Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:4 as permitted sender) client-ip=2620:137:e000::3:4;
From:   Henrik =?ISO-8859-1?Q?Lindstr=F6m?= <lindstrom515@gmail.com>
To:     Florian Westphal <fw@strlen.de>
Cc:     Florian Westphal <fw@strlen.de>, davem@davemloft.net,
        edumazet@google.com, kuba@kernel.org, pabeni@redhat.com,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: macvtap performs IP defragmentation,
 causing MTU problems for virtual machines
Date:   Thu, 05 Oct 2023 19:25:53 +0200
Message-ID: <3259970.44csPzL39Z@pc>
In-Reply-To: <20231004080037.GC15013@breakpoint.cc>
References: <CAHkKap3sdN4wZm_euAZEyt3XB4bvr6cV-oAMGtrmrm5Z8biZ_Q@mail.gmail.com>
 <2197902.NgBsaNRSFp@pc> <20231004080037.GC15013@breakpoint.cc>
MIME-Version: 1.0
Content-Transfer-Encoding: 7Bit
Content-Type: text/plain; charset="us-ascii"
Precedence: bulk

On onsdag 4 oktober 2023 10:00:37 CEST Florian Westphal wrote:
> Can you submit this formally, with proper changelog and Signed-off-by?
> See scripts/checkpatch.pl in the kernel tree.
Sure, i can give it a shot. How do i properly credit you if i submit your
patch with some small changes of my own?

> You could also mention in changelog that this is ipv4 only because
> ipv6 already considers the interface index during reassembly.
Interesting. I've been trying to understand the code and it seems like
ipv6 does defragmentation per-interface, while ipv4 does it "per-vrf"
(correct me if i'm wrong). Is there any reason for this difference? 

I also did some more testing with the diff from my previous mail. It 
looks like the problem remains for interfaces under vrfs. I think simply
doing the bcast/mcast check first fixes that though, something like this:
	if (skb->pkt_type == PACKET_BROADCAST || skb->pkt_type == PACKET_MULTICAST) {
		if (dev)
			return dev->ifindex;
	}

	return l3mdev_master_ifindex_rcu(dev);
Does that look reasonable?
The idea being that bcast/mcast packets are always defragmented
per-interface, and unicast packets always "per-vrf".

Thanks,
Henrik