virtio-net: tweak for better TX performance in NAPI mode

virtio-net: tweak for better TX performance in NAPI mode

There are several issues existed in start_xmit():

- Transmitted packets need to be freed before sending a packet, this
  introduces delay and increases the average packets transmit
  time. This also increase the time that spent in holding the TX lock.
- Notification is enabled after free_old_xmit_skbs() which will
  introduce unnecessary interrupts if TX notification happens on the
  same CPU that is doing the transmission now (actually, virtio-net
  driver are optimized for this case).

So this patch tries to avoid those issues by not cleaning transmitted
packets in start_xmit() when TX NAPI is enabled and disable
notifications even more aggressively. Notification will be since the
beginning of the start_xmit(). But we can't enable delayed
notification after TX is stopped as we will lose the
notifications. Instead, the delayed notification needs is enabled
after the virtqueue is kicked for best performance.

Performance numbers:

1) single queue 2 vcpus guest with pktgen_sample03_burst_single_flow.sh
   (burst 256) + testpmd (rxonly) on the host:

- When pinning TX IRQ to pktgen VCPU: split virtqueue PPS were
  increased 55% from 6.89 Mpps to 10.7 Mpps and 32% TX interrupts were
  eliminated. Packed virtqueue PPS were increased 50% from 7.09 Mpps to
  10.7 Mpps, 99% TX interrupts were eliminated.

- When pinning TX IRQ to VCPU other than pktgen: split virtqueue PPS
  were increased 96% from 5.29 Mpps to 10.4 Mpps and 45% TX interrupts
  were eliminated; Packed virtqueue PPS were increased 78% from 6.12
  Mpps to 10.9 Mpps and 99% TX interrupts were eliminated.

2) single queue 1 vcpu guest + vhost-net/TAP on the host: single
   session netperf from guest to host shows 82% improvement from
   31Gb/s to 58Gb/s, %stddev were reduced from 34.5% to 1.9% and 88%
   of TX interrupts were eliminated.

Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>