Re: [PATCH net] net: tighten bad gso csum offset check in virtio_net_hdr

On Tue, Sep 10, 2024 at 10:54 PM Willem de Bruijn
[off-list ref] wrote:

Willem de Bruijn wrote:

quoted

Jason Wang wrote:

quoted

On Tue, Sep 10, 2024 at 8:40 AM Willem de Bruijn
[off-list ref] wrote:

quoted

From: Willem de Bruijn <willemb@google.com>

The referenced commit drops bad input, but has false positives.
Tighten the check to avoid these.

The check detects illegal checksum offload requests, which produce
csum_start/csum_off beyond end of packet after segmentation.

But it is based on two incorrect assumptions:

1. virtio_net_hdr_to_skb with VIRTIO_NET_HDR_GSO_TCP[46] implies GSO.
True in callers that inject into the tx path, such as tap.
But false in callers that inject into rx, like virtio-net.
Here, the flags indicate GRO, and CHECKSUM_UNNECESSARY or
CHECKSUM_NONE without VIRTIO_NET_HDR_F_NEEDS_CSUM is normal.

2. TSO requires checksum offload, i.e., ip_summed == CHECKSUM_PARTIAL.
False, as tcp[46]_gso_segment will fix up csum_start and offset for
all other ip_summed by calling __tcp_v4_send_check.

Because of 2, we can limit the scope of the fix to virtio_net_hdr
that do try to set these fields, with a bogus value.

Link: https://lore.kernel.org/netdev/20240909094527.GA3048202@port70.net/ (local)
Fixes: 89add40066f9 ("net: drop bad gso csum_start and offset in virtio_net_hdr")
Signed-off-by: Willem de Bruijn <willemb@google.com>
Cc: <redacted>

---

Verified that the syzbot repro is still caught.

An equivalent alternative would be to move the check for csum_offset
to where the csum_start check is in segmentation:

-    if (unlikely(skb_checksum_start(skb) != skb_transport_header(skb)))
+    if (unlikely(skb_checksum_start(skb) != skb_transport_header(skb) ||
+                 skb->csum_offset != offsetof(struct tcphdr, check)))

Cleaner, but messier stable backport.

We'll need an equivalent patch to this for VIRTIO_NET_HDR_GSO_UDP_L4.
But that csum_offset test was in a different commit, so different

Not for this patch, but I see this in UDP_L4:

                       if (!(hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM))
                               return -EINVAL;

This seems to forbid VIRTIO_NET_HDR_F_DATA_VALID. I wonder what's the
reason for doing this.

It tests &, not == ?

Oh you mean as alternative, for receive of GRO from hypervisor.

Or it could be a physical device that can do GRO HW.

Yes, fair point.

Then we also trust a privileged process over tun, like syzkaller.
When it comes to checksums, I suppose that is fine: it cannot harm
kernel integrity.

Yes.

One missing piece is that TCP GSO will fix up non CHECKSUM_PARTIAL
skbs. UDP GSO does not have the same logic.

Thanks