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

From: Willem de Bruijn
Date: Wed Dec 25 2019 - 11:02:54 EST


On Mon, Dec 23, 2019 at 3:22 PM Alistair Delva <adelva@xxxxxxxxxx> wrote:
>
> On Mon, Dec 23, 2019 at 12:12 PM Willem de Bruijn
> <willemdebruijn.kernel@xxxxxxxxx> wrote:
> >
> > On Mon, Dec 23, 2019 at 2:56 PM Willem de Bruijn
> > <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > >
> > > 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.
> >
> > dev_disable_lro expects that NETIF_F_LRO is always configurable. Which
> > I guess is a reasonable assumption, just not necessarily the case in
> > virtio_net.
> >
> > So I think we need both patches. Correctly mark the feature as fixed
> > by removing from dev->hw_features and also ignore the request from
> > dev_disable_lro, which does not check for this.
>
> Something like this maybe:
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 4d7d5434cc5d..0556f42b0fb5 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -2560,6 +2560,9 @@ static int virtnet_set_features(struct net_device *dev,
> u64 offloads;
> int err;
>
> + if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
> + return 0;
> +
> if ((dev->features ^ features) & NETIF_F_LRO) {
> if (vi->xdp_queue_pairs)
> return -EBUSY;
> @@ -2971,6 +2974,15 @@ static int virtnet_validate(struct virtio_device *vdev)
> if (!virtnet_validate_features(vdev))
> return -EINVAL;
>
> + /* VIRTIO_NET_F_CTRL_GUEST_OFFLOADS does not work without
> + * VIRTIO_NET_F_CTRL_VQ. However the virtio spec does not
> + * specify that VIRTIO_NET_F_CTRL_GUEST_OFFLOADS depends
> + * on VIRTIO_NET_F_CTRL_VQ so devices can set the later but
> + * not the former.
> + */
> + if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
> + __virtio_clear_bit(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS);
> +
> if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
> int mtu = virtio_cread16(vdev,
> offsetof(struct virtio_net_config,

I think we need three separate patches.

- disable guest offloads if VQ is absent
- remove LRO from hw_features if guest offloads are absent

This should fix the ethtool path. But there is a separate path to
disable LRO through dev_disable_lro (from bond enslave, xdp install,
sysctl -n net.ipv.ip_forward=1, ..) that assumes LRO is always
configurable. So we cannot work around needing a patch to
virtset_set_features.

After a long detour, I think your original submission is fine for
that.

Perhaps with a comment in the commit that it is needed even
if the feature is absent from hw_features for dev_disable_lro.