);
6470                 goto out_unlock;
6471         }
6472
6473         /*
6474         * If the active LSM wants to access the inode during
6475         * d_instantiate it needs these. Smack checks to see
6476         * if the filesystem supports xattrs by looking at the
6477         * ops vector.
6478         */
6479         inode->i_op = &btrfs_special_inode_operations;
6480         init_special_inode(inode, inode->i_mode, rdev);
6481
6482         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6483         if (err)
6484                 goto out_unlock_inode;
6485
6486         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6487                         0, index);
6488         if (err) {
6489                 goto out_unlock_inode;
6490         } else {
6491                 btrfs_update_inode(trans, root, inode);
6492                 unlock_new_inode(inode);
6493                 d_instantiate(dentry, inode);
6494         }
6495
6496 out_unlock:
6497         btrfs_end_transaction(trans);
6498         btrfs_balance_delayed_items(fs_info);
6499         btrfs_btree_balance_dirty(fs_info);
6500         if (drop_inode) {
6501                 inode_dec_link_count(inode);
6502                 iput(inode);
6503         }
6504         return err;
6505
6506 out_unlock_inode:
6507         drop_inode = 1;
6508         unlock_new_inode(inode);
6509         goto out_unlock;
6510
6511 }
6512
6513 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6514                         umode_t mode, bool excl)
6515 {
6516         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6517         struct btrfs_trans_handle *trans;
6518         struct btrfs_root *root = BTRFS_I(dir)->root;
6519         struct inode *inode = NULL;
6520         int drop_inode_on_err = 0;
6521         int err;
6522         u64 objectid;
6523         u64 index = 0;
6524
6525         /*
6526          * 2 for inode item and ref
6527          * 2 for dir items
6528          * 1 for xattr if selinux is on
6529          */
6530         trans = btrfs_start_transaction(root, 5);
6531         if (IS_ERR(trans))
6532                 return PTR_ERR(trans);
6533
6534         err = btrfs_find_free_ino(root, &objectid);
6535         if (err)
6536                 goto out_unlock;
6537
6538         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6539                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6540                         mode, &index);
6541         if (IS_ERR(inode)) {
6542                 err = PTR_ERR(inode);
6543                 goto out_unlock;
6544         }
6545         drop_inode_on_err = 1;
6546         /*
6547         * If the active LSM wants to access the inode during
6548         * d_instantiate it needs these. Smack checks to see
6549         * if the filesystem supports xattrs by looking at the
6550         * ops vector.
6551         */
6552         inode->i_fop = &btrfs_file_operations;
6553         inode->i_op = &btrfs_file_inode_operations;
6554         inode->i_mapping->a_ops = &btrfs_aops;
6555
6556         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6557         if (err)
6558                 goto out_unlock_inode;
6559
6560         err = btrfs_update_inode(trans, root, inode);
6561         if (err)
6562                 goto out_unlock_inode;
6563
6564         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6565                         0, index);
6566         if (err)
6567                 goto out_unlock_inode;
6568
6569         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6570         unlock_new_inode(inode);
6571         d_instantiate(dentry, inode);
6572
6573 out_unlock:
6574         btrfs_end_transaction(trans);
6575         if (err && drop_inode_on_err) {
6576                 inode_dec_link_count(inode);
6577                 iput(inode);
6578         }
6579         btrfs_balance_delayed_items(fs_info);
6580         btrfs_btree_balance_dirty(fs_info);
6581         return err;
6582
6583 out_unlock_inode:
6584         unlock_new_inode(inode);
6585         goto out_unlock;
6586
6587 }
6588
6589 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6590                       struct dentry *dentry)
6591 {
6592         struct btrfs_trans_handle *trans = NULL;
6593         struct btrfs_root *root = BTRFS_I(dir)->root;
6594         struct inode *inode = d_inode(old_dentry);
6595         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6596         u64 index;
6597         int err;
6598         int drop_inode = 0;
6599
6600         /* do not allow sys_link's with other subvols of the same device */
6601         if (root->objectid != BTRFS_I(inode)->root->objectid)
6602                 return -EXDEV;
6603
6604         if (inode->i_nlink >= BTRFS_LINK_MAX)
6605                 return -EMLINK;
6606
6607         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6608         if (err)
6609                 goto fail;
6610
6611         /*
6612          * 2 items for inode and inode ref
6613          * 2 items for dir items
6614          * 1 item for parent inode
6615          */
6616         trans = btrfs_start_transaction(root, 5);
6617         if (IS_ERR(trans)) {
6618                 err = PTR_ERR(trans);
6619                 trans = NULL;
6620                 goto fail;
6621         }
6622
6623         /* There are several dir indexes for this inode, clear the cache. */
6624         BTRFS_I(inode)->dir_index = 0ULL;
6625         inc_nlink(inode);
6626         inode_inc_iversion(inode);
6627         inode->i_ctime = current_time(inode);
6628         ihold(inode);
6629         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6630
6631         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6632                         1, index);
6633
6634         if (err) {
6635                 drop_inode = 1;
6636         } else {
6637                 struct dentry *parent = dentry->d_parent;
6638                 err = btrfs_update_inode(trans, root, inode);
6639                 if (err)
6640                         goto fail;
6641                 if (inode->i_nlink == 1) {
6642                         /*
6643                          * If new hard link count is 1, it's a file created
6644                          * with open(2) O_TMPFILE flag.
6645                          */
6646                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6647                         if (err)
6648                                 goto fail;
6649                 }
6650                 d_instantiate(dentry, inode);
6651                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6652         }
6653
6654         btrfs_balance_delayed_items(fs_info);
6655 fail:
6656         if (trans)
6657                 btrfs_end_transaction(trans);
6658         if (drop_inode) {
6659                 inode_dec_link_count(inode);
6660                 iput(inode);
6661         }
6662         btrfs_btree_balance_dirty(fs_info);
6663         return err;
6664 }
6665
6666 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6667 {
6668         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6669         struct inode *inode = NULL;
6670         struct btrfs_trans_handle *trans;
6671         struct btrfs_root *root = BTRFS_I(dir)->root;
6672         int err = 0;
6673         int drop_on_err = 0;
6674         u64 objectid = 0;
6675         u64 index = 0;
6676
6677         /*
6678          * 2 items for inode and ref
6679          * 2 items for dir items
6680          * 1 for xattr if selinux is on
6681          */
6682         trans = btrfs_start_transaction(root, 5);
6683         if (IS_ERR(trans))
6684                 return PTR_ERR(trans);
6685
6686         err = btrfs_find_free_ino(root, &objectid);
6687         if (err)
6688                 goto out_fail;
6689
6690         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6691                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6692                         S_IFDIR | mode, &index);
6693         if (IS_ERR(inode)) {
6694                 err = PTR_ERR(inode);
6695                 goto out_fail;
6696         }
6697
6698         drop_on_err = 1;
6699         /* these must be set before we unlock the inode */
6700         inode->i_op = &btrfs_dir_inode_operations;
6701         inode->i_fop = &btrfs_dir_file_operations;
6702
6703         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6704         if (err)
6705                 goto out_fail_inode;
6706
6707         btrfs_i_size_write(BTRFS_I(inode), 0);
6708         err = btrfs_update_inode(trans, root, inode);
6709         if (err)
6710                 goto out_fail_inode;
6711
6712         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6713                         dentry->d_name.name,
6714                         dentry->d_name.len, 0, index);
6715         if (err)
6716                 goto out_fail_inode;
6717
6718         d_instantiate(dentry, inode);
6719         /*
6720          * mkdir is special.  We're unlocking after we call d_instantiate
6721          * to avoid a race with nfsd calling d_instantiate.
6722          */
6723         unlock_new_inode(inode);
6724         drop_on_err = 0;
6725
6726 out_fail:
6727         btrfs_end_transaction(trans);
6728         if (drop_on_err) {
6729                 inode_dec_link_count(inode);
6730                 iput(inode);
6731         }
6732         btrfs_balance_delayed_items(fs_info);
6733         btrfs_btree_balance_dirty(fs_info);
6734         return err;
6735
6736 out_fail_inode:
6737         unlock_new_inode(inode);
6738         goto out_fail;
6739 }
6740
6741 /* Find next extent map of a given extent map, caller needs to ensure locks */
6742 static struct extent_map *next_extent_map(struct extent_map *em)
6743 {
6744         struct rb_node *next;
6745
6746         next = rb_next(&em->rb_node);
6747         if (!next)
6748                 return NULL;
6749         return container_of(next, struct extent_map, rb_node);
6750 }
6751
6752 static struct extent_map *prev_extent_map(struct extent_map *em)
6753 {
6754         struct rb_node *prev;
6755
6756         prev = rb_prev(&em->rb_node);
6757         if (!prev)
6758                 return NULL;
6759         return container_of(prev, struct extent_map, rb_node);
6760 }
6761
6762 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6763  * the existing extent is the nearest extent to map_start,
6764  * and an extent that you want to insert, deal with overlap and insert
6765  * the best fitted new extent into the tree.
6766  */
6767 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6768                                 struct extent_map *existing,
6769                                 struct extent_map *em,
6770                                 u64 map_start)
6771 {
6772         struct extent_map *prev;
6773         struct extent_map *next;
6774         u64 start;
6775         u64 end;
6776         u64 start_diff;
6777
6778         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6779
6780         if (existing->start > map_start) {
6781                 next = existing;
6782                 prev = prev_extent_map(next);
6783         } else {
6784                 prev = existing;
6785                 next = next_extent_map(prev);
6786         }
6787
6788         start = prev ? extent_map_end(prev) : em->start;
6789         start = max_t(u64, start, em->start);
6790         end = next ? next->start : extent_map_end(em);
6791         end = min_t(u64, end, extent_map_end(em));
6792         start_diff = start - em->start;
6793         em->start = start;
6794         em->len = end - start;
6795         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6796             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6797                 em->block_start += start_diff;
6798                 em->block_len -= start_diff;
6799         }
6800         return add_extent_mapping(em_tree, em, 0);
6801 }
6802
6803 static noinline int uncompress_inline(struct btrfs_path *path,
6804                                       struct page *page,
6805                                       size_t pg_offset, u64 extent_offset,
6806                                       struct btrfs_file_extent_item *item)
6807 {
6808         int ret;
6809         struct extent_buffer *leaf = path->nodes[0];
6810         char *tmp;
6811         size_t max_size;
6812         unsigned long inline_size;
6813         unsigned long ptr;
6814         int compress_type;
6815
6816         WARN_ON(pg_offset != 0);
6817         compress_type = btrfs_file_extent_compression(leaf, item);
6818         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6819         inline_size = btrfs_file_extent_inline_item_len(leaf,
6820                                         btrfs_item_nr(path->slots[0]));
6821         tmp = kmalloc(inline_size, GFP_NOFS);
6822         if (!tmp)
6823                 return -ENOMEM;
6824         ptr = btrfs_file_extent_inline_start(item);
6825
6826         read_extent_buffer(leaf, tmp, ptr, inline_size);
6827
6828         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6829         ret = btrfs_decompress(compress_type, tmp, page,
6830                                extent_offset, inline_size, max_size);
6831
6832         /*
6833          * decompression code contains a memset to fill in any space between the end
6834          * of the uncompressed data and the end of max_size in case the decompressed
6835          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6836          * the end of an inline extent and the beginning of the next block, so we
6837          * cover that region here.
6838          */
6839
6840         if (max_size + pg_offset < PAGE_SIZE) {
6841                 char *map = kmap(page);
6842                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6843                 kunmap(page);
6844         }
6845         kfree(tmp);
6846         return ret;
6847 }
6848
6849 /*
6850  * a bit scary, this does extent mapping from logical file offset to the disk.
6851  * the ugly parts come from merging extents from the disk with the in-ram
6852  * representation.  This gets more complex because of the data=ordered code,
6853  * where the in-ram extents might be locked pending data=ordered completion.
6854  *
6855  * This also copies inline extents directly into the page.
6856  */
6857 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6858                 struct page *page,
6859             size_t pg_offset, u64 start, u64 len,
6860                 int create)
6861 {
6862         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6863         int ret;
6864         int err = 0;
6865         u64 extent_start = 0;
6866         u64 extent_end = 0;
6867         u64 objectid = btrfs_ino(inode);
6868         u32 found_type;
6869         struct btrfs_path *path = NULL;
6870         struct btrfs_root *root = inode->root;
6871         struct btrfs_file_extent_item *item;
6872         struct extent_buffer *leaf;
6873         struct btrfs_key found_key;
6874         struct extent_map *em = NULL;
6875         struct extent_map_tree *em_tree = &inode->extent_tree;
6876         struct extent_io_tree *io_tree = &inode->io_tree;
6877         struct btrfs_trans_handle *trans = NULL;
6878         const bool new_inline = !page || create;
6879
6880 again:
6881         read_lock(&em_tree->lock);
6882         em = lookup_extent_mapping(em_tree, start, len);
6883         if (em)
6884                 em->bdev = fs_info->fs_devices->latest_bdev;
6885         read_unlock(&em_tree->lock);
6886
6887         if (em) {
6888                 if (em->start > start || em->start + em->len <= start)
6889                         free_extent_map(em);
6890                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6891                         free_extent_map(em);
6892                 else
6893                         goto out;
6894         }
6895         em = alloc_extent_map();
6896         if (!em) {
6897                 err = -ENOMEM;
6898                 goto out;
6899         }
6900         em->bdev = fs_info->fs_devices->latest_bdev;
6901         em->start = EXTENT_MAP_HOLE;
6902         em->orig_start = EXTENT_MAP_HOLE;
6903         em->len = (u64)-1;
6904         em->block_len = (u64)-1;
6905
6906         if (!path) {
6907                 path = btrfs_alloc_path();
6908                 if (!path) {
6909                         err = -ENOMEM;
6910                         goto out;
6911                 }
6912                 /*
6913                  * Chances are we'll be called again, so go ahead and do
6914                  * readahead
6915                  */
6916                 path->reada = READA_FORWARD;
6917         }
6918
6919         ret = btrfs_lookup_file_extent(trans, root, path,
6920                                        objectid, start, trans != NULL);
6921         if (ret < 0) {
6922                 err = ret;
6923                 goto out;
6924         }
6925
6926         if (ret != 0) {
6927                 if (path->slots[0] == 0)
6928                         goto not_found;
6929                 path->slots[0]--;
6930         }
6931
6932         leaf = path->nodes[0];
6933         item = btrfs_item_ptr(leaf, path->slots[0],
6934                               struct btrfs_file_extent_item);
6935         /* are we inside the extent that was found? */
6936         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6937         found_type = found_key.type;
6938         if (found_key.objectid != objectid ||
6939             found_type != BTRFS_EXTENT_DATA_KEY) {
6940                 /*
6941                  * If we backup past the first extent we want to move forward
6942                  * and see if there is an extent in front of us, otherwise we'll
6943                  * say there is a hole for our whole search range which can
6944                  * cause problems.
6945                  */
6946                 extent_end = start;
6947                 goto next;
6948         }
6949
6950         found_type = btrfs_file_extent_type(leaf, item);
6951         extent_start = found_key.offset;
6952         if (found_type == BTRFS_FILE_EXTENT_REG ||
6953             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6954                 extent_end = extent_start +
6955                        btrfs_file_extent_num_bytes(leaf, item);
6956
6957                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6958                                                        extent_start);
6959         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6960                 size_t size;
6961                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6962                 extent_end = ALIGN(extent_start + size,
6963                                    fs_info->sectorsize);
6964
6965                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6966                                                       path->slots[0],
6967                                                       extent_start);
6968         }
6969 next:
6970         if (start >= extent_end) {
6971                 path->slots[0]++;
6972                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6973                         ret = btrfs_next_leaf(root, path);
6974                         if (ret < 0) {
6975                                 err = ret;
6976                                 goto out;
6977                         }
6978                         if (ret > 0)
6979                                 goto not_found;
6980                         leaf = path->nodes[0];
6981                 }
6982                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6983                 if (found_key.objectid != objectid ||
6984                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6985                         goto not_found;
6986                 if (start + len <= found_key.offset)
6987                         goto not_found;
6988                 if (start > found_key.offset)
6989                         goto next;
6990                 em->start = start;
6991                 em->orig_start = start;
6992                 em->len = found_key.offset - start;
6993                 goto not_found_em;
6994         }
6995
6996         btrfs_extent_item_to_extent_map(inode, path, item,
6997                         new_inline, em);
6998
6999         if (found_type == BTRFS_FILE_EXTENT_REG ||
7000             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7001                 goto insert;
7002         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7003                 unsigned long ptr;
7004                 char *map;
7005                 size_t size;
7006                 size_t extent_offset;
7007                 size_t copy_size;
7008
7009                 if (new_inline)
7010                         goto out;
7011
7012                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7013                 extent_offset = page_offset(page) + pg_offset - extent_start;
7014                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7015                                   size - extent_offset);
7016                 em->start = extent_start + extent_offset;
7017                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7018                 em->orig_block_len = em->len;
7019                 em->orig_start = em->start;
7020                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7021                 if (create == 0 && !PageUptodate(page)) {
7022                         if (btrfs_file_extent_compression(leaf, item) !=
7023                             BTRFS_COMPRESS_NONE) {
7024                                 ret = uncompress_inline(path, page, pg_offset,
7025                                                         extent_offset, item);
7026                                 if (ret) {
7027                                         err = ret;
7028                                         goto out;
7029                                 }
7030                         } else {
7031                                 map = kmap(page);
7032                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7033                                                    copy_size);
7034                                 if (pg_offset + copy_size < PAGE_SIZE) {
7035                                         memset(map + pg_offset + copy_size, 0,
7036                                                PAGE_SIZE - pg_offset -
7037                                                copy_size);
7038                                 }
7039                                 kunmap(page);
7040                         }
7041                         flush_dcache_page(page);
7042                 } else if (create && PageUptodate(page)) {
7043                         BUG();
7044                         if (!trans) {
7045                                 kunmap(page);
7046                                 free_extent_map(em);
7047                                 em = NULL;
7048
7049                                 btrfs_release_path(path);
7050                                 trans = btrfs_join_transaction(root);
7051
7052                                 if (IS_ERR(trans))
7053                                         return ERR_CAST(trans);
7054                                 goto again;
7055                         }
7056                         map = kmap(page);
7057                         write_extent_buffer(leaf, map + pg_offset, ptr,
7058                                             copy_size);
7059                         kunmap(page);
7060                         btrfs_mark_buffer_dirty(leaf);
7061                 }
7062                 set_extent_uptodate(io_tree, em->start,
7063                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7064                 goto insert;
7065         }
7066 not_found:
7067         em->start = start;
7068         em->orig_start = start;
7069         em->len = len;
7070 not_found_em:
7071         em->block_start = EXTENT_MAP_HOLE;
7072         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7073 insert:
7074         btrfs_release_path(path);
7075         if (em->start > start || extent_map_end(em) <= start) {
7076                 btrfs_err(fs_info,
7077                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7078                           em->start, em->len, start, len);
7079                 err = -EIO;
7080                 goto out;
7081         }
7082
7083         err = 0;
7084         write_lock(&em_tree->lock);
7085         ret = add_extent_mapping(em_tree, em, 0);
7086         /* it is possible that someone inserted the extent into the tree
7087          * while we had the lock dropped.  It is also possible that
7088          * an overlapping map exists in the tree
7089          */
7090         if (ret == -EEXIST) {
7091                 struct extent_map *existing;
7092
7093                 ret = 0;
7094
7095                 existing = search_extent_mapping(em_tree, start, len);
7096                 /*
7097                  * existing will always be non-NULL, since there must be
7098                  * extent causing the -EEXIST.
7099                  */
7100                 if (existing->start == em->start &&
7101                     extent_map_end(existing) >= extent_map_end(em) &&
7102                     em->block_start == existing->block_start) {
7103                         /*
7104                          * The existing extent map already encompasses the
7105                          * entire extent map we tried to add.
7106                          */
7107                         free_extent_map(em);
7108                         em = existing;
7109                         err = 0;
7110
7111                 } else if (start >= extent_map_end(existing) ||
7112                     start <= existing->start) {
7113                         /*
7114                          * The existing extent map is the one nearest to
7115                          * the [start, start + len) range which overlaps
7116                          */
7117                         err = merge_extent_mapping(em_tree, existing,
7118                                                    em, start);
7119                         free_extent_map(existing);
7120                         if (err) {
7121                                 free_extent_map(em);
7122                                 em = NULL;
7123                         }
7124                 } else {
7125                         free_extent_map(em);
7126                         em = existing;
7127                         err = 0;
7128                 }
7129         }
7130         write_unlock(&em_tree->lock);
7131 out:
7132
7133         trace_btrfs_get_extent(root, inode, em);
7134
7135         btrfs_free_path(path);
7136         if (trans) {
7137                 ret = btrfs_end_transaction(trans);
7138                 if (!err)
7139                         err = ret;
7140         }
7141         if (err) {
7142                 free_extent_map(em);
7143                 return ERR_PTR(err);
7144         }
7145         BUG_ON(!em); /* Error is always set */
7146         return em;
7147 }
7148
7149 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7150                 struct page *page,
7151                 size_t pg_offset, u64 start, u64 len,
7152                 int create)
7153 {
7154         struct extent_map *em;
7155         struct extent_map *hole_em = NULL;
7156         u64 range_start = start;
7157         u64 end;
7158         u64 found;
7159         u64 found_end;
7160         int err = 0;
7161
7162         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7163         if (IS_ERR(em))
7164                 return em;
7165         /*
7166          * If our em maps to:
7167          * - a hole or
7168          * - a pre-alloc extent,
7169          * there might actually be delalloc bytes behind it.
7170          */
7171         if (em->block_start != EXTENT_MAP_HOLE &&
7172             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7173                 return em;
7174         else
7175                 hole_em = em;
7176
7177         /* check to see if we've wrapped (len == -1 or similar) */
7178         end = start + len;
7179         if (end < start)
7180                 end = (u64)-1;
7181         else
7182                 end -= 1;
7183
7184         em = NULL;
7185
7186         /* ok, we didn't find anything, lets look for delalloc */
7187         found = count_range_bits(&inode->io_tree, &range_start,
7188                                  end, len, EXTENT_DELALLOC, 1);
7189         found_end = range_start + found;
7190         if (found_end < range_start)
7191                 found_end = (u64)-1;
7192
7193         /*
7194          * we didn't find anything useful, return
7195          * the original results from get_extent()
7196          */
7197         if (range_start > end || found_end <= start) {
7198                 em = hole_em;
7199                 hole_em = NULL;
7200                 goto out;
7201         }
7202
7203         /* adjust the range_start to make sure it doesn't
7204          * go backwards from the start they passed in
7205          */
7206         range_start = max(start, range_start);
7207         found = found_end - range_start;
7208
7209         if (found > 0) {
7210                 u64 hole_start = start;
7211                 u64 hole_len = len;
7212
7213                 em = alloc_extent_map();
7214                 if (!em) {
7215                         err = -ENOMEM;
7216                         goto out;
7217                 }
7218                 /*
7219                  * when btrfs_get_extent can't find anything it
7220                  * returns one huge hole
7221                  *
7222                  * make sure what it found really fits our range, and
7223                  * adjust to make sure it is based on the start from
7224                  * the caller
7225                  */
7226                 if (hole_em) {
7227                         u64 calc_end = extent_map_end(hole_em);
7228
7229                         if (calc_end <= start || (hole_em->start > end)) {
7230                                 free_extent_map(hole_em);
7231                                 hole_em = NULL;
7232                         } else {
7233                                 hole_start = max(hole_em->start, start);
7234                                 hole_len = calc_end - hole_start;
7235                         }
7236                 }
7237                 em->bdev = NULL;
7238                 if (hole_em && range_start > hole_start) {
7239                         /* our hole starts before our delalloc, so we
7240                          * have to return just the parts of the hole
7241                          * that go until  the delalloc starts
7242                          */
7243                         em->len = min(hole_len,
7244                                       range_start - hole_start);
7245                         em->start = hole_start;
7246                         em->orig_start = hole_start;
7247                         /*
7248                          * don't adjust block start at all,
7249                          * it is fixed at EXTENT_MAP_HOLE
7250                          */
7251                         em->block_start = hole_em->block_start;
7252                         em->block_len = hole_len;
7253                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7254                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7255                 } else {
7256                         em->start = range_start;
7257                         em->len = found;
7258                         em->orig_start = range_start;
7259                         em->block_start = EXTENT_MAP_DELALLOC;
7260                         em->block_len = found;
7261                 }
7262         } else if (hole_em) {
7263                 return hole_em;
7264         }
7265 out:
7266
7267         free_extent_map(hole_em);
7268         if (err) {
7269                 free_extent_map(em);
7270                 return ERR_PTR(err);
7271         }
7272         return em;
7273 }
7274
7275 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7276                                                   const u64 start,
7277                                                   const u64 len,
7278                                                   const u64 orig_start,
7279                                                   const u64 block_start,
7280                                                   const u64 block_len,
7281                                                   const u64 orig_block_len,
7282                                                   const u64 ram_bytes,
7283                                                   const int type)
7284 {
7285         struct extent_map *em = NULL;
7286         int ret;
7287
7288         if (type != BTRFS_ORDERED_NOCOW) {
7289                 em = create_io_em(inode, start, len, orig_start,
7290                                   block_start, block_len, orig_block_len,
7291                                   ram_bytes,
7292                                   BTRFS_COMPRESS_NONE, /* compress_type */
7293                                   type);
7294                 if (IS_ERR(em))
7295                         goto out;
7296         }
7297         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7298                                            len, block_len, type);
7299         if (ret) {
7300                 if (em) {
7301                         free_extent_map(em);
7302                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7303                                                 start + len - 1, 0);
7304                 }
7305                 em = ERR_PTR(ret);
7306         }
7307  out:
7308
7309         return em;
7310 }
7311
7312 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7313                                                   u64 start, u64 len)
7314 {
7315         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7316         struct btrfs_root *root = BTRFS_I(inode)->root;
7317         struct extent_map *em;
7318         struct btrfs_key ins;
7319         u64 alloc_hint;
7320         int ret;
7321
7322         alloc_hint = get_extent_allocation_hint(inode, start, len);
7323         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7324                                    0, alloc_hint, &ins, 1, 1);
7325         if (ret)
7326                 return ERR_PTR(ret);
7327
7328         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7329                                      ins.objectid, ins.offset, ins.offset,
7330                                      ins.offset, BTRFS_ORDERED_REGULAR);
7331         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7332         if (IS_ERR(em))
7333                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7334                                            ins.offset, 1);
7335
7336         return em;
7337 }
7338
7339 /*
7340  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7341  * block must be cow'd
7342  */
7343 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7344                               u64 *orig_start, u64 *orig_block_len,
7345                               u64 *ram_bytes)
7346 {
7347         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7348         struct btrfs_path *path;
7349         int ret;
7350         struct extent_buffer *leaf;
7351         struct btrfs_root *root = BTRFS_I(inode)->root;
7352         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7353         struct btrfs_file_extent_item *fi;
7354         struct btrfs_key key;
7355         u64 disk_bytenr;
7356         u64 backref_offset;
7357         u64 extent_end;
7358         u64 num_bytes;
7359         int slot;
7360         int found_type;
7361         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7362
7363         path = btrfs_alloc_path();
7364         if (!path)
7365                 return -ENOMEM;
7366
7367         ret = btrfs_lookup_file_extent(NULL, root, path,
7368                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7369         if (ret < 0)
7370                 goto out;
7371
7372         slot = path->slots[0];
7373         if (ret == 1) {
7374                 if (slot == 0) {
7375                         /* can't find the item, must cow */
7376                         ret = 0;
7377                         goto out;
7378                 }
7379                 slot--;
7380         }
7381         ret = 0;
7382         leaf = path->nodes[0];
7383         btrfs_item_key_to_cpu(leaf, &key, slot);
7384         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7385             key.type != BTRFS_EXTENT_DATA_KEY) {
7386                 /* not our file or wrong item type, must cow */
7387                 goto out;
7388         }
7389
7390         if (key.offset > offset) {
7391                 /* Wrong offset, must cow */
7392                 goto out;
7393         }
7394
7395         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7396         found_type = btrfs_file_extent_type(leaf, fi);
7397         if (found_type != BTRFS_FILE_EXTENT_REG &&
7398             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7399                 /* not a regular extent, must cow */
7400                 goto out;
7401         }
7402
7403         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7404                 goto out;
7405
7406         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7407         if (extent_end <= offset)
7408                 goto out;
7409
7410         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7411         if (disk_bytenr == 0)
7412                 goto out;
7413
7414         if (btrfs_file_extent_compression(leaf, fi) ||
7415             btrfs_file_extent_encryption(leaf, fi) ||
7416             btrfs_file_extent_other_encoding(leaf, fi))
7417                 goto out;
7418
7419         backref_offset = btrfs_file_extent_offset(leaf, fi);
7420
7421         if (orig_start) {
7422                 *orig_start = key.offset - backref_offset;
7423                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7424                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7425         }
7426
7427         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7428                 goto out;
7429
7430         num_bytes = min(offset + *len, extent_end) - offset;
7431         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7432                 u64 range_end;
7433
7434                 range_end = round_up(offset + num_bytes,
7435                                      root->fs_info->sectorsize) - 1;
7436                 ret = test_range_bit(io_tree, offset, range_end,
7437                                      EXTENT_DELALLOC, 0, NULL);
7438                 if (ret) {
7439                         ret = -EAGAIN;
7440                         goto out;
7441                 }
7442         }
7443
7444         btrfs_release_path(path);
7445
7446         /*
7447          * look for other files referencing this extent, if we
7448          * find any we must cow
7449          */
7450
7451         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7452                                     key.offset - backref_offset, disk_bytenr);
7453         if (ret) {
7454                 ret = 0;
7455                 goto out;
7456         }
7457
7458         /*
7459          * adjust disk_bytenr and num_bytes to cover just the bytes
7460          * in this extent we are about to write.  If there
7461          * are any csums in that range we have to cow in order
7462          * to keep the csums correct
7463          */
7464         disk_bytenr += backref_offset;
7465         disk_bytenr += offset - key.offset;
7466         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7467                 goto out;
7468         /*
7469          * all of the above have passed, it is safe to overwrite this extent
7470          * without cow
7471          */
7472         *len = num_bytes;
7473         ret = 1;
7474 out:
7475         btrfs_free_path(path);
7476         return ret;
7477 }
7478
7479 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7480 {
7481         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7482         int found = false;
7483         void **pagep = NULL;
7484         struct page *page = NULL;
7485         unsigned long start_idx;
7486         unsigned long end_idx;
7487
7488         start_idx = start >> PAGE_SHIFT;
7489
7490         /*
7491          * end is the last byte in the last page.  end == start is legal
7492          */
7493         end_idx = end >> PAGE_SHIFT;
7494
7495         rcu_read_lock();
7496
7497         /* Most of the code in this while loop is lifted from
7498          * find_get_page.  It's been modified to begin searching from a
7499          * page and return just the first page found in that range.  If the
7500          * found idx is less than or equal to the end idx then we know that
7501          * a page exists.  If no pages are found or if those pages are
7502          * outside of the range then we're fine (yay!) */
7503         while (page == NULL &&
7504                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7505                 page = radix_tree_deref_slot(pagep);
7506                 if (unlikely(!page))
7507                         break;
7508
7509                 if (radix_tree_exception(page)) {
7510                         if (radix_tree_deref_retry(page)) {
7511                                 page = NULL;
7512                                 continue;
7513                         }
7514                         /*
7515                          * Otherwise, shmem/tmpfs must be storing a swap entry
7516                          * here as an exceptional entry: so return it without
7517                          * attempting to raise page count.
7518                          */
7519                         page = NULL;
7520                         break; /* TODO: Is this relevant for this use case? */
7521                 }
7522
7523                 if (!page_cache_get_speculative(page)) {
7524                         page = NULL;
7525                         continue;
7526                 }
7527
7528                 /*
7529                  * Has the page moved?
7530                  * This is part of the lockless pagecache protocol. See
7531                  * include/linux/pagemap.h for details.
7532                  */
7533                 if (unlikely(page != *pagep)) {
7534                         put_page(page);
7535                         page = NULL;
7536                 }
7537         }
7538
7539         if (page) {
7540                 if (page->index <= end_idx)
7541                         found = true;
7542                 put_page(page);
7543         }
7544
7545         rcu_read_unlock();
7546         return found;
7547 }
7548
7549 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7550                               struct extent_state **cached_state, int writing)
7551 {
7552         struct btrfs_ordered_extent *ordered;
7553         int ret = 0;
7554
7555         while (1) {
7556                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7557                                  cached_state);
7558                 /*
7559                  * We're concerned with the entire range that we're going to be
7560                  * doing DIO to, so we need to make sure there's no ordered
7561                  * extents in this range.
7562                  */
7563                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7564                                                      lockend - lockstart + 1);
7565
7566                 /*
7567                  * We need to make sure there are no buffered pages in this
7568                  * range either, we could have raced between the invalidate in
7569                  * generic_file_direct_write and locking the extent.  The
7570                  * invalidate needs to happen so that reads after a write do not
7571                  * get stale data.
7572                  */
7573                 if (!ordered &&
7574                     (!writing ||
7575                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7576                         break;
7577
7578                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7579                                      cached_state, GFP_NOFS);
7580
7581                 if (ordered) {
7582                         /*
7583                          * If we are doing a DIO read and the ordered extent we
7584                          * found is for a buffered write, we can not wait for it
7585                          * to complete and retry, because if we do so we can
7586                          * deadlock with concurrent buffered writes on page
7587                          * locks. This happens only if our DIO read covers more
7588                          * than one extent map, if at this point has already
7589                          * created an ordered extent for a previous extent map
7590                          * and locked its range in the inode's io tree, and a
7591                          * concurrent write against that previous extent map's
7592                          * range and this range started (we unlock the ranges
7593                          * in the io tree only when the bios complete and
7594                          * buffered writes always lock pages before attempting
7595                          * to lock range in the io tree).
7596                          */
7597                         if (writing ||
7598                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7599                                 btrfs_start_ordered_extent(inode, ordered, 1);
7600                         else
7601                                 ret = -ENOTBLK;
7602                         btrfs_put_ordered_extent(ordered);
7603                 } else {
7604                         /*
7605                          * We could trigger writeback for this range (and wait
7606                          * for it to complete) and then invalidate the pages for
7607                          * this range (through invalidate_inode_pages2_range()),
7608                          * but that can lead us to a deadlock with a concurrent
7609                          * call to readpages() (a buffered read or a defrag call
7610                          * triggered a readahead) on a page lock due to an
7611                          * ordered dio extent we created before but did not have
7612                          * yet a corresponding bio submitted (whence it can not
7613                          * complete), which makes readpages() wait for that
7614                          * ordered extent to complete while holding a lock on
7615                          * that page.
7616                          */
7617                         ret = -ENOTBLK;
7618                 }
7619
7620                 if (ret)
7621                         break;
7622
7623                 cond_resched();
7624         }
7625
7626         return ret;
7627 }
7628
7629 /* The callers of this must take lock_extent() */
7630 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7631                                        u64 orig_start, u64 block_start,
7632                                        u64 block_len, u64 orig_block_len,
7633                                        u64 ram_bytes, int compress_type,
7634                                        int type)
7635 {
7636         struct extent_map_tree *em_tree;
7637         struct extent_map *em;
7638         struct btrfs_root *root = BTRFS_I(inode)->root;
7639         int ret;
7640
7641         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7642                type == BTRFS_ORDERED_COMPRESSED ||
7643                type == BTRFS_ORDERED_NOCOW ||
7644                type == BTRFS_ORDERED_REGULAR);
7645
7646         em_tree = &BTRFS_I(inode)->extent_tree;
7647         em = alloc_extent_map();
7648         if (!em)
7649                 return ERR_PTR(-ENOMEM);
7650
7651         em->start = start;
7652         em->orig_start = orig_start;
7653         em->len = len;
7654         em->block_len = block_len;
7655         em->block_start = block_start;
7656         em->bdev = root->fs_info->fs_devices->latest_bdev;
7657         em->orig_block_len = orig_block_len;
7658         em->ram_bytes = ram_bytes;
7659         em->generation = -1;
7660         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7661         if (type == BTRFS_ORDERED_PREALLOC) {
7662                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7663         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7664                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7665                 em->compress_type = compress_type;
7666         }
7667
7668         do {
7669                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7670                                 em->start + em->len - 1, 0);
7671                 write_lock(&em_tree->lock);
7672                 ret = add_extent_mapping(em_tree, em, 1);
7673                 write_unlock(&em_tree->lock);
7674                 /*
7675                  * The caller has taken lock_extent(), who could race with us
7676                  * to add em?
7677                  */
7678         } while (ret == -EEXIST);
7679
7680         if (ret) {
7681                 free_extent_map(em);
7682                 return ERR_PTR(ret);
7683         }
7684
7685         /* em got 2 refs now, callers needs to do free_extent_map once. */
7686         return em;
7687 }
7688
7689 static void adjust_dio_outstanding_extents(struct inode *inode,
7690                                            struct btrfs_dio_data *dio_data,
7691                                            const u64 len)
7692 {
7693         unsigned num_extents = count_max_extents(len);
7694
7695         /*
7696          * If we have an outstanding_extents count still set then we're
7697          * within our reservation, otherwise we need to adjust our inode
7698          * counter appropriately.
7699          */
7700         if (dio_data->outstanding_extents >= num_extents) {
7701                 dio_data->outstanding_extents -= num_extents;
7702         } else {
7703                 /*
7704                  * If dio write length has been split due to no large enough
7705                  * contiguous space, we need to compensate our inode counter
7706                  * appropriately.
7707                  */
7708                 u64 num_needed = num_extents - dio_data->outstanding_extents;
7709
7710                 spin_lock(&BTRFS_I(inode)->lock);
7711                 BTRFS_I(inode)->outstanding_extents += num_needed;
7712                 spin_unlock(&BTRFS_I(inode)->lock);
7713         }
7714 }
7715
7716 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7717                                    struct buffer_head *bh_result, int create)
7718 {
7719         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7720         struct extent_map *em;
7721         struct extent_state *cached_state = NULL;
7722         struct btrfs_dio_data *dio_data = NULL;
7723         u64 start = iblock << inode->i_blkbits;
7724         u64 lockstart, lockend;
7725         u64 len = bh_result->b_size;
7726         int unlock_bits = EXTENT_LOCKED;
7727         int ret = 0;
7728
7729         if (create)
7730                 unlock_bits |= EXTENT_DIRTY;
7731         else
7732                 len = min_t(u64, len, fs_info->sectorsize);
7733
7734         lockstart = start;
7735         lockend = start + len - 1;
7736
7737         if (current->journal_info) {
7738                 /*
7739                  * Need to pull our outstanding extents and set journal_info to NULL so
7740                  * that anything that needs to check if there's a transaction doesn't get
7741                  * confused.
7742                  */
7743                 dio_data = current->journal_info;
7744                 current->journal_info = NULL;
7745         }
7746
7747         /*
7748          * If this errors out it's because we couldn't invalidate pagecache for
7749          * this range and we need to fallback to buffered.
7750          */
7751         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7752                                create)) {
7753                 ret = -ENOTBLK;
7754                 goto err;
7755         }
7756
7757         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7758         if (IS_ERR(em)) {
7759                 ret = PTR_ERR(em);
7760                 goto unlock_err;
7761         }
7762
7763         /*
7764          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7765          * io.  INLINE is special, and we could probably kludge it in here, but
7766          * it's still buffered so for safety lets just fall back to the generic
7767          * buffered path.
7768          *
7769          * For COMPRESSED we _have_ to read the entire extent in so we can
7770          * decompress it, so there will be buffering required no matter what we
7771          * do, so go ahead and fallback to buffered.
7772          *
7773          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7774          * to buffered IO.  Don't blame me, this is the price we pay for using
7775          * the generic code.
7776          */
7777         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7778             em->block_start == EXTENT_MAP_INLINE) {
7779                 free_extent_map(em);
7780                 ret = -ENOTBLK;
7781                 goto unlock_err;
7782         }
7783
7784         /* Just a good old fashioned hole, return */
7785         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7786                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7787                 free_extent_map(em);
7788                 goto unlock_err;
7789         }
7790
7791         /*
7792          * We don't allocate a new extent in the following cases
7793          *
7794          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7795          * existing extent.
7796          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7797          * just use the extent.
7798          *
7799          */
7800         if (!create) {
7801                 len = min(len, em->len - (start - em->start));
7802                 lockstart = start + len;
7803                 goto unlock;
7804         }
7805
7806         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7807             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7808              em->block_start != EXTENT_MAP_HOLE)) {
7809                 int type;
7810                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7811
7812                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7813                         type = BTRFS_ORDERED_PREALLOC;
7814                 else
7815                         type = BTRFS_ORDERED_NOCOW;
7816                 len = min(len, em->len - (start - em->start));
7817                 block_start = em->block_start + (start - em->start);
7818
7819                 if (can_nocow_extent(inode, start, &len, &orig_start,
7820                                      &orig_block_len, &ram_bytes) == 1 &&
7821                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7822                         struct extent_map *em2;
7823
7824                         em2 = btrfs_create_dio_extent(inode, start, len,
7825                                                       orig_start, block_start,
7826                                                       len, orig_block_len,
7827                                                       ram_bytes, type);
7828                         btrfs_dec_nocow_writers(fs_info, block_start);
7829                         if (type == BTRFS_ORDERED_PREALLOC) {
7830                                 free_extent_map(em);
7831                                 em = em2;
7832                         }
7833                         if (em2 && IS_ERR(em2)) {
7834                                 ret = PTR_ERR(em2);
7835                                 goto unlock_err;
7836                         }
7837                         /*
7838                          * For inode marked NODATACOW or extent marked PREALLOC,
7839                          * use the existing or preallocated extent, so does not
7840                          * need to adjust btrfs_space_info's bytes_may_use.
7841                          */
7842                         btrfs_free_reserved_data_space_noquota(inode,
7843                                         start, len);
7844                         goto unlock;
7845                 }
7846         }
7847
7848         /*
7849          * this will cow the extent, reset the len in case we changed
7850          * it above
7851          */
7852         len = bh_result->b_size;
7853         free_extent_map(em);
7854         em = btrfs_new_extent_direct(inode, start, len);
7855         if (IS_ERR(em)) {
7856                 ret = PTR_ERR(em);
7857                 goto unlock_err;
7858         }
7859         len = min(len, em->len - (start - em->start));
7860 unlock:
7861         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7862                 inode->i_blkbits;
7863         bh_result->b_size = len;
7864         bh_result->b_bdev = em->bdev;
7865         set_buffer_mapped(bh_result);
7866         if (create) {
7867                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7868                         set_buffer_new(bh_result);
7869
7870                 /*
7871                  * Need to update the i_size under the extent lock so buffered
7872                  * readers will get the updated i_size when we unlock.
7873                  */
7874                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7875                         i_size_write(inode, start + len);
7876
7877                 adjust_dio_outstanding_extents(inode, dio_data, len);
7878                 WARN_ON(dio_data->reserve < len);
7879                 dio_data->reserve -= len;
7880                 dio_data->unsubmitted_oe_range_end = start + len;
7881                 current->journal_info = dio_data;
7882         }
7883
7884         /*
7885          * In the case of write we need to clear and unlock the entire range,
7886          * in the case of read we need to unlock only the end area that we
7887          * aren't using if there is any left over space.
7888          */
7889         if (lockstart < lockend) {
7890                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7891                                  lockend, unlock_bits, 1, 0,
7892                                  &cached_state, GFP_NOFS);
7893         } else {
7894                 free_extent_state(cached_state);
7895         }
7896
7897         free_extent_map(em);
7898
7899         return 0;
7900
7901 unlock_err:
7902         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7903                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7904 err:
7905         if (dio_data)
7906                 current->journal_info = dio_data;
7907         /*
7908          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7909          * write less data then expected, so that we don't underflow our inode's
7910          * outstanding extents counter.
7911          */
7912         if (create && dio_data)
7913                 adjust_dio_outstanding_extents(inode, dio_data, len);
7914
7915         return ret;
7916 }
7917
7918 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7919                                         int mirror_num)
7920 {
7921         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7922         int ret;
7923
7924         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7925
7926         bio_get(bio);
7927
7928         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7929         if (ret)
7930                 goto err;
7931
7932         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7933 err:
7934         bio_put(bio);
7935         return ret;
7936 }
7937
7938 static int btrfs_check_dio_repairable(struct inode *inode,
7939                                       struct bio *failed_bio,
7940                                       struct io_failure_record *failrec,
7941                                       int failed_mirror)
7942 {
7943         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7944         int num_copies;
7945
7946         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7947         if (num_copies == 1) {
7948                 /*
7949                  * we only have a single copy of the data, so don't bother with
7950                  * all the retry and error correction code that follows. no
7951                  * matter what the error is, it is very likely to persist.
7952                  */
7953                 btrfs_debug(fs_info,
7954                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7955                         num_copies, failrec->this_mirror, failed_mirror);
7956                 return 0;
7957         }
7958
7959         failrec->failed_mirror = failed_mirror;
7960         failrec->this_mirror++;
7961         if (failrec->this_mirror == failed_mirror)
7962                 failrec->this_mirror++;
7963
7964         if (failrec->this_mirror > num_copies) {
7965                 btrfs_debug(fs_info,
7966                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7967                         num_copies, failrec->this_mirror, failed_mirror);
7968                 return 0;
7969         }
7970
7971         return 1;
7972 }
7973
7974 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7975                         struct page *page, unsigned int pgoff,
7976                         u64 start, u64 end, int failed_mirror,
7977                         bio_end_io_t *repair_endio, void *repair_arg)
7978 {
7979         struct io_failure_record *failrec;
7980         struct bio *bio;
7981         int isector;
7982         int read_mode = 0;
7983         int ret;
7984
7985         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7986
7987         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7988         if (ret)
7989                 return ret;
7990
7991         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7992                                          failed_mirror);
7993         if (!ret) {
7994                 free_io_failure(BTRFS_I(inode), failrec);
7995                 return -EIO;
7996         }
7997
7998         if ((failed_bio->bi_vcnt > 1)
7999                 || (failed_bio->bi_io_vec->bv_len
8000                         > btrfs_inode_sectorsize(inode)))
8001                 read_mode |= REQ_FAILFAST_DEV;
8002
8003         isector = start - btrfs_io_bio(failed_bio)->logical;
8004         isector >>= inode->i_sb->s_blocksize_bits;
8005         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
8006                                 pgoff, isector, repair_endio, repair_arg);
8007         if (!bio) {
8008                 free_io_failure(BTRFS_I(inode), failrec);
8009                 return -EIO;
8010         }
8011         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
8012
8013         btrfs_debug(BTRFS_I(inode)->root->fs_info,
8014                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
8015                     read_mode, failrec->this_mirror, failrec->in_validation);
8016
8017         ret = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
8018         if (ret) {
8019                 free_io_failure(BTRFS_I(inode), failrec);
8020                 bio_put(bio);
8021         }
8022
8023         return ret;
8024 }
8025
8026 struct btrfs_retry_complete {
8027         struct completion done;
8028         struct inode *inode;
8029         u64 start;
8030         int uptodate;
8031 };
8032
8033 static void btrfs_retry_endio_nocsum(struct bio *bio)
8034 {
8035         struct btrfs_retry_complete *done = bio->bi_private;
8036         struct bio_vec *bvec;
8037         int i;
8038
8039         if (bio->bi_status)
8040                 goto end;
8041
8042         ASSERT(bio->bi_vcnt == 1);
8043         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8044
8045         done->uptodate = 1;
8046         bio_for_each_segment_all(bvec, bio, i)
8047                 clean_io_failure(BTRFS_I(done->inode), done->start,
8048                                  bvec->bv_page, 0);
8049 end:
8050         complete(&done->done);
8051         bio_put(bio);
8052 }
8053
8054 static int __btrfs_correct_data_nocsum(struct inode *inode,
8055                                        struct btrfs_io_bio *io_bio)
8056 {
8057         struct btrfs_fs_info *fs_info;
8058         struct bio_vec *bvec;
8059         struct btrfs_retry_complete done;
8060         u64 start;
8061         unsigned int pgoff;
8062         u32 sectorsize;
8063         int nr_sectors;
8064         int i;
8065         int ret;
8066
8067         fs_info = BTRFS_I(inode)->root->fs_info;
8068         sectorsize = fs_info->sectorsize;
8069
8070         start = io_bio->logical;
8071         done.inode = inode;
8072
8073         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8074                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8075                 pgoff = bvec->bv_offset;
8076
8077 next_block_or_try_again:
8078                 done.uptodate = 0;
8079                 done.start = start;
8080                 init_completion(&done.done);
8081
8082                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8083                                 pgoff, start, start + sectorsize - 1,
8084                                 io_bio->mirror_num,
8085                                 btrfs_retry_endio_nocsum, &done);
8086                 if (ret)
8087                         return ret;
8088
8089                 wait_for_completion(&done.done);
8090
8091                 if (!done.uptodate) {
8092                         /* We might have another mirror, so try again */
8093                         goto next_block_or_try_again;
8094                 }
8095
8096                 start += sectorsize;
8097
8098                 nr_sectors--;
8099                 if (nr_sectors) {
8100                         pgoff += sectorsize;
8101                         ASSERT(pgoff < PAGE_SIZE);
8102                         goto next_block_or_try_again;
8103                 }
8104         }
8105
8106         return 0;
8107 }
8108
8109 static void btrfs_retry_endio(struct bio *bio)
8110 {
8111         struct btrfs_retry_complete *done = bio->bi_private;
8112         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8113         struct bio_vec *bvec;
8114         int uptodate;
8115         int ret;
8116         int i;
8117
8118         if (bio->bi_status)
8119                 goto end;
8120
8121         uptodate = 1;
8122
8123         ASSERT(bio->bi_vcnt == 1);
8124         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8125
8126         bio_for_each_segment_all(bvec, bio, i) {
8127                 ret = __readpage_endio_check(done->inode, io_bio, i,
8128                                         bvec->bv_page, bvec->bv_offset,
8129                                         done->start, bvec->bv_len);
8130                 if (!ret)
8131                         clean_io_failure(BTRFS_I(done->inode), done->start,
8132                                         bvec->bv_page, bvec->bv_offset);
8133                 else
8134                         uptodate = 0;
8135         }
8136
8137         done->uptodate = uptodate;
8138 end:
8139         complete(&done->done);
8140         bio_put(bio);
8141 }
8142
8143 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8144                 struct btrfs_io_bio *io_bio, blk_status_t err)
8145 {
8146         struct btrfs_fs_info *fs_info;
8147         struct bio_vec *bvec;
8148         struct btrfs_retry_complete done;
8149         u64 start;
8150         u64 offset = 0;
8151         u32 sectorsize;
8152         int nr_sectors;
8153         unsigned int pgoff;
8154         int csum_pos;
8155         int i;
8156         int ret;
8157
8158         fs_info = BTRFS_I(inode)->root->fs_info;
8159         sectorsize = fs_info->sectorsize;
8160
8161         err = 0;
8162         start = io_bio->logical;
8163         done.inode = inode;
8164
8165         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8166                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8167
8168                 pgoff = bvec->bv_offset;
8169 next_block:
8170                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8171                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8172                                         bvec->bv_page, pgoff, start,
8173                                         sectorsize);
8174                 if (likely(!ret))
8175                         goto next;
8176 try_again:
8177                 done.uptodate = 0;
8178                 done.start = start;
8179                 init_completion(&done.done);
8180
8181                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8182                                 pgoff, start, start + sectorsize - 1,
8183                                 io_bio->mirror_num,
8184                                 btrfs_retry_endio, &done);
8185                 if (ret) {
8186                         err = errno_to_blk_status(ret);
8187                         goto next;
8188                 }
8189
8190                 wait_for_completion(&done.done);
8191
8192                 if (!done.uptodate) {
8193                         /* We might have another mirror, so try again */
8194                         goto try_again;
8195                 }
8196 next:
8197                 offset += sectorsize;
8198                 start += sectorsize;
8199
8200                 ASSERT(nr_sectors);
8201
8202                 nr_sectors--;
8203                 if (nr_sectors) {
8204                         pgoff += sectorsize;
8205                         ASSERT(pgoff < PAGE_SIZE);
8206                         goto next_block;
8207                 }
8208         }
8209
8210         return err;
8211 }
8212
8213 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8214                 struct btrfs_io_bio *io_bio, blk_status_t err)
8215 {
8216         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8217
8218         if (skip_csum) {
8219                 if (unlikely(err))
8220                         return __btrfs_correct_data_nocsum(inode, io_bio);
8221                 else
8222                         return 0;
8223         } else {
8224                 return __btrfs_subio_endio_read(inode, io_bio, err);
8225         }
8226 }
8227
8228 static void btrfs_endio_direct_read(struct bio *bio)
8229 {
8230         struct btrfs_dio_private *dip = bio->bi_private;
8231         struct inode *inode = dip->inode;
8232         struct bio *dio_bio;
8233         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8234         blk_status_t err = bio->bi_status;
8235
8236         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8237                 err = btrfs_subio_endio_read(inode, io_bio, err);
8238
8239         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8240                       dip->logical_offset + dip->bytes - 1);
8241         dio_bio = dip->dio_bio;
8242
8243         kfree(dip);
8244
8245         dio_bio->bi_status = bio->bi_status;
8246         dio_end_io(dio_bio);
8247
8248         if (io_bio->end_io)
8249                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8250         bio_put(bio);
8251 }
8252
8253 static void __endio_write_update_ordered(struct inode *inode,
8254                                          const u64 offset, const u64 bytes,
8255                                          const bool uptodate)
8256 {
8257         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8258         struct btrfs_ordered_extent *ordered = NULL;
8259         struct btrfs_workqueue *wq;
8260         btrfs_work_func_t func;
8261         u64 ordered_offset = offset;
8262         u64 ordered_bytes = bytes;
8263         int ret;
8264
8265         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8266                 wq = fs_info->endio_freespace_worker;
8267                 func = btrfs_freespace_write_helper;
8268         } else {
8269                 wq = fs_info->endio_write_workers;
8270                 func = btrfs_endio_write_helper;
8271         }
8272
8273 again:
8274         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8275                                                    &ordered_offset,
8276                                                    ordered_bytes,
8277                                                    uptodate);
8278         if (!ret)
8279                 goto out_test;
8280
8281         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8282         btrfs_queue_work(wq, &ordered->work);
8283 out_test:
8284         /*
8285          * our bio might span multiple ordered extents.  If we haven't
8286          * completed the accounting for the whole dio, go back and try again
8287          */
8288         if (ordered_offset < offset + bytes) {
8289                 ordered_bytes = offset + bytes - ordered_offset;
8290                 ordered = NULL;
8291                 goto again;
8292         }
8293 }
8294
8295 static void btrfs_endio_direct_write(struct bio *bio)
8296 {
8297         struct btrfs_dio_private *dip = bio->bi_private;
8298         struct bio *dio_bio = dip->dio_bio;
8299
8300         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8301                                      dip->bytes, !bio->bi_status);
8302
8303         kfree(dip);
8304
8305         dio_bio->bi_status = bio->bi_status;
8306         dio_end_io(dio_bio);
8307         bio_put(bio);
8308 }
8309
8310 static blk_status_t __btrfs_submit_bio_start_direct_io(struct inode *inode,
8311                                     struct bio *bio, int mirror_num,
8312                                     unsigned long bio_flags, u64 offset)
8313 {
8314         blk_status_t ret;
8315         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8316         BUG_ON(ret); /* -ENOMEM */
8317         return 0;
8318 }
8319
8320 static void btrfs_end_dio_bio(struct bio *bio)
8321 {
8322         struct btrfs_dio_private *dip = bio->bi_private;
8323         blk_status_t err = bio->bi_status;
8324
8325         if (err)
8326                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8327                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8328                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8329                            bio->bi_opf,
8330                            (unsigned long long)bio->bi_iter.bi_sector,
8331                            bio->bi_iter.bi_size, err);
8332
8333         if (dip->subio_endio)
8334                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8335
8336         if (err) {
8337                 dip->errors = 1;
8338
8339                 /*
8340                  * before atomic variable goto zero, we must make sure
8341                  * dip->errors is perceived to be set.
8342                  */
8343                 smp_mb__before_atomic();
8344         }
8345
8346         /* if there are more bios still pending for this dio, just exit */
8347         if (!atomic_dec_and_test(&dip->pending_bios))
8348                 goto out;
8349
8350         if (dip->errors) {
8351                 bio_io_error(dip->orig_bio);
8352         } else {
8353                 dip->dio_bio->bi_status = 0;
8354                 bio_endio(dip->orig_bio);
8355         }
8356 out:
8357         bio_put(bio);
8358 }
8359
8360 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8361                                        u64 first_sector, gfp_t gfp_flags)
8362 {
8363         struct bio *bio;
8364         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8365         if (bio)
8366                 bio_associate_current(bio);
8367         return bio;
8368 }
8369
8370 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8371                                                  struct btrfs_dio_private *dip,
8372                                                  struct bio *bio,
8373                                                  u64 file_offset)
8374 {
8375         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8376         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8377         blk_status_t ret;
8378
8379         /*
8380          * We load all the csum data we need when we submit
8381          * the first bio to reduce the csum tree search and
8382          * contention.
8383          */
8384         if (dip->logical_offset == file_offset) {
8385                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8386                                                 file_offset);
8387                 if (ret)
8388                         return ret;
8389         }
8390
8391         if (bio == dip->orig_bio)
8392                 return 0;
8393
8394         file_offset -= dip->logical_offset;
8395         file_offset >>= inode->i_sb->s_blocksize_bits;
8396         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8397
8398         return 0;
8399 }
8400
8401 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8402                                          u64 file_offset, int skip_sum,
8403                                          int async_submit)
8404 {
8405         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8406         struct btrfs_dio_private *dip = bio->bi_private;
8407         bool write = bio_op(bio) == REQ_OP_WRITE;
8408         blk_status_t ret;
8409
8410         if (async_submit)
8411                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8412
8413         bio_get(bio);
8414
8415         if (!write) {
8416                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8417                 if (ret)
8418                         goto err;
8419         }
8420
8421         if (skip_sum)
8422                 goto map;
8423
8424         if (write && async_submit) {
8425                 ret = btrfs_wq_submit_bio(fs_info, inode, bio, 0, 0,
8426                                           file_offset,
8427                                           __btrfs_submit_bio_start_direct_io,
8428                                           __btrfs_submit_bio_done);
8429                 goto err;
8430         } else if (write) {
8431                 /*
8432                  * If we aren't doing async submit, calculate the csum of the
8433                  * bio now.
8434                  */
8435                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8436                 if (ret)
8437                         goto err;
8438         } else {
8439                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8440                                                      file_offset);
8441                 if (ret)
8442                         goto err;
8443         }
8444 map:
8445         ret = btrfs_map_bio(fs_info, bio, 0, async_submit);
8446 err:
8447         bio_put(bio);
8448         return ret;
8449 }
8450
8451 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip,
8452                                     int skip_sum)
8453 {
8454         struct inode *inode = dip->inode;
8455         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8456         struct btrfs_root *root = BTRFS_I(inode)->root;
8457         struct bio *bio;
8458         struct bio *orig_bio = dip->orig_bio;
8459         struct bio_vec *bvec;
8460         u64 start_sector = orig_bio->bi_iter.bi_sector;
8461         u64 file_offset = dip->logical_offset;
8462         u64 submit_len = 0;
8463         u64 map_length;
8464         u32 blocksize = fs_info->sectorsize;
8465         int async_submit = 0;
8466         int nr_sectors;
8467         int ret;
8468         int i, j;
8469
8470         map_length = orig_bio->bi_iter.bi_size;
8471         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8472                               &map_length, NULL, 0);
8473         if (ret)
8474                 return -EIO;
8475
8476         if (map_length >= orig_bio->bi_iter.bi_size) {
8477                 bio = orig_bio;
8478                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8479                 goto submit;
8480         }
8481
8482         /* async crcs make it difficult to collect full stripe writes. */
8483         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8484                 async_submit = 0;
8485         else
8486                 async_submit = 1;
8487
8488         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8489         if (!bio)
8490                 return -ENOMEM;
8491
8492         bio->bi_opf = orig_bio->bi_opf;
8493         bio->bi_private = dip;
8494         bio->bi_end_io = btrfs_end_dio_bio;
8495         btrfs_io_bio(bio)->logical = file_offset;
8496         atomic_inc(&dip->pending_bios);
8497
8498         bio_for_each_segment_all(bvec, orig_bio, j) {
8499                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8500                 i = 0;
8501 next_block:
8502                 if (unlikely(map_length < submit_len + blocksize ||
8503                     bio_add_page(bio, bvec->bv_page, blocksize,
8504                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8505                         /*
8506                          * inc the count before we submit the bio so
8507                          * we know the end IO handler won't happen before
8508                          * we inc the count. Otherwise, the dip might get freed
8509                          * before we're done setting it up
8510                          */
8511                         atomic_inc(&dip->pending_bios);
8512                         ret = __btrfs_submit_dio_bio(bio, inode,
8513                                                      file_offset, skip_sum,
8514                                                      async_submit);
8515                         if (ret) {
8516                                 bio_put(bio);
8517                                 atomic_dec(&dip->pending_bios);
8518                                 goto out_err;
8519                         }
8520
8521                         start_sector += submit_len >> 9;
8522                         file_offset += submit_len;
8523
8524                         submit_len = 0;
8525
8526                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8527                                                   start_sector, GFP_NOFS);
8528                         if (!bio)
8529                                 goto out_err;
8530                         bio->bi_opf = orig_bio->bi_opf;
8531                         bio->bi_private = dip;
8532                         bio->bi_end_io = btrfs_end_dio_bio;
8533                         btrfs_io_bio(bio)->logical = file_offset;
8534
8535                         map_length = orig_bio->bi_iter.bi_size;
8536                         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8537                                               start_sector << 9,
8538                                               &map_length, NULL, 0);
8539                         if (ret) {
8540                                 bio_put(bio);
8541                                 goto out_err;
8542                         }
8543
8544                         goto next_block;
8545                 } else {
8546                         submit_len += blocksize;
8547                         if (--nr_sectors) {
8548                                 i++;
8549                                 goto next_block;
8550                         }
8551                 }
8552         }
8553
8554 submit:
8555         ret = __btrfs_submit_dio_bio(bio, inode, file_offset, skip_sum,
8556                                      async_submit);
8557         if (!ret)
8558                 return 0;
8559
8560         bio_put(bio);
8561 out_err:
8562         dip->errors = 1;
8563         /*
8564          * before atomic variable goto zero, we must
8565          * make sure dip->errors is perceived to be set.
8566          */
8567         smp_mb__before_atomic();
8568         if (atomic_dec_and_test(&dip->pending_bios))
8569                 bio_io_error(dip->orig_bio);
8570
8571         /* bio_end_io() will handle error, so we needn't return it */
8572         return 0;
8573 }
8574
8575 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8576                                 loff_t file_offset)
8577 {
8578         struct btrfs_dio_private *dip = NULL;
8579         struct bio *io_bio = NULL;
8580         struct btrfs_io_bio *btrfs_bio;
8581         int skip_sum;
8582         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8583         int ret = 0;
8584
8585         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8586
8587         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8588         if (!io_bio) {
8589                 ret = -ENOMEM;
8590                 goto free_ordered;
8591         }
8592
8593         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8594         if (!dip) {
8595                 ret = -ENOMEM;
8596                 goto free_ordered;
8597         }
8598
8599         dip->private = dio_bio->bi_private;
8600         dip->inode = inode;
8601         dip->logical_offset = file_offset;
8602         dip->bytes = dio_bio->bi_iter.bi_size;
8603         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8604         io_bio->bi_private = dip;
8605         dip->orig_bio = io_bio;
8606         dip->dio_bio = dio_bio;
8607         atomic_set(&dip->pending_bios, 0);
8608         btrfs_bio = btrfs_io_bio(io_bio);
8609         btrfs_bio->logical = file_offset;
8610
8611         if (write) {
8612                 io_bio->bi_end_io = btrfs_endio_direct_write;
8613         } else {
8614                 io_bio->bi_end_io = btrfs_endio_direct_read;
8615                 dip->subio_endio = btrfs_subio_endio_read;
8616         }
8617
8618         /*
8619          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8620          * even if we fail to submit a bio, because in such case we do the
8621          * corresponding error handling below and it must not be done a second
8622          * time by btrfs_direct_IO().
8623          */
8624         if (write) {
8625                 struct btrfs_dio_data *dio_data = current->journal_info;
8626
8627                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8628                         dip->bytes;
8629                 dio_data->unsubmitted_oe_range_start =
8630                         dio_data->unsubmitted_oe_range_end;
8631         }
8632
8633         ret = btrfs_submit_direct_hook(dip, skip_sum);
8634         if (!ret)
8635                 return;
8636
8637         if (btrfs_bio->end_io)
8638                 btrfs_bio->end_io(btrfs_bio, ret);
8639
8640 free_ordered:
8641         /*
8642          * If we arrived here it means either we failed to submit the dip
8643          * or we either failed to clone the dio_bio or failed to allocate the
8644          * dip. If we cloned the dio_bio and allocated the dip, we can just
8645          * call bio_endio against our io_bio so that we get proper resource
8646          * cleanup if we fail to submit the dip, otherwise, we must do the
8647          * same as btrfs_endio_direct_[write|read] because we can't call these
8648          * callbacks - they require an allocated dip and a clone of dio_bio.
8649          */
8650         if (io_bio && dip) {
8651                 io_bio->bi_status = BLK_STS_IOERR;
8652                 bio_endio(io_bio);
8653                 /*
8654                  * The end io callbacks free our dip, do the final put on io_bio
8655                  * and all the cleanup and final put for dio_bio (through
8656                  * dio_end_io()).
8657                  */
8658                 dip = NULL;
8659                 io_bio = NULL;
8660         } else {
8661                 if (write)
8662                         __endio_write_update_ordered(inode,
8663                                                 file_offset,
8664                                                 dio_bio->bi_iter.bi_size,
8665                                                 false);
8666                 else
8667                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8668                               file_offset + dio_bio->bi_iter.bi_size - 1);
8669
8670                 dio_bio->bi_status = BLK_STS_IOERR;
8671                 /*
8672                  * Releases and cleans up our dio_bio, no need to bio_put()
8673                  * nor bio_endio()/bio_io_error() against dio_bio.
8674                  */
8675                 dio_end_io(dio_bio);
8676         }
8677         if (io_bio)
8678                 bio_put(io_bio);
8679         kfree(dip);
8680 }
8681
8682 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8683                                struct kiocb *iocb,
8684                                const struct iov_iter *iter, loff_t offset)
8685 {
8686         int seg;
8687         int i;
8688         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8689         ssize_t retval = -EINVAL;
8690
8691         if (offset & blocksize_mask)
8692                 goto out;
8693
8694         if (iov_iter_alignment(iter) & blocksize_mask)
8695                 goto out;
8696
8697         /* If this is a write we don't need to check anymore */
8698         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8699                 return 0;
8700         /*
8701          * Check to make sure we don't have duplicate iov_base's in this
8702          * iovec, if so return EINVAL, otherwise we'll get csum errors
8703          * when reading back.
8704          */
8705         for (seg = 0; seg < iter->nr_segs; seg++) {
8706                 for (i = seg + 1; i < iter->nr_segs; i++) {
8707                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8708                                 goto out;
8709                 }
8710         }
8711         retval = 0;
8712 out:
8713         return retval;
8714 }
8715
8716 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8717 {
8718         struct file *file = iocb->ki_filp;
8719         struct inode *inode = file->f_mapping->host;
8720         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8721         struct btrfs_dio_data dio_data = { 0 };
8722         loff_t offset = iocb->ki_pos;
8723         size_t count = 0;
8724         int flags = 0;
8725         bool wakeup = true;
8726         bool relock = false;
8727         ssize_t ret;
8728
8729         if (check_direct_IO(fs_info, iocb, iter, offset))
8730                 return 0;
8731
8732         inode_dio_begin(inode);
8733         smp_mb__after_atomic();
8734
8735         /*
8736          * The generic stuff only does filemap_write_and_wait_range, which
8737          * isn't enough if we've written compressed pages to this area, so
8738          * we need to flush the dirty pages again to make absolutely sure
8739          * that any outstanding dirty pages are on disk.
8740          */
8741         count = iov_iter_count(iter);
8742         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8743                      &BTRFS_I(inode)->runtime_flags))
8744                 filemap_fdatawrite_range(inode->i_mapping, offset,
8745                                          offset + count - 1);
8746
8747         if (iov_iter_rw(iter) == WRITE) {
8748                 /*
8749                  * If the write DIO is beyond the EOF, we need update
8750                  * the isize, but it is protected by i_mutex. So we can
8751                  * not unlock the i_mutex at this case.
8752                  */
8753                 if (offset + count <= inode->i_size) {
8754                         dio_data.overwrite = 1;
8755                         inode_unlock(inode);
8756                         relock = true;
8757                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8758                         ret = -EAGAIN;
8759                         goto out;
8760                 }
8761                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8762                 if (ret)
8763                         goto out;
8764                 dio_data.outstanding_extents = count_max_extents(count);
8765
8766                 /*
8767                  * We need to know how many extents we reserved so that we can
8768                  * do the accounting properly if we go over the number we
8769                  * originally calculated.  Abuse current->journal_info for this.
8770                  */
8771                 dio_data.reserve = round_up(count,
8772                                             fs_info->sectorsize);
8773                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8774                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8775                 current->journal_info = &dio_data;
8776                 down_read(&BTRFS_I(inode)->dio_sem);
8777         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8778                                      &BTRFS_I(inode)->runtime_flags)) {
8779                 inode_dio_end(inode);
8780                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8781                 wakeup = false;
8782         }
8783
8784         ret = __blockdev_direct_IO(iocb, inode,
8785                                    fs_info->fs_devices->latest_bdev,
8786                                    iter, btrfs_get_blocks_direct, NULL,
8787                                    btrfs_submit_direct, flags);
8788         if (iov_iter_rw(iter) == WRITE) {
8789                 up_read(&BTRFS_I(inode)->dio_sem);
8790                 current->journal_info = NULL;
8791                 if (ret < 0 && ret != -EIOCBQUEUED) {
8792                         if (dio_data.reserve)
8793                                 btrfs_delalloc_release_space(inode, offset,
8794                                                              dio_data.reserve);
8795                         /*
8796                          * On error we might have left some ordered extents
8797                          * without submitting corresponding bios for them, so
8798                          * cleanup them up to avoid other tasks getting them
8799                          * and waiting for them to complete forever.
8800                          */
8801                         if (dio_data.unsubmitted_oe_range_start <
8802                             dio_data.unsubmitted_oe_range_end)
8803                                 __endio_write_update_ordered(inode,
8804                                         dio_data.unsubmitted_oe_range_start,
8805                                         dio_data.unsubmitted_oe_range_end -
8806                                         dio_data.unsubmitted_oe_range_start,
8807                                         false);
8808                 } else if (ret >= 0 && (size_t)ret < count)
8809                         btrfs_delalloc_release_space(inode, offset,
8810                                                      count - (size_t)ret);
8811         }
8812 out:
8813         if (wakeup)
8814                 inode_dio_end(inode);
8815         if (relock)
8816                 inode_lock(inode);
8817
8818         return ret;
8819 }
8820
8821 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8822
8823 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8824                 __u64 start, __u64 len)
8825 {
8826         int     ret;
8827
8828         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8829         if (ret)
8830                 return ret;
8831
8832         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8833 }
8834
8835 int btrfs_readpage(struct file *file, struct page *page)
8836 {
8837         struct extent_io_tree *tree;
8838         tree = &BTRFS_I(page->mapping->host)->io_tree;
8839         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8840 }
8841
8842 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8843 {
8844         struct extent_io_tree *tree;
8845         struct inode *inode = page->mapping->host;
8846         int ret;
8847
8848         if (current->flags & PF_MEMALLOC) {
8849                 redirty_page_for_writepage(wbc, page);
8850                 unlock_page(page);
8851                 return 0;
8852         }
8853
8854         /*
8855          * If we are under memory pressure we will call this directly from the
8856          * VM, we need to make sure we have the inode referenced for the ordered
8857          * extent.  If not just return like we didn't do anything.
8858          */
8859         if (!igrab(inode)) {
8860                 redirty_page_for_writepage(wbc, page);
8861                 return AOP_WRITEPAGE_ACTIVATE;
8862         }
8863         tree = &BTRFS_I(page->mapping->host)->io_tree;
8864         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8865         btrfs_add_delayed_iput(inode);
8866         return ret;
8867 }
8868
8869 static int btrfs_writepages(struct address_space *mapping,
8870                             struct writeback_control *wbc)
8871 {
8872         struct extent_io_tree *tree;
8873
8874         tree = &BTRFS_I(mapping->host)->io_tree;
8875         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8876 }
8877
8878 static int
8879 btrfs_readpages(struct file *file, struct address_space *mapping,
8880                 struct list_head *pages, unsigned nr_pages)
8881 {
8882         struct extent_io_tree *tree;
8883         tree = &BTRFS_I(mapping->host)->io_tree;
8884         return extent_readpages(tree, mapping, pages, nr_pages,
8885                                 btrfs_get_extent);
8886 }
8887 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8888 {
8889         struct extent_io_tree *tree;
8890         struct extent_map_tree *map;
8891         int ret;
8892
8893         tree = &BTRFS_I(page->mapping->host)->io_tree;
8894         map = &BTRFS_I(page->mapping->host)->extent_tree;
8895         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8896         if (ret == 1) {
8897                 ClearPagePrivate(page);
8898                 set_page_private(page, 0);
8899                 put_page(page);
8900         }
8901         return ret;
8902 }
8903
8904 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8905 {
8906         if (PageWriteback(page) || PageDirty(page))
8907                 return 0;
8908         return __btrfs_releasepage(page, gfp_flags);
8909 }
8910
8911 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8912                                  unsigned int length)
8913 {
8914         struct inode *inode = page->mapping->host;
8915         struct extent_io_tree *tree;
8916         struct btrfs_ordered_extent *ordered;
8917         struct extent_state *cached_state = NULL;
8918         u64 page_start = page_offset(page);
8919         u64 page_end = page_start + PAGE_SIZE - 1;
8920         u64 start;
8921         u64 end;
8922         int inode_evicting = inode->i_state & I_FREEING;
8923
8924         /*
8925          * we have the page locked, so new writeback can't start,
8926          * and the dirty bit won't be cleared while we are here.
8927          *
8928          * Wait for IO on this page so that we can safely clear
8929          * the PagePrivate2 bit and do ordered accounting
8930          */
8931         wait_on_page_writeback(page);
8932
8933         tree = &BTRFS_I(inode)->io_tree;
8934         if (offset) {
8935                 btrfs_releasepage(page, GFP_NOFS);
8936                 return;
8937         }
8938
8939         if (!inode_evicting)
8940                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8941 again:
8942         start = page_start;
8943         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8944                                         page_end - start + 1);
8945         if (ordered) {
8946                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8947                 /*
8948                  * IO on this page will never be started, so we need
8949                  * to account for any ordered extents now
8950                  */
8951                 if (!inode_evicting)
8952                         clear_extent_bit(tree, start, end,
8953                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8954                                          EXTENT_DELALLOC_NEW |
8955                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8956                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8957                                          GFP_NOFS);
8958                 /*
8959                  * whoever cleared the private bit is responsible
8960                  * for the finish_ordered_io
8961                  */
8962                 if (TestClearPagePrivate2(page)) {
8963                         struct btrfs_ordered_inode_tree *tree;
8964                         u64 new_len;
8965
8966                         tree = &BTRFS_I(inode)->ordered_tree;
8967
8968                         spin_lock_irq(&tree->lock);
8969                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8970                         new_len = start - ordered->file_offset;
8971                         if (new_len < ordered->truncated_len)
8972                                 ordered->truncated_len = new_len;
8973                         spin_unlock_irq(&tree->lock);
8974
8975                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8976                                                            start,
8977                                                            end - start + 1, 1))
8978                                 btrfs_finish_ordered_io(ordered);
8979                 }
8980                 btrfs_put_ordered_extent(ordered);
8981                 if (!inode_evicting) {
8982                         cached_state = NULL;
8983                         lock_extent_bits(tree, start, end,
8984                                          &cached_state);
8985                 }
8986
8987                 start = end + 1;
8988                 if (start < page_end)
8989                         goto again;
8990         }
8991
8992         /*
8993          * Qgroup reserved space handler
8994          * Page here will be either
8995          * 1) Already written to disk
8996          *    In this case, its reserved space is released from data rsv map
8997          *    and will be freed by delayed_ref handler finally.
8998          *    So even we call qgroup_free_data(), it won't decrease reserved
8999          *    space.
9000          * 2) Not written to disk
9001          *    This means the reserved space should be freed here. However,
9002          *    if a truncate invalidates the page (by clearing PageDirty)
9003          *    and the page is accounted for while allocating extent
9004          *    in btrfs_check_data_free_space() we let delayed_ref to
9005          *    free the entire extent.
9006          */
9007         if (PageDirty(page))
9008                 btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
9009         if (!inode_evicting) {
9010                 clear_extent_bit(tree, page_start, page_end,
9011                                  EXTENT_LOCKED | EXTENT_DIRTY |
9012                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
9013                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
9014                                  &cached_state, GFP_NOFS);
9015
9016                 __btrfs_releasepage(page, GFP_NOFS);
9017         }
9018
9019         ClearPageChecked(page);
9020         if (PagePrivate(page)) {
9021                 ClearPagePrivate(page);
9022                 set_page_private(page, 0);
9023                 put_page(page);
9024         }
9025 }
9026
9027 /*
9028  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
9029  * called from a page fault handler when a page is first dirtied. Hence we must
9030  * be careful to check for EOF conditions here. We set the page up correctly
9031  * for a written page which means we get ENOSPC checking when writing into
9032  * holes and correct delalloc and unwritten extent mapping on filesystems that
9033  * support these features.
9034  *
9035  * We are not allowed to take the i_mutex here so we have to play games to
9036  * protect against truncate races as the page could now be beyond EOF.  Because
9037  * vmtruncate() writes the inode size before removing pages, once we have the
9038  * page lock we can determine safely if the page is beyond EOF. If it is not
9039  * beyond EOF, then the page is guaranteed safe against truncation until we
9040  * unlock the page.
9041  */
9042 int btrfs_page_mkwrite(struct vm_fault *vmf)
9043 {
9044         struct page *page = vmf->page;
9045         struct inode *inode = file_inode(vmf->vma->vm_file);
9046         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9047         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
9048         struct btrfs_ordered_extent *ordered;
9049         struct extent_state *cached_state = NULL;
9050         char *kaddr;
9051         unsigned long zero_start;
9052         loff_t size;
9053         int ret;
9054         int reserved = 0;
9055         u64 reserved_space;
9056         u64 page_start;
9057         u64 page_end;
9058         u64 end;
9059
9060         reserved_space = PAGE_SIZE;
9061
9062         sb_start_pagefault(inode->i_sb);
9063         page_start = page_offset(page);
9064         page_end = page_start + PAGE_SIZE - 1;
9065         end = page_end;
9066
9067         /*
9068          * Reserving delalloc space after obtaining the page lock can lead to
9069          * deadlock. For example, if a dirty page is locked by this function
9070          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9071          * dirty page write out, then the btrfs_writepage() function could
9072          * end up waiting indefinitely to get a lock on the page currently
9073          * being processed by btrfs_page_mkwrite() function.
9074          */
9075         ret = btrfs_delalloc_reserve_space(inode, page_start,
9076                                            reserved_space);
9077         if (!ret) {
9078                 ret = file_update_time(vmf->vma->vm_file);
9079                 reserved = 1;
9080         }
9081         if (ret) {
9082                 if (ret == -ENOMEM)
9083                         ret = VM_FAULT_OOM;
9084                 else /* -ENOSPC, -EIO, etc */
9085                         ret = VM_FAULT_SIGBUS;
9086                 if (reserved)
9087                         goto out;
9088                 goto out_noreserve;
9089         }
9090
9091         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9092 again:
9093         lock_page(page);
9094         size = i_size_read(inode);
9095
9096         if ((page->mapping != inode->i_mapping) ||
9097             (page_start >= size)) {
9098                 /* page got truncated out from underneath us */
9099                 goto out_unlock;
9100         }
9101         wait_on_page_writeback(page);
9102
9103         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9104         set_page_extent_mapped(page);
9105
9106         /*
9107          * we can't set the delalloc bits if there are pending ordered
9108          * extents.  Drop our locks and wait for them to finish
9109          */
9110         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9111                         PAGE_SIZE);
9112         if (ordered) {
9113                 unlock_extent_cached(io_tree, page_start, page_end,
9114                                      &cached_state, GFP_NOFS);
9115                 unlock_page(page);
9116                 btrfs_start_ordered_extent(inode, ordered, 1);
9117                 btrfs_put_ordered_extent(ordered);
9118                 goto again;
9119         }
9120
9121         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9122                 reserved_space = round_up(size - page_start,
9123                                           fs_info->sectorsize);
9124                 if (reserved_space < PAGE_SIZE) {
9125                         end = page_start + reserved_space - 1;
9126                         spin_lock(&BTRFS_I(inode)->lock);
9127                         BTRFS_I(inode)->outstanding_extents++;
9128                         spin_unlock(&BTRFS_I(inode)->lock);
9129                         btrfs_delalloc_release_space(inode, page_start,
9130                                                 PAGE_SIZE - reserved_space);
9131                 }
9132         }
9133
9134         /*
9135          * page_mkwrite gets called when the page is firstly dirtied after it's
9136          * faulted in, but write(2) could also dirty a page and set delalloc
9137          * bits, thus in this case for space account reason, we still need to
9138          * clear any delalloc bits within this page range since we have to
9139          * reserve data&meta space before lock_page() (see above comments).
9140          */
9141         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9142                           EXTENT_DIRTY | EXTENT_DELALLOC |
9143                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9144                           0, 0, &cached_state, GFP_NOFS);
9145
9146         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9147                                         &cached_state, 0);
9148         if (ret) {
9149                 unlock_extent_cached(io_tree, page_start, page_end,
9150                                      &cached_state, GFP_NOFS);
9151                 ret = VM_FAULT_SIGBUS;
9152                 goto out_unlock;
9153         }
9154         ret = 0;
9155
9156         /* page is wholly or partially inside EOF */
9157         if (page_start + PAGE_SIZE > size)
9158                 zero_start = size & ~PAGE_MASK;
9159         else
9160                 zero_start = PAGE_SIZE;
9161
9162         if (zero_start != PAGE_SIZE) {
9163                 kaddr = kmap(page);
9164                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9165                 flush_dcache_page(page);
9166                 kunmap(page);
9167         }
9168         ClearPageChecked(page);
9169         set_page_dirty(page);
9170         SetPageUptodate(page);
9171
9172         BTRFS_I(inode)->last_trans = fs_info->generation;
9173         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9174         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9175
9176         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9177
9178 out_unlock:
9179         if (!ret) {
9180                 sb_end_pagefault(inode->i_sb);
9181                 return VM_FAULT_LOCKED;
9182         }
9183         unlock_page(page);
9184 out:
9185         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9186 out_noreserve:
9187         sb_end_pagefault(inode->i_sb);
9188         return ret;
9189 }
9190
9191 static int btrfs_truncate(struct inode *inode)
9192 {
9193         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9194         struct btrfs_root *root = BTRFS_I(inode)->root;
9195         struct btrfs_block_rsv *rsv;
9196         int ret = 0;
9197         int err = 0;
9198         struct btrfs_trans_handle *trans;
9199         u64 mask = fs_info->sectorsize - 1;
9200         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9201
9202         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9203                                        (u64)-1);
9204         if (ret)
9205                 return ret;
9206
9207         /*
9208          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9209          * 3 things going on here
9210          *
9211          * 1) We need to reserve space for our orphan item and the space to
9212          * delete our orphan item.  Lord knows we don't want to have a dangling
9213          * orphan item because we didn't reserve space to remove it.
9214          *
9215          * 2) We need to reserve space to update our inode.
9216          *
9217          * 3) We need to have something to cache all the space that is going to
9218          * be free'd up by the truncate operation, but also have some slack
9219          * space reserved in case it uses space during the truncate (thank you
9220          * very much snapshotting).
9221          *
9222          * And we need these to all be separate.  The fact is we can use a lot of
9223          * space doing the truncate, and we have no earthly idea how much space
9224          * we will use, so we need the truncate reservation to be separate so it
9225          * doesn't end up using space reserved for updating the inode or
9226          * removing the orphan item.  We also need to be able to stop the
9227          * transaction and start a new one, which means we need to be able to
9228          * update the inode several times, and we have no idea of knowing how
9229          * many times that will be, so we can't just reserve 1 item for the
9230          * entirety of the operation, so that has to be done separately as well.
9231          * Then there is the orphan item, which does indeed need to be held on
9232          * to for the whole operation, and we need nobody to touch this reserved
9233          * space except the orphan code.
9234          *
9235          * So that leaves us with
9236          *
9237          * 1) root->orphan_block_rsv - for the orphan deletion.
9238          * 2) rsv - for the truncate reservation, which we will steal from the
9239          * transaction reservation.
9240          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9241          * updating the inode.
9242          */
9243         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9244         if (!rsv)
9245                 return -ENOMEM;
9246         rsv->size = min_size;
9247         rsv->failfast = 1;
9248
9249         /*
9250          * 1 for the truncate slack space
9251          * 1 for updating the inode.
9252          */
9253         trans = btrfs_start_transaction(root, 2);
9254         if (IS_ERR(trans)) {
9255                 err = PTR_ERR(trans);
9256                 goto out;
9257         }
9258
9259         /* Migrate the slack space for the truncate to our reserve */
9260         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9261                                       min_size, 0);
9262         BUG_ON(ret);
9263
9264         /*
9265          * So if we truncate and then write and fsync we normally would just
9266          * write the extents that changed, which is a problem if we need to
9267          * first truncate that entire inode.  So set this flag so we write out
9268          * all of the extents in the inode to the sync log so we're completely
9269          * safe.
9270          */
9271         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9272         trans->block_rsv = rsv;
9273
9274         while (1) {
9275                 ret = btrfs_truncate_inode_items(trans, root, inode,
9276                                                  inode->i_size,
9277                                                  BTRFS_EXTENT_DATA_KEY);
9278                 if (ret != -ENOSPC && ret != -EAGAIN) {
9279                         err = ret;
9280                         break;
9281                 }
9282
9283                 trans->block_rsv = &fs_info->trans_block_rsv;
9284                 ret = btrfs_update_inode(trans, root, inode);
9285                 if (ret) {
9286                         err = ret;
9287                         break;
9288                 }
9289
9290                 btrfs_end_transaction(trans);
9291                 btrfs_btree_balance_dirty(fs_info);
9292
9293                 trans = btrfs_start_transaction(root, 2);
9294                 if (IS_ERR(trans)) {
9295                         ret = err = PTR_ERR(trans);
9296                         trans = NULL;
9297                         break;
9298                 }
9299
9300                 btrfs_block_rsv_release(fs_info, rsv, -1);
9301                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9302                                               rsv, min_size, 0);
9303                 BUG_ON(ret);    /* shouldn't happen */
9304                 trans->block_rsv = rsv;
9305         }
9306
9307         if (ret == 0 && inode->i_nlink > 0) {
9308                 trans->block_rsv = root->orphan_block_rsv;
9309                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9310                 if (ret)
9311                         err = ret;
9312         }
9313
9314         if (trans) {
9315                 trans->block_rsv = &fs_info->trans_block_rsv;
9316                 ret = btrfs_update_inode(trans, root, inode);
9317                 if (ret && !err)
9318                         err = ret;
9319
9320                 ret = btrfs_end_transaction(trans);
9321                 btrfs_btree_balance_dirty(fs_info);
9322         }
9323 out:
9324         btrfs_free_block_rsv(fs_info, rsv);
9325
9326         if (ret && !err)
9327                 err = ret;
9328
9329         return err;
9330 }
9331
9332 /*
9333  * create a new subvolume directory/inode (helper for the ioctl).
9334  */
9335 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9336                              struct btrfs_root *new_root,
9337                              struct btrfs_root *parent_root,
9338                              u64 new_dirid)
9339 {
9340         struct inode *inode;
9341         int err;
9342         u64 index = 0;
9343
9344         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9345                                 new_dirid, new_dirid,
9346                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9347                                 &index);
9348         if (IS_ERR(inode))
9349                 return PTR_ERR(inode);
9350         inode->i_op = &btrfs_dir_inode_operations;
9351         inode->i_fop = &btrfs_dir_file_operations;
9352
9353         set_nlink(inode, 1);
9354         btrfs_i_size_write(BTRFS_I(inode), 0);
9355         unlock_new_inode(inode);
9356
9357         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9358         if (err)
9359                 btrfs_err(new_root->fs_info,
9360                           "error inheriting subvolume %llu properties: %d",
9361                           new_root->root_key.objectid, err);
9362
9363         err = btrfs_update_inode(trans, new_root, inode);
9364
9365         iput(inode);
9366         return err;
9367 }
9368
9369 struct inode *btrfs_alloc_inode(struct super_block *sb)
9370 {
9371         struct btrfs_inode *ei;
9372         struct inode *inode;
9373
9374         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9375         if (!ei)
9376                 return NULL;
9377
9378         ei->root = NULL;
9379         ei->generation = 0;
9380         ei->last_trans = 0;
9381         ei->last_sub_trans = 0;
9382         ei->logged_trans = 0;
9383         ei->delalloc_bytes = 0;
9384         ei->new_delalloc_bytes = 0;
9385         ei->defrag_bytes = 0;
9386         ei->disk_i_size = 0;
9387         ei->flags = 0;
9388         ei->csum_bytes = 0;
9389         ei->index_cnt = (u64)-1;
9390         ei->dir_index = 0;
9391         ei->last_unlink_trans = 0;
9392         ei->last_log_commit = 0;
9393         ei->delayed_iput_count = 0;
9394
9395         spin_lock_init(&ei->lock);
9396         ei->outstanding_extents = 0;
9397         ei->reserved_extents = 0;
9398
9399         ei->runtime_flags = 0;
9400         ei->force_compress = BTRFS_COMPRESS_NONE;
9401
9402         ei->delayed_node = NULL;
9403
9404         ei->i_otime.tv_sec = 0;
9405         ei->i_otime.tv_nsec = 0;
9406
9407         inode = &ei->vfs_inode;
9408         extent_map_tree_init(&ei->extent_tree);
9409         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9410         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9411         ei->io_tree.track_uptodate = 1;
9412         ei->io_failure_tree.track_uptodate = 1;
9413         atomic_set(&ei->sync_writers, 0);
9414         mutex_init(&ei->log_mutex);
9415         mutex_init(&ei->delalloc_mutex);
9416         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9417         INIT_LIST_HEAD(&ei->delalloc_inodes);
9418         INIT_LIST_HEAD(&ei->delayed_iput);
9419         RB_CLEAR_NODE(&ei->rb_node);
9420         init_rwsem(&ei->dio_sem);
9421
9422         return inode;
9423 }
9424
9425 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9426 void btrfs_test_destroy_inode(struct inode *inode)
9427 {
9428         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9429         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9430 }
9431 #endif
9432
9433 static void btrfs_i_callback(struct rcu_head *head)
9434 {
9435         struct inode *inode = container_of(head, struct inode, i_rcu);
9436         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9437 }
9438
9439 void btrfs_destroy_inode(struct inode *inode)
9440 {
9441         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9442         struct btrfs_ordered_extent *ordered;
9443         struct btrfs_root *root = BTRFS_I(inode)->root;
9444
9445         WARN_ON(!hlist_empty(&inode->i_dentry));
9446         WARN_ON(inode->i_data.nrpages);
9447         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9448         WARN_ON(BTRFS_I(inode)->reserved_extents);
9449         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9450         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9451         WARN_ON(BTRFS_I(inode)->csum_bytes);
9452         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9453
9454         /*
9455          * This can happen where we create an inode, but somebody else also
9456          * created the same inode and we need to destroy the one we already
9457          * created.
9458          */
9459         if (!root)
9460                 goto free;
9461
9462         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9463                      &BTRFS_I(inode)->runtime_flags)) {
9464                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9465                            btrfs_ino(BTRFS_I(inode)));
9466                 atomic_dec(&root->orphan_inodes);
9467         }
9468
9469         while (1) {
9470                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9471                 if (!ordered)
9472                         break;
9473                 else {
9474                         btrfs_err(fs_info,
9475                                   "found ordered extent %llu %llu on inode cleanup",
9476                                   ordered->file_offset, ordered->len);
9477                         btrfs_remove_ordered_extent(inode, ordered);
9478                         btrfs_put_ordered_extent(ordered);
9479                         btrfs_put_ordered_extent(ordered);
9480                 }
9481         }
9482         btrfs_qgroup_check_reserved_leak(inode);
9483         inode_tree_del(inode);
9484         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9485 free:
9486         call_rcu(&inode->i_rcu, btrfs_i_callback);
9487 }
9488
9489 int btrfs_drop_inode(struct inode *inode)
9490 {
9491         struct btrfs_root *root = BTRFS_I(inode)->root;
9492
9493         if (root == NULL)
9494                 return 1;
9495
9496         /* the snap/subvol tree is on deleting */
9497         if (btrfs_root_refs(&root->root_item) == 0)
9498                 return 1;
9499         else
9500                 return generic_drop_inode(inode);
9501 }
9502
9503 static void init_once(void *foo)
9504 {
9505         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9506
9507         inode_init_once(&ei->vfs_inode);
9508 }
9509
9510 void btrfs_destroy_cachep(void)
9511 {
9512         /*
9513          * Make sure all delayed rcu free inodes are flushed before we
9514          * destroy cache.
9515          */
9516         rcu_barrier();
9517         kmem_cache_destroy(btrfs_inode_cachep);
9518         kmem_cache_destroy(btrfs_trans_handle_cachep);
9519         kmem_cache_destroy(btrfs_transaction_cachep);
9520         kmem_cache_destroy(btrfs_path_cachep);
9521         kmem_cache_destroy(btrfs_free_space_cachep);
9522 }
9523
9524 int btrfs_init_cachep(void)
9525 {
9526         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9527                         sizeof(struct btrfs_inode), 0,
9528                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9529                         init_once);
9530         if (!btrfs_inode_cachep)
9531                 goto fail;
9532
9533         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9534                         sizeof(struct btrfs_trans_handle), 0,
9535                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9536         if (!btrfs_trans_handle_cachep)
9537                 goto fail;
9538
9539         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9540                         sizeof(struct btrfs_transaction), 0,
9541                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9542         if (!btrfs_transaction_cachep)
9543                 goto fail;
9544
9545         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9546                         sizeof(struct btrfs_path), 0,
9547                         SLAB_MEM_SPREAD, NULL);
9548         if (!btrfs_path_cachep)
9549                 goto fail;
9550
9551         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9552                         sizeof(struct btrfs_free_space), 0,
9553                         SLAB_MEM_SPREAD, NULL);
9554         if (!btrfs_free_space_cachep)
9555                 goto fail;
9556
9557         return 0;
9558 fail:
9559         btrfs_destroy_cachep();
9560         return -ENOMEM;
9561 }
9562
9563 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9564                          u32 request_mask, unsigned int flags)
9565 {
9566         u64 delalloc_bytes;
9567         struct inode *inode = d_inode(path->dentry);
9568         u32 blocksize = inode->i_sb->s_blocksize;
9569
9570         generic_fillattr(inode, stat);
9571         stat->dev = BTRFS_I(inode)->root->anon_dev;
9572
9573         spin_lock(&BTRFS_I(inode)->lock);
9574         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9575         spin_unlock(&BTRFS_I(inode)->lock);
9576         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9577                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9578         return 0;
9579 }
9580
9581 static int btrfs_rename_exchange(struct inode *old_dir,
9582                               struct dentry *old_dentry,
9583                               struct inode *new_dir,
9584                               struct dentry *new_dentry)
9585 {
9586         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9587         struct btrfs_trans_handle *trans;
9588         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9589         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9590         struct inode *new_inode = new_dentry->d_inode;
9591         struct inode *old_inode = old_dentry->d_inode;
9592         struct timespec ctime = current_time(old_inode);
9593         struct dentry *parent;
9594         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9595         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9596         u64 old_idx = 0;
9597         u64 new_idx = 0;
9598         u64 root_objectid;
9599         int ret;
9600         bool root_log_pinned = false;
9601         bool dest_log_pinned = false;
9602
9603         /* we only allow rename subvolume link between subvolumes */
9604         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9605                 return -EXDEV;
9606
9607         /* close the race window with snapshot create/destroy ioctl */
9608         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9609                 down_read(&fs_info->subvol_sem);
9610         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9611                 down_read(&fs_info->subvol_sem);
9612
9613         /*
9614          * We want to reserve the absolute worst case amount of items.  So if
9615          * both inodes are subvols and we need to unlink them then that would
9616          * require 4 item modifications, but if they are both normal inodes it
9617          * would require 5 item modifications, so we'll assume their normal
9618          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9619          * should cover the worst case number of items we'll modify.
9620          */
9621         trans = btrfs_start_transaction(root, 12);
9622         if (IS_ERR(trans)) {
9623                 ret = PTR_ERR(trans);
9624                 goto out_notrans;
9625         }
9626
9627         /*
9628          * We need to find a free sequence number both in the source and
9629          * in the destination directory for the exchange.
9630          */
9631         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9632         if (ret)
9633                 goto out_fail;
9634         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9635         if (ret)
9636                 goto out_fail;
9637
9638         BTRFS_I(old_inode)->dir_index = 0ULL;
9639         BTRFS_I(new_inode)->dir_index = 0ULL;
9640
9641         /* Reference for the source. */
9642         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9643                 /* force full log commit if subvolume involved. */
9644                 btrfs_set_log_full_commit(fs_info, trans);
9645         } else {
9646                 btrfs_pin_log_trans(root);
9647                 root_log_pinned = true;
9648                 ret = btrfs_insert_inode_ref(trans, dest,
9649                                              new_dentry->d_name.name,
9650                                              new_dentry->d_name.len,
9651                                              old_ino,
9652                                              btrfs_ino(BTRFS_I(new_dir)),
9653                                              old_idx);
9654                 if (ret)
9655                         goto out_fail;
9656         }
9657
9658         /* And now for the dest. */
9659         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9660                 /* force full log commit if subvolume involved. */
9661                 btrfs_set_log_full_commit(fs_info, trans);
9662         } else {
9663                 btrfs_pin_log_trans(dest);
9664                 dest_log_pinned = true;
9665                 ret = btrfs_insert_inode_ref(trans, root,
9666                                              old_dentry->d_name.name,
9667                                              old_dentry->d_name.len,
9668                                              new_ino,
9669                                              btrfs_ino(BTRFS_I(old_dir)),
9670                                              new_idx);
9671                 if (ret)
9672                         goto out_fail;
9673         }
9674
9675         /* Update inode version and ctime/mtime. */
9676         inode_inc_iversion(old_dir);
9677         inode_inc_iversion(new_dir);
9678         inode_inc_iversion(old_inode);
9679         inode_inc_iversion(new_inode);
9680         old_dir->i_ctime = old_dir->i_mtime = ctime;
9681         new_dir->i_ctime = new_dir->i_mtime = ctime;
9682         old_inode->i_ctime = ctime;
9683         new_inode->i_ctime = ctime;
9684
9685         if (old_dentry->d_parent != new_dentry->d_parent) {
9686                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9687                                 BTRFS_I(old_inode), 1);
9688                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9689                                 BTRFS_I(new_inode), 1);
9690         }
9691
9692         /* src is a subvolume */
9693         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9694                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9695                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9696                                           root_objectid,
9697                                           old_dentry->d_name.name,
9698                                           old_dentry->d_name.len);
9699         } else { /* src is an inode */
9700                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9701                                            BTRFS_I(old_dentry->d_inode),
9702                                            old_dentry->d_name.name,
9703                                            old_dentry->d_name.len);
9704                 if (!ret)
9705                         ret = btrfs_update_inode(trans, root, old_inode);
9706         }
9707         if (ret) {
9708                 btrfs_abort_transaction(trans, ret);
9709                 goto out_fail;
9710         }
9711
9712         /* dest is a subvolume */
9713         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9714                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9715                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9716                                           root_objectid,
9717                                           new_dentry->d_name.name,
9718                                           new_dentry->d_name.len);
9719         } else { /* dest is an inode */
9720                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9721                                            BTRFS_I(new_dentry->d_inode),
9722                                            new_dentry->d_name.name,
9723                                            new_dentry->d_name.len);
9724                 if (!ret)
9725                         ret = btrfs_update_inode(trans, dest, new_inode);
9726         }
9727         if (ret) {
9728                 btrfs_abort_transaction(trans, ret);
9729                 goto out_fail;
9730         }
9731
9732         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9733                              new_dentry->d_name.name,
9734                              new_dentry->d_name.len, 0, old_idx);
9735         if (ret) {
9736                 btrfs_abort_transaction(trans, ret);
9737                 goto out_fail;
9738         }
9739
9740         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9741                              old_dentry->d_name.name,
9742                              old_dentry->d_name.len, 0, new_idx);
9743         if (ret) {
9744                 btrfs_abort_transaction(trans, ret);
9745                 goto out_fail;
9746         }
9747
9748         if (old_inode->i_nlink == 1)
9749                 BTRFS_I(old_inode)->dir_index = old_idx;
9750         if (new_inode->i_nlink == 1)
9751                 BTRFS_I(new_inode)->dir_index = new_idx;
9752
9753         if (root_log_pinned) {
9754                 parent = new_dentry->d_parent;
9755                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9756                                 parent);
9757                 btrfs_end_log_trans(root);
9758                 root_log_pinned = false;
9759         }
9760         if (dest_log_pinned) {
9761                 parent = old_dentry->d_parent;
9762                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9763                                 parent);
9764                 btrfs_end_log_trans(dest);
9765                 dest_log_pinned = false;
9766         }
9767 out_fail:
9768         /*
9769          * If we have pinned a log and an error happened, we unpin tasks
9770          * trying to sync the log and force them to fallback to a transaction
9771          * commit if the log currently contains any of the inodes involved in
9772          * this rename operation (to ensure we do not persist a log with an
9773          * inconsistent state for any of these inodes or leading to any
9774          * inconsistencies when replayed). If the transaction was aborted, the
9775          * abortion reason is propagated to userspace when attempting to commit
9776          * the transaction. If the log does not contain any of these inodes, we
9777          * allow the tasks to sync it.
9778          */
9779         if (ret && (root_log_pinned || dest_log_pinned)) {
9780                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9781                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9782                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9783                     (new_inode &&
9784                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9785                         btrfs_set_log_full_commit(fs_info, trans);
9786
9787                 if (root_log_pinned) {
9788                         btrfs_end_log_trans(root);
9789                         root_log_pinned = false;
9790                 }
9791                 if (dest_log_pinned) {
9792                         btrfs_end_log_trans(dest);
9793                         dest_log_pinned = false;
9794                 }
9795         }
9796         ret = btrfs_end_transaction(trans);
9797 out_notrans:
9798         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9799                 up_read(&fs_info->subvol_sem);
9800         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9801                 up_read(&fs_info->subvol_sem);
9802
9803         return ret;
9804 }
9805
9806 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9807                                      struct btrfs_root *root,
9808                                      struct inode *dir,
9809                                      struct dentry *dentry)
9810 {
9811         int ret;
9812         struct inode *inode;
9813         u64 objectid;
9814         u64 index;
9815
9816         ret = btrfs_find_free_ino(root, &objectid);
9817         if (ret)
9818                 return ret;
9819
9820         inode = btrfs_new_inode(trans, root, dir,
9821                                 dentry->d_name.name,
9822                                 dentry->d_name.len,
9823                                 btrfs_ino(BTRFS_I(dir)),
9824                                 objectid,
9825                                 S_IFCHR | WHITEOUT_MODE,
9826                                 &index);
9827
9828         if (IS_ERR(inode)) {
9829                 ret = PTR_ERR(inode);
9830                 return ret;
9831         }
9832
9833         inode->i_op = &btrfs_special_inode_operations;
9834         init_special_inode(inode, inode->i_mode,
9835                 WHITEOUT_DEV);
9836
9837         ret = btrfs_init_inode_security(trans, inode, dir,
9838                                 &dentry->d_name);
9839         if (ret)
9840                 goto out;
9841
9842         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9843                                 BTRFS_I(inode), 0, index);
9844         if (ret)
9845                 goto out;
9846
9847         ret = btrfs_update_inode(trans, root, inode);
9848 out:
9849         unlock_new_inode(inode);
9850         if (ret)
9851                 inode_dec_link_count(inode);
9852         iput(inode);
9853
9854         return ret;
9855 }
9856
9857 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9858                            struct inode *new_dir, struct dentry *new_dentry,
9859                            unsigned int flags)
9860 {
9861         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9862         struct btrfs_trans_handle *trans;
9863         unsigned int trans_num_items;
9864         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9865         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9866         struct inode *new_inode = d_inode(new_dentry);
9867         struct inode *old_inode = d_inode(old_dentry);
9868         u64 index = 0;
9869         u64 root_objectid;
9870         int ret;
9871         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9872         bool log_pinned = false;
9873
9874         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9875                 return -EPERM;
9876
9877         /* we only allow rename subvolume link between subvolumes */
9878         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9879                 return -EXDEV;
9880
9881         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9882             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9883                 return -ENOTEMPTY;
9884
9885         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9886             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9887                 return -ENOTEMPTY;
9888
9889
9890         /* check for collisions, even if the  name isn't there */
9891         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9892                              new_dentry->d_name.name,
9893                              new_dentry->d_name.len);
9894
9895         if (ret) {
9896                 if (ret == -EEXIST) {
9897                         /* we shouldn't get
9898                          * eexist without a new_inode */
9899                         if (WARN_ON(!new_inode)) {
9900                                 return ret;
9901                         }
9902                 } else {
9903                         /* maybe -EOVERFLOW */
9904                         return ret;
9905                 }
9906         }
9907         ret = 0;
9908
9909         /*
9910          * we're using rename to replace one file with another.  Start IO on it
9911          * now so  we don't add too much work to the end of the transaction
9912          */
9913         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9914                 filemap_flush(old_inode->i_mapping);
9915
9916         /* close the racy window with snapshot create/destroy ioctl */
9917         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9918                 down_read(&fs_info->subvol_sem);
9919         /*
9920          * We want to reserve the absolute worst case amount of items.  So if
9921          * both inodes are subvols and we need to unlink them then that would
9922          * require 4 item modifications, but if they are both normal inodes it
9923          * would require 5 item modifications, so we'll assume they are normal
9924          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9925          * should cover the worst case number of items we'll modify.
9926          * If our rename has the whiteout flag, we need more 5 units for the
9927          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9928          * when selinux is enabled).
9929          */
9930         trans_num_items = 11;
9931         if (flags & RENAME_WHITEOUT)
9932                 trans_num_items += 5;
9933         trans = btrfs_start_transaction(root, trans_num_items);
9934         if (IS_ERR(trans)) {
9935                 ret = PTR_ERR(trans);
9936                 goto out_notrans;
9937         }
9938
9939         if (dest != root)
9940                 btrfs_record_root_in_trans(trans, dest);
9941
9942         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9943         if (ret)
9944                 goto out_fail;
9945
9946         BTRFS_I(old_inode)->dir_index = 0ULL;
9947         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9948                 /* force full log commit if subvolume involved. */
9949                 btrfs_set_log_full_commit(fs_info, trans);
9950         } else {
9951                 btrfs_pin_log_trans(root);
9952                 log_pinned = true;
9953                 ret = btrfs_insert_inode_ref(trans, dest,
9954                                              new_dentry->d_name.name,
9955                                              new_dentry->d_name.len,
9956                                              old_ino,
9957                                              btrfs_ino(BTRFS_I(new_dir)), index);
9958                 if (ret)
9959                         goto out_fail;
9960         }
9961
9962         inode_inc_iversion(old_dir);
9963         inode_inc_iversion(new_dir);
9964         inode_inc_iversion(old_inode);
9965         old_dir->i_ctime = old_dir->i_mtime =
9966         new_dir->i_ctime = new_dir->i_mtime =
9967         old_inode->i_ctime = current_time(old_dir);
9968
9969         if (old_dentry->d_parent != new_dentry->d_parent)
9970                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9971                                 BTRFS_I(old_inode), 1);
9972
9973         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9974                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9975                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9976                                         old_dentry->d_name.name,
9977                                         old_dentry->d_name.len);
9978         } else {
9979                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9980                                         BTRFS_I(d_inode(old_dentry)),
9981                                         old_dentry->d_name.name,
9982                                         old_dentry->d_name.len);
9983                 if (!ret)
9984                         ret = btrfs_update_inode(trans, root, old_inode);
9985         }
9986         if (ret) {
9987                 btrfs_abort_transaction(trans, ret);
9988                 goto out_fail;
9989         }
9990
9991         if (new_inode) {
9992                 inode_inc_iversion(new_inode);
9993                 new_inode->i_ctime = current_time(new_inode);
9994                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9995                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9996                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9997                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9998                                                 root_objectid,
9999                                                 new_dentry->d_name.name,
10000                                                 new_dentry->d_name.len);
10001                         BUG_ON(new_inode->i_nlink == 0);
10002                 } else {
10003                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10004                                                  BTRFS_I(d_inode(new_dentry)),
10005                                                  new_dentry->d_name.name,
10006                                                  new_dentry->d_name.len);
10007                 }
10008                 if (!ret && new_inode->i_nlink == 0)
10009                         ret = btrfs_orphan_add(trans,
10010                                         BTRFS_I(d_inode(new_dentry)));
10011                 if (ret) {
10012                         btrfs_abort_transaction(trans, ret);
10013                         goto out_fail;
10014                 }
10015         }
10016
10017         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10018                              new_dentry->d_name.name,
10019                              new_dentry->d_name.len, 0, index);
10020         if (ret) {
10021                 btrfs_abort_transaction(trans, ret);
10022                 goto out_fail;
10023         }
10024
10025         if (old_inode->i_nlink == 1)
10026                 BTRFS_I(old_inode)->dir_index = index;
10027
10028         if (log_pinned) {
10029                 struct dentry *parent = new_dentry->d_parent;
10030
10031                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
10032                                 parent);
10033                 btrfs_end_log_trans(root);
10034                 log_pinned = false;
10035         }
10036
10037         if (flags & RENAME_WHITEOUT) {
10038                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10039                                                 old_dentry);
10040
10041                 if (ret) {
10042                         btrfs_abort_transaction(trans, ret);
10043                         goto out_fail;
10044                 }
10045         }
10046 out_fail:
10047         /*
10048          * If we have pinned the log and an error happened, we unpin tasks
10049          * trying to sync the log and force them to fallback to a transaction
10050          * commit if the log currently contains any of the inodes involved in
10051          * this rename operation (to ensure we do not persist a log with an
10052          * inconsistent state for any of these inodes or leading to any
10053          * inconsistencies when replayed). If the transaction was aborted, the
10054          * abortion reason is propagated to userspace when attempting to commit
10055          * the transaction. If the log does not contain any of these inodes, we
10056          * allow the tasks to sync it.
10057          */
10058         if (ret && log_pinned) {
10059                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10060                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10061                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10062                     (new_inode &&
10063                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10064                         btrfs_set_log_full_commit(fs_info, trans);
10065
10066                 btrfs_end_log_trans(root);
10067                 log_pinned = false;
10068         }
10069         btrfs_end_transaction(trans);
10070 out_notrans:
10071         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10072                 up_read(&fs_info->subvol_sem);
10073
10074         return ret;
10075 }
10076
10077 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10078                          struct inode *new_dir, struct dentry *new_dentry,
10079                          unsigned int flags)
10080 {
10081         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10082                 return -EINVAL;
10083
10084         if (flags & RENAME_EXCHANGE)
10085                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10086                                           new_dentry);
10087
10088         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10089 }
10090
10091 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10092 {
10093         struct btrfs_delalloc_work *delalloc_work;
10094         struct inode *inode;
10095
10096         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10097                                      work);
10098         inode = delalloc_work->inode;
10099         filemap_flush(inode->i_mapping);
10100         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10101                                 &BTRFS_I(inode)->runtime_flags))
10102                 filemap_flush(inode->i_mapping);
10103
10104         if (delalloc_work->delay_iput)
10105                 btrfs_add_delayed_iput(inode);
10106         else
10107                 iput(inode);
10108         complete(&delalloc_work->completion);
10109 }
10110
10111 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10112                                                     int delay_iput)
10113 {
10114         struct btrfs_delalloc_work *work;
10115
10116         work = kmalloc(sizeof(*work), GFP_NOFS);
10117         if (!work)
10118                 return NULL;
10119
10120         init_completion(&work->completion);
10121         INIT_LIST_HEAD(&work->list);
10122         work->inode = inode;
10123         work->delay_iput = delay_iput;
10124         WARN_ON_ONCE(!inode);
10125         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10126                         btrfs_run_delalloc_work, NULL, NULL);
10127
10128         return work;
10129 }
10130
10131 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10132 {
10133         wait_for_completion(&work->completion);
10134         kfree(work);
10135 }
10136
10137 /*
10138  * some fairly slow code that needs optimization. This walks the list
10139  * of all the inodes with pending delalloc and forces them to disk.
10140  */
10141 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10142                                    int nr)
10143 {
10144         struct btrfs_inode *binode;
10145         struct inode *inode;
10146         struct btrfs_delalloc_work *work, *next;
10147         struct list_head works;
10148         struct list_head splice;
10149         int ret = 0;
10150
10151         INIT_LIST_HEAD(&works);
10152         INIT_LIST_HEAD(&splice);
10153
10154         mutex_lock(&root->delalloc_mutex);
10155         spin_lock(&root->delalloc_lock);
10156         list_splice_init(&root->delalloc_inodes, &splice);
10157         while (!list_empty(&splice)) {
10158                 binode = list_entry(splice.next, struct btrfs_inode,
10159                                     delalloc_inodes);
10160
10161                 list_move_tail(&binode->delalloc_inodes,
10162                                &root->delalloc_inodes);
10163                 inode = igrab(&binode->vfs_inode);
10164                 if (!inode) {
10165                         cond_resched_lock(&root->delalloc_lock);
10166                         continue;
10167                 }
10168                 spin_unlock(&root->delalloc_lock);
10169
10170                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10171                 if (!work) {
10172                         if (delay_iput)
10173                                 btrfs_add_delayed_iput(inode);
10174                         else
10175                                 iput(inode);
10176                         ret = -ENOMEM;
10177                         goto out;
10178                 }
10179                 list_add_tail(&work->list, &works);
10180                 btrfs_queue_work(root->fs_info->flush_workers,
10181                                  &work->work);
10182                 ret++;
10183                 if (nr != -1 && ret >= nr)
10184                         goto out;
10185                 cond_resched();
10186                 spin_lock(&root->delalloc_lock);
10187         }
10188         spin_unlock(&root->delalloc_lock);
10189
10190 out:
10191         list_for_each_entry_safe(work, next, &works, list) {
10192                 list_del_init(&work->list);
10193                 btrfs_wait_and_free_delalloc_work(work);
10194         }
10195
10196         if (!list_empty_careful(&splice)) {
10197                 spin_lock(&root->delalloc_lock);
10198                 list_splice_tail(&splice, &root->delalloc_inodes);
10199                 spin_unlock(&root->delalloc_lock);
10200         }
10201         mutex_unlock(&root->delalloc_mutex);
10202         return ret;
10203 }
10204
10205 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10206 {
10207         struct btrfs_fs_info *fs_info = root->fs_info;
10208         int ret;
10209
10210         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10211                 return -EROFS;
10212
10213         ret = __start_delalloc_inodes(root, delay_iput, -1);
10214         if (ret > 0)
10215                 ret = 0;
10216         /*
10217          * the filemap_flush will queue IO into the worker threads, but
10218          * we have to make sure the IO is actually started and that
10219          * ordered extents get created before we return
10220          */
10221         atomic_inc(&fs_info->async_submit_draining);
10222         while (atomic_read(&fs_info->nr_async_submits) ||
10223                atomic_read(&fs_info->async_delalloc_pages)) {
10224                 wait_event(fs_info->async_submit_wait,
10225                            (atomic_read(&fs_info->nr_async_submits) == 0 &&
10226                             atomic_read(&fs_info->async_delalloc_pages) == 0));
10227         }
10228         atomic_dec(&fs_info->async_submit_draining);
10229         return ret;
10230 }
10231
10232 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10233                                int nr)
10234 {
10235         struct btrfs_root *root;
10236         struct list_head splice;
10237         int ret;
10238
10239         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10240                 return -EROFS;
10241
10242         INIT_LIST_HEAD(&splice);
10243
10244         mutex_lock(&fs_info->delalloc_root_mutex);
10245         spin_lock(&fs_info->delalloc_root_lock);
10246         list_splice_init(&fs_info->delalloc_roots, &splice);
10247         while (!list_empty(&splice) && nr) {
10248                 root = list_first_entry(&splice, struct btrfs_root,
10249                                         delalloc_root);
10250                 root = btrfs_grab_fs_root(root);
10251                 BUG_ON(!root);
10252                 list_move_tail(&root->delalloc_root,
10253                                &fs_info->delalloc_roots);
10254                 spin_unlock(&fs_info->delalloc_root_lock);
10255
10256                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10257                 btrfs_put_fs_root(root);
10258                 if (ret < 0)
10259                         goto out;
10260
10261                 if (nr != -1) {
10262                         nr -= ret;
10263                         WARN_ON(nr < 0);
10264                 }
10265                 spin_lock(&fs_info->delalloc_root_lock);
10266         }
10267         spin_unlock(&fs_info->delalloc_root_lock);
10268
10269         ret = 0;
10270         atomic_inc(&fs_info->async_submit_draining);
10271         while (atomic_read(&fs_info->nr_async_submits) ||
10272               atomic_read(&fs_info->async_delalloc_pages)) {
10273                 wait_event(fs_info->async_submit_wait,
10274                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10275                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10276         }
10277         atomic_dec(&fs_info->async_submit_draining);
10278 out:
10279         if (!list_empty_careful(&splice)) {
10280                 spin_lock(&fs_info->delalloc_root_lock);
10281                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10282                 spin_unlock(&fs_info->delalloc_root_lock);
10283         }
10284         mutex_unlock(&fs_info->delalloc_root_mutex);
10285         return ret;
10286 }
10287
10288 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10289                          const char *symname)
10290 {
10291         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10292         struct btrfs_trans_handle *trans;
10293         struct btrfs_root *root = BTRFS_I(dir)->root;
10294         struct btrfs_path *path;
10295         struct btrfs_key key;
10296         struct inode *inode = NULL;
10297         int err;
10298         int drop_inode = 0;
10299         u64 objectid;
10300         u64 index = 0;
10301         int name_len;
10302         int datasize;
10303         unsigned long ptr;
10304         struct btrfs_file_extent_item *ei;
10305         struct extent_buffer *leaf;
10306
10307         name_len = strlen(symname);
10308         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10309                 return -ENAMETOOLONG;
10310
10311         /*
10312          * 2 items for inode item and ref
10313          * 2 items for dir items
10314          * 1 item for updating parent inode item
10315          * 1 item for the inline extent item
10316          * 1 item for xattr if selinux is on
10317          */
10318         trans = btrfs_start_transaction(root, 7);
10319         if (IS_ERR(trans))
10320                 return PTR_ERR(trans);
10321
10322         err = btrfs_find_free_ino(root, &objectid);
10323         if (err)
10324                 goto out_unlock;
10325
10326         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10327                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10328                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10329         if (IS_ERR(inode)) {
10330                 err = PTR_ERR(inode);
10331                 goto out_unlock;
10332         }
10333
10334         /*
10335         * If the active LSM wants to access the inode during
10336         * d_instantiate it needs these. Smack checks to see
10337         * if the filesystem supports xattrs by looking at the
10338         * ops vector.
10339         */
10340         inode->i_fop = &btrfs_file_operations;
10341         inode->i_op = &btrfs_file_inode_operations;
10342         inode->i_mapping->a_ops = &btrfs_aops;
10343         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10344
10345         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10346         if (err)
10347                 goto out_unlock_inode;
10348
10349         path = btrfs_alloc_path();
10350         if (!path) {
10351                 err = -ENOMEM;
10352                 goto out_unlock_inode;
10353         }
10354         key.objectid = btrfs_ino(BTRFS_I(inode));
10355         key.offset = 0;
10356         key.type = BTRFS_EXTENT_DATA_KEY;
10357         datasize = btrfs_file_extent_calc_inline_size(name_len);
10358         err = btrfs_insert_empty_item(trans, root, path, &key,
10359                                       datasize);
10360         if (err) {
10361                 btrfs_free_path(path);
10362                 goto out_unlock_inode;
10363         }
10364         leaf = path->nodes[0];
10365         ei = btrfs_item_ptr(leaf, path->slots[0],
10366                             struct btrfs_file_extent_item);
10367         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10368         btrfs_set_file_extent_type(leaf, ei,
10369                                    BTRFS_FILE_EXTENT_INLINE);
10370         btrfs_set_file_extent_encryption(leaf, ei, 0);
10371         btrfs_set_file_extent_compression(leaf, ei, 0);
10372         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10373         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10374
10375         ptr = btrfs_file_extent_inline_start(ei);
10376         write_extent_buffer(leaf, symname, ptr, name_len);
10377         btrfs_mark_buffer_dirty(leaf);
10378         btrfs_free_path(path);
10379
10380         inode->i_op = &btrfs_symlink_inode_operations;
10381         inode_nohighmem(inode);
10382         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10383         inode_set_bytes(inode, name_len);
10384         btrfs_i_size_write(BTRFS_I(inode), name_len);
10385         err = btrfs_update_inode(trans, root, inode);
10386         /*
10387          * Last step, add directory indexes for our symlink inode. This is the
10388          * last step to avoid extra cleanup of these indexes if an error happens
10389          * elsewhere above.
10390          */
10391         if (!err)
10392                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10393                                 BTRFS_I(inode), 0, index);
10394         if (err) {
10395                 drop_inode = 1;
10396                 goto out_unlock_inode;
10397         }
10398
10399         unlock_new_inode(inode);
10400         d_instantiate(dentry, inode);
10401
10402 out_unlock:
10403         btrfs_end_transaction(trans);
10404         if (drop_inode) {
10405                 inode_dec_link_count(inode);
10406                 iput(inode);
10407         }
10408         btrfs_btree_balance_dirty(fs_info);
10409         return err;
10410
10411 out_unlock_inode:
10412         drop_inode = 1;
10413         unlock_new_inode(inode);
10414         goto out_unlock;
10415 }
10416
10417 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10418                                        u64 start, u64 num_bytes, u64 min_size,
10419                                        loff_t actual_len, u64 *alloc_hint,
10420                                        struct btrfs_trans_handle *trans)
10421 {
10422         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10423         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10424         struct extent_map *em;
10425         struct btrfs_root *root = BTRFS_I(inode)->root;
10426         struct btrfs_key ins;
10427         u64 cur_offset = start;
10428         u64 i_size;
10429         u64 cur_bytes;
10430         u64 last_alloc = (u64)-1;
10431         int ret = 0;
10432         bool own_trans = true;
10433         u64 end = start + num_bytes - 1;
10434
10435         if (trans)
10436                 own_trans = false;
10437         while (num_bytes > 0) {
10438                 if (own_trans) {
10439                         trans = btrfs_start_transaction(root, 3);
10440                         if (IS_ERR(trans)) {
10441                                 ret = PTR_ERR(trans);
10442                                 break;
10443                         }
10444                 }
10445
10446                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10447                 cur_bytes = max(cur_bytes, min_size);
10448                 /*
10449                  * If we are severely fragmented we could end up with really
10450                  * small allocations, so if the allocator is returning small
10451                  * chunks lets make its job easier by only searching for those
10452                  * sized chunks.
10453                  */
10454                 cur_bytes = min(cur_bytes, last_alloc);
10455                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10456                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10457                 if (ret) {
10458                         if (own_trans)
10459                                 btrfs_end_transaction(trans);
10460                         break;
10461                 }
10462                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10463
10464                 last_alloc = ins.offset;
10465                 ret = insert_reserved_file_extent(trans, inode,
10466                                                   cur_offset, ins.objectid,
10467                                                   ins.offset, ins.offset,
10468                                                   ins.offset, 0, 0, 0,
10469                                                   BTRFS_FILE_EXTENT_PREALLOC);
10470                 if (ret) {
10471                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10472                                                    ins.offset, 0);
10473                         btrfs_abort_transaction(trans, ret);
10474                         if (own_trans)
10475                                 btrfs_end_transaction(trans);
10476                         break;
10477                 }
10478
10479                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10480                                         cur_offset + ins.offset -1, 0);
10481
10482                 em = alloc_extent_map();
10483                 if (!em) {
10484                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10485                                 &BTRFS_I(inode)->runtime_flags);
10486                         goto next;
10487                 }
10488
10489                 em->start = cur_offset;
10490                 em->orig_start = cur_offset;
10491                 em->len = ins.offset;
10492                 em->block_start = ins.objectid;
10493                 em->block_len = ins.offset;
10494                 em->orig_block_len = ins.offset;
10495                 em->ram_bytes = ins.offset;
10496                 em->bdev = fs_info->fs_devices->latest_bdev;
10497                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10498                 em->generation = trans->transid;
10499
10500                 while (1) {
10501                         write_lock(&em_tree->lock);
10502                         ret = add_extent_mapping(em_tree, em, 1);
10503                         write_unlock(&em_tree->lock);
10504                         if (ret != -EEXIST)
10505                                 break;
10506                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10507                                                 cur_offset + ins.offset - 1,
10508                                                 0);
10509                 }
10510                 free_extent_map(em);
10511 next:
10512                 num_bytes -= ins.offset;
10513                 cur_offset += ins.offset;
10514                 *alloc_hint = ins.objectid + ins.offset;
10515
10516                 inode_inc_iversion(inode);
10517                 inode->i_ctime = current_time(inode);
10518                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10519                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10520                     (actual_len > inode->i_size) &&
10521                     (cur_offset > inode->i_size)) {
10522                         if (cur_offset > actual_len)
10523                                 i_size = actual_len;
10524                         else
10525                                 i_size = cur_offset;
10526                         i_size_write(inode, i_size);
10527                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10528                 }
10529
10530                 ret = btrfs_update_inode(trans, root, inode);
10531
10532                 if (ret) {
10533                         btrfs_abort_transaction(trans, ret);
10534                         if (own_trans)
10535                                 btrfs_end_transaction(trans);
10536                         break;
10537                 }
10538
10539                 if (own_trans)
10540                         btrfs_end_transaction(trans);
10541         }
10542         if (cur_offset < end)
10543                 btrfs_free_reserved_data_space(inode, cur_offset,
10544                         end - cur_offset + 1);
10545         return ret;
10546 }
10547
10548 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10549                               u64 start, u64 num_bytes, u64 min_size,
10550                               loff_t actual_len, u64 *alloc_hint)
10551 {
10552         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10553                                            min_size, actual_len, alloc_hint,
10554                                            NULL);
10555 }
10556
10557 int btrfs_prealloc_file_range_trans(struct inode *inode,
10558                                     struct btrfs_trans_handle *trans, int mode,
10559                                     u64 start, u64 num_bytes, u64 min_size,
10560                                     loff_t actual_len, u64 *alloc_hint)
10561 {
10562         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10563                                            min_size, actual_len, alloc_hint, trans);
10564 }
10565
10566 static int btrfs_set_page_dirty(struct page *page)
10567 {
10568         return __set_page_dirty_nobuffers(page);
10569 }
10570
10571 static int btrfs_permission(struct inode *inode, int mask)
10572 {
10573         struct btrfs_root *root = BTRFS_I(inode)->root;
10574         umode_t mode = inode->i_mode;
10575
10576         if (mask & MAY_WRITE &&
10577             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10578                 if (btrfs_root_readonly(root))
10579                         return -EROFS;
10580                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10581                         return -EACCES;
10582         }
10583         return generic_permission(inode, mask);
10584 }
10585
10586 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10587 {
10588         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10589         struct btrfs_trans_handle *trans;
10590         struct btrfs_root *root = BTRFS_I(dir)->root;
10591         struct inode *inode = NULL;
10592         u64 objectid;
10593         u64 index;
10594         int ret = 0;
10595
10596         /*
10597          * 5 units required for adding orphan entry
10598          */
10599         trans = btrfs_start_transaction(root, 5);
10600         if (IS_ERR(trans))
10601                 return PTR_ERR(trans);
10602
10603         ret = btrfs_find_free_ino(root, &objectid);
10604         if (ret)
10605                 goto out;
10606
10607         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10608                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10609         if (IS_ERR(inode)) {
10610                 ret = PTR_ERR(inode);
10611                 inode = NULL;
10612                 goto out;
10613         }
10614
10615         inode->i_fop = &btrfs_file_operations;
10616         inode->i_op = &btrfs_file_inode_operations;
10617
10618         inode->i_mapping->a_ops = &btrfs_aops;
10619         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10620
10621         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10622         if (ret)
10623                 goto out_inode;
10624
10625         ret = btrfs_update_inode(trans, root, inode);
10626         if (ret)
10627                 goto out_inode;
10628         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10629         if (ret)
10630                 goto out_inode;
10631
10632         /*
10633          * We set number of links to 0 in btrfs_new_inode(), and here we set
10634          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10635          * through:
10636          *
10637          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10638          */
10639         set_nlink(inode, 1);
10640         unlock_new_inode(inode);
10641         d_tmpfile(dentry, inode);
10642         mark_inode_dirty(inode);
10643
10644 out:
10645         btrfs_end_transaction(trans);
10646         if (ret)
10647                 iput(inode);
10648         btrfs_balance_delayed_items(fs_info);
10649         btrfs_btree_balance_dirty(fs_info);
10650         return ret;
10651
10652 out_inode:
10653         unlock_new_inode(inode);
10654         goto out;
10655
10656 }
10657
10658 __attribute__((const))
10659 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10660 {
10661         return -EAGAIN;
10662 }
10663
10664 static const struct inode_operations btrfs_dir_inode_operations = {
10665         .getattr        = btrfs_getattr,
10666         .lookup         = btrfs_lookup,
10667         .create         = btrfs_create,
10668         .unlink         = btrfs_unlink,
10669         .link           = btrfs_link,
10670         .mkdir          = btrfs_mkdir,
10671         .rmdir          = btrfs_rmdir,
10672         .rename         = btrfs_rename2,
10673         .symlink        = btrfs_symlink,
10674         .setattr        = btrfs_setattr,
10675         .mknod          = btrfs_mknod,
10676         .listxattr      = btrfs_listxattr,
10677         .permission     = btrfs_permission,
10678         .get_acl        = btrfs_get_acl,
10679         .set_acl        = btrfs_set_acl,
10680         .update_time    = btrfs_update_time,
10681         .tmpfile        = btrfs_tmpfile,
10682 };
10683 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10684         .lookup         = btrfs_lookup,
10685         .permission     = btrfs_permission,
10686         .update_time    = btrfs_update_time,
10687 };
10688
10689 static const struct file_operations btrfs_dir_file_operations = {
10690         .llseek         = generic_file_llseek,
10691         .read           = generic_read_dir,
10692         .iterate_shared = btrfs_real_readdir,
10693         .unlocked_ioctl = btrfs_ioctl,
10694 #ifdef CONFIG_COMPAT
10695         .compat_ioctl   = btrfs_compat_ioctl,
10696 #endif
10697         .release        = btrfs_release_file,
10698         .fsync          = btrfs_sync_file,
10699 };
10700
10701 static const struct extent_io_ops btrfs_extent_io_ops = {
10702         /* mandatory callbacks */
10703         .submit_bio_hook = btrfs_submit_bio_hook,
10704         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10705         .merge_bio_hook = btrfs_merge_bio_hook,
10706         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10707
10708         /* optional callbacks */
10709         .fill_delalloc = run_delalloc_range,
10710         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10711         .writepage_start_hook = btrfs_writepage_start_hook,
10712         .set_bit_hook = btrfs_set_bit_hook,
10713         .clear_bit_hook = btrfs_clear_bit_hook,
10714         .merge_extent_hook = btrfs_merge_extent_hook,
10715         .split_extent_hook = btrfs_split_extent_hook,
10716 };
10717
10718 /*
10719  * btrfs doesn't support the bmap operation because swapfiles
10720  * use bmap to make a mapping of extents in the file.  They assume
10721  * these extents won't change over the life of the file and they
10722  * use the bmap result to do IO directly to the drive.
10723  *
10724  * the btrfs bmap call would return logical addresses that aren't
10725  * suitable for IO and they also will change frequently as COW
10726  * operations happen.  So, swapfile + btrfs == corruption.
10727  *
10728  * For now we're avoiding this by dropping bmap.
10729  */
10730 static const struct address_space_operations btrfs_aops = {
10731         .readpage       = btrfs_readpage,
10732         .writepage      = btrfs_writepage,
10733         .writepages     = btrfs_writepages,
10734         .readpages      = btrfs_readpages,
10735         .direct_IO      = btrfs_direct_IO,
10736         .invalidatepage = btrfs_invalidatepage,
10737         .releasepage    = btrfs_releasepage,
10738         .set_page_dirty = btrfs_set_page_dirty,
10739         .error_remove_page = generic_error_remove_page,
10740 };
10741
10742 static const struct address_space_operations btrfs_symlink_aops = {
10743         .readpage       = btrfs_readpage,
10744         .writepage      = btrfs_writepage,
10745         .invalidatepage = btrfs_invalidatepage,
10746         .releasepage    = btrfs_releasepage,
10747 };
10748
10749 static const struct inode_operations btrfs_file_inode_operations = {
10750         .getattr        = btrfs_getattr,
10751         .setattr        = btrfs_setattr,
10752         .listxattr      = btrfs_listxattr,
10753         .permission     = btrfs_permission,
10754         .fiemap         = btrfs_fiemap,
10755         .get_acl        = btrfs_get_acl,
10756         .set_acl        = btrfs_set_acl,
10757         .update_time    = btrfs_update_time,
10758 };
10759 static const struct inode_operations btrfs_special_inode_operations = {
10760         .getattr        = btrfs_getattr,
10761         .setattr        = btrfs_setattr,
10762         .permission     = btrfs_permission,
10763         .listxattr      = btrfs_listxattr,
10764         .get_acl        = btrfs_get_acl,
10765         .set_acl        = btrfs_set_acl,
10766         .update_time    = btrfs_update_time,
10767 };
10768 static const struct inode_operations btrfs_symlink_inode_operations = {
10769         .get_link       = page_get_link,
10770         .getattr        = btrfs_getattr,
10771         .setattr        = btrfs_setattr,
10772         .permission     = btrfs_permission,
10773         .listxattr      = btrfs_listxattr,
10774         .update_time    = btrfs_update_time,
10775 };
10776
10777 const struct dentry_operations btrfs_dentry_operations = {
10778         .d_delete       = btrfs_dentry_delete,
10779         .d_release      = btrfs_dentry_release,
10780 };