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

From: Jason Gunthorpe
Date: Fri Nov 06 2020 - 07:55:10 EST


On Fri, Nov 06, 2020 at 11:27:59AM +0100, Daniel Vetter wrote:
> On Fri, Nov 6, 2020 at 11:01 AM Daniel Vetter <daniel@xxxxxxxx> wrote:
> >
> > On Fri, Nov 6, 2020 at 5:08 AM John Hubbard <jhubbard@xxxxxxxxxx> wrote:
> > >
> > > On 11/5/20 4:49 AM, Jason Gunthorpe wrote:
> > > > On Thu, Nov 05, 2020 at 10:25:24AM +0100, Daniel Vetter wrote:
> > > >>> /*
> > > >>> * 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.
> > > >
> > > > It means ioremap can't create an IO page PUD, it has to be broken up.
> > > >
> > > > Does ioremap even create anything larger than PTEs?
> >
> > gpu drivers also tend to use vmf_insert_pfn* directly, so we can do
> > on-demand paging and move buffers around. From what I glanced for
> > lowest level we to the pte_mkspecial correctly (I think I convinced
> > myself that vm_insert_pfn does that), but for pud/pmd levels it seems
> > just yolo.
>
> So I dug around a bit more and ttm sets PFN_DEV | PFN_MAP to get past
> the various pft_t_devmap checks (see e.g. vmf_insert_pfn_pmd_prot()).
> x86-64 has ARCH_HAS_PTE_DEVMAP, and gup.c seems to handle these
> specially, but frankly I got totally lost in what this does.

The fact vmf_insert_pfn_pmd_prot() has all those BUG_ON's to prevent
putting VM_PFNMAP pages into the page tables seems like a big red
flag.

The comment seems to confirm what we are talking about here:

/*
* If we had pmd_special, we could avoid all these restrictions,
* but we need to be consistent with PTEs and architectures that
* can't support a 'special' bit.
*/

ie without the ability to mark special we can't block fast gup and
anyone who does O_DIRECT on these ranges will crash the kernel when it
tries to convert a IO page into a struct page.

Should be easy enough to directly test?

Putting non-struct page PTEs into a VMA without setting VM_PFNMAP just
seems horribly wrong to me.

Jason