fix-races-of-writeback-with-punch-hole-and-zero-range

   1 ext4: fix races of writeback with punch hole and zero range
   2
   3 From: Jan Kara <jack@suse.com>
   4
   5 When doing delayed allocation, update of on-disk inode size is postponed
   6 until IO submission time. However hole punch or zero range fallocate
   7 calls can end up discarding the tail page cache page and thus on-disk
   8 inode size would never be properly updated.
   9
  10 Make sure the on-disk inode size is updated before truncating page
  11 cache.
  12
  13 Signed-off-by: Jan Kara <jack@suse.com>
  14 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
  15 ---
  16  fs/ext4/ext4.h    |  3 +++
  17  fs/ext4/extents.c |  5 +++++
  18  fs/ext4/inode.c   | 35 ++++++++++++++++++++++++++++++++++-
  19  3 files changed, 42 insertions(+), 1 deletion(-)
  20
  21 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
  22 index c19ff61ccbdf..c86546efec30 100644
  23 --- a/fs/ext4/ext4.h
  24 +++ b/fs/ext4/ext4.h
  25 @@ -2691,6 +2691,9 @@ static inline int ext4_update_inode_size(struct inode *inode, loff_t newsize)
  26         return changed;
  27  }
  28
  29 +int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
  30 +                                     loff_t len);
  31 +
  32  struct ext4_group_info {
  33         unsigned long   bb_state;
  34         struct rb_root  bb_free_root;
  35 diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
  36 index 892245a55c53..9977b557d50d 100644
  37 --- a/fs/ext4/extents.c
  38 +++ b/fs/ext4/extents.c
  39 @@ -4843,6 +4843,11 @@ static long ext4_zero_range(struct file *file, loff_t offset,
  40                  * released from page cache.
  41                  */
  42                 down_write(&EXT4_I(inode)->i_mmap_sem);
  43 +               ret = ext4_update_disksize_before_punch(inode, offset, len);
  44 +               if (ret) {
  45 +                       up_write(&EXT4_I(inode)->i_mmap_sem);
  46 +                       goto out_dio;
  47 +               }
  48                 /* Now release the pages and zero block aligned part of pages */
  49                 truncate_pagecache_range(inode, start, end - 1);
  50                 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
  51 diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
  52 index 36ad45906d26..f72212ef1fee 100644
  53 --- a/fs/ext4/inode.c
  54 +++ b/fs/ext4/inode.c
  55 @@ -3517,6 +3517,35 @@ int ext4_can_truncate(struct inode *inode)
  56  }
  57
  58  /*
  59 + * We have to make sure i_disksize gets properly updated before we truncate
  60 + * page cache due to hole punching or zero range. Otherwise i_disksize update
  61 + * can get lost as it may have been postponed to submission of writeback but
  62 + * that will never happen after we truncate page cache.
  63 + */
  64 +int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
  65 +                                     loff_t len)
  66 +{
  67 +       handle_t *handle;
  68 +       loff_t size = i_size_read(inode);
  69 +
  70 +       WARN_ON(!mutex_is_locked(&inode->i_mutex));
  71 +       if (offset > size || offset + len < size)
  72 +               return 0;
  73 +
  74 +       if (EXT4_I(inode)->i_disksize >= size)
  75 +               return 0;
  76 +
  77 +       handle = ext4_journal_start(inode, EXT4_HT_MISC, 1);
  78 +       if (IS_ERR(handle))
  79 +               return PTR_ERR(handle);
  80 +       ext4_update_i_disksize(inode, size);
  81 +       ext4_mark_inode_dirty(handle, inode);
  82 +       ext4_journal_stop(handle);
  83 +
  84 +       return 0;
  85 +}
  86 +
  87 +/*
  88   * ext4_punch_hole: punches a hole in a file by releaseing the blocks
  89   * associated with the given offset and length
  90   *
  91 @@ -3594,9 +3623,13 @@ int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
  92         last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
  93
  94         /* Now release the pages and zero block aligned part of pages*/
  95 -       if (last_block_offset > first_block_offset)
  96 +       if (last_block_offset > first_block_offset) {
  97 +               ret = ext4_update_disksize_before_punch(inode, offset, length);
  98 +               if (ret)
  99 +                       goto out_dio;
 100                 truncate_pagecache_range(inode, first_block_offset,
 101                                          last_block_offset);
 102 +       }
 103
 104         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 105                 credits = ext4_writepage_trans_blocks(inode);
 106 --
 107 2.1.4
 108
 109