Re: [net-next PATCH v2 3/5] virtio_net: Add XDP support

On 16-11-22 06:58 AM, Michael S. Tsirkin wrote:

On Tue, Nov 22, 2016 at 12:27:03AM -0800, John Fastabend wrote:

quoted

On 16-11-21 03:20 PM, Michael S. Tsirkin wrote:

quoted

On Sat, Nov 19, 2016 at 06:50:33PM -0800, John Fastabend wrote:

quoted

From: Shrijeet Mukherjee <redacted>

This adds XDP support to virtio_net. Some requirements must be
met for XDP to be enabled depending on the mode. First it will
only be supported with LRO disabled so that data is not pushed
across multiple buffers. The MTU must be less than a page size
to avoid having to handle XDP across multiple pages.

If mergeable receive is enabled this first series only supports
the case where header and data are in the same buf which we can
check when a packet is received by looking at num_buf. If the
num_buf is greater than 1 and a XDP program is loaded the packet
is dropped and a warning is thrown. When any_header_sg is set this
does not happen and both header and data is put in a single buffer
as expected so we check this when XDP programs are loaded. Note I
have only tested this with Linux vhost backend.

If big packets mode is enabled and MTU/LRO conditions above are
met then XDP is allowed.

A follow on patch can be generated to solve the mergeable receive
case with num_bufs equal to 2. Buffers greater than two may not
be handled has easily.

I would very much prefer support for other layouts without drops
before merging this.
header by itself can certainly be handled by skipping it.
People wanted to use that e.g. for zero copy.

OK fair enough I'll do this now rather than push it out.

Hi Michael,

The header skip logic however complicates the xmit handling a fair
amount. Specifically when we release the buffers after xmit then
both the hdr and data portions need to be released which requires
some tracking.

Is the header split logic actually in use somewhere today? It looks
like its not being used in Linux case. And zero copy RX is currently as
best I can tell not supported anywhere so I would prefer not to
complicate the XDP path at the moment with a possible future feature.

quoted

quoted

Anything else can be handled by copying the packet.

Any idea how to test this? At the moment I have some code to linearize
the data in all cases with more than a single buffer. But wasn't clear
to me which features I could negotiate with vhost/qemu to get more than
a single buffer in the receive path.

Thanks,
John