Re: [PATCH net] virtio_net: CTRL_GUEST_OFFLOADS depends on CTRL_VQ

[Willem de Bruijn]

00fffe0ff0 DR7: 0000000000000400
> > Call Trace:
> >  ? preempt_count_add+0x58/0xb0
> >  ? _raw_spin_lock_irqsave+0x36/0x70
> >  ? _raw_spin_unlock_irqrestore+0x1a/0x40
> >  ? __wake_up+0x70/0x190
> >  virtnet_set_features+0x90/0xf0 [virtio_net]
> >  __netdev_update_features+0x271/0x980
> >  ? nlmsg_notify+0x5b/0xa0
> >  dev_disable_lro+0x2b/0x190
> >  ? inet_netconf_notify_devconf+0xe2/0x120
> >  devinet_sysctl_forward+0x176/0x1e0
> >  proc_sys_call_handler+0x1f0/0x250
> >  proc_sys_write+0xf/0x20
> >  __vfs_write+0x3e/0x190
> >  ? __sb_start_write+0x6d/0xd0
> >  vfs_write+0xd3/0x190
> >  ksys_write+0x68/0xd0
> >  __ia32_sys_write+0x14/0x20
> >  do_fast_syscall_32+0x86/0xe0
> >  entry_SYSENTER_compat+0x7c/0x8e
> >
> > A similar crash will likely trigger when enabling XDP.
> >
> > Reported-by: Alistair Delva <adelva@xxxxxxxxxx>
> > Reported-by: Willem de Bruijn <willemdebruijn.kernel@xxxxxxxxx>
> > Fixes: 3f93522ffab2 ("virtio-net: switch off offloads on demand if possible on XDP set")
> > Signed-off-by: Michael S. Tsirkin <mst@xxxxxxxxxx>
> > ---
> >
> > Lightly tested.
> >
> > Alistair, could you please test and confirm that this resolves the
> > crash for you?
>
> This patch doesn't work. The reason is that NETIF_F_LRO is also turned
> on by TSO4/TSO6, which your patch didn't check for. So it ends up
> going through the same path and crashing in the same way.
>
>         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
>             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
>                 dev->features |= NETIF_F_LRO;
>
> It sounds like this patch is fixing something slightly differently to
> my patch fixed. virtnet_set_features() doesn't care about
> GUEST_OFFLOADS, it only tests against NETIF_F_LRO. Even if "offloads"
> is zero, it will call virtnet_set_guest_offloads(), which triggers the
> crash.


Interesting. It's surprising that it is trying to configure a flag
that is not configurable, i.e., absent from dev->hw_features
after Michael's change.

> So either we need to ensure NETIF_F_LRO is never set, or

LRO might be available, just not configurable. Indeed this was what I
observed in the past.