Re: [PATCH v2 net 3/3] wireguard: send: account for mtu=0 devices

[Eric Dumazet <eric.dumazet@gmail.com>]

On 2/14/20 10:37 AM, Jason A. Donenfeld wrote:
> On 2/14/20, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>>
>>
>> On 2/14/20 10:15 AM, Jason A. Donenfeld wrote:
>>> On Fri, Feb 14, 2020 at 6:56 PM Eric Dumazet <eric.dumazet@gmail.com>
>>> wrote:
>>>> Oh dear, can you describe what do you expect of a wireguard device with
>>>> mtu == 0 or mtu == 1
>>>>
>>>> Why simply not allowing silly configurations, instead of convoluted tests
>>>> in fast path ?
>>>>
>>>> We are speaking of tunnels adding quite a lot of headers, so we better
>>>> not try to make them
>>>> work on networks with tiny mtu. Just say no to syzbot.
>>>
>>> The idea was that wireguard might still be useful for the persistent
>>> keepalive stuff. This branch becomes very cold very fast, so I don't
>>> think it makes a difference performance wise, but if you feel strongly
>>> about it, I can get rid of it and set a non-zero min_mtu that's the
>>> smallest thing wireguard's xmit semantics will accept. It sounds like
>>> you'd prefer that?
>>>
>> Well, if you believe that wireguard in persistent keepalive
>> has some value on its own, I guess that we will have to support this mode.
> 
> Alright.
> 
>>
>> Some legacy devices can have arbitrary mtu, and this has caused headaches.
>> I was hoping that for brand new devices, we could have saner limits.
>>
>> About setting max_mtu to ~MAX_INT, does it mean wireguard will attempt
>> to send UDP datagrams bigger than 64K ? Where is the segmentation done ?
> 
> The before passings off to the udp tunnel api, we indicate that we
> support ip segmentation, and then it gets handled and fragmented
> deeper down. Check out socket.c. 

Okay. Speaking of socket.c, I found this wg_socket_reinit() snippet :

synchronize_rcu();
synchronize_net();

Which makes little sense. Please add a comment explaining why these two
calls are needed.