[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CACGkMEuTnLX7EPuOLb2UhrZT2oH2AcXPQrvq-uw2ZydYV_FAgQ@mail.gmail.com>
Date: Thu, 30 Oct 2025 10:27:55 +0800
From: Jason Wang <jasowang@...hat.com>
To: Alyssa Ross <hi@...ssa.is>
Cc: Paolo Abeni <pabeni@...hat.com>, "Michael S. Tsirkin" <mst@...hat.com>, netdev@...r.kernel.org,
virtualization@...ts.linux.dev, regressions@...ts.linux.dev
Subject: Re: [REGRESSION][BISECTED] virtio_net CSUM broken with Cloud Hypervisor
On Thu, Oct 30, 2025 at 4:03 AM Alyssa Ross <hi@...ssa.is> wrote:
>
> Since 56a06bd40fab ("virtio_net: enable gso over UDP tunnel support."),
> networking in Cloud Hypervisor is no longer working for me.
>
> I've narrowed down the problem to here:
>
> > @@ -2555,14 +2567,21 @@ static void virtnet_receive_done(struct virtnet_info *vi, struct receive_queue *
> > if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
> > virtio_skb_set_hash(&hdr->hash_v1_hdr, skb);
> >
> > - if (flags & VIRTIO_NET_HDR_F_DATA_VALID)
> > - skb->ip_summed = CHECKSUM_UNNECESSARY;
> > + hdr->hdr.flags = flags;
>
> It looks like this was added because virtio_net_handle_csum_offload()
> looks at the flags from the hdr it's given, rather than having it passed
> separately, but it appears something later on relies on the previous
> value of hdr->hdr.flags.
>
> From my tracing, hdr->hdr.flags is set to either 0 or
> VIRTIO_NET_HDR_F_NEEDS_CSUM before this assignment, and flags is always
> 0, so in some cases VIRTIO_NET_HDR_F_NEEDS_CSUM now ends up being unset.
Are you using XDP, if not there should be no change.
>
> > + if (virtio_net_handle_csum_offload(skb, &hdr->hdr, vi->rx_tnl_csum)) {
> > + net_warn_ratelimited("%s: bad csum: flags: %x, gso_type: %x rx_tnl_csum %d\n",
> > + dev->name, hdr->hdr.flags,
> > + hdr->hdr.gso_type, vi->rx_tnl_csum);
> > + goto frame_err;
> > + }
>
> If I change it to save the previous value of hdr->hdr.flags, and restore
> it again here, everything works again.
>
> Disabling offload_csum in Cloud Hypervisor is a usable workaround,
> because then hdr->hdr.flags is always 0 to begin with anyway.
>
> #regzbot introduced: 56a06bd40fab
Is mergeable rx buffer enabled, if not, I wonder if this can be fixed by:
https://marc.info/?l=linux-netdev&m=176170721926346&w=2
Thanks
Powered by blists - more mailing lists