Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [PATCH net-next] tcp: introduce skb_rbtree_walk_safe() and use it
 in tcp_clean_rtx_queue()
From:   Eric Dumazet <eric.dumazet@gmail.com>
To:     Yafang Shao <laoar.shao@gmail.com>, davem@davemloft.net,
        edumazet@google.com
Cc:     netdev@vger.kernel.org, linux-kernel@vger.kernel.org
References: <1543407379-3144-1-git-send-email-laoar.shao@gmail.com>
 <60467855-ea5f-3351-8b88-ffdfa60501e0@gmail.com>
 <408125e7-bbb0-dcfb-ee19-58089802cac7@gmail.com>
Message-ID: <b8b7bbd2-8f7c-2aef-e9d0-343f8c4c2b2d@gmail.com>
Date:   Wed, 28 Nov 2018 09:26:27 -0800
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101
 Thunderbird/52.9.1
MIME-Version: 1.0
In-Reply-To: <408125e7-bbb0-dcfb-ee19-58089802cac7@gmail.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


On 11/28/2018 08:53 AM, Eric Dumazet wrote:
> 
> 
> On 11/28/2018 06:44 AM, Eric Dumazet wrote:
>>
> 
>> Because we can break of the loop if the current skb is not fully acked.
>>
>> So your patch would add unnecessary overhead, since the extra sk_rb_next()
>> could add more extra cache line misses.
> 
> I am testing the following optimization, since we can avoid the rb_next() call
> when we reached snd_una
> 
> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> index f32397890b6dcbc34976954c4be142108efa04d8..6829e470f0c186a73c34dca414cd4a2b379baded 100644
> --- a/net/ipv4/tcp_input.c
> +++ b/net/ipv4/tcp_input.c
> @@ -3126,7 +3126,8 @@ static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
>                 if (!fully_acked)
>                         break;
>  
> -               next = skb_rb_next(skb);
> +               next = (scb->end_seq == tp->snd_una) ? NULL : skb_rb_next(skb);
> +
>                 if (unlikely(skb == tp->retransmit_skb_hint))
>                         tp->retransmit_skb_hint = NULL;
>                 if (unlikely(skb == tp->lost_skb_hint))
> 

This does not work since we use next skb after the loop :

1)   if (!skb)
          tcp_chrono_stop(sk, TCP_CHRONO_BUSY);

  This test could use tcp_rtx_queue_empty()

2) SACK reneging
    if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
            flag |= FLAG_SACK_RENEGING;

3) FLAG_SET_XMIT_TIMER
     } else if (skb && rtt_update && sack_rtt_us >= 0 && 
                sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
                                                 tcp_skb_timestamp_us(skb))) {
                flag |= FLAG_SET_XMIT_TIMER;  /* set TLP or RTO timer */
 
So this idea is not feasible.