[previously sent to the allegedly dead [email protected]]
Hi,
Through playing with tcpcp[1], I've found out about a quirk in the
kernel's handling of non-blocking connection-based sockets. The
sk_socket->state value can take on one of SS_FREE, SS_UNCONNECTED,
SS_CONNECTING, SS_CONNECTED or SS_DISCONNECTING. On a standard
*blocking*, connection-oriented socket (eg, TCP), after connect()
returns, sk_socket->state will be SS_CONNECTED.
However, if the socket is placed into non-blocking mode before the
connect() call, connect() returns immediately with EINPROGRESS, and
the sk_socket->state is set to SS_CONNECTING. When the socket
finally does connect, the application is notified via poll(), but
the state remains as SS_CONNECTING (which causes issues for tcpcp,
though doesn't appear to have any other externally visible
implications).
Werner, the author of tcpcp, suggests that the application should
call connect() on the socket a second time, after the successful
connection, to force the sk_socket->state value to SS_CONNECTED.
Should it be the kernel's responsibility to set SS_CONNECTED when
the connection is established? Or should I go file bugs and submit
patches on all the applications that use non-blocking sockets and
don't call connect() a second time?
Thanks in advance,
Bernard.
[1] http://tcpcp.sf.net/
--
Bernard Blackham <bernard at blackham dot com dot au>
On Iau, 2005-04-28 at 11:34, Bernard Blackham wrote:
> Should it be the kernel's responsibility to set SS_CONNECTED when
> the connection is established? Or should I go file bugs and submit
> patches on all the applications that use non-blocking sockets and
> don't call connect() a second time?
See posix 1003.1g drafts. I believe from the state diagram there that
you should call connect() again once it completes.