Re: [PATCH] tcp: properly update lost_cnt_hint during shifting

["Yan, Zheng" <zheng.z.yan@intel.com>]

On 09/28/2011 05:50 PM, Ilpo Järvinen wrote:
> On Wed, 28 Sep 2011, Yan, Zheng wrote:
> 
>> On 09/28/2011 04:55 PM, Yan, Zheng wrote:
>>> 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:
>>> ---
>>>
>>> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
>>> index 21fab3e..dcc2411 100644
>>> --- a/net/ipv4/tcp_input.c
>>> +++ b/net/ipv4/tcp_input.c
>>> @@ -1390,9 +1390,15 @@ 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 ((TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) &&
>>> +		    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;
>>> +	}
>>>  
>>>  	TCP_SKB_CB(prev)->end_seq += shifted;
>>>  	TCP_SKB_CB(skb)->seq += shifted;
>>> ---
>>
>> Sorry, I didn't think out the "skb before lost_skb_hint" case neither.
>> If the skb isn't sacked, tcp_sacktag_one() will increase the lost_cnt_hint.
>> So tcp_shifted_skb() shouldn't adjust the the lost_cnt_hint.
>>
>> I hope my patch is correct this time.
>>
>> ---
>> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
>> index 21fab3e..697ce5f 100644
>> --- a/net/ipv4/tcp_input.c
>> +++ b/net/ipv4/tcp_input.c
>> @@ -1390,8 +1390,8 @@ 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))
>> +	if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint == skb &&
>> +	    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
>>  		tp->lost_cnt_hint += pcount;
>>  
>>  	TCP_SKB_CB(prev)->end_seq += shifted;
> 
> Hehe, besides not spotting all this, I also made another mistake in my 
> last post. It seems that this code has been quite broken from the 
> beginning or we still lack some detail. ...But the latest change certainly 
> seems more reasonable than the previous code of mine if I've successfully 
> understood enough pieces. These hints, although providing significant 
> performance benefits, are really pain to get right :-).
> 
> But is the non-SACKed case really handled right when hint == skb by the 
> sacktag_one. We move the seqno in between and then before(x->newseq, 
> x->newseq) check returns false?
> 
you are right, thank you.

really hope my patch is correct this time :)
---
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 21fab3e..a04622e 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -1390,8 +1390,7 @@ 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))
+	if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint == skb)
 		tp->lost_cnt_hint += pcount;
 
 	TCP_SKB_CB(prev)->end_seq += shifted;