On Mon, 2015-11-30 at 19:13 +0100, Paolo Abeni wrote:

On Mon, 2015-11-30 at 08:35 -0800, Eric Dumazet wrote:

quoted

ip6_sk_dst_lookup_flow() uses sk_dst_check() anyway, so the simplest
way to fix the mess is to remove sk_dst_lock completely, as we did for
IPv4.

Probably I'm missing something here, but why we don't need to sync the
update of sk_dst_cache and of dst_cookie (i.e. put them under the same
lock)?

Can't we end up with inconsistent values after concurrent udp
sendmsg() ? 

I do not think this is an issue. A route is best effort.

If really a packet is dropped during a route flap, no big deal,
especially if this is during a fuzzer test ;)