Re: [Qemu-devel] [PATCH] virtio-net: allow increasing rx queue size

[Cornelia Huck]

On Thu, 4 Aug 2016 22:52:29 +0300
"Michael S. Tsirkin" <address@hidden> wrote:

> On Thu, Aug 04, 2016 at 09:35:15AM +0200, Cornelia Huck wrote:
> > On Thu, 4 Aug 2016 02:16:14 +0300
> > "Michael S. Tsirkin" <address@hidden> wrote:
> > 
> > > This allows increasing the rx queue size up to 1024: unlike with tx,
> > > guests don't put in huge S/G lists into RX so the risk of running into
> > > the max 1024 limitation due to some off-by-one seems small.
> > > 
> > > It's helpful for users like OVS-DPDK which don't do any buffering on the
> > > host - 1K roughly matches 500 entries in tun + 256 in the current rx
> > > queue, which seems to work reasonably well. We could probably make do
> > > with ~750 entries but virtio spec limits us to powers of two.
> > > It might be a good idea to specify an s/g size limit in a future
> > > version.
> > > 
> > > It also might be possible to make the queue size smaller down the road, 64
> > > seems like the minimal value which will still work (as guests seem to
> > > assume a queue full of 1.5K buffers is enough to process the largest
> > > incoming packet, which is ~64K).  No one actually asked for this, and
> > > with virtio 1 guests can reduce ring size without need for host
> > > configuration, so don't bother with this for now.
> > 
> > Do we need some kind of sanity check that the guest did not resize
> > below a reasonable limit?
> 
> Unfortunately the spec does not have an interface for that.
> Guests expect they can get away with any size.

Might be a good idea to add this in the future, so that the guest is
able to discover the minimum and the host can refuse to work if the
configured queue is too small.

(I can easily reject the setup ccw on virtio-ccw, but is there an
elegant way to refuse setting up the queues with virtio-pci?)

> 
> > > 
> > > Signed-off-by: Michael S. Tsirkin <address@hidden>
> > > ---
> > >  include/hw/virtio/virtio-net.h |  1 +
> > >  hw/net/virtio-net.c            | 22 +++++++++++++++++++++-
> > >  2 files changed, 22 insertions(+), 1 deletion(-)
> > > 
> > 
> > 
> > > @@ -1716,10 +1717,28 @@ static void virtio_net_device_realize(DeviceState 
> > > *dev, Error **errp)
> > >      VirtIONet *n = VIRTIO_NET(dev);
> > >      NetClientState *nc;
> > >      int i;
> > > +    int min_rx_queue_size;
> > > 
> > >      virtio_net_set_config_size(n, n->host_features);
> > >      virtio_init(vdev, "virtio-net", VIRTIO_ID_NET, n->config_size);
> > > 
> > > +    /*
> > > +     * We set a lower limit on RX queue size to what it always was.
> > > +     * Guests that want a smaller ring can always resize it without
> > > +     * help from us (using virtio 1 and up).
> > > +     */
> > > +    min_rx_queue_size = 256;
> > 
> > I'd find it more readable to introduce a #define with the old queue
> > size as the minimum size...
> > 
> > > +    if (n->net_conf.rx_queue_size < min_rx_queue_size ||
> > > +        n->net_conf.rx_queue_size > VIRTQUEUE_MAX_SIZE ||
> > > +        (n->net_conf.rx_queue_size & (n->net_conf.rx_queue_size - 1))) {
> > > +        error_setg(errp, "Invalid rx_queue_size (= %" PRIu16 "), "
> > > +                   "must be a power of 2 between %d and %d.",
> > > +                   n->net_conf.rx_queue_size, min_rx_queue_size,
> > > +                   VIRTQUEUE_MAX_SIZE);
> > > +        virtio_cleanup(vdev);
> > > +        return;
> > > +    }
> > > +
> > >      n->max_queues = MAX(n->nic_conf.peers.queues, 1);
> > >      if (n->max_queues * 2 + 1 > VIRTIO_QUEUE_MAX) {
> > >          error_setg(errp, "Invalid number of queues (= %" PRIu32 "), "
> > > @@ -1880,6 +1899,7 @@ static Property virtio_net_properties[] = {
> > >                         TX_TIMER_INTERVAL),
> > >      DEFINE_PROP_INT32("x-txburst", VirtIONet, net_conf.txburst, 
> > > TX_BURST),
> > >      DEFINE_PROP_STRING("tx", VirtIONet, net_conf.tx),
> > > +    DEFINE_PROP_UINT16("rx_queue_size", VirtIONet, 
> > > net_conf.rx_queue_size, 256),
> > 
> > ...and defaulting to that #define (or one derived from the #define
> > above) here.
> 
> These happen to be the same, but they are in fact
> unrelated: one is the default, the other is the
> min value.

Hm...

/* previously fixed value */
#define VIRTIO_NET_RX_DEFAULT_SIZE 256
/* for now, only allow larger queues; with virtio-1, guest can downsize */
#define VIRTIO_NET_RX_MIN_SIZE VIRTIO_NET_RX_DEFAULT_SIZE

This would allow getting rid of the new local variable and gets us a
speaking define in the property definition.