Re: TCP-MD5 checksum failure on x86_64 SMP

[Bhaskar Dutta <bhaskie@gmail.com>]

On Thu, May 6, 2010 at 12:23 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> Le mercredi 05 mai 2010 à 23:33 +0530, Bhaskar Dutta a écrit :
>
>> Hi,
>>
>> TSO, GSO and SG are already turned off.
>> rx/tx checksumming is on, but that shouldn't matter, right?
>>
>> # ethtool -k eth0
>> Offload parameters for eth0:
>> rx-checksumming: on
>> tx-checksumming: on
>> scatter-gather: off
>> tcp segmentation offload: off
>> udp fragmentation offload: off
>> generic segmentation offload: off
>>
>> The bad packets are very small in size, most have no data at all (<300 bytes).
>>
>> After adding some logs to kernel 2.6.31-12, it seems that
>> tcp_v4_md5_hash_skb (function that calculates the md5 hash) is
>> (might?) getting corrupt.
>>
>> The tcp4_pseudohdr (bp = &hp->md5_blk.ip4) structure's saddr, daddr
>> and len fields get modified to different values towards the end of the
>> tcp_v4_md5_hash_skb function whenever there is a checksum error.
>>
>> The tcp4_pseudohdr (bp) is within the tcp_md5sig_pool (hp), which is
>> filled up by tcp_get_md5sig_pool (which calls per_cpu_ptr).
>>
>> Using a local copy of the tcp4_pseudohdr in the same function
>> tcp_v4_md5_hash_skb (copied all fields from the original
>> tcp4_pseudohdr within the tcp_md5sig_pool) and calculating the md5
>> checksum with the local  tcp4_pseudohdr seems to solve the issue
>> (don't see bad packets for a hours in load tests, and without the
>> change I can see them instantaneously in the load tests).
>>
>> I am still unable to figure out how this is happening. Please let me
>> know if you have any pointers.
>
> I am not familiar with this code, but I suspect same per_cpu data can be
> used at both time by a sender (process context) and by a receiver
> (softirq context).
>
> To trigger this, you need at least two active md5 sockets.
>
> tcp_get_md5sig_pool() should probably disable bh to make sure current
> cpu wont be preempted by softirq processing
>
>
> Something like :
>
> diff --git a/include/net/tcp.h b/include/net/tcp.h
> index fb5c66b..e232123 100644
> --- a/include/net/tcp.h
> +++ b/include/net/tcp.h
> @@ -1221,12 +1221,15 @@ struct tcp_md5sig_pool          *tcp_get_md5sig_pool(void)
>        struct tcp_md5sig_pool *ret = __tcp_get_md5sig_pool(cpu);
>        if (!ret)
>                put_cpu();
> +       else
> +               local_bh_disable();
>        return ret;
>  }
>
>  static inline void             tcp_put_md5sig_pool(void)
>  {
>        __tcp_put_md5sig_pool();
> +       local_bh_enable();
>        put_cpu();
>  }
>
>
>

I put in the above change and ran some load tests with around 50
active TCP connections doing MD5.
I could see only 1 bad packet in 30 min (earlier the problem used to
occur instantaneously and repeatedly).

I think there is another possibility of being preempted when calling
tcp_alloc_md5sig_pool()
this function releases the spinlock when calling __tcp_alloc_md5sig_pool().

I will run some more tests after changing the  tcp_alloc_md5sig_pool
and see if the problem is completely resolved.

Thanks a lot for your help!
Bhaskar