moby--moby/network.go

727 lines
19 KiB
Go
Raw Normal View History

package docker
import (
"encoding/binary"
2013-02-25 18:45:23 +00:00
"errors"
"fmt"
"github.com/dotcloud/docker/iptables"
"github.com/dotcloud/docker/netlink"
2013-10-18 21:15:24 +00:00
"github.com/dotcloud/docker/proxy"
2013-05-14 22:37:35 +00:00
"github.com/dotcloud/docker/utils"
"log"
"net"
"strconv"
"sync"
)
const (
DefaultNetworkBridge = "docker0"
DisableNetworkBridge = "none"
portRangeStart = 49153
portRangeEnd = 65535
)
// Calculates the first and last IP addresses in an IPNet
func networkRange(network *net.IPNet) (net.IP, net.IP) {
netIP := network.IP.To4()
firstIP := netIP.Mask(network.Mask)
lastIP := net.IPv4(0, 0, 0, 0).To4()
for i := 0; i < len(lastIP); i++ {
lastIP[i] = netIP[i] | ^network.Mask[i]
}
return firstIP, lastIP
}
// Detects overlap between one IPNet and another
func networkOverlaps(netX *net.IPNet, netY *net.IPNet) bool {
firstIP, _ := networkRange(netX)
if netY.Contains(firstIP) {
return true
}
firstIP, _ = networkRange(netY)
if netX.Contains(firstIP) {
return true
}
return false
}
// Converts a 4 bytes IP into a 32 bit integer
func ipToInt(ip net.IP) int32 {
return int32(binary.BigEndian.Uint32(ip.To4()))
}
// Converts 32 bit integer into a 4 bytes IP address
2013-06-04 18:00:22 +00:00
func intToIP(n int32) net.IP {
b := make([]byte, 4)
binary.BigEndian.PutUint32(b, uint32(n))
return net.IP(b)
}
// Given a netmask, calculates the number of available hosts
func networkSize(mask net.IPMask) int32 {
m := net.IPv4Mask(0, 0, 0, 0)
for i := 0; i < net.IPv4len; i++ {
m[i] = ^mask[i]
}
return int32(binary.BigEndian.Uint32(m)) + 1
}
func checkRouteOverlaps(networks []*net.IPNet, dockerNetwork *net.IPNet) error {
for _, network := range networks {
if networkOverlaps(dockerNetwork, network) {
return fmt.Errorf("Network %s is already routed: '%s'", dockerNetwork, network)
}
}
return nil
}
func checkNameserverOverlaps(nameservers []string, dockerNetwork *net.IPNet) error {
if len(nameservers) > 0 {
for _, ns := range nameservers {
_, nsNetwork, err := net.ParseCIDR(ns)
if err != nil {
return err
}
if networkOverlaps(dockerNetwork, nsNetwork) {
return fmt.Errorf("%s overlaps nameserver %s", dockerNetwork, nsNetwork)
}
}
}
return nil
}
// CreateBridgeIface creates a network bridge interface on the host system with the name `ifaceName`,
// and attempts to configure it with an address which doesn't conflict with any other interface on the host.
// If it can't find an address which doesn't conflict, it will return an error.
func CreateBridgeIface(config *DaemonConfig) error {
addrs := []string{
// Here we don't follow the convention of using the 1st IP of the range for the gateway.
// This is to use the same gateway IPs as the /24 ranges, which predate the /16 ranges.
// In theory this shouldn't matter - in practice there's bound to be a few scripts relying
// on the internal addressing or other stupid things like that.
// The shouldn't, but hey, let's not break them unless we really have to.
"172.17.42.1/16", // Don't use 172.16.0.0/16, it conflicts with EC2 DNS 172.16.0.23
"10.0.42.1/16", // Don't even try using the entire /8, that's too intrusive
"10.1.42.1/16",
"10.42.42.1/16",
"172.16.42.1/24",
"172.16.43.1/24",
"172.16.44.1/24",
"10.0.42.1/24",
"10.0.43.1/24",
"192.168.42.1/24",
"192.168.43.1/24",
"192.168.44.1/24",
}
nameservers := []string{}
resolvConf, _ := utils.GetResolvConf()
// we don't check for an error here, because we don't really care
// if we can't read /etc/resolv.conf. So instead we skip the append
// if resolvConf is nil. It either doesn't exist, or we can't read it
// for some reason.
if resolvConf != nil {
nameservers = append(nameservers, utils.GetNameserversAsCIDR(resolvConf)...)
}
var ifaceAddr string
for _, addr := range addrs {
_, dockerNetwork, err := net.ParseCIDR(addr)
if err != nil {
return err
}
routes, err := netlink.NetworkGetRoutes()
if err != nil {
return err
}
if err := checkRouteOverlaps(routes, dockerNetwork); err == nil {
if err := checkNameserverOverlaps(nameservers, dockerNetwork); err == nil {
ifaceAddr = addr
break
}
} else {
2013-05-14 22:37:35 +00:00
utils.Debugf("%s: %s", addr, err)
}
}
if ifaceAddr == "" {
return fmt.Errorf("Could not find a free IP address range for interface '%s'. Please configure its address manually and run 'docker -b %s'", config.BridgeIface, config.BridgeIface)
}
utils.Debugf("Creating bridge %s with network %s", config.BridgeIface, ifaceAddr)
if err := netlink.NetworkLinkAdd(config.BridgeIface, "bridge"); err != nil {
return fmt.Errorf("Error creating bridge: %s", err)
}
iface, err := net.InterfaceByName(config.BridgeIface)
if err != nil {
return err
}
ipAddr, ipNet, err := net.ParseCIDR(ifaceAddr)
if err != nil {
return err
}
if netlink.NetworkLinkAddIp(iface, ipAddr, ipNet); err != nil {
return fmt.Errorf("Unable to add private network: %s", err)
}
if err := netlink.NetworkLinkUp(iface); err != nil {
return fmt.Errorf("Unable to start network bridge: %s", err)
}
if config.EnableIptables {
2013-11-05 15:33:07 +00:00
if output, err := iptables.Raw("-t", "nat", "-A", "POSTROUTING", "-s", ifaceAddr,
"!", "-d", ifaceAddr, "-j", "MASQUERADE"); err != nil {
return fmt.Errorf("Unable to enable network bridge NAT: %s", err)
2013-11-05 15:33:07 +00:00
} else if len(output) != 0 {
return fmt.Errorf("Error iptables postrouting: %s", output)
}
}
return nil
}
// Return the IPv4 address of a network interface
func getIfaceAddr(name string) (net.Addr, error) {
iface, err := net.InterfaceByName(name)
if err != nil {
return nil, err
}
addrs, err := iface.Addrs()
if err != nil {
return nil, err
}
var addrs4 []net.Addr
for _, addr := range addrs {
ip := (addr.(*net.IPNet)).IP
if ip4 := ip.To4(); len(ip4) == net.IPv4len {
addrs4 = append(addrs4, addr)
}
}
switch {
case len(addrs4) == 0:
return nil, fmt.Errorf("Interface %v has no IP addresses", name)
case len(addrs4) > 1:
fmt.Printf("Interface %v has more than 1 IPv4 address. Defaulting to using %v\n",
name, (addrs4[0].(*net.IPNet)).IP)
}
return addrs4[0], nil
}
// Port mapper takes care of mapping external ports to containers by setting
// up iptables rules.
// It keeps track of all mappings and is able to unmap at will
type PortMapper struct {
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
tcpMapping map[int]*net.TCPAddr
tcpProxies map[int]proxy.Proxy
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
udpMapping map[int]*net.UDPAddr
udpProxies map[int]proxy.Proxy
iptables *iptables.Chain
defaultIp net.IP
}
func (mapper *PortMapper) Map(ip net.IP, port int, backendAddr net.Addr) error {
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
if _, isTCP := backendAddr.(*net.TCPAddr); isTCP {
backendPort := backendAddr.(*net.TCPAddr).Port
backendIP := backendAddr.(*net.TCPAddr).IP
if mapper.iptables != nil {
if err := mapper.iptables.Forward(iptables.Add, ip, port, "tcp", backendIP.String(), backendPort); err != nil {
return err
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
}
mapper.tcpMapping[port] = backendAddr.(*net.TCPAddr)
proxy, err := proxy.NewProxy(&net.TCPAddr{IP: ip, Port: port}, backendAddr)
if err != nil {
mapper.Unmap(ip, port, "tcp")
return err
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
mapper.tcpProxies[port] = proxy
go proxy.Run()
} else {
backendPort := backendAddr.(*net.UDPAddr).Port
backendIP := backendAddr.(*net.UDPAddr).IP
if mapper.iptables != nil {
if err := mapper.iptables.Forward(iptables.Add, ip, port, "udp", backendIP.String(), backendPort); err != nil {
return err
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
}
mapper.udpMapping[port] = backendAddr.(*net.UDPAddr)
proxy, err := proxy.NewProxy(&net.UDPAddr{IP: ip, Port: port}, backendAddr)
if err != nil {
mapper.Unmap(ip, port, "udp")
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
return err
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
mapper.udpProxies[port] = proxy
go proxy.Run()
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
return nil
}
func (mapper *PortMapper) Unmap(ip net.IP, port int, proto string) error {
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
if proto == "tcp" {
backendAddr, ok := mapper.tcpMapping[port]
if !ok {
return fmt.Errorf("Port tcp/%v is not mapped", port)
}
if proxy, exists := mapper.tcpProxies[port]; exists {
proxy.Close()
delete(mapper.tcpProxies, port)
}
if mapper.iptables != nil {
if err := mapper.iptables.Forward(iptables.Delete, ip, port, proto, backendAddr.IP.String(), backendAddr.Port); err != nil {
return err
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
}
delete(mapper.tcpMapping, port)
} else {
backendAddr, ok := mapper.udpMapping[port]
if !ok {
return fmt.Errorf("Port udp/%v is not mapped", port)
}
if proxy, exists := mapper.udpProxies[port]; exists {
proxy.Close()
delete(mapper.udpProxies, port)
}
if mapper.iptables != nil {
if err := mapper.iptables.Forward(iptables.Delete, ip, port, proto, backendAddr.IP.String(), backendAddr.Port); err != nil {
return err
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
}
delete(mapper.udpMapping, port)
}
return nil
}
func newPortMapper(config *DaemonConfig) (*PortMapper, error) {
// We can always try removing the iptables
if err := iptables.RemoveExistingChain("DOCKER"); err != nil {
return nil, err
}
var chain *iptables.Chain
if config.EnableIptables {
var err error
chain, err = iptables.NewChain("DOCKER", config.BridgeIface)
if err != nil {
return nil, fmt.Errorf("Failed to create DOCKER chain: %s", err)
}
}
mapper := &PortMapper{
tcpMapping: make(map[int]*net.TCPAddr),
tcpProxies: make(map[int]proxy.Proxy),
udpMapping: make(map[int]*net.UDPAddr),
udpProxies: make(map[int]proxy.Proxy),
iptables: chain,
defaultIp: config.DefaultIp,
}
return mapper, nil
}
2013-02-25 18:45:23 +00:00
2013-08-12 17:53:06 +00:00
// Port allocator: Automatically allocate and release networking ports
type PortAllocator struct {
sync.Mutex
inUse map[int]struct{}
fountain chan int
quit chan bool
2013-02-25 18:45:23 +00:00
}
func (alloc *PortAllocator) runFountain() {
for {
for port := portRangeStart; port < portRangeEnd; port++ {
select {
case alloc.fountain <- port:
case quit := <-alloc.quit:
if quit {
return
}
}
}
2013-02-25 18:45:23 +00:00
}
}
// FIXME: Release can no longer fail, change its prototype to reflect that.
func (alloc *PortAllocator) Release(port int) error {
2013-05-14 22:37:35 +00:00
utils.Debugf("Releasing %d", port)
alloc.Lock()
delete(alloc.inUse, port)
alloc.Unlock()
return nil
}
func (alloc *PortAllocator) Acquire(port int) (int, error) {
2013-05-14 22:37:35 +00:00
utils.Debugf("Acquiring %d", port)
if port == 0 {
// Allocate a port from the fountain
for port := range alloc.fountain {
if _, err := alloc.Acquire(port); err == nil {
return port, nil
}
}
return -1, fmt.Errorf("Port generator ended unexpectedly")
2013-02-25 18:45:23 +00:00
}
alloc.Lock()
defer alloc.Unlock()
if _, inUse := alloc.inUse[port]; inUse {
return -1, fmt.Errorf("Port already in use: %d", port)
}
alloc.inUse[port] = struct{}{}
return port, nil
2013-02-25 18:45:23 +00:00
}
func (alloc *PortAllocator) Close() error {
alloc.quit <- true
close(alloc.quit)
close(alloc.fountain)
return nil
}
func newPortAllocator() (*PortAllocator, error) {
allocator := &PortAllocator{
2013-04-12 14:37:24 +00:00
inUse: make(map[int]struct{}),
fountain: make(chan int),
quit: make(chan bool),
}
go allocator.runFountain()
return allocator, nil
}
2013-08-12 17:53:06 +00:00
// IP allocator: Automatically allocate and release networking ports
type IPAllocator struct {
network *net.IPNet
queueAlloc chan allocatedIP
queueReleased chan net.IP
inUse map[int32]struct{}
quit chan bool
}
type allocatedIP struct {
ip net.IP
err error
}
func (alloc *IPAllocator) run() {
firstIP, _ := networkRange(alloc.network)
ipNum := ipToInt(firstIP)
ownIP := ipToInt(alloc.network.IP)
size := networkSize(alloc.network.Mask)
pos := int32(1)
max := size - 2 // -1 for the broadcast address, -1 for the gateway address
for {
var (
newNum int32
inUse bool
)
// Find first unused IP, give up after one whole round
for attempt := int32(0); attempt < max; attempt++ {
newNum = ipNum + pos
pos = pos%max + 1
// The network's IP is never okay to use
if newNum == ownIP {
continue
}
if _, inUse = alloc.inUse[newNum]; !inUse {
// We found an unused IP
break
}
2013-02-25 18:45:23 +00:00
}
2013-06-04 18:00:22 +00:00
ip := allocatedIP{ip: intToIP(newNum)}
if inUse {
ip.err = errors.New("No unallocated IP available")
2013-02-25 18:45:23 +00:00
}
select {
case quit := <-alloc.quit:
if quit {
return
}
case alloc.queueAlloc <- ip:
alloc.inUse[newNum] = struct{}{}
case released := <-alloc.queueReleased:
r := ipToInt(released)
delete(alloc.inUse, r)
if inUse {
// If we couldn't allocate a new IP, the released one
// will be the only free one now, so instantly use it
// next time
pos = r - ipNum
} else {
// Use same IP as last time
if pos == 1 {
pos = max
} else {
pos--
}
}
2013-02-25 18:45:23 +00:00
}
}
}
func (alloc *IPAllocator) Acquire() (net.IP, error) {
ip := <-alloc.queueAlloc
return ip.ip, ip.err
}
func (alloc *IPAllocator) Release(ip net.IP) {
alloc.queueReleased <- ip
2013-02-25 18:45:23 +00:00
}
func (alloc *IPAllocator) Close() error {
alloc.quit <- true
close(alloc.quit)
close(alloc.queueAlloc)
close(alloc.queueReleased)
return nil
}
func newIPAllocator(network *net.IPNet) *IPAllocator {
alloc := &IPAllocator{
network: network,
queueAlloc: make(chan allocatedIP),
queueReleased: make(chan net.IP),
inUse: make(map[int32]struct{}),
quit: make(chan bool),
}
go alloc.run()
return alloc
}
// Network interface represents the networking stack of a container
type NetworkInterface struct {
IPNet net.IPNet
Gateway net.IP
manager *NetworkManager
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
extPorts []*Nat
disabled bool
}
// Allocate an external port and map it to the interface
func (iface *NetworkInterface) AllocatePort(port Port, binding PortBinding) (*Nat, error) {
if iface.disabled {
return nil, fmt.Errorf("Trying to allocate port for interface %v, which is disabled", iface) // FIXME
}
ip := iface.manager.portMapper.defaultIp
if binding.HostIp != "" {
ip = net.ParseIP(binding.HostIp)
} else {
binding.HostIp = ip.String()
}
nat := &Nat{
Port: port,
Binding: binding,
}
containerPort, err := parsePort(port.Port())
if err != nil {
return nil, err
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
hostPort, _ := parsePort(nat.Binding.HostPort)
if nat.Port.Proto() == "tcp" {
extPort, err := iface.manager.tcpPortAllocator.Acquire(hostPort)
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
if err != nil {
return nil, err
}
backend := &net.TCPAddr{IP: iface.IPNet.IP, Port: containerPort}
if err := iface.manager.portMapper.Map(ip, extPort, backend); err != nil {
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
iface.manager.tcpPortAllocator.Release(extPort)
return nil, err
}
nat.Binding.HostPort = strconv.Itoa(extPort)
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
} else {
extPort, err := iface.manager.udpPortAllocator.Acquire(hostPort)
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
if err != nil {
return nil, err
}
backend := &net.UDPAddr{IP: iface.IPNet.IP, Port: containerPort}
if err := iface.manager.portMapper.Map(ip, extPort, backend); err != nil {
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
iface.manager.udpPortAllocator.Release(extPort)
return nil, err
}
nat.Binding.HostPort = strconv.Itoa(extPort)
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
iface.extPorts = append(iface.extPorts, nat)
return nat, nil
}
type Nat struct {
Port Port
Binding PortBinding
}
func (n *Nat) String() string {
return fmt.Sprintf("%s:%d:%d/%s", n.Binding.HostIp, n.Binding.HostPort, n.Port.Port(), n.Port.Proto())
}
// Release: Network cleanup - release all resources
func (iface *NetworkInterface) Release() {
if iface.disabled {
return
}
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
for _, nat := range iface.extPorts {
hostPort, err := parsePort(nat.Binding.HostPort)
if err != nil {
log.Printf("Unable to get host port: %s", err)
continue
}
ip := net.ParseIP(nat.Binding.HostIp)
utils.Debugf("Unmaping %s/%s", nat.Port.Proto, nat.Binding.HostPort)
if err := iface.manager.portMapper.Unmap(ip, hostPort, nat.Port.Proto()); err != nil {
log.Printf("Unable to unmap port %s: %s", nat, err)
}
if nat.Port.Proto() == "tcp" {
if err := iface.manager.tcpPortAllocator.Release(hostPort); err != nil {
log.Printf("Unable to release port %s", nat)
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
}
} else if err := iface.manager.udpPortAllocator.Release(hostPort); err != nil {
log.Printf("Unable to release port %s: %s", nat, err)
}
}
iface.manager.ipAllocator.Release(iface.IPNet.IP)
}
// Network Manager manages a set of network interfaces
// Only *one* manager per host machine should be used
type NetworkManager struct {
bridgeIface string
bridgeNetwork *net.IPNet
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
ipAllocator *IPAllocator
tcpPortAllocator *PortAllocator
udpPortAllocator *PortAllocator
portMapper *PortMapper
disabled bool
}
// Allocate a network interface
func (manager *NetworkManager) Allocate() (*NetworkInterface, error) {
if manager.disabled {
return &NetworkInterface{disabled: true}, nil
}
var ip net.IP
var err error
ip, err = manager.ipAllocator.Acquire()
2013-02-25 18:45:23 +00:00
if err != nil {
return nil, err
2013-02-25 18:45:23 +00:00
}
// avoid duplicate IP
ipNum := ipToInt(ip)
firstIP := manager.ipAllocator.network.IP.To4().Mask(manager.ipAllocator.network.Mask)
firstIPNum := ipToInt(firstIP) + 1
if firstIPNum == ipNum {
ip, err = manager.ipAllocator.Acquire()
if err != nil {
return nil, err
}
}
iface := &NetworkInterface{
IPNet: net.IPNet{IP: ip, Mask: manager.bridgeNetwork.Mask},
Gateway: manager.bridgeNetwork.IP,
manager: manager,
}
return iface, nil
}
func (manager *NetworkManager) Close() error {
err1 := manager.tcpPortAllocator.Close()
err2 := manager.udpPortAllocator.Close()
err3 := manager.ipAllocator.Close()
if err1 != nil {
return err1
}
if err2 != nil {
return err2
}
return err3
}
func newNetworkManager(config *DaemonConfig) (*NetworkManager, error) {
if config.BridgeIface == DisableNetworkBridge {
manager := &NetworkManager{
disabled: true,
}
return manager, nil
}
addr, err := getIfaceAddr(config.BridgeIface)
if err != nil {
// If the iface is not found, try to create it
if err := CreateBridgeIface(config); err != nil {
return nil, err
}
addr, err = getIfaceAddr(config.BridgeIface)
if err != nil {
return nil, err
}
}
network := addr.(*net.IPNet)
// Configure iptables for link support
if config.EnableIptables {
args := []string{"FORWARD", "-i", config.BridgeIface, "-o", config.BridgeIface, "-j", "DROP"}
if !config.InterContainerCommunication {
if !iptables.Exists(args...) {
utils.Debugf("Disable inter-container communication")
2013-11-05 15:33:07 +00:00
if output, err := iptables.Raw(append([]string{"-A"}, args...)...); err != nil {
return nil, fmt.Errorf("Unable to prevent intercontainer communication: %s", err)
2013-11-05 15:33:07 +00:00
} else if len(output) != 0 {
return nil, fmt.Errorf("Error enabling iptables: %s", output)
}
}
} else {
utils.Debugf("Enable inter-container communication")
iptables.Raw(append([]string{"-D"}, args...)...)
}
}
ipAllocator := newIPAllocator(network)
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
tcpPortAllocator, err := newPortAllocator()
if err != nil {
return nil, err
}
udpPortAllocator, err := newPortAllocator()
if err != nil {
return nil, err
}
portMapper, err := newPortMapper(config)
if err != nil {
return nil, err
}
manager := &NetworkManager{
bridgeIface: config.BridgeIface,
Add support for UDP (closes #33) API Changes ----------- The port notation is extended to support "/udp" or "/tcp" at the *end* of the specifier string (and defaults to tcp if "/tcp" or "/udp" are missing) `docker ps` now shows UDP ports as "frontend->backend/udp". Nothing changes for TCP ports. `docker inspect` now displays two sub-dictionaries: "Tcp" and "Udp", under "PortMapping" in "NetworkSettings". Theses changes stand true for the values returned by the HTTP API too. This changeset will definitely break tools built upon the API (or upon `docker inspect`). A less intrusive way to add UDP ports in `docker inspect` would be to simply add "/udp" for UDP ports but it will still break existing applications which tries to convert the whole field to an integer. I believe that having two TCP/UDP sub-dictionaries is better because it makes the whole thing more clear and more easy to parse right away (i.e: you don't have to check the format of the string, split it and convert the right part to an integer) Code Changes ------------ Significant changes in network.go: - A second PortAllocator is instantiated for the UDP range; - PortMapper maintains separate mapping for TCP and UDP; - The extPorts array in NetworkInterface is now an array of Nat objects (so we can know on which protocol a given port was mapped when NetworkInterface.Release() is called); - TCP proxying on localhost has been moved away in network_proxy.go. localhost proxy code rewrite in network_proxy.go: We have to proxy the traffic between localhost:frontend-port and container:backend-port because Netfilter doesn't work properly on the loopback interface and DNAT iptable rules aren't applied there. - Goroutines in the TCP proxying code are now explicitly stopped when the proxy is stopped; - UDP connection tracking using a map (more infos in [1]); - Support for IPv6 (to be more accurate, the code is transparent to the Go net package, so you can use, tcp/tcp4/tcp6/udp/udp4/udp6); - Single Proxy interface for both UDP and TCP proxying; - Full test suite. [1] https://github.com/dotcloud/docker/issues/33#issuecomment-20010400
2013-06-11 22:46:23 +00:00
bridgeNetwork: network,
ipAllocator: ipAllocator,
tcpPortAllocator: tcpPortAllocator,
udpPortAllocator: udpPortAllocator,
portMapper: portMapper,
}
return manager, nil
2013-02-25 18:45:23 +00:00
}