On Tue, Oct 27, 2015 at 11:29 AM, Hannes Frederic Sowa
[off-list ref] wrote:

On Tue, Oct 27, 2015, at 18:32, Tom Herbert wrote:

quoted

On Tue, Oct 27, 2015 at 9:44 AM, Hannes Frederic Sowa
[off-list ref] wrote:

quoted

On Tue, Oct 27, 2015, at 17:36, Tom Herbert wrote:> > -               if
(cork->length + length > maxnonfragsize - headersize) {

quoted

quoted

+       if (cork->length + length > maxnonfragsize - headersize) {
 emsgsize:
-                       ipv6_local_error(sk, EMSGSIZE, fl6,
-                                        mtu - headersize +
-                                        sizeof(struct ipv6hdr));
-                       return -EMSGSIZE;
-               }
+               ipv6_local_error(sk, EMSGSIZE, fl6,
+                                mtu - headersize +
+                                sizeof(struct ipv6hdr));
+               return -EMSGSIZE;
        }

+       /* CHECKSUM_PARTIAL only with no extension headers and when

No, please don't do this. CHECKSUM_PARTIAL should work with extension
headers as defined, so this is just disabling otherwise valid and
useful functionality. If (some) drivers have problems with this they
need to be identified and fixed.

I don't understand. The old code already didn't allow the use of
opt_flen with CHECKSUM_PARTIAL.

Then that's a problem with the old code :-). Is there any other reason
that we can't use CHECKSUM_PARTIAL with extension headers other than
lack of correct driver support?

The lack of correct driver support is a big bumper, but as I wrote, I
don't see a reason to not lift this restriction in net-next. I proposed
a new feature flag, or by looking at your series, we could probably use
the extension header okay field for that.

Okay, but why bother doing this for net? This problem has obviously
existed for a while, and even if the restriction is maintained here
there are still other paths that don't go through ip_append_data that
could trip the bug. Also, drivers are welcome to fix their issues in
net I believe.

I would be conservative in net though.

Bye,
Hannes