On Fri, Jan 27, 2023 at 06:41:51AM +0000, Long Li wrote:
> > Subject: Re: [PATCH] RDMA/mana_ib: Prevent array underflow in
> > mana_ib_create_qp_raw()
> >
> > On Thu, Jan 26, 2023 at 02:00:45PM +0300, Dan Carpenter wrote:
> > > On Thu, Jan 26, 2023 at 12:18:46PM +0200, Leon Romanovsky wrote:
> > > > On Tue, Jan 24, 2023 at 06:20:54PM +0300, Dan Carpenter wrote:
> > > > > The "port" comes from the user and if it is zero then the:
> > > > >
> > > > > ndev = mc->ports[port - 1];
> > > > >
> > > > > assignment does an out of bounds read. I have changed the if
> > > > > statement to fix this and to mirror how it is done in
> > > > > mana_ib_create_qp_rss().
> > > > >
> > > > > Fixes: 0266a177631d ("RDMA/mana_ib: Add a driver for Microsoft
> > > > > Azure Network Adapter")
> > > > > Signed-off-by: Dan Carpenter <error27@xxxxxxxxx>
> > > > > ---
> > > > > drivers/infiniband/hw/mana/qp.c | 2 +-
> > > > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/drivers/infiniband/hw/mana/qp.c
> > > > > b/drivers/infiniband/hw/mana/qp.c index ea15ec77e321..54b61930a7fd
> > > > > 100644
> > > > > --- a/drivers/infiniband/hw/mana/qp.c
> > > > > +++ b/drivers/infiniband/hw/mana/qp.c
> > > > > @@ -289,7 +289,7 @@ static int mana_ib_create_qp_raw(struct ib_qp
> > > > > *ibqp, struct ib_pd *ibpd,
> > > > >
> > > > > /* IB ports start with 1, MANA Ethernet ports start with 0 */
> > > > > port = ucmd.port;
> > > > > - if (ucmd.port > mc->num_ports)
> > > > > + if (port < 1 || port > mc->num_ports)
> > > >
> > > > Why do I see port in mana_ib_create_qp? It should come from ib_qp_init_attr.
> > >
> > > I am so confused by this question. Are you asking me?
> >
> > I asked *@microsoft folks.
> >
> > > This is the _raw function.
> >
> > _raw comes from QP type, it is not raw (basic) in a sense you imagine.
> >
> > > I'm now sure what mana_ib_create_qp() has to do with it.
> >
> > All create QP calls come through same verbs interface.
> > ib_create_qp_user->create_qp->.create_qp->mana_ib_create_qp-
> > >mana_ib_create_qp_raw
>
> MANA requires passing a port number when creating a RAW QP on a RDMA(Ethernet) port.
> At the hardware layer, the RDMA port and ethernet port share the same hardware resources,
> the port number needs to be known in advance when QP is created. If we don't' specify the port,
> the QP needs to take all the ports on the MANA device, some of them may have been assigned to
> Ethernet usage and can't be used for RDMA.
>
> The reason is that unlike Nvidia CX hardware, MANA doesn't support bifurcation (for RAW QP) at the hardware level.
> [https://www.dpdk.org/wp-content/uploads/sites/35/2018/06/Mellanox-bifurcated-driver-model.pdf]
> To support RAW QP on a hardware port, we need to know the port number before configuring it on the hardware.
> And Ethernet can't use this port if a RAW QP is created on it. The coordination needs to be done in software.
>
> In production environment, there are multiple ports on the same MANA device assigned to a VM. Customer can
> configure some of the ports for Ethernet and some for RDMA/DPDK.
>
> I have investigated using the port_num in struct ib_qp_init_attr, but it seems it can't be used for this purpose
> because the port needs to be specified by the user-mode.
And this is the most problematic part for me. I don't think that it is
RDMA user responsibility to dig in netdev port numbers.
Jason, what do you think? It looks like layering violation.
Thanks
>
> Thanks,
> Long
>
> >
> > >
> > > The port comes from ib_copy_from_udata() which is just a wrapper
> > > around copy_from_user().
> >
> > Right, and it shouldn't.
> >
> > >
> > > regards,
> > > dan carpenter
> > >
