Re: [PATCH v5 05/15] mm/frame-vector: Use FOLL_LONGTERM

[Daniel Vetter]

On Wed, Nov 04, 2020 at 10:44:56AM -0800, John Hubbard wrote:
> On 11/4/20 10:17 AM, Jason Gunthorpe wrote:
> > On Wed, Nov 04, 2020 at 04:41:19PM +0000, Christoph Hellwig wrote:
> > > On Wed, Nov 04, 2020 at 04:37:58PM +0000, Christoph Hellwig wrote:
> > > > On Wed, Nov 04, 2020 at 05:26:58PM +0100, Daniel Vetter wrote:
> > > > > What we're discussing is whether gup_fast and pup_fast also obey this,
> > > > > or fall over and can give you the struct page that's backing the
> > > > > dma_mmap_* memory. Since the _fast variant doesn't check for
> > > > > vma->vm_flags, and afaict that's the only thing which closes this gap.
> > > > > And like you restate, that would be a bit a problem. So where's that
> > > > > check which Jason&me aren't spotting?
> > > > 
> > > > remap_pte_range uses pte_mkspecial to set up the PTEs, and gup_pte_range
> > > > errors out on pte_special.  Of course this only works for the
> > > > CONFIG_ARCH_HAS_PTE_SPECIAL case, for other architectures we do have
> > > > a real problem.
> > > 
> > > Except that we don't really support pte-level gup-fast without
> > > CONFIG_ARCH_HAS_PTE_SPECIAL, and in fact all architectures selecting
> > > HAVE_FAST_GUP also select ARCH_HAS_PTE_SPECIAL, so we should be fine.

Thanks for the explainer. I guess I can go back to _fast and instead
adjust the commit message to explain why that's all fine.

> > Mm, I thought it was probably the special flag..
> > 
> > Knowing that CONFIG_HAVE_FAST_GUP can't be set without
> > CONFIG_ARCH_HAS_PTE_SPECIAL is pretty insightful, can we put that in
> > the Kconfig?
> > 
> > config HAVE_FAST_GUP
> >          depends on MMU
> >          depends on ARCH_HAS_PTE_SPECIAL
> >          bool
> > 
> Well, the !CONFIG_ARCH_HAS_PTE_SPECIAL case points out in a comment that
> gup-fast is not *completely* unavailable there, so I don't think you want
> to shut it off like that:
> 
> /*
>  * If we can't determine whether or not a pte is special, then fail immediately
>  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
>  * to be special.
>  *
>  * For a futex to be placed on a THP tail page, get_futex_key requires a
>  * get_user_pages_fast_only implementation that can pin pages. Thus it's still
>  * useful to have gup_huge_pmd even if we can't operate on ptes.
>  */

We support hugepage faults in gpu drivers since recently, and I'm not
seeing a pud_mkhugespecial anywhere. So not sure this works, but probably
just me missing something again.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch