[PATCH 3.8 06/91] tcp: TSO packets automatic sizing

[Kamal Mostafa <kamal@canonical.com>]

3.8.13.13 -stable review patch.  If anyone has any objections, please let me know.

------------------

From: Eric Dumazet <edumazet@google.com>

commit 95bd09eb27507691520d39ee1044d6ad831c1168 upstream.
commit 02cf4ebd82ff0ac7254b88e466820a290ed8289a upstream.
commit 7eec4174ff29cd42f2acfae8112f51c228545d40 upstream.

After hearing many people over past years complaining against TSO being
bursty or even buggy, we are proud to present automatic sizing of TSO
packets.

One part of the problem is that tcp_tso_should_defer() uses an heuristic
relying on upcoming ACKS instead of a timer, but more generally, having
big TSO packets makes little sense for low rates, as it tends to create
micro bursts on the network, and general consensus is to reduce the
buffering amount.

This patch introduces a per socket sk_pacing_rate, that approximates
the current sending rate, and allows us to size the TSO packets so
that we try to send one packet every ms.

This field could be set by other transports.

Patch has no impact for high speed flows, where having large TSO packets
makes sense to reach line rate.

For other flows, this helps better packet scheduling and ACK clocking.

This patch increases performance of TCP flows in lossy environments.

A new sysctl (tcp_min_tso_segs) is added, to specify the
minimal size of a TSO packet (default being 2).

A follow-up patch will provide a new packet scheduler (FQ), using
sk_pacing_rate as an input to perform optional per flow pacing.

This explains why we chose to set sk_pacing_rate to twice the current
rate, allowing 'slow start' ramp up.

sk_pacing_rate = 2 * cwnd * mss / srtt

v2: Neal Cardwell reported a suspect deferring of last two segments on
initial write of 10 MSS, I had to change tcp_tso_should_defer() to take
into account tp->xmit_size_goal_segs

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Van Jacobson <vanj@google.com>
Cc: Tom Herbert <therbert@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
[ kamal: backport to 3.8 (context) ]
Signed-off-by: Kamal Mostafa <kamal@canonical.com>
---
 Documentation/networking/ip-sysctl.txt |  9 +++++++++
 include/net/sock.h                     |  2 ++
 include/net/tcp.h                      |  1 +
 net/core/sock.c                        |  1 +
 net/ipv4/sysctl_net_ipv4.c             | 10 ++++++++++
 net/ipv4/tcp.c                         | 28 ++++++++++++++++++++++-----
 net/ipv4/tcp_input.c                   | 35 +++++++++++++++++++++++++++++++++-
 net/ipv4/tcp_output.c                  |  2 +-
 8 files changed, 81 insertions(+), 7 deletions(-)

diff --git a/Documentation/networking/ip-sysctl.txt b/Documentation/networking/ip-sysctl.txt
index dbca661..62b9a61 100644
--- a/Documentation/networking/ip-sysctl.txt
+++ b/Documentation/networking/ip-sysctl.txt
@@ -510,6 +510,15 @@ tcp_syn_retries - INTEGER
 tcp_timestamps - BOOLEAN
 	Enable timestamps as defined in RFC1323.
 
+tcp_min_tso_segs - INTEGER
+	Minimal number of segments per TSO frame.
+	Since linux-3.12, TCP does an automatic sizing of TSO frames,
+	depending on flow rate, instead of filling 64Kbytes packets.
+	For specific usages, it's possible to force TCP to build big
+	TSO frames. Note that TCP stack might split too big TSO packets
+	if available window is too small.
+	Default: 2
+
 tcp_tso_win_divisor - INTEGER
 	This allows control over what percentage of the congestion window
 	can be consumed by a single TSO frame.
diff --git a/include/net/sock.h b/include/net/sock.h
index 873abca..94871cc 100644
--- a/include/net/sock.h
+++ b/include/net/sock.h
@@ -228,6 +228,7 @@ struct cg_proto;
   *	@sk_wmem_queued: persistent queue size
   *	@sk_forward_alloc: space allocated forward
   *	@sk_allocation: allocation mode
+  *	@sk_pacing_rate: Pacing rate (if supported by transport/packet scheduler)
   *	@sk_sndbuf: size of send buffer in bytes
   *	@sk_flags: %SO_LINGER (l_onoff), %SO_BROADCAST, %SO_KEEPALIVE,
   *		   %SO_OOBINLINE settings, %SO_TIMESTAMPING settings
@@ -352,6 +353,7 @@ struct sock {
 	kmemcheck_bitfield_end(flags);
 	int			sk_wmem_queued;
 	gfp_t			sk_allocation;
+	u32			sk_pacing_rate; /* bytes per second */
 	netdev_features_t	sk_route_caps;
 	netdev_features_t	sk_route_nocaps;
 	int			sk_gso_type;
diff --git a/include/net/tcp.h b/include/net/tcp.h
index 4da2167..45f3368 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -292,6 +292,7 @@ extern int sysctl_tcp_thin_dupack;
 extern int sysctl_tcp_early_retrans;
 extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
+extern int sysctl_tcp_min_tso_segs;
 
 extern atomic_long_t tcp_memory_allocated;
 extern struct percpu_counter tcp_sockets_allocated;
diff --git a/net/core/sock.c b/net/core/sock.c
index b8af814..fc0d751 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -2258,6 +2258,7 @@ void sock_init_data(struct socket *sock, struct sock *sk)
 
 	sk->sk_stamp = ktime_set(-1L, 0);
 
+	sk->sk_pacing_rate = ~0U;
 	/*
 	 * Before updating sk_refcnt, we must commit prior changes to memory
 	 * (Documentation/RCU/rculist_nulls.txt for details)
diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c
index d22765d..ae245cb 100644
--- a/net/ipv4/sysctl_net_ipv4.c
+++ b/net/ipv4/sysctl_net_ipv4.c
@@ -29,6 +29,7 @@
 static int zero;
 static int one = 1;
 static int two = 2;
+static int gso_max_segs = GSO_MAX_SEGS;
 static int tcp_retr1_max = 255;
 static int ip_local_port_range_min[] = { 1, 1 };
 static int ip_local_port_range_max[] = { 65535, 65535 };
@@ -781,6 +782,15 @@ static struct ctl_table ipv4_table[] = {
 		.extra2		= &two,
 	},
 	{
+		.procname	= "tcp_min_tso_segs",
+		.data		= &sysctl_tcp_min_tso_segs,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= proc_dointvec_minmax,
+		.extra1		= &zero,
+		.extra2		= &gso_max_segs,
+	},
+	{
 		.procname	= "udp_mem",
 		.data		= &sysctl_udp_mem,
 		.maxlen		= sizeof(sysctl_udp_mem),
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 8e79542..c2aaeab 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -282,6 +282,8 @@
 
 int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
 
+int sysctl_tcp_min_tso_segs __read_mostly = 2;
+
 struct percpu_counter tcp_orphan_count;
 EXPORT_SYMBOL_GPL(tcp_orphan_count);
 
@@ -793,12 +795,28 @@ static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 	xmit_size_goal = mss_now;
 
 	if (large_allowed && sk_can_gso(sk)) {
-		xmit_size_goal = ((sk->sk_gso_max_size - 1) -
-				  inet_csk(sk)->icsk_af_ops->net_header_len -
-				  inet_csk(sk)->icsk_ext_hdr_len -
-				  tp->tcp_header_len);
+		u32 gso_size, hlen;
+
+		/* Maybe we should/could use sk->sk_prot->max_header here ? */
+		hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
+		       inet_csk(sk)->icsk_ext_hdr_len +
+		       tp->tcp_header_len;
+
+		/* Goal is to send at least one packet per ms,
+		 * not one big TSO packet every 100 ms.
+		 * This preserves ACK clocking and is consistent
+		 * with tcp_tso_should_defer() heuristic.
+		 */
+		gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC);
+		gso_size = max_t(u32, gso_size,
+				 sysctl_tcp_min_tso_segs * mss_now);
+
+		xmit_size_goal = min_t(u32, gso_size,
+				       sk->sk_gso_max_size - 1 - hlen);
 
-		/* TSQ : try to have two TSO segments in flight */
+		/* TSQ : try to have at least two segments in flight
+		 * (one in NIC TX ring, another in Qdisc)
+		 */
 		xmit_size_goal = min_t(u32, xmit_size_goal,
 				       sysctl_tcp_limit_output_bytes >> 1);
 
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index c834d83..b2263d8 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -705,6 +705,34 @@ static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
 	}
 }
 
+/* Set the sk_pacing_rate to allow proper sizing of TSO packets.
+ * Note: TCP stack does not yet implement pacing.
+ * FQ packet scheduler can be used to implement cheap but effective
+ * TCP pacing, to smooth the burst on large writes when packets
+ * in flight is significantly lower than cwnd (or rwin)
+ */
+static void tcp_update_pacing_rate(struct sock *sk)
+{
+	const struct tcp_sock *tp = tcp_sk(sk);
+	u64 rate;
+
+	/* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
+	rate = (u64)tp->mss_cache * 2 * (HZ << 3);
+
+	rate *= max(tp->snd_cwnd, tp->packets_out);
+
+	/* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
+	 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
+	 * We probably need usec resolution in the future.
+	 * Note: This also takes care of possible srtt=0 case,
+	 * when tcp_rtt_estimator() was not yet called.
+	 */
+	if (tp->srtt > 8 + 2)
+		do_div(rate, tp->srtt);
+
+	sk->sk_pacing_rate = min_t(u64, rate, ~0U);
+}
+
 /* Calculate rto without backoff.  This is the second half of Van Jacobson's
  * routine referred to above.
  */
@@ -3605,7 +3633,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 	u32 ack_seq = TCP_SKB_CB(skb)->seq;
 	u32 ack = TCP_SKB_CB(skb)->ack_seq;
 	bool is_dupack = false;
-	u32 prior_in_flight;
+	u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
 	u32 prior_fackets;
 	int prior_packets = tp->packets_out;
 	int prior_sacked = tp->sacked_out;
@@ -3723,6 +3751,9 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 		if (dst)
 			dst_confirm(dst);
 	}
+
+	if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
+		tcp_update_pacing_rate(sk);
 	return 1;
 
 no_queue:
@@ -6063,6 +6094,8 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
 				} else
 					tcp_init_metrics(sk);
 
+				tcp_update_pacing_rate(sk);
+
 				/* Prevent spurious tcp_cwnd_restart() on
 				 * first data packet.
 				 */
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 436d8fb..9d2de62 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -1784,7 +1784,7 @@ static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
 
 	/* If a full-sized TSO skb can be sent, do it. */
 	if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
-			   sk->sk_gso_max_segs * tp->mss_cache))
+			   tp->xmit_size_goal_segs * tp->mss_cache))
 		goto send_now;
 
 	/* Middle in queue won't get any more data, full sendable already? */
-- 
1.8.1.2