Christoph,

On Sat, Aug 27, 2011 at 02:14:09PM +0800, Christoph Hellwig wrote:

Right now ->write_inode has no way to safely return a EAGAIN without explicitly
redirtying the inode, as we would lose the dirty state otherwise.  Most
filesystems get this wrong, but XFS makes heavy use of it to avoid blocking
the flusher thread when ->write_inode hits contentended inode locks.  A
contended ilock is something XFS can hit very easibly when extending files, as
the data I/O completion handler takes the lock to update the size, and the
->write_inode call can race with it fairly easily if writing enough data
in one go so that the completion for the first write come in just before
we call ->write_inode.

Change the handling of this case to use requeue_io for a quick retry instead
of redirty_tail, which keeps moving out the dirtied_when data and thus keeps
delaying the writeout more and more with every failed attempt to get the lock.

Yeah redirty_tail() does have the problem of possibly delay inodes for
too long time. However, you know requeue_io() always has the danger of
triggering busy wb_writeback() loops in corner cases.

For example, nfs_write_inode()/nfs_commit_unstable_pages() often
redirty the inode without doing anything (hence no any progress, a
prerequisite for busy loops) depending on the in flight writes, which
unfortunately further depends on _external_ network/server states..
That means some stalled network/sever state could lead to busy loops
in NFS clients.

The alternative solution may be to firstly apply the attached patch,
and change this one to:

  -			redirty_tail(inode, wb);
  +			requeue_io_wait(inode, wb);

Thanks,
Fengguang