Re: [PATCH net-next] bonding: RCUify bond_set_rx_mode()

[Veaceslav Falico <vfalico@redhat.com>]

On Mon, Aug 05, 2013 at 12:21:56PM +0200, Nikolay Aleksandrov wrote:
>On 08/05/2013 11:26 AM, Veaceslav Falico wrote:
>> Currently, we might easily deadlock with bond_set_rx_mode() and
>> bond_hw_addr_swap(). bond_set_rx_mode() is called via dev_set_rx_mode(),
>> which already holds the netif_addr_lock_bh(bond), and inside it takes the
>> bond->curr_active_slave lock, while bond_hw_addr_swap() is called with
>> bond->curr_active_slave lock held and then takes netif_addr_lock_bh(bond),
>> which results in deadlock.
>>
>> CPU0                    CPU1
>> ----                    ----
>> lock(&bonding_netdev_addr_lock_key);
>> 			lock(&bond->curr_slave_lock);
>> 			lock(&bonding_netdev_addr_lock_key);
>> lock(&bond->curr_slave_lock);
>>
>> Fix this by using the RCU primites in bond_set_rx_mode(). We're safe wrt
>> racing of dev_?c_(un)sync() because we hold
>> lock(&bonding_netdev_addr_lock_key), and thus nobody will be able to modify
>> these lists before we finish.
>>
>Hi,
>I don't think this deadlock can actually happen because bond_hw_addr_swap() is
>called from bond_change_active_slave() only in USES_PRIMARY mode, and in such
>mode it's always called with rtnl acquired before that, and since
>dev_set_rx_mode is called with rtnl, IMO such deadlock can't happen.

Yep, indeed, missed the part with USES_PRIMARY(). So the lockdep had a
false alarm.

>Also I think bond_set_rx_mode() can work without RCU because of the held rtnl
>and converted to ASSERT_RTNL (this is optional) + rtnl_dereference for the
>curr_active_slave.

Yes, we don't need the real rcu cause we're under rtnl and everybody else
who touches it also is under rtnl. Awesome catch.

Thanks, will resubmit another patch (hard to call it v2...).

>
>Cheers,
> Nik
>
>> CC: Jay Vosburgh <fubar@us.ibm.com>
>> CC: Andy Gospodarek <andy@greyhouse.net>
>> CC: Nikolay Aleksandrov <nikolay@redhat.com>
>> Signed-off-by: Veaceslav Falico <vfalico@redhat.com>
>> ---
>