To: Felix von Leitner <felix@codeblau.de>
Subject: Re: [dnscache] problems (hanging) when lookup-cache full?
In-Reply-To: Your message of "Sun, 17 Sep 2000 21:09:30 +0200."
             <20000917210930.A7865@codeblau.de> 
--------
I wrote:
>> (For instance, Alan Cox thinks there is a good reason to be
>> incompatible with BSD when it comes to reporting of ICMP errors on
>> UDP sockets, although I think he's full of it.)

To the best of my understanding:

In Linux, when an ICMP error arrives such as "port unreachable" or
"host unreachable" in response to a UDP operation, the error will be
delivered to the last UDP socket which sent to that address, even if
that socket is not connected.  (By "the error will be delivered," I
mean the socket is woken up wrt select() for reading and writing, and
that the next operation on that socket wpill return -1 with errno set
to EHOSTUNREACH or whatever.  This is consistent with how errors are
delivered on sockets in general.)

(I don't know what happens if there is a socket connected to the
address the ICMP error came from and it is not the same as the socket
which last sent to that address.)

Every other operating system in the world only reports such errors if
the UDP socket is connected.

I believe there is a /proc switch which can make Linux perform
according to the BSD behavior, but I can't find it on my machine.

Alan Cox believes this aberration is justified because the IETF host
requirements say:

         4.1.3.3  ICMP Messages

            UDP MUST pass to the application layer all ICMP error
            messages that it receives from the IP layer.  Conceptually
            at least, this may be accomplished with an upcall to the
            ERROR_REPORT routine (see Section 4.2.4.1).

On the other hand, this text is followed with:

            DISCUSSION:
                 Note that ICMP error messages resulting from sending a
                 UDP datagram are received asynchronously.  A UDP-based
                 application that wants to receive ICMP error messages
                 is responsible for maintaining the state necessary to
                 demultiplex these messages when they arrive; for
                 example, the application may keep a pending receive
                 operation for this purpose.  The application is also
                 responsible to avoid confusion from a delayed ICMP
                 error message resulting from an earlier use of the same
                 port(s).

I (and most BSD networking people) think Alan Cox is full of it
because:

	* Requiring the socket to be connected seems to be covered
	  under "maintaining the state necessary to demultiplex these
	  messages when they arrive."

	* The Linux approach does not make it possible to tell which
	  destination an error resulted from, unless you've only been
	  sending messages to one place.  (In which case, you might as
	  well have connected the socket.)  So it's not really useful
	  behavior.