On Tue, Sep 01, 2015 at 03:38:36PM -0700, Eric Dumazet wrote:
> On Tue, 2015-09-01 at 15:25 -0700, Martin KaFai Lau wrote:
> > On Tue, Sep 01, 2015 at 02:26:58PM -0700, Eric Dumazet wrote:
> > > On Tue, 2015-09-01 at 13:55 -0700, Martin KaFai Lau wrote:
> > > > On Tue, Sep 01, 2015 at 01:14:20PM -0700, Eric Dumazet wrote:
> > > > > It should not be a problem. refcnt is taken when/if necessary (skb
> > > > > queued on a qdisc for example)
> > > > >
> > > > > We have other uses of skb_dst_set_noref()
> > > > >
> > > > > Please describe the problem ?
> > > > The current ip6_tnl_dst_get() does not take the dst refcnt.
> > > >
> > > > If the dst is released after ip6_tnl_dst_get() and before
> > > > skb_dst_set_noref(), would it cause an issue?
> > >
> > > We are under rcu here, and a dst in a cache is protected by RCU by
> > > definition.
> > >
> > > skb_dst_set_noref() has following debugging clause, does it trigger for
> > > you ?
> > >
> > > WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
> > No. I did not see this.
> >
> > I am probably missing something. Do you mean the rcu can
> > protect the followings:
> >
> >
> > ip6_tnl_dst_get()
> > dst_release()
> > dst_free() /* refcnt is 0 */
> > skb_dst_set_noref()
> >
>
> Yes, this is protected by normal rcu rules.
>
> dst wont be freed until all cpus exit their rcu read sections.
I can see skb_dst_set_noref() is safe at this point
but what happen after rcu_read_unlock()?
hmmm... Also, I don't see dst_free() is under the rcu contract, like
call_rcu() or synchronize_rcu().
Even it is (like dst_release for DST_NOCACHE dst_entry), what happen
after the rcu_read_unlock()? Would someone (like qdisc) holds a dst refcnt
to an already/to-be destroyed dst? For DST_NOCACHE, like:
rcu_read_lock()
ip6_tnl_dst_get()
dst_release() /* refcnt is 0 */
=>call_rcu(dst_destroy)
skb_dst_set_noref()
__dev_queue_xmit()
=>skb_dst_force()
=>__dev_xmit_skb()
=>q->enqueue()
rcu_read_unlock()
/* Here, I am holding a dst refcnt but
* the dst is already in the next
* rcu destroy cycle?
*/
> You should take a look at following commits for a bit of history
>
> 10e2eb878f3ca07ac2f05fa5ca5e6c4c9174a27a
> dbfc4fb7d578d4f224faa6b60deb40804dfdc1b1
> f88649721268999bdff09777847080a52004f691
> 6c7e7610ff6888ea15a901fbcb30c5d461816b34
>
Thanks for the pointers. It seems the ip_tunnel.c and ip_tunnel_core.c is
always doing skb_dst_set() also. It is something that can also be optimized
in IPv4 or the situation is different in IPv4?
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
