pmtud package

2025-08-19 20:04:22 +00:00
parent 5306e3bab1
commit e21d798f57
14 changed files with 859 additions and 0 deletions
--- a/internal/pmtud/apple_ipv4.go
+++ b/internal/pmtud/apple_ipv4.go
@@ -0,0 +1,49 @@
+package pmtud
+
+import (
+	"net"
+	"time"
+
+	"golang.org/x/net/ipv4"
+)
+
+var _ net.PacketConn = &ipv4Wrapper{}
+
+// ipv4Wrapper is a wrapper around ipv4.PacketConn to implement
+// the net.PacketConn interface. It's only used for Darwin or iOS.
+type ipv4Wrapper struct {
+	ipv4Conn *ipv4.PacketConn
+}
+
+func ipv4ToNetPacketConn(ipv4 *ipv4.PacketConn) *ipv4Wrapper {
+	return &ipv4Wrapper{ipv4Conn: ipv4}
+}
+
+func (i *ipv4Wrapper) ReadFrom(p []byte) (n int, addr net.Addr, err error) {
+	n, _, addr, err = i.ipv4Conn.ReadFrom(p)
+	return n, addr, err
+}
+
+func (i *ipv4Wrapper) WriteTo(p []byte, addr net.Addr) (n int, err error) {
+	return i.ipv4Conn.WriteTo(p, nil, addr)
+}
+
+func (i *ipv4Wrapper) Close() error {
+	return i.ipv4Conn.Close()
+}
+
+func (i *ipv4Wrapper) LocalAddr() net.Addr {
+	return i.ipv4Conn.LocalAddr()
+}
+
+func (i *ipv4Wrapper) SetDeadline(t time.Time) error {
+	return i.ipv4Conn.SetDeadline(t)
+}
+
+func (i *ipv4Wrapper) SetReadDeadline(t time.Time) error {
+	return i.ipv4Conn.SetReadDeadline(t)
+}
+
+func (i *ipv4Wrapper) SetWriteDeadline(t time.Time) error {
+	return i.ipv4Conn.SetWriteDeadline(t)
+}
--- a/internal/pmtud/check.go
+++ b/internal/pmtud/check.go
@@ -0,0 +1,83 @@
+package pmtud
+
+import (
+	"bytes"
+	"errors"
+	"fmt"
+
+	"golang.org/x/net/icmp"
+)
+
+var (
+	ErrICMPNextHopMTUTooLow  = errors.New("ICMP Next Hop MTU is too low")
+	ErrICMPNextHopMTUTooHigh = errors.New("ICMP Next Hop MTU is too high")
+)
+
+func checkMTU(mtu, minMTU, physicalLinkMTU int) (err error) {
+	switch {
+	case mtu < minMTU:
+		return fmt.Errorf("%w: %d", ErrICMPNextHopMTUTooLow, mtu)
+	case mtu > physicalLinkMTU:
+		return fmt.Errorf("%w: %d is larger than physical link MTU %d",
+			ErrICMPNextHopMTUTooHigh, mtu, physicalLinkMTU)
+	default:
+		return nil
+	}
+}
+
+func checkInvokingReplyIDMatch(icmpProtocol int, received []byte,
+	outboundMessage *icmp.Message,
+) (match bool, err error) {
+	inboundMessage, err := icmp.ParseMessage(icmpProtocol, received)
+	if err != nil {
+		return false, fmt.Errorf("parsing invoking packet: %w", err)
+	}
+	inboundBody, ok := inboundMessage.Body.(*icmp.Echo)
+	if !ok {
+		return false, fmt.Errorf("%w: %T", ErrICMPBodyUnsupported, inboundMessage.Body)
+	}
+	outboundBody := outboundMessage.Body.(*icmp.Echo) //nolint:forcetypeassert
+	return inboundBody.ID == outboundBody.ID, nil
+}
+
+var ErrICMPIDMismatch = errors.New("ICMP id mismatch")
+
+func checkEchoReply(icmpProtocol int, received []byte,
+	outboundMessage *icmp.Message, truncatedBody bool,
+) (err error) {
+	inboundMessage, err := icmp.ParseMessage(icmpProtocol, received)
+	if err != nil {
+		return fmt.Errorf("parsing invoking packet: %w", err)
+	}
+	inboundBody, ok := inboundMessage.Body.(*icmp.Echo)
+	if !ok {
+		return fmt.Errorf("%w: %T", ErrICMPBodyUnsupported, inboundMessage.Body)
+	}
+	outboundBody := outboundMessage.Body.(*icmp.Echo) //nolint:forcetypeassert
+	if inboundBody.ID != outboundBody.ID {
+		return fmt.Errorf("%w: sent id %d and received id %d",
+			ErrICMPIDMismatch, outboundBody.ID, inboundBody.ID)
+	}
+	err = checkEchoBodies(outboundBody.Data, inboundBody.Data, truncatedBody)
+	if err != nil {
+		return fmt.Errorf("checking sent and received bodies: %w", err)
+	}
+	return nil
+}
+
+var ErrICMPEchoDataMismatch = errors.New("ICMP data mismatch")
+
+func checkEchoBodies(sent, received []byte, receivedTruncated bool) (err error) {
+	if len(received) > len(sent) {
+		return fmt.Errorf("%w: sent %d bytes and received %d bytes",
+			ErrICMPEchoDataMismatch, len(sent), len(received))
+	}
+	if receivedTruncated {
+		sent = sent[:len(received)]
+	}
+	if !bytes.Equal(received, sent) {
+		return fmt.Errorf("%w: sent %x and received %x",
+			ErrICMPEchoDataMismatch, sent, received)
+	}
+	return nil
+}
--- a/internal/pmtud/df.go
+++ b/internal/pmtud/df.go
@@ -0,0 +1,10 @@
+//go:build !linux && !windows
+
+package pmtud
+
+// setDontFragment for platforms other than Linux and Windows
+// is not implemented, so we just return assuming the don't
+// fragment flag is set on IP packets.
+func setDontFragment(fd uintptr) (err error) {
+	return nil
+}
--- a/internal/pmtud/df_linux.go
+++ b/internal/pmtud/df_linux.go
@@ -0,0 +1,12 @@
+//go:build linux
+
+package pmtud
+
+import (
+	"syscall"
+)
+
+func setDontFragment(fd uintptr) (err error) {
+	return syscall.SetsockoptInt(int(fd), syscall.IPPROTO_IP,
+		syscall.IP_MTU_DISCOVER, syscall.IP_PMTUDISC_PROBE)
+}
--- a/internal/pmtud/df_windows.go
+++ b/internal/pmtud/df_windows.go
@@ -0,0 +1,13 @@
+//go:build windows
+
+package pmtud
+
+import (
+	"syscall"
+)
+
+func setDontFragment(fd uintptr) (err error) {
+	// https://docs.microsoft.com/en-us/troubleshoot/windows/win32/header-library-requirement-socket-ipproto-ip
+	// #define IP_DONTFRAGMENT        14     /* don't fragment IP datagrams */
+	return syscall.SetsockoptInt(syscall.Handle(fd), syscall.IPPROTO_IP, 14, 1)
+}
--- a/internal/pmtud/errors.go
+++ b/internal/pmtud/errors.go
@@ -0,0 +1,10 @@
+package pmtud
+
+import (
+	"errors"
+)
+
+var (
+	ErrICMPDestinationUnreachable = errors.New("ICMP destination unreachable")
+	ErrICMPBodyUnsupported        = errors.New("ICMP body type is not supported")
+)
--- a/internal/pmtud/interfaces.go
+++ b/internal/pmtud/interfaces.go
@@ -0,0 +1,7 @@
+package pmtud
+
+type Logger interface {
+	Debug(msg string)
+	Debugf(msg string, args ...any)
+	Warnf(msg string, args ...any)
+}
--- a/internal/pmtud/ipv4.go
+++ b/internal/pmtud/ipv4.go
@@ -0,0 +1,145 @@
+package pmtud
+
+import (
+	"context"
+	"encoding/binary"
+	"fmt"
+	"net"
+	"net/netip"
+	"runtime"
+	"syscall"
+	"time"
+
+	"golang.org/x/net/icmp"
+	"golang.org/x/net/ipv4"
+)
+
+const (
+	// see https://en.wikipedia.org/wiki/Maximum_transmission_unit#MTUs_for_common_media
+	minIPv4MTU     = 68
+	icmpv4Protocol = 1
+)
+
+func listenICMPv4(ctx context.Context) (conn net.PacketConn, err error) {
+	var listenConfig net.ListenConfig
+	listenConfig.Control = func(_, _ string, rawConn syscall.RawConn) error {
+		var setDFErr error
+		err := rawConn.Control(func(fd uintptr) {
+			setDFErr = setDontFragment(fd) // runs when calling ListenPacket
+		})
+		if err == nil {
+			err = setDFErr
+		}
+		return err
+	}
+
+	const listenAddress = ""
+	packetConn, err := listenConfig.ListenPacket(ctx, "ip4:icmp", listenAddress)
+	if err != nil {
+		return nil, fmt.Errorf("listening for ICMP packets: %w", err)
+	}
+
+	if runtime.GOOS == "darwin" || runtime.GOOS == "ios" {
+		packetConn = ipv4ToNetPacketConn(ipv4.NewPacketConn(packetConn))
+	}
+
+	return packetConn, nil
+}
+
+func findIPv4NextHopMTU(ctx context.Context, ip netip.Addr,
+	physicalLinkMTU int, pingTimeout time.Duration, logger Logger,
+) (mtu int, err error) {
+	if ip.Is6() {
+		panic("IP address is not v4")
+	}
+	conn, err := listenICMPv4(ctx)
+	if err != nil {
+		return 0, fmt.Errorf("listening for ICMP packets: %w", err)
+	}
+	ctx, cancel := context.WithTimeout(ctx, pingTimeout)
+	defer cancel()
+	go func() {
+		<-ctx.Done()
+		conn.Close()
+	}()
+
+	// First try to send a packet which is too big to get the maximum MTU
+	// directly.
+	outboundID, outboundMessage := buildMessageToSend("v4", physicalLinkMTU)
+	encodedMessage, err := outboundMessage.Marshal(nil)
+	if err != nil {
+		return 0, fmt.Errorf("encoding ICMP message: %w", err)
+	}
+
+	_, err = conn.WriteTo(encodedMessage, &net.IPAddr{IP: ip.AsSlice()})
+	if err != nil {
+		return 0, fmt.Errorf("writing ICMP message: %w", err)
+	}
+
+	buffer := make([]byte, physicalLinkMTU)
+
+	for { // for loop in case we read an echo reply for another ICMP request
+		// Note we need to read the whole packet in one call to ReadFrom, so the buffer
+		// must be large enough to read the entire reply packet. See:
+		// https://groups.google.com/g/golang-nuts/c/5dy2Q4nPs08/m/KmuSQAGEtG4J
+		bytesRead, _, err := conn.ReadFrom(buffer)
+		if err != nil {
+			return 0, fmt.Errorf("reading from ICMP connection: %w", err)
+		}
+		packetBytes := buffer[:bytesRead]
+		// Side note: echo reply should be at most the number of bytes
+		// sent, and can be lower, more precisely 576-ipHeader bytes,
+		// in case the next hop we are reaching replies with a destination
+		// unreachable and wants to ensure the response makes it way back
+		// by keeping a low packet size, see:
+		// https://datatracker.ietf.org/doc/html/rfc1122#page-59
+
+		inboundMessage, err := icmp.ParseMessage(icmpv4Protocol, packetBytes)
+		if err != nil {
+			return 0, fmt.Errorf("parsing message: %w", err)
+		}
+
+		switch typedBody := inboundMessage.Body.(type) {
+		case *icmp.DstUnreach:
+			const fragmentationRequiredAndDFFlagSetCode = 4
+			if inboundMessage.Code != fragmentationRequiredAndDFFlagSetCode {
+				return 0, fmt.Errorf("%w: code %d",
+					ErrICMPDestinationUnreachable, inboundMessage.Code)
+			}
+
+			// See https://datatracker.ietf.org/doc/html/rfc1191#section-4
+			// Note: the go library does not handle this NextHopMTU section.
+			nextHopMTU := packetBytes[6:8]
+			mtu = int(binary.BigEndian.Uint16(nextHopMTU))
+			err = checkMTU(mtu, minIPv4MTU, physicalLinkMTU)
+			if err != nil {
+				return 0, fmt.Errorf("checking next-hop-mtu found: %w", err)
+			}
+
+			// The code below is really for sanity checks
+			packetBytes = packetBytes[8:]
+			header, err := ipv4.ParseHeader(packetBytes)
+			if err != nil {
+				return 0, fmt.Errorf("parsing IPv4 header: %w", err)
+			}
+			packetBytes = packetBytes[header.Len:] // truncated original datagram
+
+			const truncated = true
+			err = checkEchoReply(icmpv4Protocol, packetBytes, outboundMessage, truncated)
+			if err != nil {
+				return 0, fmt.Errorf("checking echo reply: %w", err)
+			}
+			return mtu, nil
+		case *icmp.Echo:
+			inboundID := uint16(typedBody.ID) //nolint:gosec
+			if inboundID == outboundID {
+				return physicalLinkMTU, nil
+			}
+			logger.Debug("discarding received ICMP echo reply with id %d mismatching sent id %d",
+				inboundID, outboundID)
+			continue
+		default:
+			return 0, fmt.Errorf("%w: %T", ErrICMPBodyUnsupported, typedBody)
+		}
+	}
+}
--- a/internal/pmtud/ipv6.go
+++ b/internal/pmtud/ipv6.go
@@ -0,0 +1,116 @@
+package pmtud
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"net/netip"
+	"time"
+
+	"golang.org/x/net/icmp"
+	"golang.org/x/net/ipv6"
+)
+
+const (
+	minIPv6MTU     = 1280
+	icmpv6Protocol = 58
+)
+
+func listenICMPv6(ctx context.Context) (conn net.PacketConn, err error) {
+	var listenConfig net.ListenConfig
+	const listenAddress = ""
+	packetConn, err := listenConfig.ListenPacket(ctx, "ip6:ipv6-icmp", listenAddress)
+	if err != nil {
+		return nil, fmt.Errorf("listening for ICMPv6 packets: %w", err)
+	}
+	return packetConn, nil
+}
+
+func getIPv6PacketTooBig(ctx context.Context, ip netip.Addr,
+	physicalLinkMTU int, pingTimeout time.Duration, logger Logger,
+) (mtu int, err error) {
+	if ip.Is4() {
+		panic("IP address is not v6")
+	}
+	conn, err := listenICMPv6(ctx)
+	if err != nil {
+		return 0, fmt.Errorf("listening for ICMP packets: %w", err)
+	}
+	ctx, cancel := context.WithTimeout(ctx, pingTimeout)
+	defer cancel()
+	go func() {
+		<-ctx.Done()
+		conn.Close()
+	}()
+
+	// First try to send a packet which is too big to get the maximum MTU
+	// directly.
+	outboundID, outboundMessage := buildMessageToSend("v6", physicalLinkMTU)
+	encodedMessage, err := outboundMessage.Marshal(nil)
+	if err != nil {
+		return 0, fmt.Errorf("encoding ICMP message: %w", err)
+	}
+
+	_, err = conn.WriteTo(encodedMessage, &net.IPAddr{IP: ip.AsSlice(), Zone: ip.Zone()})
+	if err != nil {
+		return 0, fmt.Errorf("writing ICMP message: %w", err)
+	}
+
+	buffer := make([]byte, physicalLinkMTU)
+
+	for { // for loop if we encounter another ICMP packet with an unknown id.
+		// Note we need to read the whole packet in one call to ReadFrom, so the buffer
+		// must be large enough to read the entire reply packet. See:
+		// https://groups.google.com/g/golang-nuts/c/5dy2Q4nPs08/m/KmuSQAGEtG4J
+		bytesRead, _, err := conn.ReadFrom(buffer)
+		if err != nil {
+			return 0, fmt.Errorf("reading from ICMP connection: %w", err)
+		}
+		packetBytes := buffer[:bytesRead]
+
+		packetBytes = packetBytes[ipv6.HeaderLen:]
+
+		inboundMessage, err := icmp.ParseMessage(icmpv6Protocol, packetBytes)
+		if err != nil {
+			return 0, fmt.Errorf("parsing message: %w", err)
+		}
+
+		switch typedBody := inboundMessage.Body.(type) {
+		case *icmp.PacketTooBig:
+			// https://datatracker.ietf.org/doc/html/rfc1885#section-3.2
+			mtu = typedBody.MTU
+			err = checkMTU(mtu, minIPv6MTU, physicalLinkMTU)
+			if err != nil {
+				return 0, fmt.Errorf("checking MTU: %w", err)
+			}
+
+			// Sanity checks
+			const truncatedBody = true
+			err = checkEchoReply(icmpv6Protocol, typedBody.Data, outboundMessage, truncatedBody)
+			if err != nil {
+				return 0, fmt.Errorf("checking invoking message: %w", err)
+			}
+			return typedBody.MTU, nil
+		case *icmp.DstUnreach:
+			// https://datatracker.ietf.org/doc/html/rfc1885#section-3.1
+			idMatch, err := checkInvokingReplyIDMatch(icmpv6Protocol, packetBytes, outboundMessage)
+			if err != nil {
+				return 0, fmt.Errorf("checking invoking message id: %w", err)
+			} else if idMatch {
+				return 0, fmt.Errorf("%w", ErrICMPDestinationUnreachable)
+			}
+			logger.Debug("discarding received ICMP destination unreachable reply with an unknown id")
+			continue
+		case *icmp.Echo:
+			inboundID := uint16(typedBody.ID) //nolint:gosec
+			if inboundID == outboundID {
+				return physicalLinkMTU, nil
+			}
+			logger.Debug("discarding received ICMP echo reply with id %d mismatching sent id %d",
+				inboundID, outboundID)
+			continue
+		default:
+			return 0, fmt.Errorf("%w: %T", ErrICMPBodyUnsupported, typedBody)
+		}
+	}
+}
--- a/internal/pmtud/message.go
+++ b/internal/pmtud/message.go
@@ -0,0 +1,58 @@
+package pmtud
+
+import (
+	cryptorand "crypto/rand"
+	"encoding/binary"
+	"fmt"
+	"math/rand/v2"
+
+	"golang.org/x/net/icmp"
+	"golang.org/x/net/ipv4"
+	"golang.org/x/net/ipv6"
+)
+
+func buildMessageToSend(ipVersion string, mtu int) (id uint16, message *icmp.Message) {
+	var seed [32]byte
+	_, _ = cryptorand.Read(seed[:])
+	randomSource := rand.NewChaCha8(seed)
+
+	const uint16Bytes = 2
+	idBytes := make([]byte, uint16Bytes)
+	_, _ = randomSource.Read(idBytes)
+	id = binary.BigEndian.Uint16(idBytes)
+
+	var ipHeaderLength int
+	var icmpType icmp.Type
+	switch ipVersion {
+	case "v4":
+		ipHeaderLength = ipv4.HeaderLen
+		icmpType = ipv4.ICMPTypeEcho
+	case "v6":
+		ipHeaderLength = ipv6.HeaderLen
+		icmpType = ipv6.ICMPTypeEchoRequest
+	default:
+		panic(fmt.Sprintf("IP version %q not supported", ipVersion))
+	}
+	const pingHeaderLength = 0 +
+		1 + // type
+		1 + // code
+		2 + // checksum
+		2 + // identifier
+		2 // sequence number
+	pingBodyDataSize := mtu - ipHeaderLength - pingHeaderLength
+	messageBodyData := make([]byte, pingBodyDataSize)
+	_, _ = randomSource.Read(messageBodyData)
+
+	// See https://www.iana.org/assignments/icmp-parameters/icmp-parameters.xhtml#icmp-parameters-types
+	message = &icmp.Message{
+		Type:     icmpType, // echo request
+		Code:     0,        // no code
+		Checksum: 0,        // calculated at encoding (ipv4) or sending (ipv6)
+		Body: &icmp.Echo{
+			ID:   int(id),
+			Seq:  0, // only one packet
+			Data: messageBodyData,
+		},
+	}
+	return id, message
+}
--- a/internal/pmtud/nooplogger.go
+++ b/internal/pmtud/nooplogger.go
@@ -0,0 +1,7 @@
+package pmtud
+
+type noopLogger struct{}
+
+func (noopLogger) Debug(_ string)            {}
+func (noopLogger) Debugf(_ string, _ ...any) {}
+func (noopLogger) Warnf(_ string, _ ...any)  {}
--- a/internal/pmtud/pmtud.go
+++ b/internal/pmtud/pmtud.go
@@ -0,0 +1,272 @@
+package pmtud
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"math"
+	"net"
+	"net/netip"
+	"time"
+
+	"golang.org/x/net/icmp"
+)
+
+// PathMTUDiscover discovers the maximum MTU for the path to the given ip address.
+// If the physicalLinkMTU is zero, it defaults to 1500 which is the ethernet standard MTU.
+// If the pingTimeout is zero, it defaults to 1 second.
+// If the logger is nil, a no-op logger is used.
+func PathMTUDiscover(ctx context.Context, ip netip.Addr,
+	physicalLinkMTU int, pingTimeout time.Duration, logger Logger) (
+	mtu int, err error,
+) {
+	if physicalLinkMTU == 0 {
+		const ethernetStandardMTU = 1500
+		physicalLinkMTU = ethernetStandardMTU
+	}
+	if pingTimeout == 0 {
+		pingTimeout = time.Second
+	}
+	if logger == nil {
+		logger = &noopLogger{}
+	}
+
+	if ip.Is4() {
+		logger.Debug("finding IPv4 next hop MTU")
+		mtu, err = findIPv4NextHopMTU(ctx, ip, physicalLinkMTU, pingTimeout, logger)
+		switch {
+		case err == nil:
+			return mtu, nil
+		case errors.Is(err, net.ErrClosed): // blackhole
+		default:
+			return 0, fmt.Errorf("finding IPv4 next hop MTU: %w", err)
+		}
+	} else {
+		logger.Debug("requesting IPv6 ICMP packet-too-big reply")
+		mtu, err = getIPv6PacketTooBig(ctx, ip, physicalLinkMTU, pingTimeout, logger)
+		switch {
+		case err == nil:
+			return mtu, nil
+		case errors.Is(err, net.ErrClosed): // blackhole
+		default:
+			return 0, fmt.Errorf("getting IPv6 packet-too-big message: %w", err)
+		}
+	}
+
+	// Fall back method: send echo requests with different packet
+	// sizes and check which ones succeed to find the maximum MTU.
+	logger.Debug("falling back to sending different sized echo packets")
+	minMTU := minIPv4MTU
+	if ip.Is6() {
+		minMTU = minIPv6MTU
+	}
+	return pmtudMultiSizes(ctx, ip, minMTU, physicalLinkMTU, pingTimeout, logger)
+}
+
+type pmtudTestUnit struct {
+	mtu       int
+	echoID    uint16
+	sentBytes int
+	ok        bool
+}
+
+func pmtudMultiSizes(ctx context.Context, ip netip.Addr,
+	minMTU, maxPossibleMTU int, pingTimeout time.Duration,
+	logger Logger,
+) (maxMTU int, err error) {
+	var ipVersion string
+	var conn net.PacketConn
+	if ip.Is4() {
+		ipVersion = "v4"
+		conn, err = listenICMPv4(ctx)
+	} else {
+		ipVersion = "v6"
+		conn, err = listenICMPv6(ctx)
+	}
+	if err != nil {
+		return 0, fmt.Errorf("listening for ICMP packets: %w", err)
+	}
+
+	mtusToTest := makeMTUsToTest(minMTU, maxPossibleMTU)
+	if len(mtusToTest) == 0 {
+		return minMTU, nil
+	}
+	logger.Debugf("testing the following MTUs: %v", mtusToTest)
+
+	tests := make([]pmtudTestUnit, len(mtusToTest))
+	for i := range mtusToTest {
+		tests[i] = pmtudTestUnit{mtu: mtusToTest[i]}
+	}
+
+	timedCtx, cancel := context.WithTimeout(ctx, pingTimeout)
+	defer cancel()
+	go func() {
+		<-timedCtx.Done()
+		conn.Close()
+	}()
+
+	for i := range tests {
+		id, message := buildMessageToSend(ipVersion, tests[i].mtu)
+		tests[i].echoID = id
+
+		encodedMessage, err := message.Marshal(nil)
+		if err != nil {
+			return 0, fmt.Errorf("encoding ICMP message: %w", err)
+		}
+		tests[i].sentBytes = len(encodedMessage)
+
+		_, err = conn.WriteTo(encodedMessage, &net.IPAddr{IP: ip.AsSlice()})
+		if err != nil {
+			return 0, fmt.Errorf("writing ICMP message: %w", err)
+		}
+	}
+
+	err = collectReplies(conn, ipVersion, tests, logger)
+	switch {
+	case err == nil: // max possible MTU is working
+		return tests[len(tests)-1].mtu, nil
+	case err != nil && errors.Is(err, net.ErrClosed):
+		// we have timeouts (IPv4 testing or IPv6 PMTUD blackholes)
+		// so find the highest MTU which worked.
+		// Note we start from index len(tests) - 2 since the max MTU
+		// cannot be working if we had a timeout.
+		for i := len(tests) - 2; i >= 0; i-- { //nolint:mnd
+			if tests[i].ok {
+				return pmtudMultiSizes(ctx, ip, tests[i].mtu, tests[i+1].mtu,
+					pingTimeout, logger)
+			}
+		}
+
+		// All MTUs failed.
+		if tests[0].mtu == minMTU+1 {
+			return minMTU, nil
+		}
+		// Re-test with MTUs between the minimum MTU
+		// and the smallest next MTU we tested.
+		return pmtudMultiSizes(ctx, ip, minMTU, tests[0].mtu,
+			pingTimeout, logger)
+	case err != nil:
+		return 0, fmt.Errorf("collecting ICMP echo replies: %w", err)
+	default:
+		panic("unreachable")
+	}
+}
+
+// Create the MTU slice of length 8 such that:
+// - the first element is the minMTU plus the step
+// - the last element is the maxMTU
+// - elements in-between are separated as close to each other
+// - Don't make the minMTU part of the MTUs to test since
+// it's assumed it's already working.
+func makeMTUsToTest(minMTU, maxMTU int) (mtus []int) {
+	const mtusLength = 8
+	diff := maxMTU - minMTU
+	switch {
+	case minMTU > maxMTU:
+		panic("minMTU > maxMTU")
+	case diff <= mtusLength:
+		mtus = make([]int, 0, diff)
+		for mtu := minMTU + 1; mtu <= maxMTU; mtu++ {
+			mtus = append(mtus, mtu)
+		}
+	case diff < mtusLength*2:
+		step := float64(diff) / float64(mtusLength)
+		mtus = make([]int, 0, mtusLength)
+		for mtu := float64(minMTU) + step; len(mtus) < mtusLength-1; mtu += step {
+			mtus = append(mtus, int(math.Round(mtu)))
+		}
+		mtus = append(mtus, maxMTU) // last element is the maxMTU
+	default:
+		step := diff / mtusLength
+		mtus = make([]int, 0, mtusLength)
+		for mtu := minMTU + step; len(mtus) < mtusLength-1; mtu += step {
+			mtus = append(mtus, mtu)
+		}
+		mtus = append(mtus, maxMTU) // last element is the maxMTU
+	}
+
+	return mtus
+}
+
+func collectReplies(conn net.PacketConn, ipVersion string,
+	tests []pmtudTestUnit, logger Logger,
+) (err error) {
+	echoIDToTestIndex := make(map[uint16]int, len(tests))
+	for i, test := range tests {
+		echoIDToTestIndex[test.echoID] = i
+	}
+
+	// The theoretical limit is 4GiB for IPv6 MTU path discovery jumbograms, but that would
+	// create huge buffers which we don't really want to support anyway.
+	// The standard frame maximum MTU is 1500 bytes, and there are Jumbo frames with
+	// a conventional maximum of 9000 bytes. However, some manufacturers support up
+	// 9216-20 = 9196 bytes for the maximum MTU. We thus use buffers of size 9196 to
+	// match eventual Jumbo frames. More information at:
+	// https://en.wikipedia.org/wiki/Maximum_transmission_unit#MTUs_for_common_media
+	const maxPossibleMTU = 9196
+	buffer := make([]byte, maxPossibleMTU)
+
+	idsFound := 0
+	for idsFound < len(tests) {
+		// Note we need to read the whole packet in one call to ReadFrom, so the buffer
+		// must be large enough to read the entire reply packet. See:
+		// https://groups.google.com/g/golang-nuts/c/5dy2Q4nPs08/m/KmuSQAGEtG4J
+		bytesRead, _, err := conn.ReadFrom(buffer)
+		if err != nil {
+			return fmt.Errorf("reading from ICMP connection: %w", err)
+		}
+		packetBytes := buffer[:bytesRead]
+
+		ipPacketLength := len(packetBytes)
+
+		var icmpProtocol int
+		switch ipVersion {
+		case "v4":
+			icmpProtocol = icmpv4Protocol
+		case "v6":
+			icmpProtocol = icmpv6Protocol
+		default:
+			panic(fmt.Sprintf("unknown IP version: %s", ipVersion))
+		}
+
+		// Parse the ICMP message
+		// Note: this parsing works for a truncated 556 bytes ICMP reply packet.
+		message, err := icmp.ParseMessage(icmpProtocol, packetBytes)
+		if err != nil {
+			return fmt.Errorf("parsing message: %w", err)
+		}
+
+		echoBody, ok := message.Body.(*icmp.Echo)
+		if !ok {
+			return fmt.Errorf("%w: %T", ErrICMPBodyUnsupported, message.Body)
+		}
+
+		id := uint16(echoBody.ID) //nolint:gosec
+		testIndex, testing := echoIDToTestIndex[id]
+		if !testing { // not an id we expected so ignore it
+			logger.Warnf("ignoring ICMP reply with unexpected ID %d (type: %d, code: %d, length: %d)",
+				echoBody.ID, message.Type, message.Code, ipPacketLength)
+			continue
+		}
+		idsFound++
+		sentBytes := tests[testIndex].sentBytes
+
+		// echo reply should be at most the number of bytes sent,
+		// and can be lower, more precisely 556 bytes, in case
+		// the host we are reaching wants to stay out of trouble
+		// and ensure its echo reply goes through without
+		// fragmentation, see the following page:
+		// https://datatracker.ietf.org/doc/html/rfc1122#page-59
+		const conservativeReplyLength = 556
+		truncated := ipPacketLength < sentBytes &&
+			ipPacketLength == conservativeReplyLength
+		// Check the packet size is the same if the reply is not truncated
+		if !truncated && sentBytes != ipPacketLength {
+			return fmt.Errorf("%w: sent %dB and received %dB",
+				ErrICMPEchoDataMismatch, sentBytes, ipPacketLength)
+		}
+		// Truncated reply or matching reply size
+		tests[testIndex].ok = true
+	}
+	return nil
+}
--- a/internal/pmtud/pmtud_integration_test.go
+++ b/internal/pmtud/pmtud_integration_test.go
@@ -0,0 +1,22 @@
+//go:build integration
+
+package pmtud
+
+import (
+	"context"
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+func Test_PathMTUDiscover(t *testing.T) {
+	t.Parallel()
+	const physicalLinkMTU = 1500
+	const timeout = time.Second
+	mtu, err := PathMTUDiscover(context.Background(), netip.MustParseAddr("1.1.1.1"),
+		physicalLinkMTU, timeout, nil)
+	require.NoError(t, err)
+	t.Log("MTU found:", mtu)
+}
--- a/internal/pmtud/pmtud_test.go
+++ b/internal/pmtud/pmtud_test.go
@@ -0,0 +1,55 @@
+package pmtud
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func Test_makeMTUsToTest(t *testing.T) {
+	t.Parallel()
+
+	testCases := map[string]struct {
+		minMTU int
+		maxMTU int
+		mtus   []int
+	}{
+		"0_0": {
+			mtus: []int{},
+		},
+		"0_1": {
+			maxMTU: 1,
+			mtus:   []int{1},
+		},
+		"0_8": {
+			maxMTU: 8,
+			mtus:   []int{1, 2, 3, 4, 5, 6, 7, 8},
+		},
+		"0_12": {
+			maxMTU: 12,
+			mtus:   []int{2, 3, 5, 6, 8, 9, 11, 12},
+		},
+		"0_80": {
+			maxMTU: 80,
+			mtus:   []int{10, 20, 30, 40, 50, 60, 70, 80},
+		},
+		"0_100": {
+			maxMTU: 100,
+			mtus:   []int{12, 24, 36, 48, 60, 72, 84, 100},
+		},
+		"1280_1500": {
+			minMTU: 1280,
+			maxMTU: 1500,
+			mtus:   []int{1307, 1334, 1361, 1388, 1415, 1442, 1469, 1500},
+		},
+	}
+
+	for name, testCase := range testCases {
+		t.Run(name, func(t *testing.T) {
+			t.Parallel()
+
+			mtus := makeMTUsToTest(testCase.minMTU, testCase.maxMTU)
+			assert.Equal(t, testCase.mtus, mtus)
+		})
+	}
+}