Re: [Qemu-devel] [PATCH 5/5] virtio-net: unbreak any layout

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Jul 13, 2015 at 12:54:02PM +0200, Greg Kurz wrote:
> On Mon, 13 Jul 2015 11:22:09 +0300
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > On Mon, Jul 13, 2015 at 10:24:59AM +0300, Michael S. Tsirkin wrote:
> > > On Mon, Jul 13, 2015 at 01:46:51PM +0800, Jason Wang wrote:
> > > > Commit 032a74a1c0fcdd5fd1c69e56126b4c857ee36611
> > > > ("virtio-net: byteswap virtio-net header") breaks any layout by
> > > > requiring out_sg[0].iov_len >= n->guest_hdr_len. Fixing this by
> > > > copying header to temporary buffer and copying it back after byteswap.
> > > > 
> > > > Fixes 032a74a1c0fcdd5fd1c69e56126b4c857ee36611
> > > > ("virtio-net: byteswap virtio-net header")
> > > > Cc: qemu-stable@nongnu.org
> > > > Cc: clg@fr.ibm.com
> > > > Signed-off-by: Jason Wang <jasowang@redhat.com>
> > > > ---
> > > >  hw/net/virtio-net.c | 10 +++++++---
> > > >  1 file changed, 7 insertions(+), 3 deletions(-)
> > > > 
> > > > diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
> > > > index e3c2db3..b42af8f 100644
> > > > --- a/hw/net/virtio-net.c
> > > > +++ b/hw/net/virtio-net.c
> > > > @@ -1139,10 +1139,11 @@ static int32_t virtio_net_flush_tx(VirtIONetQueue *q)
> > > >      }
> > > >  
> > > >      while (virtqueue_pop(q->tx_vq, &elem)) {
> > > > -        ssize_t ret, len;
> > > > +        ssize_t ret, len, s;
> > > >          unsigned int out_num = elem.out_num;
> > > >          struct iovec *out_sg = &elem.out_sg[0];
> > > >          struct iovec sg[VIRTQUEUE_MAX_SIZE];
> > > > +        struct virtio_net_hdr hdr;
> > > >  
> > > >          if (out_num < 1) {
> > > >              error_report("virtio-net header not in first element");
> > > > @@ -1150,11 +1151,14 @@ static int32_t virtio_net_flush_tx(VirtIONetQueue *q)
> > > >          }
> > > >  
> > > >          if (n->has_vnet_hdr) {
> > > > -            if (out_sg[0].iov_len < n->guest_hdr_len) {
> > > > +            s = iov_to_buf(out_sg, out_num, 0, &hdr, sizeof(hdr));
> > > > +            if (s != sizeof(hdr)) {
> > > >                  error_report("virtio-net header incorrect");
> > > >                  exit(1);
> > > >              }
> > > > -            virtio_net_hdr_swap(vdev, (void *) out_sg[0].iov_base);
> > > > +            virtio_net_hdr_swap(vdev, (void *) &hdr);
> > > > +            s = iov_from_buf(out_sg, out_num, 0, &hdr, sizeof(hdr));
> > > > +            assert(s == sizeof(hdr));
> > > >          }
> > > >  
> > > >          /*
> > > 
> > > Looks like this will write into out_sg - that's violating the virtio spec.
> > 
> > And the fault lies with the original code upstream - not with these
> > patches. In my opinion the thing to do here is to completely
> 
> I fully agree: it is a hack.
> 
> > avoid using vnet header for cross endian unless host supports ioctls
> > to set endian-ness.
> > 
> 
> Makes sense. I had an *experimental* patch that would just deactivate
> the cross-endian hack if the host kernel provides the TUNSETVNETLE/BE
> ioctls. It worked but the patch was really ugly and I did not care to
> post at the time...

So my proposal would be to basically drop the hack completely:
if we are cross-endian and TUNSETVNETLE/BE are unavailable,
treat it as if has_vnet_hdr was not set.


> I'll try to work something out before I go on
> vacation.
> 
> > > > -- 
> > > > 2.1.4
> > 
> 
> Cheers.
> 
> --
> Greg