Re: [PATCH v16 12/16] IB, arm64: untag user pointers in ib_uverbs_(re)reg_mr()

[Andrey Konovalov <andreyknvl@google.com>]

On Tue, Jun 4, 2019 at 3:02 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
>
> On Tue, Jun 04, 2019 at 02:45:32PM +0200, Andrey Konovalov wrote:
> > On Tue, Jun 4, 2019 at 2:27 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
> > >
> > > On Tue, Jun 04, 2019 at 02:18:19PM +0200, Andrey Konovalov wrote:
> > > > On Mon, Jun 3, 2019 at 7:46 PM Jason Gunthorpe <jgg@ziepe.ca> wrote:
> > > > >
> > > > > On Mon, Jun 03, 2019 at 06:55:14PM +0200, Andrey Konovalov wrote:
> > > > > > This patch is a part of a series that extends arm64 kernel ABI to allow to
> > > > > > pass tagged user pointers (with the top byte set to something else other
> > > > > > than 0x00) as syscall arguments.
> > > > > >
> > > > > > ib_uverbs_(re)reg_mr() use provided user pointers for vma lookups (through
> > > > > > e.g. mlx4_get_umem_mr()), which can only by done with untagged pointers.
> > > > > >
> > > > > > Untag user pointers in these functions.
> > > > > >
> > > > > > Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
> > > > > >  drivers/infiniband/core/uverbs_cmd.c | 4 ++++
> > > > > >  1 file changed, 4 insertions(+)
> > > > > >
> > > > > > diff --git a/drivers/infiniband/core/uverbs_cmd.c b/drivers/infiniband/core/uverbs_cmd.c
> > > > > > index 5a3a1780ceea..f88ee733e617 100644
> > > > > > +++ b/drivers/infiniband/core/uverbs_cmd.c
> > > > > > @@ -709,6 +709,8 @@ static int ib_uverbs_reg_mr(struct uverbs_attr_bundle *attrs)
> > > > > >       if (ret)
> > > > > >               return ret;
> > > > > >
> > > > > > +     cmd.start = untagged_addr(cmd.start);
> > > > > > +
> > > > > >       if ((cmd.start & ~PAGE_MASK) != (cmd.hca_va & ~PAGE_MASK))
> > > > > >               return -EINVAL;
> > > > >
> > > > > I feel like we shouldn't thave to do this here, surely the cmd.start
> > > > > should flow unmodified to get_user_pages, and gup should untag it?
> > > > >
> > > > > ie, this sort of direction for the IB code (this would be a giant
> > > > > patch, so I didn't have time to write it all, but I think it is much
> > > > > saner):
> > > >
> > > > Hi Jason,
> > > >
> > > > ib_uverbs_reg_mr() passes cmd.start to mlx4_get_umem_mr(), which calls
> > > > find_vma(), which only accepts untagged addresses. Could you explain
> > > > how your patch helps?
> > >
> > > That mlx4 is just a 'weird duck', it is not the normal flow, and I
> > > don't think the core code should be making special consideration for
> > > it.
> >
> > How do you think we should do untagging (or something else) to deal
> > with this 'weird duck' case?
>
> mlx4 should handle it around the call to find_vma like other patches
> do, ideally as part of the cast from a void __user * to the unsigned
> long that find_vma needs

So essentially what we had a few versions ago
(https://lkml.org/lkml/2019/4/30/785) plus changing unsigned longs to
__user * across all IB code? I think the second part is something
that's not related to this series and needs to be done separately. I
can move untagging back to mlx4_get_umem_mr() though.

Catalin, you've initially asked to to move untagging out of
mlx4_get_umem_mr(), do you have any comments on this?

>
> Jason