Re: [PATCH net-next] tcp: prepare skbs for better sack shifting

[David Miller]

From: Eric Dumazet 
Date: Thu, 15 Sep 2016 09:33:02 -0700

> From: Eric Dumazet 
> 
> With large BDP TCP flows and lossy networks, it is very important
> to keep a low number of skbs in the write queue.
> 
> RACK and SACK processing can perform a linear scan of it.
> 
> We should avoid putting any payload in skb->head, so that SACK
> shifting can be done if needed.
> 
> With this patch, we allow to pack ~0.5 MB per skb instead of
> the 64KB initially cooked at tcp_sendmsg() time.
> 
> This gives a reduction of number of skbs in write queue by eight.
> tcp_rack_detect_loss() likes this.
> 
> We still allow payload in skb->head for first skb put in the queue,
> to not impact RPC workloads.
> 
> Signed-off-by: Eric Dumazet 
> Cc: Yuchung Cheng 

Applied.

Re: [PATCH net-next] tcp: prepare skbs for better sack shifting

[Yuchung Cheng]

On Thu, Sep 15, 2016 at 9:33 AM, Eric Dumazet  wrote:
>
> From: Eric Dumazet 
>
> With large BDP TCP flows and lossy networks, it is very important
> to keep a low number of skbs in the write queue.
>
> RACK and SACK processing can perform a linear scan of it.
>
> We should avoid putting any payload in skb->head, so that SACK
> shifting can be done if needed.
>
> With this patch, we allow to pack ~0.5 MB per skb instead of
> the 64KB initially cooked at tcp_sendmsg() time.
>
> This gives a reduction of number of skbs in write queue by eight.
> tcp_rack_detect_loss() likes this.
>
> We still allow payload in skb->head for first skb put in the queue,
> to not impact RPC workloads.
>
> Signed-off-by: Eric Dumazet 
> Cc: Yuchung Cheng 
Acked-by: Yuchung Cheng 


> ---
>  net/ipv4/tcp.c |   31 ---
>  1 file changed, 24 insertions(+), 7 deletions(-)
>
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index 
> a13fcb369f52fe85def7c9d856259bc0509f3453..7dae800092e62cec330544851289d20a68642561
>  100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -1020,17 +1020,31 @@ int tcp_sendpage(struct sock *sk, struct page *page, 
> int offset,
>  }
>  EXPORT_SYMBOL(tcp_sendpage);
>
> -static inline int select_size(const struct sock *sk, bool sg)
> +/* Do not bother using a page frag for very small frames.
> + * But use this heuristic only for the first skb in write queue.
> + *
> + * Having no payload in skb->head allows better SACK shifting
> + * in tcp_shift_skb_data(), reducing sack/rack overhead, because
> + * write queue has less skbs.
> + * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
> + * This also speeds up tso_fragment(), since it wont fallback
> + * to tcp_fragment().
> + */
> +static int linear_payload_sz(bool first_skb)
> +{
> +   if (first_skb)
> +   return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
> +   return 0;
> +}
> +
> +static int select_size(const struct sock *sk, bool sg, bool first_skb)
>  {
> const struct tcp_sock *tp = tcp_sk(sk);
> int tmp = tp->mss_cache;
>
> if (sg) {
> if (sk_can_gso(sk)) {
> -   /* Small frames wont use a full page:
> -* Payload will immediately follow tcp header.
> -*/
> -   tmp = SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
> +   tmp = linear_payload_sz(first_skb);
> } else {
> int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
>
> @@ -1161,6 +1175,8 @@ restart:
> }
>
> if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
> +   bool first_skb;
> +
>  new_segment:
> /* Allocate new segment. If the interface is SG,
>  * allocate skb fitting to single page.
> @@ -1172,10 +1188,11 @@ new_segment:
> process_backlog = false;
> goto restart;
> }
> +   first_skb = skb_queue_empty(>sk_write_queue);
> skb = sk_stream_alloc_skb(sk,
> - select_size(sk, sg),
> + select_size(sk, sg, 
> first_skb),
>   sk->sk_allocation,
> - 
> skb_queue_empty(>sk_write_queue));
> + first_skb);
> if (!skb)
> goto wait_for_memory;
>
>
>