Re: [PATCH] resolv_test.c: also cope with CONNREFUSED errors returned by recvfrom

[Samuel Thibault <samuel dot thibault at ens-lyon dot org>]

Hello,

Florian Weimer, on lun. 25 sept. 2017 10:33:52 +0200, wrote:
> On 09/25/2017 10:30 AM, Samuel Thibault wrote:
> > Florian Weimer, on lun. 25 sept. 2017 10:20:46 +0200, wrote:
> > > On 09/25/2017 09:38 AM, Samuel Thibault wrote:
> > > > Florian Weimer, on lun. 25 sept. 2017 09:31:02 +0200, wrote:
> > > > > On 09/10/2017 04:38 PM, Samuel Thibault wrote:
> > > > > > +  if (length < 0)
> > > > > > +    {
> > > > > > +      /* The other end had closed the socket, and we are notified only now. */
> > > > > > +      TEST_VERIFY_EXIT (errno == ECONNREFUSED);
> > > > > > +      return true;
> > > > > > +    }
> > > > > 
> > > > > Sorry for not replying sooner.
> > > > > 
> > > > > This UDP socket is unconnected.  If asynchronous error notifications are
> > > > > received on it, this is arguably a TCP/IP stack bug.
> > > > 
> > > > Reception of "port unreachable" icmp packets can be asynchronous with
> > > > UDP too.
> > > 
> > > Stevens says that this can happen on connected sockets only.  Both TCPv2 and
> > > UNIX Network Programming have detailed explanations.  I can summarize those
> > > if you don't have access.
> > > 
> > > The only quoted exception is Linux, but the kernel behavior changed around
> > > 2.4.0.
> > 
> > Running Linux 4.13
> > 
> > $ strace nc localhost 12345 -u
> > ...
> > sendmsg(3, {msg_name=NULL, msg_namelen=0, msg_iov=[{iov_base="kljsdf\n", iov_len=7}], msg_iovlen=1, msg_controllen=0, msg_flags=0}, 0) = 7
> > select(5, [3 4], [], NULL, NULL)        = 1 (in [3])
> > recvmsg(3, {msg_namelen=0}, 0)          = -1 ECONNREFUSED (Connection refused)
> 
> Surely this is a connected socket, otherwise the sendmsg call would fail due
> to the lack of an address to send the datagram to.

Ah, ok, yes, I misunderstood "connected" above.

> This entire discussion is specific to unconnected sockets, where the
> application has no way to match error responses to sent packets (unless
> IP_RECVERR is set and the information is received as ancillary data).

Indeed.  I have reverted my commit.

Thanks,
Samuel