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

[Michel Lespinasse <michel@lespinasse.org>]

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.

Right, I agree that having the vma copy on-stack is not ideal.

I think what really should be done, is to copy just the attributes of
the vma that will be needed during the page fault. Things like vm_mm,
vm_page_prot, vm_flags, vm_ops, vm_pgoff, vm_file, vm_private_data,
vm_policy. We definitely do not need rbtree and rmap fields such as
vm_prev, vm_next, vm_rb, rb_subtree_gap, shared, anon_vma_chain etc...

The reason I did things this way, is because changing the entire fault
handler to use attributes stored in struct vm_fault, rather than in
the original vma, would be quite intrusive. I think it would be a
reasonable step to consider once there is agreement on the rest of the
speculative fault patch set, but it's practical doing it before then.