Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [PATCH] neighbour: Disregard DEAD dst in neigh_update
Date:   Thu, 7 Jan 2021 21:36:37 -0500
From:   Your Real Name <zhutong@amazon.com>
To:     David Miller <davem@davemloft.net>
CC:     <sashal@kernel.org>, <edumazet@google.com>, <vvs@virtuozzo.com>,
        <netdev@vger.kernel.org>, <stable@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>
Message-ID: <20210108023637.GA31904@ucf43ac461c9a53.ant.amazon.com>
References: <20201230225415.GA490@ucf43ac461c9a53.ant.amazon.com>
 <20210105.160521.1279064249478522010.davem@davemloft.net>
MIME-Version: 1.0
Content-Type: text/plain; charset="utf-8"
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20210105.160521.1279064249478522010.davem@davemloft.net>
User-Agent: Mutt/1.9.4 (2018-02-28)
Precedence: bulk

On Tue, Jan 05, 2021 at 04:05:21PM -0800, David Miller wrote: 
> 
> 
> From: Tong Zhu <zhutong@amazon.com>
> Date: Wed, 30 Dec 2020 17:54:23 -0500
> 
> > In 4.x kernel a dst in DST_OBSOLETE_DEAD state is associated
> > with loopback net_device and leads to loopback neighbour. It
> > leads to an ethernet header with all zero addresses.
> >
> > A very troubling case is working with mac80211 and ath9k.
> > A packet with all zero source MAC address to mac80211 will
> > eventually fail ieee80211_find_sta_by_ifaddr in ath9k (xmit.c).
> > As result, ath9k flushes tx queue (ath_tx_complete_aggr) without
> > updating baw (block ack window), damages baw logic and disables
> > transmission.
> >
> > Signed-off-by: Tong Zhu <zhutong@amazon.com>
> 
> Please repost with an appropriate Fixes: tag.
> 
> Thanks.

I had a second thought on this. This fix should go mainline too. This is a 
case we are sending out queued packets when arp reply from the neighbour 
comes in. With 5.x kernel, a dst in DST_OBSOLETE_DEAD state leads to dropping
of this packet. It is not as bad as with 4.x kernel that may end up with an
all-zero mac address packet out to ethernet or choking up ath9k when using 
block ack. Dropping the packet is still wrong. I’ll repost as a fix to
mainline and target backport to 4.x LTS releases.

Best regards