[TCP]: Move to new TSO segmenting scheme.

Make TSO segment transmit size decisions at send time not earlier.

The basic scheme is that we try to build as large a TSO frame as
possible when pulling in the user data, but the size of the TSO frame
output to the card is determined at transmit time.

This is guided by tp->xmit_size_goal.  It is always set to a multiple
of MSS and tells sendmsg/sendpage how large an SKB to try and build.

Later, tcp_write_xmit() and tcp_push_one() chop up the packet if
necessary and conditions warrant.  These routines can also decide to
"defer" in order to wait for more ACKs to arrive and thus allow larger
TSO frames to be emitted.

A general observation is that TSO elongates the pipe, thus requiring a
larger congestion window and larger buffering especially at the sender
side.  Therefore, it is important that applications 1) get a large
enough socket send buffer (this is accomplished by our dynamic send
buffer expansion code) 2) do large enough writes.

Signed-off-by: David S. Miller <davem@davemloft.net>

4 files changed, 380 insertions, 236 deletions

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 2ba73bf..29894c7 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -615,7 +615,7 @@ static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffse
 			 size_t psize, int flags)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
-	int mss_now;
+	int mss_now, size_goal;
 	int err;
 	ssize_t copied;
 	long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
@@ -628,6 +628,7 @@ static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffse
 	clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 
 	mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
+	size_goal = tp->xmit_size_goal;
 	copied = 0;
 
 	err = -EPIPE;
@@ -641,7 +642,7 @@ static ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffse
 		int offset = poffset % PAGE_SIZE;
 		int size = min_t(size_t, psize, PAGE_SIZE - offset);
 
-		if (!sk->sk_send_head || (copy = mss_now - skb->len) <= 0) {
+		if (!sk->sk_send_head || (copy = size_goal - skb->len) <= 0) {
 new_segment:
 			if (!sk_stream_memory_free(sk))
 				goto wait_for_sndbuf;
@@ -652,7 +653,7 @@ new_segment:
 				goto wait_for_memory;
 
 			skb_entail(sk, tp, skb);
-			copy = mss_now;
+			copy = size_goal;
 		}
 
 		if (copy > size)
@@ -693,7 +694,7 @@ new_segment:
 		if (!(psize -= copy))
 			goto out;
 
-		if (skb->len != mss_now || (flags & MSG_OOB))
+		if (skb->len < mss_now || (flags & MSG_OOB))
 			continue;
 
 		if (forced_push(tp)) {
@@ -713,6 +714,7 @@ wait_for_memory:
 			goto do_error;
 
 		mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
+		size_goal = tp->xmit_size_goal;
 	}
 
 out:
@@ -754,7 +756,7 @@ ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
 
 static inline int select_size(struct sock *sk, struct tcp_sock *tp)
 {
-	int tmp = tp->mss_cache_std;
+	int tmp = tp->mss_cache;
 
 	if (sk->sk_route_caps & NETIF_F_SG) {
 		if (sk->sk_route_caps & NETIF_F_TSO)
@@ -778,7 +780,7 @@ int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct sk_buff *skb;
 	int iovlen, flags;
-	int mss_now;
+	int mss_now, size_goal;
 	int err, copied;
 	long timeo;
 
@@ -797,6 +799,7 @@ int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 	clear_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 
 	mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
+	size_goal = tp->xmit_size_goal;
 
 	/* Ok commence sending. */
 	iovlen = msg->msg_iovlen;
@@ -819,7 +822,7 @@ int tcp_sendmsg(struct kiocb *iocb, struct sock *sk, struct msghdr *msg,
 			skb = sk->sk_write_queue.prev;
 
 			if (!sk->sk_send_head ||
-			    (copy = mss_now - skb->len) <= 0) {
+			    (copy = size_goal - skb->len) <= 0) {
 
 new_segment:
 				/* Allocate new segment. If the interface is SG,
@@ -842,7 +845,7 @@ new_segment:
 					skb->ip_summed = CHECKSUM_HW;
 
 				skb_entail(sk, tp, skb);
-				copy = mss_now;
+				copy = size_goal;
 			}
 
 			/* Try to append data to the end of skb. */
@@ -937,7 +940,7 @@ new_segment:
 			if ((seglen -= copy) == 0 && iovlen == 0)
 				goto out;
 
-			if (skb->len != mss_now || (flags & MSG_OOB))
+			if (skb->len < mss_now || (flags & MSG_OOB))
 				continue;
 
 			if (forced_push(tp)) {
@@ -957,6 +960,7 @@ wait_for_memory:
 				goto do_error;
 
 			mss_now = tcp_current_mss(sk, !(flags&MSG_OOB));
+			size_goal = tp->xmit_size_goal;
 		}
 	}
 
@@ -2128,7 +2132,7 @@ void tcp_get_info(struct sock *sk, struct tcp_info *info)
 
 	info->tcpi_rto = jiffies_to_usecs(tp->rto);
 	info->tcpi_ato = jiffies_to_usecs(tp->ack.ato);
-	info->tcpi_snd_mss = tp->mss_cache_std;
+	info->tcpi_snd_mss = tp->mss_cache;
 	info->tcpi_rcv_mss = tp->ack.rcv_mss;
 
 	info->tcpi_unacked = tp->packets_out;
@@ -2178,7 +2182,7 @@ int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
 
 	switch (optname) {
 	case TCP_MAXSEG:
-		val = tp->mss_cache_std;
+		val = tp->mss_cache;
 		if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
 			val = tp->rx_opt.user_mss;
 		break;
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 2ef2f35..8de2f10 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -740,10 +740,10 @@ __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
 	__u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
 
 	if (!cwnd) {
-		if (tp->mss_cache_std > 1460)
+		if (tp->mss_cache > 1460)
 			cwnd = 2;
 		else
-			cwnd = (tp->mss_cache_std > 1095) ? 3 : 4;
+			cwnd = (tp->mss_cache > 1095) ? 3 : 4;
 	}
 	return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
 }
@@ -914,7 +914,7 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
 	if (sk->sk_route_caps & NETIF_F_TSO) {
 		sk->sk_route_caps &= ~NETIF_F_TSO;
 		sock_set_flag(sk, SOCK_NO_LARGESEND);
-		tp->mss_cache = tp->mss_cache_std;
+		tp->mss_cache = tp->mss_cache;
 	}
 
 	if (!tp->sacked_out)
@@ -1077,7 +1077,7 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
 			    (IsFack(tp) ||
 			     !before(lost_retrans,
 				     TCP_SKB_CB(skb)->ack_seq + tp->reordering *
-				     tp->mss_cache_std))) {
+				     tp->mss_cache))) {
 				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
 				tp->retrans_out -= tcp_skb_pcount(skb);
 
@@ -3334,7 +3334,7 @@ static void tcp_new_space(struct sock *sk)
 	struct tcp_sock *tp = tcp_sk(sk);
 
 	if (tcp_should_expand_sndbuf(sk, tp)) {
- 		int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache_std) +
+ 		int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
 			MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
 		    demanded = max_t(unsigned int, tp->snd_cwnd,
 						   tp->reordering + 1);
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index ebf1123..62f62bb 100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -2045,7 +2045,7 @@ static int tcp_v4_init_sock(struct sock *sk)
 	 */
 	tp->snd_ssthresh = 0x7fffffff;	/* Infinity */
 	tp->snd_cwnd_clamp = ~0;
-	tp->mss_cache_std = tp->mss_cache = 536;
+	tp->mss_cache = 536;
 
 	tp->reordering = sysctl_tcp_reordering;
 	tp->ca_ops = &tcp_init_congestion_ops;
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 0a4cd24..fd3ce38 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -49,7 +49,7 @@ int sysctl_tcp_retrans_collapse = 1;
  * will allow a single TSO frame to consume.  Building TSO frames
  * which are too large can cause TCP streams to be bursty.
  */
-int sysctl_tcp_tso_win_divisor = 8;
+int sysctl_tcp_tso_win_divisor = 3;
 
 static inline void update_send_head(struct sock *sk, struct tcp_sock *tp,
 				    struct sk_buff *skb)
@@ -403,21 +403,11 @@ static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
 		sk->sk_send_head = skb;
 }
 
-static inline void tcp_tso_set_push(struct sk_buff *skb)
-{
-	/* Force push to be on for any TSO frames to workaround
-	 * problems with busted implementations like Mac OS-X that
-	 * hold off socket receive wakeups until push is seen.
-	 */
-	if (tcp_skb_pcount(skb) > 1)
-		TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
-}
-
 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 
-	if (skb->len <= tp->mss_cache_std ||
+	if (skb->len <= tp->mss_cache ||
 	    !(sk->sk_route_caps & NETIF_F_TSO)) {
 		/* Avoid the costly divide in the normal
 		 * non-TSO case.
@@ -427,164 +417,10 @@ static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb)
 	} else {
 		unsigned int factor;
 
-		factor = skb->len + (tp->mss_cache_std - 1);
-		factor /= tp->mss_cache_std;
+		factor = skb->len + (tp->mss_cache - 1);
+		factor /= tp->mss_cache;
 		skb_shinfo(skb)->tso_segs = factor;
-		skb_shinfo(skb)->tso_size = tp->mss_cache_std;
-	}
-}
-
-/* Does SKB fit into the send window? */
-static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
-{
-	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
-
-	return !after(end_seq, tp->snd_una + tp->snd_wnd);
-}
-
-/* Can at least one segment of SKB be sent right now, according to the
- * congestion window rules?  If so, return how many segments are allowed.
- */
-static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
-{
-	u32 in_flight, cwnd;
-
-	/* Don't be strict about the congestion window for the final FIN.  */
-	if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
-		return 1;
-
-	in_flight = tcp_packets_in_flight(tp);
-	cwnd = tp->snd_cwnd;
-	if (in_flight < cwnd)
-		return (cwnd - in_flight);
-
-	return 0;
-}
-
-static inline int tcp_minshall_check(const struct tcp_sock *tp)
-{
-	return after(tp->snd_sml,tp->snd_una) &&
-		!after(tp->snd_sml, tp->snd_nxt);
-}
-
-/* Return 0, if packet can be sent now without violation Nagle's rules:
- * 1. It is full sized.
- * 2. Or it contains FIN. (already checked by caller)
- * 3. Or TCP_NODELAY was set.
- * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
- *    With Minshall's modification: all sent small packets are ACKed.
- */
-
-static inline int tcp_nagle_check(const struct tcp_sock *tp,
-				  const struct sk_buff *skb, 
-				  unsigned mss_now, int nonagle)
-{
-	return (skb->len < mss_now &&
-		((nonagle&TCP_NAGLE_CORK) ||
-		 (!nonagle &&
-		  tp->packets_out &&
-		  tcp_minshall_check(tp))));
-}
-
-/* Return non-zero if the Nagle test allows this packet to be
- * sent now.
- */
-static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
-				 unsigned int cur_mss, int nonagle)
-{
-	/* Nagle rule does not apply to frames, which sit in the middle of the
-	 * write_queue (they have no chances to get new data).
-	 *
-	 * This is implemented in the callers, where they modify the 'nonagle'
-	 * argument based upon the location of SKB in the send queue.
-	 */
-	if (nonagle & TCP_NAGLE_PUSH)
-		return 1;
-
-	/* Don't use the nagle rule for urgent data (or for the final FIN).  */
-	if (tp->urg_mode ||
-	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
-		return 1;
-
-	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
-		return 1;
-
-	return 0;
-}
-
-/* This must be invoked the first time we consider transmitting
- * SKB onto the wire.
- */
-static inline int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb)
-{
-	int tso_segs = tcp_skb_pcount(skb);
-
-	if (!tso_segs) {
-		tcp_set_skb_tso_segs(sk, skb);
-		tso_segs = tcp_skb_pcount(skb);
-	}
-	return tso_segs;
-}
-
-/* This checks if the data bearing packet SKB (usually sk->sk_send_head)
- * should be put on the wire right now.  If so, it returns the number of
- * packets allowed by the congestion window.
- */
-static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
-				 unsigned int cur_mss, int nonagle)
-{
-	struct tcp_sock *tp = tcp_sk(sk);
-	unsigned int cwnd_quota;
-
-	tcp_init_tso_segs(sk, skb);
-
-	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
-		return 0;
-
-	cwnd_quota = tcp_cwnd_test(tp, skb);
-	if (cwnd_quota &&
-	    !tcp_snd_wnd_test(tp, skb, cur_mss))
-		cwnd_quota = 0;
-
-	return cwnd_quota;
-}
-
-static inline int tcp_skb_is_last(const struct sock *sk, 
-				  const struct sk_buff *skb)
-{
-	return skb->next == (struct sk_buff *)&sk->sk_write_queue;
-}
-
-int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
-{
-	struct sk_buff *skb = sk->sk_send_head;
-
-	return (skb &&
-		tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
-			     (tcp_skb_is_last(sk, skb) ?
-			      TCP_NAGLE_PUSH :
-			      tp->nonagle)));
-}
-
-
-/* Send _single_ skb sitting at the send head. This function requires
- * true push pending frames to setup probe timer etc.
- */
-void tcp_push_one(struct sock *sk, unsigned cur_mss)
-{
-	struct tcp_sock *tp = tcp_sk(sk);
-	struct sk_buff *skb = sk->sk_send_head;
-
-	if (tcp_snd_test(sk, skb, cur_mss, TCP_NAGLE_PUSH)) {
-		/* Send it out now. */
-		TCP_SKB_CB(skb)->when = tcp_time_stamp;
-		tcp_tso_set_push(skb);
-		if (!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation))) {
-			sk->sk_send_head = NULL;
-			tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
-			tcp_packets_out_inc(sk, tp, skb);
-			return;
-		}
+		skb_shinfo(skb)->tso_size = tp->mss_cache;
 	}
 }
 
@@ -791,7 +627,7 @@ unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
 
 	/* And store cached results */
 	tp->pmtu_cookie = pmtu;
-	tp->mss_cache = tp->mss_cache_std = mss_now;
+	tp->mss_cache = mss_now;
 
 	return mss_now;
 }
@@ -803,56 +639,47 @@ unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
  * cannot be large. However, taking into account rare use of URG, this
  * is not a big flaw.
  */
-
-unsigned int tcp_current_mss(struct sock *sk, int large)
+unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct dst_entry *dst = __sk_dst_get(sk);
-	unsigned int do_large, mss_now;
+	u32 mss_now;
+	u16 xmit_size_goal;
+	int doing_tso = 0;
+
+	mss_now = tp->mss_cache;
+
+	if (large_allowed &&
+	    (sk->sk_route_caps & NETIF_F_TSO) &&
+	    !tp->urg_mode)
+		doing_tso = 1;
 
-	mss_now = tp->mss_cache_std;
 	if (dst) {
 		u32 mtu = dst_mtu(dst);
 		if (mtu != tp->pmtu_cookie)
 			mss_now = tcp_sync_mss(sk, mtu);
 	}
 
-	do_large = (large &&
-		    (sk->sk_route_caps & NETIF_F_TSO) &&
-		    !tp->urg_mode);
+	if (tp->rx_opt.eff_sacks)
+		mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
+			    (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
 
-	if (do_large) {
-		unsigned int large_mss, factor, limit;
+	xmit_size_goal = mss_now;
 
-		large_mss = 65535 - tp->af_specific->net_header_len -
+	if (doing_tso) {
+		xmit_size_goal = 65535 -
+			tp->af_specific->net_header_len -
 			tp->ext_header_len - tp->tcp_header_len;
 
-		if (tp->max_window && large_mss > (tp->max_window>>1))
-			large_mss = max((tp->max_window>>1),
-					68U - tp->tcp_header_len);
-
-		factor = large_mss / mss_now;
+		if (tp->max_window &&
+		    (xmit_size_goal > (tp->max_window >> 1)))
+			xmit_size_goal = max((tp->max_window >> 1),
+					     68U - tp->tcp_header_len);
 
-		/* Always keep large mss multiple of real mss, but
-		 * do not exceed 1/tso_win_divisor of the congestion window
-		 * so we can keep the ACK clock ticking and minimize
-		 * bursting.
-		 */
-		limit = tp->snd_cwnd;
-		if (sysctl_tcp_tso_win_divisor)
-			limit /= sysctl_tcp_tso_win_divisor;
-		limit = max(1U, limit);
-		if (factor > limit)
-			factor = limit;
-
-		tp->mss_cache = mss_now * factor;
-
-		mss_now = tp->mss_cache;
+		xmit_size_goal -= (xmit_size_goal % mss_now);
 	}
+	tp->xmit_size_goal = xmit_size_goal;
 
-	if (tp->rx_opt.eff_sacks)
-		mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
-			    (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
 	return mss_now;
 }
 
@@ -876,6 +703,251 @@ static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
 	}
 }
 
+static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
+{
+	u32 window, cwnd_len;
+
+	window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
+	cwnd_len = mss_now * cwnd;
+	return min(window, cwnd_len);
+}
+
+/* Can at least one segment of SKB be sent right now, according to the
+ * congestion window rules?  If so, return how many segments are allowed.
+ */
+static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
+{
+	u32 in_flight, cwnd;
+
+	/* Don't be strict about the congestion window for the final FIN.  */
+	if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
+		return 1;
+
+	in_flight = tcp_packets_in_flight(tp);
+	cwnd = tp->snd_cwnd;
+	if (in_flight < cwnd)
+		return (cwnd - in_flight);
+
+	return 0;
+}
+
+/* This must be invoked the first time we consider transmitting
+ * SKB onto the wire.
+ */
+static inline int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb)
+{
+	int tso_segs = tcp_skb_pcount(skb);
+
+	if (!tso_segs) {
+		tcp_set_skb_tso_segs(sk, skb);
+		tso_segs = tcp_skb_pcount(skb);
+	}
+	return tso_segs;
+}
+
+static inline int tcp_minshall_check(const struct tcp_sock *tp)
+{
+	return after(tp->snd_sml,tp->snd_una) &&
+		!after(tp->snd_sml, tp->snd_nxt);
+}
+
+/* Return 0, if packet can be sent now without violation Nagle's rules:
+ * 1. It is full sized.
+ * 2. Or it contains FIN. (already checked by caller)
+ * 3. Or TCP_NODELAY was set.
+ * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
+ *    With Minshall's modification: all sent small packets are ACKed.
+ */
+
+static inline int tcp_nagle_check(const struct tcp_sock *tp,
+				  const struct sk_buff *skb, 
+				  unsigned mss_now, int nonagle)
+{
+	return (skb->len < mss_now &&
+		((nonagle&TCP_NAGLE_CORK) ||
+		 (!nonagle &&
+		  tp->packets_out &&
+		  tcp_minshall_check(tp))));
+}
+
+/* Return non-zero if the Nagle test allows this packet to be
+ * sent now.
+ */
+static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
+				 unsigned int cur_mss, int nonagle)
+{
+	/* Nagle rule does not apply to frames, which sit in the middle of the
+	 * write_queue (they have no chances to get new data).
+	 *
+	 * This is implemented in the callers, where they modify the 'nonagle'
+	 * argument based upon the location of SKB in the send queue.
+	 */
+	if (nonagle & TCP_NAGLE_PUSH)
+		return 1;
+
+	/* Don't use the nagle rule for urgent data (or for the final FIN).  */
+	if (tp->urg_mode ||
+	    (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
+		return 1;
+
+	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
+		return 1;
+
+	return 0;
+}
+
+/* Does at least the first segment of SKB fit into the send window? */
+static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
+{
+	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
+
+	if (skb->len > cur_mss)
+		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
+
+	return !after(end_seq, tp->snd_una + tp->snd_wnd);
+}
+
+/* This checks if the data bearing packet SKB (usually sk->sk_send_head)
+ * should be put on the wire right now.  If so, it returns the number of
+ * packets allowed by the congestion window.
+ */
+static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
+				 unsigned int cur_mss, int nonagle)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	unsigned int cwnd_quota;
+
+	tcp_init_tso_segs(sk, skb);
+
+	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
+		return 0;
+
+	cwnd_quota = tcp_cwnd_test(tp, skb);
+	if (cwnd_quota &&
+	    !tcp_snd_wnd_test(tp, skb, cur_mss))
+		cwnd_quota = 0;
+
+	return cwnd_quota;
+}
+
+static inline int tcp_skb_is_last(const struct sock *sk, 
+				  const struct sk_buff *skb)
+{
+	return skb->next == (struct sk_buff *)&sk->sk_write_queue;
+}
+
+int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
+{
+	struct sk_buff *skb = sk->sk_send_head;
+
+	return (skb &&
+		tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
+			     (tcp_skb_is_last(sk, skb) ?
+			      TCP_NAGLE_PUSH :
+			      tp->nonagle)));
+}
+
+/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
+ * which is put after SKB on the list.  It is very much like
+ * tcp_fragment() except that it may make several kinds of assumptions
+ * in order to speed up the splitting operation.  In particular, we
+ * know that all the data is in scatter-gather pages, and that the
+ * packet has never been sent out before (and thus is not cloned).
+ */
+static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len)
+{
+	struct sk_buff *buff;
+	int nlen = skb->len - len;
+	u16 flags;
+
+	/* All of a TSO frame must be composed of paged data.  */
+	BUG_ON(skb->len != skb->data_len);
+
+	buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
+	if (unlikely(buff == NULL))
+		return -ENOMEM;
+
+	buff->truesize = nlen;
+	skb->truesize -= nlen;
+
+	/* Correct the sequence numbers. */
+	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
+	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
+	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
+
+	/* PSH and FIN should only be set in the second packet. */
+	flags = TCP_SKB_CB(skb)->flags;
+	TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
+	TCP_SKB_CB(buff)->flags = flags;
+
+	/* This packet was never sent out yet, so no SACK bits. */
+	TCP_SKB_CB(buff)->sacked = 0;
+
+	buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
+	skb_split(skb, buff, len);
+
+	/* Fix up tso_factor for both original and new SKB.  */
+	tcp_set_skb_tso_segs(sk, skb);
+	tcp_set_skb_tso_segs(sk, buff);
+
+	/* Link BUFF into the send queue. */
+	skb_header_release(buff);
+	__skb_append(skb, buff);
+
+	return 0;
+}
+
+/* Try to defer sending, if possible, in order to minimize the amount
+ * of TSO splitting we do.  View it as a kind of TSO Nagle test.
+ *
+ * This algorithm is from John Heffner.
+ */
+static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
+{
+	u32 send_win, cong_win, limit, in_flight;
+
+	if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
+		return 0;
+
+	in_flight = tcp_packets_in_flight(tp);
+
+	BUG_ON(tcp_skb_pcount(skb) <= 1 ||
+	       (tp->snd_cwnd <= in_flight));
+
+	send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
+
+	/* From in_flight test above, we know that cwnd > in_flight.  */
+	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
+
+	limit = min(send_win, cong_win);
+
+	/* If sk_send_head can be sent fully now, just do it.  */
+	if (skb->len <= limit)
+		return 0;
+
+	if (sysctl_tcp_tso_win_divisor) {
+		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
+
+		/* If at least some fraction of a window is available,
+		 * just use it.
+		 */
+		chunk /= sysctl_tcp_tso_win_divisor;
+		if (limit >= chunk)
+			return 0;
+	} else {
+		/* Different approach, try not to defer past a single
+		 * ACK.  Receiver should ACK every other full sized
+		 * frame, so if we have space for more than 3 frames
+		 * then send now.
+		 */
+		if (limit > tcp_max_burst(tp) * tp->mss_cache)
+			return 0;
+	}
+
+	/* Ok, it looks like it is advisable to defer.  */
+	return 1;
+}
+
 /* This routine writes packets to the network.  It advances the
  * send_head.  This happens as incoming acks open up the remote
  * window for us.
@@ -887,8 +959,8 @@ static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct sk_buff *skb;
-	unsigned int tso_segs, cwnd_quota;
-	int sent_pkts;
+	unsigned int tso_segs, sent_pkts;
+	int cwnd_quota;
 
 	/* If we are closed, the bytes will have to remain here.
 	 * In time closedown will finish, we empty the write queue and all
@@ -903,24 +975,44 @@ static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
 
 	tso_segs = tcp_init_tso_segs(sk, skb);
 	cwnd_quota = tcp_cwnd_test(tp, skb);
+	if (unlikely(!cwnd_quota))
+		goto out;
+
 	sent_pkts = 0;
+	while (likely(tcp_snd_wnd_test(tp, skb, mss_now))) {
+		BUG_ON(!tso_segs);
 
-	while (cwnd_quota >= tso_segs) {
-		if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
-					     (tcp_skb_is_last(sk, skb) ?
-					      nonagle : TCP_NAGLE_PUSH))))
-			break;
+		if (tso_segs == 1) {
+			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
+						     (tcp_skb_is_last(sk, skb) ?
+						      nonagle : TCP_NAGLE_PUSH))))
+				break;
+		} else {
+			if (tcp_tso_should_defer(sk, tp, skb))
+				break;
+		}
 
-		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
-			break;
+		if (tso_segs > 1) {
+			u32 limit = tcp_window_allows(tp, skb,
+						      mss_now, cwnd_quota);
+
+			if (skb->len < limit) {
+				unsigned int trim = skb->len % mss_now;
 
-		if (unlikely(skb->len > mss_now)) {
+				if (trim)
+					limit = skb->len - trim;
+			}
+			if (skb->len > limit) {
+				if (tso_fragment(sk, skb, limit))
+					break;
+			}
+		} else if (unlikely(skb->len > mss_now)) {
 			if (unlikely(tcp_fragment(sk, skb,  mss_now)))
 				break;
 		}
 
 		TCP_SKB_CB(skb)->when = tcp_time_stamp;
-		tcp_tso_set_push(skb);
+
 		if (unlikely(tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))))
 			break;
 
@@ -936,6 +1028,11 @@ static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
 		 * the packet above, tso_segs will no longer be valid.
 		 */
 		cwnd_quota -= tcp_skb_pcount(skb);
+
+		BUG_ON(cwnd_quota < 0);
+		if (!cwnd_quota)
+			break;
+
 		skb = sk->sk_send_head;
 		if (!skb)
 			break;
@@ -946,7 +1043,7 @@ static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
 		tcp_cwnd_validate(sk, tp);
 		return 0;
 	}
-
+out:
 	return !tp->packets_out && sk->sk_send_head;
 }
 
@@ -965,6 +1062,53 @@ void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
 	}
 }
 
+/* Send _single_ skb sitting at the send head. This function requires
+ * true push pending frames to setup probe timer etc.
+ */
+void tcp_push_one(struct sock *sk, unsigned int mss_now)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct sk_buff *skb = sk->sk_send_head;
+	unsigned int tso_segs, cwnd_quota;
+
+	BUG_ON(!skb || skb->len < mss_now);
+
+	tso_segs = tcp_init_tso_segs(sk, skb);
+	cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
+
+	if (likely(cwnd_quota)) {
+		BUG_ON(!tso_segs);
+
+		if (tso_segs > 1) {
+			u32 limit = tcp_window_allows(tp, skb,
+						      mss_now, cwnd_quota);
+
+			if (skb->len < limit) {
+				unsigned int trim = skb->len % mss_now;
+
+				if (trim)
+					limit = skb->len - trim;
+			}
+			if (skb->len > limit) {
+				if (unlikely(tso_fragment(sk, skb, limit)))
+					return;
+			}
+		} else if (unlikely(skb->len > mss_now)) {
+			if (unlikely(tcp_fragment(sk, skb, mss_now)))
+				return;
+		}
+
+		/* Send it out now. */
+		TCP_SKB_CB(skb)->when = tcp_time_stamp;
+
+		if (likely(!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation)))) {
+			update_send_head(sk, tp, skb);
+			tcp_cwnd_validate(sk, tp);
+			return;
+		}
+	}
+}
+
 /* This function returns the amount that we can raise the
  * usable window based on the following constraints
  *  
@@ -1222,7 +1366,6 @@ int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
 		if (sk->sk_route_caps & NETIF_F_TSO) {
 			sk->sk_route_caps &= ~NETIF_F_TSO;
 			sock_set_flag(sk, SOCK_NO_LARGESEND);
-			tp->mss_cache = tp->mss_cache_std;
 		}
 
 		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
@@ -1284,7 +1427,6 @@ int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
 	 * is still in somebody's hands, else make a clone.
 	 */
 	TCP_SKB_CB(skb)->when = tcp_time_stamp;
-	tcp_tso_set_push(skb);
 
 	err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
 				    pskb_copy(skb, GFP_ATOMIC):
@@ -1853,14 +1995,12 @@ int tcp_write_wakeup(struct sock *sk)
 				if (sk->sk_route_caps & NETIF_F_TSO) {
 					sock_set_flag(sk, SOCK_NO_LARGESEND);
 					sk->sk_route_caps &= ~NETIF_F_TSO;
-					tp->mss_cache = tp->mss_cache_std;
 				}
 			} else if (!tcp_skb_pcount(skb))
 				tcp_set_skb_tso_segs(sk, skb);
 
 			TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
 			TCP_SKB_CB(skb)->when = tcp_time_stamp;
-			tcp_tso_set_push(skb);
 			err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
 			if (!err) {
 				update_send_head(sk, tp, skb);