Re: [v3 PATCH 1/2] bonding: sync netpoll code with bridge

[Cong Wang <amwang@redhat.com>]

于 2010年12月09日 15:33, Cong Wang 写道:
> On 12/08/10 21:57, Neil Horman wrote:
>> On Wed, Dec 08, 2010 at 02:52:08AM -0500, Amerigo Wang wrote:
>>> -    bond_for_each_slave(bond, slave, i) {
>>> -        if ((slave->dev->priv_flags&  IFF_DISABLE_NETPOLL) ||
>>> -            !slave->dev->netdev_ops->ndo_poll_controller)
>>> -            ret = false;
>>> +    np = kmalloc(sizeof(*np), GFP_KERNEL);
>>> +    err = -ENOMEM;
>>> +    if (!np)
>>> +        goto out;
>>> +
>>> +    np->dev = slave->dev;
>>> +    err = __netpoll_setup(np);
>> Setting up our own netpoll instance on each slave worries me a bit.  The
>> implication here is that, by doing so, some frames will get entirely processed
>> by the slave.  Most notably arp frames.  That means anything that gets queued up
>> to the arp_tx queue in __netpoll_rx will get processed during that poll event,
>> and responded to with the mac of the slave device, rather than with the mac of
>> the bond device, which isn't always what you want.  I think if you go with this
>> route, you'll need to add code to netpoll_poll_dev, right before the call to
>> service_arp_queue, to check if IFF_SLAVE is set in priv_flags, and move the list
>> to the master device, or some such.
>
>
> Good point! Will fix i

Hi, Neil,

I think we should do that in bond_poll_controller() rather than netpoll_poll_dev(),
right? Since this is bond-specific. Does moving all arp_tx of slaves to their bond
address your concern?

Thanks!