Re: [PATCH v2 01/17] mm/gup: Fixup p*_access_permitted()

[Al Viro]

On Fri, Dec 15, 2017 at 06:28:36PM -0800, Linus Torvalds wrote:
> On Fri, Dec 15, 2017 at 5:25 PM, Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
> >
> > I think the reason we needed VMA and PTE checks was the
> > get_user_pages_fast() path not having a VMA.
> 
> That is indeed the point of get_user_pages_fast(): no vma lookup, no
> locking, just "do the default case as streamlined as possible".
> 
> But part of it is also that we should fall back to the slow case if
> the fast case doesn't work (eg because the page isn't there or
> whatever).
> 
> So what we could do - perhaps - is to just make get_user_pages_fast()
> check whether any of the protection key bits are set, and fail for
> that case.

FWIW, a good description of fast path in get_user_pages_fast() is
"simulate a TLB miss", the slow path being "... and go for simulated
page fault if TLB miss would have escalated to #PF".

Treating protection key bits as "escalate to page fault and let that
deal with the checks" should be fine - page fault handler must
cope with the page actually being present in page tables anyway, for
obvious reasons...