ReadBatch is named Read now

Bumps the zx2c4-dependencies group with 1 update: golang.zx2c4.com/wireguard/windows.


Updates `golang.zx2c4.com/wireguard/windows` from 0.5.3 to 0.6.1

---
updated-dependencies:
- dependency-name: golang.zx2c4.com/wireguard/windows
  dependency-version: 0.6.1
  dependency-type: direct:production
  update-type: version-update:semver-minor
  dependency-group: zx2c4-dependencies
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

.github/workflows/release.yml

@@ -209,10 +209,11 @@ jobs:
         id: create_release
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_REF_NAME: ${{ github.ref_name }}
         run: |
           cd artifacts
           gh release create \
             --verify-tag \
-            --title "Release ${{ github.ref_name }}" \
-            "${{ github.ref_name }}" \
+            --title "Release ${GITHUB_REF_NAME}" \
+            "${GITHUB_REF_NAME}" \
             SHASUM256.txt *-latest/*.zip *-latest/*.tar.gz

.github/workflows/smoke-extra.yml

@@ -30,6 +30,9 @@ jobs:
     - name: add hashicorp source
       run: wget -O- https://apt.releases.hashicorp.com/gpg | gpg --dearmor | sudo tee /usr/share/keyrings/hashicorp-archive-keyring.gpg && echo "deb [signed-by=/usr/share/keyrings/hashicorp-archive-keyring.gpg] https://apt.releases.hashicorp.com $(lsb_release -cs) main" | sudo tee /etc/apt/sources.list.d/hashicorp.list
 
+    - name: workaround AMD-V issue  # https://github.com/cri-o/packaging/pull/306
+      run: sudo rmmod kvm_amd
+
     - name: install vagrant
       run: sudo apt-get update && sudo apt-get install -y vagrant virtualbox
 

.github/workflows/smoke/smoke-vagrant.sh

@@ -29,6 +29,17 @@ docker run --name lighthouse1 --rm "$CONTAINER" -config lighthouse1.yml -test
 docker run --name host2 --rm "$CONTAINER" -config host2.yml -test
 
 vagrant up
+
+# OpenBSD: synced folders are disabled because Vagrant's rsync installer
+# uses ftp.openbsd.org which no longer hosts packages for older releases.
+# Copy build artifacts in via scp instead.
+case "$1" in
+    openbsd-*)
+        vagrant ssh -c "sudo mkdir -p /nebula" -- -T
+        tar -cf - -C build . | vagrant ssh -c "sudo tar -xf - -C /nebula && sudo chmod -R a+r /nebula" -- -T
+        ;;
+esac
+
 vagrant ssh -c "cd /nebula && /nebula/$1-nebula -config host3.yml -test" -- -T
 
 docker run --name lighthouse1 --device /dev/net/tun:/dev/net/tun --cap-add NET_ADMIN --rm "$CONTAINER" -config lighthouse1.yml 2>&1 | tee logs/lighthouse1 | sed -u 's/^/  [lighthouse1]  /' &

.github/workflows/smoke/smoke.sh

@@ -37,17 +37,18 @@ docker run --name host4 --device /dev/net/tun:/dev/net/tun --cap-add NET_ADMIN -
 sleep 1
 
 # grab tcpdump pcaps for debugging
-docker exec lighthouse1 tcpdump -i nebula1 -q -w - -U 2>logs/lighthouse1.inside.log >logs/lighthouse1.inside.pcap &
+docker exec lighthouse1 tcpdump -i tun0 -q -w - -U 2>logs/lighthouse1.inside.log >logs/lighthouse1.inside.pcap &
 docker exec lighthouse1 tcpdump -i eth0 -q -w - -U 2>logs/lighthouse1.outside.log >logs/lighthouse1.outside.pcap &
-docker exec host2 tcpdump -i nebula1 -q -w - -U 2>logs/host2.inside.log >logs/host2.inside.pcap &
+docker exec host2 tcpdump -i tun0 -q -w - -U 2>logs/host2.inside.log >logs/host2.inside.pcap &
 docker exec host2 tcpdump -i eth0 -q -w - -U 2>logs/host2.outside.log >logs/host2.outside.pcap &
-docker exec host3 tcpdump -i nebula1 -q -w - -U 2>logs/host3.inside.log >logs/host3.inside.pcap &
+docker exec host3 tcpdump -i tun0 -q -w - -U 2>logs/host3.inside.log >logs/host3.inside.pcap &
 docker exec host3 tcpdump -i eth0 -q -w - -U 2>logs/host3.outside.log >logs/host3.outside.pcap &
-docker exec host4 tcpdump -i nebula1 -q -w - -U 2>logs/host4.inside.log >logs/host4.inside.pcap &
+docker exec host4 tcpdump -i tun0 -q -w - -U 2>logs/host4.inside.log >logs/host4.inside.pcap &
 docker exec host4 tcpdump -i eth0 -q -w - -U 2>logs/host4.outside.log >logs/host4.outside.pcap &
 
 docker exec host2 ncat -nklv 0.0.0.0 2000 &
 docker exec host3 ncat -nklv 0.0.0.0 2000 &
+docker exec host4 ncat -nkluv 0.0.0.0 4000 &
 docker exec host2 ncat -e '/usr/bin/echo host2' -nkluv 0.0.0.0 3000 &
 docker exec host3 ncat -e '/usr/bin/echo host3' -nkluv 0.0.0.0 3000 &
 
@@ -119,11 +120,12 @@ echo
 echo " *** Testing conntrack"
 echo
 set -x
-# host2 can ping host3 now that host3 pinged it first
-docker exec host2 ping -c1 192.168.100.3
-# host4 can ping host2 once conntrack established
-docker exec host2 ping -c1 192.168.100.4
-docker exec host4 ping -c1 192.168.100.2
+
+# host2 speaking to host4 on UDP 4000 should allow it to reply, when firewall rules would normally not permit this
+docker exec host2 sh -c "/usr/bin/echo host2 | ncat -nuv 192.168.100.4 4000"
+docker exec host2 ncat -e '/usr/bin/echo helloagainfromhost2' -nkluv 0.0.0.0 4000 &
+sleep 1
+docker exec host4 sh -c "/usr/bin/echo host4 | ncat -nuv 192.168.100.2 4000"
 
 docker exec host4 sh -c 'kill 1'
 docker exec host3 sh -c 'kill 1'

.github/workflows/smoke/vagrant-openbsd-amd64/Vagrantfile

@@ -3,5 +3,5 @@
 Vagrant.configure("2") do |config|
   config.vm.box = "generic/openbsd7"
 
-  config.vm.synced_folder "../build", "/nebula", type: "rsync"
+  config.vm.synced_folder ".", "/vagrant", disabled: true
 end

CODEOWNERS

@@ -0,0 +1 @@
+#ECCN:Open Source

README.md

@@ -57,7 +57,7 @@ Check the [releases](https://github.com/slackhq/nebula/releases/latest) page for
     docker pull nebulaoss/nebula
     ```
 
-#### Mobile
+#### Mobile ([source code](https://github.com/DefinedNet/mobile_nebula))
 
 - [iOS](https://apps.apple.com/us/app/mobile-nebula/id1509587936?itsct=apps_box&itscg=30200)
 - [Android](https://play.google.com/store/apps/details?id=net.defined.mobile_nebula&pcampaignid=pcampaignidMKT-Other-global-all-co-prtnr-py-PartBadge-Mar2515-1)
@@ -76,6 +76,8 @@ Nebula was created to provide a mechanism for groups of hosts to communicate sec
 
 ## Getting started (quickly)
 
+**Don't want to manage your own PKI and lighthouses?** [Managed Nebula](https://www.defined.net/) from Defined Networking handles all of this for you.
+
 To set up a Nebula network, you'll need:
 
 #### 1. The [Nebula binaries](https://github.com/slackhq/nebula/releases) or [Distribution Packages](https://github.com/slackhq/nebula#distribution-packages) for your specific platform. Specifically you'll need `nebula-cert` and the specific nebula binary for each platform you use.

batch.go

@@ -0,0 +1,70 @@
+package nebula
+
+import "net/netip"
+
+// sendBatchCap is the maximum number of encrypted packets accumulated before a
+// flush is forced. TSO superpackets segment to at most ~45 packets on
+// reasonable MTUs, so 128 leaves headroom without bloating the backing
+// allocation.
+const sendBatchCap = 128
+
+// sendBatch accumulates encrypted UDP packets for a single sendmmsg flush.
+// One sendBatch is owned by each listenIn goroutine; no locking is needed.
+// The backing storage holds up to batchCap packets of slotCap bytes each;
+// bufs and dsts are parallel slices of committed slots.
+type sendBatch struct {
+	bufs     [][]byte
+	dsts     []netip.AddrPort
+	backing  []byte
+	slotCap  int
+	batchCap int
+	nextSlot int
+}
+
+func newSendBatch(batchCap, slotCap int) *sendBatch {
+	return &sendBatch{
+		bufs:     make([][]byte, 0, batchCap),
+		dsts:     make([]netip.AddrPort, 0, batchCap),
+		backing:  make([]byte, batchCap*slotCap),
+		slotCap:  slotCap,
+		batchCap: batchCap,
+	}
+}
+
+// Next returns a zero-length slice with slotCap capacity over the next unused
+// slot's backing bytes. The caller writes into the returned slice and then
+// calls Commit with the final length and destination. Next returns nil when
+// the batch is full.
+func (b *sendBatch) Next() []byte {
+	if b.nextSlot >= b.batchCap {
+		return nil
+	}
+	start := b.nextSlot * b.slotCap
+	return b.backing[start : start : start+b.slotCap]
+}
+
+// Commit records the slot just returned by Next as a packet of length n
+// destined for dst.
+func (b *sendBatch) Commit(n int, dst netip.AddrPort) {
+	start := b.nextSlot * b.slotCap
+	b.bufs = append(b.bufs, b.backing[start:start+n])
+	b.dsts = append(b.dsts, dst)
+	b.nextSlot++
+}
+
+// Reset clears committed slots; backing storage is retained for reuse.
+func (b *sendBatch) Reset() {
+	b.bufs = b.bufs[:0]
+	b.dsts = b.dsts[:0]
+	b.nextSlot = 0
+}
+
+// Len returns the number of committed packets.
+func (b *sendBatch) Len() int {
+	return len(b.bufs)
+}
+
+// Cap returns the maximum number of slots in the batch.
+func (b *sendBatch) Cap() int {
+	return b.batchCap
+}

batch_test.go

@@ -0,0 +1,137 @@
+package nebula
+
+import (
+	"net/netip"
+	"testing"
+)
+
+func TestSendBatchBookkeeping(t *testing.T) {
+	b := newSendBatch(4, 32)
+	if b.Len() != 0 || b.Cap() != 4 {
+		t.Fatalf("fresh batch: len=%d cap=%d", b.Len(), b.Cap())
+	}
+
+	ap := netip.MustParseAddrPort("10.0.0.1:4242")
+	for i := 0; i < 4; i++ {
+		slot := b.Next()
+		if slot == nil {
+			t.Fatalf("slot %d: Next returned nil before cap", i)
+		}
+		if cap(slot) != 32 || len(slot) != 0 {
+			t.Fatalf("slot %d: got len=%d cap=%d want len=0 cap=32", i, len(slot), cap(slot))
+		}
+		// Write a marker byte.
+		slot = append(slot, byte(i), byte(i+1), byte(i+2))
+		b.Commit(len(slot), ap)
+	}
+	if b.Next() != nil {
+		t.Fatalf("Next should return nil when full")
+	}
+	if b.Len() != 4 {
+		t.Fatalf("Len=%d want 4", b.Len())
+	}
+	for i, buf := range b.bufs {
+		if len(buf) != 3 || buf[0] != byte(i) {
+			t.Errorf("buf %d: %x", i, buf)
+		}
+		if b.dsts[i] != ap {
+			t.Errorf("dst %d: got %v want %v", i, b.dsts[i], ap)
+		}
+	}
+
+	// Reset returns empty and Next works again.
+	b.Reset()
+	if b.Len() != 0 {
+		t.Fatalf("after Reset Len=%d want 0", b.Len())
+	}
+	slot := b.Next()
+	if slot == nil || cap(slot) != 32 {
+		t.Fatalf("after Reset Next nil or wrong cap: %v cap=%d", slot == nil, cap(slot))
+	}
+}
+
+func TestBatchSegmentable(t *testing.T) {
+	ap := netip.MustParseAddrPort("10.0.0.1:4242")
+	other := netip.MustParseAddrPort("10.0.0.2:4242")
+
+	mk := func(addrs []netip.AddrPort, sizes []int) *sendBatch {
+		b := newSendBatch(len(addrs), 64)
+		for i, a := range addrs {
+			s := b.Next()
+			for j := 0; j < sizes[i]; j++ {
+				s = append(s, byte(j))
+			}
+			b.Commit(len(s), a)
+		}
+		return b
+	}
+
+	t.Run("uniform same dst", func(t *testing.T) {
+		b := mk([]netip.AddrPort{ap, ap, ap}, []int{10, 10, 10})
+		seg, ok := batchSegmentable(b)
+		if !ok || seg != 10 {
+			t.Fatalf("got seg=%d ok=%v", seg, ok)
+		}
+	})
+
+	t.Run("last segment short ok", func(t *testing.T) {
+		b := mk([]netip.AddrPort{ap, ap, ap}, []int{10, 10, 4})
+		seg, ok := batchSegmentable(b)
+		if !ok || seg != 10 {
+			t.Fatalf("got seg=%d ok=%v", seg, ok)
+		}
+	})
+
+	t.Run("mixed dst rejected", func(t *testing.T) {
+		b := mk([]netip.AddrPort{ap, other, ap}, []int{10, 10, 10})
+		if _, ok := batchSegmentable(b); ok {
+			t.Fatalf("expected rejection for mixed dst")
+		}
+	})
+
+	t.Run("mid-batch short rejected", func(t *testing.T) {
+		b := mk([]netip.AddrPort{ap, ap, ap}, []int{10, 4, 10})
+		if _, ok := batchSegmentable(b); ok {
+			t.Fatalf("expected rejection for short mid-batch")
+		}
+	})
+
+	t.Run("mid-batch longer rejected", func(t *testing.T) {
+		b := mk([]netip.AddrPort{ap, ap, ap}, []int{10, 11, 10})
+		if _, ok := batchSegmentable(b); ok {
+			t.Fatalf("expected rejection for longer mid-batch")
+		}
+	})
+
+	t.Run("last longer rejected", func(t *testing.T) {
+		b := mk([]netip.AddrPort{ap, ap, ap}, []int{10, 10, 11})
+		if _, ok := batchSegmentable(b); ok {
+			t.Fatalf("expected rejection for longer last segment")
+		}
+	})
+
+	t.Run("first zero rejected", func(t *testing.T) {
+		b := mk([]netip.AddrPort{ap, ap}, []int{0, 10})
+		if _, ok := batchSegmentable(b); ok {
+			t.Fatalf("expected rejection for zero first")
+		}
+	})
+}
+
+func TestSendBatchSlotsDoNotOverlap(t *testing.T) {
+	b := newSendBatch(3, 8)
+	ap := netip.MustParseAddrPort("10.0.0.1:80")
+
+	// Fill three slots, each with its own sentinel byte.
+	for i := 0; i < 3; i++ {
+		s := b.Next()
+		s = append(s, byte(0xA0+i), byte(0xB0+i))
+		b.Commit(len(s), ap)
+	}
+
+	for i, buf := range b.bufs {
+		if buf[0] != byte(0xA0+i) || buf[1] != byte(0xB0+i) {
+			t.Errorf("slot %d corrupted: %x", i, buf)
+		}
+	}
+}

boring.go

@@ -1,5 +1,4 @@
 //go:build boringcrypto
-// +build boringcrypto
 
 package nebula
 

cert/ca_pool.go

@@ -1,11 +1,14 @@
 package cert
 
 import (
+	"bufio"
+	"bytes"
+	"encoding/pem"
 	"errors"
 	"fmt"
+	"io"
 	"net/netip"
 	"slices"
-	"strings"
 	"time"
 )
 
@@ -29,22 +32,46 @@ func NewCAPool() *CAPool {
 // If the pool contains any expired certificates, an ErrExpired will be
 // returned along with the pool. The caller must handle any such errors.
 func NewCAPoolFromPEM(caPEMs []byte) (*CAPool, error) {
+	return NewCAPoolFromPEMReader(bytes.NewReader(caPEMs))
+}
+
+// NewCAPoolFromPEMReader will create a new CA pool from the provided reader.
+// The reader must contain a PEM-encoded set of nebula certificates.
+func NewCAPoolFromPEMReader(r io.Reader) (*CAPool, error) {
 	pool := NewCAPool()
-	var err error
+
 	var expired bool
-	for {
-		caPEMs, err = pool.AddCAFromPEM(caPEMs)
-		if errors.Is(err, ErrExpired) {
-			expired = true
-			err = nil
+
+	scanner := bufio.NewScanner(r)
+	scanner.Split(SplitPEM)
+
+	for scanner.Scan() {
+		pemBytes := scanner.Bytes()
+
+		block, rest := pem.Decode(pemBytes)
+		if len(bytes.TrimSpace(rest)) > 0 {
+			return nil, ErrInvalidPEMBlock
 		}
+		if block == nil {
+			return nil, ErrInvalidPEMBlock
+		}
+
+		c, err := unmarshalCertificateBlock(block)
 		if err != nil {
 			return nil, err
 		}
-		if len(caPEMs) == 0 || strings.TrimSpace(string(caPEMs)) == "" {
-			break
+
+		err = pool.AddCA(c)
+		if errors.Is(err, ErrExpired) {
+			expired = true
+			continue
+		} else if err != nil {
+			return nil, err
 		}
 	}
+	if err := scanner.Err(); err != nil {
+		return nil, ErrInvalidPEMBlock
+	}
 
 	if expired {
 		return pool, ErrExpired

cert/ca_pool_test.go

@@ -1,7 +1,10 @@
 package cert
 
 import (
+	"bytes"
+	"io"
 	"net/netip"
+	"strings"
 	"testing"
 	"time"
 
@@ -112,6 +115,60 @@ k+coOv04r+zh33ISyhbsafnYduN17p2eD7CmHvHuerguXD9f32gcxo/KsFCKEjMe
 	assert.Len(t, ppppp.CAs, 1)
 }
 
+// oneByteReader wraps a reader to return at most 1 byte per Read call,
+// exercising the streaming accumulation logic in NewCAPoolFromPEMReader.
+type oneByteReader struct {
+	r io.Reader
+}
+
+func (o *oneByteReader) Read(p []byte) (int, error) {
+	if len(p) == 0 {
+		return 0, nil
+	}
+	return o.r.Read(p[:1])
+}
+
+func TestNewCAPoolFromPEMReader_EmptyReader(t *testing.T) {
+	pool, err := NewCAPoolFromPEMReader(bytes.NewReader(nil))
+	require.NoError(t, err)
+	assert.Empty(t, pool.CAs)
+
+	pool, err = NewCAPoolFromPEMReader(strings.NewReader("   \n\t\n  "))
+	require.NoError(t, err)
+	assert.Empty(t, pool.CAs)
+}
+
+func TestNewCAPoolFromPEMReader_OneByteReads(t *testing.T) {
+	ca1, _, _, pem1 := NewTestCaCert(Version2, Curve_CURVE25519, time.Now(), time.Now().Add(time.Hour), nil, nil, nil)
+	ca2, _, _, pem2 := NewTestCaCert(Version2, Curve_CURVE25519, time.Now(), time.Now().Add(time.Hour), nil, nil, nil)
+
+	bundle := append(pem1, pem2...)
+	pool, err := NewCAPoolFromPEMReader(&oneByteReader{r: bytes.NewReader(bundle)})
+	require.NoError(t, err)
+	assert.Len(t, pool.CAs, 2)
+
+	fp1, err := ca1.Fingerprint()
+	require.NoError(t, err)
+	fp2, err := ca2.Fingerprint()
+	require.NoError(t, err)
+
+	assert.Contains(t, pool.CAs, fp1)
+	assert.Contains(t, pool.CAs, fp2)
+}
+
+func TestNewCAPoolFromPEMReader_TruncatedPEM(t *testing.T) {
+	_, err := NewCAPoolFromPEMReader(strings.NewReader("-----BEGIN NEBULA CERTIFICATE-----\npartialdata"))
+	assert.ErrorIs(t, err, ErrInvalidPEMBlock)
+}
+
+func TestNewCAPoolFromPEMReader_TrailingGarbage(t *testing.T) {
+	_, _, _, pem1 := NewTestCaCert(Version2, Curve_CURVE25519, time.Now(), time.Now().Add(time.Hour), nil, nil, nil)
+
+	bundle := append(pem1, []byte("some trailing garbage")...)
+	_, err := NewCAPoolFromPEMReader(bytes.NewReader(bundle))
+	assert.ErrorIs(t, err, ErrInvalidPEMBlock)
+}
+
 func TestCertificateV1_Verify(t *testing.T) {
 	ca, _, caKey, _ := NewTestCaCert(Version1, Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, nil)
 	c, _, _, _ := NewTestCert(Version1, Curve_CURVE25519, ca, caKey, "test cert", time.Now(), time.Now().Add(5*time.Minute), nil, nil, nil)

cert/p256/p256.go

@@ -44,7 +44,12 @@ func swap(r, s []byte) ([]byte, []byte, error) {
 	}
 	sNormalized := nMod.Nat().Sub(bigS, nMod)
 
-	return r, sNormalized.Bytes(nMod), nil
+	result := sNormalized.Bytes(nMod)
+	for len(result) > 1 && result[0] == 0 {
+		result = result[1:]
+	}
+
+	return r, result, nil
 }
 
 func Normalize(sig []byte) ([]byte, error) {

cert/pem.go

@@ -1,12 +1,66 @@
 package cert
 
 import (
+	"bytes"
 	"encoding/pem"
+	"errors"
 	"fmt"
 
 	"golang.org/x/crypto/ed25519"
 )
 
+var ErrTruncatedPEMBlock = errors.New("truncated PEM block")
+
+// SplitPEM is a split function for bufio.Scanner that returns each PEM block.
+func SplitPEM(data []byte, atEOF bool) (advance int, token []byte, err error) {
+	// Look for the start of a PEM block
+	start := bytes.Index(data, []byte("-----BEGIN "))
+	if start == -1 {
+		if atEOF && len(bytes.TrimSpace(data)) > 0 {
+			// Non-whitespace content with no PEM block
+			return 0, nil, ErrTruncatedPEMBlock
+		}
+		if atEOF {
+			return len(data), nil, nil
+		}
+		// Request more data
+		return 0, nil, nil
+	}
+
+	// Look for the end marker
+	endMarkerStart := bytes.Index(data[start:], []byte("-----END "))
+	if endMarkerStart == -1 {
+		if atEOF {
+			// Incomplete PEM block at EOF
+			return 0, nil, ErrTruncatedPEMBlock
+		}
+		// Need more data to find the end
+		return 0, nil, nil
+	}
+
+	// Find the actual end of the END line (after the newline)
+	endMarkerStart += start
+	endLineEnd := bytes.IndexByte(data[endMarkerStart:], '\n')
+	var end int
+	if endLineEnd == -1 {
+		if atEOF {
+			// END marker without newline at EOF - take it anyway
+			end = len(data)
+		} else {
+			// Need more data
+			return 0, nil, nil
+		}
+	} else {
+		end = endMarkerStart + endLineEnd + 1
+	}
+
+	// Extract the PEM block
+	pemBlock := data[start:end]
+
+	// Return the valid PEM block
+	return end, pemBlock, nil
+}
+
 const ( //cert banners
 	CertificateBanner   = "NEBULA CERTIFICATE"
 	CertificateV2Banner = "NEBULA CERTIFICATE V2"
@@ -37,19 +91,7 @@ func UnmarshalCertificateFromPEM(b []byte) (Certificate, []byte, error) {
 		return nil, r, ErrInvalidPEMBlock
 	}
 
-	var c Certificate
-	var err error
-
-	switch p.Type {
-	// Implementations must validate the resulting certificate contains valid information
-	case CertificateBanner:
-		c, err = unmarshalCertificateV1(p.Bytes, nil)
-	case CertificateV2Banner:
-		c, err = unmarshalCertificateV2(p.Bytes, nil, Curve_CURVE25519)
-	default:
-		return nil, r, ErrInvalidPEMCertificateBanner
-	}
-
+	c, err := unmarshalCertificateBlock(p)
 	if err != nil {
 		return nil, r, err
 	}
@@ -58,6 +100,20 @@ func UnmarshalCertificateFromPEM(b []byte) (Certificate, []byte, error) {
 
 }
 
+// unmarshalCertificateBlock decodes a single PEM block into a certificate.
+// It expects a Nebula certificate banner and returns ErrInvalidPEMCertificateBanner otherwise.
+func unmarshalCertificateBlock(block *pem.Block) (Certificate, error) {
+	switch block.Type {
+	// Implementations must validate the resulting certificate contains valid information
+	case CertificateBanner:
+		return unmarshalCertificateV1(block.Bytes, nil)
+	case CertificateV2Banner:
+		return unmarshalCertificateV2(block.Bytes, nil, Curve_CURVE25519)
+	default:
+		return nil, ErrInvalidPEMCertificateBanner
+	}
+}
+
 func marshalCertPublicKeyToPEM(c Certificate) []byte {
 	if c.IsCA() {
 		return MarshalSigningPublicKeyToPEM(c.Curve(), c.PublicKey())

cert/pem_test.go

@@ -1,12 +1,88 @@
 package cert
 
 import (
+	"bufio"
+	"strings"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
 
+func scanAll(t *testing.T, input string) ([]string, error) {
+	t.Helper()
+	scanner := bufio.NewScanner(strings.NewReader(input))
+	scanner.Split(SplitPEM)
+	var blocks []string
+	for scanner.Scan() {
+		blocks = append(blocks, scanner.Text())
+	}
+	return blocks, scanner.Err()
+}
+
+func TestSplitPEM_Single(t *testing.T) {
+	input := "-----BEGIN TEST-----\ndata\n-----END TEST-----\n"
+	blocks, err := scanAll(t, input)
+	require.NoError(t, err)
+	require.Len(t, blocks, 1)
+	require.Equal(t, input, blocks[0])
+}
+
+func TestSplitPEM_Multiple(t *testing.T) {
+	block1 := "-----BEGIN TEST-----\naaa\n-----END TEST-----\n"
+	block2 := "-----BEGIN TEST-----\nbbb\n-----END TEST-----\n"
+	blocks, err := scanAll(t, block1+block2)
+	require.NoError(t, err)
+	require.Len(t, blocks, 2)
+	require.Equal(t, block1, blocks[0])
+	require.Equal(t, block2, blocks[1])
+}
+
+func TestSplitPEM_CommentsAndWhitespaceBetweenBlocks(t *testing.T) {
+	input := "# comment\n\n-----BEGIN TEST-----\naaa\n-----END TEST-----\n\n# another comment\n\n-----BEGIN TEST-----\nbbb\n-----END TEST-----\n"
+	blocks, err := scanAll(t, input)
+	require.NoError(t, err)
+	require.Len(t, blocks, 2)
+}
+
+func TestSplitPEM_Empty(t *testing.T) {
+	blocks, err := scanAll(t, "")
+	require.NoError(t, err)
+	require.Empty(t, blocks)
+}
+
+func TestSplitPEM_WhitespaceOnly(t *testing.T) {
+	blocks, err := scanAll(t, "  \n\t\n  ")
+	require.NoError(t, err)
+	require.Empty(t, blocks)
+}
+
+func TestSplitPEM_TrailingGarbage(t *testing.T) {
+	input := "-----BEGIN TEST-----\ndata\n-----END TEST-----\ngarbage"
+	blocks, err := scanAll(t, input)
+	require.ErrorIs(t, err, ErrTruncatedPEMBlock)
+	require.Len(t, blocks, 1)
+}
+
+func TestSplitPEM_TruncatedBlock(t *testing.T) {
+	input := "-----BEGIN TEST-----\npartial data with no end"
+	_, err := scanAll(t, input)
+	require.ErrorIs(t, err, ErrTruncatedPEMBlock)
+}
+
+func TestSplitPEM_NoEndNewline(t *testing.T) {
+	input := "-----BEGIN TEST-----\ndata\n-----END TEST-----"
+	blocks, err := scanAll(t, input)
+	require.NoError(t, err)
+	require.Len(t, blocks, 1)
+	require.Equal(t, input, blocks[0])
+}
+
+func TestSplitPEM_GarbageOnly(t *testing.T) {
+	_, err := scanAll(t, "this is not PEM data")
+	require.ErrorIs(t, err, ErrTruncatedPEMBlock)
+}
+
 func TestUnmarshalCertificateFromPEM(t *testing.T) {
 	goodCert := []byte(`
 # A good cert

cmd/nebula-cert/verify.go

@@ -6,7 +6,6 @@ import (
 	"fmt"
 	"io"
 	"os"
-	"strings"
 	"time"
 
 	"github.com/slackhq/nebula/cert"
@@ -40,21 +39,15 @@ func verify(args []string, out io.Writer, errOut io.Writer) error {
 		return err
 	}
 
-	rawCACert, err := os.ReadFile(*vf.caPath)
+	caFile, err := os.Open(*vf.caPath)
 	if err != nil {
 		return fmt.Errorf("error while reading ca: %w", err)
 	}
+	defer caFile.Close()
 
-	caPool := cert.NewCAPool()
-	for {
-		rawCACert, err = caPool.AddCAFromPEM(rawCACert)
-		if err != nil {
-			return fmt.Errorf("error while adding ca cert to pool: %w", err)
-		}
-
-		if rawCACert == nil || len(rawCACert) == 0 || strings.TrimSpace(string(rawCACert)) == "" {
-			break
-		}
+	caPool, err := cert.NewCAPoolFromPEMReader(caFile)
+	if err != nil && !errors.Is(err, cert.ErrExpired) {
+		return fmt.Errorf("error while adding ca cert to pool: %w", err)
 	}
 
 	rawCert, err := os.ReadFile(*vf.certPath)

cmd/nebula-cert/verify_test.go

@@ -64,7 +64,7 @@ func Test_verify(t *testing.T) {
 	err = verify([]string{"-ca", caFile.Name(), "-crt", "does_not_exist"}, ob, eb)
 	assert.Empty(t, ob.String())
 	assert.Empty(t, eb.String())
-	require.EqualError(t, err, "error while adding ca cert to pool: input did not contain a valid PEM encoded block")
+	require.ErrorIs(t, err, cert.ErrInvalidPEMBlock)
 
 	// make a ca for later
 	caPub, caPriv, _ := ed25519.GenerateKey(rand.Reader)

cmd/nebula-service/main.go

@@ -78,8 +78,21 @@ func main() {
 	}
 
 	if !*configTest {
-		ctrl.Start()
-		ctrl.ShutdownBlock()
+		wait, err := ctrl.Start()
+		if err != nil {
+			util.LogWithContextIfNeeded("Error while running", err, l)
+			os.Exit(1)
+		}
+
+		go ctrl.ShutdownBlock()
+
+		if err := wait(); err != nil {
+			l.WithError(err).Error("Nebula stopped due to fatal error")
+			l.Info("Goodbye")
+			os.Exit(2)
+		}
+
+		l.Info("Goodbye")
 	}
 
 	os.Exit(0)

cmd/nebula/main.go

@@ -72,9 +72,22 @@ func main() {
 	}
 
 	if !*configTest {
-		ctrl.Start()
+		wait, err := ctrl.Start()
+		if err != nil {
+			util.LogWithContextIfNeeded("Error while running", err, l)
+			os.Exit(1)
+		}
+
+		go ctrl.ShutdownBlock()
 		notifyReady(l)
-		ctrl.ShutdownBlock()
+
+		if err := wait(); err != nil {
+			l.WithError(err).Error("Nebula stopped due to fatal error")
+			l.Info("Goodbye")
+			os.Exit(2)
+		}
+
+		l.Info("Goodbye")
 	}
 
 	os.Exit(0)

connection_manager_test.go

@@ -10,6 +10,7 @@ import (
 	"github.com/flynn/noise"
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay"
 	"github.com/slackhq/nebula/test"
 	"github.com/slackhq/nebula/udp"
 	"github.com/stretchr/testify/assert"
@@ -52,7 +53,7 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 	lh := newTestLighthouse()
 	ifce := &Interface{
 		hostMap:          hostMap,
-		inside:           &test.NoopTun{},
+		inside:           &overlay.NoopTun{},
 		outside:          &udp.NoopConn{},
 		firewall:         &Firewall{},
 		lightHouse:       lh,
@@ -135,7 +136,7 @@ func Test_NewConnectionManagerTest2(t *testing.T) {
 	lh := newTestLighthouse()
 	ifce := &Interface{
 		hostMap:          hostMap,
-		inside:           &test.NoopTun{},
+		inside:           &overlay.NoopTun{},
 		outside:          &udp.NoopConn{},
 		firewall:         &Firewall{},
 		lightHouse:       lh,
@@ -220,7 +221,7 @@ func Test_NewConnectionManager_DisconnectInactive(t *testing.T) {
 	lh := newTestLighthouse()
 	ifce := &Interface{
 		hostMap:          hostMap,
-		inside:           &test.NoopTun{},
+		inside:           &overlay.NoopTun{},
 		outside:          &udp.NoopConn{},
 		firewall:         &Firewall{},
 		lightHouse:       lh,
@@ -347,7 +348,7 @@ func Test_NewConnectionManagerTest_DisconnectInvalid(t *testing.T) {
 	lh := newTestLighthouse()
 	ifce := &Interface{
 		hostMap:          hostMap,
-		inside:           &test.NoopTun{},
+		inside:           &overlay.NoopTun{},
 		outside:          &udp.NoopConn{},
 		firewall:         &Firewall{},
 		lightHouse:       lh,

control.go

@@ -2,9 +2,11 @@ package nebula
 
 import (
 	"context"
+	"errors"
 	"net/netip"
 	"os"
 	"os/signal"
+	"sync"
 	"syscall"
 
 	"github.com/sirupsen/logrus"
@@ -13,6 +15,16 @@ import (
 	"github.com/slackhq/nebula/overlay"
 )
 
+type RunState int
+
+const (
+	Stopped  RunState = 0 // The control has yet to be started
+	Started  RunState = 1 // The control has been started
+	Stopping RunState = 2 // The control is stopping
+)
+
+var ErrAlreadyStarted = errors.New("nebula is already started")
+
 // Every interaction here needs to take extra care to copy memory and not return or use arguments "as is" when touching
 // core. This means copying IP objects, slices, de-referencing pointers and taking the actual value, etc
 
@@ -26,6 +38,9 @@ type controlHostLister interface {
 }
 
 type Control struct {
+	stateLock sync.Mutex
+	state     RunState
+
 	f                      *Interface
 	l                      *logrus.Logger
 	ctx                    context.Context
@@ -49,10 +64,24 @@ type ControlHostInfo struct {
 	CurrentRelaysThroughMe []netip.Addr     `json:"currentRelaysThroughMe"`
 }
 
-// Start actually runs nebula, this is a nonblocking call. To block use Control.ShutdownBlock()
-func (c *Control) Start() {
+// Start actually runs nebula, this is a nonblocking call.
+// The returned function blocks until nebula has fully stopped and returns the
+// first fatal reader error (if any). A nil error means nebula shut down
+// gracefully; a non-nil error means a reader hit an unexpected failure that
+// triggered the shutdown.
+func (c *Control) Start() (func() error, error) {
+	c.stateLock.Lock()
+	if c.state != Stopped {
+		c.stateLock.Unlock()
+		return nil, ErrAlreadyStarted
+	}
+
 	// Activate the interface
-	c.f.activate()
+	err := c.f.activate()
+	if err != nil {
+		c.stateLock.Unlock()
+		return nil, err
+	}
 
 	// Call all the delayed funcs that waited patiently for the interface to be created.
 	if c.sshStart != nil {
@@ -71,16 +100,36 @@ func (c *Control) Start() {
 		c.lighthouseStart()
 	}
 
+	c.f.triggerShutdown = c.Stop
+
 	// Start reading packets.
-	c.f.run()
+	c.state = Started
+	c.stateLock.Unlock()
+	return c.f.run()
+}
+
+func (c *Control) State() RunState {
+	c.stateLock.Lock()
+	defer c.stateLock.Unlock()
+	return c.state
 }
 
 func (c *Control) Context() context.Context {
 	return c.ctx
 }
 
-// Stop signals nebula to shutdown and close all tunnels, returns after the shutdown is complete
+// Stop is a non-blocking call that signals nebula to close all tunnels and shut down
 func (c *Control) Stop() {
+	c.stateLock.Lock()
+	if c.state != Started {
+		c.stateLock.Unlock()
+		// We are stopping or stopped already
+		return
+	}
+
+	c.state = Stopping
+	c.stateLock.Unlock()
+
 	// Stop the handshakeManager (and other services), to prevent new tunnels from
 	// being created while we're shutting them all down.
 	c.cancel()
@@ -89,7 +138,9 @@ func (c *Control) Stop() {
 	if err := c.f.Close(); err != nil {
 		c.l.WithError(err).Error("Close interface failed")
 	}
-	c.l.Info("Goodbye")
+	c.stateLock.Lock()
+	c.state = Stopped
+	c.stateLock.Unlock()
 }
 
 // ShutdownBlock will listen for and block on term and interrupt signals, calling Control.Stop() once signalled

control_tester.go

@@ -1,5 +1,4 @@
 //go:build e2e_testing
-// +build e2e_testing
 
 package nebula
 

dist/wireshark/nebula.lua

@@ -84,30 +84,24 @@ end
 
 function nebula.prefs_changed()
     if default_settings.all_ports == nebula.prefs.all_ports and default_settings.port == nebula.prefs.port then
-        -- Nothing changed, bail
         return
     end
 
-    -- Remove our old dissector
+    -- Remove all existing registrations
     DissectorTable.get("udp.port"):remove_all(nebula)
 
-    if nebula.prefs.all_ports and default_settings.all_ports ~= nebula.prefs.all_ports then
-        default_settings.all_port = nebula.prefs.all_ports
-
+    if nebula.prefs.all_ports then
+        -- Register on every port for hole punch capture
         for i=0, 65535 do
             DissectorTable.get("udp.port"):add(i, nebula)
         end
-
-        -- no need to establish again on specific ports
-        return
+    else
+        -- Register on the configured port only
+        DissectorTable.get("udp.port"):add(nebula.prefs.port, nebula)
     end
 
-
-    if default_settings.all_ports ~= nebula.prefs.all_ports then
-        -- Add our new port dissector
-        default_settings.port = nebula.prefs.port
-        DissectorTable.get("udp.port"):add(default_settings.port, nebula)
-    end
+    default_settings.all_ports = nebula.prefs.all_ports
+    default_settings.port = nebula.prefs.port
 end
 
 DissectorTable.get("udp.port"):add(default_settings.port, nebula)

e2e/handshake_manager_test.go

@@ -0,0 +1,565 @@
+//go:build e2e_testing
+// +build e2e_testing
+
+package e2e
+
+import (
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/e2e/router"
+	"github.com/slackhq/nebula/header"
+	"github.com/slackhq/nebula/udp"
+	"github.com/stretchr/testify/assert"
+)
+
+// makeHandshakePacket creates a handshake packet with the given parameters.
+func makeHandshakePacket(from, to netip.AddrPort, subtype header.MessageSubType, remoteIndex uint32, counter uint64) *udp.Packet {
+	data := make([]byte, 200)
+	header.Encode(data, header.Version, header.Handshake, subtype, remoteIndex, counter)
+	for i := header.Len; i < len(data); i++ {
+		data[i] = byte(i)
+	}
+	return &udp.Packet{To: to, From: from, Data: data}
+}
+
+func TestHandshakeRetransmitDuplicate(t *testing.T) {
+	// Verify the responder correctly handles receiving the same msg1 multiple times
+	// (retransmission). The duplicate goes through CheckAndComplete -> ErrAlreadySeen
+	// and the cached response is resent.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	myControl.Start()
+	theirControl.Start()
+
+	r := router.NewR(t, myControl, theirControl)
+	defer r.RenderFlow()
+
+	t.Log("Trigger handshake from me to them")
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi"))
+
+	t.Log("Grab my msg1")
+	msg1 := myControl.GetFromUDP(true)
+
+	t.Log("Inject msg1 into them, first time")
+	theirControl.InjectUDPPacket(msg1)
+	_ = theirControl.GetFromUDP(true)
+
+	t.Log("Inject the SAME msg1 again, tests ErrAlreadySeen path")
+	theirControl.InjectUDPPacket(msg1)
+	resp2 := theirControl.GetFromUDP(true)
+	assert.NotNil(t, resp2, "should get cached response on duplicate msg1")
+
+	t.Log("Complete handshake with cached response")
+	myControl.InjectUDPPacket(resp2)
+	myControl.WaitForType(1, 0, theirControl)
+
+	t.Log("Drain cached packet and verify tunnel works")
+	cachedPacket := theirControl.GetFromTun(true)
+	assertUdpPacket(t, []byte("Hi"), cachedPacket, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), 80, 80)
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	t.Log("Verify only one tunnel exists on each side")
+	assert.Len(t, myControl.ListHostmapHosts(false), 1)
+	assert.Len(t, theirControl.ListHostmapHosts(false), 1)
+
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+func TestHandshakeTruncatedPacketRecovery(t *testing.T) {
+	// Verify that a truncated handshake packet is ignored and the real
+	// packet can still complete the handshake.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	myControl.Start()
+	theirControl.Start()
+
+	r := router.NewR(t, myControl, theirControl)
+	defer r.RenderFlow()
+
+	t.Log("Trigger handshake")
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi"))
+
+	t.Log("Get msg1 and deliver to responder")
+	msg1 := myControl.GetFromUDP(true)
+	theirControl.InjectUDPPacket(msg1)
+
+	t.Log("Get the real response")
+	realResp := theirControl.GetFromUDP(true)
+
+	t.Log("Truncate the response and inject, should be ignored")
+	truncResp := realResp.Copy()
+	truncResp.Data = truncResp.Data[:header.Len]
+	myControl.InjectUDPPacket(truncResp)
+
+	t.Log("Verify pending handshake survived the truncated packet")
+	assert.NotEmpty(t, myControl.ListHostmapHosts(true), "pending handshake should still exist")
+
+	t.Log("Inject real response, should complete handshake")
+	myControl.InjectUDPPacket(realResp)
+	myControl.WaitForType(1, 0, theirControl)
+
+	t.Log("Drain and verify tunnel")
+	cachedPacket := theirControl.GetFromTun(true)
+	assertUdpPacket(t, []byte("Hi"), cachedPacket, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), 80, 80)
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+func TestHandshakeOrphanedMsg2Dropped(t *testing.T) {
+	// A msg2 arriving with no matching pending index should be silently dropped
+	// with no response sent and no state changes.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	myControl.Start()
+	theirControl.Start()
+
+	r := router.NewR(t, myControl, theirControl)
+	defer r.RenderFlow()
+
+	t.Log("Complete a normal handshake")
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi"))
+	r.RouteForAllUntilTxTun(theirControl)
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	t.Log("Record hostmap state")
+	myIndexes := len(myControl.ListHostmapIndexes(false))
+
+	t.Log("Inject a fake msg2 with unknown RemoteIndex")
+	myControl.InjectUDPPacket(makeHandshakePacket(theirUdpAddr, myUdpAddr, header.HandshakeIXPSK0, 0xDEADBEEF, 2))
+
+	t.Log("Verify no new indexes created")
+	assert.Equal(t, myIndexes, len(myControl.ListHostmapIndexes(false)))
+
+	t.Log("Verify no UDP response was sent")
+	time.Sleep(100 * time.Millisecond)
+	assert.Nil(t, myControl.GetFromUDP(false), "should not send a response to orphaned msg2")
+
+	t.Log("Verify existing tunnel still works")
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+func TestHandshakeUnknownMessageCounter(t *testing.T) {
+	// A handshake packet with an unexpected message counter should be silently
+	// dropped with no side effects and no UDP response.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, _, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+
+	myControl.Start()
+	theirControl.Start()
+
+	t.Log("Inject handshake with MessageCounter=3")
+	myControl.InjectUDPPacket(makeHandshakePacket(theirUdpAddr, myUdpAddr, header.HandshakeIXPSK0, 0, 3))
+
+	t.Log("Inject handshake with MessageCounter=99")
+	myControl.InjectUDPPacket(makeHandshakePacket(theirUdpAddr, myUdpAddr, header.HandshakeIXPSK0, 0, 99))
+
+	t.Log("Verify no tunnels or pending handshakes")
+	assert.Empty(t, myControl.ListHostmapHosts(false))
+	assert.Empty(t, myControl.ListHostmapHosts(true))
+
+	t.Log("Verify no UDP response was sent")
+	time.Sleep(100 * time.Millisecond)
+	assert.Nil(t, myControl.GetFromUDP(false))
+
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+func TestHandshakeUnknownSubtype(t *testing.T) {
+	// A handshake packet with an unknown subtype should be silently dropped.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, _, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
+	theirControl, _, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.Start()
+	theirControl.Start()
+
+	t.Log("Inject handshake with unknown subtype 99")
+	myControl.InjectUDPPacket(makeHandshakePacket(theirUdpAddr, myUdpAddr, header.MessageSubType(99), 0, 1))
+
+	t.Log("Verify no tunnels or pending handshakes")
+	assert.Empty(t, myControl.ListHostmapHosts(false))
+	assert.Empty(t, myControl.ListHostmapHosts(true))
+
+	t.Log("Verify no UDP response was sent")
+	time.Sleep(100 * time.Millisecond)
+	assert.Nil(t, myControl.GetFromUDP(false))
+
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+func TestHandshakeLateResponse(t *testing.T) {
+	// After a handshake times out, a late response should be silently ignored
+	// with no new tunnels created.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", m{
+		"handshakes": m{
+			"try_interval": "200ms",
+			"retries":      2,
+		},
+	})
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+
+	myControl.Start()
+	theirControl.Start()
+
+	t.Log("Trigger handshake from me")
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi"))
+
+	t.Log("Grab msg1 but don't deliver")
+	msg1 := myControl.GetFromUDP(true)
+
+	t.Log("Wait for handshake to time out")
+	for i := 0; i < 5; i++ {
+		time.Sleep(300 * time.Millisecond)
+		myControl.GetFromUDP(false)
+	}
+
+	t.Log("Confirm no pending handshakes remain")
+	assert.Empty(t, myControl.ListHostmapHosts(true))
+
+	t.Log("Deliver old msg1 to them, they create a tunnel")
+	theirControl.InjectUDPPacket(msg1)
+	resp := theirControl.GetFromUDP(true)
+	assert.NotNil(t, resp)
+
+	t.Log("Inject late response into me, should be ignored")
+	myControl.InjectUDPPacket(resp)
+
+	t.Log("No tunnel should exist on my side")
+	assert.Empty(t, myControl.ListHostmapHosts(false))
+	assert.Empty(t, myControl.ListHostmapHosts(true))
+
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+func TestHandshakeSelfConnectionRejected(t *testing.T) {
+	// Verify that a node rejects a handshake containing its own VPN IP in the
+	// peer cert. We do this by sending the initiator's own msg1 back to itself.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
+
+	// Need a lighthouse entry to trigger a handshake
+	myControl.InjectLightHouseAddr(netip.MustParseAddr("10.128.0.2"), netip.MustParseAddrPort("10.0.0.2:4242"))
+
+	myControl.Start()
+
+	t.Log("Trigger handshake from me")
+	myControl.InjectTunUDPPacket(netip.MustParseAddr("10.128.0.2"), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi"))
+	msg1 := myControl.GetFromUDP(true)
+
+	t.Log("Drain any handshake retransmits before injecting")
+	time.Sleep(100 * time.Millisecond)
+	for myControl.GetFromUDP(false) != nil {
+	}
+
+	t.Log("Feed my own msg1 back to me as if it came from someone else")
+	selfMsg := msg1.Copy()
+	selfMsg.From = netip.MustParseAddrPort("10.0.0.99:4242")
+	selfMsg.To = myUdpAddr
+	myControl.InjectUDPPacket(selfMsg)
+
+	t.Log("Verify no response was sent (self-connection rejected)")
+	time.Sleep(100 * time.Millisecond)
+	// Drain any further retransmits from the original handshake, then check
+	// that none of them are a handshake response (MessageCounter=2)
+	h := &header.H{}
+	for {
+		p := myControl.GetFromUDP(false)
+		if p == nil {
+			break
+		}
+		_ = h.Parse(p.Data)
+		assert.NotEqual(t, uint64(2), h.MessageCounter,
+			"should not send a stage 2 response to self-connection")
+	}
+
+	t.Log("Verify no tunnel to myself was created")
+	assert.Nil(t, myControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false))
+
+	myControl.Stop()
+}
+
+func TestHandshakeMessageCounter0Dropped(t *testing.T) {
+	// MessageCounter=0 is not a valid handshake message and should be dropped.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, _, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
+	_, _, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.Start()
+
+	t.Log("Inject handshake with MessageCounter=0")
+	myControl.InjectUDPPacket(makeHandshakePacket(theirUdpAddr, myUdpAddr, header.HandshakeIXPSK0, 0, 0))
+
+	time.Sleep(100 * time.Millisecond)
+	assert.Empty(t, myControl.ListHostmapHosts(false))
+	assert.Empty(t, myControl.ListHostmapHosts(true))
+	assert.Nil(t, myControl.GetFromUDP(false))
+
+	myControl.Stop()
+}
+
+func TestHandshakeRemoteAllowList(t *testing.T) {
+	// Verify that a handshake from a blocked underlay IP is dropped with no
+	// response and no state changes. Then verify the same packet from an
+	// allowed IP succeeds.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", m{
+		"lighthouse": m{
+			"remote_allow_list": m{
+				"10.0.0.0/8": true,
+				"0.0.0.0/0":  false,
+			},
+		},
+	})
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	myControl.Start()
+	theirControl.Start()
+
+	r := router.NewR(t, myControl, theirControl)
+	defer r.RenderFlow()
+
+	t.Log("Trigger handshake from them")
+	theirControl.InjectTunUDPPacket(myVpnIpNet[0].Addr(), 80, theirVpnIpNet[0].Addr(), 80, []byte("Hi"))
+	msg1 := theirControl.GetFromUDP(true)
+
+	t.Log("Rewrite the source to a blocked IP and inject")
+	blockedMsg := msg1.Copy()
+	blockedMsg.From = netip.MustParseAddrPort("192.168.1.1:4242")
+	myControl.InjectUDPPacket(blockedMsg)
+
+	t.Log("Verify no tunnel, no pending, no response from blocked source")
+	time.Sleep(100 * time.Millisecond)
+	assert.Empty(t, myControl.ListHostmapHosts(false))
+	assert.Empty(t, myControl.ListHostmapHosts(true))
+	assert.Nil(t, myControl.GetFromUDP(false), "should not respond to blocked source")
+
+	t.Log("Now inject the real packet from the allowed source")
+	myControl.InjectUDPPacket(msg1)
+
+	t.Log("Verify handshake completes from allowed source")
+	resp := myControl.GetFromUDP(true)
+	assert.NotNil(t, resp)
+	theirControl.InjectUDPPacket(resp)
+	theirControl.WaitForType(1, 0, myControl)
+
+	t.Log("Drain cached packet and verify tunnel works")
+	cachedPacket := myControl.GetFromTun(true)
+	assertUdpPacket(t, []byte("Hi"), cachedPacket, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), 80, 80)
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+func TestHandshakeAlreadySeenPreferredRemote(t *testing.T) {
+	// When a duplicate msg1 arrives via ErrAlreadySeen, verify the tunnel
+	// remains functional and hostmap index count is stable.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", nil)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", nil)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	myControl.Start()
+	theirControl.Start()
+
+	r := router.NewR(t, myControl, theirControl)
+	defer r.RenderFlow()
+
+	t.Log("Complete a normal handshake via the router")
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi"))
+	r.RouteForAllUntilTxTun(theirControl)
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	t.Log("Record hostmap state")
+	theirIndexes := len(theirControl.ListHostmapIndexes(false))
+	hi := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+	assert.NotNil(t, hi)
+	originalRemote := hi.CurrentRemote
+
+	t.Log("Re-trigger traffic to cause a new handshake attempt (ErrAlreadySeen)")
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("roam"))
+	r.RouteForAllUntilTxTun(theirControl)
+
+	t.Log("Verify tunnel still works")
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	t.Log("Verify remote is still valid and index count is stable")
+	hi2 := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+	assert.NotNil(t, hi2)
+	assert.Equal(t, originalRemote, hi2.CurrentRemote)
+	assert.Equal(t, theirIndexes, len(theirControl.ListHostmapIndexes(false)),
+		"no extra indexes should be created from ErrAlreadySeen")
+
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+func TestHandshakeWrongResponderPacketStore(t *testing.T) {
+	// Verify that when the wrong host responds, the cached packets are
+	// transferred to the new handshake, the evil tunnel is closed, evil's
+	// address is blocked, and the correct tunnel is eventually established.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.100/24", nil)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.99/24", nil)
+	evilControl, evilVpnIpNet, evilUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "evil", "10.128.0.2/24", nil)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), evilUdpAddr)
+
+	r := router.NewR(t, myControl, theirControl, evilControl)
+	defer r.RenderFlow()
+
+	myControl.Start()
+	theirControl.Start()
+	evilControl.Start()
+
+	t.Log("Send multiple packets to them (cached during handshake)")
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("packet1"))
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("packet2"))
+
+	t.Log("Route until evil tunnel is closed")
+	h := &header.H{}
+	r.RouteForAllExitFunc(func(p *udp.Packet, c *nebula.Control) router.ExitType {
+		if err := h.Parse(p.Data); err != nil {
+			panic(err)
+		}
+		if h.Type == header.CloseTunnel && p.To == evilUdpAddr {
+			return router.RouteAndExit
+		}
+		return router.KeepRouting
+	})
+
+	t.Log("Verify evil's address is blocked in the new pending handshake")
+	pendingHI := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), true)
+	if pendingHI != nil {
+		assert.NotContains(t, pendingHI.RemoteAddrs, evilUdpAddr,
+			"evil's address should be blocked")
+	}
+
+	t.Log("Inject correct lighthouse addr for them")
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+
+	t.Log("Route until cached packets arrive at the real them")
+	p := r.RouteForAllUntilTxTun(theirControl)
+	assert.NotNil(t, p, "a cached packet should be delivered to the correct host")
+
+	t.Log("Verify the correct host has a tunnel")
+	assertHostInfoPair(t, myUdpAddr, theirUdpAddr, myVpnIpNet, theirVpnIpNet, myControl, theirControl)
+
+	t.Log("Verify no hostinfo artifacts from evil remain")
+	assert.Nil(t, myControl.GetHostInfoByVpnAddr(evilVpnIpNet[0].Addr(), true),
+		"no pending hostinfo for evil")
+	assert.Nil(t, myControl.GetHostInfoByVpnAddr(evilVpnIpNet[0].Addr(), false),
+		"no main hostinfo for evil")
+
+	myControl.Stop()
+	theirControl.Stop()
+	evilControl.Stop()
+}
+
+func TestHandshakeRelayComplete(t *testing.T) {
+	// Verify that a relay handshake completes correctly and relay state is
+	// properly maintained on all three nodes.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
+	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "relay", "10.128.0.128/24", m{"relay": m{"am_relay": true}})
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", m{"relay": m{"use_relays": true}})
+
+	myControl.InjectLightHouseAddr(relayVpnIpNet[0].Addr(), relayUdpAddr)
+	myControl.InjectRelays(theirVpnIpNet[0].Addr(), []netip.Addr{relayVpnIpNet[0].Addr()})
+	relayControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+
+	r := router.NewR(t, myControl, relayControl, theirControl)
+	defer r.RenderFlow()
+
+	myControl.Start()
+	relayControl.Start()
+	theirControl.Start()
+
+	t.Log("Trigger handshake via relay")
+	myControl.InjectTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi via relay"))
+
+	p := r.RouteForAllUntilTxTun(theirControl)
+	assertUdpPacket(t, []byte("Hi via relay"), p, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), 80, 80)
+
+	t.Log("Verify bidirectional tunnel via relay")
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	t.Log("Verify relay state on my side shows relay-to-me")
+	myHI := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false)
+	assert.NotNil(t, myHI)
+	assert.NotEmpty(t, myHI.CurrentRelaysToMe, "should have relay-to-me for them")
+
+	t.Log("Verify relay state on their side shows relay-to-me")
+	theirHI := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+	assert.NotNil(t, theirHI)
+	assert.NotEmpty(t, theirHI.CurrentRelaysToMe, "should have relay-to-me for me")
+
+	t.Log("Verify relay node shows through-me relays")
+	relayHI := relayControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+	assert.NotNil(t, relayHI)
+
+	myControl.Stop()
+	relayControl.Stop()
+	theirControl.Stop()
+}
+
+// NOTE: Relay V1 cert + IPv6 rejection is not tested here because
+// InjectTunUDPPacket from a V4 node to a V6 address panics in the test
+// framework. The check is in handshake_manager.go handleOutbound relay
+// logic (lines ~304-313): if the relay host has a V1 cert and either
+// address is IPv6, the relay is skipped.
+
+// NOTE: Relay reestablishment (Disestablished state transition) is covered
+// by the existing TestReestablishRelays in handshakes_test.go.

e2e/tunnels_test.go

@@ -12,6 +12,8 @@ import (
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/cert_test"
 	"github.com/slackhq/nebula/e2e/router"
+	"github.com/slackhq/nebula/header"
+	"github.com/slackhq/nebula/udp"
 	"github.com/stretchr/testify/assert"
 	"gopkg.in/yaml.v3"
 )
@@ -365,3 +367,106 @@ func TestCrossStackRelaysWork(t *testing.T) {
 	//theirControl.Stop()
 	//relayControl.Stop()
 }
+
+func TestCloseTunnelAuthenticated(t *testing.T) {
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "me", "10.128.0.1/24", m{"tunnels": m{"drop_inactive": true, "inactivity_timeout": "5s"}})
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them", "10.128.0.2/24", m{"tunnels": m{"drop_inactive": true, "inactivity_timeout": "10m"}})
+
+	// Share our underlay information
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	// Start the servers
+	myControl.Start()
+	theirControl.Start()
+
+	r := router.NewR(t, myControl, theirControl)
+
+	r.Log("Assert the tunnel between me and them works")
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+
+	r.Log("Close the tunnel")
+	myControl.CloseTunnel(theirVpnIpNet[0].Addr(), false)
+	r.FlushAll()
+
+	waitStart := time.Now()
+	for {
+		myIndexes := len(myControl.GetHostmap().Indexes)
+		theirIndexes := len(theirControl.GetHostmap().Indexes)
+		if myIndexes == 0 && theirIndexes == 0 {
+			break
+		}
+
+		since := time.Since(waitStart)
+		r.Logf("my tunnels: %v; their tunnels: %v; duration: %v", myIndexes, theirIndexes, since)
+		if since > time.Second*6 {
+			t.Fatal("Tunnel should have been declared inactive after 2 seconds and before 6 seconds")
+		}
+
+		time.Sleep(1 * time.Second)
+		//r.FlushAll()
+	}
+
+	r.Logf("Happy path success, tunnels were dropped within %v", time.Since(waitStart))
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+	r.Log("Assert another tunnel between me and them works")
+	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
+	hi := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false)
+	if hi == nil {
+		t.Fatal("There is no hostinfo for this tunnel")
+	}
+	myHi := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+	if myHi == nil {
+		t.Fatal("There is no hostinfo for my tunnel")
+	}
+	r.Log("It does")
+
+	buf := make([]byte, 1024)
+	hdr := header.H{
+		Version:        1,
+		Type:           header.CloseTunnel,
+		Subtype:        0,
+		Reserved:       0,
+		RemoteIndex:    hi.RemoteIndex,
+		MessageCounter: 5,
+	}
+	out, err := hdr.Encode(buf)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	pkt := &udp.Packet{
+		To:   hi.CurrentRemote,
+		From: myHi.CurrentRemote,
+		Data: out,
+	}
+	r.InjectUDPPacket(myControl, theirControl, pkt)
+	r.Log("Injected bogus close tunnel. Let's see!")
+	waitStart = time.Now()
+	for {
+		myIndexes := len(myControl.GetHostmap().Indexes)
+		theirIndexes := len(theirControl.GetHostmap().Indexes)
+		if myIndexes == 0 {
+			t.Fatal("myIndexes should not be 0")
+		}
+		if theirIndexes == 0 {
+			t.Fatal("theirIndexes should not be 0, they should have rejected this bogus packet")
+		}
+
+		since := time.Since(waitStart)
+		r.Logf("my tunnels: %v; their tunnels: %v; duration: %v", myIndexes, theirIndexes, since)
+		if since > time.Second*4 {
+			t.Log("The tunnel would have been gone by now")
+			break
+		}
+
+		time.Sleep(1 * time.Second)
+		r.FlushAll()
+	}
+
+	myControl.Stop()
+	theirControl.Stop()
+}

examples/config.yml

@@ -204,6 +204,12 @@ punchy:
   # Trusted SSH CA public keys. These are the public keys of the CAs that are allowed to sign SSH keys for access.
   #trusted_cas:
     #- "ssh public key string"
+  # sandbox_dir restricts file paths for profiling commands (start-cpu-profile, save-heap-profile,
+  # save-mutex-profile) to the specified directory. Relative paths will be resolved within this directory,
+  # and absolute paths outside of it will be rejected. Default is $TMP/nebula-debug.
+  # The directory is NOT automatically created.
+  # Overriding this to "" is the same as "/" and will allow overwriting any path on the host.
+  #sandbox_dir: /var/tmp/nebula-debug
 
 # EXPERIMENTAL: relay support for networks that can't establish direct connections.
 relay:
@@ -382,8 +388,8 @@ firewall:
   # Rules are comprised of a protocol, port, and one or more of host, group, or CIDR
   # Logical evaluation is roughly: port AND proto AND (ca_sha OR ca_name) AND (host OR group OR groups OR cidr) AND (local cidr)
   # - port: Takes `0` or `any` as any, a single number `80`, a range `200-901`, or `fragment` to match second and further fragments of fragmented packets (since there is no port available).
-  #   code: same as port but makes more sense when talking about ICMP, TODO: this is not currently implemented in a way that works, use `any`
   #   proto: `any`, `tcp`, `udp`, or `icmp`
+  #          a port specification is ignored if proto is `icmp`
   #   host: `any` or a literal hostname, ie `test-host`
   #   group: `any` or a literal group name, ie `default-group`
   #   groups: Same as group but accepts a list of values. Multiple values are AND'd together and a certificate would have to contain all groups to pass

firewall.go

@@ -230,7 +230,7 @@ func NewFirewallFromConfig(l *logrus.Logger, cs *CertState, c *config.C) (*Firew
 	case "drop":
 		fw.OutSendReject = false
 	default:
-		l.WithField("action", inboundAction).Warn("invalid firewall.outbound_action, defaulting to `drop`")
+		l.WithField("action", outboundAction).Warn("invalid firewall.outbound_action, defaulting to `drop`")
 		fw.OutSendReject = false
 	}
 
@@ -249,20 +249,6 @@ func NewFirewallFromConfig(l *logrus.Logger, cs *CertState, c *config.C) (*Firew
 
 // AddRule properly creates the in memory rule structure for a firewall table.
 func (f *Firewall) AddRule(incoming bool, proto uint8, startPort int32, endPort int32, groups []string, host string, cidr, localCidr, caName string, caSha string) error {
-	// We need this rule string because we generate a hash. Removing this will break firewall reload.
-	ruleString := fmt.Sprintf(
-		"incoming: %v, proto: %v, startPort: %v, endPort: %v, groups: %v, host: %v, ip: %v, localIp: %v, caName: %v, caSha: %s",
-		incoming, proto, startPort, endPort, groups, host, cidr, localCidr, caName, caSha,
-	)
-	f.rules += ruleString + "\n"
-
-	direction := "incoming"
-	if !incoming {
-		direction = "outgoing"
-	}
-	f.l.WithField("firewallRule", m{"direction": direction, "proto": proto, "startPort": startPort, "endPort": endPort, "groups": groups, "host": host, "cidr": cidr, "localCidr": localCidr, "caName": caName, "caSha": caSha}).
-		Info("Firewall rule added")
-
 	var (
 		ft *FirewallTable
 		fp firewallPort
@@ -280,6 +266,12 @@ func (f *Firewall) AddRule(incoming bool, proto uint8, startPort int32, endPort
 	case firewall.ProtoUDP:
 		fp = ft.UDP
 	case firewall.ProtoICMP, firewall.ProtoICMPv6:
+		//ICMP traffic doesn't have ports, so we always coerce to "any", even if a value is provided
+		if startPort != firewall.PortAny {
+			f.l.WithField("startPort", startPort).Warn("ignoring port specification for ICMP firewall rule")
+		}
+		startPort = firewall.PortAny
+		endPort = firewall.PortAny
 		fp = ft.ICMP
 	case firewall.ProtoAny:
 		fp = ft.AnyProto
@@ -287,6 +279,20 @@ func (f *Firewall) AddRule(incoming bool, proto uint8, startPort int32, endPort
 		return fmt.Errorf("unknown protocol %v", proto)
 	}
 
+	// We need this rule string because we generate a hash. Removing this will break firewall reload.
+	ruleString := fmt.Sprintf(
+		"incoming: %v, proto: %v, startPort: %v, endPort: %v, groups: %v, host: %v, ip: %v, localIp: %v, caName: %v, caSha: %s",
+		incoming, proto, startPort, endPort, groups, host, cidr, localCidr, caName, caSha,
+	)
+	f.rules += ruleString + "\n"
+
+	direction := "incoming"
+	if !incoming {
+		direction = "outgoing"
+	}
+	f.l.WithField("firewallRule", m{"direction": direction, "proto": proto, "startPort": startPort, "endPort": endPort, "groups": groups, "host": host, "cidr": cidr, "localCidr": localCidr, "caName": caName, "caSha": caSha}).
+		Info("Firewall rule added")
+
 	return fp.addRule(f, startPort, endPort, groups, host, cidr, localCidr, caName, caSha)
 }
 
@@ -349,24 +355,31 @@ func AddFirewallRulesFromConfig(l *logrus.Logger, inbound bool, c *config.C, fw
 			sPort = r.Port
 		}
 
-		startPort, endPort, err := parsePort(sPort)
-		if err != nil {
-			return fmt.Errorf("%s rule #%v; %s %s", table, i, errPort, err)
-		}
-
 		var proto uint8
+		var startPort, endPort int32
 		switch r.Proto {
 		case "any":
 			proto = firewall.ProtoAny
+			startPort, endPort, err = parsePort(sPort)
 		case "tcp":
 			proto = firewall.ProtoTCP
+			startPort, endPort, err = parsePort(sPort)
 		case "udp":
 			proto = firewall.ProtoUDP
+			startPort, endPort, err = parsePort(sPort)
 		case "icmp":
 			proto = firewall.ProtoICMP
+			startPort = firewall.PortAny
+			endPort = firewall.PortAny
+			if sPort != "" {
+				l.WithField("port", sPort).Warn("ignoring port specification for ICMP firewall rule")
+			}
 		default:
 			return fmt.Errorf("%s rule #%v; proto was not understood; `%s`", table, i, r.Proto)
 		}
+		if err != nil {
+			return fmt.Errorf("%s rule #%v; %s %s", table, i, errPort, err)
+		}
 
 		if r.Cidr != "" && r.Cidr != "any" {
 			_, err = netip.ParsePrefix(r.Cidr)
@@ -467,7 +480,7 @@ func (f *Firewall) metrics(incoming bool) firewallMetrics {
 	}
 }
 
-// Destroy cleans up any known cyclical references so the object can be free'd my GC. This should be called if a new
+// Destroy cleans up any known cyclical references so the object can be freed by GC. This should be called if a new
 // firewall object is created
 func (f *Firewall) Destroy() {
 	//TODO: clean references if/when needed
@@ -660,6 +673,13 @@ func (fp firewallPort) match(p firewall.Packet, incoming bool, c *cert.CachedCer
 		return false
 	}
 
+	// this branch is here to catch traffic from FirewallTable.Any.match and FirewallTable.ICMP.match
+	if p.Protocol == firewall.ProtoICMP || p.Protocol == firewall.ProtoICMPv6 {
+		// port numbers are re-used for connection tracking of ICMP,
+		// but we don't want to actually filter on them.
+		return fp[firewall.PortAny].match(p, c, caPool)
+	}
+
 	var port int32
 
 	if p.Fragment {
@@ -804,10 +824,8 @@ func (fr *FirewallRule) isAny(groups []string, host string, cidr string) bool {
 		return true
 	}
 
-	for _, group := range groups {
-		if group == "any" {
-			return true
-		}
+	if slices.Contains(groups, "any") {
+		return true
 	}
 
 	if host == "any" {
@@ -1018,54 +1036,56 @@ func (r *rule) sanity() error {
 		}
 	}
 
+	if r.Code != "" {
+		return fmt.Errorf("code specified as [%s]. Support for 'code' will be dropped in a future release, as it has never been functional", r.Code)
+	}
+
 	//todo alert on cidr-any
 
 	return nil
 }
 
-func parsePort(s string) (startPort, endPort int32, err error) {
+func parsePort(s string) (int32, int32, error) {
+	var err error
+	const notAPort int32 = -2
 	if s == "any" {
-		startPort = firewall.PortAny
-		endPort = firewall.PortAny
-
-	} else if s == "fragment" {
-		startPort = firewall.PortFragment
-		endPort = firewall.PortFragment
-
-	} else if strings.Contains(s, `-`) {
-		sPorts := strings.SplitN(s, `-`, 2)
-		sPorts[0] = strings.Trim(sPorts[0], " ")
-		sPorts[1] = strings.Trim(sPorts[1], " ")
-
-		if len(sPorts) != 2 || sPorts[0] == "" || sPorts[1] == "" {
-			return 0, 0, fmt.Errorf("appears to be a range but could not be parsed; `%s`", s)
-		}
-
-		rStartPort, err := strconv.Atoi(sPorts[0])
-		if err != nil {
-			return 0, 0, fmt.Errorf("beginning range was not a number; `%s`", sPorts[0])
-		}
-
-		rEndPort, err := strconv.Atoi(sPorts[1])
-		if err != nil {
-			return 0, 0, fmt.Errorf("ending range was not a number; `%s`", sPorts[1])
-		}
-
-		startPort = int32(rStartPort)
-		endPort = int32(rEndPort)
-
-		if startPort == firewall.PortAny {
-			endPort = firewall.PortAny
-		}
-
-	} else {
+		return firewall.PortAny, firewall.PortAny, nil
+	}
+	if s == "fragment" {
+		return firewall.PortFragment, firewall.PortFragment, nil
+	}
+	if !strings.Contains(s, `-`) {
 		rPort, err := strconv.Atoi(s)
 		if err != nil {
-			return 0, 0, fmt.Errorf("was not a number; `%s`", s)
+			return notAPort, notAPort, fmt.Errorf("was not a number; `%s`", s)
 		}
-		startPort = int32(rPort)
-		endPort = startPort
+		return int32(rPort), int32(rPort), nil
 	}
 
-	return
+	sPorts := strings.SplitN(s, `-`, 2)
+	for i := range sPorts {
+		sPorts[i] = strings.Trim(sPorts[i], " ")
+	}
+	if len(sPorts) != 2 || sPorts[0] == "" || sPorts[1] == "" {
+		return notAPort, notAPort, fmt.Errorf("appears to be a range but could not be parsed; `%s`", s)
+	}
+
+	rStartPort, err := strconv.Atoi(sPorts[0])
+	if err != nil {
+		return notAPort, notAPort, fmt.Errorf("beginning range was not a number; `%s`", sPorts[0])
+	}
+
+	rEndPort, err := strconv.Atoi(sPorts[1])
+	if err != nil {
+		return notAPort, notAPort, fmt.Errorf("ending range was not a number; `%s`", sPorts[1])
+	}
+
+	startPort := int32(rStartPort)
+	endPort := int32(rEndPort)
+
+	if startPort == firewall.PortAny {
+		endPort = firewall.PortAny
+	}
+
+	return startPort, endPort, nil
 }

firewall/packet.go

@@ -22,7 +22,10 @@ const (
 type Packet struct {
 	LocalAddr  netip.Addr
 	RemoteAddr netip.Addr
-	LocalPort  uint16
+	// LocalPort is the destination port for incoming traffic, or the source port for outgoing. Zero for ICMP.
+	LocalPort uint16
+	// RemotePort is the source port for incoming traffic, or the destination port for outgoing.
+	// For ICMP, it's the "identifier". This is only used for connection tracking, actual firewall rules will not filter on ICMP identifier
 	RemotePort uint16
 	Protocol   uint8
 	Fragment   bool
@@ -46,6 +49,8 @@ func (fp Packet) MarshalJSON() ([]byte, error) {
 		proto = "tcp"
 	case ProtoICMP:
 		proto = "icmp"
+	case ProtoICMPv6:
+		proto = "icmpv6"
 	case ProtoUDP:
 		proto = "udp"
 	default:

firewall_test.go

@@ -87,9 +87,10 @@ func TestFirewall_AddRule(t *testing.T) {
 
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, c)
 	require.NoError(t, fw.AddRule(true, firewall.ProtoICMP, 1, 1, []string{}, "h1", "", "", "", ""))
-	assert.Nil(t, fw.InRules.ICMP[1].Any.Any)
-	assert.Empty(t, fw.InRules.ICMP[1].Any.Groups)
-	assert.Contains(t, fw.InRules.ICMP[1].Any.Hosts, "h1")
+	//no matter what port is given for icmp, it should end up as "any"
+	assert.Nil(t, fw.InRules.ICMP[firewall.PortAny].Any.Any)
+	assert.Empty(t, fw.InRules.ICMP[firewall.PortAny].Any.Groups)
+	assert.Contains(t, fw.InRules.ICMP[firewall.PortAny].Any.Hosts, "h1")
 
 	fw = NewFirewall(l, time.Second, time.Minute, time.Hour, c)
 	require.NoError(t, fw.AddRule(false, firewall.ProtoAny, 1, 1, []string{}, "", ti.String(), "", "", ""))
@@ -734,6 +735,150 @@ func TestFirewall_DropConntrackReload(t *testing.T) {
 	assert.Equal(t, fw.Drop(p, false, &h, cp, nil), ErrNoMatchingRule)
 }
 
+func TestFirewall_ICMPPortBehavior(t *testing.T) {
+	l := test.NewLogger()
+	ob := &bytes.Buffer{}
+	l.SetOutput(ob)
+	myVpnNetworksTable := new(bart.Lite)
+	myVpnNetworksTable.Insert(netip.MustParsePrefix("1.1.1.1/8"))
+
+	network := netip.MustParsePrefix("1.2.3.4/24")
+
+	c := cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:     "host1",
+			networks: []netip.Prefix{network},
+			groups:   []string{"default-group"},
+			issuer:   "signer-shasum",
+		},
+		InvertedGroups: map[string]struct{}{"default-group": {}},
+	}
+	h := HostInfo{
+		ConnectionState: &ConnectionState{
+			peerCert: &c,
+		},
+		vpnAddrs: []netip.Addr{network.Addr()},
+	}
+	h.buildNetworks(myVpnNetworksTable, c.Certificate)
+
+	cp := cert.NewCAPool()
+
+	templ := firewall.Packet{
+		LocalAddr:  netip.MustParseAddr("1.2.3.4"),
+		RemoteAddr: netip.MustParseAddr("1.2.3.4"),
+		Protocol:   firewall.ProtoICMP,
+		Fragment:   false,
+	}
+
+	t.Run("ICMP allowed", func(t *testing.T) {
+		fw := NewFirewall(l, time.Second, time.Minute, time.Hour, c.Certificate)
+		require.NoError(t, fw.AddRule(true, firewall.ProtoICMP, 0, 0, []string{"any"}, "", "", "", "", ""))
+		t.Run("zero ports", func(t *testing.T) {
+			p := templ.Copy()
+			p.LocalPort = 0
+			p.RemotePort = 0
+			// Drop outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+			// Allow inbound
+			resetConntrack(fw)
+			require.NoError(t, fw.Drop(*p, true, &h, cp, nil))
+			//now also allow outbound
+			require.NoError(t, fw.Drop(*p, false, &h, cp, nil))
+		})
+
+		t.Run("nonzero ports", func(t *testing.T) {
+			p := templ.Copy()
+			p.LocalPort = 0xabcd
+			p.RemotePort = 0x1234
+			// Drop outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+			// Allow inbound
+			resetConntrack(fw)
+			require.NoError(t, fw.Drop(*p, true, &h, cp, nil))
+			//now also allow outbound
+			require.NoError(t, fw.Drop(*p, false, &h, cp, nil))
+		})
+	})
+
+	t.Run("Any proto, some ports allowed", func(t *testing.T) {
+		fw := NewFirewall(l, time.Second, time.Minute, time.Hour, c.Certificate)
+		require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 80, 444, []string{"any"}, "", "", "", "", ""))
+		t.Run("zero ports, still blocked", func(t *testing.T) {
+			p := templ.Copy()
+			p.LocalPort = 0
+			p.RemotePort = 0
+			// Drop outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+			// Allow inbound
+			resetConntrack(fw)
+			assert.Equal(t, fw.Drop(*p, true, &h, cp, nil), ErrNoMatchingRule)
+			//now also allow outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+		})
+
+		t.Run("nonzero ports, still blocked", func(t *testing.T) {
+			p := templ.Copy()
+			p.LocalPort = 0xabcd
+			p.RemotePort = 0x1234
+			// Drop outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+			// Allow inbound
+			resetConntrack(fw)
+			assert.Equal(t, fw.Drop(*p, true, &h, cp, nil), ErrNoMatchingRule)
+			//now also allow outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+		})
+
+		t.Run("nonzero, matching ports, still blocked", func(t *testing.T) {
+			p := templ.Copy()
+			p.LocalPort = 80
+			p.RemotePort = 80
+			// Drop outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+			// Allow inbound
+			resetConntrack(fw)
+			assert.Equal(t, fw.Drop(*p, true, &h, cp, nil), ErrNoMatchingRule)
+			//now also allow outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+		})
+	})
+	t.Run("Any proto, any port", func(t *testing.T) {
+		fw := NewFirewall(l, time.Second, time.Minute, time.Hour, c.Certificate)
+		require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
+		t.Run("zero ports, allowed", func(t *testing.T) {
+			resetConntrack(fw)
+			p := templ.Copy()
+			p.LocalPort = 0
+			p.RemotePort = 0
+			// Drop outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+			// Allow inbound
+			resetConntrack(fw)
+			require.NoError(t, fw.Drop(*p, true, &h, cp, nil))
+			//now also allow outbound
+			require.NoError(t, fw.Drop(*p, false, &h, cp, nil))
+		})
+
+		t.Run("nonzero ports, allowed", func(t *testing.T) {
+			resetConntrack(fw)
+			p := templ.Copy()
+			p.LocalPort = 0xabcd
+			p.RemotePort = 0x1234
+			// Drop outbound
+			assert.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+			// Allow inbound
+			resetConntrack(fw)
+			require.NoError(t, fw.Drop(*p, true, &h, cp, nil))
+			//now also allow outbound
+			require.NoError(t, fw.Drop(*p, false, &h, cp, nil))
+			//different ID is blocked
+			p.RemotePort++
+			require.Equal(t, fw.Drop(*p, false, &h, cp, nil), ErrNoMatchingRule)
+		})
+	})
+
+}
+
 func TestFirewall_DropIPSpoofing(t *testing.T) {
 	l := test.NewLogger()
 	ob := &bytes.Buffer{}
@@ -919,11 +1064,11 @@ func TestNewFirewallFromConfig(t *testing.T) {
 
 	// Test code/port error
 	conf = config.NewC(l)
-	conf.Settings["firewall"] = map[string]any{"outbound": []any{map[string]any{"code": "a", "host": "testh"}}}
+	conf.Settings["firewall"] = map[string]any{"outbound": []any{map[string]any{"code": "a", "host": "testh", "proto": "any"}}}
 	_, err = NewFirewallFromConfig(l, cs, conf)
 	require.EqualError(t, err, "firewall.outbound rule #0; code was not a number; `a`")
 
-	conf.Settings["firewall"] = map[string]any{"outbound": []any{map[string]any{"port": "a", "host": "testh"}}}
+	conf.Settings["firewall"] = map[string]any{"outbound": []any{map[string]any{"port": "a", "host": "testh", "proto": "any"}}}
 	_, err = NewFirewallFromConfig(l, cs, conf)
 	require.EqualError(t, err, "firewall.outbound rule #0; port was not a number; `a`")
 
@@ -973,7 +1118,14 @@ func TestAddFirewallRulesFromConfig(t *testing.T) {
 	mf = &mockFirewall{}
 	conf.Settings["firewall"] = map[string]any{"outbound": []any{map[string]any{"port": "1", "proto": "icmp", "host": "a"}}}
 	require.NoError(t, AddFirewallRulesFromConfig(l, false, conf, mf))
-	assert.Equal(t, addRuleCall{incoming: false, proto: firewall.ProtoICMP, startPort: 1, endPort: 1, groups: nil, host: "a", ip: "", localIp: ""}, mf.lastCall)
+	assert.Equal(t, addRuleCall{incoming: false, proto: firewall.ProtoICMP, startPort: firewall.PortAny, endPort: firewall.PortAny, groups: nil, host: "a", ip: "", localIp: ""}, mf.lastCall)
+
+	// Test adding icmp rule no port
+	conf = config.NewC(l)
+	mf = &mockFirewall{}
+	conf.Settings["firewall"] = map[string]any{"outbound": []any{map[string]any{"proto": "icmp", "host": "a"}}}
+	require.NoError(t, AddFirewallRulesFromConfig(l, false, conf, mf))
+	assert.Equal(t, addRuleCall{incoming: false, proto: firewall.ProtoICMP, startPort: firewall.PortAny, endPort: firewall.PortAny, groups: nil, host: "a", ip: "", localIp: ""}, mf.lastCall)
 
 	// Test adding any rule
 	conf = config.NewC(l)

go.mod

@@ -1,6 +1,6 @@
 module github.com/slackhq/nebula
 
-go 1.25
+go 1.25.0
 
 require (
 	dario.cat/mergo v1.0.2
@@ -13,8 +13,8 @@ require (
 	github.com/gogo/protobuf v1.3.2
 	github.com/google/gopacket v1.1.19
 	github.com/kardianos/service v1.2.4
-	github.com/miekg/dns v1.1.70
-	github.com/miekg/pkcs11 v1.1.2-0.20231115102856-9078ad6b9d4b
+	github.com/miekg/dns v1.1.72
+	github.com/miekg/pkcs11 v1.1.2
 	github.com/nbrownus/go-metrics-prometheus v0.0.0-20210712211119-974a6260965f
 	github.com/prometheus/client_golang v1.23.2
 	github.com/rcrowley/go-metrics v0.0.0-20201227073835-cf1acfcdf475
@@ -24,15 +24,15 @@ require (
 	github.com/stretchr/testify v1.11.1
 	github.com/vishvananda/netlink v1.3.1
 	go.yaml.in/yaml/v3 v3.0.4
-	golang.org/x/crypto v0.47.0
+	golang.org/x/crypto v0.50.0
 	golang.org/x/exp v0.0.0-20230725093048-515e97ebf090
-	golang.org/x/net v0.49.0
-	golang.org/x/sync v0.19.0
-	golang.org/x/sys v0.40.0
-	golang.org/x/term v0.39.0
+	golang.org/x/net v0.52.0
+	golang.org/x/sync v0.20.0
+	golang.org/x/sys v0.43.0
+	golang.org/x/term v0.42.0
 	golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2
 	golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b
-	golang.zx2c4.com/wireguard/windows v0.5.3
+	golang.zx2c4.com/wireguard/windows v0.6.1
 	google.golang.org/protobuf v1.36.11
 	gopkg.in/yaml.v3 v3.0.1
 	gvisor.dev/gvisor v0.0.0-20240423190808-9d7a357edefe
@@ -50,7 +50,7 @@ require (
 	github.com/prometheus/procfs v0.16.1 // indirect
 	github.com/vishvananda/netns v0.0.5 // indirect
 	go.yaml.in/yaml/v2 v2.4.2 // indirect
-	golang.org/x/mod v0.31.0 // indirect
+	golang.org/x/mod v0.34.0 // indirect
 	golang.org/x/time v0.5.0 // indirect
-	golang.org/x/tools v0.40.0 // indirect
+	golang.org/x/tools v0.43.0 // indirect
 )

go.sum

@@ -85,10 +85,10 @@ github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
 github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc=
 github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=
 github.com/matttproud/golang_protobuf_extensions v1.0.1/go.mod h1:D8He9yQNgCq6Z5Ld7szi9bcBfOoFv/3dc6xSMkL2PC0=
-github.com/miekg/dns v1.1.70 h1:DZ4u2AV35VJxdD9Fo9fIWm119BsQL5cZU1cQ9s0LkqA=
-github.com/miekg/dns v1.1.70/go.mod h1:+EuEPhdHOsfk6Wk5TT2CzssZdqkmFhf8r+aVyDEToIs=
-github.com/miekg/pkcs11 v1.1.2-0.20231115102856-9078ad6b9d4b h1:J/AzCvg5z0Hn1rqZUJjpbzALUmkKX0Zwbc/i4fw7Sfk=
-github.com/miekg/pkcs11 v1.1.2-0.20231115102856-9078ad6b9d4b/go.mod h1:XsNlhZGX73bx86s2hdc/FuaLm2CPZJemRLMA+WTFxgs=
+github.com/miekg/dns v1.1.72 h1:vhmr+TF2A3tuoGNkLDFK9zi36F2LS+hKTRW0Uf8kbzI=
+github.com/miekg/dns v1.1.72/go.mod h1:+EuEPhdHOsfk6Wk5TT2CzssZdqkmFhf8r+aVyDEToIs=
+github.com/miekg/pkcs11 v1.1.2 h1:/VxmeAX5qU6Q3EwafypogwWbYryHFmF2RpkJmw3m4MQ=
+github.com/miekg/pkcs11 v1.1.2/go.mod h1:XsNlhZGX73bx86s2hdc/FuaLm2CPZJemRLMA+WTFxgs=
 github.com/modern-go/concurrent v0.0.0-20180228061459-e0a39a4cb421/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/reflect2 v0.0.0-20180701023420-4b7aa43c6742/go.mod h1:bx2lNnkwVCuqBIxFjflWJWanXIb3RllmbCylyMrvgv0=
@@ -164,16 +164,16 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
 golang.org/x/crypto v0.0.0-20210322153248-0c34fe9e7dc2/go.mod h1:T9bdIzuCu7OtxOm1hfPfRQxPLYneinmdGuTeoZ9dtd4=
-golang.org/x/crypto v0.47.0 h1:V6e3FRj+n4dbpw86FJ8Fv7XVOql7TEwpHapKoMJ/GO8=
-golang.org/x/crypto v0.47.0/go.mod h1:ff3Y9VzzKbwSSEzWqJsJVBnWmRwRSHt/6Op5n9bQc4A=
+golang.org/x/crypto v0.50.0 h1:zO47/JPrL6vsNkINmLoo/PH1gcxpls50DNogFvB5ZGI=
+golang.org/x/crypto v0.50.0/go.mod h1:3muZ7vA7PBCE6xgPX7nkzzjiUq87kRItoJQM1Yo8S+Q=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090 h1:Di6/M8l0O2lCLc6VVRWhgCiApHV8MnQurBnFSHsQtNY=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090/go.mod h1:FXUEEKJgO7OQYeo8N01OfiKP8RXMtf6e8aTskBGqWdc=
 golang.org/x/lint v0.0.0-20200302205851-738671d3881b/go.mod h1:3xt1FjdF8hUf6vQPIChWIBhFzV8gjjsPE/fR3IyQdNY=
 golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee/go.mod h1:QqPTAvyqsEbceGzBzNggFXnrqF1CaUcvgkdR5Ot7KZg=
 golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
-golang.org/x/mod v0.31.0 h1:HaW9xtz0+kOcWKwli0ZXy79Ix+UW/vOfmWI5QVd2tgI=
-golang.org/x/mod v0.31.0/go.mod h1:43JraMp9cGx1Rx3AqioxrbrhNsLl2l/iNAvuBkrezpg=
+golang.org/x/mod v0.34.0 h1:xIHgNUUnW6sYkcM5Jleh05DvLOtwc6RitGHbDk4akRI=
+golang.org/x/mod v0.34.0/go.mod h1:ykgH52iCZe79kzLLMhyCUzhMci+nQj+0XkbXpNYtVjY=
 golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20181114220301-adae6a3d119a/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190108225652-1e06a53dbb7e/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
@@ -184,8 +184,8 @@ golang.org/x/net v0.0.0-20200226121028-0de0cce0169b/go.mod h1:z5CRVTTTmAJ677TzLL
 golang.org/x/net v0.0.0-20200625001655-4c5254603344/go.mod h1:/O7V0waA8r7cgGh81Ro3o1hOxt32SMVPicZroKQ2sZA=
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
-golang.org/x/net v0.49.0 h1:eeHFmOGUTtaaPSGNmjBKpbng9MulQsJURQUAfUwY++o=
-golang.org/x/net v0.49.0/go.mod h1:/ysNB2EvaqvesRkuLAyjI1ycPZlQHM3q01F02UY/MV8=
+golang.org/x/net v0.52.0 h1:He/TN1l0e4mmR3QqHMT2Xab3Aj3L9qjbhRm78/6jrW0=
+golang.org/x/net v0.52.0/go.mod h1:R1MAz7uMZxVMualyPXb+VaqGSa3LIaUqk0eEt3w36Sw=
 golang.org/x/oauth2 v0.0.0-20190226205417-e64efc72b421/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20181221193216-37e7f081c4d4/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -193,8 +193,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201207232520-09787c993a3a/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.19.0 h1:vV+1eWNmZ5geRlYjzm2adRgW2/mcpevXNg50YZtPCE4=
-golang.org/x/sync v0.19.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
+golang.org/x/sync v0.20.0 h1:e0PTpb7pjO8GAtTs2dQ6jYa5BWYlMuX047Dco/pItO4=
+golang.org/x/sync v0.20.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.0.0-20180905080454-ebe1bf3edb33/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20181116152217-5ac8a444bdc5/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
@@ -210,11 +210,11 @@ golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c/go.mod h1:h1NjWce9XRLGQEsW7w
 golang.org/x/sys v0.0.0-20210603081109-ebe580a85c40/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.2.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.40.0 h1:DBZZqJ2Rkml6QMQsZywtnjnnGvHza6BTfYFWY9kjEWQ=
-golang.org/x/sys v0.40.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
+golang.org/x/sys v0.43.0 h1:Rlag2XtaFTxp19wS8MXlJwTvoh8ArU6ezoyFsMyCTNI=
+golang.org/x/sys v0.43.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.39.0 h1:RclSuaJf32jOqZz74CkPA9qFuVTX7vhLlpfj/IGWlqY=
-golang.org/x/term v0.39.0/go.mod h1:yxzUCTP/U+FzoxfdKmLaA0RV1WgE0VY7hXBwKtY/4ww=
+golang.org/x/term v0.42.0 h1:UiKe+zDFmJobeJ5ggPwOshJIVt6/Ft0rcfrXZDLWAWY=
+golang.org/x/term v0.42.0/go.mod h1:Dq/D+snpsbazcBG5+F9Q1n2rXV8Ma+71xEjTRufARgY=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.2/go.mod h1:bEr9sfX3Q8Zfm5fL9x+3itogRgK3+ptLWKqgva+5dAk=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
@@ -225,8 +225,8 @@ golang.org/x/tools v0.0.0-20191119224855-298f0cb1881e/go.mod h1:b+2E5dAYhXwXZwtn
 golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapKDTa4BR/hXlZSLoq2Wpct/0txZ28=
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
-golang.org/x/tools v0.40.0 h1:yLkxfA+Qnul4cs9QA3KnlFu0lVmd8JJfoq+E41uSutA=
-golang.org/x/tools v0.40.0/go.mod h1:Ik/tzLRlbscWpqqMRjyWYDisX8bG13FrdXp3o4Sr9lc=
+golang.org/x/tools v0.43.0 h1:12BdW9CeB3Z+J/I/wj34VMl8X+fEXBxVR90JeMX5E7s=
+golang.org/x/tools v0.43.0/go.mod h1:uHkMso649BX2cZK6+RpuIPXS3ho2hZo4FVwfoy1vIk0=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
@@ -235,8 +235,8 @@ golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2 h1:B82qJJgjvYKsXS9jeu
 golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2/go.mod h1:deeaetjYA+DHMHg+sMSMI58GrEteJUUzzw7en6TJQcI=
 golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b h1:J1CaxgLerRR5lgx3wnr6L04cJFbWoceSK9JWBdglINo=
 golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b/go.mod h1:tqur9LnfstdR9ep2LaJT4lFUl0EjlHtge+gAjmsHUG4=
-golang.zx2c4.com/wireguard/windows v0.5.3 h1:On6j2Rpn3OEMXqBq00QEDC7bWSZrPIHKIus8eIuExIE=
-golang.zx2c4.com/wireguard/windows v0.5.3/go.mod h1:9TEe8TJmtwyQebdFwAkEWOPr3prrtqm+REGFifP60hI=
+golang.zx2c4.com/wireguard/windows v0.6.1 h1:XMaKojH1Hs/raMrmnir4n35nTvzvWj7NmSYzHn2F4qU=
+golang.zx2c4.com/wireguard/windows v0.6.1/go.mod h1:04aqInu5GYuTFvMuDw/rKBAF7mHrltW/3rekpfbbZDM=
 google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4=
 google.golang.org/protobuf v0.0.0-20200109180630-ec00e32a8dfd/go.mod h1:DFci5gLYBciE7Vtevhsrf46CRTquxDuWsQurQQe4oz8=
 google.golang.org/protobuf v0.0.0-20200221191635-4d8936d0db64/go.mod h1:kwYJMbMJ01Woi6D6+Kah6886xMZcty6N08ah7+eCXa0=

handshake_manager.go

@@ -590,7 +590,7 @@ func (hm *HandshakeManager) allocateIndex(hh *HandshakeHostInfo) error {
 	hm.Lock()
 	defer hm.Unlock()
 
-	for i := 0; i < 32; i++ {
+	for range 32 {
 		index, err := generateIndex(hm.l)
 		if err != nil {
 			return err

hostmap_tester.go

@@ -1,5 +1,4 @@
 //go:build e2e_testing
-// +build e2e_testing
 
 package nebula
 

inside.go

@@ -11,7 +11,7 @@ import (
 	"github.com/slackhq/nebula/routing"
 )
 
-func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache) {
+func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb []byte, batch *sendBatch, rejectBuf []byte, q int, localCache firewall.ConntrackCache) {
 	err := newPacket(packet, false, fwPacket)
 	if err != nil {
 		if f.l.Level >= logrus.DebugLevel {
@@ -33,7 +33,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 		// routes packets from the Nebula addr to the Nebula addr through the Nebula
 		// TUN device.
 		if immediatelyForwardToSelf {
-			_, err := f.readers[q].Write(packet)
+			_, err := f.readers[q].WriteReject(packet)
 			if err != nil {
 				f.l.WithError(err).Error("Failed to forward to tun")
 			}
@@ -53,7 +53,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	})
 
 	if hostinfo == nil {
-		f.rejectInside(packet, out, q)
+		f.rejectInside(packet, rejectBuf, q)
 		if f.l.Level >= logrus.DebugLevel {
 			f.l.WithField("vpnAddr", fwPacket.RemoteAddr).
 				WithField("fwPacket", fwPacket).
@@ -68,10 +68,10 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 
 	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason == nil {
-		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q)
+		f.sendInsideMessage(hostinfo, packet, nb, batch, rejectBuf, q)
 
 	} else {
-		f.rejectInside(packet, out, q)
+		f.rejectInside(packet, rejectBuf, q)
 		if f.l.Level >= logrus.DebugLevel {
 			hostinfo.logger(f.l).
 				WithField("fwPacket", fwPacket).
@@ -81,6 +81,63 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	}
 }
 
+// sendInsideMessage encrypts a firewall-approved inside packet into the
+// caller's batch slot for later sendmmsg flush. When hostinfo.remote is not
+// valid we fall through to the relay slow path via the unbatched sendNoMetrics
+// so relay behavior is unchanged.
+func (f *Interface) sendInsideMessage(hostinfo *HostInfo, p, nb []byte, batch *sendBatch, rejectBuf []byte, q int) {
+	ci := hostinfo.ConnectionState
+	if ci.eKey == nil {
+		return
+	}
+
+	if !hostinfo.remote.IsValid() {
+		// Slow path: relay fallback. Reuse rejectBuf as the ciphertext
+		// scratch; sendNoMetrics arranges header space for SendVia.
+		f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, p, nb, rejectBuf, q)
+		return
+	}
+
+	scratch := batch.Next()
+	if scratch == nil {
+		// Batch full: bypass batching and send this packet directly so we
+		// never drop traffic on over-subscribed iterations.
+		f.sendNoMetrics(header.Message, 0, ci, hostinfo, netip.AddrPort{}, p, nb, rejectBuf, q)
+		return
+	}
+
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Lock()
+	}
+	c := ci.messageCounter.Add(1)
+
+	out := header.Encode(scratch, header.Version, header.Message, 0, hostinfo.remoteIndexId, c)
+	f.connectionManager.Out(hostinfo)
+
+	if hostinfo.lastRebindCount != f.rebindCount {
+		//NOTE: there is an update hole if a tunnel isn't used and exactly 256 rebinds occur before the tunnel is
+		// finally used again. This tunnel would eventually be torn down and recreated if this action didn't help.
+		f.lightHouse.QueryServer(hostinfo.vpnAddrs[0])
+		hostinfo.lastRebindCount = f.rebindCount
+		if f.l.Level >= logrus.DebugLevel {
+			f.l.WithField("vpnAddrs", hostinfo.vpnAddrs).Debug("Lighthouse update triggered for punch due to rebind counter")
+		}
+	}
+
+	out, err := ci.eKey.EncryptDanger(out, out, p, c, nb)
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Unlock()
+	}
+	if err != nil {
+		hostinfo.logger(f.l).WithError(err).
+			WithField("udpAddr", hostinfo.remote).WithField("counter", c).
+			Error("Failed to encrypt outgoing packet")
+		return
+	}
+
+	batch.Commit(len(out), hostinfo.remote)
+}
+
 func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 	if !f.firewall.InSendReject {
 		return
@@ -91,7 +148,7 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 		return
 	}
 
-	_, err := f.readers[q].Write(out)
+	_, err := f.readers[q].WriteReject(out)
 	if err != nil {
 		f.l.WithError(err).Error("Failed to write to tun")
 	}

inside_bsd.go

@@ -1,5 +1,4 @@
 //go:build darwin || dragonfly || freebsd || netbsd || openbsd
-// +build darwin dragonfly freebsd netbsd openbsd
 
 package nebula
 

inside_generic.go

@@ -1,5 +1,4 @@
 //go:build !darwin && !dragonfly && !freebsd && !netbsd && !openbsd
-// +build !darwin,!dragonfly,!freebsd,!netbsd,!openbsd
 
 package nebula
 

interface.go

@@ -4,10 +4,8 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"io"
 	"net/netip"
-	"os"
-	"runtime"
+	"sync"
 	"sync/atomic"
 	"time"
 
@@ -87,7 +85,18 @@ type Interface struct {
 	conntrackCacheTimeout time.Duration
 
 	writers []udp.Conn
-	readers []io.ReadWriteCloser
+	readers []overlay.Queue
+	// tunCoalescers is one tcpCoalescer per tun queue, wrapping readers[i].
+	// decryptToTun sends plaintext into the coalescer; listenOut calls its
+	// Flush at the end of each UDP recvmmsg batch.
+	tunCoalescers []*tcpCoalescer
+	wg            sync.WaitGroup
+
+	// fatalErr holds the first unexpected reader error that caused shutdown.
+	// nil means "no fatal error" (yet)
+	fatalErr atomic.Pointer[error]
+	// triggerShutdown is a function that will be run exactly once, when onFatal swaps something non-nil into fatalErr
+	triggerShutdown func()
 
 	metricHandshakes    metrics.Histogram
 	messageMetrics      *MessageMetrics
@@ -178,7 +187,8 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		routines:              c.routines,
 		version:               c.version,
 		writers:               make([]udp.Conn, c.routines),
-		readers:               make([]io.ReadWriteCloser, c.routines),
+		readers:               make([]overlay.Queue, c.routines),
+		tunCoalescers:         make([]*tcpCoalescer, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -210,7 +220,7 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 // activate creates the interface on the host. After the interface is created, any
 // other services that want to bind listeners to its IP may do so successfully. However,
 // the interface isn't going to process anything until run() is called.
-func (f *Interface) activate() {
+func (f *Interface) activate() error {
 	// actually turn on tun dev
 
 	addr, err := f.outside.LocalAddr()
@@ -233,38 +243,63 @@ func (f *Interface) activate() {
 	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))
 
 	// Prepare n tun queues
-	var reader io.ReadWriteCloser = f.inside
+	var reader overlay.Queue = f.inside
 	for i := 0; i < f.routines; i++ {
 		if i > 0 {
 			reader, err = f.inside.NewMultiQueueReader()
 			if err != nil {
-				f.l.Fatal(err)
+				return err
 			}
 		}
 		f.readers[i] = reader
+		f.tunCoalescers[i] = newTCPCoalescer(reader)
 	}
 
-	if err := f.inside.Activate(); err != nil {
+	f.wg.Add(1) // for us to wait on Close() to return
+	if err = f.inside.Activate(); err != nil {
 		f.inside.Close()
-		f.l.Fatal(err)
+		return err
 	}
+
+	return nil
 }
 
-func (f *Interface) run() {
+func (f *Interface) run() (func() error, error) {
 	// Launch n queues to read packets from udp
 	for i := 0; i < f.routines; i++ {
-		go f.listenOut(i)
+		f.wg.Go(func() {
+			f.listenOut(i)
+		})
 	}
 
 	// Launch n queues to read packets from tun dev
 	for i := 0; i < f.routines; i++ {
-		go f.listenIn(f.readers[i], i)
+		f.wg.Go(func() {
+			f.listenIn(f.readers[i], i)
+		})
+	}
+
+	return func() error {
+		f.wg.Wait()
+		if e := f.fatalErr.Load(); e != nil {
+			return *e
+		}
+		return nil
+	}, nil
+}
+
+// onFatal stores the first fatal reader error, and calls triggerShutdown if it was the first one
+func (f *Interface) onFatal(err error) {
+	swapped := f.fatalErr.CompareAndSwap(nil, &err)
+	if !swapped {
+		return
+	}
+	if f.triggerShutdown != nil {
+		f.triggerShutdown()
 	}
 }
 
 func (f *Interface) listenOut(i int) {
-	runtime.LockOSThread()
-
 	var li udp.Conn
 	if i > 0 {
 		li = f.writers[i]
@@ -274,40 +309,119 @@ func (f *Interface) listenOut(i int) {
 
 	ctCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
 	lhh := f.lightHouse.NewRequestHandler()
-	plaintext := make([]byte, udp.MTU)
 	h := &header.H{}
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
-	li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
-		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get(f.l))
+	// plaintexts is a ring of decrypt scratches, one per packet in a UDP
+	// recvmmsg batch. The coalescer borrows payload slices from here and
+	// requires they stay valid until Flush — so we rotate each packet and
+	// reset only in the batch-end flush callback.
+	var plaintexts [][]byte
+	idx := 0
+	coalescer := f.tunCoalescers[i]
+	err := li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
+		if idx >= len(plaintexts) {
+			plaintexts = append(plaintexts, make([]byte, udp.MTU))
+		}
+		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintexts[idx][:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get(f.l))
+		idx++
+	}, func() {
+		if err := coalescer.Flush(); err != nil {
+			f.l.WithError(err).Error("Failed to flush tun coalescer")
+		}
+		idx = 0
 	})
+
+	if err != nil && !f.closed.Load() {
+		f.l.WithError(err).Error("Error while reading inbound packet, closing")
+		f.onFatal(err)
+	}
+
+	f.l.Infof("underlay reader %v is done", i)
 }
 
-func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
-	runtime.LockOSThread()
-
-	packet := make([]byte, mtu)
-	out := make([]byte, mtu)
+func (f *Interface) listenIn(reader overlay.Queue, i int) {
+	rejectBuf := make([]byte, mtu)
+	batch := newSendBatch(sendBatchCap, udp.MTU+32)
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
 	conntrackCache := firewall.NewConntrackCacheTicker(f.conntrackCacheTimeout)
 
 	for {
-		n, err := reader.Read(packet)
+		pkts, err := reader.Read()
 		if err != nil {
-			if errors.Is(err, os.ErrClosed) && f.closed.Load() {
-				return
+			if !f.closed.Load() {
+				f.l.WithError(err).WithField("reader", i).Error("Error while reading outbound packet, closing")
+				f.onFatal(err)
 			}
-
-			f.l.WithError(err).Error("Error while reading outbound packet")
-			// This only seems to happen when something fatal happens to the fd, so exit.
-			os.Exit(2)
+			break
 		}
 
-		f.consumeInsidePacket(packet[:n], fwPacket, nb, out, i, conntrackCache.Get(f.l))
+		batch.Reset()
+		for _, pkt := range pkts {
+			if batch.Len() >= batch.Cap() {
+				f.flushBatch(batch, i)
+				batch.Reset()
+			}
+			f.consumeInsidePacket(pkt, fwPacket, nb, batch, rejectBuf, i, conntrackCache.Get(f.l))
+		}
+		if batch.Len() > 0 {
+			f.flushBatch(batch, i)
+		}
 	}
+
+	f.l.Infof("overlay reader %v is done", i)
+}
+
+func (f *Interface) flushBatch(batch *sendBatch, q int) {
+	//if len(batch.bufs) == 1 {
+	//	if err := f.writers[q].WriteTo(batch.bufs[0], batch.dsts[0]); err != nil {
+	//		f.l.WithError(err).WithField("writer", q).Error("Failed to write outgoing single-batch")
+	//	}
+	//	return
+	//}
+	w := f.writers[q]
+	if w.SupportsGSO() {
+		if segSize, ok := batchSegmentable(batch); ok {
+			if err := w.WriteSegmented(batch.bufs, batch.dsts[0], segSize); err != nil {
+				f.l.WithError(err).WithField("writer", q).Error("Failed to write outgoing GSO batch")
+			}
+			return
+		}
+	}
+	if err := w.WriteBatch(batch.bufs, batch.dsts); err != nil {
+		f.l.WithError(err).WithField("writer", q).Error("Failed to write outgoing batch")
+	}
+}
+
+// batchSegmentable reports whether a batch can be emitted as a single UDP GSO
+// superpacket: all packets go to the same destination, and every packet
+// except possibly the last has the same length. Returns the segment size on
+// success. The single-packet case is handled in flushBatch before this runs.
+func batchSegmentable(b *sendBatch) (int, bool) {
+	segSize := len(b.bufs[0])
+	if segSize == 0 {
+		return 0, false
+	}
+	dst := b.dsts[0]
+	last := len(b.bufs) - 1
+	for i := 1; i <= last; i++ {
+		if b.dsts[i] != dst {
+			return 0, false
+		}
+		if i < last {
+			if len(b.bufs[i]) != segSize {
+				return 0, false
+			}
+		} else {
+			if len(b.bufs[i]) == 0 || len(b.bufs[i]) > segSize {
+				return 0, false
+			}
+		}
+	}
+	return segSize, true
 }
 
 func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
@@ -482,23 +596,19 @@ func (f *Interface) GetCertState() *CertState {
 }
 
 func (f *Interface) Close() error {
+	var err error
 	f.closed.Store(true)
 
-	for _, u := range f.writers {
-		err := u.Close()
+	// Release the udp readers
+	for i, u := range f.writers {
+		err = u.Close()
 		if err != nil {
-			f.l.WithError(err).Error("Error while closing udp socket")
-		}
-	}
-	for i, r := range f.readers {
-		if i == 0 {
-			continue // f.readers[0] is f.inside, which we want to save for last
-		}
-		if err := r.Close(); err != nil {
-			f.l.WithError(err).Error("Error while closing tun reader")
+			f.l.WithError(err).WithField("writer", i).Error("Error while closing udp socket")
 		}
 	}
 
-	// Release the tun device
-	return f.inside.Close()
+	// Release the tun device (closing the tun also closes all readers)
+	err = f.inside.Close()
+	f.wg.Done()
+	return err
 }

lighthouse.go

@@ -713,21 +713,14 @@ func (lh *LightHouse) unlockedShouldAddV6(vpnAddr netip.Addr, to *V6AddrPort) bo
 
 func (lh *LightHouse) IsLighthouseAddr(vpnAddr netip.Addr) bool {
 	l := lh.GetLighthouses()
-	for i := range l {
-		if l[i] == vpnAddr {
-			return true
-		}
-	}
-	return false
+	return slices.Contains(l, vpnAddr)
 }
 
 func (lh *LightHouse) IsAnyLighthouseAddr(vpnAddrs []netip.Addr) bool {
 	l := lh.GetLighthouses()
 	for i := range vpnAddrs {
-		for j := range l {
-			if l[j] == vpnAddrs[i] {
-				return true
-			}
+		if slices.Contains(l, vpnAddrs[i]) {
+			return true
 		}
 	}
 	return false

main.go

@@ -105,11 +105,7 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 		// deprecated and undocumented
 		tunQueues := c.GetInt("tun.routines", 1)
 		udpQueues := c.GetInt("listen.routines", 1)
-		if tunQueues > udpQueues {
-			routines = tunQueues
-		} else {
-			routines = udpQueues
-		}
+		routines = max(tunQueues, udpQueues)
 		if routines != 1 {
 			l.WithField("routines", routines).Warn("Setting tun.routines and listen.routines is deprecated. Use `routines` instead")
 		}
@@ -292,15 +288,15 @@ func Main(c *config.C, configTest bool, buildVersion string, logger *logrus.Logg
 	}
 
 	return &Control{
-		ifce,
-		l,
-		ctx,
-		cancel,
-		sshStart,
-		statsStart,
-		dnsStart,
-		lightHouse.StartUpdateWorker,
-		connManager.Start,
+		f:                      ifce,
+		l:                      l,
+		ctx:                    ctx,
+		cancel:                 cancel,
+		sshStart:               sshStart,
+		statsStart:             statsStart,
+		dnsStart:               dnsStart,
+		lighthouseStart:        lightHouse.StartUpdateWorker,
+		connectionManagerStart: connManager.Start,
 	}, nil
 }
 

notboring.go

@@ -1,5 +1,4 @@
 //go:build !boringcrypto
-// +build !boringcrypto
 
 package nebula
 

outside.go

@@ -190,6 +190,13 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 		if !f.handleEncrypted(ci, via, h) {
 			return
 		}
+		_, err = f.decrypt(hostinfo, h.MessageCounter, out, packet, h, nb)
+		if err != nil {
+			hostinfo.logger(f.l).WithError(err).WithField("from", via).
+				WithField("packet", packet).
+				Error("Failed to decrypt CloseTunnel packet")
+			return
+		}
 
 		hostinfo.logger(f.l).WithField("from", via).
 			Info("Close tunnel received, tearing down.")
@@ -327,13 +334,29 @@ func parseV6(data []byte, incoming bool, fp *firewall.Packet) error {
 		proto := layers.IPProtocol(data[protoAt])
 
 		switch proto {
-		case layers.IPProtocolICMPv6, layers.IPProtocolESP, layers.IPProtocolNoNextHeader:
+		case layers.IPProtocolESP, layers.IPProtocolNoNextHeader:
 			fp.Protocol = uint8(proto)
 			fp.RemotePort = 0
 			fp.LocalPort = 0
 			fp.Fragment = false
 			return nil
 
+		case layers.IPProtocolICMPv6:
+			if dataLen < offset+6 {
+				return ErrIPv6PacketTooShort
+			}
+			fp.Protocol = uint8(proto)
+			fp.LocalPort = 0 //incoming vs outgoing doesn't matter for icmpv6
+			icmptype := data[offset+1]
+			switch icmptype {
+			case layers.ICMPv6TypeEchoRequest, layers.ICMPv6TypeEchoReply:
+				fp.RemotePort = binary.BigEndian.Uint16(data[offset+4 : offset+6]) //identifier
+			default:
+				fp.RemotePort = 0
+			}
+			fp.Fragment = false
+			return nil
+
 		case layers.IPProtocolTCP, layers.IPProtocolUDP:
 			if dataLen < offset+4 {
 				return ErrIPv6PacketTooShort
@@ -423,34 +446,38 @@ func parseV4(data []byte, incoming bool, fp *firewall.Packet) error {
 
 	// Accounting for a variable header length, do we have enough data for our src/dst tuples?
 	minLen := ihl
-	if !fp.Fragment && fp.Protocol != firewall.ProtoICMP {
-		minLen += minFwPacketLen
+	if !fp.Fragment {
+		if fp.Protocol == firewall.ProtoICMP {
+			minLen += minFwPacketLen + 2
+		} else {
+			minLen += minFwPacketLen
+		}
 	}
+
 	if len(data) < minLen {
 		return ErrIPv4InvalidHeaderLength
 	}
 
-	// Firewall packets are locally oriented
-	if incoming {
+	if incoming { // Firewall packets are locally oriented
 		fp.RemoteAddr, _ = netip.AddrFromSlice(data[12:16])
 		fp.LocalAddr, _ = netip.AddrFromSlice(data[16:20])
-		if fp.Fragment || fp.Protocol == firewall.ProtoICMP {
-			fp.RemotePort = 0
-			fp.LocalPort = 0
-		} else {
-			fp.RemotePort = binary.BigEndian.Uint16(data[ihl : ihl+2])
-			fp.LocalPort = binary.BigEndian.Uint16(data[ihl+2 : ihl+4])
-		}
 	} else {
 		fp.LocalAddr, _ = netip.AddrFromSlice(data[12:16])
 		fp.RemoteAddr, _ = netip.AddrFromSlice(data[16:20])
-		if fp.Fragment || fp.Protocol == firewall.ProtoICMP {
-			fp.RemotePort = 0
-			fp.LocalPort = 0
-		} else {
-			fp.LocalPort = binary.BigEndian.Uint16(data[ihl : ihl+2])
-			fp.RemotePort = binary.BigEndian.Uint16(data[ihl+2 : ihl+4])
-		}
+	}
+
+	if fp.Fragment {
+		fp.RemotePort = 0
+		fp.LocalPort = 0
+	} else if fp.Protocol == firewall.ProtoICMP { //note that orientation doesn't matter on ICMP
+		fp.RemotePort = binary.BigEndian.Uint16(data[ihl+4 : ihl+6]) //identifier
+		fp.LocalPort = 0                                             //code would be uint16(data[ihl+1])
+	} else if incoming {
+		fp.RemotePort = binary.BigEndian.Uint16(data[ihl : ihl+2])  //src port
+		fp.LocalPort = binary.BigEndian.Uint16(data[ihl+2 : ihl+4]) //dst port
+	} else {
+		fp.LocalPort = binary.BigEndian.Uint16(data[ihl : ihl+2])    //src port
+		fp.RemotePort = binary.BigEndian.Uint16(data[ihl+2 : ihl+4]) //dst port
 	}
 
 	return nil
@@ -508,7 +535,7 @@ func (f *Interface) decryptToTun(hostinfo *HostInfo, messageCounter uint64, out
 	}
 
 	f.connectionManager.In(hostinfo)
-	_, err = f.readers[q].Write(out)
+	err = f.tunCoalescers[q].Add(out)
 	if err != nil {
 		f.l.WithError(err).Error("Failed to write to tun")
 	}

outside_test.go

@@ -155,6 +155,7 @@ func Test_newPacket_v6(t *testing.T) {
 	// next layer, missing length byte
 	err = newPacket(buffer.Bytes()[:49], true, p)
 	require.ErrorIs(t, err, ErrIPv6CouldNotFindPayload)
+	err = nil
 
 	// A good ICMP packet
 	ip = layers.IPv6{
@@ -165,20 +166,26 @@ func Test_newPacket_v6(t *testing.T) {
 		DstIP:      net.IPv6linklocalallnodes,
 	}
 
-	icmp := layers.ICMPv6{}
-
-	buffer.Clear()
-	err = gopacket.SerializeLayers(buffer, opt, &ip, &icmp)
-	if err != nil {
-		panic(err)
+	icmp := layers.ICMPv6{
+		TypeCode: layers.ICMPv6TypeEchoRequest,
+		Checksum: 0x1234,
 	}
 
-	err = newPacket(buffer.Bytes(), true, p)
-	require.NoError(t, err)
+	buffer.Clear()
+	require.NoError(t, gopacket.SerializeLayers(buffer, opt, &ip, &icmp))
+	require.Error(t, newPacket(buffer.Bytes(), true, p))
+
+	buffer.Clear()
+	echo := layers.ICMPv6Echo{
+		Identifier: 0xabcd,
+		SeqNumber:  1234,
+	}
+	require.NoError(t, gopacket.SerializeLayers(buffer, opt, &ip, &icmp, &echo))
+	require.NoError(t, newPacket(buffer.Bytes(), true, p))
 	assert.Equal(t, uint8(layers.IPProtocolICMPv6), p.Protocol)
 	assert.Equal(t, netip.MustParseAddr("ff02::2"), p.RemoteAddr)
 	assert.Equal(t, netip.MustParseAddr("ff02::1"), p.LocalAddr)
-	assert.Equal(t, uint16(0), p.RemotePort)
+	assert.Equal(t, uint16(0xabcd), p.RemotePort)
 	assert.Equal(t, uint16(0), p.LocalPort)
 	assert.False(t, p.Fragment)
 
@@ -574,7 +581,7 @@ func BenchmarkParseV6(b *testing.B) {
 	}
 
 	evilBytes := buffer.Bytes()
-	for i := 0; i < 200; i++ {
+	for range 200 {
 		evilBytes = append(evilBytes, hopHeader...)
 	}
 	evilBytes = append(evilBytes, lastHopHeader...)

overlay/device.go

@@ -7,12 +7,63 @@ import (
 	"github.com/slackhq/nebula/routing"
 )
 
+// defaultBatchBufSize is the per-Queue scratch size for Read on backends
+// that don't do TSO segmentation. 65535 covers any single IP packet.
+const defaultBatchBufSize = 65535
+
+// Queue is a readable/writable tun queue. One Queue is driven by a single
+// read goroutine plus concurrent writers (see Write / WriteReject below).
+type Queue interface {
+	io.Closer
+
+	// Read returns one or more packets. The returned slices are borrowed
+	// from the Queue's internal buffer and are only valid until the next
+	// Read or Close on this Queue — callers must encrypt or copy each
+	// slice before the next call. Not safe for concurrent Reads; exactly
+	// one goroutine per Queue reads.
+	Read() ([][]byte, error)
+
+	// Write emits a single packet on the plaintext (outside→inside)
+	// delivery path. May run concurrently with WriteReject on the same
+	// Queue, but not with itself.
+	Write(p []byte) (int, error)
+
+	// WriteReject writes a single packet that originated from the inside
+	// path (reject replies or self-forward) using scratch state distinct
+	// from Write, so it can run concurrently with Write on the same Queue
+	// without a data race. On backends without a shared-scratch Write, a
+	// trivial delegation to Write is acceptable.
+	WriteReject(p []byte) (int, error)
+}
+
 type Device interface {
-	io.ReadWriteCloser
+	Queue
 	Activate() error
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
 	SupportsMultiqueue() bool
-	NewMultiQueueReader() (io.ReadWriteCloser, error)
+	NewMultiQueueReader() (Queue, error)
+}
+
+// GSOWriter is implemented by Queues that can emit a TCP TSO superpacket
+// assembled from a header prefix plus one or more borrowed payload
+// fragments, in a single vectored write (writev with a leading
+// virtio_net_hdr). This lets the coalescer avoid copying payload bytes
+// between the caller's decrypt buffer and the TUN. Backends without GSO
+// support return false from GSOSupported and coalescing is skipped.
+//
+// hdr contains the IPv4/IPv6 + TCP header prefix (mutable — callers will
+// have filled in total length and pseudo-header partial). pays are
+// non-overlapping payload fragments whose concatenation is the full
+// superpacket payload; they are read-only from the writer's perspective
+// and must remain valid until the call returns. gsoSize is the MSS:
+// every segment except possibly the last is exactly that many bytes.
+// csumStart is the byte offset where the TCP header begins within hdr.
+//
+// hdr's TCP checksum field must already hold the pseudo-header partial
+// sum (single-fold, not inverted), per virtio NEEDS_CSUM semantics.
+type GSOWriter interface {
+	WriteGSO(hdr []byte, pays [][]byte, gsoSize uint16, isV6 bool, csumStart uint16) error
+	GSOSupported() bool
 }

overlay/noop.go

@@ -1,8 +1,7 @@
-package test
+package overlay
 
 import (
 	"errors"
-	"io"
 	"net/netip"
 
 	"github.com/slackhq/nebula/routing"
@@ -26,19 +25,23 @@ func (NoopTun) Name() string {
 	return "noop"
 }
 
-func (NoopTun) Read([]byte) (int, error) {
-	return 0, nil
+func (NoopTun) Read() ([][]byte, error) {
+	return nil, nil
 }
 
 func (NoopTun) Write([]byte) (int, error) {
 	return 0, nil
 }
 
+func (NoopTun) WriteReject(p []byte) (int, error) {
+	return 0, nil
+}
+
 func (NoopTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (NoopTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (NoopTun) NewMultiQueueReader() (Queue, error) {
 	return nil, errors.New("unsupported")
 }
 

overlay/tun_android.go

@@ -18,12 +18,39 @@ import (
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	fd          int
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *logrus.Logger
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.rwc.Read(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) Write(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) Close() error {
+	return t.rwc.Close()
 }
 
 func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
@@ -32,10 +59,10 @@ func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []net
 	file := os.NewFile(uintptr(deviceFd), "/dev/net/tun")
 
 	t := &tun{
-		ReadWriteCloser: file,
-		fd:              deviceFd,
-		vpnNetworks:     vpnNetworks,
-		l:               l,
+		rwc:         file,
+		fd:          deviceFd,
+		vpnNetworks: vpnNetworks,
+		l:           l,
 	}
 
 	err := t.reload(c, true)
@@ -99,6 +126,6 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for android")
 }

overlay/tun_darwin.go

@@ -23,7 +23,7 @@ import (
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	Device      string
 	vpnNetworks []netip.Prefix
 	DefaultMTU  int
@@ -34,6 +34,9 @@ type tun struct {
 
 	// cache out buffer since we need to prepend 4 bytes for tun metadata
 	out []byte
+
+	readBuf  []byte
+	batchRet [1][]byte
 }
 
 type ifReq struct {
@@ -124,11 +127,11 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, _ bool) (
 	}
 
 	t := &tun{
-		ReadWriteCloser: os.NewFile(uintptr(fd), ""),
-		Device:          name,
-		vpnNetworks:     vpnNetworks,
-		DefaultMTU:      c.GetInt("tun.mtu", DefaultMTU),
-		l:               l,
+		rwc:         os.NewFile(uintptr(fd), ""),
+		Device:      name,
+		vpnNetworks: vpnNetworks,
+		DefaultMTU:  c.GetInt("tun.mtu", DefaultMTU),
+		l:           l,
 	}
 
 	err = t.reload(c, true)
@@ -158,8 +161,8 @@ func newTunFromFd(_ *config.C, _ *logrus.Logger, _ int, _ []netip.Prefix) (*tun,
 }
 
 func (t *tun) Close() error {
-	if t.ReadWriteCloser != nil {
-		return t.ReadWriteCloser.Close()
+	if t.rwc != nil {
+		return t.rwc.Close()
 	}
 	return nil
 }
@@ -503,15 +506,31 @@ func delRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	buf := make([]byte, len(to)+4)
 
-	n, err := t.ReadWriteCloser.Read(buf)
+	n, err := t.rwc.Read(buf)
 
 	copy(to, buf[4:])
 	return n - 4, err
 }
 
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
+}
+
 // Write is only valid for single threaded use
 func (t *tun) Write(from []byte) (int, error) {
 	buf := t.out
@@ -537,7 +556,7 @@ func (t *tun) Write(from []byte) (int, error) {
 
 	copy(buf[4:], from)
 
-	n, err := t.ReadWriteCloser.Write(buf)
+	n, err := t.rwc.Write(buf)
 	return n - 4, err
 }
 
@@ -553,6 +572,6 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for darwin")
 }

overlay/tun_disabled.go

@@ -20,6 +20,23 @@ type disabledTun struct {
 	tx metrics.Counter
 	rx metrics.Counter
 	l  *logrus.Logger
+
+	batchRet [1][]byte
+}
+
+func (t *disabledTun) Read() ([][]byte, error) {
+	r, ok := <-t.read
+	if !ok {
+		return nil, io.EOF
+	}
+
+	t.tx.Inc(1)
+	if t.l.Level >= logrus.DebugLevel {
+		t.l.WithField("raw", prettyPacket(r)).Debugf("Write payload")
+	}
+
+	t.batchRet[0] = r
+	return t.batchRet[:], nil
 }
 
 func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled bool, l *logrus.Logger) *disabledTun {
@@ -56,24 +73,6 @@ func (*disabledTun) Name() string {
 	return "disabled"
 }
 
-func (t *disabledTun) Read(b []byte) (int, error) {
-	r, ok := <-t.read
-	if !ok {
-		return 0, io.EOF
-	}
-
-	if len(r) > len(b) {
-		return 0, fmt.Errorf("packet larger than mtu: %d > %d bytes", len(r), len(b))
-	}
-
-	t.tx.Inc(1)
-	if t.l.Level >= logrus.DebugLevel {
-		t.l.WithField("raw", prettyPacket(r)).Debugf("Write payload")
-	}
-
-	return copy(b, r), nil
-}
-
 func (t *disabledTun) handleICMPEchoRequest(b []byte) bool {
 	out := make([]byte, len(b))
 	out = iputil.CreateICMPEchoResponse(b, out)
@@ -105,11 +104,15 @@ func (t *disabledTun) Write(b []byte) (int, error) {
 	return len(b), nil
 }
 
+func (t *disabledTun) WriteReject(b []byte) (int, error) {
+	return t.Write(b)
+}
+
 func (t *disabledTun) SupportsMultiqueue() bool {
 	return true
 }
 
-func (t *disabledTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *disabledTun) NewMultiQueueReader() (Queue, error) {
 	return t, nil
 }
 

overlay/tun_freebsd.go

@@ -7,7 +7,6 @@ import (
 	"bytes"
 	"errors"
 	"fmt"
-	"io"
 	"io/fs"
 	"net/netip"
 	"sync/atomic"
@@ -94,9 +93,28 @@ type tun struct {
 	linkAddr    *netroute.LinkAddr
 	l           *logrus.Logger
 	devFd       int
+
+	readBuf  []byte
+	batchRet [1][]byte
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
+}
+
+func (t *tun) readOne(to []byte) (int, error) {
 	// use readv() to read from the tunnel device, to eliminate the need for copying the buffer
 	if t.devFd < 0 {
 		return -1, syscall.EINVAL
@@ -454,7 +472,7 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for freebsd")
 }
 

overlay/tun_ios.go

@@ -21,11 +21,38 @@ import (
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *logrus.Logger
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.rwc.Read(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) Write(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) Close() error {
+	return t.rwc.Close()
 }
 
 func newTun(_ *config.C, _ *logrus.Logger, _ []netip.Prefix, _ bool) (*tun, error) {
@@ -35,9 +62,9 @@ func newTun(_ *config.C, _ *logrus.Logger, _ []netip.Prefix, _ bool) (*tun, erro
 func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
 	file := os.NewFile(uintptr(deviceFd), "/dev/tun")
 	t := &tun{
-		vpnNetworks:     vpnNetworks,
-		ReadWriteCloser: &tunReadCloser{f: file},
-		l:               l,
+		vpnNetworks: vpnNetworks,
+		rwc:         &tunReadCloser{f: file},
+		l:           l,
 	}
 
 	err := t.reload(c, true)
@@ -155,6 +182,6 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for ios")
 }

overlay/tun_linux.go

@@ -4,14 +4,17 @@
 package overlay
 
 import (
+	"encoding/binary"
 	"fmt"
 	"io"
 	"net"
 	"net/netip"
 	"os"
+	"runtime"
 	"strings"
 	"sync"
 	"sync/atomic"
+	"syscall"
 	"time"
 	"unsafe"
 
@@ -24,9 +27,450 @@ import (
 	"golang.org/x/sys/unix"
 )
 
+// tunFile wraps a TUN file descriptor with poll-based reads. The FD provided will be changed to non-blocking.
+// A shared eventfd allows Close to wake all readers blocked in poll.
+type tunFile struct {
+	fd         int
+	shutdownFd int
+	lastOne    bool
+	readPoll   [2]unix.PollFd
+	writePoll  [2]unix.PollFd
+	closed     bool
+
+	// vnetHdr is true when this fd was opened with IFF_VNET_HDR and the
+	// kernel successfully accepted TUNSETOFFLOAD. Reads include a leading
+	// virtio_net_hdr and may carry a TSO superpacket we must segment;
+	// writes must prepend a zeroed virtio_net_hdr.
+	vnetHdr   bool
+	readBuf   []byte        // scratch for a single raw read (virtio hdr + superpacket)
+	segBuf    []byte        // backing store for segmented output
+	segOff    int           // cursor into segBuf for the current Read drain
+	pending   [][]byte      // segments returned from the most recent Read
+	writeIovs [2]unix.Iovec // preallocated iovecs for Write (coalescer passthrough); iovs[0] is fixed to validVnetHdr
+	// rejectIovs is a second preallocated iovec scratch used exclusively by
+	// WriteReject (reject + self-forward from the inside path). It mirrors
+	// writeIovs but lets listenIn goroutines emit reject packets without
+	// racing with the listenOut coalescer that owns writeIovs.
+	rejectIovs [2]unix.Iovec
+
+	// gsoHdrBuf is a per-queue 10-byte scratch for the virtio_net_hdr emitted
+	// by WriteGSO. Separate from validVnetHdr so a concurrent non-GSO Write on
+	// another queue never observes a half-written header.
+	gsoHdrBuf [virtioNetHdrLen]byte
+	// gsoIovs is the writev iovec scratch for WriteGSO. Sized to hold the
+	// virtio header + IP/TCP header + up to gsoInitialPayIovs payload
+	// fragments; grown on demand if a coalescer pushes more.
+	gsoIovs []unix.Iovec
+}
+
+// gsoInitialPayIovs is the starting capacity (in payload fragments) of
+// tunFile.gsoIovs. Sized to cover the default coalesce segment cap without
+// any reallocations.
+const gsoInitialPayIovs = 66
+
+// validVnetHdr is the 10-byte virtio_net_hdr we prepend to every non-GSO TUN
+// write. Only flag set is VIRTIO_NET_HDR_F_DATA_VALID, which marks the skb
+// CHECKSUM_UNNECESSARY so the receiving network stack skips L4 checksum
+// verification. All packets that reach the plain Write / WriteReject paths
+// already carry a valid L4 checksum (either supplied by a remote peer whose
+// ciphertext we AEAD-authenticated, or produced by finishChecksum during TSO
+// segmentation, or built locally by CreateRejectPacket), so trusting them is
+// safe.
+var validVnetHdr = [virtioNetHdrLen]byte{unix.VIRTIO_NET_HDR_F_DATA_VALID}
+
+// newFriend makes a tunFile for a MultiQueueReader that copies the shutdown eventfd from the parent tun
+func (r *tunFile) newFriend(fd int) (*tunFile, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
+	}
+	out := &tunFile{
+		fd:         fd,
+		shutdownFd: r.shutdownFd,
+		vnetHdr:    r.vnetHdr,
+		readBuf:    make([]byte, tunReadBufSize),
+		readPoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLIN},
+			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
+		},
+		writePoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLOUT},
+			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
+		},
+	}
+	if r.vnetHdr {
+		out.segBuf = make([]byte, tunSegBufCap)
+		out.writeIovs[0].Base = &validVnetHdr[0]
+		out.writeIovs[0].SetLen(virtioNetHdrLen)
+		out.rejectIovs[0].Base = &validVnetHdr[0]
+		out.rejectIovs[0].SetLen(virtioNetHdrLen)
+		out.gsoIovs = make([]unix.Iovec, 2, 2+gsoInitialPayIovs)
+		out.gsoIovs[0].Base = &out.gsoHdrBuf[0]
+		out.gsoIovs[0].SetLen(virtioNetHdrLen)
+	}
+	return out, nil
+}
+
+func newTunFd(fd int, vnetHdr bool) (*tunFile, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
+	}
+
+	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create eventfd: %w", err)
+	}
+
+	out := &tunFile{
+		fd:         fd,
+		shutdownFd: shutdownFd,
+		lastOne:    true,
+		vnetHdr:    vnetHdr,
+		readBuf:    make([]byte, tunReadBufSize),
+		readPoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLIN},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+		writePoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLOUT},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+	}
+	if vnetHdr {
+		out.segBuf = make([]byte, tunSegBufCap)
+		out.writeIovs[0].Base = &validVnetHdr[0]
+		out.writeIovs[0].SetLen(virtioNetHdrLen)
+		out.rejectIovs[0].Base = &validVnetHdr[0]
+		out.rejectIovs[0].SetLen(virtioNetHdrLen)
+		out.gsoIovs = make([]unix.Iovec, 2, 2+gsoInitialPayIovs)
+		out.gsoIovs[0].Base = &out.gsoHdrBuf[0]
+		out.gsoIovs[0].SetLen(virtioNetHdrLen)
+	}
+
+	return out, nil
+}
+
+func (r *tunFile) blockOnRead() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(r.readPoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	//always reset these!
+	tunEvents := r.readPoll[0].Revents
+	shutdownEvents := r.readPoll[1].Revents
+	r.readPoll[0].Revents = 0
+	r.readPoll[1].Revents = 0
+	//do the err check before trusting the potentially bogus bits we just got
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	} else if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (r *tunFile) blockOnWrite() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(r.writePoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	//always reset these!
+	tunEvents := r.writePoll[0].Revents
+	shutdownEvents := r.writePoll[1].Revents
+	r.writePoll[0].Revents = 0
+	r.writePoll[1].Revents = 0
+	//do the err check before trusting the potentially bogus bits we just got
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	} else if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (r *tunFile) readRaw(buf []byte) (int, error) {
+	for {
+		if n, err := unix.Read(r.fd, buf); err == nil {
+			return n, nil
+		} else if err == unix.EAGAIN {
+			if err = r.blockOnRead(); err != nil {
+				return 0, err
+			}
+			continue
+		} else {
+			return 0, err
+		}
+	}
+}
+
+// Read reads one or more superpackets from the tun and returns the
+// resulting packets. The first read blocks via poll; once the fd is known
+// readable we drain additional packets non-blocking until the kernel queue
+// is empty (EAGAIN), we've collected tunDrainCap packets, or we're out of
+// segBuf headroom. This amortizes the poll wake over bursts of small
+// packets (e.g. TCP ACKs). Slices point into the tunFile's internal buffers
+// and are only valid until the next Read or Close on this Queue.
+func (r *tunFile) Read() ([][]byte, error) {
+	r.pending = r.pending[:0]
+	r.segOff = 0
+
+	// Initial (blocking) read. Retry on decode errors so a single bad
+	// packet does not stall the reader.
+	for {
+		n, err := r.readRaw(r.readBuf)
+		if err != nil {
+			return nil, err
+		}
+		if !r.vnetHdr {
+			r.pending = append(r.pending, r.readBuf[:n])
+			// Non-vnetHdr mode shares one readBuf so we can't drain safely
+			// without copying; return the single packet as before.
+			return r.pending, nil
+		}
+		if err := r.decodeRead(n); err != nil {
+			// Drop and read again — a bad packet should not kill the reader.
+			continue
+		}
+		break
+	}
+
+	// Drain: non-blocking reads until the kernel queue is empty, the drain
+	// cap is reached, or segBuf no longer has room for another worst-case
+	// superpacket.
+	for len(r.pending) < tunDrainCap && tunSegBufCap-r.segOff >= tunSegBufSize {
+		n, err := unix.Read(r.fd, r.readBuf)
+		if err != nil {
+			// EAGAIN / EINTR / anything else: stop draining. We already
+			// have a valid batch from the first read.
+			break
+		}
+		if n <= 0 {
+			break
+		}
+		if err := r.decodeRead(n); err != nil {
+			// Drop this packet and stop the drain; we'd rather hand off
+			// what we have than keep spinning here.
+			break
+		}
+	}
+
+	return r.pending, nil
+}
+
+// decodeRead decodes the virtio header plus payload in r.readBuf[:n], appends
+// the segments to r.pending, and advances r.segOff by the total scratch used.
+// Caller must have already ensured r.vnetHdr is true.
+func (r *tunFile) decodeRead(n int) error {
+	if n < virtioNetHdrLen {
+		return fmt.Errorf("short tun read: %d < %d", n, virtioNetHdrLen)
+	}
+	var hdr virtioNetHdr
+	hdr.decode(r.readBuf[:virtioNetHdrLen])
+	before := len(r.pending)
+	if err := segmentInto(r.readBuf[virtioNetHdrLen:n], hdr, &r.pending, r.segBuf[r.segOff:]); err != nil {
+		return err
+	}
+	for k := before; k < len(r.pending); k++ {
+		r.segOff += len(r.pending[k])
+	}
+	return nil
+}
+
+func (r *tunFile) Write(buf []byte) (int, error) {
+	return r.writeWithScratch(buf, &r.writeIovs)
+}
+
+// WriteReject emits a packet using a dedicated iovec scratch (rejectIovs)
+// distinct from the one used by the coalescer's Write path. This avoids a
+// data race between the inside (listenIn) goroutine emitting reject or
+// self-forward packets and the outside (listenOut) goroutine flushing TCP
+// coalescer passthroughs on the same tunFile.
+func (r *tunFile) WriteReject(buf []byte) (int, error) {
+	return r.writeWithScratch(buf, &r.rejectIovs)
+}
+
+func (r *tunFile) writeWithScratch(buf []byte, iovs *[2]unix.Iovec) (int, error) {
+	if !r.vnetHdr {
+		for {
+			if n, err := unix.Write(r.fd, buf); err == nil {
+				return n, nil
+			} else if err == unix.EAGAIN {
+				if err = r.blockOnWrite(); err != nil {
+					return 0, err
+				}
+				continue
+			} else if err == unix.EINTR {
+				continue
+			} else {
+				return 0, err
+			}
+		}
+	}
+
+	if len(buf) == 0 {
+		return 0, nil
+	}
+	// Point the payload iovec at the caller's buffer. iovs[0] is pre-wired
+	// to validVnetHdr during tunFile construction so we don't rebuild it here.
+	iovs[1].Base = &buf[0]
+	iovs[1].SetLen(len(buf))
+	iovPtr := uintptr(unsafe.Pointer(&iovs[0]))
+	// The TUN fd is non-blocking (set in newTunFd / newFriend), so writev
+	// either completes promptly or returns EAGAIN — it cannot park the
+	// goroutine inside the kernel. That lets us use syscall.RawSyscall and
+	// skip the runtime.entersyscall / exitsyscall bookkeeping on every
+	// packet; we only pay that cost when we fall through to blockOnWrite.
+	for {
+		n, _, errno := syscall.RawSyscall(unix.SYS_WRITEV, uintptr(r.fd), iovPtr, 2)
+		if errno == 0 {
+			runtime.KeepAlive(buf)
+			if int(n) < virtioNetHdrLen {
+				return 0, io.ErrShortWrite
+			}
+			return int(n) - virtioNetHdrLen, nil
+		}
+		if errno == unix.EAGAIN {
+			runtime.KeepAlive(buf)
+			if err := r.blockOnWrite(); err != nil {
+				return 0, err
+			}
+			continue
+		}
+		if errno == unix.EINTR {
+			continue
+		}
+		runtime.KeepAlive(buf)
+		return 0, errno
+	}
+}
+
+// GSOSupported reports whether this queue was opened with IFF_VNET_HDR and
+// can accept WriteGSO. When false, callers should fall back to per-segment
+// Write calls.
+func (r *tunFile) GSOSupported() bool { return r.vnetHdr }
+
+// WriteGSO emits a TCP TSO superpacket in a single writev. hdr is the
+// IPv4/IPv6 + TCP header prefix (already finalized — total length, IP csum,
+// and TCP pseudo-header partial set by the caller). pays are payload
+// fragments whose concatenation forms the full coalesced payload; each
+// slice is read-only and must stay valid until return. gsoSize is the MSS;
+// every segment except possibly the last is exactly gsoSize bytes.
+// csumStart is the byte offset where the TCP header begins within hdr.
+func (r *tunFile) WriteGSO(hdr []byte, pays [][]byte, gsoSize uint16, isV6 bool, csumStart uint16) error {
+	if !r.vnetHdr {
+		return fmt.Errorf("WriteGSO called on tun without IFF_VNET_HDR")
+	}
+	if len(hdr) == 0 || len(pays) == 0 {
+		return nil
+	}
+
+	// Build the virtio_net_hdr. When pays total to <= gsoSize the kernel
+	// would produce a single segment; keep NEEDS_CSUM semantics but skip
+	// the GSO type so the kernel doesn't spuriously mark this as TSO.
+	vhdr := virtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		HdrLen:     uint16(len(hdr)),
+		GSOSize:    gsoSize,
+		CsumStart:  csumStart,
+		CsumOffset: 16, // TCP checksum field lives 16 bytes into the TCP header
+	}
+	var totalPay int
+	for _, p := range pays {
+		totalPay += len(p)
+	}
+	if totalPay > int(gsoSize) {
+		if isV6 {
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV6
+		} else {
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV4
+		}
+	} else {
+		vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+		vhdr.GSOSize = 0
+	}
+	vhdr.encode(r.gsoHdrBuf[:])
+
+	// Build the iovec array: [virtio_hdr, hdr, pays...]. r.gsoIovs[0] is
+	// wired to gsoHdrBuf at construction and never changes.
+	need := 2 + len(pays)
+	if cap(r.gsoIovs) < need {
+		grown := make([]unix.Iovec, need)
+		grown[0] = r.gsoIovs[0]
+		r.gsoIovs = grown
+	} else {
+		r.gsoIovs = r.gsoIovs[:need]
+	}
+	r.gsoIovs[1].Base = &hdr[0]
+	r.gsoIovs[1].SetLen(len(hdr))
+	for i, p := range pays {
+		r.gsoIovs[2+i].Base = &p[0]
+		r.gsoIovs[2+i].SetLen(len(p))
+	}
+
+	iovPtr := uintptr(unsafe.Pointer(&r.gsoIovs[0]))
+	iovCnt := uintptr(len(r.gsoIovs))
+	for {
+		n, _, errno := syscall.RawSyscall(unix.SYS_WRITEV, uintptr(r.fd), iovPtr, iovCnt)
+		if errno == 0 {
+			runtime.KeepAlive(hdr)
+			runtime.KeepAlive(pays)
+			if int(n) < virtioNetHdrLen {
+				return io.ErrShortWrite
+			}
+			return nil
+		}
+		if errno == unix.EAGAIN {
+			runtime.KeepAlive(hdr)
+			runtime.KeepAlive(pays)
+			if err := r.blockOnWrite(); err != nil {
+				return err
+			}
+			continue
+		}
+		if errno == unix.EINTR {
+			continue
+		}
+		runtime.KeepAlive(hdr)
+		runtime.KeepAlive(pays)
+		return errno
+	}
+}
+
+func (r *tunFile) wakeForShutdown() error {
+	var buf [8]byte
+	binary.NativeEndian.PutUint64(buf[:], 1)
+	_, err := unix.Write(int(r.readPoll[1].Fd), buf[:])
+	return err
+}
+
+func (r *tunFile) Close() error {
+	if r.closed { // avoid closing more than once. Technically a fd could get re-used, which would be a problem
+		return nil
+	}
+	r.closed = true
+	if r.lastOne {
+		_ = unix.Close(r.shutdownFd)
+	}
+	return unix.Close(r.fd)
+}
+
 type tun struct {
-	io.ReadWriteCloser
-	fd          int
+	*tunFile
+	readers     []*tunFile
+	closeLock   sync.Mutex
 	Device      string
 	vpnNetworks []netip.Prefix
 	MaxMTU      int
@@ -72,9 +516,9 @@ type ifreqQLEN struct {
 }
 
 func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
-	file := os.NewFile(uintptr(deviceFd), "/dev/net/tun")
-
-	t, err := newTunGeneric(c, l, file, vpnNetworks)
+	// We don't know what flags the caller opened this fd with and can't turn
+	// on IFF_VNET_HDR after TUNSETIFF, so skip offload on inherited fds.
+	t, err := newTunGeneric(c, l, deviceFd, false, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -84,46 +528,83 @@ func newTunFromFd(c *config.C, l *logrus.Logger, deviceFd int, vpnNetworks []net
 	return t, nil
 }
 
-func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+// openTunDev opens /dev/net/tun, creating the device node first if it's
+// missing (docker containers occasionally omit it).
+func openTunDev() (int, error) {
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
-	if err != nil {
-		// If /dev/net/tun doesn't exist, try to create it (will happen in docker)
-		if os.IsNotExist(err) {
-			err = os.MkdirAll("/dev/net", 0755)
-			if err != nil {
-				return nil, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
-			}
-			err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200)))
-			if err != nil {
-				return nil, fmt.Errorf("failed to create /dev/net/tun: %w", err)
-			}
-
-			fd, err = unix.Open("/dev/net/tun", os.O_RDWR, 0)
-			if err != nil {
-				return nil, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
-			}
-		} else {
-			return nil, err
-		}
+	if err == nil {
+		return fd, nil
 	}
+	if !os.IsNotExist(err) {
+		return -1, err
+	}
+	if err = os.MkdirAll("/dev/net", 0755); err != nil {
+		return -1, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
+	}
+	if err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200))); err != nil {
+		return -1, fmt.Errorf("failed to create /dev/net/tun: %w", err)
+	}
+	fd, err = unix.Open("/dev/net/tun", os.O_RDWR, 0)
+	if err != nil {
+		return -1, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
+	}
+	return fd, nil
+}
 
+// tunSetIff runs TUNSETIFF with the given flags and returns the kernel-chosen
+// device name on success.
+func tunSetIff(fd int, name string, flags uint16) (string, error) {
 	var req ifReq
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI)
+	req.Flags = flags
+	copy(req.Name[:], name)
+	if err := ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+		return "", err
+	}
+	return strings.Trim(string(req.Name[:]), "\x00"), nil
+}
+
+// tsoOffloadFlags are the TUN_F_* bits we ask the kernel to enable when a
+// TSO-capable TUN is available. CSUM is required as a prerequisite for TSO.
+const tsoOffloadFlags = unix.TUN_F_CSUM | unix.TUN_F_TSO4 | unix.TUN_F_TSO6
+
+func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+	baseFlags := uint16(unix.IFF_TUN | unix.IFF_NO_PI)
 	if multiqueue {
-		req.Flags |= unix.IFF_MULTI_QUEUE
+		baseFlags |= unix.IFF_MULTI_QUEUE
 	}
 	nameStr := c.GetString("tun.dev", "")
-	copy(req.Name[:], nameStr)
-	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
-		return nil, &NameError{
-			Name:       nameStr,
-			Underlying: err,
+
+	// First try to open with IFF_VNET_HDR + TUNSETOFFLOAD so we can receive
+	// TSO superpackets. If either step fails (older kernel, unprivileged
+	// container, etc.) we close and fall back to a plain TUN.
+	fd, err := openTunDev()
+	if err != nil {
+		return nil, err
+	}
+	vnetHdr := true
+	name, err := tunSetIff(fd, nameStr, baseFlags|unix.IFF_VNET_HDR|unix.IFF_NAPI)
+	if err != nil {
+		_ = unix.Close(fd)
+		vnetHdr = false
+	} else if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+		l.WithError(err).Warn("Failed to enable TUN offload (TSO); proceeding without virtio headers")
+		_ = unix.Close(fd)
+		vnetHdr = false
+	}
+
+	if !vnetHdr {
+		fd, err = openTunDev()
+		if err != nil {
+			return nil, err
+		}
+		name, err = tunSetIff(fd, nameStr, baseFlags)
+		if err != nil {
+			_ = unix.Close(fd)
+			return nil, &NameError{Name: nameStr, Underlying: err}
 		}
 	}
-	name := strings.Trim(string(req.Name[:]), "\x00")
 
-	file := os.NewFile(uintptr(fd), "/dev/net/tun")
-	t, err := newTunGeneric(c, l, file, vpnNetworks)
+	t, err := newTunGeneric(c, l, fd, vnetHdr, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -133,10 +614,17 @@ func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, multiqueu
 	return t, nil
 }
 
-func newTunGeneric(c *config.C, l *logrus.Logger, file *os.File, vpnNetworks []netip.Prefix) (*tun, error) {
+// newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
+func newTunGeneric(c *config.C, l *logrus.Logger, fd int, vnetHdr bool, vpnNetworks []netip.Prefix) (*tun, error) {
+	tfd, err := newTunFd(fd, vnetHdr)
+	if err != nil {
+		_ = unix.Close(fd)
+		return nil, err
+	}
 	t := &tun{
-		ReadWriteCloser:           file,
-		fd:                        int(file.Fd()),
+		tunFile:                   tfd,
+		readers:                   []*tunFile{tfd},
+		closeLock:                 sync.Mutex{},
 		vpnNetworks:               vpnNetworks,
 		TXQueueLen:                c.GetInt("tun.tx_queue", 500),
 		useSystemRoutes:           c.GetBool("tun.use_system_route_table", false),
@@ -145,8 +633,8 @@ func newTunGeneric(c *config.C, l *logrus.Logger, file *os.File, vpnNetworks []n
 		l:                         l,
 	}
 
-	err := t.reload(c, true)
-	if err != nil {
+	if err = t.reload(c, true); err != nil {
+		_ = t.Close()
 		return nil, err
 	}
 
@@ -238,22 +726,40 @@ func (t *tun) SupportsMultiqueue() bool {
 	return true
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (Queue, error) {
+	t.closeLock.Lock()
+	defer t.closeLock.Unlock()
+
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
 		return nil, err
 	}
 
-	var req ifReq
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
-	copy(req.Name[:], t.Device)
-	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+	flags := uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
+	if t.vnetHdr {
+		flags |= unix.IFF_VNET_HDR | unix.IFF_NAPI
+	}
+	if _, err = tunSetIff(fd, t.Device, flags); err != nil {
+		_ = unix.Close(fd)
 		return nil, err
 	}
 
-	file := os.NewFile(uintptr(fd), "/dev/net/tun")
+	if t.vnetHdr {
+		if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+			_ = unix.Close(fd)
+			return nil, fmt.Errorf("failed to enable offload on multiqueue tun fd: %w", err)
+		}
+	}
 
-	return file, nil
+	out, err := t.tunFile.newFriend(fd)
+	if err != nil {
+		_ = unix.Close(fd)
+		return nil, err
+	}
+
+	t.readers = append(t.readers, out)
+
+	return out, nil
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
@@ -261,29 +767,6 @@ func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
 	return r
 }
 
-func (t *tun) Write(b []byte) (int, error) {
-	var nn int
-	maximum := len(b)
-
-	for {
-		n, err := unix.Write(t.fd, b[nn:maximum])
-		if n > 0 {
-			nn += n
-		}
-		if nn == len(b) {
-			return nn, err
-		}
-
-		if err != nil {
-			return nn, err
-		}
-
-		if n == 0 {
-			return nn, io.ErrUnexpectedEOF
-		}
-	}
-}
-
 func (t *tun) deviceBytes() (o [16]byte) {
 	for i, c := range t.Device {
 		o[i] = byte(c)
@@ -707,18 +1190,40 @@ func (t *tun) updateRoutes(r netlink.RouteUpdate) {
 }
 
 func (t *tun) Close() error {
+	t.closeLock.Lock()
+	defer t.closeLock.Unlock()
+
 	if t.routeChan != nil {
 		close(t.routeChan)
+		t.routeChan = nil
 	}
 
-	if t.ReadWriteCloser != nil {
-		_ = t.ReadWriteCloser.Close()
-	}
+	// Signal all readers blocked in poll to wake up and exit
+	_ = t.tunFile.wakeForShutdown()
 
 	if t.ioctlFd > 0 {
-		_ = os.NewFile(t.ioctlFd, "ioctlFd").Close()
+		_ = unix.Close(int(t.ioctlFd))
 		t.ioctlFd = 0
 	}
 
-	return nil
+	for i := range t.readers {
+		if i == 0 {
+			continue //we want to close the zeroth reader last
+		}
+		err := t.readers[i].Close()
+		if err != nil {
+			t.l.WithField("reader", i).WithError(err).Error("error closing tun reader")
+		} else {
+			t.l.WithField("reader", i).Info("closed tun reader")
+		}
+	}
+
+	//this is t.readers[0] too
+	err := t.tunFile.Close()
+	if err != nil {
+		t.l.WithField("reader", 0).WithError(err).Error("error closing tun reader")
+	} else {
+		t.l.WithField("reader", 0).Info("closed tun reader")
+	}
+	return err
 }

overlay/tun_linux_offload.go

@@ -0,0 +1,331 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package overlay
+
+import (
+	"encoding/binary"
+	"fmt"
+
+	"golang.org/x/sys/unix"
+)
+
+// Size of the legacy struct virtio_net_hdr that the kernel prepends/expects on
+// a TUN opened with IFF_VNET_HDR (TUNSETVNETHDRSZ not set).
+const virtioNetHdrLen = 10
+
+// Maximum size we accept for a single read from a TUN with IFF_VNET_HDR. A
+// TSO superpacket can be up to 64KiB of payload plus a single L2/L3/L4 header
+// prefix plus the virtio header.
+const tunReadBufSize = 65535
+
+// Space for segmented output. Worst case is many small segments, each paying
+// an IP+TCP header. 128KiB comfortably covers the 64KiB payload ceiling.
+const tunSegBufSize = 131072
+
+// tunSegBufCap is the total size we allocate for the per-reader segment
+// buffer. It is sized as one worst-case TSO superpacket (tunSegBufSize) plus
+// the same again as drain headroom so a Read wake can accumulate
+// additional packets after an initial big read without overflowing.
+const tunSegBufCap = tunSegBufSize * 2
+
+// tunDrainCap caps how many packets a single Read will accumulate via
+// the post-wake drain loop. Sized to soak up a burst of small ACKs while
+// bounding how much work a single caller holds before handing off.
+const tunDrainCap = 64
+
+type virtioNetHdr struct {
+	Flags      uint8
+	GSOType    uint8
+	HdrLen     uint16
+	GSOSize    uint16
+	CsumStart  uint16
+	CsumOffset uint16
+}
+
+// decode reads a virtio_net_hdr in host byte order (TUN default; we never
+// call TUNSETVNETLE so the kernel matches our endianness).
+func (h *virtioNetHdr) decode(b []byte) {
+	h.Flags = b[0]
+	h.GSOType = b[1]
+	h.HdrLen = binary.NativeEndian.Uint16(b[2:4])
+	h.GSOSize = binary.NativeEndian.Uint16(b[4:6])
+	h.CsumStart = binary.NativeEndian.Uint16(b[6:8])
+	h.CsumOffset = binary.NativeEndian.Uint16(b[8:10])
+}
+
+// encode is the inverse of decode: writes the virtio_net_hdr fields into b
+// (must be at least virtioNetHdrLen bytes). Used to emit a TSO superpacket
+// on egress.
+func (h *virtioNetHdr) encode(b []byte) {
+	b[0] = h.Flags
+	b[1] = h.GSOType
+	binary.NativeEndian.PutUint16(b[2:4], h.HdrLen)
+	binary.NativeEndian.PutUint16(b[4:6], h.GSOSize)
+	binary.NativeEndian.PutUint16(b[6:8], h.CsumStart)
+	binary.NativeEndian.PutUint16(b[8:10], h.CsumOffset)
+}
+
+// segmentInto splits a TUN-side packet described by hdr into one or more
+// IP packets, each appended to *out as a slice of scratch. scratch must be
+// sized to hold every segment (including replicated headers).
+func segmentInto(pkt []byte, hdr virtioNetHdr, out *[][]byte, scratch []byte) error {
+	// When RSC_INFO is set the csum_start/csum_offset fields are repurposed to
+	// carry coalescing info rather than checksum offsets. A TUN writing via
+	// IFF_VNET_HDR should never emit this, but if it did we would silently
+	// miscompute the segment checksums — refuse the packet instead.
+	if hdr.Flags&unix.VIRTIO_NET_HDR_F_RSC_INFO != 0 {
+		return fmt.Errorf("virtio RSC_INFO flag not supported on TUN reads")
+	}
+
+	switch hdr.GSOType {
+	case unix.VIRTIO_NET_HDR_GSO_NONE:
+		if len(pkt) > len(scratch) {
+			return fmt.Errorf("packet larger than segment buffer: %d > %d", len(pkt), len(scratch))
+		}
+		copy(scratch, pkt)
+		seg := scratch[:len(pkt)]
+		if hdr.Flags&unix.VIRTIO_NET_HDR_F_NEEDS_CSUM != 0 {
+			if err := finishChecksum(seg, hdr); err != nil {
+				return err
+			}
+		}
+		*out = append(*out, seg)
+		return nil
+
+	case unix.VIRTIO_NET_HDR_GSO_TCPV4, unix.VIRTIO_NET_HDR_GSO_TCPV6:
+		return segmentTCP(pkt, hdr, out, scratch)
+
+	default:
+		return fmt.Errorf("unsupported virtio gso type: %d", hdr.GSOType)
+	}
+}
+
+// finishChecksum computes the L4 checksum for a non-GSO packet that the kernel
+// handed us with NEEDS_CSUM set. csum_start / csum_offset point at the 16-bit
+// checksum field; we zero it, fold a full sum (the field was pre-loaded with
+// the pseudo-header partial sum by the kernel), and store the result.
+func finishChecksum(seg []byte, hdr virtioNetHdr) error {
+	cs := int(hdr.CsumStart)
+	co := int(hdr.CsumOffset)
+	if cs+co+2 > len(seg) {
+		return fmt.Errorf("csum offsets out of range: start=%d offset=%d len=%d", cs, co, len(seg))
+	}
+	// The kernel stores a partial pseudo-header sum at [cs+co:]; sum over the
+	// L4 region starting at cs, folding the prior partial in as the seed.
+	partial := uint32(binary.BigEndian.Uint16(seg[cs+co : cs+co+2]))
+	seg[cs+co] = 0
+	seg[cs+co+1] = 0
+	sum := checksumBytes(seg[cs:], partial)
+	binary.BigEndian.PutUint16(seg[cs+co:cs+co+2], checksumFold(sum))
+	return nil
+}
+
+// segmentTCP software-segments a TSO superpacket into one IP packet per MSS
+// chunk. The caller guarantees hdr.GSOType is TCPV4 or TCPV6.
+//
+// Hot-path shape: the per-segment loop only sums the payload chunk. The TCP
+// header, the IPv4 header, and the pseudo-header src/dst/proto contributions
+// are each summed once up front — every segment reuses those three pre-folded
+// uint32 values and combines them with small per-segment deltas (seq, flags,
+// tcpLen, ip_id, total_len) that are cheap to fold in.
+func segmentTCP(pkt []byte, hdr virtioNetHdr, out *[][]byte, scratch []byte) error {
+	if hdr.GSOSize == 0 {
+		return fmt.Errorf("gso_size is zero")
+	}
+	if int(hdr.HdrLen) > len(pkt) || hdr.HdrLen == 0 {
+		return fmt.Errorf("hdr_len %d out of range (pkt %d)", hdr.HdrLen, len(pkt))
+	}
+	if hdr.CsumStart == 0 || hdr.CsumStart >= hdr.HdrLen {
+		return fmt.Errorf("csum_start %d out of range (hdr_len %d)", hdr.CsumStart, hdr.HdrLen)
+	}
+
+	isV4 := hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_TCPV4
+	headerLen := int(hdr.HdrLen)
+	csumStart := int(hdr.CsumStart)
+
+	if isV4 && csumStart < 20 {
+		return fmt.Errorf("csum_start %d too small for IPv4", csumStart)
+	}
+	if !isV4 && csumStart < 40 {
+		return fmt.Errorf("csum_start %d too small for IPv6", csumStart)
+	}
+	tcpHdrLen := headerLen - csumStart
+	if tcpHdrLen < 20 {
+		return fmt.Errorf("tcp header region too small: %d", tcpHdrLen)
+	}
+
+	payload := pkt[headerLen:]
+	payLen := len(payload)
+	gso := int(hdr.GSOSize)
+	numSeg := (payLen + gso - 1) / gso
+	if numSeg == 0 {
+		numSeg = 1
+	}
+
+	need := numSeg*headerLen + payLen
+	if need > len(scratch) {
+		return fmt.Errorf("scratch too small for %d segments: need %d have %d", numSeg, need, len(scratch))
+	}
+
+	origSeq := binary.BigEndian.Uint32(pkt[csumStart+4 : csumStart+8])
+	origFlags := pkt[csumStart+13]
+	const tcpFinPsh = 0x09 // FIN(0x01) | PSH(0x08)
+
+	// Precompute the TCP header sum with seq/flags/csum zeroed. The max TCP
+	// header is 60 bytes; copy onto the stack, zero the per-segment-varying
+	// fields, sum once.
+	var tmp [60]byte
+	copy(tmp[:tcpHdrLen], pkt[csumStart:headerLen])
+	tmp[4], tmp[5], tmp[6], tmp[7] = 0, 0, 0, 0 // seq
+	tmp[13] = 0                                 // flags
+	tmp[16], tmp[17] = 0, 0                     // csum
+	baseTcpHdrSum := checksumBytes(tmp[:tcpHdrLen], 0)
+
+	// Pseudo-header src+dst+proto contribution (tcpLen varies per segment).
+	var baseProtoSum uint32
+	if isV4 {
+		baseProtoSum = checksumBytes(pkt[12:16], 0)
+		baseProtoSum = checksumBytes(pkt[16:20], baseProtoSum)
+	} else {
+		baseProtoSum = checksumBytes(pkt[8:24], 0)
+		baseProtoSum = checksumBytes(pkt[24:40], baseProtoSum)
+	}
+	baseProtoSum += uint32(unix.IPPROTO_TCP)
+
+	// Precompute IPv4 header sum with total_len/id/csum zeroed.
+	var origIPID uint16
+	var ihl int
+	var baseIPHdrSum uint32
+	if isV4 {
+		origIPID = binary.BigEndian.Uint16(pkt[4:6])
+		ihl = int(pkt[0]&0x0f) * 4
+		if ihl < 20 || ihl > csumStart {
+			return fmt.Errorf("bad IPv4 IHL: %d", ihl)
+		}
+		var ipTmp [60]byte
+		copy(ipTmp[:ihl], pkt[:ihl])
+		ipTmp[2], ipTmp[3] = 0, 0   // total_len
+		ipTmp[4], ipTmp[5] = 0, 0   // id
+		ipTmp[10], ipTmp[11] = 0, 0 // checksum
+		baseIPHdrSum = checksumBytes(ipTmp[:ihl], 0)
+	}
+
+	off := 0
+	for i := 0; i < numSeg; i++ {
+		segStart := i * gso
+		segEnd := segStart + gso
+		if segEnd > payLen {
+			segEnd = payLen
+		}
+		segPayLen := segEnd - segStart
+
+		copy(scratch[off:], pkt[:headerLen])
+		copy(scratch[off+headerLen:], payload[segStart:segEnd])
+		seg := scratch[off : off+headerLen+segPayLen]
+		off += headerLen + segPayLen
+
+		segSeq := origSeq + uint32(segStart)
+		segFlags := origFlags
+		if i != numSeg-1 {
+			segFlags = origFlags &^ tcpFinPsh
+		}
+		totalLen := headerLen + segPayLen
+
+		// Patch IP header and write the v4 header checksum from the precomputed base.
+		if isV4 {
+			segID := origIPID + uint16(i)
+			binary.BigEndian.PutUint16(seg[2:4], uint16(totalLen))
+			binary.BigEndian.PutUint16(seg[4:6], segID)
+			ipSum := baseIPHdrSum + uint32(totalLen) + uint32(segID)
+			binary.BigEndian.PutUint16(seg[10:12], checksumFold(ipSum))
+		} else {
+			// IPv6 payload length excludes the 40-byte fixed header but
+			// includes any extension headers between [40:csumStart].
+			binary.BigEndian.PutUint16(seg[4:6], uint16(headerLen-40+segPayLen))
+		}
+
+		// Patch TCP header.
+		binary.BigEndian.PutUint32(seg[csumStart+4:csumStart+8], segSeq)
+		seg[csumStart+13] = segFlags
+		// (csum is written below; its prior contents in `seg` don't affect the
+		// computation since we never sum over the segment's own header.)
+
+		tcpLen := tcpHdrLen + segPayLen
+		paySum := checksumBytes(payload[segStart:segEnd], 0)
+
+		// Combine pre-folded uint32s into a wider accumulator, then fold. Using
+		// uint64 guards against overflow when segSeq's high bits set.
+		wide := uint64(baseTcpHdrSum) + uint64(paySum) + uint64(baseProtoSum)
+		wide += uint64(segSeq) + uint64(segFlags) + uint64(tcpLen)
+		wide = (wide & 0xffffffff) + (wide >> 32)
+		wide = (wide & 0xffffffff) + (wide >> 32)
+		binary.BigEndian.PutUint16(seg[csumStart+16:csumStart+18], checksumFold(uint32(wide)))
+
+		*out = append(*out, seg)
+	}
+
+	return nil
+}
+
+// checksumBytes returns the Internet-checksum partial sum of b, seeded with
+// initial. Result is a 32-bit accumulator; the caller folds to 16.
+//
+// Each 4-byte load is added directly into a 64-bit accumulator. Two parallel
+// accumulators break the serial dependency through `sum` and let the CPU
+// overlap independent adds. The final fold from 64 → 32 → 16 handles the
+// carries that accumulated across the 32-bit lane boundary.
+func checksumBytes(b []byte, initial uint32) uint32 {
+	s0 := uint64(initial)
+	var s1 uint64
+	for len(b) >= 32 {
+		s0 += uint64(binary.BigEndian.Uint32(b[0:4]))
+		s1 += uint64(binary.BigEndian.Uint32(b[4:8]))
+		s0 += uint64(binary.BigEndian.Uint32(b[8:12]))
+		s1 += uint64(binary.BigEndian.Uint32(b[12:16]))
+		s0 += uint64(binary.BigEndian.Uint32(b[16:20]))
+		s1 += uint64(binary.BigEndian.Uint32(b[20:24]))
+		s0 += uint64(binary.BigEndian.Uint32(b[24:28]))
+		s1 += uint64(binary.BigEndian.Uint32(b[28:32]))
+		b = b[32:]
+	}
+	sum := s0 + s1
+	for len(b) >= 4 {
+		sum += uint64(binary.BigEndian.Uint32(b[:4]))
+		b = b[4:]
+	}
+	if len(b) >= 2 {
+		sum += uint64(binary.BigEndian.Uint16(b[:2]))
+		b = b[2:]
+	}
+	if len(b) == 1 {
+		sum += uint64(b[0]) << 8
+	}
+	sum = (sum & 0xffffffff) + (sum >> 32)
+	sum = (sum & 0xffffffff) + (sum >> 32)
+	return uint32(sum)
+}
+
+func checksumFold(sum uint32) uint16 {
+	for sum>>16 != 0 {
+		sum = (sum & 0xffff) + (sum >> 16)
+	}
+	return ^uint16(sum)
+}
+
+func pseudoHeaderIPv4(src, dst []byte, proto byte, tcpLen int) uint32 {
+	sum := checksumBytes(src, 0)
+	sum = checksumBytes(dst, sum)
+	sum += uint32(proto)
+	sum += uint32(tcpLen)
+	return sum
+}
+
+func pseudoHeaderIPv6(src, dst []byte, proto byte, tcpLen int) uint32 {
+	sum := checksumBytes(src, 0)
+	sum = checksumBytes(dst, sum)
+	sum += uint32(tcpLen >> 16)
+	sum += uint32(tcpLen & 0xffff)
+	sum += uint32(proto)
+	return sum
+}

overlay/tun_linux_offload_test.go

@@ -0,0 +1,333 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package overlay
+
+import (
+	"encoding/binary"
+	"os"
+	"testing"
+
+	"golang.org/x/sys/unix"
+)
+
+// verifyChecksum confirms that the one's-complement sum across `b`, optionally
+// seeded with a pseudo-header sum, folds to all-ones (valid).
+func verifyChecksum(b []byte, pseudo uint32) bool {
+	sum := checksumBytes(b, pseudo)
+	for sum>>16 != 0 {
+		sum = (sum & 0xffff) + (sum >> 16)
+	}
+	return uint16(sum) == 0xffff
+}
+
+// buildTSOv4 builds a synthetic IPv4/TCP TSO superpacket with a payload of
+// `payLen` bytes split at `mss`.
+func buildTSOv4(t *testing.T, payLen, mss int) ([]byte, virtioNetHdr) {
+	t.Helper()
+	const ipLen = 20
+	const tcpLen = 20
+	pkt := make([]byte, ipLen+tcpLen+payLen)
+
+	// IPv4 header
+	pkt[0] = 0x45 // version 4, IHL 5
+	// total length is meaningless for TSO but set it anyway
+	binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+payLen))
+	binary.BigEndian.PutUint16(pkt[4:6], 0x4242) // original ID
+	pkt[8] = 64                                  // TTL
+	pkt[9] = unix.IPPROTO_TCP
+	copy(pkt[12:16], []byte{10, 0, 0, 1}) // src
+	copy(pkt[16:20], []byte{10, 0, 0, 2}) // dst
+
+	// TCP header
+	binary.BigEndian.PutUint16(pkt[20:22], 12345) // sport
+	binary.BigEndian.PutUint16(pkt[22:24], 80)    // dport
+	binary.BigEndian.PutUint32(pkt[24:28], 10000) // seq
+	binary.BigEndian.PutUint32(pkt[28:32], 20000) // ack
+	pkt[32] = 0x50                                // data offset 5 words
+	pkt[33] = 0x18                                // ACK | PSH
+	binary.BigEndian.PutUint16(pkt[34:36], 65535) // window
+
+	// payload
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+tcpLen+i] = byte(i & 0xff)
+	}
+
+	return pkt, virtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV4,
+		HdrLen:     uint16(ipLen + tcpLen),
+		GSOSize:    uint16(mss),
+		CsumStart:  uint16(ipLen),
+		CsumOffset: 16,
+	}
+}
+
+func TestSegmentTCPv4(t *testing.T) {
+	const mss = 100
+	const numSeg = 3
+	pkt, hdr := buildTSOv4(t, mss*numSeg, mss)
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != numSeg {
+		t.Fatalf("expected %d segments, got %d", numSeg, len(out))
+	}
+
+	for i, seg := range out {
+		if len(seg) != 40+mss {
+			t.Errorf("seg %d: unexpected len %d", i, len(seg))
+		}
+		totalLen := binary.BigEndian.Uint16(seg[2:4])
+		if totalLen != uint16(40+mss) {
+			t.Errorf("seg %d: total_len=%d want %d", i, totalLen, 40+mss)
+		}
+		id := binary.BigEndian.Uint16(seg[4:6])
+		if id != 0x4242+uint16(i) {
+			t.Errorf("seg %d: ip id=%#x want %#x", i, id, 0x4242+uint16(i))
+		}
+		seq := binary.BigEndian.Uint32(seg[24:28])
+		wantSeq := uint32(10000 + i*mss)
+		if seq != wantSeq {
+			t.Errorf("seg %d: seq=%d want %d", i, seq, wantSeq)
+		}
+		flags := seg[33]
+		wantFlags := byte(0x10) // ACK only, PSH cleared
+		if i == numSeg-1 {
+			wantFlags = 0x18 // ACK | PSH preserved on last
+		}
+		if flags != wantFlags {
+			t.Errorf("seg %d: flags=%#x want %#x", i, flags, wantFlags)
+		}
+		// IPv4 header checksum must verify against itself.
+		if !verifyChecksum(seg[:20], 0) {
+			t.Errorf("seg %d: bad IPv4 header checksum", i)
+		}
+		// TCP checksum must verify against the pseudo-header.
+		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+mss)
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentTCPv4OddTail(t *testing.T) {
+	// Payload of 250 bytes with MSS 100 → segments of 100, 100, 50.
+	pkt, hdr := buildTSOv4(t, 250, 100)
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != 3 {
+		t.Fatalf("want 3 segments, got %d", len(out))
+	}
+	wantPayLens := []int{100, 100, 50}
+	for i, seg := range out {
+		if len(seg)-40 != wantPayLens[i] {
+			t.Errorf("seg %d: pay len %d want %d", i, len(seg)-40, wantPayLens[i])
+		}
+		if !verifyChecksum(seg[:20], 0) {
+			t.Errorf("seg %d: bad IPv4 header checksum", i)
+		}
+		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+wantPayLens[i])
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentTCPv6(t *testing.T) {
+	const ipLen = 40
+	const tcpLen = 20
+	const mss = 120
+	const numSeg = 2
+	payLen := mss * numSeg
+	pkt := make([]byte, ipLen+tcpLen+payLen)
+
+	// IPv6 header
+	pkt[0] = 0x60 // version 6
+	binary.BigEndian.PutUint16(pkt[4:6], uint16(tcpLen+payLen))
+	pkt[6] = unix.IPPROTO_TCP
+	pkt[7] = 64
+	// src/dst fe80::1 / fe80::2
+	pkt[8] = 0xfe
+	pkt[9] = 0x80
+	pkt[23] = 1
+	pkt[24] = 0xfe
+	pkt[25] = 0x80
+	pkt[39] = 2
+
+	// TCP header
+	binary.BigEndian.PutUint16(pkt[40:42], 12345)
+	binary.BigEndian.PutUint16(pkt[42:44], 80)
+	binary.BigEndian.PutUint32(pkt[44:48], 7)
+	binary.BigEndian.PutUint32(pkt[48:52], 99)
+	pkt[52] = 0x50
+	pkt[53] = 0x19 // FIN | ACK | PSH — exercise FIN clearing too
+	binary.BigEndian.PutUint16(pkt[54:56], 65535)
+
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+tcpLen+i] = byte(i)
+	}
+
+	hdr := virtioNetHdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV6,
+		HdrLen:     uint16(ipLen + tcpLen),
+		GSOSize:    uint16(mss),
+		CsumStart:  uint16(ipLen),
+		CsumOffset: 16,
+	}
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentTCP: %v", err)
+	}
+	if len(out) != numSeg {
+		t.Fatalf("want %d segments, got %d", numSeg, len(out))
+	}
+
+	for i, seg := range out {
+		if len(seg) != ipLen+tcpLen+mss {
+			t.Errorf("seg %d: len %d want %d", i, len(seg), ipLen+tcpLen+mss)
+		}
+		pl := binary.BigEndian.Uint16(seg[4:6])
+		if pl != uint16(tcpLen+mss) {
+			t.Errorf("seg %d: payload_length=%d want %d", i, pl, tcpLen+mss)
+		}
+		seq := binary.BigEndian.Uint32(seg[44:48])
+		if seq != uint32(7+i*mss) {
+			t.Errorf("seg %d: seq=%d want %d", i, seq, 7+i*mss)
+		}
+		flags := seg[53]
+		// Original flags = 0x19 (FIN|ACK|PSH). FIN(0x01)+PSH(0x08) should be
+		// cleared on all but the last; ACK(0x10) always preserved.
+		wantFlags := byte(0x10)
+		if i == numSeg-1 {
+			wantFlags = 0x19
+		}
+		if flags != wantFlags {
+			t.Errorf("seg %d: flags=%#x want %#x", i, flags, wantFlags)
+		}
+		psum := pseudoHeaderIPv6(seg[8:24], seg[24:40], unix.IPPROTO_TCP, tcpLen+mss)
+		if !verifyChecksum(seg[ipLen:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+func TestSegmentGSONonePassesThrough(t *testing.T) {
+	pkt, hdr := buildTSOv4(t, 100, 100)
+	hdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+	hdr.Flags = 0 // no NEEDS_CSUM, leave packet untouched
+
+	scratch := make([]byte, tunSegBufSize)
+	var out [][]byte
+	if err := segmentInto(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentInto: %v", err)
+	}
+	if len(out) != 1 {
+		t.Fatalf("want 1 segment, got %d", len(out))
+	}
+	if len(out[0]) != len(pkt) {
+		t.Fatalf("unexpected length: %d vs %d", len(out[0]), len(pkt))
+	}
+}
+
+func TestSegmentRejectsUDP(t *testing.T) {
+	hdr := virtioNetHdr{GSOType: unix.VIRTIO_NET_HDR_GSO_UDP}
+	var out [][]byte
+	if err := segmentInto(nil, hdr, &out, nil); err == nil {
+		t.Fatalf("expected rejection for UDP GSO")
+	}
+}
+
+func BenchmarkSegmentTCPv4(b *testing.B) {
+	sizes := []struct {
+		name   string
+		payLen int
+		mss    int
+	}{
+		{"64KiB_MSS1460", 65000, 1460},
+		{"16KiB_MSS1460", 16384, 1460},
+		{"4KiB_MSS1460", 4096, 1460},
+	}
+	for _, sz := range sizes {
+		b.Run(sz.name, func(b *testing.B) {
+			const ipLen = 20
+			const tcpLen = 20
+			pkt := make([]byte, ipLen+tcpLen+sz.payLen)
+			pkt[0] = 0x45
+			binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+sz.payLen))
+			binary.BigEndian.PutUint16(pkt[4:6], 0x4242)
+			pkt[8] = 64
+			pkt[9] = unix.IPPROTO_TCP
+			copy(pkt[12:16], []byte{10, 0, 0, 1})
+			copy(pkt[16:20], []byte{10, 0, 0, 2})
+			binary.BigEndian.PutUint16(pkt[20:22], 12345)
+			binary.BigEndian.PutUint16(pkt[22:24], 80)
+			binary.BigEndian.PutUint32(pkt[24:28], 10000)
+			binary.BigEndian.PutUint32(pkt[28:32], 20000)
+			pkt[32] = 0x50
+			pkt[33] = 0x18
+			binary.BigEndian.PutUint16(pkt[34:36], 65535)
+			for i := 0; i < sz.payLen; i++ {
+				pkt[ipLen+tcpLen+i] = byte(i)
+			}
+			hdr := virtioNetHdr{
+				Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+				GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV4,
+				HdrLen:     uint16(ipLen + tcpLen),
+				GSOSize:    uint16(sz.mss),
+				CsumStart:  uint16(ipLen),
+				CsumOffset: 16,
+			}
+
+			scratch := make([]byte, tunSegBufSize)
+			out := make([][]byte, 0, 64)
+
+			b.SetBytes(int64(len(pkt)))
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				out = out[:0]
+				if err := segmentTCP(pkt, hdr, &out, scratch); err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}
+
+// TestTunFileWriteVnetHdrNoAlloc verifies the IFF_VNET_HDR fast-path write is
+// allocation-free. We write to /dev/null so every call succeeds synchronously.
+func TestTunFileWriteVnetHdrNoAlloc(t *testing.T) {
+	fd, err := unix.Open("/dev/null", os.O_WRONLY, 0)
+	if err != nil {
+		t.Fatalf("open /dev/null: %v", err)
+	}
+	t.Cleanup(func() { _ = unix.Close(fd) })
+
+	tf := &tunFile{fd: fd, vnetHdr: true}
+	tf.writeIovs[0].Base = &validVnetHdr[0]
+	tf.writeIovs[0].SetLen(virtioNetHdrLen)
+
+	payload := make([]byte, 1400)
+	// Warm up (first call may trigger one-time internal allocations elsewhere).
+	if _, err := tf.Write(payload); err != nil {
+		t.Fatalf("Write: %v", err)
+	}
+
+	allocs := testing.AllocsPerRun(1000, func() {
+		if _, err := tf.Write(payload); err != nil {
+			t.Fatalf("Write: %v", err)
+		}
+	})
+	if allocs != 0 {
+		t.Fatalf("Write allocated %.1f times per call, want 0", allocs)
+	}
+}

overlay/tun_netbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"net/netip"
 	"os"
 	"regexp"
@@ -66,6 +65,25 @@ type tun struct {
 	l           *logrus.Logger
 	f           *os.File
 	fd          int
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
 }
 
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -141,7 +159,7 @@ func (t *tun) Close() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	rc, err := t.f.SyscallConn()
 	if err != nil {
 		return 0, fmt.Errorf("failed to get syscall conn for tun: %w", err)
@@ -394,7 +412,7 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for netbsd")
 }
 

overlay/tun_openbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"net/netip"
 	"os"
 	"regexp"
@@ -59,6 +58,25 @@ type tun struct {
 	fd          int
 	// cache out buffer since we need to prepend 4 bytes for tun metadata
 	out []byte
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *tun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.readOne(t.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *tun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
 }
 
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -124,7 +142,7 @@ func (t *tun) Close() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) readOne(to []byte) (int, error) {
 	buf := make([]byte, len(to)+4)
 
 	n, err := t.f.Read(buf)
@@ -314,7 +332,7 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() (Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for openbsd")
 }
 

overlay/tun_tester.go

@@ -26,6 +26,17 @@ type TestTun struct {
 	closed    atomic.Bool
 	rxPackets chan []byte // Packets to receive into nebula
 	TxPackets chan []byte // Packets transmitted outside by nebula
+
+	batchRet [1][]byte
+}
+
+func (t *TestTun) Read() ([][]byte, error) {
+	p, ok := <-t.rxPackets
+	if !ok {
+		return nil, os.ErrClosed
+	}
+	t.batchRet[0] = p
+	return t.batchRet[:], nil
 }
 
 func newTun(c *config.C, l *logrus.Logger, vpnNetworks []netip.Prefix, _ bool) (*TestTun, error) {
@@ -115,6 +126,10 @@ func (t *TestTun) Write(b []byte) (n int, err error) {
 	return len(b), nil
 }
 
+func (t *TestTun) WriteReject(b []byte) (int, error) {
+	return t.Write(b)
+}
+
 func (t *TestTun) Close() error {
 	if t.closed.CompareAndSwap(false, true) {
 		close(t.rxPackets)
@@ -123,19 +138,10 @@ func (t *TestTun) Close() error {
 	return nil
 }
 
-func (t *TestTun) Read(b []byte) (int, error) {
-	p, ok := <-t.rxPackets
-	if !ok {
-		return 0, os.ErrClosed
-	}
-	copy(b, p)
-	return len(p), nil
-}
-
 func (t *TestTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *TestTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *TestTun) NewMultiQueueReader() (Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented")
 }

overlay/tun_windows.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"crypto"
 	"fmt"
-	"io"
 	"net/netip"
 	"os"
 	"path/filepath"
@@ -36,6 +35,25 @@ type winTun struct {
 	l           *logrus.Logger
 
 	tun *wintun.NativeTun
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (t *winTun) Read() ([][]byte, error) {
+	if t.readBuf == nil {
+		t.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := t.tun.Read(t.readBuf, 0)
+	if err != nil {
+		return nil, err
+	}
+	t.batchRet[0] = t.readBuf[:n]
+	return t.batchRet[:], nil
+}
+
+func (t *winTun) WriteReject(p []byte) (int, error) {
+	return t.Write(p)
 }
 
 func newTunFromFd(_ *config.C, _ *logrus.Logger, _ int, _ []netip.Prefix) (Device, error) {
@@ -229,10 +247,6 @@ func (t *winTun) Name() string {
 	return t.Device
 }
 
-func (t *winTun) Read(b []byte) (int, error) {
-	return t.tun.Read(b, 0)
-}
-
 func (t *winTun) Write(b []byte) (int, error) {
 	return t.tun.Write(b, 0)
 }
@@ -241,7 +255,7 @@ func (t *winTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *winTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *winTun) NewMultiQueueReader() (Queue, error) {
 	return nil, fmt.Errorf("TODO: multiqueue not implemented for windows")
 }
 

overlay/user.go

@@ -34,6 +34,21 @@ type UserDevice struct {
 
 	inboundReader *io.PipeReader
 	inboundWriter *io.PipeWriter
+
+	readBuf  []byte
+	batchRet [1][]byte
+}
+
+func (d *UserDevice) Read() ([][]byte, error) {
+	if d.readBuf == nil {
+		d.readBuf = make([]byte, defaultBatchBufSize)
+	}
+	n, err := d.outboundReader.Read(d.readBuf)
+	if err != nil {
+		return nil, err
+	}
+	d.batchRet[0] = d.readBuf[:n]
+	return d.batchRet[:], nil
 }
 
 func (d *UserDevice) Activate() error {
@@ -50,7 +65,7 @@ func (d *UserDevice) SupportsMultiqueue() bool {
 	return true
 }
 
-func (d *UserDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (d *UserDevice) NewMultiQueueReader() (Queue, error) {
 	return d, nil
 }
 
@@ -58,12 +73,12 @@ func (d *UserDevice) Pipe() (*io.PipeReader, *io.PipeWriter) {
 	return d.inboundReader, d.outboundWriter
 }
 
-func (d *UserDevice) Read(p []byte) (n int, err error) {
-	return d.outboundReader.Read(p)
-}
 func (d *UserDevice) Write(p []byte) (n int, err error) {
 	return d.inboundWriter.Write(p)
 }
+func (d *UserDevice) WriteReject(p []byte) (n int, err error) {
+	return d.Write(p)
+}
 func (d *UserDevice) Close() error {
 	d.inboundWriter.Close()
 	d.outboundWriter.Close()

pki.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"io"
 	"net"
 	"net/netip"
 	"os"
@@ -487,25 +488,25 @@ func loadCertificate(b []byte) (cert.Certificate, []byte, error) {
 }
 
 func loadCAPoolFromConfig(l *logrus.Logger, c *config.C) (*cert.CAPool, error) {
-	var rawCA []byte
-	var err error
-
 	caPathOrPEM := c.GetString("pki.ca", "")
 	if caPathOrPEM == "" {
 		return nil, errors.New("no pki.ca path or PEM data provided")
 	}
 
-	if strings.Contains(caPathOrPEM, "-----BEGIN") {
-		rawCA = []byte(caPathOrPEM)
+	var caReader io.ReadCloser
+	var err error
 
+	if strings.Contains(caPathOrPEM, "-----BEGIN") {
+		caReader = io.NopCloser(strings.NewReader(caPathOrPEM))
 	} else {
-		rawCA, err = os.ReadFile(caPathOrPEM)
+		caReader, err = os.Open(caPathOrPEM)
 		if err != nil {
 			return nil, fmt.Errorf("unable to read pki.ca file %s: %s", caPathOrPEM, err)
 		}
 	}
+	defer caReader.Close()
 
-	caPool, err := cert.NewCAPoolFromPEM(rawCA)
+	caPool, err := cert.NewCAPoolFromPEMReader(caReader)
 	if errors.Is(err, cert.ErrExpired) {
 		var expired int
 		for _, crt := range caPool.CAs {

pki_hup_benchmark_test.go

@@ -0,0 +1,121 @@
+package nebula
+
+import (
+	"bytes"
+	"fmt"
+	"net/netip"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula/cert"
+	cert_test "github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/test"
+	"github.com/stretchr/testify/require"
+)
+
+func BenchmarkReloadConfigWithCAs(b *testing.B) {
+	prevProcs := runtime.GOMAXPROCS(1)
+	b.Cleanup(func() { runtime.GOMAXPROCS(prevProcs) })
+
+	for _, size := range []int{100, 250, 500, 1000, 5000} {
+		b.Run(fmt.Sprintf("%dCAs", size), func(b *testing.B) {
+			l := test.NewLogger()
+			dir := b.TempDir()
+
+			ca, caKey, caBundle := buildCABundle(b, size)
+			caPath, certPath, keyPath := writePKIFiles(b, dir, ca, caKey, caBundle)
+
+			configBody := fmt.Sprintf(`pki:
+  ca: %s
+  cert: %s
+  key: %s
+`, caPath, certPath, keyPath)
+
+			configPath := filepath.Join(dir, "config.yml")
+			require.NoError(b, os.WriteFile(configPath, []byte(configBody), 0o600))
+
+			c := config.NewC(l)
+			require.NoError(b, c.Load(dir))
+
+			_, err := NewPKIFromConfig(l, c)
+			require.NoError(b, err)
+
+			b.ReportAllocs()
+			b.ResetTimer()
+
+			for b.Loop() {
+				c.ReloadConfig()
+			}
+		})
+	}
+}
+
+func buildCABundle(b *testing.B, count int) (cert.Certificate, []byte, []byte) {
+	b.Helper()
+	require.GreaterOrEqual(b, count, 1)
+
+	before := time.Now().Add(-24 * time.Hour)
+	after := time.Now().Add(24 * time.Hour)
+
+	ca, _, caKey, pem := cert_test.NewTestCaCert(
+		cert.Version2,
+		cert.Curve_CURVE25519,
+		before,
+		after,
+		nil,
+		nil,
+		nil,
+	)
+
+	buf := bytes.NewBuffer(pem)
+	buf.Write([]byte("\n# a comment!\n"))
+
+	for i := 1; i < count; i++ {
+		_, _, _, extraPEM := cert_test.NewTestCaCert(
+			cert.Version2,
+			cert.Curve_CURVE25519,
+			time.Now(),
+			time.Now().Add(time.Hour),
+			nil,
+			nil,
+			nil,
+		)
+		buf.Write([]byte("\n# a comment!\n"))
+		buf.Write(extraPEM)
+	}
+
+	return ca, caKey, buf.Bytes()
+}
+
+func writePKIFiles(b *testing.B, dir string, ca cert.Certificate, caKey []byte, caBundle []byte) (string, string, string) {
+	b.Helper()
+
+	networks := []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")}
+
+	_, _, keyPEM, certPEM := cert_test.NewTestCert(
+		cert.Version2,
+		cert.Curve_CURVE25519,
+		ca,
+		caKey,
+		"reload-benchmark",
+		time.Now(),
+		time.Now().Add(time.Hour),
+		networks,
+		nil,
+		nil,
+	)
+
+	caPath := filepath.Join(dir, "ca.pem")
+	certPath := filepath.Join(dir, "cert.pem")
+	keyPath := filepath.Join(dir, "key.pem")
+
+	require.NoError(b, os.WriteFile(caPath, caBundle, 0o600))
+	require.NoError(b, os.WriteFile(certPath, certPEM, 0o600))
+	require.NoError(b, os.WriteFile(keyPath, keyPEM, 0o600))
+
+	return caPath, certPath, keyPath
+}

relay_manager.go

@@ -55,7 +55,7 @@ func (rm *relayManager) setAmRelay(v bool) {
 func AddRelay(l *logrus.Logger, relayHostInfo *HostInfo, hm *HostMap, vpnIp netip.Addr, remoteIdx *uint32, relayType int, state int) (uint32, error) {
 	hm.Lock()
 	defer hm.Unlock()
-	for i := 0; i < 32; i++ {
+	for range 32 {
 		index, err := generateIndex(l)
 		if err != nil {
 			return 0, err

remote_list.go

@@ -404,12 +404,7 @@ func (r *RemoteList) Rebuild(preferredRanges []netip.Prefix) {
 
 // unlockedIsBad assumes you have the write lock and checks if the remote matches any entry in the blocked address list
 func (r *RemoteList) unlockedIsBad(remote netip.AddrPort) bool {
-	for _, v := range r.badRemotes {
-		if v == remote {
-			return true
-		}
-	}
-	return false
+	return slices.Contains(r.badRemotes, remote)
 }
 
 // unlockedSetLearnedV4 assumes you have the write lock and sets the current learned address for this owner and marks the

service/service.go

@@ -44,7 +44,10 @@ type Service struct {
 }
 
 func New(control *nebula.Control) (*Service, error) {
-	control.Start()
+	wait, err := control.Start()
+	if err != nil {
+		return nil, err
+	}
 
 	ctx := control.Context()
 	eg, ctx := errgroup.WithContext(ctx)
@@ -141,6 +144,12 @@ func New(control *nebula.Control) (*Service, error) {
 		}
 	})
 
+	// Add the nebula wait function to the group so a fatal reader error
+	// propagates out through errgroup.Wait().
+	eg.Go(func() error {
+		return wait()
+	})
+
 	return &s, nil
 }
 

ssh.go

@@ -6,9 +6,11 @@ import (
 	"errors"
 	"flag"
 	"fmt"
+	"maps"
 	"net"
 	"net/netip"
 	"os"
+	"path/filepath"
 	"reflect"
 	"runtime"
 	"runtime/pprof"
@@ -187,6 +189,12 @@ func configSSH(l *logrus.Logger, ssh *sshd.SSHServer, c *config.C) (func(), erro
 }
 
 func attachCommands(l *logrus.Logger, c *config.C, ssh *sshd.SSHServer, f *Interface) {
+	// sandboxDir defaults to a dir in temp. The intention is that end user will
+	// create this dir as needed. Overriding this config value to "" allows
+	// writing to anywhere in the system.
+	defaultDir := filepath.Join(os.TempDir(), "nebula-debug")
+	sandboxDir := c.GetString("sshd.sandbox_dir", defaultDir)
+
 	ssh.RegisterCommand(&sshd.Command{
 		Name:             "list-hostmap",
 		ShortDescription: "List all known previously connected hosts",
@@ -245,7 +253,9 @@ func attachCommands(l *logrus.Logger, c *config.C, ssh *sshd.SSHServer, f *Inter
 	ssh.RegisterCommand(&sshd.Command{
 		Name:             "start-cpu-profile",
 		ShortDescription: "Starts a cpu profile and write output to the provided file, ex: `cpu-profile.pb.gz`",
-		Callback:         sshStartCpuProfile,
+		Callback: func(fs any, a []string, w sshd.StringWriter) error {
+			return sshStartCpuProfile(sandboxDir, fs, a, w)
+		},
 	})
 
 	ssh.RegisterCommand(&sshd.Command{
@@ -260,7 +270,9 @@ func attachCommands(l *logrus.Logger, c *config.C, ssh *sshd.SSHServer, f *Inter
 	ssh.RegisterCommand(&sshd.Command{
 		Name:             "save-heap-profile",
 		ShortDescription: "Saves a heap profile to the provided path, ex: `heap-profile.pb.gz`",
-		Callback:         sshGetHeapProfile,
+		Callback: func(fs any, a []string, w sshd.StringWriter) error {
+			return sshGetHeapProfile(sandboxDir, fs, a, w)
+		},
 	})
 
 	ssh.RegisterCommand(&sshd.Command{
@@ -272,7 +284,9 @@ func attachCommands(l *logrus.Logger, c *config.C, ssh *sshd.SSHServer, f *Inter
 	ssh.RegisterCommand(&sshd.Command{
 		Name:             "save-mutex-profile",
 		ShortDescription: "Saves a mutex profile to the provided path, ex: `mutex-profile.pb.gz`",
-		Callback:         sshGetMutexProfile,
+		Callback: func(fs any, a []string, w sshd.StringWriter) error {
+			return sshGetMutexProfile(sandboxDir, fs, a, w)
+		},
 	})
 
 	ssh.RegisterCommand(&sshd.Command{
@@ -505,13 +519,43 @@ func sshListLighthouseMap(lightHouse *LightHouse, a any, w sshd.StringWriter) er
 	return nil
 }
 
-func sshStartCpuProfile(fs any, a []string, w sshd.StringWriter) error {
+// sshSanitizeFilePath validates that the given file path is within the sandbox directory.
+// If sandboxDir is empty, the path is returned as-is for backwards compatibility.
+func sshSanitizeFilePath(sandboxDir, filePath string) (string, error) {
+	if sandboxDir == "" {
+		return filePath, nil
+	}
+
+	// Clean and resolve the path relative to the sandbox directory
+	if !filepath.IsAbs(filePath) {
+		filePath = filepath.Join(sandboxDir, filePath)
+	}
+	cleaned := filepath.Clean(filePath)
+
+	// Ensure the resolved path is within the sandbox directory
+	cleanedSandbox := filepath.Clean(sandboxDir)
+	if cleaned == cleanedSandbox {
+		return "", fmt.Errorf("path %q resolves to the sandbox directory itself %q", filePath, sandboxDir)
+	}
+	if !strings.HasPrefix(cleaned, cleanedSandbox+string(filepath.Separator)) {
+		return "", fmt.Errorf("path %q is outside the sandbox directory %q", filePath, sandboxDir)
+	}
+
+	return cleaned, nil
+}
+
+func sshStartCpuProfile(sandboxDir string, fs any, a []string, w sshd.StringWriter) error {
 	if len(a) == 0 {
 		err := w.WriteLine("No path to write profile provided")
 		return err
 	}
 
-	file, err := os.Create(a[0])
+	filePath, err := sshSanitizeFilePath(sandboxDir, a[0])
+	if err != nil {
+		return w.WriteLine(err.Error())
+	}
+
+	file, err := os.Create(filePath)
 	if err != nil {
 		err = w.WriteLine(fmt.Sprintf("Unable to create profile file: %s", err))
 		return err
@@ -675,12 +719,17 @@ func sshChangeRemote(ifce *Interface, fs any, a []string, w sshd.StringWriter) e
 	return w.WriteLine("Changed")
 }
 
-func sshGetHeapProfile(fs any, a []string, w sshd.StringWriter) error {
+func sshGetHeapProfile(sandboxDir string, fs any, a []string, w sshd.StringWriter) error {
 	if len(a) == 0 {
 		return w.WriteLine("No path to write profile provided")
 	}
 
-	file, err := os.Create(a[0])
+	filePath, err := sshSanitizeFilePath(sandboxDir, a[0])
+	if err != nil {
+		return w.WriteLine(err.Error())
+	}
+
+	file, err := os.Create(filePath)
 	if err != nil {
 		err = w.WriteLine(fmt.Sprintf("Unable to create profile file: %s", err))
 		return err
@@ -711,12 +760,17 @@ func sshMutexProfileFraction(fs any, a []string, w sshd.StringWriter) error {
 	return w.WriteLine(fmt.Sprintf("New value: %d. Old value: %d", newRate, oldRate))
 }
 
-func sshGetMutexProfile(fs any, a []string, w sshd.StringWriter) error {
+func sshGetMutexProfile(sandboxDir string, fs any, a []string, w sshd.StringWriter) error {
 	if len(a) == 0 {
 		return w.WriteLine("No path to write profile provided")
 	}
 
-	file, err := os.Create(a[0])
+	filePath, err := sshSanitizeFilePath(sandboxDir, a[0])
+	if err != nil {
+		return w.WriteLine(err.Error())
+	}
+
+	file, err := os.Create(filePath)
 	if err != nil {
 		return w.WriteLine(fmt.Sprintf("Unable to create profile file: %s", err))
 	}
@@ -831,9 +885,7 @@ func sshPrintRelays(ifce *Interface, fs any, a []string, w sshd.StringWriter) er
 
 	relays := map[uint32]*HostInfo{}
 	ifce.hostMap.Lock()
-	for k, v := range ifce.hostMap.Relays {
-		relays[k] = v
-	}
+	maps.Copy(relays, ifce.hostMap.Relays)
 	ifce.hostMap.Unlock()
 
 	type RelayFor struct {

tcp_coalesce.go

@@ -0,0 +1,484 @@
+package nebula
+
+import (
+	"bytes"
+	"encoding/binary"
+	"io"
+
+	"github.com/slackhq/nebula/overlay"
+)
+
+// ipProtoTCP is the IANA protocol number for TCP. Hardcoded instead of
+// reaching for golang.org/x/sys/unix — that package doesn't define the
+// constant on Windows, which would break cross-compiles even though this
+// file runs unchanged on every platform.
+const ipProtoTCP = 6
+
+// tcpCoalesceBufSize caps total bytes per superpacket. Mirrors the kernel's
+// sk_gso_max_size of ~64KiB; anything beyond this would be rejected anyway.
+const tcpCoalesceBufSize = 65535
+
+// tcpCoalesceMaxSegs caps how many segments we'll coalesce into a single
+// superpacket. Keeping this well below the kernel's TSO ceiling bounds
+// latency.
+const tcpCoalesceMaxSegs = 64
+
+// tcpCoalesceHdrCap is the scratch space we copy a seed's IP+TCP header
+// into. IPv6 (40) + TCP with full options (60) = 100 bytes.
+const tcpCoalesceHdrCap = 100
+
+// initialSlots is the starting capacity of the slot pool. One flow per
+// packet is the worst case so this matches a typical UDP recvmmsg batch.
+const initialSlots = 64
+
+// flowKey identifies a TCP flow by {src, dst, sport, dport, family}.
+// Comparable, so linear scans over the slot list stay tight.
+type flowKey struct {
+	src, dst     [16]byte
+	sport, dport uint16
+	isV6         bool
+}
+
+// coalesceSlot is one entry in the coalescer's ordered event queue. When
+// passthrough is true the slot holds a single borrowed packet that must be
+// emitted verbatim (non-TCP, non-admissible TCP, or oversize seed). When
+// passthrough is false the slot is an in-progress coalesced superpacket:
+// hdrBuf is a mutable copy of the seed's IP+TCP header (we patch total
+// length and pseudo-header partial at flush), and payIovs are *borrowed*
+// slices from the caller's plaintext buffers — no payload is ever copied.
+// The caller (listenOut) must keep those buffers alive until Flush.
+type coalesceSlot struct {
+	passthrough bool
+	rawPkt      []byte // borrowed when passthrough
+
+	fk       flowKey
+	hdrBuf   [tcpCoalesceHdrCap]byte
+	hdrLen   int
+	ipHdrLen int
+	isV6     bool
+	gsoSize  int
+	numSeg   int
+	totalPay int
+	nextSeq  uint32
+	// psh closes the chain: set when the last-accepted segment had PSH or
+	// was sub-gsoSize. No further appends after that.
+	psh     bool
+	payIovs [][]byte
+}
+
+// tcpCoalescer accumulates adjacent in-flow TCP data segments across
+// multiple concurrent flows and emits each flow's run as a single TSO
+// superpacket via overlay.GSOWriter. All output — coalesced or not — is
+// deferred until Flush so arrival order is preserved on the wire. Owns
+// no locks; one coalescer per TUN write queue.
+type tcpCoalescer struct {
+	plainW io.Writer
+	gsoW   overlay.GSOWriter // nil when the queue doesn't support TSO
+
+	// slots is the ordered event queue. Flush walks it once and emits each
+	// entry as either a WriteGSO (coalesced) or a plainW.Write (passthrough).
+	slots []*coalesceSlot
+	// openSlots maps a flow key to its most recent non-sealed slot, so new
+	// segments can extend an in-progress superpacket in O(1). Slots are
+	// removed from this map when they close (PSH or short-last-segment),
+	// when a non-admissible packet for that flow arrives, or in Flush.
+	openSlots map[flowKey]*coalesceSlot
+	pool      []*coalesceSlot // free list for reuse
+}
+
+func newTCPCoalescer(w io.Writer) *tcpCoalescer {
+	c := &tcpCoalescer{
+		plainW:    w,
+		slots:     make([]*coalesceSlot, 0, initialSlots),
+		openSlots: make(map[flowKey]*coalesceSlot, initialSlots),
+		pool:      make([]*coalesceSlot, 0, initialSlots),
+	}
+	if gw, ok := w.(overlay.GSOWriter); ok && gw.GSOSupported() {
+		c.gsoW = gw
+	}
+	return c
+}
+
+// parsedTCP holds the fields extracted from a single parse so later steps
+// (admission, slot lookup, canAppend) don't re-walk the header.
+type parsedTCP struct {
+	fk        flowKey
+	ipHdrLen  int
+	tcpHdrLen int
+	hdrLen    int
+	payLen    int
+	seq       uint32
+	flags     byte
+}
+
+// parseTCPBase extracts the flow key and IP/TCP offsets for any TCP packet,
+// regardless of whether it's admissible for coalescing. Returns ok=false
+// for non-TCP or malformed input. Accepts IPv4 (no options, no fragmentation)
+// and IPv6 (no extension headers).
+func parseTCPBase(pkt []byte) (parsedTCP, bool) {
+	var p parsedTCP
+	if len(pkt) < 20 {
+		return p, false
+	}
+	v := pkt[0] >> 4
+	switch v {
+	case 4:
+		ihl := int(pkt[0]&0x0f) * 4
+		if ihl != 20 {
+			return p, false
+		}
+		if pkt[9] != ipProtoTCP {
+			return p, false
+		}
+		// Reject actual fragmentation (MF or non-zero frag offset).
+		if binary.BigEndian.Uint16(pkt[6:8])&0x3fff != 0 {
+			return p, false
+		}
+		totalLen := int(binary.BigEndian.Uint16(pkt[2:4]))
+		if totalLen > len(pkt) || totalLen < ihl {
+			return p, false
+		}
+		p.ipHdrLen = 20
+		p.fk.isV6 = false
+		copy(p.fk.src[:4], pkt[12:16])
+		copy(p.fk.dst[:4], pkt[16:20])
+		pkt = pkt[:totalLen]
+	case 6:
+		if len(pkt) < 40 {
+			return p, false
+		}
+		if pkt[6] != ipProtoTCP {
+			return p, false
+		}
+		payloadLen := int(binary.BigEndian.Uint16(pkt[4:6]))
+		if 40+payloadLen > len(pkt) {
+			return p, false
+		}
+		p.ipHdrLen = 40
+		p.fk.isV6 = true
+		copy(p.fk.src[:], pkt[8:24])
+		copy(p.fk.dst[:], pkt[24:40])
+		pkt = pkt[:40+payloadLen]
+	default:
+		return p, false
+	}
+
+	if len(pkt) < p.ipHdrLen+20 {
+		return p, false
+	}
+	tcpOff := int(pkt[p.ipHdrLen+12]>>4) * 4
+	if tcpOff < 20 || tcpOff > 60 {
+		return p, false
+	}
+	if len(pkt) < p.ipHdrLen+tcpOff {
+		return p, false
+	}
+	p.tcpHdrLen = tcpOff
+	p.hdrLen = p.ipHdrLen + tcpOff
+	p.payLen = len(pkt) - p.hdrLen
+	p.seq = binary.BigEndian.Uint32(pkt[p.ipHdrLen+4 : p.ipHdrLen+8])
+	p.flags = pkt[p.ipHdrLen+13]
+	p.fk.sport = binary.BigEndian.Uint16(pkt[p.ipHdrLen : p.ipHdrLen+2])
+	p.fk.dport = binary.BigEndian.Uint16(pkt[p.ipHdrLen+2 : p.ipHdrLen+4])
+	return p, true
+}
+
+// coalesceable reports whether a parsed TCP segment is eligible for
+// coalescing. Accepts only ACK or ACK|PSH with a non-empty payload.
+func (p parsedTCP) coalesceable() bool {
+	const ack = 0x10
+	const psh = 0x08
+	if p.flags&^(ack|psh) != 0 || p.flags&ack == 0 {
+		return false
+	}
+	return p.payLen > 0
+}
+
+// Add borrows pkt. The caller must keep pkt valid until the next Flush,
+// whether or not the packet was coalesced — passthrough (non-admissible)
+// packets are queued and written at Flush time, not synchronously.
+func (c *tcpCoalescer) Add(pkt []byte) error {
+	if c.gsoW == nil {
+		c.addPassthrough(pkt)
+		return nil
+	}
+
+	info, ok := parseTCPBase(pkt)
+	if !ok {
+		// Non-TCP or malformed — can't possibly collide with an open flow.
+		c.addPassthrough(pkt)
+		return nil
+	}
+	if !info.coalesceable() {
+		// TCP but not admissible (SYN/FIN/RST/URG/CWR/ECE or zero-payload).
+		// Seal this flow's open slot so later in-flow packets don't extend
+		// it and accidentally reorder past this passthrough.
+		delete(c.openSlots, info.fk)
+		c.addPassthrough(pkt)
+		return nil
+	}
+
+	if open := c.openSlots[info.fk]; open != nil {
+		if c.canAppend(open, pkt, info) {
+			c.appendPayload(open, pkt, info)
+			if open.psh {
+				delete(c.openSlots, info.fk)
+			}
+			return nil
+		}
+		// Can't extend — seal it and fall through to seed a fresh slot.
+		delete(c.openSlots, info.fk)
+	}
+	c.seed(pkt, info)
+	return nil
+}
+
+// Flush emits every queued event in arrival order. Coalesced slots go out
+// via WriteGSO; passthrough slots go out via plainW.Write. Returns the
+// first error observed; keeps draining so one bad packet doesn't hold up
+// the rest. After Flush returns, borrowed payload slices may be recycled.
+func (c *tcpCoalescer) Flush() error {
+	var first error
+	for _, s := range c.slots {
+		var err error
+		if s.passthrough {
+			_, err = c.plainW.Write(s.rawPkt)
+		} else {
+			err = c.flushSlot(s)
+		}
+		if err != nil && first == nil {
+			first = err
+		}
+		c.release(s)
+	}
+	for i := range c.slots {
+		c.slots[i] = nil
+	}
+	c.slots = c.slots[:0]
+	for k := range c.openSlots {
+		delete(c.openSlots, k)
+	}
+	return first
+}
+
+func (c *tcpCoalescer) addPassthrough(pkt []byte) {
+	s := c.take()
+	s.passthrough = true
+	s.rawPkt = pkt
+	c.slots = append(c.slots, s)
+}
+
+func (c *tcpCoalescer) seed(pkt []byte, info parsedTCP) {
+	if info.hdrLen > tcpCoalesceHdrCap || info.hdrLen+info.payLen > tcpCoalesceBufSize {
+		// Pathological shape — can't fit our scratch, emit as-is.
+		c.addPassthrough(pkt)
+		return
+	}
+	s := c.take()
+	s.passthrough = false
+	s.rawPkt = nil
+	copy(s.hdrBuf[:], pkt[:info.hdrLen])
+	s.hdrLen = info.hdrLen
+	s.ipHdrLen = info.ipHdrLen
+	s.isV6 = info.fk.isV6
+	s.fk = info.fk
+	s.gsoSize = info.payLen
+	s.numSeg = 1
+	s.totalPay = info.payLen
+	s.nextSeq = info.seq + uint32(info.payLen)
+	s.psh = info.flags&0x08 != 0
+	s.payIovs = append(s.payIovs[:0], pkt[info.hdrLen:info.hdrLen+info.payLen])
+	c.slots = append(c.slots, s)
+	if !s.psh {
+		c.openSlots[info.fk] = s
+	}
+}
+
+// canAppend reports whether info's packet extends the slot's seed: same
+// header shape and stable contents, adjacent seq, not oversized, chain not
+// closed.
+func (c *tcpCoalescer) canAppend(s *coalesceSlot, pkt []byte, info parsedTCP) bool {
+	if s.psh {
+		return false
+	}
+	if info.hdrLen != s.hdrLen {
+		return false
+	}
+	if info.seq != s.nextSeq {
+		return false
+	}
+	if s.numSeg >= tcpCoalesceMaxSegs {
+		return false
+	}
+	if info.payLen > s.gsoSize {
+		return false
+	}
+	if s.hdrLen+s.totalPay+info.payLen > tcpCoalesceBufSize {
+		return false
+	}
+	if !headersMatch(s.hdrBuf[:s.hdrLen], pkt[:info.hdrLen], s.isV6, s.ipHdrLen) {
+		return false
+	}
+	return true
+}
+
+func (c *tcpCoalescer) appendPayload(s *coalesceSlot, pkt []byte, info parsedTCP) {
+	s.payIovs = append(s.payIovs, pkt[info.hdrLen:info.hdrLen+info.payLen])
+	s.numSeg++
+	s.totalPay += info.payLen
+	s.nextSeq = info.seq + uint32(info.payLen)
+	if info.payLen < s.gsoSize || info.flags&0x08 != 0 {
+		s.psh = true
+	}
+}
+
+func (c *tcpCoalescer) take() *coalesceSlot {
+	if n := len(c.pool); n > 0 {
+		s := c.pool[n-1]
+		c.pool[n-1] = nil
+		c.pool = c.pool[:n-1]
+		return s
+	}
+	return &coalesceSlot{}
+}
+
+func (c *tcpCoalescer) release(s *coalesceSlot) {
+	s.passthrough = false
+	s.rawPkt = nil
+	for i := range s.payIovs {
+		s.payIovs[i] = nil
+	}
+	s.payIovs = s.payIovs[:0]
+	s.numSeg = 0
+	s.totalPay = 0
+	s.psh = false
+	c.pool = append(c.pool, s)
+}
+
+// flushSlot patches the header and calls WriteGSO. Does not remove the
+// slot from c.slots.
+func (c *tcpCoalescer) flushSlot(s *coalesceSlot) error {
+	total := s.hdrLen + s.totalPay
+	l4Len := total - s.ipHdrLen
+	hdr := s.hdrBuf[:s.hdrLen]
+
+	if s.isV6 {
+		binary.BigEndian.PutUint16(hdr[4:6], uint16(l4Len))
+	} else {
+		binary.BigEndian.PutUint16(hdr[2:4], uint16(total))
+		hdr[10] = 0
+		hdr[11] = 0
+		binary.BigEndian.PutUint16(hdr[10:12], ipv4HdrChecksum(hdr[:s.ipHdrLen]))
+	}
+
+	var psum uint32
+	if s.isV6 {
+		psum = pseudoSumIPv6(hdr[8:24], hdr[24:40], ipProtoTCP, l4Len)
+	} else {
+		psum = pseudoSumIPv4(hdr[12:16], hdr[16:20], ipProtoTCP, l4Len)
+	}
+	tcsum := s.ipHdrLen + 16
+	binary.BigEndian.PutUint16(hdr[tcsum:tcsum+2], foldOnceNoInvert(psum))
+
+	return c.gsoW.WriteGSO(hdr, s.payIovs, uint16(s.gsoSize), s.isV6, uint16(s.ipHdrLen))
+}
+
+// headersMatch compares two IP+TCP header prefixes for byte-for-byte
+// equality on every field that must be identical across coalesced
+// segments. Size/IPID/IPCsum/seq/flags/tcpCsum are masked out.
+func headersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	if isV6 {
+		// IPv6: bytes [0:4] = version/TC/flow-label, [6:8] = next_hdr/hop,
+		// [8:40] = src+dst. Skip [4:6] payload length.
+		if !bytes.Equal(a[0:4], b[0:4]) {
+			return false
+		}
+		if !bytes.Equal(a[6:40], b[6:40]) {
+			return false
+		}
+	} else {
+		// IPv4: [0:2] version/IHL/TOS, [6:10] flags/fragoff/TTL/proto,
+		// [12:20] src+dst. Skip [2:4] total len, [4:6] id, [10:12] csum.
+		if !bytes.Equal(a[0:2], b[0:2]) {
+			return false
+		}
+		if !bytes.Equal(a[6:10], b[6:10]) {
+			return false
+		}
+		if !bytes.Equal(a[12:20], b[12:20]) {
+			return false
+		}
+	}
+	// TCP: compare [0:4] ports, [8:13] ack+dataoff, [14:16] window,
+	// [18:tcpHdrLen] options (incl. urgent).
+	tcp := ipHdrLen
+	if !bytes.Equal(a[tcp:tcp+4], b[tcp:tcp+4]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+8:tcp+13], b[tcp+8:tcp+13]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+14:tcp+16], b[tcp+14:tcp+16]) {
+		return false
+	}
+	if !bytes.Equal(a[tcp+18:], b[tcp+18:]) {
+		return false
+	}
+	return true
+}
+
+// ipv4HdrChecksum computes the IPv4 header checksum over hdr (which must
+// already have its checksum field zeroed) and returns the folded/inverted
+// 16-bit value to store.
+func ipv4HdrChecksum(hdr []byte) uint16 {
+	var sum uint32
+	for i := 0; i+1 < len(hdr); i += 2 {
+		sum += uint32(binary.BigEndian.Uint16(hdr[i : i+2]))
+	}
+	if len(hdr)%2 == 1 {
+		sum += uint32(hdr[len(hdr)-1]) << 8
+	}
+	for sum>>16 != 0 {
+		sum = (sum & 0xffff) + (sum >> 16)
+	}
+	return ^uint16(sum)
+}
+
+// pseudoSumIPv4 / pseudoSumIPv6 build the TCP pseudo-header partial sum
+// expected by the virtio NEEDS_CSUM kernel path: the 32-bit accumulator
+// before folding.
+func pseudoSumIPv4(src, dst []byte, proto byte, l4Len int) uint32 {
+	var sum uint32
+	sum += uint32(binary.BigEndian.Uint16(src[0:2]))
+	sum += uint32(binary.BigEndian.Uint16(src[2:4]))
+	sum += uint32(binary.BigEndian.Uint16(dst[0:2]))
+	sum += uint32(binary.BigEndian.Uint16(dst[2:4]))
+	sum += uint32(proto)
+	sum += uint32(l4Len)
+	return sum
+}
+
+func pseudoSumIPv6(src, dst []byte, proto byte, l4Len int) uint32 {
+	var sum uint32
+	for i := 0; i < 16; i += 2 {
+		sum += uint32(binary.BigEndian.Uint16(src[i : i+2]))
+		sum += uint32(binary.BigEndian.Uint16(dst[i : i+2]))
+	}
+	sum += uint32(l4Len >> 16)
+	sum += uint32(l4Len & 0xffff)
+	sum += uint32(proto)
+	return sum
+}
+
+// foldOnceNoInvert folds the 32-bit accumulator to 16 bits and returns it
+// unchanged (no one's complement). This is what virtio NEEDS_CSUM wants in
+// the L4 checksum field — the kernel will add the payload sum and invert.
+func foldOnceNoInvert(sum uint32) uint16 {
+	for sum>>16 != 0 {
+		sum = (sum & 0xffff) + (sum >> 16)
+	}
+	return uint16(sum)
+}

tcp_coalesce_test.go

@@ -0,0 +1,576 @@
+package nebula
+
+import (
+	"encoding/binary"
+	"testing"
+)
+
+// fakeTunWriter records plain Writes and WriteGSO calls without touching a
+// real TUN fd. WriteGSO preserves the split between hdr and borrowed pays
+// so tests can inspect each independently.
+type fakeTunWriter struct {
+	gsoEnabled bool
+	writes     [][]byte
+	gsoWrites  []fakeGSOWrite
+}
+
+type fakeGSOWrite struct {
+	hdr       []byte
+	pays      [][]byte
+	gsoSize   uint16
+	isV6      bool
+	csumStart uint16
+}
+
+// total returns hdrLen + sum of pay lens.
+func (g fakeGSOWrite) total() int {
+	n := len(g.hdr)
+	for _, p := range g.pays {
+		n += len(p)
+	}
+	return n
+}
+
+// payLen sums the pays.
+func (g fakeGSOWrite) payLen() int {
+	var n int
+	for _, p := range g.pays {
+		n += len(p)
+	}
+	return n
+}
+
+func (w *fakeTunWriter) Write(p []byte) (int, error) {
+	buf := make([]byte, len(p))
+	copy(buf, p)
+	w.writes = append(w.writes, buf)
+	return len(p), nil
+}
+
+func (w *fakeTunWriter) WriteGSO(hdr []byte, pays [][]byte, gsoSize uint16, isV6 bool, csumStart uint16) error {
+	hcopy := make([]byte, len(hdr))
+	copy(hcopy, hdr)
+	paysCopy := make([][]byte, len(pays))
+	for i, p := range pays {
+		pc := make([]byte, len(p))
+		copy(pc, p)
+		paysCopy[i] = pc
+	}
+	w.gsoWrites = append(w.gsoWrites, fakeGSOWrite{
+		hdr:       hcopy,
+		pays:      paysCopy,
+		gsoSize:   gsoSize,
+		isV6:      isV6,
+		csumStart: csumStart,
+	})
+	return nil
+}
+
+func (w *fakeTunWriter) GSOSupported() bool { return w.gsoEnabled }
+
+// buildTCPv4 constructs a minimal IPv4+TCP packet with the given payload,
+// seq, and flags. Assumes no IP options and a 20-byte TCP header.
+func buildTCPv4(seq uint32, flags byte, payload []byte) []byte {
+	return buildTCPv4Ports(1000, 2000, seq, flags, payload)
+}
+
+// buildTCPv4Ports is buildTCPv4 with caller-specified ports so tests can
+// build distinct flows.
+func buildTCPv4Ports(sport, dport uint16, seq uint32, flags byte, payload []byte) []byte {
+	const ipHdrLen = 20
+	const tcpHdrLen = 20
+	total := ipHdrLen + tcpHdrLen + len(payload)
+	pkt := make([]byte, total)
+
+	pkt[0] = 0x45
+	pkt[1] = 0x00
+	binary.BigEndian.PutUint16(pkt[2:4], uint16(total))
+	binary.BigEndian.PutUint16(pkt[4:6], 0)
+	binary.BigEndian.PutUint16(pkt[6:8], 0x4000)
+	pkt[8] = 64
+	pkt[9] = ipProtoTCP
+	copy(pkt[12:16], []byte{10, 0, 0, 1})
+	copy(pkt[16:20], []byte{10, 0, 0, 2})
+
+	binary.BigEndian.PutUint16(pkt[20:22], sport)
+	binary.BigEndian.PutUint16(pkt[22:24], dport)
+	binary.BigEndian.PutUint32(pkt[24:28], seq)
+	binary.BigEndian.PutUint32(pkt[28:32], 12345)
+	pkt[32] = 0x50
+	pkt[33] = flags
+	binary.BigEndian.PutUint16(pkt[34:36], 0xffff)
+
+	copy(pkt[40:], payload)
+	return pkt
+}
+
+const (
+	tcpAck    = 0x10
+	tcpPsh    = 0x08
+	tcpSyn    = 0x02
+	tcpFin    = 0x01
+	tcpAckPsh = tcpAck | tcpPsh
+)
+
+func TestCoalescerPassthroughWhenGSOUnavailable(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: false}
+	c := newTCPCoalescer(w)
+	pkt := buildTCPv4(1000, tcpAck, []byte("hello"))
+	if err := c.Add(pkt); err != nil {
+		t.Fatal(err)
+	}
+	// No sync write — passthrough is deferred to Flush.
+	if len(w.writes) != 0 || len(w.gsoWrites) != 0 {
+		t.Fatalf("no Add-time writes: got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("want single plain write, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerNonTCPPassthrough(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pkt := make([]byte, 28)
+	pkt[0] = 0x45
+	binary.BigEndian.PutUint16(pkt[2:4], 28)
+	pkt[9] = 1
+	copy(pkt[12:16], []byte{10, 0, 0, 1})
+	copy(pkt[16:20], []byte{10, 0, 0, 2})
+	if err := c.Add(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("ICMP should pass through unchanged")
+	}
+}
+
+func TestCoalescerSeedThenFlushAlone(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pkt := buildTCPv4(1000, tcpAck, make([]byte, 1000))
+	if err := c.Add(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 0 || len(w.gsoWrites) != 0 {
+		t.Fatalf("unexpected output before flush")
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Single-segment flush now goes through WriteGSO with GSO_NONE
+	// (virtio NEEDS_CSUM lets the kernel fill in the L4 csum).
+	if len(w.gsoWrites) != 1 || len(w.writes) != 0 {
+		t.Fatalf("single-seg flush: writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+	g := w.gsoWrites[0]
+	if g.total() != 40+1000 {
+		t.Errorf("super total=%d want %d", g.total(), 40+1000)
+	}
+	if g.payLen() != 1000 {
+		t.Errorf("payLen=%d want 1000", g.payLen())
+	}
+}
+
+func TestCoalescerCoalescesAdjacentACKs(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(2200, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(3400, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 1 {
+		t.Fatalf("want 1 gso write, got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
+	}
+	g := w.gsoWrites[0]
+	if g.gsoSize != 1200 {
+		t.Errorf("gsoSize=%d want 1200", g.gsoSize)
+	}
+	if len(g.hdr) != 40 {
+		t.Errorf("hdrLen=%d want 40", len(g.hdr))
+	}
+	if g.csumStart != 20 {
+		t.Errorf("csumStart=%d want 20", g.csumStart)
+	}
+	if len(g.pays) != 3 {
+		t.Errorf("pay count=%d want 3", len(g.pays))
+	}
+	if g.total() != 40+3*1200 {
+		t.Errorf("superpacket len=%d want %d", g.total(), 40+3*1200)
+	}
+	if tot := binary.BigEndian.Uint16(g.hdr[2:4]); int(tot) != g.total() {
+		t.Errorf("ip total_length=%d want %d", tot, g.total())
+	}
+}
+
+func TestCoalescerRejectsSeqGap(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(3000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Each packet flushes as its own single-segment WriteGSO now.
+	if len(w.gsoWrites) != 2 || len(w.writes) != 0 {
+		t.Fatalf("seq gap: want 2 gso writes got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsFlagMismatch(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// SYN|ACK is non-admissible. Must flush matching flow's slot (gso)
+	// and then plain-write the SYN packet itself.
+	syn := buildTCPv4(2200, tcpSyn|tcpAck, pay)
+	if err := c.Add(syn); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 1 || len(w.gsoWrites) != 1 {
+		t.Fatalf("flag mismatch: want 1 plain + 1 gso, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsFIN(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	fin := buildTCPv4(1000, tcpAck|tcpFin, []byte("x"))
+	if err := c.Add(fin); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// FIN isn't admissible — passthrough as plain, no slot, no gso.
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("FIN should be passthrough, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerShortLastSegmentClosesChain(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	full := make([]byte, 1200)
+	half := make([]byte, 500)
+	if err := c.Add(buildTCPv4(1000, tcpAck, full)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(2200, tcpAck, half)); err != nil {
+		t.Fatal(err)
+	}
+	// Chain now closed; next packet seeds a new slot on the same flow
+	// after flushing the old one.
+	if err := c.Add(buildTCPv4(2700, tcpAck, full)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Expect two gso writes: the first two packets coalesced, then the
+	// third flushed alone (single-seg via GSO_NONE).
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want 0 plain writes got %d", len(w.writes))
+	}
+	if w.gsoWrites[0].gsoSize != 1200 {
+		t.Errorf("gsoSize=%d want 1200", w.gsoWrites[0].gsoSize)
+	}
+	if got, want := w.gsoWrites[0].total(), 40+1200+500; got != want {
+		t.Errorf("super len=%d want %d", got, want)
+	}
+}
+
+func TestCoalescerPSHFinalizesChain(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(2200, tcpAckPsh, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4(3400, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// First two coalesce; the third seeds a fresh slot that flushes alone.
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want 0 plain writes got %d", len(w.writes))
+	}
+}
+
+func TestCoalescerRejectsDifferentFlow(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 1200)
+	p1 := buildTCPv4(1000, tcpAck, pay)
+	p2 := buildTCPv4(2200, tcpAck, pay)
+	binary.BigEndian.PutUint16(p2[20:22], 9999)
+	if err := c.Add(p1); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(p2); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Two independent flows, each flushes its own single-segment WriteGSO.
+	if len(w.gsoWrites) != 2 || len(w.writes) != 0 {
+		t.Fatalf("diff flow: want 2 gso writes got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerRejectsIPOptions(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 500)
+	pkt := buildTCPv4(1000, tcpAck, pay)
+	// Bump IHL to 6 to simulate 4 bytes of IP options. Don't actually add
+	// bytes — parser should bail before it matters.
+	pkt[0] = 0x46
+	if err := c.Add(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Non-admissible parse → passthrough as plain.
+	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
+		t.Fatalf("IP options should passthrough, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestCoalescerCapBySegments(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 512)
+	seq := uint32(1000)
+	for i := 0; i < tcpCoalesceMaxSegs+5; i++ {
+		if err := c.Add(buildTCPv4(seq, tcpAck, pay)); err != nil {
+			t.Fatal(err)
+		}
+		seq += uint32(len(pay))
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	for _, g := range w.gsoWrites {
+		segs := len(g.pays)
+		if segs > tcpCoalesceMaxSegs {
+			t.Fatalf("super exceeded seg cap: %d > %d", segs, tcpCoalesceMaxSegs)
+		}
+	}
+}
+
+// TestCoalescerMultipleFlowsInSameBatch proves two interleaved bulk TCP
+// flows coalesce independently in a single Flush.
+func TestCoalescerMultipleFlowsInSameBatch(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 1200)
+
+	// Flow A: sport 1000. Flow B: sport 3000.
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 1300, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 1700, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 2500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 2900, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes (one per flow), got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 0 {
+		t.Fatalf("want no plain writes, got %d", len(w.writes))
+	}
+	// Each superpacket should carry 3 segments.
+	for i, g := range w.gsoWrites {
+		if len(g.pays) != 3 {
+			t.Errorf("gso[%d]: segs=%d want 3", i, len(g.pays))
+		}
+		if g.gsoSize != 1200 {
+			t.Errorf("gso[%d]: gsoSize=%d want 1200", i, g.gsoSize)
+		}
+	}
+	// Verify each superpacket carries the source port it was seeded with.
+	seenSports := map[uint16]bool{}
+	for _, g := range w.gsoWrites {
+		sp := binary.BigEndian.Uint16(g.hdr[20:22])
+		seenSports[sp] = true
+	}
+	if !seenSports[1000] || !seenSports[3000] {
+		t.Errorf("expected superpackets for sports 1000 and 3000, got %v", seenSports)
+	}
+}
+
+// TestCoalescerPreservesArrivalOrder confirms that with passthrough and
+// coalesced events both queued, Flush emits them in Add order rather than
+// writing passthrough packets synchronously.
+func TestCoalescerPreservesArrivalOrder(t *testing.T) {
+	w := &orderedFakeWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	// Sequence: coalesceable TCP, ICMP (passthrough), coalesceable TCP on
+	// a different flow. Expected emit order: gso(X), plain(ICMP), gso(Y).
+	pay := make([]byte, 1200)
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	icmp := make([]byte, 28)
+	icmp[0] = 0x45
+	binary.BigEndian.PutUint16(icmp[2:4], 28)
+	icmp[9] = 1
+	copy(icmp[12:16], []byte{10, 0, 0, 1})
+	copy(icmp[16:20], []byte{10, 0, 0, 3})
+	if err := c.Add(icmp); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Nothing should have hit the writer synchronously.
+	if len(w.events) != 0 {
+		t.Fatalf("Add emitted events synchronously: %v", w.events)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if got, want := w.events, []string{"gso", "plain", "gso"}; !stringSliceEq(got, want) {
+		t.Fatalf("flush order=%v want %v", got, want)
+	}
+}
+
+// orderedFakeWriter records only the sequence of call types so tests can
+// assert arrival order without inspecting bytes.
+type orderedFakeWriter struct {
+	gsoEnabled bool
+	events     []string
+}
+
+func (w *orderedFakeWriter) Write(p []byte) (int, error) {
+	w.events = append(w.events, "plain")
+	return len(p), nil
+}
+
+func (w *orderedFakeWriter) WriteGSO(hdr []byte, pays [][]byte, gsoSize uint16, isV6 bool, csumStart uint16) error {
+	w.events = append(w.events, "gso")
+	return nil
+}
+
+func (w *orderedFakeWriter) GSOSupported() bool { return w.gsoEnabled }
+
+func stringSliceEq(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}
+
+// TestCoalescerInterleavedFlowsPreserveOrdering checks that a non-admissible
+// packet (SYN) mid-flow only flushes its own flow, not others.
+func TestCoalescerInterleavedFlowsPreserveOrdering(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := newTCPCoalescer(w)
+	pay := make([]byte, 1200)
+
+	// Flow A two segments.
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 100, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(1000, 2000, 1300, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Flow B two segments.
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 500, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 1700, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	// Flow A SYN (non-admissible) — must flush only flow A's slot.
+	syn := buildTCPv4Ports(1000, 2000, 9999, tcpSyn|tcpAck, pay)
+	if err := c.Add(syn); err != nil {
+		t.Fatal(err)
+	}
+	// Flow B continues — should still be coalesced with its seed.
+	if err := c.Add(buildTCPv4Ports(3000, 2000, 2900, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+
+	// Expected:
+	//   - 1 gso for flow A (first 2 segments)
+	//   - 1 plain for flow A SYN
+	//   - 1 gso for flow B (3 segments)
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes, got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 1 {
+		t.Fatalf("want 1 plain write (SYN), got %d", len(w.writes))
+	}
+	// Find the 3-segment gso (flow B) and the 2-segment gso (flow A).
+	var segCounts []int
+	for _, g := range w.gsoWrites {
+		segCounts = append(segCounts, len(g.pays))
+	}
+	if !(segCounts[0] == 2 && segCounts[1] == 3) && !(segCounts[0] == 3 && segCounts[1] == 2) {
+		t.Errorf("unexpected segment counts: %v (want 2 and 3)", segCounts)
+	}
+}

timeout_test.go

@@ -134,7 +134,7 @@ func TestTimerWheel_Purge(t *testing.T) {
 	assert.True(t, tw.lastTick.After(lastTick))
 
 	// Make sure we get all 4 packets back
-	for i := 0; i < 4; i++ {
+	for i := range 4 {
 		p, has := tw.Purge()
 		assert.True(t, has)
 		assert.Equal(t, fps[i], p)
@@ -149,7 +149,7 @@ func TestTimerWheel_Purge(t *testing.T) {
 	// Make sure we cached the free'd items
 	assert.Equal(t, 4, tw.itemsCached)
 	ci := tw.itemCache
-	for i := 0; i < 4; i++ {
+	for range 4 {
 		assert.NotNil(t, ci)
 		ci = ci.Next
 	}

udp/conn.go

@@ -8,6 +8,12 @@ import (
 
 const MTU = 9001
 
+// MaxWriteBatch is the largest batch any Conn.WriteBatch implementation is
+// required to accept. Callers SHOULD NOT pass more than this per call; Linux
+// backends preallocate sendmmsg scratch sized to this value, so exceeding it
+// only costs a chunked retry.
+const MaxWriteBatch = 128
+
 type EncReader func(
 	addr netip.AddrPort,
 	payload []byte,
@@ -16,8 +22,29 @@ type EncReader func(
 type Conn interface {
 	Rebind() error
 	LocalAddr() (netip.AddrPort, error)
-	ListenOut(r EncReader)
+	// ListenOut invokes r for each received packet. On batch-capable
+	// backends (recvmmsg), flush is called after each batch is fully
+	// delivered — callers use it to flush per-batch accumulators such as
+	// TUN write coalescers. Single-packet backends call flush after each
+	// packet. flush must not be nil.
+	ListenOut(r EncReader, flush func()) error
 	WriteTo(b []byte, addr netip.AddrPort) error
+	// WriteBatch sends a contiguous batch of packets, each with its own
+	// destination. bufs and addrs must have the same length. Linux uses
+	// sendmmsg(2) for a single syscall; other backends fall back to a
+	// WriteTo loop. Returns on the first error; callers may observe a
+	// partial send if some packets went out before the error.
+	WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error
+	// WriteSegmented sends bufs as a single UDP GSO sendmsg when the kernel
+	// supports it: all bufs go to the same addr, each must be exactly segSize
+	// bytes except the last which may be shorter. The kernel emits one
+	// datagram per buf on the wire. Backends / kernels without GSO support
+	// fall back to a per-packet WriteTo loop. Returns on the first error.
+	WriteSegmented(bufs [][]byte, addr netip.AddrPort, segSize int) error
+	// SupportsGSO reports whether WriteSegmented takes the single-syscall
+	// GSO path. Callers use this to decide at batch-assembly time whether
+	// the uniform-size / same-dst check is worth running.
+	SupportsGSO() bool
 	ReloadConfig(c *config.C)
 	SupportsMultipleReaders() bool
 	Close() error
@@ -31,8 +58,8 @@ func (NoopConn) Rebind() error {
 func (NoopConn) LocalAddr() (netip.AddrPort, error) {
 	return netip.AddrPort{}, nil
 }
-func (NoopConn) ListenOut(_ EncReader) {
-	return
+func (NoopConn) ListenOut(_ EncReader, _ func()) error {
+	return nil
 }
 func (NoopConn) SupportsMultipleReaders() bool {
 	return false
@@ -40,6 +67,15 @@ func (NoopConn) SupportsMultipleReaders() bool {
 func (NoopConn) WriteTo(_ []byte, _ netip.AddrPort) error {
 	return nil
 }
+func (NoopConn) WriteBatch(_ [][]byte, _ []netip.AddrPort) error {
+	return nil
+}
+func (NoopConn) WriteSegmented(_ [][]byte, _ netip.AddrPort, _ int) error {
+	return nil
+}
+func (NoopConn) SupportsGSO() bool {
+	return false
+}
 func (NoopConn) ReloadConfig(_ *config.C) {
 	return
 }

udp/udp_darwin.go

@@ -140,6 +140,26 @@ func (u *StdConn) WriteTo(b []byte, ap netip.AddrPort) error {
 	}
 }
 
+func (u *StdConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *StdConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, _ int) error {
+	for _, b := range bufs {
+		if err := u.WriteTo(b, addr); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *StdConn) SupportsGSO() bool { return false }
+
 func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
 	a := u.UDPConn.LocalAddr()
 
@@ -165,7 +185,7 @@ func NewUDPStatsEmitter(udpConns []Conn) func() {
 	return func() {}
 }
 
-func (u *StdConn) ListenOut(r EncReader) {
+func (u *StdConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
 	for {
@@ -173,14 +193,14 @@ func (u *StdConn) ListenOut(r EncReader) {
 		n, rua, err := u.ReadFromUDPAddrPort(buffer)
 		if err != nil {
 			if errors.Is(err, net.ErrClosed) {
-				u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
-				return
+				return err
 			}
 
 			u.l.WithError(err).Error("unexpected udp socket receive error")
 		}
 
 		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
+		flush()
 	}
 }
 

udp/udp_generic.go

@@ -10,11 +10,9 @@ package udp
 
 import (
 	"context"
-	"errors"
 	"fmt"
 	"net"
 	"net/netip"
-	"time"
 
 	"github.com/sirupsen/logrus"
 	"github.com/slackhq/nebula/config"
@@ -44,6 +42,26 @@ func (u *GenericConn) WriteTo(b []byte, addr netip.AddrPort) error {
 	return err
 }
 
+func (u *GenericConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if _, err := u.UDPConn.WriteToUDPAddrPort(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *GenericConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, _ int) error {
+	for _, b := range bufs {
+		if _, err := u.UDPConn.WriteToUDPAddrPort(b, addr); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *GenericConn) SupportsGSO() bool { return false }
+
 func (u *GenericConn) LocalAddr() (netip.AddrPort, error) {
 	a := u.UDPConn.LocalAddr()
 
@@ -73,28 +91,18 @@ type rawMessage struct {
 	Len uint32
 }
 
-func (u *GenericConn) ListenOut(r EncReader) {
+func (u *GenericConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
-	var lastRecvErr time.Time
-
 	for {
 		// Just read one packet at a time
 		n, rua, err := u.ReadFromUDPAddrPort(buffer)
 		if err != nil {
-			if errors.Is(err, net.ErrClosed) {
-				u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
-				return
-			}
-			// Dampen unexpected message warns to once per minute
-			if lastRecvErr.IsZero() || time.Since(lastRecvErr) > time.Minute {
-				lastRecvErr = time.Now()
-				u.l.WithError(err).Warn("unexpected udp socket receive error")
-			}
-			continue
+			return err
 		}
 
 		r(netip.AddrPortFrom(rua.Addr().Unmap(), rua.Port()), buffer[:n])
+		flush()
 	}
 }
 

udp/udp_linux.go

@@ -4,6 +4,7 @@
 package udp
 
 import (
+	"context"
 	"encoding/binary"
 	"fmt"
 	"net"
@@ -18,58 +19,136 @@ import (
 )
 
 type StdConn struct {
-	sysFd int
-	isV4  bool
-	l     *logrus.Logger
-	batch int
+	udpConn *net.UDPConn
+	rawConn syscall.RawConn
+	isV4    bool
+	l       *logrus.Logger
+	batch   int
+
+	// sendmmsg scratch. Each queue has its own StdConn, so no locking is
+	// needed. Sized to MaxWriteBatch at construction; WriteBatch chunks
+	// larger inputs.
+	writeMsgs  []rawMessage
+	writeIovs  []iovec
+	writeNames [][]byte
+
+	// Preallocated closure + in/out slots for sendmmsg, so the hot path
+	// does not heap-allocate a fresh closure per call.
+	writeChunk int
+	writeSent  int
+	writeErrno syscall.Errno
+	writeFunc  func(fd uintptr) bool
+
+	// UDP GSO (sendmsg with UDP_SEGMENT cmsg) support. gsoSupported is
+	// probed once at socket creation. When true, WriteSegmented takes a
+	// single-syscall GSO path; otherwise it falls back to a WriteTo loop.
+	gsoSupported bool
+	gsoMsg       msghdr
+	gsoIovs      []iovec
+	gsoName      []byte // SizeofSockaddrInet6
+	gsoCmsg      []byte // CmsgSpace(2)
+	gsoSent      int
+	gsoErrno     syscall.Errno
+	gsoFunc      func(fd uintptr) bool
 }
 
-func maybeIPV4(ip net.IP) (net.IP, bool) {
-	ip4 := ip.To4()
-	if ip4 != nil {
-		return ip4, true
+func setReusePort(network, address string, c syscall.RawConn) error {
+	var opErr error
+	err := c.Control(func(fd uintptr) {
+		opErr = unix.SetsockoptInt(int(fd), unix.SOL_SOCKET, unix.SO_REUSEPORT, 1)
+		//CloseOnExec already set by the runtime
+	})
+	if err != nil {
+		return err
 	}
-	return ip, false
+	return opErr
 }
 
 func NewListener(l *logrus.Logger, ip netip.Addr, port int, multi bool, batch int) (Conn, error) {
-	af := unix.AF_INET6
-	if ip.Is4() {
-		af = unix.AF_INET
+	listen := netip.AddrPortFrom(ip, uint16(port))
+	lc := net.ListenConfig{}
+	if multi {
+		lc.Control = setReusePort
 	}
-	syscall.ForkLock.RLock()
-	fd, err := unix.Socket(af, unix.SOCK_DGRAM, unix.IPPROTO_UDP)
-	if err == nil {
-		unix.CloseOnExec(fd)
-	}
-	syscall.ForkLock.RUnlock()
-
+	//this context is only used during the bind operation, you can't cancel it to kill the socket
+	pc, err := lc.ListenPacket(context.Background(), "udp", listen.String())
 	if err != nil {
-		unix.Close(fd)
 		return nil, fmt.Errorf("unable to open socket: %s", err)
 	}
-
-	if multi {
-		if err = unix.SetsockoptInt(fd, unix.SOL_SOCKET, unix.SO_REUSEPORT, 1); err != nil {
-			return nil, fmt.Errorf("unable to set SO_REUSEPORT: %s", err)
-		}
+	udpConn := pc.(*net.UDPConn)
+	rawConn, err := udpConn.SyscallConn()
+	if err != nil {
+		_ = udpConn.Close()
+		return nil, err
+	}
+	//gotta find out if we got an AF_INET6 socket or not:
+	out := &StdConn{
+		udpConn: udpConn,
+		rawConn: rawConn,
+		l:       l,
+		batch:   batch,
 	}
 
-	var sa unix.Sockaddr
-	if ip.Is4() {
-		sa4 := &unix.SockaddrInet4{Port: port}
-		sa4.Addr = ip.As4()
-		sa = sa4
-	} else {
-		sa6 := &unix.SockaddrInet6{Port: port}
-		sa6.Addr = ip.As16()
-		sa = sa6
-	}
-	if err = unix.Bind(fd, sa); err != nil {
-		return nil, fmt.Errorf("unable to bind to socket: %s", err)
+	af, err := out.getSockOptInt(unix.SO_DOMAIN)
+	if err != nil {
+		_ = out.Close()
+		return nil, err
 	}
+	out.isV4 = af == unix.AF_INET
 
-	return &StdConn{sysFd: fd, isV4: ip.Is4(), l: l, batch: batch}, err
+	out.prepareWriteMessages(MaxWriteBatch)
+	out.writeFunc = out.sendmmsgRawWrite
+
+	out.prepareGSO()
+
+	return out, nil
+}
+
+// maxGSOSegments caps the per-sendmsg GSO fan-out. Linux kernels have
+// historically capped UDP_MAX_SEGMENTS at 64; newer kernels raise it to 128
+// but we stay conservative so the same code works everywhere.
+const maxGSOSegments = 64
+
+// maxGSOBytes bounds the total payload per sendmsg() when UDP_SEGMENT is
+// set. The kernel stitches all iovecs into a single skb whose length the
+// UDP length field can represent, and also enforces sk_gso_max_size (which
+// on most devices is 65536). We use 65535 so ciphertext + headers always
+// fits, avoiding EMSGSIZE on large TSO superpackets.
+const maxGSOBytes = 65535
+
+// prepareGSO probes UDP_SEGMENT support and, on success, sets up the
+// reusable sendmsg scratch (iovecs, sockaddr, cmsg) plus the preallocated
+// raw-write closure used to avoid heap allocations on the hot path.
+func (u *StdConn) prepareGSO() {
+	var probeErr error
+	if err := u.rawConn.Control(func(fd uintptr) {
+		probeErr = unix.SetsockoptInt(int(fd), unix.IPPROTO_UDP, unix.UDP_SEGMENT, 0)
+	}); err != nil {
+		return
+	}
+	if probeErr != nil {
+		return
+	}
+	u.gsoSupported = true
+	u.gsoIovs = make([]iovec, maxGSOSegments)
+	u.gsoName = make([]byte, unix.SizeofSockaddrInet6)
+	u.gsoCmsg = make([]byte, unix.CmsgSpace(2))
+
+	// Wire up the static pieces of gsoMsg. Iovlen / Controllen / Namelen /
+	// cmsg contents get refreshed per call; Iov, Name, Control pointers are
+	// fixed because the scratch slices never move.
+	u.gsoMsg.Iov = &u.gsoIovs[0]
+	u.gsoMsg.Name = &u.gsoName[0]
+	u.gsoMsg.Control = &u.gsoCmsg[0]
+
+	// Prepopulate the cmsg header. Len/Level/Type are constant for our use;
+	// only the 2-byte gso_size payload changes per call.
+	cmsghdr := (*unix.Cmsghdr)(unsafe.Pointer(&u.gsoCmsg[0]))
+	cmsghdr.Level = unix.SOL_UDP
+	cmsghdr.Type = unix.UDP_SEGMENT
+	setCmsgLen(cmsghdr, unix.CmsgLen(2))
+
+	u.gsoFunc = u.sendmsgRawWriteGSO
 }
 
 func (u *StdConn) SupportsMultipleReaders() bool {
@@ -80,62 +159,134 @@ func (u *StdConn) Rebind() error {
 	return nil
 }
 
+func (u *StdConn) getSockOptInt(opt int) (int, error) {
+	if u.rawConn == nil {
+		return 0, fmt.Errorf("no UDP connection")
+	}
+	var out int
+	var opErr error
+	err := u.rawConn.Control(func(fd uintptr) {
+		out, opErr = unix.GetsockoptInt(int(fd), unix.SOL_SOCKET, opt)
+	})
+	if err != nil {
+		return 0, err
+	}
+	return out, opErr
+}
+
+func (u *StdConn) setSockOptInt(opt int, n int) error {
+	if u.rawConn == nil {
+		return fmt.Errorf("no UDP connection")
+	}
+	var opErr error
+	err := u.rawConn.Control(func(fd uintptr) {
+		opErr = unix.SetsockoptInt(int(fd), unix.SOL_SOCKET, opt, n)
+	})
+	if err != nil {
+		return err
+	}
+	return opErr
+}
+
 func (u *StdConn) SetRecvBuffer(n int) error {
-	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_RCVBUFFORCE, n)
+	return u.setSockOptInt(unix.SO_RCVBUFFORCE, n)
 }
 
 func (u *StdConn) SetSendBuffer(n int) error {
-	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_SNDBUFFORCE, n)
+	return u.setSockOptInt(unix.SO_SNDBUFFORCE, n)
 }
 
 func (u *StdConn) SetSoMark(mark int) error {
-	return unix.SetsockoptInt(u.sysFd, unix.SOL_SOCKET, unix.SO_MARK, mark)
+	return u.setSockOptInt(unix.SO_MARK, mark)
 }
 
 func (u *StdConn) GetRecvBuffer() (int, error) {
-	return unix.GetsockoptInt(int(u.sysFd), unix.SOL_SOCKET, unix.SO_RCVBUF)
+	return u.getSockOptInt(unix.SO_RCVBUF)
 }
 
 func (u *StdConn) GetSendBuffer() (int, error) {
-	return unix.GetsockoptInt(int(u.sysFd), unix.SOL_SOCKET, unix.SO_SNDBUF)
+	return u.getSockOptInt(unix.SO_SNDBUF)
 }
 
 func (u *StdConn) GetSoMark() (int, error) {
-	return unix.GetsockoptInt(int(u.sysFd), unix.SOL_SOCKET, unix.SO_MARK)
+	return u.getSockOptInt(unix.SO_MARK)
 }
 
 func (u *StdConn) LocalAddr() (netip.AddrPort, error) {
-	sa, err := unix.Getsockname(u.sysFd)
-	if err != nil {
-		return netip.AddrPort{}, err
-	}
+	a := u.udpConn.LocalAddr()
 
-	switch sa := sa.(type) {
-	case *unix.SockaddrInet4:
-		return netip.AddrPortFrom(netip.AddrFrom4(sa.Addr), uint16(sa.Port)), nil
-
-	case *unix.SockaddrInet6:
-		return netip.AddrPortFrom(netip.AddrFrom16(sa.Addr), uint16(sa.Port)), nil
+	switch v := a.(type) {
+	case *net.UDPAddr:
+		addr, ok := netip.AddrFromSlice(v.IP)
+		if !ok {
+			return netip.AddrPort{}, fmt.Errorf("LocalAddr returned invalid IP address: %s", v.IP)
+		}
+		return netip.AddrPortFrom(addr, uint16(v.Port)), nil
 
 	default:
-		return netip.AddrPort{}, fmt.Errorf("unsupported sock type: %T", sa)
+		return netip.AddrPort{}, fmt.Errorf("LocalAddr returned: %#v", a)
 	}
 }
 
-func (u *StdConn) ListenOut(r EncReader) {
+func recvmmsg(fd uintptr, msgs []rawMessage) (int, bool, error) {
+	var errno syscall.Errno
+	n, _, errno := unix.Syscall6(
+		unix.SYS_RECVMMSG,
+		fd,
+		uintptr(unsafe.Pointer(&msgs[0])),
+		uintptr(len(msgs)),
+		unix.MSG_WAITFORONE,
+		0,
+		0,
+	)
+	if errno == syscall.EAGAIN || errno == syscall.EWOULDBLOCK {
+		// No data available, block for I/O and try again.
+		return int(n), false, nil
+	}
+	if errno != 0 {
+		return int(n), true, &net.OpError{Op: "recvmmsg", Err: errno}
+	}
+	return int(n), true, nil
+}
+
+func (u *StdConn) listenOutSingle(r EncReader, flush func()) error {
+	var err error
+	var n int
+	var from netip.AddrPort
+	buffer := make([]byte, MTU)
+
+	for {
+		n, from, err = u.udpConn.ReadFromUDPAddrPort(buffer)
+		if err != nil {
+			return err
+		}
+		from = netip.AddrPortFrom(from.Addr().Unmap(), from.Port())
+		r(from, buffer[:n])
+		flush()
+	}
+}
+
+func (u *StdConn) listenOutBatch(r EncReader, flush func()) error {
 	var ip netip.Addr
+	var n int
+	var operr error
 
 	msgs, buffers, names := u.PrepareRawMessages(u.batch)
-	read := u.ReadMulti
-	if u.batch == 1 {
-		read = u.ReadSingle
+
+	//reader needs to capture variables from this function, since it's used as a lambda with rawConn.Read
+	//defining it outside the loop so it gets re-used
+	reader := func(fd uintptr) (done bool) {
+		n, done, operr = recvmmsg(fd, msgs)
+		return done
 	}
 
 	for {
-		n, err := read(msgs)
+		err := u.rawConn.Read(reader)
 		if err != nil {
-			u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
-			return
+			return err
+		}
+		if operr != nil {
+			return operr
 		}
 
 		for i := 0; i < n; i++ {
@@ -147,109 +298,254 @@ func (u *StdConn) ListenOut(r EncReader) {
 			}
 			r(netip.AddrPortFrom(ip.Unmap(), binary.BigEndian.Uint16(names[i][2:4])), buffers[i][:msgs[i].Len])
 		}
+		// End-of-batch: let callers (e.g. TUN write coalescer) flush any
+		// state they accumulated across this batch.
+		flush()
 	}
 }
 
-func (u *StdConn) ReadSingle(msgs []rawMessage) (int, error) {
-	for {
-		n, _, err := unix.Syscall6(
-			unix.SYS_RECVMSG,
-			uintptr(u.sysFd),
-			uintptr(unsafe.Pointer(&(msgs[0].Hdr))),
-			0,
-			0,
-			0,
-			0,
-		)
-
-		if err != 0 {
-			return 0, &net.OpError{Op: "recvmsg", Err: err}
-		}
-
-		msgs[0].Len = uint32(n)
-		return 1, nil
-	}
-}
-
-func (u *StdConn) ReadMulti(msgs []rawMessage) (int, error) {
-	for {
-		n, _, err := unix.Syscall6(
-			unix.SYS_RECVMMSG,
-			uintptr(u.sysFd),
-			uintptr(unsafe.Pointer(&msgs[0])),
-			uintptr(len(msgs)),
-			unix.MSG_WAITFORONE,
-			0,
-			0,
-		)
-
-		if err != 0 {
-			return 0, &net.OpError{Op: "recvmmsg", Err: err}
-		}
-
-		return int(n), nil
+func (u *StdConn) ListenOut(r EncReader, flush func()) error {
+	if u.batch == 1 {
+		return u.listenOutSingle(r, flush)
+	} else {
+		return u.listenOutBatch(r, flush)
 	}
 }
 
 func (u *StdConn) WriteTo(b []byte, ip netip.AddrPort) error {
-	if u.isV4 {
-		return u.writeTo4(b, ip)
-	}
-	return u.writeTo6(b, ip)
+	_, err := u.udpConn.WriteToUDPAddrPort(b, ip)
+	return err
 }
 
-func (u *StdConn) writeTo6(b []byte, ip netip.AddrPort) error {
-	var rsa unix.RawSockaddrInet6
-	rsa.Family = unix.AF_INET6
-	rsa.Addr = ip.Addr().As16()
-	binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa.Port))[:], ip.Port())
-
-	for {
-		_, _, err := unix.Syscall6(
-			unix.SYS_SENDTO,
-			uintptr(u.sysFd),
-			uintptr(unsafe.Pointer(&b[0])),
-			uintptr(len(b)),
-			uintptr(0),
-			uintptr(unsafe.Pointer(&rsa)),
-			uintptr(unix.SizeofSockaddrInet6),
-		)
-
-		if err != 0 {
-			return &net.OpError{Op: "sendto", Err: err}
+// WriteBatch sends bufs via sendmmsg(2) using the preallocated scratch on
+// StdConn. Chunks larger than the scratch are processed in multiple syscalls.
+// If sendmmsg returns a fatal error mid-chunk we fall back to single WriteTo
+// calls for the remainder so the caller still gets best-effort delivery.
+func (u *StdConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	if len(bufs) != len(addrs) {
+		return fmt.Errorf("WriteBatch: len(bufs)=%d != len(addrs)=%d", len(bufs), len(addrs))
+	}
+	//u.l.WithField("bufs", len(bufs)).Info("WriteBatch")
+	i := 0
+	for i < len(bufs) {
+		chunk := len(bufs) - i
+		if chunk > len(u.writeMsgs) {
+			chunk = len(u.writeMsgs)
 		}
 
-		return nil
+		for k := 0; k < chunk; k++ {
+			b := bufs[i+k]
+			if len(b) == 0 {
+				// sendmmsg with an empty iovec is legal but pointless; fall
+				// through after filling the slot so Base is still valid.
+				u.writeIovs[k].Base = nil
+				setIovLen(&u.writeIovs[k], 0)
+			} else {
+				u.writeIovs[k].Base = &b[0]
+				setIovLen(&u.writeIovs[k], len(b))
+			}
+			nlen, err := writeSockaddr(u.writeNames[k], addrs[i+k], u.isV4)
+			if err != nil {
+				return err
+			}
+			u.writeMsgs[k].Hdr.Namelen = uint32(nlen)
+		}
+
+		sent, serr := u.sendmmsg(chunk)
+		if serr != nil {
+			if sent <= 0 {
+				// nothing went out; fall back to WriteTo for this chunk.
+				for k := 0; k < chunk; k++ {
+					if err := u.WriteTo(bufs[i+k], addrs[i+k]); err != nil {
+						return err
+					}
+				}
+				i += chunk
+				continue
+			}
+			// partial: treat as success for the sent packets and retry the
+			// remainder on the next outer-loop iteration.
+		}
+		if sent == 0 {
+			return fmt.Errorf("sendmmsg made no progress")
+		}
+		i += sent
 	}
+	return nil
 }
 
-func (u *StdConn) writeTo4(b []byte, ip netip.AddrPort) error {
-	if !ip.Addr().Is4() {
-		return ErrInvalidIPv6RemoteForSocket
+// sendmmsgRawWrite is the preallocated callback passed to rawConn.Write. It
+// reads its input (u.writeChunk) and writes its outputs (u.writeSent,
+// u.writeErrno) through StdConn fields so the closure itself does not
+// capture per-call locals and therefore does not heap-allocate.
+func (u *StdConn) sendmmsgRawWrite(fd uintptr) bool {
+	r1, _, errno := unix.Syscall6(
+		unix.SYS_SENDMMSG,
+		fd,
+		uintptr(unsafe.Pointer(&u.writeMsgs[0])),
+		uintptr(u.writeChunk),
+		0,
+		0,
+		0,
+	)
+	if errno == syscall.EAGAIN || errno == syscall.EWOULDBLOCK {
+		return false
 	}
+	u.writeSent = int(r1)
+	u.writeErrno = errno
+	return true
+}
 
-	var rsa unix.RawSockaddrInet4
-	rsa.Family = unix.AF_INET
-	rsa.Addr = ip.Addr().As4()
-	binary.BigEndian.PutUint16((*[2]byte)(unsafe.Pointer(&rsa.Port))[:], ip.Port())
-
-	for {
-		_, _, err := unix.Syscall6(
-			unix.SYS_SENDTO,
-			uintptr(u.sysFd),
-			uintptr(unsafe.Pointer(&b[0])),
-			uintptr(len(b)),
-			uintptr(0),
-			uintptr(unsafe.Pointer(&rsa)),
-			uintptr(unix.SizeofSockaddrInet4),
-		)
-
-		if err != 0 {
-			return &net.OpError{Op: "sendto", Err: err}
-		}
+func (u *StdConn) SupportsGSO() bool {
+	return u.gsoSupported
+}
 
+// WriteSegmented sends bufs to addr as a UDP GSO superpacket. The kernel
+// emits one datagram per iovec on the wire; all iovecs except the last must
+// be exactly segSize bytes. Non-GSO kernels hit the WriteTo fallback.
+// Called with len(bufs) >= 1. len(bufs) > maxGSOSegments is chunked.
+func (u *StdConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, segSize int) error {
+	if len(bufs) == 0 {
 		return nil
 	}
+	if !u.gsoSupported {
+		for _, b := range bufs {
+			if err := u.WriteTo(b, addr); err != nil {
+				return err
+			}
+		}
+		return nil
+	}
+
+	nlen, err := writeSockaddr(u.gsoName, addr, u.isV4)
+	if err != nil {
+		return err
+	}
+	u.gsoMsg.Namelen = uint32(nlen)
+	setMsgControllen(&u.gsoMsg, unix.CmsgSpace(2))
+
+	// Cap the per-syscall fan-out by both segment count and total bytes.
+	// Kernel rejects sendmsg with EMSGSIZE when segCount*segSize would
+	// exceed sk_gso_max_size (typically 65536). For segSize > maxGSOBytes
+	// we can't use GSO at all and must fall back per-packet.
+	segsByBytes := maxGSOBytes / segSize
+	if segsByBytes == 0 {
+		for _, b := range bufs {
+			if werr := u.WriteTo(b, addr); werr != nil {
+				return werr
+			}
+		}
+		return nil
+	}
+	maxChunk := maxGSOSegments
+	if segsByBytes < maxChunk {
+		maxChunk = segsByBytes
+	}
+
+	i := 0
+	for i < len(bufs) {
+		chunk := len(bufs) - i
+		if chunk > maxChunk {
+			chunk = maxChunk
+		}
+		for k := 0; k < chunk; k++ {
+			b := bufs[i+k]
+			if len(b) == 0 {
+				u.gsoIovs[k].Base = nil
+				setIovLen(&u.gsoIovs[k], 0)
+			} else {
+				u.gsoIovs[k].Base = &b[0]
+				setIovLen(&u.gsoIovs[k], len(b))
+			}
+		}
+		setMsgIovlen(&u.gsoMsg, chunk)
+		binary.NativeEndian.PutUint16(u.gsoCmsg[unix.CmsgLen(0):unix.CmsgLen(0)+2], uint16(segSize))
+
+		if serr := u.sendmsgGSO(); serr != nil {
+			// Fall back to a per-packet loop for the remainder of the
+			// batch. Dropping the GSO call entirely is safer than
+			// returning mid-superpacket and losing bytes.
+			for k := 0; k < chunk; k++ {
+				if werr := u.WriteTo(bufs[i+k], addr); werr != nil {
+					return werr
+				}
+			}
+		}
+		i += chunk
+	}
+	return nil
+}
+
+// sendmsgRawWriteGSO is the preallocated rawConn.Write callback for the GSO
+// path. Reads the prebuilt u.gsoMsg and writes u.gsoSent / u.gsoErrno.
+func (u *StdConn) sendmsgRawWriteGSO(fd uintptr) bool {
+	r1, _, errno := unix.Syscall(
+		unix.SYS_SENDMSG,
+		fd,
+		uintptr(unsafe.Pointer(&u.gsoMsg)),
+		0,
+	)
+	if errno == syscall.EAGAIN || errno == syscall.EWOULDBLOCK {
+		return false
+	}
+	u.gsoSent = int(r1)
+	u.gsoErrno = errno
+	return true
+}
+
+func (u *StdConn) sendmsgGSO() error {
+	u.gsoSent = 0
+	u.gsoErrno = 0
+	if err := u.rawConn.Write(u.gsoFunc); err != nil {
+		return err
+	}
+	if u.gsoErrno != 0 {
+		return &net.OpError{Op: "sendmsg", Err: u.gsoErrno}
+	}
+	return nil
+}
+
+func (u *StdConn) sendmmsg(n int) (int, error) {
+	u.writeChunk = n
+	u.writeSent = 0
+	u.writeErrno = 0
+	if err := u.rawConn.Write(u.writeFunc); err != nil {
+		return u.writeSent, err
+	}
+	if u.writeErrno != 0 {
+		return u.writeSent, &net.OpError{Op: "sendmmsg", Err: u.writeErrno}
+	}
+	return u.writeSent, nil
+}
+
+// writeSockaddr encodes addr into buf (which must be at least
+// SizeofSockaddrInet6 bytes). Returns the number of bytes used. If isV4 is
+// true and addr is not a v4 (or v4-in-v6) address, returns an error.
+func writeSockaddr(buf []byte, addr netip.AddrPort, isV4 bool) (int, error) {
+	ap := addr.Addr().Unmap()
+	if isV4 {
+		if !ap.Is4() {
+			return 0, ErrInvalidIPv6RemoteForSocket
+		}
+		// struct sockaddr_in: { sa_family_t(2), in_port_t(2, BE), in_addr(4), zero(8) }
+		// sa_family is host endian.
+		binary.NativeEndian.PutUint16(buf[0:2], unix.AF_INET)
+		binary.BigEndian.PutUint16(buf[2:4], addr.Port())
+		ip4 := ap.As4()
+		copy(buf[4:8], ip4[:])
+		for j := 8; j < 16; j++ {
+			buf[j] = 0
+		}
+		return unix.SizeofSockaddrInet4, nil
+	}
+	// struct sockaddr_in6: { sa_family_t(2), in_port_t(2, BE), flowinfo(4), in6_addr(16), scope_id(4) }
+	binary.NativeEndian.PutUint16(buf[0:2], unix.AF_INET6)
+	binary.BigEndian.PutUint16(buf[2:4], addr.Port())
+	binary.NativeEndian.PutUint32(buf[4:8], 0)
+	ip6 := addr.Addr().As16()
+	copy(buf[8:24], ip6[:])
+	binary.NativeEndian.PutUint32(buf[24:28], 0)
+	return unix.SizeofSockaddrInet6, nil
 }
 
 func (u *StdConn) ReloadConfig(c *config.C) {
@@ -302,15 +598,28 @@ func (u *StdConn) ReloadConfig(c *config.C) {
 
 func (u *StdConn) getMemInfo(meminfo *[unix.SK_MEMINFO_VARS]uint32) error {
 	var vallen uint32 = 4 * unix.SK_MEMINFO_VARS
-	_, _, err := unix.Syscall6(unix.SYS_GETSOCKOPT, uintptr(u.sysFd), uintptr(unix.SOL_SOCKET), uintptr(unix.SO_MEMINFO), uintptr(unsafe.Pointer(meminfo)), uintptr(unsafe.Pointer(&vallen)), 0)
-	if err != 0 {
+
+	if u.rawConn == nil {
+		return fmt.Errorf("no UDP connection")
+	}
+	var opErr error
+	err := u.rawConn.Control(func(fd uintptr) {
+		_, _, syserr := unix.Syscall6(unix.SYS_GETSOCKOPT, fd, uintptr(unix.SOL_SOCKET), uintptr(unix.SO_MEMINFO), uintptr(unsafe.Pointer(meminfo)), uintptr(unsafe.Pointer(&vallen)), 0)
+		if syserr != 0 {
+			opErr = syserr
+		}
+	})
+	if err != nil {
 		return err
 	}
-	return nil
+	return opErr
 }
 
 func (u *StdConn) Close() error {
-	return syscall.Close(u.sysFd)
+	if u.udpConn != nil {
+		return u.udpConn.Close()
+	}
+	return nil
 }
 
 func NewUDPStatsEmitter(udpConns []Conn) func() {

udp/udp_linux_32.go

@@ -52,3 +52,35 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 	return msgs, buffers, names
 }
+
+// prepareWriteMessages allocates one Mmsghdr/iovec/sockaddr scratch per slot,
+// wired up so each writeMsgs[i] already points at writeIovs[i] and
+// writeNames[i]. Callers fill in the iovec Base/Len, the sockaddr bytes, and
+// Namelen before each sendmmsg.
+func (u *StdConn) prepareWriteMessages(n int) {
+	u.writeMsgs = make([]rawMessage, n)
+	u.writeIovs = make([]iovec, n)
+	u.writeNames = make([][]byte, n)
+	for i := range u.writeMsgs {
+		u.writeNames[i] = make([]byte, unix.SizeofSockaddrInet6)
+		u.writeMsgs[i].Hdr.Iov = &u.writeIovs[i]
+		u.writeMsgs[i].Hdr.Iovlen = 1
+		u.writeMsgs[i].Hdr.Name = &u.writeNames[i][0]
+	}
+}
+
+func setIovLen(v *iovec, n int) {
+	v.Len = uint32(n)
+}
+
+func setMsgIovlen(m *msghdr, n int) {
+	m.Iovlen = uint32(n)
+}
+
+func setMsgControllen(m *msghdr, n int) {
+	m.Controllen = uint32(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint32(n)
+}

udp/udp_linux_64.go

@@ -55,3 +55,35 @@ func (u *StdConn) PrepareRawMessages(n int) ([]rawMessage, [][]byte, [][]byte) {
 
 	return msgs, buffers, names
 }
+
+// prepareWriteMessages allocates one Mmsghdr/iovec/sockaddr scratch per slot,
+// wired up so each writeMsgs[i] already points at writeIovs[i] and
+// writeNames[i]. Callers fill in the iovec Base/Len, the sockaddr bytes, and
+// Namelen before each sendmmsg.
+func (u *StdConn) prepareWriteMessages(n int) {
+	u.writeMsgs = make([]rawMessage, n)
+	u.writeIovs = make([]iovec, n)
+	u.writeNames = make([][]byte, n)
+	for i := range u.writeMsgs {
+		u.writeNames[i] = make([]byte, unix.SizeofSockaddrInet6)
+		u.writeMsgs[i].Hdr.Iov = &u.writeIovs[i]
+		u.writeMsgs[i].Hdr.Iovlen = 1
+		u.writeMsgs[i].Hdr.Name = &u.writeNames[i][0]
+	}
+}
+
+func setIovLen(v *iovec, n int) {
+	v.Len = uint64(n)
+}
+
+func setMsgIovlen(m *msghdr, n int) {
+	m.Iovlen = uint64(n)
+}
+
+func setMsgControllen(m *msghdr, n int) {
+	m.Controllen = uint64(n)
+}
+
+func setCmsgLen(h *unix.Cmsghdr, n int) {
+	h.Len = uint64(n)
+}

udp/udp_rio_windows.go

@@ -140,7 +140,7 @@ func (u *RIOConn) bind(l *logrus.Logger, sa windows.Sockaddr) error {
 	return nil
 }
 
-func (u *RIOConn) ListenOut(r EncReader) {
+func (u *RIOConn) ListenOut(r EncReader, flush func()) error {
 	buffer := make([]byte, MTU)
 
 	var lastRecvErr time.Time
@@ -151,8 +151,7 @@ func (u *RIOConn) ListenOut(r EncReader) {
 
 		if err != nil {
 			if errors.Is(err, net.ErrClosed) {
-				u.l.WithError(err).Debug("udp socket is closed, exiting read loop")
-				return
+				return err
 			}
 			// Dampen unexpected message warns to once per minute
 			if lastRecvErr.IsZero() || time.Since(lastRecvErr) > time.Minute {
@@ -163,6 +162,7 @@ func (u *RIOConn) ListenOut(r EncReader) {
 		}
 
 		r(netip.AddrPortFrom(netip.AddrFrom16(rua.Addr).Unmap(), (rua.Port>>8)|((rua.Port&0xff)<<8)), buffer[:n])
+		flush()
 	}
 }
 
@@ -317,6 +317,26 @@ func (u *RIOConn) WriteTo(buf []byte, ip netip.AddrPort) error {
 	return winrio.SendEx(u.rq, dataBuffer, 1, nil, addressBuffer, nil, nil, 0, 0)
 }
 
+func (u *RIOConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *RIOConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, _ int) error {
+	for _, b := range bufs {
+		if err := u.WriteTo(b, addr); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *RIOConn) SupportsGSO() bool { return false }
+
 func (u *RIOConn) LocalAddr() (netip.AddrPort, error) {
 	sa, err := windows.Getsockname(u.sock)
 	if err != nil {

udp/udp_tester.go

@@ -6,6 +6,7 @@ package udp
 import (
 	"io"
 	"net/netip"
+	"os"
 	"sync/atomic"
 
 	"github.com/sirupsen/logrus"
@@ -106,13 +107,34 @@ func (u *TesterConn) WriteTo(b []byte, addr netip.AddrPort) error {
 	return nil
 }
 
-func (u *TesterConn) ListenOut(r EncReader) {
+func (u *TesterConn) WriteBatch(bufs [][]byte, addrs []netip.AddrPort) error {
+	for i, b := range bufs {
+		if err := u.WriteTo(b, addrs[i]); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *TesterConn) WriteSegmented(bufs [][]byte, addr netip.AddrPort, _ int) error {
+	for _, b := range bufs {
+		if err := u.WriteTo(b, addr); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (u *TesterConn) SupportsGSO() bool { return false }
+
+func (u *TesterConn) ListenOut(r EncReader, flush func()) error {
 	for {
 		p, ok := <-u.RxPackets
 		if !ok {
-			return
+			return os.ErrClosed
 		}
 		r(p.From, p.Data)
+		flush()
 	}
 }