Re: luasocket invalidates its internal state by closing the socket when setsockname() fails

[Ted Unangst <ted.unangst@...>]

On Fri, Mar 19, 2010 at 8:41 PM, Sam Roberts <vieuxtech@gmail.com> wrote:
> On Fri, Mar 19, 2010 at 4:22 PM, Ted Unangst <ted.unangst@gmail.com> wrote:
>> On Wed, Feb 24, 2010 at 6:34 PM, Sam Roberts <vieuxtech@gmail.com> wrote:
>>> This is purely a luasocket behaviour. The system calls allow bind() to
>>> fail, leaving the socket valid.
>>
>> Only on some systems.
>
> a) If such a system exists, wouldn't this be considered an unfortunate
> bug in that system?

Well, it's been like that for 20 years...  But I may be mistaken, I
was thinking of connect, not bind.  It's been a few years since I
actually had this problem.

>>> This behaviour of luasocket is particularly anti-social when
>>> attempting to repeatedly bind to local ports, to search for one that
>>> is free.
>>
>> You're supposed to create a new socket after a failed bind.  Read your
>> Stevens. :)
>
> b) Do you have a page or section reference?
>
> I find no mention of this in "Unix Network Progamming - Networking
> APIs: Sockets and XTI", 2ed, under the bind or udp docs, but its a big
> book. I have copies of TCP/IP illustrated, too, if you find this claim
> in them.
>
> Also, those books are old. In none of  these do I see a claim that
> returning an error invalidates the descriptor:
>
> http://developer.apple.com/Mac/library/documentation/Darwin/Reference/ManPages/man2/bind.2.html
>
> http://www.freebsd.org/cgi/man.cgi?query=bind&apropos=0&sektion=2&manpath=FreeBSD+8.0-RELEASE&format=html
>
> http://docs.sun.com/app/docs/doc/816-5170/bind-3socket?a=view
>
> I do see "A process with a connnected UDP socket can call connect
> again for that socket." (p.226, also the freebsd man pages). This
> isn't supported by luasocket, either, connected sockets can only be
> disconnected:

Ah, sorry, no reference, other than the man pages (for connect, not bind, oops).

For UDP, you're certainly right.  I was thinking of the stream behavior.

"Generally, stream sockets may successfully connect() only once;"  I
think this is often interpreted to mean "one successful connect" but
the implementation is actually "only call connect() once,
successfully".

The problem exists in OpenBSD and FreeBSD.  Possibly Mac, I didn't
test at the time.  Sorry for the hazy memory.