Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Sat, 16 Apr 2022 09:33:20 +0200
From:   Jakub Kicinski <kuba@kernel.org>
To:     Peilin Ye <yepeilin.cs@gmail.com>
Cc:     "David S. Miller" <davem@davemloft.net>,
        Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>,
        David Ahern <dsahern@kernel.org>,
        Peilin Ye <peilin.ye@bytedance.com>,
        Cong Wang <cong.wang@bytedance.com>,
        Feng Zhou <zhoufeng.zf@bytedance.com>, netdev@vger.kernel.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH net 2/2] ip6_gre: Fix skb_under_panic in __gre6_xmit()
Message-ID: <20220416093320.13f4ba1d@kernel.org>
In-Reply-To: <20220416065633.GA10882@bytedance>
References: <c5b7dc6020c93a1e7b40bc472fcdb6429999473e.1649715555.git.peilin.ye@bytedance.com>
        <9cd9ca4ac2c19be288cb8734a86eb30e4d9e2050.1649715555.git.peilin.ye@bytedance.com>
        <20220414131424.744aa842@kernel.org>
        <20220414200854.GA2729@bytedance>
        <20220415191133.0597a79a@kernel.org>
        <20220416065633.GA10882@bytedance>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Fri, 15 Apr 2022 23:56:33 -0700 Peilin Ye wrote:
> On Fri, Apr 15, 2022 at 07:11:33PM +0200, Jakub Kicinski wrote:
> > > Could you explain this a bit more?  It seems that commit 77a5196a804e
> > > ("gre: add sequence number for collect md mode.") added this
> > > intentionally.  
> > 
> > Interesting. Maybe a better way of dealing with the problem would be
> > rejecting SEQ if it's not set on the device itself.  
> 
> According to ip-link(8), the 'external' option is mutually exclusive
> with the '[o]seq' option.  In other words, a collect_md mode IP6GRETAP
> device should always have the TUNNEL_SEQ flag off in its
> 'tunnel->parms.o_flags'.
> 
> (However, I just tried:
> 
>   $ ip link add dev ip6gretap11 type ip6gretap oseq external
> 					       ^^^^ ^^^^^^^^
>  ...and my 'ip' executed it with no error.  I will take a closer look at
>  iproute2 later; maybe it's undefined behavior...)
> 
> How about:
> 
> 1. If 'external', then 'oseq' means "always turn off NETIF_F_LLTX, so
> it's okay to set TUNNEL_SEQ in e.g. eBPF";
> 
> 2. Otherwise, if 'external' but NOT 'oseq', then whenever we see a
> TUNNEL_SEQ in skb's tunnel info, we do something like WARN_ONCE() then
> return -EINVAL.

Maybe pr_warn_once(), no need for a stacktrace.

> > When the device is set up without the SEQ bit enabled it disables Tx
> > locking (look for LLTX). This means that multiple CPUs can try to do
> > the tunnel->o_seqno++ in parallel. Not catastrophic but racy for sure.  
> 
> Thanks for the explanation!  At first glance, I was wondering why don't
> we make 'o_seqno' atomic until I found commit b790e01aee74 ("ip_gre:
> lockless xmit").  I quote:
> 
> """
> Even using an atomic_t o_seq, we would increase chance for packets being
> out of order at receiver.
> """
> 
> I don't fully understand this out-of-order yet, but it seems that making
> 'o_seqno' atomic is not an option?

atomic_t would also work (if it has enough bits). Whatever is simplest
TBH. It's just about correctness, I don't think seq is widely used.

Thanks!