If segmentation offload is enabled by the host, we currently allocate
maximum sized packet buffers and pass them to the host. This uses up
20 ring entries, allowing us to supply only 20 packet buffers to the
host with a 256 entry ring. This is a huge overhead when receiving
small packets, and is most keenly felt when receiving MTU sized
packets from off-host.
The VIRTIO_NET_F_MRG_RXBUF feature flag is set by hosts which support
using receive buffers which are smaller than the maximum packet size.
In order to transfer large packets to the guest, the host merges
together multiple receive buffers to form a larger logical buffer.
The number of merged buffers is returned to the guest via a field in
the virtio_net_hdr.
Make use of this support by supplying single page receive buffers to
the host. On receive, we extract the virtio_net_hdr, copy 128 bytes of
the payload to the skb's linear data buffer and adjust the fragment
offset to point to the remaining data. This ensures proper alignment
and allows us to not use any paged data for small packets. If the
payload occupies multiple pages, we simply append those pages as
fragments and free the associated skbs.
This scheme allows us to be efficient in our use of ring entries
while still supporting large packets. Benchmarking using netperf from
an external machine to a guest over a 10Gb/s network shows a 100%
improvement from ~1Gb/s to ~2Gb/s. With a local host->guest benchmark
with GSO disabled on the host side, throughput was seen to increase
from 700Mb/s to 1.7Gb/s.
Based on a patch from Herbert Xu.
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> (use netdev_priv)
Signed-off-by: David S. Miller <davem@davemloft.net>
We want others to implement and use virtio, so it makes sense to BSD
license the non-__KERNEL__ parts of the headers to make this crystal
clear.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Acked-by: Christian Borntraeger <borntraeger@de.ibm.com>
Acked-by: Mark McLoughlin <markmc@redhat.com>
Acked-by: Ryan Harper <ryanh@us.ibm.com>
Acked-by: Eric Van Hensbergen <ericvh@gmail.com>
Acked-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
So, we previously had a 'VIRTIO_NET_F_GSO' bit which meant that 'the
host can handle csum offload, and any TSO (v4&v6 incl ECN) or UFO
packets you might want to send. I thought this was good enough for
Linux, but it actually isn't, since we don't do UFO in software.
So, add separate feature bits for what the host can handle. Add
equivalent ones for the guest to say what it can handle, because LRO
is coming too (thanks Herbert!).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
1) Turn GSO on virtio net into an all-or-nothing (keep checksumming
separate). Having multiple bits is a pain: if you can't support something
you should handle it in software, which is still a performance win.
2) Make VIRTIO_NET_HDR_GSO_ECN a flag in the header, so it can apply to
IPv6 or v4.
3) Rename VIRTIO_NET_F_NO_CSUM to VIRTIO_NET_F_CSUM (ie. means we do
checksumming).
4) Add csum and gso params to virtio_net to allow more testing.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
It's far easier to deal with packets if we don't have to parse the
packet to figure out the header length to know how much to pull into
the skb data. Add the field to the virtio_net_hdr struct (and fix the
spaces that somehow crept in there).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Previously we used a type/len pair within the config space, but this
seems overkill. We now simply define a structure which represents the
layout in the config space: the config space can now only be extended
at the end.
The main driver-visible changes:
1) We indicate what fields are present with an explicit feature bit.
2) Virtqueues are explicitly numbered, and not in the config space.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The network driver uses two virtqueues: one for input packets and one
for output packets. This has nice locking properties (ie. we don't do
any for recv vs send).
TODO:
1) Big packets.
2) Multi-client devices (maybe separate driver?).
3) Resolve freeing of old xmit skbs (Christian Borntraeger)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Christian Borntraeger <borntraeger@de.ibm.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: netdev@vger.kernel.org
Mark McLoughlin
Rusty Russell