On 11/17/20 1:53 PM, Mahesh Bandewar (महेश बंडेवार) wrote:
> On Tue, Nov 17, 2020 at 9:18 AM Ido Schimmel <ido...@idosch.org> wrote:
>>
>> On Mon, Nov 16, 2020 at 01:03:32PM -0800, Mahesh Bandewar (महेश बंडेवार)
>> wrote:
>>> On Mon, Nov 16, 2020 at 12:34 PM Jakub Kicinski <k...@kernel.org> wrote:
>>>>
>>>> On Mon, 16 Nov 2020 12:02:48 -0800 Mahesh Bandewar (महेश बंडेवार) wrote:
>>>>>>> diff --git a/drivers/net/loopback.c b/drivers/net/loopback.c
>>>>>>> index a1c77cc00416..76dc92ac65a2 100644
>>>>>>> --- a/drivers/net/loopback.c
>>>>>>> +++ b/drivers/net/loopback.c
>>>>>>> @@ -219,6 +219,13 @@ static __net_init int loopback_net_init(struct net
>>>>>>> *net)
>>>>>>>
>>>>>>> BUG_ON(dev->ifindex != LOOPBACK_IFINDEX);
>>>>>>> net->loopback_dev = dev;
>>>>>>> +
>>>>>>> + if (sysctl_netdev_loopback_state) {
>>>>>>> + /* Bring loopback device UP */
>>>>>>> + rtnl_lock();
>>>>>>> + dev_open(dev, NULL);
>>>>>>> + rtnl_unlock();
>>>>>>> + }
>>>>>>
>>>>>> The only concern I have here is that it breaks notification ordering.
>>>>>> Is there precedent for NETDEV_UP to be generated before all pernet ops
>>>>>> ->init was called?
>>>>> I'm not sure if any and didn't see any issues in our usage / tests.
>>>>> I'm not even sure anyone is watching/monitoring for lo status as such.
>>>>
>>>> Ido, David, how does this sound to you?
>>>>
>>>> I can't think of any particular case where bringing the device up (and
>>>> populating it's addresses) before per netns init is finished could be
>>>> problematic. But if this is going to make kernel coding harder the
>>>> minor convenience of the knob is probably not worth it.
>>>
>>> +Eric Dumazet
>>>
>>> I'm not sure why kernel coding should get harder, but happy to listen
>>> to the opinions.
>>
>> Hi,
>>
>> Sorry for the delay. Does not occur to me as a problematic change. I ran
>> various tests with 'sysctl -qw net.core.netdev_loopback_state=1' and a
>> debug config. Looks OK.
>
> Thanks for the confirmation Ido. I think Jian is getting powerpc
> config build fixed to address the build-bots findings and then he can
> push the updated version.
>
If there is no harm in just creating lo in the up state, why not just do
it vs relying on a sysctl? It only affects 'local' networking so no real
impact to containers that do not do networking (ie., packets can't
escape). Linux has a lot of sysctl options; is this one really needed?
