[PATCH RFC] tcp: Allow sk_wmem_alloc to exceed sysctl_tcp_limit_output_bytes

[Jonathan Davies <jonathan.davies@citrix.com>]

Network drivers with slow TX completion can experience poor network transmit
throughput, limited by hitting the sk_wmem_alloc limit check in tcp_write_xmit.
The limit is 128 KB (by default), which means we are limited to two 64 KB skbs
in-flight. This has been observed to limit transmit throughput with xen-netfront
because its TX completion can be relatively slow compared to physical NIC
drivers.

There have been several modifications to the calculation of the sk_wmem_alloc
limit in the past. Here is a brief history:

 * Since TSQ was introduced, the queue size limit was
       sysctl_tcp_limit_output_bytes.

 * Commit c9eeec26 ("tcp: TSQ can use a dynamic limit") made the limit
       max(skb->truesize, sk->sk_pacing_rate >> 10).
   This allows more packets in-flight according to the estimated rate.

 * Commit 98e09386 ("tcp: tsq: restore minimal amount of queueing") made the
   limit
       max_t(unsigned int, sysctl_tcp_limit_output_bytes,
                           sk->sk_pacing_rate >> 10).
   This ensures at least sysctl_tcp_limit_output_bytes in flight but allowed
   more if rate estimation shows this to be worthwhile.

 * Commit 605ad7f1 ("tcp: refine TSO autosizing") made the limit
       min_t(u32, max(2 * skb->truesize, sk->sk_pacing_rate >> 10),
                  sysctl_tcp_limit_output_bytes).
   This meant that the limit can never exceed sysctl_tcp_limit_output_bytes,
   regardless of what rate estimation suggests. It's not clear from the commit
   message why this significant change was justified, changing
   sysctl_tcp_limit_output_bytes from being a lower bound to an upper bound.

This patch restores the behaviour that allows the limit to grow above
sysctl_tcp_limit_output_bytes according to the rate estimation.

This has been measured to improve xen-netfront throughput from a domU to dom0
from 5.5 Gb/s to 8.0 Gb/s. Or, in the case of transmitting from one domU to
another (on the same host), throughput rose from 2.8 Gb/s to 8.0 Gb/s. In the
latter case, TX completion is especially slow, explaining the large improvement.
These values were measured against 4.0-rc5 using "iperf -c <ip> -i 1" using
CentOS 7.0 VM(s) on Citrix XenServer 6.5 on a Dell R730 host with a pair of Xeon
E5-2650 v3 CPUs.

Fixes: 605ad7f184b6 ("tcp: refine TSO autosizing")
Signed-off-by: Jonathan Davies <jonathan.davies@citrix.com>
---
 net/ipv4/tcp_output.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 1db253e..3a49af8 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -2052,7 +2052,7 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 		 * One example is wifi aggregation (802.11 AMPDU)
 		 */
 		limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
-		limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
+		limit = max_t(u32, limit, sysctl_tcp_limit_output_bytes);
 
 		if (atomic_read(&sk->sk_wmem_alloc) > limit) {
 			set_bit(TSQ_THROTTLED, &tp->tsq_flags);
-- 
1.9.1