Re: WARNING in ip_recv_error

From: Paolo Abeni
Date: Thu May 24 2018 - 03:07:48 EST


On Wed, 2018-05-23 at 11:40 -0400, Willem de Bruijn wrote:
> On Sun, May 20, 2018 at 7:13 PM, Willem de Bruijn
> <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > On Fri, May 18, 2018 at 2:59 PM, Willem de Bruijn
> > <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > > On Fri, May 18, 2018 at 2:46 PM, Willem de Bruijn
> > > <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > > > On Fri, May 18, 2018 at 2:44 PM, Willem de Bruijn
> > > > <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > > > > On Fri, May 18, 2018 at 1:09 PM, Willem de Bruijn
> > > > > <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > > > > > On Fri, May 18, 2018 at 11:44 AM, David Miller <davem@xxxxxxxxxxxxx> wrote:
> > > > > > > From: Eric Dumazet <eric.dumazet@xxxxxxxxx>
> > > > > > > Date: Fri, 18 May 2018 08:30:43 -0700
> > > > > > >
> > > > > > > > We probably need to revert Willem patch (7ce875e5ecb8562fd44040f69bda96c999e38bbc)
> > > > > > >
> > > > > > > Is it really valid to reach ip_recv_err with an ipv6 socket?
> > > > > >
> > > > > > I guess the issue is that setsockopt IPV6_ADDRFORM is not an
> > > > > > atomic operation, so that the socket is neither fully ipv4 nor fully
> > > > > > ipv6 by the time it reaches ip_recv_error.
> > > > > >
> > > > > > sk->sk_socket->ops = &inet_dgram_ops;
> > > > > > < HERE >
> > > > > > sk->sk_family = PF_INET;
> > > > > >
> > > > > > Even calling inet_recv_error to demux would not necessarily help.
> > > > > >
> > > > > > Safest would be to look up by skb->protocol, similar to what
> > > > > > ipv6_recv_error does to handle v4-mapped-v6.
> > > > > >
> > > > > > Or to make that function safe with PF_INET and swap the order
> > > > > > of the above two operations.
> > > > > >
> > > > > > All sound needlessly complicated for this rare socket option, but
> > > > > > I don't have a better idea yet. Dropping on the floor is not nice,
> > > > > > either.
> > > > >
> > > > > Ensuring that ip_recv_error correctly handles packets from either
> > > > > socket and removing the warning should indeed be good.
> > > > >
> > > > > It is robust against v4-mapped packets from an AF_INET6 socket,
> > > > > but see caveat on reconnect below.
> > > > >
> > > > > The code between ipv6_recv_error for v4-mapped addresses and
> > > > > ip_recv_error is essentially the same, the main difference being
> > > > > whether to return network headers as sockaddr_in with SOL_IP
> > > > > or sockaddr_in6 with SOL_IPV6.
> > > > >
> > > > > There are very few other locations in the stack that explicitly test
> > > > > sk_family in this way and thus would be vulnerable to races with
> > > > > IPV6_ADDRFORM.
> > > > >
> > > > > I'm not sure whether it is possible for a udpv6 socket to queue a
> > > > > real ipv6 packet on the error queue, disconnect, connect to an
> > > > > ipv4 address, call IPV6_ADDRFORM and then call ip_recv_error
> > > > > on a true ipv6 packet. That would return buggy data, e.g., in
> > > > > msg_name.
> > > >
> > > > In do_ipv6_setsockopt IPV6_ADDRFORM we can test that the
> > > > error queue is empty, and then take its lock for the duration of the
> > > > operation.
> > >
> > > Actually, no reason to hold the lock. This setsockopt holds the socket
> > > lock, which connect would need, too. So testing that the queue
> > > is empty after testing that it is connected to a v4 address is
> > > sufficient to ensure that no ipv6 packets are queued for reception.
> > >
> > > diff --git a/net/ipv6/ipv6_sockglue.c b/net/ipv6/ipv6_sockglue.c
> > > index 4d780c7f0130..a975d6311341 100644
> > > --- a/net/ipv6/ipv6_sockglue.c
> > > +++ b/net/ipv6/ipv6_sockglue.c
> > > @@ -199,6 +199,11 @@ static int do_ipv6_setsockopt(struct sock *sk,
> > > int level, int optname,
> > >
> > > if (ipv6_only_sock(sk) ||
> > > !ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
> > > retv = -EADDRNOTAVAIL;
> > > break;
> > > }
> > >
> > > + if (!skb_queue_empty(&sk->sk_error_queue)) {
> > > + retv = -EBUSY;
> > > + break;
> > > + }
> > > +
> > > fl6_free_socklist(sk);
> > > __ipv6_sock_mc_close(sk);
> > >
> > > After this it should be safe to remove the warning in ip_recv_error.
> >
> > Hmm.. nope.
> >
> > This ensures that the socket cannot produce any new true v6 packets.
> > But it does not guarantee that they are not already in the system, e.g.
> > queued in tc, and will find their way to the error queue later.
> >
> > We'll have to just be able to handle ipv6 packets in ip_recv_error.
> > Since IPV6_ADDRFORM is used to pass to legacy v4-only
> > processes and those likely are only confused by SOL_IPV6
> > error messages, it is probably best to just drop them and perhaps
> > WARN_ONCE.
>
> Even more fun, this is not limited to the error queue.
>
> I can queue a v6 packet for reception on a socket, connect to a v4
> address, call IPV6_ADDRFORM and then a regular recvfrom will
> return a partial v6 address as AF_INET.
>
> We definitely do not want to have to add a check
>
> if (skb->protocol == htons(ETH_P_IPV6)) {
> kfree_skb(skb);
> goto try_again;
> }
>
> to the normal recvmsg path.
>
> An alternative may be to tighten the check on when to allow
> IPV6_ADDRFORM. Not only return EBUSY if a packet is pending,
> but also if any sk_{rmem, omem, wmem}_alloc is non-zero. Only,
> these tightened constraints could break a legacy application.

I fear that condition will be very restrictive: for UDP sockets sk_rmem
can be zero only occasionally, after the first packet has been
received, due to the peculiar memory accounting - commit 6b229cf77d68
("This computer thing still completely fool me").

Cheers,

Paolo