netdev - Re: [PATCH] af_packet: Fix softirq mismatch in tpacket

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <2c72e9ad-657b-a498-0fe5-1224a4b3dc1a@candelatech.com>
Date: Fri, 20 Sep 2024 09:16:52 -0700
From: Ben Greear <greearb@...delatech.com>
To: David Ahern <dsahern@...nel.org>,
 Willem de Bruijn <willemdebruijn.kernel@...il.com>, netdev@...r.kernel.org,
 Ido Schimmel <idosch@...dia.com>
Subject: Re: [PATCH] af_packet: Fix softirq mismatch in tpacket_rcv

On 9/20/24 08:01, David Ahern wrote:
> On 9/20/24 2:57 AM, Willem de Bruijn wrote:
>> Ben Greear wrote:
>>> On 9/19/24 13:00, David Ahern wrote:
>>>> On 9/19/24 10:44 AM, Willem de Bruijn wrote:
>>>>> Yes, it seems that VRF calls dev_queue_xmit_nit without the same BH
>>>>> protections that it expects.
>>>>>
>>>>> I suspect that the fix is in VRF, to disable BH the same way that
>>>>> __dev_queue_xmit does, before calling dev_queue_xmit_nit.
>>>>>
>>>>
>>>> commit 504fc6f4f7f681d2a03aa5f68aad549d90eab853 removed the bh around
>>>> dev_queue_xmit_nit:
>>>>
>>>> diff --git a/drivers/net/vrf.c b/drivers/net/vrf.c
>>>> index 6043e63b42f9..43f374444684 100644
>>>> --- a/drivers/net/vrf.c
>>>> +++ b/drivers/net/vrf.c
>>>> @@ -638,9 +638,7 @@ static void vrf_finish_direct(struct sk_buff *skb)
>>>>                   eth_zero_addr(eth->h_dest);
>>>>                   eth->h_proto = skb->protocol;
>>>>
>>>> -               rcu_read_lock_bh();
>>>>                   dev_queue_xmit_nit(skb, vrf_dev);
>>>> -               rcu_read_unlock_bh();
>>>>
>>>>                   skb_pull(skb, ETH_HLEN);
>>>>           }
>>>
>>> So I guess we should revert this?
>>
>> Looks like it to me.
>>
>> In which case good to not just revert, but explain why, and probably
>> copy the comment that is present in __dev_queue_xmit.
>>
> 
> Ben: does it resolve the problem you were investigating?
> 
> It would be good to add a selftest that sets up a VRF, attaches tcpdump
> and then sends a few seconds of iperf3 traffic through it. That should
> be similar to the use case here and I expect it to create a similar
> crash. That should help prevent a regression in addition to the comment.

We'll test in next day or two and let you know, but at least the patch
I posted previously 'fixed' things, so likely the revert will as well.

And, I think you need a 'real' eth port in the VRF, as original likely broken
commit claimed to test with veth and tcpdump successfully, probably because
veth rcv path is not in soft-irq or something?

Thanks,
Ben

-- 
Ben Greear <greearb@...delatech.com>
Candela Technologies Inc  http://www.candelatech.com