[PATCH net-next] tcp: batch calls to sk_flush_backlog()

[PATCH net-next] tcp: batch calls to sk_flush_backlog()

[Eric Dumazet]

Starting from commit d41a69f1d390 ("tcp: make tcp_sendmsg() aware of socket backlog")
loopback flows got hurt, because for each skb sent, the socket receives an
immediate ACK and sk_flush_backlog() causes extra work.

Intent was to not let the backlog grow too much, but we went a bit too far.

We can check the backlog every 16 skbs (about 1MB chunks)
to increase TCP over loopback performance by about 15 %

Note that the call to sk_flush_backlog() handles a single ACK,
thanks to coalescing done on backlog, but cleans the 16 skbs
found in rtx rb-tree.

Reported-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
---
 net/ipv4/tcp.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index a0a66321c0ee99918b2080219dbaefcf3c398e13..f8fa1686f7f3e64f5d4ea8163e7f87538cc0d672 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -1162,7 +1162,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 	struct sockcm_cookie sockc;
 	int flags, err, copied = 0;
 	int mss_now = 0, size_goal, copied_syn = 0;
-	bool process_backlog = false;
+	int process_backlog = 0;
 	bool zc = false;
 	long timeo;
 
@@ -1254,9 +1254,10 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 			if (!sk_stream_memory_free(sk))
 				goto wait_for_sndbuf;
 
-			if (process_backlog && sk_flush_backlog(sk)) {
-				process_backlog = false;
-				goto restart;
+			if (unlikely(process_backlog >= 16)) {
+				process_backlog = 0;
+				if (sk_flush_backlog(sk))
+					goto restart;
 			}
 			first_skb = tcp_rtx_and_write_queues_empty(sk);
 			skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
@@ -1264,7 +1265,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 			if (!skb)
 				goto wait_for_memory;
 
-			process_backlog = true;
+			process_backlog++;
 			skb->ip_summed = CHECKSUM_PARTIAL;
 
 			skb_entail(sk, skb);
-- 
2.23.0.rc1.153.gdeed80330f-goog

Re: [PATCH net-next] tcp: batch calls to sk_flush_backlog()

[Soheil Hassas Yeganeh]

On Fri, Aug 9, 2019 at 8:04 AM Eric Dumazet <edumazet@google.com> wrote:
>
> Starting from commit d41a69f1d390 ("tcp: make tcp_sendmsg() aware of socket backlog")
> loopback flows got hurt, because for each skb sent, the socket receives an
> immediate ACK and sk_flush_backlog() causes extra work.
>
> Intent was to not let the backlog grow too much, but we went a bit too far.
>
> We can check the backlog every 16 skbs (about 1MB chunks)
> to increase TCP over loopback performance by about 15 %
>
> Note that the call to sk_flush_backlog() handles a single ACK,
> thanks to coalescing done on backlog, but cleans the 16 skbs
> found in rtx rb-tree.
>
> Reported-by: Soheil Hassas Yeganeh <soheil@google.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>

Acked-by: Soheil Hassas Yeganeh <soheil@google.com>

Thank you very much, Eric!

> ---
>  net/ipv4/tcp.c | 11 ++++++-----
>  1 file changed, 6 insertions(+), 5 deletions(-)
>
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index a0a66321c0ee99918b2080219dbaefcf3c398e13..f8fa1686f7f3e64f5d4ea8163e7f87538cc0d672 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -1162,7 +1162,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
>         struct sockcm_cookie sockc;
>         int flags, err, copied = 0;
>         int mss_now = 0, size_goal, copied_syn = 0;
> -       bool process_backlog = false;
> +       int process_backlog = 0;
>         bool zc = false;
>         long timeo;
>
> @@ -1254,9 +1254,10 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
>                         if (!sk_stream_memory_free(sk))
>                                 goto wait_for_sndbuf;
>
> -                       if (process_backlog && sk_flush_backlog(sk)) {
> -                               process_backlog = false;
> -                               goto restart;
> +                       if (unlikely(process_backlog >= 16)) {
> +                               process_backlog = 0;
> +                               if (sk_flush_backlog(sk))
> +                                       goto restart;
>                         }
>                         first_skb = tcp_rtx_and_write_queues_empty(sk);
>                         skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
> @@ -1264,7 +1265,7 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
>                         if (!skb)
>                                 goto wait_for_memory;
>
> -                       process_backlog = true;
> +                       process_backlog++;
>                         skb->ip_summed = CHECKSUM_PARTIAL;
>
>                         skb_entail(sk, skb);
> --
> 2.23.0.rc1.153.gdeed80330f-goog
>

Re: [PATCH net-next] tcp: batch calls to sk_flush_backlog()

[David Miller]

From: Eric Dumazet <edumazet@google.com>
Date: Fri,  9 Aug 2019 05:04:47 -0700

> Starting from commit d41a69f1d390 ("tcp: make tcp_sendmsg() aware of socket backlog")
> loopback flows got hurt, because for each skb sent, the socket receives an
> immediate ACK and sk_flush_backlog() causes extra work.
> 
> Intent was to not let the backlog grow too much, but we went a bit too far.
> 
> We can check the backlog every 16 skbs (about 1MB chunks)
> to increase TCP over loopback performance by about 15 %
> 
> Note that the call to sk_flush_backlog() handles a single ACK,
> thanks to coalescing done on backlog, but cleans the 16 skbs
> found in rtx rb-tree.
> 
> Reported-by: Soheil Hassas Yeganeh <soheil@google.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>

Applied.