From a50e885121d18a64638f7714148d96a5381b0d58 Mon Sep 17 00:00:00 2001
From: Santhosh Manohar <santhosh@docker.com>
Date: Thu, 13 Jul 2017 16:50:56 -0700
Subject: [PATCH] When the gc_thresh3 value is reached kenel might remove
 existing neighbor entries. On an l3 miss try to reprogram the neighbor entry
 if the peer is valid. Its a best effort attempt because if the arp table is
 still at gc_thresh3 value, addition will fail.

Signed-off-by: Santhosh Manohar <santhosh@docker.com>
---
 libnetwork/drivers/overlay/ov_network.go | 65 ++++++++++++++++++++----
 1 file changed, 54 insertions(+), 11 deletions(-)

diff --git a/libnetwork/drivers/overlay/ov_network.go b/libnetwork/drivers/overlay/ov_network.go
index 3a4ea41bfc..01f6287bed 100644
--- a/libnetwork/drivers/overlay/ov_network.go
+++ b/libnetwork/drivers/overlay/ov_network.go
@@ -12,6 +12,7 @@ import (
 	"strings"
 	"sync"
 	"syscall"
+	"time"
 
 	"github.com/Sirupsen/logrus"
 	"github.com/docker/docker/pkg/reexec"
@@ -705,6 +706,7 @@ func (n *network) initSandbox(restore bool) error {
 }
 
 func (n *network) watchMiss(nlSock *nl.NetlinkSocket) {
+	t := time.Now()
 	for {
 		msgs, err := nlSock.Receive()
 		if err != nil {
@@ -757,23 +759,55 @@ func (n *network) watchMiss(nlSock *nl.NetlinkSocket) {
 				continue
 			}
 
-			if !n.driver.isSerfAlive() {
-				continue
-			}
+			if n.driver.isSerfAlive() {
+				mac, IPmask, vtep, err := n.driver.resolvePeer(n.id, ip)
+				if err != nil {
+					logrus.Errorf("could not resolve peer %q: %v", ip, err)
+					continue
+				}
 
-			mac, IPmask, vtep, err := n.driver.resolvePeer(n.id, ip)
-			if err != nil {
-				logrus.Errorf("could not resolve peer %q: %v", ip, err)
-				continue
-			}
-
-			if err := n.driver.peerAdd(n.id, "dummy", ip, IPmask, mac, vtep, true, l2Miss, l3Miss); err != nil {
-				logrus.Errorf("could not add neighbor entry for missed peer %q: %v", ip, err)
+				if err := n.driver.peerAdd(n.id, "dummy", ip, IPmask, mac, vtep, true, l2Miss, l3Miss); err != nil {
+					logrus.Errorf("could not add neighbor entry for missed peer %q: %v", ip, err)
+				}
+			} else {
+				// If the gc_thresh values are lower kernel might knock off the neighor entries.
+				// When we get a L3 miss check if its a valid peer and reprogram the neighbor
+				// entry again. Rate limit it to once attempt every 500ms, just in case a faulty
+				// container sends a flood of packets to invalid peers
+				if !l3Miss {
+					continue
+				}
+				if time.Since(t) > 500*time.Millisecond {
+					t = time.Now()
+					n.programNeighbor(ip)
+				}
 			}
 		}
 	}
 }
 
+func (n *network) programNeighbor(ip net.IP) {
+	peerMac, _, _, err := n.driver.peerDbSearch(n.id, ip)
+	if err != nil {
+		logrus.Errorf("Reprogramming on L3 miss failed for %s, no peer entry", ip)
+		return
+	}
+	s := n.getSubnetforIPAddr(ip)
+	if s == nil {
+		logrus.Errorf("Reprogramming on L3 miss failed for %s, not a valid subnet", ip)
+		return
+	}
+	sbox := n.sandbox()
+	if sbox == nil {
+		logrus.Errorf("Reprogramming on L3 miss failed for %s, overlay sandbox missing", ip)
+		return
+	}
+	if err := sbox.AddNeighbor(ip, peerMac, true, sbox.NeighborOptions().LinkName(s.vxlanName)); err != nil {
+		logrus.Errorf("Reprogramming on L3 miss failed for %s: %v", ip, err)
+		return
+	}
+}
+
 func (d *driver) addNetwork(n *network) {
 	d.Lock()
 	d.networks[n.id] = n
@@ -1052,6 +1086,15 @@ func (n *network) contains(ip net.IP) bool {
 	return false
 }
 
+func (n *network) getSubnetforIPAddr(ip net.IP) *subnet {
+	for _, s := range n.subnets {
+		if s.subnetIP.Contains(ip) {
+			return s
+		}
+	}
+	return nil
+}
+
 // getSubnetforIP returns the subnet to which the given IP belongs
 func (n *network) getSubnetforIP(ip *net.IPNet) *subnet {
 	for _, s := range n.subnets {