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