diff options
author | Ian Gudger <igudger@google.com> | 2020-06-23 19:14:05 -0700 |
---|---|---|
committer | gVisor bot <gvisor-bot@google.com> | 2020-06-23 19:15:38 -0700 |
commit | 2141013dcef04de0591e36462ea997be1b6374d7 (patch) | |
tree | 7d9eea880528202a3125f614c97dd5eb35c54d98 /pkg | |
parent | 0f328beb0d830a6e3b88b9f2cc1c9cb48f55c752 (diff) |
Add support for SO_REUSEADDR to TCP sockets/endpoints.
For TCP sockets, SO_REUSEADDR relaxes the rules for binding addresses.
gVisor/netstack already supported a behavior similar to SO_REUSEADDR, but did
not allow disabling it. This change brings the SO_REUSEADDR behavior closer to
the behavior implemented by Linux and adds a new SO_REUSEADDR disabled
behavior. Like Linux, SO_REUSEADDR is now disabled by default.
PiperOrigin-RevId: 317984380
Diffstat (limited to 'pkg')
-rw-r--r-- | pkg/tcpip/ports/ports.go | 206 | ||||
-rw-r--r-- | pkg/tcpip/ports/ports_test.go | 58 | ||||
-rw-r--r-- | pkg/tcpip/stack/stack.go | 6 | ||||
-rw-r--r-- | pkg/tcpip/stack/transport_demuxer.go | 65 | ||||
-rw-r--r-- | pkg/tcpip/transport/tcp/accept.go | 140 | ||||
-rw-r--r-- | pkg/tcpip/transport/tcp/endpoint.go | 100 | ||||
-rw-r--r-- | pkg/tcpip/transport/tcp/endpoint_state.go | 31 | ||||
-rw-r--r-- | pkg/tcpip/transport/tcp/tcp_test.go | 18 | ||||
-rw-r--r-- | pkg/tcpip/transport/udp/endpoint.go | 8 |
9 files changed, 464 insertions, 168 deletions
diff --git a/pkg/tcpip/ports/ports.go b/pkg/tcpip/ports/ports.go index edc29ad27..f6d592eb5 100644 --- a/pkg/tcpip/ports/ports.go +++ b/pkg/tcpip/ports/ports.go @@ -52,6 +52,9 @@ type Flags struct { // // LoadBalanced takes precidence over MostRecent. LoadBalanced bool + + // TupleOnly represents TCP SO_REUSEADDR. + TupleOnly bool } // Bits converts the Flags to their bitset form. @@ -63,6 +66,9 @@ func (f Flags) Bits() BitFlags { if f.LoadBalanced { rf |= LoadBalancedFlag } + if f.TupleOnly { + rf |= TupleOnlyFlag + } return rf } @@ -98,6 +104,9 @@ const ( // LoadBalancedFlag represents Flags.LoadBalanced. LoadBalancedFlag + // TupleOnlyFlag represents Flags.TupleOnly. + TupleOnlyFlag + // nextFlag is the value that the next added flag will have. // // It is used to calculate FlagMask below. It is also the number of @@ -106,6 +115,10 @@ const ( // FlagMask is a bit mask for BitFlags. FlagMask = nextFlag - 1 + + // MultiBindFlagMask contains the flags that allow binding the same + // tuple multiple times. + MultiBindFlagMask = MostRecentFlag | LoadBalancedFlag ) // ToFlags converts the bitset into a Flags struct. @@ -113,6 +126,7 @@ func (f BitFlags) ToFlags() Flags { return Flags{ MostRecent: f&MostRecentFlag != 0, LoadBalanced: f&LoadBalancedFlag != 0, + TupleOnly: f&TupleOnlyFlag != 0, } } @@ -175,9 +189,54 @@ func (c FlagCounter) IntersectionRefs() BitFlags { return intersection } +type destination struct { + addr tcpip.Address + port uint16 +} + +func makeDestination(a tcpip.FullAddress) destination { + return destination{ + a.Addr, + a.Port, + } +} + +// portNode is never empty. When it has no elements, it is removed from the +// map that references it. +type portNode map[destination]FlagCounter + +// intersectionRefs calculates the intersection of flag bit values which affect +// the specified destination. +// +// If no destinations are present, all flag values are returned as there are no +// entries to limit possible flag values of a new entry. +// +// In addition to the intersection, the number of intersecting refs is +// returned. +func (p portNode) intersectionRefs(dst destination) (BitFlags, int) { + intersection := FlagMask + var count int + + for d, f := range p { + if d == dst { + intersection &= f.IntersectionRefs() + count++ + continue + } + // Wildcard destinations affect all destinations for TupleOnly. + if d.addr == anyIPAddress || dst.addr == anyIPAddress { + // Only bitwise and the TupleOnlyFlag. + intersection &= ((^TupleOnlyFlag) | f.IntersectionRefs()) + count++ + } + } + + return intersection, count +} + // deviceNode is never empty. When it has no elements, it is removed from the // map that references it. -type deviceNode map[tcpip.NICID]FlagCounter +type deviceNode map[tcpip.NICID]portNode // isAvailable checks whether binding is possible by device. If not binding to a // device, check against all FlagCounters. If binding to a specific device, check @@ -186,17 +245,15 @@ type deviceNode map[tcpip.NICID]FlagCounter // 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 { +func (d deviceNode) isAvailable(flags Flags, bindToDevice tcpip.NICID, dst destination) bool { flagBits := flags.Bits() if bindToDevice == 0 { - // Trying to binding all devices. - if flagBits == 0 { - // Can't bind because the (addr,port) is already bound. - return false - } intersection := FlagMask for _, p := range d { - i := p.IntersectionRefs() + i, c := p.intersectionRefs(dst) + if c == 0 { + continue + } intersection &= i if intersection&flagBits == 0 { // Can't bind because the (addr,port) was @@ -210,16 +267,17 @@ func (d deviceNode) isAvailable(flags Flags, bindToDevice tcpip.NICID) bool { intersection := FlagMask if p, ok := d[0]; ok { - intersection = p.IntersectionRefs() - if intersection&flagBits == 0 { + var c int + intersection, c = p.intersectionRefs(dst) + if c > 0 && intersection&flagBits == 0 { return false } } if p, ok := d[bindToDevice]; ok { - i := p.IntersectionRefs() + i, c := p.intersectionRefs(dst) intersection &= i - if intersection&flagBits == 0 { + if c > 0 && intersection&flagBits == 0 { return false } } @@ -233,12 +291,12 @@ 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, flags Flags, bindToDevice tcpip.NICID) bool { +func (b bindAddresses) isAvailable(addr tcpip.Address, flags Flags, bindToDevice tcpip.NICID, dst destination) 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(flags, bindToDevice) { + if !d.isAvailable(flags, bindToDevice, dst) { return false } } @@ -247,14 +305,14 @@ func (b bindAddresses) isAvailable(addr tcpip.Address, flags Flags, bindToDevice // Check that there is no conflict with the "any" address. if d, ok := b[anyIPAddress]; ok { - if !d.isAvailable(flags, bindToDevice) { + if !d.isAvailable(flags, bindToDevice, dst) { return false } } // Check that this is no conflict with the provided address. if d, ok := b[addr]; ok { - if !d.isAvailable(flags, bindToDevice) { + if !d.isAvailable(flags, bindToDevice, dst) { return false } } @@ -320,17 +378,17 @@ 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, flags Flags, bindToDevice tcpip.NICID) bool { +func (s *PortManager) IsPortAvailable(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID, dest tcpip.FullAddress) bool { s.mu.Lock() defer s.mu.Unlock() - return s.isPortAvailableLocked(networks, transport, addr, port, flags, bindToDevice) + return s.isPortAvailableLocked(networks, transport, addr, port, flags, bindToDevice, makeDestination(dest)) } -func (s *PortManager) isPortAvailableLocked(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID) bool { +func (s *PortManager) isPortAvailableLocked(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID, dst destination) bool { for _, network := range networks { desc := portDescriptor{network, transport, port} if addrs, ok := s.allocatedPorts[desc]; ok { - if !addrs.isAvailable(addr, flags, bindToDevice) { + if !addrs.isAvailable(addr, flags, bindToDevice, dst) { return false } } @@ -342,14 +400,16 @@ 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, flags Flags, 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, dest tcpip.FullAddress) (reservedPort uint16, err *tcpip.Error) { s.mu.Lock() defer s.mu.Unlock() + dst := makeDestination(dest) + // If a port is specified, just try to reserve it for all network // protocols. if port != 0 { - if !s.reserveSpecificPort(networks, transport, addr, port, flags, bindToDevice) { + if !s.reserveSpecificPort(networks, transport, addr, port, flags, bindToDevice, dst) { return 0, tcpip.ErrPortInUse } return port, nil @@ -357,15 +417,16 @@ 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, flags, bindToDevice), nil + return s.reserveSpecificPort(networks, transport, addr, p, flags, bindToDevice, dst), 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, flags Flags, bindToDevice tcpip.NICID) bool { - if !s.isPortAvailableLocked(networks, transport, addr, port, flags, bindToDevice) { +func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID, dst destination) bool { + if !s.isPortAvailableLocked(networks, transport, addr, port, flags, bindToDevice, dst) { return false } + flagBits := flags.Bits() // Reserve port on all network protocols. @@ -381,9 +442,65 @@ func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber d = make(deviceNode) m[addr] = d } - n := d[bindToDevice] + p := d[bindToDevice] + if p == nil { + p = make(portNode) + } + n := p[dst] n.AddRef(flagBits) - d[bindToDevice] = n + p[dst] = n + d[bindToDevice] = p + } + + return true +} + +// ReserveTuple adds a port reservation for the tuple on all given protocol. +func (s *PortManager) ReserveTuple(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID, dest tcpip.FullAddress) bool { + flagBits := flags.Bits() + dst := makeDestination(dest) + + s.mu.Lock() + defer s.mu.Unlock() + + // It is easier to undo the entire reservation, so if we find that the + // tuple can't be fully added, finish and undo the whole thing. + undo := false + + // Reserve port on all network protocols. + for _, network := range networks { + desc := portDescriptor{network, transport, port} + m, ok := s.allocatedPorts[desc] + if !ok { + m = make(bindAddresses) + s.allocatedPorts[desc] = m + } + d, ok := m[addr] + if !ok { + d = make(deviceNode) + m[addr] = d + } + p := d[bindToDevice] + if p == nil { + p = make(portNode) + } + + n := p[dst] + if n.TotalRefs() != 0 && n.IntersectionRefs()&flagBits == 0 { + // Tuple already exists. + undo = true + } + n.AddRef(flagBits) + p[dst] = n + d[bindToDevice] = p + } + + if undo { + // releasePortLocked decrements the counts (rather than setting + // them to zero), so it will undo the incorrect incrementing + // above. + s.releasePortLocked(networks, transport, addr, port, flagBits, bindToDevice, dst) + return false } return true @@ -391,12 +508,14 @@ 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, flags Flags, bindToDevice tcpip.NICID) { +func (s *PortManager) ReleasePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags Flags, bindToDevice tcpip.NICID, dest tcpip.FullAddress) { s.mu.Lock() defer s.mu.Unlock() - flagBits := flags.Bits() + s.releasePortLocked(networks, transport, addr, port, flags.Bits(), bindToDevice, makeDestination(dest)) +} +func (s *PortManager) releasePortLocked(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, flags BitFlags, bindToDevice tcpip.NICID, dst destination) { for _, network := range networks { desc := portDescriptor{network, transport, port} if m, ok := s.allocatedPorts[desc]; ok { @@ -404,21 +523,32 @@ func (s *PortManager) ReleasePort(networks []tcpip.NetworkProtocolNumber, transp if !ok { continue } - n, ok := d[bindToDevice] + p, ok := d[bindToDevice] if !ok { continue } - n.refs[flagBits]-- - d[bindToDevice] = n - if n.TotalRefs() == 0 { - delete(d, bindToDevice) + n, ok := p[dst] + if !ok { + continue + } + n.DropRef(flags) + if n.TotalRefs() > 0 { + p[dst] = n + continue } - if len(d) == 0 { - delete(m, addr) + delete(p, dst) + if len(p) > 0 { + continue + } + delete(d, bindToDevice) + if len(d) > 0 { + continue } - if len(m) == 0 { - delete(s.allocatedPorts, desc) + delete(m, addr) + if len(m) > 0 { + continue } + delete(s.allocatedPorts, desc) } } } diff --git a/pkg/tcpip/ports/ports_test.go b/pkg/tcpip/ports/ports_test.go index d6969d050..58db5868c 100644 --- a/pkg/tcpip/ports/ports_test.go +++ b/pkg/tcpip/ports/ports_test.go @@ -36,6 +36,7 @@ type portReserveTestAction struct { flags Flags release bool device tcpip.NICID + dest tcpip.FullAddress } func TestPortReservation(t *testing.T) { @@ -272,6 +273,54 @@ func TestPortReservation(t *testing.T) { {port: 24, ip: fakeIPAddress, flags: Flags{MostRecent: true, LoadBalanced: true}, want: nil}, {port: 24, ip: fakeIPAddress, flags: Flags{MostRecent: true}, want: tcpip.ErrPortInUse}, }, + }, { + tname: "bind tuple with reuseaddr, and then wildcard with reuseaddr", + actions: []portReserveTestAction{ + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 24}, want: nil}, + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{}, want: nil}, + }, + }, { + tname: "bind tuple with reuseaddr, and then wildcard", + actions: []portReserveTestAction{ + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 24}, want: nil}, + {port: 24, ip: fakeIPAddress, want: tcpip.ErrPortInUse}, + }, + }, { + tname: "bind wildcard with reuseaddr, and then tuple with reuseaddr", + actions: []portReserveTestAction{ + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{}, want: nil}, + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 24}, want: nil}, + }, + }, { + tname: "bind tuple with reuseaddr, and then wildcard", + actions: []portReserveTestAction{ + {port: 24, ip: fakeIPAddress, want: nil}, + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 24}, want: tcpip.ErrPortInUse}, + }, + }, { + tname: "bind two tuples with reuseaddr", + actions: []portReserveTestAction{ + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 24}, want: nil}, + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 25}, want: nil}, + }, + }, { + tname: "bind two tuples", + actions: []portReserveTestAction{ + {port: 24, ip: fakeIPAddress, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 24}, want: nil}, + {port: 24, ip: fakeIPAddress, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 25}, want: nil}, + }, + }, { + tname: "bind wildcard, and then tuple with reuseaddr", + actions: []portReserveTestAction{ + {port: 24, ip: fakeIPAddress, dest: tcpip.FullAddress{}, want: nil}, + {port: 24, ip: fakeIPAddress, flags: Flags{TupleOnly: true}, dest: tcpip.FullAddress{Addr: fakeIPAddress, Port: 24}, want: tcpip.ErrPortInUse}, + }, + }, { + tname: "bind wildcard twice with reuseaddr", + actions: []portReserveTestAction{ + {port: 24, ip: anyIPAddress, flags: Flags{TupleOnly: true}, want: nil}, + {port: 24, ip: anyIPAddress, flags: Flags{TupleOnly: true}, want: nil}, + }, }, } { t.Run(test.tname, func(t *testing.T) { @@ -280,19 +329,18 @@ func TestPortReservation(t *testing.T) { for _, test := range test.actions { if test.release { - pm.ReleasePort(net, fakeTransNumber, test.ip, test.port, test.flags, test.device) + pm.ReleasePort(net, fakeTransNumber, test.ip, test.port, test.flags, test.device, test.dest) continue } - gotPort, err := pm.ReservePort(net, fakeTransNumber, test.ip, test.port, test.flags, test.device) + gotPort, err := pm.ReservePort(net, fakeTransNumber, test.ip, test.port, test.flags, test.device, test.dest) if err != test.want { - t.Fatalf("ReservePort(.., .., %s, %d, %+v, %d) = %v, want %v", test.ip, test.port, test.flags, test.device, err, test.want) + t.Fatalf("ReservePort(.., .., %s, %d, %+v, %d, %v) = %v, want %v", test.ip, test.port, test.flags, test.device, test.dest, err, test.want) } if test.port == 0 && (gotPort == 0 || gotPort < FirstEphemeral) { - t.Fatalf("ReservePort(.., .., .., 0) = %d, want port number >= %d to be picked", gotPort, FirstEphemeral) + t.Fatalf("ReservePort(.., .., .., 0, ..) = %d, want port number >= %d to be picked", gotPort, FirstEphemeral) } } }) - } } diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go index 51abe32a7..e92ec0c24 100644 --- a/pkg/tcpip/stack/stack.go +++ b/pkg/tcpip/stack/stack.go @@ -1408,6 +1408,12 @@ func (s *Stack) RegisterTransportEndpoint(nicID tcpip.NICID, netProtos []tcpip.N return s.demux.registerEndpoint(netProtos, protocol, id, ep, flags, bindToDevice) } +// CheckRegisterTransportEndpoint checks if an endpoint can be registered with +// the stack transport dispatcher. +func (s *Stack) CheckRegisterTransportEndpoint(nicID tcpip.NICID, netProtos []tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, flags ports.Flags, bindToDevice tcpip.NICID) *tcpip.Error { + return s.demux.checkEndpoint(netProtos, protocol, id, flags, bindToDevice) +} + // UnregisterTransportEndpoint removes the endpoint with the given id from the // stack transport dispatcher. func (s *Stack) UnregisterTransportEndpoint(nicID tcpip.NICID, netProtos []tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, ep TransportEndpoint, flags ports.Flags, bindToDevice tcpip.NICID) { diff --git a/pkg/tcpip/stack/transport_demuxer.go b/pkg/tcpip/stack/transport_demuxer.go index 118b449d5..b902c6ca9 100644 --- a/pkg/tcpip/stack/transport_demuxer.go +++ b/pkg/tcpip/stack/transport_demuxer.go @@ -221,6 +221,18 @@ func (epsByNIC *endpointsByNIC) registerEndpoint(d *transportDemuxer, netProto t return multiPortEp.singleRegisterEndpoint(t, flags) } +func (epsByNIC *endpointsByNIC) checkEndpoint(d *transportDemuxer, netProto tcpip.NetworkProtocolNumber, transProto tcpip.TransportProtocolNumber, flags ports.Flags, bindToDevice tcpip.NICID) *tcpip.Error { + epsByNIC.mu.RLock() + defer epsByNIC.mu.RUnlock() + + multiPortEp, ok := epsByNIC.endpoints[bindToDevice] + if !ok { + return nil + } + + return multiPortEp.singleCheckEndpoint(flags) +} + // unregisterEndpoint returns true if endpointsByNIC has to be unregistered. func (epsByNIC *endpointsByNIC) unregisterEndpoint(bindToDevice tcpip.NICID, t TransportEndpoint, flags ports.Flags) bool { epsByNIC.mu.Lock() @@ -289,6 +301,17 @@ func (d *transportDemuxer) registerEndpoint(netProtos []tcpip.NetworkProtocolNum return nil } +// checkEndpoint checks if an endpoint can be registered with the dispatcher. +func (d *transportDemuxer) checkEndpoint(netProtos []tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, flags ports.Flags, bindToDevice tcpip.NICID) *tcpip.Error { + for _, n := range netProtos { + if err := d.singleCheckEndpoint(n, protocol, id, flags, bindToDevice); err != nil { + return err + } + } + + return nil +} + // multiPortEndpoint is a container for TransportEndpoints which are bound to // the same pair of address and port. endpointsArr always has at least one // element. @@ -380,7 +403,7 @@ func (ep *multiPortEndpoint) singleRegisterEndpoint(t TransportEndpoint, flags p ep.mu.Lock() defer ep.mu.Unlock() - bits := flags.Bits() + bits := flags.Bits() & ports.MultiBindFlagMask if len(ep.endpoints) != 0 { // If it was previously bound, we need to check if we can bind again. @@ -395,6 +418,22 @@ func (ep *multiPortEndpoint) singleRegisterEndpoint(t TransportEndpoint, flags p return nil } +func (ep *multiPortEndpoint) singleCheckEndpoint(flags ports.Flags) *tcpip.Error { + ep.mu.RLock() + defer ep.mu.RUnlock() + + bits := flags.Bits() & ports.MultiBindFlagMask + + if len(ep.endpoints) != 0 { + // If it was previously bound, we need to check if we can bind again. + if ep.flags.TotalRefs() > 0 && bits&ep.flags.IntersectionRefs() == 0 { + return tcpip.ErrPortInUse + } + } + + return nil +} + // unregisterEndpoint returns true if multiPortEndpoint has to be unregistered. func (ep *multiPortEndpoint) unregisterEndpoint(t TransportEndpoint, flags ports.Flags) bool { ep.mu.Lock() @@ -406,7 +445,7 @@ func (ep *multiPortEndpoint) unregisterEndpoint(t TransportEndpoint, flags ports ep.endpoints[len(ep.endpoints)-1] = nil ep.endpoints = ep.endpoints[:len(ep.endpoints)-1] - ep.flags.DropRef(flags.Bits()) + ep.flags.DropRef(flags.Bits() & ports.MultiBindFlagMask) break } } @@ -439,6 +478,28 @@ func (d *transportDemuxer) singleRegisterEndpoint(netProto tcpip.NetworkProtocol return epsByNIC.registerEndpoint(d, netProto, protocol, ep, flags, bindToDevice) } +func (d *transportDemuxer) singleCheckEndpoint(netProto tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, flags ports.Flags, bindToDevice tcpip.NICID) *tcpip.Error { + if id.RemotePort != 0 { + // SO_REUSEPORT only applies to bound/listening endpoints. + flags.LoadBalanced = false + } + + eps, ok := d.protocol[protocolIDs{netProto, protocol}] + if !ok { + return tcpip.ErrUnknownProtocol + } + + eps.mu.RLock() + defer eps.mu.RUnlock() + + epsByNIC, ok := eps.endpoints[id] + if !ok { + return nil + } + + return epsByNIC.checkEndpoint(d, netProto, protocol, flags, bindToDevice) +} + // unregisterEndpoint unregisters the endpoint with the given id such that it // won't receive any more packets. func (d *transportDemuxer) unregisterEndpoint(netProtos []tcpip.NetworkProtocolNumber, protocol tcpip.TransportProtocolNumber, id TransportEndpointID, ep TransportEndpoint, flags ports.Flags, bindToDevice tcpip.NICID) { diff --git a/pkg/tcpip/transport/tcp/accept.go b/pkg/tcpip/transport/tcp/accept.go index 7679fe169..6e00e5526 100644 --- a/pkg/tcpip/transport/tcp/accept.go +++ b/pkg/tcpip/transport/tcp/accept.go @@ -27,7 +27,6 @@ import ( "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" - "gvisor.dev/gvisor/pkg/tcpip/ports" "gvisor.dev/gvisor/pkg/tcpip/seqnum" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/waiter" @@ -199,9 +198,8 @@ func (l *listenContext) isCookieValid(id stack.TransportEndpointID, cookie seqnu } // createConnectingEndpoint creates a new endpoint in a connecting state, with -// the connection parameters given by the arguments. The endpoint is returned -// with n.mu held. -func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, irs seqnum.Value, rcvdSynOpts *header.TCPSynOptions, queue *waiter.Queue) (*endpoint, *tcpip.Error) { +// the connection parameters given by the arguments. +func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, irs seqnum.Value, rcvdSynOpts *header.TCPSynOptions, queue *waiter.Queue) *endpoint { // Create a new endpoint. netProto := l.netProto if netProto == 0 { @@ -227,22 +225,7 @@ func (l *listenContext) createConnectingEndpoint(s *segment, iss seqnum.Value, i // window to grow to a really large value. n.rcvAutoParams.prevCopied = n.initialReceiveWindow() - // Lock the endpoint before registering to ensure that no out of - // band changes are possible due to incoming packets etc till - // the endpoint is done initializing. - n.mu.Lock() - - // Register new endpoint so that packets are routed to it. - if err := n.stack.RegisterTransportEndpoint(n.boundNICID, n.effectiveNetProtos, ProtocolNumber, n.ID, n, ports.Flags{LoadBalanced: n.reusePort}, n.boundBindToDevice); err != nil { - n.mu.Unlock() - n.Close() - return nil, err - } - - n.isRegistered = true - n.registeredReusePort = n.reusePort - - return n, nil + return n } // createEndpointAndPerformHandshake creates a new endpoint in connected state @@ -253,10 +236,12 @@ func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *head // Create new endpoint. irs := s.sequenceNumber isn := generateSecureISN(s.id, l.stack.Seed()) - ep, err := l.createConnectingEndpoint(s, isn, irs, opts, queue) - if err != nil { - return nil, err - } + ep := l.createConnectingEndpoint(s, isn, irs, opts, queue) + + // Lock the endpoint before registering to ensure that no out of + // band changes are possible due to incoming packets etc till + // the endpoint is done initializing. + ep.mu.Lock() ep.owner = owner // listenEP is nil when listenContext is used by tcp.Forwarder. @@ -264,18 +249,13 @@ func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *head if l.listenEP != nil { l.listenEP.mu.Lock() if l.listenEP.EndpointState() != StateListen { + l.listenEP.mu.Unlock() // Ensure we release any registrations done by the newly // created endpoint. ep.mu.Unlock() ep.Close() - // Wake up any waiters. This is strictly not required normally - // as a socket that was never accepted can't really have any - // registered waiters except when stack.Wait() is called which - // waits for all registered endpoints to stop and expects an - // EventHUp. - ep.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) return nil, tcpip.ErrConnectionAborted } l.addPendingEndpoint(ep) @@ -284,21 +264,44 @@ func (l *listenContext) createEndpointAndPerformHandshake(s *segment, opts *head // to the newly created endpoint. l.listenEP.propagateInheritableOptionsLocked(ep) + if !ep.reserveTupleLocked() { + ep.mu.Unlock() + ep.Close() + + if l.listenEP != nil { + l.removePendingEndpoint(ep) + l.listenEP.mu.Unlock() + } + + return nil, tcpip.ErrConnectionAborted + } + deferAccept = l.listenEP.deferAccept l.listenEP.mu.Unlock() } + // Register new endpoint so that packets are routed to it. + if err := ep.stack.RegisterTransportEndpoint(ep.boundNICID, ep.effectiveNetProtos, ProtocolNumber, ep.ID, ep, ep.boundPortFlags, ep.boundBindToDevice); err != nil { + ep.mu.Unlock() + ep.Close() + + if l.listenEP != nil { + l.removePendingEndpoint(ep) + } + + ep.drainClosingSegmentQueue() + + return nil, err + } + + ep.isRegistered = true + // Perform the 3-way handshake. h := newPassiveHandshake(ep, seqnum.Size(ep.initialReceiveWindow()), isn, irs, opts, deferAccept) if err := h.execute(); err != nil { ep.mu.Unlock() ep.Close() - // Wake up any waiters. This is strictly not required normally - // as a socket that was never accepted can't really have any - // registered waiters except when stack.Wait() is called which - // waits for all registered endpoints to stop and expects an - // EventHUp. - ep.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) + ep.notifyAborted() if l.listenEP != nil { l.removePendingEndpoint(ep) @@ -374,6 +377,43 @@ func (e *endpoint) deliverAccepted(n *endpoint) { // Precondition: e.mu and n.mu must be held. func (e *endpoint) propagateInheritableOptionsLocked(n *endpoint) { n.userTimeout = e.userTimeout + n.portFlags = e.portFlags + n.boundBindToDevice = e.boundBindToDevice + n.boundPortFlags = e.boundPortFlags +} + +// reserveTupleLocked reserves an accepted endpoint's tuple. +// +// Preconditions: +// * propagateInheritableOptionsLocked has been called. +// * e.mu is held. +func (e *endpoint) reserveTupleLocked() bool { + dest := tcpip.FullAddress{Addr: e.ID.RemoteAddress, Port: e.ID.RemotePort} + if !e.stack.ReserveTuple( + e.effectiveNetProtos, + ProtocolNumber, + e.ID.LocalAddress, + e.ID.LocalPort, + e.boundPortFlags, + e.boundBindToDevice, + dest, + ) { + return false + } + + e.isPortReserved = true + e.boundDest = dest + return true +} + +// notifyAborted wakes up any waiters on registered, but not accepted +// endpoints. +// +// This is strictly not required normally as a socket that was never accepted +// can't really have any registered waiters except when stack.Wait() is called +// which waits for all registered endpoints to stop and expects an EventHUp. +func (e *endpoint) notifyAborted() { + e.waiterQueue.Notify(waiter.EventHUp | waiter.EventErr | waiter.EventIn | waiter.EventOut) } // handleSynSegment is called in its own goroutine once the listening endpoint @@ -568,16 +608,34 @@ func (e *endpoint) handleListenSegment(ctx *listenContext, s *segment) { rcvdSynOptions.TSEcr = s.parsedOptions.TSEcr } - n, err := ctx.createConnectingEndpoint(s, iss, irs, rcvdSynOptions, &waiter.Queue{}) - if err != nil { + n := ctx.createConnectingEndpoint(s, iss, irs, rcvdSynOptions, &waiter.Queue{}) + + n.mu.Lock() + + // Propagate any inheritable options from the listening endpoint + // to the newly created endpoint. + e.propagateInheritableOptionsLocked(n) + + if !n.reserveTupleLocked() { + n.mu.Unlock() + n.Close() + e.stack.Stats().TCP.FailedConnectionAttempts.Increment() e.stats.FailedConnectionAttempts.Increment() return } - // Propagate any inheritable options from the listening endpoint - // to the newly created endpoint. - e.propagateInheritableOptionsLocked(n) + // Register new endpoint so that packets are routed to it. + if err := n.stack.RegisterTransportEndpoint(n.boundNICID, n.effectiveNetProtos, ProtocolNumber, n.ID, n, n.boundPortFlags, n.boundBindToDevice); err != nil { + n.mu.Unlock() + n.Close() + + e.stack.Stats().TCP.FailedConnectionAttempts.Increment() + e.stats.FailedConnectionAttempts.Increment() + return + } + + n.isRegistered = true // clear the tsOffset for the newly created // endpoint as the Timestamp was already diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go index 10df2bcd5..1e4c2f507 100644 --- a/pkg/tcpip/transport/tcp/endpoint.go +++ b/pkg/tcpip/transport/tcp/endpoint.go @@ -396,7 +396,8 @@ type endpoint struct { mu sync.Mutex `state:"nosave"` ownedByUser uint32 - // state must be read/set using the EndpointState()/setEndpointState() methods. + // state must be read/set using the EndpointState()/setEndpointState() + // methods. state EndpointState `state:".(EndpointState)"` // origEndpointState is only used during a restore phase to save the @@ -405,8 +406,8 @@ type endpoint struct { origEndpointState EndpointState `state:"nosave"` isPortReserved bool `state:"manual"` - isRegistered bool - boundNICID tcpip.NICID `state:"manual"` + isRegistered bool `state:"manual"` + boundNICID tcpip.NICID route stack.Route `state:"manual"` ttl uint8 v6only bool @@ -415,10 +416,14 @@ type endpoint struct { // disabling SO_BROADCAST, albeit as a NOOP. broadcast bool + // portFlags stores the current values of port related flags. + portFlags ports.Flags + // Values used to reserve a port or register a transport endpoint // (which ever happens first). boundBindToDevice tcpip.NICID boundPortFlags ports.Flags + boundDest tcpip.FullAddress // effectiveNetProtos contains the network protocols actually in use. In // most cases it will only contain "netProto", but in cases like IPv6 @@ -426,7 +431,7 @@ type endpoint struct { // protocols (e.g., IPv6 and IPv4) or a single different protocol (e.g., // IPv4 when IPv6 endpoint is bound or connected to an IPv4 mapped // address). - effectiveNetProtos []tcpip.NetworkProtocolNumber `state:"manual"` + effectiveNetProtos []tcpip.NetworkProtocolNumber // workerRunning specifies if a worker goroutine is running. workerRunning bool @@ -462,13 +467,6 @@ type endpoint struct { // sack holds TCP SACK related information for this endpoint. sack SACKInfo - // reusePort is set to true if SO_REUSEPORT is enabled. - reusePort bool - - // registeredReusePort is set if the current endpoint registration was - // done with SO_REUSEPORT enabled. - registeredReusePort bool - // bindToDevice is set to the NIC on which to bind or disabled if 0. bindToDevice tcpip.NICID @@ -488,7 +486,6 @@ type endpoint struct { // The options below aren't implemented, but we remember the user // settings because applications expect to be able to set/query these // options. - reuseAddr bool // slowAck holds the negated state of quick ack. It is stubbed out and // does nothing. @@ -838,7 +835,6 @@ func newEndpoint(s *stack.Stack, netProto tcpip.NetworkProtocolNumber, waiterQue rcvBufSize: DefaultReceiveBufferSize, sndBufSize: DefaultSendBufferSize, sndMTU: int(math.MaxInt32), - reuseAddr: true, keepalive: keepalive{ // Linux defaults. idle: 2 * time.Hour, @@ -1025,15 +1021,15 @@ func (e *endpoint) closeNoShutdownLocked() { // in Listen() when trying to register. if e.EndpointState() == StateListen && e.isPortReserved { if e.isRegistered { - e.stack.StartTransportEndpointCleanup(e.boundNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e, ports.Flags{LoadBalanced: e.registeredReusePort}, e.boundBindToDevice) + e.stack.StartTransportEndpointCleanup(e.boundNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e, e.boundPortFlags, e.boundBindToDevice) e.isRegistered = false - e.registeredReusePort = false } - e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice, e.boundDest) e.isPortReserved = false e.boundBindToDevice = 0 e.boundPortFlags = ports.Flags{} + e.boundDest = tcpip.FullAddress{} } // Mark endpoint as closed. @@ -1091,17 +1087,17 @@ func (e *endpoint) cleanupLocked() { e.workerCleanup = false if e.isRegistered { - e.stack.StartTransportEndpointCleanup(e.boundNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e, ports.Flags{LoadBalanced: e.registeredReusePort}, e.boundBindToDevice) + e.stack.StartTransportEndpointCleanup(e.boundNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e, e.boundPortFlags, e.boundBindToDevice) e.isRegistered = false - e.registeredReusePort = false } if e.isPortReserved { - e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice, e.boundDest) e.isPortReserved = false } e.boundBindToDevice = 0 e.boundPortFlags = ports.Flags{} + e.boundDest = tcpip.FullAddress{} e.route.Release() e.stack.CompleteTransportEndpointCleanup(e) @@ -1522,12 +1518,12 @@ func (e *endpoint) SetSockOptBool(opt tcpip.SockOptBool, v bool) *tcpip.Error { case tcpip.ReuseAddressOption: e.LockUser() - e.reuseAddr = v + e.portFlags.TupleOnly = v e.UnlockUser() case tcpip.ReusePortOption: e.LockUser() - e.reusePort = v + e.portFlags.LoadBalanced = v e.UnlockUser() case tcpip.V6OnlyOption: @@ -1831,14 +1827,14 @@ func (e *endpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { case tcpip.ReuseAddressOption: e.LockUser() - v := e.reuseAddr + v := e.portFlags.TupleOnly e.UnlockUser() return v, nil case tcpip.ReusePortOption: e.LockUser() - v := e.reusePort + v := e.portFlags.LoadBalanced e.UnlockUser() return v, nil @@ -2091,8 +2087,6 @@ func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) *tc } defer r.Release() - origID := e.ID - netProtos := []tcpip.NetworkProtocolNumber{netProto} e.ID.LocalAddress = r.LocalAddress e.ID.RemoteAddress = r.RemoteAddress @@ -2100,11 +2094,10 @@ func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) *tc if e.ID.LocalPort != 0 { // The endpoint is bound to a port, attempt to register it. - err := e.stack.RegisterTransportEndpoint(nicID, netProtos, ProtocolNumber, e.ID, e, ports.Flags{LoadBalanced: e.reusePort}, e.boundBindToDevice) + err := e.stack.RegisterTransportEndpoint(nicID, netProtos, ProtocolNumber, e.ID, e, e.boundPortFlags, e.boundBindToDevice) if err != nil { return err } - e.registeredReusePort = e.reusePort } else { // The endpoint doesn't have a local port yet, so try to get // one. Make sure that it isn't one that will result in the same @@ -2128,40 +2121,33 @@ func (e *endpoint) connect(addr tcpip.FullAddress, handshake bool, run bool) *tc if sameAddr && p == e.ID.RemotePort { return false, nil } - // 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, ports.Flags{LoadBalanced: false}, e.bindToDevice) { + if _, err := e.stack.ReservePort(netProtos, ProtocolNumber, e.ID.LocalAddress, p, e.portFlags, e.bindToDevice, addr); err != nil { return false, nil } id := e.ID id.LocalPort = p - switch e.stack.RegisterTransportEndpoint(nicID, netProtos, ProtocolNumber, id, e, ports.Flags{LoadBalanced: e.reusePort}, e.bindToDevice) { - case nil: - // Port picking successful. Save the details of - // the selected port. - e.ID = id - e.boundBindToDevice = e.bindToDevice - e.registeredReusePort = e.reusePort - return true, nil - case tcpip.ErrPortInUse: - return false, nil - default: + if err := e.stack.RegisterTransportEndpoint(nicID, netProtos, ProtocolNumber, id, e, e.portFlags, e.bindToDevice); err != nil { + e.stack.ReleasePort(netProtos, ProtocolNumber, e.ID.LocalAddress, p, e.portFlags, e.bindToDevice, addr) + if err == tcpip.ErrPortInUse { + return false, nil + } return false, err } + + // Port picking successful. Save the details of + // the selected port. + e.ID = id + e.isPortReserved = true + e.boundBindToDevice = e.bindToDevice + e.boundPortFlags = e.portFlags + e.boundDest = addr + return true, nil }); err != nil { return err } } - // Remove the port reservation. This can happen when Bind is called - // 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.boundPortFlags, e.boundBindToDevice) - e.isPortReserved = false - } - e.isRegistered = true e.setEndpointState(StateConnecting) e.route = r.Clone() @@ -2340,13 +2326,12 @@ func (e *endpoint) listen(backlog int) *tcpip.Error { } // Register the endpoint. - if err := e.stack.RegisterTransportEndpoint(e.boundNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e, ports.Flags{LoadBalanced: e.reusePort}, e.boundBindToDevice); err != nil { + if err := e.stack.RegisterTransportEndpoint(e.boundNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e, e.boundPortFlags, e.boundBindToDevice); err != nil { return err } e.isRegistered = true e.setEndpointState(StateListen) - e.registeredReusePort = e.reusePort // The channel may be non-nil when we're restoring the endpoint, and it // may be pre-populated with some previously accepted (but not Accepted) @@ -2433,16 +2418,13 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) (err *tcpip.Error) { } } - flags := ports.Flags{ - LoadBalanced: e.reusePort, - } - port, err := e.stack.ReservePort(netProtos, ProtocolNumber, addr.Addr, addr.Port, flags, e.bindToDevice) + port, err := e.stack.ReservePort(netProtos, ProtocolNumber, addr.Addr, addr.Port, e.portFlags, e.bindToDevice, tcpip.FullAddress{}) if err != nil { return err } e.boundBindToDevice = e.bindToDevice - e.boundPortFlags = flags + e.boundPortFlags = e.portFlags e.isPortReserved = true e.effectiveNetProtos = netProtos e.ID.LocalPort = port @@ -2450,7 +2432,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) (err *tcpip.Error) { // Any failures beyond this point must remove the port registration. defer func(portFlags ports.Flags, bindToDevice tcpip.NICID) { if err != nil { - e.stack.ReleasePort(netProtos, ProtocolNumber, addr.Addr, port, portFlags, bindToDevice) + e.stack.ReleasePort(netProtos, ProtocolNumber, addr.Addr, port, portFlags, bindToDevice, tcpip.FullAddress{}) e.isPortReserved = false e.effectiveNetProtos = nil e.ID.LocalPort = 0 @@ -2473,6 +2455,10 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) (err *tcpip.Error) { e.ID.LocalAddress = addr.Addr } + if err := e.stack.CheckRegisterTransportEndpoint(e.boundNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e.boundPortFlags, e.boundBindToDevice); err != nil { + return err + } + // Mark endpoint as bound. e.setEndpointState(StateBound) diff --git a/pkg/tcpip/transport/tcp/endpoint_state.go b/pkg/tcpip/transport/tcp/endpoint_state.go index 0bebec2d1..8258c0ecc 100644 --- a/pkg/tcpip/transport/tcp/endpoint_state.go +++ b/pkg/tcpip/transport/tcp/endpoint_state.go @@ -93,10 +93,6 @@ func (e *endpoint) beforeSave() { if e.waiterQueue != nil && !e.waiterQueue.IsEmpty() { panic("endpoint still has waiters upon save") } - - if e.EndpointState() != StateClose && !((e.EndpointState() == StateBound || e.EndpointState() == StateListen) == e.isPortReserved) { - panic("endpoints which are not in the closed state must have a reserved port IFF they are in bound or listen state") - } } // saveAcceptedChan is invoked by stateify. @@ -198,14 +194,17 @@ func (e *endpoint) Resume(s *stack.Stack) { } bind := func() { - if len(e.BindAddr) == 0 { - e.BindAddr = e.ID.LocalAddress + addr, _, err := e.checkV4MappedLocked(tcpip.FullAddress{Addr: e.BindAddr, Port: e.ID.LocalPort}) + if err != nil { + panic("unable to parse BindAddr: " + err.String()) } - addr := e.BindAddr - port := e.ID.LocalPort - if err := e.Bind(tcpip.FullAddress{Addr: addr, Port: port}); err != nil { - panic(fmt.Sprintf("endpoint binding [%v]:%d failed: %v", addr, port, err)) + if ok := e.stack.ReserveTuple(e.effectiveNetProtos, ProtocolNumber, addr.Addr, addr.Port, e.boundPortFlags, e.boundBindToDevice, e.boundDest); !ok { + panic(fmt.Sprintf("unable to re-reserve tuple (%v, %q, %d, %+v, %d, %v)", e.effectiveNetProtos, addr.Addr, addr.Port, e.boundPortFlags, e.boundBindToDevice, e.boundDest)) } + e.isPortReserved = true + + // Mark endpoint as bound. + e.setEndpointState(StateBound) } switch { @@ -277,17 +276,7 @@ func (e *endpoint) Resume(s *stack.Stack) { tcpip.AsyncLoading.Done() }() case epState == StateClose: - if e.isPortReserved { - tcpip.AsyncLoading.Add(1) - go func() { - connectedLoading.Wait() - listenLoading.Wait() - connectingLoading.Wait() - bind() - e.setEndpointState(StateClose) - tcpip.AsyncLoading.Done() - }() - } + e.isPortReserved = false e.state = StateClose e.stack.CompleteTransportEndpointCleanup(e) tcpip.DeleteDanglingEndpoint(e) diff --git a/pkg/tcpip/transport/tcp/tcp_test.go b/pkg/tcpip/transport/tcp/tcp_test.go index aca6a7951..2632a3c67 100644 --- a/pkg/tcpip/transport/tcp/tcp_test.go +++ b/pkg/tcpip/transport/tcp/tcp_test.go @@ -3879,6 +3879,9 @@ func TestReusePort(t *testing.T) { if err != nil { t.Fatalf("NewEndpoint failed; %s", err) } + if err := c.EP.SetSockOptBool(tcpip.ReuseAddressOption, true); err != nil { + t.Fatalf("SetSockOptBool ReuseAddressOption failed: %s", err) + } if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { t.Fatalf("Bind failed: %s", err) } @@ -3888,6 +3891,9 @@ func TestReusePort(t *testing.T) { if err != nil { t.Fatalf("NewEndpoint failed; %s", err) } + if err := c.EP.SetSockOptBool(tcpip.ReuseAddressOption, true); err != nil { + t.Fatalf("SetSockOptBool ReuseAddressOption failed: %s", err) + } if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { t.Fatalf("Bind failed: %s", err) } @@ -3898,6 +3904,9 @@ func TestReusePort(t *testing.T) { if err != nil { t.Fatalf("NewEndpoint failed; %s", err) } + if err := c.EP.SetSockOptBool(tcpip.ReuseAddressOption, true); err != nil { + t.Fatalf("SetSockOptBool ReuseAddressOption failed: %s", err) + } if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { t.Fatalf("Bind failed: %s", err) } @@ -3910,6 +3919,9 @@ func TestReusePort(t *testing.T) { if err != nil { t.Fatalf("NewEndpoint failed; %s", err) } + if err := c.EP.SetSockOptBool(tcpip.ReuseAddressOption, true); err != nil { + t.Fatalf("SetSockOptBool ReuseAddressOption failed: %s", err) + } if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { t.Fatalf("Bind failed: %s", err) } @@ -3920,6 +3932,9 @@ func TestReusePort(t *testing.T) { if err != nil { t.Fatalf("NewEndpoint failed; %s", err) } + if err := c.EP.SetSockOptBool(tcpip.ReuseAddressOption, true); err != nil { + t.Fatalf("SetSockOptBool ReuseAddressOption failed: %s", err) + } if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { t.Fatalf("Bind failed: %s", err) } @@ -3932,6 +3947,9 @@ func TestReusePort(t *testing.T) { if err != nil { t.Fatalf("NewEndpoint failed; %s", err) } + if err := c.EP.SetSockOptBool(tcpip.ReuseAddressOption, true); err != nil { + t.Fatalf("SetSockOptBool ReuseAddressOption failed: %s", err) + } if err := c.EP.Bind(tcpip.FullAddress{Port: context.StackPort}); err != nil { t.Fatalf("Bind failed: %s", err) } diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go index 40d66ef09..6ea212093 100644 --- a/pkg/tcpip/transport/udp/endpoint.go +++ b/pkg/tcpip/transport/udp/endpoint.go @@ -231,7 +231,7 @@ func (e *endpoint) Close() { switch e.state { case StateBound, StateConnected: e.stack.UnregisterTransportEndpoint(e.RegisterNICID, e.effectiveNetProtos, ProtocolNumber, e.ID, e, e.boundPortFlags, e.boundBindToDevice) - e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, e.boundPortFlags, e.boundBindToDevice, tcpip.FullAddress{}) e.boundBindToDevice = 0 e.boundPortFlags = ports.Flags{} } @@ -1047,7 +1047,7 @@ 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, boundPortFlags, e.boundBindToDevice) + e.stack.ReleasePort(e.effectiveNetProtos, ProtocolNumber, e.ID.LocalAddress, e.ID.LocalPort, boundPortFlags, e.boundBindToDevice, tcpip.FullAddress{}) e.boundPortFlags = ports.Flags{} } e.state = StateInitial @@ -1198,7 +1198,7 @@ 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.portFlags, e.bindToDevice) + port, err := e.stack.ReservePort(netProtos, ProtocolNumber, id.LocalAddress, id.LocalPort, e.portFlags, e.bindToDevice, tcpip.FullAddress{}) if err != nil { return id, e.bindToDevice, err } @@ -1208,7 +1208,7 @@ func (e *endpoint) registerWithStack(nicID tcpip.NICID, netProtos []tcpip.Networ err := e.stack.RegisterTransportEndpoint(nicID, netProtos, ProtocolNumber, id, e, e.boundPortFlags, e.bindToDevice) if err != nil { - e.stack.ReleasePort(netProtos, ProtocolNumber, id.LocalAddress, id.LocalPort, e.boundPortFlags, e.bindToDevice) + e.stack.ReleasePort(netProtos, ProtocolNumber, id.LocalAddress, id.LocalPort, e.boundPortFlags, e.bindToDevice, tcpip.FullAddress{}) e.boundPortFlags = ports.Flags{} } return id, e.bindToDevice, err |