Hi Eric,
On Tue, 27 Sep 2016 14:27:13 -0700 Eric Dumazet [off-list ref] wrote:
Since this runs lockless, another cpu might change m->tcfm_eaction in
the middle, and you could call dev_queue_xmit(skb2) while the skb2 was
prepared for the opposite action.
Well, seem members of 'struct tcf_mirred' are out of sync wrt to each
other, even in existing code, regadless this patch:
- 'tcfm_dev' may be assigned, but 'tcfm_ok_push' not yet updated,
may result in skb_push_rcsum being called/not called
- 'tcfm_eaction' is changed, in between "mirror is always swallowed" to
the final 'out:' label,
may result in wrong tc_verd assigned (or lack of assignment)
Seems the whole "params" need be rcu_dereferenced, like in
tunnel_key_act, or like your suggestion in
https://patchwork.ozlabs.org/patch/667680/.
I'm gonna fix the new problem you pointed out, by reading-once
'tcfm_eaction' early (right when tcfm_dev is dereferenced) knowing this
is just "keeping things as is wrt running lockless", without introducing
any new non-coherent code.
Thanks,
Shmulik