[ofa-general] [PATCH] ipoib: refresh path when remote lid changes

[Yossi Etigin]

On 02/08/09 14:26, Hal Rosenstock wrote:
> 
> By handled correctly, you mean that the ARP request gets to the remote
> node, is responded to, and the response makes it back and that is
> treated as valid path indication, right ?
> 
> If so, is the original ARP request unicast or broadcast ?
> 
> If the request is unicast, couldn't it be sent using the wrong static
> rate as isn't it using the original path parameters ?
> 
> Even if it is broadcast, if the original path parameters are still
> used (like rate, etc.) at the local node, doesn't this assume a
> homogeneous subnet ?

By handled correctly, I mean that:
- If the LID is not changed, the mechanism will not trigger path refresh.
      (The first patch without any LMC handling does not satisfy this)
- If the LID is changed, the mechanism will trigger a path refresh (eventually)

The ARP stuff works this way: Remote LID changes. In some point, either the remote
node will send an ARP reply (gratuitous), or (more likely) the local network stack
will start sending solicited ARPs, unicast, using the invalid path. They will fail,
so the stack will send broadcast ARP. Then, the remote node will answer, and IPoIB
will see a different slid than expected. This will trigger path refresh.

The broadcast ARP is sent with the AH of the broadcast group (which is joined when
IPoIB interface goes up), and not the parameters of the path to any specific node.

--Yossi