Re: [RFC PATCH 11/37] x86/mm: attempt speculative mm faults first

[Michel Lespinasse]

On Wed, Apr 07, 2021 at 04:35:28PM +0100, Matthew Wilcox wrote:
> On Wed, Apr 07, 2021 at 04:48:44PM +0200, Peter Zijlstra wrote:
> > On Tue, Apr 06, 2021 at 06:44:36PM -0700, Michel Lespinasse wrote:
> > > --- a/arch/x86/mm/fault.c
> > > +++ b/arch/x86/mm/fault.c
> > > @@ -1219,6 +1219,8 @@ void do_user_addr_fault(struct pt_regs *regs,
> > >  	struct mm_struct *mm;
> > >  	vm_fault_t fault;
> > >  	unsigned int flags = FAULT_FLAG_DEFAULT;
> > > +	struct vm_area_struct pvma;
> > 
> > That's 200 bytes on-stack... I suppose that's just about acceptible, but
> > perhaps we need a comment in struct vm_area_struct to make people aware
> > this things lives on-stack and size really is an issue now.
> 
> Michel's gone off and done his own thing here.

I don't think that is an entirely fair representation. First we are
both aware of each other's work, there is no working in dark caves here.
But also, I don't even consider this patchset to be entirely my thing;
most of the main building blocks come from prior proposals before mine.

> The rest of us (Laurent, Liam & I) are working on top of the maple tree
> which shrinks vm_area_struct by five pointers, so just 160 bytes.

The idea of evaluating maple tree and speculative faults as a bundle
is actually worrying to me. I think both ideas are interesting and
worth looking into on their own, but I'm not convinced that they have
much to do with each other.

> Also, our approach doesn't involve copying VMAs in order to handle a fault.

See my other reply to Peter's message - copying VMAs is a convenient
way to reduce the size of the patchset, but it's not fundamental to
the approach at all.