Re: [PATCH v2 33/35] arm64/mm: attempt speculative mm faults first

[Mike Rapoport]

On Mon, Feb 07, 2022 at 09:39:19AM -0800, Michel Lespinasse wrote:
> On Tue, Feb 01, 2022 at 10:58:03AM +0200, Mike Rapoport wrote:
> > On Mon, Jan 31, 2022 at 12:07:29AM -0800, Michel Lespinasse wrote:
> > > On Sun, Jan 30, 2022 at 11:13:26AM +0200, Mike Rapoport wrote:
> > > > The speculative page fault implementation here (and for PowerPC as well)
> > > > looks very similar to x86. Can we factor it our rather than copy 3 (or
> > > > more) times?
> > > 
> > > In each arch, the speculative code was written along the lines of the
> > > existing non-speculative code, so that behavior would be unchanged
> > > when speculation succeeds.
> > > 
> > > Now each arch's existing, non-speculative code paths are quite similar,
> > > but they do have small differences as to how they implement various
> > > permission checks, protection keys and the like. The same small
> > > differences end up being reflected in the new speculative code paths.
> > > 
> > > I agree it would be nice if this code could be unified between archs,
> > > but IMO this should start with the existing non-speculative code -
> > > I don't think it would make sense to try unifying the new speculative
> > > code while trying to follow the behavior of the non-unified old
> > > non-speculative code paths...
> > 
> > Then maybe this unification can be done as the ground work for the
> > speculative page fault handling?
> 
> I feel like this is quite unrelated, and that introducing such
> artificial dependencies is a bad work habit we have here in linux MM...

The reduction of the code duplication in page fault handlers per se is
indeed not very related to SPF work, but since the SPF patches increase
the code duplication, I believe that the refactoring that prevents this
additional code duplication is related and is in scope of this work.
 
> That said, unifying the PF code between archs would be an interesting
> project on its own. The way I see it, there could be a unified page
> fault handler, with some arch specific parts defined as inline
> functions.  I can see myself making an x86/arm64/powerpc initial
> proposal if there is enough interest for it, but I'm not sure how
> extending it to more exotic archs would go - I think this would have
> to involve arch maintainers at least for testing purposes, and I'm not
> sure if they'd have any bandwidth for such a project...

There is no need to convert all architectures and surely not at once.
The parts of page fault handler that are shared by several architectures
can live under #ifdef ARCH_WANTS_GENERIC_PAGE_FAULT or something like this.

> --
> Michel "walken" Lespinasse

-- 
Sincerely yours,
Mike.