Re: [PATCH net-next v18 07/14] memory-provider: dmabuf devmem memory provider

On Tue, Aug 13, 2024 at 4:39 AM Mina Almasry [off-list ref] wrote:

On Mon, Aug 12, 2024 at 8:15 PM Jakub Kicinski [off-list ref] wrote:

quoted

BTW, Mina, the core should probably also check that XDP isn't installed
before / while the netmem is bound to a queue.

Sorry if noob question, but what is the proper check for this? I tried
adding this to net_devmem_bind_dmabuf_to_queue():

if (xdp_rxq_info_is_reg(&rxq->xdp_rxq))
                 return -EEXIST;

But quickly found out that in  netif_alloc_rx_queues() we initialize
all the rxq->xdp_rxq to state REGISTERED regardless whether xdp is
installed or not, so this check actually fails.

Worthy of note is that GVE holds an instance of xdp_rxq_info in
gve_rx_ring, and seems to use that for its xdp information, not the
one that hangs off of netdev_rx_queue in core.

To elaborate further, in order to disable binding dmabuf and XDP on
the same rx queue for GVE, AFAIT the check would need to be inside of
GVE. Inside of GVE I'd check if gve_priv->xdp_prog is installed, and
check if the gve_rx_ring->xdp_info is registered. If so, then the rx
queue is XDP enabled, and should not be bound to dmabuf. I think that
would work.

At the moment I can't think of a check inside of core that would be
compatible with GVE, but above you clearly are specifically asking for
a check in core. Any pointers to what you have in mind would be
appreciated here, but I'll try to take a deeper look.

Additionally, my understanding of XDP is limited, but why do we want
to disable it? My understanding is that XDP is a kernel bypass that
hands the data directly to userspace. In theory at least there should
be no issue binding dmabuf to a queue, then getting the data in the
queue via an XDP program instead of via TCP sockets or io uring. Is
there some fundamental reason why dmabuf and XDP are incompatible?

--
Thanks,
Mina

-- 
Thanks,
Mina