Re: [PATCH] mm/x86/pat: Only untrack the pfn range if unmap region

[David Hildenbrand]

On 23.07.24 23:44, Jason Gunthorpe wrote:
> On Tue, Jul 23, 2024 at 11:36:51PM +0200, David Hildenbrand wrote:
> > I wonder if we could just let relevant users do the PAT handling manually:
> > I'm also not sure how many remap_pfn_range users end up triggering VM_PAT
> > code although they don't really have to (just because they happen to cover a
> > full VMA)?
> >
> > One nasty thing is fork(), I was wondering if relevant users really rely on
> > that or if we could force these VMAs to simply not get copied during fork.
> > During fork() we have to "duplicate" the reservation ...
>
> I admit I barely understand what x86 uses this PAT stuff for -
> allowing WC mappings is part of it?

I'm confused by all of that, but I think yes :)

> If yes, then RDMA would expect WC PFN MAP VMAs to copy their WCness
> during fork..

Okay, that answers that question: we have to support fork().

During fork() we call vm_ops->open when cloning the VMA. So the backend 
can realize when all users (VMAs, whatsoever) are gone to undo any 
registration. Maybe that's a path forward. But still no clue on details 
... one important step might be figuring out who exactly really relies 
on that VM_PAT handling.

--
Cheers,

David / dhildenb