Re: [RFC PATCH] powerpc/pseries/mobility: notify network peers after migration

[Thomas Falcon <tlfalcon@linux.ibm.com>]

On 11/6/19 4:14 PM, Nathan Lynch wrote:
> Hi Tom,
>
> Thomas Falcon <tlfalcon@linux.ibm.com> writes:
>> After a migration, it is necessary to send a gratuitous ARP
>> from all running interfaces so that the rest of the network
>> is aware of its new location. However, some supported network
>> devices are unaware that they have been migrated. To avoid network
>> interruptions and other unwanted behavior, force a GARP on all
>> valid, running interfaces as part of the post_mobility_fixup
>> routine.
> [...]
>
>> @@ -331,6 +334,8 @@ void post_mobility_fixup(void)
>>   {
>>   	int rc;
>>   	int activate_fw_token;
>> +	struct net_device *netdev;
>> +	struct net *net;
>>   
>>   	activate_fw_token = rtas_token("ibm,activate-firmware");
>>   	if (activate_fw_token == RTAS_UNKNOWN_SERVICE) {
>> @@ -371,6 +376,21 @@ void post_mobility_fixup(void)
>>   	/* Possibly switch to a new RFI flush type */
>>   	pseries_setup_rfi_flush();
>>   
>> +	/* need to force a gratuitous ARP on running interfaces */
>> +	rtnl_lock();
>> +	for_each_net(net) {
>> +		for_each_netdev(net, netdev) {
>> +			if (netif_device_present(netdev) &&
>> +			    netif_running(netdev) &&
>> +			    !(netdev->flags & (IFF_NOARP | IFF_LOOPBACK)))
>> +				call_netdevice_notifiers(NETDEV_NOTIFY_PEERS,
>> +							 netdev);
>> +				call_netdevice_notifiers(NETDEV_RESEND_IGMP,
>> +							 netdev);
>> +		}
>> +	}
>> +	rtnl_unlock();
>> +
> This isn't an outright nak, but this is not nice. It illustrates the
> need to rethink the pseries partition migration code. There is no
> mechanism for drivers and other interested code to prepare for a
> migration or to adjust to the destination. So post_mobility_fixup() will
> continue to grow into a fragile collection of calls into unrelated
> subsystems until there is a better design -- either a pseries-specific
> notification/callback mechanism, or something based on the pm framework.
>
> My understanding is that this is needed specifically for ibmveth and,
> unlike ibmvnic, the platform does not provide any notification to that
> driver that a migration has occurred, right?

Correct, the ibmveth device, unlike ibmvnic, receives no signal or 
notification at all in the event of a partition migration, so it can not 
handle it or send a gratuitous ARP because from the driver's perspective 
nothing has changed.  As you've described, there is no existing notifier 
in the kernel to inform interested parties that the system has migrated 
or is about to migrate. Without adding the needed infrastructure to do 
that, I'm not sure how else to fix this.

Tom