[Cluster-devel] [PATCH v5 00/12] gfs2: Fix mmap + page fault deadlocks

[Andreas Gruenbacher]

On Wed, Aug 18, 2021 at 11:49 PM Linus Torvalds
<torvalds at linux-foundation.org> wrote:
> [ Sorry for the delay, I was on the road and this fell through the cracks ]

No harm done, I was busy enough implementing your previous suggestions.

> On Mon, Aug 16, 2021 at 12:14 PM Andreas Gruenbacher
> <agruenba at redhat.com> wrote:
> >
> > On Tue, Aug 3, 2021 at 9:45 PM Linus Torvalds
> > <torvalds at linux-foundation.org> wrote:
> > >
> > > Hmm. Have you tried to figure out why that "still returns 0" happens?
> >
> > The call stack is:
> >
> > gup_pte_range
> > gup_pmd_range
> > gup_pud_range
> > gup_p4d_range
> > gup_pgd_range
> > lockless_pages_from_mm
> > internal_get_user_pages_fast
> > get_user_pages_fast
> > iov_iter_get_pages
> > __bio_iov_iter_get_pages
> > bio_iov_iter_get_pages
> > iomap_dio_bio_actor
> > iomap_dio_actor
> > iomap_apply
> > iomap_dio_rw
> > gfs2_file_direct_write
> >
> > In gup_pte_range, pte_special(pte) is true and so we return 0.
>
> Ok, so that is indeed something that the fast-case can't handle,
> because some of the special code wants to have the mm_lock so that it
> can look at the vma flags (eg "vm_normal_page()" and friends.
>
> That said, some of these cases even the full GUP won't ever handle,
> simply because a mapping doesn't necessarily even _have_ a 'struct
> page' associated with it if it's a VM_IO mapping.
>
> So it turns out that you can't just always do
> fault_in_iov_iter_readable() and then assume that you can do
> iov_iter_get_pages() and repeat until successful.
>
> We could certainly make get_user_pages_fast() handle a few more cases,
> but I get the feeling that we need to have separate error cases for
> EFAULT - no page exists - and the "page exists, but cannot be mapped
> as a 'struct page'" case.

Hmm, what if GUP is made to skip VM_IO vmas without adding anything to
the pages array? That would match fault_in_iov_iter_writeable, which
is modeled after __mm_populate and which skips VM_IO and VM_PFNMAP
vmas.

The other strategy I've added is to scale back the page fault windows
to a single page if faulting in multiple pages didn't help, and to
give up if the I/O operation still fails after that. So pathological
cases won't loop indefinitely anymore at least.

> I also do still think that even regardless of that, we want to just
> add a FOLL_NOFAULT flag that just disables calling handle_mm_fault(),
> and then you can use the regular get_user_pages().
>
> That at least gives us the full _normal_ page handling stuff.

And it does fix the generic/208 failure.

Thanks,
Andreas