GUP guarantees wrt to userspace mappings redesign

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Mon, May 02, 2016 at 01:15:13PM +0200, Jerome Glisse wrote:
> On Mon, May 02, 2016 at 01:41:19PM +0300, Kirill A. Shutemov wrote:
> > Other thing I would like to discuss is if there's a problem on vfio side.
> > To me it looks like vfio expects guarantee from get_user_pages() which it
> > doesn't provide: obtaining pin on the page doesn't guarantee that the page
> > is going to remain mapped into userspace until the pin is gone.
> > 
> > Even with THP COW regressing fixed, vfio would stay fragile: any
> > MADV_DONTNEED/fork()/mremap()/whatever what would make vfio expectation
> > broken.
> > 
> 
> Well i don't think it is fair/accurate assessment of get_user_pages(), page
> must remain mapped to same virtual address until pin is gone. I am ignoring
> mremap() as it is a scient decision from userspace and while virtual address
> change in that case, the pined page behind should move with the mapping.
> Same of MADV_DONTNEED. I agree that get_user_pages() is broken after fork()
> but this have been the case since dawn of time, so it is something expected.
> 
> If not vfio, then direct-io, have been expecting this kind of behavior for
> long time, so i see this as part of get_user_pages() guarantee.
> 
> Concerning vfio, not providing this guarantee will break countless number of
> workload. Thing like qemu/kvm allocate anonymous memory and hand it over to
> the guest kernel which presents it as memory. Now a device driver inside the
> guest kernel need to get bus mapping for a given (guest) page, which from
> host point of view means a mapping from anonymous page to bus mapping but
> for guest to keep accessing the same page the anonymous mapping (ie a
> specific virtual address on the host side) must keep pointing to the same
> page. This have been the case with get_user_pages() until now, so whether
> we like it or not we must keep that guarantee.
> 
> This kind of workload knows that they can't do mremap()/fork()/... and keep
> that guarantee but they at expect existing guarantee and i don't think we
> can break that.

Quick look around:

 - I don't see any check page_count() around __replace_page() in uprobes,
   so it can easily replace pinned page.

 - KSM has the page_count() check, there's still race wrt GUP_fast: it can
   take the pin between the check and establishing new pte entry.

 - khugepaged: the same story as with KSM.

I don't see how we can deliver on the guarantee, especially with lockless
GUP_fast.

Or am I missing something important?

-- 
 Kirill A. Shutemov