Re: [PATCH net] ip_tunnel: fix over-mtu packet send fail without TUNNEL_DONT_FRAGMENT flags

From: David Ahern <dsahern@gmail.com>
To: Jakub Kicinski <kuba@kernel.org>, wenxu <wenxu@ucloud.cn>
Cc: netdev@vger.kernel.org, Stefano Brivio <sbrivio@redhat.com>,
	David Ahern <dsahern@kernel.org>
Subject: Re: [PATCH net] ip_tunnel: fix over-mtu packet send fail without TUNNEL_DONT_FRAGMENT flags
Date: Tue, 27 Oct 2020 08:51:07 -0600	[thread overview]
Message-ID: <8e24e490-b3bf-5268-4bd5-98b598b36b36@gmail.com> (raw)
In-Reply-To: <20201026135626.23684484@kicinski-fedora-PC1C0HJN.hsd1.ca.comcast.net>

On 10/26/20 2:56 PM, Jakub Kicinski wrote:
> On Mon, 26 Oct 2020 16:23:29 +0800 wenxu wrote:
>> On 10/24/2020 5:12 AM, Jakub Kicinski wrote:
>>> On Wed, 21 Oct 2020 17:21:55 +0800 wenxu@ucloud.cn wrote:  
>>>> From: wenxu <wenxu@ucloud.cn>
>>>>
>>>> The TUNNEL_DONT_FRAGMENT flags specific the tunnel outer ip can do
>>>> fragment or not in the md mode. Without the TUNNEL_DONT_FRAGMENT
>>>> should always do fragment. So it should not care the frag_off in
>>>> inner ip.  
>>> Can you describe the use case better? My understanding is that we
>>> should propagate DF in normally functioning networks, and let PMTU 
>>> do its job.  
>>
>> Sorry for relying so late.  ip_md_tunnel_xmit send packet in the collect_md mode.
>>
>> For OpenVswitch example, ovs set the gre port with flags df_default=false which will not
>>
>> set TUNNEL_DONT_FRAGMENT for tun_flags.
>>
>> And the mtu of virtual machine is 1500 with default. And the tunnel underlay device mtu
>>
>> is 1500 default too. So if the size of packet send from vm +  underlay length > underlay device mtu.
>>
>> The packet always be dropped if the ip header of  packet set flags with DF.
>>
>> In the collect_md the outer packet can fragment or not should depends on the tun_flags but not inner
>>
>> ip header like vxlan device did.
> 
> Is this another incarnation of 4cb47a8644cc ("tunnels: PMTU discovery
> support for directly bridged IP packets")? Sounds like non-UDP tunnels
> need the same treatment to make PMTUD work.
> 
> RFC2003 seems to clearly forbid ignoring the inner DF:

I was looking at this patch Sunday night. To me it seems odd that
packets flowing through the overlay affect decisions in the underlay
which meant I agree with the proposed change.

ip_md_tunnel_xmit is inconsistent right now. tnl_update_pmtu is called
based on the TUNNEL_DONT_FRAGMENT flag, so why let it be changed later
based on the inner header? Or, if you agree with RFC 2003 and the DF
should be propagated outer to inner, then it seems like the df reset
needs to be moved up before the call to tnl_update_pmtu