Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next

Pull networking updates from David Miller:
 "Highlights:

   1) Support TX_RING in AF_PACKET TPACKET_V3 mode, from Sowmini
      Varadhan.

   2) Simplify classifier state on sk_buff in order to shrink it a bit.
      From Willem de Bruijn.

   3) Introduce SIPHASH and it's usage for secure sequence numbers and
      syncookies. From Jason A. Donenfeld.

   4) Reduce CPU usage for ICMP replies we are going to limit or
      suppress, from Jesper Dangaard Brouer.

   5) Introduce Shared Memory Communications socket layer, from Ursula
      Braun.

   6) Add RACK loss detection and allow it to actually trigger fast
      recovery instead of just assisting after other algorithms have
      triggered it. From Yuchung Cheng.

   7) Add xmit_more and BQL support to mvneta driver, from Simon Guinot.

   8) skb_cow_data avoidance in esp4 and esp6, from Steffen Klassert.

   9) Export MPLS packet stats via netlink, from Robert Shearman.

  10) Significantly improve inet port bind conflict handling, especially
      when an application is restarted and changes it's setting of
      reuseport. From Josef Bacik.

  11) Implement TX batching in vhost_net, from Jason Wang.

  12) Extend the dummy device so that VF (virtual function) features,
      such as configuration, can be more easily tested. From Phil
      Sutter.

  13) Avoid two atomic ops per page on x86 in bnx2x driver, from Eric
      Dumazet.

  14) Add new bpf MAP, implementing a longest prefix match trie. From
      Daniel Mack.

  15) Packet sample offloading support in mlxsw driver, from Yotam Gigi.

  16) Add new aquantia driver, from David VomLehn.

  17) Add bpf tracepoints, from Daniel Borkmann.

  18) Add support for port mirroring to b53 and bcm_sf2 drivers, from
      Florian Fainelli.

  19) Remove custom busy polling in many drivers, it is done in the core
      networking since 4.5 times. From Eric Dumazet.

  20) Support XDP adjust_head in virtio_net, from John Fastabend.

  21) Fix several major holes in neighbour entry confirmation, from
      Julian Anastasov.

  22) Add XDP support to bnxt_en driver, from Michael Chan.

  23) VXLAN offloads for enic driver, from Govindarajulu Varadarajan.

  24) Add IPVTAP driver (IP-VLAN based tap driver) from Sainath Grandhi.

  25) Support GRO in IPSEC protocols, from Steffen Klassert"

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1764 commits)
  Revert "ath10k: Search SMBIOS for OEM board file extension"
  net: socket: fix recvmmsg not returning error from sock_error
  bnxt_en: use eth_hw_addr_random()
  bpf: fix unlocking of jited image when module ronx not set
  arch: add ARCH_HAS_SET_MEMORY config
  net: napi_watchdog() can use napi_schedule_irqoff()
  tcp: Revert "tcp: tcp_probe: use spin_lock_bh()"
  net/hsr: use eth_hw_addr_random()
  net: mvpp2: enable building on 64-bit platforms
  net: mvpp2: switch to build_skb() in the RX path
  net: mvpp2: simplify MVPP2_PRS_RI_* definitions
  net: mvpp2: fix indentation of MVPP2_EXT_GLOBAL_CTRL_DEFAULT
  net: mvpp2: remove unused register definitions
  net: mvpp2: simplify mvpp2_bm_bufs_add()
  net: mvpp2: drop useless fields in mvpp2_bm_pool and related code
  net: mvpp2: remove unused 'tx_skb' field of 'struct mvpp2_tx_queue'
  net: mvpp2: release reference to txq_cpu[] entry after unmapping
  net: mvpp2: handle too large value in mvpp2_rx_time_coal_set()
  net: mvpp2: handle too large value handling in mvpp2_rx_pkts_coal_set()
  net: mvpp2: remove useless arguments in mvpp2_rx_{pkts, time}_coal_set
  ...

1 files changed, 111 insertions, 37 deletions

diff --git a/net/ipv4/tcp_recovery.c b/net/ipv4/tcp_recovery.c
index e36df4fcfeba..4ecb38ae8504 100644
--- a/net/ipv4/tcp_recovery.c
+++ b/net/ipv4/tcp_recovery.c
@@ -1,9 +1,32 @@
 #include <linux/tcp.h>
 #include <net/tcp.h>
 
-int sysctl_tcp_recovery __read_mostly = TCP_RACK_LOST_RETRANS;
+int sysctl_tcp_recovery __read_mostly = TCP_RACK_LOSS_DETECTION;
 
-/* Marks a packet lost, if some packet sent later has been (s)acked.
+static void tcp_rack_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+
+	tcp_skb_mark_lost_uncond_verify(tp, skb);
+	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
+		/* Account for retransmits that are lost again */
+		TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
+		tp->retrans_out -= tcp_skb_pcount(skb);
+		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
+	}
+}
+
+static bool tcp_rack_sent_after(const struct skb_mstamp *t1,
+				const struct skb_mstamp *t2,
+				u32 seq1, u32 seq2)
+{
+	return skb_mstamp_after(t1, t2) ||
+	       (t1->v64 == t2->v64 && after(seq1, seq2));
+}
+
+/* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01):
+ *
+ * Marks a packet lost, if some packet sent later has been (s)acked.
  * The underlying idea is similar to the traditional dupthresh and FACK
  * but they look at different metrics:
  *
@@ -16,31 +39,26 @@ int sysctl_tcp_recovery __read_mostly = TCP_RACK_LOST_RETRANS;
  * is being more resilient to reordering by simply allowing some
  * "settling delay", instead of tweaking the dupthresh.
  *
- * The current version is only used after recovery starts but can be
- * easily extended to detect the first loss.
+ * When tcp_rack_detect_loss() detects some packets are lost and we
+ * are not already in the CA_Recovery state, either tcp_rack_reo_timeout()
+ * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will
+ * make us enter the CA_Recovery state.
  */
-int tcp_rack_mark_lost(struct sock *sk)
+static void tcp_rack_detect_loss(struct sock *sk, const struct skb_mstamp *now,
+				 u32 *reo_timeout)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct sk_buff *skb;
-	u32 reo_wnd, prior_retrans = tp->retrans_out;
-
-	if (inet_csk(sk)->icsk_ca_state < TCP_CA_Recovery || !tp->rack.advanced)
-		return 0;
-
-	/* Reset the advanced flag to avoid unnecessary queue scanning */
-	tp->rack.advanced = 0;
+	u32 reo_wnd;
 
+	*reo_timeout = 0;
 	/* To be more reordering resilient, allow min_rtt/4 settling delay
 	 * (lower-bounded to 1000uS). We use min_rtt instead of the smoothed
 	 * RTT because reordering is often a path property and less related
 	 * to queuing or delayed ACKs.
-	 *
-	 * TODO: measure and adapt to the observed reordering delay, and
-	 * use a timer to retransmit like the delayed early retransmit.
 	 */
 	reo_wnd = 1000;
-	if (tp->rack.reord && tcp_min_rtt(tp) != ~0U)
+	if ((tp->rack.reord || !tp->lost_out) && tcp_min_rtt(tp) != ~0U)
 		reo_wnd = max(tcp_min_rtt(tp) >> 2, reo_wnd);
 
 	tcp_for_write_queue(skb, sk) {
@@ -54,20 +72,29 @@ int tcp_rack_mark_lost(struct sock *sk)
 		    scb->sacked & TCPCB_SACKED_ACKED)
 			continue;
 
-		if (skb_mstamp_after(&tp->rack.mstamp, &skb->skb_mstamp)) {
+		if (tcp_rack_sent_after(&tp->rack.mstamp, &skb->skb_mstamp,
+					tp->rack.end_seq, scb->end_seq)) {
+			/* Step 3 in draft-cheng-tcpm-rack-00.txt:
+			 * A packet is lost if its elapsed time is beyond
+			 * the recent RTT plus the reordering window.
+			 */
+			u32 elapsed = skb_mstamp_us_delta(now,
+							  &skb->skb_mstamp);
+			s32 remaining = tp->rack.rtt_us + reo_wnd - elapsed;
 
-			if (skb_mstamp_us_delta(&tp->rack.mstamp,
-						&skb->skb_mstamp) <= reo_wnd)
+			if (remaining < 0) {
+				tcp_rack_mark_skb_lost(sk, skb);
 				continue;
-
-			/* skb is lost if packet sent later is sacked */
-			tcp_skb_mark_lost_uncond_verify(tp, skb);
-			if (scb->sacked & TCPCB_SACKED_RETRANS) {
-				scb->sacked &= ~TCPCB_SACKED_RETRANS;
-				tp->retrans_out -= tcp_skb_pcount(skb);
-				NET_INC_STATS(sock_net(sk),
-					      LINUX_MIB_TCPLOSTRETRANSMIT);
 			}
+
+			/* Skip ones marked lost but not yet retransmitted */
+			if ((scb->sacked & TCPCB_LOST) &&
+			    !(scb->sacked & TCPCB_SACKED_RETRANS))
+				continue;
+
+			/* Record maximum wait time (+1 to avoid 0) */
+			*reo_timeout = max_t(u32, *reo_timeout, 1 + remaining);
+
 		} else if (!(scb->sacked & TCPCB_RETRANS)) {
 			/* Original data are sent sequentially so stop early
 			 * b/c the rest are all sent after rack_sent
@@ -75,20 +102,43 @@ int tcp_rack_mark_lost(struct sock *sk)
 			break;
 		}
 	}
-	return prior_retrans - tp->retrans_out;
 }
 
-/* Record the most recently (re)sent time among the (s)acked packets */
-void tcp_rack_advance(struct tcp_sock *tp,
-		      const struct skb_mstamp *xmit_time, u8 sacked)
+void tcp_rack_mark_lost(struct sock *sk, const struct skb_mstamp *now)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	u32 timeout;
+
+	if (!tp->rack.advanced)
+		return;
+
+	/* Reset the advanced flag to avoid unnecessary queue scanning */
+	tp->rack.advanced = 0;
+	tcp_rack_detect_loss(sk, now, &timeout);
+	if (timeout) {
+		timeout = usecs_to_jiffies(timeout + TCP_REO_TIMEOUT_MIN);
+		inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT,
+					  timeout, inet_csk(sk)->icsk_rto);
+	}
+}
+
+/* Record the most recently (re)sent time among the (s)acked packets
+ * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
+ * draft-cheng-tcpm-rack-00.txt
+ */
+void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
+		      const struct skb_mstamp *xmit_time,
+		      const struct skb_mstamp *ack_time)
 {
+	u32 rtt_us;
+
 	if (tp->rack.mstamp.v64 &&
-	    !skb_mstamp_after(xmit_time, &tp->rack.mstamp))
+	    !tcp_rack_sent_after(xmit_time, &tp->rack.mstamp,
+				 end_seq, tp->rack.end_seq))
 		return;
 
+	rtt_us = skb_mstamp_us_delta(ack_time, xmit_time);
 	if (sacked & TCPCB_RETRANS) {
-		struct skb_mstamp now;
-
 		/* If the sacked packet was retransmitted, it's ambiguous
 		 * whether the retransmission or the original (or the prior
 		 * retransmission) was sacked.
@@ -99,11 +149,35 @@ void tcp_rack_advance(struct tcp_sock *tp,
 		 * so it's at least one RTT (i.e., retransmission is at least
 		 * an RTT later).
 		 */
-		skb_mstamp_get(&now);
-		if (skb_mstamp_us_delta(&now, xmit_time) < tcp_min_rtt(tp))
+		if (rtt_us < tcp_min_rtt(tp))
 			return;
 	}
-
+	tp->rack.rtt_us = rtt_us;
 	tp->rack.mstamp = *xmit_time;
+	tp->rack.end_seq = end_seq;
 	tp->rack.advanced = 1;
 }
+
+/* We have waited long enough to accommodate reordering. Mark the expired
+ * packets lost and retransmit them.
+ */
+void tcp_rack_reo_timeout(struct sock *sk)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct skb_mstamp now;
+	u32 timeout, prior_inflight;
+
+	skb_mstamp_get(&now);
+	prior_inflight = tcp_packets_in_flight(tp);
+	tcp_rack_detect_loss(sk, &now, &timeout);
+	if (prior_inflight != tcp_packets_in_flight(tp)) {
+		if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) {
+			tcp_enter_recovery(sk, false);
+			if (!inet_csk(sk)->icsk_ca_ops->cong_control)
+				tcp_cwnd_reduction(sk, 1, 0);
+		}
+		tcp_xmit_retransmit_queue(sk);
+	}
+	if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS)
+		tcp_rearm_rto(sk);
+}