portmap plugin should flush previous udp connections

conntrack does not have any way to track UDP connections, so
it relies on timers to delete a connection.
The problem is that UDP is connectionless, so a client will keep
sending traffic despite the server has gone, thus renewing the
conntrack entries.
Pods that use portmaps to expose UDP services need to flush the existing
conntrack entries on the port exposed when they are created,
otherwise conntrack will keep sending the traffic to the previous IP
until the connection age (the client stops sending traffic)

Signed-off-by: Antonio Ojea <aojea@redhat.com>
This commit is contained in:
Antonio Ojea
2020-11-17 23:34:59 +01:00
parent c41c78b600
commit 108c2aebd4
6 changed files with 401 additions and 25 deletions

View File

@ -28,12 +28,14 @@ package main
import (
"encoding/json"
"fmt"
"log"
"net"
"github.com/containernetworking/cni/pkg/skel"
"github.com/containernetworking/cni/pkg/types"
"github.com/containernetworking/cni/pkg/types/current"
"github.com/containernetworking/cni/pkg/version"
"golang.org/x/sys/unix"
bv "github.com/containernetworking/plugins/pkg/utils/buildversion"
)
@ -89,12 +91,24 @@ func cmdAdd(args *skel.CmdArgs) error {
if err := forwardPorts(netConf, netConf.ContIPv4); err != nil {
return err
}
// Delete conntrack entries for UDP to avoid conntrack blackholing traffic
// due to stale connections. We do that after the iptables rules are set, so
// the new traffic uses them. Failures are informative only.
if err := deletePortmapStaleConnections(netConf.RuntimeConfig.PortMaps, unix.AF_INET); err != nil {
log.Printf("failed to delete stale UDP conntrack entries for %s: %v", netConf.ContIPv4.IP, err)
}
}
if netConf.ContIPv6.IP != nil {
if err := forwardPorts(netConf, netConf.ContIPv6); err != nil {
return err
}
// Delete conntrack entries for UDP to avoid conntrack blackholing traffic
// due to stale connections. We do that after the iptables rules are set, so
// the new traffic uses them. Failures are informative only.
if err := deletePortmapStaleConnections(netConf.RuntimeConfig.PortMaps, unix.AF_INET6); err != nil {
log.Printf("failed to delete stale UDP conntrack entries for %s: %v", netConf.ContIPv6.IP, err)
}
}
// Pass through the previous result

View File

@ -19,10 +19,12 @@ import (
"net"
"sort"
"strconv"
"strings"
"github.com/containernetworking/plugins/pkg/utils"
"github.com/containernetworking/plugins/pkg/utils/sysctl"
"github.com/coreos/go-iptables/iptables"
"github.com/vishvananda/netlink"
)
// This creates the chains to be added to iptables. The basic structure is
@ -42,10 +44,12 @@ import (
// The names of the top-level summary chains.
// These should never be changed, or else upgrading will require manual
// intervention.
const TopLevelDNATChainName = "CNI-HOSTPORT-DNAT"
const SetMarkChainName = "CNI-HOSTPORT-SETMARK"
const MarkMasqChainName = "CNI-HOSTPORT-MASQ"
const OldTopLevelSNATChainName = "CNI-HOSTPORT-SNAT"
const (
TopLevelDNATChainName = "CNI-HOSTPORT-DNAT"
SetMarkChainName = "CNI-HOSTPORT-SETMARK"
MarkMasqChainName = "CNI-HOSTPORT-MASQ"
OldTopLevelSNATChainName = "CNI-HOSTPORT-SNAT"
)
// forwardPorts establishes port forwarding to a given container IP.
// containerNet.IP can be either v4 or v6.
@ -113,7 +117,6 @@ func forwardPorts(config *PortMapConf, containerNet net.IPNet) error {
}
func checkPorts(config *PortMapConf, containerNet net.IPNet) error {
dnatChain := genDnatChain(config.Name, config.ContainerID)
fillDnatRules(&dnatChain, config, containerNet)
@ -189,7 +192,7 @@ func fillDnatRules(c *chain, config *PortMapConf, containerNet net.IPNet) {
setMarkChainName = *config.ExternalSetMarkChain
}
//Generate the dnat entry rules. We'll use multiport, but it ony accepts
// Generate the dnat entry rules. We'll use multiport, but it ony accepts
// up to 15 rules, so partition the list if needed.
// Do it in a stable order for testing
protoPorts := groupByProto(entries)
@ -243,7 +246,8 @@ func fillDnatRules(c *chain, config *PortMapConf, containerNet net.IPNet) {
ruleBase := []string{
"-p", entry.Protocol,
"--dport", strconv.Itoa(entry.HostPort)}
"--dport", strconv.Itoa(entry.HostPort),
}
if addRuleBaseDst {
ruleBase = append(ruleBase,
"-d", entry.HostIP)
@ -406,3 +410,19 @@ func maybeGetIptables(isV6 bool) *iptables.IPTables {
return ipt
}
// deletePortmapStaleConnections delete the UDP conntrack entries on the specified IP family
// from the ports mapped to the container
func deletePortmapStaleConnections(portMappings []PortMapEntry, family netlink.InetFamily) error {
for _, pm := range portMappings {
// skip if is not UDP
if strings.ToLower(pm.Protocol) != "udp" {
continue
}
err := utils.DeleteConntrackEntriesForDstPort(uint16(pm.HostPort), utils.PROTOCOL_UDP, family)
if err != nil {
return err
}
}
return nil
}

View File

@ -184,16 +184,16 @@ var _ = Describe("portmap integration tests", func() {
Expect(cmd.Run()).To(Succeed())
// Sanity check: verify that the container is reachable directly
contOK := testEchoServer(contIP.String(), containerPort, "")
contOK := testEchoServer(contIP.String(), "tcp", containerPort, "")
// Verify that a connection to the forwarded port works
dnatOK := testEchoServer(hostIP, hostPort, "")
dnatOK := testEchoServer(hostIP, "tcp", hostPort, "")
// Verify that a connection to localhost works
snatOK := testEchoServer("127.0.0.1", hostPort, "")
snatOK := testEchoServer("127.0.0.1", "tcp", hostPort, "")
// verify that hairpin works
hairpinOK := testEchoServer(hostIP, hostPort, targetNS.Path())
hairpinOK := testEchoServer(hostIP, "tcp", hostPort, targetNS.Path())
// Cleanup
session.Terminate()
@ -219,21 +219,216 @@ var _ = Describe("portmap integration tests", func() {
}
close(done)
}, TIMEOUT*9)
It("forwards a UDP port on ipv4 and keep working after creating a second container with the same HostPort", func(done Done) {
var err error
hostPort := rand.Intn(10000) + 1025
runtimeConfig := libcni.RuntimeConf{
ContainerID: fmt.Sprintf("unit-test-%d", hostPort),
NetNS: targetNS.Path(),
IfName: "eth0",
CapabilityArgs: map[string]interface{}{
"portMappings": []map[string]interface{}{
{
"hostPort": hostPort,
"containerPort": containerPort,
"protocol": "udp",
},
},
},
}
// Make delete idempotent, so we can clean up on failure
netDeleted := false
deleteNetwork := func() error {
if netDeleted {
return nil
}
netDeleted = true
return cniConf.DelNetworkList(context.TODO(), configList, &runtimeConfig)
}
// Create the network
resI, err := cniConf.AddNetworkList(context.TODO(), configList, &runtimeConfig)
Expect(err).NotTo(HaveOccurred())
defer deleteNetwork()
// Undo Docker's forwarding policy
cmd := exec.Command("iptables", "-t", "filter",
"-P", "FORWARD", "ACCEPT")
cmd.Stderr = GinkgoWriter
err = cmd.Run()
Expect(err).NotTo(HaveOccurred())
result, err := current.GetResult(resI)
Expect(err).NotTo(HaveOccurred())
var contIP net.IP
for _, ip := range result.IPs {
intfIndex := *ip.Interface
if result.Interfaces[intfIndex].Sandbox == "" {
continue
}
contIP = ip.Address.IP
}
if contIP == nil {
Fail("could not determine container IP")
}
hostIP := getLocalIP()
fmt.Fprintf(GinkgoWriter, "First container hostIP: %s:%d, contIP: %s:%d\n",
hostIP, hostPort, contIP, containerPort)
// dump iptables-save output for debugging
cmd = exec.Command("iptables-save")
cmd.Stderr = GinkgoWriter
cmd.Stdout = GinkgoWriter
Expect(cmd.Run()).To(Succeed())
// dump ip routes output for debugging
cmd = exec.Command("ip", "route")
cmd.Stderr = GinkgoWriter
cmd.Stdout = GinkgoWriter
Expect(cmd.Run()).To(Succeed())
// dump ip addresses output for debugging
cmd = exec.Command("ip", "addr")
cmd.Stderr = GinkgoWriter
cmd.Stdout = GinkgoWriter
Expect(cmd.Run()).To(Succeed())
// Sanity check: verify that the container is reachable directly
fmt.Fprintln(GinkgoWriter, "Connect to container:", contIP.String(), containerPort)
contOK := testEchoServer(contIP.String(), "udp", containerPort, "")
// Verify that a connection to the forwarded port works
fmt.Fprintln(GinkgoWriter, "Connect to host:", hostIP, hostPort)
dnatOK := testEchoServer(hostIP, "udp", hostPort, "")
// Cleanup
session.Terminate()
err = deleteNetwork()
Expect(err).NotTo(HaveOccurred())
// Check that everything succeeded *after* we clean up the network
if !contOK {
Fail("connection direct to " + contIP.String() + " failed")
}
if !dnatOK {
Fail("Connection to " + hostIP + " was not forwarded")
}
// Create a second container
targetNS2, err := testutils.NewNS()
Expect(err).NotTo(HaveOccurred())
fmt.Fprintln(GinkgoWriter, "namespace:", targetNS2.Path())
// Start an echo server and get the port
containerPort, session2, err := StartEchoServerInNamespace(targetNS2)
Expect(err).NotTo(HaveOccurred())
runtimeConfig2 := libcni.RuntimeConf{
ContainerID: fmt.Sprintf("unit-test2-%d", hostPort),
NetNS: targetNS2.Path(),
IfName: "eth0",
CapabilityArgs: map[string]interface{}{
"portMappings": []map[string]interface{}{
{
"hostPort": hostPort,
"containerPort": containerPort,
"protocol": "udp",
},
},
},
}
// Make delete idempotent, so we can clean up on failure
net2Deleted := false
deleteNetwork2 := func() error {
if net2Deleted {
return nil
}
net2Deleted = true
return cniConf.DelNetworkList(context.TODO(), configList, &runtimeConfig2)
}
// Create the network
resI2, err := cniConf.AddNetworkList(context.TODO(), configList, &runtimeConfig2)
Expect(err).NotTo(HaveOccurred())
defer deleteNetwork2()
result2, err := current.GetResult(resI2)
Expect(err).NotTo(HaveOccurred())
var contIP2 net.IP
for _, ip := range result2.IPs {
intfIndex := *ip.Interface
if result2.Interfaces[intfIndex].Sandbox == "" {
continue
}
contIP2 = ip.Address.IP
}
if contIP2 == nil {
Fail("could not determine container IP")
}
fmt.Fprintf(GinkgoWriter, "Second container: hostIP: %s:%d, contIP: %s:%d\n",
hostIP, hostPort, contIP2, containerPort)
// dump iptables-save output for debugging
cmd = exec.Command("iptables-save")
cmd.Stderr = GinkgoWriter
cmd.Stdout = GinkgoWriter
Expect(cmd.Run()).To(Succeed())
// dump ip routes output for debugging
cmd = exec.Command("ip", "route")
cmd.Stderr = GinkgoWriter
cmd.Stdout = GinkgoWriter
Expect(cmd.Run()).To(Succeed())
// dump ip addresses output for debugging
cmd = exec.Command("ip", "addr")
cmd.Stderr = GinkgoWriter
cmd.Stdout = GinkgoWriter
Expect(cmd.Run()).To(Succeed())
// Sanity check: verify that the container is reachable directly
fmt.Fprintln(GinkgoWriter, "Connect to container:", contIP2.String(), containerPort)
cont2OK := testEchoServer(contIP2.String(), "udp", containerPort, "")
// Verify that a connection to the forwarded port works
fmt.Fprintln(GinkgoWriter, "Connect to host:", hostIP, hostPort)
dnat2OK := testEchoServer(hostIP, "udp", hostPort, "")
// Cleanup
session2.Terminate()
err = deleteNetwork2()
Expect(err).NotTo(HaveOccurred())
// Check that everything succeeded *after* we clean up the network
if !cont2OK {
Fail("connection direct to " + contIP2.String() + " failed")
}
if !dnat2OK {
Fail("Connection to " + hostIP + " was not forwarded")
}
close(done)
}, TIMEOUT*9)
})
})
// testEchoServer returns true if we found an echo server on the port
func testEchoServer(address string, port int, netns string) bool {
func testEchoServer(address, protocol string, port int, netns string) bool {
message := "'Aliquid melius quam pessimum optimum non est.'"
var cmd *exec.Cmd
if netns != "" {
netns = filepath.Base(netns)
cmd = exec.Command("ip", "netns", "exec", netns, echoClientBinaryPath, "--target", fmt.Sprintf("%s:%d", address, port), "--message", message)
cmd = exec.Command("ip", "netns", "exec", netns, echoClientBinaryPath, "--target", fmt.Sprintf("%s:%d", address, port), "--message", message, "--protocol", protocol)
} else {
cmd = exec.Command(echoClientBinaryPath, "--target", fmt.Sprintf("%s:%d", address, port), "--message", message)
cmd = exec.Command(echoClientBinaryPath, "--target", fmt.Sprintf("%s:%d", address, port), "--message", message, "--protocol", protocol)
}
cmd.Stdin = bytes.NewBufferString(message)
cmd.Stderr = GinkgoWriter