diff options
-rw-r--r-- | pkg/tcpip/ports/ports.go | 148 | ||||
-rwxr-xr-x | pkg/tcpip/ports/ports_state_autogen.go | 20 | ||||
-rw-r--r-- | pkg/tcpip/transport/tcp/endpoint.go | 25 | ||||
-rwxr-xr-x | pkg/tcpip/transport/tcp/tcp_state_autogen.go | 2 | ||||
-rw-r--r-- | pkg/tcpip/transport/udp/endpoint.go | 19 | ||||
-rwxr-xr-x | pkg/tcpip/transport/udp/udp_state_autogen.go | 2 |
6 files changed, 174 insertions, 42 deletions
diff --git a/pkg/tcpip/ports/ports.go b/pkg/tcpip/ports/ports.go index 30cea8996..6c5e19e8f 100644 --- a/pkg/tcpip/ports/ports.go +++ b/pkg/tcpip/ports/ports.go @@ -41,6 +41,30 @@ type portDescriptor struct { port uint16 } +// Flags represents the type of port reservation. +// +// +stateify savable +type Flags struct { + // MostRecent represents UDP SO_REUSEADDR. + MostRecent bool + + // LoadBalanced indicates SO_REUSEPORT. + // + // LoadBalanced takes precidence over MostRecent. + LoadBalanced bool +} + +func (f Flags) bits() reuseFlag { + var rf reuseFlag + if f.MostRecent { + rf |= mostRecentFlag + } + if f.LoadBalanced { + rf |= loadBalancedFlag + } + return rf +} + // PortManager manages allocating, reserving and releasing ports. type PortManager struct { mu sync.RWMutex @@ -54,9 +78,59 @@ type PortManager struct { hint uint32 } +type reuseFlag int + +const ( + mostRecentFlag reuseFlag = 1 << iota + loadBalancedFlag + nextFlag + + flagMask = nextFlag - 1 +) + type portNode struct { - reuse bool - refs int + // refs stores the count for each possible flag combination. + refs [nextFlag]int +} + +func (p portNode) totalRefs() int { + var total int + for _, r := range p.refs { + total += r + } + return total +} + +// flagRefs returns the number of references with all specified flags. +func (p portNode) flagRefs(flags reuseFlag) int { + var total int + for i, r := range p.refs { + if reuseFlag(i)&flags == flags { + total += r + } + } + return total +} + +// allRefsHave returns if all references have all specified flags. +func (p portNode) allRefsHave(flags reuseFlag) bool { + for i, r := range p.refs { + if reuseFlag(i)&flags == flags && r > 0 { + return false + } + } + return true +} + +// intersectionRefs returns the set of flags shared by all references. +func (p portNode) intersectionRefs() reuseFlag { + intersection := flagMask + for i, r := range p.refs { + if r > 0 { + intersection &= reuseFlag(i) + } + } + return intersection } // deviceNode is never empty. When it has no elements, it is removed from the @@ -66,30 +140,44 @@ type deviceNode map[tcpip.NICID]portNode // isAvailable checks whether binding is possible by device. If not binding to a // device, check against all portNodes. If binding to a specific device, check // against the unspecified device and the provided device. -func (d deviceNode) isAvailable(reuse bool, bindToDevice tcpip.NICID) bool { +// +// If either of the port reuse flags is enabled on any of the nodes, all nodes +// sharing a port must share at least one reuse flag. This matches Linux's +// behavior. +func (d deviceNode) isAvailable(flags Flags, bindToDevice tcpip.NICID) bool { + flagBits := flags.bits() if bindToDevice == 0 { // Trying to binding all devices. - if !reuse { + if flagBits == 0 { // Can't bind because the (addr,port) is already bound. return false } + intersection := flagMask for _, p := range d { - if !p.reuse { - // Can't bind because the (addr,port) was previously bound without reuse. + i := p.intersectionRefs() + intersection &= i + if intersection&flagBits == 0 { + // Can't bind because the (addr,port) was + // previously bound without reuse. return false } } return true } + intersection := flagMask + if p, ok := d[0]; ok { - if !reuse || !p.reuse { + intersection = p.intersectionRefs() + if intersection&flagBits == 0 { return false } } if p, ok := d[bindToDevice]; ok { - if !reuse || !p.reuse { + i := p.intersectionRefs() + intersection &= i + if intersection&flagBits == 0 { return false } } @@ -103,12 +191,12 @@ type bindAddresses map[tcpip.Address]deviceNode // isAvailable checks whether an IP address is available to bind to. If the // address is the "any" address, check all other addresses. Otherwise, just // check against the "any" address and the provided address. -func (b bindAddresses) isAvailable(addr tcpip.Address, reuse bool, bindToDevice tcpip.NICID) bool { +func (b bindAddresses) isAvailable(addr tcpip.Address, flags Flags, bindToDevice tcpip.NICID) bool { if addr == anyIPAddress { // If binding to the "any" address then check that there are no conflicts // with all addresses. for _, d := range b { - if !d.isAvailable(reuse, bindToDevice) { + if !d.isAvailable(flags, bindToDevice) { return false } } @@ -117,14 +205,14 @@ func (b bindAddresses) isAvailable(addr tcpip.Address, reuse bool, bindToDevice // Check that there is no conflict with the "any" address. if d, ok := b[anyIPAddress]; ok { - if !d.isAvailable(reuse, bindToDevice) { + if !d.isAvailable(flags, bindToDevice) { return false } } // Check that this is no conflict with the provided address. if d, ok := b[addr]; ok { - if !d.isAvailable(reuse, bindToDevice) { + if !d.isAvailable(flags, bindToDevice) { return false } } @@ -190,17 +278,17 @@ func (s *PortManager) pickEphemeralPort(offset, count uint32, testPort func(p ui } // IsPortAvailable tests if the given port is available on all given protocols. -func (s *PortManager) IsPortAvailable(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool, bindToDevice tcpip.NICID) bool { +func (s *PortManager) IsPortAvailable(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID) bool { s.mu.Lock() defer s.mu.Unlock() - return s.isPortAvailableLocked(networks, transport, addr, port, reuse, bindToDevice) + return s.isPortAvailableLocked(networks, transport, addr, port, flags, bindToDevice) } -func (s *PortManager) isPortAvailableLocked(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool, bindToDevice tcpip.NICID) bool { +func (s *PortManager) isPortAvailableLocked(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID) bool { for _, network := range networks { desc := portDescriptor{network, transport, port} if addrs, ok := s.allocatedPorts[desc]; ok { - if !addrs.isAvailable(addr, reuse, bindToDevice) { + if !addrs.isAvailable(addr, flags, bindToDevice) { return false } } @@ -212,14 +300,14 @@ func (s *PortManager) isPortAvailableLocked(networks []tcpip.NetworkProtocolNumb // reserved by another endpoint. If port is zero, ReservePort will search for // an unreserved ephemeral port and reserve it, returning its value in the // "port" return value. -func (s *PortManager) ReservePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool, bindToDevice tcpip.NICID) (reservedPort uint16, err *tcpip.Error) { +func (s *PortManager) ReservePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID) (reservedPort uint16, err *tcpip.Error) { s.mu.Lock() defer s.mu.Unlock() // If a port is specified, just try to reserve it for all network // protocols. if port != 0 { - if !s.reserveSpecificPort(networks, transport, addr, port, reuse, bindToDevice) { + if !s.reserveSpecificPort(networks, transport, addr, port, flags, bindToDevice) { return 0, tcpip.ErrPortInUse } return port, nil @@ -227,15 +315,16 @@ func (s *PortManager) ReservePort(networks []tcpip.NetworkProtocolNumber, transp // A port wasn't specified, so try to find one. return s.PickEphemeralPort(func(p uint16) (bool, *tcpip.Error) { - return s.reserveSpecificPort(networks, transport, addr, p, reuse, bindToDevice), nil + return s.reserveSpecificPort(networks, transport, addr, p, flags, bindToDevice), nil }) } // reserveSpecificPort tries to reserve the given port on all given protocols. -func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool, bindToDevice tcpip.NICID) bool { - if !s.isPortAvailableLocked(networks, transport, addr, port, reuse, bindToDevice) { +func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID) bool { + if !s.isPortAvailableLocked(networks, transport, addr, port, flags, bindToDevice) { return false } + flagBits := flags.bits() // Reserve port on all network protocols. for _, network := range networks { @@ -250,12 +339,9 @@ func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber d = make(deviceNode) m[addr] = d } - if n, ok := d[bindToDevice]; ok { - n.refs++ - d[bindToDevice] = n - } else { - d[bindToDevice] = portNode{reuse: reuse, refs: 1} - } + n := d[bindToDevice] + n.refs[flagBits]++ + d[bindToDevice] = n } return true @@ -263,10 +349,12 @@ func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber // ReleasePort releases the reservation on a port/IP combination so that it can // be reserved by other endpoints. -func (s *PortManager) ReleasePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, bindToDevice tcpip.NICID) { +func (s *PortManager) ReleasePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID) { s.mu.Lock() defer s.mu.Unlock() + flagBits := flags.bits() + for _, network := range networks { desc := portDescriptor{network, transport, port} if m, ok := s.allocatedPorts[desc]; ok { @@ -278,9 +366,9 @@ func (s *PortManager) ReleasePort(networks []tcpip.NetworkProtocolNumber, transp if !ok { continue } - n.refs-- + n.refs[flagBits]-- d[bindToDevice] = n - if n.refs == 0 { + if n.refs == [nextFlag]int{} { delete(d, bindToDevice) } if len(d) == 0 { diff --git a/pkg/tcpip/ports/ports_state_autogen.go b/pkg/tcpip/ports/ports_state_autogen.go index 664cc3e71..badae1397 100755 --- a/pkg/tcpip/ports/ports_state_autogen.go +++ b/pkg/tcpip/ports/ports_state_autogen.go @@ -2,3 +2,23 @@ package ports +import ( + "gvisor.dev/gvisor/pkg/state" +) + +func (x *Flags) beforeSave() {} +func (x *Flags) save(m state.Map) { + x.beforeSave() + m.Save("MostRecent", &x.MostRecent) + m.Save("LoadBalanced", &x.LoadBalanced) +} + +func (x *Flags) afterLoad() {} +func (x *Flags) load(m state.Map) { + m.Load("MostRecent", &x.MostRecent) + m.Load("LoadBalanced", &x.LoadBalanced) +} + +func init() { + state.Register("ports.Flags", (*Flags)(nil), state.Fns{Save: (*Flags).save, Load: (*Flags).load}) +} diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index 9d4a87e30..4861ab513 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -30,6 +30,7 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/hash/jenkins" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/ports" "gvisor.dev/gvisor/pkg/tcpip/seqnum" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/tmutex" @@ -343,6 +344,7 @@ type endpoint struct { // Values used to reserve a port or register a transport endpoint // (which ever happens first). boundBindToDevice tcpip.NICID + boundPortFlags ports.Flags // effectiveNetProtos contains the network protocols actually in use. In // most cases it will only contain "netProto", but in cases like IPv6 @@ -737,9 +739,10 @@ func (e *endpoint) Close() { e.isRegistered = false } - e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice) e.isPortReserved = false e.boundBindToDevice = 0 + e.boundPortFlags = ports.Flags{} } // Mark endpoint as closed. @@ -800,10 +803,11 @@ func (e *endpoint) cleanupLocked() { } if e.isPortReserved { - e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice) e.isPortReserved = false } e.boundBindToDevice = 0 + e.boundPortFlags = ports.Flags{} e.route.Release() e.stack.CompleteTransportEndpointCleanup(e) @@ -1775,7 +1779,7 @@ func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) *tc } // reusePort is false below because connect cannot reuse a port even if // reusePort was set. - if !e.stack.IsPortAvailable(netProtos, ProtocolNumber, e.ID.LocalAddress, p, false /* reusePort */, e.bindToDevice) { + if !e.stack.IsPortAvailable(netProtos, ProtocolNumber, e.ID.LocalAddress, p, ports.Flags{LoadBalanced: false}, e.bindToDevice) { return false, nil } @@ -1802,7 +1806,7 @@ func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) *tc // before Connect: in such a case we don't want to hold on to // reservations anymore. if e.isPortReserved { - e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, origID.LocalAddress, origID.LocalPort, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, origID.LocalAddress, origID.LocalPort, e.boundPortFlags, e.boundBindToDevice) e.isPortReserved = false } @@ -2034,28 +2038,33 @@ func (e *endpoint) Bind(addr tcpip.FullAddress) (err *tcpip.Error) { } } - port, err := e.stack.ReservePort(netProtos, ProtocolNumber, addr.Addr, addr.Port, e.reusePort, e.bindToDevice) + flags := ports.Flags{ + LoadBalanced: e.reusePort, + } + port, err := e.stack.ReservePort(netProtos, ProtocolNumber, addr.Addr, addr.Port, flags, e.bindToDevice) if err != nil { return err } e.boundBindToDevice = e.bindToDevice + e.boundPortFlags = flags e.isPortReserved = true e.effectiveNetProtos = netProtos e.ID.LocalPort = port // Any failures beyond this point must remove the port registration. - defer func(bindToDevice tcpip.NICID) { + defer func(portFlags ports.Flags, bindToDevice tcpip.NICID) { if err != nil { - e.stack.ReleasePort(netProtos, ProtocolNumber, addr.Addr, port, bindToDevice) + e.stack.ReleasePort(netProtos, ProtocolNumber, addr.Addr, port, portFlags, bindToDevice) e.isPortReserved = false e.effectiveNetProtos = nil e.ID.LocalPort = 0 e.ID.LocalAddress = "" e.boundNICID = 0 e.boundBindToDevice = 0 + e.boundPortFlags = ports.Flags{} } - }(e.boundBindToDevice) + }(e.boundPortFlags, e.boundBindToDevice) // If an address is specified, we must ensure that it's one of our // local addresses. diff --git a/pkg/tcpip/transport/tcp/tcp_state_autogen.go b/pkg/tcpip/transport/tcp/tcp_state_autogen.go index 9d16e3526..cc3f5c931 100755 --- a/pkg/tcpip/transport/tcp/tcp_state_autogen.go +++ b/pkg/tcpip/transport/tcp/tcp_state_autogen.go @@ -112,6 +112,7 @@ func (x *endpoint) save(m state.Map) { m.Save("isConnectNotified", &x.isConnectNotified) m.Save("broadcast", &x.broadcast) m.Save("boundBindToDevice", &x.boundBindToDevice) + m.Save("boundPortFlags", &x.boundPortFlags) m.Save("workerRunning", &x.workerRunning) m.Save("workerCleanup", &x.workerCleanup) m.Save("sendTSOk", &x.sendTSOk) @@ -165,6 +166,7 @@ func (x *endpoint) load(m state.Map) { m.Load("isConnectNotified", &x.isConnectNotified) m.Load("broadcast", &x.broadcast) m.Load("boundBindToDevice", &x.boundBindToDevice) + m.Load("boundPortFlags", &x.boundPortFlags) m.Load("workerRunning", &x.workerRunning) m.Load("workerCleanup", &x.workerCleanup) m.Load("sendTSOk", &x.sendTSOk) diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go index 4b161e404..1ac4705af 100644 --- a/pkg/tcpip/transport/udp/endpoint.go +++ b/pkg/tcpip/transport/udp/endpoint.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/buffer" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/iptables" + "gvisor.dev/gvisor/pkg/tcpip/ports" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/waiter" ) @@ -107,6 +108,7 @@ type endpoint struct { // Values used to reserve a port or register a transport endpoint. // (which ever happens first). boundBindToDevice tcpip.NICID + boundPortFlags ports.Flags // sendTOS represents IPv4 TOS or IPv6 TrafficClass, // applied while sending packets. Defaults to 0 as on Linux. @@ -180,8 +182,9 @@ func (e *endpoint) Close() { switch e.state { case StateBound, StateConnected: e.stack.UnregisterTransportEndpoint(e.RegisterNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e, e.boundBindToDevice) - e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice) e.boundBindToDevice = 0 + e.boundPortFlags = ports.Flags{} } for _, mem := range e.multicastMemberships { @@ -895,7 +898,8 @@ func (e *endpoint) Disconnect() *tcpip.Error { } else { if e.ID.LocalPort != 0 { // Release the ephemeral port. - e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice) + e.boundPortFlags = ports.Flags{} } e.state = StateInitial } @@ -1042,16 +1046,23 @@ func (*endpoint) Accept() (tcpip.Endpoint, *waiter.Queue, *tcpip.Error) { func (e *endpoint) registerWithStack(nicID tcpip.NICID, netProtos []tcpip.NetworkProtocolNumber, id stack.TransportEndpointID) (stack.TransportEndpointID, tcpip.NICID, *tcpip.Error) { if e.ID.LocalPort == 0 { - port, err := e.stack.ReservePort(netProtos, ProtocolNumber, id.LocalAddress, id.LocalPort, e.reusePort, e.bindToDevice) + flags := ports.Flags{ + LoadBalanced: e.reusePort, + // FIXME(b/129164367): Support SO_REUSEADDR. + MostRecent: false, + } + port, err := e.stack.ReservePort(netProtos, ProtocolNumber, id.LocalAddress, id.LocalPort, flags, e.bindToDevice) if err != nil { return id, e.bindToDevice, err } + e.boundPortFlags = flags id.LocalPort = port } err := e.stack.RegisterTransportEndpoint(nicID, netProtos, ProtocolNumber, id, e, e.reusePort, e.bindToDevice) if err != nil { - e.stack.ReleasePort(netProtos, ProtocolNumber, id.LocalAddress, id.LocalPort, e.bindToDevice) + e.stack.ReleasePort(netProtos, ProtocolNumber, id.LocalAddress, id.LocalPort, e.boundPortFlags, e.bindToDevice) + e.boundPortFlags = ports.Flags{} } return id, e.bindToDevice, err } diff --git a/pkg/tcpip/transport/udp/udp_state_autogen.go b/pkg/tcpip/transport/udp/udp_state_autogen.go index 48a3bfc79..5c4b43214 100755 --- a/pkg/tcpip/transport/udp/udp_state_autogen.go +++ b/pkg/tcpip/transport/udp/udp_state_autogen.go @@ -49,6 +49,7 @@ func (x *endpoint) save(m state.Map) { m.Save("bindToDevice", &x.bindToDevice) m.Save("broadcast", &x.broadcast) m.Save("boundBindToDevice", &x.boundBindToDevice) + m.Save("boundPortFlags", &x.boundPortFlags) m.Save("sendTOS", &x.sendTOS) m.Save("shutdownFlags", &x.shutdownFlags) m.Save("multicastMemberships", &x.multicastMemberships) @@ -76,6 +77,7 @@ func (x *endpoint) load(m state.Map) { m.Load("bindToDevice", &x.bindToDevice) m.Load("broadcast", &x.broadcast) m.Load("boundBindToDevice", &x.boundBindToDevice) + m.Load("boundPortFlags", &x.boundPortFlags) m.Load("sendTOS", &x.sendTOS) m.Load("shutdownFlags", &x.shutdownFlags) m.Load("multicastMemberships", &x.multicastMemberships) |