diff --git a/libnetwork/drivers/overlay/bpf_linux_test.go b/libnetwork/drivers/overlay/bpf_linux_test.go
new file mode 100644
index 0000000000..e20065169b
--- /dev/null
+++ b/libnetwork/drivers/overlay/bpf_linux_test.go
@@ -0,0 +1,227 @@
+package overlay
+
+import (
+	"bytes"
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"net"
+	"net/netip"
+	"testing"
+	"time"
+
+	"golang.org/x/net/bpf"
+	"golang.org/x/net/ipv4"
+	"golang.org/x/sys/unix"
+)
+
+func TestVNIMatchBPF(t *testing.T) {
+	// The BPF filter program under test uses Linux extensions which are not
+	// emulated by any user-space BPF interpreters. It is also classic BPF,
+	// which cannot be tested in-kernel using the bpf(BPF_PROG_RUN) syscall.
+	// The best we can do without actually programming it into an iptables
+	// rule and end-to-end testing it is to attach it as a socket filter to
+	// a raw socket and test which loopback packets make it through.
+	//
+	// Modern kernels transpile cBPF programs into eBPF for execution, so a
+	// possible future direction would be to extract the transpiler and
+	// convert the program under test to eBPF so it could be loaded and run
+	// using the bpf(2) syscall.
+	// https://elixir.bootlin.com/linux/v6.2/source/net/core/filter.c#L559
+	// Though the effort would be better spent on adding nftables support to
+	// libnetwork so this whole BPF program could be replaced with a native
+	// nftables '@th' match expression.
+	//
+	// The filter could be manually e2e-tested for both IPv4 and IPv6 by
+	// programming ip[6]tables rules which log matching packets and sending
+	// test packets loopback using netcat. All the necessary information
+	// (bytecode and an acceptable test vector) is logged by this test.
+	//
+	//     $ sudo ip6tables -A INPUT -p udp -s ::1 -d ::1 -m bpf \
+	//         --bytecode "${bpf_program_under_test}" \
+	//         -j LOG --log-prefix '[IPv6 VNI match]:'
+	//     $ <<<"${udp_payload_hexdump}" xxd -r -p | nc -u -6 localhost 30000
+	//     $ sudo dmesg
+
+	loopback := net.IPv4(127, 0, 0, 1)
+
+	// Reserve an ephemeral UDP port for loopback testing.
+	// Binding to a TUN device would be more hermetic, but is much more effort to set up.
+	reservation, err := net.ListenUDP("udp", &net.UDPAddr{IP: loopback, Port: 0})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer reservation.Close()
+	daddr := reservation.LocalAddr().(*net.UDPAddr).AddrPort()
+
+	sender, err := net.DialUDP("udp", nil, reservation.LocalAddr().(*net.UDPAddr))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer sender.Close()
+	saddr := sender.LocalAddr().(*net.UDPAddr).AddrPort()
+
+	// There doesn't seem to be a way to receive the entire Layer-3 IPv6
+	// packet including the fixed IP header using the portable raw sockets
+	// API. That can only be done from an AF_PACKET socket, and it is
+	// unclear whether 'ld poff' would behave the same in a BPF program
+	// attached to such a socket as in an xt_bpf match.
+	c, err := net.ListenIP("ip4:udp", &net.IPAddr{IP: loopback})
+	if err != nil {
+		if errors.Is(err, unix.EPERM) {
+			t.Skip("test requires CAP_NET_RAW")
+		}
+		t.Fatal(err)
+	}
+	defer c.Close()
+
+	pc := ipv4.NewPacketConn(c)
+
+	testvectors := []uint32{
+		0,
+		1,
+		0x08,
+		42,
+		0x80,
+		0xfe,
+		0xff,
+		0x100,
+		0xfff,  // 4095
+		0x1000, // 4096
+		0x1001,
+		0x10000,
+		0xfffffe,
+		0xffffff, // Max VNI
+	}
+	for _, vni := range []uint32{1, 42, 0x100, 0x1000, 0xfffffe, 0xffffff} {
+		t.Run(fmt.Sprintf("vni=%d", vni), func(t *testing.T) {
+			setBPF(t, pc, vniMatchBPF(vni))
+
+			for _, v := range testvectors {
+				pkt := appendVXLANHeader(nil, v)
+				pkt = append(pkt, []byte{0xde, 0xad, 0xbe, 0xef}...)
+				if _, err := sender.Write(pkt); err != nil {
+					t.Fatal(err)
+				}
+
+				rpkt, ok := readUDPPacketFromRawSocket(t, pc, saddr, daddr)
+				// Sanity check: the only packets readUDPPacketFromRawSocket
+				// should return are ones we sent.
+				if ok && !bytes.Equal(pkt, rpkt) {
+					t.Fatalf("received unexpected packet: % x", rpkt)
+				}
+				if ok != (v == vni) {
+					t.Errorf("unexpected packet tagged with vni=%d (got %v, want %v)", v, ok, v == vni)
+				}
+			}
+		})
+	}
+}
+
+func appendVXLANHeader(b []byte, vni uint32) []byte {
+	// https://tools.ietf.org/html/rfc7348#section-5
+	b = append(b, []byte{0x08, 0x00, 0x00, 0x00}...)
+	return binary.BigEndian.AppendUint32(b, vni<<8)
+}
+
+func setBPF(t *testing.T, c *ipv4.PacketConn, fprog []bpf.RawInstruction) {
+	// https://natanyellin.com/posts/ebpf-filtering-done-right/
+	blockall, _ := bpf.Assemble([]bpf.Instruction{bpf.RetConstant{Val: 0}})
+	if err := c.SetBPF(blockall); err != nil {
+		t.Fatal(err)
+	}
+	ms := make([]ipv4.Message, 100)
+	for {
+		n, err := c.ReadBatch(ms, unix.MSG_DONTWAIT)
+		if err != nil {
+			if errors.Is(err, unix.EAGAIN) {
+				break
+			}
+			t.Fatal(err)
+		}
+		if n == 0 {
+			break
+		}
+	}
+
+	t.Logf("setting socket filter: %v", marshalXTBPF(fprog))
+	if err := c.SetBPF(fprog); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// readUDPPacketFromRawSocket reads raw IP packets from pc until a UDP packet
+// which matches the (src, dst) 4-tuple is found or the receive buffer is empty,
+// and returns the payload of the UDP packet.
+func readUDPPacketFromRawSocket(t *testing.T, pc *ipv4.PacketConn, src, dst netip.AddrPort) ([]byte, bool) {
+	t.Helper()
+
+	ms := []ipv4.Message{
+		{Buffers: [][]byte{make([]byte, 1500)}},
+	}
+
+	// Set a time limit to prevent an infinite loop if there is a lot of
+	// loopback traffic being captured which prevents the buffer from
+	// emptying.
+	deadline := time.Now().Add(1 * time.Second)
+	for time.Now().Before(deadline) {
+		n, err := pc.ReadBatch(ms, unix.MSG_DONTWAIT)
+		if err != nil {
+			if !errors.Is(err, unix.EAGAIN) {
+				t.Fatal(err)
+			}
+			break
+		}
+		if n == 0 {
+			break
+		}
+		pkt := ms[0].Buffers[0][:ms[0].N]
+		psrc, pdst, payload, ok := parseUDP(pkt)
+		// Discard captured packets which belong to other unrelated flows.
+		if !ok || psrc != src || pdst != dst {
+			t.Logf("discarding packet:\n% x", pkt)
+			continue
+		}
+		t.Logf("received packet (%v -> %v):\n% x", psrc, pdst, payload)
+		// While not strictly required, copy payload into a new
+		// slice which does not share a backing array with pkt
+		// so the IP and UDP headers can be garbage collected.
+		return append([]byte(nil), payload...), true
+	}
+	return nil, false
+}
+
+func parseIPv4(b []byte) (src, dst netip.Addr, protocol byte, payload []byte, ok bool) {
+	if len(b) < 20 {
+		return netip.Addr{}, netip.Addr{}, 0, nil, false
+	}
+	hlen := int(b[0]&0x0f) * 4
+	if hlen < 20 {
+		return netip.Addr{}, netip.Addr{}, 0, nil, false
+	}
+	src, _ = netip.AddrFromSlice(b[12:16])
+	dst, _ = netip.AddrFromSlice(b[16:20])
+	protocol = b[9]
+	payload = b[hlen:]
+	return src, dst, protocol, payload, true
+}
+
+// parseUDP parses the IP and UDP headers from the raw Layer-3 packet data in b.
+func parseUDP(b []byte) (src, dst netip.AddrPort, payload []byte, ok bool) {
+	srcip, dstip, protocol, ippayload, ok := parseIPv4(b)
+	if !ok {
+		return netip.AddrPort{}, netip.AddrPort{}, nil, false
+	}
+	if protocol != 17 {
+		return netip.AddrPort{}, netip.AddrPort{}, nil, false
+	}
+	if len(ippayload) < 8 {
+		return netip.AddrPort{}, netip.AddrPort{}, nil, false
+	}
+	sport := binary.BigEndian.Uint16(ippayload[0:2])
+	dport := binary.BigEndian.Uint16(ippayload[2:4])
+	src = netip.AddrPortFrom(srcip, sport)
+	dst = netip.AddrPortFrom(dstip, dport)
+	payload = ippayload[8:]
+	return src, dst, payload, true
+}