Jason Gunthorpe wrote:
> On Wed, Aug 26, 2009 at 12:07:51PM +0300, Or Gerlitz wrote:
>
>> isn't Jason's approach enough for the bonding case?! I saw that your
>> patch ("bonding: clean muticast addresses when device changes type"
>
> I think working versions of all three patches are required:
> 1) Fix the bonding driver. Otherwise the right groups might not be
> joined.
> 2) Check the address format, to protect against 'ip maddr add' and
> other wakkyness
> 3) Fix the timeout handling, so mlid exhaustion and other SA side
> errors are handled elegantly.
>
> All are bugs..
>
>> and maybe also in mainline .31-rcX . However, it has the
>> down-side-effect of e.g loosing routes already set for the the bond
>> while adding the underline IPoIB devices, so if Jason's patch is
>> enough
>
> Is this true? That is pretty ugly, but probably manageable..
>
Yes it's true but I'm not sure it's ugly. Changing device type is not a common
event and requires device ops change which I think is better to do when the
device is closed. Unfortunately, losing routes is a side effect of closing the
device but it might be necessary.
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
