Re: [PATCH 01/45] writeback: reduce calls to global_page_state in balance_dirty_pages()

[Wu Fengguang <fengguang.wu@intel.com>]

On Mon, Oct 12, 2009 at 05:07:10PM +0800, Peter Zijlstra wrote:
> On Mon, 2009-10-12 at 09:26 +0800, Wu Fengguang wrote:
> > On Sun, Oct 11, 2009 at 07:25:17PM +0800, Peter Zijlstra wrote:
> > > On Sun, 2009-10-11 at 18:50 +0800, Wu Fengguang wrote:
> > > > 
> > > > Sorry for the confusion, but I mean, filesystems have to limit
> > > > nr_writeback (directly or indirectly via the block io queue),
> > > > otherwise it either hit nr_dirty to 0 (with the loop), or let
> > > > nr_writeback grow out of control (without the loop).
> > > 
> > > Doesn't this require the writeback queue to have a limit < dirty_thresh?
> > 
> > Yes, this is the key (open) issue. For now we have nothing to limit
> > 
> >         nr_writeback < dirty_thresh
> > 
> > > Or more specifically, for the bdi case:
> > > 
> > >  bdi_dirty + bdi_writeback + bdi_unstable <= bdi_thresh
> > > 
> > > we require that the writeback queue be smaller than bdi_thresh, which
> > > could be quite difficult, since bdi_thresh can easily be 0.
> > 
> > We could apply a MIN_BDI_DIRTY_THRESH. Because the bdi threshold is
> > estimated from writeback events, so bdi_thresh must be non-zero to
> > allow some writeback pages in flight :)
> 
> Not really, suppose you have 1000 NFS clients, of which you only use a
> hand full at a time.
> 
> Then the bdi_thresh will be 0 for most of them, and only when you switch
> to one it'll start growing. But it's perfectly reasonable to expect
> bdi_thresh=0 to work. It just reverts to sync behaviour, we write out
> everything and block until they're all gone from writeback state.

Ah I see. We still do writeback when bdi_thresh=0, with any
application blocked in balance_dirty_pages().

> MIN_BDI_DIRTY_THRESH != 0, will have a side effect of imposing a max
> number of BDIs on the system, I'm not sure you want to go there.

OK that's not a good idea.

> > > Without observing the bdi_thresh constraint we can have:
> > > 
> > >   \Sum_(over bdis) writeback_queue_size
> > > 
> > > dirty pages outstanding, which could be significantly higher than
> > > dirty_thresh.
> >  
> > Yes.  Maybe we could do some per-bdi and/or global writeback wait
> > queue (ie. some generalized version of the patch 20: NFS: introduce
> > writeback wait queue).
> > 
> > The per-bdi writeback queue size should ideally be proportional to its
> > available writeback bandwidth. MIN_BDI_DIRTY_THRESH could be defined
> > to (2*bdi_writeback_bandwidth) or something close. And if the resulted
> > bdi limits turn out to be too large for a small memory system, we just
> > let the global limit kick in. For such small memory systems, it is
> > very likely there are only one bdi. So it is not likely to lose
> > fairness to base its limits on available memory instead of device
> > capability.
> 
> I'm not seeing why. By simply keeping that loop we're good again, and
> can have a writeback queue that works well in the saturated case.

OK it looks better to keep the loop. The memory tight systems may go
into the nr_dirty=0 situation, but it may not be an urgent problem
(its nr_dirty will be small anyway).

Thanks,
Fengguang