fix-reservation-release-on-invalidatepage-for-delalloc-fs

   1 ext4: fix reservation release on invalidatepage for delalloc fs
   2
   3 From: Lukas Czerner <lczerner@redhat.com>
   4
   5 On delalloc enabled file system on invalidatepage operation
   6 in ext4_da_page_release_reservation() we want to clear the delayed
   7 buffer and remove the extent covering the delayed buffer from the extent
   8 status tree.
   9
  10 However currently there is a bug where on the systems with page size >
  11 block size we will always remove extents from the start of the page
  12 regardless where the actual delayed buffers are positioned in the page.
  13 This leads to the errors like this:
  14
  15 EXT4-fs warning (device loop0): ext4_da_release_space:1225:
  16 ext4_da_release_space: ino 13, to_free 1 with only 0 reserved data
  17 blocks
  18
  19 This however can cause data loss on writeback time if the file system is
  20 in ENOSPC condition because we're releasing reservation for someones
  21 else delayed buffer.
  22
  23 Fix this by only removing extents that corresponds to the part of the
  24 page we want to invalidate.
  25
  26 This problem is reproducible by the following fio receipt (however I was
  27 only able to reproduce it with fio-2.1 or older.
  28
  29 [global]
  30 bs=8k
  31 iodepth=1024
  32 iodepth_batch=60
  33 randrepeat=1
  34 size=1m
  35 directory=/mnt/test
  36 numjobs=20
  37 [job1]
  38 ioengine=sync
  39 bs=1k
  40 direct=1
  41 rw=randread
  42 filename=file1:file2
  43 [job2]
  44 ioengine=libaio
  45 rw=randwrite
  46 direct=1
  47 filename=file1:file2
  48 [job3]
  49 bs=1k
  50 ioengine=posixaio
  51 rw=randwrite
  52 direct=1
  53 filename=file1:file2
  54 [job5]
  55 bs=1k
  56 ioengine=sync
  57 rw=randread
  58 filename=file1:file2
  59 [job7]
  60 ioengine=libaio
  61 rw=randwrite
  62 filename=file1:file2
  63 [job8]
  64 ioengine=posixaio
  65 rw=randwrite
  66 filename=file1:file2
  67 [job10]
  68 ioengine=mmap
  69 rw=randwrite
  70 bs=1k
  71 filename=file1:file2
  72 [job11]
  73 ioengine=mmap
  74 rw=randwrite
  75 direct=1
  76 filename=file1:file2
  77
  78 Signed-off-by: Lukas Czerner <lczerner@redhat.com>
  79 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
  80 Reviewed-by: Jan Kara <jack@suse.cz>
  81 Cc: stable@vger.kernel.org
  82 ---
  83  fs/ext4/inode.c | 15 ++++++++++++---
  84  1 file changed, 12 insertions(+), 3 deletions(-)
  85
  86 diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
  87 index 0554b0b..46f4a49 100644
  88 --- a/fs/ext4/inode.c
  89 +++ b/fs/ext4/inode.c
  90 @@ -1342,7 +1342,7 @@ static void ext4_da_page_release_reservation(struct page *page,
  91                                              unsigned int offset,
  92                                              unsigned int length)
  93  {
  94 -       int to_release = 0;
  95 +       int to_release = 0, contiguous_blks = 0;
  96         struct buffer_head *head, *bh;
  97         unsigned int curr_off = 0;
  98         struct inode *inode = page->mapping->host;
  99 @@ -1363,14 +1363,23 @@ static void ext4_da_page_release_reservation(struct page *page,
 100
 101                 if ((offset <= curr_off) && (buffer_delay(bh))) {
 102                         to_release++;
 103 +                       contiguous_blks++;
 104                         clear_buffer_delay(bh);
 105 +               } else if (contiguous_blks) {
 106 +                       lblk = page->index <<
 107 +                              (PAGE_CACHE_SHIFT - inode->i_blkbits);
 108 +                       lblk += (curr_off >> inode->i_blkbits) -
 109 +                               contiguous_blks;
 110 +                       ext4_es_remove_extent(inode, lblk, contiguous_blks);
 111 +                       contiguous_blks = 0;
 112                 }
 113                 curr_off = next_off;
 114         } while ((bh = bh->b_this_page) != head);
 115
 116 -       if (to_release) {
 117 +       if (contiguous_blks) {
 118                 lblk = page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
 119 -               ext4_es_remove_extent(inode, lblk, to_release);
 120 +               lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
 121 +               ext4_es_remove_extent(inode, lblk, contiguous_blks);
 122         }
 123
 124         /* If we have released all the blocks belonging to a cluster, then we
 125 --
 126 1.8.3.1
 127
 128 --
 129 To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
 130 the body of a message to majordomo@vger.kernel.org
 131 More majordomo info at  http://vger.kernel.org/majordomo-info.html
 132