releases/2.6.32.17/0038-ext4-don-t-scan-accumulate-more-pages-than-mballoc-w.patch

   1 From 2f4283aff3e5415fa36cbf81aa2a6247bfbb0527 Mon Sep 17 00:00:00 2001
   2 From: Eric Sandeen <sandeen@redhat.com>
   3 Date: Sun, 30 May 2010 22:49:52 -0400
   4 Subject: ext4: don't scan/accumulate more pages than mballoc will allocate
   5
   6 commit c445e3e0a5c2804524dec6e55f66d63f6bc5bc3e upstream (as of v2.6.34-git13)
   7
   8 There was a bug reported on RHEL5 that a 10G dd on a 12G box
   9 had a very, very slow sync after that.
  10
  11 At issue was the loop in write_cache_pages scanning all the way
  12 to the end of the 10G file, even though the subsequent call
  13 to mpage_da_submit_io would only actually write a smallish amt; then
  14 we went back to the write_cache_pages loop ... wasting tons of time
  15 in calling __mpage_da_writepage for thousands of pages we would
  16 just revisit (many times) later.
  17
  18 Upstream it's not such a big issue for sys_sync because we get
  19 to the loop with a much smaller nr_to_write, which limits the loop.
  20
  21 However, talking with Aneesh he realized that fsync upstream still
  22 gets here with a very large nr_to_write and we face the same problem.
  23
  24 This patch makes mpage_add_bh_to_extent stop the loop after we've
  25 accumulated 2048 pages, by setting mpd->io_done = 1; which ultimately
  26 causes the write_cache_pages loop to break.
  27
  28 Repeating the test with a dirty_ratio of 80 (to leave something for
  29 fsync to do), I don't see huge IO performance gains, but the reduction
  30 in cpu usage is striking: 80% usage with stock, and 2% with the
  31 below patch.  Instrumenting the loop in write_cache_pages clearly
  32 shows that we are wasting time here.
  33
  34 Eventually we need to change mpage_da_map_pages() also submit its I/O
  35 to the block layer, subsuming mpage_da_submit_io(), and then change it
  36 call ext4_get_blocks() multiple times.
  37
  38 Signed-off-by: Eric Sandeen <sandeen@redhat.com>
  39 Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
  40 Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
  41 ---
  42  fs/ext4/inode.c |    9 +++++++++
  43  1 file changed, 9 insertions(+)
  44
  45 --- a/fs/ext4/inode.c
  46 +++ b/fs/ext4/inode.c
  47 @@ -2361,6 +2361,15 @@ static void mpage_add_bh_to_extent(struc
  48         sector_t next;
  49         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
  50
  51 +       /*
  52 +        * XXX Don't go larger than mballoc is willing to allocate
  53 +        * This is a stopgap solution.  We eventually need to fold
  54 +        * mpage_da_submit_io() into this function and then call
  55 +        * ext4_get_blocks() multiple times in a loop
  56 +        */
  57 +       if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
  58 +               goto flush_it;
  59 +
  60         /* check if thereserved journal credits might overflow */
  61         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
  62                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {