Re: Stable regression with 'tcp: allow splice() to build full TSO packets'

[Willy Tarreau <w@1wt.eu>]

Hi Eric,

On Thu, May 17, 2012 at 02:18:00PM +0200, Willy Tarreau wrote:
> Hi Eric,
> 
> I'm facing a regression in stable 3.2.17 and 3.0.31 which is
> exhibited by your patch 'tcp: allow splice() to build full TSO
> packets' which unfortunately I am very interested in !
> 
> What I'm observing is that TCP transmits using splice() stall
> quite quickly if I'm using pipes larger than 64kB (even 65537
> is enough to reliably observe the stall).
(...)

I managed to fix the issue and I really think that the fix makes sense.
I'm appending the patch, please could you review it and if approved,
push it for inclusion ?

BTW, your patch significantly improves performance here. On this
machine I was reaching max 515 Mbps of proxied traffic, and with
the patch I reach 665 Mbps with the same test ! I think you managed
to fix what caused splice() to always be slightly slower than
recv/send() for a long time in my tests with a number of NICs !

Thanks,
Willy

-------

>From 6da6a21798d0156e647a993c31782eec739fa5df Mon Sep 17 00:00:00 2001
From: Willy Tarreau <w@1wt.eu>
Date: Thu, 17 May 2012 16:48:56 +0200
Subject: [PATCH] tcp: force push data out when buffers are missing

Commit 2f533844242 (tcp: allow splice() to build full TSO packets)
significantly improved splice() performance for some workloads but
caused stalls when pipe buffers were larger than socket buffers.

The issue seems to happen when no data can be copied at all due to
lack of buffers, which results in pending data never being pushed.

This change checks if all pending data has been pushed or not and
pushes them when waiting for send buffers.
---
 net/ipv4/tcp.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 80b988f..f6d9e5f 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -850,7 +850,7 @@ new_segment:
 wait_for_sndbuf:
 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 wait_for_memory:
-		if (copied)
+		if (tcp_sk(sk)->pushed_seq != tcp_sk(sk)->write_seq)
 			tcp_push(sk, flags & ~MSG_MORE, mss_now, TCP_NAGLE_PUSH);
 
 		if ((err = sk_stream_wait_memory(sk, &timeo)) != 0)
-- 
1.7.2.1.45.g54fbc