[openib-general] [PATCH][RFC] CMA automatic port number assignment

[Steve Wise]

On Fri, 2006-02-24 at 12:55 -0500, James Lentini wrote:
> The RDMA CM does not automatically assign a port number to the active 
> consumer on IB.
> 
> My expectation was that if the active consumer called 
> 
>  rdma_create_id()
>  rdma_resolve_addr()
>  rdma_create_qp()
>  rdma_resolve_route()
>  rdma_connect()
> 
> analogous to the BSD sockets sequence
> 
>  socket(2)
>  connect(2)
> 
> the consumer would automatically be assigned a local TCP/UDP/SCTP port 
> number.
> 
> The simple patch below solves the problem by picking a random 
> ephemeral port number (where ephemeral is defined as any port > 1024).
> 
> This fix is not perfect.
> 
> The choice of port number does not coordinate with the relevant 
> transport stack (TCP/UDP/SDP). Given that the passive side CMA code 
> has the same behavior, this follows an established precedent.
> 
> Also, this method needs to be reviewed for security issues. There is 
> no attempt to ensure that multiple (src addr, dest addr, dest port) 
> triplets do not simultaneously use the same src port or to ensure that 
> the same port is not reused before some given timeout (e.g. 
> TIME_WAIT). Given that the src port is chosen at random, I would 
> expect both of these occurrences to be extremely unlikely.
> 
> As a reference, the actual TCP code makes this assignment in 
> tcp_v4_hash_connect() using the sysctl_local_port_range array to 
> determine the range of ephemeral ports and the 
> secure_tcp_port_ephemeral() function to choose a secure point to start 
> searching for an available port. Unfortunately neither of these are 
> available to module code.
> 

I don't think this is going to work for iWARP providers.  From an iWARP
perspective, either the CMA needs to track the entire port number space
across all providers (ie replicate what the native stack does), or it
needs to track port spaces per provider, or it needs to query the
provider to get an ephemeral port.  The problem is that the provider
might already have that port in use from a previous connection. 



> Signed-off-by: James Lentini <jlentini at netapp.com>
> 
> Index: core/cma.c
> ===================================================================
> --- core/cma.c	(revision 5489)
> +++ core/cma.c	(working copy)
> @@ -1361,6 +1361,14 @@ err:
>  }
>  EXPORT_SYMBOL(rdma_bind_addr);
>  
> +static u16 cma_generate_ephemeral_port(void)
> +{
> +	u16 port;
> +
> +	get_random_bytes(&port, sizeof port);
> +	return cpu_to_be16(port | 1024U);
> +}
> +
>  static void cma_format_hdr(void *hdr, enum rdma_port_space ps,
>  			   struct rdma_route *route)
>  {
> @@ -1371,6 +1379,9 @@ static void cma_format_hdr(void *hdr, en
>  	src4 = (struct sockaddr_in *) &route->addr.src_addr;
>  	dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
>  
> +	if (!src4->sin_port)
> +		src4->sin_port = cma_generate_ephemeral_port();
> +
>  	switch (ps) {
>  	case RDMA_PS_SDP:
>  		sdp_hdr = hdr;
> _______________________________________________
> openib-general mailing list
> openib-general at openib.org
> http://openib.org/mailman/listinfo/openib-general
> 
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general