On Fri, Jun 07, 2013 at 11:42:43AM +0930, Rusty Russell wrote:

"Michael S. Tsirkin" [off-list ref] writes:

quoted

For small packets we can simplify xmit processing by linearizing buffers
with the header: most packets seem to have enough head room we can use
for this purpose.

Since some older hypervisors (e.g. qemu before version 1.5)
required that header is the first s/g element,
we need a feature bit for this.

OK, we know this is horrible.  But I will sleep better knowing that we
this feature need never make it into a final 1.0 spec, since it can be
assumed at that point...

Nod. Though if we want to require this for all devices,
virtio-blk scsi command passthrough will need to change -
I sent a spec patch a while ago
	virtio-spec: add field for scsi command size
any comments on it?

quoted

 	pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
+	if (vi->mergeable_rx_bufs)
+		hdr_len = sizeof hdr->mhdr;
+	else
+		hdr_len = sizeof hdr->hdr;
+
+	can_push = vi->any_header_sg &&
+		!((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
+		!skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;

Idle thought: how often does this fail?

I think it's mostly doesn't fail in my testing.

It's probably a good idea to add a counter here, then
if it starts triggering we can optimize.

I think things like skb_header_cloned depend on guest config
really, e.g. tcpdump running on the interface in guest can cause this.

 Would it suck if we copied
headers which didn't let us prepend data?

I think it will - copies are generally best avoided,
and header is easily 1K of data.

 Or could we bump
dev->hard_header_len appropriately?

Needs some thought, though from experience it's a pain.

Thanks,
Rusty.