ext4_i_disksize_lock_race_fix.patch

   1 ext4: Properly update i_disksize.
   2
   3 From: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
   4
   5 With delayed allocation we use i_data_sem to update i_disksize.  We need
   6 to update i_disksize only if the new size specified is greater than the
   7 current value and we need to make sure we don't race with other
   8 i_disksize update.  With delayed allocation we will switch to the
   9 write_begin function for non-delayed allocation if we are low on free
  10 blocks.  That means write_begin function for non-delayed allocation also
  11 needs to use the same locking.
  12
  13 We also need to check and update i_disksize even if the new size is less
  14 that inode.i_size because of delayed allocation.
  15
  16 Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
  17 Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
  18 ---
  19  fs/ext4/ext4.h    |   11 ++++++++++
  20  fs/ext4/extents.c |    9 ++++----
  21  fs/ext4/inode.c   |   61 ++++++++++++++++++++++++++++++++---------------------
  22  3 files changed, 53 insertions(+), 28 deletions(-)
  23
  24 Index: linux-2.6.27-rc5/fs/ext4/ext4.h
  25 ===================================================================
  26 --- linux-2.6.27-rc5.orig/fs/ext4/ext4.h        2008-08-28 23:25:05.000000000 -0700
  27 +++ linux-2.6.27-rc5/fs/ext4/ext4.h     2008-08-28 23:28:51.000000000 -0700
  28 @@ -1219,6 +1219,17 @@ do {                                                             \
  29  #define EXT4_FREEBLOCKS_WATERMARK 0
  30  #endif
  31
  32 +static inline void ext4_update_i_disksize(struct inode *inode, loff_t newsize)
  33 +{
  34 +       /*
  35 +        * XXX: replace with spinlock if seen contended -bzzz
  36 +        */
  37 +       down_write(&EXT4_I(inode)->i_data_sem);
  38 +       if (newsize > EXT4_I(inode)->i_disksize)
  39 +               EXT4_I(inode)->i_disksize = newsize;
  40 +       up_write(&EXT4_I(inode)->i_data_sem);
  41 +       return ;
  42 +}
  43
  44  /*
  45   * Inodes and files operations
  46 Index: linux-2.6.27-rc5/fs/ext4/extents.c
  47 ===================================================================
  48 --- linux-2.6.27-rc5.orig/fs/ext4/extents.c     2008-08-28 23:25:09.000000000 -0700
  49 +++ linux-2.6.27-rc5/fs/ext4/extents.c  2008-08-28 23:28:51.000000000 -0700
  50 @@ -2877,10 +2877,11 @@ static void ext4_falloc_update_inode(str
  51          * Update only when preallocation was requested beyond
  52          * the file size.
  53          */
  54 -       if (!(mode & FALLOC_FL_KEEP_SIZE) &&
  55 -                               new_size > i_size_read(inode)) {
  56 -               i_size_write(inode, new_size);
  57 -               EXT4_I(inode)->i_disksize = new_size;
  58 +       if (!(mode & FALLOC_FL_KEEP_SIZE)) {
  59 +               if (new_size > i_size_read(inode))
  60 +                       i_size_write(inode, new_size);
  61 +               if (new_size > EXT4_I(inode)->i_disksize)
  62 +                       ext4_update_i_disksize(inode, new_size);
  63         }
  64
  65  }
  66 Index: linux-2.6.27-rc5/fs/ext4/inode.c
  67 ===================================================================
  68 --- linux-2.6.27-rc5.orig/fs/ext4/inode.c       2008-08-28 23:28:42.000000000 -0700
  69 +++ linux-2.6.27-rc5/fs/ext4/inode.c    2008-08-28 23:28:51.000000000 -0700
  70 @@ -1434,16 +1434,18 @@ static int ext4_ordered_write_end(struct
  71         ret = ext4_jbd2_file_inode(handle, inode);
  72
  73         if (ret == 0) {
  74 -               /*
  75 -                * generic_write_end() will run mark_inode_dirty() if i_size
  76 -                * changes.  So let's piggyback the i_disksize mark_inode_dirty
  77 -                * into that.
  78 -                */
  79                 loff_t new_i_size;
  80
  81                 new_i_size = pos + copied;
  82 -               if (new_i_size > EXT4_I(inode)->i_disksize)
  83 -                       EXT4_I(inode)->i_disksize = new_i_size;
  84 +               if (new_i_size > EXT4_I(inode)->i_disksize) {
  85 +                       ext4_update_i_disksize(inode, new_i_size);
  86 +                       /* We need to mark inode dirty even if
  87 +                        * new_i_size is less that inode->i_size
  88 +                        * bu greater than i_disksize.(hint delalloc)
  89 +                        */
  90 +                       ext4_mark_inode_dirty(handle, inode);
  91 +               }
  92 +
  93                 ret2 = generic_write_end(file, mapping, pos, len, copied,
  94                                                         page, fsdata);
  95                 copied = ret2;
  96 @@ -1468,8 +1470,14 @@ static int ext4_writeback_write_end(stru
  97         loff_t new_i_size;
  98
  99         new_i_size = pos + copied;
 100 -       if (new_i_size > EXT4_I(inode)->i_disksize)
 101 -               EXT4_I(inode)->i_disksize = new_i_size;
 102 +       if (new_i_size > EXT4_I(inode)->i_disksize) {
 103 +               ext4_update_i_disksize(inode, new_i_size);
 104 +               /* We need to mark inode dirty even if
 105 +                * new_i_size is less that inode->i_size
 106 +                * bu greater than i_disksize.(hint delalloc)
 107 +                */
 108 +               ext4_mark_inode_dirty(handle, inode);
 109 +       }
 110
 111         ret2 = generic_write_end(file, mapping, pos, len, copied,
 112                                                         page, fsdata);
 113 @@ -1494,6 +1502,7 @@ static int ext4_journalled_write_end(str
 114         int ret = 0, ret2;
 115         int partial = 0;
 116         unsigned from, to;
 117 +       loff_t new_i_size;
 118
 119         from = pos & (PAGE_CACHE_SIZE - 1);
 120         to = from + len;
 121 @@ -1508,11 +1517,12 @@ static int ext4_journalled_write_end(str
 122                                 to, &partial, write_end_fn);
 123         if (!partial)
 124                 SetPageUptodate(page);
 125 -       if (pos+copied > inode->i_size)
 126 +       new_i_size = pos + copied;
 127 +       if (new_i_size > inode->i_size)
 128                 i_size_write(inode, pos+copied);
 129         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
 130 -       if (inode->i_size > EXT4_I(inode)->i_disksize) {
 131 -               EXT4_I(inode)->i_disksize = inode->i_size;
 132 +       if (new_i_size > EXT4_I(inode)->i_disksize) {
 133 +               ext4_update_i_disksize(inode, new_i_size);
 134                 ret2 = ext4_mark_inode_dirty(handle, inode);
 135                 if (!ret)
 136                         ret = ret2;
 137 @@ -2227,18 +2237,9 @@ static int ext4_da_get_block_write(struc
 138                 if (disksize > i_size_read(inode))
 139                         disksize = i_size_read(inode);
 140                 if (disksize > EXT4_I(inode)->i_disksize) {
 141 -                       /*
 142 -                        * XXX: replace with spinlock if seen contended -bzzz
 143 -                        */
 144 -                       down_write(&EXT4_I(inode)->i_data_sem);
 145 -                       if (disksize > EXT4_I(inode)->i_disksize)
 146 -                               EXT4_I(inode)->i_disksize = disksize;
 147 -                       up_write(&EXT4_I(inode)->i_data_sem);
 148 -
 149 -                       if (EXT4_I(inode)->i_disksize == disksize) {
 150 -                               ret = ext4_mark_inode_dirty(handle, inode);
 151 -                               return ret;
 152 -                       }
 153 +                       ext4_update_i_disksize(inode, disksize);
 154 +                       ret = ext4_mark_inode_dirty(handle, inode);
 155 +                       return ret;
 156                 }
 157                 ret = 0;
 158         }
 159 @@ -2567,6 +2568,13 @@ retry:
 160                 unlock_page(page);
 161                 ext4_journal_stop(handle);
 162                 page_cache_release(page);
 163 +               /*
 164 +                * block_write_begin may have instantiated a few blocks
 165 +                * outside i_size.  Trim these off again. Don't need
 166 +                * i_size_read because we hold i_mutex.
 167 +                */
 168 +               if (pos + len > inode->i_size)
 169 +                       vmtruncate(inode, inode->i_size);
 170         }
 171
 172         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
 173 @@ -2647,6 +2655,11 @@ static int ext4_da_write_end(struct file
 174                                 EXT4_I(inode)->i_disksize = new_i_size;
 175                         }
 176                         up_write(&EXT4_I(inode)->i_data_sem);
 177 +                       /* We need to mark inode dirty even if
 178 +                        * new_i_size is less that inode->i_size
 179 +                        * bu greater than i_disksize.(hint delalloc)
 180 +                        */
 181 +                       ext4_mark_inode_dirty(handle, inode);
 182                 }
 183         }
 184         ret2 = generic_write_end(file, mapping, pos, len, copied,