Experimental Physics and
Industrial Control System

Torsten Bögershausen <[email protected]> · Fri, 10 Oct 2014 18:52:04 +0200

On 10/10/14 15:39, J. Lewis Muir wrote:

On 10/10/14 3:37 AM, Torsten Bögershausen wrote:

There is now way for an application to generate a "TCP packet with a
zero length data payload" and deliver that to the receiver.

Hi, Torsten.

I stand corrected.  I tried yesterday on Mac OS X Mavericks using the
write function on a socket file descriptor, and it didn't send anything.
It didn't report an error but just returned 0.  So, if anything, the
behavior might be classified as undefined.  But you're probably 100%
right that there's no way to send a TCP packet with zero length from a
send or write function call.  (A small side of me still wonders if there
might be some crazy TCP implementation out there somewhere that does,
but probably not.)  I wish this were specified explicitly somewhere,
though.  That is, I wish some documentation said that trying to write
data of length 0 on a TCP socket will return an error.  Or, if not an
error, will return 0 and not write anything.

We need to remember that the send() is a generic interface function.
It allows access to different kind of protocols which the kernel
implements.
So the send(socket, buf, 0, 0) on a TCP/IP socket is only one
possible use case.
And "man send" does not cover any protocol specific information.

Other protocols can do this, but not TCP.

Note that TCP is a stream protocol, when the sender runs

send(socket, buff, 5, 0);
send(socket, buff2, 3, 0);

The receiver runs
recv(socket, buff, sizeof(buf), 0);
and may receive either the 5 bytes or the 8 bytes.

Knowing this, a
ret = send(socket, buff, 0, 0);
yields ret == 0 on my Mac OS X box, but the spec doesn't mention it as
a valid operation.
And I would treat anything not mentioned as "not portable", avoid it.
In the same sense as we avoid to pick up the phone and order 0 pizzas
somewhere ;-)

Right, well, I was just noting it doesn't say it's invalid either.  It
just doesn't say anything about this case, unfortunately.

The same thing applies for
ret = recv(socket, buf, 0, 0);
What should the function return ?
0 because you asked to receive 0 bytes ?
-1 because this is illegal ?

I would guess it would return 0, but I don't know.

I did a test on Mac OS X and Linux and cygwin: all return 0

A different thing is when
ret = recv(socket, buf, sizeof(buf), 0);
we find ret == 0.
Then this is an EOF indicator, saying that the peer has shutdown his
write side, or closed the connection

Right, and that's what I noted would make it impossible to differentiate
a connection close from a 0 length TCP send.  But again, I see now that
a 0 length send is likely not even possible on any TCP implementation.

Of course, there is probably software that will let one construct a
special TCP packet with data of length 0 and send it, but that's not
what we're worried about here.

TCP uses these packets internally, to send an ACK to the peer.

Regards,

Lewis

/Torsten

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System