releases/2.6.22.14/ocfs2-fix-write-performance-regression.patch

   1 From mark.fasheh@oracle.com  Mon Nov 19 09:43:53 2007
   2 From: Mark Fasheh <mark.fasheh@oracle.com>
   3 Date: Wed, 14 Nov 2007 13:33:27 -0800
   4 Subject: ocfs2: fix write() performance regression
   5 To: gregkh@suse.de
   6 Cc: stable@kernel.org
   7 Message-ID: <20071114213327.GN28607@ca-server1.us.oracle.com>
   8 Content-Disposition: inline
   9
  10
  11 From: Mark Fasheh <mark.fasheh@oracle.com>
  12
  13 ocfs2: fix write() performance regression
  14
  15 patch 4e9563fd55ff4479f2b118d0757d121dd0cfc39c in mainline.
  16
  17 On file systems which don't support sparse files, Ocfs2_map_page_blocks()
  18 was reading blocks on appending writes. This caused write performance to
  19 suffer dramatically. Fix this by detecting an appending write on a nonsparse
  20 fs and skipping the read.
  21
  22 Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
  23 Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
  24
  25 ---
  26  fs/ocfs2/aops.c |   24 +++++++++++++++++++++++-
  27  1 file changed, 23 insertions(+), 1 deletion(-)
  28
  29 --- a/fs/ocfs2/aops.c
  30 +++ b/fs/ocfs2/aops.c
  31 @@ -661,6 +661,27 @@ static void ocfs2_clear_page_regions(str
  32  }
  33
  34  /*
  35 + * Nonsparse file systems fully allocate before we get to the write
  36 + * code. This prevents ocfs2_write() from tagging the write as an
  37 + * allocating one, which means ocfs2_map_page_blocks() might try to
  38 + * read-in the blocks at the tail of our file. Avoid reading them by
  39 + * testing i_size against each block offset.
  40 + */
  41 +static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
  42 +                                unsigned int block_start)
  43 +{
  44 +       u64 offset = page_offset(page) + block_start;
  45 +
  46 +       if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
  47 +               return 1;
  48 +
  49 +       if (i_size_read(inode) > offset)
  50 +               return 1;
  51 +
  52 +       return 0;
  53 +}
  54 +
  55 +/*
  56   * Some of this taken from block_prepare_write(). We already have our
  57   * mapping by now though, and the entire write will be allocating or
  58   * it won't, so not much need to use BH_New.
  59 @@ -711,7 +732,8 @@ int ocfs2_map_page_blocks(struct page *p
  60                         if (!buffer_uptodate(bh))
  61                                 set_buffer_uptodate(bh);
  62                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
  63 -                    (block_start < from || block_end > to)) {
  64 +                          ocfs2_should_read_blk(inode, page, block_start) &&
  65 +                          (block_start < from || block_end > to)) {
  66                         ll_rw_block(READ, 1, &bh);
  67                         *wait_bh++=bh;
  68                 }