[ofa-general] implicit address binding by rdma_resolve_addr

[Alexander Mikheev]

It is a problem when you want implement something like the UDP server
that
handles all request on the single socket:

bind(fd, ...);
listen(fd);
while(1) {
	read_request(fd, ...);
	do_work();
	write_responce(fd, ...);
}

Correct me if I am wrong, but implementing this scenario for UD will
look like this: 
/* accept handling not shown... */
rdma_bind_addr(id, ...);
rdma_listen(id)
while(1) {
	read_request();
	do_work();
	if(new_client) { /* check if we already connected to client */
		/* following is not going to work because local address
is already bound */
		rdma_resolve_addr(my_local_already_bound_addr,
client_addr);
		do_connect();
	}
	write_responce();
}

If rdma_bind is done for the specific local ip and port, then I would
like to be able to resolve address from the specific ip/port to the
client.

-----Original Message-----
From: Sean Hefty [mailto:mshefty at ichips.intel.com] 
Sent: Monday, August 06, 2007 8:15 PM
To: Or Gerlitz
Cc: Sean Hefty; Alexander Mikheev; Alex Rosenbaum;
general at lists.openfabrics.org
Subject: Re: [ofa-general] implicit address binding by rdma_resolve_addr

> We have noticed that when rdma_resolve_addr is provided with src
> address, the rdma cm does an implicit --address-- (=ip/port) binding,
> which for example has the side effect of disallowing two processes to
> call rdma_resolve_addr() with the same src port.
> 
> Does this implicit binding mimics the network stack behavior
> under similar flow when socket are used?

Specifying a source address with rdma_resolve_addr is similar to calling

socket bind.  I can't think of a reason why an app would want to specify

a source port when calling rdma_resolve_addr.  Specifying a source 
address (wild card port) allows binding to a specific RDMA device.

- Sean