Patchwork tcp: properly update lost_cnt_hint during shifting

On 09/28/2011 04:17 PM, Ilpo Järvinen wrote:
> On Wed, 28 Sep 2011, Yan, Zheng wrote:
> 
>> lost_skb_hint is used by tcp_mark_head_lost() to mark the first
>> unhandled skb. lost_cnt_hint is the number of sacked packets before
>> the lost_skb_hint. tcp_shifted_skb() shouldn't increase lost_cnt_hint
>> when shifting a sacked skb that is before the lost_skb_hint, because
>> packets in it are already counted.
>>
>> Signed-off-by: Zheng Yan <zheng.z.yan@intel.com>
>> ---
>> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
>> index 21fab3e..f712ace 100644
>> --- a/net/ipv4/tcp_input.c
>> +++ b/net/ipv4/tcp_input.c
>> @@ -1390,9 +1390,14 @@ static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
>>  	BUG_ON(!pcount);
>>  
>>  	/* Tweak before seqno plays */
>> -	if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
>> -	    !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
>> -		tp->lost_cnt_hint += pcount;
>> +	if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint) {
>> +		if (skb == tp->lost_skb_hint)
>> +			tp->lost_cnt_hint += pcount;
>> +		else if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) &&
>> +			 before(TCP_SKB_CB(skb)->seq,
>> +				TCP_SKB_CB(tp->lost_skb_hint)->seq))
>> +			tp->lost_cnt_hint += pcount;
>> +	}
> 
> Ah right, the hole filled case which shifts not only the newly SACKed 
> skb but also the next, already SACKed skb?
> 
> I fail to see why you needed to change !before into two checks though:
>  skb == tp->lost_skb_hint and before(params reversed) ? Shouldn't the 
> equality that is provided by the negation cover for the == check (and the 
> params reversion isn't necessary in any case)? In fact, isn't the skb == 
> tp->lost_skb_hint check strictly wrong without the same TCPCB_SACKED_ACKED 
> guard (though I'm not sure, I didn't check, if the hint can ever point to 
> such a segment in the first place)?

Thanks you for your reply.

skb == tp->lost_skb_hint is special.

If the skb is sacked and we shift 'pcount' packets to previous skb,
these packets will not be counted by future tcp_mark_head_lost() call.
So we should increase lost_cnt_hint.

If the skb is not sacked, the skb will be sacked soon by tcp_sacktag_one(),
So we should not increase lost_cnt_hint.

I didn't think out the second case. I think the correct patch should be:
---

---


> 
> Added Cc to Nandita as they're hunting (possibly other) bug in 
> tcp_mark_head_lost.
> 

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html