virtio_net: don't free buffers in xmit ring

virtio_net: don't free buffers in xmit ring

The virtio_net driver is complicated by the two methods of freeing old
xmit buffers (in addition to freeing old ones at the start of the xmit
path).

The original code used a 1/10 second timer attached to xmit_free(),
reset on every xmit.  Before we orphaned skbs on xmit, the
transmitting userspace could block with a full socket until the timer
fired, the skb destructor was called, and they were re-woken.

So we added the VIRTIO_F_NOTIFY_ON_EMPTY feature: supporting devices
send an interrupt (even if normally suppressed) on an empty xmit ring
which makes us schedule xmit_tasklet().  This was a benchmark win.

Unfortunately, VIRTIO_F_NOTIFY_ON_EMPTY makes quite a lot of work: a
host which is faster than the guest will fire the interrupt every xmit
packet (slowing the guest down further).  Attempting mitigation in the
host adds overhead of userspace timers (possibly with the additional
pain of signals), and risks increasing latency anyway if you get it
wrong.

In practice, this effect was masked by benchmarks which take advantage
of GSO (with its inherent transmit batching), but it's still there.

Now we orphan xmitted skbs, the pressure is off: remove both paths and
no longer request VIRTIO_F_NOTIFY_ON_EMPTY.  Note that the current
QEMU will notify us even if we don't negotiate this feature (legal,
but suboptimal); a patch is outstanding to improve that.

Move the skb_orphan/nf_reset to after we've done the send and notified
the other end, for a slight optimization.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Mark McLoughlin <markmc@redhat.com>