Re: [RFC PATCH v0 1/2] net: bridge: propagate FDB table into hardware

[John Fastabend <john.r.fastabend@intel.com>]

On 2/8/2012 8:36 PM, Stephen Hemminger wrote:
> On Wed, 08 Feb 2012 19:22:06 -0800
> John Fastabend <john.r.fastabend@intel.com> wrote:
> 
>> Propagate software FDB table into hardware uc, mc lists when
>> the NETIF_F_HW_FDB is set.
>>
>> This resolves the case below where an embedded switch is used
>> in hardware to do inter-VF or VF-PF switching. This patch
>> pushes the FDB entry (specifically the MAC address) into the
>> embedded switch with dev_add_uc and dev_add_mc so the switch
>> "learns" about the software bridge.
>>
>>
>>           veth0  veth2
>>             |      |
>>           ------------
>>           |  bridge0 |   <---- software bridging
>>           ------------
>>                /
>>                /
>>   ethx.y      ethx
>>     VF         PF
>>      \         \          <---- propagate FDB entries to HW
>>      \         \
>>   --------------------
>>   |  Embedded Bridge |    <---- hardware offloaded switching
>>   --------------------
>>
>> This is only an RFC couple more changes are needed.
>>
>> (1) Optimize HW FDB set/del to only walk list if an FDB offloaded
>>     device is attached. Or decide it doesn't matter from unlikely()
>>     path.
>>
>> (2) Is it good enough to just call dev_uc_{add|del} or
>>     dev_mc_{add|del}? Or do some devices really need a new netdev
>>     callback to do this operation correctly. I think it should be
>>     good enough as is.
>>
>> (3) wrapped list walk in rcu_read_lock() just in case maybe every
>>     case is already inside rcu_read_lock()/unlock().
>>
>> Also this is in response to this thread regarding the macvlan and
>> exposing rx filters posting now to see if folks think this is the
>> right idea and if it will resolve at least the bridge case.
>>
>> http://lists.openwall.net/netdev/2011/11/08/135
>>
>> Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
>> ---
>>
>>  include/linux/netdev_features.h |    2 ++
>>  net/bridge/br_fdb.c             |   34 ++++++++++++++++++++++++++++++++++
>>  2 files changed, 36 insertions(+), 0 deletions(-)
>>
>> diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h
>> index 77f5202..5936fae 100644
> 
> Rather than yet another device feature, I would rather use netlink_notifier
> callback. The notifier is more general and generic without messing with internals
> of bridge.
> 

But the device features makes it easy for user space to learn that the device
supports this sort of offload. Now if all SR-IOV devices support this then it
doesn't matter but I thought there were SR-IOV devices that didn't do any
switching? I'll dig through the SR-IOV drivers to check there are not too
many of them.

By netlink_notifier do you mean adding a notifier_block and using atomic_notifier_call_chain()
probably in rtnl_notify()? Then drivers could register with the notifier chain with
atomic_notifier_chain_register() and receive the events correctly. Or did I miss
some notifier chain that already exists?

Thanks,
John