Re: Inconsistency in packet drop due to MTU (eth vs veth)

[Fredrik Markstrom]

/F


 ---- On Tue, 31 Jan 2017 17:27:09 +0100 Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote ---- 
 > On Tue, 2017-01-31 at 14:32 +0100, Fredrik Markstrom wrote: 
 > >  ---- On Thu, 19 Jan 2017 19:53:47 +0100 Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote ----  
 > >  > On Thu, 2017-01-19 at 17:41 +0100, Fredrik Markstrom wrote:  
 > >  > > Hello,  
 > >  > >   
 > >  > > I've noticed an inconsistency between how physical ethernet and 
 > > veth handles mtu.  
 > >  > >   
 > >  > > If I setup two physical interfaces (directly connected) with 
 > > different mtu:s, only the size of the outgoing packets are limited by 
 > > the mtu. But with veth a packet is dropped if the mtu of the receiving 
 > > interface is smaller then the packet size.   
 > >  > >   
 > >  > > This seems inconsistent to me, but maybe there is a reason for 
 > > it ?   
 > >  > >   
 > >  > > Can someone confirm if it's a deliberate inconsistency or just a 
 > > side effect of using dev_forward_skb() ?  
 > >  >   
 > >  > It looks this was added in commit  
 > >  > 38d408152a86598a50680a82fe3353b506630409  
 > >  > ("veth: Allow setting the L3 MTU")  
 > >  >   
 > >  > But what was really needed here was a way to change MRU :(  
 > >  
 > > Ok, do we consider this correct and/or something we need to be 
 > > backwards compatible with ? Is it insane to believe that we can fix 
 > > this "inconsistency" by removing the check ? 
 > >  
 > > The commit message reads "For consistency I drop packets on the 
 > > receive side when they are larger than the MTU", do we know what it's 
 > > supposed 
 > > to be consistent with or is that lost in history ? 
 >  
 > There is no consistency among existing Ethernet drivers. 
 >  
 > Many ethernet drivers size the buffers they post in RX ring buffer 
 > according to MTU. 
 >  
 > If MTU is set to 1500, RX buffers are sized to be about 1536 bytes, 
 > so you wont be able to receive a 1700 bytes frame. 
 >  
 > I guess that you could add a specific veth attribute to precisely 
 > control MRU, that would not break existing applications. 

Ok, I will propose a patch shortly. And thanks, your response time is
awesome !

/Fredrik

 >  
 >  
 >  
 >