[openib-general] Re: [IBAT] interface numbering assumption

[James Lentini]

On Fri, 23 Sep 2005, Hal Rosenstock wrote:

> > Is there a better way to enumerate all of the network inferaces? I 
> > believe that is what this for loop is attempting to accomplish.
> 
> Yes. I think that the net_device list from dev_base could be walked
> instead and that would resolve this issue.

Can you help we understand the logic in at.c:resolve_ip()? Here is my 
assumption of what this function does:

 1) consults the IP routing table for an interface 
    device using ip_route_output_key

 2) if the device does not meet certain criteria, return an error

 3) if the device is a loopback device, search for another device 
    that is an INFINIBAND device and is UP.

 4) ...

I've included a small patch below to fix the problem I observed in #3. 
It walks the dev_base list as you described.

However I don't understand why the device returned in step #1 isn't 
always used as I assumpe this is the interface the routing table says 
to use. That makes me think I've misinterpreted the purpose of 
ip_route_output_key. What am I missing?

--
Signed-off-by: James Lentini <jlentini at netapp.com>

Index: core/at.c
===================================================================
--- core/at.c	(revision 3528)
+++ core/at.c	(working copy)
@@ -430,7 +430,7 @@
 			}
 		},
 	};
-	int i, r;
+	int r;
 
 	DEBUG("dst ip %08x src ip %08x tos %d", dst_ip, src_ip, tos);
 
@@ -490,13 +490,15 @@
 	src->gw  = rt->rt_gateway;
 	src->ip = rt->rt_src;     /* true source IP address */
 
-	if (ipoib_dev->flags & IFF_LOOPBACK)
-		for (i = 1; (ipoib_dev = dev_get_by_index(i)); i++) {
-			dev_put(ipoib_dev);
+	if (ipoib_dev->flags & IFF_LOOPBACK) {
+		read_lock(&dev_base_lock);
+		for(ipoib_dev = dev_base; ipoib_dev; 
+		    ipoib_dev = ipoib_dev->next)
 			if (ARPHRD_INFINIBAND == ipoib_dev->type &&
 			    (ipoib_dev->flags & IFF_UP))
 				break;
-		}
+		read_unlock(&dev_base_lock);
+	}
 
 	if (!ipoib_dev) {
 		WARN("No device for IB comm <%s:%08x:%08x>",