On Wed, Jul 18, 2018 at 3:05 AM Paolo Abeni <pab...@redhat.com> wrote:
>
> Hi,
>
> On Tue, 2018-07-17 at 10:24 -0700, Cong Wang wrote:
> > If you goal is to get rid of skb_clone(), why not just do the following?
> >
> > if (tcf_mirred_is_act_redirect(m_eaction)) {
> > skb2 = skb;
> > } else {
> > skb2 = skb_clone(skb, GFP_ATOMIC);
> > if (!skb2)
> > goto out;
> > }
> >
> > For redirect, we return TC_ACT_SHOT, so upper layer should not
> > touch the skb after that.
> >
> > What am I missing here?
>
> With ACT_SHOT caller/upper layer will free the skb, too. We will have
> an use after free (from either the upper layer and the xmit device).
> Similar issues with STOLEN, TRAP, etc.
>
> In the past, Changli Gao attempted to avoid the clone incrementing the
> skb usage count:
>
> commit 210d6de78c5d7c785fc532556cea340e517955e1
> Author: Changli Gao <xiao...@gmail.com>
> Date: Thu Jun 24 16:25:12 2010 +0000
>
> act_mirred: don't clone skb when skb isn't shared
>
> but some/many device drivers expect an skb usage count of 1, and that
> caused ooops and was revered.
Interesting, I wasn't aware of the above commit and its revert.
First, I didn't use skb_get() above.
Second, I think the caller of dev_queue_xmit() should not
touch the skb after it, the skb is either freed by dev_queue_xmit()
or successfully transmitted, in either case, the ownership belongs
to dev_queue_xmit(). So, I think we should skip the qdisc_drop()
for this case.
Not sure about netif_receive_skb() case, given veth calls in its
xmit too, I speculate the rule is probably same.
Not sure about other ACT_SHOT case than act_mirred...
>
> I think the only other option (beyond re-using ACT_MIRROR) is adding
> another action value, and let the upper layer re-inject the packet
> while handling such action (similar to what ACT_MIRROR currently does,
> but preserving the current mirred semantic).
Maybe if you mean to avoid breaking ACT_SHOT.
Thanks.
