Re: [patch 3/7] mm: make page writeback obey cpuset constraints

[David Rientjes <rientjes@google.com>]

On Thu, 30 Oct 2008, Peter Zijlstra wrote:

> On Tue, 2008-10-28 at 12:18 -0700, David Rientjes wrote:
> > On Tue, 28 Oct 2008, Peter Zijlstra wrote:
> > 
> > > > +	is_subset = cpuset_populate_dirty_limits(dl, &dirtyable_memory,
> > > > +						 &nr_mapped, nodes);
> > > > +	if (!is_subset) {
> > > > +		dl->nr_dirty = global_page_state(NR_FILE_DIRTY);
> > > > +		dl->nr_unstable = global_page_state(NR_UNSTABLE_NFS);
> > > > +		dl->nr_writeback = global_page_state(NR_WRITEBACK);
> > > > +		dirtyable_memory = determine_dirtyable_memory();
> > > > +		nr_mapped = global_page_state(NR_FILE_MAPPED) +
> > > > +			global_page_state(NR_ANON_PAGES);
> > > > +	} else
> > > > +		dirtyable_memory -= highmem_dirtyable_memory(nodes,
> > > > +							dirtyable_memory);
> > > 
> > > Why not fold that all into cpuset_populate_dirty_limits() ?
> > > 
> > 
> > cpuset_populate_dirty_limits() is a no-op on !CONFIG_CPUSETS kernels.
> 
> Right, humm. Maybe introduce a populate_dirty_limits() and differentiate
> that between CONFIG_CPUSETS and not, and make it do everything.
> 
> That would get rid of this fudge I think, no?
> 

I agree it would look much cleaner and I actually did that originally, but 
it requires adding #ifdef's for CONFIG_CPUSETS to mm/page-writeback.c 
since highmem_dirtyable_memory() is static.

There's actually nothing cpuset-specific about 
cpuset_populate_dirty_limits() and it could be used for !CONFIG_CPUSETS 
kernels, but there's a performance benefit of using the global ZVC values 
for the page stats as opposed to iterating over each node and adding their 
respective values.  So is_subset is only non-zero when the passed nodemask 
excludes system nodes (the cpuset case), and we must iterate the node ZVC 
values.

So perhaps the solution is to introduce populate_nodemask_dirty_limits() 
and populate_global_dirty_limits() both in mm/page-writeback.c?