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

[Yossi Etigin]

On 31/07/09 20:50, Hal Rosenstock wrote:
> 
> 
> On Fri, Jul 31, 2009 at 12:13 PM, Yossi Etigin <yosefe at voltaire.com
> <mailto:yosefe at voltaire.com>> wrote:
> 
>     On 29/07/09 19:35, Hal Rosenstock wrote:
>     >
>     >
>     > On Wed, Jul 29, 2009 at 10:22 AM, Moni Shoua <monis at voltaire.com
>     <mailto:monis at voltaire.com>
>     > <mailto:monis at voltaire.com <mailto:monis at voltaire.com>>> wrote:
>     >
>     >     Hal Rosenstock wrote:
>     >     >
>     >     >
>     >     > On Wed, Jul 29, 2009 at 8:16 AM, Moni Shoua
>     <monis at voltaire.com <mailto:monis at voltaire.com>
>     >     <mailto:monis at voltaire.com <mailto:monis at voltaire.com>>
>     >     > <mailto:monis at voltaire.com <mailto:monis at voltaire.com>
>     <mailto:monis at voltaire.com <mailto:monis at voltaire.com>>>> wrote:
>     >     >
>     >     >
>     >     >     >     > Is the LMC in the above from the local port ?
>     >     Unfortunately, LMC
>     >     >     >     is not
>     >     >     >     > required to be uniform across the subnet so the
>     remote
>     >     >     port's LMC may
>     >     >     >     > not be the same as that on the local port.
>     >     >     >
>     >     >     >     opensm configures the same LMC to all endports
>     (CA) in the
>     >     >     fabric so
>     >     >     >     in which case do you suspect that
>     >     >     >     it will be different?
>     >     >     >
>     >     >     >
>     >     >     > I wasn't talking about the simplification that OpenSM uses
>     >     but rather
>     >     >     > what IBA allows. There may be other SMs which do this
>     or OpenSM
>     >     >     could be
>     >     >     > extended (with additional configuration for this).
>     >     >     >
>     >     >     > -- Hal
>     >     >
>     >     >     So I guess a possible solutions for that is to query each
>     >     suspected
>     >     >     node before making a decision to flush it.
>     >     >     When getting the node info response the true LMC can be
>     put in the
>     >     >     LID check function.
>     >     >
>     >     >
>     >     > I'm not following how you are saying to determine the true
>     LMC (of the
>     >     > remote port).
>     >     >
>     >     Send PortInfo query for that port
>     >
>     >
>     > from the IPoIB client ? I think this ends up needing to contact the SA
>     > to get this info rather than the port directly. Isn't that what
>     you were
>     > trying to avoid originally ? If that's the case, one way would've been
>     > if ARPs carried the LMC as well as the LID.
>     >
> 
>     What if we query the remote port LMC once, when the path is
>     resolved, and then
>     use it to mask the LID until the path is refreshed again?
> 
>  
> Just like the LID, remote LMC could change.
> 

Yes, but AFAIK the only "bad" case is if the LID stays the same but LMC changes to a lower
value. In this case the path refresh will not happen when it is supposed to. If LMC changes to a higher value it might trigger an unneeded path refresh, which is not bad because it will refresh the LMC as well.