Virtualization - May 2009 - [RFC] virtio: orphan skbs if we're relying on timer to free them

We check for finished xmit skbs on every xmit, or on a timer (unless
the host promises to force an interrupt when the xmit ring is empty).
This can penalize userspace tasks which fill their sockbuf.  Not much
difference with TSO, but measurable with large numbers of packets.

There are a finite number of packets which can be in the transmission
queue.  We could fire the timer more than every 100ms, but that would
just hurt performance for a corner case.  This seems neatest.

With interrupt when Tx ring empty:
				Seconds	TxPkts	TxIRQs
 1G TCP Guest->Host:		3.76	32833	32758
 1M normal pings:		111	1000008	997463
 1M 1k pings (-l 120):		55	1000007	488920

Without interrupt, without orphaning:
 1G TCP Guest->Host:		3.64	32806	1
 1M normal pings:		106	1000008	1
 1M 1k pings (-l 120):		68	1000005	1

With orphaning:
 1G TCP Guest->Host:		3.86	32821	1
 1M normal pings:		102	1000007	1
 1M 1k pings (-l 120):		43	1000005	1

Signed-off-by: Rusty Russell <rusty at rustcorp.com.au>
---
 drivers/net/virtio_net.c |    5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -522,6 +522,11 @@ static int start_xmit(struct sk_buff *sk
 {
 	struct virtnet_info *vi = netdev_priv(dev);
 
+	/* We queue a limited number; don't let that delay writers if
+	 * we are slow in getting tx interrupt. */
+	if (!vi->free_in_tasklet)
+		skb_orphan(skb);
+
 again:
 	/* Free up any pending old buffers before queueing new ones. */
 	free_old_xmit_skbs(vi);

From: Rusty Russell <rusty at rustcorp.com.au>
Date: Mon, 18 May 2009 22:18:47 +0930
> We check for finished xmit skbs on every xmit, or on a timer (unless
> the host promises to force an interrupt when the xmit ring is empty).
> This can penalize userspace tasks which fill their sockbuf.  Not much
> difference with TSO, but measurable with large numbers of packets.
> 
> There are a finite number of packets which can be in the transmission
> queue.  We could fire the timer more than every 100ms, but that would
> just hurt performance for a corner case.  This seems neatest.
 ...
> Signed-off-by: Rusty Russell <rusty at rustcorp.com.au>
If this is so great for virtio it would also be a great idea
universally, but we don't do it.

What you're doing by orphan'ing is creating a situation where a single
UDP socket can loop doing sends and monopolize the TX queue of a
device.  The only control we have over a sender for fairness in
datagram protocols is that send buffer allocation.

I'm guilty of doing this too in the NIU driver, also because there I
lack a "TX queue empty" interrupt and this can keep TCP sockets from
getting stuck.

I think we need a generic solution to this issue because it is getting
quite common to see cases where the packets in the TX queue of a
device can sit there indefinitely.