Re: [PATCH rdma-next] RDMA/restrack: Delay QP deletion till all users are gone

[Leon Romanovsky <leon@kernel.org>]

On Tue, Apr 20, 2021 at 12:25:41PM -0300, Jason Gunthorpe wrote:
> On Tue, Apr 20, 2021 at 04:06:03PM +0300, Leon Romanovsky wrote:
> > On Tue, Apr 20, 2021 at 09:39:50AM -0300, Jason Gunthorpe wrote:
> > > On Sun, Apr 18, 2021 at 04:37:35PM +0300, Leon Romanovsky wrote:
> > > > From: Shay Drory <shayd@nvidia.com>
> > > > 
> > > > Currently, in case of QP, the following use-after-free is possible:
> > > > 
> > > > 	cpu0				cpu1
> > > >  res_get_common_dumpit()
> > > >  rdma_restrack_get()
> > > >  fill_res_qp_entry()
> > > > 				ib_destroy_qp_user()
> > > > 				 rdma_restrack_del()
> > > > 				 qp->device->ops.destroy_qp()
> > > >   ib_query_qp()
> > > >   qp->device->ops.query_qp()
> > > > 
> > > > This is because rdma_restrack_del(), in case of QP, isn't waiting until
> > > > all users are gone.
> > > > 
> > > > Fix it by making rdma_restrack_del() wait until all users are gone for
> > > > QPs as well.
> > > > 
> > > > Fixes: 13ef5539def7 ("RDMA/restrack: Count references to the verbs objects")
> > > > Signed-off-by: Shay Drory <shayd@nvidia.com>
> > > > Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> > > >  drivers/infiniband/core/restrack.c | 2 +-
> > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > 
> > > > diff --git a/drivers/infiniband/core/restrack.c b/drivers/infiniband/core/restrack.c
> > > > index ffabaf327242..def0c5b0efe9 100644
> > > > +++ b/drivers/infiniband/core/restrack.c
> > > > @@ -340,7 +340,7 @@ void rdma_restrack_del(struct rdma_restrack_entry *res)
> > > >  	rt = &dev->res[res->type];
> > > >  
> > > >  	old = xa_erase(&rt->xa, res->id);
> > > > -	if (res->type == RDMA_RESTRACK_MR || res->type == RDMA_RESTRACK_QP)
> > > > +	if (res->type == RDMA_RESTRACK_MR)
> > > >  		return;
> > > 
> > > Why is MR skipping this?
> > 
> > I don't remember the justification for RDMA_RESTRACK_MR || RDMA_RESTRACK_QP check.
> > My guess that it is related to the allocation flow (both are not converted) and
> > have internal objects to the drivers.
> 
> Well, I don't understand why QP has a bug but MR would not have the
> same one??

I didn't understand when reviewed either, but decided to post it anyway
to get possible explanation for this RDMA_RESTRACK_MR || RDMA_RESTRACK_QP
check.

Thanks

> 
> Jason