Re: [RFC patch net-next-2.6] net: allow multiple rx_handler registration

[Ben Greear <greearb@candelatech.com>]

On 06/30/2011 11:28 AM, Jiri Pirko wrote:
> Thu, Jun 30, 2011 at 08:13:59PM CEST, greearb@candelatech.com wrote:
>> On 06/30/2011 08:16 AM, Jiri Pirko wrote:
>>> For some net topos it is necessary to have multiple "soft-net-devices"
>>> hooked on one netdev. For example very common is to have
>>> eth<->(br+vlan). Vlan is not using rh_handler (yet) but also for example
>>> macvlan would be useful to have hooked on same netdev as br.
>>>
>>> This patch introduces rx_handler list. size struct net_device stays
>>> intact. Measured performance regression on eth-br topo is ~1% (on received
>>> pkts generated by pktgen) and on eth-bond topo it is ~0.25%
>>>
>>> On br I think that the performance can be brought back maybe by using per-cpu
>>> variables to store port in rx_path (I must check this)
>>
>>> +enum rx_handler_prio {
>>> +	RX_HANDLER_PRIO_BRIDGE,
>>> +	RX_HANDLER_PRIO_BOND,
>>> +	RX_HANDLER_PRIO_MACVLAN,
>>> +};
>>
>> Maybe add RX_HANDLER_PRIO_LATER, RX_HANDLER_PRIO_FIRST
>> for other modules that want to link
>> here, but don't have specific ordering other than before
>> or after these specific types?
>>
>> Or maybe start PRIO_BRIDGE at 100, BOND 110, MACVLAN 120
>> to leave gaps.
>
> I was thinking about this. But that would be useful only for out of the
> tree drivers. For in tree drivers, new prio would be just inserted
> whereever.

There might be a reason to add a hook at various places, depending
on user prefs.  I don't know of one now, so it could stay as is and
we could change it later if the need came up.  I assume this PRIO stuff is
purely private to the kernel, so it should be easy to change...

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com