[ltt-dev] [PATCH] Fix dirty page accounting in redirty_page_for_writepage()

[Christoph Lameter]

On Sat, 2 May 2009, Mathieu Desnoyers wrote:

> Modifying mmzone.h "struct per_cpu_pageset" so it uses percpu dynamic
> allocation seems to be the far-reaching part of the modification (of a
> subsystem I'm not completely familiar with, including NUMA special
> cases). Is there any patch already doing this kind of modification
> floating around ?

Yes. Its was part of the original cpu alloc patchset. One reason to do cpu
alloc was to clean up the current problem with a having to allocate for
NR_CPUS in the page allocator. That in turn would result in the ability to
use the fast percpu ops to optimize the allocator paths. Hopefully we can
end up with page allocator alloc/free operations that do not need to
disable interrupts.

> I'd be glad to give a try at some percpu_add_return counters experiments
> if percpu struct per_cpu_pageset allocation happens to be already
> available.

Look at the cpu alloc patchset following the link that I posted.