RE: [PATCH rdma-rc] RDMA/core: Fix pkey and port assignment in get_new_pps

["Marciniszyn, Mike" <mike.marciniszyn@intel.com>]

> On Thu, Feb 27, 2020 at 02:07:10PM +0000, Marciniszyn, Mike wrote:
> > >
> > > From: Maor Gottlieb <maorg@mellanox.com>
> > >
> > > When port is part of the modify mask, then we should take
> > > it from the qp_attr and not from the old pps. Same for PKEY.
> > >
> > > Cc: <stable@vger.kernel.org>
> > > Fixes: 1dd017882e01 ("RDMA/core: Fix protection fault in
> > > get_pkey_idx_qp_list")
> > > Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
> > > Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
> > > ---
> > >  drivers/infiniband/core/security.c | 12 ++++++++----
> > >  1 file changed, 8 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/drivers/infiniband/core/security.c
> > > b/drivers/infiniband/core/security.c
> > > index b9a36ea244d4..2d5608315dc8 100644
> > > --- a/drivers/infiniband/core/security.c
> > > +++ b/drivers/infiniband/core/security.c
> > > @@ -340,11 +340,15 @@ static struct ib_ports_pkeys
> *get_new_pps(const
> > > struct ib_qp *qp,
> > >  		return NULL;
> > >
> > >  	if (qp_attr_mask & IB_QP_PORT)
> > > -		new_pps->main.port_num =
> > > -			(qp_pps) ? qp_pps->main.port_num : qp_attr-
> > > >port_num;
> > > +		new_pps->main.port_num = qp_attr->port_num;
> > > +	else if (qp_pps)
> > > +		new_pps->main.port_num = qp_pps->main.port_num;
> > > +
> > >  	if (qp_attr_mask & IB_QP_PKEY_INDEX)
> > > -		new_pps->main.pkey_index = (qp_pps) ? qp_pps-
> > > >main.pkey_index :
> > > -						      qp_attr->pkey_index;
> > > +		new_pps->main.pkey_index = qp_attr->pkey_index;
> > > +	else if (qp_pps)
> > > +		new_pps->main.pkey_index = qp_pps->main.pkey_index;
> > > +
> > >  	if ((qp_attr_mask & IB_QP_PKEY_INDEX) && (qp_attr_mask &
> > > IB_QP_PORT))
> > >  		new_pps->main.state = IB_PORT_PKEY_VALID;
> > >
> >
> > I agree with this aspect of the patch and it does fix the panic, because the
> correct unit
> > is gotten from qp_pps.
> >
> > My issue is that the new_pps->main.state will come back as 0, and the
> insert routine will drop any new pkey index update.
> >
> > The sequence I'm concerned about is:
> >
> > 0x71 attr mask with both pkey index and port.
> >
> > A ulp decides to change the pkey index and does an 0x51 modify without
> setting the unit.
> >
> > I see new_pps->main.state being returned as 0 and port_pkey_list_insert()
> will early out.
> 
> I see, maybe we can store the main.state in qps and restore it from there?
> 

I don't think we need to go that far.

I think all we need to do is 

	if ((qp_attr_mask & IB_QP_PKEY_INDEX) && (qp_attr_mask & IB_QP_PORT))
		new_pps->main.state = IB_PORT_PKEY_VALID;
	 else if ((qp_attr_mask & (IB_QP_PKEY_INDEX | IB_QP_PORT)) && qp_pps) {
                             /*
		 * one of the attributes modified and already copied above.
		 *
		 * correct the state based on qp_pps state
		 */
		if (qp_pps->main.state != IB_PORT_PKEY_NOT_VALID)
			new_pps->main.state = IB_PORT_PKEY_VALID;
	}

I'm ok will a follow-up patch after Maor's patch to fix remaining issues.

Right now, all of our upstream testing (5.6 rc3+, stables) is dead because of 1dd017882e01.

Mike