On Tue, Sep 3, 2019 at 11:52 AM Shmulik Ladkani
[off-list ref] wrote:

On Sun, 1 Sep 2019 16:05:48 -0400
Willem de Bruijn [off-list ref] wrote:

quoted

One quick fix is to disable sg and thus revert to copying in this
case. Not ideal, but better than a kernel splat:

@@ -3714,6 +3714,9 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb,
        sg = !!(features & NETIF_F_SG);
        csum = !!can_checksum_protocol(features, proto);

+       if (list_skb && skb_headlen(list_skb) && !list_skb->head_frag)
+               sg = false;
+

Thanks Willem.

I followed this approach, and further refined it based on the conditions
that lead to this BUG_ON:

 - existance of frag_list
 - mangled gso_size (using SKB_GSO_DODGY as a hint)
 - some frag in the frag_list has a linear part that is NOT head_frag,
   or length not equal to the requested gso_size

BTW, doing so allowed me to refactor a loop that tests for similar
conditions in the !(features & NETIF_F_GSO_PARTIAL) case, where an
attempt to execute partial splitting at the frag_list pointer (see
07b26c9454a2 and 43170c4e0ba7).

I've tested this using the reproducer, with various linear skbs in
the frag_list and different gso_size mangling. All resulting 'segs'
looked correct. Did not test on a live system yet.

Comments are welcome.

specifically, I would like to know whether we can
 - better refine the condition where this "sg=false fallback" needs
   to be applied
 - consolidate my new 'list_skb && (type & SKB_GSO_DODGY)' case with
   the existing '!(features & NETIF_F_GSO_PARTIAL)' case

This is a lot more code change. Especially for stable fixes that need
to be backported, a smaller patch is preferable.

My suggestion only tested the first frag_skb length. If a list can be
created where the first frag_skb is head_frag but a later one is not,
it will fail short. I kind of doubt that.

By default skb_gro_receive builds GSO skbs that can be segmented
along the original gso_size boundaries. We have so far only observed
this issue when messing with gso_size.

We can easily refine the test to fall back on to copying only if
skb_headlen(list_skb) != mss. Alternatively, only on SKB_GSO_DODGY is fine, too.

I suggest we stick with the two-liner.