On 03/05/2016 03:44 PM, Xin Long wrote:
> On Thu, Mar 3, 2016 at 8:29 PM, Nikolay Aleksandrov
> <niko...@cumulusnetworks.com> wrote:
>>
>> This is incorrect because you don't have rtnl here, bridge device sysfs
>> options take care of rtnl only on per-option basis and they obtain and
>> release it themselves, so you won't have rtnl held when you call
>> netdev_state_change. While I agree that this is needed, a larger change
>> would be necessary for br_sysfs_br.c.
> Sorry, I can't follow you, cause I didn't see any held in dev_ioctl, like:
> ipip6_tunnel_ioctl
> ipip6_tunnel_update
> netdev_state_change
>
> why sysfs have to hold rtnl ?
>
See the comment above dev_ifsioc:
/*
* Perform the SIOCxIFxxx calls, inside rtnl_lock()
*/
static int dev_ifsioc(struct net *net, struct ifreq *ifr, unsigned int cmd)
{
...
it is usually called like:
rtnl_lock();
ret = dev_ifsioc(net, &ifr, cmd);
rtnl_unlock();
And also you cannot be calling netdevice notifiers without RTNL. So in any
case you do need it here as well, in fact you'll surely hit the ASSERT_RTNL();
in call_netdevice_notifiers_info if you do so, thus I'm not sure how this
patch was actually tested.
>> Off-topic: I've been looking into factoring out the bond option API and
>> reusing
>> it here as it already has all of this handled, but I won't have time to
>> finish
>> it before the next merge window, so if you fix the issue here I'm okay with
>> this as interim solution.
>>
>> Cheers,
>> Nik
>>
