[netstack] Implement MSG_ERRQUEUE flag for recvmsg(2).

Introduces the per-socket error queue and the necessary cmsg mechanisms.

PiperOrigin-RevId: 348028508

pkg/abi/linux/BUILD

@@ -21,6 +21,7 @@ go_library(
         "epoll_amd64.go",
         "epoll_arm64.go",
         "errors.go",
+        "errqueue.go",
         "eventfd.go",
         "exec.go",
         "fadvise.go",

pkg/abi/linux/errqueue.go

@@ -0,0 +1,93 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.dev/gvisor/pkg/marshal"
+)
+
+// Socket error origin codes as defined in include/uapi/linux/errqueue.h.
+const (
+	SO_EE_ORIGIN_NONE  = 0
+	SO_EE_ORIGIN_LOCAL = 1
+	SO_EE_ORIGIN_ICMP  = 2
+	SO_EE_ORIGIN_ICMP6 = 3
+)
+
+// SockExtendedErr represents struct sock_extended_err in Linux defined in
+// include/uapi/linux/errqueue.h.
+//
+// +marshal
+type SockExtendedErr struct {
+	Errno  uint32
+	Origin uint8
+	Type   uint8
+	Code   uint8
+	Pad    uint8
+	Info   uint32
+	Data   uint32
+}
+
+// SockErrCMsg represents the IP*_RECVERR control message.
+type SockErrCMsg interface {
+	marshal.Marshallable
+
+	CMsgLevel() uint32
+	CMsgType() uint32
+}
+
+// SockErrCMsgIPv4 is the IP_RECVERR control message used in
+// recvmsg(MSG_ERRQUEUE) by ipv4 sockets. This is equilavent to `struct errhdr`
+// defined in net/ipv4/ip_sockglue.c:ip_recv_error().
+//
+// +marshal
+type SockErrCMsgIPv4 struct {
+	SockExtendedErr
+	Offender SockAddrInet
+}
+
+var _ SockErrCMsg = (*SockErrCMsgIPv4)(nil)
+
+// CMsgLevel implements SockErrCMsg.CMsgLevel.
+func (*SockErrCMsgIPv4) CMsgLevel() uint32 {
+	return SOL_IP
+}
+
+// CMsgType implements SockErrCMsg.CMsgType.
+func (*SockErrCMsgIPv4) CMsgType() uint32 {
+	return IP_RECVERR
+}
+
+// SockErrCMsgIPv6 is the IPV6_RECVERR control message used in
+// recvmsg(MSG_ERRQUEUE) by ipv6 sockets. This is equilavent to `struct errhdr`
+// defined in net/ipv6/datagram.c:ipv6_recv_error().
+//
+// +marshal
+type SockErrCMsgIPv6 struct {
+	SockExtendedErr
+	Offender SockAddrInet6
+}
+
+var _ SockErrCMsg = (*SockErrCMsgIPv6)(nil)
+
+// CMsgLevel implements SockErrCMsg.CMsgLevel.
+func (*SockErrCMsgIPv6) CMsgLevel() uint32 {
+	return SOL_IPV6
+}
+
+// CMsgType implements SockErrCMsg.CMsgType.
+func (*SockErrCMsgIPv6) CMsgType() uint32 {
+	return IPV6_RECVERR
+}

pkg/sentry/socket/control/control.go

@@ -371,6 +371,17 @@ func PackOriginalDstAddress(t *kernel.Task, originalDstAddress linux.SockAddr, b
 		buf, level, optType, t.Arch().Width(), originalDstAddress)
 }
 
+// PackSockExtendedErr packs an IP*_RECVERR socket control message.
+func PackSockExtendedErr(t *kernel.Task, sockErr linux.SockErrCMsg, buf []byte) []byte {
+	return putCmsgStruct(
+		buf,
+		sockErr.CMsgLevel(),
+		sockErr.CMsgType(),
+		t.Arch().Width(),
+		sockErr,
+	)
+}
+
 // PackControlMessages packs control messages into the given buffer.
 //
 // We skip control messages specific to Unix domain sockets.
@@ -403,6 +414,10 @@ func PackControlMessages(t *kernel.Task, cmsgs socket.ControlMessages, buf []byt
 		buf = PackOriginalDstAddress(t, cmsgs.IP.OriginalDstAddress, buf)
 	}
 
+	if cmsgs.IP.SockErr != nil {
+		buf = PackSockExtendedErr(t, cmsgs.IP.SockErr, buf)
+	}
+
 	return buf
 }
 
@@ -440,6 +455,10 @@ func CmsgsSpace(t *kernel.Task, cmsgs socket.ControlMessages) int {
 		space += cmsgSpace(t, cmsgs.IP.OriginalDstAddress.SizeBytes())
 	}
 
+	if cmsgs.IP.SockErr != nil {
+		space += cmsgSpace(t, cmsgs.IP.SockErr.SizeBytes())
+	}
+
 	return space
 }
 
@@ -546,6 +565,16 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con
 				cmsgs.IP.OriginalDstAddress = &addr
 				i += binary.AlignUp(length, width)
 
+			case linux.IP_RECVERR:
+				var errCmsg linux.SockErrCMsgIPv4
+				if length < errCmsg.SizeBytes() {
+					return socket.ControlMessages{}, syserror.EINVAL
+				}
+
+				errCmsg.UnmarshalBytes(buf[i : i+errCmsg.SizeBytes()])
+				cmsgs.IP.SockErr = &errCmsg
+				i += binary.AlignUp(length, width)
+
 			default:
 				return socket.ControlMessages{}, syserror.EINVAL
 			}
@@ -568,6 +597,16 @@ func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (socket.Con
 				cmsgs.IP.OriginalDstAddress = &addr
 				i += binary.AlignUp(length, width)
 
+			case linux.IPV6_RECVERR:
+				var errCmsg linux.SockErrCMsgIPv6
+				if length < errCmsg.SizeBytes() {
+					return socket.ControlMessages{}, syserror.EINVAL
+				}
+
+				errCmsg.UnmarshalBytes(buf[i : i+errCmsg.SizeBytes()])
+				cmsgs.IP.SockErr = &errCmsg
+				i += binary.AlignUp(length, width)
+
 			default:
 				return socket.ControlMessages{}, syserror.EINVAL
 			}

pkg/sentry/socket/hostinet/socket.go

@@ -450,11 +450,7 @@ func (s *socketOpsCommon) recvMsgFromHost(iovs []syscall.Iovec, flags int, sende
 // RecvMsg implements socket.Socket.RecvMsg.
 func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlLen uint64) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) {
 	// Only allow known and safe flags.
-	//
-	// FIXME(jamieliu): We can't support MSG_ERRQUEUE because it uses ancillary
-	// messages that gvisor/pkg/tcpip/transport/unix doesn't understand. Kill the
-	// Socket interface's dependence on netstack.
-	if flags&^(syscall.MSG_DONTWAIT|syscall.MSG_PEEK|syscall.MSG_TRUNC) != 0 {
+	if flags&^(syscall.MSG_DONTWAIT|syscall.MSG_PEEK|syscall.MSG_TRUNC|syscall.MSG_ERRQUEUE) != 0 {
 		return 0, 0, nil, 0, socket.ControlMessages{}, syserr.ErrInvalidArgument
 	}
 
@@ -488,7 +484,8 @@ func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags
 
 	var ch chan struct{}
 	n, err := copyToDst()
-	if flags&syscall.MSG_DONTWAIT == 0 {
+	// recv*(MSG_ERRQUEUE) never blocks, even without MSG_DONTWAIT.
+	if flags&(syscall.MSG_DONTWAIT|syscall.MSG_ERRQUEUE) == 0 {
 		for err == syserror.ErrWouldBlock {
 			// We only expect blocking to come from the actual syscall, in which
 			// case it can't have returned any data.
@@ -551,6 +548,11 @@ func parseUnixControlMessages(unixControlMessages []unix.SocketControlMessage) s
 				var addr linux.SockAddrInet
 				binary.Unmarshal(unixCmsg.Data[:addr.SizeBytes()], usermem.ByteOrder, &addr)
 				controlMessages.IP.OriginalDstAddress = &addr
+
+			case syscall.IP_RECVERR:
+				var errCmsg linux.SockErrCMsgIPv4
+				errCmsg.UnmarshalBytes(unixCmsg.Data)
+				controlMessages.IP.SockErr = &errCmsg
 			}
 
 		case linux.SOL_IPV6:
@@ -563,6 +565,11 @@ func parseUnixControlMessages(unixControlMessages []unix.SocketControlMessage) s
 				var addr linux.SockAddrInet6
 				binary.Unmarshal(unixCmsg.Data[:addr.SizeBytes()], usermem.ByteOrder, &addr)
 				controlMessages.IP.OriginalDstAddress = &addr
+
+			case syscall.IPV6_RECVERR:
+				var errCmsg linux.SockErrCMsgIPv6
+				errCmsg.UnmarshalBytes(unixCmsg.Data)
+				controlMessages.IP.SockErr = &errCmsg
 			}
 
 		case linux.SOL_TCP:

pkg/sentry/socket/netstack/netstack.go

@@ -2772,6 +2772,8 @@ func (s *socketOpsCommon) controlMessages() socket.ControlMessages {
 		IP: socket.IPControlMessages{
 			HasTimestamp:       s.readCM.HasTimestamp && s.sockOptTimestamp,
 			Timestamp:          s.readCM.Timestamp,
+			HasInq:             s.readCM.HasInq,
+			Inq:                s.readCM.Inq,
 			HasTOS:             s.readCM.HasTOS,
 			TOS:                s.readCM.TOS,
 			HasTClass:          s.readCM.HasTClass,
@@ -2779,6 +2781,7 @@ func (s *socketOpsCommon) controlMessages() socket.ControlMessages {
 			HasIPPacketInfo:    s.readCM.HasIPPacketInfo,
 			PacketInfo:         s.readCM.PacketInfo,
 			OriginalDstAddress: s.readCM.OriginalDstAddress,
+			SockErr:            s.readCM.SockErr,
 		},
 	}
 }
@@ -2795,9 +2798,49 @@ func (s *socketOpsCommon) updateTimestamp() {
 	}
 }
 
+// addrFamilyFromNetProto returns the address family identifier for the given
+// network protocol.
+func addrFamilyFromNetProto(net tcpip.NetworkProtocolNumber) int {
+	switch net {
+	case header.IPv4ProtocolNumber:
+		return linux.AF_INET
+	case header.IPv6ProtocolNumber:
+		return linux.AF_INET6
+	default:
+		panic(fmt.Sprintf("invalid net proto for addr family inference: %d", net))
+	}
+}
+
+// recvErr handles MSG_ERRQUEUE for recvmsg(2).
+// This is analogous to net/ipv4/ip_sockglue.c:ip_recv_error().
+func (s *socketOpsCommon) recvErr(t *kernel.Task, dst usermem.IOSequence) (int, int, linux.SockAddr, uint32, socket.ControlMessages, *syserr.Error) {
+	sockErr := s.Endpoint.SocketOptions().DequeueErr()
+	if sockErr == nil {
+		return 0, 0, nil, 0, socket.ControlMessages{}, syserr.ErrTryAgain
+	}
+
+	// The payload of the original packet that caused the error is passed as
+	// normal data via msg_iovec.  -- recvmsg(2)
+	msgFlags := linux.MSG_ERRQUEUE
+	if int(dst.NumBytes()) < len(sockErr.Payload) {
+		msgFlags |= linux.MSG_TRUNC
+	}
+	n, err := dst.CopyOut(t, sockErr.Payload)
+
+	// The original destination address of the datagram that caused the error is
+	// supplied via msg_name.  -- recvmsg(2)
+	dstAddr, dstAddrLen := socket.ConvertAddress(addrFamilyFromNetProto(sockErr.NetProto), sockErr.Dst)
+	cmgs := socket.ControlMessages{IP: socket.NewIPControlMessages(s.family, tcpip.ControlMessages{SockErr: sockErr})}
+	return n, msgFlags, dstAddr, dstAddrLen, cmgs, syserr.FromError(err)
+}
+
 // RecvMsg implements the linux syscall recvmsg(2) for sockets backed by
 // tcpip.Endpoint.
 func (s *socketOpsCommon) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, msgFlags int, senderAddr linux.SockAddr, senderAddrLen uint32, controlMessages socket.ControlMessages, err *syserr.Error) {
+	if flags&linux.MSG_ERRQUEUE != 0 {
+		return s.recvErr(t, dst)
+	}
+
 	trunc := flags&linux.MSG_TRUNC != 0
 	peek := flags&linux.MSG_PEEK != 0
 	dontWait := flags&linux.MSG_DONTWAIT != 0

pkg/sentry/socket/socket.go

@@ -56,6 +56,57 @@ func packetInfoToLinux(packetInfo tcpip.IPPacketInfo) linux.ControlMessageIPPack
 	return p
 }
 
+// errOriginToLinux maps tcpip socket origin to Linux socket origin constants.
+func errOriginToLinux(origin tcpip.SockErrOrigin) uint8 {
+	switch origin {
+	case tcpip.SockExtErrorOriginNone:
+		return linux.SO_EE_ORIGIN_NONE
+	case tcpip.SockExtErrorOriginLocal:
+		return linux.SO_EE_ORIGIN_LOCAL
+	case tcpip.SockExtErrorOriginICMP:
+		return linux.SO_EE_ORIGIN_ICMP
+	case tcpip.SockExtErrorOriginICMP6:
+		return linux.SO_EE_ORIGIN_ICMP6
+	default:
+		panic(fmt.Sprintf("unknown socket origin: %d", origin))
+	}
+}
+
+// sockErrCmsgToLinux converts SockError control message from tcpip format to
+// Linux format.
+func sockErrCmsgToLinux(sockErr *tcpip.SockError) linux.SockErrCMsg {
+	if sockErr == nil {
+		return nil
+	}
+
+	ee := linux.SockExtendedErr{
+		Errno:  uint32(syserr.TranslateNetstackError(sockErr.Err).ToLinux().Number()),
+		Origin: errOriginToLinux(sockErr.ErrOrigin),
+		Type:   sockErr.ErrType,
+		Code:   sockErr.ErrCode,
+		Info:   sockErr.ErrInfo,
+	}
+
+	switch sockErr.NetProto {
+	case header.IPv4ProtocolNumber:
+		errMsg := &linux.SockErrCMsgIPv4{SockExtendedErr: ee}
+		if len(sockErr.Offender.Addr) > 0 {
+			addr, _ := ConvertAddress(linux.AF_INET, sockErr.Offender)
+			errMsg.Offender = *addr.(*linux.SockAddrInet)
+		}
+		return errMsg
+	case header.IPv6ProtocolNumber:
+		errMsg := &linux.SockErrCMsgIPv6{SockExtendedErr: ee}
+		if len(sockErr.Offender.Addr) > 0 {
+			addr, _ := ConvertAddress(linux.AF_INET6, sockErr.Offender)
+			errMsg.Offender = *addr.(*linux.SockAddrInet6)
+		}
+		return errMsg
+	default:
+		panic(fmt.Sprintf("invalid net proto for creating SockErrCMsg: %d", sockErr.NetProto))
+	}
+}
+
 // NewIPControlMessages converts the tcpip ControlMessgaes (which does not
 // have Linux specific format) to Linux format.
 func NewIPControlMessages(family int, cmgs tcpip.ControlMessages) IPControlMessages {
@@ -75,6 +126,7 @@ func NewIPControlMessages(family int, cmgs tcpip.ControlMessages) IPControlMessa
 		HasIPPacketInfo:    cmgs.HasIPPacketInfo,
 		PacketInfo:         packetInfoToLinux(cmgs.PacketInfo),
 		OriginalDstAddress: orgDstAddr,
+		SockErr:            sockErrCmsgToLinux(cmgs.SockErr),
 	}
 }
 
@@ -117,6 +169,9 @@ type IPControlMessages struct {
 	// OriginalDestinationAddress holds the original destination address
 	// and port of the incoming packet.
 	OriginalDstAddress linux.SockAddr
+
+	// SockErr is the dequeued socket error on recvmsg(MSG_ERRQUEUE).
+	SockErr linux.SockErrCMsg
 }
 
 // Release releases Unix domain socket credentials and rights.

pkg/sentry/syscalls/linux/sys_socket.go

@@ -749,11 +749,6 @@ func recvSingleMsg(t *kernel.Task, s socket.Socket, msgPtr usermem.Addr, flags i
 		return 0, err
 	}
 
-	// FIXME(b/63594852): Pretend we have an empty error queue.
-	if flags&linux.MSG_ERRQUEUE != 0 {
-		return 0, syserror.EAGAIN
-	}
-
 	// Fast path when no control message nor name buffers are provided.
 	if msg.ControlLen == 0 && msg.NameLen == 0 {
 		n, mflags, _, _, cms, err := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, false, 0)

pkg/sentry/syscalls/linux/vfs2/socket.go

@@ -752,11 +752,6 @@ func recvSingleMsg(t *kernel.Task, s socket.SocketVFS2, msgPtr usermem.Addr, fla
 		return 0, err
 	}
 
-	// FIXME(b/63594852): Pretend we have an empty error queue.
-	if flags&linux.MSG_ERRQUEUE != 0 {
-		return 0, syserror.EAGAIN
-	}
-
 	// Fast path when no control message nor name buffers are provided.
 	if msg.ControlLen == 0 && msg.NameLen == 0 {
 		n, mflags, _, _, cms, err := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, false, 0)

pkg/tcpip/BUILD

@@ -1,10 +1,24 @@
 load("//tools:defs.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
 
 package(licenses = ["notice"])
 
+go_template_instance(
+    name = "sock_err_list",
+    out = "sock_err_list.go",
+    package = "tcpip",
+    prefix = "sockError",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Element": "*SockError",
+        "Linker": "*SockError",
+    },
+)
+
 go_library(
     name = "tcpip",
     srcs = [
+        "sock_err_list.go",
         "socketops.go",
         "tcpip.go",
         "time_unsafe.go",

pkg/tcpip/socketops.go

@@ -104,7 +104,7 @@ type SocketOptions struct {
 	keepAliveEnabled uint32
 
 	// multicastLoopEnabled determines whether multicast packets sent over a
-	// non-loopback interface will be looped back. Analogous to inet->mc_loop.
+	// non-loopback interface will be looped back.
 	multicastLoopEnabled uint32
 
 	// receiveTOSEnabled is used to specify if the TOS ancillary message is
@@ -145,6 +145,10 @@ type SocketOptions struct {
 	// the incoming packet should be returned as an ancillary message.
 	receiveOriginalDstAddress uint32
 
+	// errQueue is the per-socket error queue. It is protected by errQueueMu.
+	errQueueMu sync.Mutex `state:"nosave"`
+	errQueue   sockErrorList
+
 	// mu protects the access to the below fields.
 	mu sync.Mutex `state:"nosave"`
 
@@ -362,3 +366,60 @@ func (so *SocketOptions) SetLinger(linger LingerOption) {
 	so.linger = linger
 	so.mu.Unlock()
 }
+
+// SockErrOrigin represents the constants for error origin.
+type SockErrOrigin uint8
+
+const (
+	// SockExtErrorOriginNone represents an unknown error origin.
+	SockExtErrorOriginNone SockErrOrigin = iota
+
+	// SockExtErrorOriginLocal indicates a local error.
+	SockExtErrorOriginLocal
+
+	// SockExtErrorOriginICMP indicates an IPv4 ICMP error.
+	SockExtErrorOriginICMP
+
+	// SockExtErrorOriginICMP6 indicates an IPv6 ICMP error.
+	SockExtErrorOriginICMP6
+)
+
+// SockError represents a queue entry in the per-socket error queue.
+//
+// +stateify savable
+type SockError struct {
+	sockErrorEntry
+
+	// Err is the error caused by the errant packet.
+	Err *Error
+	// ErrOrigin indicates the error origin.
+	ErrOrigin SockErrOrigin
+	// ErrType is the type in the ICMP header.
+	ErrType uint8
+	// ErrCode is the code in the ICMP header.
+	ErrCode uint8
+	// ErrInfo is additional info about the error.
+	ErrInfo uint32
+
+	// Payload is the errant packet's payload.
+	Payload []byte
+	// Dst is the original destination address of the errant packet.
+	Dst FullAddress
+	// Offender is the original sender address of the errant packet.
+	Offender FullAddress
+	// NetProto is the network protocol being used to transmit the packet.
+	NetProto NetworkProtocolNumber
+}
+
+// DequeueErr dequeues a socket extended error from the error queue and returns
+// it. Returns nil if queue is empty.
+func (so *SocketOptions) DequeueErr() *SockError {
+	so.errQueueMu.Lock()
+	defer so.errQueueMu.Unlock()
+
+	err := so.errQueue.Front()
+	if err != nil {
+		so.errQueue.Remove(err)
+	}
+	return err
+}

pkg/tcpip/tcpip.go

@@ -500,6 +500,9 @@ type ControlMessages struct {
 	// OriginalDestinationAddress holds the original destination address
 	// and port of the incoming packet.
 	OriginalDstAddress FullAddress
+
+	// SockErr is the dequeued socket error on recvmsg(MSG_ERRQUEUE).
+	SockErr *SockError
 }
 
 // PacketOwner is used to get UID and GID of the packet.