On Thu 04-08-11 08:19:16, Christoph Hellwig wrote:
> On Thu, Aug 04, 2011 at 02:07:24PM +0200, Jan Kara wrote:
> > Hmm, BTW, shouldn't the call to xfs_flush_pages() in
> > xfs_file_buffered_aio_write() be converted to an asynchronous one? I don't
> > quite see a point in waiting for io completion... Generally, flushing of
> > the inode there seems of limited usefulness to me since that inode could be
> > just a tiny victim not holding much delayallocated blocks.
> This comes from commit
> xfs: make inode flush at ENOSPC synchronous
> from Dave - before that it was asynchronous and in weird context, so
> it seems we defintively need it to be synchronous.
From the changelog it seems it needs to be synchronous in the sense that
we don't offload flushing to a different thread as we used to. Also the
reason why previously flushing didn't work was that we held page locks and
IO lock but it's not the case in xfs_file_buffered_aio_write() anymore. So
filemap_flush() still looks like an appropriate thing to me.
> I agree that just flushing this inode seems like a rather odd handling
> for ENOSPC. It's even more odd as we already use the big hammer before
> in when we git ENOSPC in ->write_begin. The only thing I can imagine is
> that this is the last attempt to get anything freed.
OK, I'll leave it there then. I just wonder whether I should convert it
to filemap_flush() or to filemap_write_and_wait()...
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR