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

[Christoph Lameter]

On Thu, 30 Apr 2009, Ingo Molnar wrote:

> > Yes but sometimes you are already irq safe and such a fallback
> > would create significant irq/enable/disable stack operations etc
> > overhead for architectures that are using the fallback.
>
> It's a fallback slowpath - non-x86 architectures should still fill
> in a real implementation of course.

Arch code cannot provide an effective implementation since they
always have to assume that interupts need to be disabled if we stay with
the current implementation.

> So we first have to see the list of architectures that _cannot_
> implement an irq-safe op here via a single machine instruction.
> x86, ia64 and powerpc should be fine.

Look at Ia64, sparc, s/390, powerpc. They can fall back to atomic ops but
those are very ineffective on some of these platforms. Since these are
performance critical they will need to be optimized depending on the
context of their use in the core.