Re: [PATCH 6/7] writeback: separate starting of sync vs opportunistic writeback

[Chris Mason <chris.mason@oracle.com>]

On Tue, Sep 15, 2009 at 04:01:45PM +0200, Jan Kara wrote:
> On Tue 15-09-09 09:08:29, Christoph Hellwig wrote:
> > On Tue, Sep 15, 2009 at 03:04:19PM +0200, Jens Axboe wrote:
> > > 
> > > >   Let's have a look at the flags in wbc:
> > > >   nonblocking - Currently only set by direct callers of ->writepage() BUT
> > > >                 originally wb_kupdate() and background_writeout() also
> > > >                 set this flag. Since filesystems and write_cache_pages()
> > > > 		use the flag we should set it for equivalent writeouts as
> > > >                 well. This should be fixed...
> > > 
> > > Since this is all handled by the dedicated thread now, dropping the
> > > nonblocking bit was on purpose. What would the point be, except for
> > > stopping pdflush being blocked on request allocation?
> > 
> > Note that this flag just caused utter mess traditionally.  btrfs decided
> > to ignore it completely and ext4 partially.  Removing this check in
> > XFS increases large bufferd write loads massively.
> > 
> > Just half-removing it is a bad idea, though - if you don't set it
> > anymore please kill it entirely.
>   The nonblocking flag is still set for writeback done for memory reclaim.
> OTOH the only real consumer of this flag now seems to be
> __block_write_full_page() which does trylock_buffer() in case of
> nonblocking writeback. I'm undecided whether it makes sence or not.

Ugh, making sense is tricky to say.  If __block_write_full_page
does a lock_buffer() instead of a trylock_buffer(), and ext3 is mounted in
data=ordered mode then it is very possible that we'll end up with a
dirty page with locked buffers.

The buffers will have been locked by ext3 data=ordered writeback and
they won't unlock until the IO is done.

We probably don't want kswapd waiting on that writeback.

-chris