Re: a problem tcp_v4_err()

[Eric Dumazet <eric.dumazet@gmail.com>]

Le vendredi 12 novembre 2010 à 19:21 +0100, Eric Dumazet a écrit :
> Le vendredi 12 novembre 2010 à 19:12 +0100, Eric Dumazet a écrit :
> > Le vendredi 12 novembre 2010 à 20:57 +0300, Alexey Kuznetsov a écrit :
> > > Hello!
> > > 
> > > I looked at tcp_v4_err() and found something strange. Quite non-trivial operations
> > > are performed on unlocked sockets. It looks like at least this BUG_ON():
> > > 
> > >                 skb = tcp_write_queue_head(sk);
> > >                 BUG_ON(!skb);
> > > 
> > > can be easily triggered.
> > > 
> > > Do I miss something?
> > > 
> > 
> > Hi Alexey !
> > 
> > I see socket is locked around line 368,
> > 
> >         bh_lock_sock(sk);
> >         /* If too many ICMPs get dropped on busy
> >          * servers this needs to be solved differently.
> >          */
> >         if (sock_owned_by_user(sk))
> >                 NET_INC_STATS_BH(net, LINUX_MIB_LOCKDROPPEDICMPS);
> > 
> > 
> > Hmm, maybe some goto is missing ;)
> > 
> 
> Well, goto is not missing.
> 
> Why do you think BUG_ON(!skb) can be triggered ?
> 
> We test before :
> 
> 	if (seq != tp->snd_una  || !icsk->icsk_retransmits ||
> 		!icsk->icsk_backoff)
> 		break;
> 
> So a concurrent user only can add new skb(s) in the (non empty) queue ?
> 
> 

Oh well, it seems you are right (backlog processing)

Bug was introduced in commit f1ecd5d9e736660 (Revert Backoff [v3]:
Revert RTO on ICMP destination unreachable) from Damian Lukowski